Re: [PATCH 2/4] vhost-user: Interface for migration state transfer

[Eugenio Pérez]

On Wed, 2023-04-19 at 07:21 -0400, Stefan Hajnoczi wrote:
> On Wed, 19 Apr 2023 at 07:10, Hanna Czenczek <hreitz@redhat.com> wrote:
> > On 18.04.23 09:54, Eugenio Perez Martin wrote:
> > > On Mon, Apr 17, 2023 at 9:21 PM Stefan Hajnoczi <stefanha@gmail.com>
> > > wrote:
> > > > On Mon, 17 Apr 2023 at 15:08, Eugenio Perez Martin <eperezma@redhat.com>
> > > > wrote:
> > > > > On Mon, Apr 17, 2023 at 7:14 PM Stefan Hajnoczi <stefanha@redhat.com>
> > > > > wrote:
> > > > > > On Thu, Apr 13, 2023 at 12:14:24PM +0200, Eugenio Perez Martin
> > > > > > wrote:
> > > > > > > On Wed, Apr 12, 2023 at 11:06 PM Stefan Hajnoczi <
> > > > > > > stefanha@redhat.com> wrote:
> > > > > > > > On Tue, Apr 11, 2023 at 05:05:13PM +0200, Hanna Czenczek wrote:
> > > > > > > > > So-called "internal" virtio-fs migration refers to
> > > > > > > > > transporting the
> > > > > > > > > back-end's (virtiofsd's) state through qemu's migration
> > > > > > > > > stream.  To do
> > > > > > > > > this, we need to be able to transfer virtiofsd's internal
> > > > > > > > > state to and
> > > > > > > > > from virtiofsd.
> > > > > > > > > 
> > > > > > > > > Because virtiofsd's internal state will not be too large, we
> > > > > > > > > believe it
> > > > > > > > > is best to transfer it as a single binary blob after the
> > > > > > > > > streaming
> > > > > > > > > phase.  Because this method should be useful to other vhost-
> > > > > > > > > user
> > > > > > > > > implementations, too, it is introduced as a general-purpose
> > > > > > > > > addition to
> > > > > > > > > the protocol, not limited to vhost-user-fs.
> > > > > > > > > 
> > > > > > > > > These are the additions to the protocol:
> > > > > > > > > - New vhost-user protocol feature
> > > > > > > > > VHOST_USER_PROTOCOL_F_MIGRATORY_STATE:
> > > > > > > > >    This feature signals support for transferring state, and is
> > > > > > > > > added so
> > > > > > > > >    that migration can fail early when the back-end has no
> > > > > > > > > support.
> > > > > > > > > 
> > > > > > > > > - SET_DEVICE_STATE_FD function: Front-end and back-end
> > > > > > > > > negotiate a pipe
> > > > > > > > >    over which to transfer the state.  The front-end sends an
> > > > > > > > > FD to the
> > > > > > > > >    back-end into/from which it can write/read its state, and
> > > > > > > > > the back-end
> > > > > > > > >    can decide to either use it, or reply with a different FD
> > > > > > > > > for the
> > > > > > > > >    front-end to override the front-end's choice.
> > > > > > > > >    The front-end creates a simple pipe to transfer the state,
> > > > > > > > > but maybe
> > > > > > > > >    the back-end already has an FD into/from which it has to
> > > > > > > > > write/read
> > > > > > > > >    its state, in which case it will want to override the
> > > > > > > > > simple pipe.
> > > > > > > > >    Conversely, maybe in the future we find a way to have the
> > > > > > > > > front-end
> > > > > > > > >    get an immediate FD for the migration stream (in some
> > > > > > > > > cases), in which
> > > > > > > > >    case we will want to send this to the back-end instead of
> > > > > > > > > creating a
> > > > > > > > >    pipe.
> > > > > > > > >    Hence the negotiation: If one side has a better idea than a
> > > > > > > > > plain
> > > > > > > > >    pipe, we will want to use that.
> > > > > > > > > 
> > > > > > > > > - CHECK_DEVICE_STATE: After the state has been transferred
> > > > > > > > > through the
> > > > > > > > >    pipe (the end indicated by EOF), the front-end invokes this
> > > > > > > > > function
> > > > > > > > >    to verify success.  There is no in-band way (through the
> > > > > > > > > pipe) to
> > > > > > > > >    indicate failure, so we need to check explicitly.
> > > > > > > > > 
> > > > > > > > > Once the transfer pipe has been established via
> > > > > > > > > SET_DEVICE_STATE_FD
> > > > > > > > > (which includes establishing the direction of transfer and
> > > > > > > > > migration
> > > > > > > > > phase), the sending side writes its data into the pipe, and
> > > > > > > > > the reading
> > > > > > > > > side reads it until it sees an EOF.  Then, the front-end will
> > > > > > > > > check for
> > > > > > > > > success via CHECK_DEVICE_STATE, which on the destination side
> > > > > > > > > includes
> > > > > > > > > checking for integrity (i.e. errors during deserialization).
> > > > > > > > > 
> > > > > > > > > Suggested-by: Stefan Hajnoczi <stefanha@redhat.com>
> > > > > > > > > Signed-off-by: Hanna Czenczek <hreitz@redhat.com>
> > > > > > > > > ---
> > > > > > > > >   include/hw/virtio/vhost-backend.h |  24 +++++
> > > > > > > > >   include/hw/virtio/vhost.h         |  79 ++++++++++++++++
> > > > > > > > >   hw/virtio/vhost-user.c            | 147
> > > > > > > > > ++++++++++++++++++++++++++++++
> > > > > > > > >   hw/virtio/vhost.c                 |  37 ++++++++
> > > > > > > > >   4 files changed, 287 insertions(+)
> > > > > > > > > 
> > > > > > > > > diff --git a/include/hw/virtio/vhost-backend.h
> > > > > > > > > b/include/hw/virtio/vhost-backend.h
> > > > > > > > > index ec3fbae58d..5935b32fe3 100644
> > > > > > > > > --- a/include/hw/virtio/vhost-backend.h
> > > > > > > > > +++ b/include/hw/virtio/vhost-backend.h
> > > > > > > > > @@ -26,6 +26,18 @@ typedef enum VhostSetConfigType {
> > > > > > > > >       VHOST_SET_CONFIG_TYPE_MIGRATION = 1,
> > > > > > > > >   } VhostSetConfigType;
> > > > > > > > > 
> > > > > > > > > +typedef enum VhostDeviceStateDirection {
> > > > > > > > > +    /* Transfer state from back-end (device) to front-end */
> > > > > > > > > +    VHOST_TRANSFER_STATE_DIRECTION_SAVE = 0,
> > > > > > > > > +    /* Transfer state from front-end to back-end (device) */
> > > > > > > > > +    VHOST_TRANSFER_STATE_DIRECTION_LOAD = 1,
> > > > > > > > > +} VhostDeviceStateDirection;
> > > > > > > > > +
> > > > > > > > > +typedef enum VhostDeviceStatePhase {
> > > > > > > > > +    /* The device (and all its vrings) is stopped */
> > > > > > > > > +    VHOST_TRANSFER_STATE_PHASE_STOPPED = 0,
> > > > > > > > > +} VhostDeviceStatePhase;
> > > > > > > > vDPA has:
> > > > > > > > 
> > > > > > > >    /* Suspend a device so it does not process virtqueue requests
> > > > > > > > anymore
> > > > > > > >     *
> > > > > > > >     * After the return of ioctl the device must preserve all the
> > > > > > > > necessary state
> > > > > > > >     * (the virtqueue vring base plus the possible device
> > > > > > > > specific states) that is
> > > > > > > >     * required for restoring in the future. The device must not
> > > > > > > > change its
> > > > > > > >     * configuration after that point.
> > > > > > > >     */
> > > > > > > >    #define VHOST_VDPA_SUSPEND      _IO(VHOST_VIRTIO, 0x7D)
> > > > > > > > 
> > > > > > > >    /* Resume a device so it can resume processing virtqueue
> > > > > > > > requests
> > > > > > > >     *
> > > > > > > >     * After the return of this ioctl the device will have
> > > > > > > > restored all the
> > > > > > > >     * necessary states and it is fully operational to continue
> > > > > > > > processing the
> > > > > > > >     * virtqueue descriptors.
> > > > > > > >     */
> > > > > > > >    #define VHOST_VDPA_RESUME       _IO(VHOST_VIRTIO, 0x7E)
> > > > > > > > 
> > > > > > > > I wonder if it makes sense to import these into vhost-user so
> > > > > > > > that the
> > > > > > > > difference between kernel vhost and vhost-user is minimized.
> > > > > > > > It's okay
> > > > > > > > if one of them is ahead of the other, but it would be nice to
> > > > > > > > avoid
> > > > > > > > overlapping/duplicated functionality.
> > > > > > > > 
> > > > > > > That's what I had in mind in the first versions. I proposed
> > > > > > > VHOST_STOP
> > > > > > > instead of VHOST_VDPA_STOP for this very reason. Later it did
> > > > > > > change
> > > > > > > to SUSPEND.
> > > > > > > 
> > > > > > > Generally it is better if we make the interface less parametrized
> > > > > > > and
> > > > > > > we trust in the messages and its semantics in my opinion. In other
> > > > > > > words, instead of
> > > > > > > vhost_set_device_state_fd_op(VHOST_TRANSFER_STATE_PHASE_STOPPED),
> > > > > > > send
> > > > > > > individually the equivalent of VHOST_VDPA_SUSPEND vhost-user
> > > > > > > command.
> > > > > > > 
> > > > > > > Another way to apply this is with the "direction" parameter. Maybe
> > > > > > > it
> > > > > > > is better to split it into "set_state_fd" and "get_state_fd"?
> > > > > > > 
> > > > > > > In that case, reusing the ioctls as vhost-user messages would be
> > > > > > > ok.
> > > > > > > But that puts this proposal further from the VFIO code, which uses
> > > > > > > "migration_set_state(state)", and maybe it is better when the
> > > > > > > number
> > > > > > > of states is high.
> > > > > > Hi Eugenio,
> > > > > > Another question about vDPA suspend/resume:
> > > > > > 
> > > > > >    /* Host notifiers must be enabled at this point. */
> > > > > >    void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev,
> > > > > > bool vrings)
> > > > > >    {
> > > > > >        int i;
> > > > > > 
> > > > > >        /* should only be called after backend is connected */
> > > > > >        assert(hdev->vhost_ops);
> > > > > >        event_notifier_test_and_clear(
> > > > > >            &hdev-
> > > > > > >vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier);
> > > > > >        event_notifier_test_and_clear(&vdev->config_notifier);
> > > > > > 
> > > > > >        trace_vhost_dev_stop(hdev, vdev->name, vrings);
> > > > > > 
> > > > > >        if (hdev->vhost_ops->vhost_dev_start) {
> > > > > >            hdev->vhost_ops->vhost_dev_start(hdev, false);
> > > > > >            ^^^ SUSPEND ^^^
> > > > > >        }
> > > > > >        if (vrings) {
> > > > > >            vhost_dev_set_vring_enable(hdev, false);
> > > > > >        }
> > > > > >        for (i = 0; i < hdev->nvqs; ++i) {
> > > > > >            vhost_virtqueue_stop(hdev,
> > > > > >                                 vdev,
> > > > > >                                 hdev->vqs + i,
> > > > > >                                 hdev->vq_index + i);
> > > > > >          ^^^ fetch virtqueue state from kernel ^^^
> > > > > >        }
> > > > > >        if (hdev->vhost_ops->vhost_reset_status) {
> > > > > >            hdev->vhost_ops->vhost_reset_status(hdev);
> > > > > >            ^^^ reset device^^^
> > > > > > 
> > > > > > I noticed the QEMU vDPA code resets the device in vhost_dev_stop()
> > > > > > ->
> > > > > > vhost_reset_status(). The device's migration code runs after
> > > > > > vhost_dev_stop() and the state will have been lost.
> > > > > > 
> > > > > vhost_virtqueue_stop saves the vq state (indexes, vring base) in the
> > > > > qemu VirtIONet device model. This is for all vhost backends.
> > > > > 
> > > > > Regarding the state like mac or mq configuration, SVQ runs for all the
> > > > > VM run in the CVQ. So it can track all of that status in the device
> > > > > model too.
> > > > > 
> > > > > When a migration effectively occurs, all the frontend state is
> > > > > migrated as a regular emulated device. To route all of the state in a
> > > > > normalized way for qemu is what leaves open the possibility to do
> > > > > cross-backends migrations, etc.
> > > > > 
> > > > > Does that answer your question?
> > > > I think you're confirming that changes would be necessary in order for
> > > > vDPA to support the save/load operation that Hanna is introducing.
> > > > 
> > > Yes, this first iteration was centered on net, with an eye on block,
> > > where state can be routed through classical emulated devices. This is
> > > how vhost-kernel and vhost-user do classically. And it allows
> > > cross-backend, to not modify qemu migration state, etc.
> > > 
> > > To introduce this opaque state to qemu, that must be fetched after the
> > > suspend and not before, requires changes in vhost protocol, as
> > > discussed previously.
> > > 
> > > > > > It looks like vDPA changes are necessary in order to support
> > > > > > stateful
> > > > > > devices even though QEMU already uses SUSPEND. Is my understanding
> > > > > > correct?
> > > > > > 
> > > > > Changes are required elsewhere, as the code to restore the state
> > > > > properly in the destination has not been merged.
> > > > I'm not sure what you mean by elsewhere?
> > > > 
> > > I meant for vdpa *net* devices the changes are not required in vdpa
> > > ioctls, but mostly in qemu.
> > > 
> > > If you meant stateful as "it must have a state blob that it must be
> > > opaque to qemu", then I think the straightforward action is to fetch
> > > state blob about the same time as vq indexes. But yes, changes (at
> > > least a new ioctl) is needed for that.
> > > 
> > > > I'm asking about vDPA ioctls. Right now the sequence is SUSPEND and
> > > > then VHOST_VDPA_SET_STATUS 0.
> > > > 
> > > > In order to save device state from the vDPA device in the future, it
> > > > will be necessary to defer the VHOST_VDPA_SET_STATUS 0 call so that
> > > > the device state can be saved before the device is reset.
> > > > 
> > > > Does that sound right?
> > > > 
> > > The split between suspend and reset was added recently for that very
> > > reason. In all the virtio devices, the frontend is initialized before
> > > the backend, so I don't think it is a good idea to defer the backend
> > > cleanup. Especially if we have already set the state is small enough
> > > to not needing iterative migration from virtiofsd point of view.
> > > 
> > > If fetching that state at the same time as vq indexes is not valid,
> > > could it follow the same model as the "in-flight descriptors"?
> > > vhost-user follows them by using a shared memory region where their
> > > state is tracked [1]. This allows qemu to survive vhost-user SW
> > > backend crashes, and does not forbid the cross-backends live migration
> > > as all the information is there to recover them.
> > > 
> > > For hw devices this is not convenient as it occupies PCI bandwidth. So
> > > a possibility is to synchronize this memory region after a
> > > synchronization point, being the SUSPEND call or GET_VRING_BASE. HW
> > > devices are not going to crash in the software sense, so all use cases
> > > remain the same to qemu. And that shared memory information is
> > > recoverable after vhost_dev_stop.
> > > 
> > > Does that sound reasonable to virtiofsd? To offer a shared memory
> > > region where it dumps the state, maybe only after the
> > > set_state(STATE_PHASE_STOPPED)?
> > 
> > I don’t think we need the set_state() call, necessarily, if SUSPEND is
> > mandatory anyway.
> > 
> > As for the shared memory, the RFC before this series used shared memory,
> > so it’s possible, yes.  But “shared memory region” can mean a lot of
> > things – it sounds like you’re saying the back-end (virtiofsd) should
> > provide it to the front-end, is that right?  That could work like this:
> > 
> > On the source side:
> > 
> > S1. SUSPEND goes to virtiofsd
> > S2. virtiofsd maybe double-checks that the device is stopped, then
> > serializes its state into a newly allocated shared memory area[1]
> > S3. virtiofsd responds to SUSPEND
> > S4. front-end requests shared memory, virtiofsd responds with a handle,
> > maybe already closes its reference
> > S5. front-end saves state, closes its handle, freeing the SHM
> > 
> > [1] Maybe virtiofsd can correctly size the serialized state’s size, then
> > it can immediately allocate this area and serialize directly into it;
> > maybe it can’t, then we’ll need a bounce buffer.  Not really a
> > fundamental problem, but there are limitations around what you can do
> > with serde implementations in Rust…
> > 
> > On the destination side:
> > 
> > D1. Optional SUSPEND goes to virtiofsd that hasn’t yet done much;
> > virtiofsd would serialize its empty state into an SHM area, and respond
> > to SUSPEND
> > D2. front-end reads state from migration stream into an SHM it has allocated
> > D3. front-end supplies this SHM to virtiofsd, which discards its
> > previous area, and now uses this one
> > D4. RESUME goes to virtiofsd, which deserializes the state from the SHM
> > 
> > Couple of questions:
> > 
> > A. Stefan suggested D1, but it does seem wasteful now.  But if SUSPEND
> > would imply to deserialize a state, and the state is to be transferred
> > through SHM, this is what would need to be done.  So maybe we should
> > skip SUSPEND on the destination?
> > B. You described that the back-end should supply the SHM, which works
> > well on the source.  On the destination, only the front-end knows how
> > big the state is, so I’ve decided above that it should allocate the SHM
> > (D2) and provide it to the back-end.  Is that feasible or is it
> > important (e.g. for real hardware) that the back-end supplies the SHM?
> > (In which case the front-end would need to tell the back-end how big the
> > state SHM needs to be.)
> 
> How does this work for iterative live migration?
> 

A pipe will always fit better for iterative from qemu POV, that's for sure. 
Especially if we want to keep that opaqueness.

But  we will need to communicate with the HW device using shared memory sooner
or later for big states.  If we don't transform it in qemu, we will need to do
it in the kernel.  Also, the pipe will not support daemon crashes.

Again I'm just putting this on the table, just in case it fits better or it is
convenient.  I missed the previous patch where SHM was proposed too, so maybe I
missed some feedback useful here.  I think the pipe is a better solution in the
long run because of the iterative part.

Thanks!