[RFC PATCH] mm: avoid access flag update TLB flush for retried page fault

[Yang Shi]

On 7/8/20 1:00 AM, Will Deacon wrote:
> On Wed, Jul 08, 2020 at 02:54:32AM +0800, Yang Shi wrote:
>> Recently we found regression when running will_it_scale/page_fault3 test
>> on ARM64.  Over 70% down for the multi processes cases and over 20% down
>> for the multi threads cases.  It turns out the regression is caused by commit
>> 89b15332af7c0312a41e50846819ca6613b58b4c ("mm: drop mmap_sem before
>> calling balance_dirty_pages() in write fault").
>>
>> The test mmaps a memory size file then write to the mapping, this would
>> make all memory dirty and trigger dirty pages throttle, that upstream
>> commit would release mmap_sem then retry the page fault.  The retried
>> page fault would see correct PTEs installed by the first try then update
>> access flags and flush TLBs.  The regression is caused by the excessive
>> TLB flush.  It is fine on x86 since x86 doesn't need flush TLB for
>> access flag update.
>>
>> The page fault would be retried due to:
>> 1. Waiting for page readahead
>> 2. Waiting for page swapped in
>> 3. Waiting for dirty pages throttling
>>
>> The first two cases don't have PTEs set up at all, so the retried page
>> fault would install the PTEs, so they don't reach there.  But the #3
>> case usually has PTEs installed, the retried page fault would reach the
>> access flag update.  But it seems not necessary to update access flags
>> for #3 since retried page fault is not real "second access", so it
>> sounds safe to skip access flag update for retried page fault.
>>
>> With this fix the test result get back to normal.
>>
>> Reported-by: Xu Yu <xuyu at linux.alibaba.com>
>> Debugged-by: Xu Yu <xuyu at linux.alibaba.com>
>> Tested-by: Xu Yu <xuyu at linux.alibaba.com>
>> Signed-off-by: Yang Shi <yang.shi at linux.alibaba.com>
>> ---
>> I'm not sure if this is safe for non-x86 machines, we did some tests on arm64, but
>> there may be still corner cases not covered.
>>
>>   mm/memory.c | 7 ++++++-
>>   1 file changed, 6 insertions(+), 1 deletion(-)
>>
>> diff --git a/mm/memory.c b/mm/memory.c
>> index 87ec87c..3d4e671 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -4241,8 +4241,13 @@ static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
>>   	if (vmf->flags & FAULT_FLAG_WRITE) {
>>   		if (!pte_write(entry))
>>   			return do_wp_page(vmf);
>> -		entry = pte_mkdirty(entry);
>>   	}
>> +
>> +	if ((vmf->flags & FAULT_FLAG_WRITE) && !(vmf->flags & FAULT_FLAG_TRIED))
>> +		entry = pte_mkdirty(entry);
>> +	else if (vmf->flags & FAULT_FLAG_TRIED)
>> +		goto unlock;
>> +
> Can you rewrite this as:
>
> 	if (vmf->flags & FAULT_FLAG_TRIED)
> 		goto unlock;
>
> 	if (vmf->flags & FAULT_FLAG_WRITE)
> 		entry = pte_mkdirty(entry);

Yes, it does the same.

>
> ? (I'm half-asleep this morning and there are people screaming and shouting
> outside my window, so this might be rubbish)
>
> If you _can_make that change, then I don't understand why the existing
> pte_mkdirty() line needs to move at all. Couldn't you just add:
>
> 	if (vmf->flags & FAULT_FLAG_TRIED)
> 		goto unlock;
>
> after the existing "vmf->flags & FAULT_FLAG_WRITE" block?

The intention is to not set dirty bit if it is in retried page fault 
since the bit should be already set in the first try. And, I'm not quite 
sure if TLB needs to be flushed on non-x86 if dirty bit is set. If it is 
unnecessary, then the above change does make sense.

>
> Will