Re: [PATCH] locking/rwsem: Add up_write_non_owner() for percpu_up_write()

[Oleg Nesterov]

On 04/09, Waiman Long wrote:
>
> On 04/09/2018 07:20 AM, Oleg Nesterov wrote:
> > Hmm. Can you look at lockdep_sb_freeze_release() and lockdep_sb_freeze_acquire()?
>
> These 2 functions are there to deal with the lockdep code.

Plus they clearly document why sem->owner check is not right when it comes
to super_block->s_writers[]. Not only freeze and thaw can be called by
different processes, we need to return to user-space with rwsem held for
writing.

> > At first glance, it would be much better to set sem->owner = current in
> > percpu_rwsem_acquire(), no?
>
> The primary purpose of the owner field is to enable optimistic spinning
> to improve locking performance. So it needs to be set during an
> up_write() call.

Unless, again, the "owner" has to do lockdep_sb_freeze_release() for any
reason.

And please note that percpu_rwsem_release() already clears rw_sem.owner.
It checks CONFIG_RWSEM_SPIN_ON_OWNER, but this is simply because
rw_semaphore->owner doesn't exist otherwise.

> My rwsem debug patch does use it also to check for consistency in the
> use of lock/unlock call. Anyway, I don't think it is right to set it
> again in percpu_rwsem_acquire() if there is no guarantee that the task
> that call percpu_rwsem_acquire will be the one that will do the unlock.

Hmm. Perhaps I missed something, but I think this should be true.

Of course, you need to check "if (!read)", but again, this is what
percpu_rwsem_release() already does.

> I am wondering if it makes sense to do optimistic spinning in the case
> of percpu_rwsem where the unlocker may be a different task.

Again, perhaps I missed something, but see above. percpu_rwsem does not
really differ from the regular rwsem, however its usage in sb->s_writers[]
differs.

Oleg.