Re: [PATCH 06/19] quota: Remove state_lock

[Jan Kara <jack@suse.cz>]

On Tue 23-11-10 00:31:56, Dmitry wrote:
> On Mon, 22 Nov 2010 22:12:09 +0100, Jan Kara <jack@suse.cz> wrote:
> > On Thu 11-11-10 15:14:25, Dmitry Monakhov wrote:
> > > The only reader which use state_lock is dqget(), and is it serialized
> > > with quota_disable via SRCU. And state_lock doesn't guaranties anything
> > > for that case. All methods which modify quota flags already protected by
> > > dqonoff_mutex. Get rid of useless state_lock.
> >   dq_state_lock has two properties you miss here:
> > a) It guarantees we read a consistent value from a quota state variable.
>   So what, it is not guaranteed anything. Quota make becomes disabled
>   one nanosecond after check.
>   And with srcu_read_lock we can be guarantee what quota will be available
>   for the caller. Even in case of stale value.
  Well, that's not completely my point. The point is that when we
manipulate flags in an non-atomic way (and using &=, |=, ... is a
non-atomic way), we are not really guaranteed what values a reader will see
while the change happens. So in theory the reader could see that all quotas
are disabled during a transition from usrquota enabled to usrquota+grpquota
enabled (I admit a stupid example but you get the point).

> > b) We read an uptodate value from quota state variable.
>    All callers which modify the state already holds onoff_mutex so
>    implicit mem barriers always guaranteed.
  Yes, so eventually the quota state is guaranteed to have a consistent
value but it says nothing about when and what reader not using onoff_mutex
sees.

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR