Re: [bug report] irq_work: Allow irq_work_sync() to sleep if irq_work() no IRQ support.

[Sebastian Andrzej Siewior <bigeasy@linutronix.de>]

+peterz, tglx

On 2022-01-19 16:37:29 [+0300], Dan Carpenter wrote:
> Hello Sebastian Andrzej Siewior,
> 
> The patch 810979682ccc: "irq_work: Allow irq_work_sync() to sleep if
> irq_work() no IRQ support." from Oct 6, 2021, leads to the following
> Smatch static checker warning:
> 
> 	kernel/events/core.c:4955 _free_event()
> 	warn: sleeping in atomic context
> 
> There are two potential call trees where preemption is disabled.
> 
> perf_remove_from_context() <- disables preempt
 raw_spin_lock_irq(&ctx->lock)

> __perf_event_exit_context() <- disables preempt
don't see additional one.

> -> __perf_remove_from_context()
>    -> perf_group_detach()
>       -> perf_put_aux_event()
>          -> put_event()
>             -> _free_event()
> 
> _free_event() calls irq_work_sync() which can now sleep.

That is true. There is also 
|	if (event->rb) 
|		mutex_lock(&event->mmap_mutex);

Can we somehow delay the whole thing? There are also kfree() and
spinlock_t usage which are not very RT compatible with disabled
preemption.

> regards,
> dan carpenter

Sebastian