在 2024/3/27 上午10:10, Heng Qi 写道:
在 2024/3/26 下午11:18, Dan Jurgens 写道:
From: Heng Qi <hengqi@xxxxxxxxxxxxxxxxx>
Sent: Tuesday, March 26, 2024 3:55 AM
To: Dan Jurgens <danielj@xxxxxxxxxx>; netdev@xxxxxxxxxxxxxxx
Cc: mst@xxxxxxxxxx; jasowang@xxxxxxxxxx; xuanzhuo@xxxxxxxxxxxxxxxxx;
virtualization@xxxxxxxxxxxxxxx; davem@xxxxxxxxxxxxx;
edumazet@xxxxxxxxxx; kuba@xxxxxxxxxx; pabeni@xxxxxxxxxx; Jiri Pirko
<jiri@xxxxxxxxxx>
Subject: Re: [PATCH net-next 4/4] virtio_net: Remove rtnl lock
protection of
command buffers
在 2024/3/26 上午5:49, Daniel Jurgens 写道:
The rtnl lock is no longer needed to protect the control buffer and
command VQ.
Signed-off-by: Daniel Jurgens <danielj@xxxxxxxxxx>
Reviewed-by: Jiri Pirko <jiri@xxxxxxxxxx>
---
drivers/net/virtio_net.c | 27 +++++----------------------
1 file changed, 5 insertions(+), 22 deletions(-)
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c index
41f8dc16ff38..d09ea20b16be 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -2639,14 +2639,12 @@ static void virtnet_stats(struct net_device
*dev,
static void virtnet_ack_link_announce(struct virtnet_info *vi)
{
- rtnl_lock();
if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
VIRTIO_NET_CTRL_ANNOUNCE_ACK,
NULL))
dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
- rtnl_unlock();
}
-static int _virtnet_set_queues(struct virtnet_info *vi, u16
queue_pairs)
+static int virtnet_set_queues(struct virtnet_info *vi, u16
+queue_pairs)
{
struct virtio_net_ctrl_mq *mq __free(kfree) = NULL;
struct scatterlist sg;
@@ -2677,16 +2675,6 @@ static int _virtnet_set_queues(struct
virtnet_info *vi, u16 queue_pairs)
return 0;
}
-static int virtnet_set_queues(struct virtnet_info *vi, u16
queue_pairs) -{
- int err;
-
- rtnl_lock();
- err = _virtnet_set_queues(vi, queue_pairs);
- rtnl_unlock();
- return err;
-}
-
static int virtnet_close(struct net_device *dev)
{
struct virtnet_info *vi = netdev_priv(dev); @@ -3268,7
+3256,7 @@
static int virtnet_set_channels(struct net_device *dev,
return -EINVAL;
cpus_read_lock();
- err = _virtnet_set_queues(vi, queue_pairs);
+ err = virtnet_set_queues(vi, queue_pairs);
if (err) {
cpus_read_unlock();
goto err;
@@ -3558,14 +3546,11 @@ static void virtnet_rx_dim_work(struct
work_struct *work)
struct dim_cq_moder update_moder;
int i, qnum, err;
- if (!rtnl_trylock())
- return;
-
Does this guarantee that the synchronization is completely correct?
The purpose of this patch set is to add a separate lock for ctrlq
rather than
reusing the RTNL lock.
But for dim workers, it not only involves the use of ctrlq, but also
involves
reading shared variables in interfaces such as .set_coalesce,
.get_coalesce,
etc.
It looks like there is a risk of a dirty read in the get (usecs
updated, but not max_packets).
Also dim_enabled.
And later I need to asynchronousize the dim cmds, which means that
different dim workers will operate a shared linked list.
So we need a lock.
After removing the loop, maybe READ_ONCE/WRITE_ONCE will be enough?
In the set it will return -EINVAL if trying to adjust the settings
aside from DIM enabled. I can add a lock for this if you think it's
needed, but it doesn't seem like a major problem for debug info.
Not just for debug info, but future extensions as well.
These desynchronizations can introduce more trouble in the future.
Regards,
Heng
In addition, assuming there are 10 queues, each queue is scheduled
with its
own dim worker at the same time, then these 10 workers may issue
parameters to rxq0 10 times in parallel, just because the RTNL lock is
removed here.
Therefore, when the RTNL lock is removed, a 'for loop' is no longer
needed in
virtnet_rx_dim_work, and the dim worker of each queue only
configures its
own parameters.
Good point. I'll add a new patch to remove the for loop.
Alternatively, please keep RTNL lock here.
Regards,
Heng
/* Each rxq's work is queued by "net_dim()->schedule_work()"
* in response to NAPI traffic changes. Note that
dim->profile_ix
* for each rxq is updated prior to the queuing action.
* So we only need to traverse and update profiles for all rxqs
- * in the work which is holding rtnl_lock.
+ * in the work.
*/
for (i = 0; i < vi->curr_queue_pairs; i++) {
rq = &vi->rq[i];
@@ -3587,8 +3572,6 @@ static void virtnet_rx_dim_work(struct
work_struct *work)
dim->state = DIM_START_MEASURE;
}
}
-
- rtnl_unlock();
}
static int virtnet_coal_params_supported(struct ethtool_coalesce
*ec) @@ -4036,7 +4019,7 @@ static int virtnet_xdp_set(struct
net_device
*dev, struct bpf_prog *prog,
synchronize_net();
}
- err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
+ err = virtnet_set_queues(vi, curr_qp + xdp_qp);
if (err)
goto err;
netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp); @@ -
4852,7
+4835,7 @@ static int virtnet_probe(struct virtio_device *vdev)
virtio_device_ready(vdev);
- _virtnet_set_queues(vi, vi->curr_queue_pairs);
+ virtnet_set_queues(vi, vi->curr_queue_pairs);
/* a random MAC address has been assigned, notify the device.
* We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not
there
