On Mon, Sep 10, 2018 at 03:22:52PM +0200, Valentin Vidic wrote:
> On Mon, Sep 10, 2018 at 02:45:31PM +0200, Lars Ellenberg wrote:
> > On Sat, Sep 08, 2018 at 09:34:32AM +0200, Valentin Vidic wrote:
> > > On Fri, Sep 07, 2018 at 07:14:59PM +0200, Valentin Vidic wrote:
> > > > In fact the first one is the original code path before I modified
> > > > blkback. The problem is it gets executed async from workqueue so
> > > > it might not always run before the call to drbdadm secondary.
> > >
> > > As the DRBD device gets released only when the last IO request
> > > has finished, I found a way to check and wait for this in the
> > > block-drbd script:
> >
> > > --- block-drbd.orig 2018-09-08 09:07:23.499648515 +0200
> > > +++ block-drbd 2018-09-08 09:28:12.892193649 +0200
> > > @@ -230,6 +230,24 @@
> > > and so cannot be mounted ${m2}${when}."
> > > }
> > >
> > > +wait_for_inflight()
> > > +{
> > > + local dev="$1"
> > > + local inflight="/sys/block/${dev#/dev/}/inflight"
> > > + local rd wr
> > > +
> > > + if ! [ -f "$inflight" ]; then
> > > + return
> > > + fi
> > > +
> > > + while true; do
> > > + read rd wr < $inflight
> > > + if [ "$rd" = "0" -a "$wr" = "0" ]; then
> >
> > If it is "idle" now, but still "open",
> > this will not sleep, and still fail the demotion below.
>
> True, but in this case blkback is holding it open until all
> the writes have finished and the last write closes the device.
> Since fuser can't check blkback this is an approximation that
> seems to work because I don't get any failed drbdadm calls now.
>
> > You try to help it by "waiting forever until it appears to be idle".
> > I suggest to at least limit the retries by iteration or time.
> > And also (or, instead; but you'd potentially get a number of
> > "scary messages" in the logs) add something like:
>
> Ok, should I open a PR to discuss this change further?
>
> > Or, well, yes, fix blkback to not "defer" the final close "too long",
> > if at all possible.
>
> blkback needs to finish the writes on shutdown or I get a fsck errors
> on next boot. Ideally XenbusStateClosed should be delayed until the
> device release but currently it does not seem possible without breaking
> other things.
I can try to take a look at this and attempt to make sure the state is
only changed to closed in blkback _after_ the device has been
released, but it might take me a couple of days to get you a patch.
I'm afraid that other hotplug scripts will also have issues with such
behavior, and we shouldn't force all users of hotplug scripts to add
such workarounds.
Roger.
