[00/10] SUNRPC: remainder of srv queueing work

Message

NeilBrown Aug. 15, 2023, 1:54 a.m. UTC

This series is on top of topic-sunrpc-thread-scheduling (4fa42add1a7)

It includes a fixup for an earlier patch in that topic, revised versions
of patches that have been sent but are yet to land, and the remainder of
what I have been working on.

With the full set enqueueing and dequeueing of the list of idle threads
is completely lockless following a LIFO discipline.  Threads are woken
gradually without the possibility of a thundering herd.  If lots of work
becomes ready only one thread is woken and once it dequeues the first
work item, it will dequeue another thread if there is more work to do.

Queueing and dequeueing for work-items for these threads is partially
lockless following a FIFO discipline.  Enqueueing, which can happen in
bh context, is lockless. Dequeueing in the server threads still
requires a spinlock, but does not need to disable 'bh'.

NeilBrown

 [PATCH 01/10] SQUASH: SUNRPC: rename and refactor svc_get_next_xprt()
 [PATCH 02/10] SUNRPC: add list of idle threads
 [PATCH 03/10] SUNRPC: discard SP_CONGESTED
 [PATCH 04/10] SUNRPC: change service idle list to be an llist
 [PATCH 05/10] SUNRPC: only have one thread waking up at a time
 [PATCH 06/10] SUNRPC/svc: add light-weight queuing mechanism.
 [PATCH 07/10] SUNRPC: use lwq for sp_sockets - renamed to sp_xprts
 [PATCH 08/10] SUNRPC: change sp_nrthreads to atomic_t
 [PATCH 09/10] SUNRPC: discard sp_lock
 [PATCH 10/10] SUNRPC: change the back-channel queue to lwq