Re: [PATCH 3/5] blk-mq: add helper of blk_mq_global_quiesce_wait()

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, Dec 08, 2021 at 02:49:25PM +0200, Sagi Grimberg wrote:
> 
> 
> On 11/30/21 4:33 AM, Ming Lei wrote:
> > On Tue, Nov 23, 2021 at 11:00:45AM +0200, Sagi Grimberg wrote:
> > > 
> > > > > > > > Add helper of blk_mq_global_quiesce_wait() for supporting to quiesce
> > > > > > > > queues in parallel, then we can just wait once if global quiesce wait
> > > > > > > > is allowed.
> > > > > > > 
> > > > > > > blk_mq_global_quiesce_wait() is a poor name... global is scope-less and
> > > > > > > obviously it has a scope.
> > > > > > 
> > > > > > How about blk_mq_shared_quiesce_wait()? or any suggestion?
> > > > > 
> > > > > Shared between what?
> > > > 
> > > > All request queues in one host-wide, both scsi and nvme has such
> > > > requirement.
> > > > 
> > > > > 
> > > > > Maybe if the queue has a non-blocking tagset, it can have a "quiesced"
> > > > > flag that is cleared in unquiesce? then the callers can just continue
> > > > > to iterate but will only wait the rcu grace period once.
> > > > 
> > > > Yeah, that is what these patches try to implement.
> > > 
> > > I was suggesting to "hide" it in the interface.
> > > Maybe something like:
> > > --
> > > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > > index 8799fa73ef34..627b631db1f9 100644
> > > --- a/block/blk-mq.c
> > > +++ b/block/blk-mq.c
> > > @@ -263,14 +263,18 @@ void blk_mq_wait_quiesce_done(struct request_queue *q)
> > >          unsigned int i;
> > >          bool rcu = false;
> > > 
> > > +       if (!q->has_srcu && q->quiesced)
> > > +               return;
> > >          queue_for_each_hw_ctx(q, hctx, i) {
> > >                  if (hctx->flags & BLK_MQ_F_BLOCKING)
> > >                          synchronize_srcu(hctx->srcu);
> > >                  else
> > >                          rcu = true;
> > >          }
> > > -       if (rcu)
> > > +       if (rcu) {
> > >                  synchronize_rcu();
> > > +               q->quiesced = true;
> > > +       }
> > >   }
> > >   EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
> > > 
> > > @@ -308,6 +312,7 @@ void blk_mq_unquiesce_queue(struct request_queue *q)
> > >          } else if (!--q->quiesce_depth) {
> > >                  blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
> > >                  run_queue = true;
> > > +               q->quiesced = false;
> > 
> > Different request queues are passed to blk_mq_wait_quiesce_done() during
> > the iteration, so marking 'quiesced' doesn't make any difference here.
> 
> I actually meant q->tag_set->quiesced, such that the flag will be
> used in the tag_set reference. This way this sharing will be kept
> hidden from the callers.

That looks easy, but not true really from API viewpoint, q->tag_set has different
lifetime which is covered by driver, not mention other request queues may touch
q->tag_set->quiesced from other code path, so things will become much complicated,
also what is the meaning of quiesced state for tagset wide?

Here I prefer to philosophy of 'Worse is better'[1], and simplicity over perfection
by adding one simple helper of blk_mq_shared_quiesce_wait() for improving the case
very easily.

[1] ihttps://ipfs.fleek.co/ipfs/QmXoypizjW3WknFiJnKLwHCnL72vedxjQkDDP1mXWo6uco/wiki/Unix_philosophy.html


Thanks,
Ming




[Index of Archives]     [Linux RAID]     [Linux SCSI]     [Linux ATA RAID]     [IDE]     [Linux Wireless]     [Linux Kernel]     [ATH6KL]     [Linux Bluetooth]     [Linux Netdev]     [Kernel Newbies]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Device Mapper]

  Powered by Linux