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

[Breno Leitao <leitao@debian.org>]

On Mon, Mar 31, 2025 at 08:29:53AM -0700, Stanislav Fomichev wrote:
> On 03/31, Breno Leitao wrote:
> > 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@mail.gmail.com/
> > > Fixes: 004b5008016a ("eth: bnxt: remove most dependencies on RTNL")
> > > Reported-by: Taehee Yoo <ap420073@gmail.com>
> > > Signed-off-by: Stanislav Fomichev <sdf@fomichev.me>
> > 
> > Tested-by: Breno Leitao <leitao@debian.org>
> > 
> > > ---
> > >  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);
> > 			...
> > 		  }
> 
> - netif_close adds dev->close_list to the list (if it was up)

Right, but that list has only one net_device entry, right?

netif_close() instanciates a single list and merges it into `dev->close_list`

> - __dev_close_many walks over that list, so your new assert should
>   trigger as well

Why? Isn't the list only contain the dev that is already protected by
netdev_lock()?

> But also in general, it would be nice to keep existing
> rtnl+instance_lock scheme for now (except were we want to explicitly opt
> out, as in queue apis); we can follow up later to un-rtnl the rest.

I am just wondering if the code as-is is already safe from a locking
perspecting, and just the warning (ASSERT_RTNL) is wrong.