On Thu, 30 Jun 2022 08:11:34 -0700 Jakub Kicinski wrote:

On Thu, 30 Jun 2022 12:08:39 +0200 Piotr Skajewski wrote:

quoted

On Tue, 28 Jun 2022 10:27:07 -0700 Jakub Kicinski wrote:

quoted

On Tue, 28 Jun 2022 09:53:46 -0700 Tony Nguyen wrote:  

quoted

+	spin_lock_irqsave(&adapter->vfs_lock, flags);
+
 	/* set num VFs to 0 to prevent access to vfinfo */
 	adapter->num_vfs = 0;
 

@@ -228,6 +231,8 @@ int ixgbe_disable_sriov(struct ixgbe_adapter *adapter)
 	kfree(adapter->mv_list);
 	adapter->mv_list = NULL;
 
+	spin_unlock_irqrestore(&adapter->vfs_lock, flags);  

There's a pci_dev_put() in there, are you sure it won't sleep?  

Thank Jakub for your notice, during development we were aware about this
and tests we've made on this particular case, did not report any problems
that could be related to might_sleep in conjunction with spinlock.

To be on the safe side how about we protect adapter->num_vfs instead 
of adapter->vfinfo ?

You can hold the lock just around setting adapter->num_vfs to zero,
and then inside ixgbe_msg_task() you don't have to add the new if()
because the loop bound is already adapter->num_vfs.

Smaller change, and safer.

Yes sounds good, I will test it and prepare the patch.