Re: [PATCH net] bnxt_en: bring back rtnl lock in bnxt_shutdown

[Breno Leitao]

Hello Stanislav,

On Fri, Mar 28, 2025 at 10:42:16AM -0700, Stanislav Fomichev wrote:
> Taehee reports missing rtnl from bnxt_shutdown path:
> 
> inetdev_event (./include/linux/inetdevice.h:256 net/ipv4/devinet.c:1585)
> notifier_call_chain (kernel/notifier.c:85)
> __dev_close_many (net/core/dev.c:1732 (discriminator 3))
> kernel/locking/mutex.c:713 kernel/locking/mutex.c:732)
> dev_close_many (net/core/dev.c:1786)
> netif_close (./include/linux/list.h:124 ./include/linux/list.h:215
> bnxt_shutdown (drivers/net/ethernet/broadcom/bnxt/bnxt.c:16707) bnxt_en
> pci_device_shutdown (drivers/pci/pci-driver.c:511)
> device_shutdown (drivers/base/core.c:4820)
> kernel_restart (kernel/reboot.c:271 kernel/reboot.c:285)

I've got this issue as well.

> 
> Bring back the rtnl lock.
> 
> Link: https://lore.kernel.org/netdev/CAMArcTV4P8PFsc6O2tSgzRno050DzafgqkLA2b7t=Fv_SY=brw@xxxxxxxxxxxxxx/
> Fixes: 004b5008016a ("eth: bnxt: remove most dependencies on RTNL")
> Reported-by: Taehee Yoo <ap420073@xxxxxxxxx>
> Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxxx>

Tested-by: Breno Leitao <leitao@xxxxxxxxxx>

> ---
>  drivers/net/ethernet/broadcom/bnxt/bnxt.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
> index 934ba9425857..1a70605fad38 100644
> --- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c
> +++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
> @@ -16698,6 +16698,7 @@ static void bnxt_shutdown(struct pci_dev *pdev)
>  	if (!dev)
>  		return;
>  
> +	rtnl_lock();
>  	netdev_lock(dev);

can't we leverage the `struct net_device->lock` for the shutdown.
Basically we have the lock the single device we are turning it down.

I am wondering if we really need the big RTNL lock. This is my
understanding of what is happening:

pci_device_shutdown() is called for a single device
 - netdev_lock(dev)
 - netif_close(dev);
    - dev_close_many(&single, true);
      - __dev_close_many()
        - ASSERT_RTNL();

Basically we ware only closing one device, and the net_device->lock
is already held. Shouldn't it be enough?

Can we do something like this (from my naive point of view):

	 static void __dev_close_many(struct list_head *head)
	  {
		  struct net_device *dev;

	-         ASSERT_RTNL();
		  might_sleep();

		  list_for_each_entry(dev, head, close_list) {
	+	  	ASSERT_RTNL_NET(dev);
			...
		  }

Thanks
--breno