2016-04-04 21:00:37 +08:00
|
|
|
/* Client connection-specific management code.
|
|
|
|
*
|
|
|
|
* Copyright (C) 2016 Red Hat, Inc. All Rights Reserved.
|
|
|
|
* Written by David Howells (dhowells@redhat.com)
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public Licence
|
|
|
|
* as published by the Free Software Foundation; either version
|
|
|
|
* 2 of the Licence, or (at your option) any later version.
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
*
|
|
|
|
*
|
|
|
|
* Client connections need to be cached for a little while after they've made a
|
|
|
|
* call so as to handle retransmitted DATA packets in case the server didn't
|
|
|
|
* receive the final ACK or terminating ABORT we sent it.
|
|
|
|
*
|
|
|
|
* Client connections can be in one of a number of cache states:
|
|
|
|
*
|
|
|
|
* (1) INACTIVE - The connection is not held in any list and may not have been
|
|
|
|
* exposed to the world. If it has been previously exposed, it was
|
|
|
|
* discarded from the idle list after expiring.
|
|
|
|
*
|
|
|
|
* (2) WAITING - The connection is waiting for the number of client conns to
|
|
|
|
* drop below the maximum capacity. Calls may be in progress upon it from
|
|
|
|
* when it was active and got culled.
|
|
|
|
*
|
|
|
|
* The connection is on the rxrpc_waiting_client_conns list which is kept
|
|
|
|
* in to-be-granted order. Culled conns with waiters go to the back of
|
|
|
|
* the queue just like new conns.
|
|
|
|
*
|
|
|
|
* (3) ACTIVE - The connection has at least one call in progress upon it, it
|
|
|
|
* may freely grant available channels to new calls and calls may be
|
|
|
|
* waiting on it for channels to become available.
|
|
|
|
*
|
|
|
|
* The connection is on the rxrpc_active_client_conns list which is kept
|
|
|
|
* in activation order for culling purposes.
|
|
|
|
*
|
|
|
|
* rxrpc_nr_active_client_conns is held incremented also.
|
|
|
|
*
|
|
|
|
* (4) CULLED - The connection got summarily culled to try and free up
|
|
|
|
* capacity. Calls currently in progress on the connection are allowed to
|
|
|
|
* continue, but new calls will have to wait. There can be no waiters in
|
|
|
|
* this state - the conn would have to go to the WAITING state instead.
|
|
|
|
*
|
|
|
|
* (5) IDLE - The connection has no calls in progress upon it and must have
|
|
|
|
* been exposed to the world (ie. the EXPOSED flag must be set). When it
|
|
|
|
* expires, the EXPOSED flag is cleared and the connection transitions to
|
|
|
|
* the INACTIVE state.
|
|
|
|
*
|
|
|
|
* The connection is on the rxrpc_idle_client_conns list which is kept in
|
|
|
|
* order of how soon they'll expire.
|
|
|
|
*
|
|
|
|
* There are flags of relevance to the cache:
|
|
|
|
*
|
|
|
|
* (1) EXPOSED - The connection ID got exposed to the world. If this flag is
|
|
|
|
* set, an extra ref is added to the connection preventing it from being
|
|
|
|
* reaped when it has no calls outstanding. This flag is cleared and the
|
|
|
|
* ref dropped when a conn is discarded from the idle list.
|
|
|
|
*
|
|
|
|
* This allows us to move terminal call state retransmission to the
|
|
|
|
* connection and to discard the call immediately we think it is done
|
|
|
|
* with. It also give us a chance to reuse the connection.
|
|
|
|
*
|
|
|
|
* (2) DONT_REUSE - The connection should be discarded as soon as possible and
|
|
|
|
* should not be reused. This is set when an exclusive connection is used
|
|
|
|
* or a call ID counter overflows.
|
|
|
|
*
|
|
|
|
* The caching state may only be changed if the cache lock is held.
|
|
|
|
*
|
|
|
|
* There are two idle client connection expiry durations. If the total number
|
|
|
|
* of connections is below the reap threshold, we use the normal duration; if
|
|
|
|
* it's above, we use the fast duration.
|
2016-04-04 21:00:37 +08:00
|
|
|
*/
|
|
|
|
|
|
|
|
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
|
|
|
|
|
|
|
|
#include <linux/slab.h>
|
|
|
|
#include <linux/idr.h>
|
|
|
|
#include <linux/timer.h>
|
|
|
|
#include "ar-internal.h"
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
__read_mostly unsigned int rxrpc_max_client_connections = 1000;
|
|
|
|
__read_mostly unsigned int rxrpc_reap_client_connections = 900;
|
|
|
|
__read_mostly unsigned int rxrpc_conn_idle_client_expiry = 2 * 60 * HZ;
|
|
|
|
__read_mostly unsigned int rxrpc_conn_idle_client_fast_expiry = 2 * HZ;
|
|
|
|
|
|
|
|
static unsigned int rxrpc_nr_client_conns;
|
|
|
|
static unsigned int rxrpc_nr_active_client_conns;
|
|
|
|
static __read_mostly bool rxrpc_kill_all_client_conns;
|
|
|
|
|
|
|
|
static DEFINE_SPINLOCK(rxrpc_client_conn_cache_lock);
|
|
|
|
static DEFINE_SPINLOCK(rxrpc_client_conn_discard_mutex);
|
|
|
|
static LIST_HEAD(rxrpc_waiting_client_conns);
|
|
|
|
static LIST_HEAD(rxrpc_active_client_conns);
|
|
|
|
static LIST_HEAD(rxrpc_idle_client_conns);
|
|
|
|
|
2016-04-04 21:00:37 +08:00
|
|
|
/*
|
|
|
|
* We use machine-unique IDs for our client connections.
|
|
|
|
*/
|
|
|
|
DEFINE_IDR(rxrpc_client_conn_ids);
|
|
|
|
static DEFINE_SPINLOCK(rxrpc_conn_id_lock);
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
static void rxrpc_cull_active_client_conns(void);
|
|
|
|
static void rxrpc_discard_expired_client_conns(struct work_struct *);
|
|
|
|
|
|
|
|
static DECLARE_DELAYED_WORK(rxrpc_client_conn_reap,
|
|
|
|
rxrpc_discard_expired_client_conns);
|
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
const char rxrpc_conn_cache_states[RXRPC_CONN__NR_CACHE_STATES][5] = {
|
|
|
|
[RXRPC_CONN_CLIENT_INACTIVE] = "Inac",
|
|
|
|
[RXRPC_CONN_CLIENT_WAITING] = "Wait",
|
|
|
|
[RXRPC_CONN_CLIENT_ACTIVE] = "Actv",
|
|
|
|
[RXRPC_CONN_CLIENT_CULLED] = "Cull",
|
|
|
|
[RXRPC_CONN_CLIENT_IDLE] = "Idle",
|
|
|
|
};
|
|
|
|
|
2016-04-04 21:00:37 +08:00
|
|
|
/*
|
|
|
|
* Get a connection ID and epoch for a client connection from the global pool.
|
|
|
|
* The connection struct pointer is then recorded in the idr radix tree. The
|
2016-09-04 20:14:46 +08:00
|
|
|
* epoch doesn't change until the client is rebooted (or, at least, unless the
|
|
|
|
* module is unloaded).
|
2016-04-04 21:00:37 +08:00
|
|
|
*/
|
2016-04-04 21:00:40 +08:00
|
|
|
static int rxrpc_get_client_connection_id(struct rxrpc_connection *conn,
|
|
|
|
gfp_t gfp)
|
2016-04-04 21:00:37 +08:00
|
|
|
{
|
|
|
|
int id;
|
|
|
|
|
|
|
|
_enter("");
|
|
|
|
|
|
|
|
idr_preload(gfp);
|
|
|
|
spin_lock(&rxrpc_conn_id_lock);
|
|
|
|
|
2016-09-04 20:14:46 +08:00
|
|
|
id = idr_alloc_cyclic(&rxrpc_client_conn_ids, conn,
|
|
|
|
1, 0x40000000, GFP_NOWAIT);
|
|
|
|
if (id < 0)
|
|
|
|
goto error;
|
2016-04-04 21:00:37 +08:00
|
|
|
|
|
|
|
spin_unlock(&rxrpc_conn_id_lock);
|
|
|
|
idr_preload_end();
|
|
|
|
|
2016-09-04 20:14:46 +08:00
|
|
|
conn->proto.epoch = rxrpc_epoch;
|
2016-04-04 21:00:37 +08:00
|
|
|
conn->proto.cid = id << RXRPC_CIDSHIFT;
|
|
|
|
set_bit(RXRPC_CONN_HAS_IDR, &conn->flags);
|
2016-09-04 20:14:46 +08:00
|
|
|
_leave(" [CID %x]", conn->proto.cid);
|
2016-04-04 21:00:37 +08:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
error:
|
|
|
|
spin_unlock(&rxrpc_conn_id_lock);
|
|
|
|
idr_preload_end();
|
|
|
|
_leave(" = %d", id);
|
|
|
|
return id;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Release a connection ID for a client connection from the global pool.
|
|
|
|
*/
|
2016-06-30 17:45:22 +08:00
|
|
|
static void rxrpc_put_client_connection_id(struct rxrpc_connection *conn)
|
2016-04-04 21:00:37 +08:00
|
|
|
{
|
|
|
|
if (test_bit(RXRPC_CONN_HAS_IDR, &conn->flags)) {
|
|
|
|
spin_lock(&rxrpc_conn_id_lock);
|
|
|
|
idr_remove(&rxrpc_client_conn_ids,
|
|
|
|
conn->proto.cid >> RXRPC_CIDSHIFT);
|
|
|
|
spin_unlock(&rxrpc_conn_id_lock);
|
|
|
|
}
|
|
|
|
}
|
2016-06-27 17:32:02 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Destroy the client connection ID tree.
|
|
|
|
*/
|
|
|
|
void rxrpc_destroy_client_conn_ids(void)
|
|
|
|
{
|
|
|
|
struct rxrpc_connection *conn;
|
|
|
|
int id;
|
|
|
|
|
|
|
|
if (!idr_is_empty(&rxrpc_client_conn_ids)) {
|
|
|
|
idr_for_each_entry(&rxrpc_client_conn_ids, conn, id) {
|
|
|
|
pr_err("AF_RXRPC: Leaked client conn %p {%d}\n",
|
|
|
|
conn, atomic_read(&conn->usage));
|
|
|
|
}
|
|
|
|
BUG();
|
|
|
|
}
|
|
|
|
|
|
|
|
idr_destroy(&rxrpc_client_conn_ids);
|
|
|
|
}
|
2016-04-04 21:00:40 +08:00
|
|
|
|
|
|
|
/*
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
* Allocate a client connection.
|
2016-04-04 21:00:40 +08:00
|
|
|
*/
|
|
|
|
static struct rxrpc_connection *
|
|
|
|
rxrpc_alloc_client_connection(struct rxrpc_conn_parameters *cp, gfp_t gfp)
|
|
|
|
{
|
|
|
|
struct rxrpc_connection *conn;
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
_enter("");
|
|
|
|
|
|
|
|
conn = rxrpc_alloc_connection(gfp);
|
|
|
|
if (!conn) {
|
|
|
|
_leave(" = -ENOMEM");
|
|
|
|
return ERR_PTR(-ENOMEM);
|
|
|
|
}
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
atomic_set(&conn->usage, 1);
|
2016-09-30 05:37:15 +08:00
|
|
|
if (cp->exclusive)
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
__set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
|
|
|
|
|
2016-04-04 21:00:40 +08:00
|
|
|
conn->params = *cp;
|
|
|
|
conn->out_clientflag = RXRPC_CLIENT_INITIATED;
|
|
|
|
conn->state = RXRPC_CONN_CLIENT;
|
|
|
|
|
|
|
|
ret = rxrpc_get_client_connection_id(conn, gfp);
|
|
|
|
if (ret < 0)
|
|
|
|
goto error_0;
|
|
|
|
|
|
|
|
ret = rxrpc_init_client_conn_security(conn);
|
|
|
|
if (ret < 0)
|
|
|
|
goto error_1;
|
|
|
|
|
|
|
|
ret = conn->security->prime_packet_security(conn);
|
|
|
|
if (ret < 0)
|
|
|
|
goto error_2;
|
|
|
|
|
|
|
|
write_lock(&rxrpc_connection_lock);
|
2016-08-24 14:30:52 +08:00
|
|
|
list_add_tail(&conn->proc_link, &rxrpc_connection_proc_list);
|
2016-04-04 21:00:40 +08:00
|
|
|
write_unlock(&rxrpc_connection_lock);
|
|
|
|
|
|
|
|
/* We steal the caller's peer ref. */
|
|
|
|
cp->peer = NULL;
|
|
|
|
rxrpc_get_local(conn->params.local);
|
|
|
|
key_get(conn->params.key);
|
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_conn(conn, rxrpc_conn_new_client, atomic_read(&conn->usage),
|
|
|
|
__builtin_return_address(0));
|
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_alloc);
|
2016-04-04 21:00:40 +08:00
|
|
|
_leave(" = %p", conn);
|
|
|
|
return conn;
|
|
|
|
|
|
|
|
error_2:
|
|
|
|
conn->security->clear(conn);
|
|
|
|
error_1:
|
|
|
|
rxrpc_put_client_connection_id(conn);
|
|
|
|
error_0:
|
|
|
|
kfree(conn);
|
|
|
|
_leave(" = %d", ret);
|
|
|
|
return ERR_PTR(ret);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
* Determine if a connection may be reused.
|
2016-04-04 21:00:40 +08:00
|
|
|
*/
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
static bool rxrpc_may_reuse_conn(struct rxrpc_connection *conn)
|
|
|
|
{
|
|
|
|
int id_cursor, id, distance, limit;
|
|
|
|
|
|
|
|
if (test_bit(RXRPC_CONN_DONT_REUSE, &conn->flags))
|
|
|
|
goto dont_reuse;
|
|
|
|
|
|
|
|
if (conn->proto.epoch != rxrpc_epoch)
|
|
|
|
goto mark_dont_reuse;
|
|
|
|
|
|
|
|
/* The IDR tree gets very expensive on memory if the connection IDs are
|
|
|
|
* widely scattered throughout the number space, so we shall want to
|
|
|
|
* kill off connections that, say, have an ID more than about four
|
|
|
|
* times the maximum number of client conns away from the current
|
|
|
|
* allocation point to try and keep the IDs concentrated.
|
|
|
|
*/
|
|
|
|
id_cursor = READ_ONCE(rxrpc_client_conn_ids.cur);
|
|
|
|
id = conn->proto.cid >> RXRPC_CIDSHIFT;
|
|
|
|
distance = id - id_cursor;
|
|
|
|
if (distance < 0)
|
|
|
|
distance = -distance;
|
|
|
|
limit = round_up(rxrpc_max_client_connections, IDR_SIZE) * 4;
|
|
|
|
if (distance > limit)
|
|
|
|
goto mark_dont_reuse;
|
|
|
|
|
|
|
|
return true;
|
|
|
|
|
|
|
|
mark_dont_reuse:
|
|
|
|
set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
|
|
|
|
dont_reuse:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Create or find a client connection to use for a call.
|
|
|
|
*
|
|
|
|
* If we return with a connection, the call will be on its waiting list. It's
|
|
|
|
* left to the caller to assign a channel and wake up the call.
|
|
|
|
*/
|
|
|
|
static int rxrpc_get_client_conn(struct rxrpc_call *call,
|
|
|
|
struct rxrpc_conn_parameters *cp,
|
|
|
|
struct sockaddr_rxrpc *srx,
|
|
|
|
gfp_t gfp)
|
2016-04-04 21:00:40 +08:00
|
|
|
{
|
|
|
|
struct rxrpc_connection *conn, *candidate = NULL;
|
|
|
|
struct rxrpc_local *local = cp->local;
|
|
|
|
struct rb_node *p, **pp, *parent;
|
|
|
|
long diff;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
int ret = -ENOMEM;
|
2016-04-04 21:00:40 +08:00
|
|
|
|
|
|
|
_enter("{%d,%lx},", call->debug_id, call->user_call_ID);
|
|
|
|
|
|
|
|
cp->peer = rxrpc_lookup_peer(cp->local, srx, gfp);
|
|
|
|
if (!cp->peer)
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
goto error;
|
2016-04-04 21:00:40 +08:00
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/* If the connection is not meant to be exclusive, search the available
|
|
|
|
* connections to see if the connection we want to use already exists.
|
|
|
|
*/
|
2016-04-04 21:00:40 +08:00
|
|
|
if (!cp->exclusive) {
|
|
|
|
_debug("search 1");
|
|
|
|
spin_lock(&local->client_conns_lock);
|
|
|
|
p = local->client_conns.rb_node;
|
|
|
|
while (p) {
|
|
|
|
conn = rb_entry(p, struct rxrpc_connection, client_node);
|
|
|
|
|
|
|
|
#define cmp(X) ((long)conn->params.X - (long)cp->X)
|
|
|
|
diff = (cmp(peer) ?:
|
|
|
|
cmp(key) ?:
|
|
|
|
cmp(security_level));
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
#undef cmp
|
|
|
|
if (diff < 0) {
|
2016-04-04 21:00:40 +08:00
|
|
|
p = p->rb_left;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
} else if (diff > 0) {
|
2016-04-04 21:00:40 +08:00
|
|
|
p = p->rb_right;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
} else {
|
|
|
|
if (rxrpc_may_reuse_conn(conn) &&
|
|
|
|
rxrpc_get_connection_maybe(conn))
|
|
|
|
goto found_extant_conn;
|
|
|
|
/* The connection needs replacing. It's better
|
|
|
|
* to effect that when we have something to
|
|
|
|
* replace it with so that we don't have to
|
|
|
|
* rebalance the tree twice.
|
|
|
|
*/
|
|
|
|
break;
|
|
|
|
}
|
2016-04-04 21:00:40 +08:00
|
|
|
}
|
|
|
|
spin_unlock(&local->client_conns_lock);
|
|
|
|
}
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/* There wasn't a connection yet or we need an exclusive connection.
|
|
|
|
* We need to create a candidate and then potentially redo the search
|
|
|
|
* in case we're racing with another thread also trying to connect on a
|
|
|
|
* shareable connection.
|
|
|
|
*/
|
|
|
|
_debug("new conn");
|
2016-04-04 21:00:40 +08:00
|
|
|
candidate = rxrpc_alloc_client_connection(cp, gfp);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
if (IS_ERR(candidate)) {
|
|
|
|
ret = PTR_ERR(candidate);
|
|
|
|
goto error_peer;
|
2016-04-04 21:00:40 +08:00
|
|
|
}
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/* Add the call to the new connection's waiting list in case we're
|
|
|
|
* going to have to wait for the connection to come live. It's our
|
|
|
|
* connection, so we want first dibs on the channel slots. We would
|
|
|
|
* normally have to take channel_lock but we do this before anyone else
|
|
|
|
* can see the connection.
|
|
|
|
*/
|
|
|
|
list_add_tail(&call->chan_wait_link, &candidate->waiting_calls);
|
|
|
|
|
2016-04-04 21:00:40 +08:00
|
|
|
if (cp->exclusive) {
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
call->conn = candidate;
|
2016-09-07 22:19:25 +08:00
|
|
|
call->security_ix = candidate->security_ix;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
_leave(" = 0 [exclusive %d]", candidate->debug_id);
|
|
|
|
return 0;
|
2016-04-04 21:00:40 +08:00
|
|
|
}
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/* Publish the new connection for userspace to find. We need to redo
|
|
|
|
* the search before doing this lest we race with someone else adding a
|
|
|
|
* conflicting instance.
|
2016-04-04 21:00:40 +08:00
|
|
|
*/
|
|
|
|
_debug("search 2");
|
|
|
|
spin_lock(&local->client_conns_lock);
|
|
|
|
|
|
|
|
pp = &local->client_conns.rb_node;
|
|
|
|
parent = NULL;
|
|
|
|
while (*pp) {
|
|
|
|
parent = *pp;
|
|
|
|
conn = rb_entry(parent, struct rxrpc_connection, client_node);
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
#define cmp(X) ((long)conn->params.X - (long)candidate->params.X)
|
2016-04-04 21:00:40 +08:00
|
|
|
diff = (cmp(peer) ?:
|
|
|
|
cmp(key) ?:
|
|
|
|
cmp(security_level));
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
#undef cmp
|
|
|
|
if (diff < 0) {
|
2016-04-04 21:00:40 +08:00
|
|
|
pp = &(*pp)->rb_left;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
} else if (diff > 0) {
|
2016-04-04 21:00:40 +08:00
|
|
|
pp = &(*pp)->rb_right;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
} else {
|
|
|
|
if (rxrpc_may_reuse_conn(conn) &&
|
|
|
|
rxrpc_get_connection_maybe(conn))
|
|
|
|
goto found_extant_conn;
|
|
|
|
/* The old connection is from an outdated epoch. */
|
|
|
|
_debug("replace conn");
|
|
|
|
clear_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags);
|
|
|
|
rb_replace_node(&conn->client_node,
|
|
|
|
&candidate->client_node,
|
|
|
|
&local->client_conns);
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_replace);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
goto candidate_published;
|
|
|
|
}
|
2016-04-04 21:00:40 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
_debug("new conn");
|
2016-06-30 17:45:22 +08:00
|
|
|
rb_link_node(&candidate->client_node, parent, pp);
|
|
|
|
rb_insert_color(&candidate->client_node, &local->client_conns);
|
2016-04-04 21:00:40 +08:00
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
candidate_published:
|
|
|
|
set_bit(RXRPC_CONN_IN_CLIENT_CONNS, &candidate->flags);
|
|
|
|
call->conn = candidate;
|
2016-09-07 22:19:25 +08:00
|
|
|
call->security_ix = candidate->security_ix;
|
2016-04-04 21:00:40 +08:00
|
|
|
spin_unlock(&local->client_conns_lock);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
_leave(" = 0 [new %d]", candidate->debug_id);
|
|
|
|
return 0;
|
2016-04-04 21:00:40 +08:00
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/* We come here if we found a suitable connection already in existence.
|
|
|
|
* Discard any candidate we may have allocated, and try to get a
|
|
|
|
* channel on this one.
|
|
|
|
*/
|
|
|
|
found_extant_conn:
|
|
|
|
_debug("found conn");
|
|
|
|
spin_unlock(&local->client_conns_lock);
|
2016-04-04 21:00:40 +08:00
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
if (candidate) {
|
|
|
|
trace_rxrpc_client(candidate, -1, rxrpc_client_duplicate);
|
|
|
|
rxrpc_put_connection(candidate);
|
|
|
|
candidate = NULL;
|
|
|
|
}
|
2016-04-04 21:00:40 +08:00
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
spin_lock(&conn->channel_lock);
|
|
|
|
call->conn = conn;
|
2016-09-07 22:19:25 +08:00
|
|
|
call->security_ix = conn->security_ix;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
list_add(&call->chan_wait_link, &conn->waiting_calls);
|
2016-04-04 21:00:40 +08:00
|
|
|
spin_unlock(&conn->channel_lock);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
_leave(" = 0 [extant %d]", conn->debug_id);
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
error_peer:
|
2016-04-04 21:00:40 +08:00
|
|
|
rxrpc_put_peer(cp->peer);
|
|
|
|
cp->peer = NULL;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
error:
|
|
|
|
_leave(" = %d", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
2016-04-04 21:00:40 +08:00
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/*
|
|
|
|
* Activate a connection.
|
|
|
|
*/
|
|
|
|
static void rxrpc_activate_conn(struct rxrpc_connection *conn)
|
|
|
|
{
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_to_active);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
conn->cache_state = RXRPC_CONN_CLIENT_ACTIVE;
|
|
|
|
rxrpc_nr_active_client_conns++;
|
|
|
|
list_move_tail(&conn->cache_link, &rxrpc_active_client_conns);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Attempt to animate a connection for a new call.
|
|
|
|
*
|
|
|
|
* If it's not exclusive, the connection is in the endpoint tree, and we're in
|
|
|
|
* the conn's list of those waiting to grab a channel. There is, however, a
|
|
|
|
* limit on the number of live connections allowed at any one time, so we may
|
|
|
|
* have to wait for capacity to become available.
|
|
|
|
*
|
|
|
|
* Note that a connection on the waiting queue might *also* have active
|
|
|
|
* channels if it has been culled to make space and then re-requested by a new
|
|
|
|
* call.
|
|
|
|
*/
|
|
|
|
static void rxrpc_animate_client_conn(struct rxrpc_connection *conn)
|
|
|
|
{
|
|
|
|
unsigned int nr_conns;
|
|
|
|
|
|
|
|
_enter("%d,%d", conn->debug_id, conn->cache_state);
|
|
|
|
|
|
|
|
if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
spin_lock(&rxrpc_client_conn_cache_lock);
|
|
|
|
|
|
|
|
nr_conns = rxrpc_nr_client_conns;
|
2016-09-17 17:49:14 +08:00
|
|
|
if (!test_and_set_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
|
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_count);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
rxrpc_nr_client_conns = nr_conns + 1;
|
2016-09-17 17:49:14 +08:00
|
|
|
}
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
|
|
|
|
switch (conn->cache_state) {
|
|
|
|
case RXRPC_CONN_CLIENT_ACTIVE:
|
|
|
|
case RXRPC_CONN_CLIENT_WAITING:
|
|
|
|
break;
|
|
|
|
|
|
|
|
case RXRPC_CONN_CLIENT_INACTIVE:
|
|
|
|
case RXRPC_CONN_CLIENT_CULLED:
|
|
|
|
case RXRPC_CONN_CLIENT_IDLE:
|
|
|
|
if (nr_conns >= rxrpc_max_client_connections)
|
|
|
|
goto wait_for_capacity;
|
|
|
|
goto activate_conn;
|
|
|
|
|
|
|
|
default:
|
|
|
|
BUG();
|
2016-06-30 17:45:22 +08:00
|
|
|
}
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
out_unlock:
|
|
|
|
spin_unlock(&rxrpc_client_conn_cache_lock);
|
|
|
|
out:
|
|
|
|
_leave(" [%d]", conn->cache_state);
|
|
|
|
return;
|
2016-04-04 21:00:40 +08:00
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
activate_conn:
|
|
|
|
_debug("activate");
|
|
|
|
rxrpc_activate_conn(conn);
|
|
|
|
goto out_unlock;
|
|
|
|
|
|
|
|
wait_for_capacity:
|
|
|
|
_debug("wait");
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
|
|
|
|
list_move_tail(&conn->cache_link, &rxrpc_waiting_client_conns);
|
|
|
|
goto out_unlock;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Deactivate a channel.
|
|
|
|
*/
|
|
|
|
static void rxrpc_deactivate_one_channel(struct rxrpc_connection *conn,
|
|
|
|
unsigned int channel)
|
|
|
|
{
|
|
|
|
struct rxrpc_channel *chan = &conn->channels[channel];
|
|
|
|
|
|
|
|
rcu_assign_pointer(chan->call, NULL);
|
|
|
|
conn->active_chans &= ~(1 << channel);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Assign a channel to the call at the front of the queue and wake the call up.
|
|
|
|
* We don't increment the callNumber counter until this number has been exposed
|
|
|
|
* to the world.
|
|
|
|
*/
|
|
|
|
static void rxrpc_activate_one_channel(struct rxrpc_connection *conn,
|
|
|
|
unsigned int channel)
|
|
|
|
{
|
|
|
|
struct rxrpc_channel *chan = &conn->channels[channel];
|
|
|
|
struct rxrpc_call *call = list_entry(conn->waiting_calls.next,
|
|
|
|
struct rxrpc_call, chan_wait_link);
|
|
|
|
u32 call_id = chan->call_counter + 1;
|
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, channel, rxrpc_client_chan_activate);
|
|
|
|
|
2016-09-04 20:10:10 +08:00
|
|
|
write_lock_bh(&call->state_lock);
|
|
|
|
call->state = RXRPC_CALL_CLIENT_SEND_REQUEST;
|
|
|
|
write_unlock_bh(&call->state_lock);
|
|
|
|
|
2016-08-30 16:49:29 +08:00
|
|
|
rxrpc_see_call(call);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
list_del_init(&call->chan_wait_link);
|
|
|
|
conn->active_chans |= 1 << channel;
|
|
|
|
call->peer = rxrpc_get_peer(conn->params.peer);
|
|
|
|
call->cid = conn->proto.cid | channel;
|
|
|
|
call->call_id = call_id;
|
|
|
|
|
|
|
|
_net("CONNECT call %08x:%08x as call %d on conn %d",
|
|
|
|
call->cid, call->call_id, call->debug_id, conn->debug_id);
|
|
|
|
|
|
|
|
/* Paired with the read barrier in rxrpc_wait_for_channel(). This
|
|
|
|
* orders cid and epoch in the connection wrt to call_id without the
|
|
|
|
* need to take the channel_lock.
|
|
|
|
*
|
|
|
|
* We provisionally assign a callNumber at this point, but we don't
|
|
|
|
* confirm it until the call is about to be exposed.
|
|
|
|
*
|
|
|
|
* TODO: Pair with a barrier in the data_ready handler when that looks
|
|
|
|
* at the call ID through a connection channel.
|
|
|
|
*/
|
|
|
|
smp_wmb();
|
|
|
|
chan->call_id = call_id;
|
|
|
|
rcu_assign_pointer(chan->call, call);
|
|
|
|
wake_up(&call->waitq);
|
|
|
|
}
|
|
|
|
|
2016-09-30 05:37:15 +08:00
|
|
|
/*
|
|
|
|
* Assign channels and callNumbers to waiting calls with channel_lock
|
|
|
|
* held by caller.
|
|
|
|
*/
|
|
|
|
static void rxrpc_activate_channels_locked(struct rxrpc_connection *conn)
|
|
|
|
{
|
|
|
|
u8 avail, mask;
|
|
|
|
|
|
|
|
switch (conn->cache_state) {
|
|
|
|
case RXRPC_CONN_CLIENT_ACTIVE:
|
|
|
|
mask = RXRPC_ACTIVE_CHANS_MASK;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
while (!list_empty(&conn->waiting_calls) &&
|
|
|
|
(avail = ~conn->active_chans,
|
|
|
|
avail &= mask,
|
|
|
|
avail != 0))
|
|
|
|
rxrpc_activate_one_channel(conn, __ffs(avail));
|
|
|
|
}
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/*
|
|
|
|
* Assign channels and callNumbers to waiting calls.
|
|
|
|
*/
|
|
|
|
static void rxrpc_activate_channels(struct rxrpc_connection *conn)
|
|
|
|
{
|
|
|
|
_enter("%d", conn->debug_id);
|
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_activate_chans);
|
|
|
|
|
2016-09-30 05:37:15 +08:00
|
|
|
if (conn->active_chans == RXRPC_ACTIVE_CHANS_MASK)
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
return;
|
|
|
|
|
|
|
|
spin_lock(&conn->channel_lock);
|
2016-09-30 05:37:15 +08:00
|
|
|
rxrpc_activate_channels_locked(conn);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
spin_unlock(&conn->channel_lock);
|
|
|
|
_leave("");
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Wait for a callNumber and a channel to be granted to a call.
|
|
|
|
*/
|
|
|
|
static int rxrpc_wait_for_channel(struct rxrpc_call *call, gfp_t gfp)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
|
|
|
|
_enter("%d", call->debug_id);
|
|
|
|
|
|
|
|
if (!call->call_id) {
|
|
|
|
DECLARE_WAITQUEUE(myself, current);
|
2016-04-04 21:00:40 +08:00
|
|
|
|
|
|
|
if (!gfpflags_allow_blocking(gfp)) {
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
ret = -EAGAIN;
|
|
|
|
goto out;
|
2016-04-04 21:00:40 +08:00
|
|
|
}
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
add_wait_queue_exclusive(&call->waitq, &myself);
|
2016-04-04 21:00:40 +08:00
|
|
|
for (;;) {
|
|
|
|
set_current_state(TASK_INTERRUPTIBLE);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
if (call->call_id)
|
|
|
|
break;
|
|
|
|
if (signal_pending(current)) {
|
|
|
|
ret = -ERESTARTSYS;
|
2016-04-04 21:00:40 +08:00
|
|
|
break;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
}
|
2016-04-04 21:00:40 +08:00
|
|
|
schedule();
|
|
|
|
}
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
remove_wait_queue(&call->waitq, &myself);
|
2016-04-04 21:00:40 +08:00
|
|
|
__set_current_state(TASK_RUNNING);
|
|
|
|
}
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/* Paired with the write barrier in rxrpc_activate_one_channel(). */
|
|
|
|
smp_rmb();
|
|
|
|
|
|
|
|
out:
|
|
|
|
_leave(" = %d", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* find a connection for a call
|
|
|
|
* - called in process context with IRQs enabled
|
|
|
|
*/
|
|
|
|
int rxrpc_connect_call(struct rxrpc_call *call,
|
|
|
|
struct rxrpc_conn_parameters *cp,
|
|
|
|
struct sockaddr_rxrpc *srx,
|
|
|
|
gfp_t gfp)
|
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
_enter("{%d,%lx},", call->debug_id, call->user_call_ID);
|
|
|
|
|
|
|
|
rxrpc_discard_expired_client_conns(NULL);
|
|
|
|
rxrpc_cull_active_client_conns();
|
|
|
|
|
|
|
|
ret = rxrpc_get_client_conn(call, cp, srx, gfp);
|
|
|
|
if (ret < 0)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
rxrpc_animate_client_conn(call->conn);
|
|
|
|
rxrpc_activate_channels(call->conn);
|
|
|
|
|
|
|
|
ret = rxrpc_wait_for_channel(call, gfp);
|
|
|
|
if (ret < 0)
|
|
|
|
rxrpc_disconnect_client_call(call);
|
|
|
|
|
|
|
|
_leave(" = %d", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Note that a connection is about to be exposed to the world. Once it is
|
|
|
|
* exposed, we maintain an extra ref on it that stops it from being summarily
|
|
|
|
* discarded before it's (a) had a chance to deal with retransmission and (b)
|
|
|
|
* had a chance at re-use (the per-connection security negotiation is
|
|
|
|
* expensive).
|
|
|
|
*/
|
2016-09-17 17:49:14 +08:00
|
|
|
static void rxrpc_expose_client_conn(struct rxrpc_connection *conn,
|
|
|
|
unsigned int channel)
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
{
|
2016-09-17 17:49:14 +08:00
|
|
|
if (!test_and_set_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
|
|
|
|
trace_rxrpc_client(conn, channel, rxrpc_client_exposed);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
rxrpc_get_connection(conn);
|
2016-09-17 17:49:14 +08:00
|
|
|
}
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Note that a call, and thus a connection, is about to be exposed to the
|
|
|
|
* world.
|
|
|
|
*/
|
|
|
|
void rxrpc_expose_client_call(struct rxrpc_call *call)
|
|
|
|
{
|
2016-09-17 17:49:14 +08:00
|
|
|
unsigned int channel = call->cid & RXRPC_CHANNELMASK;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
struct rxrpc_connection *conn = call->conn;
|
2016-09-17 17:49:14 +08:00
|
|
|
struct rxrpc_channel *chan = &conn->channels[channel];
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
|
|
|
|
if (!test_and_set_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
|
|
|
|
/* Mark the call ID as being used. If the callNumber counter
|
|
|
|
* exceeds ~2 billion, we kill the connection after its
|
|
|
|
* outstanding calls have finished so that the counter doesn't
|
|
|
|
* wrap.
|
|
|
|
*/
|
|
|
|
chan->call_counter++;
|
|
|
|
if (chan->call_counter >= INT_MAX)
|
|
|
|
set_bit(RXRPC_CONN_DONT_REUSE, &conn->flags);
|
2016-09-17 17:49:14 +08:00
|
|
|
rxrpc_expose_client_conn(conn, channel);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Disconnect a client call.
|
|
|
|
*/
|
|
|
|
void rxrpc_disconnect_client_call(struct rxrpc_call *call)
|
|
|
|
{
|
|
|
|
unsigned int channel = call->cid & RXRPC_CHANNELMASK;
|
|
|
|
struct rxrpc_connection *conn = call->conn;
|
|
|
|
struct rxrpc_channel *chan = &conn->channels[channel];
|
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, channel, rxrpc_client_chan_disconnect);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
call->conn = NULL;
|
|
|
|
|
2016-04-04 21:00:40 +08:00
|
|
|
spin_lock(&conn->channel_lock);
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/* Calls that have never actually been assigned a channel can simply be
|
|
|
|
* discarded. If the conn didn't get used either, it will follow
|
|
|
|
* immediately unless someone else grabs it in the meantime.
|
|
|
|
*/
|
|
|
|
if (!list_empty(&call->chan_wait_link)) {
|
|
|
|
_debug("call is waiting");
|
|
|
|
ASSERTCMP(call->call_id, ==, 0);
|
|
|
|
ASSERT(!test_bit(RXRPC_CALL_EXPOSED, &call->flags));
|
|
|
|
list_del_init(&call->chan_wait_link);
|
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, channel, rxrpc_client_chan_unstarted);
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
/* We must deactivate or idle the connection if it's now
|
|
|
|
* waiting for nothing.
|
|
|
|
*/
|
|
|
|
spin_lock(&rxrpc_client_conn_cache_lock);
|
|
|
|
if (conn->cache_state == RXRPC_CONN_CLIENT_WAITING &&
|
|
|
|
list_empty(&conn->waiting_calls) &&
|
|
|
|
!conn->active_chans)
|
|
|
|
goto idle_connection;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
ASSERTCMP(rcu_access_pointer(chan->call), ==, call);
|
|
|
|
|
|
|
|
/* If a client call was exposed to the world, we save the result for
|
|
|
|
* retransmission.
|
|
|
|
*
|
|
|
|
* We use a barrier here so that the call number and abort code can be
|
|
|
|
* read without needing to take a lock.
|
|
|
|
*
|
|
|
|
* TODO: Make the incoming packet handler check this and handle
|
|
|
|
* terminal retransmission without requiring access to the call.
|
|
|
|
*/
|
|
|
|
if (test_bit(RXRPC_CALL_EXPOSED, &call->flags)) {
|
2016-08-30 16:49:28 +08:00
|
|
|
_debug("exposed %u,%u", call->call_id, call->abort_code);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
__rxrpc_disconnect_call(conn, call);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* See if we can pass the channel directly to another call. */
|
|
|
|
if (conn->cache_state == RXRPC_CONN_CLIENT_ACTIVE &&
|
|
|
|
!list_empty(&conn->waiting_calls)) {
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
rxrpc_activate_one_channel(conn, channel);
|
|
|
|
goto out_2;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Things are more complex and we need the cache lock. We might be
|
|
|
|
* able to simply idle the conn or it might now be lurking on the wait
|
|
|
|
* list. It might even get moved back to the active list whilst we're
|
|
|
|
* waiting for the lock.
|
|
|
|
*/
|
|
|
|
spin_lock(&rxrpc_client_conn_cache_lock);
|
|
|
|
|
|
|
|
switch (conn->cache_state) {
|
|
|
|
case RXRPC_CONN_CLIENT_ACTIVE:
|
|
|
|
if (list_empty(&conn->waiting_calls)) {
|
|
|
|
rxrpc_deactivate_one_channel(conn, channel);
|
|
|
|
if (!conn->active_chans) {
|
|
|
|
rxrpc_nr_active_client_conns--;
|
|
|
|
goto idle_connection;
|
|
|
|
}
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, channel, rxrpc_client_chan_pass);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
rxrpc_activate_one_channel(conn, channel);
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
case RXRPC_CONN_CLIENT_CULLED:
|
|
|
|
rxrpc_deactivate_one_channel(conn, channel);
|
|
|
|
ASSERT(list_empty(&conn->waiting_calls));
|
|
|
|
if (!conn->active_chans)
|
|
|
|
goto idle_connection;
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
case RXRPC_CONN_CLIENT_WAITING:
|
|
|
|
rxrpc_deactivate_one_channel(conn, channel);
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
default:
|
|
|
|
BUG();
|
|
|
|
}
|
2016-04-04 21:00:40 +08:00
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
out:
|
|
|
|
spin_unlock(&rxrpc_client_conn_cache_lock);
|
|
|
|
out_2:
|
|
|
|
spin_unlock(&conn->channel_lock);
|
2016-04-04 21:00:40 +08:00
|
|
|
rxrpc_put_connection(conn);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
_leave("");
|
|
|
|
return;
|
|
|
|
|
|
|
|
idle_connection:
|
|
|
|
/* As no channels remain active, the connection gets deactivated
|
|
|
|
* immediately or moved to the idle list for a short while.
|
|
|
|
*/
|
|
|
|
if (test_bit(RXRPC_CONN_EXPOSED, &conn->flags)) {
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, channel, rxrpc_client_to_idle);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
conn->idle_timestamp = jiffies;
|
|
|
|
conn->cache_state = RXRPC_CONN_CLIENT_IDLE;
|
|
|
|
list_move_tail(&conn->cache_link, &rxrpc_idle_client_conns);
|
|
|
|
if (rxrpc_idle_client_conns.next == &conn->cache_link &&
|
|
|
|
!rxrpc_kill_all_client_conns)
|
|
|
|
queue_delayed_work(rxrpc_workqueue,
|
|
|
|
&rxrpc_client_conn_reap,
|
|
|
|
rxrpc_conn_idle_client_expiry);
|
|
|
|
} else {
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, channel, rxrpc_client_to_inactive);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
|
|
|
|
list_del_init(&conn->cache_link);
|
|
|
|
}
|
|
|
|
goto out;
|
2016-04-04 21:00:40 +08:00
|
|
|
}
|
2016-06-30 17:45:22 +08:00
|
|
|
|
|
|
|
/*
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
* Clean up a dead client connection.
|
2016-06-30 17:45:22 +08:00
|
|
|
*/
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
static struct rxrpc_connection *
|
|
|
|
rxrpc_put_one_client_conn(struct rxrpc_connection *conn)
|
2016-06-30 17:45:22 +08:00
|
|
|
{
|
2016-09-17 17:49:12 +08:00
|
|
|
struct rxrpc_connection *next = NULL;
|
2016-06-30 17:45:22 +08:00
|
|
|
struct rxrpc_local *local = conn->params.local;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
unsigned int nr_conns;
|
2016-06-30 17:45:22 +08:00
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_cleanup);
|
|
|
|
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
if (test_bit(RXRPC_CONN_IN_CLIENT_CONNS, &conn->flags)) {
|
|
|
|
spin_lock(&local->client_conns_lock);
|
|
|
|
if (test_and_clear_bit(RXRPC_CONN_IN_CLIENT_CONNS,
|
|
|
|
&conn->flags))
|
|
|
|
rb_erase(&conn->client_node, &local->client_conns);
|
|
|
|
spin_unlock(&local->client_conns_lock);
|
|
|
|
}
|
2016-06-30 17:45:22 +08:00
|
|
|
|
|
|
|
rxrpc_put_client_connection_id(conn);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
|
|
|
|
ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_INACTIVE);
|
|
|
|
|
2016-09-17 17:49:12 +08:00
|
|
|
if (test_bit(RXRPC_CONN_COUNTED, &conn->flags)) {
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_uncount);
|
2016-09-17 17:49:12 +08:00
|
|
|
spin_lock(&rxrpc_client_conn_cache_lock);
|
|
|
|
nr_conns = --rxrpc_nr_client_conns;
|
|
|
|
|
|
|
|
if (nr_conns < rxrpc_max_client_connections &&
|
|
|
|
!list_empty(&rxrpc_waiting_client_conns)) {
|
|
|
|
next = list_entry(rxrpc_waiting_client_conns.next,
|
|
|
|
struct rxrpc_connection, cache_link);
|
|
|
|
rxrpc_get_connection(next);
|
|
|
|
rxrpc_activate_conn(next);
|
|
|
|
}
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
|
2016-09-17 17:49:12 +08:00
|
|
|
spin_unlock(&rxrpc_client_conn_cache_lock);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
rxrpc_kill_connection(conn);
|
|
|
|
if (next)
|
|
|
|
rxrpc_activate_channels(next);
|
|
|
|
|
|
|
|
/* We need to get rid of the temporary ref we took upon next, but we
|
|
|
|
* can't call rxrpc_put_connection() recursively.
|
|
|
|
*/
|
|
|
|
return next;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Clean up a dead client connections.
|
|
|
|
*/
|
|
|
|
void rxrpc_put_client_conn(struct rxrpc_connection *conn)
|
|
|
|
{
|
2016-09-17 17:49:14 +08:00
|
|
|
const void *here = __builtin_return_address(0);
|
|
|
|
int n;
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
|
|
|
|
do {
|
2016-09-17 17:49:14 +08:00
|
|
|
n = atomic_dec_return(&conn->usage);
|
|
|
|
trace_rxrpc_conn(conn, rxrpc_conn_put_client, n, here);
|
|
|
|
if (n > 0)
|
|
|
|
return;
|
|
|
|
ASSERTCMP(n, >=, 0);
|
|
|
|
|
|
|
|
conn = rxrpc_put_one_client_conn(conn);
|
|
|
|
} while (conn);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Kill the longest-active client connections to make room for new ones.
|
|
|
|
*/
|
|
|
|
static void rxrpc_cull_active_client_conns(void)
|
|
|
|
{
|
|
|
|
struct rxrpc_connection *conn;
|
|
|
|
unsigned int nr_conns = rxrpc_nr_client_conns;
|
|
|
|
unsigned int nr_active, limit;
|
|
|
|
|
|
|
|
_enter("");
|
|
|
|
|
|
|
|
ASSERTCMP(nr_conns, >=, 0);
|
|
|
|
if (nr_conns < rxrpc_max_client_connections) {
|
|
|
|
_leave(" [ok]");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
limit = rxrpc_reap_client_connections;
|
|
|
|
|
|
|
|
spin_lock(&rxrpc_client_conn_cache_lock);
|
|
|
|
nr_active = rxrpc_nr_active_client_conns;
|
|
|
|
|
|
|
|
while (nr_active > limit) {
|
|
|
|
ASSERT(!list_empty(&rxrpc_active_client_conns));
|
|
|
|
conn = list_entry(rxrpc_active_client_conns.next,
|
|
|
|
struct rxrpc_connection, cache_link);
|
|
|
|
ASSERTCMP(conn->cache_state, ==, RXRPC_CONN_CLIENT_ACTIVE);
|
|
|
|
|
|
|
|
if (list_empty(&conn->waiting_calls)) {
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_to_culled);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
conn->cache_state = RXRPC_CONN_CLIENT_CULLED;
|
|
|
|
list_del_init(&conn->cache_link);
|
|
|
|
} else {
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_to_waiting);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
conn->cache_state = RXRPC_CONN_CLIENT_WAITING;
|
|
|
|
list_move_tail(&conn->cache_link,
|
|
|
|
&rxrpc_waiting_client_conns);
|
|
|
|
}
|
|
|
|
|
|
|
|
nr_active--;
|
|
|
|
}
|
|
|
|
|
|
|
|
rxrpc_nr_active_client_conns = nr_active;
|
|
|
|
spin_unlock(&rxrpc_client_conn_cache_lock);
|
|
|
|
ASSERTCMP(nr_active, >=, 0);
|
|
|
|
_leave(" [culled]");
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Discard expired client connections from the idle list. Each conn in the
|
|
|
|
* idle list has been exposed and holds an extra ref because of that.
|
|
|
|
*
|
|
|
|
* This may be called from conn setup or from a work item so cannot be
|
|
|
|
* considered non-reentrant.
|
|
|
|
*/
|
|
|
|
static void rxrpc_discard_expired_client_conns(struct work_struct *work)
|
|
|
|
{
|
|
|
|
struct rxrpc_connection *conn;
|
|
|
|
unsigned long expiry, conn_expires_at, now;
|
|
|
|
unsigned int nr_conns;
|
|
|
|
bool did_discard = false;
|
|
|
|
|
|
|
|
_enter("%c", work ? 'w' : 'n');
|
|
|
|
|
|
|
|
if (list_empty(&rxrpc_idle_client_conns)) {
|
|
|
|
_leave(" [empty]");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Don't double up on the discarding */
|
|
|
|
if (!spin_trylock(&rxrpc_client_conn_discard_mutex)) {
|
|
|
|
_leave(" [already]");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* We keep an estimate of what the number of conns ought to be after
|
|
|
|
* we've discarded some so that we don't overdo the discarding.
|
|
|
|
*/
|
|
|
|
nr_conns = rxrpc_nr_client_conns;
|
|
|
|
|
|
|
|
next:
|
|
|
|
spin_lock(&rxrpc_client_conn_cache_lock);
|
|
|
|
|
|
|
|
if (list_empty(&rxrpc_idle_client_conns))
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
conn = list_entry(rxrpc_idle_client_conns.next,
|
|
|
|
struct rxrpc_connection, cache_link);
|
|
|
|
ASSERT(test_bit(RXRPC_CONN_EXPOSED, &conn->flags));
|
|
|
|
|
|
|
|
if (!rxrpc_kill_all_client_conns) {
|
|
|
|
/* If the number of connections is over the reap limit, we
|
|
|
|
* expedite discard by reducing the expiry timeout. We must,
|
|
|
|
* however, have at least a short grace period to be able to do
|
|
|
|
* final-ACK or ABORT retransmission.
|
|
|
|
*/
|
|
|
|
expiry = rxrpc_conn_idle_client_expiry;
|
|
|
|
if (nr_conns > rxrpc_reap_client_connections)
|
|
|
|
expiry = rxrpc_conn_idle_client_fast_expiry;
|
|
|
|
|
|
|
|
conn_expires_at = conn->idle_timestamp + expiry;
|
|
|
|
|
|
|
|
now = READ_ONCE(jiffies);
|
|
|
|
if (time_after(conn_expires_at, now))
|
|
|
|
goto not_yet_expired;
|
|
|
|
}
|
|
|
|
|
2016-09-17 17:49:14 +08:00
|
|
|
trace_rxrpc_client(conn, -1, rxrpc_client_discard);
|
rxrpc: Improve management and caching of client connection objects
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
2016-08-24 14:30:52 +08:00
|
|
|
if (!test_and_clear_bit(RXRPC_CONN_EXPOSED, &conn->flags))
|
|
|
|
BUG();
|
|
|
|
conn->cache_state = RXRPC_CONN_CLIENT_INACTIVE;
|
|
|
|
list_del_init(&conn->cache_link);
|
|
|
|
|
|
|
|
spin_unlock(&rxrpc_client_conn_cache_lock);
|
|
|
|
|
|
|
|
/* When we cleared the EXPOSED flag, we took on responsibility for the
|
|
|
|
* reference that that had on the usage count. We deal with that here.
|
|
|
|
* If someone re-sets the flag and re-gets the ref, that's fine.
|
|
|
|
*/
|
|
|
|
rxrpc_put_connection(conn);
|
|
|
|
did_discard = true;
|
|
|
|
nr_conns--;
|
|
|
|
goto next;
|
|
|
|
|
|
|
|
not_yet_expired:
|
|
|
|
/* The connection at the front of the queue hasn't yet expired, so
|
|
|
|
* schedule the work item for that point if we discarded something.
|
|
|
|
*
|
|
|
|
* We don't worry if the work item is already scheduled - it can look
|
|
|
|
* after rescheduling itself at a later time. We could cancel it, but
|
|
|
|
* then things get messier.
|
|
|
|
*/
|
|
|
|
_debug("not yet");
|
|
|
|
if (!rxrpc_kill_all_client_conns)
|
|
|
|
queue_delayed_work(rxrpc_workqueue,
|
|
|
|
&rxrpc_client_conn_reap,
|
|
|
|
conn_expires_at - now);
|
|
|
|
|
|
|
|
out:
|
|
|
|
spin_unlock(&rxrpc_client_conn_cache_lock);
|
|
|
|
spin_unlock(&rxrpc_client_conn_discard_mutex);
|
|
|
|
_leave("");
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Preemptively destroy all the client connection records rather than waiting
|
|
|
|
* for them to time out
|
|
|
|
*/
|
|
|
|
void __exit rxrpc_destroy_all_client_connections(void)
|
|
|
|
{
|
|
|
|
_enter("");
|
|
|
|
|
|
|
|
spin_lock(&rxrpc_client_conn_cache_lock);
|
|
|
|
rxrpc_kill_all_client_conns = true;
|
|
|
|
spin_unlock(&rxrpc_client_conn_cache_lock);
|
|
|
|
|
|
|
|
cancel_delayed_work(&rxrpc_client_conn_reap);
|
|
|
|
|
|
|
|
if (!queue_delayed_work(rxrpc_workqueue, &rxrpc_client_conn_reap, 0))
|
|
|
|
_debug("destroy: queue failed");
|
|
|
|
|
|
|
|
_leave("");
|
2016-06-30 17:45:22 +08:00
|
|
|
}
|