Am 08.03.2012 18:55, schrieb Avi Kivity: > On 03/08/2012 07:15 PM, Paolo Bonzini wrote: >> While the formats/protocols are free to implement a discard operation >> that does *not* zero the data, providing a common view to the guests is >> the only way to avoid configuration and migration nightmares. (We don't >> want the admin to manually set up discard_zeroes_data/discard_granularity!) >> >> This has a couple of drawbacks: >> >> 1) QEMU effectively will not try to use discard unless discard_zeroes_data >> is true. To do this we could add a flag such as BDRV_ALWAYS_DISCARD, >> which the device models could use if their discard_zeroes_data property is >> set to false. However, I haven't done this, estimating that no strictly >> positive integer could represent the number of people using it. >> >> 2) it may turn a thin-provisioning operation into a full preallocation, >> which is not nice. >> >> 3) it may cost memory for a bounce buffer that only needs to be filled >> with zeroes. >> >> While (3) can be worked around, the only way around the other two, >> unfortunately, is support in the formats and protocols. We will still >> provide device options to opt out of this, but with raw and qed covered >> (+ qcow2 without backing file, and qcow3 in the future) it should not >> be too bad. > > Can't qcow2 with a backing file also be supported? Zero out the first > cluster, and remember it. The following discards can reuse this zero > cluster, as long as it hasn't been overwritten.
qcow2 can't handle clusters that are referenced twice from the same L1 table. This would require a reverse lookup to adjust the QCOW_O_COPIED flags in the L2 tables containing the other references. Kevin
