Re: regression in page writeback

[Wu Fengguang <fengguang.wu@intel.com>]

On Wed, Sep 23, 2009 at 09:59:41AM +0800, Andrew Morton wrote:
> On Wed, 23 Sep 2009 09:45:00 +0800 Wu Fengguang <fengguang.wu@intel.com> wrote:
> 
> > On Wed, Sep 23, 2009 at 09:28:32AM +0800, Andrew Morton wrote:
> > > On Wed, 23 Sep 2009 09:17:58 +0800 Wu Fengguang <fengguang.wu@intel.com> wrote:
> > > 
> > > > On Wed, Sep 23, 2009 at 08:54:52AM +0800, Andrew Morton wrote:
> > > > > On Wed, 23 Sep 2009 08:22:20 +0800 Wu Fengguang <fengguang.wu@intel.com> wrote:
> > > > > 
> > > > > > Jens' per-bdi writeback has another improvement. In 2.6.31, when
> > > > > > superblocks A and B both have 100000 dirty pages, it will first
> > > > > > exhaust A's 100000 dirty pages before going on to sync B's.
> > > > > 
> > > > > That would only be true if someone broke 2.6.31.  Did they?
> > > > > 
> > > > > SYSCALL_DEFINE0(sync)
> > > > > {
> > > > > 	wakeup_pdflush(0);
> > > > > 	sync_filesystems(0);
> > > > > 	sync_filesystems(1);
> > > > > 	if (unlikely(laptop_mode))
> > > > > 		laptop_sync_completion();
> > > > > 	return 0;
> > > > > }
> > > > > 
> > > > > the sync_filesystems(0) is supposed to non-blockingly start IO against
> > > > > all devices.  It used to do that correctly.  But people mucked with it
> > > > > so perhaps it no longer does.
> > > > 
> > > > I'm referring to writeback_inodes(). Each invocation of which (to sync
> > > > 4MB) will do the same iteration over superblocks A => B => C ... So if
> > > > A has dirty pages, it will always be served first.
> > > > 
> > > > So if wbc->bdi == NULL (which is true for kupdate/background sync), it
> > > > will have to first exhaust A before going on to B and C.
> > > 
> > > But that works OK.  We fill the first device's queue, then it gets
> > > congested and sync_sb_inodes() does nothing and we advance to the next
> > > queue.
> > 
> > So in common cases "exhaust" is a bit exaggerated, but A does receive
> > much more opportunity than B. Computation resources for IO submission
> > are unbalanced for A, and there are pointless overheads in rechecking A.
> 
> That's unquantified handwaving.  One CPU can do a *lot* of IO.

Yes.. I had the impression that the writeback submission can be pretty slow.
It should be because of the congestion_wait. Now that it is removed,
things are going faster when queue is not full.

> > > If a device has more than a queue's worth of dirty data then we'll
> > > probably leave some of that dirty memory un-queued, so there's some
> > > lack of concurrency in that situation.
> > 
> > Good insight.
> 
> It was wrong.  See the other email.

No your first insight is correct. Because the (unnecessary) teeny
sleeps is independent of the A=>B=>C traversing order. Only queue
congestion could help skip A.

> > That possibly explains one major factor of the
> > performance gains of Jens' per-bdi writeback.
> 
> I've yet to see any believable and complete explanation for these
> gains.  I've asked about these things multiple times and nothing happened.
 
The per-bdi writeback threads does make things more straight.
But given that Jens also piggy backed some other behavior changes,
it's hard to judge the pure gain of the per-bdi writeback.

> I suspect that what happened over time was that previously-working code
> got broken, then later people noticed the breakage but failed to
> analyse and fix it in favour of simply ripping everything out and
> starting again.
>
> So for the want of analysing and fixing several possible regressions,
> we've tossed away some very sensitive core kernel code which had tens
> of millions of machine-years testing.  I find this incredibly rash.

Sorry..