On Mon, Dec 04, 2023 at 06:13:36PM -0300, Fabiano Rosas wrote:
> Steve Sistare <steven.sist...@oracle.com> writes:
> > Create a separate bootfile for the outgoing and incoming vm, so the block
> > layer can lock the file during the background migration test.  Otherwise,
> > the test fails with:
> >   "Failed to get "write" lock.  Is another process using the image
> >    [/tmp/migration-test-WAKPD2/bootsect]?"
> Hm.. what is the background migration even trying to access on the boot
> disk? @Peter?

I didn't yet notice this patch until you asked, but background snapshot is
not designed to be used like this, afaict.

It should normally be used when someone would like to use "savevm", then
background snapshot makes that snapshot save happen with VM running (live)
and mostly as performant as "savevm" due to page write protections (IOW, it
is not dirty tracking, but wr-protect each page so not writtable at all
until unprotected).

Another difference (from "savevm") is, instead of storing that image onto
the block images, it stores that image also separately just like migrating
with "file:" as of now.

When the dest QEMU starts it'll try to grab the image lock already because
it should never run with src running, just like when "loadvm" QEMU doesn't
assume the QEMU that ran "savevm" will be running.

> This might be a good use for the -snapshot option. It should stop any
> attempt to get the write lock. Not a lot of difference, but slightly
> simpler.

We don't yet have a background-snapshot test case.  If we ever need,
that'll need to be done in two steps: start src, save snapshot into file,
start dest, load from snapshot file.  We just shouldn't boot two together.

Now after two years when I re-read the snapshot code a bit, I didn't even
find where QEMU took the disk snapshots.. logically it should be done at
the start of live background snapshot when VM was dumping device states,
something like bdrv_all_can_snapshot() orshould be needed to make sure all
images support snapshot on its own or it should already fail, and take
snapshots to match the image.

IOW, I don't even think current raw disk would be able to support
background snapshot at all, otherwise if VM is live I don't see a way to
match the image (which is still lively updated by the running VM) to a live
snapshot taken.  Copy the author, Andrey, for this question.

Before that is confirmed, maybe the easiest way is we can go without a
background snapshot test case for suspend vm scenario.


Peter Xu

Reply via email to