Re: [PATCH v3 10/12] mm/gup: /proc/vmstat: pin_user_pages (FOLL_PIN) reporting

[John Hubbard]

On 2/3/20 1:34 PM, John Hubbard wrote:
> On 2/3/20 1:30 PM, Kirill A. Shutemov wrote:
>> On Mon, Feb 03, 2020 at 01:04:04PM -0800, John Hubbard wrote:
>>> On 2/3/20 5:53 AM, Kirill A. Shutemov wrote:
>>>> On Fri, Jan 31, 2020 at 07:40:27PM -0800, John Hubbard wrote:
>>>>> diff --git a/mm/gup.c b/mm/gup.c
>>>>> index c10d0d051c5b..9fe61d15fc0e 100644
>>>>> --- a/mm/gup.c
>>>>> +++ b/mm/gup.c
>>>>> @@ -29,6 +29,19 @@ struct follow_page_context {
>>>>>  	unsigned int page_mask;
>>>>>  };
>>>>>  
>>>>> +#ifdef CONFIG_DEBUG_VM
>>>>
>>>> Why under CONFIG_DEBUG_VM? There's nothing about this in the cover letter.
>>>>
>>>
>>> Early on, gup_benchmark showed a really significant slowdown from using these 
>>> counters. And I don't doubt that it's still the case.
>>>
>>> I'll re-measure and add a short summary and a few numbers to the patch commit
>>> description, and to the v4 cover letter.
>>
>> Looks like you'll show zeros for these counters if debug is off. It can be
>> confusing to the user. I think these counters should go away if you don't
>> count them.
>>
> 
> OK, that's a good point. (And in fact, the counters==0 situation already led me
> astray briefly while debugging with Leon R, even. heh.) I'll remove them entirely for
> the !CONFIG_DEBUG_VM case.
> 

On second thought, let me do some more careful performance testing. I don't recall
now if I was just removing every possible perf slowdown item, when I made this decision.
It could be that the perf is not affected, and I could just leave this feature enabled
at all times, which would be nicer.

And after all, these counters were designed for pretty hot-path items. I'll report back
with results...


thanks,
-- 
John Hubbard
NVIDIA