]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/commitdiff
projects / mirror_ubuntu-hirsute-kernel.git / commitdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: 2140a99)
x86/asm: Optimize clear_page()
authorBorislav Petkov <bp@suse.de>
Thu, 9 Feb 2017 00:34:49 +0000 (01:34 +0100)
committerIngo Molnar <mingo@kernel.org>
Tue, 7 Mar 2017 07:28:00 +0000 (08:28 +0100)
Currently, we CALL clear_page() which then JMPs to the proper function
chosen by the alternatives.

What we should do instead is CALL the proper function directly. (This
was something Ingo suggested a while ago). So let's do that.

Measuring our favourite kernel build workload shows that there are no
significant changes in performance.

AMD
===
  -- /tmp/before 2017-02-09 18:01:46.451961188 +0100
  ++ /tmp/after  2017-02-09 18:01:54.883961175 +0100
  @@ -1,15 +1,15 @@
    Performance counter stats for 'system wide' (5 runs):

  -    1028960.373643      cpu-clock (msec)          #    6.000 CPUs utilized            ( +-  1.41% )
  +    1023086.018961      cpu-clock (msec)          #    6.000 CPUs utilized            ( +-  1.20% )
  -           518,744      context-switches          #    0.504 K/sec                    ( +-  1.04% )
  +           518,254      context-switches          #    0.507 K/sec                    ( +-  1.01% )
  -            38,112      cpu-migrations            #    0.037 K/sec                    ( +-  1.95% )
  +            37,917      cpu-migrations            #    0.037 K/sec                    ( +-  1.02% )
  -        20,874,266      page-faults               #    0.020 M/sec                    ( +-  0.07% )
  +        20,918,897      page-faults               #    0.020 M/sec                    ( +-  0.18% )
  - 2,043,646,230,667      cycles                    #    1.986 GHz                      ( +-  0.14% )  (66.67%)
  + 2,045,305,584,032      cycles                    #    1.999 GHz                      ( +-  0.16% )  (66.67%)
  -   553,698,855,431      stalled-cycles-frontend   #   27.09% frontend cycles idle     ( +-  0.07% )  (66.67%)
  +   555,099,401,413      stalled-cycles-frontend   #   27.14% frontend cycles idle     ( +-  0.13% )  (66.67%)
  -   621,544,286,390      stalled-cycles-backend    #   30.41% backend cycles idle      ( +-  0.39% )  (66.67%)
  +   621,371,430,254      stalled-cycles-backend    #   30.38% backend cycles idle      ( +-  0.32% )  (66.67%)
  - 1,738,364,431,659      instructions              #    0.85  insn per cycle
  + 1,739,895,771,901      instructions              #    0.85  insn per cycle
  -                                                  #    0.36  stalled cycles per insn  ( +-  0.11% )  (66.67%)
  +                                                  #    0.36  stalled cycles per insn  ( +-  0.13% )  (66.67%)
  -   391,170,943,850      branches                  #  380.161 M/sec                    ( +-  0.13% )  (66.67%)
  +   391,398,551,757      branches                  #  382.567 M/sec                    ( +-  0.13% )  (66.67%)
  -    22,567,810,411      branch-misses             #    5.77% of all branches          ( +-  0.11% )  (66.67%)
  +    22,574,726,683      branch-misses             #    5.77% of all branches          ( +-  0.13% )  (66.67%)

  -     171.480741921 seconds time elapsed                                          ( +-  1.41% )
  +     170.509229451 seconds time elapsed                                          ( +-  1.20% )

Intel
=====

  -- /tmp/before 2017-02-09 20:36:19.851947473 +0100
  ++ /tmp/after  2017-02-09 20:36:30.151947458 +0100
  @@ -1,15 +1,15 @@
    Performance counter stats for 'system wide' (5 runs):

  -    2207248.598126      cpu-clock (msec)          #    8.000 CPUs utilized            ( +-  0.69% )
  +    2213300.106631      cpu-clock (msec)          #    8.000 CPUs utilized            ( +-  0.73% )
  -           899,342      context-switches          #    0.407 K/sec                    ( +-  0.68% )
  +           898,381      context-switches          #    0.406 K/sec                    ( +-  0.79% )
  -            80,553      cpu-migrations            #    0.036 K/sec                    ( +-  1.13% )
  +            80,979      cpu-migrations            #    0.037 K/sec                    ( +-  1.11% )
  -        36,171,148      page-faults               #    0.016 M/sec                    ( +-  0.02% )
  +        36,179,791      page-faults               #    0.016 M/sec                    ( +-  0.02% )
  - 6,665,288,826,484      cycles                    #    3.020 GHz                      ( +-  0.07% )  (83.33%)
  + 6,671,638,410,799      cycles                    #    3.014 GHz                      ( +-  0.06% )  (83.33%)
  - 5,065,975,115,197      stalled-cycles-frontend   #   76.01% frontend cycles idle     ( +-  0.11% )  (83.33%)
  + 5,076,835,183,223      stalled-cycles-frontend   #   76.10% frontend cycles idle     ( +-  0.11% )  (83.33%)
  - 3,841,556,350,614      stalled-cycles-backend    #   57.64% backend cycles idle      ( +-  0.13% )  (66.67%)
  + 3,852,823,974,333      stalled-cycles-backend    #   57.75% backend cycles idle      ( +-  0.12% )  (66.67%)
  - 4,148,398,171,079      instructions              #    0.62  insn per cycle
  + 4,148,997,156,059      instructions              #    0.62  insn per cycle
  -                                                  #    1.22  stalled cycles per insn  ( +-  0.10% )  (83.33%)
  +                                                  #    1.22  stalled cycles per insn  ( +-  0.11% )  (83.33%)
  -   887,187,118,591      branches                  #  401.943 M/sec                    ( +-  0.09% )  (83.33%)
  +   887,271,341,121      branches                  #  400.882 M/sec                    ( +-  0.11% )  (83.33%)
  -    30,139,439,034      branch-misses             #    3.40% of all branches          ( +-  0.09% )  (83.33%)
  +    30,134,864,997      branch-misses             #    3.40% of all branches          ( +-  0.06% )  (83.33%)

  -     275.904405540 seconds time elapsed                                          ( +-  0.69% )
  +     276.660352016 seconds time elapsed                                          ( +-  0.73% )

allmodconfig vmlinux size grows by a ~1Kb but that's fine - we optimize
our calling of the clear_page variants.

     text    data     bss     dec     hex filename
  9051979 23067670        27009024        59128673        3863b61 vmlinux
  9053000 23067670        27009024        59129694        3863f5e vmlinux.clear_page

Reported-by: kernel test robot <fengguang.wu@intel.com>
Tested-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20170215111927.emdgxf2pide3kwro@pd.tnic
Signed-off-by: Ingo Molnar <mingo@kernel.org>
arch/x86/include/asm/page_64.h patch | blob | blame | history
arch/x86/lib/clear_page_64.S patch | blob | blame | history

diff --git a/arch/x86/include/asm/page_64.h b/arch/x86/include/asm/page_64.h
index b3bebf9e57466985cfcf6a54dd5d4410dd5417c8..b4a0d43248cf3d6f2c7ef042200c31a25f29f122 100644 (file)
--- a/arch/x86/include/asm/page_64.h
+++ b/arch/x86/include/asm/page_64.h
@@ -4,6 +4,7 @@
 #include <asm/page_64_types.h>
 
 #ifndef __ASSEMBLY__
+#include <asm/alternative.h>
 
 /* duplicated to the one in bootmem.h */
 extern unsigned long max_pfn;
@@ -34,7 +35,20 @@ extern unsigned long __phys_addr_symbol(unsigned long);
 #define pfn_valid(pfn)          ((pfn) < max_pfn)
 #endif
 
-void clear_page(void *page);
+void clear_page_orig(void *page);
+void clear_page_rep(void *page);
+void clear_page_erms(void *page);
+
+static inline void clear_page(void *page)
+{
+       alternative_call_2(clear_page_orig,
+                          clear_page_rep, X86_FEATURE_REP_GOOD,
+                          clear_page_erms, X86_FEATURE_ERMS,
+                          "=D" (page),
+                          "0" (page)
+                          : "memory", "rax", "rcx");
+}
+
 void copy_page(void *to, void *from);
 
 #endif /* !__ASSEMBLY__ */
diff --git a/arch/x86/lib/clear_page_64.S b/arch/x86/lib/clear_page_64.S
index 5e2af3a88cf5e47e2505926938d3f239d55f2ff3..81b1635d67dee9624de37e1f4577a36e9b7065cb 100644 (file)
--- a/arch/x86/lib/clear_page_64.S
+++ b/arch/x86/lib/clear_page_64.S
@@ -14,20 +14,15 @@
  * Zero a page.
  * %rdi        - page
  */
-ENTRY(clear_page)
-
-       ALTERNATIVE_2 "jmp clear_page_orig", "", X86_FEATURE_REP_GOOD, \
-                     "jmp clear_page_c_e", X86_FEATURE_ERMS
-
+ENTRY(clear_page_rep)
        movl $4096/8,%ecx
        xorl %eax,%eax
        rep stosq
        ret
-ENDPROC(clear_page)
-EXPORT_SYMBOL(clear_page)
+ENDPROC(clear_page_rep)
+EXPORT_SYMBOL_GPL(clear_page_rep)
 
 ENTRY(clear_page_orig)
-
        xorl   %eax,%eax
        movl   $4096/64,%ecx
        .p2align 4
@@ -47,10 +42,12 @@ ENTRY(clear_page_orig)
        nop
        ret
 ENDPROC(clear_page_orig)
+EXPORT_SYMBOL_GPL(clear_page_orig)
 
-ENTRY(clear_page_c_e)
+ENTRY(clear_page_erms)
        movl $4096,%ecx
        xorl %eax,%eax
        rep stosb
        ret
-ENDPROC(clear_page_c_e)
+ENDPROC(clear_page_erms)
+EXPORT_SYMBOL_GPL(clear_page_erms)
Ubuntu Hirsute Linux kernel mirror