From: Johannes Berg <johannes@sipsolutions.net> Date: 2021-05-14 08:26:43
On Fri, 2021-05-14 at 01:07 +0200, Maximilian Luz wrote:
Following commit a05829a7222e ("cfg80211: avoid holding the RTNL when
calling the driver"), the mwifiex_pcie module fails to unload. This also
prevents the device from rebooting / shutting down.
Attempting to unload the module
[ 245.504764] Tainted: G C OE 5.11.0-1-surface-dev #2
so I have no idea what kernel you're using, because 5.11 did *not*
contain commit a05829a7222e ("cfg80211: avoid holding the RTNL when
calling the driver"). If you backported the bug you get to be
responsible for backporting the fixes too?
If that's all not solving the issue then please try to resolve with gdb
what line of code "cfg80211_netdev_notifier_call+0x12a" is, and please
also clarify exactly what (upstream!) kernel you're using.
johannes
From: Maximilian Luz <luzmaximilian@gmail.com> Date: 2021-05-14 11:41:03
On 5/14/21 10:26 AM, Johannes Berg wrote:
On Fri, 2021-05-14 at 01:07 +0200, Maximilian Luz wrote:
quoted
Following commit a05829a7222e ("cfg80211: avoid holding the RTNL when
calling the driver"), the mwifiex_pcie module fails to unload. This also
prevents the device from rebooting / shutting down.
Attempting to unload the module
I'm *guessing* that you're attempting to unload the module while the
interface is still up, i.e. you didn't "ip link set wlan0 down" first?
I did not. Doing so indeed allows unloading of the module.
Thanks for pointing this out. That's backported to 5.12.2, right?
Unfortunately the error still persists there (as on all other post 5.12
kernels, as mentioned below).
However, your log says:
quoted
[ 245.504764] Tainted: G C OE 5.11.0-1-surface-dev #2
so I have no idea what kernel you're using, because 5.11 did *not*
contain commit a05829a7222e ("cfg80211: avoid holding the RTNL when
calling the driver"). If you backported the bug you get to be
responsible for backporting the fixes too?
This is a log from a05829a7222e checked out, which was pre 5.12-rc1, so the
last tag is still 5.11 which sets the package version in my packaging scripts
to that as well. It's not an actual 5.11 kernel, sorry for the confusion.
There is no backporting going on, just bisecting.
If that's all not solving the issue then please try to resolve with gdb
what line of code "cfg80211_netdev_notifier_call+0x12a" is, and please
also clarify exactly what (upstream!) kernel you're using.
The issue exists on 5.12 through 5.12.4, as well as 5.13-rc1 (which is
why I didn't bother specifying a version, sorry again).
If you really want me to, I'll try to find some time to learn GDB kernel
debugging (never done that before, so might take a bit) however, I think
it's fairly clear what's going wrong and why the fix you've linked below
doesn't apply in this case:
Your fix is fixing
cfg80211_destroy_iface_wk() takes wiphy_lock
-> cfg80211_destroy_ifaces()
-> ieee80211_del_iface
-> ieeee80211_if_remove
-> cfg80211_unregister_wdev
-> unregister_netdevice_queue
-> dev_close_many
-> __dev_close_many
-> raw_notifier_call_chain
-> cfg80211_netdev_notifier_call
by addressing this in cfg80211_destroy_iface{s,_wk}(). The trace from my
log shows
mwifiex_uninit_sw() takes wiphy_lock
-> mwifiex_del_virtual_intf
-> cfg80211_unregister_netdevice()
-> cfg80211_unregister_wdev()
-> _cfg80211_unregister_wdev() has lockdep_assert_held(&rdev->wiphy.mtx)
-> unregister_netdevice_queue
-> dev_close_many
-> __dev_close_many
-> raw_notifier_call_chain
-> cfg80211_netdev_notifier_call attempts to take wiphy_lock again
So your fix does not address this particular issue. It doesn't even
touch any of the affected code path. I believe it is instead fixing one
symptom of the same underlying problem.
While the last parts of the trace are the same (specifically following
cfg80211_unregister_wdev()), the lock is initially taken in different
functions. Your fix addresses this by changing cfg80211_destroy_ifaces(),
and cfg80211_destroy_iface_wk() which, however, were never called on the
path that's causing _this_ issue.
Furthermore, if you go through that trace, there's only one notifier
call in __dev_close_many(), which is
call_netdevice_notifiers(NETDEV_GOING_DOWN, dev)
which I believe I have linked in my previous mail. Since the state value
is NETDEV_GOING_DOWN, this has to be case [3] (unless I'm missing
something).
Regards,
Max
[3]: https://elixir.bootlin.com/linux/v5.13-rc1/source/net/wireless/core.c#L1428
From: Maximilian Luz <luzmaximilian@gmail.com> Date: 2021-05-14 13:46:18
On 5/14/21 1:07 AM, Maximilian Luz wrote:
Following commit a05829a7222e ("cfg80211: avoid holding the RTNL when
calling the driver"), the mwifiex_pcie module fails to unload. This also
prevents the device from rebooting / shutting down.
Attempting to unload the module produces the log pasted below. Upon
further investigation, this looks like a deadlock inside
_cfg80211_unregister_wdev():
- According to [1], this function expects the rdev->wiphy.mtx to be
held.
- Down the line, this function (through some indirections, see third
trace in log below) calls call_netdevice_notifiers(NETDEV_GOING_DOWN,
...) [2].
- One of the registered notifiers seems to be
cfg80211_netdev_notifier_call(), which attempts to lock
rdev->wiphy.mtx again [3], completing the deadlock.
From: Brian Norris <briannorris@chromium.org> Date: 2021-05-15 02:44:47
On Fri, May 14, 2021 at 10:26:25AM +0200, Johannes Berg wrote:
If that's all not solving the issue then please try to resolve with gdb
what line of code "cfg80211_netdev_notifier_call+0x12a" is, and please
also clarify exactly what (upstream!) kernel you're using.
I can reproduce on v5.12 and v5.12.4 as well. With v5.12.4, I'm at:
net/wireless/core.c:
1428 wiphy_lock(&rdev->wiphy);
include/net/cfg80211.h:
5269 mutex_lock(&wiphy->mtx);
i.e.,
static int cfg80211_netdev_notifier_call(struct notifier_block *nb,
unsigned long state, void *ptr)
{
...
case NETDEV_GOING_DOWN:
wiphy_lock(&rdev->wiphy); <--- right here
cfg80211_leave(rdev, wdev);
wiphy_unlock(&rdev->wiphy);
...
It would seem like _anyone_ that calls cfg80211_unregister_wdev() with
an interface up will hit this -- not unique to mwifiex. In fact, apart
from the fact that all his line numbers are wrong, Maximilian's original
email points out exactly where the deadlock is.
cfg80211_unregister_wdev() holds the wiphy lock, and the GOING_DOWN
notification also tries to grab it.
It does happen that in many other paths, you've already ensured that you
bring the interface down, so e.g., mac80211 drivers don't tend to hit
this. But I wouldn't be surprised if a few other cfg80211 drivers hit
this too.
The best solution I could figure was to do a similar lock dance done in
nl80211_del_interface() -- close the netdev without holding the wiphy
lock. I'll send out a patch shortly.
Brian
From: Maximilian Luz <luzmaximilian@gmail.com> Date: 2021-05-15 11:25:12
On 5/15/21 4:44 AM, Brian Norris wrote:
It would seem like _anyone_ that calls cfg80211_unregister_wdev() with
an interface up will hit this -- not unique to mwifiex. In fact, apart
from the fact that all his line numbers are wrong, Maximilian's original
email points out exactly where the deadlock is.
cfg80211_unregister_wdev() holds the wiphy lock, and the GOING_DOWN
notification also tries to grab it.
It does happen that in many other paths, you've already ensured that you
bring the interface down, so e.g., mac80211 drivers don't tend to hit
this. But I wouldn't be surprised if a few other cfg80211 drivers hit
this too.
The best solution I could figure was to do a similar lock dance done in
nl80211_del_interface() -- close the netdev without holding the wiphy
lock. I'll send out a patch shortly.
I believe that if we're going to fix that in the individual drivers,
there should be at least some sort of warning/documentation on
cfg80211_unregister_wdev().
Also someone might want to look at other WiFi drivers calling
cfg80211_unregister_wdev(). For example, I can see a locked call in the
brcm80211 driver, but no previous dev_close() call (see [1]). Haven't
looked in detail though, so I might just be wrong.
I can't help but think that this should maybe be addressed in that
common part instead. I know too little of that subsystem to tell if that
might be infeasible though.
Regards,
Max
[1]: https://elixir.bootlin.com/linux/v5.13-rc1/source/drivers/net/wireless/broadcom/brcm80211/brcmfmac/p2p.c#L2445