Hi again,

So, here's another stab at fixing this.  This patch is very much an RFC,
so do not pull it into anything bound for Linus.  ;-)  For those new to
this topic, here is the original posting:  http://lkml.org/lkml/2010/4/1/344

The basic problem is that, when running iozone on smallish files (up to
8MB in size) and including fsync in the timings, deadline outperforms
CFQ by a factor of about 5 for 64KB files, and by about 10% for 8MB
files.  From examining the blktrace data, it appears that iozone will
issue an fsync() call, and will have to wait until it's CFQ timeslice
has expired before the journal thread can run to actually commit data to
disk.

The approach below puts an explicit call into the filesystem-specific
fsync code to yield the disk so that the jbd[2] process has a chance to
issue I/O.  This bring performance of CFQ in line with deadline.

There is one outstanding issue with the patch that Vivek pointed out.
Basically, this could starve out the sync-noidle workload if there is a
lot of fsync-ing going on.  I'll address that in a follow-on patch.  For
now, I wanted to get the idea out there for others to comment on.

Thanks a ton to Vivek for spotting the problem with the initial
approach, and for his continued review.

Cheers,
Jeff

# git diff --stat
 block/blk-core.c            |   15 +++++++++++++++
 block/cfq-iosched.c         |   38 +++++++++++++++++++++++++++++++++++++-
 block/elevator.c            |    8 ++++++++
 fs/ext3/fsync.c             |    5 ++++-
 fs/ext4/fsync.c             |    6 +++++-
 include/linux/backing-dev.h |   13 +++++++++++++
 include/linux/blkdev.h      |    1 +
 include/linux/elevator.h    |    3 +++
 8 files changed, 86 insertions(+), 3 deletions(-)


diff --git a/block/blk-core.c b/block/blk-core.c
index 9fe174d..1be9413 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -323,6 +323,19 @@ void blk_unplug(struct request_queue *q)
 }
 EXPORT_SYMBOL(blk_unplug);
 
+void blk_backing_dev_yield(struct ...
[ message continues ]

Thanks Jeff. Conceptually this appraoch makes lot of sense to me. Higher
layers explicitly telling CFQ not to idle/yield the slice.

My firefox timing test is perfoming much better now.

real	0m15.957s
user	0m0.608s
sys	0m0.165s

real	0m12.984s
user	0m0.602s
sys	0m0.148s

real	0m13.057s
user	0m0.624s
sys	0m0.145s

So we got to take care of two issues now.

- Make it work with dm/md devices also. Somehow shall have to propogate
  this yield semantic down the stack.

- The issue of making sure we don't yield if we are servicing sync-noidle
  tree and there is other IO going on which relies on sync-noidle tree
  idling (as you have already mentioned).



I think it would be good if we also check that cfqq is empty or not. If cfqq
is not empty we don't want to give up slice? But this is a minor point. At
least in case of fsync, we seem to be waiting for all the iozone request
to finish and then calling yield. So cfqq should be empty at this point of
time. 

Vivek
--

[ message continues ]

The way that Jeff set it up, it's completely parallel to eg congestion


That's a bit of a problem. Say the queue has one request left, it'll
service that and the start idling. But we already got this hint that we
should yield, but as it happened before we reached the idling state,
we'll not yield. Perhaps it would be better to mark the cfqq as yielding
if it isn't ready to yield just yet.

-- 
Jens Axboe

--

[ message continues ]

Ok, so various dm targets now need to define "yield_fn" and propogate the

That makes sense. So another cfqq flag in place. As soon as all the
request from cfqq are dispatched, yield the slice instead of idling. And
clear the flag when slice is expiring.

Vivek
--

[ message continues ]

Precisely. The next question would be how to control the yielding. In
this particular case, you want to be yielding to a specific cfqq. IOW,
you essentially want to pass your slide on to that queue. The way the
above is implemented, you could easily just switch to another unrelated
queue. And if that is done, fairness is skewed without helping the
yielding process at all (which was the intention).

-- 
Jens Axboe

--

[ message continues ]

On Thu, Apr 08, 2010 at 04:09:44PM +0200, Jens Axboe wrote:


True. I guess this is relatively simple yield implementation where we are
telling IO scheduler that there is no more IO coming on this context so
don't waste time idling. That's a different thing that after giving up
slice, cfq might schedule in another sequential reader instead of
journalling thread.

Ideally it would be better to be also able to specify who to transfer
remaining slice to. I am not sure if there is an easy way to pass this
info CFQ.

So this implementation might be good first step.

Vivek
--

[ message continues ]

Well, that's true in part.  Prior to this patch, the process would idle,
keeping all other cfq_queues on the system from making progress.  With
this patch, at least *somebody* else makes progress, getting you closer
to running the journal thread that you're blocked on.  Ideally, you'd
want the thread you're waiting on to get disk time next, sure.  You
would have to pass the process information down to the I/O scheduler for
that, and I'm not sure that the file system code knows which process to
hand off to.  Does it?

Do we really want to go down this road at all?  I'm not convinced.

Cheers,
Jeff
--

[ message continues ]

Don't get me wrong, neither am I. I'm just thinking out loud and
pondering. As a general mechanism, yield to a specific cfqq is going to
be tricky and doing a generic yield to signal that _someone_ else must
make progress before we can is better than nothing.

Continuing that train of thought, I don't think we'll ever need full
'yield to X' functionality where 'X' is a really specific target. But
for this fsync case, we know at least what we are yielding to and it
seems like a shame to throw away that information. So you could include
a hint of what to yield to, which cfq could factor in.

Dunno, I need to think a bit about the cleanliness of such an approach.
We can definitely use your patch as a starting point.

-- 
Jens Axboe

--

[ message continues ]

To do so doesn't DM (and MD) need a blk_queue_yield() setter to
establish its own yield_fn?  The established dm_yield_fn would call
blk_yield() for all real devices in a given DM target.  Something like
how blk_queue_merge_bvec() or blk_queue_make_request() allow DM to
provide functional extensions.

I'm not seeing such a yield_fn hook for stacking drivers to use. And
as is, jbd and jbd2 just call blk_yield() directly and there is no way
for the block layer to call into DM.

What am I missing?

Thanks,
Mike
--

[ message continues ]

Nothing, it is I who am missing something (extra code).  When I send out
the next version, I'll add the setter function and ensure that
queue->yield_fn is called from blk_yield.  Hopefully that's not viewed
as upside down.  We'll see.

Thanks for the review, Mike!

-Jeff
--

[ message continues ]

I like the concept, it's definitely useful (and your results amply
demonstrate that). I was thinking if there was a way in through the ioc
itself, rather than bdi -> queue and like you are doing. But I can't


spin_lock_irq() is sufficient here.

-- 
Jens Axboe

--

[ message continues ]

I think, one issue with ioc based approach will be that it will then call
yield operation on all the devices in the system where this context has ever
done any IO. With bdi based approach this call will remain limited to
a smaller set of devices.

Thanks
Vivek
--

[ message continues ]

Oh, you'd want the bdi as well. And as I said, I don't think it was
workable, just trying to think it over and consider potentially other
ways to accomplish this.

At one point I had a patch that did the equivalant of this yield on
being scheduled out on the CPU side, which is probably why I was in the
ioc mindset.

-- 
Jens Axboe

--

[ message continues ]

Which actually brings up the question of whether this needs some
knowledge of whether the journal is on the same device as the file
system!  In such a case, we need not yield.  I think I'll stick my head
in the sand for this one.  ;-)

Cheers,
Jeff
--

[ message continues ]

Yes, that is true. But that should be relatively easy to check.

-- 
Jens Axboe

--

[ message continues ]

Jeff even if journal is not on same device, what harm yielding could do?
Anyway there is no IO on that queue and we are idling. Only side affect is
that yielding process could lose a bit if after fsync it immediately submits
more IO. Because this process has yielded it slice, it is back in the queue
instead of doing more IO in the current slice immediately.

Vivek
--

[ message continues ]

What happens if the journal is on a super fast device, and finishes up
very quickly allowing our process to initiate more I/O within the idle
window?

Cheers,
Jeff
--

[ message continues ]

You lose. :-) But at the same time if journalling devices is not fast
enough, and you if can't submit next IO in idling window, then you are 
unnecessarily keeping the disk idle and preventing others from making
progress.

Vivek
--

[ message continues ]

Probably not needed for this incarnation of the patch, though previous
iterations would Oops on boot w/o this.  If you look through all of the
code in this code path, cfqg == NULL seems to be handled, so it's

OK, thanks!

Jeff
--

[ message continues ]