Re: [PATCH 2/8] md/raid5: Ensure a batch member is not handled prematurely.

[Shaohua Li <shli@kernel.org>]

On Sat, May 23, 2015 at 10:26:40AM +1000, NeilBrown wrote:
> On Fri, 22 May 2015 16:44:02 -0700 Shaohua Li <shli@kernel.org> wrote:
> 
> > On Fri, May 22, 2015 at 03:30:58PM +1000, NeilBrown wrote:
> > > If a stripe is a member of a batch, but not the head, it must
> > > not be handled separately from the rest of the batch.
> > > 
> > > 'clear_batch_ready()' handles this requirement to some
> > > extent but not completely.  If a member is passed to handle_stripe()
> > > a second time it returns '0' indicating the stripe can be handled,
> > > which is wrong.
> > > So add an extra test.
> > > 
> > > Signed-off-by: NeilBrown <neilb@suse.de>
> > > ---
> > >  drivers/md/raid5.c |    6 +++++-
> > >  1 file changed, 5 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
> > > index c3ccefbd4fe7..9a803b735848 100644
> > > --- a/drivers/md/raid5.c
> > > +++ b/drivers/md/raid5.c
> > > @@ -4192,9 +4192,13 @@ static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
> > >  
> > >  static int clear_batch_ready(struct stripe_head *sh)
> > >  {
> > > +	/* Return '1' if this is a member of batch, or
> > > +	 * '0' if it is a lone stripe or a head which can now be
> > > +	 * handled.
> > > +	 */
> > >  	struct stripe_head *tmp;
> > >  	if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
> > > -		return 0;
> > > +		return (sh->batch_head && sh->batch_head != sh);
> > >  	spin_lock(&sh->stripe_lock);
> > >  	if (!sh->batch_head) {
> > >  		spin_unlock(&sh->stripe_lock);
> > 
> > which case can this happen in?
> 
> It definitely happens as I had reliable problems until I added this fix.
> 'retry_aligned_read()' can call handle_stripe() on any stripe at any time,
> but I doubt that would apply.  I might try putting a warn-on there and see if
> it provides any hints.
> 
> > 
> > Patches look good. But I'm not in Fusionio any more, so can't check the
> > performance in big raid array with fast flash cards. I'm doing some tests here.
> > I hit a warning in break_stripe_batch_list, STRIPE_BIT_DELAY is set in the
> > stripe state. I'm checking the reason, but if you have thoughts I can try
> > immediately, please let me know.
> 
> I got STRIPE_BIT_DELAY a few times.  That was the main reason for
> 
>   md/raid5: ensure whole batch is delayed for all required bitmap updates.
> 
> and they went away after I got that patch right.
> 
> Maybe there is a race in there..
> 
> If you can reproduce it, maybe WARN whenever STRIPE_BIT_DELAY gets set on a
> stripe with ->batch_head.

Ok, there is a race in add_stripe_bio(). We unlocked the stripe_lock to set the
BIT_DELAY. After the unlock, the stripe might be added to a batch,
stripe_add_to_batch_list didn't clear the bit. Holding the lock in
add_stripe_bio and checking ->batch_head again when we set the bit should fix
the issue.

And STRIPE_ON_UNPLUG_LIST and STRIPE_ON_RELEASE_LIST are set is legit in
break_stripe_batch_list(), they should be removed from the WARN_ON_ONCE().

Thanks,
Shaohua