Re: [Qemu-devel] [PATCH] migration: avoid copying ignore-shared ramblock when in incoming migration

[Dr. David Alan Gilbert]

* Catherine Ho (address@hidden) wrote:
> On Tue, 2 Apr 2019 at 20:37, Peter Xu <address@hidden> wrote:
> 
> > On Tue, Apr 02, 2019 at 05:06:15PM +0800, Catherine Ho wrote:
> > > On Tue, 2 Apr 2019 at 15:58, Peter Xu <address@hidden> wrote:
> > >
> > > > On Tue, Apr 02, 2019 at 03:47:16PM +0800, Catherine Ho wrote:
> > > > > Hi Peter Maydell
> > > > >
> > > > > On Tue, 2 Apr 2019 at 11:05, Peter Maydell <address@hidden
> > >
> > > > wrote:
> > > > >
> > > > > > On Tue, 2 Apr 2019 at 09:57, Catherine Ho <
> > address@hidden>
> > > > > > wrote:
> > > > > > > The root cause is the used idx is moved forward after 1st time
> > > > incoming,
> > > > > > and in 2nd time incoming,
> > > > > > > the last_avail_idx will be incorrectly restored from the saved
> > device
> > > > > > state file(not in the ram).
> > > > > > >
> > > > > > > I watched this even on x86 for a virtio-scsi disk
> > > > > > >
> > > > > > > Any ideas for supporting 2nd time, 3rd time... incoming
> > restoring?
> > > > > >
> > > > > > Does the destination end go through reset between the 1st and 2nd
> > > > > >
> > > > > seems not, please see my step below
> > > > >
> > > > > > incoming attempts? I'm not a migration expert, but I thought that
> > > > > > devices were allowed to assume that their state is "state of the
> > > > > > device following QEMU reset" before the start of an incoming
> > > > > > migration attempt.
> > > > > >
> > > > >
> > > > > Here  is my step:
> > > > > 1. start guest normal by qemu with shared memory-backend file
> > > > > 2. stop the vm. save the device state to another file via monitor
> > migrate
> > > > > "exec: cat>..."
> > > > > 3. quit the vm
> > > > > 4. retore the vm by qemu -incoming "exec:cat ..."
> > > > > 5. continue the vm via monito, the 1st incoming works fine
> > > > > 6. quit the vm
> > > > > 7. retore the vm by qemu -incoming "exec:cat ..." for 2nd time
> > > > > 8. continue   -> error happened
> > > > > Actually, this can be fixed by forcely restore the idx by
> > > > > virtio_queue_restore_last_avail_idx(）
> > > > > But I am sure whether it is reasonable.
> > > >
> > > > Yeah I really suspect its validity.
> > > >
> > > > IMHO normal migration streams keep the device state and RAM data
> > > > together in the dumped file, so they always match.
> > > >
> > > > In your shared case, the device states are in the dumped file however
> > > > the RAM data is located somewhere else.  After you quit the VM from
> > > > the 1st incoming migration the RAM is new (because that's a shared
> > > > memory file) and the device data is still old.  They do not match
> > > > already, then I'd say you can't migrate with that any more.
> > > >
> > > > If you want to do that, you'd better take snapshot of the RAM backend
> > > > file if your filesystem supports (or even simpler, to back it up
> > > > before hand) before you start any incoming migration.  Then with the
> > > > dumped file (which contains the device states) and that snapshot file
> > > > (which contains the exact RAM data that matches the device states)
> > > > you'll alway be able to migrate for as many times as you want.
> > > >
> > >
> > > Understood, thanks Peter Xu
> > > Is there any feasible way to indicate the snapshot of the RAM backend
> > file
> > > is
> > > matched with the device data?
> > > >VQ 2 size 0x400 < last_avail_idx 0x1639 - used_idx 0x2688
> > > >Failed to load virtio-scsi:virtio
> > >
> > > Because I thought reporting above error is not so friendly. Could we add
> > a
> > > version id in both RAM backend file and device date file?
> >
> > It would be non-trivial I'd say - AFAIK we don't have an existing way
> > to tag the memory-backend-file content (IIUC that's what you use).
> >
> > And since you mentioned about versioning of these states, I just
> > remembered that even with this you may not be able to get a complete
> > matched state of the VM, because AFAICT actually besides RAM state &
> > device state, you probably also need to consider the disk state as
> > well.  After you started the VM of the 1st incoming, there could be
> > data flushed to the VM backend disk and then that state is changed as
> > well.  So here even if you snapshot the RAM file you'll still lose the
> > disk state IIUC so it could still be broken.  In other words, to make
> > a migration/snapshot to work you'll need to make all these three
> > states to match.
> >
> 
> Yes, thanks
> 
> >
> > Before we discuss further on the topic... could you share me with your
> > requirement first?  I started to get a bit confused now since when I
> > thought about shared mem I was thinking about migrating within the
> > same host to e.g. upgrade the hypervisor but that obviously does not
> > need you to do incoming migration for multiple times.  Then what do
> > you finally want to achieve?
> >
> Actually, I am investigating the ignore-shared capability case support on
> arm64. This feature is used in Kata containers project as "vm template"
>  The rom reset failure is the first bug.
> Ok, now I can confirm that doing incoming migration for multiple times is
> supported. Thanks for the detailed explanation :)

I think the kata case should be OK;  it's just you've got to be
careful of what you're doing once you start splitting the state into
separate chunks.  That's always been true with the disk/ram split and
it just gets a bit hairier when you split the RAM/devices.

Dave

> B.R.
> Catherine
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK