Re: [RFC] mm: Generalize notify_page_fault()

[Anshuman Khandual <anshuman.khandual@arm.com>]

On 05/30/2019 04:36 PM, Matthew Wilcox wrote:
> On Thu, May 30, 2019 at 11:25:13AM +0530, Anshuman Khandual wrote:
>> 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()
>> must not be wrapped again within CONFIG_KPROBES. Trap number argument can
> 
> This is a funny quirk of the English language.  "must not" means "is not
> allowed to be", not "does not have to be".

You are right. Noted for future. Thanks !

> 
>> @@ -141,6 +142,19 @@ static int __init init_zero_pfn(void)
>>  core_initcall(init_zero_pfn);
>>  
>>  
>> +int __kprobes notify_page_fault(struct pt_regs *regs, unsigned int trap)
>> +{
>> +	int ret = 0;
>> +
>> +	if (kprobes_built_in() && !user_mode(regs)) {
>> +		preempt_disable();
>> +		if (kprobe_running() && kprobe_fault_handler(regs, trap))
>> +			ret = 1;
>> +		preempt_enable();
>> +	}
>> +	return ret;
>> +}
>> +
>>  #if defined(SPLIT_RSS_COUNTING)
> 
> Comparing this to the canonical implementation (ie x86), it looks similar.
> 
> static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
> {
>         if (!kprobes_built_in())
>                 return 0;
>         if (user_mode(regs))
>                 return 0;
>         /*
>          * To be potentially processing a kprobe fault and to be allowed to call
>          * kprobe_running(), we have to be non-preemptible.
>          */
>         if (preemptible())
>                 return 0;
>         if (!kprobe_running())
>                 return 0;
>         return kprobe_fault_handler(regs, X86_TRAP_PF);
> }
> 
> The two handle preemption differently.  Why is x86 wrong and this one
> correct?

Here it expects context to be already non-preemptible where as the proposed
generic function makes it non-preemptible with a preempt_[disable|enable]()
pair for the required code section, irrespective of it's present state. Is
not this better ?