Re: [PATCH] vfs: make evict() use smp_mb__after_spinlock instead of smp_mb
From: Mateusz Guzik
Date: Wed Nov 13 2024 - 11:29:32 EST
On Wed, Nov 13, 2024 at 5:17 PM Jan Kara <jack@xxxxxxx> wrote:
>
> On Wed 13-11-24 16:51:03, Mateusz Guzik wrote:
> > It literally directly follows a spin_lock() call.
> >
> > This whacks an explicit barrier on x86-64.
> >
> > Signed-off-by: Mateusz Guzik <mjguzik@xxxxxxxxx>
>
> Looks good. Feel free to add:
>
> Reviewed-by: Jan Kara <jack@xxxxxxx>
>
thanks
> > This plausibly can go away altogether, but I could not be arsed to
> > convince myself that's correct. Individuals willing to put in time are
> > welcome :)
>
> AFAICS there's nothing else really guaranteeing the last store to
> inode->i_state cannot be reordered up to after the wake up so I think the
> barrier should be there.
>
There is a bunch of lock round trips in this routine alone, including
on i_lock itself, but that aside:
I *suspect* something like spin_wait_unlocked(&inode->i_state)
shipping with a full fence at the beginning of the routine would
correctly allow to check all the possible waiter et al flags without
acquiring the lock anymore, shaving off at least 2 lock trips in the
common case.
However, I don't see such a routine as is and I'm definitely not going
to flame about adding it for the time being.
> Honza
> >
> > fs/inode.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/fs/inode.c b/fs/inode.c
> > index e5a60084a7a9..b3db1234737f 100644
> > --- a/fs/inode.c
> > +++ b/fs/inode.c
> > @@ -817,7 +817,7 @@ static void evict(struct inode *inode)
> > * ___wait_var_event() either sees the bit cleared or
> > * waitqueue_active() check in wake_up_var() sees the waiter.
> > */
> > - smp_mb();
> > + smp_mb__after_spinlock();
> > inode_wake_up_bit(inode, __I_NEW);
> > BUG_ON(inode->i_state != (I_FREEING | I_CLEAR));
> > spin_unlock(&inode->i_lock);
> > --
> > 2.43.0
> >
> --
> Jan Kara <jack@xxxxxxxx>
> SUSE Labs, CR
--
Mateusz Guzik <mjguzik gmail.com>