On 08/08, Minchan Kim wrote:
>On Mon, Aug 07, 2017 at 10:51:00PM -0700, Nadav Amit wrote:
>> Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
>>
>> > Minchan Kim <minchan@xxxxxxxxxx> wrote:
>> >
>> >> Hi,
>> >>
>> >> On Tue, Aug 08, 2017 at 09:19:23AM +0800, kernel test robot wrote:
>> >>> Greeting,
>> >>>
>> >>> FYI, we noticed a -19.3% regression of will-it-scale.per_process_ops due to commit:
>> >>>
>> >>>
>> >>> commit: 76742700225cad9df49f05399381ac3f1ec3dc60 ("mm: fix MADV_[FREE|DONTNEED] TLB flush miss problem")
>> >>> url: https://github.com/0day-ci/linux/commits/Nadav-Amit/mm-migrate-prevent-racy-access-to-tlb_flush_pending/20170802-205715
>> >>>
>> >>>
>> >>> in testcase: will-it-scale
>> >>> on test machine: 88 threads Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz with 64G memory
>> >>> with following parameters:
>> >>>
>> >>> nr_task: 16
>> >>> mode: process
>> >>> test: brk1
>> >>> cpufreq_governor: performance
>> >>>
>> >>> test-description: Will It Scale takes a testcase and runs it from 1 through to n parallel copies to see if the testcase will scale. It builds both a process and threads based test in order to see any differences between the two.
>> >>> test-url: https://github.com/antonblanchard/will-it-scale
>> >>
>> >> Thanks for the report.
>> >> Could you explain what kinds of workload you are testing?
>> >>
>> >> Does it calls frequently madvise(MADV_DONTNEED) in parallel on multiple
>> >> threads?
>> >
>> > According to the description it is "testcase:brk increase/decrease of one
>> > page”. According to the mode it spawns multiple processes, not threads.
>> >
>> > Since a single page is unmapped each time, and the iTLB-loads increase
>> > dramatically, I would suspect that for some reason a full TLB flush is
>> > caused during do_munmap().
>> >
>> > If I find some free time, I’ll try to profile the workload - but feel free
>> > to beat me to it.
>>
>> The root-cause appears to be that tlb_finish_mmu() does not call
>> dec_tlb_flush_pending() - as it should. Any chance you can take care of it?
>
>Oops, but with second looking, it seems it's not my fault. ;-)
>https://marc.info/?l=linux-mm&m=150156699114088&w=2
>
>Anyway, thanks for the pointing out.
>xiaolong.ye, could you retest with this fix?
>
I've queued tests for 5 times and results show this patch (e8f682574e4 "mm:
decrease tlb flush pending count in tlb_finish_mmu") does help recover the
performance back.
378005bdbac0a2ec 76742700225cad9df49f053993 e8f682574e45b6406dadfffeb4
---------------- -------------------------- --------------------------
%stddev change %stddev change %stddev
\ | \ | \
3405093 -19% 2747088 -2% 3348752 will-it-scale.per_process_ops
1280 ± 3% -2% 1257 ± 3% -6% 1207 vmstat.system.cs
2702 ± 18% 11% 3002 ± 19% 17% 3156 ± 18% numa-vmstat.node0.nr_mapped
10765 ± 18% 11% 11964 ± 19% 17% 12588 ± 18% numa-meminfo.node0.Mapped
0.00 ± 47% -40% 0.00 ± 45% -84% 0.00 ± 42% mpstat.cpu.soft%
Thanks,
Xiaolong
>From 83012114c9cd9304f0d55d899bb4b9329d0e22ac Mon Sep 17 00:00:00 2001
>From: Minchan Kim <minchan@xxxxxxxxxx>
>Date: Tue, 8 Aug 2017 17:05:19 +0900
>Subject: [PATCH] mm: decrease tlb flush pending count in tlb_finish_mmu
>
>The tlb pending count increased by tlb_gather_mmu should be decreased
>at tlb_finish_mmu. Otherwise, A lot of TLB happens which makes
>performance regression.
>
>Signed-off-by: Minchan Kim <minchan@xxxxxxxxxx>
>---
> mm/memory.c | 1 +
> 1 file changed, 1 insertion(+)
>
>diff --git a/mm/memory.c b/mm/memory.c
>index 34b1fcb829e4..ad2617552f55 100644
>--- a/mm/memory.c
>+++ b/mm/memory.c
>@@ -423,6 +423,7 @@ void tlb_finish_mmu(struct mmu_gather *tlb,
> bool force = mm_tlb_flush_nested(tlb->mm);
>
> arch_tlb_finish_mmu(tlb, start, end, force);
>+ dec_tlb_flush_pending(tlb->mm);
> }
>
> /*
>--
>2.7.4
>
