On Thu, Feb 27, 2025 at 6:13 AM Michael S. Tsirkin [off-list ref] wrote:

On Wed, Feb 26, 2025 at 03:27:42PM -0500, Joe Damato wrote:

quoted

On Wed, Feb 26, 2025 at 01:08:49PM -0500, Joe Damato wrote:

quoted

On Wed, Feb 26, 2025 at 01:03:09PM -0500, Joe Damato wrote:

quoted

On Wed, Feb 26, 2025 at 01:48:50PM +0800, Jason Wang wrote:

quoted

On Tue, Feb 25, 2025 at 10:05 AM Joe Damato [off-list ref] wrote:

quoted

Use netif_queue_set_napi to map NAPIs to queue IDs so that the mapping
can be accessed by user apps, taking care to hold RTNL as needed.

I may miss something but I wonder whether letting the caller hold the
lock is better.

Hmm...

Double checking all the paths over again, here's what I see:
  - refill_work, delayed work that needs RTNL so this change seems
    right?

  - virtnet_disable_queue_pair, called from virtnet_open and
    virtnet_close. When called via NDO these are safe and hold RTNL,
    but they can be called from power management and need RTNL.

  - virtnet_enable_queue_pair called from virtnet_open, safe when
    used via NDO but needs RTNL when used via power management.

  - virtnet_rx_pause called in both paths as you mentioned, one
    which needs RTNL and one which doesn't.

Sorry, I missed more paths:

    - virtnet_rx_resume
    - virtnet_tx_pause and virtnet_tx_resume

which are similar to path you mentioned (virtnet_rx_pause) and need
rtnl in one of two different paths.

Let me know if I missed any paths and what your preferred way to fix
this would be?

I think both options below are possible and I have no strong
preference.

OK, my apologies. I read your message and the code wrong. Sorry for
the back-to-back emails from me.

Please ignore my message above... I think after re-reading the code,
here's where I've arrived:

  - refill_work needs to hold RTNL (as in the existing patch)

  - virtnet_rx_pause, virtnet_rx_resume, virtnet_tx_pause,
    virtnet_tx_resume -- all do NOT need to hold RTNL because it is
    already held in the ethtool resize path and the XSK path, as you
    explained, but I mis-read (sorry).

  - virtnet_disable_queue_pair and virtnet_enable_queue_pair both
    need to hold RTNL only when called via power management, but not
    when called via ndo_open or ndo_close

Is my understanding correct and does it match your understanding?

If so, that means there are two issues:

  1. Fixing the hardcoded bools in rx_pause, rx_resume, tx_pause,
     tx_resume (all should be false, RTNL is not needed).

  2. Handling the power management case which calls virtnet_open and
     virtnet_close.

I made a small diff included below as an example of a possible
solution:

  1. Modify virtnet_disable_queue_pair and virtnet_enable_queue_pair
     to take a "bool need_rtnl" and pass it through to the helpers
     they call.

  2. Create two helpers, virtnet_do_open and virt_do_close both of
     which take struct net_device *dev, bool need_rtnl. virtnet_open
     and virtnet_close are modified to call the helpers and pass
     false for need_rtnl. The power management paths call the
     helpers and pass true for need_rtnl. (fixes issue 2 above)

  3. Fix the bools for rx_pause, rx_resume, tx_pause, tx_resume to
     pass false since all paths that I could find that lead to these
     functions hold RTNL. (fixes issue 1 above)

See the diff below (which can be applied on top of patch 3) to see
what it looks like.

If you are OK with this approach, I will send a v5 where patch 3
includes the changes shown in this diff.

Please let me know what you think:

Looks ok I think.

quoted

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 13bb4a563073..76ecb8f3ce9a 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -3098,14 +3098,16 @@ static int virtnet_poll(struct napi_struct *napi, int budget)
      return received;
 }

-static void virtnet_disable_queue_pair(struct virtnet_info *vi, int qp_index)
+static void virtnet_disable_queue_pair(struct virtnet_info *vi, int qp_index,
+                                    bool need_rtnl)
 {
-     virtnet_napi_tx_disable(&vi->sq[qp_index], false);
-     virtnet_napi_disable(&vi->rq[qp_index], false);
+     virtnet_napi_tx_disable(&vi->sq[qp_index], need_rtnl);
+     virtnet_napi_disable(&vi->rq[qp_index], need_rtnl);
      xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
 }

-static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
+static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index,
+                                  bool need_rtnl)
 {
      struct net_device *dev = vi->dev;
      int err;

@@ -3120,8 +3122,8 @@ static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
      if (err < 0)
              goto err_xdp_reg_mem_model;

-     virtnet_napi_enable(&vi->rq[qp_index], false);
-     virtnet_napi_tx_enable(&vi->sq[qp_index], false);
+     virtnet_napi_enable(&vi->rq[qp_index], need_rtnl);
+     virtnet_napi_tx_enable(&vi->sq[qp_index], need_rtnl);

      return 0;

@@ -3156,7 +3158,7 @@ static void virtnet_update_settings(struct virtnet_info *vi)
              vi->duplex = duplex;
 }

-static int virtnet_open(struct net_device *dev)
+static int virtnet_do_open(struct net_device *dev, bool need_rtnl)
 {
      struct virtnet_info *vi = netdev_priv(dev);
      int i, err;

@@ -3169,7 +3171,7 @@ static int virtnet_open(struct net_device *dev)
                      if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
                              schedule_delayed_work(&vi->refill, 0);

-             err = virtnet_enable_queue_pair(vi, i);
+             err = virtnet_enable_queue_pair(vi, i, need_rtnl);
              if (err < 0)
                      goto err_enable_qp;
      }

@@ -3190,13 +3192,18 @@ static int virtnet_open(struct net_device *dev)
      cancel_delayed_work_sync(&vi->refill);

      for (i--; i >= 0; i--) {
-             virtnet_disable_queue_pair(vi, i);
+             virtnet_disable_queue_pair(vi, i, need_rtnl);
              virtnet_cancel_dim(vi, &vi->rq[i].dim);
      }

      return err;
 }

+static int virtnet_open(struct net_device *dev)
+{
+     return virtnet_do_open(dev, false);
+}
+
 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
 {
      struct send_queue *sq = container_of(napi, struct send_queue, napi);

@@ -3373,7 +3380,7 @@ static void virtnet_rx_pause(struct virtnet_info *vi, struct receive_queue *rq)
      bool running = netif_running(vi->dev);

      if (running) {
-             virtnet_napi_disable(rq, true);
+             virtnet_napi_disable(rq, false);
              virtnet_cancel_dim(vi, &rq->dim);
      }
 }

@@ -3386,7 +3393,7 @@ static void virtnet_rx_resume(struct virtnet_info *vi, struct receive_queue *rq)
              schedule_delayed_work(&vi->refill, 0);

      if (running)
-             virtnet_napi_enable(rq, true);
+             virtnet_napi_enable(rq, false);
 }

 static int virtnet_rx_resize(struct virtnet_info *vi,

@@ -3415,7 +3422,7 @@ static void virtnet_tx_pause(struct virtnet_info *vi, struct send_queue *sq)
      qindex = sq - vi->sq;

      if (running)
-             virtnet_napi_tx_disable(sq, true);
+             virtnet_napi_tx_disable(sq, false);

      txq = netdev_get_tx_queue(vi->dev, qindex);

@@ -3449,7 +3456,7 @@ static void virtnet_tx_resume(struct virtnet_info *vi, struct send_queue *sq)
      __netif_tx_unlock_bh(txq);

      if (running)
-             virtnet_napi_tx_enable(sq, true);
+             virtnet_napi_tx_enable(sq, false);

Instead of this, it looks to me it would be much simpler if we can
just hold the rtnl lock in freeze/restore.

Thanks