block: fix overlapping multiwrite requests

When request A is a strict superset of request B:

  AAAAAAAA
    BBBB

multiwrite_merge() merges them as follows:

  AABBBB

The tail of request A should have been included:

  AABBBBAA

This patch fixes data loss but this code path is probably rare.  Since
guests cannot assume ordering between in-flight requests, few
applications submit overlapping write requests.

Reported-by: Slava Pestov <sviatoslav.pestov@gmail.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Benoit Canet <benoit@irqsave.net>

diff --git a/block.c b/block.c
index 9c5566b..cb670fd 100644 (file)
--- a/block.c
+++ b/block.c
@@ -4553,6 +4553,12 @@ static int multiwrite_merge(BlockDriverState *bs, BlockRequest *reqs,
             // Add the second request
             qemu_iovec_concat(qiov, reqs[i].qiov, 0, reqs[i].qiov->size);
 
+            // Add tail of first request, if necessary
+            if (qiov->size < reqs[outidx].qiov->size) {
+                qemu_iovec_concat(qiov, reqs[outidx].qiov, qiov->size,
+                                  reqs[outidx].qiov->size - qiov->size);
+            }
+
             reqs[outidx].nb_sectors = qiov->size >> 9;
             reqs[outidx].qiov = qiov;