Patchwork [0.15.0] qcow2: Fix L1 table size after bdrv_snapshot_goto

login
register
mail settings
Submitter Kevin Wolf
Date Aug. 4, 2011, 4:17 p.m.
Message ID <1312474673-2705-1-git-send-email-kwolf@redhat.com>
Download mbox | patch
Permalink /patch/108536/
State New
Headers show

Comments

Kevin Wolf - Aug. 4, 2011, 4:17 p.m.
When loading an internal snapshot whose L1 table is smaller than the current L1
table, the size of the current L1 would be shrunk to the snapshot's L1 size in
memory, but not on disk. This lead to incorrect refcount updates and eventuelly
to image corruption.

Instead of writing the new L1 size to disk, this simply retains the bigger L1
size that is currently in use and makes sure that the unused part is zeroed.

Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---

Philipp, I think this should fix your corruption. Please give it a try.

Anthony, this must go into 0.15. Given the short time until -rc2, do you prefer
to pick it up directly or should I send a pull request tomorrow? The patch
looks obvious, is tested with the given testcase and survives a basic
qemu-iotests run (though qemu-iotests doesn't exercise snapshots a lot)

Stefan, please review :-)

 block/qcow2-snapshot.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

Patch

diff --git a/block/qcow2-snapshot.c b/block/qcow2-snapshot.c
index 74823a5..4a8868a 100644
--- a/block/qcow2-snapshot.c
+++ b/block/qcow2-snapshot.c
@@ -330,8 +330,9 @@  int qcow2_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
     if (qcow2_grow_l1_table(bs, sn->l1_size, true) < 0)
         goto fail;
 
-    s->l1_size = sn->l1_size;
+    memset(s->l1_table + sn->l1_size, 0, s->l1_size - sn->l1_size);
     l1_size2 = s->l1_size * sizeof(uint64_t);
+
     /* copy the snapshot l1 table to the current l1 table */
     if (bdrv_pread(bs->file, sn->l1_table_offset,
                    s->l1_table, l1_size2) != l1_size2)