Merge tag 'rxrpc-next-
20221201-b' of git://git./linux/kernel/git/dhowells/linux-fs
David Howells says:
====================
rxrpc: Increasing SACK size and moving away from softirq, parts 2 & 3
Here are the second and third parts of patches in the process of moving
rxrpc from doing a lot of its stuff in softirq context to doing it in an
I/O thread in process context and thereby making it easier to support a
larger SACK table.
The full description is in the description for the first part[1] which is
already in net-next.
The second part includes some cleanups, adds some testing and overhauls
some tracing:
(1) Remove declaration of rxrpc_kernel_call_is_complete() as the
definition is no longer present.
(2) Remove the knet() and kproto() macros in favour of using tracepoints.
(3) Remove handling of duplicate packets from recvmsg. The input side
isn't now going to insert overlapping/duplicate packets into the
recvmsg queue.
(4) Don't use the rxrpc_conn_parameters struct in the rxrpc_connection or
rxrpc_bundle structs - rather put the members in directly.
(5) Extract the abort code from a received abort packet right up front
rather than doing it in multiple places later.
(6) Use enums and symbol lists rather than __builtin_return_address() to
indicate where a tracepoint was triggered for local, peer, conn, call
and skbuff tracing.
(7) Add a refcount tracepoint for the rxrpc_bundle struct.
(8) Implement an in-kernel server for the AFS rxperf testing program to
talk to (enabled by a Kconfig option).
This is tagged as rxrpc-next-
20221201-a.
The third part introduces the I/O thread and switches various bits over to
running there:
(1) Fix call timers and call and connection workqueues to not hold refs on
the rxrpc_call and rxrpc_connection structs to thereby avoid messy
cleanup when the last ref is put in softirq mode.
(2) Split input.c so that the call packet processing bits are separate
from the received packet distribution bits. Call packet processing
gets bumped over to the call event handler.
(3) Create a per-local endpoint I/O thread. Barring some tiny bits that
still get done in softirq context, all packet reception, processing
and transmission is done in this thread. That will allow a load of
locking to be removed.
(4) Perform packet processing and error processing from the I/O thread.
(5) Provide a mechanism to process call event notifications in the I/O
thread rather than queuing a work item for that call.
(6) Move data and ACK transmission into the I/O thread. ACKs can then be
transmitted at the point they're generated rather than getting
delegated from softirq context to some process context somewhere.
(7) Move call and local processor event handling into the I/O thread.
(8) Move cwnd degradation to after packets have been transmitted so that
they don't shorten the window too quickly.
A bunch of simplifications can then be done:
(1) The input_lock is no longer necessary as exclusion is achieved by
running the code in the I/O thread only.
(2) Don't need to use sk->sk_receive_queue.lock to guard socket state
changes as the socket mutex should suffice.
(3) Don't take spinlocks in RCU callback functions as they get run in
softirq context and thus need _bh annotations.
(4) RCU is then no longer needed for the peer's error_targets list.
(5) Simplify the skbuff handling in the receive path by dropping the ref
in the basic I/O thread loop and getting an extra ref as and when we
need to queue the packet for recvmsg or another context.
(6) Get the peer address earlier in the input process and pass it to the
users so that we only do it once.
This is tagged as rxrpc-next-
20221201-b.
Changes:
========
ver #2)
- Added a patch to change four assertions into warnings in rxrpc_read()
and fixed a checker warning from a __user annotation that should have
been removed..
- Change a min() to min_t() in rxperf as PAGE_SIZE doesn't seem to match
type size_t on i386.
- Three error handling issues in rxrpc_new_incoming_call():
- If not DATA or not seq #1, should drop the packet, not abort.
- Fix a goto that went to the wrong place, dropping a non-held lock.
- Fix an rcu_read_lock that should've been an unlock.
Tested-by: Marc Dionne <marc.dionne@auristor.com>
Tested-by: kafs-testing+fedora36_64checkkafs-build-144@auristor.com
Link: https://lore.kernel.org/r/166794587113.2389296.16484814996876530222.stgit@warthog.procyon.org.uk/
Link: https://lore.kernel.org/r/166982725699.621383.2358362793992993374.stgit@warthog.procyon.org.uk/
====================
Signed-off-by: David S. Miller <davem@davemloft.net>