Re: [RFC PATCH v3 06/12] xen-blkfront: add callbacks for PM suspend and hibernation

From: Roger Pau MonnÃ
Date: Thu Feb 20 2020 - 03:46:24 EST


Thanks for this work, please see below.

On Wed, Feb 19, 2020 at 06:04:24PM +0000, Anchal Agarwal wrote:
> On Tue, Feb 18, 2020 at 10:16:11AM +0100, Roger Pau Monnà wrote:
> > On Mon, Feb 17, 2020 at 11:05:53PM +0000, Anchal Agarwal wrote:
> > > On Mon, Feb 17, 2020 at 11:05:09AM +0100, Roger Pau Monnà wrote:
> > > > On Fri, Feb 14, 2020 at 11:25:34PM +0000, Anchal Agarwal wrote:
> > > > > From: Munehisa Kamata <kamatam@xxxxxxxxxx
> > > > >
> > > > > Add freeze, thaw and restore callbacks for PM suspend and hibernation
> > > > > support. All frontend drivers that needs to use PM_HIBERNATION/PM_SUSPEND
> > > > > events, need to implement these xenbus_driver callbacks.
> > > > > The freeze handler stops a block-layer queue and disconnect the
> > > > > frontend from the backend while freeing ring_info and associated resources.
> > > > > The restore handler re-allocates ring_info and re-connect to the
> > > > > backend, so the rest of the kernel can continue to use the block device
> > > > > transparently. Also, the handlers are used for both PM suspend and
> > > > > hibernation so that we can keep the existing suspend/resume callbacks for
> > > > > Xen suspend without modification. Before disconnecting from backend,
> > > > > we need to prevent any new IO from being queued and wait for existing
> > > > > IO to complete.
> > > >
> > > > This is different from Xen (xenstore) initiated suspension, as in that
> > > > case Linux doesn't flush the rings or disconnects from the backend.
> > > Yes, AFAIK in xen initiated suspension backend takes care of it.
> >
> > No, in Xen initiated suspension backend doesn't take care of flushing
> > the rings, the frontend has a shadow copy of the ring contents and it
> > re-issues the requests on resume.
> >
> Yes, I meant suspension in general where both xenstore and backend knows
> system is going under suspension and not flushing of rings.

backend has no idea the guest is going to be suspended. Backend code
is completely agnostic to suspension/resume.

> That happens
> in frontend when backend indicates that state is closing and so on.
> I may have written it in wrong context.

I'm afraid I'm not sure I fully understand this last sentence.

> > > > > +static int blkfront_freeze(struct xenbus_device *dev)
> > > > > +{
> > > > > + unsigned int i;
> > > > > + struct blkfront_info *info = dev_get_drvdata(&dev->dev);
> > > > > + struct blkfront_ring_info *rinfo;
> > > > > + /* This would be reasonable timeout as used in xenbus_dev_shutdown() */
> > > > > + unsigned int timeout = 5 * HZ;
> > > > > + int err = 0;
> > > > > +
> > > > > + info->connected = BLKIF_STATE_FREEZING;
> > > > > +
> > > > > + blk_mq_freeze_queue(info->rq);
> > > > > + blk_mq_quiesce_queue(info->rq);
> > > > > +
> > > > > + for (i = 0; i < info->nr_rings; i++) {
> > > > > + rinfo = &info->rinfo[i];
> > > > > +
> > > > > + gnttab_cancel_free_callback(&rinfo->callback);
> > > > > + flush_work(&rinfo->work);
> > > > > + }
> > > > > +
> > > > > + /* Kick the backend to disconnect */
> > > > > + xenbus_switch_state(dev, XenbusStateClosing);
> > > >
> > > > Are you sure this is safe?
> > > >
> > > In my testing running multiple fio jobs, other test scenarios running
> > > a memory loader works fine. I did not came across a scenario that would
> > > have failed resume due to blkfront issues unless you can sugest some?
> >
> > AFAICT you don't wait for the in-flight requests to be finished, and
> > just rely on blkback to finish processing those. I'm not sure all
> > blkback implementations out there can guarantee that.
> >
> > The approach used by Xen initiated suspension is to re-issue the
> > in-flight requests when resuming. I have to admit I don't think this
> > is the best approach, but I would like to keep both the Xen and the PM
> > initiated suspension using the same logic, and hence I would request
> > that you try to re-use the existing resume logic (blkfront_resume).
> >
> > > > I don't think you wait for all requests pending on the ring to be
> > > > finished by the backend, and hence you might loose requests as the
> > > > ones on the ring would not be re-issued by blkfront_restore AFAICT.
> > > >
> > > AFAIU, blk_mq_freeze_queue/blk_mq_quiesce_queue should take care of no used
> > > request on the shared ring. Also, we I want to pause the queue and flush all
> > > the pending requests in the shared ring before disconnecting from backend.
> >
> > Oh, so blk_mq_freeze_queue does wait for in-flight requests to be
> > finished. I guess it's fine then.
> >
> Ok.
> > > Quiescing the queue seemed a better option here as we want to make sure ongoing
> > > requests dispatches are totally drained.
> > > I should accept that some of these notion is borrowed from how nvme freeze/unfreeze
> > > is done although its not apple to apple comparison.
> >
> > That's fine, but I would still like to requests that you use the same
> > logic (as much as possible) for both the Xen and the PM initiated
> > suspension.
> >
> > So you either apply this freeze/unfreeze to the Xen suspension (and
> > drop the re-issuing of requests on resume) or adapt the same approach
> > as the Xen initiated suspension. Keeping two completely different
> > approaches to suspension / resume on blkfront is not suitable long
> > term.
> >
> I agree with you on overhaul of xen suspend/resume wrt blkfront is a good
> idea however, IMO that is a work for future and this patch series should
> not be blocked for it. What do you think?

It's not so much that I think an overhaul of suspend/resume in
blkfront is needed, it's just that I don't want to have two completely
different suspend/resume paths inside blkfront.

So from my PoV I think the right solution is to either use the same
code (as much as possible) as it's currently used by Xen initiated
suspend/resume, or to also switch Xen initiated suspension to use the
newly introduced code.

Having two different approaches to suspend/resume in the same driver
is a recipe for disaster IMO: it adds complexity by forcing developers
to take into account two different suspend/resume approaches when
there's no need for it.

Thanks, Roger.