[PATCH] block: fix bounce_end_io

[Ming Lei]

When bio bounce is involved, one new bio and its io vector are
cloned from the coming bio, which can be one fast-cloned bio
and its io vector can be shared with another bio too, especially
after bio_split() is introduced.

So it is obviously wrong to assume the start index of the original
bio's io vector is zero, which can be any value between 0 and
(bi_max_vecs - 1), especially in case of bio split.

Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxxxxx>
---
 block/bounce.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/block/bounce.c b/block/bounce.c
index 0611aea..1cb5dd3 100644
--- a/block/bounce.c
+++ b/block/bounce.c
@@ -128,12 +128,14 @@ static void bounce_end_io(struct bio *bio, mempool_t *pool)
 	struct bio *bio_orig = bio->bi_private;
 	struct bio_vec *bvec, *org_vec;
 	int i;
+	int start = bio_orig->bi_iter.bi_idx;
 
 	/*
 	 * free up bounce indirect pages used
 	 */
 	bio_for_each_segment_all(bvec, bio, i) {
-		org_vec = bio_orig->bi_io_vec + i;
+		org_vec = bio_orig->bi_io_vec + i + start;
+
 		if (bvec->bv_page == org_vec->bv_page)
 			continue;
 
-- 
1.9.1

> But inlined here so we can continue on list:
> (In reply to Josh Boyer from comment #40)
> > The function that was fixed in 4.2 doesn't exist any longer in
> > 4.3.0-0.rc0.git6.1.fc24.  That kernel corresponds to Linux
> > v4.2-6105-gdd5cdb48edfd which contains commit
> > 8ae126660fddbeebb9251a174e6fa45b6ad8f932, which removed it completely.  So
> > whatever fix was made in dm_merge_bvec doesn't seem to have made it to
> > whatever replaced it.
> 
> The dm core fix to dm_merge_bvec was commit bd4aaf8f9b ("dm: fix
> dm_merge_bvec regression on 32 bit systems").  But I'm not sure there is
> a clear equivalent in the late bio splitting code that replaced block
> core's merge_bvec logic.
> 
> merge_bvec was all about limiting bios (by asking "can/should this page
> be added to this bio?") whereas the late bio splitting is more "build
> the bios as large as possible and worry about splitting later".

IMO, given one vector can only point to one page, there shouldn't
have difference between the two.

> 
> Regardless, this regression needs to be reported to Ming Lin
> <ming.l@xxxxxxxxxxxxxxx>, Jens Axboe and the others involved in
> maintaining the late bio splitting changes in block core.
> 
> Josh and/or Adam: it would _really_ help if the regression test you guys
> are using could be handed-over and/or explained to us.  Is it as simple
> as loading a 32bit with a particular config?  Can you share the guest
> image if it is small enough?

Josh, Adam, would you mind testing the above patch to see if it can fix
your issue?

Thanks,
Ming

> 
> Mike
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/