-----Original Message----- From: Dexuan Cui decui@microsoft.com Sent: Tuesday, January 30, 2024 3:13 PM To: Shradha Gupta shradhagupta@linux.microsoft.com; KY Srinivasan
@@ -2205,8 +2209,11 @@ static int netvsc_vf_join(struct net_device *vf_netdev, ndev->name, ret); goto upper_link_failed; }
- schedule_delayed_work(&ndev_ctx->vf_takeover,
VF_TAKEOVER_INT);
- /* If this registration is called from probe context vf_takeover
* is taken care of later in probe itself.I suspect "later in probe itself" is not accurate. If 'context' is VF_REG_IN_PROBE, I suppose what happens here is: after netvsc_probe() finishes, the netvsc interface becomes available, so the user space will ifup it, and netvsc_open() will UP the VF interface, and netvsc_netdev_event() is called for the VF with event == NETDEV_POST_INIT (?) and NETDEV_CHANGE, and the data path is switched to the VF.
In register_netdevice(), NETDEV_POST_INIT is earlier than NETDEV_REGISTER. This case: netvsc_open >> dev_open(vf) >> NETDEV_UP >> netvsc_vf_changed(event_dev, event);
If my understanding is correct, I think in the case of 'context' == VF_REG_IN_PROBE, I suspect the "Align MTU of VF with master" and the "sync address list from ndev to VF" in __netvsc_vf_setup() are omitted? If so, should this be fixed? e.g. Not sure if the below is an issue or not:
- a VF is bound to a NetVSC interface, and a user sets the MTUs to 1024.
- rmmod hv_netvsc
- modprobe hv_netvsc
- the netvsc interface uses MTU=1500 (the default), and the VF still
uses 1024.
__netvsc_vf_setup() is skipped from the netvsc_register_vf >> netvsc_vf_join(), but called from netvsc_probe(), so the VF mtu is sync-ed to 1500. I verified mtu sync in test.
Thanks, - Haiyang