Re: [RFC V2] mm: Generalize notify_page_fault()

[Anshuman Khandual]

On 06/05/2019 04:49 PM, Michael Ellerman wrote:
> Anshuman Khandual <anshuman.khandual@xxxxxxx> writes:
>> Similar notify_page_fault() definitions are being used by architectures
>> duplicating much of the same code. This attempts to unify them into a
>> single implementation, generalize it and then move it to a common place.
>> kprobes_built_in() can detect CONFIG_KPROBES, hence notify_page_fault()
>> need not be wrapped again within CONFIG_KPROBES. Trap number argument can
>> now contain upto an 'unsigned int' accommodating all possible platforms.
> ...
>> diff --git a/arch/arm/mm/fault.c b/arch/arm/mm/fault.c
>> index 58f69fa..1bc3b18 100644
>> --- a/arch/arm/mm/fault.c
>> +++ b/arch/arm/mm/fault.c
>> @@ -30,28 +30,6 @@
>>  
>>  #ifdef CONFIG_MMU
>>  
>> -#ifdef CONFIG_KPROBES
>> -static inline int notify_page_fault(struct pt_regs *regs, unsigned int fsr)
>> -{
>> -	int ret = 0;
>> -
>> -	if (!user_mode(regs)) {
>> -		/* kprobe_running() needs smp_processor_id() */
>> -		preempt_disable();
>> -		if (kprobe_running() && kprobe_fault_handler(regs, fsr))
>> -			ret = 1;
>> -		preempt_enable();
>> -	}
>> -
>> -	return ret;
>> -}
>> -#else
> 
> You've changed several of the architectures from something like above,
> where it disables preemption around the call into the below:
> 
>> +int __kprobes notify_page_fault(struct pt_regs *regs, unsigned int trap)
>> +{
>> +	int ret = 0;
>> +
>> +	/*
>> +	 * To be potentially processing a kprobe fault and to be allowed
>> +	 * to call kprobe_running(), we have to be non-preemptible.
>> +	 */
>> +	if (kprobes_built_in() && !preemptible() && !user_mode(regs)) {
>> +		if (kprobe_running() && kprobe_fault_handler(regs, trap))
>> +			ret = 1;
>> +	}
>> +	return ret;
>> +}
> 
> Which skips everything if we're preemptible. Is that an equivalent

Right.

> change? If so can you please explain why in more detail.

It is probably not an equivalent change. The following explanation is extracted from
RFC V1 discussions (https://patchwork.kernel.org/patch/10968273/). Will explain the
rational for this behavior change in the commit message next time around.

----------------------------
a980c0ef9f6d ("x86/kprobes: Refactor kprobes_fault() like kprobe_exceptions_notify()")
b506a9d08bae ("x86: code clarification patch to Kprobes arch code")

In particular the later one (b506a9d08bae). It explains how the invoking context
in itself should be non-preemptible for the kprobes processing context irrespective
of whether kprobe_running() or perhaps smp_processor_id() is safe or not. Hence it
does not make much sense to continue when original invoking context is preemptible.
Instead just bail out earlier. This seems to be making more sense than preempt
disable-enable pair. If there are no concerns about this change from other platforms,
I will change the preemption behavior in proposed generic function next time around.
----------------------------

Do you see any concern changing preempt behavior in the x86 way ?

> 
> Also why not have it return bool?

Just that all architectures (except powerpc) had 'int' as return type. But we can
change that to 'bool'.