On Tue, Mar 25, 2014 at 09:05:04AM -0700, Eric Sandeen wrote:
> >> Out of curiosity, is there any major reason we don't use 0 here
> >> unconditionally? Are we worried about I/O completing before we have a
> >> chance to decrement the reference?
> > I think this should unconditionally avoid the schedule, and while we're
> > at it we should kill _xfs_buf_ioend and opencode it here and at the
> > other callsite.
> And then remove the flag from xfs_buf_ioend which is always 0 at that
> point ...
Is it? xfs_buf_bio_end_io should stil be passing 1, the bio end_io
handler is the place we really need the workqueue for anyway.
> Yeah I have a patch to do that as well; I wanted to separate the
> bugfix from the more invasive cleanup, though - and I wanted to
> get the fix out for review sooner.
Sure, feel free to leave all the cleanups to another patch.
> But yeah, I was unsure about whether or not to schedule at all here.
> We come here from a lot of callsites and I'm honestly not sure what
> the implications are yet.
I think the the delayed completion is always wrong from the submission
path. The error path is just a special case of a completion happening
before _xfs_buf_ioapply returns. The combination of incredibly fast
hardware and bad preemption could cause the same bug you observed.