lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAEmg6AVrJ0A9QsnDZApdnq4gu=x0_1soqrhNv9oBMQKNGtsKqw@mail.gmail.com>
Date: Wed, 2 Jul 2025 17:00:14 +0800
From: Xavier Xia <xavier.qyxia@...il.com>
To: Catalin Marinas <catalin.marinas@....com>
Cc: ryan.roberts@....com, will@...nel.org, 21cnbao@...il.com, 
	ioworker0@...il.com, dev.jain@....com, akpm@...ux-foundation.org, 
	david@...hat.com, gshan@...hat.com, linux-arm-kernel@...ts.infradead.org, 
	linux-kernel@...r.kernel.org, linux-mm@...ck.org, willy@...radead.org, 
	xavier_qy@....com, ziy@...dia.com, Barry Song <baohua@...nel.org>
Subject: Re: [PATCH v7] arm64/mm: Optimize loop to reduce redundant operations
 of contpte_ptep_get

Hi Catalin,


On Tue, Jul 1, 2025 at 9:59 PM Catalin Marinas <catalin.marinas@....com> wrote:
>
> On Tue, Jun 24, 2025 at 11:25:49PM +0800, Xavier Xia wrote:
> > This commit optimizes the contpte_ptep_get and contpte_ptep_get_lockless
> > function by adding early termination logic. It checks if the dirty and
> > young bits of orig_pte are already set and skips redundant bit-setting
> > operations during the loop. This reduces unnecessary iterations and
> > improves performance.
> >
> > In order to verify the optimization performance, a test function has been
> > designed. The function's execution time and instruction statistics have
> > been traced using perf, and the following are the operation results on a
> > certain Qualcomm mobile phone chip:
> >
> > Test Code:
> >       #include <stdlib.h>
> >       #include <sys/mman.h>
> >       #include <stdio.h>
> >
> >       #define PAGE_SIZE 4096
> >       #define CONT_PTES 16
> >       #define TEST_SIZE (4096* CONT_PTES * PAGE_SIZE)
> >       #define YOUNG_BIT 8
> >       void rwdata(char *buf)
> >       {
> >               for (size_t i = 0; i < TEST_SIZE; i += PAGE_SIZE) {
> >                       buf[i] = 'a';
> >                       volatile char c = buf[i];
> >               }
> >       }
> >       void clear_young_dirty(char *buf)
> >       {
> >               if (madvise(buf, TEST_SIZE, MADV_FREE) == -1) {
> >                       perror("madvise free failed");
> >                       free(buf);
> >                       exit(EXIT_FAILURE);
> >               }
> >               if (madvise(buf, TEST_SIZE, MADV_COLD) == -1) {
> >                       perror("madvise free failed");
> >                       free(buf);
> >                       exit(EXIT_FAILURE);
> >               }
> >       }
> >       void set_one_young(char *buf)
> >       {
> >               for (size_t i = 0; i < TEST_SIZE; i += CONT_PTES * PAGE_SIZE) {
> >                       volatile char c = buf[i + YOUNG_BIT * PAGE_SIZE];
> >               }
> >       }
> >
> >       void test_contpte_perf() {
> >               char *buf;
> >               int ret = posix_memalign((void **)&buf, CONT_PTES * PAGE_SIZE,
> >                               TEST_SIZE);
> >               if ((ret != 0) || ((unsigned long)buf % CONT_PTES * PAGE_SIZE)) {
> >                       perror("posix_memalign failed");
> >                       exit(EXIT_FAILURE);
> >               }
> >
> >               rwdata(buf);
> >       #if TEST_CASE2 || TEST_CASE3
> >               clear_young_dirty(buf);
> >       #endif
> >       #if TEST_CASE2
> >               set_one_young(buf);
> >       #endif
> >
> >               for (int j = 0; j < 500; j++) {
> >                       mlock(buf, TEST_SIZE);
> >
> >                       munlock(buf, TEST_SIZE);
> >               }
> >               free(buf);
> >       }
> >
> >       int main(void)
> >       {
> >               test_contpte_perf();
> >               return 0;
> >       }
> >
> >       Descriptions of three test scenarios
> >
> > Scenario 1
> >       The data of all 16 PTEs are both dirty and young.
> >       #define TEST_CASE2 0
> >       #define TEST_CASE3 0
> >
> > Scenario 2
> >       Among the 16 PTEs, only the 8th one is young, and there are no dirty ones.
> >       #define TEST_CASE2 1
> >       #define TEST_CASE3 0
> >
> > Scenario 3
> >       Among the 16 PTEs, there are neither young nor dirty ones.
> >       #define TEST_CASE2 0
> >       #define TEST_CASE3 1
> >
> > Test results
> >
> > |Scenario 1         |       Original|       Optimized|
> > |-------------------|---------------|----------------|
> > |instructions       |    37912436160|     18731580031|
> > |test time          |         4.2797|          2.2949|
> > |overhead of        |               |                |
> > |contpte_ptep_get() |         21.31%|           4.80%|
> >
> > |Scenario 2         |       Original|       Optimized|
> > |-------------------|---------------|----------------|
> > |instructions       |    36701270862|     36115790086|
> > |test time          |         3.2335|          3.0874|
> > |Overhead of        |               |                |
> > |contpte_ptep_get() |         32.26%|          33.57%|
> >
> > |Scenario 3         |       Original|       Optimized|
> > |-------------------|---------------|----------------|
> > |instructions       |    36706279735|     36750881878|
> > |test time          |         3.2008|          3.1249|
> > |Overhead of        |               |                |
> > |contpte_ptep_get() |         31.94%|          34.59%|
> >
> > For Scenario 1, optimized code can achieve an instruction benefit of 50.59%
> > and a time benefit of 46.38%.
> > For Scenario 2, optimized code can achieve an instruction count benefit of
> > 1.6% and a time benefit of 4.5%.
> > For Scenario 3, since all the PTEs have neither the young nor the dirty
> > flag, the branches taken by optimized code should be the same as those of
> > the original code. In fact, the test results of optimized code seem to be
> > closer to those of the original code.
> >
> > Ryan re-ran these tests on Apple M2 with 4K base pages + 64K mTHP.
> >
> > Scenario 1: reduced to 56% of baseline execution time
> > Scenario 2: reduced to 89% of baseline execution time
> > Scenario 3: reduced to 91% of baseline execution time
>
> Still not keen on microbenchmarks to justify such change but at least
> the code is more readable than the macro approach in some earlier
> version.
>
> Do you have any numbers to see how it compares with your v1:
>
> https://lore.kernel.org/all/20250407092243.2207837-1-xavier_qy@163.com/
>
> That patch was a lot simpler.
>

You can check the comparison data via:

https://lore.kernel.org/all/3d338f91.8c71.1965cd8b1b8.Coremail.xavier_qy@163.com/

The v1 only optimizes Scenario 1 case (where all PTEs are both young and dirty),
but it degrades performance in other scenarios. Although the current
version increases
code complexity, its optimization results are notably significant.

--

Thanks,
Xavier

Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ