From: Tim Van Patten <[email protected]>
__thaw_task() was recently updated to warn if the task being thawed was
part of a freezer cgroup that is still currently freezing:
void __thaw_task(struct task_struct *p)
{
...
if (WARN_ON_ONCE(freezing(p)))
goto unlock;
This has exposed a bug in cgroup1 freezing where when CGROUP_FROZEN is
asserted, the CGROUP_FREEZING bits are not also cleared at the same
time. Meaning, when a cgroup is marked FROZEN it continues to be marked
FREEZING as well. This causes the WARNING to trigger, because
cgroup_freezing() thinks the cgroup is still freezing.
There are two ways to fix this:
1. Whenever FROZEN is set, clear FREEZING for the cgroup and all
children cgroups.
2. Update cgroup_freezing() to also verify that FROZEN is not set.
This patch implements option (2), since it's smaller and more
straightforward.
Signed-off-by: Tim Van Patten <[email protected]>
---
kernel/cgroup/legacy_freezer.c | 8 +++++++-
1 file changed, 7 insertions(+), 1 deletion(-)
diff --git a/kernel/cgroup/legacy_freezer.c b/kernel/cgroup/legacy_freezer.c
index 122dacb3a443..66d1708042a7 100644
--- a/kernel/cgroup/legacy_freezer.c
+++ b/kernel/cgroup/legacy_freezer.c
@@ -66,9 +66,15 @@ static struct freezer *parent_freezer(struct freezer *freezer)
bool cgroup_freezing(struct task_struct *task)
{
bool ret;
+ unsigned int state;
rcu_read_lock();
- ret = task_freezer(task)->state & CGROUP_FREEZING;
+ /* Check if the cgroup is still FREEZING, but not FROZEN. The extra
+ * !FROZEN check is required, because the FREEZING bit is not cleared
+ * when the state FROZEN is reached.
+ */
+ state = task_freezer(task)->state;
+ ret = (state & CGROUP_FREEZING) && !(state & CGROUP_FROZEN);
rcu_read_unlock();
return ret;
--
2.43.0.rc0.421.g78406f8d94-goog
>
> From: Tim Van Patten <[email protected]>
>
> __thaw_task() was recently updated to warn if the task being thawed was
> part of a freezer cgroup that is still currently freezing:
>
> void __thaw_task(struct task_struct *p)
> {
> ...
> if (WARN_ON_ONCE(freezing(p)))
> goto unlock;
>
> This has exposed a bug in cgroup1 freezing where when CGROUP_FROZEN is
> asserted, the CGROUP_FREEZING bits are not also cleared at the same
> time. Meaning, when a cgroup is marked FROZEN it continues to be marked
> FREEZING as well. This causes the WARNING to trigger, because
> cgroup_freezing() thinks the cgroup is still freezing.
>
> There are two ways to fix this:
>
> 1. Whenever FROZEN is set, clear FREEZING for the cgroup and all
> children cgroups.
> 2. Update cgroup_freezing() to also verify that FROZEN is not set.
>
> This patch implements option (2), since it's smaller and more
> straightforward.
>
> Signed-off-by: Tim Van Patten <[email protected]>
> ---
>
> kernel/cgroup/legacy_freezer.c | 8 +++++++-
> 1 file changed, 7 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/cgroup/legacy_freezer.c b/kernel/cgroup/legacy_freezer.c
> index 122dacb3a443..66d1708042a7 100644
> --- a/kernel/cgroup/legacy_freezer.c
> +++ b/kernel/cgroup/legacy_freezer.c
> @@ -66,9 +66,15 @@ static struct freezer *parent_freezer(struct freezer *freezer)
> bool cgroup_freezing(struct task_struct *task)
> {
> bool ret;
> + unsigned int state;
>
> rcu_read_lock();
> - ret = task_freezer(task)->state & CGROUP_FREEZING;
> + /* Check if the cgroup is still FREEZING, but not FROZEN. The extra
> + * !FROZEN check is required, because the FREEZING bit is not cleared
> + * when the state FROZEN is reached.
> + */
> + state = task_freezer(task)->state;
> + ret = (state & CGROUP_FREEZING) && !(state & CGROUP_FROZEN);
> rcu_read_unlock();
>
> return ret;
> --
Tested-by: Mark Hasemeyer <[email protected]>
On Wed, Nov 15, 2023 at 09:20:43AM -0700, Tim Van Patten wrote:
> From: Tim Van Patten <[email protected]>
>
> __thaw_task() was recently updated to warn if the task being thawed was
> part of a freezer cgroup that is still currently freezing:
>
> void __thaw_task(struct task_struct *p)
> {
> ...
> if (WARN_ON_ONCE(freezing(p)))
> goto unlock;
>
> This has exposed a bug in cgroup1 freezing where when CGROUP_FROZEN is
> asserted, the CGROUP_FREEZING bits are not also cleared at the same
> time. Meaning, when a cgroup is marked FROZEN it continues to be marked
> FREEZING as well. This causes the WARNING to trigger, because
> cgroup_freezing() thinks the cgroup is still freezing.
>
> There are two ways to fix this:
>
> 1. Whenever FROZEN is set, clear FREEZING for the cgroup and all
> children cgroups.
> 2. Update cgroup_freezing() to also verify that FROZEN is not set.
>
> This patch implements option (2), since it's smaller and more
> straightforward.
>
> Signed-off-by: Tim Van Patten <[email protected]>
Applied to cgroup/for-6.7-fixes.
Thanks.
--
tejun
> Applied to cgroup/for-6.7-fixes.
>
> Thanks.
>
> --
> tejun
Thanks Tejun!
As this hasn't been merged to Linus's tree yet, do you think you could
Cc: [email protected]?
On Mon, Nov 27, 2023 at 11:19:52AM -0700, Mark Hasemeyer wrote:
> > Applied to cgroup/for-6.7-fixes.
> >
> > Thanks.
> >
> > --
> > tejun
>
> Thanks Tejun!
> As this hasn't been merged to Linus's tree yet, do you think you could
> Cc: [email protected]?
Yeah, I can do that. That'd be for v6.1+ and fix f5d39b020809
("freezer,sched: Rewrite core freezer logic"), right?
Thanks.
--
tejun
> Yeah, I can do that. That'd be for v6.1+ and fix f5d39b020809
> ("freezer,sched: Rewrite core freezer logic"), right?
>
> Thanks.
>
> --
> tejun
Correct. Thank you!