[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

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



On Tue, Jun 23, 2020 at 10:19:03AM +0200, Roger Pau Monné wrote:
> CAUTION: This email originated from outside of the organization. Do not click 
> links or open attachments unless you can confirm the sender and know the 
> content is safe.
> 
> 
> 
> On Tue, Jun 23, 2020 at 12:43:14AM +0000, Anchal Agarwal wrote:
> > On Mon, Jun 22, 2020 at 10:38:46AM +0200, Roger Pau Monné wrote:
> > > CAUTION: This email originated from outside of the organization. Do not 
> > > click links or open attachments unless you can confirm the sender and 
> > > know the content is safe.
> > >
> > >
> > >
> > > On Fri, Jun 19, 2020 at 11:43:12PM +0000, Anchal Agarwal wrote:
> > > > On Wed, Jun 17, 2020 at 10:35:28AM +0200, Roger Pau Monné wrote:
> > > > > CAUTION: This email originated from outside of the organization. Do 
> > > > > not click links or open attachments unless you can confirm the sender 
> > > > > and know the content is safe.
> > > > >
> > > > >
> > > > >
> > > > > On Tue, Jun 16, 2020 at 09:49:25PM +0000, Anchal Agarwal wrote:
> > > > > > On Thu, Jun 04, 2020 at 09:05:48AM +0200, Roger Pau Monné wrote:
> > > > > > > CAUTION: This email originated from outside of the organization. 
> > > > > > > Do not click links or open attachments unless you can confirm the 
> > > > > > > sender and know the content is safe.
> > > > > > > On Wed, Jun 03, 2020 at 11:33:52PM +0000, Agarwal, Anchal wrote:
> > > > > > > >  CAUTION: This email originated from outside of the 
> > > > > > > > organization. Do not click links or open attachments unless you 
> > > > > > > > can confirm the sender and know the content is safe.
> > > > > > > >     > +             xenbus_dev_error(dev, err, "Freezing timed 
> > > > > > > > out;"
> > > > > > > >     > +                              "the device may become 
> > > > > > > > inconsistent state");
> > > > > > > >
> > > > > > > >     Leaving the device in this state is quite bad, as it's in a 
> > > > > > > > closed
> > > > > > > >     state and with the queues frozen. You should make an 
> > > > > > > > attempt to
> > > > > > > >     restore things to a working state.
> > > > > > > >
> > > > > > > > You mean if backend closed after timeout? Is there a way to 
> > > > > > > > know that? I understand it's not good to
> > > > > > > > leave it in this state however, I am still trying to find if 
> > > > > > > > there is a good way to know if backend is still connected after 
> > > > > > > > timeout.
> > > > > > > > Hence the message " the device may become inconsistent state".  
> > > > > > > > I didn't see a timeout not even once on my end so that's why
> > > > > > > > I may be looking for an alternate perspective here. may be need 
> > > > > > > > to thaw everything back intentionally is one thing I could 
> > > > > > > > think of.
> > > > > > >
> > > > > > > You can manually force this state, and then check that it will 
> > > > > > > behave
> > > > > > > correctly. I would expect that on a failure to disconnect from the
> > > > > > > backend you should switch the frontend to the 'Init' state in 
> > > > > > > order to
> > > > > > > try to reconnect to the backend when possible.
> > > > > > >
> > > > > > From what I understand forcing manually is, failing the freeze 
> > > > > > without
> > > > > > disconnect and try to revive the connection by unfreezing the
> > > > > > queues->reconnecting to backend [which never got diconnected]. May 
> > > > > > be even
> > > > > > tearing down things manually because I am not sure what state will 
> > > > > > frontend
> > > > > > see if backend fails to to disconnect at any point in time. I 
> > > > > > assumed connected.
> > > > > > Then again if its "CONNECTED" I may not need to tear down 
> > > > > > everything and start
> > > > > > from Initialising state because that may not work.
> > > > > >
> > > > > > So I am not so sure about backend's state so much, lets say if  
> > > > > > xen_blkif_disconnect fail,
> > > > > > I don't see it getting handled in the backend then what will be 
> > > > > > backend's state?
> > > > > > Will it still switch xenbus state to 'Closed'? If not what will 
> > > > > > frontend see,
> > > > > > if it tries to read backend's state through 
> > > > > > xenbus_read_driver_state ?
> > > > > >
> > > > > > So the flow be like:
> > > > > > Front end marks XenbusStateClosing
> > > > > > Backend marks its state as XenbusStateClosing
> > > > > >     Frontend marks XenbusStateClosed
> > > > > >     Backend disconnects calls xen_blkif_disconnect
> > > > > >        Backend fails to disconnect, the above function returns EBUSY
> > > > > >        What will be state of backend here?
> > > > >
> > > > > Backend should stay in state 'Closing' then, until it can finish
> > > > > tearing down.
> > > > >
> > > > It disconnects the ring after switching to connected state too.
> > > > > >        Frontend did not tear down the rings if backend does not 
> > > > > > switches the
> > > > > >        state to 'Closed' in case of failure.
> > > > > >
> > > > > > If backend stays in CONNECTED state, then even if we mark it 
> > > > > > Initialised in frontend, backend
> > > > >
> > > > > Backend will stay in state 'Closing' I think.
> > > > >
> > > > > > won't be calling connect(). {From reading code in frontend_changed}
> > > > > > IMU, Initialising will fail since backend dev->state != 
> > > > > > XenbusStateClosed plus
> > > > > > we did not tear down anything so calling talk_to_blkback may not be 
> > > > > > needed
> > > > > >
> > > > > > Does that sound correct?
> > > > >
> > > > > I think switching to the initial state in order to try to attempt a
> > > > > reconnection would be our best bet here.
> > > > >
> > > > It does not seems to work correctly, I get hung tasks all over and all 
> > > > the
> > > > requests to filesystem gets stuck. Backend does shows the state as 
> > > > connected
> > > > after xenbus_dev_suspend fails but I think there may be something 
> > > > missing.
> > > > I don't seem to get IO interrupts thereafter i.e hitting the function 
> > > > blkif_interrupts.
> > > > I think just marking it initialised may not be the only thing.
> > > > Here is a short description of what I am trying to do:
> > > > So, on timeout:
> > > >     Switch XenBusState to "Initialized"
> > > >     unquiesce/unfreeze the queues and return
> > > >     mark info->connected = BLKIF_STATE_CONNECTED
> > >
> > > If xenbus state is Initialized isn't it wrong to set info->connected
> > > == CONNECTED?
> > >
> > Yes, you are right earlier I was marking it explicitly but that was not 
> > right,
> > the connect path for blkfront will do that.
> > > You should tear down all the internal state (like a proper close)?
> > >
> > Isn't that similar to disconnecting in the first place that failed during
> > freeze? Do you mean re-try to close but this time re-connect after close
> > basically do everything you would at "restore"?
> 
> Last time I checked blkfront supported reconnections (ie: disconnect
> from a backend and connect again). I was assuming we could apply the
> same here on timeout, and just follow the same path where the frontend
> waits indefinitely for the backend to close and then attempts to
> reconnect.
> 
> > Also, I experimented with that and it works intermittently. I want to take a
> > step back on this issue and ask few questions here:
> > 1. Is fixing this recovery a blocker for me sending in a V2 version?
> 
> At the end of day it's your feature. I would certainly prefer for it
> to work as good as possible, this being a recovery in case of failure
> just make sure it does something sane (ie: crash/close the frontend)
> and add a TODO note.
> 
> > 2. In our 2-3 years of supporting this feature at large scale we haven't 
> > seen this issue
> > where backend fails to disconnect. What we are trying to do here is create a
> > hypothetical situation where we leave backend in Closing state and try and 
> > see how it
> > recovers. The reason why I think it "may not" occur and the timeout of 5HZ 
> > is
> > sufficient is because we haven't come across even a single use-case where it
> > caused hibernation to fail.
> > The reason why I think "it may" occur is if we are running a really memory
> > intensive workload and ring is busy and is unable to complete all the 
> > requests
> > in the given timeout. This is very unlikely though.
> 
> As said above I would generally prefer for code to handle possible
> failures the best way, and hence I think here it would be nice to
> fallback to the normal disconnect path and just wait for the backend
> to close.
>
Do you mind throwing some light in here, what that path may be, if its
straight forward to fix I would like to debug it a bit more. May be I am
missing some of the context here.

I was of the view we may just want to mark frontend closed which should do 
the job of freeing resources and then following the same flow as
blkfront_restore. That does not seems to work correctly 100% of the time.

> You likely have this very well tuned to your own environment and
> workloads, since this will now be upstream others might have more
> contended systems where it could start to fail.
> 
I agree, however, this is also from the testing I did with 100 of runs 
outside of EC2 running few tests of my own. 
> > 3) Also, I do not think this may be straight forward to fix and expect
> > hibernation to work flawlessly in subsequent invocations. I am open to
> > all suggestions.
> 
> Right, adding a TODO would seem appropriate then.
>
Just to double check, I will send in a V2 with this marked as TO-DO?
> Roger.

Thanks,
Anchal



 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.