On Thu, Jun 01, 2023 at 05:35:07PM +0100, Daniel P. Berrangé wrote:
> If non-live the guest won't have dirtied any pages, so I wasn't
> confident there would be sufficient amount of dirty ram to send
> to trigger this. Possibly that's being too paranoid though
I think it makes sense to keep compression tests in white list, as I
mentioned the compressor does have issue with buffer being modified during
compressing. IIRC we used to hit that, so live test makes sense even for
code path coverage. See comment in do_compress_ram_page(), and also
34ab9e9743 ("migration: detect compression and decompression errors").
If we want to start doing this, imho we should make it strict and any live
use case needs to be very well justifed or otherwise should be non-live in
the qest, so we don't easily go back to square 1.
I see that in the latest post that part is still missing. I can do on top
to prepare a patch to document why each use cases need live, and emphasize
that.
--
Peter Xu
