Re: [PATCH 4/4] change sb_writers to use percpu_rw_semaphore

[Oleg Nesterov <oleg@redhat.com>]

Hi Jan,

Thanks for your review and sorry for delay, I was on vacation.

On 07/28, Jan Kara wrote:
>
> On Wed 22-07-15 23:15:41, Oleg Nesterov wrote:
> >
> > Perhaps we should also cleanup the usage of ->frozen. It would be
> > better to set/clear (say) SB_FREEZE_WRITE with the corresponding
> > write-lock held. Currently freeze_super() has to set SB_FREEZE_WRITE
> > before sb_wait_write(SB_FREEZE_WRITE) to avoid the race with itself,
> > we can add another state. The "From now on, no new normal writers
> > can start" removed by this patch was not really correct.
>
> The patch looks good, just one question: Why wasn't the above comment
> really correct?

It is not that I think it was wrong, just not 100% accurate even before
this change. "w_writers.frozen = SB_FREEZE_WRITE" itself can't guarantee
that "no new normal writers can start". We do not know when other CPU's
will see the result of this STORE.

> Do you mean it wouldn't be correct after your changes? I
> agree with that.

Yes, yes, this was the actual reason to remove this comment. Sorry for
confusion.

> Also when you'd like to "cleanup the usage of ->frozen", you have to be
> careful no only about races with freeze_super() itself but also about races
> with remount (that's one of the reasons why we use s_umount for protecting
> modifications of ->frozen). So I'm not sure how much we can actually
> improve on code readability...

Yes, me too. Probably I should simply remove this (confusing) part of the
changelog.

> Anyway, you can add:
>
> Reviewed-by: Jan Kara <jack@suse.com>

Thanks!

OK. Now I'll try to actually test this all. Hopefully this week.

Oleg.