Re: [PATCH v2] xfs: serialize unaligned dio writes against all other dio writes

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Tue, Apr 16, 2019 at 02:18:31PM -0400, Brian Foster wrote:
> On Tue, Apr 16, 2019 at 08:14:34AM -0700, Darrick J. Wong wrote:
> > On Mon, Mar 25, 2019 at 01:24:48PM -0400, Brian Foster wrote:
> > > XFS applies more strict serialization constraints to unaligned
> > > direct writes to accommodate things like direct I/O layer zeroing,
> > > unwritten extent conversion, etc. Unaligned submissions acquire the
> > > exclusive iolock and wait for in-flight dio to complete to ensure
> > > multiple submissions do not race on the same block and cause data
> > > corruption.
> > > 
> > > This generally works in the case of an aligned dio followed by an
> > > unaligned dio, but the serialization is lost if I/Os occur in the
> > > opposite order. If an unaligned write is submitted first and
> > > immediately followed by an overlapping, aligned write, the latter
> > > submits without the typical unaligned serialization barriers because
> > > there is no indication of an unaligned dio still in-flight. This can
> > > lead to unpredictable results.
> > > 
> > > To provide proper unaligned dio serialization, require that such
> > > direct writes are always the only dio allowed in-flight at one time
> > > for a particular inode. We already acquire the exclusive iolock and
> > > drain pending dio before submitting the unaligned dio. Wait once
> > > more after the dio submission to hold the iolock across the I/O and
> > > prevent further submissions until the unaligned I/O completes. This
> > > is heavy handed, but consistent with the current pre-submission
> > > serialization for unaligned direct writes.
> > > 
> > > Signed-off-by: Brian Foster <bfoster@redhat.com>
> > > Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
> > > ---
> > > 
> > > v2:
> > > - Use dio return value instead of I/O type in wait logic.
> > > - Drop spurious else logic and fix up comments.
> > > v1: https://marc.info/?l=linux-xfs&m=155327356800415&w=2
> > > 
> > >  fs/xfs/xfs_file.c | 27 +++++++++++++++++----------
> > >  1 file changed, 17 insertions(+), 10 deletions(-)
> > > 
> > > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> > > index 770cc2edf777..933d9c467f56 100644
> > > --- a/fs/xfs/xfs_file.c
> > > +++ b/fs/xfs/xfs_file.c
> > > @@ -529,18 +529,17 @@ xfs_file_dio_aio_write(
> > >  	count = iov_iter_count(from);
> > >  
> > >  	/*
> > > -	 * If we are doing unaligned IO, wait for all other IO to drain,
> > > -	 * otherwise demote the lock if we had to take the exclusive lock
> > > -	 * for other reasons in xfs_file_aio_write_checks.
> > > +	 * If we are doing unaligned IO, we can't allow any other overlapping IO
> > > +	 * in-flight at the same time or we risk data corruption. Wait for all
> > > +	 * other IO to drain before we submit. If the IO is aligned, demote the
> > > +	 * iolock if we had to take the exclusive lock in
> > > +	 * xfs_file_aio_write_checks() for other reasons.
> > >  	 */
> > >  	if (unaligned_io) {
> > > -		/* If we are going to wait for other DIO to finish, bail */
> > > -		if (iocb->ki_flags & IOCB_NOWAIT) {
> > > -			if (atomic_read(&inode->i_dio_count))
> > > -				return -EAGAIN;
> > > -		} else {
> > > -			inode_dio_wait(inode);
> > > -		}
> > > +		/* unaligned dio always waits, bail */
> > > +		if (iocb->ki_flags & IOCB_NOWAIT)
> > > +			return -EAGAIN;
> > 
> > Hmm, Dave pointed out on IRC that this looks like we're bailing out with
> > *iolock held.  I took another look at the function and wondered why
> > wouldn't we bail out as soon as we know that we're doing unaligned
> > nowait directio, which is before we take all the locks and such?
> > 
> 
> Yeah, though it doesn't look like that's due to the patch above (though
> I wish I noticed it then :P)..? The above patch basically just removed
> the dio count check in that first hunk.

Yes, the leak was there before this patch came along.  My first impulse
was simply to change it to "ret = -EAGAIN; goto out;" but then I noticed
that now that we no longer have that i_dio_count conditional we might as
well fail without bothering to take any locks at all.

> 
> > --D
> > 
> > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> > index cdcc75735521..c586fd9f244c 100644
> > --- a/fs/xfs/xfs_file.c
> > +++ b/fs/xfs/xfs_file.c
> > @@ -517,7 +517,8 @@ xfs_file_dio_aio_write(
> >         }
> >  
> >         if (iocb->ki_flags & IOCB_NOWAIT) {
> > -               if (!xfs_ilock_nowait(ip, iolock))
> > +               /* unaligned dio always waits, bail */
> > +               if (unaligned_io || !xfs_ilock_nowait(ip, iolock))
> 
> I'd prefer to see the lock on a line of its own, but otherwise I think
> that's reasonable. After the patch above, an unaligned dio is by
> definition going to wait on itself at the very least.

Ok.  The tricky part here is that the side effects are different with
this change -- now we won't break layouts, update mtime, or cancel
security privileges before failing.  Seeing as the write doesn't go
through I don't think that's a big deal, but who knows...?

--D

> Brian
> 
> >                         return -EAGAIN;
> >         } else {
> >                 xfs_ilock(ip, iolock);
> > @@ -536,9 +537,6 @@ xfs_file_dio_aio_write(
> >          * xfs_file_aio_write_checks() for other reasons.
> >          */
> >         if (unaligned_io) {
> > -               /* unaligned dio always waits, bail */
> > -               if (iocb->ki_flags & IOCB_NOWAIT)
> > -                       return -EAGAIN;
> >                 inode_dio_wait(inode);
> >         } else if (iolock == XFS_IOLOCK_EXCL) {
> >                 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
> > 
> > 
> > > +		inode_dio_wait(inode);
> > >  	} else if (iolock == XFS_IOLOCK_EXCL) {
> > >  		xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
> > >  		iolock = XFS_IOLOCK_SHARED;
> > > @@ -548,6 +547,14 @@ xfs_file_dio_aio_write(
> > >  
> > >  	trace_xfs_file_direct_write(ip, count, iocb->ki_pos);
> > >  	ret = iomap_dio_rw(iocb, from, &xfs_iomap_ops, xfs_dio_write_end_io);
> > > +
> > > +	/*
> > > +	 * If unaligned, this is the only IO in-flight. If it has not yet
> > > +	 * completed, wait on it before we release the iolock to prevent
> > > +	 * subsequent overlapping IO.
> > > +	 */
> > > +	if (ret == -EIOCBQUEUED && unaligned_io)
> > > +		inode_dio_wait(inode);
> > >  out:
> > >  	xfs_iunlock(ip, iolock);
> > >  
> > > -- 
> > > 2.17.2
> > >