Re: [PATCH for-6.1? v2 6/7] mirror: Check job_is_cancelled() earlier

[Max Reitz]

On 03.08.21 16:34, Kevin Wolf wrote:
> Am 26.07.2021 um 16:46 hat Max Reitz geschrieben:
> > We must check whether the job is force-cancelled early in our main loop,
> > most importantly before any `continue` statement.  For example, we used
> > to have `continue`s before our current checking location that are
> > triggered by `mirror_flush()` failing.  So, if `mirror_flush()` kept
> > failing, force-cancelling the job would not terminate it.
> >
> > A job being force-cancelled should be treated the same as the job having
> > failed, so put the check in the same place where we check `s->ret < 0`.
> >
> > Buglink: https://gitlab.com/qemu-project/qemu/-/issues/462
> > Signed-off-by: Max Reitz <mreitz@redhat.com>
> > ---
> >   block/mirror.c | 7 +------
> >   1 file changed, 1 insertion(+), 6 deletions(-)
> >
> > diff --git a/block/mirror.c b/block/mirror.c
> > index 72e02fa34e..46d1a1e5a2 100644
> > --- a/block/mirror.c
> > +++ b/block/mirror.c
> > @@ -993,7 +993,7 @@ static int coroutine_fn mirror_run(Job *job, Error **errp)
> >               mirror_wait_for_any_operation(s, true);
> >           }
> >   
> > -        if (s->ret < 0) {
> > +        if (s->ret < 0 || job_is_cancelled(&s->common.job)) {
> >               ret = s->ret;
> >               goto immediate_exit;
> >           }
> > @@ -1078,8 +1078,6 @@ static int coroutine_fn mirror_run(Job *job, Error **errp)
> >               break;
> >           }
> >   
> > -        ret = 0;
> > -
> >           if (job_is_ready(&s->common.job) && !should_complete) {
> >               delay_ns = (s->in_flight == 0 &&
> >                           cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
> > @@ -1087,9 +1085,6 @@ static int coroutine_fn mirror_run(Job *job, Error **errp)
> >           trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
> >                                     delay_ns);
> >           job_sleep_ns(&s->common.job, delay_ns);
> > -        if (job_is_cancelled(&s->common.job)) {
> > -            break;
> > -        }
> I think it was intentional that the check is here because it means
> skipping the job_sleep_ns() and instead cancelling immediately, and we
> probably still want that. Between your check above and here, the
> coroutine can yield, so cancellation could have been newly requested.

I’m afraid I don’t quite understand.  If cancel is requested in 
job_sleep_ns(), then we will go back to the top of the loop, wait for 
in-flight active requests and then break.  Waiting for the in-flight 
requests seems unnecessary, but does it really make a difference in 
practice?  We don’t start new requests, so it should be legal to wait 
for existing ones to settle, and also I believe someone will have to 
wait for those in-flight requests anyway (when the mirror top node is 
removed).  (The only thing we could do is to cancel the in-flight 
requests, but that is what mirror_cancel() does.)

Looking more at the whole loop, there are a couple of places that can 
yield.  Of course we can check whether the job has been cancelled after 
every single one of them, but that would be a bit strange.  We only 
really need to check before we initiate new requests or want to change 
the state.  I believe the right place to do the check would be after the 
job_pause_point().

And perhaps the active write functions (bdrv_mirror_top_do_write() and 
bdrv_mirror_top_pwritev()) should stop copying to the target if the job 
has been cancelled.

Max

> So have the check in both places, I guess? And a comment to explain why
> neither is redundant.
>
> >           s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
> >       }
> Kevin