| Message ID | 20181011111652.31840-1-vincent.whitchurch@axis.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | ARM: Optimise copy_{from/to}_user for !CPU_USE_DOMAINS | expand |
On Thu, 11 Oct 2018, Vincent Whitchurch wrote: > ARMv6+ processors do not use CONFIG_CPU_USE_DOMAINS and use privileged > ldr/str instructions in copy_{from/to}_user. They are currently > unnecessarily using single ldr/str instructions and can use ldm/stm > instructions instead like memcpy does (but with appropriate fixup > tables). > > function old new delta > arm_copy_from_user 924 824 -100 > __copy_to_user_std 936 824 -112 > > This speeds up a "dd if=foo of=bar bs=32k" on a tmpfs filesystem by > about 4% on my Cortex-A9. > > before:134217728 bytes (128.0MB) copied, 0.543848 seconds, 235.4MB/s > before:134217728 bytes (128.0MB) copied, 0.538610 seconds, 237.6MB/s > before:134217728 bytes (128.0MB) copied, 0.544356 seconds, 235.1MB/s > before:134217728 bytes (128.0MB) copied, 0.544364 seconds, 235.1MB/s > before:134217728 bytes (128.0MB) copied, 0.537130 seconds, 238.3MB/s > before:134217728 bytes (128.0MB) copied, 0.533443 seconds, 240.0MB/s > before:134217728 bytes (128.0MB) copied, 0.545691 seconds, 234.6MB/s > before:134217728 bytes (128.0MB) copied, 0.534695 seconds, 239.4MB/s > before:134217728 bytes (128.0MB) copied, 0.540561 seconds, 236.8MB/s > before:134217728 bytes (128.0MB) copied, 0.541025 seconds, 236.6MB/s > > after:134217728 bytes (128.0MB) copied, 0.520445 seconds, 245.9MB/s > after:134217728 bytes (128.0MB) copied, 0.527846 seconds, 242.5MB/s > after:134217728 bytes (128.0MB) copied, 0.519510 seconds, 246.4MB/s > after:134217728 bytes (128.0MB) copied, 0.527231 seconds, 242.8MB/s > after:134217728 bytes (128.0MB) copied, 0.525030 seconds, 243.8MB/s > after:134217728 bytes (128.0MB) copied, 0.524236 seconds, 244.2MB/s > after:134217728 bytes (128.0MB) copied, 0.523659 seconds, 244.4MB/s > after:134217728 bytes (128.0MB) copied, 0.525018 seconds, 243.8MB/s > after:134217728 bytes (128.0MB) copied, 0.519249 seconds, 246.5MB/s > after:134217728 bytes (128.0MB) copied, 0.518527 seconds, 246.9MB/s > > Signed-off-by: Vincent Whitchurch <vincent.whitchurch@axis.com> > --- [...] > index a826df3d3814..7751072d409a 100644 > --- a/arch/arm/lib/copy_from_user.S > +++ b/arch/arm/lib/copy_from_user.S > @@ -34,13 +34,19 @@ > * Number of bytes NOT copied. > */ > > +#ifdef CONFIG_CPU_USE_DOMAINS > #ifndef CONFIG_THUMB2_KERNEL > #define LDR1W_SHIFT 0 > #else > #define LDR1W_SHIFT 1 > #endif > +#else > +#define LDR1W_SHIFT 0 > +#endif > + > #define STR1W_SHIFT 0 > > +#ifdef CONFIG_CPU_USE_DOMAINS > .macro ldr1w ptr reg abort I'd suggest you do the grouping differently. The LDR1W_SHIFT and STR1W_SHIFT definitions are direct properties of the ldr1w and str1w macro implementations, and not necessarily whether you have CONFIG_CPU_USE_DOMAINS defined. In other words, I'd prefer to see: #ifdef CONFIG_CPU_USE_DOMAINS #define LDR1W_SHIFT ... #define STR1W_SHIFT ... .macro ldr1w ... ... #else #define LDR1W_SHIFT ... #define STR1W_SHIFT ... .macro ldr1w ... ... #endif Nicolas
diff --git a/arch/arm/include/asm/assembler.h b/arch/arm/include/asm/assembler.h index b17ee03d280b..da16d31c7ef9 100644 --- a/arch/arm/include/asm/assembler.h +++ b/arch/arm/include/asm/assembler.h @@ -243,13 +243,15 @@ .endm #endif -#define USER(x...) \ +#define USERL(l, x...) \ 9999: x; \ .pushsection __ex_table,"a"; \ .align 3; \ - .long 9999b,9001f; \ + .long 9999b,l; \ .popsection +#define USER(x...) USERL(9001f, x) + #ifdef CONFIG_SMP #define ALT_SMP(instr...) \ 9998: instr diff --git a/arch/arm/lib/copy_from_user.S b/arch/arm/lib/copy_from_user.S index a826df3d3814..7751072d409a 100644 --- a/arch/arm/lib/copy_from_user.S +++ b/arch/arm/lib/copy_from_user.S @@ -34,13 +34,19 @@ * Number of bytes NOT copied. */ +#ifdef CONFIG_CPU_USE_DOMAINS #ifndef CONFIG_THUMB2_KERNEL #define LDR1W_SHIFT 0 #else #define LDR1W_SHIFT 1 #endif +#else +#define LDR1W_SHIFT 0 +#endif + #define STR1W_SHIFT 0 +#ifdef CONFIG_CPU_USE_DOMAINS .macro ldr1w ptr reg abort ldrusr \reg, \ptr, 4, abort=\abort .endm @@ -60,6 +66,23 @@ .macro ldr1b ptr reg cond=al abort ldrusr \reg, \ptr, 1, \cond, abort=\abort .endm +#else + .macro ldr1w ptr reg abort + USERL(\abort, W(ldr) \reg, [\ptr], #4) + .endm + + .macro ldr4w ptr reg1 reg2 reg3 reg4 abort + USERL(\abort, ldmia \ptr!, {\reg1, \reg2, \reg3, \reg4}) + .endm + + .macro ldr8w ptr reg1 reg2 reg3 reg4 reg5 reg6 reg7 reg8 abort + USERL(\abort, ldmia \ptr!, {\reg1, \reg2, \reg3, \reg4, \reg5, \reg6, \reg7, \reg8}) + .endm + + .macro ldr1b ptr reg cond=al abort + USERL(\abort, ldr\cond\()b \reg, [\ptr], #1) + .endm +#endif .macro str1w ptr reg abort W(str) \reg, [\ptr], #4 diff --git a/arch/arm/lib/copy_to_user.S b/arch/arm/lib/copy_to_user.S index caf5019d8161..3149446bc21e 100644 --- a/arch/arm/lib/copy_to_user.S +++ b/arch/arm/lib/copy_to_user.S @@ -35,11 +35,16 @@ */ #define LDR1W_SHIFT 0 + +#ifdef CONFIG_CPU_USE_DOMAINS #ifndef CONFIG_THUMB2_KERNEL #define STR1W_SHIFT 0 #else #define STR1W_SHIFT 1 #endif +#else +#define STR1W_SHIFT 0 +#endif .macro ldr1w ptr reg abort W(ldr) \reg, [\ptr], #4 @@ -57,6 +62,7 @@ ldr\cond\()b \reg, [\ptr], #1 .endm +#ifdef CONFIG_CPU_USE_DOMAINS .macro str1w ptr reg abort strusr \reg, \ptr, 4, abort=\abort .endm @@ -75,6 +81,19 @@ .macro str1b ptr reg cond=al abort strusr \reg, \ptr, 1, \cond, abort=\abort .endm +#else + .macro str1w ptr reg abort + USERL(\abort, W(str) \reg, [\ptr], #4) + .endm + + .macro str8w ptr reg1 reg2 reg3 reg4 reg5 reg6 reg7 reg8 abort + USERL(\abort, stmia \ptr!, {\reg1, \reg2, \reg3, \reg4, \reg5, \reg6, \reg7, \reg8}) + .endm + + .macro str1b ptr reg cond=al abort + USERL(\abort, str\cond\()b \reg, [\ptr], #1) + .endm +#endif .macro enter reg1 reg2 mov r3, #0
ARMv6+ processors do not use CONFIG_CPU_USE_DOMAINS and use privileged ldr/str instructions in copy_{from/to}_user. They are currently unnecessarily using single ldr/str instructions and can use ldm/stm instructions instead like memcpy does (but with appropriate fixup tables). function old new delta arm_copy_from_user 924 824 -100 __copy_to_user_std 936 824 -112 This speeds up a "dd if=foo of=bar bs=32k" on a tmpfs filesystem by about 4% on my Cortex-A9. before:134217728 bytes (128.0MB) copied, 0.543848 seconds, 235.4MB/s before:134217728 bytes (128.0MB) copied, 0.538610 seconds, 237.6MB/s before:134217728 bytes (128.0MB) copied, 0.544356 seconds, 235.1MB/s before:134217728 bytes (128.0MB) copied, 0.544364 seconds, 235.1MB/s before:134217728 bytes (128.0MB) copied, 0.537130 seconds, 238.3MB/s before:134217728 bytes (128.0MB) copied, 0.533443 seconds, 240.0MB/s before:134217728 bytes (128.0MB) copied, 0.545691 seconds, 234.6MB/s before:134217728 bytes (128.0MB) copied, 0.534695 seconds, 239.4MB/s before:134217728 bytes (128.0MB) copied, 0.540561 seconds, 236.8MB/s before:134217728 bytes (128.0MB) copied, 0.541025 seconds, 236.6MB/s after:134217728 bytes (128.0MB) copied, 0.520445 seconds, 245.9MB/s after:134217728 bytes (128.0MB) copied, 0.527846 seconds, 242.5MB/s after:134217728 bytes (128.0MB) copied, 0.519510 seconds, 246.4MB/s after:134217728 bytes (128.0MB) copied, 0.527231 seconds, 242.8MB/s after:134217728 bytes (128.0MB) copied, 0.525030 seconds, 243.8MB/s after:134217728 bytes (128.0MB) copied, 0.524236 seconds, 244.2MB/s after:134217728 bytes (128.0MB) copied, 0.523659 seconds, 244.4MB/s after:134217728 bytes (128.0MB) copied, 0.525018 seconds, 243.8MB/s after:134217728 bytes (128.0MB) copied, 0.519249 seconds, 246.5MB/s after:134217728 bytes (128.0MB) copied, 0.518527 seconds, 246.9MB/s Signed-off-by: Vincent Whitchurch <vincent.whitchurch@axis.com> --- arch/arm/include/asm/assembler.h | 6 ++++-- arch/arm/lib/copy_from_user.S | 23 +++++++++++++++++++++++ arch/arm/lib/copy_to_user.S | 19 +++++++++++++++++++ 3 files changed, 46 insertions(+), 2 deletions(-)