powerpc: POWER7 optimised copy_page using VMX and enhanced prefetch
authorAnton Blanchard <anton@samba.org>
Tue, 29 May 2012 19:33:12 +0000 (19:33 +0000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Tue, 3 Jul 2012 04:14:44 +0000 (14:14 +1000)
Implement a POWER7 optimised copy_page using VMX and enhanced
prefetch instructions. We use enhanced prefetch hints to prefetch
both the load and store side. We copy a cacheline at a time and
fall back to regular loads and stores if we are unable to use VMX
(eg we are in an interrupt).

The following microbenchmark was used to assess the impact of
the patch:

http://ozlabs.org/~anton/junkcode/page_fault_file.c

We test MAP_PRIVATE page faults across a 1GB file, 100 times:

# time ./page_fault_file -p -l 1G -i 100

Before: 22.25s
After:  18.89s

17% faster

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/lib/Makefile
arch/powerpc/lib/copypage_64.S
arch/powerpc/lib/copypage_power7.S [new file with mode: 0644]
arch/powerpc/lib/vmx-helper.c

index 1eb94c7..873805e 100644 (file)
@@ -17,7 +17,7 @@ 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 string_64.o
+                          copyuser_power7.o string_64.o copypage_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
index 53dcb6b..9f9434a 100644 (file)
@@ -17,7 +17,11 @@ PPC64_CACHES:
         .section        ".text"
 
 _GLOBAL(copy_page)
+BEGIN_FTR_SECTION
        lis     r5,PAGE_SIZE@h
+FTR_SECTION_ELSE
+       b       .copypage_power7
+ALT_FTR_SECTION_END_IFCLR(CPU_FTR_VMX_COPY)
        ori     r5,r5,PAGE_SIZE@l
 BEGIN_FTR_SECTION
        ld      r10,PPC64_CACHES@toc(r2)
diff --git a/arch/powerpc/lib/copypage_power7.S b/arch/powerpc/lib/copypage_power7.S
new file mode 100644 (file)
index 0000000..01e2b5d
--- /dev/null
@@ -0,0 +1,168 @@
+/*
+ * 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, 2012
+ *
+ * Author: Anton Blanchard <anton@au.ibm.com>
+ */
+#include <asm/page.h>
+#include <asm/ppc_asm.h>
+
+#define STACKFRAMESIZE 256
+#define STK_REG(i)     (112 + ((i)-14)*8)
+
+_GLOBAL(copypage_power7)
+       /*
+        * We prefetch both the source and destination using enhanced touch
+        * instructions. We use a stream ID of 0 for the load side and
+        * 1 for the store side. Since source and destination are page
+        * aligned we don't need to clear the bottom 7 bits of either
+        * address.
+        */
+       ori     r9,r3,1         /* stream=1 */
+
+#ifdef CONFIG_PPC_64K_PAGES
+       lis     r7,0x0E01       /* depth=7, units=512 */
+#else
+       lis     r7,0x0E00       /* depth=7 */
+       ori     r7,r7,0x1000    /* units=32 */
+#endif
+       ori     r10,r7,1        /* stream=1 */
+
+       lis     r8,0x8000       /* GO=1 */
+       clrldi  r8,r8,32
+
+.machine push
+.machine "power4"
+       dcbt    r0,r4,0b01000
+       dcbt    r0,r7,0b01010
+       dcbtst  r0,r9,0b01000
+       dcbtst  r0,r10,0b01010
+       eieio
+       dcbt    r0,r8,0b01010   /* GO */
+.machine pop
+
+#ifdef CONFIG_ALTIVEC
+       mflr    r0
+       std     r3,48(r1)
+       std     r4,56(r1)
+       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)
+       mtlr    r0
+
+       li      r0,(PAGE_SIZE/128)
+       mtctr   r0
+
+       beq     .Lnonvmx_copy
+
+       addi    r1,r1,STACKFRAMESIZE
+
+       li      r6,16
+       li      r7,32
+       li      r8,48
+       li      r9,64
+       li      r10,80
+       li      r11,96
+       li      r12,112
+
+       .align  5
+1:     lvx     vr7,r0,r4
+       lvx     vr6,r4,r6
+       lvx     vr5,r4,r7
+       lvx     vr4,r4,r8
+       lvx     vr3,r4,r9
+       lvx     vr2,r4,r10
+       lvx     vr1,r4,r11
+       lvx     vr0,r4,r12
+       addi    r4,r4,128
+       stvx    vr7,r0,r3
+       stvx    vr6,r3,r6
+       stvx    vr5,r3,r7
+       stvx    vr4,r3,r8
+       stvx    vr3,r3,r9
+       stvx    vr2,r3,r10
+       stvx    vr1,r3,r11
+       stvx    vr0,r3,r12
+       addi    r3,r3,128
+       bdnz    1b
+
+       b       .exit_vmx_copy          /* tail call optimise */
+
+#else
+       li      r0,(PAGE_SIZE/128)
+       mtctr   r0
+
+       stdu    r1,-STACKFRAMESIZE(r1)
+#endif
+
+.Lnonvmx_copy:
+       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)
+
+1:     ld      r0,0(r4)
+       ld      r5,8(r4)
+       ld      r6,16(r4)
+       ld      r7,24(r4)
+       ld      r8,32(r4)
+       ld      r9,40(r4)
+       ld      r10,48(r4)
+       ld      r11,56(r4)
+       ld      r12,64(r4)
+       ld      r14,72(r4)
+       ld      r15,80(r4)
+       ld      r16,88(r4)
+       ld      r17,96(r4)
+       ld      r18,104(r4)
+       ld      r19,112(r4)
+       ld      r20,120(r4)
+       addi    r4,r4,128
+       std     r0,0(r3)
+       std     r5,8(r3)
+       std     r6,16(r3)
+       std     r7,24(r3)
+       std     r8,32(r3)
+       std     r9,40(r3)
+       std     r10,48(r3)
+       std     r11,56(r3)
+       std     r12,64(r3)
+       std     r14,72(r3)
+       std     r15,80(r3)
+       std     r16,88(r3)
+       std     r17,96(r3)
+       std     r18,104(r3)
+       std     r19,112(r3)
+       std     r20,120(r3)
+       addi    r3,r3,128
+       bdnz    1b
+
+       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)
+       addi    r1,r1,STACKFRAMESIZE
+       blr
index 753a839..3cf529c 100644 (file)
@@ -49,3 +49,26 @@ int exit_vmx_usercopy(void)
        pagefault_enable();
        return 0;
 }
+
+int enter_vmx_copy(void)
+{
+       if (in_interrupt())
+               return 0;
+
+       preempt_disable();
+
+       enable_kernel_altivec();
+
+       return 1;
+}
+
+/*
+ * All calls to this function will be optimised into tail calls. We are
+ * passed a pointer to the destination which we return as required by a
+ * memcpy implementation.
+ */
+void *exit_vmx_copy(void *dest)
+{
+       preempt_enable();
+       return dest;
+}