Re: [PATCH -mm] mm: Clear to access sub-page last when clearing huge page

[Huang\, Ying]

Hi, Andrew,

Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> writes:

> On Mon,  7 Aug 2017 15:21:31 +0800 "Huang, Ying" <ying.huang@xxxxxxxxx> wrote:
>
>> From: Huang Ying <ying.huang@xxxxxxxxx>
>> 
>> Huge page helps to reduce TLB miss rate, but it has higher cache
>> footprint, sometimes this may cause some issue.  For example, when
>> clearing huge page on x86_64 platform, the cache footprint is 2M.  But
>> on a Xeon E5 v3 2699 CPU, there are 18 cores, 36 threads, and only 45M
>> LLC (last level cache).  That is, in average, there are 2.5M LLC for
>> each core and 1.25M LLC for each thread.  If the cache pressure is
>> heavy when clearing the huge page, and we clear the huge page from the
>> begin to the end, it is possible that the begin of huge page is
>> evicted from the cache after we finishing clearing the end of the huge
>> page.  And it is possible for the application to access the begin of
>> the huge page after clearing the huge page.
>> 
>> To help the above situation, in this patch, when we clear a huge page,
>> the order to clear sub-pages is changed.  In quite some situation, we
>> can get the address that the application will access after we clear
>> the huge page, for example, in a page fault handler.  Instead of
>> clearing the huge page from begin to end, we will clear the sub-pages
>> farthest from the the sub-page to access firstly, and clear the
>> sub-page to access last.  This will make the sub-page to access most
>> cache-hot and sub-pages around it more cache-hot too.  If we cannot
>> know the address the application will access, the begin of the huge
>> page is assumed to be the the address the application will access.
>> 
>> With this patch, the throughput increases ~28.3% in vm-scalability
>> anon-w-seq test case with 72 processes on a 2 socket Xeon E5 v3 2699
>> system (36 cores, 72 threads).  The test case creates 72 processes,
>> each process mmap a big anonymous memory area and writes to it from
>> the begin to the end.  For each process, other processes could be seen
>> as other workload which generates heavy cache pressure.  At the same
>> time, the cache miss rate reduced from ~33.4% to ~31.7%, the
>> IPC (instruction per cycle) increased from 0.56 to 0.74, and the time
>> spent in user space is reduced ~7.9%
>> 
>> Thanks Andi Kleen to propose to use address to access to determine the
>> order of sub-pages to clear.
>> 
>> The hugetlbfs access address could be improved, will do that in
>> another patch.
>
> I agree with what others said, plus...
>
>> @@ -4374,9 +4374,31 @@ void clear_huge_page(struct page *page,
>>  	}
>>  
>>  	might_sleep();
>> -	for (i = 0; i < pages_per_huge_page; i++) {
>> +	VM_BUG_ON(clamp(addr_hint, addr, addr +
>> +			(pages_per_huge_page << PAGE_SHIFT)) != addr_hint);
>> +	n = (addr_hint - addr) / PAGE_SIZE;
>> +	if (2 * n <= pages_per_huge_page) {
>> +		base = 0;
>> +		l = n;
>> +		for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
>> +			cond_resched();
>> +			clear_user_highpage(page + i, addr + i * PAGE_SIZE);
>> +		}
>> +	} else {
>> +		base = 2 * n - pages_per_huge_page;
>> +		l = pages_per_huge_page - n;
>> +		for (i = 0; i < base; i++) {
>> +			cond_resched();
>> +			clear_user_highpage(page + i, addr + i * PAGE_SIZE);
>> +		}
>> +	}
>> +	for (i = 0; i < l; i++) {
>> +		cond_resched();
>> +		clear_user_highpage(page + base + i,
>> +				    addr + (base + i) * PAGE_SIZE);
>>  		cond_resched();
>> -		clear_user_highpage(page + i, addr + i * PAGE_SIZE);
>> +		clear_user_highpage(page + base + 2 * l - 1 - i,
>> +				    addr + (base + 2 * l - 1 - i) * PAGE_SIZE);
>
> Please document this design with a carefully written code comment.
> For example, why was "2 * n" chosen?  What is it trying to achieve?

Sure.

"2 * n" here is to determine whether addr_hint is in the first half (2 *
n <= pages_per_huge_page) or the second half (2 * n >
pages_per_huge_page) of the huge page.

> Also, the final clearing loop "for (i = 0; i < l; i++)" might cause
> eviction of data which was cached in the previous loop.  Perhaps some
> additional gains will be made by clearing the hugepage in a
> left-right-left-right "start from the ends and work inwards" manner, if
> you see what I mean.  So the 4k pages immediately surrounding addr_hint
> are the most-recently-cleared.  Although accesses to the data at lower
> addresses than addr_hint are probably somewhat rare (and may be
> nonexistent in your synthetic test case).

Yes.  I think I have done exactly this in the patch.  For each iteration
of the loop, two sub-pages will be cleared: base + i, and base + 2 * l -
1 - i, that is, the left and right of the fault sub-page, and finally
reach the fault sub-page as the last sub-page to clear.

Best Regards,
Huang, Ying