Re: [patch V4 15/36] rseq: Record interrupt from user space

[Thomas Gleixner]

On Tue, Sep 09 2025 at 09:53, Mathieu Desnoyers wrote:
>>   static __always_inline void rseq_exit_to_user_mode(void)
>>   {
>> -	if (IS_ENABLED(CONFIG_DEBUG_RSEQ)) {
>> -		if (WARN_ON_ONCE(current->rseq.event.has_rseq &&
>> -				 current->rseq.event.events))
>> -			current->rseq.event.events = 0;
>> -	}
>> +	struct rseq_event *ev = &current->rseq.event;
>> +
>> +	if (IS_ENABLED(CONFIG_DEBUG_RSEQ))
>> +		WARN_ON_ONCE(ev->sched_switch);
>
> OK. Now I'm confused.
>
> It is perfectly legal to issue a system call from userspace as long
> as it's not from within an rseq critical section.
>
> That system call can be scheduled out, and can set the ev->sched_switch.
>
> This would cause the rseq_exit_to_user_mode from system call to
> hit this.
>
> What is disallowed is only issuing a system call from a rseq critical
> section. The other parts of rseq (updates of cpu id and mm cid) still
> have to happen when returning from a system call.
>
> What am I missing ?

The fact that any setting of ev->sched_switch has to be handled on the
way out independent of user interrupt or not as MM CID can change
obviously.

This is not any different from the state before this patch. Just that it
now only looks at sched_switch instead of the full event as that might
contain a set user_irq bit w/o sched_switch being set, no?

Thanks,

        tglx