[RESEND] Re: [PATCH 1/2] fs/epoll: mark epoll_wait tasks TASK_FREEZABLE

[Sadrul Chowdhury]

On Thu, Sep 24, 2026 at 8:21 AM Sadrul Chowdhury <sadrul@xxxxxxxxxx> wrote:
>
> On Wed, Sep 23, 2026 at 7:24 AM Jan Kara <jack@xxxxxxx> wrote:
> >
> > On Tue 22-09-26 21:15:58, Sadrul Chowdhury wrote:
> > > During system suspend (freeze_processes) and resume (thaw_processes),
> > > threads blocked in epoll_wait() sleep in TASK_INTERRUPTIBLE. When
> > > try_to_freeze_tasks() executes, it must wake up every sleeping thread
> > > with a synthetic signal so the thread can schedule, run try_to_freeze(),
> > > enter TASK_FROZEN, and sleep again. Upon resume, the freezer wakes them
> > > all up simultaneously, only for each thread to re-enter ep_poll(),
> > > observe no ready I/O events, and call schedule() to go back to sleep.
> > >
> > > On systems with numerous event-driven daemons (where hundreds of
> > > threads idle in event loops), this causes a large wakeup stampede and
> > > thousands of redundant context switches per suspend/resume cycle.
> > >
> > > An earlier attempt to make epoll_wait() freezable in commit 1c441e921201
> > > ("epoll: use freezable blocking call") was reverted in commit
> > > c511851de162 ("Revert "epoll: use freezable blocking call"") due to
> > > reports of userspace corruption (Bugzilla #61781). Under the legacy
> > > freezer design, freezable tasks could wake up prematurely while
> > > hardware and peripheral drivers were still suspended.
> > >
> > > With the 2022 core freezer rewrite (commit f5d39b020809 ("freezer,sched:
> > > Rewrite core freezer logic")), the legacy fake-signal mechanism was
> > > replaced with the TASK_FROZEN state. Tasks sleeping in TASK_FREEZABLE
> > > are safely frozen in-place without waking them up, and the scheduler
> > > guarantees they cannot be woken until thaw_processes() explicitly clears
> > > the frozen state.
> > >
> > > Mark sleeping tasks in ep_poll() as TASK_FREEZABLE. Benchmarks on an
> > > ARM64 test device demonstrate that combining this with freezable
> > > select/poll reduces thaw latency by 63% (from 20.64 ms down to 7.55 ms
> > > mean at fixed CPU frequency) and reduces context switches during
> > > freeze/thaw by 87%.
> > >
> > > Signed-off-by: Sadrul Chowdhury <sadrul@xxxxxxxxxx>
> >
> > With similar caveat as on patch 2/2 feel free to add:
> >
> > Reviewed-by: Jan Kara <jack@xxxxxxx>
> >
>
> Thank you for the review!
>
> Regards,
> Sadrul
>
> >                                                                 Honza
> >
> > > ---
> > >  fs/eventpoll.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/fs/eventpoll.c b/fs/eventpoll.c
> > > index 6cbef8b4c2e0..35992f6457bf 100644
> > > --- a/fs/eventpoll.c
> > > +++ b/fs/eventpoll.c
> > > @@ -39,6 +39,7 @@
> > >  #include <linux/rculist.h>
> > >  #include <linux/capability.h>
> > >  #include <net/busy_poll.h>
> > > +#include <linux/freezer.h>
> > >
> > >  #include <trace/hooks/fs.h>
> > >
> > > @@ -1987,7 +1988,7 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
> > >                * the same lock on wakeup ep_poll_callback() side, so it
> > >                * is safe to avoid an explicit barrier.
> > >                */
> > > -             __set_current_state(TASK_INTERRUPTIBLE);
> > > +             __set_current_state(TASK_INTERRUPTIBLE | TASK_FREEZABLE);
> > >
> > >               /*
> > >                * Do the final check under the lock. ep_start/done_scan()
> > > --
> > > 2.55.0.1082.g2b9226bbc0-goog
> > >
> > --
> > Jan Kara <jack@xxxxxxxx>
> > SUSE Labs, CR
From 5cabe92784535d8cd5cc9d4b76d184f3533848b3 Mon Sep 17 00:00:00 2001
From: Sadrul Chowdhury <sadrul@google.com>
Date: Tue, 13 Oct 2026 20:08:34 +0000
Subject: [PATCH 1/2] fs/epoll: mark epoll_wait tasks TASK_FREEZABLE

During system suspend (freeze_processes) and resume (thaw_processes),
threads blocked in epoll_wait() sleep in TASK_INTERRUPTIBLE. When
try_to_freeze_tasks() executes, it must wake up every sleeping thread
with a synthetic signal so the thread can schedule, run try_to_freeze(),
enter TASK_FROZEN, and sleep again. Upon resume, the freezer wakes them
all up simultaneously, only for each thread to re-enter ep_poll(),
observe no ready I/O events, and call schedule() to go back to sleep.

On systems with numerous event-driven daemons (where hundreds of
threads idle in event loops), this causes a large wakeup stampede and
thousands of redundant context switches per suspend/resume cycle.

An earlier attempt to make epoll_wait() freezable in commit 1c441e921201
("epoll: use freezable blocking call") was reverted in commit
c511851de162 ("Revert "epoll: use freezable blocking call"") due to
reports of userspace corruption (Bugzilla #61781). Under the legacy
freezer design, freezable tasks could wake up prematurely while
hardware and peripheral drivers were still suspended.

With the 2022 core freezer rewrite (commit f5d39b020809 ("freezer,sched:
Rewrite core freezer logic")), the legacy fake-signal mechanism was
replaced with the TASK_FROZEN state. Tasks sleeping in TASK_FREEZABLE
are safely frozen in-place without waking them up, and the scheduler
guarantees they cannot be woken until thaw_processes() explicitly clears
the frozen state.

Mark sleeping tasks in ep_poll() as TASK_FREEZABLE. Benchmarks on an
ARM64 test device demonstrate that combining this with freezable
select/poll reduces thaw latency by 63% (from 20.64 ms down to 7.55 ms
mean at fixed CPU frequency) and reduces context switches during
freeze/thaw by 87%.

Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Sadrul Chowdhury <sadrul@google.com>
---
 fs/eventpoll.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index 6cbef8b4c2e0..35992f6457bf 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -39,6 +39,7 @@
 #include <linux/rculist.h>
 #include <linux/capability.h>
 #include <net/busy_poll.h>
+#include <linux/freezer.h>
 
 #include <trace/hooks/fs.h>
 
@@ -1987,7 +1988,7 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
 		 * the same lock on wakeup ep_poll_callback() side, so it
 		 * is safe to avoid an explicit barrier.
 		 */
-		__set_current_state(TASK_INTERRUPTIBLE);
+		__set_current_state(TASK_INTERRUPTIBLE | TASK_FREEZABLE);
 
 		/*
 		 * Do the final check under the lock. ep_start/done_scan()
-- 
2.56.0.385.gd3acb90ef8-goog

From 23b093f0f9f266a80d367227a8efce9953d90a7c Mon Sep 17 00:00:00 2001
From: Sadrul Chowdhury <sadrul@google.com>
Date: Tue, 13 Oct 2026 20:08:35 +0000
Subject: [PATCH 2/2] fs/select: mark select and poll tasks TASK_FREEZABLE

Similar to epoll_wait(), threads sleeping in do_select() and do_poll()
sleep in TASK_INTERRUPTIBLE via poll_schedule_timeout(). During system
suspend and resume, these threads are woken up by the freezer solely to
transition to TASK_FROZEN, and woken again upon thaw only to re-evaluate
file descriptors with no pending activity and sleep again.

A previous attempt to make select() and poll() freezable in commit
9745cdb36da8 ("select: use freezable blocking call") was reverted in
commit 59612d187912 ("Revert "select: use freezable blocking call"")
alongside the epoll revert due to issues caused by the legacy freezer
architecture.

Following the freezer overhaul in commit f5d39b020809 ("freezer,sched:
Rewrite core freezer logic"), tasks in TASK_FREEZABLE are safely frozen
in-place into TASK_FROZEN without running userspace code or experiencing
premature wakeups before resume completes.

Pass TASK_FREEZABLE to poll_schedule_timeout() in do_select() and
do_poll() (TASK_INTERRUPTIBLE | TASK_FREEZABLE) so that polling threads
remain asleep undisturbed across suspend/resume cycles.

Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Sadrul Chowdhury <sadrul@google.com>
---
 fs/select.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/select.c b/fs/select.c
index 8943f20e8300..5cbbe86135bc 100644
--- a/fs/select.c
+++ b/fs/select.c
@@ -601,7 +601,7 @@ static noinline_for_stack int do_select(int n, fd_set_bits *fds, struct timespec
 			to = &expire;
 		}
 
-		if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE,
+		if (!poll_schedule_timeout(&table, TASK_INTERRUPTIBLE | TASK_FREEZABLE,
 					   to, slack))
 			timed_out = 1;
 	}
@@ -961,7 +961,7 @@ static int do_poll(struct poll_list *list, struct poll_wqueues *wait,
 			to = &expire;
 		}
 
-		if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack))
+		if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE | TASK_FREEZABLE, to, slack))
 			timed_out = 1;
 	}
 	return count;
-- 
2.56.0.385.gd3acb90ef8-goog