The rxrpc_input_packet() function and its call tree was built around the
assumption that data_ready() handler called from UDP to inform a kernel
service that there is data to be had was non-reentrant. This means that
certain locking could be dispensed with.
This, however, turns out not to be the case with a multi-queue network card
that can deliver packets to multiple cpus simultaneously. Each of those
cpus can be in the rxrpc_input_packet() function at the same time.
Fix by adding or changing some structure members:
(1) Add peer->rtt_input_lock to serialise access to the RTT buffer.
(2) Make conn->service_id into a 32-bit variable so that it can be
cmpxchg'd on all arches.
(3) Add call->input_lock to serialise access to the Rx/Tx state. Note
that although the Rx and Tx states are (almost) entirely separate,
there's no point completing the separation and having separate locks
since it's a bi-phasal RPC protocol rather than a bi-direction
streaming protocol. Data transmission and data reception do not take
place simultaneously on any particular call.
and making the following functional changes:
(1) In rxrpc_input_data(), hold call->input_lock around the core to
prevent simultaneous producing of packets into the Rx ring and
updating of tracking state for a particular call.
(2) In rxrpc_input_ping_response(), only read call->ping_serial once, and
check it before checking RXRPC_CALL_PINGING as that's a cheaper test.
The bit test and bit clear can then be combined. No further locking
is needed here.
(3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of
the ACK packet. The superseded ACK check is then done both before and
after the lock is taken.
The handing of ackinfo data is split, parsing before the lock is taken
and processing with it held. This is keyed on rxMTU being non-zero.
Congestion management is also done within the locked section.
(4) In rxrpc_input_ackall(), take call->input_lock around the Tx window
rotation. The ACKALL packet carries no information and is only really
useful after all packets have been transmitted since it's imprecise.
(5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to
prevent calls being simultaneously implicitly ended on two cpus and
also to prevent any races with incoming call setup.
(6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade
on a connection. It is only permitted to happen once for a
connection.
(7) In rxrpc_new_incoming_call(), we have to recheck the routing inside
rx->incoming_lock to see if someone else set up the call, connection
or peer whilst we were getting there. We can't trust the values from
the earlier routing check unless we pin refs on them - which we want
to avoid.
Further, we need to allow for an incoming call to have its state
changed on another CPU between us making it live and us adjusting it
because the conn is now in the RXRPC_CONN_SERVICE state.
(8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access
to the RTT buffer. Don't need to lock around setting peer->rtt.
For reference, the inventory of state-accessing or state-altering functions
used by the packet input procedure is:
> rxrpc_input_packet()
* PACKET CHECKING
* ROUTING
> rxrpc_post_packet_to_local()
> rxrpc_find_connection_rcu() - uses RCU
> rxrpc_lookup_peer_rcu() - uses RCU
> rxrpc_find_service_conn_rcu() - uses RCU
> idr_find() - uses RCU
* CONNECTION-LEVEL PROCESSING
- Service upgrade
- Can only happen once per conn
! Changed to use cmpxchg
> rxrpc_post_packet_to_conn()
- Setting conn->hi_serial
- Probably safe not using locks
- Maybe use cmpxchg
* CALL-LEVEL PROCESSING
> Old-call checking
> rxrpc_input_implicit_end_call()
> rxrpc_call_completed()
> rxrpc_queue_call()
! Need to take rx->incoming_lock
> __rxrpc_disconnect_call()
> rxrpc_notify_socket()
> rxrpc_new_incoming_call()
- Uses rx->incoming_lock for the entire process
- Might be able to drop this earlier in favour of the call lock
> rxrpc_incoming_call()
! Conflicts with rxrpc_input_implicit_end_call()
> rxrpc_send_ping()
- Don't need locks to check rtt state
> rxrpc_propose_ACK
* PACKET DISTRIBUTION
> rxrpc_input_call_packet()
> rxrpc_input_data()
* QUEUE DATA PACKET ON CALL
> rxrpc_reduce_call_timer()
- Uses timer_reduce()
! Needs call->input_lock()
> rxrpc_receiving_reply()
! Needs locking around ack state
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_proto_abort()
> rxrpc_input_dup_data()
- Fills the Rx buffer
- rxrpc_propose_ACK()
- rxrpc_notify_socket()
> rxrpc_input_ack()
* APPLY ACK PACKET TO CALL AND DISCARD PACKET
> rxrpc_input_ping_response()
- Probably doesn't need any extra locking
! Need READ_ONCE() on call->ping_serial
> rxrpc_input_check_for_lost_ack()
- Takes call->lock to consult Tx buffer
> rxrpc_peer_add_rtt()
! Needs to take a lock (peer->rtt_input_lock)
! Could perhaps manage with cmpxchg() and xadd() instead
> rxrpc_input_requested_ack
- Consults Tx buffer
! Probably needs a lock
> rxrpc_peer_add_rtt()
> rxrpc_propose_ack()
> rxrpc_input_ackinfo()
- Changes call->tx_winsize
! Use cmpxchg to handle change
! Should perhaps track serial number
- Uses peer->lock to record MTU specification changes
> rxrpc_proto_abort()
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_input_soft_acks()
- Consults the Tx buffer
> rxrpc_congestion_management()
- Modifies the Tx annotations
! Needs call->input_lock()
> rxrpc_queue_call()
> rxrpc_input_abort()
* APPLY ABORT PACKET TO CALL AND DISCARD PACKET
> rxrpc_set_call_completion()
> rxrpc_notify_socket()
> rxrpc_input_ackall()
* APPLY ACKALL PACKET TO CALL AND DISCARD PACKET
! Need to take call->input_lock
> rxrpc_rotate_tx_window()
> rxrpc_end_tx_phase()
> rxrpc_reject_packet()
There are some functions used by the above that queue the packet, after
which the procedure is terminated:
- rxrpc_post_packet_to_local()
- local->event_queue is an sk_buff_head
- local->processor is a work_struct
- rxrpc_post_packet_to_conn()
- conn->rx_queue is an sk_buff_head
- conn->processor is a work_struct
- rxrpc_reject_packet()
- local->reject_queue is an sk_buff_head
- local->processor is a work_struct
And some that offload processing to process context:
- rxrpc_notify_socket()
- Uses RCU lock
- Uses call->notify_lock to call call->notify_rx
- Uses call->recvmsg_lock to queue recvmsg side
- rxrpc_queue_call()
- call->processor is a work_struct
- rxrpc_propose_ACK()
- Uses call->lock to wrap __rxrpc_propose_ACK()
And a bunch that complete a call, all of which use call->state_lock to
protect the call state:
- rxrpc_call_completed()
- rxrpc_set_call_completion()
- rxrpc_abort_call()
- rxrpc_proto_abort()
- Also uses rxrpc_queue_call()
Fixes: 17926a7932 ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both")
Signed-off-by: David Howells <dhowells@redhat.com>
c1e15b4944