On 12/17/25 09:58, Jason Wang wrote:
On Wed, Dec 17, 2025 at 12:23 AM Bui Quang Minh
<[email protected]> wrote:
On 12/16/25 11:16, Jason Wang wrote:
On Fri, Dec 12, 2025 at 11:28 PM Bui Quang Minh
<[email protected]> wrote:
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 <[email protected]>
Closes:
https://netdev-ctrl.bots.linux.dev/logs/vmksft/drv-hw-dbg/results/400961/3-xdp-py/stderr
Cc: [email protected]
Signed-off-by: Bui Quang Minh <[email protected]>
---
Changes in v2:
- Move try_fill_recv() before rx napi_enable()
- Link to v1:
https://lore.kernel.org/netdev/[email protected]/
---
drivers/net/virtio_net.c | 71 +++++++++++++++++++++++++---------------
1 file changed, 45 insertions(+), 26 deletions(-)
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 8e04adb57f52..4e08880a9467 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -3214,21 +3214,31 @@ static void virtnet_update_settings(struct virtnet_info
*vi)
static int virtnet_open(struct net_device *dev)
{
struct virtnet_info *vi = netdev_priv(dev);
+ bool schedule_refill = false;
int i, err;
- enable_delayed_refill(vi);
-
+ /* - We must call try_fill_recv before enabling napi of the same receive
+ * queue so that it doesn't race with the call in virtnet_receive.
+ * - We must enable and schedule delayed refill work only when we have
+ * enabled all the receive queue's napi. Otherwise, in refill_work, we
+ * have a deadlock when calling napi_disable on an already disabled
+ * napi.
+ */
for (i = 0; i < vi->max_queue_pairs; i++) {
if (i < vi->curr_queue_pairs)
/* Make sure we have some buffers: if oom use wq. */
if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
- schedule_delayed_work(&vi->refill, 0);
+ schedule_refill = true;
err = virtnet_enable_queue_pair(vi, i);
if (err < 0)
goto err_enable_qp;
}
So NAPI could be scheduled and it may want to refill but since refill
is not enabled, there would be no refill work.
Is this a problem?
You are right. It is indeed a problem.
I think we can unconditionally schedule the delayed refill after
enabling all the RX NAPIs (don't check the boolean schedule_refill
anymore) to ensure that we will have refill work. We can still keep the
try_fill_recv here to fill the receive buffer earlier in normal case.
What do you think?
Or we can have a reill_pending
Okay, let me implement this in the next version.
but basically I think we need something
that is much more simple. That is, using a per rq work instead of a
global one?
I think we can leave this in a net-next patch later.
Thanks,
Quang Minh
Thanks
+ enable_delayed_refill(vi);
+ if (schedule_refill)
+ schedule_delayed_work(&vi->refill, 0);
+
if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
if (vi->status & VIRTIO_NET_S_LINK_UP)
netif_carrier_on(vi->dev);
@@ -3463,39 +3473,48 @@ static void virtnet_rx_pause(struct virtnet_info *vi,
struct receive_queue *rq)
__virtnet_rx_pause(vi, rq);
}
Thanks
Thanks,
Quang Minh.