| Message ID | 1360153926-9492-30-git-send-email-benoit@irqsave.net |
|---|---|
| State | New |
| Headers | show |
On Wed, Feb 06, 2013 at 01:32:02PM +0100, Benoît Canet wrote: > In the case of a race condition between two writes a l2 entry can be written > without QCOW_OFLAG_COPIED before the first write fill it. > This patch simply check if the l2 entry has the correct offset without > QCOW_OFLAG_COPIED and do nothing. > > Signed-off-by: Benoit Canet <benoit@irqsave.net> > --- > block/qcow2-cluster.c | 5 +++++ > 1 file changed, 5 insertions(+) Is this a general qcow2 bug (without deduplication)? I thought there is L2Meta wait code in qcow2_co_writev() that prevents overlapping writes from executing at the same time?
diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c index fedcf57..c016e85 100644 --- a/block/qcow2-cluster.c +++ b/block/qcow2-cluster.c @@ -763,6 +763,11 @@ int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m) for (i = 0; i < m->nb_clusters; i++) { uint64_t flags = 0; uint64_t offset = cluster_offset + (i << s->cluster_bits); + + if (be64_to_cpu(l2_table[l2_index + i]) == offset) { + continue; + } + /* if two concurrent writes happen to the same unallocated cluster * each write allocates separate cluster and writes data concurrently. * The first one to complete updates l2 table with pointer to its
In the case of a race condition between two writes a l2 entry can be written without QCOW_OFLAG_COPIED before the first write fill it. This patch simply check if the l2 entry has the correct offset without QCOW_OFLAG_COPIED and do nothing. Signed-off-by: Benoit Canet <benoit@irqsave.net> --- block/qcow2-cluster.c | 5 +++++ 1 file changed, 5 insertions(+)