Re: "BUG: Invalid wait context" in invalidate_batched_entropy

[Sebastian Andrzej Siewior]

On 2022-01-28 17:04:13 [+0100], Jason A. Donenfeld wrote:
> Hi Sebastian/Jonathan,
> 
> On Fri, Jan 28, 2022 at 9:35 AM Sebastian Andrzej Siewior
> <bigeasy@xxxxxxxxxxxxx> wrote:
> > This report is due to CONFIG_PROVE_LOCKING=y _and_
> > CONFIG_PROVE_RAW_LOCK_NESTING=y. It reports a nesting problem
> > (raw_spinlock_t -> spinlock_t lock ordering) which becomes a real
> > problem on PREEMPT_RT.
> 
> Hmm, I'm still having a tough time reproducing this. I'm trying to
> understand your intuition. Is the problem you see that something else
> in the IRQ path uses a raw_spinlock_t, and then with that lock still
> held, we call invalidate_batched_entropy(), which takes an ordinary
> spinlock_t, non-raw? And taking a spinlock-t while holding a
> raw_spinlock_t is illegal?

Correct. You must not acquire a spinlock_t while holding a
raw_spinlock_t. This is because on PREEMPT_RT the spinlock_t is a
sleeping lock while raw_spinlock_t disables preemption/ interrupts and
sleeping is not possible.
	Documentation/locking/locktypes.rst

On non-PREEMPT both lock types (spinlock_t & raw_spinlock_t) behave in
the same way but lockdep can tell them apart with
CONFIG_PROVE_RAW_LOCK_NESTING=y and show code that is problematic on RT.

> Jason

Sebastian