Re: [PATCH RFC 2/3] locking/rwsem: clear reader-owner on unlock to reduce false positives

[Lance Yang]

On 2025/6/24 08:26, Masami Hiramatsu (Google) wrote:
> On Thu, 12 Jun 2025 12:19:25 +0800
> Lance Yang <ioworker0@xxxxxxxxx> wrote:
>
> > From: Lance Yang <lance.yang@xxxxxxxxx>
> >
> > When CONFIG_DETECT_HUNG_TASK_BLOCKER is enabled, a stale owner pointer in a
> > reader-owned rwsem can lead to false positives in blocker tracking.
> >
> > To mitigate this, let’s try to clear the owner field on unlock, as a NULL
> > owner is better than a stale one for diagnostics.
>
> Can we merge this to [PATCH 1/3]? It seems that you removed #ifdef and
> remove it. This means in anyway we need the feature enabled by DEBUG_RWSEMS.

Thanks for the feedback! I see your point about the dependency ;)

Personlly, I'd perfer to keep them separate. The reasoning is that
they addreess two distinct things, and I think splitting them makes
this series clearer and easier to review ;)

Patch #1 focuses on "ownership tracking": Its only job is to make
the existing owner-related helpers (rwsem_owner(), is_rwsem_reader_owned())
globally available when blocker tracking is enabled.

Patch #2, on the other hand, is about "reader-owner cleanup": It
introduces a functional change to the unlock path, trying to clear
the owner field for reader-owned rwsems.

Does this reasoning make sense to you?

Thanks,
Lance

> Thanks,
>
> > Signed-off-by: Lance Yang <lance.yang@xxxxxxxxx>
> > ---
> >   kernel/locking/rwsem.c | 10 ++++------
> >   1 file changed, 4 insertions(+), 6 deletions(-)
> >
> > diff --git a/kernel/locking/rwsem.c b/kernel/locking/rwsem.c
> > index 6cb29442d4fc..a310eb9896de 100644
> > --- a/kernel/locking/rwsem.c
> > +++ b/kernel/locking/rwsem.c
> > @@ -205,14 +205,12 @@ bool is_rwsem_reader_owned(struct rw_semaphore *sem)
> >   		return false;
> >   	return rwsem_test_oflags(sem, RWSEM_READER_OWNED);
> >   }
> > -#endif
> >   
> > -#ifdef CONFIG_DEBUG_RWSEMS
> >   /*
> > - * With CONFIG_DEBUG_RWSEMS configured, it will make sure that if there
> > - * is a task pointer in owner of a reader-owned rwsem, it will be the
> > - * real owner or one of the real owners. The only exception is when the
> > - * unlock is done by up_read_non_owner().
> > + * With CONFIG_DEBUG_RWSEMS or CONFIG_DETECT_HUNG_TASK_BLOCKER configured,
> > + * it will make sure that the owner field of a reader-owned rwsem either
> > + * points to a real reader-owner(s) or gets cleared. The only exception is
> > + * when the unlock is done by up_read_non_owner().
> >    */
> >   static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
> >   {
> > --
> > 2.49.0