Re: about mirror cancel

[Max Reitz <mreitz@redhat.com>]

On 15.04.21 20:46, Vladimir Sementsov-Ogievskiy wrote:
> Hi all!
> 
> Recently I've implemented fast-cancelling of mirror job: do 
> bdrv_cancel_in_flight() in mirror_cancel().
> 
> Now I'm in doubt: is it a correct thing? I heard, that mirror-cancel is 
> a kind of valid mirror completion..
> 
> Looking at documentation:
> 
> # Note that if you issue 'block-job-cancel' after 'drive-mirror' has 
> indicated
> # (via the event BLOCK_JOB_READY) that the source and destination are
> # synchronized, then the event triggered by this command changes to
> # BLOCK_JOB_COMPLETED, to indicate that the mirroring has ended and the
> # destination now has a point-in-time copy tied to the time of the 
> cancellation.
> 
> So, in other words, do we guarantee something to the user, if it calls 
> mirror-cancel in ready state? Does this abuse exist in libvirt?

How is it abuse it if it’s documented?  AFAIR it does exist, because 
libvirt’s blockcopy always uses mirror (with --pivot it’s allowed to 
complete, without it is cancelled).

(And the point of course is that if you want mirror to create a copy 
without pivoting, you need this behavior, because otherwise the target 
may be in an inconsistent state.)

> ====
> 
> Note, that if cancelling all in-flight requests on target is wrong on 
> mirror cancel, we still don't have real bug, as the only implementation 
> of .bdrv_cancel_in_flight is stopping reconnect waiting in nbd driver. 
> So, we'll cancel requests only if connection is already lost anyway.
> 
> But that probably means, that correct name of the handler would be 
> .bdrv_cancel_in_fligth_requests_that_will_most_probably_fail_anyway()..

It’s documentation states that it should cancel all in-flight requests, 
so it’s more likely it just isn’t implemented where it could be.

> And it also means, that abuse of mirror-cancel as valid completion is a 
> bad idea, as we can't distinguish the cases when user calls job-cancel 
> to have a kind of point-in-time copy, or just to cancel job (and being 
> not interested in the final state of target).
> 
> So, probably we need an option boolean argument for blockjob-cancel, 
> like "hard", that will cancel in-flight writes on target node..

There is @force.  See commit b76e4458b1eb3c3.

Max