* [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX @ 2011-12-08 5:02 Anton Blanchard 2011-12-08 5:44 ` Kumar Gala 0 siblings, 1 reply; 11+ messages in thread From: Anton Blanchard @ 2011-12-08 5:02 UTC (permalink / raw) To: benh, paulus, mikey, sukadev, hbabu; +Cc: linuxppc-dev Implement a POWER7 optimised copy_to_user/copy_from_user using VMX. For large aligned copies this new loop is over 10% faster, and for large unaligned copies it is over 200% faster. If we take a fault we fall back to the old version, this keeps things relatively simple and easy to verify. On POWER7 unaligned stores rarely slow down - they only flush when a store crosses a 4KB page boundary. Furthermore this flush is handled completely in hardware and should be 20-30 cycles. Unaligned loads on the other hand flush much more often - whenever crossing a 128 byte cache line, or a 32 byte sector if either sector is an L1 miss. Considering this information we really want to get the loads aligned and not worry about the alignment of the stores. Microbenchmarks confirm that this approach is much faster than the current unaligned copy loop that uses shifts and rotates to ensure both loads and stores are aligned. We also want to try and do the stores in cacheline aligned, cacheline sized chunks. If the store queue is unable to merge an entire cacheline of stores then the L2 cache will have to do a read/modify/write. Even worse, we will serialise this with the stores in the next iteration of the copy loop since both iterations hit the same cacheline. Based on this, the new loop does the following things: 1 - 127 bytes Get the source 8 byte aligned and use 8 byte loads and stores. Pretty boring and similar to how the current loop works. 128 - 4095 bytes Get the source 8 byte aligned and use 8 byte loads and stores, 1 cacheline at a time. We aren't doing the stores in cacheline aligned chunks so we will potentially serialise once per cacheline. Even so it is much better than the loop we have today. 4096 - bytes If both source and destination have the same alignment get them both 16 byte aligned, then get the destination cacheline aligned. Do cacheline sized loads and stores using VMX. If source and destination do not have the same alignment, we get the destination cacheline aligned, and use permute to do aligned loads. In both cases the VMX loop should be optimal - we always do aligned loads and stores and are always doing stores in cacheline aligned, cacheline sized chunks. To be able to use VMX we must be careful about interrupts and sleeping. We don't use the VMX loop when in an interrupt (which should be rare anyway) and we wrap the VMX loop in disable/enable_pagefault and fall back to the existing copy_tofrom_user loop if we do need to sleep. The VMX breakpoint of 4096 bytes was chosen using this microbenchmark: http://ozlabs.org/~anton/junkcode/copy_to_user.c Since we are using VMX and there is a cost to saving and restoring the user VMX state there are two broad cases we need to benchmark: - Best case - userspace never uses VMX - Worst case - userspace always uses VMX In reality a userspace process will sit somewhere between these two extremes. Since we need to test both aligned and unaligned copies we end up with 4 combinations. The point at which the VMX loop begins to win is: 0% VMX aligned 2048 bytes unaligned 2048 bytes 100% VMX aligned 16384 bytes unaligned 8192 bytes Considering this is a microbenchmark, the data is hot in cache and the VMX loop has better store queue merging properties we set the breakpoint to 4096 bytes, a little below the unaligned breakpoints. Some future optimisations we can look at: - Looking at the perf data, a significant part of the cost when a task is always using VMX is the extra exception we take to restore the VMX state. As such we should do something similar to the x86 optimisation that restores FPU state for heavy users. ie: /* * If the task has used fpu the last 5 timeslices, just do a full * restore of the math state immediately to avoid the trap; the * chances of needing FPU soon are obviously high now */ preload_fpu = tsk_used_math(next_p) && next_p->fpu_counter > 5; and /* * fpu_counter contains the number of consecutive context switches * that the FPU is used. If this is over a threshold, the lazy fpu * saving becomes unlazy to save the trap. This is an unsigned char * so that after 256 times the counter wraps and the behavior turns * lazy again; this to deal with bursty apps that only use FPU for * a short time */ - We could create a paca bit to mirror the VMX enabled MSR bit and check that first, avoiding multiple calls to calling enable_kernel_altivec. That should help with iovec based system calls like readv. - We could have two VMX breakpoints, one for when we know the user VMX state is loaded into the registers and one when it isn't. This could be a second bit in the paca so we can calculate the break points quickly. - One suggestion from Ben was to save and restore the VSX registers we use inline instead of using enable_kernel_altivec. Signed-off-by: Anton Blanchard <anton@samba.org> --- Index: linux-build/arch/powerpc/include/asm/cputable.h =================================================================== --- linux-build.orig/arch/powerpc/include/asm/cputable.h 2011-09-07 15:15:49.096458526 +1000 +++ linux-build/arch/powerpc/include/asm/cputable.h 2011-12-08 15:38:46.627313507 +1100 @@ -201,6 +201,7 @@ extern const char *powerpc_base_platform #define CPU_FTR_POPCNTB LONG_ASM_CONST(0x0400000000000000) #define CPU_FTR_POPCNTD LONG_ASM_CONST(0x0800000000000000) #define CPU_FTR_ICSWX LONG_ASM_CONST(0x1000000000000000) +#define CPU_FTR_POWER7 LONG_ASM_CONST(0x2000000000000000) #ifndef __ASSEMBLY__ @@ -425,7 +426,7 @@ extern const char *powerpc_base_platform CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \ CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYM_SMT | \ CPU_FTR_STCX_CHECKS_ADDRESS | CPU_FTR_POPCNTB | CPU_FTR_POPCNTD | \ - CPU_FTR_ICSWX | CPU_FTR_CFAR | CPU_FTR_HVMODE) + CPU_FTR_ICSWX | CPU_FTR_CFAR | CPU_FTR_HVMODE | CPU_FTR_POWER7) #define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \ CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \ CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \ Index: linux-build/arch/powerpc/lib/copyuser_64.S =================================================================== --- linux-build.orig/arch/powerpc/lib/copyuser_64.S 2011-09-07 15:15:49.146459439 +1000 +++ linux-build/arch/powerpc/lib/copyuser_64.S 2011-12-08 15:38:42.491238635 +1100 @@ -11,6 +11,12 @@ .align 7 _GLOBAL(__copy_tofrom_user) +BEGIN_FTR_SECTION + nop +FTR_SECTION_ELSE + b __copy_tofrom_user_power7 +ALT_FTR_SECTION_END_IFCLR(CPU_FTR_POWER7) +_GLOBAL(__copy_tofrom_user_base) /* first check for a whole page copy on a page boundary */ cmpldi cr1,r5,16 cmpdi cr6,r5,4096 Index: linux-build/arch/powerpc/lib/copyuser_power7.S =================================================================== --- /dev/null 1970-01-01 00:00:00.000000000 +0000 +++ linux-build/arch/powerpc/lib/copyuser_power7.S 2011-12-08 15:38:44.655277808 +1100 @@ -0,0 +1,669 @@ +/* + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. + * + * Copyright (C) IBM Corporation, 2011 + * + * Author: Anton Blanchard <anton@au.ibm.com> + */ +#include <asm/ppc_asm.h> + +#define STACKFRAMESIZE 256 +#define STK_REG(i) (112 + ((i)-14)*8) + + .macro err1 +100: + .section __ex_table,"a" + .align 3 + .llong 100b,.Ldo_err1 + .previous + .endm + + .macro err2 +200: + .section __ex_table,"a" + .align 3 + .llong 200b,.Ldo_err2 + .previous + .endm + + .macro err3 +300: + .section __ex_table,"a" + .align 3 + .llong 300b,.Ldo_err3 + .previous + .endm + + .macro err4 +400: + .section __ex_table,"a" + .align 3 + .llong 400b,.Ldo_err4 + .previous + .endm + + +.Ldo_err4: + ld r16,STK_REG(r16)(r1) + ld r15,STK_REG(r15)(r1) + ld r14,STK_REG(r14)(r1) +.Ldo_err3: + bl .exit_vmx_copy + ld r0,STACKFRAMESIZE+16(r1) + mtlr r0 + b .Lexit + +.Ldo_err2: + ld r22,STK_REG(r22)(r1) + ld r21,STK_REG(r21)(r1) + ld r20,STK_REG(r20)(r1) + ld r19,STK_REG(r19)(r1) + ld r18,STK_REG(r18)(r1) + ld r17,STK_REG(r17)(r1) + ld r16,STK_REG(r16)(r1) + ld r15,STK_REG(r15)(r1) + ld r14,STK_REG(r14)(r1) +.Lexit: + addi r1,r1,STACKFRAMESIZE +.Ldo_err1: + ld r3,48(r1) + ld r4,56(r1) + ld r5,64(r1) + b __copy_tofrom_user_base + + +_GLOBAL(__copy_tofrom_user_power7) + cmpldi r5,16 + cmpldi cr1,r5,4096 + + std r3,48(r1) + std r4,56(r1) + std r5,64(r1) + + blt .Lshort_copy + bgt cr1,.Lvmx_copy + +.Lnonvmx_copy: + /* Get the source 8B aligned */ + neg r6,r4 + mtocrf 0x01,r6 + clrldi r6,r6,(64-3) + + bf cr7*4+3,1f +err1; lbz r0,0(r4) + addi r4,r4,1 +err1; stb r0,0(r3) + addi r3,r3,1 + +1: bf cr7*4+2,2f +err1; lhz r0,0(r4) + addi r4,r4,2 +err1; sth r0,0(r3) + addi r3,r3,2 + +2: bf cr7*4+1,3f +err1; lwz r0,0(r4) + addi r4,r4,4 +err1; stw r0,0(r3) + addi r3,r3,4 + +3: sub r5,r5,r6 + cmpldi r5,128 + blt 5f + + mflr r0 + stdu r1,-STACKFRAMESIZE(r1) + std r14,STK_REG(r14)(r1) + std r15,STK_REG(r15)(r1) + std r16,STK_REG(r16)(r1) + std r17,STK_REG(r17)(r1) + std r18,STK_REG(r18)(r1) + std r19,STK_REG(r19)(r1) + std r20,STK_REG(r20)(r1) + std r21,STK_REG(r21)(r1) + std r22,STK_REG(r22)(r1) + std r0,STACKFRAMESIZE+16(r1) + + srdi r6,r5,7 + mtctr r6 + + /* Now do cacheline (128B) sized loads and stores. */ + .align 5 +4: +err2; ld r0,0(r4) +err2; ld r6,8(r4) +err2; ld r7,16(r4) +err2; ld r8,24(r4) +err2; ld r9,32(r4) +err2; ld r10,40(r4) +err2; ld r11,48(r4) +err2; ld r12,56(r4) +err2; ld r14,64(r4) +err2; ld r15,72(r4) +err2; ld r16,80(r4) +err2; ld r17,88(r4) +err2; ld r18,96(r4) +err2; ld r19,104(r4) +err2; ld r20,112(r4) +err2; ld r21,120(r4) + addi r4,r4,128 +err2; std r0,0(r3) +err2; std r6,8(r3) +err2; std r7,16(r3) +err2; std r8,24(r3) +err2; std r9,32(r3) +err2; std r10,40(r3) +err2; std r11,48(r3) +err2; std r12,56(r3) +err2; std r14,64(r3) +err2; std r15,72(r3) +err2; std r16,80(r3) +err2; std r17,88(r3) +err2; std r18,96(r3) +err2; std r19,104(r3) +err2; std r20,112(r3) +err2; std r21,120(r3) + addi r3,r3,128 + bdnz 4b + + clrldi r5,r5,(64-7) + + ld r14,STK_REG(r14)(r1) + ld r15,STK_REG(r15)(r1) + ld r16,STK_REG(r16)(r1) + ld r17,STK_REG(r17)(r1) + ld r18,STK_REG(r18)(r1) + ld r19,STK_REG(r19)(r1) + ld r20,STK_REG(r20)(r1) + ld r21,STK_REG(r21)(r1) + ld r22,STK_REG(r22)(r1) + addi r1,r1,STACKFRAMESIZE + + /* Up to 127B to go */ +5: srdi r6,r5,4 + mtocrf 0x01,r6 + +6: bf cr7*4+1,7f +err1; ld r0,0(r4) +err1; ld r6,8(r4) +err1; ld r7,16(r4) +err1; ld r8,24(r4) +err1; ld r9,32(r4) +err1; ld r10,40(r4) +err1; ld r11,48(r4) +err1; ld r12,56(r4) + addi r4,r4,64 +err1; std r0,0(r3) +err1; std r6,8(r3) +err1; std r7,16(r3) +err1; std r8,24(r3) +err1; std r9,32(r3) +err1; std r10,40(r3) +err1; std r11,48(r3) +err1; std r12,56(r3) + addi r3,r3,64 + + /* Up to 63B to go */ +7: bf cr7*4+2,8f +err1; ld r0,0(r4) +err1; ld r6,8(r4) +err1; ld r7,16(r4) +err1; ld r8,24(r4) + addi r4,r4,32 +err1; std r0,0(r3) +err1; std r6,8(r3) +err1; std r7,16(r3) +err1; std r8,24(r3) + addi r3,r3,32 + + /* Up to 31B to go */ +8: bf cr7*4+3,9f +err1; ld r0,0(r4) +err1; ld r6,8(r4) + addi r4,r4,16 +err1; std r0,0(r3) +err1; std r6,8(r3) + addi r3,r3,16 + +9: clrldi r5,r5,(64-4) + + /* Up to 15B to go */ +.Lshort_copy: + mtocrf 0x01,r5 + bf cr7*4+0,12f +err1; lwz r0,0(r4) /* Less chance of a reject with word ops */ +err1; lwz r6,4(r4) + addi r4,r4,8 +err1; stw r0,0(r3) +err1; stw r6,4(r3) + addi r3,r3,8 + +12: bf cr7*4+1,13f +err1; lwz r0,0(r4) + addi r4,r4,4 +err1; stw r0,0(r3) + addi r3,r3,4 + +13: bf cr7*4+2,14f +err1; lhz r0,0(r4) + addi r4,r4,2 +err1; sth r0,0(r3) + addi r3,r3,2 + +14: bf cr7*4+3,15f +err1; lbz r0,0(r4) +err1; stb r0,0(r3) + +15: li r3,0 + blr + +.Lunwind_stack_nonvmx_copy: + addi r1,r1,STACKFRAMESIZE + b .Lnonvmx_copy + +.Lvmx_copy: + mflr r0 + std r0,16(r1) + stdu r1,-STACKFRAMESIZE(r1) + bl .enter_vmx_copy + cmpwi r3,0 + ld r0,STACKFRAMESIZE+16(r1) + ld r3,STACKFRAMESIZE+48(r1) + ld r4,STACKFRAMESIZE+56(r1) + ld r5,STACKFRAMESIZE+64(r1) + mtlr r0 + + beq .Lunwind_stack_nonvmx_copy + + /* + * If source and destination are not relatively aligned we use a + * slower permute loop. + */ + xor r6,r4,r3 + rldicl. r6,r6,0,(64-4) + bne .Lvmx_unaligned_copy + + /* Get the destination 16B aligned */ + neg r6,r3 + mtocrf 0x01,r6 + clrldi r6,r6,(64-4) + + bf cr7*4+3,1f +err3; lbz r0,0(r4) + addi r4,r4,1 +err3; stb r0,0(r3) + addi r3,r3,1 + +1: bf cr7*4+2,2f +err3; lhz r0,0(r4) + addi r4,r4,2 +err3; sth r0,0(r3) + addi r3,r3,2 + +2: bf cr7*4+1,3f +err3; lwz r0,0(r4) + addi r4,r4,4 +err3; stw r0,0(r3) + addi r3,r3,4 + +3: bf cr7*4+0,4f +err3; ld r0,0(r4) + addi r4,r4,8 +err3; std r0,0(r3) + addi r3,r3,8 + +4: sub r5,r5,r6 + + /* Get the desination 128B aligned */ + neg r6,r3 + srdi r7,r6,4 + mtocrf 0x01,r7 + clrldi r6,r6,(64-7) + + li r9,16 + li r10,32 + li r11,48 + + bf cr7*4+3,5f +err3; lvx vr1,r0,r4 + addi r4,r4,16 +err3; stvx vr1,r0,r3 + addi r3,r3,16 + +5: bf cr7*4+2,6f +err3; lvx vr1,r0,r4 +err3; lvx vr0,r4,r9 + addi r4,r4,32 +err3; stvx vr1,r0,r3 +err3; stvx vr0,r3,r9 + addi r3,r3,32 + +6: bf cr7*4+1,7f +err3; lvx vr3,r0,r4 +err3; lvx vr2,r4,r9 +err3; lvx vr1,r4,r10 +err3; lvx vr0,r4,r11 + addi r4,r4,64 +err3; stvx vr3,r0,r3 +err3; stvx vr2,r3,r9 +err3; stvx vr1,r3,r10 +err3; stvx vr0,r3,r11 + addi r3,r3,64 + +7: sub r5,r5,r6 + srdi r6,r5,7 + + std r14,STK_REG(r14)(r1) + std r15,STK_REG(r15)(r1) + std r16,STK_REG(r16)(r1) + + li r12,64 + li r14,80 + li r15,96 + li r16,112 + + mtctr r6 + + /* + * Now do cacheline sized loads and stores. By this stage the + * cacheline stores are also cacheline aligned. + */ + .align 5 +8: +err4; lvx vr7,r0,r4 +err4; lvx vr6,r4,r9 +err4; lvx vr5,r4,r10 +err4; lvx vr4,r4,r11 +err4; lvx vr3,r4,r12 +err4; lvx vr2,r4,r14 +err4; lvx vr1,r4,r15 +err4; lvx vr0,r4,r16 + addi r4,r4,128 +err4; stvx vr7,r0,r3 +err4; stvx vr6,r3,r9 +err4; stvx vr5,r3,r10 +err4; stvx vr4,r3,r11 +err4; stvx vr3,r3,r12 +err4; stvx vr2,r3,r14 +err4; stvx vr1,r3,r15 +err4; stvx vr0,r3,r16 + addi r3,r3,128 + bdnz 8b + + ld r14,STK_REG(r14)(r1) + ld r15,STK_REG(r15)(r1) + ld r16,STK_REG(r16)(r1) + + /* Up to 127B to go */ + clrldi r5,r5,(64-7) + srdi r6,r5,4 + mtocrf 0x01,r6 + + bf cr7*4+1,9f +err3; lvx vr3,r0,r4 +err3; lvx vr2,r4,r9 +err3; lvx vr1,r4,r10 +err3; lvx vr0,r4,r11 + addi r4,r4,64 +err3; stvx vr3,r0,r3 +err3; stvx vr2,r3,r9 +err3; stvx vr1,r3,r10 +err3; stvx vr0,r3,r11 + addi r3,r3,64 + +9: bf cr7*4+2,10f +err3; lvx vr1,r0,r4 +err3; lvx vr0,r4,r9 + addi r4,r4,32 +err3; stvx vr1,r0,r3 +err3; stvx vr0,r3,r9 + addi r3,r3,32 + +10: bf cr7*4+3,11f +err3; lvx vr1,r0,r4 + addi r4,r4,16 +err3; stvx vr1,r0,r3 + addi r3,r3,16 + + /* Up to 15B to go */ +11: clrldi r5,r5,(64-4) + mtocrf 0x01,r5 + bf cr7*4+0,12f +err3; ld r0,0(r4) + addi r4,r4,8 +err3; std r0,0(r3) + addi r3,r3,8 + +12: bf cr7*4+1,13f +err3; lwz r0,0(r4) + addi r4,r4,4 +err3; stw r0,0(r3) + addi r3,r3,4 + +13: bf cr7*4+2,14f +err3; lhz r0,0(r4) + addi r4,r4,2 +err3; sth r0,0(r3) + addi r3,r3,2 + +14: bf cr7*4+3,15f +err3; lbz r0,0(r4) +err3; stb r0,0(r3) + +15: addi r1,r1,STACKFRAMESIZE + b .exit_vmx_copy /* tail call optimise */ + +.Lvmx_unaligned_copy: + /* Get the destination 16B aligned */ + neg r6,r3 + mtocrf 0x01,r6 + clrldi r6,r6,(64-4) + + bf cr7*4+3,1f +err3; lbz r0,0(r4) + addi r4,r4,1 +err3; stb r0,0(r3) + addi r3,r3,1 + +1: bf cr7*4+2,2f +err3; lhz r0,0(r4) + addi r4,r4,2 +err3; sth r0,0(r3) + addi r3,r3,2 + +2: bf cr7*4+1,3f +err3; lwz r0,0(r4) + addi r4,r4,4 +err3; stw r0,0(r3) + addi r3,r3,4 + +3: bf cr7*4+0,4f +err3; lwz r0,0(r4) /* Less chance of a reject with word ops */ +err3; lwz r7,4(r4) + addi r4,r4,8 +err3; stw r0,0(r3) +err3; stw r7,4(r3) + addi r3,r3,8 + +4: sub r5,r5,r6 + + /* Get the desination 128B aligned */ + neg r6,r3 + srdi r7,r6,4 + mtocrf 0x01,r7 + clrldi r6,r6,(64-7) + + li r9,16 + li r10,32 + li r11,48 + + lvsl vr16,0,r4 /* Setup permute control vector */ +err3; lvx vr0,0,r4 + addi r4,r4,16 + + bf cr7*4+3,5f +err3; lvx vr1,r0,r4 + vperm vr8,vr0,vr1,vr16 + addi r4,r4,16 +err3; stvx vr8,r0,r3 + addi r3,r3,16 + vor vr0,vr1,vr1 + +5: bf cr7*4+2,6f +err3; lvx vr1,r0,r4 + vperm vr8,vr0,vr1,vr16 +err3; lvx vr0,r4,r9 + vperm vr9,vr1,vr0,vr16 + addi r4,r4,32 +err3; stvx vr8,r0,r3 +err3; stvx vr9,r3,r9 + addi r3,r3,32 + +6: bf cr7*4+1,7f +err3; lvx vr3,r0,r4 + vperm vr8,vr0,vr3,vr16 +err3; lvx vr2,r4,r9 + vperm vr9,vr3,vr2,vr16 +err3; lvx vr1,r4,r10 + vperm vr10,vr2,vr1,vr16 +err3; lvx vr0,r4,r11 + vperm vr11,vr1,vr0,vr16 + addi r4,r4,64 +err3; stvx vr8,r0,r3 +err3; stvx vr9,r3,r9 +err3; stvx vr10,r3,r10 +err3; stvx vr11,r3,r11 + addi r3,r3,64 + +7: sub r5,r5,r6 + srdi r6,r5,7 + + std r14,STK_REG(r14)(r1) + std r15,STK_REG(r15)(r1) + std r16,STK_REG(r16)(r1) + + li r12,64 + li r14,80 + li r15,96 + li r16,112 + + mtctr r6 + + /* + * Now do cacheline sized loads and stores. By this stage the + * cacheline stores are also cacheline aligned. + */ + .align 5 +8: +err4; lvx vr7,r0,r4 + vperm vr8,vr0,vr7,vr16 +err4; lvx vr6,r4,r9 + vperm vr9,vr7,vr6,vr16 +err4; lvx vr5,r4,r10 + vperm vr10,vr6,vr5,vr16 +err4; lvx vr4,r4,r11 + vperm vr11,vr5,vr4,vr16 +err4; lvx vr3,r4,r12 + vperm vr12,vr4,vr3,vr16 +err4; lvx vr2,r4,r14 + vperm vr13,vr3,vr2,vr16 +err4; lvx vr1,r4,r15 + vperm vr14,vr2,vr1,vr16 +err4; lvx vr0,r4,r16 + vperm vr15,vr1,vr0,vr16 + addi r4,r4,128 +err4; stvx vr8,r0,r3 +err4; stvx vr9,r3,r9 +err4; stvx vr10,r3,r10 +err4; stvx vr11,r3,r11 +err4; stvx vr12,r3,r12 +err4; stvx vr13,r3,r14 +err4; stvx vr14,r3,r15 +err4; stvx vr15,r3,r16 + addi r3,r3,128 + bdnz 8b + + ld r14,STK_REG(r14)(r1) + ld r15,STK_REG(r15)(r1) + ld r16,STK_REG(r16)(r1) + + /* Up to 127B to go */ + clrldi r5,r5,(64-7) + srdi r6,r5,4 + mtocrf 0x01,r6 + + bf cr7*4+1,9f +err3; lvx vr3,r0,r4 + vperm vr8,vr0,vr3,vr16 +err3; lvx vr2,r4,r9 + vperm vr9,vr3,vr2,vr16 +err3; lvx vr1,r4,r10 + vperm vr10,vr2,vr1,vr16 +err3; lvx vr0,r4,r11 + vperm vr11,vr1,vr0,vr16 + addi r4,r4,64 +err3; stvx vr8,r0,r3 +err3; stvx vr9,r3,r9 +err3; stvx vr10,r3,r10 +err3; stvx vr11,r3,r11 + addi r3,r3,64 + +9: bf cr7*4+2,10f +err3; lvx vr1,r0,r4 + vperm vr8,vr0,vr1,vr16 +err3; lvx vr0,r4,r9 + vperm vr9,vr1,vr0,vr16 + addi r4,r4,32 +err3; stvx vr8,r0,r3 +err3; stvx vr9,r3,r9 + addi r3,r3,32 + +10: bf cr7*4+3,11f +err3; lvx vr1,r0,r4 + vperm vr8,vr0,vr1,vr16 + addi r4,r4,16 +err3; stvx vr8,r0,r3 + addi r3,r3,16 + + /* Up to 15B to go */ +11: clrldi r5,r5,(64-4) + addi r4,r4,-16 /* Unwind the +16 load offset */ + mtocrf 0x01,r5 + bf cr7*4+0,12f +err3; lwz r0,0(r4) /* Less chance of a reject with word ops */ +err3; lwz r6,4(r4) + addi r4,r4,8 +err3; stw r0,0(r3) +err3; stw r6,4(r3) + addi r3,r3,8 + +12: bf cr7*4+1,13f +err3; lwz r0,0(r4) + addi r4,r4,4 +err3; stw r0,0(r3) + addi r3,r3,4 + +13: bf cr7*4+2,14f +err3; lhz r0,0(r4) + addi r4,r4,2 +err3; sth r0,0(r3) + addi r3,r3,2 + +14: bf cr7*4+3,15f +err3; lbz r0,0(r4) +err3; stb r0,0(r3) + +15: addi r1,r1,STACKFRAMESIZE + b .exit_vmx_copy /* tail call optimise */ Index: linux-build/arch/powerpc/lib/Makefile =================================================================== --- linux-build.orig/arch/powerpc/lib/Makefile 2011-09-07 15:15:49.146459439 +1000 +++ linux-build/arch/powerpc/lib/Makefile 2011-12-08 15:17:29.000000000 +1100 @@ -16,7 +16,8 @@ obj-$(CONFIG_HAS_IOMEM) += devres.o obj-$(CONFIG_PPC64) += copypage_64.o copyuser_64.o \ memcpy_64.o usercopy_64.o mem_64.o string.o \ - checksum_wrappers_64.o hweight_64.o + checksum_wrappers_64.o hweight_64.o \ + copyuser_power7.o copyuser_power7_vmx.o obj-$(CONFIG_XMON) += sstep.o ldstfp.o obj-$(CONFIG_KPROBES) += sstep.o ldstfp.o obj-$(CONFIG_HAVE_HW_BREAKPOINT) += sstep.o ldstfp.o Index: linux-build/arch/powerpc/lib/copyuser_power7_vmx.c =================================================================== --- /dev/null 1970-01-01 00:00:00.000000000 +0000 +++ linux-build/arch/powerpc/lib/copyuser_power7_vmx.c 2011-12-08 15:38:44.639277518 +1100 @@ -0,0 +1,44 @@ +/* + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. + * + * Copyright (C) IBM Corporation, 2011 + * + * Authors: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com> + * Anton Blanchard <anton@au.ibm.com> + */ +#include <linux/uaccess.h> +#include <linux/hardirq.h> + +int enter_vmx_copy(void) +{ + if (in_interrupt()) + return 0; + + enable_kernel_altivec(); + + pagefault_disable(); + + return 1; +} + +/* + * This function must return 0 because we tail call optimise when calling + * from __copy_tofrom_user_power7 which returns 0 on success. + */ +int exit_vmx_copy(void) +{ + pagefault_enable(); + return 0; +} ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-08 5:02 [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX Anton Blanchard @ 2011-12-08 5:44 ` Kumar Gala 2011-12-08 5:54 ` Michael Neuling 0 siblings, 1 reply; 11+ messages in thread From: Kumar Gala @ 2011-12-08 5:44 UTC (permalink / raw) To: Anton Blanchard; +Cc: mikey, paulus, sukadev, linuxppc-dev On Dec 7, 2011, at 11:02 PM, Anton Blanchard wrote: > Index: linux-build/arch/powerpc/include/asm/cputable.h > =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D > --- linux-build.orig/arch/powerpc/include/asm/cputable.h = 2011-09-07 15:15:49.096458526 +1000 > +++ linux-build/arch/powerpc/include/asm/cputable.h 2011-12-08 = 15:38:46.627313507 +1100 > @@ -201,6 +201,7 @@ extern const char *powerpc_base_platform > #define CPU_FTR_POPCNTB = LONG_ASM_CONST(0x0400000000000000) > #define CPU_FTR_POPCNTD = LONG_ASM_CONST(0x0800000000000000) > #define CPU_FTR_ICSWX = LONG_ASM_CONST(0x1000000000000000) > +#define CPU_FTR_POWER7 = LONG_ASM_CONST(0x2000000000000000) >=20 > #ifndef __ASSEMBLY__ >=20 > @@ -425,7 +426,7 @@ extern const char *powerpc_base_platform > CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \ > CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYM_SMT | \ > CPU_FTR_STCX_CHECKS_ADDRESS | CPU_FTR_POPCNTB | = CPU_FTR_POPCNTD | \ > - CPU_FTR_ICSWX | CPU_FTR_CFAR | CPU_FTR_HVMODE) > + CPU_FTR_ICSWX | CPU_FTR_CFAR | CPU_FTR_HVMODE | = CPU_FTR_POWER7) > #define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \ > CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \ > CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \ > Index: linux-build/arch/powerpc/lib/copyuser_64.S > =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D > --- linux-build.orig/arch/powerpc/lib/copyuser_64.S 2011-09-07 = 15:15:49.146459439 +1000 > +++ linux-build/arch/powerpc/lib/copyuser_64.S 2011-12-08 = 15:38:42.491238635 +1100 > @@ -11,6 +11,12 @@ >=20 > .align 7 > _GLOBAL(__copy_tofrom_user) > +BEGIN_FTR_SECTION > + nop > +FTR_SECTION_ELSE > + b __copy_tofrom_user_power7 > +ALT_FTR_SECTION_END_IFCLR(CPU_FTR_POWER7) > +_GLOBAL(__copy_tofrom_user_base) > /* first check for a whole page copy on a page boundary */ > cmpldi cr1,r5,16 > cmpdi cr6,r5,4096 Can we find a means to do the fixup that does NOT require a FTR bit? I = have the feeling FSL will want to have various optimized copy functions = for our different cores and I hate to blow features bits just for this. - k= ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-08 5:44 ` Kumar Gala @ 2011-12-08 5:54 ` Michael Neuling 2011-12-08 6:04 ` Anton Blanchard 2011-12-08 6:11 ` Anton Blanchard 0 siblings, 2 replies; 11+ messages in thread From: Michael Neuling @ 2011-12-08 5:54 UTC (permalink / raw) To: Kumar Gala; +Cc: paulus, Anton Blanchard, sukadev, linuxppc-dev <snip> > > +#define CPU_FTR_POWER7 = LONG_ASM_CONST(0x2000000000000000) <snip> > Can we find a means to do the fixup that does NOT require a FTR bit? I > have the feeling FSL will want to have various optimized copy functions > for our different cores and I hate to blow features bits just for this. +1 I hate the idea of having a POWER7 FTR bit. Every loon will (and has tried to in the past) attach every POWER7 related thing to it, rather than thinking about what the feature really is for. What about other processors which could also benefit from this copy loop? Turning on CPU_FTR_POWER7 for them is gonna look a bit silly. Mikey ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-08 5:54 ` Michael Neuling @ 2011-12-08 6:04 ` Anton Blanchard 2011-12-08 13:31 ` Segher Boessenkool 2011-12-08 19:40 ` Benjamin Herrenschmidt 2011-12-08 6:11 ` Anton Blanchard 1 sibling, 2 replies; 11+ messages in thread From: Anton Blanchard @ 2011-12-08 6:04 UTC (permalink / raw) To: Michael Neuling; +Cc: paulus, sukadev, linuxppc-dev Hi, > I hate the idea of having a POWER7 FTR bit. Every loon will (and has > tried to in the past) attach every POWER7 related thing to it, rather > than thinking about what the feature really is for. > > What about other processors which could also benefit from this copy > loop? Turning on CPU_FTR_POWER7 for them is gonna look a bit silly. As we discussed online, we could call it CPU_FTR_VMX_COPY and start thinking about a better way to solve the CPU feature bit mess. One idea would be to have a structure of function pointers for each CPU that gets runtime patched into the right places, similar to how we do some of the MMU fixups. Anton ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-08 6:04 ` Anton Blanchard @ 2011-12-08 13:31 ` Segher Boessenkool 2011-12-08 14:02 ` David Laight 2011-12-08 19:40 ` Benjamin Herrenschmidt 1 sibling, 1 reply; 11+ messages in thread From: Segher Boessenkool @ 2011-12-08 13:31 UTC (permalink / raw) To: Anton Blanchard; +Cc: linuxppc-dev, Michael Neuling, sukadev, paulus >> I hate the idea of having a POWER7 FTR bit. Every loon will (and has >> tried to in the past) attach every POWER7 related thing to it, rather >> than thinking about what the feature really is for. >> >> What about other processors which could also benefit from this copy >> loop? Turning on CPU_FTR_POWER7 for them is gonna look a bit silly. > > As we discussed online, we could call it CPU_FTR_VMX_COPY and start > thinking about a better way to solve the CPU feature bit mess. But then, most CPUs with VMX will not want that, because it is slower code for them. For things like copy loops it makes perfect sense to have them tuned per CPU core. For example, this code likes to use unaligned stores over more complicated shift-and-combine stuff; that works great on POWER7, but not on much else. Maybe you should have the various kinds of loop ("source aligned, dest unaligned, using unaligned stores, 64 bytes") as asm routines, have some higher level code (which can be runtime patched) select which to run. > One idea would be to have a structure of function pointers for each > CPU that gets runtime patched into the right places, similar to how we > do some of the MMU fixups. Sounds good to me :-) Segher ^ permalink raw reply [flat|nested] 11+ messages in thread
* RE: [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-08 13:31 ` Segher Boessenkool @ 2011-12-08 14:02 ` David Laight 2011-12-08 14:12 ` Segher Boessenkool 0 siblings, 1 reply; 11+ messages in thread From: David Laight @ 2011-12-08 14:02 UTC (permalink / raw) To: Segher Boessenkool, Anton Blanchard Cc: paulus, Michael Neuling, sukadev, linuxppc-dev =20 > > One idea would be to have a structure of function pointers for each > > CPU that gets runtime patched into the right places,=20 > > similar to how we do some of the MMU fixups. >=20 > Sounds good to me :-) Except the indirect jump/call is almost certainly never predicted - so will be slow. You might want to patch jump instructions instead. The same is true for in-kernel memcpy() and other similar operations. I actually wonder sometimes what the typical lengths are for these sort of functions, and whether, in fact, small lengths dominate - where the fixed costs matter. David ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-08 14:02 ` David Laight @ 2011-12-08 14:12 ` Segher Boessenkool 0 siblings, 0 replies; 11+ messages in thread From: Segher Boessenkool @ 2011-12-08 14:12 UTC (permalink / raw) To: David Laight Cc: paulus, Michael Neuling, sukadev, linuxppc-dev, Anton Blanchard >>> One idea would be to have a structure of function pointers for each >>> CPU that gets runtime patched into the right places, >>> similar to how we do some of the MMU fixups. >> >> Sounds good to me :-) > > Except the indirect jump/call is almost certainly > never predicted - so will be slow. What indirect jump? He said "runtime patched" :-) Also, that "slow" is largely a myth. Segher ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-08 6:04 ` Anton Blanchard 2011-12-08 13:31 ` Segher Boessenkool @ 2011-12-08 19:40 ` Benjamin Herrenschmidt 1 sibling, 0 replies; 11+ messages in thread From: Benjamin Herrenschmidt @ 2011-12-08 19:40 UTC (permalink / raw) To: Anton Blanchard; +Cc: Michael Neuling, paulus, sukadev, linuxppc-dev On Thu, 2011-12-08 at 17:04 +1100, Anton Blanchard wrote: > Hi, > > > I hate the idea of having a POWER7 FTR bit. Every loon will (and has > > tried to in the past) attach every POWER7 related thing to it, rather > > than thinking about what the feature really is for. > > > > What about other processors which could also benefit from this copy > > loop? Turning on CPU_FTR_POWER7 for them is gonna look a bit silly. > > As we discussed online, we could call it CPU_FTR_VMX_COPY and start > thinking about a better way to solve the CPU feature bit mess. > > One idea would be to have a structure of function pointers for each > CPU that gets runtime patched into the right places, similar to how we > do some of the MMU fixups. Or the vdso... we have a table of some sort which is used to patch symbols. But it's keyed off cpu features. I'm reluctant to adding another table of PVRs, it needs to deal with the pseudo-PVRs from pHyp, and things will get out of sync. CPU features are the way to go, tho we can use them to key off a branch patching mechanism if we want to. It's easy to add new bitmasks for in-kernel features at least (it's the user features which are more nasty but they are a separate thing). So if we have to we can split cpu feature into another separate mask like I did for mmu features with the same macros etc... For example, the debug features could be moved out, or whatever. Cheers, Ben. ^ permalink raw reply [flat|nested] 11+ messages in thread
* [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-08 5:54 ` Michael Neuling 2011-12-08 6:04 ` Anton Blanchard @ 2011-12-08 6:11 ` Anton Blanchard 2011-12-19 3:00 ` Benjamin Herrenschmidt 1 sibling, 1 reply; 11+ messages in thread From: Anton Blanchard @ 2011-12-08 6:11 UTC (permalink / raw) To: Michael Neuling; +Cc: paulus, sukadev, linuxppc-dev Implement a POWER7 optimised copy_to_user/copy_from_user using VMX. For large aligned copies this new loop is over 10% faster, and for large unaligned copies it is over 200% faster. If we take a fault we fall back to the old version, this keeps things relatively simple and easy to verify. On POWER7 unaligned stores rarely slow down - they only flush when a store crosses a 4KB page boundary. Furthermore this flush is handled completely in hardware and should be 20-30 cycles. Unaligned loads on the other hand flush much more often - whenever crossing a 128 byte cache line, or a 32 byte sector if either sector is an L1 miss. Considering this information we really want to get the loads aligned and not worry about the alignment of the stores. Microbenchmarks confirm that this approach is much faster than the current unaligned copy loop that uses shifts and rotates to ensure both loads and stores are aligned. We also want to try and do the stores in cacheline aligned, cacheline sized chunks. If the store queue is unable to merge an entire cacheline of stores then the L2 cache will have to do a read/modify/write. Even worse, we will serialise this with the stores in the next iteration of the copy loop since both iterations hit the same cacheline. Based on this, the new loop does the following things: 1 - 127 bytes Get the source 8 byte aligned and use 8 byte loads and stores. Pretty boring and similar to how the current loop works. 128 - 4095 bytes Get the source 8 byte aligned and use 8 byte loads and stores, 1 cacheline at a time. We aren't doing the stores in cacheline aligned chunks so we will potentially serialise once per cacheline. Even so it is much better than the loop we have today. 4096 - bytes If both source and destination have the same alignment get them both 16 byte aligned, then get the destination cacheline aligned. Do cacheline sized loads and stores using VMX. If source and destination do not have the same alignment, we get the destination cacheline aligned, and use permute to do aligned loads. In both cases the VMX loop should be optimal - we always do aligned loads and stores and are always doing stores in cacheline aligned, cacheline sized chunks. To be able to use VMX we must be careful about interrupts and sleeping. We don't use the VMX loop when in an interrupt (which should be rare anyway) and we wrap the VMX loop in disable/enable_pagefault and fall back to the existing copy_tofrom_user loop if we do need to sleep. The VMX breakpoint of 4096 bytes was chosen using this microbenchmark: http://ozlabs.org/~anton/junkcode/copy_to_user.c Since we are using VMX and there is a cost to saving and restoring the user VMX state there are two broad cases we need to benchmark: - Best case - userspace never uses VMX - Worst case - userspace always uses VMX In reality a userspace process will sit somewhere between these two extremes. Since we need to test both aligned and unaligned copies we end up with 4 combinations. The point at which the VMX loop begins to win is: 0% VMX aligned 2048 bytes unaligned 2048 bytes 100% VMX aligned 16384 bytes unaligned 8192 bytes Considering this is a microbenchmark, the data is hot in cache and the VMX loop has better store queue merging properties we set the breakpoint to 4096 bytes, a little below the unaligned breakpoints. Some future optimisations we can look at: - Looking at the perf data, a significant part of the cost when a task is always using VMX is the extra exception we take to restore the VMX state. As such we should do something similar to the x86 optimisation that restores FPU state for heavy users. ie: /* * If the task has used fpu the last 5 timeslices, just do a full * restore of the math state immediately to avoid the trap; the * chances of needing FPU soon are obviously high now */ preload_fpu = tsk_used_math(next_p) && next_p->fpu_counter > 5; and /* * fpu_counter contains the number of consecutive context switches * that the FPU is used. If this is over a threshold, the lazy fpu * saving becomes unlazy to save the trap. This is an unsigned char * so that after 256 times the counter wraps and the behavior turns * lazy again; this to deal with bursty apps that only use FPU for * a short time */ - We could create a paca bit to mirror the VMX enabled MSR bit and check that first, avoiding multiple calls to calling enable_kernel_altivec. That should help with iovec based system calls like readv. - We could have two VMX breakpoints, one for when we know the user VMX state is loaded into the registers and one when it isn't. This could be a second bit in the paca so we can calculate the break points quickly. - One suggestion from Ben was to save and restore the VSX registers we use inline instead of using enable_kernel_altivec. Signed-off-by: Anton Blanchard <anton@samba.org> --- Index: linux-build/arch/powerpc/include/asm/cputable.h =================================================================== --- linux-build.orig/arch/powerpc/include/asm/cputable.h 2011-09-07 15:15:49.096458526 +1000 +++ linux-build/arch/powerpc/include/asm/cputable.h 2011-12-08 15:38:46.627313507 +1100 @@ -201,6 +201,7 @@ extern const char *powerpc_base_platform #define CPU_FTR_POPCNTB LONG_ASM_CONST(0x0400000000000000) #define CPU_FTR_POPCNTD LONG_ASM_CONST(0x0800000000000000) #define CPU_FTR_ICSWX LONG_ASM_CONST(0x1000000000000000) +#define CPU_FTR_VMX_COPY LONG_ASM_CONST(0x2000000000000000) #ifndef __ASSEMBLY__ @@ -425,7 +426,7 @@ extern const char *powerpc_base_platform CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \ CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYM_SMT | \ CPU_FTR_STCX_CHECKS_ADDRESS | CPU_FTR_POPCNTB | CPU_FTR_POPCNTD | \ - CPU_FTR_ICSWX | CPU_FTR_CFAR | CPU_FTR_HVMODE) + CPU_FTR_ICSWX | CPU_FTR_CFAR | CPU_FTR_HVMODE | CPU_FTR_VMX_COPY) #define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \ CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \ CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \ Index: linux-build/arch/powerpc/lib/copyuser_64.S =================================================================== --- linux-build.orig/arch/powerpc/lib/copyuser_64.S 2011-09-07 15:15:49.146459439 +1000 +++ linux-build/arch/powerpc/lib/copyuser_64.S 2011-12-08 15:38:42.491238635 +1100 @@ -11,6 +11,12 @@ .align 7 _GLOBAL(__copy_tofrom_user) +BEGIN_FTR_SECTION + nop +FTR_SECTION_ELSE + b __copy_tofrom_user_power7 +ALT_FTR_SECTION_END_IFCLR(CPU_FTR_VMX_COPY) +_GLOBAL(__copy_tofrom_user_base) /* first check for a whole page copy on a page boundary */ cmpldi cr1,r5,16 cmpdi cr6,r5,4096 Index: linux-build/arch/powerpc/lib/copyuser_power7.S =================================================================== --- /dev/null 1970-01-01 00:00:00.000000000 +0000 +++ linux-build/arch/powerpc/lib/copyuser_power7.S 2011-12-08 15:38:44.655277808 +1100 @@ -0,0 +1,669 @@ +/* + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. + * + * Copyright (C) IBM Corporation, 2011 + * + * Author: Anton Blanchard <anton@au.ibm.com> + */ +#include <asm/ppc_asm.h> + +#define STACKFRAMESIZE 256 +#define STK_REG(i) (112 + ((i)-14)*8) + + .macro err1 +100: + .section __ex_table,"a" + .align 3 + .llong 100b,.Ldo_err1 + .previous + .endm + + .macro err2 +200: + .section __ex_table,"a" + .align 3 + .llong 200b,.Ldo_err2 + .previous + .endm + + .macro err3 +300: + .section __ex_table,"a" + .align 3 + .llong 300b,.Ldo_err3 + .previous + .endm + + .macro err4 +400: + .section __ex_table,"a" + .align 3 + .llong 400b,.Ldo_err4 + .previous + .endm + + +.Ldo_err4: + ld r16,STK_REG(r16)(r1) + ld r15,STK_REG(r15)(r1) + ld r14,STK_REG(r14)(r1) +.Ldo_err3: + bl .exit_vmx_copy + ld r0,STACKFRAMESIZE+16(r1) + mtlr r0 + b .Lexit + +.Ldo_err2: + ld r22,STK_REG(r22)(r1) + ld r21,STK_REG(r21)(r1) + ld r20,STK_REG(r20)(r1) + ld r19,STK_REG(r19)(r1) + ld r18,STK_REG(r18)(r1) + ld r17,STK_REG(r17)(r1) + ld r16,STK_REG(r16)(r1) + ld r15,STK_REG(r15)(r1) + ld r14,STK_REG(r14)(r1) +.Lexit: + addi r1,r1,STACKFRAMESIZE +.Ldo_err1: + ld r3,48(r1) + ld r4,56(r1) + ld r5,64(r1) + b __copy_tofrom_user_base + + +_GLOBAL(__copy_tofrom_user_power7) + cmpldi r5,16 + cmpldi cr1,r5,4096 + + std r3,48(r1) + std r4,56(r1) + std r5,64(r1) + + blt .Lshort_copy + bgt cr1,.Lvmx_copy + +.Lnonvmx_copy: + /* Get the source 8B aligned */ + neg r6,r4 + mtocrf 0x01,r6 + clrldi r6,r6,(64-3) + + bf cr7*4+3,1f +err1; lbz r0,0(r4) + addi r4,r4,1 +err1; stb r0,0(r3) + addi r3,r3,1 + +1: bf cr7*4+2,2f +err1; lhz r0,0(r4) + addi r4,r4,2 +err1; sth r0,0(r3) + addi r3,r3,2 + +2: bf cr7*4+1,3f +err1; lwz r0,0(r4) + addi r4,r4,4 +err1; stw r0,0(r3) + addi r3,r3,4 + +3: sub r5,r5,r6 + cmpldi r5,128 + blt 5f + + mflr r0 + stdu r1,-STACKFRAMESIZE(r1) + std r14,STK_REG(r14)(r1) + std r15,STK_REG(r15)(r1) + std r16,STK_REG(r16)(r1) + std r17,STK_REG(r17)(r1) + std r18,STK_REG(r18)(r1) + std r19,STK_REG(r19)(r1) + std r20,STK_REG(r20)(r1) + std r21,STK_REG(r21)(r1) + std r22,STK_REG(r22)(r1) + std r0,STACKFRAMESIZE+16(r1) + + srdi r6,r5,7 + mtctr r6 + + /* Now do cacheline (128B) sized loads and stores. */ + .align 5 +4: +err2; ld r0,0(r4) +err2; ld r6,8(r4) +err2; ld r7,16(r4) +err2; ld r8,24(r4) +err2; ld r9,32(r4) +err2; ld r10,40(r4) +err2; ld r11,48(r4) +err2; ld r12,56(r4) +err2; ld r14,64(r4) +err2; ld r15,72(r4) +err2; ld r16,80(r4) +err2; ld r17,88(r4) +err2; ld r18,96(r4) +err2; ld r19,104(r4) +err2; ld r20,112(r4) +err2; ld r21,120(r4) + addi r4,r4,128 +err2; std r0,0(r3) +err2; std r6,8(r3) +err2; std r7,16(r3) +err2; std r8,24(r3) +err2; std r9,32(r3) +err2; std r10,40(r3) +err2; std r11,48(r3) +err2; std r12,56(r3) +err2; std r14,64(r3) +err2; std r15,72(r3) +err2; std r16,80(r3) +err2; std r17,88(r3) +err2; std r18,96(r3) +err2; std r19,104(r3) +err2; std r20,112(r3) +err2; std r21,120(r3) + addi r3,r3,128 + bdnz 4b + + clrldi r5,r5,(64-7) + + ld r14,STK_REG(r14)(r1) + ld r15,STK_REG(r15)(r1) + ld r16,STK_REG(r16)(r1) + ld r17,STK_REG(r17)(r1) + ld r18,STK_REG(r18)(r1) + ld r19,STK_REG(r19)(r1) + ld r20,STK_REG(r20)(r1) + ld r21,STK_REG(r21)(r1) + ld r22,STK_REG(r22)(r1) + addi r1,r1,STACKFRAMESIZE + + /* Up to 127B to go */ +5: srdi r6,r5,4 + mtocrf 0x01,r6 + +6: bf cr7*4+1,7f +err1; ld r0,0(r4) +err1; ld r6,8(r4) +err1; ld r7,16(r4) +err1; ld r8,24(r4) +err1; ld r9,32(r4) +err1; ld r10,40(r4) +err1; ld r11,48(r4) +err1; ld r12,56(r4) + addi r4,r4,64 +err1; std r0,0(r3) +err1; std r6,8(r3) +err1; std r7,16(r3) +err1; std r8,24(r3) +err1; std r9,32(r3) +err1; std r10,40(r3) +err1; std r11,48(r3) +err1; std r12,56(r3) + addi r3,r3,64 + + /* Up to 63B to go */ +7: bf cr7*4+2,8f +err1; ld r0,0(r4) +err1; ld r6,8(r4) +err1; ld r7,16(r4) +err1; ld r8,24(r4) + addi r4,r4,32 +err1; std r0,0(r3) +err1; std r6,8(r3) +err1; std r7,16(r3) +err1; std r8,24(r3) + addi r3,r3,32 + + /* Up to 31B to go */ +8: bf cr7*4+3,9f +err1; ld r0,0(r4) +err1; ld r6,8(r4) + addi r4,r4,16 +err1; std r0,0(r3) +err1; std r6,8(r3) + addi r3,r3,16 + +9: clrldi r5,r5,(64-4) + + /* Up to 15B to go */ +.Lshort_copy: + mtocrf 0x01,r5 + bf cr7*4+0,12f +err1; lwz r0,0(r4) /* Less chance of a reject with word ops */ +err1; lwz r6,4(r4) + addi r4,r4,8 +err1; stw r0,0(r3) +err1; stw r6,4(r3) + addi r3,r3,8 + +12: bf cr7*4+1,13f +err1; lwz r0,0(r4) + addi r4,r4,4 +err1; stw r0,0(r3) + addi r3,r3,4 + +13: bf cr7*4+2,14f +err1; lhz r0,0(r4) + addi r4,r4,2 +err1; sth r0,0(r3) + addi r3,r3,2 + +14: bf cr7*4+3,15f +err1; lbz r0,0(r4) +err1; stb r0,0(r3) + +15: li r3,0 + blr + +.Lunwind_stack_nonvmx_copy: + addi r1,r1,STACKFRAMESIZE + b .Lnonvmx_copy + +.Lvmx_copy: + mflr r0 + std r0,16(r1) + stdu r1,-STACKFRAMESIZE(r1) + bl .enter_vmx_copy + cmpwi r3,0 + ld r0,STACKFRAMESIZE+16(r1) + ld r3,STACKFRAMESIZE+48(r1) + ld r4,STACKFRAMESIZE+56(r1) + ld r5,STACKFRAMESIZE+64(r1) + mtlr r0 + + beq .Lunwind_stack_nonvmx_copy + + /* + * If source and destination are not relatively aligned we use a + * slower permute loop. + */ + xor r6,r4,r3 + rldicl. r6,r6,0,(64-4) + bne .Lvmx_unaligned_copy + + /* Get the destination 16B aligned */ + neg r6,r3 + mtocrf 0x01,r6 + clrldi r6,r6,(64-4) + + bf cr7*4+3,1f +err3; lbz r0,0(r4) + addi r4,r4,1 +err3; stb r0,0(r3) + addi r3,r3,1 + +1: bf cr7*4+2,2f +err3; lhz r0,0(r4) + addi r4,r4,2 +err3; sth r0,0(r3) + addi r3,r3,2 + +2: bf cr7*4+1,3f +err3; lwz r0,0(r4) + addi r4,r4,4 +err3; stw r0,0(r3) + addi r3,r3,4 + +3: bf cr7*4+0,4f +err3; ld r0,0(r4) + addi r4,r4,8 +err3; std r0,0(r3) + addi r3,r3,8 + +4: sub r5,r5,r6 + + /* Get the desination 128B aligned */ + neg r6,r3 + srdi r7,r6,4 + mtocrf 0x01,r7 + clrldi r6,r6,(64-7) + + li r9,16 + li r10,32 + li r11,48 + + bf cr7*4+3,5f +err3; lvx vr1,r0,r4 + addi r4,r4,16 +err3; stvx vr1,r0,r3 + addi r3,r3,16 + +5: bf cr7*4+2,6f +err3; lvx vr1,r0,r4 +err3; lvx vr0,r4,r9 + addi r4,r4,32 +err3; stvx vr1,r0,r3 +err3; stvx vr0,r3,r9 + addi r3,r3,32 + +6: bf cr7*4+1,7f +err3; lvx vr3,r0,r4 +err3; lvx vr2,r4,r9 +err3; lvx vr1,r4,r10 +err3; lvx vr0,r4,r11 + addi r4,r4,64 +err3; stvx vr3,r0,r3 +err3; stvx vr2,r3,r9 +err3; stvx vr1,r3,r10 +err3; stvx vr0,r3,r11 + addi r3,r3,64 + +7: sub r5,r5,r6 + srdi r6,r5,7 + + std r14,STK_REG(r14)(r1) + std r15,STK_REG(r15)(r1) + std r16,STK_REG(r16)(r1) + + li r12,64 + li r14,80 + li r15,96 + li r16,112 + + mtctr r6 + + /* + * Now do cacheline sized loads and stores. By this stage the + * cacheline stores are also cacheline aligned. + */ + .align 5 +8: +err4; lvx vr7,r0,r4 +err4; lvx vr6,r4,r9 +err4; lvx vr5,r4,r10 +err4; lvx vr4,r4,r11 +err4; lvx vr3,r4,r12 +err4; lvx vr2,r4,r14 +err4; lvx vr1,r4,r15 +err4; lvx vr0,r4,r16 + addi r4,r4,128 +err4; stvx vr7,r0,r3 +err4; stvx vr6,r3,r9 +err4; stvx vr5,r3,r10 +err4; stvx vr4,r3,r11 +err4; stvx vr3,r3,r12 +err4; stvx vr2,r3,r14 +err4; stvx vr1,r3,r15 +err4; stvx vr0,r3,r16 + addi r3,r3,128 + bdnz 8b + + ld r14,STK_REG(r14)(r1) + ld r15,STK_REG(r15)(r1) + ld r16,STK_REG(r16)(r1) + + /* Up to 127B to go */ + clrldi r5,r5,(64-7) + srdi r6,r5,4 + mtocrf 0x01,r6 + + bf cr7*4+1,9f +err3; lvx vr3,r0,r4 +err3; lvx vr2,r4,r9 +err3; lvx vr1,r4,r10 +err3; lvx vr0,r4,r11 + addi r4,r4,64 +err3; stvx vr3,r0,r3 +err3; stvx vr2,r3,r9 +err3; stvx vr1,r3,r10 +err3; stvx vr0,r3,r11 + addi r3,r3,64 + +9: bf cr7*4+2,10f +err3; lvx vr1,r0,r4 +err3; lvx vr0,r4,r9 + addi r4,r4,32 +err3; stvx vr1,r0,r3 +err3; stvx vr0,r3,r9 + addi r3,r3,32 + +10: bf cr7*4+3,11f +err3; lvx vr1,r0,r4 + addi r4,r4,16 +err3; stvx vr1,r0,r3 + addi r3,r3,16 + + /* Up to 15B to go */ +11: clrldi r5,r5,(64-4) + mtocrf 0x01,r5 + bf cr7*4+0,12f +err3; ld r0,0(r4) + addi r4,r4,8 +err3; std r0,0(r3) + addi r3,r3,8 + +12: bf cr7*4+1,13f +err3; lwz r0,0(r4) + addi r4,r4,4 +err3; stw r0,0(r3) + addi r3,r3,4 + +13: bf cr7*4+2,14f +err3; lhz r0,0(r4) + addi r4,r4,2 +err3; sth r0,0(r3) + addi r3,r3,2 + +14: bf cr7*4+3,15f +err3; lbz r0,0(r4) +err3; stb r0,0(r3) + +15: addi r1,r1,STACKFRAMESIZE + b .exit_vmx_copy /* tail call optimise */ + +.Lvmx_unaligned_copy: + /* Get the destination 16B aligned */ + neg r6,r3 + mtocrf 0x01,r6 + clrldi r6,r6,(64-4) + + bf cr7*4+3,1f +err3; lbz r0,0(r4) + addi r4,r4,1 +err3; stb r0,0(r3) + addi r3,r3,1 + +1: bf cr7*4+2,2f +err3; lhz r0,0(r4) + addi r4,r4,2 +err3; sth r0,0(r3) + addi r3,r3,2 + +2: bf cr7*4+1,3f +err3; lwz r0,0(r4) + addi r4,r4,4 +err3; stw r0,0(r3) + addi r3,r3,4 + +3: bf cr7*4+0,4f +err3; lwz r0,0(r4) /* Less chance of a reject with word ops */ +err3; lwz r7,4(r4) + addi r4,r4,8 +err3; stw r0,0(r3) +err3; stw r7,4(r3) + addi r3,r3,8 + +4: sub r5,r5,r6 + + /* Get the desination 128B aligned */ + neg r6,r3 + srdi r7,r6,4 + mtocrf 0x01,r7 + clrldi r6,r6,(64-7) + + li r9,16 + li r10,32 + li r11,48 + + lvsl vr16,0,r4 /* Setup permute control vector */ +err3; lvx vr0,0,r4 + addi r4,r4,16 + + bf cr7*4+3,5f +err3; lvx vr1,r0,r4 + vperm vr8,vr0,vr1,vr16 + addi r4,r4,16 +err3; stvx vr8,r0,r3 + addi r3,r3,16 + vor vr0,vr1,vr1 + +5: bf cr7*4+2,6f +err3; lvx vr1,r0,r4 + vperm vr8,vr0,vr1,vr16 +err3; lvx vr0,r4,r9 + vperm vr9,vr1,vr0,vr16 + addi r4,r4,32 +err3; stvx vr8,r0,r3 +err3; stvx vr9,r3,r9 + addi r3,r3,32 + +6: bf cr7*4+1,7f +err3; lvx vr3,r0,r4 + vperm vr8,vr0,vr3,vr16 +err3; lvx vr2,r4,r9 + vperm vr9,vr3,vr2,vr16 +err3; lvx vr1,r4,r10 + vperm vr10,vr2,vr1,vr16 +err3; lvx vr0,r4,r11 + vperm vr11,vr1,vr0,vr16 + addi r4,r4,64 +err3; stvx vr8,r0,r3 +err3; stvx vr9,r3,r9 +err3; stvx vr10,r3,r10 +err3; stvx vr11,r3,r11 + addi r3,r3,64 + +7: sub r5,r5,r6 + srdi r6,r5,7 + + std r14,STK_REG(r14)(r1) + std r15,STK_REG(r15)(r1) + std r16,STK_REG(r16)(r1) + + li r12,64 + li r14,80 + li r15,96 + li r16,112 + + mtctr r6 + + /* + * Now do cacheline sized loads and stores. By this stage the + * cacheline stores are also cacheline aligned. + */ + .align 5 +8: +err4; lvx vr7,r0,r4 + vperm vr8,vr0,vr7,vr16 +err4; lvx vr6,r4,r9 + vperm vr9,vr7,vr6,vr16 +err4; lvx vr5,r4,r10 + vperm vr10,vr6,vr5,vr16 +err4; lvx vr4,r4,r11 + vperm vr11,vr5,vr4,vr16 +err4; lvx vr3,r4,r12 + vperm vr12,vr4,vr3,vr16 +err4; lvx vr2,r4,r14 + vperm vr13,vr3,vr2,vr16 +err4; lvx vr1,r4,r15 + vperm vr14,vr2,vr1,vr16 +err4; lvx vr0,r4,r16 + vperm vr15,vr1,vr0,vr16 + addi r4,r4,128 +err4; stvx vr8,r0,r3 +err4; stvx vr9,r3,r9 +err4; stvx vr10,r3,r10 +err4; stvx vr11,r3,r11 +err4; stvx vr12,r3,r12 +err4; stvx vr13,r3,r14 +err4; stvx vr14,r3,r15 +err4; stvx vr15,r3,r16 + addi r3,r3,128 + bdnz 8b + + ld r14,STK_REG(r14)(r1) + ld r15,STK_REG(r15)(r1) + ld r16,STK_REG(r16)(r1) + + /* Up to 127B to go */ + clrldi r5,r5,(64-7) + srdi r6,r5,4 + mtocrf 0x01,r6 + + bf cr7*4+1,9f +err3; lvx vr3,r0,r4 + vperm vr8,vr0,vr3,vr16 +err3; lvx vr2,r4,r9 + vperm vr9,vr3,vr2,vr16 +err3; lvx vr1,r4,r10 + vperm vr10,vr2,vr1,vr16 +err3; lvx vr0,r4,r11 + vperm vr11,vr1,vr0,vr16 + addi r4,r4,64 +err3; stvx vr8,r0,r3 +err3; stvx vr9,r3,r9 +err3; stvx vr10,r3,r10 +err3; stvx vr11,r3,r11 + addi r3,r3,64 + +9: bf cr7*4+2,10f +err3; lvx vr1,r0,r4 + vperm vr8,vr0,vr1,vr16 +err3; lvx vr0,r4,r9 + vperm vr9,vr1,vr0,vr16 + addi r4,r4,32 +err3; stvx vr8,r0,r3 +err3; stvx vr9,r3,r9 + addi r3,r3,32 + +10: bf cr7*4+3,11f +err3; lvx vr1,r0,r4 + vperm vr8,vr0,vr1,vr16 + addi r4,r4,16 +err3; stvx vr8,r0,r3 + addi r3,r3,16 + + /* Up to 15B to go */ +11: clrldi r5,r5,(64-4) + addi r4,r4,-16 /* Unwind the +16 load offset */ + mtocrf 0x01,r5 + bf cr7*4+0,12f +err3; lwz r0,0(r4) /* Less chance of a reject with word ops */ +err3; lwz r6,4(r4) + addi r4,r4,8 +err3; stw r0,0(r3) +err3; stw r6,4(r3) + addi r3,r3,8 + +12: bf cr7*4+1,13f +err3; lwz r0,0(r4) + addi r4,r4,4 +err3; stw r0,0(r3) + addi r3,r3,4 + +13: bf cr7*4+2,14f +err3; lhz r0,0(r4) + addi r4,r4,2 +err3; sth r0,0(r3) + addi r3,r3,2 + +14: bf cr7*4+3,15f +err3; lbz r0,0(r4) +err3; stb r0,0(r3) + +15: addi r1,r1,STACKFRAMESIZE + b .exit_vmx_copy /* tail call optimise */ Index: linux-build/arch/powerpc/lib/Makefile =================================================================== --- linux-build.orig/arch/powerpc/lib/Makefile 2011-09-07 15:15:49.146459439 +1000 +++ linux-build/arch/powerpc/lib/Makefile 2011-12-08 15:17:29.000000000 +1100 @@ -16,7 +16,8 @@ obj-$(CONFIG_HAS_IOMEM) += devres.o obj-$(CONFIG_PPC64) += copypage_64.o copyuser_64.o \ memcpy_64.o usercopy_64.o mem_64.o string.o \ - checksum_wrappers_64.o hweight_64.o + checksum_wrappers_64.o hweight_64.o \ + copyuser_power7.o copyuser_power7_vmx.o obj-$(CONFIG_XMON) += sstep.o ldstfp.o obj-$(CONFIG_KPROBES) += sstep.o ldstfp.o obj-$(CONFIG_HAVE_HW_BREAKPOINT) += sstep.o ldstfp.o Index: linux-build/arch/powerpc/lib/copyuser_power7_vmx.c =================================================================== --- /dev/null 1970-01-01 00:00:00.000000000 +0000 +++ linux-build/arch/powerpc/lib/copyuser_power7_vmx.c 2011-12-08 15:38:44.639277518 +1100 @@ -0,0 +1,44 @@ +/* + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. + * + * Copyright (C) IBM Corporation, 2011 + * + * Authors: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com> + * Anton Blanchard <anton@au.ibm.com> + */ +#include <linux/uaccess.h> +#include <linux/hardirq.h> + +int enter_vmx_copy(void) +{ + if (in_interrupt()) + return 0; + + enable_kernel_altivec(); + + pagefault_disable(); + + return 1; +} + +/* + * This function must return 0 because we tail call optimise when calling + * from __copy_tofrom_user_power7 which returns 0 on success. + */ +int exit_vmx_copy(void) +{ + pagefault_enable(); + return 0; +} ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-08 6:11 ` Anton Blanchard @ 2011-12-19 3:00 ` Benjamin Herrenschmidt 2011-12-19 3:19 ` Benjamin Herrenschmidt 0 siblings, 1 reply; 11+ messages in thread From: Benjamin Herrenschmidt @ 2011-12-19 3:00 UTC (permalink / raw) To: Anton Blanchard; +Cc: Michael Neuling, paulus, sukadev, linuxppc-dev On Thu, 2011-12-08 at 17:11 +1100, Anton Blanchard wrote: > Implement a POWER7 optimised copy_to_user/copy_from_user using VMX. > For large aligned copies this new loop is over 10% faster, and for > large unaligned copies it is over 200% faster. Breaks !CONFIG_ALTIVEC build an pops some WARN's with preempt & lockdep, this seems to fix them: diff --git a/arch/powerpc/lib/Makefile b/arch/powerpc/lib/Makefile index b90b3e7..7735a2c 100644 --- a/arch/powerpc/lib/Makefile +++ b/arch/powerpc/lib/Makefile @@ -17,13 +17,14 @@ obj-$(CONFIG_HAS_IOMEM) += devres.o obj-$(CONFIG_PPC64) += copypage_64.o copyuser_64.o \ memcpy_64.o usercopy_64.o mem_64.o string.o \ checksum_wrappers_64.o hweight_64.o \ - copyuser_power7.o copyuser_power7_vmx.o + copyuser_power7.o obj-$(CONFIG_XMON) += sstep.o ldstfp.o obj-$(CONFIG_KPROBES) += sstep.o ldstfp.o obj-$(CONFIG_HAVE_HW_BREAKPOINT) += sstep.o ldstfp.o ifeq ($(CONFIG_PPC64),y) obj-$(CONFIG_SMP) += locks.o +obj-$(CONFIG_ALTIVEC) += copyuser_power7_vmx.o endif obj-$(CONFIG_PPC_LIB_RHEAP) += rheap.o diff --git a/arch/powerpc/lib/copyuser_power7.S b/arch/powerpc/lib/copyuser_power7.S index 4395939..9a21b08 100644 --- a/arch/powerpc/lib/copyuser_power7.S +++ b/arch/powerpc/lib/copyuser_power7.S @@ -85,6 +85,7 @@ _GLOBAL(__copy_tofrom_user_power7) +#ifdef CONFIG_ALTIVEC cmpldi r5,16 cmpldi cr1,r5,4096 @@ -94,6 +95,15 @@ _GLOBAL(__copy_tofrom_user_power7) blt .Lshort_copy bgt cr1,.Lvmx_copy +#else + cmpldi r5,16 + + std r3,48(r1) + std r4,56(r1) + std r5,64(r1) + + blt .Lshort_copy +#endif .Lnonvmx_copy: /* Get the source 8B aligned */ @@ -273,6 +283,7 @@ err1; stb r0,0(r3) addi r1,r1,STACKFRAMESIZE b .Lnonvmx_copy +#ifdef CONFIG_ALTIVEC .Lvmx_copy: mflr r0 std r0,16(r1) @@ -667,3 +678,4 @@ err3; stb r0,0(r3) 15: addi r1,r1,STACKFRAMESIZE b .exit_vmx_copy /* tail call optimise */ +#endif /* CONFiG_ALTIVEC */ diff --git a/arch/powerpc/lib/copyuser_power7_vmx.c b/arch/powerpc/lib/copyuser_power7_vmx.c index c37b949..6e1efad 100644 --- a/arch/powerpc/lib/copyuser_power7_vmx.c +++ b/arch/powerpc/lib/copyuser_power7_vmx.c @@ -26,10 +26,16 @@ int enter_vmx_copy(void) if (in_interrupt()) return 0; - enable_kernel_altivec(); - + /* This acts as preempt_disable() as well and will make + * enable_kernel_altivec(). We need to disable page faults + * as they can call schedule and thus make us lose the VMX + * context. So on page faults, we just fail which will cause + * a fallback to the normal non-vmx copy. + */ pagefault_disable(); + enable_kernel_altivec(); + return 1; } ^ permalink raw reply related [flat|nested] 11+ messages in thread
* Re: [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 2011-12-19 3:00 ` Benjamin Herrenschmidt @ 2011-12-19 3:19 ` Benjamin Herrenschmidt 0 siblings, 0 replies; 11+ messages in thread From: Benjamin Herrenschmidt @ 2011-12-19 3:19 UTC (permalink / raw) To: Anton Blanchard; +Cc: Michael Neuling, paulus, sukadev, linuxppc-dev On Mon, 2011-12-19 at 14:00 +1100, Benjamin Herrenschmidt wrote: > On Thu, 2011-12-08 at 17:11 +1100, Anton Blanchard wrote: > > Implement a POWER7 optimised copy_to_user/copy_from_user using VMX. > > For large aligned copies this new loop is over 10% faster, and for > > large unaligned copies it is over 200% faster. > > Breaks !CONFIG_ALTIVEC build an pops some WARN's with preempt & lockdep, > this seems to fix them: And it's missing another ifdef actually for !CONFIG_ALTIVEC, I'll fix it locally and will put a fixed version in -next. Cheers, Ben. ^ permalink raw reply [flat|nested] 11+ messages in thread
end of thread, other threads:[~2011-12-19 3:19 UTC | newest] Thread overview: 11+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2011-12-08 5:02 [PATCH] powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX Anton Blanchard 2011-12-08 5:44 ` Kumar Gala 2011-12-08 5:54 ` Michael Neuling 2011-12-08 6:04 ` Anton Blanchard 2011-12-08 13:31 ` Segher Boessenkool 2011-12-08 14:02 ` David Laight 2011-12-08 14:12 ` Segher Boessenkool 2011-12-08 19:40 ` Benjamin Herrenschmidt 2011-12-08 6:11 ` Anton Blanchard 2011-12-19 3:00 ` Benjamin Herrenschmidt 2011-12-19 3:19 ` Benjamin Herrenschmidt
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).