Re: [PATCH] mm: memory: check userfaultfd_wp() in vmf_orig_pte_uffd_wp()

[Kefeng Wang <wangkefeng.wang@huawei.com>]

On 2024/4/18 2:34, Peter Xu wrote:
> Hi, Kefeng,
> 
> On Wed, Apr 17, 2024 at 05:30:40PM +0800, Kefeng Wang wrote:
>>
>>
>> On 2024/4/17 16:23, Kefeng Wang wrote:
>>> Directly call vmf_orig_pte_uffd_wp() in do_anonymous_page() and
>>> set_pte_range() to save a uffd_wp and add userfaultfd_wp() check
>>> in vmf_orig_pte_uffd_wp() to avoid the unnecessary function calls
>>> in the most sense, lat_pagefault testcase does show improvement
>>> though very small(~1%).
> 
> I'm ok with the change if that helps as big as 1%, but I'm a bit surprised
> to see such a difference, because for file pte_marker_uffd_wp() should
> check first on pte_none() then it should return already if uffd not even
> registered for the vma, while orig_pte should be hot too if valid.

Yes, retest, not as big as 1%, but the perf shows vmf_orig_pte_uffd_wp
is eliminated,

[root@localhost]# perf report -i perf.data.old |grep vmf
      0.17%     0.13%  lat_pagefault  [kernel.kallsyms]      [k] 
vmf_orig_pte_uffd_wp.part.0.isra.0
[root@localhost]# perf report -i perf.data |grep vmf

> 
>>>
>>> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
>>> ---
>>>    mm/memory.c | 9 +++++----
>>>    1 file changed, 5 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 5ae2409d3cb9..a6afc96001e6 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -117,6 +117,9 @@ static bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
>>
>>
>> 	 if (!IS_ENABLED(CONFIG_PTE_MARKER_UFFD_WP))
>>          	return false;
>>
>> Will add config check too,
> 
> pte_marker_uffd_wp() returns false when !PTE_MARKER_UFFD_WP, so kind of
> imply this.  I assume you meant to avoid checking ORIG_PTE_VALID flag, but

Just to avoid checking ORIG_PTE_VALID and the new userfaultfd_wp() since
it is not supported on most archs.

> the flags is pretty hot too.  Again, just want to double check with you on
> whether it can have such a huge difference, e.g., how that compares with
> the current code v.s. original patch v.s. this squashed.

I will change the changelog to show different about vmf_orig_pte_uffd_wp 
  from perf data.

Thanks.