Re: [syzbot] [wireless?] BUG: sleeping function called from invalid context in wext_netdev_notifier_call

[Kuniyuki Iwashima <kuniyu@amazon.com>]

From: Johannes Berg <johannes@sipsolutions.net>
Date: Fri, 13 Dec 2024 09:39:04 +0100
> On Fri, 2024-12-13 at 17:36 +0900, Kuniyuki Iwashima wrote:
> > From: Johannes Berg <johannes@sipsolutions.net>
> > Date: Thu, 12 Dec 2024 09:52:44 +0100
> > > On Wed, 2024-12-11 at 20:41 -0800, syzbot wrote:
> > > > CPU: 1 UID: 0 PID: 12894 Comm: kworker/u8:18 Not tainted 6.13.0-rc1-syzkaller-00183-g4c49f38e20a5 #0
> > > > Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 09/13/2024
> > > > Workqueue: bond0 bond_mii_monitor
> > > > Call Trace:
> > > >  <TASK>
> > > >  __dump_stack lib/dump_stack.c:94 [inline]
> > > >  dump_stack_lvl+0x241/0x360 lib/dump_stack.c:120
> > > >  __might_resched+0x5d4/0x780 kernel/sched/core.c:8758
> > > >  down_read+0x8e/0xa40 kernel/locking/rwsem.c:1523
> > > >  wireless_nlevent_flush net/wireless/wext-core.c:351 [inline]
> > > >  wext_netdev_notifier_call+0x1f/0x120 net/wireless/wext-core.c:371
> > > >  notifier_call_chain+0x1a5/0x3f0 kernel/notifier.c:85
> > > >  netdev_state_change+0x11f/0x1a0 net/core/dev.c:1378
> > > >  linkwatch_do_dev+0x112/0x170 net/core/link_watch.c:182
> > > >  ethtool_op_get_link+0x15/0x60 net/ethtool/ioctl.c:62
> > > >  bond_check_dev_link+0x1f1/0x3f0 drivers/net/bonding/bond_main.c:873
> > > >  bond_miimon_inspect drivers/net/bonding/bond_main.c:2740 [inline]
> > > >  bond_mii_monitor+0x49a/0x3170 drivers/net/bonding/bond_main.c:2962
> > > 
> > > Yeah this isn't new. I thought we were going to just squash this with
> > > 
> > > https://lore.kernel.org/netdev/2730097.1721581672@famine/
> > > 
> > > Whatever came of that?
> > 
> > Now I remember I forgot to respin this patch.
> > https://lore.kernel.org/linux-wireless/20241014205543.94787-4-kuniyu@amazon.com/
> > 
> > If the issue above is still not fixed, I will respin it on
> > wireless.git with Fixes tag.
> 
> Wait, that's not related at all? The bonding issue still is that it
> calls ethtool ops under RCU but ethtool ops can sleep. That issue has
> nothing to do with the wext nlevent namespace-ification?

Ah okay, I thought removing the mutex there was another option
to silence it, but I didn't check other places.

> 
> 
> We can still do the namespace thing for wext nlevent, but it's not going
> to fix this issue, and I don't think we need to on wireless (rather than
> wireless-next)

I see, will post to -next.

Thanks!