Re: [RFC PATCH v1 1/3] locking/lockdep: Fix false warning of check_wait_context()

From: Boqun Feng
Date: Mon Jul 12 2021 - 05:09:16 EST


On Mon, Jul 12, 2021 at 04:18:36PM +0800, Xiongwei Song wrote:
> On Mon, Jul 12, 2021 at 12:43 AM Waiman Long <llong@xxxxxxxxxx> wrote:
> >
> > On 7/11/21 10:14 AM, Xiongwei Song wrote:
> > > From: Xiongwei Song <sxwjean@xxxxxxxxx>
> > >
> > > We now always get a "Invalid wait context" warning with
> > > CONFIG_PROVE_RAW_LOCK_NESTING=y, see the full warning below:
> > >
> > > [ 0.705900] =============================
> > > [ 0.706002] [ BUG: Invalid wait context ]
> > > [ 0.706180] 5.13.0+ #4 Not tainted
> > > [ 0.706349] -----------------------------
> >
> > I believe the purpose of CONFIG_PROVE_RAW_LOCK_NESTING is experimental
> > and it is turned off by default. Turning it on can cause problem as
> > shown in your lockdep splat. Limiting it to just PREEMPT_RT will defeat
> > its purpose to find potential spinlock nesting problem in non-PREEMPT_RT
> > kernel.
> As far as I know, a spinlock can nest another spinlock. In
> non-PREEMPT_RT kernel
> spin_lock and raw_spin_lock are same , so here acquiring a spin_lock in hardirq
> context is acceptable, the warning is not needed. My knowledge on this
> is not enough,
> Will dig into this.
>

You may find this useful: https://lwn.net/Articles/146861/ ;-)

The thing is that most of the irq handlers will run in process contexts
in PREEMPT_RT kernel (threaded irq), while the rest continues to run in
hardirq contexts. spinlock_t is allowed int threaded irqs but not in
hardirq contexts for PREEMPT_RT, because spinlock_t will become
sleeplable locks.

Regards,
Boqun

> > The point is to fix the issue found,
> Agree. I thought there was a spinlock usage issue, but by checking
> deactivate_slab context,
> looks like the spinlock usage is well. Maybe I'm missing something?
>
> > not hiding it from appearing.
> I'm not trying to hiding it, according to the code context, the fix is
> reasonable from my point of
> view. Let me check again.
>
> Thank you for the comments.
>
> Regards,
> Xiongwei
> >
> > Cheers,
> > Longman
> >