Re: [Qemu-devel][PATCH,RFC] Zero cluster dedup

[Kevin Wolf]

Laurent Vivier schrieb:
> Le mercredi 03 septembre 2008 à 00:35 -0700, Shahar Frank a écrit :
>>> Is it really needed to have a shared zero page ?
>>>
>>> When I read qcow_read() and qcow_aio_read_cb() I see:
>>>
>>> if (!cluster_offset) {
>>> ...
>>> memset(buf, 0, 512 * n);
>>> ...
>>> }
>>> and so I think you have just to clear the l2_table entry for the
>> given
>>> clusters (and free them).
>> You are correct if this is a standalone image i.e. without backing
>> file. If the image has a backing file, a zero mapping is interpreted
>> as "read it from the backing file at the same (logical) offset". 
> 
> I think backing file is also qcow2, so if you write all zeros the
> behavior will be the same, and read from the backing file will use also
> memset().

Not if the cluster is already used in the backing file. Then you would
write 0 to the L2 table of the image itself but not to the L2 table of
the backing file (if you did, it would be a bug). So a read would
operate on the old allocation of the backing file rather than on zeros.

Kevin