Re: [PATCH 9/9] x86/clear_huge_page: make clear_contig_region() preemptible

[Ankur Arora <ankur.a.arora@oracle.com>]

Peter Zijlstra <peterz@infradead.org> writes:

> On Sun, Apr 02, 2023 at 10:22:33PM -0700, Ankur Arora wrote:
>> clear_contig_region() can be used to clear up to a huge-page (2MB/1GB)
>> chunk Allow preemption in the irqentry_exit path to make sure we don't
>> hold on to the CPU for an arbitrarily long period.
>>
>> Performance: vm-scalability/case-anon-w-seq-hugetlb mmaps an anonymous
>> hugetlb-2mb region, and then writes sequentially to the region, demand
>> faulting pages on the way.
>>
>> This test, with a CONFIG_VOLUNTARY config shows the effects of this
>> change: stime drops (~18% on Icelakex, ~5% on Milan), while the utime
>> goes up (~15% on Icelakex, ~13% on Milan.)
>>
>>   *Icelakex*                  mm/clear_huge_page   x86/clear_huge_page   change
>>   (mem=4GB/task, tasks=128)
>>
>>   stime                           293.02 +- .49%        239.39 +- .83%   -18.30%
>>   utime                           440.11 +- .28%        508.74 +- .60%   +15.59%
>>   wall-clock                        5.96 +- .33%          6.27 +-2.23%   + 5.20%
>>
>>
>>
>>   *Milan*                     mm/clear_huge_page   x86/clear_huge_page   change
>>   (mem=1GB/task, tasks=512)
>>
>>   stime                          490.95 +- 3.55%       466.90 +- 4.79%   - 4.89%
>>   utime                          276.43 +- 2.85%       311.97 +- 5.15%   +12.85%
>>   wall-clock                       3.74 +- 6.41%         3.58 +- 7.82%   - 4.27%
>>
>> The drop in stime is due to REP; STOS being more efficient for bigger
>> extents.  The increase in utime is due to cache effects of that change:
>> mm/clear_huge_page() clears page-at-a-time, while narrowing towards the
>> faulting page; while x86/clear_huge_page only optimizes for cache
>> locality in the local neighbourhood of the faulting address.
>>
>> This effect on utime is visible via the increased L1-dcache-load-misses
>> and LLC-load* and an increased backend boundedness for perf user-stat
>> --all-user on Icelakex. The effect is slight but given the heavy cache
>> pressure generated by the test, shows up in the drop in user IPC:
>>
>>     -  9,455,243,414,829      instructions                     #    2.75  insn per cycle              ( +- 14.14% )  (46.17%)
>>     -  2,367,920,864,112      L1-dcache-loads                  #    1.054 G/sec                       ( +- 14.14% )  (69.24%)
>>     -     42,075,182,813      L1-dcache-load-misses            #    2.96% of all L1-dcache accesses   ( +- 14.14% )  (69.24%)
>>     -         20,365,688      LLC-loads                        #    9.064 K/sec                       ( +- 13.98% )  (69.24%)
>>     -            890,382      LLC-load-misses                  #    7.18% of all LL-cache accesses    ( +- 14.91% )  (69.24%)
>>
>>     +  9,467,796,660,698      instructions                     #    2.37  insn per cycle              ( +- 14.14% )  (46.16%)
>>     +  2,369,973,307,561      L1-dcache-loads                  #    1.027 G/sec                       ( +- 14.14% )  (69.24%)
>>     +     42,155,621,201      L1-dcache-load-misses            #    2.96% of all L1-dcache accesses   ( +- 14.14% )  (69.24%)
>>     +         22,116,300      LLC-loads                        #    9.588 K/sec                       ( +- 14.20% )  (69.24%)
>>     +          1,355,607      LLC-load-misses                  #   10.29% of all LL-cache accesses    ( +- 15.49% )  (69.25%)
>>
>> Given the fact that the stime improves for all loads using this path,
>> while the utime drop is load dependent add this change.
>
> Either I really need sleep, or *NONE* of the above is actually relevant
> to what the patch below actually does!

Yeah, you are right about the relevance.

I wanted to provide two sets of stats: the raw memory BW stats and the
relevant vm-scalability workload. The vm-scalability workload needs a
more reasonable scheduling model than what's present until this patch
and so it seemed to make sense to put here for that reason.
But yeah it doesn't really fit here.

> The above talks about the glories of using large clears, while the patch
> allows reschedules which are about latency.

Yeah, let me find a more reasonable way to present these.

Ankur

>> Signed-off-by: Ankur Arora <ankur.a.arora@oracle.com>
>> ---
>>  arch/x86/mm/hugetlbpage.c | 10 ++++++++++
>>  1 file changed, 10 insertions(+)
>>
>> diff --git a/arch/x86/mm/hugetlbpage.c b/arch/x86/mm/hugetlbpage.c
>> index 4294b77c4f18..c8564b0552e5 100644
>> --- a/arch/x86/mm/hugetlbpage.c
>> +++ b/arch/x86/mm/hugetlbpage.c
>> @@ -158,7 +158,17 @@ hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
>>  static void clear_contig_region(struct page *page, unsigned long vaddr,
>>  				unsigned int npages)
>>  {
>> +	might_sleep();
>> +
>> +	/*
>> +	 * We might be clearing a large region.
>> +	 * Allow rescheduling.
>> +	 */
>> +	allow_resched();
>>  	clear_user_pages(page_address(page), vaddr, page, npages);
>> +	disallow_resched();
>> +
>> +	cond_resched();
>>  }
>>
>>  void clear_huge_page(struct page *page,
>> --
>> 2.31.1
>>


--
ankur