Re: [PATCH net-next 3/3] virtio_net: Map NAPIs to queues

On Mon, Jan 13, 2025 at 02:04:46PM -0800, Jakub Kicinski wrote:

On Mon, 13 Jan 2025 09:30:20 -0800 Joe Damato wrote:

quoted

quoted

quoted

 static void virtnet_napi_enable_lock(struct virtqueue *vq,
-                                    struct napi_struct *napi)
+                                    struct napi_struct *napi,
+                                    bool need_rtnl)
 {
+       struct virtnet_info *vi = vq->vdev->priv;
+       int q = vq2rxq(vq);
+
        virtnet_napi_do_enable(vq, napi);
+
+       if (q < vi->curr_queue_pairs) {
+               if (need_rtnl)
+                       rtnl_lock();  

Can we tweak the caller to call rtnl_lock() instead to avoid this trick?  

The major problem is that if the caller calls rtnl_lock() before
calling virtnet_napi_enable_lock, then virtnet_napi_do_enable (and
thus napi_enable) happen under the lock.

Jakub mentioned in a recent change [1] that napi_enable may soon
need to sleep.

Given the above constraints, the only way to avoid the "need_rtnl"
would be to refactor the code much more, placing calls (or wrappers)
to netif_queue_set_napi in many locations.

IMHO: This implementation seemed cleaner than putting calls to
netif_queue_set_napi throughout the driver.

Please let me know how you'd like to proceed on this.

[1]: https://lore.kernel.org/netdev/20250111024742.3680902-1-kuba@kernel.org/ (local)

I'm going to make netif_queue_set_napi() take netdev->lock, and remove
the rtnl_lock requirement ~this week. If we need conditional locking
perhaps we're better off waiting?

That seems reasonable to me and I can wait.

Please CC me on that series so I can take a look and I'll adjust the
v2 of this series to avoid the locking once your series is merged.

Thanks!