Re: [PATCH 0/6] migration: bring savevm/loadvm/delvm over to QMP

[Daniel P . Berrangé]

On Mon, Jul 06, 2020 at 06:21:56PM +0200, Kevin Wolf wrote:
> Am 06.07.2020 um 18:03 hat Daniel P. Berrangé geschrieben:
> > On Mon, Jul 06, 2020 at 05:50:11PM +0200, Kevin Wolf wrote:
> > > Am 06.07.2020 um 17:29 hat Daniel P. Berrangé geschrieben:
> > > > On Mon, Jul 06, 2020 at 05:27:01PM +0200, Kevin Wolf wrote:
> > > > > Am 03.07.2020 um 19:29 hat Denis V. Lunev geschrieben:
> > > > > > On 7/3/20 8:22 PM, Daniel P. Berrangé wrote:
> > > > > > > On Fri, Jul 03, 2020 at 08:15:44PM +0300, Denis V. Lunev wrote:
> > > > > > >> On 7/2/20 8:57 PM, Daniel P. Berrangé wrote:
> > > > > > >>> When QMP was first introduced some 10+ years ago now, the 
> > > > > > >>> snapshot
> > > > > > >>> related commands (savevm/loadvm/delvm) were not converted. This 
> > > > > > >>> was
> > > > > > >>> primarily because their implementation causes blocking of the 
> > > > > > >>> thread
> > > > > > >>> running the monitor commands. This was (and still is) considered
> > > > > > >>> undesirable behaviour both in HMP and QMP.
> > > > > > >>>
> > > > > > >>> In theory someone was supposed to fix this flaw at some point 
> > > > > > >>> in the
> > > > > > >>> past 10 years and bring them into the QMP world. Sadly, thus 
> > > > > > >>> far it
> > > > > > >>> hasn't happened as people always had more important things to 
> > > > > > >>> work
> > > > > > >>> on. Enterprise apps were much more interested in external 
> > > > > > >>> snapshots
> > > > > > >>> than internal snapshots as they have many more features.
> > > > > > >>>
> > > > > > >>> Meanwhile users still want to use internal snapshots as there is
> > > > > > >>> a certainly simplicity in having everything self-contained in 
> > > > > > >>> one
> > > > > > >>> image, even though it has limitations. Thus the apps that end up
> > > > > > >>> executing the savevm/loadvm/delvm via the 
> > > > > > >>> "human-monitor-command"
> > > > > > >>> QMP command.
> > > > > > >>>
> > > > > > >>>
> > > > > > >>> IOW, the problematic blocking behaviour that was one of the 
> > > > > > >>> reasons
> > > > > > >>> for not having savevm/loadvm/delvm in QMP is experienced by 
> > > > > > >>> applications
> > > > > > >>> regardless. By not portting the commands to QMP due to one 
> > > > > > >>> design flaw,
> > > > > > >>> we've forced apps and users to suffer from other design flaws 
> > > > > > >>> of HMP (
> > > > > > >>> bad error reporting, strong type checking of args, no 
> > > > > > >>> introspection) for
> > > > > > >>> an additional 10 years. This feels rather sub-optimal :-(
> > > > > > >>>
> > > > > > >>> In practice users don't appear to care strongly about the fact 
> > > > > > >>> that these
> > > > > > >>> commands block the VM while they run. I might have seen one bug 
> > > > > > >>> report
> > > > > > >>> about it, but it certainly isn't something that comes up as a 
> > > > > > >>> frequent
> > > > > > >>> topic except among us QEMU maintainers. Users do care about 
> > > > > > >>> having
> > > > > > >>> access to the snapshot feature.
> > > > > > >>>
> > > > > > >>> Where I am seeing frequent complaints is wrt the use of OVMF 
> > > > > > >>> combined
> > > > > > >>> with snapshots which has some serious pain points. This is 
> > > > > > >>> getting worse
> > > > > > >>> as the push to ditch legacy BIOS in favour of UEFI gain 
> > > > > > >>> momentum both
> > > > > > >>> across OS vendors and mgmt apps. Solving it requires new 
> > > > > > >>> parameters to
> > > > > > >>> the commands, but doing this in HMP is super unappealing.
> > > > > > >>>
> > > > > > >>>
> > > > > > >>>
> > > > > > >>> After 10 years, I think it is time for us to be a little 
> > > > > > >>> pragmatic about
> > > > > > >>> our handling of snapshots commands. My desire is that libvirt 
> > > > > > >>> should never
> > > > > > >>> use "human-monitor-command" under any circumstances, because of 
> > > > > > >>> the
> > > > > > >>> inherant flaws in HMP as a protocol for machine consumption. If 
> > > > > > >>> there
> > > > > > >>> are flaws in QMP commands that's fine. If we fix them in 
> > > > > > >>> future, we can
> > > > > > >>> deprecate the current QMP commands and remove them not too long 
> > > > > > >>> after,
> > > > > > >>> without being locked in forever.
> > > > > > >>>
> > > > > > >>>
> > > > > > >>> Thus in this series I'm proposing a direct 1-1 mapping of the 
> > > > > > >>> existing
> > > > > > >>> HMP commands for savevm/loadvm/delvm into QMP as a first step. 
> > > > > > >>> This does
> > > > > > >>> not solve the blocking thread problem, but it does eliminate 
> > > > > > >>> the error
> > > > > > >>> reporting, type checking and introspection problems inherant to 
> > > > > > >>> HMP.
> > > > > > >>> We're winning on 3 out of the 4 long term problems.
> > > > > > >>>
> > > > > > >>> If someone can suggest a easy way to fix the thread blocking 
> > > > > > >>> problem
> > > > > > >>> too, I'd be interested to hear it. If it involves a major 
> > > > > > >>> refactoring
> > > > > > >>> then I think user are better served by unlocking what look like 
> > > > > > >>> easy
> > > > > > >>> wins today.
> > > > > > >>>
> > > > > > >>> With a QMP variant, we reasonably deal with the problems 
> > > > > > >>> related to OVMF:
> > > > > > >>>
> > > > > > >>>  - The logic to pick which disk to store the vmstate in is not
> > > > > > >>>    satsifactory.
> > > > > > >>>
> > > > > > >>>    The first block driver state cannot be assumed to be the 
> > > > > > >>> root disk
> > > > > > >>>    image, it might be OVMF varstore and we don't want to store 
> > > > > > >>> vmstate
> > > > > > >>>    in there.
> > > > > > >>>
> > > > > > >>>  - The logic to decide which disks must be snapshotted is 
> > > > > > >>> hardwired
> > > > > > >>>    to all disks which are writable
> > > > > > >>>
> > > > > > >>>    Again with OVMF there might be a writable varstore, but this 
> > > > > > >>> can be
> > > > > > >>>    raw rather than qcow2 format, and thus unable to be 
> > > > > > >>> snapshotted.
> > > > > > >>>    While users might wish to snapshot their varstore, in 
> > > > > > >>> some/many/most
> > > > > > >>>    cases it is entirely uneccessary. Users are blocked from 
> > > > > > >>> snapshotting
> > > > > > >>>    their VM though due to this varstore.
> > > > > > >>>
> > > > > > >>> These are solved by adding two parameters to the commands. The 
> > > > > > >>> first is
> > > > > > >>> a block device node name that identifies the image to store 
> > > > > > >>> vmstate in,
> > > > > > >>> and the second is a list of node names to exclude from 
> > > > > > >>> snapshots.
> > > > > > >>>
> > > > > > >>> In the block code I've only dealt with node names for block 
> > > > > > >>> devices, as
> > > > > > >>> IIUC, this is all that libvirt should need in the -blockdev 
> > > > > > >>> world it now
> > > > > > >>> lives in. IOW, I've made not attempt to cope with people 
> > > > > > >>> wanting to use
> > > > > > >>> these QMP commands in combination with -drive args.
> > > > > > >>>
> > > > > > >>> I've done some minimal work in libvirt to start to make use of 
> > > > > > >>> the new
> > > > > > >>> commands to validate their functionality, but this isn't 
> > > > > > >>> finished yet.
> > > > > > >>>
> > > > > > >>> My ultimate goal is to make the GNOME Boxes maintainer happy 
> > > > > > >>> again by
> > > > > > >>> having internal snapshots work with OVMF:
> > > > > > >>>
> > > > > > >>>   
> > > > > > >>> https://gitlab.gnome.org/GNOME/gnome-boxes/-/commit/c486da262f6566326fbcb5e=
> > > > > > >>> f45c5f64048f16a6e
> > > > > > >>>
> > > > > > >>> Daniel P. Berrang=C3=A9 (6):
> > > > > > >>>   migration: improve error reporting of block driver state name
> > > > > > >>>   migration: introduce savevm, loadvm, delvm QMP commands
> > > > > > >>>   block: add ability to filter out blockdevs during snapshot
> > > > > > >>>   block: allow specifying name of block device for vmstate 
> > > > > > >>> storage
> > > > > > >>>   migration: support excluding block devs in QMP snapshot 
> > > > > > >>> commands
> > > > > > >>>   migration: support picking vmstate disk in QMP snapshot 
> > > > > > >>> commands
> > > > > > >>>
> > > > > > >>>  block/monitor/block-hmp-cmds.c |  4 +-
> > > > > > >>>  block/snapshot.c               | 68 +++++++++++++++++++------
> > > > > > >>>  include/block/snapshot.h       | 21 +++++---
> > > > > > >>>  include/migration/snapshot.h   | 10 +++-
> > > > > > >>>  migration/savevm.c             | 71 +++++++++++++++++++-------
> > > > > > >>>  monitor/hmp-cmds.c             | 20 ++------
> > > > > > >>>  qapi/migration.json            | 91 
> > > > > > >>> ++++++++++++++++++++++++++++++++++
> > > > > > >>>  replay/replay-snapshot.c       |  4 +-
> > > > > > >>>  softmmu/vl.c                   |  2 +-
> > > > > > >>>  9 files changed, 228 insertions(+), 63 deletions(-)
> > > > > > >> I have tried to work in this interface in 2016. That time
> > > > > > >> we have struggled with the idea that this QMP interface should
> > > > > > >> be ready to work asynchronously.
> > > > > > >>
> > > > > > >> Write-protect userfaultfd was merged into vanilla Linux
> > > > > > >> thus it is time to async savevm interface, which will also
> > > > > > >> bring async loadvm and some rework for state storing.
> > > > > > >>
> > > > > > >> Thus I think that with the introduction of the QMP interface
> > > > > > >> we should at least run save VM not from the main
> > > > > > >> thread but from the background with the event at the end.
> > > > > > > spawning a thread in which to invoke save_snapshot() and 
> > > > > > > load_snapshot()
> > > > > > > is easy enough.  I'm not at all clear on what we need in the way 
> > > > > > > of
> > > > > > > mutex locking though, to make those methods safe to run in a 
> > > > > > > thread
> > > > > > > that isn't the main event loop.
> > > > > > 
> > > > > > I am unsure that this is so easy. We need to be protected from other
> > > > > > operations
> > > > > > coming through QMP interface. Right now parallel operations are not 
> > > > > > allowed.
> > > > > > 
> > > > > > > Even with savevm/loadvm being blocking, we could introduce a QMP 
> > > > > > > event
> > > > > > > straight away, and document that users shouldn't assume the 
> > > > > > > operation
> > > > > > > is complete until they see the event. That would let us make the 
> > > > > > > commands
> > > > > > > non-blocking later with same documented semantics.
> > > > > > OK. Let us assume that you have added QMP savevm as proposed. It is
> > > > > > sync now. Sooner or later (I hope sooner) we will have to 
> > > > > > re-implement
> > > > > > this command with async version of the command, which will bring
> > > > > > again event etc and thus you will have to add compat layers to the
> > > > > > libvirt.
> > > > > > 
> > > > > > I think that it would be cleaner to start with the interface 
> > > > > > suitable for
> > > > > > further (coming) features and not copy obsolete implementation.
> > > > > > Yes, unfortunately, this is much more complex :(
> > > > > 
> > > > > Should we make this a job (may or may not be a block job) that just
> > > > > happens to block the VM and return completion immediately with the
> > > > > simple implementation we can have today? Then moving it later to a
> > > > > truly async operation mode should become transparent to the QMP 
> > > > > client.
> > > > 
> > > > What would making it a job / block job need from a QMP design POV ?
> > > 
> > > The actual QMP syntax for the command wouldn't look much different (I
> > > think just a new option 'job-id'), but the difference would be that it's
> > > not documented as performing the whole action, but just starting the
> > > job. The expectation would then be that it can be managed with the
> > > job-* commands and that it emits the job status events.
> > > 
> > > This may sound complicated, but most of it is actually covered by the
> > > generic job infrastructure.
> > > 
> > > The simplest job that we have is blockdev-create, which is implemented
> > > in block/create.c (99 lines including the license header). I think this
> > > would be a good model for our new case.
> > 
> > The QMP design and internal API looks simple enough, but I'm wondering
> > what implications come with the job infra wrt locking/thread safety. In
> > particular I see the "job_start" command runs the impl in a coroutine.
> > I can't tell if that's going to cause any interactions wrto the current
> > loadvm/savevm impl and its assumptions about blocking execution while
> > running.
> 
> Yes, the job infrastructure is build on coroutines and we'd need to
> check that this is safe. But both loadvm and savevm call both vm_stop()
> and bdrv_drain_all_begin/end(), so not much should be going on in
> parallel.
> 
> If this doesn't easily work out, there is still a simple solution for
> our sync implementation with an async interface: Just leave coroutine
> context immediately again by scheduling a BH that does the actual work.

Ok, I'm going to try out use of the job framework and post a v2.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|