On 07/30/2012 07:57 AM, Paolo Bonzini wrote:
> Il 30/07/2012 15:51, Richard W.M. Jones ha scritto:
>>>> Or just finish up discard support and use the existing fstrim command of
>>>> qemu-ga.  :)
>> What's actually involved to do this?  I noticed that a virtio-scsi on
>> qcow2 v3 device exported to the guest does not appear to support TRIM
>> at all (see attached test script and output).
> 
> The most basic thing to do is to pass -device discard_granularity=NNN,
> where NNN should be the cluster size of the device.  On top of this you
> need to add the ability to do TRIM asynchronously, quite some testing,
> and perhaps improving the disk formats so that they support
> sector-granularity discard.
> 
> Patches for asynchronous TRIM are in the works (the threadpool patches
> on the list are the first step).
> 
> Long term I would like the ability to distinguish
> discard-for-thin-provisioning ("may cause fragmentation, all subsequent
> accesses could be slower") from discard-for-wear-leveling ("only the
> first subsequent access would be slowed down").  This requires changes
> at all levels (host kernel, QEMU, management, guest kernel, and possibly
> guest applications).

Ooh, nice bullet point to add to my upcoming presentation at Linux
Plumber's Conference in a month:
http://summit.linuxplumbersconf.org/lpc-2012/meeting/33/lpc2012-ref-improved-virt-disk-handling/

Does anyone else have some annoyances about large sparse file handling
where improving the kernel would make our life easier, that I should
incorporate into my discussion?

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