mbox series

[v2,0/7] arch/mm/fault: accelerate pagefault when badaccess

Message ID 20240403083805.1818160-1-wangkefeng.wang@huawei.com (mailing list archive)
Headers show
Series arch/mm/fault: accelerate pagefault when badaccess | expand

Message

Kefeng Wang April 3, 2024, 8:37 a.m. UTC 
After VMA lock-based page fault handling enabled, if bad access met
under per-vma lock, it will fallback to mmap_lock-based handling,
so it leads to unnessary mmap lock and vma find again. A test from
lmbench shows 34% improve after this changes on arm64,

  lat_sig -P 1 prot lat_sig 0.29194 -> 0.19198

Only build test on other archs except arm64.

v2: 
- a better changelog, and describe the counting changes, suggested by
  Suren Baghdasaryan
- add RB

Kefeng Wang (7):
  arm64: mm: cleanup __do_page_fault()
  arm64: mm: accelerate pagefault when VM_FAULT_BADACCESS
  arm: mm: accelerate pagefault when VM_FAULT_BADACCESS
  powerpc: mm: accelerate pagefault when badaccess
  riscv: mm: accelerate pagefault when badaccess
  s390: mm: accelerate pagefault when badaccess
  x86: mm: accelerate pagefault when badaccess

 arch/arm/mm/fault.c     |  4 +++-
 arch/arm64/mm/fault.c   | 31 ++++++++++---------------------
 arch/powerpc/mm/fault.c | 33 ++++++++++++++++++++-------------
 arch/riscv/mm/fault.c   |  5 ++++-
 arch/s390/mm/fault.c    |  3 ++-
 arch/x86/mm/fault.c     | 23 ++++++++++++++---------
 6 files changed, 53 insertions(+), 46 deletions(-)

Comments

Andrew Morton April 3, 2024, 8:45 p.m. UTC | #1 
On Wed, 3 Apr 2024 16:37:58 +0800 Kefeng Wang <wangkefeng.wang@huawei.com> wrote:

> After VMA lock-based page fault handling enabled, if bad access met
> under per-vma lock, it will fallback to mmap_lock-based handling,
> so it leads to unnessary mmap lock and vma find again. A test from
> lmbench shows 34% improve after this changes on arm64,
> 
>   lat_sig -P 1 prot lat_sig 0.29194 -> 0.19198
> 
> Only build test on other archs except arm64.

Thanks.  So we now want a bunch of architectures to runtime test this.  Do
we have a selftest in place which will adequately do this?
Kefeng Wang April 7, 2024, 7:49 a.m. UTC | #2 
On 2024/4/4 4:45, Andrew Morton wrote:
> On Wed, 3 Apr 2024 16:37:58 +0800 Kefeng Wang <wangkefeng.wang@huawei.com> wrote:
> 
>> After VMA lock-based page fault handling enabled, if bad access met
>> under per-vma lock, it will fallback to mmap_lock-based handling,
>> so it leads to unnessary mmap lock and vma find again. A test from
>> lmbench shows 34% improve after this changes on arm64,
>>
>>    lat_sig -P 1 prot lat_sig 0.29194 -> 0.19198
>>
>> Only build test on other archs except arm64.
> 
> Thanks.  So we now want a bunch of architectures to runtime test this.  Do
> we have a selftest in place which will adequately do this?

I don't find such selftest, and badaccess would lead to coredump, the
performance should not affect most scene, so no selftest is acceptable.
lmbench is easy to use to measure the performance.
Heiko Carstens April 7, 2024, 5:19 p.m. UTC | #3 
On Sun, Apr 07, 2024 at 03:49:53PM +0800, Kefeng Wang wrote:
> On 2024/4/4 4:45, Andrew Morton wrote:
> > On Wed, 3 Apr 2024 16:37:58 +0800 Kefeng Wang <wangkefeng.wang@huawei.com> wrote:
> > 
> > > After VMA lock-based page fault handling enabled, if bad access met
> > > under per-vma lock, it will fallback to mmap_lock-based handling,
> > > so it leads to unnessary mmap lock and vma find again. A test from
> > > lmbench shows 34% improve after this changes on arm64,
> > > 
> > >    lat_sig -P 1 prot lat_sig 0.29194 -> 0.19198
> > > 
> > > Only build test on other archs except arm64.
> > 
> > Thanks.  So we now want a bunch of architectures to runtime test this.  Do
> > we have a selftest in place which will adequately do this?
> 
> I don't find such selftest, and badaccess would lead to coredump, the
> performance should not affect most scene, so no selftest is acceptable.
> lmbench is easy to use to measure the performance.

The rationale for this series (performance improvement) is a bit odd,
since I would expect that the changed code is usually never executed.