Fix the rxrpc call expiration timeouts and make them settable from
userspace. By analogy with other rx implementations, there should be three
timeouts:
(1) "Normal timeout"
This is set for all calls and is triggered if we haven't received any
packets from the peer in a while. It is measured from the last time
we received any packet on that call. This is not reset by any
connection packets (such as CHALLENGE/RESPONSE packets).
If a service operation takes a long time, the server should generate
PING ACKs at a duration that's substantially less than the normal
timeout so is to keep both sides alive. This is set at 1/6 of normal
timeout.
(2) "Idle timeout"
This is set only for a service call and is triggered if we stop
receiving the DATA packets that comprise the request data. It is
measured from the last time we received a DATA packet.
(3) "Hard timeout"
This can be set for a call and specified the maximum lifetime of that
call. It should not be specified by default. Some operations (such
as volume transfer) take a long time.
Allow userspace to set/change the timeouts on a call with sendmsg, using a
control message:
RXRPC_SET_CALL_TIMEOUTS
The data to the message is a number of 32-bit words, not all of which need
be given:
u32 hard_timeout; /* sec from first packet */
u32 idle_timeout; /* msec from packet Rx */
u32 normal_timeout; /* msec from data Rx */
This can be set in combination with any other sendmsg() that affects a
call.
Signed-off-by: David Howells <dhowells@redhat.com>
Fix call expiry handling in the following ways
(1) If all the request data from a client call is acked, don't send a
follow up IDLE ACK with firstPacket == 1 and previousPacket == 0 as
this appears to fool some servers into thinking everything has been
accepted.
(2) Never send an abort back to the server once it has ACK'd all the
request packets; rather just try to reuse the channel for the next
call. The first request DATA packet of the next call on the same
channel will implicitly ACK the entire reply of the dead call - even
if we haven't transmitted it yet.
(3) Don't send RX_CALL_TIMEOUT in an ABORT packet, librx uses abort codes
to pass local errors to the caller in addition to remote errors, and
this is meant to be local only.
The following also need to be addressed in future patches:
(4) Service calls should send PING ACKs as 'keep alives' if the server is
still processing the call.
(5) VERSION REPLY packets should be sent to the peers of service
connections to act as keep-alives. This is used to keep firewall
routes in place. The AFS CM should enable this.
Signed-off-by: David Howells <dhowells@redhat.com>
In preparation to enabling -Wimplicit-fallthrough, mark switch cases
where we are expecting to fall through.
Signed-off-by: Gustavo A. R. Silva <garsilva@embeddedor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Make it possible for a client to use AuriStor's service upgrade facility.
The client does this by adding an RXRPC_UPGRADE_SERVICE control message to
the first sendmsg() of a call. This takes no parameters.
When recvmsg() starts returning data from the call, the service ID field in
the returned msg_name will reflect the result of the upgrade attempt. If
the upgrade was ignored, srx_service will match what was set in the
sendmsg(); if the upgrade happened the srx_service will be altered to
indicate the service the server upgraded to.
Note that:
(1) The choice of upgrade service is up to the server
(2) Further client calls to the same server that would share a connection
are blocked if an upgrade probe is in progress.
(3) This should only be used to probe the service. Clients should then
use the returned service ID in all subsequent communications with that
server (and not set the upgrade). Note that the kernel will not
retain this information should the connection expire from its cache.
(4) If a server that supports upgrading is replaced by one that doesn't,
whilst a connection is live, and if the replacement is running, say,
OpenAFS 1.6.4 or older or an older IBM AFS, then the replacement
server will not respond to packets sent to the upgraded connection.
At this point, calls will time out and the server must be reprobed.
Signed-off-by: David Howells <dhowells@redhat.com>
Add a tracepoint (rxrpc_rx_rwind_change) to log changes in a call's receive
window size as imposed by the peer through an ACK packet.
Signed-off-by: David Howells <dhowells@redhat.com>
Add a tracepoint (rxrpc_rx_proto) to record protocol errors in received
packets. The following changes are made:
(1) Add a function, __rxrpc_abort_eproto(), to note a protocol error on a
call and mark the call aborted. This is wrapped by
rxrpc_abort_eproto() that makes the why string usable in trace.
(2) Add trace_rxrpc_rx_proto() or rxrpc_abort_eproto() to protocol error
generation points, replacing rxrpc_abort_call() with the latter.
(3) Only send an abort packet in rxkad_verify_packet*() if we actually
managed to abort the call.
Note that a trace event is also emitted if a kernel user (e.g. afs) tries
to send data through a call when it's not in the transmission phase, though
it's not technically a receive event.
Signed-off-by: David Howells <dhowells@redhat.com>
Use negative error codes in struct rxrpc_call::error because that's what
the kernel normally deals with and to make the code consistent. We only
turn them positive when transcribing into a cmsg for userspace recvmsg.
Signed-off-by: David Howells <dhowells@redhat.com>
The RxRPC ACK packet may contain an extension that includes the peer's
current Rx window size for this call. We adjust the local Tx window size
to match. However, the transmitter can stall if the receive window is
reduced to 0 by the peer and then reopened.
This is because the normal way that the transmitter is re-energised is by
dropping something out of our Tx queue and thus making space. When a
single gap is made, the transmitter is woken up. However, because there's
nothing in the Tx queue at this point, this doesn't happen.
To fix this, perform a wake_up() any time we see the peer's Rx window size
increasing.
The observable symptom is that calls start failing on ETIMEDOUT and the
following:
kAFS: SERVER DEAD state=-62
appears in dmesg.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The call state may be changed at any time by the data-ready routine in
response to received packets, so if the call state is to be read and acted
upon several times in a function, READ_ONCE() must be used unless the call
state lock is held.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
All the routines by which rxrpc is accessed from the outside are serialised
by means of the socket lock (sendmsg, recvmsg, bind,
rxrpc_kernel_begin_call(), ...) and this presents a problem:
(1) If a number of calls on the same socket are in the process of
connection to the same peer, a maximum of four concurrent live calls
are permitted before further calls need to wait for a slot.
(2) If a call is waiting for a slot, it is deep inside sendmsg() or
rxrpc_kernel_begin_call() and the entry function is holding the socket
lock.
(3) sendmsg() and recvmsg() or the in-kernel equivalents are prevented
from servicing the other calls as they need to take the socket lock to
do so.
(4) The socket is stuck until a call is aborted and makes its slot
available to the waiter.
Fix this by:
(1) Provide each call with a mutex ('user_mutex') that arbitrates access
by the users of rxrpc separately for each specific call.
(2) Make rxrpc_sendmsg() and rxrpc_recvmsg() unlock the socket as soon as
they've got a call and taken its mutex.
Note that I'm returning EWOULDBLOCK from recvmsg() if MSG_DONTWAIT is
set but someone else has the lock. Should I instead only return
EWOULDBLOCK if there's nothing currently to be done on a socket, and
sleep in this particular instance because there is something to be
done, but we appear to be blocked by the interrupt handler doing its
ping?
(3) Make rxrpc_new_client_call() unlock the socket after allocating a new
call, locking its user mutex and adding it to the socket's call tree.
The call is returned locked so that sendmsg() can add data to it
immediately.
From the moment the call is in the socket tree, it is subject to
access by sendmsg() and recvmsg() - even if it isn't connected yet.
(4) Lock new service calls in the UDP data_ready handler (in
rxrpc_new_incoming_call()) because they may already be in the socket's
tree and the data_ready handler makes them live immediately if a user
ID has already been preassigned.
Note that the new call is locked before any notifications are sent
that it is live, so doing mutex_trylock() *ought* to always succeed.
Userspace is prevented from doing sendmsg() on calls that are in a
too-early state in rxrpc_do_sendmsg().
(5) Make rxrpc_new_incoming_call() return the call with the user mutex
held so that a ping can be scheduled immediately under it.
Note that it might be worth moving the ping call into
rxrpc_new_incoming_call() and then we can drop the mutex there.
(6) Make rxrpc_accept_call() take the lock on the call it is accepting and
release the socket after adding the call to the socket's tree. This
is slightly tricky as we've dequeued the call by that point and have
to requeue it.
Note that requeuing emits a trace event.
(7) Make rxrpc_kernel_send_data() and rxrpc_kernel_recv_data() take the
new mutex immediately and don't bother with the socket mutex at all.
This patch has the nice bonus that calls on the same socket are now to some
extent parallelisable.
Note that we might want to move rxrpc_service_prealloc() calls out from the
socket lock and give it its own lock, so that we don't hang progress in
other calls because we're waiting for the allocator.
We probably also want to avoid calling rxrpc_notify_socket() from within
the socket lock (rxrpc_accept_call()).
Signed-off-by: David Howells <dhowells@redhat.com>
Tested-by: Marc Dionne <marc.c.dionne@auristor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add the following extra tracing information:
(1) Modify the rxrpc_transmit tracepoint to record the Tx window size as
this is varied by the slow-start algorithm.
(2) Modify the rxrpc_rx_ack tracepoint to record more information from
received ACK packets.
(3) Add an rxrpc_rx_data tracepoint to record the information in DATA
packets.
(4) Add an rxrpc_disconnect_call tracepoint to record call disconnection,
including the reason the call was disconnected.
(5) Add an rxrpc_improper_term tracepoint to record implicit termination
of a call by a client either by starting a new call on a particular
connection channel without first transmitting the final ACK for the
previous call.
Signed-off-by: David Howells <dhowells@redhat.com>
Fix the way enum values are translated into strings in AF_RXRPC
tracepoints. The problem with just doing a lookup in a normal flat array
of strings or chars is that external tracing infrastructure can't find it.
Rather, TRACE_DEFINE_ENUM must be used.
Also sort the enums and string tables to make it easier to keep them in
order so that a future patch to __print_symbolic() can be optimised to try
a direct lookup into the table first before iterating over it.
A couple of _proto() macro calls are removed because they refered to tables
that got moved to the tracing infrastructure. The relevant data can be
found by way of tracing.
Signed-off-by: David Howells <dhowells@redhat.com>
A new argument is added to __skb_recv_datagram to provide
an explicit skb destructor, invoked under the receive queue
lock.
The UDP protocol uses such argument to perform memory
reclaiming on dequeue, so that the UDP protocol does not
set anymore skb->desctructor.
Instead explicit memory reclaiming is performed at close() time and
when skbs are removed from the receive queue.
The in kernel UDP protocol users now need to call a
skb_recv_udp() variant instead of skb_recv_datagram() to
properly perform memory accounting on dequeue.
Overall, this allows acquiring only once the receive queue
lock on dequeue.
Tested using pktgen with random src port, 64 bytes packet,
wire-speed on a 10G link as sender and udp_sink as the receiver,
using an l4 tuple rxhash to stress the contention, and one or more
udp_sink instances with reuseport.
nr sinks vanilla patched
1 440 560
3 2150 2300
6 3650 3800
9 4450 4600
12 6250 6450
v1 -> v2:
- do rmem and allocated memory scheduling under the receive lock
- do bulk scheduling in first_packet_length() and in udp_destruct_sock()
- avoid the typdef for the dequeue callback
Suggested-by: Eric Dumazet <edumazet@google.com>
Acked-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
OpenAFS doesn't always correctly terminate client calls that it makes -
this includes calls the OpenAFS servers make to the cache manager service.
It should end the client call with either:
(1) An ACK that has firstPacket set to one greater than the seq number of
the reply DATA packet with the LAST_PACKET flag set (thereby
hard-ACK'ing all packets). nAcks should be 0 and acks[] should be
empty (ie. no soft-ACKs).
(2) An ACKALL packet.
OpenAFS, though, may send an ACK packet with firstPacket set to the last
seq number or less and soft-ACKs listed for all packets up to and including
the last DATA packet.
The transmitter, however, is obliged to keep the call live and the
soft-ACK'd DATA packets around until they're hard-ACK'd as the receiver is
permitted to drop any merely soft-ACK'd packet and request retransmission
by sending an ACK packet with a NACK in it.
Further, OpenAFS will also terminate a client call by beginning the next
client call on the same connection channel. This implicitly completes the
previous call.
This patch handles implicit ACK of a call on a channel by the reception of
the first packet of the next call on that channel.
If another call doesn't come along to implicitly ACK a call, then we have
to time the call out. There are some bugs there that will be addressed in
subsequent patches.
Signed-off-by: David Howells <dhowells@redhat.com>
Separate the output of PING ACKs from the output of other sorts of ACK so
that if we receive a PING ACK and schedule transmission of a PING RESPONSE
ACK, the response doesn't get cancelled by a PING ACK we happen to be
scheduling transmission of at the same time.
If a PING RESPONSE gets lost, the other side might just sit there waiting
for it and refuse to proceed otherwise.
Signed-off-by: David Howells <dhowells@redhat.com>
When a reply is deemed lost, we send a ping to find out the other end
received all the request data packets we sent. This should be limited to
client calls and we shouldn't do this on service calls.
Signed-off-by: David Howells <dhowells@redhat.com>
Keep that call timeouts as ktimes rather than jiffies so that they can be
expressed as functions of RTT.
Signed-off-by: David Howells <dhowells@redhat.com>
When we receive an ACK from the peer that tells us what the peer's receive
window (rwind) is, we should reduce ssthresh to rwind if rwind is smaller
than ssthresh.
Signed-off-by: David Howells <dhowells@redhat.com>
Switch to Congestion Avoidance mode at cwnd == ssthresh rather than relying
on cwnd getting incremented beyond ssthresh and the window size, the mode
being shifted and then cwnd being corrected.
We need to make sure we switch into CA mode so that we stop marking every
packet for ACK.
Signed-off-by: David Howells <dhowells@redhat.com>
Note the serial number of the packet being ACK'd in the congestion
management trace rather than the serial number of the ACK packet. Whilst
the serial number of the ACK packet is useful for matching ACK packet in
the output of wireshark, the serial number that the ACK is in response to
is of more use in working out how different trace lines relate.
Signed-off-by: David Howells <dhowells@redhat.com>
Implement RxRPC slow-start, which is similar to RFC 5681 for TCP. A
tracepoint is added to log the state of the congestion management algorithm
and the decisions it makes.
Notes:
(1) Since we send fixed-size DATA packets (apart from the final packet in
each phase), counters and calculations are in terms of packets rather
than bytes.
(2) The ACK packet carries the equivalent of TCP SACK.
(3) The FLIGHT_SIZE calculation in RFC 5681 doesn't seem particularly
suited to SACK of a small number of packets. It seems that, almost
inevitably, by the time three 'duplicate' ACKs have been seen, we have
narrowed the loss down to one or two missing packets, and the
FLIGHT_SIZE calculation ends up as 2.
(4) In rxrpc_resend(), if there was no data that apparently needed
retransmission, we transmit a PING ACK to ask the peer to tell us what
its Rx window state is.
Signed-off-by: David Howells <dhowells@redhat.com>
If we've sent all the request data in a client call but haven't seen any
sign of the reply data yet, schedule an ACK to be sent to the server to
find out if the reply data got lost.
If the server hasn't yet hard-ACK'd the request data, we send a PING ACK to
demand a response to find out whether we need to retransmit.
If the server says it has received all of the data, we send an IDLE ACK to
tell the server that we haven't received anything in the receive phase as
yet.
To make this work, a non-immediate PING ACK must carry a delay. I've chosen
the same as the IDLE ACK for the moment.
Signed-off-by: David Howells <dhowells@redhat.com>
Generate a summary of the Tx buffer packet state when an ACK is received
for use in a later patch that does congestion management.
Signed-off-by: David Howells <dhowells@redhat.com>
Clear the ACK reason, ACK timer and resend timer when entering the client
reply phase when the first DATA packet is received. New ACKs will be
proposed once the data is queued.
The resend timer is no longer relevant and we need to cancel ACKs scheduled
to probe for a lost reply.
Signed-off-by: David Howells <dhowells@redhat.com>
Send an immediate ACK if we fill in a hole in the buffer left by an
out-of-sequence packet. This may allow the congestion management in the peer
to avoid a retransmission if packets got reordered on the wire.
Signed-off-by: David Howells <dhowells@redhat.com>
Add a tracepoint to log proposed ACKs, including whether the proposal is
used to update a pending ACK or is discarded in favour of an easlier,
higher priority ACK.
Whilst we're at it, get rid of the rxrpc_acks() function and access the
name array directly. We do, however, need to validate the ACK reason
number given to trace_rxrpc_rx_ack() to make sure we don't overrun the
array.
Signed-off-by: David Howells <dhowells@redhat.com>
When the last packet of data to be transmitted on a call is queued, tx_top
is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this
leaves a race in the ACK processing side of things because the flag affects
the interpretation of tx_top and also allows us to start receiving reply
data before we've finished transmitting.
To fix this, make the following changes:
(1) rxrpc_queue_packet() now sets a marker in the annotation buffer
instead of setting the RXRPC_CALL_TX_LAST flag.
(2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the
same context as the routines that use it.
(3) rxrpc_end_tx_phase() is simplified to just shift the call state.
The Tx window must have been rotated before calling to discard the
last packet.
(4) rxrpc_receiving_reply() is added to handle the arrival of the first
DATA packet of a reply to a client call (which is an implicit ACK of
the Tx phase).
(5) The last part of rxrpc_input_ack() is reordered to perform Tx
rotation, then soft-ACK application and then to end the phase if we've
rotated the last packet. In the event of a terminal ACK, the soft-ACK
application will be skipped as nAcks should be 0.
(6) rxrpc_input_ackall() now has to rotate as well as ending the phase.
In addition:
(7) Alter the transmit tracepoint to log the rotation of the last packet.
(8) Remove the no-longer relevant queue_reqack tracepoint note. The
ACK-REQUESTED packet header flag is now set as needed when we actually
transmit the packet and may vary by retransmission.
Signed-off-by: David Howells <dhowells@redhat.com>
When rxrpc_input_soft_acks() is parsing the soft-ACKs from an ACK packet,
it updates the Tx packet annotations in the annotation buffer. If a
soft-ACK is an ACK, then we overwrite unack'd, nak'd or to-be-retransmitted
states and that is fine; but if the soft-ACK is an NACK, we overwrite the
to-be-retransmitted with a nak - which isn't.
Instead, we need to let any scheduled retransmission stand if the packet
was NAK'd.
Note that we don't reissue a resend if the annotation is in the
to-be-retransmitted state because someone else must've scheduled the
resend already.
Signed-off-by: David Howells <dhowells@redhat.com>
before_eq() and friends should be used to compare serial numbers (when not
checking for (non)equality) rather than casting to int, subtracting and
checking the result.
Signed-off-by: David Howells <dhowells@redhat.com>
We don't want to send a PING ACK for every new incoming call as that just
adds to the network traffic. Instead, we send a PING ACK to the first
three that we receive and then once per second thereafter.
This could probably be made adjustable in future.
Signed-off-by: David Howells <dhowells@redhat.com>
In addition to sending a PING ACK to gain RTT data, we can set the
RXRPC_REQUEST_ACK flag on a DATA packet and get a REQUESTED-ACK ACK. The
ACK packet contains the serial number of the packet it is in response to,
so we can look through the Tx buffer for a matching DATA packet.
This requires that the data packets be stamped with the time of
transmission as a ktime rather than having the resend_at time in jiffies.
This further requires the resend code to do the resend determination in
ktimes and convert to jiffies to set the timer.
Signed-off-by: David Howells <dhowells@redhat.com>
Send a PING ACK packet to the peer when we get a new incoming call from a
peer we don't have a record for. The PING RESPONSE ACK packet will tell us
the following about the peer:
(1) its receive window size
(2) its MTU sizes
(3) its support for jumbo DATA packets
(4) if it supports slow start (similar to RFC 5681)
(5) an estimate of the RTT
This is necessary because the peer won't normally send us an ACK until it
gets to the Rx phase and we send it a packet, but we would like to know
some of this information before we start sending packets.
A pair of tracepoints are added so that RTT determination can be observed.
Signed-off-by: David Howells <dhowells@redhat.com>
Add a Tx-phase annotation for packet buffers to indicate that a buffer has
already been retransmitted. This will be used by future congestion
management. Re-retransmissions of a packet don't affect the congestion
window managment in the same way as initial retransmissions.
Signed-off-by: David Howells <dhowells@redhat.com>
Add a configuration option to inject packet loss by discarding
approximately every 8th packet received and approximately every 8th DATA
packet transmitted.
Note that no locking is used, but it shouldn't really matter.
Signed-off-by: David Howells <dhowells@redhat.com>
Improve sk_buff tracing within AF_RXRPC by the following means:
(1) Use an enum to note the event type rather than plain integers and use
an array of event names rather than a big multi ?: list.
(2) Distinguish Rx from Tx packets and account them separately. This
requires the call phase to be tracked so that we know what we might
find in rxtx_buffer[].
(3) Add a parameter to rxrpc_{new,see,get,free}_skb() to indicate the
event type.
(4) A pair of 'rotate' events are added to indicate packets that are about
to be rotated out of the Rx and Tx windows.
(5) A pair of 'lost' events are added, along with rxrpc_lose_skb() for
packet loss injection recording.
Signed-off-by: David Howells <dhowells@redhat.com>
Add a tracepoint to follow the insertion of a packet into the transmit
buffer, its transmission and its rotation out of the buffer.
Signed-off-by: David Howells <dhowells@redhat.com>
The soft-ACK parser doesn't increment the pointer into the soft-ACK list,
resulting in the first ACK/NACK value being applied to all the relevant
packets in the Tx queue. This has the potential to miss retransmissions
and cause excessive retransmissions.
Fix this by incrementing the pointer.
Signed-off-by: David Howells <dhowells@redhat.com>
The code for determining the last packet in rxrpc_recvmsg_data() has been
using the RXRPC_CALL_RX_LAST flag to determine if the rx_top pointer points
to the last packet or not. This isn't a good idea, however, as the input
code may be running simultaneously on another CPU and that sets the flag
*before* updating the top pointer.
Fix this by the following means:
(1) Restrict the use of RXRPC_CALL_RX_LAST to the input routines only.
There's otherwise a synchronisation problem between detecting the flag
and checking tx_top. This could probably be dealt with by appropriate
application of memory barriers, but there's a simpler way.
(2) Set RXRPC_CALL_RX_LAST after setting rx_top.
(3) Make rxrpc_rotate_rx_window() consult the flags header field of the
DATA packet it's about to discard to see if that was the last packet.
Use this as the basis for ending the Rx phase. This shouldn't be a
problem because the recvmsg side of things is guaranteed to see the
packets in order.
(4) Make rxrpc_recvmsg_data() return 1 to indicate the end of the data if:
(a) the packet it has just processed is marked as RXRPC_LAST_PACKET
(b) the call's Rx phase has been ended.
Signed-off-by: David Howells <dhowells@redhat.com>
call->rx_winsize should be initialised to the sysctl setting and the sysctl
setting should be limited to the maximum we want to permit. Further, we
need to place this in the ACK info instead of the sysctl setting.
Furthermore, discard the idea of accepting the subpackets of a jumbo packet
that lie beyond the receive window when the first packet of the jumbo is
within the window. Just discard the excess subpackets instead. This
allows the receive window to be opened up right to the buffer size less one
for the dead slot.
Signed-off-by: David Howells <dhowells@redhat.com>
Allow tx_winsize to grow when the ACK info packet shows a larger receive
window at the other end rather than only permitting it to shrink.
Signed-off-by: David Howells <dhowells@redhat.com>
skb->len should be used rather than skb->data_len when referring to the
amount of data in a packet. This will only cause a malfunction in the
following cases:
(1) We receive a jumbo packet (validation and splitting both are wrong).
(2) We see if there's extra ACK info in an ACK packet (we think it's not
there and just ignore it).
Signed-off-by: David Howells <dhowells@redhat.com>
We need to wake up the sender when Tx window rotation due to an incoming
ACK makes space in the buffer otherwise the sender is liable to just hang
endlessly.
This problem isn't noticeable if the Tx phase transfers no more than will
fit in a single window or the Tx window rotates fast enough that it doesn't
get full.
Signed-off-by: David Howells <dhowells@redhat.com>
Rewrite the data and ack handling code such that:
(1) Parsing of received ACK and ABORT packets and the distribution and the
filing of DATA packets happens entirely within the data_ready context
called from the UDP socket. This allows us to process and discard ACK
and ABORT packets much more quickly (they're no longer stashed on a
queue for a background thread to process).
(2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead
keep track of the offset and length of the content of each packet in
the sk_buff metadata. This means we don't do any allocation in the
receive path.
(3) Jumbo DATA packet parsing is now done in data_ready context. Rather
than cloning the packet once for each subpacket and pulling/trimming
it, we file the packet multiple times with an annotation for each
indicating which subpacket is there. From that we can directly
calculate the offset and length.
(4) A call's receive queue can be accessed without taking locks (memory
barriers do have to be used, though).
(5) Incoming calls are set up from preallocated resources and immediately
made live. They can than have packets queued upon them and ACKs
generated. If insufficient resources exist, DATA packet #1 is given a
BUSY reply and other DATA packets are discarded).
(6) sk_buffs no longer take a ref on their parent call.
To make this work, the following changes are made:
(1) Each call's receive buffer is now a circular buffer of sk_buff
pointers (rxtx_buffer) rather than a number of sk_buff_heads spread
between the call and the socket. This permits each sk_buff to be in
the buffer multiple times. The receive buffer is reused for the
transmit buffer.
(2) A circular buffer of annotations (rxtx_annotations) is kept parallel
to the data buffer. Transmission phase annotations indicate whether a
buffered packet has been ACK'd or not and whether it needs
retransmission.
Receive phase annotations indicate whether a slot holds a whole packet
or a jumbo subpacket and, if the latter, which subpacket. They also
note whether the packet has been decrypted in place.
(3) DATA packet window tracking is much simplified. Each phase has just
two numbers representing the window (rx_hard_ack/rx_top and
tx_hard_ack/tx_top).
The hard_ack number is the sequence number before base of the window,
representing the last packet the other side says it has consumed.
hard_ack starts from 0 and the first packet is sequence number 1.
The top number is the sequence number of the highest-numbered packet
residing in the buffer. Packets between hard_ack+1 and top are
soft-ACK'd to indicate they've been received, but not yet consumed.
Four macros, before(), before_eq(), after() and after_eq() are added
to compare sequence numbers within the window. This allows for the
top of the window to wrap when the hard-ack sequence number gets close
to the limit.
Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also
to indicate when rx_top and tx_top point at the packets with the
LAST_PACKET bit set, indicating the end of the phase.
(4) Calls are queued on the socket 'receive queue' rather than packets.
This means that we don't need have to invent dummy packets to queue to
indicate abnormal/terminal states and we don't have to keep metadata
packets (such as ABORTs) around
(5) The offset and length of a (sub)packet's content are now passed to
the verify_packet security op. This is currently expected to decrypt
the packet in place and validate it.
However, there's now nowhere to store the revised offset and length of
the actual data within the decrypted blob (there may be a header and
padding to skip) because an sk_buff may represent multiple packets, so
a locate_data security op is added to retrieve these details from the
sk_buff content when needed.
(6) recvmsg() now has to handle jumbo subpackets, where each subpacket is
individually secured and needs to be individually decrypted. The code
to do this is broken out into rxrpc_recvmsg_data() and shared with the
kernel API. It now iterates over the call's receive buffer rather
than walking the socket receive queue.
Additional changes:
(1) The timers are condensed to a single timer that is set for the soonest
of three timeouts (delayed ACK generation, DATA retransmission and
call lifespan).
(2) Transmission of ACK and ABORT packets is effected immediately from
process-context socket ops/kernel API calls that cause them instead of
them being punted off to a background work item. The data_ready
handler still has to defer to the background, though.
(3) A shutdown op is added to the AF_RXRPC socket so that the AFS
filesystem can shut down the socket and flush its own work items
before closing the socket to deal with any in-progress service calls.
Future additional changes that will need to be considered:
(1) Make sure that a call doesn't hog the front of the queue by receiving
data from the network as fast as userspace is consuming it to the
exclusion of other calls.
(2) Transmit delayed ACKs from within recvmsg() when we've consumed
sufficiently more packets to avoid the background work item needing to
run.
Signed-off-by: David Howells <dhowells@redhat.com>
Make it possible for the data_ready handler called from the UDP transport
socket to completely instantiate an rxrpc_call structure and make it
immediately live by preallocating all the memory it might need. The idea
is to cut out the background thread usage as much as possible.
[Note that the preallocated structs are not actually used in this patch -
that will be done in a future patch.]
If insufficient resources are available in the preallocation buffers, it
will be possible to discard the DATA packet in the data_ready handler or
schedule a BUSY packet without the need to schedule an attempt at
allocation in a background thread.
To this end:
(1) Preallocate rxrpc_peer, rxrpc_connection and rxrpc_call structs to a
maximum number each of the listen backlog size. The backlog size is
limited to a maxmimum of 32. Only this many of each can be in the
preallocation buffer.
(2) For userspace sockets, the preallocation is charged initially by
listen() and will be recharged by accepting or rejecting pending
new incoming calls.
(3) For kernel services {,re,dis}charging of the preallocation buffers is
handled manually. Two notifier callbacks have to be provided before
kernel_listen() is invoked:
(a) An indication that a new call has been instantiated. This can be
used to trigger background recharging.
(b) An indication that a call is being discarded. This is used when
the socket is being released.
A function, rxrpc_kernel_charge_accept() is called by the kernel
service to preallocate a single call. It should be passed the user ID
to be used for that call and a callback to associate the rxrpc call
with the kernel service's side of the ID.
(4) Discard the preallocation when the socket is closed.
(5) Temporarily bump the refcount on the call allocated in
rxrpc_incoming_call() so that rxrpc_release_call() can ditch the
preallocation ref on service calls unconditionally. This will no
longer be necessary once the preallocation is used.
Note that this does not yet control the number of active service calls on a
client - that will come in a later patch.
A future development would be to provide a setsockopt() call that allows a
userspace server to manually charge the preallocation buffer. This would
allow user call IDs to be provided in advance and the awkward manual accept
stage to be bypassed.
Signed-off-by: David Howells <dhowells@redhat.com>
Add two tracepoints:
(1) Record the RxRPC protocol header of packets retrieved from the UDP
socket by the data_ready handler.
(2) Record the outcome of the data_ready handler.
Signed-off-by: David Howells <dhowells@redhat.com>
David Howells