On 2026/10/07 3:39, Peter Xu wrote:
On Tue, Oct 06, 2026 at 06:19:29PM +0300, Andrey Drobyshev wrote:
On 10/6/26 5:14 PM, Peter Xu wrote:
On Tue, Oct 06, 2026 at 04:58:39PM +0300, Andrey Drobyshev wrote:
A device may take over a resource from the source in post_load(), e.g.
vhost-vsock acquires the device ownership on a CPR target. If the
incoming migration fails afterwards, there's currently no way for the
device to give that resource back before the source is told about the
failure, over the return path or via the MIGRATION event, and resumes,
or before the target exits.
Add a notifier list for that, called from the failure path of
process_incoming_migration_co() before the failure is published.
Signed-off-by: Andrey Drobyshev <[email protected]>
Should we better reuse migration_state_notifiers[] and introduce
MIG_EVENT_INCOMING_FAILED?
So far I can't see why this wouldn't work. Should be functionally
identical to what I've done in this and subsequent patches. So if you
think adding a separate event is cleaner - I can certainly go this way.
Yes I think it's cleaner.
Not only because I want to reduce possible global list of notifies only to
be used in migration (we still have another global notifier for postcopy
that only vhost-user uses..), but also in this case you didn't introduce
per-mode notifiers, stop matching with what the source notifiers can do.
When reused, it should then only register it against CPR_TRANSFER|CPR_EXEC,
IIUC.
The logic in the current series looks sound, and I don't see any gaps. I
look forward to reviewing the next version with the suggested
refactoring, and will likely give my Reviewed-by for the entire series then.
Regards,
Akihiko Odaki
