[Qemu-devel] [PATCH] Fix (at least one cause of) qcow2 corruption.

[Qemu-devel] [PATCH] Fix (at least one cause of) qcow2 corruption.

[Nolan]

qcow2's get_cluster_offset() scans forward in the l2 table to find other
clusters that have the same allocation status as the first cluster.
This is used by (among others) qcow_is_allocated().

Unfortunately, it was not checking to be sure that it didn't fall off
the end of the l2 table.  This patch adds that check.

The symptom that motivated me to look into this was that
bdrv_is_allocated() was returning false when there was in fact data
there.  This is one of many ways this bug could lead to data corruption.

I checked the other place that scans for consecutive unallocated blocks
(alloc_cluster_offset()) and it appears to be OK:
    nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
appears to prevent the same problem from occurring.

Signed-off-by: Nolan Leake <nolan <at> sigbus.net>

Index: block-qcow2.c
===================================================================
--- block-qcow2.c	(revision 6963)
+++ block-qcow2.c	(working copy)
@@ -761,6 +761,10 @@
 
     nb_available = (nb_available >> 9) + index_in_cluster;
 
+    if (nb_needed > nb_available) {
+        nb_needed = nb_available;
+    }
+
     cluster_offset = 0;
 
     /* seek the the l2 offset in the l1 table */

Re: [Qemu-devel] [PATCH] Fix (at least one cause of) qcow2 corruption.

[Anthony Liguori]

Nolan wrote:
> qcow2's get_cluster_offset() scans forward in the l2 table to find other
> clusters that have the same allocation status as the first cluster.
> This is used by (among others) qcow_is_allocated().
>
> Unfortunately, it was not checking to be sure that it didn't fall off
> the end of the l2 table.  This patch adds that check.
>
> The symptom that motivated me to look into this was that
> bdrv_is_allocated() was returning false when there was in fact data
> there.  This is one of many ways this bug could lead to data corruption.
>
> I checked the other place that scans for consecutive unallocated blocks
> (alloc_cluster_offset()) and it appears to be OK:
>     nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
> appears to prevent the same problem from occurring.
>
> Signed-off-by: Nolan Leake <nolan <at> sigbus.net>
>   

Applied.  Very good work.  Thanks!

Regards,

Anthony Liguori

Re: [Qemu-devel] [PATCH] Fix (at least one cause of) qcow2 corruption.

[Anthony Liguori]

Nolan wrote:
> qcow2's get_cluster_offset() scans forward in the l2 table to find other
> clusters that have the same allocation status as the first cluster.
> This is used by (among others) qcow_is_allocated().
>
> Unfortunately, it was not checking to be sure that it didn't fall off
> the end of the l2 table.  This patch adds that check.
>
> The symptom that motivated me to look into this was that
> bdrv_is_allocated() was returning false when there was in fact data
> there.  This is one of many ways this bug could lead to data corruption.
>
> I checked the other place that scans for consecutive unallocated blocks
> (alloc_cluster_offset()) and it appears to be OK:
>     nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
> appears to prevent the same problem from occurring.
>
> Signed-off-by: Nolan Leake <nolan <at> sigbus.net>
>   
Applied to stable.  Thanks

-- 
Regards,

Anthony Liguori

Re: [Qemu-devel] [PATCH] Fix (at least one cause of) qcow2 corruption.

[Christoph Hellwig]

On Tue, Mar 31, 2009 at 03:40:54PM -0700, Nolan wrote:
> qcow2's get_cluster_offset() scans forward in the l2 table to find other
> clusters that have the same allocation status as the first cluster.
> This is used by (among others) qcow_is_allocated().
> 
> Unfortunately, it was not checking to be sure that it didn't fall off
> the end of the l2 table.  This patch adds that check.
> 
> The symptom that motivated me to look into this was that
> bdrv_is_allocated() was returning false when there was in fact data
> there.  This is one of many ways this bug could lead to data corruption.
> 
> I checked the other place that scans for consecutive unallocated blocks
> (alloc_cluster_offset()) and it appears to be OK:
>     nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
> appears to prevent the same problem from occurring.

Do you have a specific I/O patter nthat would trigger this issue?