fs/epoll: Remove unnecessary wakeups of nested epoll that in ET mode
PATCH RESEND v4 1/1 fs/epoll: Remove unnecessary wakeups of nested epoll that in ET mode
From: Heiher <[email protected]>
Take the case where we have:
t0
| (ew)
e0
| (et)
e1
| (lt)
s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew: epoll_wait
et: edge-trigger
lt: level-trigger
We only need to wakeup nested epoll fds if something has been queued to the overflow list, since the ep_poll() traverses the rdllist during recursive poll and thus events on the overflow list may not be visible yet.
Test code:
#include <unistd.h>
#include <sys/epoll.h>
#include <sys/socket.h>
int main(int argc, char *argv[])
{
int sfd[2];
int efd[2];
struct epoll_event e;
if (socketpair(AF_UNIX, SOCK_STREAM, 0, sfd) < 0)
goto out;
efd[0] = epoll_create(1);
if (efd[0] < 0)
goto out;
efd[1] = epoll_create(1);
if (efd[1] < 0)
goto out;
e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, sfd[0], &e) < 0)
goto out;
e.events = EPOLLIN | EPOLLET;
if (epoll_ctl(efd[0], EPOLL_CTL_ADD, efd[1], &e) < 0)
goto out;
if (write(sfd[1], "w", 1) != 1)
goto out;
if (epoll_wait(efd[0], &e, 1, 0) != 1)
goto out;
if (epoll_wait(efd[0], &e, 1, 0) != 0)
goto out;
close(efd[0]);
close(efd[1]);
close(sfd[0]);
close(sfd[1]);
return 0;
out:
return -1;
}
More tests:
https://github.com/heiher/epoll-wakeup
--- fs/eventpoll.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)
fs/eventpoll.c +11 -2
@@ -704,12 +704,21 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep, res = (*sproc)(ep, &txlist, priv);write_lock_irq (&ep->lock); + nepi =READ_ONCE (ep->ovflist); +/* +* We only need to wakeup nested epoll fds if something has been queued +* to the overflow list, since the ep_poll() traverses the rdllist +* during recursive poll and thus events on the overflow list may not be +* visible yet. +*/ +if (nepi !=NULL ) + pwake++;/* * During the time we spent inside the "sproc" callback, some * other events might have been queued by the poll callback. * We re-insert them inside the main ready-list here. */ -for (nepi =READ_ONCE (ep->ovflist); (epi = nepi) !=NULL ; +for (; (epi = nepi) !=NULL ; nepi = epi->next, epi->next =EP_UNACTIVE_PTR ) {/* * We need to check if the item is already in the list.
@@ -755,7 +764,7 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep,mutex_unlock (&ep->mtx);/* We have to call this outside the lock */ -if (pwake) +if (pwake ==2 )ep_poll_safewake (&ep->poll_wait);return res;
Signature · 1 line
2.23.0
Quoted 20 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0 | (ew) e0 | (et) e1 | (lt) s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-trigger
Quoted 5 lines
We only need to wakeup nested epoll fds if something has been queued to the overflow list, since the
ep_poll()traverses the rdllist during recursive poll and thus events on the overflow list may not be visible yet.Test code:
Look sane to me. Do you have any performance testing results which show a benefit?
epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
Quoted 26 lines · Andrew Morton
Quoted 24 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0 | (ew) e0 | (et) e1 | (lt) s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-triggerWe only need to wakeup nested epoll fds if something has been queued to the overflow list, since the
ep_poll()traverses the rdllist during recursive poll and thus events on the overflow list may not be visible yet.
Quoted 7 lines
Quoted 1 line
Test code:
Look sane to me. Do you have any performance testing results which show a benefit?
epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
Yes, I can revisit this once more next week.
Heiher, mind to prepare a patchset with your test suit and make it a part of kselftest? I hope nobody has any objections.
--
Roman
Quoted 26 lines · Andrew Morton
Quoted 24 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0 | (ew) e0 | (et) e1 | (lt) s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-triggerWe only need to wakeup nested epoll fds if something has been queued to the overflow list, since the
ep_poll()traverses the rdllist during recursive poll and thus events on the overflow list may not be visible yet.
Quoted 7 lines
Quoted 1 line
Test code:
Look sane to me. Do you have any performance testing results which show a benefit?
epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
So here is my observation: current patch does not fix the described problem (double wakeup) for the case, when new event comes exactly to the ->ovflist. According to the patch this is the desired intention:
/*
* We only need to wakeup nested epoll fds if something has been
queued
* to the overflow list, since the ep_poll() traverses the rdllist
* during recursive poll and thus events on the overflow list may not
be
* visible yet.
*/
if (nepi != NULL)
pwake++;
....
if (pwake == 2)
ep_poll_safewake(&ep->poll_wait);but this actually means that we repeat the same behavior (double wakeup) but only for the case, when event comes to the ->ovflist.
How to reproduce? Can be easily done (ok, not so easy but it is possible to try): to the given userspace test we need to add one more socket and immediately fire the event:
e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s2fd[0], &e) < 0)
goto out;
/*
* Signal any fd to let epoll_wait() to call ep_scan_ready_list()
* in order to "catch" it there and add new event to ->ovflist.
*/
if (write(s2fd[1], "w", 1) != 1)
goto out;That is done in order to let the following epoll_wait() call to invoke ep_scan_ready_list(), where we can "catch" and insert new event exactly to the ->ovflist. In order to insert event exactly in the correct list I introduce artificial delay.
Modified test and kernel patch is below. Here is the output of the testing tool with some debug lines from kernel:
# ~/devel/test/edge-bug
[ 59.263178] ### sleep 2 >> write to sock [ 61.318243] ### done sleep [ 61.318991] !!!!!!!!!!! ep_poll_safewake(&ep->poll_wait); events_in_rdllist=1, events_in_ovflist=1 [ 61.321204] ### sleep 2 [ 63.398325] ### done sleep error: What?! Again?!
First epoll_wait() call (ep_scan_ready_list()) observes 2 events (see "!!!!!!!!!!! ep_poll_safewake" output line), exactly what we wanted to achieve, so eventually ep_poll_safewake() is called again which leads to double wakeup.
In my opinion current patch as it is should be dropped, it does not fix the described problem but just hides it.
--
Roman
######### USERSPACE ##########
#include <unistd.h>
#include <sys/epoll.h>
#include <sys/socket.h>
#include <stdio.h>
#include <pthread.h>
static void *do_thread(void *arg)
{
int s = *(int *)arg;
sleep(1);
printf(">> write to sock\n");
write(s, "w", 1);
}int main(int argc, char *argv[])
{
int s1fd[2];
int s2fd[2];
int efd[2];
struct epoll_event e;
if (socketpair(AF_UNIX, SOCK_STREAM, 0, s1fd) < 0)
goto out;
if (socketpair(AF_UNIX, SOCK_STREAM, 0, s2fd) < 0)
goto out;
efd[0] = epoll_create(1);
if (efd[0] < 0)
goto out;
efd[1] = epoll_create(1);
if (efd[1] < 0)
goto out;
e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s1fd[0], &e) < 0)
goto out;
e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s2fd[0], &e) < 0)
goto out;
e.events = EPOLLIN | EPOLLET;
if (epoll_ctl(efd[0], EPOLL_CTL_ADD, efd[1], &e) < 0)
goto out;
/*
* Signal any fd to let epoll_wait() to call ep_scan_ready_list()
* in order to "catch" it there and add new event to ->ovflist.
*/
if (write(s2fd[1], "w", 1) != 1)
goto out;
pthread_t thr;
pthread_create(&thr, NULL, do_thread, &s1fd[1]);
if (epoll_wait(efd[0], &e, 1, 0) != 1) {
goto out;
}
pthread_join(thr, NULL);
if (epoll_wait(efd[0], &e, 1, 0) != 0) {
printf("error: What?! Again?!\n");
goto out;
}
return 0;out:
return -1; }
######### KERNEL ##########
fs/eventpoll.c +46 -3
@@ -39,6 +39,8 @@#include <linux/rculist.h> #include <net/busy_poll.h> +static bool is_send_events_call; +/* * LOCKING: * There are three level of locking required by epoll :
@@ -672,6 +674,8 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep,__poll_t res;int pwake =0 ;struct epitem *epi, *nepi; +unsigned events_in_rdllist =0 ; +unsigned events_in_ovflist =0 ;LIST_HEAD (txlist);lockdep_assert_irqs_enabled ();
@@ -693,23 +697,52 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep,* in a lockless way. */ write_lock_irq (&ep->lock); + +/* XXX Count events */ +if (!strcmp ("edge-bug" , current->comm) && depth) { +struct list_head *l; +list_for_each (l, &ep->rdllist) + events_in_rdllist++; + }list_splice_init (&ep->rdllist, &txlist);WRITE_ONCE (ep->ovflist,NULL );write_unlock_irq (&ep->lock); +if (!strcmp ("edge-bug" , current->comm) && depth && is_send_events_call) { +/* +* XXX Introduce delay to let userspace fire event +* XXX directly to ovflist. +*/ +pr_err ("### sleep 2\n" ); +msleep (2000 ); +pr_err ("### done sleep\n" ); + } + +/* * Now call the callback function. */ res = (*sproc)(ep, &txlist, priv);write_lock_irq (&ep->lock); + nepi =READ_ONCE (ep->ovflist); +/* +* We only need to wakeup nested epoll fds if something has been queued +* to the overflow list, since the ep_poll() traverses the rdllist +* during recursive poll and thus events on the overflow list may not be +* visible yet. +*/ +if (nepi !=NULL ) + pwake++;/* * During the time we spent inside the "sproc" callback, some * other events might have been queued by the poll callback. * We re-insert them inside the main ready-list here. */ -for (nepi =READ_ONCE (ep->ovflist); (epi = nepi) !=NULL ; +for (; (epi = nepi) !=NULL ; nepi = epi->next, epi->next =EP_UNACTIVE_PTR ) { +/* XXX Count events */ + events_in_ovflist++;/* * We need to check if the item is already in the list. * During the "sproc" callback execution time, items are
@@ -754,8 +787,11 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep,mutex_unlock (&ep->mtx);/* We have to call this outside the lock */ -if (pwake) +if (pwake ==2 ) { +pr_err ("!!!!!!!!!!! ep_poll_safewake(&ep->poll_wait); events_in_rdllist=%d, events_in_ovflist=%d\n" , + events_in_rdllist, events_in_ovflist);ep_poll_safewake (&ep->poll_wait); + }return res; }
@@ -1925,9 +1961,16 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,* there's still timeout left over, we go trying again in search of * more luck. */ + +/* XXX Catch only ep_scan_ready_list() called from here */ +if (!strcmp ("edge-bug" , current->comm)) + is_send_events_call =1 ;if (!res && eavail && - !(res =ep_send_events (ep, events, maxevents)) && !timed_out) + !(res =ep_send_events (ep, events, maxevents)) && !timed_out) { + is_send_events_call =0 ; goto fetch_events; + } + is_send_events_call =0 ;if (waiter) {spin_lock_irq (&ep->wq.lock);
Quoted 92 lines · Roman Penyaev
Quoted 28 lines · Andrew Morton
Quoted 26 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0
| (ew)
e0
| (et)
e1
| (lt)
s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-triggerWe only need to wakeup nested epoll fds if something has been queued to the
overflow list, since theep_poll()traverses the rdllist during recursive poll
and thus events on the overflow list may not be visible yet.Quoted 7 lines
Quoted 1 line
Test code:
Look sane to me. Do you have any performance testing results which show a benefit?
epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
So here is my observation: current patch does not fix the described problem (double wakeup) for the case, when new event comes exactly to the ->ovflist. According to the patch this is the desired intention:
/*
* We only need to wakeup nested epoll fds if something has been queued
* to the overflow list, since theep_poll()traverses the rdllist
* during recursive poll and thus events on the overflow list may not be
* visible yet.
*/
if (nepi != NULL)
pwake++;....
if (pwake == 2)
ep_poll_safewake(&ep->poll_wait);but this actually means that we repeat the same behavior (double wakeup) but only for the case, when event comes to the ->ovflist.
How to reproduce? Can be easily done (ok, not so easy but it is possible to try): to the given userspace test we need to add one more socket and immediately fire the event:
e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s2fd[0], &e)< 0) goto out;/*
* Signal any fd to letepoll_wait()to callep_scan_ready_list()
* in order to "catch" it there and add new event to ->ovflist. */
if (write(s2fd[1], "w", 1)!= 1)
goto out;That is done in order to let the following
epoll_wait()call to invokeep_scan_ready_list(), where we can "catch" and insert new event exactly to the ->ovflist. In order to insert event exactly in the correct list I introduce artificial delay.Modified test and kernel patch is below. Here is the output of the testing tool with some debug lines from kernel:
# ~/devel/test/edge-bug
[ 59.263178] ### sleep 2
>> write to sock
[ 61.318243] ### done sleep
[ 61.318991] !!!!!!!!!!!ep_poll_safewake(&ep->poll_wait); events_in_rdllist=1, events_in_ovflist=1
[ 61.321204] ### sleep 2
[ 63.398325] ### done sleep
error: What?! Again?!
Quoted 9 lines
First
epoll_wait()call (ep_scan_ready_list()) observes 2 events (see "!!!!!!!!!!!ep_poll_safewake" output line), exactly what we wanted to achieve, so eventuallyep_poll_safewake()is called again which leads to double wakeup.In my opinion current patch as it is should be dropped, it does not fix the described problem but just hides it.
Yes, there are 2 wakeups in the test case you describe, but if the second event (write to s1fd) gets queued after the first call to epoll_wait(), we are going to get 2 wakeups anyways. So yes, there may be a slightly bigger window with this patch for 2 wakeups, but its small and I tried to be conservative with the patch - I'd rather get an occasional 2nd wakeup then miss one. Trying to debug missing wakeups isn't always fun...
That said, the reason for propagating events that end up on the overflow list was to prevent the race of the wakee not seeing events because they were still on the overflow list. In the testcase, imagine if there was a thread doing epoll_wait() on efd[0], and then a write happends on s1fd. I thought it was possible then that a 2nd thread doing epoll_wait() on efd[1], wakes up, checks efd[0] and sees no events because they are still potentially on the overflow list. However, I think that case is not possible because the thread doing epoll_wait() on efd[0] is going to have the ep->mtx, and thus when the thread wakes up on efd[1], its going to have to be ordered because its also grabbing the ep->mtx associated with efd[0].
So I think its safe to do the following if we want to go further than the proposed patch, which is what you suggested earlier in the thread (minus keeping the wakeup on ep->wq).
diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index c4159bc..61d653d1 100644
fs/eventpoll.c +0 -1
@@ -671,7 +671,6 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep, void *priv, int depth, bool ep_locked) { __poll_t res; -int pwake =0 ;struct epitem *epi, *nepi;LIST_HEAD (txlist);
@@ -741,23 +740,16 @@ static __poll_t ep_scan_ready_list(struct
eventpoll *ep,
if (!list_empty(&ep->rdllist)) {
/*- * Wake up (if active) both the eventpoll wait list and
- * the ->poll() wait list (delayed after we release the
lock).
+ * Wake up (if active) the eventpoll wait list
*/
if (waitqueue_active(&ep->wq))
wake_up(&ep->wq);
- if (waitqueue_active(&ep->poll_wait))
- pwake++;
}
write_unlock_irq(&ep->lock);
if (!ep_locked)
mutex_unlock(&ep->mtx);- /* We have to call this outside the lock */
- if (pwake)
- ep_poll_safewake(&ep->poll_wait);
-
return res;
}
Thanks,
-Jason
Quoted 98 lines · Jason Baron
Quoted 97 lines · Roman Penyaev
Quoted 28 lines · Andrew Morton
Quoted 26 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0
| (ew)
e0
| (et)
e1
| (lt)
s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-triggerWe only need to wakeup nested epoll fds if something has been queued to the
overflow list, since theep_poll()traverses the rdllist during recursive poll
and thus events on the overflow list may not be visible yet.Quoted 7 lines
Quoted 1 line
Test code:
Look sane to me. Do you have any performance testing results which show a benefit?
epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
So here is my observation: current patch does not fix the described problem (double wakeup) for the case, when new event comes exactly to the ->ovflist. According to the patch this is the desired intention:
/*
* We only need to wakeup nested epoll fds if something has been queued
* to the overflow list, since theep_poll()traverses the rdllist
* during recursive poll and thus events on the overflow list may not be
* visible yet.
*/
if (nepi != NULL)
pwake++;....
if (pwake == 2)
ep_poll_safewake(&ep->poll_wait);but this actually means that we repeat the same behavior (double wakeup) but only for the case, when event comes to the ->ovflist.
How to reproduce? Can be easily done (ok, not so easy but it is possible to try): to the given userspace test we need to add one more socket and immediately fire the event:
e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s2fd[0], &e)< 0) goto out;/*
* Signal any fd to letepoll_wait()to callep_scan_ready_list()
* in order to "catch" it there and add new event to ->ovflist. */
if (write(s2fd[1], "w", 1)!= 1)
goto out;That is done in order to let the following
epoll_wait()call to invokeep_scan_ready_list(), where we can "catch" and insert new event exactly to the ->ovflist. In order to insert event exactly in the correct list I introduce artificial delay.Modified test and kernel patch is below. Here is the output of the testing tool with some debug lines from kernel:
# ~/devel/test/edge-bug
[ 59.263178] ### sleep 2
>> write to sock
[ 61.318243] ### done sleep
[ 61.318991] !!!!!!!!!!!ep_poll_safewake(&ep->poll_wait); events_in_rdllist=1, events_in_ovflist=1
[ 61.321204] ### sleep 2
[ 63.398325] ### done sleep
error: What?! Again?!First
epoll_wait()call (ep_scan_ready_list()) observes 2 events (see "!!!!!!!!!!!ep_poll_safewake" output line), exactly what we wanted to achieve, so eventuallyep_poll_safewake()is called again which leads to double wakeup.
Quoted 4 lines
Quoted 4 lines
In my opinion current patch as it is should be dropped, it does not fix the described problem but just hides it.
--
Hi Jason,
Quoted 4 lines
Yes, there are 2 wakeups in the test case you describe, but if the second event (write to s1fd) gets queued after the first call to
epoll_wait(), we are going to get 2 wakeups anyways.
Yes, exactly, for this reason I print out the number of events observed on first wait, there should be 1 (rdllist) and 1 (ovflist), otherwise this is another case, when second event comes exactly after first wait, which is legitimate wakeup.
Quoted 18 lines
So yes, there may
be a slightly bigger window with this patch for 2 wakeups, but its small and I tried to be conservative with the patch - I'd rather get an occasional 2nd wakeup then miss one. Trying to debug missing wakeups isn't always fun...That said, the reason for propagating events that end up on the overflow list was to prevent the race of the wakee not seeing events because they were still on the overflow list. In the testcase, imagine if there was a thread doing
epoll_wait()on efd[0], and then a write happends on s1fd. I thought it was possible then that a 2nd thread doingepoll_wait()on efd[1], wakes up, checks efd[0] and sees no events because they are still potentially on the overflow list. However, I think that case is not possible because the thread doingepoll_wait()on efd[0] is going to have theep->mtx, and thus when the thread wakes up on efd[1], its going to have to be ordered because its also grabbing theep->mtxassociated with efd[0].
Quoted 3 lines
So I think its safe to do the following if we want to go further than the proposed patch, which is what you suggested earlier in the thread (minus keeping the wakeup on
ep->wq).
Then I do not understand why we need to keep ep->wq wakeup? @wq and @poll_wait are almost the same with only one difference:
@wq is used when you sleep on it inside epoll_wait() and the other is used when you nest epoll fd inside epoll fd. Either you wake both up either you don't this at all.
ep_poll_callback() does wakeup explicitly, ep_insert() and ep_modify() do wakeup explicitly, so what are the cases when we need to do wakeups from ep_scan_ready_list()?
I would still remove the whole branch:
fs/eventpoll.c +0 -16
@@ -671,7 +671,6 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep, void *priv, int depth, bool ep_locked) { __poll_t res; -int pwake =0 ;struct epitem *epi, *nepi;LIST_HEAD (txlist);
@@ -738,26 +737,11 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep,*/ list_splice (&txlist, &ep->rdllist);__pm_relax (ep->ws); - -if (!list_empty (&ep->rdllist)) { -/* -* Wake up (if active) both the eventpoll wait list and -* the ->poll() wait list (delayed after we release the lock). -*/ -if (waitqueue_active (&ep->wq)) -wake_up (&ep->wq); -if (waitqueue_active (&ep->poll_wait)) - pwake++; - }write_unlock_irq (&ep->lock);if (!ep_locked)mutex_unlock (&ep->mtx); -/* We have to call this outside the lock */ -if (pwake) -ep_poll_safewake (&ep->poll_wait); -return res; }
--
Roman
Quoted 146 lines · Roman Penyaev
Quoted 100 lines · Jason Baron
Quoted 99 lines · Roman Penyaev
Quoted 28 lines · Andrew Morton
Quoted 26 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0
| (ew)
e0
| (et)
e1
| (lt)
s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-triggerWe only need to wakeup nested epoll fds if something has been queued to the
overflow list, since theep_poll()traverses the rdllist during recursive poll
and thus events on the overflow list may not be visible yet.Quoted 7 lines
Quoted 1 line
Test code:
Look sane to me. Do you have any performance testing results which show a benefit?
epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
So here is my observation: current patch does not fix the described problem (double wakeup) for the case, when new event comes exactly to the ->ovflist. According to the patch this is the desired intention:
/*
* We only need to wakeup nested epoll fds if something has been queued
* to the overflow list, since theep_poll()traverses the rdllist
* during recursive poll and thus events on the overflow list may not be
* visible yet.
*/
if (nepi != NULL)
pwake++;....
if (pwake == 2)
ep_poll_safewake(&ep->poll_wait);but this actually means that we repeat the same behavior (double wakeup) but only for the case, when event comes to the ->ovflist.
How to reproduce? Can be easily done (ok, not so easy but it is possible to try): to the given userspace test we need to add one more socket and immediately fire the event:
e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s2fd[0], &e)< 0) goto out;/*
* Signal any fd to letepoll_wait()to callep_scan_ready_list()
* in order to "catch" it there and add new event to ->ovflist. */
if (write(s2fd[1], "w", 1)!= 1)
goto out;That is done in order to let the following
epoll_wait()call to invokeep_scan_ready_list(), where we can "catch" and insert new event exactly to the ->ovflist. In order to insert event exactly in the correct list I introduce artificial delay.Modified test and kernel patch is below. Here is the output of the testing tool with some debug lines from kernel:
# ~/devel/test/edge-bug
[ 59.263178] ### sleep 2
>> write to sock
[ 61.318243] ### done sleep
[ 61.318991] !!!!!!!!!!!ep_poll_safewake(&ep->poll_wait); events_in_rdllist=1, events_in_ovflist=1
[ 61.321204] ### sleep 2
[ 63.398325] ### done sleep
error: What?! Again?!First
epoll_wait()call (ep_scan_ready_list()) observes 2 events (see "!!!!!!!!!!!ep_poll_safewake" output line), exactly what we wanted to achieve, so eventuallyep_poll_safewake()is called again which leads to double wakeup.Quoted 4 lines
Quoted 4 lines
In my opinion current patch as it is should be dropped, it does not fix the described problem but just hides it.
Hi Jason,
Quoted 4 lines
Yes, there are 2 wakeups in the test case you describe, but if the second event (write to s1fd) gets queued after the first call to
epoll_wait(), we are going to get 2 wakeups anyways.Yes, exactly, for this reason I print out the number of events observed on first wait, there should be 1 (rdllist) and 1 (ovflist), otherwise this is another case, when second event comes exactly after first wait, which is legitimate wakeup.
Quoted 18 lines
So yes, there may
be a slightly bigger window with this patch for 2 wakeups, but its small and I tried to be conservative with the patch - I'd rather get an occasional 2nd wakeup then miss one. Trying to debug missing wakeups isn't always fun...That said, the reason for propagating events that end up on the overflow list was to prevent the race of the wakee not seeing events because they were still on the overflow list. In the testcase, imagine if there was a thread doing
epoll_wait()on efd[0], and then a write happends on s1fd. I thought it was possible then that a 2nd thread doingepoll_wait()on efd[1], wakes up, checks efd[0] and sees no events because they are still potentially on the overflow list. However, I think that case is not possible because the thread doingepoll_wait()on efd[0] is going to have theep->mtx, and thus when the thread wakes up on efd[1], its going to have to be ordered because its also grabbing theep->mtxassociated with efd[0].Quoted 3 lines
So I think its safe to do the following if we want to go further than the proposed patch, which is what you suggested earlier in the thread (minus keeping the wakeup on
ep->wq).Then I do not understand why we need to keep
ep->wqwakeup? @wq and @poll_wait are almost the same with only one difference:
@wq is used when you sleep on it insideepoll_wait()and the other is used when you nest epoll fd inside epoll fd. Either you wake both up either you don't this at all.
Quoted 3 lines
ep_poll_callback()does wakeup explicitly,ep_insert()andep_modify()do wakeup explicitly, so what are the cases when we need to do wakeups fromep_scan_ready_list()?
Hi Roman,
So the reason I was saying not to drop the ep->wq wakeup was that I was thinking about a usecase where you have multi-threads say thread A and thread B which are doing epoll_wait() on the same epfd. Now, the threads both call epoll_wait() and are added as exclusive to ep->wq. Now a bunch of events happen and thread A is woken up. However, thread A may only process a subset of the events due to its 'maxevents' parameter. In that case, I was thinking that the wakeup on ep->wq might be helpful, because in the absence of subsequent events, thread B can now start processing the rest, instead of waiting for the next event to be queued.
However, I was thinking about the state of things before:
86c0517 fs/epoll: deal with wait_queue only once
Before that patch, thread A would have been removed from eq->wq before the wakeup call, thus waking up thread B. However, now that thread A stays on the queue during the call to ep_send_events(), I believe the wakeup would only target thread A, which doesn't help since its already checking for events. So given the state of things I think you are right in that its not needed. However, I wonder if not removing from the ep->wq affects the multi-threaded case I described. Its been around since 5.0, so probably not, but it would be a more subtle performance difference.
Thanks,
-Jason
Quoted 49 lines
I would still remove the whole branch:
fs/eventpoll.c +0
-6@@ -671,7 +671,6 @@ static __poll_t ep_scan_ready_list(struct eventpoll *ep, void *priv, int depth, bool ep_locked) { __poll_t res; - int pwake = 0; struct epitem *epi, *nepi; LIST_HEAD(txlist);@@ -738,26 +737,11 @@ static __poll_t ep_scan_ready_list(structeventpoll *ep, */ list_splice(&txlist, &ep->rdllist); __pm_relax(ep->ws); - - if (!list_empty(&ep->rdllist)) { - /* -* Wake up (if active) both the eventpoll wait list and -* the ->poll() wait list (delayed after we release the lock).
- */
- if (waitqueue_active(&ep->wq)) -wake_up(&ep->wq);
- if (waitqueue_active(&ep->poll_wait)) - pwake++;
- }
write_unlock_irq(&ep->lock);if (!ep_locked)
mutex_unlock(&ep->mtx);- /* We have to call this outside the lock */ - if (pwake)
-ep_poll_safewake(&ep->poll_wait); -
return res;
}Signature · 1 line
Roman
Quoted 166 lines · Jason Baron
Quoted 146 lines · Roman Penyaev
Quoted 100 lines · Jason Baron
Quoted 99 lines · Roman Penyaev
Quoted 28 lines · Andrew Morton
Quoted 26 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0
| (ew)
e0
| (et)
e1
| (lt)
s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-triggerWe only need to wakeup nested epoll fds if something has been queued to the
overflow list, since theep_poll()traverses the rdllist during recursive poll
and thus events on the overflow list may not be visible yet.Quoted 7 lines
Quoted 1 line
Test code:
Look sane to me. Do you have any performance testing results which show a benefit?
epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
So here is my observation: current patch does not fix the described problem (double wakeup) for the case, when new event comes exactly to the ->ovflist. According to the patch this is the desired intention:
/*
* We only need to wakeup nested epoll fds if something has been queued
* to the overflow list, since theep_poll()traverses the rdllist
* during recursive poll and thus events on the overflow list may not be
* visible yet.
*/
if (nepi != NULL)
pwake++;....
if (pwake == 2)
ep_poll_safewake(&ep->poll_wait);but this actually means that we repeat the same behavior (double wakeup) but only for the case, when event comes to the ->ovflist.
How to reproduce? Can be easily done (ok, not so easy but it is possible to try): to the given userspace test we need to add one more socket and immediately fire the event:
e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s2fd[0], &e)< 0) goto out;/*
* Signal any fd to letepoll_wait()to callep_scan_ready_list()
* in order to "catch" it there and add new event to ->ovflist. */
if (write(s2fd[1], "w", 1)!= 1)
goto out;That is done in order to let the following
epoll_wait()call to invokeep_scan_ready_list(), where we can "catch" and insert new event exactly to the ->ovflist. In order to insert event exactly in the correct list I introduce artificial delay.Modified test and kernel patch is below. Here is the output of the testing tool with some debug lines from kernel:
# ~/devel/test/edge-bug
[ 59.263178] ### sleep 2
>> write to sock
[ 61.318243] ### done sleep
[ 61.318991] !!!!!!!!!!!ep_poll_safewake(&ep->poll_wait); events_in_rdllist=1, events_in_ovflist=1
[ 61.321204] ### sleep 2
[ 63.398325] ### done sleep
error: What?! Again?!First
epoll_wait()call (ep_scan_ready_list()) observes 2 events (see "!!!!!!!!!!!ep_poll_safewake" output line), exactly what we wanted to achieve, so eventuallyep_poll_safewake()is called again which leads to double wakeup.Quoted 4 lines
Quoted 4 lines
In my opinion current patch as it is should be dropped, it does not fix the described problem but just hides it.
--
Hi Jason,
Quoted 4 lines
Yes, there are 2 wakeups in the test case you describe, but if the second event (write to s1fd) gets queued after the first call to
epoll_wait(), we are going to get 2 wakeups anyways.Yes, exactly, for this reason I print out the number of events observed on first wait, there should be 1 (rdllist) and 1 (ovflist), otherwise this is another case, when second event comes exactly after first wait, which is legitimate wakeup.
Quoted 18 lines
So yes, there may
be a slightly bigger window with this patch for 2 wakeups, but its small and I tried to be conservative with the patch - I'd rather get an occasional 2nd wakeup then miss one. Trying to debug missing wakeups isn't always fun...That said, the reason for propagating events that end up on the overflow list was to prevent the race of the wakee not seeing events because they were still on the overflow list. In the testcase, imagine if there was a thread doing
epoll_wait()on efd[0], and then a write happends on s1fd. I thought it was possible then that a 2nd thread doingepoll_wait()on efd[1], wakes up, checks efd[0] and sees no events because they are still potentially on the overflow list. However, I think that case is not possible because the thread doingepoll_wait()on efd[0] is going to have theep->mtx, and thus when the thread wakes up on efd[1], its going to have to be ordered because its also grabbing theep->mtxassociated with efd[0].Quoted 3 lines
So I think its safe to do the following if we want to go further than the proposed patch, which is what you suggested earlier in the thread (minus keeping the wakeup on
ep->wq).Then I do not understand why we need to keep
ep->wqwakeup? @wq and @poll_wait are almost the same with only one difference:
@wq is used when you sleep on it insideepoll_wait()and the other is used when you nest epoll fd inside epoll fd. Either you wake both up either you don't this at all.Quoted 3 lines
ep_poll_callback()does wakeup explicitly,ep_insert()andep_modify()do wakeup explicitly, so what are the cases when we need to do wakeups fromep_scan_ready_list()?Hi Roman,
So the reason I was saying not to drop the
ep->wqwakeup was that I was thinking about a usecase where you have multi-threads say thread A and thread B which are doingepoll_wait()on the same epfd. Now, the threads both callepoll_wait()and are added as exclusive toep->wq. Now a bunch of events happen and thread A is woken up. However, thread A may only process a subset of the events due to its 'maxevents' parameter. In that case, I was thinking that the wakeup onep->wqmight be helpful, because in the absence of subsequent events, thread B can now start processing the rest, instead of waiting for the next event to be queued.However, I was thinking about the state of things before:
86c0517 fs/epoll: deal withwait_queueonly once
Quoted 9 lines
Before that patch, thread A would have been removed from
eq->wqbefore the wakeup call, thus waking up thread B. However, now that thread A stays on the queue during the call toep_send_events(), I believe the wakeup would only target thread A, which doesn't help since its already checking for events. So given the state of things I think you are right in that its not needed. However, I wonder if not removing from theep->wqaffects the multi-threaded case I described. Its been around since 5.0, so probably not, but it would be a more subtle performance difference.
Now I understand what you mean. You want to prevent "idling" of events, while thread A is busy with the small portion of events (portion is equal to maxevents). On next iteration thread A will pick up the rest, no doubts, but would be nice to give a chance to thread B immediately to deal with the rest. Ok, makes sense.
But what if to make this wakeup explicit if we have more events to process? (nothing is tested, just a guess)
@@ -255,6 +255,7 @@ struct ep_pqueue {
struct ep_send_events_data {
int maxevents;
struct epoll_event __user *events;
+ bool have_more;
int res; };
@@ -1783,14 +1768,17 @@ static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head
}
static int ep_send_events(struct eventpoll *ep,
- struct epoll_event __user *events, int maxevents)
+ struct epoll_event __user *events, int maxevents,
+ bool *have_more)
{
- struct ep_send_events_data esed;
-
- esed.maxevents = maxevents;
- esed.events = events;
+ struct ep_send_events_data esed = {+ .maxevents = maxevents,
+ .events = events,
+ };
ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
+ *have_more = esed.have_more;
+
return esed.res;}
@@ -1827,7 +1815,7 @@ 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;
+ bool waiter = false, have_more;
wait_queue_entry_t wait;
ktime_t expires, *to = NULL;@@ -1927,7 +1915,8 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
* more luck.
*/
if (!res && eavail &&
- !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
+ !(res = ep_send_events(ep, events, maxevents, &have_more)) &&
+ !timed_out)
goto fetch_events;
if (waiter) {
@@ -1935,6 +1924,12 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
__remove_wait_queue(&ep->wq, &wait);
spin_unlock_irq(&ep->wq.lock);
}
+ /*
+ * We were not able to process all the events, so immediately
+ * wakeup other waiter.
+ */+ if (res > 0 && have_more && waitqueue_active(&ep->wq))
+ wake_up(&ep->wq);
return res;
}
PS. So what we decide with the original patch? Remove the whole branch?
--
Roman
Quoted 188 lines · Roman Penyaev
Quoted 177 lines · Jason Baron
Quoted 157 lines · Roman Penyaev
Quoted 105 lines · Jason Baron
Quoted 104 lines · Roman Penyaev
Quoted 28 lines · Andrew Morton
Quoted 26 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0
| (ew)
e0
| (et)
e1
| (lt)
s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-triggerWe only need to wakeup nested epoll fds if something has been queued to the
overflow list, since theep_poll()traverses the rdllist during recursive poll
and thus events on the overflow list may not be visible yet.Quoted 7 lines
Quoted 1 line
Test code:
Look sane to me. Do you have any performance testing results which show a benefit?
epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
So here is my observation: current patch does not fix the described problem (double wakeup) for the case, when new event comes exactly to the ->ovflist. According to the patch this is the desired intention:
/*
* We only need to wakeup nested epoll fds if something has been queued
* to the overflow list, since theep_poll()traverses the rdllist
* during recursive poll and thus events on the overflow list may not be
* visible yet.
*/
if (nepi != NULL)
pwake++;....
if (pwake == 2)
ep_poll_safewake(&ep->poll_wait);but this actually means that we repeat the same behavior (double wakeup)
but only for the case, when event comes to the ->ovflist.How to reproduce? Can be easily done (ok, not so easy but it is possible
to try): to the given userspace test we need to add one more socket and
immediately fire the event:e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s2fd[0], &e)< 0) goto out;/*
* Signal any fd to letepoll_wait()to callep_scan_ready_list()
* in order to "catch" it there and add new event to ->ovflist. */
if (write(s2fd[1], "w", 1)!= 1)
goto out;That is done in order to let the following
epoll_wait()call to invokeep_scan_ready_list(), where we can "catch" and insert new event exactly
to the ->ovflist. In order to insert event exactly in the correct list I introduce artificial delay.Modified test and kernel patch is below. Here is the output of the testing tool with some debug lines from kernel:
# ~/devel/test/edge-bug
[ 59.263178] ### sleep 2
>> write to sock
[ 61.318243] ### done sleep
[ 61.318991] !!!!!!!!!!!ep_poll_safewake(&ep->poll_wait); events_in_rdllist=1, events_in_ovflist=1
[ 61.321204] ### sleep 2
[ 63.398325] ### done sleep
error: What?! Again?!First
epoll_wait()call (ep_scan_ready_list()) observes 2 events (see "!!!!!!!!!!!ep_poll_safewake" output line), exactly what we wanted to achieve, so eventuallyep_poll_safewake()is called again which leads to double wakeup.Quoted 4 lines
Quoted 4 lines
In my opinion current patch as it is should be dropped, it does not fix the described problem but just hides it.
Hi Jason,
Quoted 4 lines
Yes, there are 2 wakeups in the test case you describe, but if the second event (write to s1fd) gets queued after the first call to
epoll_wait(), we are going to get 2 wakeups anyways.Yes, exactly, for this reason I print out the number of events observed on first wait, there should be 1 (rdllist) and 1 (ovflist), otherwise this is another case, when second event comes exactly after first wait, which is legitimate wakeup.
Quoted 24 lines
So yes, there may
be a slightly bigger window with this patch for 2 wakeups, but its small
and I tried to be conservative with the patch - I'd rather get an occasional 2nd wakeup then miss one. Trying to debug missing wakeups isn't always fun...That said, the reason for propagating events that end up on the overflow
list was to prevent the race of the wakee not seeing events because they
were still on the overflow list. In the testcase, imagine if there was a
thread doingepoll_wait()on efd[0], and then a write happends on s1fd. I thought it was possible then that a 2nd thread doingepoll_wait()on efd[1], wakes up, checks efd[0] and sees no events because they are still potentially on the overflow list. However, I think that case is not possible because the thread doingepoll_wait()on efd[0] is going to
have theep->mtx, and thus when the thread wakes up on efd[1], its going
to have to be ordered because its also grabbing theep->mtxassociated with efd[0].Quoted 3 lines
So I think its safe to do the following if we want to go further than the proposed patch, which is what you suggested earlier in the thread (minus keeping the wakeup on
ep->wq).Then I do not understand why we need to keep
ep->wqwakeup? @wq and @poll_wait are almost the same with only one difference:
@wq is used when you sleep on it insideepoll_wait()and the other is used when you nest epoll fd inside epoll fd. Either you wake both up either you don't this at all.Quoted 3 lines
ep_poll_callback()does wakeup explicitly,ep_insert()andep_modify()do wakeup explicitly, so what are the cases when we need to do wakeups fromep_scan_ready_list()?Hi Roman,
So the reason I was saying not to drop the
ep->wqwakeup was that I was thinking about a usecase where you have multi-threads say thread A and thread B which are doingepoll_wait()on the same epfd. Now, the threads both callepoll_wait()and are added as exclusive toep->wq. Now a bunch of events happen and thread A is woken up. However, thread A may only process a subset of the events due to its 'maxevents' parameter. In that case, I was thinking that the wakeup onep->wqmight be helpful, because in the absence of subsequent events, thread B can now start processing the rest, instead of waiting for the next event to be queued.However, I was thinking about the state of things before:
86c0517 fs/epoll: deal withwait_queueonly onceQuoted 9 lines
Before that patch, thread A would have been removed from
eq->wqbefore the wakeup call, thus waking up thread B. However, now that thread A stays on the queue during the call toep_send_events(), I believe the wakeup would only target thread A, which doesn't help since its already checking for events. So given the state of things I think you are right in that its not needed. However, I wonder if not removing from theep->wqaffects the multi-threaded case I described. Its been around since 5.0, so probably not, but it would be a more subtle performance difference.
Quoted 6 lines
Now I understand what you mean. You want to prevent "idling" of events, while thread A is busy with the small portion of events (portion is equal to maxevents). On next iteration thread A will pick up the rest, no doubts,
but would be nice to give a chance to thread B immediately to deal with the rest. Ok, makes sense.
Exactly, I don't believe its racy as is - but it seems like it would be good to wakeup other threads that may be waiting. That said, this logic was removed as I pointed out. So I'm not sure we need to tie this change to the current one - but it may be a nice addition.
Quoted 69 lines
But what if to make this wakeup explicit if we have more events to process? (nothing is tested, just a guess)
(quoted diff) +1
-0@@ -255,6 +255,7 @@ struct ep_pqueue { struct ep_send_events_data { int maxevents; struct epoll_event __user *events; + bool have_more; int res; };@@ -1783,14 +1768,17 @@ static __poll_t ep_send_events_proc(structeventpoll *ep, struct
list_head*head
}static int ep_send_events(struct eventpoll *ep,
- structepoll_event__user*events, int maxevents) + structepoll_event__user*events, int maxevents, + bool*have_more)
{
- structep_send_events_dataesed;
-
- esed.maxevents = maxevents;
- esed.events = events;
+ structep_send_events_dataesed = {
+ .maxevents = maxevents,
+ .events = events,
+ };
ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false); +*have_more= esed.have_more;
+
return esed.res;
}(quoted diff) +1
-1@@ -1827,7 +1815,7 @@ 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; + bool waiter = false, have_more; wait_queue_entry_t wait; ktime_t expires, *to = NULL;@@ -1927,7 +1915,8 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, * more luck. */if (!res && eavail &&
- !(res =ep_send_events(ep, events, maxevents)) && !timed_out) + !(res =ep_send_events(ep, events, maxevents, &have_more)) && + !timed_out)
gotofetch_events;if (waiter) {
(quoted diff) +6
-0@@ -1935,6 +1924,12 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, __remove_wait_queue(&ep->wq, &wait); spin_unlock_irq(&ep->wq.lock); } + /* + * We were not able to process all the events, so immediately + * wakeup other waiter. + */ + if (res > 0 && have_more && waitqueue_active(&ep->wq)) + wake_up(&ep->wq); return res;
Quoted 3 lines
}
Ok, yeah I like making it explicit. Looks like you are missing the changes to ep_scan_ready_list(), but I think the general approach makes sense. Although I really didn't have a test case that motivated this - its just was sort of noting this change in behavior while reviewing the current change.
Quoted 2 lines
PS. So what we decide with the original patch? Remove the whole branch?
For fwiw, I'm ok removing the whole branch as you proposed. And I think the above change can go in separately (if we decide we want it). I don't think they need to be tied together. I also want to make sure this change gets a full linux-next cycle, so I think it should target 5.5 at this point.
Thanks,
-Jason
Quoted 274 lines · Jason Baron
Quoted 189 lines · Roman Penyaev
Quoted 178 lines · Jason Baron
Quoted 158 lines · Roman Penyaev
Quoted 106 lines · Jason Baron
Quoted 105 lines · Roman Penyaev
Quoted 28 lines · Andrew Morton
Quoted 26 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0
| (ew)
e0
| (et)
e1
| (lt)
s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-triggerWe only need to wakeup nested epoll fds if something has been queued to the
overflow list, since theep_poll()traverses the rdllist during recursive poll
and thus events on the overflow list may not be visible yet.Quoted 7 lines
Quoted 1 line
Test code:
Look sane to me. Do you have any performance testing results which show a benefit?
epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
So here is my observation: current patch does not fix the described problem (double wakeup) for the case, when new event comes exactly to the ->ovflist. According to the patch this is the desired intention:
/*
* We only need to wakeup nested epoll fds if something has been queued
* to the overflow list, since theep_poll()traverses the rdllist
* during recursive poll and thus events on the overflow list may not be
* visible yet.
*/
if (nepi != NULL)
pwake++;....
if (pwake == 2)
ep_poll_safewake(&ep->poll_wait);but this actually means that we repeat the same behavior (double wakeup)
but only for the case, when event comes to the ->ovflist.How to reproduce? Can be easily done (ok, not so easy but it is possible
to try): to the given userspace test we need to add one more socket and
immediately fire the event:e.events = EPOLLIN;
if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s2fd[0], &e)< 0) goto out;/*
* Signal any fd to letepoll_wait()to callep_scan_ready_list()
* in order to "catch" it there and add new event to ->ovflist.
*/
if (write(s2fd[1], "w", 1)!= 1)
goto out;That is done in order to let the following
epoll_wait()call to invokeep_scan_ready_list(), where we can "catch" and insert new event exactly
to the ->ovflist. In order to insert event exactly in the correct list I introduce artificial delay.Modified test and kernel patch is below. Here is the output of the testing tool with some debug lines from kernel:
# ~/devel/test/edge-bug
[ 59.263178] ### sleep 2
>> write to sock
[ 61.318243] ### done sleep
[ 61.318991] !!!!!!!!!!!ep_poll_safewake(&ep->poll_wait); events_in_rdllist=1, events_in_ovflist=1
[ 61.321204] ### sleep 2
[ 63.398325] ### done sleep
error: What?! Again?!First
epoll_wait()call (ep_scan_ready_list()) observes 2 events (see "!!!!!!!!!!!ep_poll_safewake" output line), exactly what we wanted to achieve, so eventuallyep_poll_safewake()is called again which leads to double wakeup.Quoted 4 lines
Quoted 4 lines
In my opinion current patch as it is should be dropped, it does not fix the described problem but just hides it.
--
Hi Jason,
Quoted 4 lines
Yes, there are 2 wakeups in the test case you describe, but if the second event (write to s1fd) gets queued after the first call to
epoll_wait(), we are going to get 2 wakeups anyways.Yes, exactly, for this reason I print out the number of events observed on first wait, there should be 1 (rdllist) and 1 (ovflist), otherwise this is another case, when second event comes exactly after first wait, which is legitimate wakeup.
Quoted 24 lines
So yes, there may
be a slightly bigger window with this patch for 2 wakeups, but its small
and I tried to be conservative with the patch - I'd rather get an occasional 2nd wakeup then miss one. Trying to debug missing wakeups isn't always fun...That said, the reason for propagating events that end up on the overflow
list was to prevent the race of the wakee not seeing events because they
were still on the overflow list. In the testcase, imagine if there was a
thread doingepoll_wait()on efd[0], and then a write happends on s1fd. I thought it was possible then that a 2nd thread doingepoll_wait()on efd[1], wakes up, checks efd[0] and sees no events because they are still potentially on the overflow list. However, I think that case is not possible because the thread doingepoll_wait()on efd[0] is going to
have theep->mtx, and thus when the thread wakes up on efd[1], its going
to have to be ordered because its also grabbing theep->mtxassociated with efd[0].Quoted 3 lines
So I think its safe to do the following if we want to go further than the proposed patch, which is what you suggested earlier in the thread (minus keeping the wakeup on
ep->wq).Then I do not understand why we need to keep
ep->wqwakeup? @wq and @poll_wait are almost the same with only one difference:
@wq is used when you sleep on it insideepoll_wait()and the other is used when you nest epoll fd inside epoll fd. Either you wake both up either you don't this at all.Quoted 3 lines
ep_poll_callback()does wakeup explicitly,ep_insert()andep_modify()do wakeup explicitly, so what are the cases when we need to do wakeups fromep_scan_ready_list()?Hi Roman,
So the reason I was saying not to drop the
ep->wqwakeup was that I was thinking about a usecase where you have multi-threads say thread A and thread B which are doingepoll_wait()on the same epfd. Now, the threads both callepoll_wait()and are added as exclusive toep->wq. Now a bunch of events happen and thread A is woken up. However, thread A may only process a subset of the events due to its 'maxevents' parameter. In that case, I was thinking that the wakeup onep->wqmight be helpful, because in the absence of subsequent events, thread B can now start processing the rest, instead of waiting for the next event to be queued.However, I was thinking about the state of things before:
86c0517 fs/epoll: deal withwait_queueonly onceQuoted 9 lines
Before that patch, thread A would have been removed from
eq->wqbefore the wakeup call, thus waking up thread B. However, now that thread A stays on the queue during the call toep_send_events(), I believe the wakeup would only target thread A, which doesn't help since its already checking for events. So given the state of things I think you are right in that its not needed. However, I wonder if not removing from theep->wqaffects the multi-threaded case I described. Its been around since 5.0, so probably not, but it would be a more subtle performance difference.Quoted 6 lines
Now I understand what you mean. You want to prevent "idling" of events, while thread A is busy with the small portion of events (portion is equal to maxevents). On next iteration thread A will pick up the rest, no doubts,
but would be nice to give a chance to thread B immediately to deal with the rest. Ok, makes sense.Exactly, I don't believe its racy as is - but it seems like it would be good to wakeup other threads that may be waiting. That said, this logic was removed as I pointed out. So I'm not sure we need to tie this change to the current one - but it may be a nice addition.
Quoted 72 lines
But what if to make this wakeup explicit if we have more events to process? (nothing is tested, just a guess)
(quoted diff) +1
-0@@ -255,6 +255,7 @@ struct ep_pqueue { struct ep_send_events_data { int maxevents; struct epoll_event __user *events; + bool have_more; int res; };@@ -1783,14 +1768,17 @@ static __poll_t ep_send_events_proc(structeventpoll *ep, struct
list_head*head
}static int ep_send_events(struct eventpoll *ep,
- structepoll_event__user*events, int maxevents)
+ structepoll_event__user*events, int maxevents,
+ bool*have_more)
{
- structep_send_events_dataesed;
-
- esed.maxevents = maxevents;
- esed.events = events;
+ structep_send_events_dataesed = {
+ .maxevents = maxevents,
+ .events = events,
+ };
ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false); +*have_more= esed.have_more;
+
return esed.res;
}(quoted diff) +1
-1@@ -1827,7 +1815,7 @@ 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; + bool waiter = false, have_more; wait_queue_entry_t wait; ktime_t expires, *to = NULL;@@ -1927,7 +1915,8 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, * more luck. */if (!res && eavail &&
- !(res =ep_send_events(ep, events, maxevents)) && !timed_out)
+ !(res =ep_send_events(ep, events, maxevents, &have_more)) &&
+ !timed_out)
gotofetch_events;if (waiter) {
(quoted diff) +6
-0@@ -1935,6 +1924,12 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, __remove_wait_queue(&ep->wq, &wait); spin_unlock_irq(&ep->wq.lock); } + /* + * We were not able to process all the events, so immediately + * wakeup other waiter. + */ + if (res > 0 && have_more && waitqueue_active(&ep->wq)) + wake_up(&ep->wq);
Quoted 8 lines
Quoted 4 lines
fs/eventpoll.c +0
-0return res;}
Ok, yeah I like making it explicit. Looks like you are missing the changes to
ep_scan_ready_list(), but I think the general approach makes sense.
Yeah, missed the hunk:
@@ -1719,8 +1704,10 @@ static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head
lockdep_assert_held(&ep->mtx);
list_for_each_entry_safe(epi, tmp, head, rdllink) {
- if (esed->res >= esed->maxevents)
+ if (esed->res >= esed->maxevents) {
+ esed->have_more = true;
break;+ }
Quoted 8 lines
Although I really didn't have a test case that motivated this - its just was sort of noting this change in behavior while reviewing the current change.
Quoted 2 lines
PS. So what we decide with the original patch? Remove the whole branch?
For fwiw, I'm ok removing the whole branch as you proposed.
Then probably Heiher could resend once more. Heiher, can you, please?
Quoted 4 lines
And I think the above change can go in separately (if we decide we want it). I don't think they need to be tied together. I also want to make sure this change gets a full linux-next cycle, so I think it should target 5.5 at this point.
Sure, this explicit ->wq wakeup is a separate patch, which should be covered with some benchmarks. I can try to cook something out in order to get numbers.
--
Roman
Hi,
Quoted 341 lines · Roman Penyaev
Quoted 313 lines · Jason Baron
Quoted 223 lines · Roman Penyaev
Quoted 209 lines · Jason Baron
Quoted 180 lines · Roman Penyaev
Quoted 118 lines · Jason Baron
Quoted 117 lines · Roman Penyaev
Quoted 29 lines · Andrew Morton
Quoted 27 lines · On Wed, 25 Sep 2019 09:56:03 +0800 hev <[email protected]> wrote:
From: Heiher <[email protected]>
Take the case where we have:
t0 | (ew) e0 | (et) e1 | (lt) s0t0: thread 0
e0: epoll fd 0
e1: epoll fd 1
s0: socket fd 0
ew:epoll_wait
et: edge-trigger
lt: level-triggerWe only need to wakeup nested epoll fds if something has been queued
to the
overflow list, since theep_poll()traverses the rdllist during recursive poll
and thus events on the overflow list may not be visible yet.Quoted 8 lines
Quoted 1 line
Test code:
Look sane to me. Do you have any performance testing results which
show a benefit?epoll maintainership isn't exactly a hive of activity nowadays :( Roman, would you please have time to review this?
So here is my observation: current patch does not fix the described
problem (double wakeup) for the case, when new event comes exactly to the ->ovflist. According to the patch this is the desired intention:/*
* We only need to wakeup nested epoll fds if something has
been
queued
* to the overflow list, since theep_poll()traverses the
rdllist
* during recursive poll and thus events on the overflow list
may
not be* visible yet. */ if (nepi != NULL) pwake++; .... if (pwake == 2) ep_poll_safewake(&ep->poll_wait);but this actually means that we repeat the same behavior (double wakeup)
but only for the case, when event comes to the ->ovflist.How to reproduce? Can be easily done (ok, not so easy but it is possible
to try): to the given userspace test we need to add one more socket
and
immediately fire the event:e.events = EPOLLIN; if (epoll_ctl(efd[1], EPOLL_CTL_ADD, s2fd[0], &e) < 0) goto out; /* * Signal any fd to let epoll_wait() to call ep_scan_ready_list() * in order to "catch" it there and add new event to ->ovflist. */ if (write(s2fd[1], "w", 1) != 1) goto out;That is done in order to let the following
epoll_wait()call to invokeep_scan_ready_list(), where we can "catch" and insert new event exactly
to the ->ovflist. In order to insert event exactly in the correct list
I introduce artificial delay.Modified test and kernel patch is below. Here is the output of the
testing tool with some debug lines from kernel:# ~/devel/test/edge-bug
[ 59.263178] ### sleep 2 >> write to sock [ 61.318243] ### done sleep [ 61.318991] !!!!!!!!!!! ep_poll_safewake(&ep->poll_wait); events_in_rdllist=1, events_in_ovflist=1 [ 61.321204] ### sleep 2 [ 63.398325] ### done sleep error: What?! Again?!First
epoll_wait()call (ep_scan_ready_list()) observes 2 events (see "!!!!!!!!!!!ep_poll_safewake" output line), exactly what we wanted to achieve, so eventuallyep_poll_safewake()is called again
which leads to double wakeup.Quoted 5 lines
Quoted 5 lines
In my opinion current patch as it is should be dropped, it does not
fix the described problem but just hides it.--
Hi Jason,
Quoted 4 lines
Yes, there are 2 wakeups in the test case you describe, but if the second event (write to s1fd) gets queued after the first call to
epoll_wait(), we are going to get 2 wakeups anyways.Yes, exactly, for this reason I print out the number of events observed
on first wait, there should be 1 (rdllist) and 1 (ovflist), otherwise
this is another case, when second event comes exactly after first wait, which is legitimate wakeup.Quoted 29 lines
So yes, there may
be a slightly bigger window with this patch for 2 wakeups, but its small
and I tried to be conservative with the patch - I'd rather get an occasional 2nd wakeup then miss one. Trying to debug missing wakeups
isn't always fun...That said, the reason for propagating events that end up on the overflow
list was to prevent the race of the wakee not seeing events because they
were still on the overflow list. In the testcase, imagine if there was a
thread doingepoll_wait()on efd[0], and then a write happends on s1fd.
I thought it was possible then that a 2nd thread doingepoll_wait()on
efd[1], wakes up, checks efd[0] and sees no events because they are still potentially on the overflow list. However, I think that case is
not possible because the thread doingepoll_wait()on efd[0] is going to
have theep->mtx, and thus when the thread wakes up on efd[1], its going
to have to be ordered because its also grabbing theep->mtxassociated
with efd[0].Quoted 5 lines
So I think its safe to do the following if we want to go further than
the proposed patch, which is what you suggested earlier in the thread
(minus keeping the wakeup onep->wq).Then I do not understand why we need to keep
ep->wqwakeup? @wq and @poll_wait are almost the same with only one difference:
@wq is used when you sleep on it insideepoll_wait()and the other is used when you nest epoll fd inside epoll fd. Either you wake both up either you don't this at all.Quoted 5 lines
ep_poll_callback()does wakeup explicitly,ep_insert()andep_modify()
do wakeup explicitly, so what are the cases when we need to do wakeups
fromep_scan_ready_list()?Hi Roman,
So the reason I was saying not to drop the
ep->wqwakeup was that I was
thinking about a usecase where you have multi-threads say thread A and
thread B which are doingepoll_wait()on the same epfd. Now, the threads
both callepoll_wait()and are added as exclusive toep->wq. Now a bunch
of events happen and thread A is woken up. However, thread A may only process a subset of the events due to its 'maxevents' parameter. In that
case, I was thinking that the wakeup onep->wqmight be helpful, because
in the absence of subsequent events, thread B can now start processing
the rest, instead of waiting for the next event to be queued.However, I was thinking about the state of things before:
86c0517 fs/epoll: deal withwait_queueonly onceQuoted 12 lines
Before that patch, thread A would have been removed from
eq->wqbefore
the wakeup call, thus waking up thread B. However, now that thread A stays on the queue during the call toep_send_events(), I believe the wakeup would only target thread A, which doesn't help since its already
checking for events. So given the state of things I think you are right
in that its not needed. However, I wonder if not removing from theep->wqaffects the multi-threaded case I described. Its been around since 5.0, so probably not, but it would be a more subtle performance difference.Quoted 9 lines
Now I understand what you mean. You want to prevent "idling" of events,
while thread A is busy with the small portion of events (portion is equal
to maxevents). On next iteration thread A will pick up the rest, no doubts,
but would be nice to give a chance to thread B immediately to deal with the
rest. Ok, makes sense.Exactly, I don't believe its racy as is - but it seems like it would be good to wakeup other threads that may be waiting. That said, this logic was removed as I pointed out. So I'm not sure we need to tie this change
to the current one - but it may be a nice addition.Quoted 73 lines
But what if to make this wakeup explicit if we have more events to process?
(nothing is tested, just a guess)(quoted diff) +1
-0@@ -255,6 +255,7 @@ struct ep_pqueue { struct ep_send_events_data { int maxevents; struct epoll_event __user *events; + bool have_more; int res; };@@ -1783,14 +1768,17 @@ static __poll_t ep_send_events_proc(structeventpoll *ep, struct
list_head*head
}static int ep_send_events(struct eventpoll *ep,
- structepoll_event__user*events, int
maxevents)
+ structepoll_event__user*events, int
maxevents,(quoted diff) +9
-6+ bool *have_more) { - struct ep_send_events_data esed; - - esed.maxevents = maxevents; - esed.events = events; + struct ep_send_events_data esed = { + .maxevents = maxevents, + .events = events, + }; ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false); + *have_more = esed.have_more; + return esed.res; }@@ -1827,7 +1815,7 @@ 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; + bool waiter = false, have_more; wait_queue_entry_t wait; ktime_t expires, *to = NULL;@@ -1927,7 +1915,8 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, * more luck. */ if (!res && eavail && - !(res = ep_send_events(ep, events, maxevents)) && !timed_out) + !(res = ep_send_events(ep, events, maxevents, &have_more))&&
(quoted diff) +7
-0+ !timed_out) goto fetch_events; if (waiter) {@@ -1935,6 +1924,12 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, __remove_wait_queue(&ep->wq, &wait); spin_unlock_irq(&ep->wq.lock); } + /* + * We were not able to process all the events, so immediately + * wakeup other waiter. + */ + if (res > 0 && have_more && waitqueue_active(&ep->wq)) + wake_up(&ep->wq);Quoted 8 lines
Quoted 4 lines
fs/eventpoll.c +0
-0return res;}
Ok, yeah I like making it explicit. Looks like you are missing the changes to
ep_scan_ready_list(), but I think the general approach makes sense.Yeah, missed the hunk:
(quoted diff) +0
-0@@ -1719,8 +1704,10 @@ static __poll_t ep_send_events_proc(structeventpoll *ep, struct
list_head*headlockdep_assert_held(&ep->mtx); list_for_each_entry_safe(epi, tmp, head, rdllink) {(quoted diff) +3
-1- if (esed->res >= esed->maxevents) + if (esed->res >= esed->maxevents) { + esed->have_more = true; break; + }Quoted 5 lines
Although I really didn't have a test case that motivated this - its just was sort of noting this change in behavior while reviewing the current change.
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
PS. So what we decide with the original patch? Remove the whole branch?
For fwiw, I'm ok removing the whole branch as you proposed.
Then probably Heiher could resend once more. Heiher, can you, please?
Sorry for delay.
Thank you for your help! That's ok, I will re-send the patch and add unit-tests to
kselftests in the later patches.
Quoted 7 lines
Quoted 6 lines
And I think the above change can go in separately (if we decide we want it). I don't
think they need to be tied together. I also want to make sure this change gets a full linux-next cycle, so I think it should target 5.5 at this point.
Quoted 7 lines
Sure, this explicit ->wq wakeup is a separate patch, which should be covered with some benchmarks. I can try to cook something out in order to get numbers.
--
Roman
--
Best regards!
Hev
https://hev.cc
On 2019-10-07 20:43, Jason Baron wrote:
[...]
Quoted 76 lines
Quoted 75 lines
But what if to make this wakeup explicit if we have more events to process? (nothing is tested, just a guess)
(quoted diff) +1
-0@@ -255,6 +255,7 @@ struct ep_pqueue { struct ep_send_events_data { int maxevents; struct epoll_event __user *events; + bool have_more; int res; };@@ -1783,14 +1768,17 @@ static __poll_t ep_send_events_proc(structeventpoll *ep, struct
list_head*head
}static int ep_send_events(struct eventpoll *ep,
- structepoll_event__user*events, int maxevents)
+ structepoll_event__user*events, int maxevents,
+ bool*have_more)
{
- structep_send_events_dataesed;
-
- esed.maxevents = maxevents;
- esed.events = events;
+ structep_send_events_dataesed = {
+ .maxevents = maxevents,
+ .events = events,
+ };
ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false); +*have_more= esed.have_more;
+
return esed.res;
}(quoted diff) +1
-1@@ -1827,7 +1815,7 @@ 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; + bool waiter = false, have_more; wait_queue_entry_t wait; ktime_t expires, *to = NULL;@@ -1927,7 +1915,8 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, * more luck. */if (!res && eavail &&
- !(res =ep_send_events(ep, events, maxevents)) && !timed_out)
+ !(res =ep_send_events(ep, events, maxevents, &have_more)) &&
+ !timed_out)
gotofetch_events;if (waiter) {
(quoted diff) +6
-0@@ -1935,6 +1924,12 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, __remove_wait_queue(&ep->wq, &wait); spin_unlock_irq(&ep->wq.lock); } + /* + * We were not able to process all the events, so immediately + * wakeup other waiter. + */ + if (res > 0 && have_more && waitqueue_active(&ep->wq)) + wake_up(&ep->wq); return res;}
[...]
Quoted 1 line
And I think the above change can go in separately (if we decide we want it).
Hi Jason,
I did measurements using Eric's test http://yhbt.net/eponeshotmt.c
(8 writers, 8 waiters; 1 writer, 8 waiters) and tested the impact of outrunning wakeup: I do not see any difference. Since write events are constantly coming, next waiter will be woken up anyway by the following write event. In order to have some perf gain probably writes should happen with some interval: produce bunch of events, sleep, produce bunch of events, sleep, etc, which seems can bring something only if writer is accidentally synchronized with waiters. Not a clean way of perf improvement.
--
Roman