releases/3.4.56/virtio-race-2of2.diff

   1 From: Michael S. Tsirkin <mst@redhat.com>
   2 Subject: virtio_net: fix race in RX VQ processing
   3
   4 From: Michael S. Tsirkin <mst@redhat.com>
   5
   6 commit cbdadbbf0c790f79350a8f36029208944c5487d0 upstream
   7
   8 virtio net called virtqueue_enable_cq on RX path after napi_complete, so
   9 with NAPI_STATE_SCHED clear - outside the implicit napi lock.
  10 This violates the requirement to synchronize virtqueue_enable_cq wrt
  11 virtqueue_add_buf.  In particular, used event can move backwards,
  12 causing us to lose interrupts.
  13 In a debug build, this can trigger panic within START_USE.
  14
  15 Jason Wang reports that he can trigger the races artificially,
  16 by adding udelay() in virtqueue_enable_cb() after virtio_mb().
  17
  18 However, we must call napi_complete to clear NAPI_STATE_SCHED before
  19 polling the virtqueue for used buffers, otherwise napi_schedule_prep in
  20 a callback will fail, causing us to lose RX events.
  21
  22 To fix, call virtqueue_enable_cb_prepare with NAPI_STATE_SCHED
  23 set (under napi lock), later call virtqueue_poll with
  24 NAPI_STATE_SCHED clear (outside the lock).
  25
  26 Reported-by: Jason Wang <jasowang@redhat.com>
  27 Tested-by: Jason Wang <jasowang@redhat.com>
  28 Acked-by: Jason Wang <jasowang@redhat.com>
  29 Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
  30 Signed-off-by: David S. Miller <davem@davemloft.net>
  31 [wg: Backported to 3.2]
  32 Signed-off-by: Wolfram Gloger <wmglo@dent.med.uni-muenchen.de>
  33 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
  34 ---
  35 ---
  36  drivers/net/virtio_net.c |    5 +++--
  37  1 file changed, 3 insertions(+), 2 deletions(-)
  38
  39 --- a/drivers/net/virtio_net.c
  40 +++ b/drivers/net/virtio_net.c
  41 @@ -518,7 +518,7 @@ static int virtnet_poll(struct napi_stru
  42  {
  43         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
  44         void *buf;
  45 -       unsigned int len, received = 0;
  46 +       unsigned int r, len, received = 0;
  47
  48  again:
  49         while (received < budget &&
  50 @@ -535,8 +535,9 @@ again:
  51
  52         /* Out of packets? */
  53         if (received < budget) {
  54 +               r = virtqueue_enable_cb_prepare(vi->rvq);
  55                 napi_complete(napi);
  56 -               if (unlikely(!virtqueue_enable_cb(vi->rvq)) &&
  57 +               if (unlikely(virtqueue_poll(vi->rvq, r)) &&
  58                     napi_schedule_prep(napi)) {
  59                         virtqueue_disable_cb(vi->rvq);
  60                         __napi_schedule(napi);