On Tue, Sep 07, 2010 at 11:04:59AM +0200, Tejun Heo wrote:
> On 09/07/2010 09:29 AM, Dave Chinner wrote:
> > 1. I have had xfstests deadlock twice via #3, once on 2.6.36-rc2,
> > and once on 2.6.36-rc3. This is clearly a regression, but it is not
> > caused by any XFS changes since 2.6.35. From what I can tell from
> > the backtraces I saw was that it appears that the delaying of the
> > data IO completion processing by requeuing does not allow the
> > workqueue to move off the kworker thread. As a result, any work that
> > is still queued on that kworker queue appears to be starved, and
> > hence we never get the log workqueue processed that would allow data
> > IO completion processing to make progress.
> This is puzzling. Queueing order shouldn't have changed. Maybe I
> screwed up queueing order handling of delayed works. Which workqueue
> is this?
The three workqueues are initialised in
They do not use delayed works, the requeuing of interest here
occurs in .../xfs_aops.c::xfs_end_io via
.../xfs_aops.c:xfs_finish_ioend() onto the xfsdatad_workqueue.
> Or better, can you give me a small test case which
> reproduces the problem?
I've seen it twice in about 100 xfstests runs in the past week.
I can't remember the test that tripped over it - 078 I think did
once, and it was a different test the first time - only some tests
use the loopback device. We've never had a reliable reproducer
because of the complexity of the race condition that leads to
> > 2. I have circumstantial evidence that #4 is contributing to
> > several minute long livelocks. This is intertwined with memory
> > reclaim and lock contention, but fundamentally log IO completion
> > processing is being blocked for extremely long periods of time
> > waiting for a kworker thread to start processing them. In this
> > case, I'm creating close to 100,000 inodes every second, and they
> > are getting written to disk. There is a burst of log IO every 3s or
> > so, so the log Io completion is getting queued behind at least tens
> > of thousands of inode IO completion work items. These work
> > completion items are generating lock contention which slows down
> > processing even further. The transaciton subsystem stalls completely
> > while it waits for log IO completion to be processed. AFAICT, this
> > did not happen on 2.6.35.
> Creating the workqueue for log completion w/ WQ_HIGHPRI should solve
So what you are saying is that we need to change the workqueue
creation interface to use alloc_workqueue() with some special set of
flags to make the workqueue behave as we want, and that each
workqueue will require a different configuration? Where can I find
the interface documentation that describes how the different flags
affect the workqueue behaviour?
> > XFS has used workqueues for these "separate processing threads"
> > because they were a simple primitve that provided the separation and
> > isolation guarantees that XFS IO completion processing required.
> > That is, work deferred from one processing queue to another would
> > not block the original queue, and queues can be blocked
> > independently of the processing of other queues.
> Semantically, that property is (or should be) preserved. The
> scheduling properties change tho and if the code has been depending on
> more subtile aspects of work scheduling, it will definitely need to be
> >>From what I can tell of the new kworker thread based implementation,
> > I cannot see how it provides the same work queue separation,
> > blocking and isolation guarantees. If we block during work
> > processing, then anything on the queue for that thread appears to be
> > blocked from processing until the work is unblocked.
> I fail to follow here. Can you elaborate a bit?
Here's what the work function does:
-> run @work
-> trylock returned EAGAIN
-> delay(1); // to stop workqueue spinning chewing up CPU
So basically I'm seeing a kworker thread blocked in delay(1) - it's
appears to be making progress by processing the same work item over and over
again with delay(1) calls between them. The queued log IO completion
is not being processed, even though it is sitting in a queue
> > Hence my main concern is that the new work queue implementation does
> > not provide the same semantics as the old workqueues, and as such
> > re-introduces a class of problems that will cause random hangs and
> > other bad behaviours on XFS filesystems under heavy load.
> I don't think it has that level of fundamental design flaw.
> > Hence, I'd like to know if my reading of the new workqueue code is
> > correct and:
> Probably not.
> > a) if not, understand why the workqueues are deadlocking;
> Yeah, let's track this one down.
> > c) understand how we can prioritise log IO completion
> > processing over data, metadata and unwritten extent IO
> > completion processing; and
> As I wrote above, WQ_HIGHPRI is there for you.
> > d) what can be done before 2.6.36 releases.
> To preserve the original behavior, create_workqueue() and friends
> create workqueues with @max_active of 1, which is pretty silly and bad
> for latency. Aside from fixing the above problems, it would be nice
> to find out better values for @max_active for xfs workqueues. For
Um, call me clueless, but WTF does max_active actually do? It's not
described anywhere, it's clamped to magic numbers ("I really like
512"), etc. AFAICT, it determines whether the work is queued as
delayed work or whether it is put on an active worklist straight
away. However, the lack of documentation describing the behaviour of
the workqueues and why I might want to set a value other than 1 or
the default makes it pretty hard to work out anything for sure...
> most users, using the pretty high default value is okay as they
> usually have much stricter constraint elsewhere (like limited number
> of work_struct), but last time I tried xfs allocated work_structs and
> fired them as fast as it could, so it looked like it definitely needed
> some kind of resasonable capping value.
What part of XFS fired work structures as fast as it could? Queuing
rates are determined completely by the IO completion rates...