[Qemu-devel] savevm/loadvm

[Alexey Kardashevskiy <aik@ozlabs.ru>]

Hi!

I need the community help with savevm/loadvm.

I run QEMU like this:

./qemu-system-ppc64 \
 -drive file=virtimg/fc19_16GB.qcow2 \
 -nodefaults \
 -m "2048" \
 -machine "pseries" \
 -nographic \
 -vga "none" \
 -enable-kvm


The disk image is an 16GB qcow2 image.

Now I start the guest and do "savevm 1" and "loadvm 1" from the qemu
console. Everything works. Then I exit qemu, make sure that the snapshot is
there and run QEMU as above plus "-loadvm 1". It fails with:

qemu-system-ppc64: qcow2: Loading snapshots with different disk size is not
implemented
qemu-system-ppc64: Error -95 while activating snapshot '2' on 'scsi0-hd0'

The check is added by commit 90b277593df873d3a2480f002e2eb5fe1f8e5277
"qcow2: Save disk size in snapshot header".

As I cannot realize the whole idea of the patch, I looked a bit deeper.
This is the check:

int qcow2_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
{
[...]
    if (sn->disk_size != bs->total_sectors * BDRV_SECTOR_SIZE) {
        error_report("qcow2: Loading snapshots with different disk "
            "size is not implemented");
        ret = -ENOTSUP;
        goto fail;
    }


My understanding of the patch was that the disk_size should remain 16GB
(0x4.0000.0000) as it uses bs->total_sectors and never changes it. And
bs->growable is 0 for qcow2 image because it is not really growable. At
least the total_sectors value from the qcow2 file header does not change
between QEMU starts.

However qcow2_save_vmstate() sets bs->growable to 1 for a short time
(commit 178e08a58f40dd5aef2ce774fe0850f5d0e56918 from 2009) and this
triggers a branch in bdrv_co_do_writev() which changes bs->total_sectors.
So when QEMU writes snapshots to the file, the disk_size field of a
snapshot has bigger value (for example 0x4.007b.8180).

And the check above fails. It does not fail if to do "loadvm"
_in_the_same_run_ after "savevm" because QEMU operates with the updated
bs->total_sectors.

What the proper fix would be? Or it is not a bug at all and I should be
using something else for "-loadvm"? Thanks.



-- 
Alexey