Re: [PATCH] block: flush writeback dwork before detaching a bdev inode from it

[Tejun Heo <tj@kernel.org>]

Hello,

On Mon, Jun 20, 2016 at 03:38:41PM +0200, Dmitry Vyukov wrote:
> > Sorry for the late reply but now when thinking about the patch I don't
> > think it is quite right. Writeback can happen from other contexts than just
> > the worker one (e.g. kswapd can do writeback, or in some out-of-memory
> > situations we may punt to doing writeback directly instead of calling the
> > worker, or sync(2) calls fdatawrite() for block device inode directly when
> > iterating through blockdev superblock). So flushing the workqueue IMHO is
> > not covering 100% of cases.

Hmmm, yeah, the patch undoes what the cgroup writeback changes added
but it looks like the addition was exposing the existing problem more
rather than causing a new one.

> > I wanted to suggest to use inode_wait_for_writeback() which will make sure
> > writeback is done with the inode. However we effectively already do this
> > by calling bdev_write_inode() which calls writeback_single_inode() in
> > WB_SYNC_ALL mode. So by the time that call completes we are sure writeback
> > code is not looking at the inode. *However* what I think is happening is
> > that sync_blockdev() writes all the dirty pages, bdev_write_inode() writes
> > the inode and clears all dirty bits, however the inode still stays in the
> > b_dirty / b_io list of the wb because it has I_DIRTY_TIME set. Subsequently
> > once flusher work runs, it finds the inode, looks at it and boom. And the
> > problem seems to be that write_inode_now(inode, true) does not result in
> > I_DIRTY_TIME being cleared.
> >
> > Attached patch should fix this issue - it is compile-tested only. Dmitry,
> > can you check whether this patch fixes the issue for you as well?
> 
> I can't directly test it because crash happened very infrequently.
> If Tejun/Ted agree that it is the right way to fix it, then I can
> patch it, restart the fuzzer and leave it running for a while.

Yes, please try out the patch.

Thanks a lot.

-- 
tejun