Re: [LKP] Re: [mm/gup] 57efa1fe59: will-it-scale.per_thread_ops -9.2% regression

[Xing Zhengjun]

Hi Linus,

On 5/25/2021 11:11 AM, Linus Torvalds wrote:
> On Mon, May 24, 2021 at 5:00 PM kernel test robot <oliver.sang@xxxxxxxxx> wrote:
> > FYI, we noticed a -9.2% regression of will-it-scale.per_thread_ops due to commit:
> > commit: 57efa1fe5957694fa541c9062de0a127f0b9acb0 ("mm/gup: prevent gup_fast from racing with COW during fork")
> Hmm. This looks like one of those "random fluctuations" things.
>
> It would be good to hear if other test-cases also bisect to the same
> thing, but this report already says:
>
> > In addition to that, the commit also has significant impact on the following tests:
> >
> > +------------------+---------------------------------------------------------------------------------+
> > | testcase: change | will-it-scale: will-it-scale.per_thread_ops 3.7% improvement                    |
> which does kind of reinforce that "this benchmark gives unstable numbers".
>
> The perf data doesn't even mention any of the GUP paths, and on the
> pure fork path the biggest impact would be:
>
>   (a) maybe "struct mm_struct" changed in size or had a different cache layout
I move "write_protect_seq" to the tail of the "struct mm_struct", the 
regression reduced to -3.6%. The regression should relate to the cache 
layout.
=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/nr_task/mode/test/cpufreq_governor/ucode:
lkp-icl-2sp1/will-it-scale/debian-10.4-x86_64-20200603.cgz/x86_64-rhel-8.3/gcc-9/50%/thread/mmap1/performance/0xb000280

commit:
  c28b1fc70390df32e29991eedd52bd86e7aba080
  57efa1fe5957694fa541c9062de0a127f0b9acb0
  f6a9c27882d51ff551e15522992d3725c342372d  (the test patch)

c28b1fc70390df32 57efa1fe5957694fa541c9062de f6a9c27882d51ff551e15522992
---------------- --------------------------- ---------------------------
         %stddev     %change         %stddev     %change %stddev
             \          |                \          | \
    341938            -9.0%     311218 ±  2%      -3.6% 329513        
will-it-scale.48.threads
      7123            -9.0%       6483 ±  2%      -3.6% 6864        
will-it-scale.per_thread_ops
    341938            -9.0%     311218 ±  2%      -3.6% 329513        
will-it-scale.workload

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 915f4f100383..34bb2a01806c 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -447,13 +447,6 @@ struct mm_struct {
                 */
                atomic_t has_pinned;

-               /**
-                * @write_protect_seq: Locked when any thread is write
-                * protecting pages mapped by this mm to enforce a later 
COW,
-                * for instance during page table copying for fork().
-                */
-               seqcount_t write_protect_seq;
-
 #ifdef CONFIG_MMU
                atomic_long_t pgtables_bytes;   /* PTE page table pages */
 #endif
@@ -564,6 +557,12 @@ struct mm_struct {
 #ifdef CONFIG_IOMMU_SUPPORT
                u32 pasid;
 #endif
+                /**
+                 * @write_protect_seq: Locked when any thread is write
+                 * protecting pages mapped by this mm to enforce a 
later COW,
+                 * for instance during page table copying for fork().
+                 */
+                seqcount_t write_protect_seq;
        } __randomize_layout;

        /*

>   (b) two added (nonatomic) increment operations in the fork path due
> to the seqcount
>
> and I'm not seeing what would cause that 9% change. Obviously cache
> placement has done it before.
>
> If somebody else sees something that I'm missing, please holler. But
> I'll ignore this as "noise" otherwise.
>
>              Linus
> _______________________________________________
> LKP mailing list -- lkp@xxxxxxxxxxxx
> To unsubscribe send an email to lkp-leave@xxxxxxxxxxxx

--
Zhengjun Xing