[v2,14/18] block: protect against "torn reads" for guest_block_size > host_block_size

Submitted by Paolo Bonzini on Jan. 26, 2012, 5:22 p.m.

Details

Message ID 1327598569-5199-15-git-send-email-pbonzini@redhat.com
State New
Headers show

Commit Message

Paolo Bonzini Jan. 26, 2012, 5:22 p.m.
When the guest sees a higher alignment than the host, writes may be
done in multiple steps.  So, reads have to be serialized against
overlapping writes, so that the writes look atomic to the guest.
This is true even when O_DIRECT is not in use.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 block.c |   22 ++++++++++++++++++++++
 1 files changed, 22 insertions(+), 0 deletions(-)

Patch hide | download patch | download mbox

diff --git a/block.c b/block.c
index c78ca47..683d4a3 100644
--- a/block.c
+++ b/block.c
@@ -1613,6 +1613,16 @@  static int coroutine_fn bdrv_co_do_readv(BlockDriverState *bs,
                                       get_cluster_size(bs), false);
     }
 
+    /* When the guest sees a higher alignment than the host, writes may be
+     * done in multiple steps.  So, reads have to be serialized against
+     * overlapping writes, so that the writes look atomic to the guest,
+     * even when O_DIRECT is not in use.
+     */
+    if (bs->guest_block_size > bs->host_block_size) {
+        wait_for_overlapping_requests(bs, sector_num, nb_sectors,
+                                      bs->guest_block_size, true);
+    }
+
     tracked_request_begin(&req, bs, sector_num, nb_sectors, false);
 
     if (flags & BDRV_REQ_COPY_ON_READ) {
@@ -3629,6 +3639,18 @@  BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
 void bdrv_set_guest_block_size(BlockDriverState *bs, int align)
 {
     bs->guest_block_size = align;
+    if ((bs->open_flags & BDRV_O_RDWR) &&
+        bs->host_block_size < bs->guest_block_size) {
+        error_report("Host block size is %d, guest block size is %d.  Due to partially\n"
+                     "written sectors, power failures may cause data corruption.%s",
+                     bs->host_block_size, bs->guest_block_size,
+
+                     /* The host block size might not be detected correctly if
+                      * the guest is not using O_DIRECT.  */
+                     (bs->open_flags & BDRV_O_NOCACHE) ? "" : "\n"
+                     "If you think this message is wrong, start the guest with cache=none\n"
+                     "and see if it disappears.");
+    }
 }
 
 void *qemu_blockalign(BlockDriverState *bs, size_t size)