PATCH: Altivec support for RAID-6
Subsystems (2)
This patch adds Altivec support for RAID-6, if appropriately configured on the ppc or ppc64 architectures. Note that it changes the compile flags for ppc64 in order to handle -maltivec correctly; this change was vetted on the ppc64 mailing list and OK'd by paulus.
Index: raid6/arch/ppc64/Makefile
===================================================================
RCS file: /home/hpa/kernel/bkcvs/linux-2.5/arch/ppc64/Makefile,v retrieving revision 1.42
arch/ppc64/Makefile +4 -0
@@ -35,7 +35,11 @@ CFLAGS += -msoft-float -pipe -mminimal-toc -mtraceback=none ifeq ($(CONFIG_POWER4_ONLY),y) +ifeq ($(CONFIG_ALTIVEC),y) + CFLAGS += $(call cc-option,-mcpu=970) +else CFLAGS += $(call cc-option,-mcpu=power4) +endif else CFLAGS += $(call cc-option,-mtune=power4) endif
Index: raid6/drivers/md/Makefile
===================================================================
RCS file: /home/hpa/kernel/bkcvs/linux-2.5/drivers/md/Makefile,v retrieving revision 1.23
drivers/md/Makefile +30 -0
@@ -9,6 +9,8 @@ raid6-objs := raid6main.o raid6algos.o raid6recov.o raid6tables.o \ raid6int1.o raid6int2.o raid6int4.o \ raid6int8.o raid6int16.o raid6int32.o \ + raid6altivec1.o raid6altivec2.o raid6altivec4.o \ + raid6altivec8.o \ raid6mmx.o raid6sse1.o raid6sse2.o hostprogs-y := mktables
@@ -36,6 +38,10 @@ cmd_unroll = $(PERL) $(srctree)/$(src)/unroll.pl $(UNROLL) \ < $< > $@ || ( rm -f $@ && exit 1 ) +ifeq ($(CONFIG_ALTIVEC),y) +altivec_flags := -maltivec -mabi=altivec +endif + targets += raid6int1.c $(obj)/raid6int1.c: UNROLL := 1 $(obj)/raid6int1.c: $(src)/raid6int.uc $(src)/unroll.pl FORCE
@@ -66,6 +72,30 @@ $(obj)/raid6int32.c: $(src)/raid6int.uc $(src)/unroll.pl FORCE $(call if_changed,unroll) +CFLAGS_raid6altivec1.o += $(altivec_flags) +targets += raid6altivec1.c +$(obj)/raid6altivec1.c: UNROLL := 1 +$(obj)/raid6altivec1.c: $(src)/raid6altivec.uc $(src)/unroll.pl FORCE + $(call if_changed,unroll) + +CFLAGS_raid6altivec2.o += $(altivec_flags) +targets += raid6altivec2.c +$(obj)/raid6altivec2.c: UNROLL := 2 +$(obj)/raid6altivec2.c: $(src)/raid6altivec.uc $(src)/unroll.pl FORCE + $(call if_changed,unroll) + +CFLAGS_raid6altivec4.o += $(altivec_flags) +targets += raid6altivec4.c +$(obj)/raid6altivec4.c: UNROLL := 4 +$(obj)/raid6altivec4.c: $(src)/raid6altivec.uc $(src)/unroll.pl FORCE + $(call if_changed,unroll) + +CFLAGS_raid6altivec8.o += $(altivec_flags) +targets += raid6altivec8.c +$(obj)/raid6altivec8.c: UNROLL := 8 +$(obj)/raid6altivec8.c: $(src)/raid6altivec.uc $(src)/unroll.pl FORCE + $(call if_changed,unroll) + quiet_cmd_mktable = TABLE $@ cmd_mktable = $(obj)/mktables > $@ || ( rm -f $@ && exit 1 )
Index: raid6/drivers/md/raid6algos.c
===================================================================
RCS file: /home/hpa/kernel/bkcvs/linux-2.5/drivers/md/raid6algos.c,v retrieving revision 1.4
drivers/md/raid6algos.c +10 -0
@@ -37,6 +37,10 @@extern const struct raid6_calls raid6_sse2x1;extern const struct raid6_calls raid6_sse2x2;extern const struct raid6_calls raid6_sse2x4; +extern const struct raid6_calls raid6_altivec1; +extern const struct raid6_calls raid6_altivec2; +extern const struct raid6_calls raid6_altivec4; +extern const struct raid6_calls raid6_altivec8;const struct raid6_calls *const raid6_algos[] = { &raid6_intx1,
@@ -60,6 +64,12 @@ &raid6_sse2x2, &raid6_sse2x4, #endif +#ifdef CONFIG_ALTIVEC + &raid6_altivec1, + &raid6_altivec2, + &raid6_altivec4, + &raid6_altivec8, +#endifNULL };
Index: raid6/drivers/md/raid6altivec.uc
===================================================================
RCS file: raid6/drivers/md/raid6altivec.uc
drivers/md/raid6altivec.uc new file +122 -0
122 lines not on this page · raw message
Quoted 2 lines · H. Peter Anvin
$(obj)/raid6int1.c: $(src)/raid6int.uc $(src)/unroll.pl FORCE
So we require that raid6int.uc propagate through the system with the x bit set.
I wonder if it would be safer to stick a $(SHELL) in there?
Quoted 7 lines · Andrew Morton
Quoted 2 lines · H. Peter Anvin
$(obj)/raid6int1.c: $(src)/raid6int.uc $(src)/unroll.pl FORCESo we require that raid6int.uc propagate through the system with the x bit set.
err, I think I misread that. Ignore.
Quoted 7 lines · H. Peter Anvin
This patch adds Altivec support for RAID-6, if appropriately configured on the ppc or ppc64 architectures. Note that it changes the compile flags for ppc64 in order to handle -maltivec correctly; this change was vetted on the ppc64 mailing list and OK'd by paulus.
Signed-off-by H. Peter Anvin <[email protected]>
hpa, are you aware of any other routines which should benefit from altivec?
Signature · 3 lines
Greetz, Antonio Vargas aka winden of network Las cosas no son lo que parecen, excepto cuando parecen lo que si son.
Followup to: <[email protected]> By author: Antonio Vargas <[email protected]> In newsgroup: linux.dev.kernel
Quoted 3 lines
hpa, are you aware of any other routines which should benefit from altivec?
Presumably the XOR code used by RAID-5, and quite possibly some of the cryptography stuff. Unlike most SIMD instruction sets, it should be possible to write AES using Altivec.
-hpa
Quoted 5 lines · Antonio Vargas
Quoted 4 lines · H. Peter Anvin
Followup to: <[email protected]> By author: Antonio Vargas <[email protected]> In newsgroup: linux.dev.kernelQuoted 1 line
Quoted 9 lines
Quoted 6 lines
Quoted 2 lines
hpa, are you aware of any other routines which should benefit from altivec?
Presumably the XOR code used by RAID-5, and quite possibly some of the cryptography stuff. Unlike most SIMD instruction sets, it should be possible to write AES using Altivec.
I'll take a crack at the RAID-5 stuff first then.
Here we go... this file is compile-tested on gcc 3.3 from userspace, it can serve as an starting point for testing.
Signature · 3 lines
Greetz, Antonio Vargas aka winden of network Las cosas no son lo que parecen, excepto cuando parecen lo que si son.
/*
* include/asm-ppc/xor.h
*
* Altivec optimized RAID-5 checksumming functions.
* Copyright 2004 Antonio Vargas, [email protected]
*
* Based on include/asm-generic/xor.h and drivers/md/raid6altivec.uc *
* 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, version 2.
*
* You should have received a copy of the GNU General Public License
* (for example /usr/src/linux/COPYING); if not, write to the Free
* Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. */
#ifdef KERNEL
#include <asm-generic/xor.h>
#endif
#ifdef CONFIG_ALTIVEC
#ifdef KERNEL
#include <altivec.h>
#include <asm/cputable.h>
#include <asm/processor.h>
#include <asm/system.h>
#else
#define prefetch(x)
#define prefetchw(x)
void enable_kernel_altivec(){};
void preempt_enable(){};
void preempt_disable(){};
#endif
typedef vector unsigned char unative_t;
#define BODY\
long lines = bytes / (4 * sizeof(unative_t)) - 1;\
int i;\
preempt_disable();\
enable_kernel_altivec();\
PREF(0)\
do {\
PREF(4);\
CALC(0);\
CALC(1);\
CALC(2);\
CALC(3);\
INC;\
} while (--lines > 0);\
for(i = 0 ; i < 4 ; i++)\
CALC(i);\
preempt_enable();#define PREF2(x) prefetchw(p1 + x); prefetch(p2 + x);
#define PREF3(x) PREF2(x) prefetch(p3 + x); #define PREF4(x) PREF3(x) prefetch(p4 + x); #define PREF5(x) PREF4(x) prefetch(p5 + x);
#define CALC2(x) p1[x] ^= p2[x]
#define CALC3(x) CALC2(x) ^ p3[x]
#define CALC4(x) CALC3(x) ^ p4[x]
#define CALC5(x) CALC4(x) ^ p5[x]
#define INC2 p1 += 4; p2 += 4;
#define INC3 INC2 p3 += 4;
#define INC4 INC3 p4 += 4;
#define INC5 INC4 p5 += 4;
static voidxor_altivec2(unsigned long bytes, unative_t *p1, unative_t *p2) {
#undef PREF
#undef CALC
#undef INC
#define PREF(x) PREF2(x)
#define CALC(x) CALC2(x)
#define INC INC2
BODY
}
static voidxor_altivec3(unsigned long bytes, unative_t *p1, unative_t *p2, unative_t *p3) {
#undef PREF
#undef CALC
#undef INC
#define PREF(x) PREF3(x)
#define CALC(x) CALC3(x)
#define INC INC3
BODY
}
static voidxor_altivec4(unsigned long bytes, unative_t *p1, unative_t *p2, unative_t *p3, unative_t *p4) {
#undef PREF
#undef CALC
#undef INC
#define PREF(x) PREF4(x)
#define CALC(x) CALC4(x)
#define INC INC4
BODY
}
static voidxor_altivec5(unsigned long bytes, unative_t *p1, unative_t *p2, unative_t *p3, unative_t *p4, unative_t *p5) {
#undef PREF
#undef CALC
#undef INC
#define PREF(x) PREF5(x)
#define CALC(x) CALC5(x)
#define INC INC5
BODY
}
#ifdef KERNEL
static struct xor_block_template xor_block_altivec = {
.name = "altivec",
.do_2 = xor_altivec2,
.do_3 = xor_altivec3,
.do_4 = xor_altivec4,
.do_5 = xor_altivec5,
};#undef XOR_TRY_TEMPLATES
#define XOR_TRY_TEMPLATES \
do { \
xor_speed(&xor_block_8regs); \
xor_speed(&xor_block_8regs_p); \
xor_speed(&xor_block_32regs); \
xor_speed(&xor_block_32regs_p); \
xor_speed(&xor_block_altivec); \
} while (0)#endif
#endif