powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX
authorAnton Blanchard <anton@samba.org>
Wed, 7 Dec 2011 20:11:45 +0000 (20:11 +0000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Mon, 19 Dec 2011 03:40:40 +0000 (14:40 +1100)
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.

[BenH: Fixed a problem with preempt and fixed build without CONFIG_ALTIVEC]

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/include/asm/cputable.h
arch/powerpc/lib/Makefile
arch/powerpc/lib/copyuser_64.S
arch/powerpc/lib/copyuser_power7.S [new file with mode: 0644]
arch/powerpc/lib/copyuser_power7_vmx.c [new file with mode: 0644]

index 7044233124ba4fd11996c4176d910273f1b75673..ad55a1ccb9fb76fbe98c53005239df303efa5dc1 100644 (file)
@@ -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 166a6a0ad54406858f66cbb1cdd093650ceb9678..7735a2c2e6d9052db505d50b02fc8ed09750fbc4 100644 (file)
@@ -16,13 +16,15 @@ obj-$(CONFIG_HAS_IOMEM)     += devres.o
 
 obj-$(CONFIG_PPC64)    += copypage_64.o copyuser_64.o \
                           memcpy_64.o usercopy_64.o mem_64.o string.o \
-                          checksum_wrappers_64.o hweight_64.o
+                          checksum_wrappers_64.o hweight_64.o \
+                          copyuser_power7.o
 obj-$(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
index 578b625d6a3ca96692db05f62796593a11824348..773d38f90aaa27a46a0ee2e1593322b4fc947b25 100644 (file)
 
        .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
diff --git a/arch/powerpc/lib/copyuser_power7.S b/arch/powerpc/lib/copyuser_power7.S
new file mode 100644 (file)
index 0000000..497db7b
--- /dev/null
@@ -0,0 +1,683 @@
+/*
+ * 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
+
+#ifdef CONFIG_ALTIVEC
+       .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
+#endif /* CONFIG_ALTIVEC */
+
+.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)
+#ifdef CONFIG_ALTIVEC
+       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
+#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 */
+       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
+
+#ifdef CONFIG_ALTIVEC
+.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 */
+#endif /* CONFiG_ALTIVEC */
diff --git a/arch/powerpc/lib/copyuser_power7_vmx.c b/arch/powerpc/lib/copyuser_power7_vmx.c
new file mode 100644 (file)
index 0000000..6e1efad
--- /dev/null
@@ -0,0 +1,50 @@
+/*
+ * 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;
+
+       /* 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;
+}
+
+/*
+ * 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;
+}