Re: [PATCH] splice : two smp_mb() can be omitted

[Jens Axboe <jens.axboe@oracle.com>]

On Tue, Oct 31 2006, Eric Dumazet wrote:
> Jens Axboe a écrit :
> >On Tue, Oct 31 2006, Eric Dumazet wrote:
> >>This patch deletes two calls to smp_mb() that were done after 
> >>mutex_unlock() that contains an implicit memory barrier.
> >>
> >>The first one in splice_to_pipe(), where 'do_wakeup' is set to true only 
> >>if pipe->inode is set (and in this case the
> >>if (pipe->inode)
> >>   mutex_unlock(&pipe->inode->i_mutex);
> >>is done too)
> >>
> >>The second one in link_pipe(), following inode_double_unlock() that 
> >>contains calls to mutex_unlock() too.
> >
> >NAK on that patch, the smp_mb() follows the waitqueue_active(). If you
> >later change the code and move the locks or whatnot, you have lost that
> >connection.
> >
> >If you change the patch to insert a comment, then it may be more
> >applicable.
> >
> 
> Hum... I read fs/pipe.c and see no smp_mb() there, but I suspect same 
> semantics are/were used.
> 
> Should we add comments on fs/pipe.c too ?

fs/pipe.c looks different:

        if (do_wakeup) {
                wake_up_interruptible_sync(&pipe->wait);
                ...
        }

The smp_mb() is not needed if you call wake_up() directly, only if
checking via waitqueue_active().

-- 
Jens Axboe