On Thu, May 29, 2025 at 09:45:10AM -0700, Stanislav Fomichev wrote:

On 05/29, Jakub Kicinski wrote:

quoted

On Thu, 29 May 2025 08:59:43 -0700 Stanislav Fomichev wrote:

quoted

So this is internal WQ entry lock that is being reordered with rtnl
lock. But looking at process_one_work, I don't see actual locks, mostly
lock_map_acquire/lock_map_release calls to enforce some internal WQ
invariants. Not sure what to do with it, will try to read more.

Basically a flush_work() happens while holding rtnl_lock,
but the work itself takes that lock. It's a driver bug.

e400c7444d84 ("e1000: Hold RTNL when e1000_down can be called") ?
I think similar things (but wrt netdev instance lock) are happening
with iavf: iavf_remove calls cancel_work_sync while holding the
instance lock and the work callbacks grab the instance lock as well :-/

I think this is probably the same thread as:

 https://lore.kernel.org/netdev/CAP=Rh=OEsn4y_2LvkO3UtDWurKcGPnZ_NPSXK=FbgygNXL37Sw@mail.gmail.com/ (local)

I posted a response there about how to possibly avoid the problem
(based on my rough reading of the driver code), but am still
thinking more on this.