Re: [Qemu-devel] [PATCH 05/21] qcow2: Refcount overflow and qcow2_alloc_bytes()

[Eric Blake]

On 11/11/2014 01:22 AM, Max Reitz wrote:
> On 2014-11-10 at 22:12, Eric Blake wrote:
>> On 11/10/2014 06:45 AM, Max Reitz wrote:
>>> qcow2_alloc_bytes() may reuse a cluster multiple times, in which case
>>> the refcount is increased accordingly. However, if this would lead to an
>>> overflow the function should instead just not reuse this cluster and
>>> allocate a new one.
>> So if recount_order is 1 (2 bits per refcount, max refcount of 4
> 
> *max refcount of 3 (0b11)

Oh right, because 0 is special.  Although I think I figured that out...

> 
>> ), and
>> we encounter the same cluster 6 times (say by 5 back-to-back internal
>> snapshots), does this code optimize to only 2 clusters (both with
>> refcount 3) or does it result in each of the last 3 clusters spilling to

...when talking about 3 shares of a cluster.

>> its own 1-ref cluster for a total of 4 clusters?  Short of Benoit's work
>> on deduplication, is there even a way to avoid inefficient use of
>> spilled clusters?
> 
> I'm not sure what you're referring to; maybe I should add that
> qcow2_alloc_bytes() is used for allocating compressed clusters (which
> ideally don't take up a full host cluster), so "reuse" in this context
> just means that several compressed clusters share one host cluster.

No, I was thinking about internal snapshots rather than compressed
clusters (although there's probably some overlap on what happens).

> 
> Maybe you're referring to the following situation: We have the default
> cluster size of 64k. Now we're trying to allocate 16k for each of the
> compressed clusters A, B, C and D. D won't fit into that cluster because
> the maximum refcount is three, so it will be put into a newly allocated
> host cluster. Finally, we're trying to allocate 32k for a compressed
> cluster E, which will then be put into the same cluster as D. We
> therefore have the following allocation (each sub-box representing 16k):
> 
> +---+---+---+---+   +---+---+---+---+
> |A |B | C |   |   | D |   E | |
> +---+---+---+---+   +---+---+---+---+
> 
> whereas the ideal allocation would be:
> 
> +---+---+---+---+   +---+---+---+---+
> |A |B |   E   |   | C | D | | |
> +---+---+---+---+   +---+---+---+---+
> 
> This is a problem, but I think first it's a minor one (just use a
> sufficiently large refcount width if you're going to use compressed
> clusters) and second it's about compressed clusters, whose performance I
> could hardly care less about, frankly.

No, I was envisioning that we have a brand new image with one cluster
allocated (cluster 1 has refcount 1), then 5 times in a row we do
'savevm' to take an internal snapshot.  If I understand your code
correctly, the first two snapshots increase the refcount, so cluster 1
has a refcount of 3. Then the next snapshot can't increase the refcount,
so it instead copies the contents to cluster 2.  The fourth and fifth
snapshots also see that cluster 1 is full, and allocate cluster 3 and 4;
whereas a more efficient usage would increase the refcount of cluster 2
instead of allocating.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

signature.asc
Description: OpenPGP digital signature