From 1e1c15839df084f4011825fee922aa976c9159dc Mon Sep 17 00:00:00 2001 From: Nicholas Piggin <npiggin@gmail.com> Date: Tue, 7 Sep 2021 20:00:00 -0700 Subject: fs/epoll: use a per-cpu counter for user's watches count This counter tracks the number of watches a user has, to compare against the 'max_user_watches' limit. This causes a scalability bottleneck on SPECjbb2015 on large systems as there is only one user. Changing to a per-cpu counter increases throughput of the benchmark by about 30% on a 16-socket, > 1000 thread system. [rdunlap@infradead.org: fix build errors in kernel/user.c when CONFIG_EPOLL=n] [npiggin@gmail.com: move ifdefs into wrapper functions, slightly improve panic message] Link: https://lkml.kernel.org/r/1628051945.fens3r99ox.astroid@bobo.none [akpm@linux-foundation.org: tweak user_epoll_alloc(), per Guenter] Link: https://lkml.kernel.org/r/20210804191421.GA1900577@roeck-us.net Link: https://lkml.kernel.org/r/20210802032013.2751916-1-npiggin@gmail.com Signed-off-by: Nicholas Piggin <npiggin@gmail.com> Reported-by: Anton Blanchard <anton@ozlabs.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> --- fs/eventpoll.c | 18 ++++++++++-------- 1 file changed, 10 insertions(+), 8 deletions(-) (limited to 'fs') diff --git a/fs/eventpoll.c b/fs/eventpoll.c index 1e596e1d0bba..648ed77f4164 100644 --- a/fs/eventpoll.c +++ b/fs/eventpoll.c @@ -723,7 +723,7 @@ static int ep_remove(struct eventpoll *ep, struct epitem *epi) */ call_rcu(&epi->rcu, epi_rcu_free); - atomic_long_dec(&ep->user->epoll_watches); + percpu_counter_dec(&ep->user->epoll_watches); return 0; } @@ -1439,7 +1439,6 @@ static int ep_insert(struct eventpoll *ep, const struct epoll_event *event, { int error, pwake = 0; __poll_t revents; - long user_watches; struct epitem *epi; struct ep_pqueue epq; struct eventpoll *tep = NULL; @@ -1449,11 +1448,15 @@ static int ep_insert(struct eventpoll *ep, const struct epoll_event *event, lockdep_assert_irqs_enabled(); - user_watches = atomic_long_read(&ep->user->epoll_watches); - if (unlikely(user_watches >= max_user_watches)) + if (unlikely(percpu_counter_compare(&ep->user->epoll_watches, + max_user_watches) >= 0)) return -ENOSPC; - if (!(epi = kmem_cache_zalloc(epi_cache, GFP_KERNEL))) + percpu_counter_inc(&ep->user->epoll_watches); + + if (!(epi = kmem_cache_zalloc(epi_cache, GFP_KERNEL))) { + percpu_counter_dec(&ep->user->epoll_watches); return -ENOMEM; + } /* Item initialization follow here ... */ INIT_LIST_HEAD(&epi->rdllink); @@ -1466,17 +1469,16 @@ static int ep_insert(struct eventpoll *ep, const struct epoll_event *event, mutex_lock_nested(&tep->mtx, 1); /* Add the current item to the list of active epoll hook for this file */ if (unlikely(attach_epitem(tfile, epi) < 0)) { - kmem_cache_free(epi_cache, epi); if (tep) mutex_unlock(&tep->mtx); + kmem_cache_free(epi_cache, epi); + percpu_counter_dec(&ep->user->epoll_watches); return -ENOMEM; } if (full_check && !tep) list_file(tfile); - atomic_long_inc(&ep->user->epoll_watches); - /* * Add the current item to the RB tree. All RB tree operations are * protected by "mtx", and ep_insert() is called with "mtx" held. -- cgit v1.2.3