On 2024/4/21 21:53, Peter Xu wrote:
> On Sat, Apr 20, 2024 at 12:05:04PM +0800, Kefeng Wang wrote:
>> The latency of lat_pagefault increased a lot than the old kernel(vs 5.10),
>> except mm counter updating, the another obvious difference
>> shown from perf graph is the new vmf_orig_pte_uffd_wp().
> 
> Curious how different it is.
> 

It is not big, as shown in perf, only 0.1x% in the whole test, but it is
new added compared with old kernel.

Please check attached perf_0417_pagefault_x86.svg [with batch mm counter],


> I wanted to give it a quick shot over lmbench but fails on missing rpc.h,
> at least for the Intel repo.  I think it's because of the libc change to
> drop that.  Are you using a separate repo that fixed all things up?

yum install libtirpc-devel, I can build with it.

> 
>> Moving it upper is better, and maybe add __always_inline to
>> vmf_orig_pte_uffd_wp() to make set_pte_range() only check VM_UFFD_WP from
>> vm_flags?
> 
> Sounds good here, thanks.

OK, will update it too.

Thanks.
>