On Tue, Mar 27, 2012 at 12:19:41PM -0400, Vivek Goyal wrote:
> On Tue, Mar 27, 2012 at 12:03:00PM -0400, Christoph Hellwig wrote:
> > On Tue, Mar 27, 2012 at 11:57:59AM -0400, Vivek Goyal wrote:
> > > On Tue, Mar 27, 2012 at 10:31:27AM -0400, Christoph Hellwig wrote:
> > > > Vivek, does CFQ still need any hints for this sort of handoff?
> > > >
> > >
> > > Christoph, I don't understand the issue enough to comment on it.
> > >
> > > Had a quick look at the patch. Looks like some action (writing log), has
> > > been moved to a worker thread. And in some cases (log force triggered
> > > flush, whatever it is), we seem to prefer to do it from the submitter's
> > > context.
> > Yes. This is to workaround the old problem of cfq getting utterly
> > confused if cooperating I/O beeing submitted from different threads.
> > The case in the previous version of this patch was:
> > - thread doing the fsync will write out data, and wait for it
> > - then we'd force the log by kicking a workqueue and waiting for it
> > quite similar to the ext3/4 fsync issues that we had long discussions
> > about.
> Ok, then I think that fundamental issue still remains with CFQ. And there
> is no general solution to recognizing dependency between processes.
> But a specific workaround for ext3/ext4 fsync problem was put by corrado
> long back.
> commit 749ef9f8423054e326f3a246327ed2db4b6d395f
> Author: Corrado Zoccolo <czoccolo@xxxxxxxxx>
> Date: Mon Sep 20 15:24:50 2010 +0200
> cfq: improve fsync performance for small files
> Basically, I think previously journal commits were "WRITE" and were
> showing most likely on async IO tree. And "fsync" IO was synchronous
> and probably showing up on "sync-noidle" tree. CFQ does idling before
> it switches between trees hence transition from one process to other
> was slow.
> Now corrado, changed the IO type from journaling thread to "WRITE_SYNC"
> which makes writes synchronous and sets the REQ_NOIDLE flag. Hence forcing
> "journal" thread to show up on "sync-noidle" tree. I think "fsync" was
> already there so effectively both the processes are on same service tree
> and we don't idle between processes when they are on "sync-noidle" tree.
Oh, that hack. Have a look at the patch Jan Kara just posted to the
ext4 list (jbd: Refine commit writeout logic,
he removes that WRITE_SYNC from the ext4 journal commit to reduce
read latencies in the presence of writes by an order of magnitude.
IOWs, we need to revert the hack that ext4 uses to work around the
CFQ-caused fsync latency problem because it causes much worse read
latency problems on CFQ.
That is, we simply can't fix CFQ idling heuristic deficiencies at the
filesystem level simply by changing the IO classification. All we
can chosen from is bad performance on workload A or bad performance
on workload B, neither of which are particularly appealing.
> So xfs either need to resort to similar optimizaiton where IO type from
> both the process context is of same type or try to do all the IO from
> one process context.
Submiting all the IO from the same context is exactly what this
patch does - it avoids the whole steaming pile of CFQ problems that