Skip to content

L
linux
Switch branch/tag
  1. 20 May, 2022 7 commits
  3. 19 May, 2022 30 commits
  5. 18 May, 2022 3 commits
    • Linus Torvalds's avatar
      Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost · db1fd3fc
      Linus Torvalds authored 
      Pull mlx5 fix from Michael Tsirkin:
 "One last minute fixup

  The patch has been on list for a while but as it was posted as part of
  a thread it was missed"

* tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost:
  vdpa/mlx5: Use consistent RQT size
      db1fd3fc
    • Al Viro's avatar
      Fix double fget() in vhost_net_set_backend() · fb4554c2
      Al Viro authored 
      Descriptor table is a shared resource; two fget() on the same descriptor
may return different struct file references.  get_tap_ptr_ring() is
called after we'd found (and pinned) the socket we'll be using and it
tries to find the private tun/tap data structures associated with it.
Redoing the lookup by the same file descriptor we'd used to get the
socket is racy - we need to same struct file.

Thanks to Jason for spotting a braino in the original variant of patch -
I'd missed the use of fd == -1 for disabling backend, and in that case
we can end up with sock == NULL and sock != oldsock.

Cc: stable@kernel.org
Acked-by: default avatarMichael S. Tsirkin <mst@redhat.com>
Signed-off-by: default avatarJason Wang <jasowang@redhat.com>
Signed-off-by: default avatarAl Viro <viro@zeniv.linux.org.uk>
      fb4554c2
    • Eli Cohen's avatar
      vdpa/mlx5: Use consistent RQT size · acde3929
      Eli Cohen authored 
      The current code evaluates RQT size based on the configured number of
virtqueues. This can raise an issue in the following scenario:

Assume MQ was negotiated.
1. mlx5_vdpa_set_map() gets called.
2. handle_ctrl_mq() is called setting cur_num_vqs to some value, lower
   than the configured max VQs.
3. A second set_map gets called, but now a smaller number of VQs is used
   to evaluate the size of the RQT.
4. handle_ctrl_mq() is called with a value larger than what the RQT can
   hold. This will emit errors and the driver state is compromised.

To fix this, we use a new field in struct mlx5_vdpa_net to hold the
required number of entries in the RQT. This value is evaluated in
mlx5_vdpa_set_driver_features() where we have the negotiated features
all set up.

In addition to that, we take into consideration the max capability of RQT
entries early when the device is added so we don't need to take consider
it when creating the RQT.

Last, we remove the use of mlx5_vdpa_max_qps() which just returns the
max_vas / 2 and make the code clearer.

Fixes: 52893733 ("vdpa/mlx5: Add multiqueue support")
Acked-by: default avatarJason Wang <jasowang@redhat.com>
Signed-off-by: default avatarEli Cohen <elic@nvidia.com>
Signed-off-by: default avatarMichael S. Tsirkin <mst@redhat.com>
      acde3929
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7