Re: [Qemu-devel] [PATCH v4 08/11] virtio: event suppression support for packed ring

[Wei Xu <wexu@redhat.com>]

On Tue, Feb 19, 2019 at 03:19:58PM +0800, Jason Wang wrote:
> 
> On 2019/2/14 下午12:26, wexu@redhat.com wrote:
> >From: Wei Xu <wexu@redhat.com>
> >
> >Difference between 'avail_wrap_counter' and 'last_avail_wrap_counter':
> >For Tx(guest transmitting), they are the same after each pop of a desc.
> >
> >For Rx(guest receiving), they are also the same when there are enough
> >descriptors to carry the payload for a packet(e.g. usually 16 descs are
> >needed for a 64k packet in typical iperf tcp connection with tso enabled),
> >however, when the ring is running out of descriptors while there are
> >still a few free ones, e.g. 6 descriptors are available which is not
> >enough to carry an entire packet which needs 16 descriptors, in this
> >case the 'avail_wrap_counter' should be set as the first one pending
> >being handled by guest driver in order to get a notification, and the
> >'last_avail_wrap_counter' should stay unchanged to the head of available
> >descriptors, like below:
> >
> >Mark meaning:
> >     | | -- available
> >     |*| -- used
> >
> >A Snapshot of the queue:
> >                                       last_avail_idx = 253
> >                                       last_avail_wrap_counter = 1
> >                                              |
> >     +---------------------------------------------+
> >  0  | | | |*|*|*|*|*|*|*|*|*|*|*|*|*|*|*|*|*| | | | 255
> >     +---------------------------------------------+
> >            |
> >   shadow_avail_idx = 3
> >   avail_wrap_counter = 0
> 
> 
> Well this might not be the good place to describe the difference between
> shadow_avail_idx and last_avail_idx. And the comments above their definition
> looks good enough?

Sorry, I do not get it, can you elaborate more? 

This is one of the buggy parts of packed ring, it is good to make it clear here.

> 
>     /* Next head to pop */
>     uint16_t last_avail_idx;
> 
>     /* Last avail_idx read from VQ. */
>     uint16_t shadow_avail_idx;
> 

What is the meaning of these comment? Do you mean I should replace put them 
to the comments also? thanks.

> Instead, how or why need event suppress is not mentioned ...

Yes, but presumably this knowledge has been acquired from reading through the
spec, so I skipped this part.

Wei

> 
> 
> 
> >
> >Signed-off-by: Wei Xu <wexu@redhat.com>
> >---
> >  hw/virtio/virtio.c | 137 +++++++++++++++++++++++++++++++++++++++++++++++++----
> >  1 file changed, 128 insertions(+), 9 deletions(-)
> >
> >diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c
> >index 7e276b4..8cfc7b6 100644
> >--- a/hw/virtio/virtio.c
> >+++ b/hw/virtio/virtio.c
> >@@ -234,6 +234,34 @@ static void vring_desc_read(VirtIODevice *vdev, VRingDesc *desc,
> >      virtio_tswap16s(vdev, &desc->next);
> >  }
> >+static void vring_packed_event_read(VirtIODevice *vdev,
> >+                            MemoryRegionCache *cache, VRingPackedDescEvent *e)
> >+{
> >+    address_space_read_cached(cache, 0, e, sizeof(*e));
> >+    virtio_tswap16s(vdev, &e->off_wrap);
> >+    virtio_tswap16s(vdev, &e->flags);
> >+}
> >+
> >+static void vring_packed_off_wrap_write(VirtIODevice *vdev,
> >+                            MemoryRegionCache *cache, uint16_t off_wrap)
> >+{
> >+    virtio_tswap16s(vdev, &off_wrap);
> >+    address_space_write_cached(cache, offsetof(VRingPackedDescEvent, off_wrap),
> >+                            &off_wrap, sizeof(off_wrap));
> >+    address_space_cache_invalidate(cache,
> >+                offsetof(VRingPackedDescEvent, off_wrap), sizeof(off_wrap));
> >+}
> >+
> >+static void vring_packed_flags_write(VirtIODevice *vdev,
> >+                            MemoryRegionCache *cache, uint16_t flags)
> >+{
> >+    virtio_tswap16s(vdev, &flags);
> >+    address_space_write_cached(cache, offsetof(VRingPackedDescEvent, flags),
> >+                            &flags, sizeof(flags));
> >+    address_space_cache_invalidate(cache,
> >+                        offsetof(VRingPackedDescEvent, flags), sizeof(flags));
> >+}
> >+
> >  static VRingMemoryRegionCaches *vring_get_region_caches(struct VirtQueue *vq)
> >  {
> >      VRingMemoryRegionCaches *caches = atomic_rcu_read(&vq->vring.caches);
> >@@ -340,14 +368,8 @@ static inline void vring_set_avail_event(VirtQueue *vq, uint16_t val)
> >      address_space_cache_invalidate(&caches->used, pa, sizeof(val));
> >  }
> >-void virtio_queue_set_notification(VirtQueue *vq, int enable)
> >+static void virtio_queue_set_notification_split(VirtQueue *vq, int enable)
> >  {
> >-    vq->notification = enable;
> >-
> >-    if (!vq->vring.desc) {
> >-        return;
> >-    }
> >-
> >      rcu_read_lock();
> >      if (virtio_vdev_has_feature(vq->vdev, VIRTIO_RING_F_EVENT_IDX)) {
> >          vring_set_avail_event(vq, vring_avail_idx(vq));
> >@@ -363,6 +385,57 @@ void virtio_queue_set_notification(VirtQueue *vq, int enable)
> >      rcu_read_unlock();
> >  }
> >+static void virtio_queue_set_notification_packed(VirtQueue *vq, int enable)
> >+{
> >+    VRingPackedDescEvent e;
> >+    VRingMemoryRegionCaches *caches;
> >+
> >+    rcu_read_lock();
> >+    caches  = vring_get_region_caches(vq);
> >+    vring_packed_event_read(vq->vdev, &caches->used, &e);
> >+
> >+    if (!enable) {
> >+        if (virtio_vdev_has_feature(vq->vdev, VIRTIO_RING_F_EVENT_IDX)) {
> >+            /* no need to write device area since this is outdated. */
> 
> 
> We should advertise off and wrap in this case as well, otherwise we may get
> notifications earlier than expected.

Is it necessary? Supposing offset & wrap_counter are always set before update
notification flags, it is reliable to skip this for disabling, isn't it?

While the logic here is unclear, I did a concise one like below
which doesn't use the 'vq->notification' as in your comment for v2,
I think this should work for packed ring as well, anything I missed?

    if (!enable) {
        e.flags = VRING_PACKED_EVENT_FLAG_DISABLE;
    } else if (virtio_vdev_has_feature(vq->vdev, VIRTIO_RING_F_EVENT_IDX)) {
        uint16_t off_wrap;

        off_wrap = vq->shadow_avail_idx | vq->shadow_avail_wrap_counter << 15;
        vring_packed_off_wrap_write(vq->vdev, &caches->used, off_wrap);

        /* Make sure off_wrap is wrote before flags */
        smp_wmb();
        e.flags = VRING_PACKED_EVENT_FLAG_DESC;
    } else {
        e.flags = VRING_PACKED_EVENT_FLAG_ENABLE;
    }

    vring_packed_flags_write(vq->vdev, &caches->used, e.flags);

> 
> 
> >+            goto out;
> >+        }
> >+
> >+        e.flags = VRING_PACKED_EVENT_FLAG_DISABLE;
> >+        goto update;
> >+    }
> >+
> >+    e.flags = VRING_PACKED_EVENT_FLAG_ENABLE;
> 
> 
> Here and the above goto could be eliminated by:
> 
> if (even idx) {
> 
> ...
> 
> } else if (enable) {
> 
> ...
> 
> } else {
> 
> ...
> 
> }
> 

thanks, I have removed it in above snippet.

Wei
> 
> Thanks
> 
> 
> >+    if (virtio_vdev_has_feature(vq->vdev, VIRTIO_RING_F_EVENT_IDX)) {
> >+        uint16_t off_wrap = vq->shadow_avail_idx | vq->avail_wrap_counter << 15;
> >+
> >+        vring_packed_off_wrap_write(vq->vdev, &caches->used, off_wrap);
> >+        /* Make sure off_wrap is wrote before flags */
> >+        smp_wmb();
> >+
> >+        e.flags = VRING_PACKED_EVENT_FLAG_DESC;
> >+    }
> >+
> >+update:
> >+    vring_packed_flags_write(vq->vdev, &caches->used, e.flags);
> >+out:
> >+    rcu_read_unlock();
> >+}
> >+
> >+void virtio_queue_set_notification(VirtQueue *vq, int enable)
> >+{
> >+    vq->notification = enable;
> >+
> >+    if (!vq->vring.desc) {
> >+        return;
> >+    }
> >+
> >+    if (virtio_vdev_has_feature(vq->vdev, VIRTIO_F_RING_PACKED)) {
> >+        virtio_queue_set_notification_packed(vq, enable);
> >+    } else {
> >+        virtio_queue_set_notification_split(vq, enable);
> >+    }
> >+}
> >+
> >  int virtio_queue_ready(VirtQueue *vq)
> >  {
> >      return vq->vring.avail != 0;
> >@@ -2117,8 +2190,7 @@ static void virtio_set_isr(VirtIODevice *vdev, int value)
> >      }
> >  }
> >-/* Called within rcu_read_lock().  */
> >-static bool virtio_should_notify(VirtIODevice *vdev, VirtQueue *vq)
> >+static bool virtio_split_should_notify(VirtIODevice *vdev, VirtQueue *vq)
> >  {
> >      uint16_t old, new;
> >      bool v;
> >@@ -2141,6 +2213,53 @@ static bool virtio_should_notify(VirtIODevice *vdev, VirtQueue *vq)
> >      return !v || vring_need_event(vring_get_used_event(vq), new, old);
> >  }
> >+static bool vring_packed_need_event(VirtQueue *vq, bool wrap,
> >+                            uint16_t off_wrap, uint16_t new, uint16_t old)
> >+{
> >+    int off = off_wrap & ~(1 << 15);
> >+
> >+    if (wrap != off_wrap >> 15) {
> >+        off -= vq->vring.num;
> >+    }
> >+
> >+    return vring_need_event(off, new, old);
> >+}
> >+
> >+static bool virtio_packed_should_notify(VirtIODevice *vdev, VirtQueue *vq)
> >+{
> >+    VRingPackedDescEvent e;
> >+    uint16_t old, new;
> >+    bool v;
> >+    VRingMemoryRegionCaches *caches;
> >+
> >+    caches  = vring_get_region_caches(vq);
> >+    vring_packed_event_read(vdev, &caches->avail, &e);
> >+
> >+    old = vq->signalled_used;
> >+    new = vq->signalled_used = vq->used_idx;
> >+    v = vq->signalled_used_valid;
> >+    vq->signalled_used_valid = true;
> >+
> >+    if (e.flags == VRING_PACKED_EVENT_FLAG_DISABLE) {
> >+        return false;
> >+    } else if (e.flags == VRING_PACKED_EVENT_FLAG_ENABLE) {
> >+        return true;
> >+    }
> >+
> >+    return !v || vring_packed_need_event(vq,
> >+        vq->used_wrap_counter, e.off_wrap, new, old);
> >+}
> >+
> >+/* Called within rcu_read_lock().  */
> >+static bool virtio_should_notify(VirtIODevice *vdev, VirtQueue *vq)
> >+{
> >+    if (virtio_vdev_has_feature(vdev, VIRTIO_F_RING_PACKED)) {
> >+        return virtio_packed_should_notify(vdev, vq);
> >+    } else {
> >+        return virtio_split_should_notify(vdev, vq);
> >+    }
> >+}
> >+
> >  void virtio_notify_irqfd(VirtIODevice *vdev, VirtQueue *vq)
> >  {
> >      bool should_notify;
>