Re: [Qemu-devel] Re: [PATCH] qcow2/virtio corruption: Don't allocate the same cluster twice

[Gleb Natapov]

On Wed, May 06, 2009 at 07:52:44PM +0200, Kevin Wolf wrote:
> Avi Kivity schrieb:
> > Kevin Wolf wrote:
> >> Avi Kivity schrieb:
> >>> Also, the second request now depends on the first to update its 
> >>> metadata.  But if the first request fail, it will not update its 
> >>> metadata, and the second request will complete without error and also 
> >>> without updating its metadata.
> >>>     
> >> Hm, right. Need to think about this...
> >>   
> > 
> > I suggest retaining the part where you use inflight l2metas to layout 
> > data contiguously, but change alloc_cluster_link_l2() not to rely on 
> > n_start and nb_available but instead recompute them on completion. 
> > m->nb_clusters should never be zeroed for this to work.
> 
> Is there even a reason why we need to copy the unmodified sectors in
> alloc_cluster_link_l2() and cannot do that in alloc_cluster_offset()
> before we write the new data? Then the callback wouldn't need to mess
> around with figuring out which part must be overwritten and which one
> mustn't.
> 
The reason we need to copy unmodified sectors in alloc_cluster_link_l2()
is exactly to handle concurrent writes into the same cluster. This is
essentially RMW. I don't see why concurrent writes should not work with
the logic in place. There is a bug there currently of cause :) Can
somebody check this patch:


diff --git a/block-qcow2.c b/block-qcow2.c
index 7840634..801d26d 100644
--- a/block-qcow2.c
+++ b/block-qcow2.c
@@ -995,8 +995,8 @@ static int alloc_cluster_link_l2(BlockDriverState *bs, 
uint64_t cluster_offset,
         if(l2_table[l2_index + i] != 0)
             old_cluster[j++] = l2_table[l2_index + i];
 
-        l2_table[l2_index + i] = cpu_to_be64((cluster_offset +
-                    (i << s->cluster_bits)) | QCOW_OFLAG_COPIED);
+        l2_table[l2_index + i] = cpu_to_be64(((cluster_offset +
+                    (i << s->cluster_bits)) | QCOW_OFLAG_COPIED));
      }
 
     if (bdrv_pwrite(s->hd, l2_offset + l2_index * sizeof(uint64_t),
@@ -1005,7 +1005,8 @@ static int alloc_cluster_link_l2(BlockDriverState *bs, 
uint64_t cluster_offset,
         goto err;
 
     for (i = 0; i < j; i++)
-        free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1);
+        free_any_clusters(bs, be64_to_cpu(old_cluster[i]) & ~QCOW_OFLAG_COPIED,
+                          1);
 
     ret = 0;
 err:

--
                        Gleb.