Messages in this thread | | | Date | Tue, 24 Mar 2009 11:18:47 +0100 | From | Jens Axboe <> | Subject | Re: [PATCH] block: forbid to re-enable I/O stat accounting |
| |
On Sat, Mar 21 2009, Andrew Morton wrote: > > Please fix your email headers - what I received was > > From: Jerome Marchand <jmarchan@redhat.com> > To: unlisted-recipients:;;@imap1.linux-foundation.org (no To-header on input) > Cc: Jens Axboe <axboe@kernel.dk> > > > On Thu, 19 Mar 2009 11:36:50 +0100 Jerome Marchand <jmarchan@redhat.com> wrote: > > > When we stop I/O stat accounting we stop to update the in-flight > > requests counter and we need this counter to be reliable for > > accounting I/O stats. Unfortunately updating in_flight field may > > affect performance. So, until we have a better solution, just forbid > > to re-enable I/O stat accounting after it has been disabled. > > hm. Is it really so hard to just quiesce the device until all in-flight > requests have drained? freeze_bdev() might be a suitable starting point?
freeze_bdev() is too far up, I think. It would be better to just reuse the quiscing code we use for switching io schedulers on-the-fly. Basically something ala:
spin_lock_irq(q->queue_lock); queue_flag_set(QUEUE_FLAG_ELVSWITCH, q); elv_drain_elevator(q); while (q->rq.elvpriv) { blk_start_queueing(q); spin_unlock_irq(q->queue_lock); msleep(10); spin_lock_irq(q->queue_lock); elv_drain_elevator(q); }
re-enable IO accounting; spin_unlock_irq(q->queue_lock);
and then don't account requests that don't have REQ_ELVPRIV set. I think that should be acceptable. I was pretty sure I had a patch for something else that yanked this logic into start/stop helpers... Ah, found it:
http://git.kernel.dk/?p=linux-2.6-block.git;a=commitdiff;h=8cb9f793b82024cd43ed5d4583a5287d002abdf8
-- Jens Axboe
| |