Re: [Qemu-devel] Strange virtio regression on mainline and stable-0.10

[Avi Kivity <avi@redhat.com>]

Kevin Wolf wrote:
> Avi Kivity schrieb:
>   
>> Avi Kivity wrote:
>>     
>>> Running the Fedora 10 installer on a virtio disk on current master and 
>>> on v0.10.3 will cause the installer to complain when mounting the 
>>> freshly formatted filesystems.
>>>       
>> The problem is that qcow2 does a read-modify-write on 
>> non-cluster-aligned writes.  So the following sequence triggers the bug:
>> [...]
>>
>> This could be solved by maintaining a hash table of refcounted RMW 
>> copies for the disk.  When reading for a RMW, look up the hash table, if 
>> there's a copy there, use it instead of reading it yourself.
>>
>> We should also avoid the RMW for non-compressed, non-encrypted clusters, 
>> as virtually ALL writes will be misaligned.
>>     
>
> I don't think there is a RMW except for the COW case which is
> unavoidable and obviously happens only once for each cluster. Do you see
> any other places where this happens?
>   

No, I misread the code.  I think the real problem is two parallel writes 
for the same cluster (but different sectors) started concurrently, so 
get_cluster_offset() places them in different clusters.  When the second 
write completes we get unexpected results since the metadata now 
contains a block where on the start of the operation it was unallocated.

We could probably get away with serializing only writes that hit the 
same cluster.  A better approach may be to try to place parallel 
sequential writes contiguously in the allocating case.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.