On Friday, 27 October 2006 03:38, David Chinner wrote:
> On Thu, Oct 26, 2006 at 11:11:29AM +0200, Rafael J. Wysocki wrote:
> > On Thursday, 26 October 2006 10:57, David Chinner wrote:
> > > On Thu, Oct 26, 2006 at 06:18:29PM +1000, Nigel Cunningham wrote:
> > > > As you have them at the moment, the threads seem to be freezing fine.
> > > > The issue I've seen in the past related not to threads but to timer
> > > > based activity. Admittedly it was 2.6.14 when I last looked at it, but
> > > > there used to be a possibility for XFS to submit I/O from a timer when
> > > > the threads are frozen but the bdev isn't frozen. Has that changed?
> > >
> > > I didn't think we've ever done that - periodic or delayed operations
> > > are passed off to the kernel threads to execute. A stack trace
> > > (if you still have it) would be really help here.
> > >
> > > Hmmm - we have a couple of per-cpu work queues as well that are
> > > used on I/O completion and that can, in some circumstances,
> > > trigger new transactions. If we are only flush metadata, then
> > > I don't think that any more I/o will be issued, but I could be
> > > wrong (maze of twisty passages).
> > Well, I think this exactly is the problem, because worker_threads run with
> > PF_NOFREEZE set (as I've just said in another message).
> Ok, so freezing the filesystem is the only way you can prevent
> this as the workqueues are flushed as part of quiescing the filesystem.
Yes, I think so.
Now at last I know what the problem actually is and why we need the freezing
of filesystems, so thanks for helping me understand that. :-)
You never change things by fighting the existing reality.
R. Buckminster Fuller