Re: [PATCH 1/2] vdpa/mlx5: Avoid unnecessary query virtqueue

On Tue, Feb 2, 2021 at 9:16 PM Jason Wang [off-list ref] wrote:

On 2021/2/3 上午1:54, Si-Wei Liu wrote:

quoted

On Tue, Feb 2, 2021 at 1:23 AM Eli Cohen [off-list ref] wrote:

quoted

On Tue, Feb 02, 2021 at 12:38:51AM -0800, Si-Wei Liu wrote:

quoted

Thanks Eli and Jason for clarifications. See inline.

On Mon, Feb 1, 2021 at 11:06 PM Eli Cohen [off-list ref] wrote:

quoted

On Tue, Feb 02, 2021 at 02:02:25PM +0800, Jason Wang wrote:

quoted

On 2021/2/2 下午12:15, Si-Wei Liu wrote:

quoted

On Mon, Feb 1, 2021 at 7:13 PM Jason Wang [off-list ref] wrote:

quoted

On 2021/2/2 上午3:17, Si-Wei Liu wrote:

quoted

On Mon, Feb 1, 2021 at 10:51 AM Si-Wei Liu [off-list ref] wrote:

quoted

On Thu, Jan 28, 2021 at 5:46 AM Eli Cohen [off-list ref] wrote:

quoted

suspend_vq should only suspend the VQ on not save the current available
index. This is done when a change of map occurs when the driver calls
save_channel_info().

Hmmm, suspend_vq() is also called by teardown_vq(), the latter of
which doesn't save the available index as save_channel_info() doesn't
get called in that path at all. How does it handle the case that
aget_vq_state() is called from userspace (e.g. QEMU) while the
hardware VQ object was torn down, but userspace still wants to access
the queue index?

Refer to https://lore.kernel.org/netdev/1601583511-15138-1-git-send-email-si-wei.liu@oracle.com/ (local)

vhost VQ 0 ring restore failed: -1: Resource temporarily unavailable (11)
vhost VQ 1 ring restore failed: -1: Resource temporarily unavailable (11)

QEMU will complain with the above warning while VM is being rebooted
or shut down.

Looks to me either the kernel driver should cover this requirement, or
the userspace has to bear the burden in saving the index and not call
into kernel if VQ is destroyed.

Actually, the userspace doesn't have the insights whether virt queue
will be destroyed if just changing the device status via set_status().
Looking at other vdpa driver in tree i.e. ifcvf it doesn't behave like
so. Hence this still looks to me to be Mellanox specifics and
mlx5_vdpa implementation detail that shouldn't expose to userspace.

So I think we can simply drop this patch?

Yep, I think so. To be honest I don't know why it has anything to do
with the memory hotplug issue.

Eli may know more, my understanding is that, during memory hotplut, qemu
need to propagate new memory mappings via set_map(). For mellanox, it means
it needs to rebuild memory keys, so the virtqueue needs to be suspended.

I think Siwei was asking why the first patch was related to the hotplug
issue.

I was thinking how consistency is assured when saving/restoring this
h/w avail_index against the one in the virtq memory, particularly in
the region_add/.region_del() context (e.g. the hotplug case). Problem
is I don't see explicit memory barrier when guest thread updates the
avail_index, how does the device make sure the h/w avail_index is
uptodate while guest may race with updating the virtq's avail_index in
the mean while? Maybe I completely miss something obvious?

DKIM-Signature: v1; arsa-sha256; crelaxed/relaxed; dnvidia.com; sn1;
         t 12257780; bhHnB0z4VEKwRS3WGY8d836MJgxu5Eln/jbFZlNXVxc08;
         hX-PGP-Universal:Date:From:To:CC:Subject:Message-ID:References:
          MIME-Version:Content-Type:Content-Disposition:
          Content-Transfer-Encoding:In-Reply-To:User-Agent:X-Originating-IP:
          X-ClientProxiedBy;
         bgGmb8+rcn3/rKzKQ/7QzSnghWzZ+FAU0XntsRZYGQ66sFvT7zsYPHogG3LIWNY77t
          wNHPw7GCJrNaH3nEXPbOp0FMOZw4Kv4W7UPuYPobbLeTkvuPAidjB8dM42vz+1X61t
          9IVQT9X4hnAxRjI5CqZOo41GS4Tl1X+ykGoA+VE80BR/R/+nQ3tXDVULfppzeB+vu3
          TWnnpaZ2GyoNyPlMiyVRkHdXzDVgA4uQHxwHn7otGK5J4lzyu8KrFyQtiP+f6hfu5v
          crJkYS8e9A+rfzUmKWuyHcKcmhPhAVJ4XdpzZcDXXlMHVxG7nR1o88xttC6D1+oNIP
          9xHI3DkNBpEuA
If you're asking about syncronization upon hot plug of memory, the
hardware always goes to read the available index from memory when a new
hardware object is associted with a virtqueue. You can argue then that
you don't need to restore the available index and you may be right but
this is the currect inteface to the firmware.


If you're asking on generally how sync is assured when the guest updates
the available index, can you please send a pointer to the code where you
see the update without a memory barrier?

This is a snippet of virtqueue_add_split() where avail_index gets
updated by guest:

         /* Put entry in available array (but don't update avail->idx until they
          * do sync). */
         avail = vq->split.avail_idx_shadow & (vq->split.vring.num - 1);
         vq->split.vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);

         /* Descriptors and available array need to be set before we expose the
          * new available array entries. */
         virtio_wmb(vq->weak_barriers);
         vq->split.avail_idx_shadow++;
         vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
                                                 vq->split.avail_idx_shadow);
         vq->num_added++;

There's memory barrier to make sure the update to descriptor and
available ring is seen before writing to the avail->idx, but there
seems no gurantee that this update would flush to the memory
immmedately either before or after the mlx5-vdpa is suspened and gets
the old avail_index value stashed somewhere. In this case, how does
the hardware ensure the consistency between the guest virtio and host
mlx5-vdpa? Or, it completly relies on guest to update the avail_index
once the next buffer is available, so that the index will be in sync
again?

I'm not sure I get the question but notice that the driver should check
and notify virtqueue when device want a notification. So there's a
virtio_wmb() e.g in:

static bool virtqueue_kick_prepare_split(struct virtqueue *_vq)
{
     struct vring_virtqueue *vq = to_vvq(_vq);
     u16 new, old;
     bool needs_kick;

     START_USE(vq);
     /* We need to expose available array entries before checking avail
      * event. */
     virtio_mb(vq->weak_barries);

     old = vq->split.avail_idx_shadow - vq->num_added;
     new = vq->split.avail_idx_shadow;
     vq->num_added = 0;

(See the comment above virtio_mb()). So the avail idx is guaranteed to
be committed to memroy(cache hierarchy) before the check of
notification. I think we sync through this.

Thanks for pointing it out! Indeed, the avail index is synced before
kicking the device. However, even so I think the race might still be
possible, see below:


  mlx5_vdpa device                       virtio driver
-------------------------------------------------------------------------
                                  virtqueue_add_split
                                    (bumped up avail_idx1 to
                                    avail_idx2, but update
                                    not yet committed to mem)

(hot plug memory...)
mlx5_vdpa_set_map
  mlx5_vdpa_change_map
    suspend_vqs
      suspend_vq
        (avail_idx1 was saved)
    save_channels_info
    :
    :                             virtqueue_kick_prepare_split
    :                               (avail_idx2 committed to memory)
    restore_channels_info
    setup_driver
      ...
        create_virtqueue
          (saved avail_idx1
          getting restored;
          whereas current
          avail_idx2 in
          memory is ignored)
:
:
                                   virtqueue_notify
                                     vp_notify
mlx5_vdpa_kick_vq
  (device processing up to
  avail_idx1 rather than
  avail_idx2?)


According to Eli, the vdpa firmware does not load the latest value,
i.e. avail_idx2 from memory when re-creating the virtqueue object.
Instead it starts with a stale avail_idx1 that was saved in mlx5_vdpa
driver private structure before the memory update is made. That is the
way how the current firmware interface is designed. The thing I'm not
sure though is if the firmware/device will resync and get to the
latest avail_idx2 from memory while processing the kick request with
mlx5_vdpa_kick_vq? If not, a few avail entries will be missing in the
kick, which could cause odd behavior or implicit failure.

But if the firmware will resync on kick_vq (and get_vq_state), I think
I completely lost the point of saving and restoring avail_idx when
changing the memory map...

Thanks,
-Siwei

Thanks

quoted

Thanks,
-Siwei

quoted

quoted

Thanks,
-Siwei

quoted

But you're correct. When memory is added, I get a new memory map. This
requires me to build a new memory key object which covers the new memory
map. Since the virtqueue objects are referencing this memory key, I need
to destroy them and build new ones referncing the new memory key.

quoted

Thanks

quoted

-Siwei

quoted

Thanks

quoted

quoted

-Siwei

quoted

Signed-off-by: Eli Cohen <redacted>
---
    drivers/vdpa/mlx5/net/mlx5_vnet.c | 8 --------
    1 file changed, 8 deletions(-)

diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
index 88dde3455bfd..549ded074ff3 100644
--- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
+++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c

@@ -1148,8 +1148,6 @@ static int setup_vq(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)

    static void suspend_vq(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq)
    {
-       struct mlx5_virtq_attr attr;
-
           if (!mvq->initialized)
                   return;

@@ -1158,12 +1156,6 @@ static void suspend_vq(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *m

           if (modify_virtqueue(ndev, mvq, MLX5_VIRTIO_NET_Q_OBJECT_STATE_SUSPEND))
                   mlx5_vdpa_warn(&ndev->mvdev, "modify to suspend failed\n");
-
-       if (query_virtqueue(ndev, mvq, &attr)) {
-               mlx5_vdpa_warn(&ndev->mvdev, "failed to query virtqueue\n");
-               return;
-       }
-       mvq->avail_idx = attr.available_index;
    }

    static void suspend_vqs(struct mlx5_vdpa_net *ndev)
--

2.29.2