Re: [PATCH 2/2] epoll: atomically remove wait entry on wake up
On 4/30/20 9:03 AM, Roman Penyaev wrote:
> This patch does two things:
>
> 1. fixes lost wakeup introduced by:
> 339ddb53d373 ("fs/epoll: remove unnecessary wakeups of nested epoll")
>
> 2. improves performance for events delivery.
>
> The description of the problem is the following: if N (>1) threads
> are waiting on ep->wq for new events and M (>1) events come, it is
> quite likely that >1 wakeups hit the same wait queue entry, because
> there is quite a big window between __add_wait_queue_exclusive() and
> the following __remove_wait_queue() calls in ep_poll() function. This
> can lead to lost wakeups, because thread, which was woken up, can
> handle not all the events in ->rdllist. (in better words the problem
> is described here: https://lkml.org/lkml/2019/10/7/905)
>
> The idea of the current patch is to use init_wait() instead of
> init_waitqueue_entry(). Internally init_wait() sets
> autoremove_wake_function as a callback, which removes the wait entry
> atomically (under the wq locks) from the list, thus the next coming
> wakeup hits the next wait entry in the wait queue, thus preventing
> lost wakeups.
>
> Problem is very well reproduced by the epoll60 test case [1].
>
> Wait entry removal on wakeup has also performance benefits, because
> there is no need to take a ep->lock and remove wait entry from the
> queue after the successful wakeup. Here is the timing output of
> the epoll60 test case:
>
> With explicit wakeup from ep_scan_ready_list() (the state of the
> code prior 339ddb53d373):
>
> real0m6.970s
> user0m49.786s
> sys 0m0.113s
>
> After this patch:
>
>real0m5.220s
>user0m36.879s
>sys 0m0.019s
>
> The other testcase is the stress-epoll [2], where one thread consumes
> all the events and other threads produce many events:
>
> With explicit wakeup from ep_scan_ready_list() (the state of the
> code prior 339ddb53d373):
>
> threads events/ms run-time ms
> 8 5427 1474
> 16 6163 2596
> 32 6824 4689
> 64 7060 9064
> 128 699118309
>
> After this patch:
>
> threads events/ms run-time ms
> 8 5598 1429
> 16 7073 2262
> 32 7502 4265
> 64 7640 8376
> 128 763416767
>
> (number of "events/ms" represents event bandwidth, thus higher is
> better; number of "run-time ms" represents overall time spent
> doing the benchmark, thus lower is better)
>
> [1] tools/testing/selftests/filesystems/epoll/epoll_wakeup_test.c
> [2] https://github.com/rouming/test-tools/blob/master/stress-epoll.c
>
> Signed-off-by: Roman Penyaev
> Cc: Andrew Morton
> Cc: Khazhismel Kumykov
> Cc: Alexander Viro
> Cc: Heiher
> Cc: Jason Baron
> Cc: sta...@vger.kernel.org
> Cc: linux-fsde...@vger.kernel.org
> Cc: linux-kernel@vger.kernel.org
> ---
> fs/eventpoll.c | 43 ---
> 1 file changed, 24 insertions(+), 19 deletions(-)
>
Looks good to me and nice speedups.
Reviewed-by: Jason Baron
Thanks,
-Jason
Re: [PATCH 2/2] epoll: atomically remove wait entry on wake up
Hi
[This is an automated email]
This commit has been processed because it contains a -stable tag.
The stable tag indicates that it's relevant for the following trees: all
The bot has tested the following trees: v5.6.7, v5.4.35, v4.19.118, v4.14.177,
v4.9.220, v4.4.220.
v5.6.7: Build OK!
v5.4.35: Build OK!
v4.19.118: Failed to apply! Possible dependencies:
1b53734bd0b2 ("epoll: fix possible lost wakeup on epoll_ctl() path")
35cff1a6e023 ("fs/epoll: rename check_events label to send_events")
86c051793b4c ("fs/epoll: deal with wait_queue only once")
abc610e01c66 ("fs/epoll: avoid barrier after an epoll_wait(2) timeout")
c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
v4.14.177: Failed to apply! Possible dependencies:
002b343669c4 ("fs/epoll: loosen irq safety in ep_scan_ready_list()")
35cff1a6e023 ("fs/epoll: rename check_events label to send_events")
37b5e5212a44 ("epoll: remove ep_call_nested() from ep_eventpoll_poll()")
679abf381a18 ("fs/eventpoll.c: loosen irq safety in ep_poll()")
69112736e2f0 ("eventpoll: no need to mask the result of epi_item_poll()
again")
86c051793b4c ("fs/epoll: deal with wait_queue only once")
bec1a502d34d ("eventpoll: constify struct epoll_event pointers")
c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
d85e2aa2e34d ("annotate ep_scan_ready_list()")
ee8ef0a4b167 ("epoll: use the waitqueue lock to protect ep->wq")
v4.9.220: Failed to apply! Possible dependencies:
86c051793b4c ("fs/epoll: deal with wait_queue only once")
89d947561077 ("sd: Implement support for ZBC devices")
ac6424b981bc ("sched/wait: Rename wait_queue_t => wait_queue_entry_t")
bd166ef183c2 ("blk-mq-sched: add framework for MQ capable IO schedulers")
cf43e6be865a ("block: add scalable completion tracking of requests")
e806402130c9 ("block: split out request-only flags into a new namespace")
v4.4.220: Failed to apply! Possible dependencies:
27489a3c827b ("blk-mq: turn hctx->run_work into a regular work struct")
511cbce2ff8b ("irq_poll: make blk-iopoll available outside the block layer")
86c051793b4c ("fs/epoll: deal with wait_queue only once")
8d354f133e86 ("blk-mq: improve layout of blk_mq_hw_ctx")
9467f85960a3 ("blk-mq/cpu-notif: Convert to new hotplug state machine")
ac6424b981bc ("sched/wait: Rename wait_queue_t => wait_queue_entry_t")
bd166ef183c2 ("blk-mq-sched: add framework for MQ capable IO schedulers")
cf43e6be865a ("block: add scalable completion tracking of requests")
d9d8c5c489f4 ("block: convert is_sync helpers to use REQ_OPs.")
da8b44d5a9f8 ("timer: convert timer_slack_ns from unsigned long to u64")
e57690fe009b ("blk-mq: don't overwrite rq->mq_ctx")
e6a40b096e28 ("block: prepare request creation/destruction code to use
REQ_OPs")
e806402130c9 ("block: split out request-only flags into a new namespace")
NOTE: The patch will not be queued to stable trees until it is upstream.
How should we proceed with this patch?
--
Thanks
Sasha
[PATCH 2/2] epoll: atomically remove wait entry on wake up
This patch does two things:
1. fixes lost wakeup introduced by:
339ddb53d373 ("fs/epoll: remove unnecessary wakeups of nested epoll")
2. improves performance for events delivery.
The description of the problem is the following: if N (>1) threads
are waiting on ep->wq for new events and M (>1) events come, it is
quite likely that >1 wakeups hit the same wait queue entry, because
there is quite a big window between __add_wait_queue_exclusive() and
the following __remove_wait_queue() calls in ep_poll() function. This
can lead to lost wakeups, because thread, which was woken up, can
handle not all the events in ->rdllist. (in better words the problem
is described here: https://lkml.org/lkml/2019/10/7/905)
The idea of the current patch is to use init_wait() instead of
init_waitqueue_entry(). Internally init_wait() sets
autoremove_wake_function as a callback, which removes the wait entry
atomically (under the wq locks) from the list, thus the next coming
wakeup hits the next wait entry in the wait queue, thus preventing
lost wakeups.
Problem is very well reproduced by the epoll60 test case [1].
Wait entry removal on wakeup has also performance benefits, because
there is no need to take a ep->lock and remove wait entry from the
queue after the successful wakeup. Here is the timing output of
the epoll60 test case:
With explicit wakeup from ep_scan_ready_list() (the state of the
code prior 339ddb53d373):
real0m6.970s
user0m49.786s
sys 0m0.113s
After this patch:
real0m5.220s
user0m36.879s
sys 0m0.019s
The other testcase is the stress-epoll [2], where one thread consumes
all the events and other threads produce many events:
With explicit wakeup from ep_scan_ready_list() (the state of the
code prior 339ddb53d373):
threads events/ms run-time ms
8 5427 1474
16 6163 2596
32 6824 4689
64 7060 9064
128 699118309
After this patch:
threads events/ms run-time ms
8 5598 1429
16 7073 2262
32 7502 4265
64 7640 8376
128 763416767
(number of "events/ms" represents event bandwidth, thus higher is
better; number of "run-time ms" represents overall time spent
doing the benchmark, thus lower is better)
[1] tools/testing/selftests/filesystems/epoll/epoll_wakeup_test.c
[2] https://github.com/rouming/test-tools/blob/master/stress-epoll.c
Signed-off-by: Roman Penyaev
Cc: Andrew Morton
Cc: Khazhismel Kumykov
Cc: Alexander Viro
Cc: Heiher
Cc: Jason Baron
Cc: sta...@vger.kernel.org
Cc: linux-fsde...@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
---
fs/eventpoll.c | 43 ---
1 file changed, 24 insertions(+), 19 deletions(-)
diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index d6ba0e52439b..aba03ee749f8 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -1822,7 +1822,6 @@ static int ep_poll(struct eventpoll *ep, struct
epoll_event __user *events,
{
int res = 0, eavail, timed_out = 0;
u64 slack = 0;
- bool waiter = false;
wait_queue_entry_t wait;
ktime_t expires, *to = NULL;
@@ -1867,21 +1866,23 @@ static int ep_poll(struct eventpoll *ep, struct
epoll_event __user *events,
*/
ep_reset_busy_poll_napi_id(ep);
- /*
-* We don't have any available event to return to the caller. We need
-* to sleep here, and we will be woken by ep_poll_callback() when events
-* become available.
-*/
- if (!waiter) {
- waiter = true;
- init_waitqueue_entry(, current);
-
+ do {
+ /*
+* Internally init_wait() uses autoremove_wake_function(),
+* thus wait entry is removed from the wait queue on each
+* wakeup. Why it is important? In case of several waiters
+* each new wakeup will hit the next waiter, giving it the
+* chance to harvest new event. Otherwise wakeup can be
+* lost. This is also good performance-wise, because on
+* normal wakeup path no need to call __remove_wait_queue()
+* explicitly, thus ep->lock is not taken, which halts the
+* event delivery.
+*/
+ init_wait();
write_lock_irq(>lock);
__add_wait_queue_exclusive(>wq, );
write_unlock_irq(>lock);
- }
- for (;;) {
/*
* We don't want to sleep if the ep_poll_callback() sends us
* a wakeup in between. That's why we set the task state
@@ -1911,10 +1912,20 @@ static int ep_poll(struct eventpoll *ep, struct
epoll_event __user *events,
timed_out = 1;
break;
