leandrof/linux
1
0
Fork 0
forked from Minki/linux
Code Issues Pull Requests Packages Projects Releases Wiki Activity
d0c2e4e467
linux/arch/powerpc/lib/string_64.S

203 lines
3.4 KiB
ArmAsm
Raw Normal View History

powerpc: 64bit optimised __clear_user I noticed __clear_user high up in a profile of one of my RAID stress tests. The testcase was doing a dd from /dev/zero which ends up calling __clear_user. __clear_user is basically a loop with a single 4 byte store which is horribly slow. We can do much better by aligning the desination and doing 32 bytes of 8 byte stores in a loop. The following testcase was used to verify the patch: http://ozlabs.org/~anton/junkcode/stress_clear_user.c To show the improvement in performance I ran a dd from /dev/zero to /dev/null on a POWER7 box: Before: # dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 3.72379 s, 2.8 GB/s After: # time dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 0.728318 s, 14.4 GB/s Over 5x faster. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-05-27 19:54:03 +00:00
/*
* 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/ppc_asm.h>
powerpc: Optimise the 64bit optimised __clear_user I blame Mikey for this. He elevated my slightly dubious testcase: to benchmark status. And naturally we need to be number 1 at creating zeros. So lets improve __clear_user some more. As Paul suggests we can use dcbz for large lengths. This patch gets the destination cacheline aligned then uses dcbz on whole cachelines. Before: 10485760000 bytes (10 GB) copied, 0.414744 s, 25.3 GB/s After: 10485760000 bytes (10 GB) copied, 0.268597 s, 39.0 GB/s 39 GB/s, a new record. Signed-off-by: Anton Blanchard <anton@samba.org> Tested-by: Olof Johansson <olof@lixom.net> Acked-by: Olof Johansson <olof@lixom.net> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-06-04 16:02:22 +00:00
#include <asm/asm-offsets.h>
.section ".toc","aw"
PPC64_CACHES:
.tc ppc64_caches[TC],ppc64_caches
.section ".text"
powerpc: 64bit optimised __clear_user I noticed __clear_user high up in a profile of one of my RAID stress tests. The testcase was doing a dd from /dev/zero which ends up calling __clear_user. __clear_user is basically a loop with a single 4 byte store which is horribly slow. We can do much better by aligning the desination and doing 32 bytes of 8 byte stores in a loop. The following testcase was used to verify the patch: http://ozlabs.org/~anton/junkcode/stress_clear_user.c To show the improvement in performance I ran a dd from /dev/zero to /dev/null on a POWER7 box: Before: # dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 3.72379 s, 2.8 GB/s After: # time dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 0.728318 s, 14.4 GB/s Over 5x faster. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-05-27 19:54:03 +00:00
/**
* __clear_user: - Zero a block of memory in user space, with less checking.
* @to: Destination address, in user space.
* @n: Number of bytes to zero.
*
* Zero a block of memory in user space. Caller must check
* the specified block with access_ok() before calling this function.
*
* Returns number of bytes that could not be cleared.
* On success, this will be zero.
*/
.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
.Ldo_err1:
mr r3,r8
.Ldo_err2:
mtctr r4
1:
err3; stb r0,0(r3)
addi r3,r3,1
addi r4,r4,-1
bdnz 1b
.Ldo_err3:
mr r3,r4
blr
_GLOBAL(__clear_user)
cmpdi r4,32
neg r6,r3
li r0,0
blt .Lshort_clear
mr r8,r3
mtocrf 0x01,r6
clrldi r6,r6,(64-3)
/* Get the destination 8 byte aligned */
bf cr7*4+3,1f
err1; stb r0,0(r3)
addi r3,r3,1
1: bf cr7*4+2,2f
err1; sth r0,0(r3)
addi r3,r3,2
2: bf cr7*4+1,3f
err1; stw r0,0(r3)
addi r3,r3,4
3: sub r4,r4,r6
powerpc: Optimise the 64bit optimised __clear_user I blame Mikey for this. He elevated my slightly dubious testcase: to benchmark status. And naturally we need to be number 1 at creating zeros. So lets improve __clear_user some more. As Paul suggests we can use dcbz for large lengths. This patch gets the destination cacheline aligned then uses dcbz on whole cachelines. Before: 10485760000 bytes (10 GB) copied, 0.414744 s, 25.3 GB/s After: 10485760000 bytes (10 GB) copied, 0.268597 s, 39.0 GB/s 39 GB/s, a new record. Signed-off-by: Anton Blanchard <anton@samba.org> Tested-by: Olof Johansson <olof@lixom.net> Acked-by: Olof Johansson <olof@lixom.net> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-06-04 16:02:22 +00:00
powerpc: 64bit optimised __clear_user I noticed __clear_user high up in a profile of one of my RAID stress tests. The testcase was doing a dd from /dev/zero which ends up calling __clear_user. __clear_user is basically a loop with a single 4 byte store which is horribly slow. We can do much better by aligning the desination and doing 32 bytes of 8 byte stores in a loop. The following testcase was used to verify the patch: http://ozlabs.org/~anton/junkcode/stress_clear_user.c To show the improvement in performance I ran a dd from /dev/zero to /dev/null on a POWER7 box: Before: # dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 3.72379 s, 2.8 GB/s After: # time dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 0.728318 s, 14.4 GB/s Over 5x faster. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-05-27 19:54:03 +00:00
cmpdi r4,32
powerpc: Optimise the 64bit optimised __clear_user I blame Mikey for this. He elevated my slightly dubious testcase: to benchmark status. And naturally we need to be number 1 at creating zeros. So lets improve __clear_user some more. As Paul suggests we can use dcbz for large lengths. This patch gets the destination cacheline aligned then uses dcbz on whole cachelines. Before: 10485760000 bytes (10 GB) copied, 0.414744 s, 25.3 GB/s After: 10485760000 bytes (10 GB) copied, 0.268597 s, 39.0 GB/s 39 GB/s, a new record. Signed-off-by: Anton Blanchard <anton@samba.org> Tested-by: Olof Johansson <olof@lixom.net> Acked-by: Olof Johansson <olof@lixom.net> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-06-04 16:02:22 +00:00
cmpdi cr1,r4,512
powerpc: 64bit optimised __clear_user I noticed __clear_user high up in a profile of one of my RAID stress tests. The testcase was doing a dd from /dev/zero which ends up calling __clear_user. __clear_user is basically a loop with a single 4 byte store which is horribly slow. We can do much better by aligning the desination and doing 32 bytes of 8 byte stores in a loop. The following testcase was used to verify the patch: http://ozlabs.org/~anton/junkcode/stress_clear_user.c To show the improvement in performance I ran a dd from /dev/zero to /dev/null on a POWER7 box: Before: # dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 3.72379 s, 2.8 GB/s After: # time dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 0.728318 s, 14.4 GB/s Over 5x faster. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-05-27 19:54:03 +00:00
blt .Lshort_clear
powerpc: Optimise the 64bit optimised __clear_user I blame Mikey for this. He elevated my slightly dubious testcase: to benchmark status. And naturally we need to be number 1 at creating zeros. So lets improve __clear_user some more. As Paul suggests we can use dcbz for large lengths. This patch gets the destination cacheline aligned then uses dcbz on whole cachelines. Before: 10485760000 bytes (10 GB) copied, 0.414744 s, 25.3 GB/s After: 10485760000 bytes (10 GB) copied, 0.268597 s, 39.0 GB/s 39 GB/s, a new record. Signed-off-by: Anton Blanchard <anton@samba.org> Tested-by: Olof Johansson <olof@lixom.net> Acked-by: Olof Johansson <olof@lixom.net> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-06-04 16:02:22 +00:00
bgt cr1,.Llong_clear
.Lmedium_clear:
srdi r6,r4,5
powerpc: 64bit optimised __clear_user I noticed __clear_user high up in a profile of one of my RAID stress tests. The testcase was doing a dd from /dev/zero which ends up calling __clear_user. __clear_user is basically a loop with a single 4 byte store which is horribly slow. We can do much better by aligning the desination and doing 32 bytes of 8 byte stores in a loop. The following testcase was used to verify the patch: http://ozlabs.org/~anton/junkcode/stress_clear_user.c To show the improvement in performance I ran a dd from /dev/zero to /dev/null on a POWER7 box: Before: # dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 3.72379 s, 2.8 GB/s After: # time dd if=/dev/zero of=/dev/null bs=1M count=10000 10485760000 bytes (10 GB) copied, 0.728318 s, 14.4 GB/s Over 5x faster. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-05-27 19:54:03 +00:00
mtctr r6
/* Do 32 byte chunks */
4:
err2; std r0,0(r3)
err2; std r0,8(r3)
err2; std r0,16(r3)
err2; std r0,24(r3)
addi r3,r3,32
addi r4,r4,-32
bdnz 4b
.Lshort_clear:
/* up to 31 bytes to go */
cmpdi r4,16
blt 6f
err2; std r0,0(r3)
err2; std r0,8(r3)
addi r3,r3,16
addi r4,r4,-16
/* Up to 15 bytes to go */
6: mr r8,r3
clrldi r4,r4,(64-4)
mtocrf 0x01,r4
bf cr7*4+0,7f
err1; std r0,0(r3)
addi r3,r3,8
7: bf cr7*4+1,8f
err1; stw r0,0(r3)
addi r3,r3,4
8: bf cr7*4+2,9f
err1; sth r0,0(r3)
addi r3,r3,2
9: bf cr7*4+3,10f
err1; stb r0,0(r3)
10: li r3,0
blr
powerpc: Optimise the 64bit optimised __clear_user I blame Mikey for this. He elevated my slightly dubious testcase: to benchmark status. And naturally we need to be number 1 at creating zeros. So lets improve __clear_user some more. As Paul suggests we can use dcbz for large lengths. This patch gets the destination cacheline aligned then uses dcbz on whole cachelines. Before: 10485760000 bytes (10 GB) copied, 0.414744 s, 25.3 GB/s After: 10485760000 bytes (10 GB) copied, 0.268597 s, 39.0 GB/s 39 GB/s, a new record. Signed-off-by: Anton Blanchard <anton@samba.org> Tested-by: Olof Johansson <olof@lixom.net> Acked-by: Olof Johansson <olof@lixom.net> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-06-04 16:02:22 +00:00
.Llong_clear:
ld r5,PPC64_CACHES@toc(r2)
bf cr7*4+0,11f
err2; std r0,0(r3)
addi r3,r3,8
addi r4,r4,-8
/* Destination is 16 byte aligned, need to get it cacheline aligned */
11: lwz r7,DCACHEL1LOGLINESIZE(r5)
lwz r9,DCACHEL1LINESIZE(r5)
/*
* With worst case alignment the long clear loop takes a minimum
* of 1 byte less than 2 cachelines.
*/
sldi r10,r9,2
cmpd r4,r10
blt .Lmedium_clear
neg r6,r3
addi r10,r9,-1
and. r5,r6,r10
beq 13f
srdi r6,r5,4
mtctr r6
mr r8,r3
12:
err1; std r0,0(r3)
err1; std r0,8(r3)
addi r3,r3,16
bdnz 12b
sub r4,r4,r5
13: srd r6,r4,r7
mtctr r6
mr r8,r3
14:
err1; dcbz r0,r3
add r3,r3,r9
bdnz 14b
and r4,r4,r10
cmpdi r4,32
blt .Lshort_clear
b .Lmedium_clear
Reference in New Issue Copy Permalink