On 12/11/25 14:27, Jason Wang wrote:

On Wed, Dec 10, 2025 at 11:33 PM Bui Quang Minh
[off-list ref] wrote:

quoted

On 12/10/25 12:45, Jason Wang wrote:

quoted

On Tue, Dec 9, 2025 at 11:23 PM Bui Quang Minh [off-list ref] wrote:

quoted

On 12/9/25 11:30, Jason Wang wrote:

quoted

On Mon, Dec 8, 2025 at 11:35 PM Bui Quang Minh [off-list ref] wrote:

quoted

Calling napi_disable() on an already disabled napi can cause the
deadlock. In commit 4bc12818b363 ("virtio-net: disable delayed refill
when pausing rx"), to avoid the deadlock, when pausing the RX in
virtnet_rx_pause[_all](), we disable and cancel the delayed refill work.
However, in the virtnet_rx_resume_all(), we enable the delayed refill
work too early before enabling all the receive queue napis.

The deadlock can be reproduced by running
selftests/drivers/net/hw/xsk_reconfig.py with multiqueue virtio-net
device and inserting a cond_resched() inside the for loop in
virtnet_rx_resume_all() to increase the success rate. Because the worker
processing the delayed refilled work runs on the same CPU as
virtnet_rx_resume_all(), a reschedule is needed to cause the deadlock.
In real scenario, the contention on netdev_lock can cause the
reschedule.

This fixes the deadlock by ensuring all receive queue's napis are
enabled before we enable the delayed refill work in
virtnet_rx_resume_all() and virtnet_open().

Fixes: 4bc12818b363 ("virtio-net: disable delayed refill when pausing rx")
Reported-by: Paolo Abeni <pabeni@redhat.com>
Closes: https://netdev-ctrl.bots.linux.dev/logs/vmksft/drv-hw-dbg/results/400961/3-xdp-py/stderr
Signed-off-by: Bui Quang Minh <redacted>
---
    drivers/net/virtio_net.c | 59 +++++++++++++++++++---------------------
    1 file changed, 28 insertions(+), 31 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 8e04adb57f52..f2b1ea65767d 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -2858,6 +2858,20 @@ static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
           return err != -ENOMEM;
    }

+static void virtnet_rx_refill_all(struct virtnet_info *vi)
+{
+       bool schedule_refill = false;
+       int i;
+
+       enable_delayed_refill(vi);

This seems to be still racy?

For example, in virtnet_open() we had:

static int virtnet_open(struct net_device *dev)
{
           struct virtnet_info *vi = netdev_priv(dev);
           int i, err;

           for (i = 0; i < vi->max_queue_pairs; i++) {
                   err = virtnet_enable_queue_pair(vi, i);
                   if (err < 0)
                           goto err_enable_qp;
           }

           virtnet_rx_refill_all(vi);

So NAPI and refill work is enabled in this case, so the refill work
could be scheduled and run at the same time?

Yes, that's what we expect. We must ensure that refill work is scheduled
only when all NAPIs are enabled. The deadlock happens when refill work
is scheduled but there are still disabled RX NAPIs.

Just to make sure we are on the same page, I meant, after refill work
is enabled, rq0 is NAPI is enabled, in this case the refill work could
be triggered by the rq0's NAPI so we may end up in the refill work
that it tries to disable rq1's NAPI while holding the netdev lock.

I don't quite get your point. The current deadlock scenario is this

virtnet_rx_resume_all
napi_enable(rq0) (the rq1 napi is still disabled)
enable_refill_work

refill_work
napi_disable(rq0) -> still okay
napi_enable(rq0) -> still okay
napi_disable(rq1)
-> hold netdev_lock
      -> stuck inside the while loop in napi_disable_locked
              while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) {
                  usleep_range(20, 200);
                  val = READ_ONCE(n->state);
              }


napi_enable(rq1)
-> stuck while trying to acquire the netdev_lock

The problem is that we must not call napi_disable() on an already
disabled NAPI (rq1's NAPI in the example).

In the new virtnet_open

static int virtnet_open(struct net_device *dev)
{
           struct virtnet_info *vi = netdev_priv(dev);
           int i, err;

           // Note that at this point, refill work is still disabled, vi->refill_enabled == false,
           // so even if virtnet_receive is called, the refill_work will not be scheduled.
           for (i = 0; i < vi->max_queue_pairs; i++) {
                   err = virtnet_enable_queue_pair(vi, i);
                   if (err < 0)
                           goto err_enable_qp;
           }

           // Here all RX NAPIs are enabled so it's safe to enable refill work again
           virtnet_rx_refill_all(vi);

I meant this part:

+static void virtnet_rx_refill_all(struct virtnet_info *vi)
+{
+       bool schedule_refill = false;
+       int i;
+
+       enable_delayed_refill(vi);

refill_work could run here.

I don't see how this can trigger the current deadlock race. However, I 
see that this code is racy, the try_fill_recv function is not safe to 
concurrently executed on the same receive queue. So there is a 
requirement that we need to call try_fill_recv before enabling napi. Is 
it what you mean?

+       for (i = 0; i < vi->curr_queue_pairs; i++)
+               if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
+                       schedule_refill = true;
+

I think it can be fixed by moving enable_delayed_refill() here.

+       if (schedule_refill)
+               schedule_delayed_work(&vi->refill, 0);
+}

Thanks,
Quang Minh.