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

[Lance Yang]

On 2025/6/24 11:53, Masami Hiramatsu (Google) wrote:
> On Tue, 24 Jun 2025 09:44:55 +0800
> Lance Yang <lance.yang@xxxxxxxxx> wrote:
>
> > 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.
>
> But without clearing the owner, the owner information can be
> broken, right? Since CONFIG_DEBUG_RWSEMS is working as it is,

You're right, the owner info would be broken without the cleanup logic
in patch #2. But ...

> I think those cannot be decoupled. For example, comparing the
> result of both DETECT_HUNG_TASK_BLOCKER and DEBUG_RWSEMS are
> enabled and only DETECT_HUNG_TASK_BLOCKER is enabled, the
> result is different.

The actual blocker tracking for rwsems is only turned on in patch #3.
So, there's no case where the feature is active without the cleanup
logic already being in place.

> > Does this reasoning make sense to you?
>
> Sorry, no. I think "reader-owner cleanup" is a part of "ownership
> tracking" as DEBUG_RWSEMS does (and that keeps consistency of
> the ownership tracking behavior same as DEBUG_RWSEM).

I thought this step-by-step approach was a bit cleaner, since there are
currently only two users for these owner helpers (DEBUG_RWSEMS and
DETECT_HUNG_TASK_BLOCKER).

Anyway, if you still feel strongly that they should be merged, I'm happy
to rework the series as you suggested ;p

Thanks,
Lance

> Thank 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