Re: [PATCH 4/5] ath10k: detect conf_mutex held ath10k_drain_tx() calls

[Shuah Khan]

On 2/11/21 4:20 AM, Kalle Valo wrote:
> Shuah Khan <skhan@xxxxxxxxxxxxxxxxxxx> writes:
>
> > On 2/10/21 1:25 AM, Kalle Valo wrote:
> > > Shuah Khan <skhan@xxxxxxxxxxxxxxxxxxx> writes:
> > >
> > > > ath10k_drain_tx() must not be called with conf_mutex held as workers can
> > > > use that also. Add check to detect conf_mutex held calls.
> > > >
> > > > Signed-off-by: Shuah Khan <skhan@xxxxxxxxxxxxxxxxxxx>
> > >
> > > The commit log does not answer to "Why?". How did you find this? What
> > > actual problem are you trying to solve?
> >
> > I came across the comment block above the ath10k_drain_tx() as I was
> > reviewing at conf_mutex holds while I was debugging the conf_mutex
> > lock assert in ath10k_debug_fw_stats_request().
> >
> > My reasoning is that having this will help detect incorrect usages
> > of ath10k_drain_tx() while holding conf_mutex which could lead to
> > locking problems when async worker routines try to call this routine.
>
> Ok, makes sense. I prefer having this background info in the commit log,
> for example "found by code review" or something like that. Or just copy
> what you wrote above :)

Thanks. I will do that.

> > > > --- a/drivers/net/wireless/ath/ath10k/mac.c
> > > > +++ b/drivers/net/wireless/ath/ath10k/mac.c
> > > > @@ -4566,6 +4566,7 @@ static void
> > > > ath10k_mac_op_wake_tx_queue(struct ieee80211_hw *hw,
> > > >    /* Must not be called with conf_mutex held as workers can use that also. */
> > > >    void ath10k_drain_tx(struct ath10k *ar)
> > > >    {
> > > > +	WARN_ON(lockdep_is_held(&ar->conf_mutex));
> > >
> > > Empty line after WARN_ON().
> >
> > Will do.
> >
> > > Shouldn't this check debug_locks similarly lockdep_assert_held() does?
> > >
> > > #define lockdep_assert_held(l)	do {				\
> > > 		WARN_ON(debug_locks && !lockdep_is_held(l));	\
> > > 	} while (0)
> > >
> > > And I suspect you need #ifdef CONFIG_LOCKDEP which should fix the kbuild
> > > bot error.
> >
> > Yes.
> >
> > > But honestly I would prefer to have lockdep_assert_not_held() in
> > > include/linux/lockdep.h, much cleaner that way. Also
> > > i915_gem_object_lookup_rcu() could then use the same macro.
> >
> > Right. This is the right way to go. That was first instinct and
> > decided to have the discussion evolve in that direction. Now that
> > it has, I will combine this change with
> > include/linux/lockdep.h and add lockdep_assert_not_held()
> >
> > I think we might have other places in the kernel that could use
> > lockdep_assert_not_held() in addition to i915_gem_object_lookup_rcu()

I looked at i915_gem_object_lookup_rcu(). The following can be replaced
by lockdep_assert_held().

#ifdef CONFIG_LOCKDEP
        WARN_ON(debug_locks && !lock_is_held(&rcu_lock_map));
#endif

> Great, thank you. The only problem is that lockdep.h changes have to go
> via some other tree, I just don't know which :) I think it would be
> easiest if also the ath10k patch goes via that other tree, I can ack the
> ath10k changes.
>
> Another option is that I'll apply the ath10k patch after the lockdep.h
> change has trickled down to my tree, but that usually happens only after
> the merge window and means weeks of waiting. Either is fine for me.

I will send the include/linux/lockdep.h and ath10k patch together and
we will take it from there.

thanks,
-- Shuah