Skip to contentSkip to conversation

pass dio_complete proper offset from finished_one_bio

PATCH/RFC Eric Sandeen 2 participants 3 messages Nov 26, 2006 – Nov 28, 2006 VFS

lore
Share & tools
lore
#0 lore raw

We saw problems w/ xfs doing AIO+DIO into a sparse file.
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=217098

It seemed that xfs was doing "extent conversion" at the wrong offsets, so written regions came up as unwritten (zeros) and stale data was exposed in the region after the write. Thanks to Peter Backes for the very nice testcase.

This also broke xen with blktap over xfs.

Here's what I found:
finished_one_bio calls dio_complete calls xfs_end_io_direct with an offset, but:

	offset = dio->iocb->ki_pos;

so this is the -current- post-IO position, not the IO start point that dio_complete expects. So, xfs converts the the wrong region. Ouch!

XFS seems to be the only filesystem that uses the offset passed to the end_io function, so only it is affected by this as near as I can tell.

However, the "short read" case is probably also wrong, as it is checking:

                       if ((dio->rw == READ) &&
                           ((offset + transferred) > dio->i_size))
                               transferred = dio->i_size - offset;

but offset is the ending IO position in this case too.

This patch seems to fix it up.

Comments?

Thanks,
-Eric

fs/direct-io.c +0 -0
Index: linux-2.6.18/fs/direct-io.c
===================================================================
@@ -244,12 +244,12 @@ static void finished_one_bio(struct dio 
			 */
			spin_unlock_irqrestore(&dio->bio_lock, flags);

- /* Check for short read case */

			transferred = dio->result;

- offset = dio->iocb->ki_pos;
+ offset = dio->iocb->ki_pos - transferred;

+ /* Check for short read case */

			if ((dio->rw == READ) &&
-			    ((offset + transferred) > dio->i_size))
+			    (dio->iocb->ki_pos > dio->i_size))
				transferred = dio->i_size - offset;

			/* check for error in completion path */
#1 lore raw

Good find, Eric.
Looks good to me. I guess, that can explain some problems I saw on SLES10, and never had time to dig in it.
Seems to me, it's been that way for quite a while, though. This code was introduced in 2.6.12. Prior to that, the code in finished_one_bio() did following:

                       dio_complete(dio, dio->block_in_file << 
dio->blkbits,
                                       dio->result);

where block_in_file is "current offset into the underlying file in dio_block units", which is still the post-IO position, not the IO start offset, isn't it?

- Felix

Quoted 59 lines · Eric Sandeen

We saw problems w/ xfs doing AIO+DIO into a sparse file.
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=217098

It seemed that xfs was doing "extent conversion" at the wrong offsets, so written regions came up as unwritten (zeros) and stale data was exposed in the region after the write. Thanks to Peter Backes for the very nice testcase.

This also broke xen with blktap over xfs.

Here's what I found:
finished_one_bio calls dio_complete calls xfs_end_io_direct with an offset, but:

    offset = dio->iocb->ki_pos;

so this is the -current- post-IO position, not the IO start point that dio_complete expects. So, xfs converts the the wrong region. Ouch!

XFS seems to be the only filesystem that uses the offset passed to the end_io function, so only it is affected by this as near as I can tell.

However, the "short read" case is probably also wrong, as it is checking:

                       if ((dio->rw == READ) &&
                           ((offset + transferred) > dio->i_size))
                               transferred = dio->i_size - offset;

but offset is the ending IO position in this case too.

This patch seems to fix it up.

Comments?

Thanks,
-Eric

Index: linux-2.6.18/fs/direct-io.c
===================================================================

fs/direct-io.c +3 -3
@@ -244,12 +244,12 @@ static void finished_one_bio(struct dio 
             */
            spin_unlock_irqrestore(&dio->bio_lock, flags);
 
-            /* Check for short read case */
            transferred = dio->result;
-            offset = dio->iocb->ki_pos;
+            offset = dio->iocb->ki_pos - transferred;
 
+            /* Check for short read case */
            if ((dio->rw == READ) &&
-                ((offset + transferred) > dio->i_size))
+                (dio->iocb->ki_pos > dio->i_size))
                transferred = dio->i_size - offset;
 
            /* check for error in completion path */
#2 lore raw
In reply to Eric Sandeen
Quoted 9 lines · Eric Sandeen

We saw problems w/ xfs doing AIO+DIO into a sparse file.
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=217098

It seemed that xfs was doing "extent conversion" at the wrong offsets, so written regions came up as unwritten (zeros) and stale data was exposed in the region after the write. Thanks to Peter Backes for the very nice testcase.

This also broke xen with blktap over xfs.

Hrmph. Zach's changes in -mm magically made this go away... I was about to submit a proper patch against -mm but it seem to be not needed.

So, now digging around to see why that is, and what exactly "fixed" things.

-Eric