Re: [linux-next:master] [mm/hugetlb_vmemmap] 875fa64577: vm-scalability.throughput -34.3% regression

[Oliver Sang <oliver.sang@intel.com>]

hi, Muchun Song,

On Mon, Jul 15, 2024 at 10:40:43AM +0800, Muchun Song wrote:
> 
> 
> > On Jul 14, 2024, at 20:26, Oliver Sang <oliver.sang@intel.com> wrote:
> > 
> > hi, Yu Zhao,
> > 
> > On Wed, Jul 10, 2024 at 12:22:40AM -0600, Yu Zhao wrote:
> >> On Mon, Jul 8, 2024 at 11:11 PM kernel test robot <oliver.sang@intel.com> wrote:
> >>> 
> >>> Hello,
> >>> 
> >>> kernel test robot noticed a -34.3% regression of vm-scalability.throughput on:
> >>> 
> >>> 
> >>> commit: 875fa64577da9bc8e9963ee14fef8433f20653e7 ("mm/hugetlb_vmemmap: fix race with speculative PFN walkers")
> >> 
> >> This is likely caused by synchronize_rcu() wandering into the
> >> allocation path. I'll patch that up soon.
> >> 
> > 
> > we noticed this commit has already been merged into mainline
> > 
> > [bd225530a4c717714722c3731442b78954c765b3] mm/hugetlb_vmemmap: fix race with speculative PFN walkers
> > branch: linus/master
> 
> Did you test with HVO enabled (there are two ways to enable HVO: 1) adding cmdline with "hugetlb_free_vmemmap=on"
> or 2) write 1 to /proc/sys/vm/hugetlb_optimize_vmemmap)? I want to confirm if the regression is related
> to HVO routine.

we found a strange thing, after adding 'hugetlb_free_vmemmap=on', the data
become unstable by run to run (we use kexec from previous job to next one).
below is for 875fa64577 + 'hugetlb_free_vmemmap=on'

  "vm-scalability.throughput": [
    611622,
    645261,
    705923,
    833589,
    840140,
    884010
  ],


as a comparison, without 'hugetlb_free_vmemmap=on', for 875fa64577:

  "vm-scalability.throughput": [
    4597606,
    4357960,
    4385331,
    4631803,
    4554570,
    4462691
  ],

for 73236245e0 (parent of 875fa64577):

  "vm-scalability.throughput": [
    6866441,
    6769773,
    6942991,
    6877124,
    6785790,
    6812001
  ],

> 
> Thanks.
> 
> > 
> > and the regression still exists in our tests. do you want us to test your
> > patch? Thanks!
>