The Linux Kernel Mailing List
 help / color / mirror / Atom feed
* [PATCH] mm/damon/core: fix damos_walk() hang with zero watermarks interval
@ 2026-11-18 23:50 Yifan Feng
  2026-11-19  1:42 ` SJ Park
  0 siblings, 1 reply; 4+ messages in thread
From: Yifan Feng @ 2026-11-18 23:50 UTC (permalink / raw)
  To: SJ Park, Andrew Morton; +Cc: damon, linux-mm, linux-kernel, Yifan Feng, stable

kdamond_wait_activation() cancels damos_walk() requests while all
schemes are deactivated by their watermarks.  But it returns without
waiting when the watermarks check interval of a deactivated scheme is
zero, which is the default of the DAMON sysfs interface.  The kdamond
main loop then runs with all schemes deactivated.

In that state, kdamond_apply_schemes() finds no scheme to apply and
returns before it calls damos_walk_complete().  The walk request is
neither completed nor cancelled, so damos_walk() waits until the
watermarks activate a scheme again, possibly forever.

For example, with a scheme that has the free_mem_rate metric, zero
interval_us and a high watermark below the current free memory rate,
writing update_schemes_tried_regions to the kdamond state file never
returns and cannot be killed.  The writer holds damon_sysfs_lock, so all
other DAMON sysfs writes including "off" fail with -EBUSY meanwhile.

Do not skip the deactivated schemes when deciding whether
kdamond_apply_schemes() has something to do.  damon_do_apply_schemes()
still skips them, so no action is applied, and the walk is completed
with no region for them, as it already is when another scheme is
activated.

Fixes: bf0eaba0ff9c ("mm/damon/core: implement damos_walk()")
Cc: <stable@vger.kernel.org> # 6.14.x
Signed-off-by: Yifan Feng <evanfengff@gmail.com>
---
QEMU: the write hangs before, returns after; kunit and 13 selftests pass on both (2 need drgn, not run).
Found and written with the help of an AI coding assistant.

 mm/damon/core.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/damon/core.c b/mm/damon/core.c
index b8678a53b0a9..12578fc3280c 100644
--- a/mm/damon/core.c
+++ b/mm/damon/core.c
@@ -3498,11 +3498,11 @@ static noinline_for_stack void kdamond_apply_schemes(struct damon_ctx *c)
 		if (time_before(c->passed_sample_intervals, s->next_apply_sis))
 			continue;
 
+		has_schemes_to_apply = true;
+
 		if (!s->wmarks.activated)
 			continue;
 
-		has_schemes_to_apply = true;
-
 		damos_adjust_quota(c, s);
 	}
 

base-commit: 4405c18c95d2d0f99f6c581ef0ee8591a1aa6ccd
-- 
2.39.3 (Apple Git-146)


^ permalink raw reply related	[flat|nested] 4+ messages in thread

* Re: [PATCH] mm/damon/core: fix damos_walk() hang with zero watermarks interval
  2026-11-18 23:50 [PATCH] mm/damon/core: fix damos_walk() hang with zero watermarks interval Yifan Feng
@ 2026-11-19  1:42 ` SJ Park
  2026-11-20  0:13   ` Yifan Feng
  0 siblings, 1 reply; 4+ messages in thread
From: SJ Park @ 2026-11-19  1:42 UTC (permalink / raw)
  To: Yifan Feng; +Cc: SJ Park, Andrew Morton, damon, linux-mm, linux-kernel, stable

Hello Yifan,

On Wed, 18 Nov 2026 15:50:12 -0800 Yifan Feng <evanfengff@gmail.com> wrote:

> kdamond_wait_activation() cancels damos_walk() requests while all
> schemes are deactivated by their watermarks.  But it returns without
> waiting when the watermarks check interval of a deactivated scheme is
> zero, which is the default of the DAMON sysfs interface.  The kdamond
> main loop then runs with all schemes deactivated.
> 
> In that state, kdamond_apply_schemes() finds no scheme to apply and
> returns before it calls damos_walk_complete().  The walk request is
> neither completed nor cancelled, so damos_walk() waits until the
> watermarks activate a scheme again, possibly forever.
> 
> For example, with a scheme that has the free_mem_rate metric, zero
> interval_us and a high watermark below the current free memory rate,
> writing update_schemes_tried_regions to the kdamond state file never
> returns and cannot be killed.  The writer holds damon_sysfs_lock, so all
> other DAMON sysfs writes including "off" fail with -EBUSY meanwhile.

Thank you for finding and sharing the issue.

> 
> Do not skip the deactivated schemes when deciding whether
> kdamond_apply_schemes() has something to do.  damon_do_apply_schemes()
> still skips them, so no action is applied, and the walk is completed
> with no region for them, as it already is when another scheme is
> activated.
> 
> Fixes: bf0eaba0ff9c ("mm/damon/core: implement damos_walk()")
> Cc: <stable@vger.kernel.org> # 6.14.x
> Signed-off-by: Yifan Feng <evanfengff@gmail.com>
> ---
> QEMU: the write hangs before, returns after; kunit and 13 selftests pass on both (2 need drgn, not run).
> Found and written with the help of an AI coding assistant.
> 
>  mm/damon/core.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/damon/core.c b/mm/damon/core.c
> index b8678a53b0a9..12578fc3280c 100644
> --- a/mm/damon/core.c
> +++ b/mm/damon/core.c
> @@ -3498,11 +3498,11 @@ static noinline_for_stack void kdamond_apply_schemes(struct damon_ctx *c)
>  		if (time_before(c->passed_sample_intervals, s->next_apply_sis))
>  			continue;
>  
> +		has_schemes_to_apply = true;
> +
>  		if (!s->wmarks.activated)
>  			continue;
>  
> -		has_schemes_to_apply = true;
> -
>  		damos_adjust_quota(c, s);
>  	}

But, this could make has_schemes_to_apply has false positive.  How about
calling damos_Walk_cancel() when has_schemes_to_apply is false, like below?

'''
--- a/mm/damon/core.c
+++ b/mm/damon/core.c
@@ -3849,8 +3849,10 @@ static noinline_for_stack void kdamond_apply_schemes(struct damon_ctx *c)
                damos_adjust_quota(c, s);
        }

-       if (!has_schemes_to_apply)
+       if (!has_schemes_to_apply) {
+               damos_walk_cancel(c);
                return;
+       }

        max_region_sz = damon_region_sz_limit(c);
        mutex_lock(&c->walk_control_lock);
'''


Thanks,
SJ

[...]

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH] mm/damon/core: fix damos_walk() hang with zero watermarks interval
  2026-11-19  1:42 ` SJ Park
@ 2026-11-20  0:13   ` Yifan Feng
  2026-11-20  0:49     ` SJ Park
  0 siblings, 1 reply; 4+ messages in thread
From: Yifan Feng @ 2026-11-20  0:13 UTC (permalink / raw)
  To: SJ Park; +Cc: Andrew Morton, damon, linux-mm, linux-kernel, stable

On Wed, 18 Nov 2026 17:42:46 -0800 SJ Park <sj@kernel.org> wrote:

[...]
> But, this could make has_schemes_to_apply has false positive.  How about
> calling damos_Walk_cancel() when has_schemes_to_apply is false, like below?

Thanks for the review.  You are right about the false positive.

I tried your change.  It fixes the hang, but has_schemes_to_apply is also
false whenever no scheme has reached its next_apply_sis yet, so it cancels
the walks of activated schemes too.  In my test (one stat scheme, default
intervals) 18 of 20 update_schemes_tried_regions writes got -ECANCELED,
and two selftests fail: sysfs_update_schemes_tried_regions_wss_estimation.py
and sysfs_update_schemes_tried_regions_hang.py.

What do you think about cancelling only when no scheme is activated?

'''
--- a/mm/damon/core.c
+++ b/mm/damon/core.c
@@ -3492,9 +3492,13 @@ static noinline_for_stack void kdamond_apply_schemes(struct damon_ctx *c)
 	struct damon_target *t;
 	struct damos *s;
 	bool has_schemes_to_apply = false;
+	bool has_activated_schemes = false;
 	unsigned long max_region_sz;
 
 	damon_for_each_scheme(s, c) {
+		if (s->wmarks.activated)
+			has_activated_schemes = true;
+
 		if (time_before(c->passed_sample_intervals, s->next_apply_sis))
 			continue;
 
@@ -3506,8 +3510,12 @@ static noinline_for_stack void kdamond_apply_schemes(struct damon_ctx *c)
 		damos_adjust_quota(c, s);
 	}
 
-	if (!has_schemes_to_apply)
+	if (!has_schemes_to_apply) {
+		/* The walk cannot be completed while no scheme is activated */
+		if (!has_activated_schemes)
+			damos_walk_cancel(c);
 		return;
+	}
 
 	max_region_sz = damon_region_sz_limit(c);
 	mutex_lock(&c->walk_control_lock);
'''

With this the hang is gone, 20 of 20 writes succeed while the scheme is
activated, and kunit and selftests results are the same as on the base
kernel.

I will send v2 with this if you are ok with it.

Thanks,
Yifan

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH] mm/damon/core: fix damos_walk() hang with zero watermarks interval
  2026-11-20  0:13   ` Yifan Feng
@ 2026-11-20  0:49     ` SJ Park
  0 siblings, 0 replies; 4+ messages in thread
From: SJ Park @ 2026-11-20  0:49 UTC (permalink / raw)
  To: Yifan Feng; +Cc: SJ Park, Andrew Morton, damon, linux-mm, linux-kernel, stable

On Thu, 19 Nov 2026 16:13:00 -0800 Yifan Feng <evanfengff@gmail.com> wrote:

> On Wed, 18 Nov 2026 17:42:46 -0800 SJ Park <sj@kernel.org> wrote:
> 
> [...]
> > But, this could make has_schemes_to_apply has false positive.  How about
> > calling damos_Walk_cancel() when has_schemes_to_apply is false, like below?
> 
> Thanks for the review.  You are right about the false positive.
> 
> I tried your change.  It fixes the hang, but has_schemes_to_apply is also
> false whenever no scheme has reached its next_apply_sis yet, so it cancels
> the walks of activated schemes too.  In my test (one stat scheme, default
> intervals) 18 of 20 update_schemes_tried_regions writes got -ECANCELED,
> and two selftests fail: sysfs_update_schemes_tried_regions_wss_estimation.py
> and sysfs_update_schemes_tried_regions_hang.py.

You're right.  Thanks for further testing and sharing the results.

> 
> What do you think about cancelling only when no scheme is activated?
> 
> '''
> --- a/mm/damon/core.c
> +++ b/mm/damon/core.c
> @@ -3492,9 +3492,13 @@ static noinline_for_stack void kdamond_apply_schemes(struct damon_ctx *c)
>  	struct damon_target *t;
>  	struct damos *s;
>  	bool has_schemes_to_apply = false;
> +	bool has_activated_schemes = false;
>  	unsigned long max_region_sz;
>  
>  	damon_for_each_scheme(s, c) {
> +		if (s->wmarks.activated)
> +			has_activated_schemes = true;
> +
>  		if (time_before(c->passed_sample_intervals, s->next_apply_sis))
>  			continue;
>  
> @@ -3506,8 +3510,12 @@ static noinline_for_stack void kdamond_apply_schemes(struct damon_ctx *c)
>  		damos_adjust_quota(c, s);
>  	}
>  
> -	if (!has_schemes_to_apply)
> +	if (!has_schemes_to_apply) {
> +		/* The walk cannot be completed while no scheme is activated */
> +		if (!has_activated_schemes)
> +			damos_walk_cancel(c);
>  		return;
> +	}
>  
>  	max_region_sz = damon_region_sz_limit(c);
>  	mutex_lock(&c->walk_control_lock);
> '''

Looks good to me.

> 
> With this the hang is gone, 20 of 20 writes succeed while the scheme is
> activated, and kunit and selftests results are the same as on the base
> kernel.
> 
> I will send v2 with this if you are ok with it.

Yes, I'm ok with it.  Please wait ~1 day before posting the v2, so that others
can also add comments if they have.


Thanks,
SJ

[...]

^ permalink raw reply	[flat|nested] 4+ messages in thread

end of thread, other threads:[~2026-11-20  0:49 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-18 23:50 [PATCH] mm/damon/core: fix damos_walk() hang with zero watermarks interval Yifan Feng
2026-11-19  1:42 ` SJ Park
2026-11-20  0:13   ` Yifan Feng
2026-11-20  0:49     ` SJ Park

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox