Re: [RFC 09/10] vhost: Route guest->host notification through shadow virtqueue

[Jason Wang <jasowang@redhat.com>]

On 2021/2/2 下午6:08, Eugenio Perez Martin wrote:
> On Mon, Feb 1, 2021 at 7:29 AM Jason Wang <jasowang@redhat.com> wrote:
>>
>> On 2021/1/30 上午4:54, Eugenio Pérez wrote:
>>> Shadow virtqueue notifications forwarding is disabled when vhost_dev
>>> stops.
>>>
>>> Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
>>> ---
>>>    hw/virtio/vhost-shadow-virtqueue.h |   5 ++
>>>    include/hw/virtio/vhost.h          |   4 +
>>>    hw/virtio/vhost-shadow-virtqueue.c | 123 +++++++++++++++++++++++++-
>>>    hw/virtio/vhost.c                  | 135 ++++++++++++++++++++++++++++-
>>>    4 files changed, 264 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/hw/virtio/vhost-shadow-virtqueue.h b/hw/virtio/vhost-shadow-virtqueue.h
>>> index 6cc18d6acb..466f8ae595 100644
>>> --- a/hw/virtio/vhost-shadow-virtqueue.h
>>> +++ b/hw/virtio/vhost-shadow-virtqueue.h
>>> @@ -17,6 +17,11 @@
>>>
>>>    typedef struct VhostShadowVirtqueue VhostShadowVirtqueue;
>>>
>>> +bool vhost_shadow_vq_start_rcu(struct vhost_dev *dev,
>>> +                               VhostShadowVirtqueue *svq);
>>> +void vhost_shadow_vq_stop_rcu(struct vhost_dev *dev,
>>> +                              VhostShadowVirtqueue *svq);
>>> +
>>>    VhostShadowVirtqueue *vhost_shadow_vq_new(struct vhost_dev *dev, int idx);
>>>
>>>    void vhost_shadow_vq_free(VhostShadowVirtqueue *vq);
>>> diff --git a/include/hw/virtio/vhost.h b/include/hw/virtio/vhost.h
>>> index 2be782cefd..732a4b2a2b 100644
>>> --- a/include/hw/virtio/vhost.h
>>> +++ b/include/hw/virtio/vhost.h
>>> @@ -55,6 +55,8 @@ struct vhost_iommu {
>>>        QLIST_ENTRY(vhost_iommu) iommu_next;
>>>    };
>>>
>>> +typedef struct VhostShadowVirtqueue VhostShadowVirtqueue;
>>> +
>>>    typedef struct VhostDevConfigOps {
>>>        /* Vhost device config space changed callback
>>>         */
>>> @@ -83,7 +85,9 @@ struct vhost_dev {
>>>        uint64_t backend_cap;
>>>        bool started;
>>>        bool log_enabled;
>>> +    bool sw_lm_enabled;
>>>        uint64_t log_size;
>>> +    VhostShadowVirtqueue **shadow_vqs;
>>>        Error *migration_blocker;
>>>        const VhostOps *vhost_ops;
>>>        void *opaque;
>>> diff --git a/hw/virtio/vhost-shadow-virtqueue.c b/hw/virtio/vhost-shadow-virtqueue.c
>>> index c0c967a7c5..908c36c66d 100644
>>> --- a/hw/virtio/vhost-shadow-virtqueue.c
>>> +++ b/hw/virtio/vhost-shadow-virtqueue.c
>>> @@ -8,15 +8,129 @@
>>>     */
>>>
>>>    #include "hw/virtio/vhost-shadow-virtqueue.h"
>>> +#include "hw/virtio/vhost.h"
>>> +#include "hw/virtio/virtio-access.h"
>>> +
>>> +#include "standard-headers/linux/vhost_types.h"
>>> +#include "standard-headers/linux/virtio_ring.h"
>>>
>>>    #include "qemu/error-report.h"
>>> -#include "qemu/event_notifier.h"
>>> +#include "qemu/main-loop.h"
>>>
>>>    typedef struct VhostShadowVirtqueue {
>>>        EventNotifier kick_notifier;
>>>        EventNotifier call_notifier;
>>> +    const struct vhost_virtqueue *hvq;
>>> +    VirtIODevice *vdev;
>>> +    VirtQueue *vq;
>>>    } VhostShadowVirtqueue;
>>
>> So instead of doing things at virtio level, how about do the shadow
>> stuffs at vhost level?
>>
>> It works like:
>>
>> virtio -> [shadow vhost backend] -> vhost backend
>>
>> Then the QMP is used to plug the shadow vhost backend in the middle or not.
>>
>> It looks kind of easier since we don't need to deal with virtqueue
>> handlers etc.. Instead, we just need to deal with eventfd stuffs:
>>
>> When shadow vhost mode is enabled, we just intercept the host_notifiers
>> and guest_notifiers. When it was disabled, we just pass the host/guest
>> notifiers to the real vhost backends?
>>
> Hi Jason.
>
> Sure we can try that model, but it seems to me that it comes with a
> different set of problems.
>
> For example, there are code in vhost.c that checks if implementations
> are available in vhost_ops, like:
>
> if (dev->vhost_ops->vhost_vq_get_addr) {
>          r = dev->vhost_ops->vhost_vq_get_addr(dev, &addr, vq);
>          ...
> }
>
> I can count 14 of these, checking:
>
> dev->vhost_ops->vhost_backend_can_merge
> dev->vhost_ops->vhost_backend_mem_section_filter
> dev->vhost_ops->vhost_force_iommu
> dev->vhost_ops->vhost_requires_shm_log
> dev->vhost_ops->vhost_set_backend_cap
> dev->vhost_ops->vhost_set_vring_busyloop_timeout
> dev->vhost_ops->vhost_vq_get_addr
> hdev->vhost_ops->vhost_dev_start
> hdev->vhost_ops->vhost_get_config
> hdev->vhost_ops->vhost_get_inflight_fd
> hdev->vhost_ops->vhost_net_set_backend
> hdev->vhost_ops->vhost_set_config
> hdev->vhost_ops->vhost_set_inflight_fd
> hdev->vhost_ops->vhost_set_iotlb_callback
>
> So we should Implement all of the vhost_ops callbacks, forwarding them
> to actual vhost_backed, and delete conditionally these ones? In other
> words, dynamically generate the new shadow vq vhost_ops? If a new
> callback is added to any vhost backend in the future, do we have to
> force the adding / checking for NULL in shadow backend vhost_ops?
> Would this be a good moment to check if all backends implement these
> and delete the checks?


I think it won't be easy if we want to support all kinds of vhost 
backends from the start. So we can go with vhost-vdpa one first.

Actually how it work might be something like (no need to switch 
vhost_ops, we can do everything silently in the ops)

1) when device to switch to shadow vq (e.g via QMP)
2) vhost-vdpa will stop and sync state (last_avail_idx) internally
3) reset vhost-vdpa, clean call and kick eventfd
4) allocate vqs for vhost-vdpa, new call and kick eventfd, restart 
vhost-vdpa
5) start the shadow vq (make it start for last_avail_idx)
6) intercept ioeventfd and forward the request to callfd
7) intercept callfd and forward the request to irqfd
8) forward request between shadow virtqueue and vhost-vdpa


>
> There are also checks like:
>
> if (dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER)
>
> How would shadow_vq backend expose itself? (I guess as the actual used backend).
>
> I can modify this patchset to not relay the guest->host notifications
> on vq handlers but on eventfd handlers. Although this will make it
> independent of the actual virtio device kind used, I can see two
> drawbacks:
> * The actual fact that it makes it independent of virtio device kind.
> If a device does not use the notifiers and poll the ring by itself, it
> has no chance of knowing that it should stop. What happens if
> virtio-net tx timer is armed when we start shadow vq?.


So if we do that in vhost level, it's a vhost backend from the virtio 
layer. Then we don't need to worry about tx timer stuffs.


> * The fixes (current and future) in vq notifications, like the one
> currently implemented in virtio_notify_irqfd for windows drivers
> regarding ISR bit 0. I think this one in particular is OK not to
> carry, but I think many changes affecting any of the functions will
> have to be mirrored in the other.


Consider we behave like a vhost, it just work as in the past for other 
type of vhost backends when MSI-X is not enabled?

Thanks


>
> Thoughts on this?
>
> Thanks!
>
>> Thanks
>>
>>
>>> +static uint16_t vhost_shadow_vring_used_flags(VhostShadowVirtqueue *svq)
>>> +{
>>> +    const struct vring_used *used = svq->hvq->used;
>>> +    return virtio_tswap16(svq->vdev, used->flags);
>>> +}
>>> +
>>> +static bool vhost_shadow_vring_should_kick(VhostShadowVirtqueue *vq)
>>> +{
>>> +    return !(vhost_shadow_vring_used_flags(vq) & VRING_USED_F_NO_NOTIFY);
>>> +}
>>> +
>>> +static void vhost_shadow_vring_kick(VhostShadowVirtqueue *vq)
>>> +{
>>> +    if (vhost_shadow_vring_should_kick(vq)) {
>>> +        event_notifier_set(&vq->kick_notifier);
>>> +    }
>>> +}
>>> +
>>> +static void handle_shadow_vq(VirtIODevice *vdev, VirtQueue *vq)
>>> +{
>>> +    struct vhost_dev *hdev = vhost_dev_from_virtio(vdev);
>>> +    uint16_t idx = virtio_get_queue_index(vq);
>>> +
>>> +    VhostShadowVirtqueue *svq = hdev->shadow_vqs[idx];
>>> +
>>> +    vhost_shadow_vring_kick(svq);
>>> +}
>>> +
>>> +/*
>>> + * Start shadow virtqueue operation.
>>> + * @dev vhost device
>>> + * @svq Shadow Virtqueue
>>> + *
>>> + * Run in RCU context
>>> + */
>>> +bool vhost_shadow_vq_start_rcu(struct vhost_dev *dev,
>>> +                               VhostShadowVirtqueue *svq)
>>> +{
>>> +    const VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(dev->vdev);
>>> +    EventNotifier *vq_host_notifier = virtio_queue_get_host_notifier(svq->vq);
>>> +    unsigned idx = virtio_queue_get_idx(svq->vdev, svq->vq);
>>> +    struct vhost_vring_file kick_file = {
>>> +        .index = idx,
>>> +        .fd = event_notifier_get_fd(&svq->kick_notifier),
>>> +    };
>>> +    int r;
>>> +    bool ok;
>>> +
>>> +    /* Check that notifications are still going directly to vhost dev */
>>> +    assert(virtio_queue_host_notifier_status(svq->vq));
>>> +
>>> +    ok = k->set_vq_handler(dev->vdev, idx, handle_shadow_vq);
>>> +    if (!ok) {
>>> +        error_report("Couldn't set the vq handler");
>>> +        goto err_set_kick_handler;
>>> +    }
>>> +
>>> +    r = dev->vhost_ops->vhost_set_vring_kick(dev, &kick_file);
>>> +    if (r != 0) {
>>> +        error_report("Couldn't set kick fd: %s", strerror(errno));
>>> +        goto err_set_vring_kick;
>>> +    }
>>> +
>>> +    event_notifier_set_handler(vq_host_notifier,
>>> +                               virtio_queue_host_notifier_read);
>>> +    virtio_queue_set_host_notifier_enabled(svq->vq, false);
>>> +    virtio_queue_host_notifier_read(vq_host_notifier);
>>> +
>>> +    return true;
>>> +
>>> +err_set_vring_kick:
>>> +    k->set_vq_handler(dev->vdev, idx, NULL);
>>> +
>>> +err_set_kick_handler:
>>> +    return false;
>>> +}
>>> +
>>> +/*
>>> + * Stop shadow virtqueue operation.
>>> + * @dev vhost device
>>> + * @svq Shadow Virtqueue
>>> + *
>>> + * Run in RCU context
>>> + */
>>> +void vhost_shadow_vq_stop_rcu(struct vhost_dev *dev,
>>> +                              VhostShadowVirtqueue *svq)
>>> +{
>>> +    const VirtioDeviceClass *k = VIRTIO_DEVICE_GET_CLASS(svq->vdev);
>>> +    unsigned idx = virtio_queue_get_idx(svq->vdev, svq->vq);
>>> +    EventNotifier *vq_host_notifier = virtio_queue_get_host_notifier(svq->vq);
>>> +    struct vhost_vring_file kick_file = {
>>> +        .index = idx,
>>> +        .fd = event_notifier_get_fd(vq_host_notifier),
>>> +    };
>>> +    int r;
>>> +
>>> +    /* Restore vhost kick */
>>> +    r = dev->vhost_ops->vhost_set_vring_kick(dev, &kick_file);
>>> +    /* Cannot do a lot of things */
>>> +    assert(r == 0);
>>> +
>>> +    event_notifier_set_handler(vq_host_notifier, NULL);
>>> +    virtio_queue_set_host_notifier_enabled(svq->vq, true);
>>> +    k->set_vq_handler(svq->vdev, idx, NULL);
>>> +}
>>> +
>>>    /*
>>>     * Creates vhost shadow virtqueue, and instruct vhost device to use the shadow
>>>     * methods and file descriptors.
>>> @@ -24,8 +138,13 @@ typedef struct VhostShadowVirtqueue {
>>>    VhostShadowVirtqueue *vhost_shadow_vq_new(struct vhost_dev *dev, int idx)
>>>    {
>>>        g_autofree VhostShadowVirtqueue *svq = g_new0(VhostShadowVirtqueue, 1);
>>> +    int vq_idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + idx);
>>>        int r;
>>>
>>> +    svq->vq = virtio_get_queue(dev->vdev, vq_idx);
>>> +    svq->hvq = &dev->vqs[idx];
>>> +    svq->vdev = dev->vdev;
>>> +
>>>        r = event_notifier_init(&svq->kick_notifier, 0);
>>>        if (r != 0) {
>>>            error_report("Couldn't create kick event notifier: %s",
>>> @@ -40,7 +159,7 @@ VhostShadowVirtqueue *vhost_shadow_vq_new(struct vhost_dev *dev, int idx)
>>>            goto err_init_call_notifier;
>>>        }
>>>
>>> -    return svq;
>>> +    return g_steal_pointer(&svq);
>>>
>>>    err_init_call_notifier:
>>>        event_notifier_cleanup(&svq->kick_notifier);
>>> diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c
>>> index 42836e45f3..bde688f278 100644
>>> --- a/hw/virtio/vhost.c
>>> +++ b/hw/virtio/vhost.c
>>> @@ -25,6 +25,7 @@
>>>    #include "exec/address-spaces.h"
>>>    #include "hw/virtio/virtio-bus.h"
>>>    #include "hw/virtio/virtio-access.h"
>>> +#include "hw/virtio/vhost-shadow-virtqueue.h"
>>>    #include "migration/blocker.h"
>>>    #include "migration/qemu-file-types.h"
>>>    #include "sysemu/dma.h"
>>> @@ -945,6 +946,82 @@ static void vhost_log_global_stop(MemoryListener *listener)
>>>        }
>>>    }
>>>
>>> +static int vhost_sw_live_migration_stop(struct vhost_dev *dev)
>>> +{
>>> +    int idx;
>>> +
>>> +    WITH_RCU_READ_LOCK_GUARD() {
>>> +        dev->sw_lm_enabled = false;
>>> +
>>> +        for (idx = 0; idx < dev->nvqs; ++idx) {
>>> +            vhost_shadow_vq_stop_rcu(dev, dev->shadow_vqs[idx]);
>>> +        }
>>> +    }
>>> +
>>> +    for (idx = 0; idx < dev->nvqs; ++idx) {
>>> +        vhost_shadow_vq_free(dev->shadow_vqs[idx]);
>>> +    }
>>> +
>>> +    g_free(dev->shadow_vqs);
>>> +    dev->shadow_vqs = NULL;
>>> +    return 0;
>>> +}
>>> +
>>> +static int vhost_sw_live_migration_start(struct vhost_dev *dev)
>>> +{
>>> +    int idx;
>>> +
>>> +    dev->shadow_vqs = g_new0(VhostShadowVirtqueue *, dev->nvqs);
>>> +    for (idx = 0; idx < dev->nvqs; ++idx) {
>>> +        dev->shadow_vqs[idx] = vhost_shadow_vq_new(dev, idx);
>>> +        if (unlikely(dev->shadow_vqs[idx] == NULL)) {
>>> +            goto err;
>>> +        }
>>> +    }
>>> +
>>> +    WITH_RCU_READ_LOCK_GUARD() {
>>> +        for (idx = 0; idx < dev->nvqs; ++idx) {
>>> +            int stop_idx = idx;
>>> +            bool ok = vhost_shadow_vq_start_rcu(dev,
>>> +                                                dev->shadow_vqs[idx]);
>>> +
>>> +            if (!ok) {
>>> +                while (--stop_idx >= 0) {
>>> +                    vhost_shadow_vq_stop_rcu(dev, dev->shadow_vqs[stop_idx]);
>>> +                }
>>> +
>>> +                goto err;
>>> +            }
>>> +        }
>>> +    }
>>> +
>>> +    dev->sw_lm_enabled = true;
>>> +    return 0;
>>> +
>>> +err:
>>> +    for (; idx >= 0; --idx) {
>>> +        vhost_shadow_vq_free(dev->shadow_vqs[idx]);
>>> +    }
>>> +    g_free(dev->shadow_vqs[idx]);
>>> +
>>> +    return -1;
>>> +}
>>> +
>>> +static int vhost_sw_live_migration_enable(struct vhost_dev *dev,
>>> +                                          bool enable_lm)
>>> +{
>>> +    int r;
>>> +
>>> +    if (enable_lm == dev->sw_lm_enabled) {
>>> +        return 0;
>>> +    }
>>> +
>>> +    r = enable_lm ? vhost_sw_live_migration_start(dev)
>>> +                  : vhost_sw_live_migration_stop(dev);
>>> +
>>> +    return r;
>>> +}
>>> +
>>>    static void vhost_log_start(MemoryListener *listener,
>>>                                MemoryRegionSection *section,
>>>                                int old, int new)
>>> @@ -1389,6 +1466,7 @@ int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
>>>        hdev->log = NULL;
>>>        hdev->log_size = 0;
>>>        hdev->log_enabled = false;
>>> +    hdev->sw_lm_enabled = false;
>>>        hdev->started = false;
>>>        memory_listener_register(&hdev->memory_listener, &address_space_memory);
>>>        QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
>>> @@ -1816,6 +1894,11 @@ void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
>>>            hdev->vhost_ops->vhost_dev_start(hdev, false);
>>>        }
>>>        for (i = 0; i < hdev->nvqs; ++i) {
>>> +        if (hdev->sw_lm_enabled) {
>>> +            vhost_shadow_vq_stop_rcu(hdev, hdev->shadow_vqs[i]);
>>> +            vhost_shadow_vq_free(hdev->shadow_vqs[i]);
>>> +        }
>>> +
>>>            vhost_virtqueue_stop(hdev,
>>>                                 vdev,
>>>                                 hdev->vqs + i,
>>> @@ -1829,6 +1912,8 @@ void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
>>>            memory_listener_unregister(&hdev->iommu_listener);
>>>        }
>>>        vhost_log_put(hdev, true);
>>> +    g_free(hdev->shadow_vqs);
>>> +    hdev->sw_lm_enabled = false;
>>>        hdev->started = false;
>>>        hdev->vdev = NULL;
>>>    }
>>> @@ -1845,5 +1930,53 @@ int vhost_net_set_backend(struct vhost_dev *hdev,
>>>
>>>    void qmp_x_vhost_enable_shadow_vq(const char *name, bool enable, Error **errp)
>>>    {
>>> -    error_setg(errp, "Shadow virtqueue still not implemented.");
>>> +    struct vhost_dev *hdev;
>>> +    const char *err_cause = NULL;
>>> +    const VirtioDeviceClass *k;
>>> +    int r;
>>> +    ErrorClass err_class = ERROR_CLASS_GENERIC_ERROR;
>>> +
>>> +    QLIST_FOREACH(hdev, &vhost_devices, entry) {
>>> +        if (hdev->vdev && 0 == strcmp(hdev->vdev->name, name)) {
>>> +            break;
>>> +        }
>>> +    }
>>> +
>>> +    if (!hdev) {
>>> +        err_class = ERROR_CLASS_DEVICE_NOT_FOUND;
>>> +        err_cause = "Device not found";
>>> +        goto err;
>>> +    }
>>> +
>>> +    if (!hdev->started) {
>>> +        err_cause = "Device is not started";
>>> +        goto err;
>>> +    }
>>> +
>>> +    if (hdev->acked_features & BIT_ULL(VIRTIO_F_RING_PACKED)) {
>>> +        err_cause = "Use packed vq";
>>> +        goto err;
>>> +    }
>>> +
>>> +    if (vhost_dev_has_iommu(hdev)) {
>>> +        err_cause = "Device use IOMMU";
>>> +        goto err;
>>> +    }
>>> +
>>> +    k = VIRTIO_DEVICE_GET_CLASS(hdev->vdev);
>>> +    if (!k->set_vq_handler) {
>>> +        err_cause = "Virtio device type does not support reset of vq handler";
>>> +        goto err;
>>> +    }
>>> +
>>> +    r = vhost_sw_live_migration_enable(hdev, enable);
>>> +    if (unlikely(r)) {
>>> +        err_cause = "Error enabling (see monitor)";
>>> +    }
>>> +
>>> +err:
>>> +    if (err_cause) {
>>> +        error_set(errp, err_class,
>>> +                  "Can't enable shadow vq on %s: %s", name, err_cause);
>>> +    }
>>>    }