aboutsummaryrefslogtreecommitdiffstats
path: root/lib
diff options
context:
space:
mode:
authorDavid Woodhouse <David.Woodhouse@intel.com>2010-08-09 05:36:44 -0400
committerDavid Woodhouse <David.Woodhouse@intel.com>2010-08-09 05:36:44 -0400
commit2144381da478cc4aa3a29ee29b0c5e6ddaaced14 (patch)
tree380a29fec86b537ed602d12f4050654b1c66c27c /lib
parent45d7f32c7a43cbb9592886d38190e379e2eb2226 (diff)
parente5d84970a554d5c0072043a7b9f0f5b88b5fdfe1 (diff)
Merge branch 'async' of macbook:git/btrfs-unstable
Conflicts: drivers/md/Makefile lib/raid6/unroll.pl
Diffstat (limited to 'lib')
-rw-r--r--lib/Kconfig3
-rw-r--r--lib/Makefile1
-rw-r--r--lib/raid6/Makefile78
-rw-r--r--lib/raid6/mktables.c132
-rw-r--r--lib/raid6/raid6algos.c154
-rw-r--r--lib/raid6/raid6altivec.uc130
-rw-r--r--lib/raid6/raid6int.uc117
-rw-r--r--lib/raid6/raid6mmx.c142
-rw-r--r--lib/raid6/raid6recov.c132
-rw-r--r--lib/raid6/raid6sse1.c162
-rw-r--r--lib/raid6/raid6sse2.c262
-rw-r--r--lib/raid6/raid6test/Makefile75
-rw-r--r--lib/raid6/raid6test/test.c124
-rw-r--r--lib/raid6/raid6x86.h61
-rw-r--r--lib/raid6/unroll.awk20
15 files changed, 1593 insertions, 0 deletions
diff --git a/lib/Kconfig b/lib/Kconfig
index 5b916bc0fbae..fa9bf2c06199 100644
--- a/lib/Kconfig
+++ b/lib/Kconfig
@@ -7,6 +7,9 @@ config BINARY_PRINTF
7 7
8menu "Library routines" 8menu "Library routines"
9 9
10config RAID6_PQ
11 tristate
12
10config BITREVERSE 13config BITREVERSE
11 tristate 14 tristate
12 15
diff --git a/lib/Makefile b/lib/Makefile
index 0bfabba1bb32..e6a3763b8212 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -69,6 +69,7 @@ obj-$(CONFIG_ZLIB_DEFLATE) += zlib_deflate/
69obj-$(CONFIG_REED_SOLOMON) += reed_solomon/ 69obj-$(CONFIG_REED_SOLOMON) += reed_solomon/
70obj-$(CONFIG_LZO_COMPRESS) += lzo/ 70obj-$(CONFIG_LZO_COMPRESS) += lzo/
71obj-$(CONFIG_LZO_DECOMPRESS) += lzo/ 71obj-$(CONFIG_LZO_DECOMPRESS) += lzo/
72obj-$(CONFIG_RAID6_PQ) += raid6/
72 73
73lib-$(CONFIG_DECOMPRESS_GZIP) += decompress_inflate.o 74lib-$(CONFIG_DECOMPRESS_GZIP) += decompress_inflate.o
74lib-$(CONFIG_DECOMPRESS_BZIP2) += decompress_bunzip2.o 75lib-$(CONFIG_DECOMPRESS_BZIP2) += decompress_bunzip2.o
diff --git a/lib/raid6/Makefile b/lib/raid6/Makefile
new file mode 100644
index 000000000000..19bf32da644f
--- /dev/null
+++ b/lib/raid6/Makefile
@@ -0,0 +1,78 @@
1obj-$(CONFIG_RAID6_PQ) += raid6_pq.o
2
3raid6_pq-y += raid6algos.o raid6recov.o raid6tables.o \
4 raid6int1.o raid6int2.o raid6int4.o \
5 raid6int8.o raid6int16.o raid6int32.o \
6 raid6altivec1.o raid6altivec2.o raid6altivec4.o \
7 raid6altivec8.o \
8 raid6mmx.o raid6sse1.o raid6sse2.o
9hostprogs-y += mktables
10
11quiet_cmd_unroll = UNROLL $@
12 cmd_unroll = $(AWK) -f$(srctree)/$(src)/unroll.awk -vN=$(UNROLL) \
13 < $< > $@ || ( rm -f $@ && exit 1 )
14
15ifeq ($(CONFIG_ALTIVEC),y)
16altivec_flags := -maltivec -mabi=altivec
17endif
18
19targets += raid6int1.c
20$(obj)/raid6int1.c: UNROLL := 1
21$(obj)/raid6int1.c: $(src)/raid6int.uc $(src)/unroll.awk FORCE
22 $(call if_changed,unroll)
23
24targets += raid6int2.c
25$(obj)/raid6int2.c: UNROLL := 2
26$(obj)/raid6int2.c: $(src)/raid6int.uc $(src)/unroll.awk FORCE
27 $(call if_changed,unroll)
28
29targets += raid6int4.c
30$(obj)/raid6int4.c: UNROLL := 4
31$(obj)/raid6int4.c: $(src)/raid6int.uc $(src)/unroll.awk FORCE
32 $(call if_changed,unroll)
33
34targets += raid6int8.c
35$(obj)/raid6int8.c: UNROLL := 8
36$(obj)/raid6int8.c: $(src)/raid6int.uc $(src)/unroll.awk FORCE
37 $(call if_changed,unroll)
38
39targets += raid6int16.c
40$(obj)/raid6int16.c: UNROLL := 16
41$(obj)/raid6int16.c: $(src)/raid6int.uc $(src)/unroll.awk FORCE
42 $(call if_changed,unroll)
43
44targets += raid6int32.c
45$(obj)/raid6int32.c: UNROLL := 32
46$(obj)/raid6int32.c: $(src)/raid6int.uc $(src)/unroll.awk FORCE
47 $(call if_changed,unroll)
48
49CFLAGS_raid6altivec1.o += $(altivec_flags)
50targets += raid6altivec1.c
51$(obj)/raid6altivec1.c: UNROLL := 1
52$(obj)/raid6altivec1.c: $(src)/raid6altivec.uc $(src)/unroll.awk FORCE
53 $(call if_changed,unroll)
54
55CFLAGS_raid6altivec2.o += $(altivec_flags)
56targets += raid6altivec2.c
57$(obj)/raid6altivec2.c: UNROLL := 2
58$(obj)/raid6altivec2.c: $(src)/raid6altivec.uc $(src)/unroll.awk FORCE
59 $(call if_changed,unroll)
60
61CFLAGS_raid6altivec4.o += $(altivec_flags)
62targets += raid6altivec4.c
63$(obj)/raid6altivec4.c: UNROLL := 4
64$(obj)/raid6altivec4.c: $(src)/raid6altivec.uc $(src)/unroll.awk FORCE
65 $(call if_changed,unroll)
66
67CFLAGS_raid6altivec8.o += $(altivec_flags)
68targets += raid6altivec8.c
69$(obj)/raid6altivec8.c: UNROLL := 8
70$(obj)/raid6altivec8.c: $(src)/raid6altivec.uc $(src)/unroll.awk FORCE
71 $(call if_changed,unroll)
72
73quiet_cmd_mktable = TABLE $@
74 cmd_mktable = $(obj)/mktables > $@ || ( rm -f $@ && exit 1 )
75
76targets += raid6tables.c
77$(obj)/raid6tables.c: $(obj)/mktables FORCE
78 $(call if_changed,mktable)
diff --git a/lib/raid6/mktables.c b/lib/raid6/mktables.c
new file mode 100644
index 000000000000..3b1500843bba
--- /dev/null
+++ b/lib/raid6/mktables.c
@@ -0,0 +1,132 @@
1/* -*- linux-c -*- ------------------------------------------------------- *
2 *
3 * Copyright 2002-2007 H. Peter Anvin - All Rights Reserved
4 *
5 * This file is part of the Linux kernel, and is made available under
6 * the terms of the GNU General Public License version 2 or (at your
7 * option) any later version; incorporated herein by reference.
8 *
9 * ----------------------------------------------------------------------- */
10
11/*
12 * mktables.c
13 *
14 * Make RAID-6 tables. This is a host user space program to be run at
15 * compile time.
16 */
17
18#include <stdio.h>
19#include <string.h>
20#include <inttypes.h>
21#include <stdlib.h>
22#include <time.h>
23
24static uint8_t gfmul(uint8_t a, uint8_t b)
25{
26 uint8_t v = 0;
27
28 while (b) {
29 if (b & 1)
30 v ^= a;
31 a = (a << 1) ^ (a & 0x80 ? 0x1d : 0);
32 b >>= 1;
33 }
34
35 return v;
36}
37
38static uint8_t gfpow(uint8_t a, int b)
39{
40 uint8_t v = 1;
41
42 b %= 255;
43 if (b < 0)
44 b += 255;
45
46 while (b) {
47 if (b & 1)
48 v = gfmul(v, a);
49 a = gfmul(a, a);
50 b >>= 1;
51 }
52
53 return v;
54}
55
56int main(int argc, char *argv[])
57{
58 int i, j, k;
59 uint8_t v;
60 uint8_t exptbl[256], invtbl[256];
61
62 printf("#include <linux/raid/pq.h>\n");
63
64 /* Compute multiplication table */
65 printf("\nconst u8 __attribute__((aligned(256)))\n"
66 "raid6_gfmul[256][256] =\n"
67 "{\n");
68 for (i = 0; i < 256; i++) {
69 printf("\t{\n");
70 for (j = 0; j < 256; j += 8) {
71 printf("\t\t");
72 for (k = 0; k < 8; k++)
73 printf("0x%02x,%c", gfmul(i, j + k),
74 (k == 7) ? '\n' : ' ');
75 }
76 printf("\t},\n");
77 }
78 printf("};\n");
79 printf("#ifdef __KERNEL__\n");
80 printf("EXPORT_SYMBOL(raid6_gfmul);\n");
81 printf("#endif\n");
82
83 /* Compute power-of-2 table (exponent) */
84 v = 1;
85 printf("\nconst u8 __attribute__((aligned(256)))\n"
86 "raid6_gfexp[256] =\n" "{\n");
87 for (i = 0; i < 256; i += 8) {
88 printf("\t");
89 for (j = 0; j < 8; j++) {
90 exptbl[i + j] = v;
91 printf("0x%02x,%c", v, (j == 7) ? '\n' : ' ');
92 v = gfmul(v, 2);
93 if (v == 1)
94 v = 0; /* For entry 255, not a real entry */
95 }
96 }
97 printf("};\n");
98 printf("#ifdef __KERNEL__\n");
99 printf("EXPORT_SYMBOL(raid6_gfexp);\n");
100 printf("#endif\n");
101
102 /* Compute inverse table x^-1 == x^254 */
103 printf("\nconst u8 __attribute__((aligned(256)))\n"
104 "raid6_gfinv[256] =\n" "{\n");
105 for (i = 0; i < 256; i += 8) {
106 printf("\t");
107 for (j = 0; j < 8; j++) {
108 invtbl[i + j] = v = gfpow(i + j, 254);
109 printf("0x%02x,%c", v, (j == 7) ? '\n' : ' ');
110 }
111 }
112 printf("};\n");
113 printf("#ifdef __KERNEL__\n");
114 printf("EXPORT_SYMBOL(raid6_gfinv);\n");
115 printf("#endif\n");
116
117 /* Compute inv(2^x + 1) (exponent-xor-inverse) table */
118 printf("\nconst u8 __attribute__((aligned(256)))\n"
119 "raid6_gfexi[256] =\n" "{\n");
120 for (i = 0; i < 256; i += 8) {
121 printf("\t");
122 for (j = 0; j < 8; j++)
123 printf("0x%02x,%c", invtbl[exptbl[i + j] ^ 1],
124 (j == 7) ? '\n' : ' ');
125 }
126 printf("};\n");
127 printf("#ifdef __KERNEL__\n");
128 printf("EXPORT_SYMBOL(raid6_gfexi);\n");
129 printf("#endif\n");
130
131 return 0;
132}
diff --git a/lib/raid6/raid6algos.c b/lib/raid6/raid6algos.c
new file mode 100644
index 000000000000..1f8784bfd44d
--- /dev/null
+++ b/lib/raid6/raid6algos.c
@@ -0,0 +1,154 @@
1/* -*- linux-c -*- ------------------------------------------------------- *
2 *
3 * Copyright 2002 H. Peter Anvin - All Rights Reserved
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
8 * Boston MA 02111-1307, USA; either version 2 of the License, or
9 * (at your option) any later version; incorporated herein by reference.
10 *
11 * ----------------------------------------------------------------------- */
12
13/*
14 * raid6algos.c
15 *
16 * Algorithm list and algorithm selection for RAID-6
17 */
18
19#include <linux/raid/pq.h>
20#include <linux/gfp.h>
21#ifndef __KERNEL__
22#include <sys/mman.h>
23#include <stdio.h>
24#else
25#if !RAID6_USE_EMPTY_ZERO_PAGE
26/* In .bss so it's zeroed */
27const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
28EXPORT_SYMBOL(raid6_empty_zero_page);
29#endif
30#endif
31
32struct raid6_calls raid6_call;
33EXPORT_SYMBOL_GPL(raid6_call);
34
35const struct raid6_calls * const raid6_algos[] = {
36 &raid6_intx1,
37 &raid6_intx2,
38 &raid6_intx4,
39 &raid6_intx8,
40#if defined(__ia64__)
41 &raid6_intx16,
42 &raid6_intx32,
43#endif
44#if defined(__i386__) && !defined(__arch_um__)
45 &raid6_mmxx1,
46 &raid6_mmxx2,
47 &raid6_sse1x1,
48 &raid6_sse1x2,
49 &raid6_sse2x1,
50 &raid6_sse2x2,
51#endif
52#if defined(__x86_64__) && !defined(__arch_um__)
53 &raid6_sse2x1,
54 &raid6_sse2x2,
55 &raid6_sse2x4,
56#endif
57#ifdef CONFIG_ALTIVEC
58 &raid6_altivec1,
59 &raid6_altivec2,
60 &raid6_altivec4,
61 &raid6_altivec8,
62#endif
63 NULL
64};
65
66#ifdef __KERNEL__
67#define RAID6_TIME_JIFFIES_LG2 4
68#else
69/* Need more time to be stable in userspace */
70#define RAID6_TIME_JIFFIES_LG2 9
71#define time_before(x, y) ((x) < (y))
72#endif
73
74/* Try to pick the best algorithm */
75/* This code uses the gfmul table as convenient data set to abuse */
76
77int __init raid6_select_algo(void)
78{
79 const struct raid6_calls * const * algo;
80 const struct raid6_calls * best;
81 char *syndromes;
82 void *dptrs[(65536/PAGE_SIZE)+2];
83 int i, disks;
84 unsigned long perf, bestperf;
85 int bestprefer;
86 unsigned long j0, j1;
87
88 disks = (65536/PAGE_SIZE)+2;
89 for ( i = 0 ; i < disks-2 ; i++ ) {
90 dptrs[i] = ((char *)raid6_gfmul) + PAGE_SIZE*i;
91 }
92
93 /* Normal code - use a 2-page allocation to avoid D$ conflict */
94 syndromes = (void *) __get_free_pages(GFP_KERNEL, 1);
95
96 if ( !syndromes ) {
97 printk("raid6: Yikes! No memory available.\n");
98 return -ENOMEM;
99 }
100
101 dptrs[disks-2] = syndromes;
102 dptrs[disks-1] = syndromes + PAGE_SIZE;
103
104 bestperf = 0; bestprefer = 0; best = NULL;
105
106 for ( algo = raid6_algos ; *algo ; algo++ ) {
107 if ( !(*algo)->valid || (*algo)->valid() ) {
108 perf = 0;
109
110 preempt_disable();
111 j0 = jiffies;
112 while ( (j1 = jiffies) == j0 )
113 cpu_relax();
114 while (time_before(jiffies,
115 j1 + (1<<RAID6_TIME_JIFFIES_LG2))) {
116 (*algo)->gen_syndrome(disks, PAGE_SIZE, dptrs);
117 perf++;
118 }
119 preempt_enable();
120
121 if ( (*algo)->prefer > bestprefer ||
122 ((*algo)->prefer == bestprefer &&
123 perf > bestperf) ) {
124 best = *algo;
125 bestprefer = best->prefer;
126 bestperf = perf;
127 }
128 printk("raid6: %-8s %5ld MB/s\n", (*algo)->name,
129 (perf*HZ) >> (20-16+RAID6_TIME_JIFFIES_LG2));
130 }
131 }
132
133 if (best) {
134 printk("raid6: using algorithm %s (%ld MB/s)\n",
135 best->name,
136 (bestperf*HZ) >> (20-16+RAID6_TIME_JIFFIES_LG2));
137 raid6_call = *best;
138 } else
139 printk("raid6: Yikes! No algorithm found!\n");
140
141 free_pages((unsigned long)syndromes, 1);
142
143 return best ? 0 : -EINVAL;
144}
145
146static void raid6_exit(void)
147{
148 do { } while (0);
149}
150
151subsys_initcall(raid6_select_algo);
152module_exit(raid6_exit);
153MODULE_LICENSE("GPL");
154MODULE_DESCRIPTION("RAID6 Q-syndrome calculations");
diff --git a/lib/raid6/raid6altivec.uc b/lib/raid6/raid6altivec.uc
new file mode 100644
index 000000000000..2654d5c854be
--- /dev/null
+++ b/lib/raid6/raid6altivec.uc
@@ -0,0 +1,130 @@
1/* -*- linux-c -*- ------------------------------------------------------- *
2 *
3 * Copyright 2002-2004 H. Peter Anvin - All Rights Reserved
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
8 * Boston MA 02111-1307, USA; either version 2 of the License, or
9 * (at your option) any later version; incorporated herein by reference.
10 *
11 * ----------------------------------------------------------------------- */
12
13/*
14 * raid6altivec$#.c
15 *
16 * $#-way unrolled portable integer math RAID-6 instruction set
17 *
18 * This file is postprocessed using unroll.awk
19 *
20 * <benh> hpa: in process,
21 * you can just "steal" the vec unit with enable_kernel_altivec() (but
22 * bracked this with preempt_disable/enable or in a lock)
23 */
24
25#include <linux/raid/pq.h>
26
27#ifdef CONFIG_ALTIVEC
28
29#include <altivec.h>
30#ifdef __KERNEL__
31# include <asm/system.h>
32# include <asm/cputable.h>
33#endif
34
35/*
36 * This is the C data type to use. We use a vector of
37 * signed char so vec_cmpgt() will generate the right
38 * instruction.
39 */
40
41typedef vector signed char unative_t;
42
43#define NBYTES(x) ((vector signed char) {x,x,x,x, x,x,x,x, x,x,x,x, x,x,x,x})
44#define NSIZE sizeof(unative_t)
45
46/*
47 * The SHLBYTE() operation shifts each byte left by 1, *not*
48 * rolling over into the next byte
49 */
50static inline __attribute_const__ unative_t SHLBYTE(unative_t v)
51{
52 return vec_add(v,v);
53}
54
55/*
56 * The MASK() operation returns 0xFF in any byte for which the high
57 * bit is 1, 0x00 for any byte for which the high bit is 0.
58 */
59static inline __attribute_const__ unative_t MASK(unative_t v)
60{
61 unative_t zv = NBYTES(0);
62
63 /* vec_cmpgt returns a vector bool char; thus the need for the cast */
64 return (unative_t)vec_cmpgt(zv, v);
65}
66
67
68/* This is noinline to make damned sure that gcc doesn't move any of the
69 Altivec code around the enable/disable code */
70static void noinline
71raid6_altivec$#_gen_syndrome_real(int disks, size_t bytes, void **ptrs)
72{
73 u8 **dptr = (u8 **)ptrs;
74 u8 *p, *q;
75 int d, z, z0;
76
77 unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
78 unative_t x1d = NBYTES(0x1d);
79
80 z0 = disks - 3; /* Highest data disk */
81 p = dptr[z0+1]; /* XOR parity */
82 q = dptr[z0+2]; /* RS syndrome */
83
84 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
85 wq$$ = wp$$ = *(unative_t *)&dptr[z0][d+$$*NSIZE];
86 for ( z = z0-1 ; z >= 0 ; z-- ) {
87 wd$$ = *(unative_t *)&dptr[z][d+$$*NSIZE];
88 wp$$ = vec_xor(wp$$, wd$$);
89 w2$$ = MASK(wq$$);
90 w1$$ = SHLBYTE(wq$$);
91 w2$$ = vec_and(w2$$, x1d);
92 w1$$ = vec_xor(w1$$, w2$$);
93 wq$$ = vec_xor(w1$$, wd$$);
94 }
95 *(unative_t *)&p[d+NSIZE*$$] = wp$$;
96 *(unative_t *)&q[d+NSIZE*$$] = wq$$;
97 }
98}
99
100static void raid6_altivec$#_gen_syndrome(int disks, size_t bytes, void **ptrs)
101{
102 preempt_disable();
103 enable_kernel_altivec();
104
105 raid6_altivec$#_gen_syndrome_real(disks, bytes, ptrs);
106
107 preempt_enable();
108}
109
110int raid6_have_altivec(void);
111#if $# == 1
112int raid6_have_altivec(void)
113{
114 /* This assumes either all CPUs have Altivec or none does */
115# ifdef __KERNEL__
116 return cpu_has_feature(CPU_FTR_ALTIVEC);
117# else
118 return 1;
119# endif
120}
121#endif
122
123const struct raid6_calls raid6_altivec$# = {
124 raid6_altivec$#_gen_syndrome,
125 raid6_have_altivec,
126 "altivecx$#",
127 0
128};
129
130#endif /* CONFIG_ALTIVEC */
diff --git a/lib/raid6/raid6int.uc b/lib/raid6/raid6int.uc
new file mode 100644
index 000000000000..d1e276a14fab
--- /dev/null
+++ b/lib/raid6/raid6int.uc
@@ -0,0 +1,117 @@
1/* -*- linux-c -*- ------------------------------------------------------- *
2 *
3 * Copyright 2002-2004 H. Peter Anvin - All Rights Reserved
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
8 * Boston MA 02111-1307, USA; either version 2 of the License, or
9 * (at your option) any later version; incorporated herein by reference.
10 *
11 * ----------------------------------------------------------------------- */
12
13/*
14 * raid6int$#.c
15 *
16 * $#-way unrolled portable integer math RAID-6 instruction set
17 *
18 * This file is postprocessed using unroll.awk
19 */
20
21#include <linux/raid/pq.h>
22
23/*
24 * This is the C data type to use
25 */
26
27/* Change this from BITS_PER_LONG if there is something better... */
28#if BITS_PER_LONG == 64
29# define NBYTES(x) ((x) * 0x0101010101010101UL)
30# define NSIZE 8
31# define NSHIFT 3
32# define NSTRING "64"
33typedef u64 unative_t;
34#else
35# define NBYTES(x) ((x) * 0x01010101U)
36# define NSIZE 4
37# define NSHIFT 2
38# define NSTRING "32"
39typedef u32 unative_t;
40#endif
41
42
43
44/*
45 * IA-64 wants insane amounts of unrolling. On other architectures that
46 * is just a waste of space.
47 */
48#if ($# <= 8) || defined(__ia64__)
49
50
51/*
52 * These sub-operations are separate inlines since they can sometimes be
53 * specially optimized using architecture-specific hacks.
54 */
55
56/*
57 * The SHLBYTE() operation shifts each byte left by 1, *not*
58 * rolling over into the next byte
59 */
60static inline __attribute_const__ unative_t SHLBYTE(unative_t v)
61{
62 unative_t vv;
63
64 vv = (v << 1) & NBYTES(0xfe);
65 return vv;
66}
67
68/*
69 * The MASK() operation returns 0xFF in any byte for which the high
70 * bit is 1, 0x00 for any byte for which the high bit is 0.
71 */
72static inline __attribute_const__ unative_t MASK(unative_t v)
73{
74 unative_t vv;
75
76 vv = v & NBYTES(0x80);
77 vv = (vv << 1) - (vv >> 7); /* Overflow on the top bit is OK */
78 return vv;
79}
80
81
82static void raid6_int$#_gen_syndrome(int disks, size_t bytes, void **ptrs)
83{
84 u8 **dptr = (u8 **)ptrs;
85 u8 *p, *q;
86 int d, z, z0;
87
88 unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
89
90 z0 = disks - 3; /* Highest data disk */
91 p = dptr[z0+1]; /* XOR parity */
92 q = dptr[z0+2]; /* RS syndrome */
93
94 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
95 wq$$ = wp$$ = *(unative_t *)&dptr[z0][d+$$*NSIZE];
96 for ( z = z0-1 ; z >= 0 ; z-- ) {
97 wd$$ = *(unative_t *)&dptr[z][d+$$*NSIZE];
98 wp$$ ^= wd$$;
99 w2$$ = MASK(wq$$);
100 w1$$ = SHLBYTE(wq$$);
101 w2$$ &= NBYTES(0x1d);
102 w1$$ ^= w2$$;
103 wq$$ = w1$$ ^ wd$$;
104 }
105 *(unative_t *)&p[d+NSIZE*$$] = wp$$;
106 *(unative_t *)&q[d+NSIZE*$$] = wq$$;
107 }
108}
109
110const struct raid6_calls raid6_intx$# = {
111 raid6_int$#_gen_syndrome,
112 NULL, /* always valid */
113 "int" NSTRING "x$#",
114 0
115};
116
117#endif
diff --git a/lib/raid6/raid6mmx.c b/lib/raid6/raid6mmx.c
new file mode 100644
index 000000000000..e7f6c13132bf
--- /dev/null
+++ b/lib/raid6/raid6mmx.c
@@ -0,0 +1,142 @@
1/* -*- linux-c -*- ------------------------------------------------------- *
2 *
3 * Copyright 2002 H. Peter Anvin - All Rights Reserved
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
8 * Boston MA 02111-1307, USA; either version 2 of the License, or
9 * (at your option) any later version; incorporated herein by reference.
10 *
11 * ----------------------------------------------------------------------- */
12
13/*
14 * raid6mmx.c
15 *
16 * MMX implementation of RAID-6 syndrome functions
17 */
18
19#if defined(__i386__) && !defined(__arch_um__)
20
21#include <linux/raid/pq.h>
22#include "raid6x86.h"
23
24/* Shared with raid6sse1.c */
25const struct raid6_mmx_constants {
26 u64 x1d;
27} raid6_mmx_constants = {
28 0x1d1d1d1d1d1d1d1dULL,
29};
30
31static int raid6_have_mmx(void)
32{
33 /* Not really "boot_cpu" but "all_cpus" */
34 return boot_cpu_has(X86_FEATURE_MMX);
35}
36
37/*
38 * Plain MMX implementation
39 */
40static void raid6_mmx1_gen_syndrome(int disks, size_t bytes, void **ptrs)
41{
42 u8 **dptr = (u8 **)ptrs;
43 u8 *p, *q;
44 int d, z, z0;
45
46 z0 = disks - 3; /* Highest data disk */
47 p = dptr[z0+1]; /* XOR parity */
48 q = dptr[z0+2]; /* RS syndrome */
49
50 kernel_fpu_begin();
51
52 asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
53 asm volatile("pxor %mm5,%mm5"); /* Zero temp */
54
55 for ( d = 0 ; d < bytes ; d += 8 ) {
56 asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
57 asm volatile("movq %mm2,%mm4"); /* Q[0] */
58 for ( z = z0-1 ; z >= 0 ; z-- ) {
59 asm volatile("movq %0,%%mm6" : : "m" (dptr[z][d]));
60 asm volatile("pcmpgtb %mm4,%mm5");
61 asm volatile("paddb %mm4,%mm4");
62 asm volatile("pand %mm0,%mm5");
63 asm volatile("pxor %mm5,%mm4");
64 asm volatile("pxor %mm5,%mm5");
65 asm volatile("pxor %mm6,%mm2");
66 asm volatile("pxor %mm6,%mm4");
67 }
68 asm volatile("movq %%mm2,%0" : "=m" (p[d]));
69 asm volatile("pxor %mm2,%mm2");
70 asm volatile("movq %%mm4,%0" : "=m" (q[d]));
71 asm volatile("pxor %mm4,%mm4");
72 }
73
74 kernel_fpu_end();
75}
76
77const struct raid6_calls raid6_mmxx1 = {
78 raid6_mmx1_gen_syndrome,
79 raid6_have_mmx,
80 "mmxx1",
81 0
82};
83
84/*
85 * Unrolled-by-2 MMX implementation
86 */
87static void raid6_mmx2_gen_syndrome(int disks, size_t bytes, void **ptrs)
88{
89 u8 **dptr = (u8 **)ptrs;
90 u8 *p, *q;
91 int d, z, z0;
92
93 z0 = disks - 3; /* Highest data disk */
94 p = dptr[z0+1]; /* XOR parity */
95 q = dptr[z0+2]; /* RS syndrome */
96
97 kernel_fpu_begin();
98
99 asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
100 asm volatile("pxor %mm5,%mm5"); /* Zero temp */
101 asm volatile("pxor %mm7,%mm7"); /* Zero temp */
102
103 for ( d = 0 ; d < bytes ; d += 16 ) {
104 asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
105 asm volatile("movq %0,%%mm3" : : "m" (dptr[z0][d+8]));
106 asm volatile("movq %mm2,%mm4"); /* Q[0] */
107 asm volatile("movq %mm3,%mm6"); /* Q[1] */
108 for ( z = z0-1 ; z >= 0 ; z-- ) {
109 asm volatile("pcmpgtb %mm4,%mm5");
110 asm volatile("pcmpgtb %mm6,%mm7");
111 asm volatile("paddb %mm4,%mm4");
112 asm volatile("paddb %mm6,%mm6");
113 asm volatile("pand %mm0,%mm5");
114 asm volatile("pand %mm0,%mm7");
115 asm volatile("pxor %mm5,%mm4");
116 asm volatile("pxor %mm7,%mm6");
117 asm volatile("movq %0,%%mm5" : : "m" (dptr[z][d]));
118 asm volatile("movq %0,%%mm7" : : "m" (dptr[z][d+8]));
119 asm volatile("pxor %mm5,%mm2");
120 asm volatile("pxor %mm7,%mm3");
121 asm volatile("pxor %mm5,%mm4");
122 asm volatile("pxor %mm7,%mm6");
123 asm volatile("pxor %mm5,%mm5");
124 asm volatile("pxor %mm7,%mm7");
125 }
126 asm volatile("movq %%mm2,%0" : "=m" (p[d]));
127 asm volatile("movq %%mm3,%0" : "=m" (p[d+8]));
128 asm volatile("movq %%mm4,%0" : "=m" (q[d]));
129 asm volatile("movq %%mm6,%0" : "=m" (q[d+8]));
130 }
131
132 kernel_fpu_end();
133}
134
135const struct raid6_calls raid6_mmxx2 = {
136 raid6_mmx2_gen_syndrome,
137 raid6_have_mmx,
138 "mmxx2",
139 0
140};
141
142#endif
diff --git a/lib/raid6/raid6recov.c b/lib/raid6/raid6recov.c
new file mode 100644
index 000000000000..2609f00e0d61
--- /dev/null
+++ b/lib/raid6/raid6recov.c
@@ -0,0 +1,132 @@
1/* -*- linux-c -*- ------------------------------------------------------- *
2 *
3 * Copyright 2002 H. Peter Anvin - All Rights Reserved
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
8 * Boston MA 02111-1307, USA; either version 2 of the License, or
9 * (at your option) any later version; incorporated herein by reference.
10 *
11 * ----------------------------------------------------------------------- */
12
13/*
14 * raid6recov.c
15 *
16 * RAID-6 data recovery in dual failure mode. In single failure mode,
17 * use the RAID-5 algorithm (or, in the case of Q failure, just reconstruct
18 * the syndrome.)
19 */
20
21#include <linux/raid/pq.h>
22
23/* Recover two failed data blocks. */
24void raid6_2data_recov(int disks, size_t bytes, int faila, int failb,
25 void **ptrs)
26{
27 u8 *p, *q, *dp, *dq;
28 u8 px, qx, db;
29 const u8 *pbmul; /* P multiplier table for B data */
30 const u8 *qmul; /* Q multiplier table (for both) */
31
32 p = (u8 *)ptrs[disks-2];
33 q = (u8 *)ptrs[disks-1];
34
35 /* Compute syndrome with zero for the missing data pages
36 Use the dead data pages as temporary storage for
37 delta p and delta q */
38 dp = (u8 *)ptrs[faila];
39 ptrs[faila] = (void *)raid6_empty_zero_page;
40 ptrs[disks-2] = dp;
41 dq = (u8 *)ptrs[failb];
42 ptrs[failb] = (void *)raid6_empty_zero_page;
43 ptrs[disks-1] = dq;
44
45 raid6_call.gen_syndrome(disks, bytes, ptrs);
46
47 /* Restore pointer table */
48 ptrs[faila] = dp;
49 ptrs[failb] = dq;
50 ptrs[disks-2] = p;
51 ptrs[disks-1] = q;
52
53 /* Now, pick the proper data tables */
54 pbmul = raid6_gfmul[raid6_gfexi[failb-faila]];
55 qmul = raid6_gfmul[raid6_gfinv[raid6_gfexp[faila]^raid6_gfexp[failb]]];
56
57 /* Now do it... */
58 while ( bytes-- ) {
59 px = *p ^ *dp;
60 qx = qmul[*q ^ *dq];
61 *dq++ = db = pbmul[px] ^ qx; /* Reconstructed B */
62 *dp++ = db ^ px; /* Reconstructed A */
63 p++; q++;
64 }
65}
66EXPORT_SYMBOL_GPL(raid6_2data_recov);
67
68/* Recover failure of one data block plus the P block */
69void raid6_datap_recov(int disks, size_t bytes, int faila, void **ptrs)
70{
71 u8 *p, *q, *dq;
72 const u8 *qmul; /* Q multiplier table */
73
74 p = (u8 *)ptrs[disks-2];
75 q = (u8 *)ptrs[disks-1];
76
77 /* Compute syndrome with zero for the missing data page
78 Use the dead data page as temporary storage for delta q */
79 dq = (u8 *)ptrs[faila];
80 ptrs[faila] = (void *)raid6_empty_zero_page;
81 ptrs[disks-1] = dq;
82
83 raid6_call.gen_syndrome(disks, bytes, ptrs);
84
85 /* Restore pointer table */
86 ptrs[faila] = dq;
87 ptrs[disks-1] = q;
88
89 /* Now, pick the proper data tables */
90 qmul = raid6_gfmul[raid6_gfinv[raid6_gfexp[faila]]];
91
92 /* Now do it... */
93 while ( bytes-- ) {
94 *p++ ^= *dq = qmul[*q ^ *dq];
95 q++; dq++;
96 }
97}
98EXPORT_SYMBOL_GPL(raid6_datap_recov);
99
100#ifndef __KERNEL__
101/* Testing only */
102
103/* Recover two failed blocks. */
104void raid6_dual_recov(int disks, size_t bytes, int faila, int failb, void **ptrs)
105{
106 if ( faila > failb ) {
107 int tmp = faila;
108 faila = failb;
109 failb = tmp;
110 }
111
112 if ( failb == disks-1 ) {
113 if ( faila == disks-2 ) {
114 /* P+Q failure. Just rebuild the syndrome. */
115 raid6_call.gen_syndrome(disks, bytes, ptrs);
116 } else {
117 /* data+Q failure. Reconstruct data from P,
118 then rebuild syndrome. */
119 /* NOT IMPLEMENTED - equivalent to RAID-5 */
120 }
121 } else {
122 if ( failb == disks-2 ) {
123 /* data+P failure. */
124 raid6_datap_recov(disks, bytes, faila, ptrs);
125 } else {
126 /* data+data failure. */
127 raid6_2data_recov(disks, bytes, faila, failb, ptrs);
128 }
129 }
130}
131
132#endif
diff --git a/lib/raid6/raid6sse1.c b/lib/raid6/raid6sse1.c
new file mode 100644
index 000000000000..b274dd5eab8f
--- /dev/null
+++ b/lib/raid6/raid6sse1.c
@@ -0,0 +1,162 @@
1/* -*- linux-c -*- ------------------------------------------------------- *
2 *
3 * Copyright 2002 H. Peter Anvin - All Rights Reserved
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
8 * Boston MA 02111-1307, USA; either version 2 of the License, or
9 * (at your option) any later version; incorporated herein by reference.
10 *
11 * ----------------------------------------------------------------------- */
12
13/*
14 * raid6sse1.c
15 *
16 * SSE-1/MMXEXT implementation of RAID-6 syndrome functions
17 *
18 * This is really an MMX implementation, but it requires SSE-1 or
19 * AMD MMXEXT for prefetch support and a few other features. The
20 * support for nontemporal memory accesses is enough to make this
21 * worthwhile as a separate implementation.
22 */
23
24#if defined(__i386__) && !defined(__arch_um__)
25
26#include <linux/raid/pq.h>
27#include "raid6x86.h"
28
29/* Defined in raid6mmx.c */
30extern const struct raid6_mmx_constants {
31 u64 x1d;
32} raid6_mmx_constants;
33
34static int raid6_have_sse1_or_mmxext(void)
35{
36 /* Not really boot_cpu but "all_cpus" */
37 return boot_cpu_has(X86_FEATURE_MMX) &&
38 (boot_cpu_has(X86_FEATURE_XMM) ||
39 boot_cpu_has(X86_FEATURE_MMXEXT));
40}
41
42/*
43 * Plain SSE1 implementation
44 */
45static void raid6_sse11_gen_syndrome(int disks, size_t bytes, void **ptrs)
46{
47 u8 **dptr = (u8 **)ptrs;
48 u8 *p, *q;
49 int d, z, z0;
50
51 z0 = disks - 3; /* Highest data disk */
52 p = dptr[z0+1]; /* XOR parity */
53 q = dptr[z0+2]; /* RS syndrome */
54
55 kernel_fpu_begin();
56
57 asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
58 asm volatile("pxor %mm5,%mm5"); /* Zero temp */
59
60 for ( d = 0 ; d < bytes ; d += 8 ) {
61 asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
62 asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
63 asm volatile("prefetchnta %0" : : "m" (dptr[z0-1][d]));
64 asm volatile("movq %mm2,%mm4"); /* Q[0] */
65 asm volatile("movq %0,%%mm6" : : "m" (dptr[z0-1][d]));
66 for ( z = z0-2 ; z >= 0 ; z-- ) {
67 asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
68 asm volatile("pcmpgtb %mm4,%mm5");
69 asm volatile("paddb %mm4,%mm4");
70 asm volatile("pand %mm0,%mm5");
71 asm volatile("pxor %mm5,%mm4");
72 asm volatile("pxor %mm5,%mm5");
73 asm volatile("pxor %mm6,%mm2");
74 asm volatile("pxor %mm6,%mm4");
75 asm volatile("movq %0,%%mm6" : : "m" (dptr[z][d]));
76 }
77 asm volatile("pcmpgtb %mm4,%mm5");
78 asm volatile("paddb %mm4,%mm4");
79 asm volatile("pand %mm0,%mm5");
80 asm volatile("pxor %mm5,%mm4");
81 asm volatile("pxor %mm5,%mm5");
82 asm volatile("pxor %mm6,%mm2");
83 asm volatile("pxor %mm6,%mm4");
84
85 asm volatile("movntq %%mm2,%0" : "=m" (p[d]));
86 asm volatile("movntq %%mm4,%0" : "=m" (q[d]));
87 }
88
89 asm volatile("sfence" : : : "memory");
90 kernel_fpu_end();
91}
92
93const struct raid6_calls raid6_sse1x1 = {
94 raid6_sse11_gen_syndrome,
95 raid6_have_sse1_or_mmxext,
96 "sse1x1",
97 1 /* Has cache hints */
98};
99
100/*
101 * Unrolled-by-2 SSE1 implementation
102 */
103static void raid6_sse12_gen_syndrome(int disks, size_t bytes, void **ptrs)
104{
105 u8 **dptr = (u8 **)ptrs;
106 u8 *p, *q;
107 int d, z, z0;
108
109 z0 = disks - 3; /* Highest data disk */
110 p = dptr[z0+1]; /* XOR parity */
111 q = dptr[z0+2]; /* RS syndrome */
112
113 kernel_fpu_begin();
114
115 asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
116 asm volatile("pxor %mm5,%mm5"); /* Zero temp */
117 asm volatile("pxor %mm7,%mm7"); /* Zero temp */
118
119 /* We uniformly assume a single prefetch covers at least 16 bytes */
120 for ( d = 0 ; d < bytes ; d += 16 ) {
121 asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
122 asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
123 asm volatile("movq %0,%%mm3" : : "m" (dptr[z0][d+8])); /* P[1] */
124 asm volatile("movq %mm2,%mm4"); /* Q[0] */
125 asm volatile("movq %mm3,%mm6"); /* Q[1] */
126 for ( z = z0-1 ; z >= 0 ; z-- ) {
127 asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
128 asm volatile("pcmpgtb %mm4,%mm5");
129 asm volatile("pcmpgtb %mm6,%mm7");
130 asm volatile("paddb %mm4,%mm4");
131 asm volatile("paddb %mm6,%mm6");
132 asm volatile("pand %mm0,%mm5");
133 asm volatile("pand %mm0,%mm7");
134 asm volatile("pxor %mm5,%mm4");
135 asm volatile("pxor %mm7,%mm6");
136 asm volatile("movq %0,%%mm5" : : "m" (dptr[z][d]));
137 asm volatile("movq %0,%%mm7" : : "m" (dptr[z][d+8]));
138 asm volatile("pxor %mm5,%mm2");
139 asm volatile("pxor %mm7,%mm3");
140 asm volatile("pxor %mm5,%mm4");
141 asm volatile("pxor %mm7,%mm6");
142 asm volatile("pxor %mm5,%mm5");
143 asm volatile("pxor %mm7,%mm7");
144 }
145 asm volatile("movntq %%mm2,%0" : "=m" (p[d]));
146 asm volatile("movntq %%mm3,%0" : "=m" (p[d+8]));
147 asm volatile("movntq %%mm4,%0" : "=m" (q[d]));
148 asm volatile("movntq %%mm6,%0" : "=m" (q[d+8]));
149 }
150
151 asm volatile("sfence" : :: "memory");
152 kernel_fpu_end();
153}
154
155const struct raid6_calls raid6_sse1x2 = {
156 raid6_sse12_gen_syndrome,
157 raid6_have_sse1_or_mmxext,
158 "sse1x2",
159 1 /* Has cache hints */
160};
161
162#endif
diff --git a/lib/raid6/raid6sse2.c b/lib/raid6/raid6sse2.c
new file mode 100644
index 000000000000..6ed6c6c0389f
--- /dev/null
+++ b/lib/raid6/raid6sse2.c
@@ -0,0 +1,262 @@
1/* -*- linux-c -*- ------------------------------------------------------- *
2 *
3 * Copyright 2002 H. Peter Anvin - All Rights Reserved
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
8 * Boston MA 02111-1307, USA; either version 2 of the License, or
9 * (at your option) any later version; incorporated herein by reference.
10 *
11 * ----------------------------------------------------------------------- */
12
13/*
14 * raid6sse2.c
15 *
16 * SSE-2 implementation of RAID-6 syndrome functions
17 *
18 */
19
20#if (defined(__i386__) || defined(__x86_64__)) && !defined(__arch_um__)
21
22#include <linux/raid/pq.h>
23#include "raid6x86.h"
24
25static const struct raid6_sse_constants {
26 u64 x1d[2];
27} raid6_sse_constants __attribute__((aligned(16))) = {
28 { 0x1d1d1d1d1d1d1d1dULL, 0x1d1d1d1d1d1d1d1dULL },
29};
30
31static int raid6_have_sse2(void)
32{
33 /* Not really boot_cpu but "all_cpus" */
34 return boot_cpu_has(X86_FEATURE_MMX) &&
35 boot_cpu_has(X86_FEATURE_FXSR) &&
36 boot_cpu_has(X86_FEATURE_XMM) &&
37 boot_cpu_has(X86_FEATURE_XMM2);
38}
39
40/*
41 * Plain SSE2 implementation
42 */
43static void raid6_sse21_gen_syndrome(int disks, size_t bytes, void **ptrs)
44{
45 u8 **dptr = (u8 **)ptrs;
46 u8 *p, *q;
47 int d, z, z0;
48
49 z0 = disks - 3; /* Highest data disk */
50 p = dptr[z0+1]; /* XOR parity */
51 q = dptr[z0+2]; /* RS syndrome */
52
53 kernel_fpu_begin();
54
55 asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
56 asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */
57
58 for ( d = 0 ; d < bytes ; d += 16 ) {
59 asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
60 asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d])); /* P[0] */
61 asm volatile("prefetchnta %0" : : "m" (dptr[z0-1][d]));
62 asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */
63 asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z0-1][d]));
64 for ( z = z0-2 ; z >= 0 ; z-- ) {
65 asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
66 asm volatile("pcmpgtb %xmm4,%xmm5");
67 asm volatile("paddb %xmm4,%xmm4");
68 asm volatile("pand %xmm0,%xmm5");
69 asm volatile("pxor %xmm5,%xmm4");
70 asm volatile("pxor %xmm5,%xmm5");
71 asm volatile("pxor %xmm6,%xmm2");
72 asm volatile("pxor %xmm6,%xmm4");
73 asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z][d]));
74 }
75 asm volatile("pcmpgtb %xmm4,%xmm5");
76 asm volatile("paddb %xmm4,%xmm4");
77 asm volatile("pand %xmm0,%xmm5");
78 asm volatile("pxor %xmm5,%xmm4");
79 asm volatile("pxor %xmm5,%xmm5");
80 asm volatile("pxor %xmm6,%xmm2");
81 asm volatile("pxor %xmm6,%xmm4");
82
83 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
84 asm volatile("pxor %xmm2,%xmm2");
85 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
86 asm volatile("pxor %xmm4,%xmm4");
87 }
88
89 asm volatile("sfence" : : : "memory");
90 kernel_fpu_end();
91}
92
93const struct raid6_calls raid6_sse2x1 = {
94 raid6_sse21_gen_syndrome,
95 raid6_have_sse2,
96 "sse2x1",
97 1 /* Has cache hints */
98};
99
100/*
101 * Unrolled-by-2 SSE2 implementation
102 */
103static void raid6_sse22_gen_syndrome(int disks, size_t bytes, void **ptrs)
104{
105 u8 **dptr = (u8 **)ptrs;
106 u8 *p, *q;
107 int d, z, z0;
108
109 z0 = disks - 3; /* Highest data disk */
110 p = dptr[z0+1]; /* XOR parity */
111 q = dptr[z0+2]; /* RS syndrome */
112
113 kernel_fpu_begin();
114
115 asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
116 asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */
117 asm volatile("pxor %xmm7,%xmm7"); /* Zero temp */
118
119 /* We uniformly assume a single prefetch covers at least 32 bytes */
120 for ( d = 0 ; d < bytes ; d += 32 ) {
121 asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
122 asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d])); /* P[0] */
123 asm volatile("movdqa %0,%%xmm3" : : "m" (dptr[z0][d+16])); /* P[1] */
124 asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */
125 asm volatile("movdqa %xmm3,%xmm6"); /* Q[1] */
126 for ( z = z0-1 ; z >= 0 ; z-- ) {
127 asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
128 asm volatile("pcmpgtb %xmm4,%xmm5");
129 asm volatile("pcmpgtb %xmm6,%xmm7");
130 asm volatile("paddb %xmm4,%xmm4");
131 asm volatile("paddb %xmm6,%xmm6");
132 asm volatile("pand %xmm0,%xmm5");
133 asm volatile("pand %xmm0,%xmm7");
134 asm volatile("pxor %xmm5,%xmm4");
135 asm volatile("pxor %xmm7,%xmm6");
136 asm volatile("movdqa %0,%%xmm5" : : "m" (dptr[z][d]));
137 asm volatile("movdqa %0,%%xmm7" : : "m" (dptr[z][d+16]));
138 asm volatile("pxor %xmm5,%xmm2");
139 asm volatile("pxor %xmm7,%xmm3");
140 asm volatile("pxor %xmm5,%xmm4");
141 asm volatile("pxor %xmm7,%xmm6");
142 asm volatile("pxor %xmm5,%xmm5");
143 asm volatile("pxor %xmm7,%xmm7");
144 }
145 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
146 asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
147 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
148 asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
149 }
150
151 asm volatile("sfence" : : : "memory");
152 kernel_fpu_end();
153}
154
155const struct raid6_calls raid6_sse2x2 = {
156 raid6_sse22_gen_syndrome,
157 raid6_have_sse2,
158 "sse2x2",
159 1 /* Has cache hints */
160};
161
162#endif
163
164#if defined(__x86_64__) && !defined(__arch_um__)
165
166/*
167 * Unrolled-by-4 SSE2 implementation
168 */
169static void raid6_sse24_gen_syndrome(int disks, size_t bytes, void **ptrs)
170{
171 u8 **dptr = (u8 **)ptrs;
172 u8 *p, *q;
173 int d, z, z0;
174
175 z0 = disks - 3; /* Highest data disk */
176 p = dptr[z0+1]; /* XOR parity */
177 q = dptr[z0+2]; /* RS syndrome */
178
179 kernel_fpu_begin();
180
181 asm volatile("movdqa %0,%%xmm0" :: "m" (raid6_sse_constants.x1d[0]));
182 asm volatile("pxor %xmm2,%xmm2"); /* P[0] */
183 asm volatile("pxor %xmm3,%xmm3"); /* P[1] */
184 asm volatile("pxor %xmm4,%xmm4"); /* Q[0] */
185 asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */
186 asm volatile("pxor %xmm6,%xmm6"); /* Q[1] */
187 asm volatile("pxor %xmm7,%xmm7"); /* Zero temp */
188 asm volatile("pxor %xmm10,%xmm10"); /* P[2] */
189 asm volatile("pxor %xmm11,%xmm11"); /* P[3] */
190 asm volatile("pxor %xmm12,%xmm12"); /* Q[2] */
191 asm volatile("pxor %xmm13,%xmm13"); /* Zero temp */
192 asm volatile("pxor %xmm14,%xmm14"); /* Q[3] */
193 asm volatile("pxor %xmm15,%xmm15"); /* Zero temp */
194
195 for ( d = 0 ; d < bytes ; d += 64 ) {
196 for ( z = z0 ; z >= 0 ; z-- ) {
197 /* The second prefetch seems to improve performance... */
198 asm volatile("prefetchnta %0" :: "m" (dptr[z][d]));
199 asm volatile("prefetchnta %0" :: "m" (dptr[z][d+32]));
200 asm volatile("pcmpgtb %xmm4,%xmm5");
201 asm volatile("pcmpgtb %xmm6,%xmm7");
202 asm volatile("pcmpgtb %xmm12,%xmm13");
203 asm volatile("pcmpgtb %xmm14,%xmm15");
204 asm volatile("paddb %xmm4,%xmm4");
205 asm volatile("paddb %xmm6,%xmm6");
206 asm volatile("paddb %xmm12,%xmm12");
207 asm volatile("paddb %xmm14,%xmm14");
208 asm volatile("pand %xmm0,%xmm5");
209 asm volatile("pand %xmm0,%xmm7");
210 asm volatile("pand %xmm0,%xmm13");
211 asm volatile("pand %xmm0,%xmm15");
212 asm volatile("pxor %xmm5,%xmm4");
213 asm volatile("pxor %xmm7,%xmm6");
214 asm volatile("pxor %xmm13,%xmm12");
215 asm volatile("pxor %xmm15,%xmm14");
216 asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
217 asm volatile("movdqa %0,%%xmm7" :: "m" (dptr[z][d+16]));
218 asm volatile("movdqa %0,%%xmm13" :: "m" (dptr[z][d+32]));
219 asm volatile("movdqa %0,%%xmm15" :: "m" (dptr[z][d+48]));
220 asm volatile("pxor %xmm5,%xmm2");
221 asm volatile("pxor %xmm7,%xmm3");
222 asm volatile("pxor %xmm13,%xmm10");
223 asm volatile("pxor %xmm15,%xmm11");
224 asm volatile("pxor %xmm5,%xmm4");
225 asm volatile("pxor %xmm7,%xmm6");
226 asm volatile("pxor %xmm13,%xmm12");
227 asm volatile("pxor %xmm15,%xmm14");
228 asm volatile("pxor %xmm5,%xmm5");
229 asm volatile("pxor %xmm7,%xmm7");
230 asm volatile("pxor %xmm13,%xmm13");
231 asm volatile("pxor %xmm15,%xmm15");
232 }
233 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
234 asm volatile("pxor %xmm2,%xmm2");
235 asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
236 asm volatile("pxor %xmm3,%xmm3");
237 asm volatile("movntdq %%xmm10,%0" : "=m" (p[d+32]));
238 asm volatile("pxor %xmm10,%xmm10");
239 asm volatile("movntdq %%xmm11,%0" : "=m" (p[d+48]));
240 asm volatile("pxor %xmm11,%xmm11");
241 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
242 asm volatile("pxor %xmm4,%xmm4");
243 asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
244 asm volatile("pxor %xmm6,%xmm6");
245 asm volatile("movntdq %%xmm12,%0" : "=m" (q[d+32]));
246 asm volatile("pxor %xmm12,%xmm12");
247 asm volatile("movntdq %%xmm14,%0" : "=m" (q[d+48]));
248 asm volatile("pxor %xmm14,%xmm14");
249 }
250
251 asm volatile("sfence" : : : "memory");
252 kernel_fpu_end();
253}
254
255const struct raid6_calls raid6_sse2x4 = {
256 raid6_sse24_gen_syndrome,
257 raid6_have_sse2,
258 "sse2x4",
259 1 /* Has cache hints */
260};
261
262#endif
diff --git a/lib/raid6/raid6test/Makefile b/lib/raid6/raid6test/Makefile
new file mode 100644
index 000000000000..2874cbef529d
--- /dev/null
+++ b/lib/raid6/raid6test/Makefile
@@ -0,0 +1,75 @@
1#
2# This is a simple Makefile to test some of the RAID-6 code
3# from userspace.
4#
5
6CC = gcc
7OPTFLAGS = -O2 # Adjust as desired
8CFLAGS = -I.. -I ../../../include -g $(OPTFLAGS)
9LD = ld
10AWK = awk
11AR = ar
12RANLIB = ranlib
13
14.c.o:
15 $(CC) $(CFLAGS) -c -o $@ $<
16
17%.c: ../%.c
18 cp -f $< $@
19
20%.uc: ../%.uc
21 cp -f $< $@
22
23all: raid6.a raid6test
24
25raid6.a: raid6int1.o raid6int2.o raid6int4.o raid6int8.o raid6int16.o \
26 raid6int32.o \
27 raid6mmx.o raid6sse1.o raid6sse2.o \
28 raid6altivec1.o raid6altivec2.o raid6altivec4.o raid6altivec8.o \
29 raid6recov.o raid6algos.o \
30 raid6tables.o
31 rm -f $@
32 $(AR) cq $@ $^
33 $(RANLIB) $@
34
35raid6test: test.c raid6.a
36 $(CC) $(CFLAGS) -o raid6test $^
37
38raid6altivec1.c: raid6altivec.uc ../unroll.awk
39 $(AWK) ../unroll.awk -vN=1 < raid6altivec.uc > $@
40
41raid6altivec2.c: raid6altivec.uc ../unroll.awk
42 $(AWK) ../unroll.awk -vN=2 < raid6altivec.uc > $@
43
44raid6altivec4.c: raid6altivec.uc ../unroll.awk
45 $(AWK) ../unroll.awk -vN=4 < raid6altivec.uc > $@
46
47raid6altivec8.c: raid6altivec.uc ../unroll.awk
48 $(AWK) ../unroll.awk -vN=8 < raid6altivec.uc > $@
49
50raid6int1.c: raid6int.uc ../unroll.awk
51 $(AWK) ../unroll.awk -vN=1 < raid6int.uc > $@
52
53raid6int2.c: raid6int.uc ../unroll.awk
54 $(AWK) ../unroll.awk -vN=2 < raid6int.uc > $@
55
56raid6int4.c: raid6int.uc ../unroll.awk
57 $(AWK) ../unroll.awk -vN=4 < raid6int.uc > $@
58
59raid6int8.c: raid6int.uc ../unroll.awk
60 $(AWK) ../unroll.awk -vN=8 < raid6int.uc > $@
61
62raid6int16.c: raid6int.uc ../unroll.awk
63 $(AWK) ../unroll.awk -vN=16 < raid6int.uc > $@
64
65raid6int32.c: raid6int.uc ../unroll.awk
66 $(AWK) ../unroll.awk -vN=32 < raid6int.uc > $@
67
68raid6tables.c: mktables
69 ./mktables > raid6tables.c
70
71clean:
72 rm -f *.o *.a mktables mktables.c raid6int.uc raid6*.c raid6test
73
74spotless: clean
75 rm -f *~
diff --git a/lib/raid6/raid6test/test.c b/lib/raid6/raid6test/test.c
new file mode 100644
index 000000000000..7a930318b17d
--- /dev/null
+++ b/lib/raid6/raid6test/test.c
@@ -0,0 +1,124 @@
1/* -*- linux-c -*- ------------------------------------------------------- *
2 *
3 * Copyright 2002-2007 H. Peter Anvin - All Rights Reserved
4 *
5 * This file is part of the Linux kernel, and is made available under
6 * the terms of the GNU General Public License version 2 or (at your
7 * option) any later version; incorporated herein by reference.
8 *
9 * ----------------------------------------------------------------------- */
10
11/*
12 * raid6test.c
13 *
14 * Test RAID-6 recovery with various algorithms
15 */
16
17#include <stdlib.h>
18#include <stdio.h>
19#include <string.h>
20#include <linux/raid/pq.h>
21
22#define NDISKS 16 /* Including P and Q */
23
24const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
25struct raid6_calls raid6_call;
26
27char *dataptrs[NDISKS];
28char data[NDISKS][PAGE_SIZE];
29char recovi[PAGE_SIZE], recovj[PAGE_SIZE];
30
31static void makedata(void)
32{
33 int i, j;
34
35 for (i = 0; i < NDISKS; i++) {
36 for (j = 0; j < PAGE_SIZE; j++)
37 data[i][j] = rand();
38
39 dataptrs[i] = data[i];
40 }
41}
42
43static char disk_type(int d)
44{
45 switch (d) {
46 case NDISKS-2:
47 return 'P';
48 case NDISKS-1:
49 return 'Q';
50 default:
51 return 'D';
52 }
53}
54
55static int test_disks(int i, int j)
56{
57 int erra, errb;
58
59 memset(recovi, 0xf0, PAGE_SIZE);
60 memset(recovj, 0xba, PAGE_SIZE);
61
62 dataptrs[i] = recovi;
63 dataptrs[j] = recovj;
64
65 raid6_dual_recov(NDISKS, PAGE_SIZE, i, j, (void **)&dataptrs);
66
67 erra = memcmp(data[i], recovi, PAGE_SIZE);
68 errb = memcmp(data[j], recovj, PAGE_SIZE);
69
70 if (i < NDISKS-2 && j == NDISKS-1) {
71 /* We don't implement the DQ failure scenario, since it's
72 equivalent to a RAID-5 failure (XOR, then recompute Q) */
73 erra = errb = 0;
74 } else {
75 printf("algo=%-8s faila=%3d(%c) failb=%3d(%c) %s\n",
76 raid6_call.name,
77 i, disk_type(i),
78 j, disk_type(j),
79 (!erra && !errb) ? "OK" :
80 !erra ? "ERRB" :
81 !errb ? "ERRA" : "ERRAB");
82 }
83
84 dataptrs[i] = data[i];
85 dataptrs[j] = data[j];
86
87 return erra || errb;
88}
89
90int main(int argc, char *argv[])
91{
92 const struct raid6_calls *const *algo;
93 int i, j;
94 int err = 0;
95
96 makedata();
97
98 for (algo = raid6_algos; *algo; algo++) {
99 if (!(*algo)->valid || (*algo)->valid()) {
100 raid6_call = **algo;
101
102 /* Nuke syndromes */
103 memset(data[NDISKS-2], 0xee, 2*PAGE_SIZE);
104
105 /* Generate assumed good syndrome */
106 raid6_call.gen_syndrome(NDISKS, PAGE_SIZE,
107 (void **)&dataptrs);
108
109 for (i = 0; i < NDISKS-1; i++)
110 for (j = i+1; j < NDISKS; j++)
111 err += test_disks(i, j);
112 }
113 printf("\n");
114 }
115
116 printf("\n");
117 /* Pick the best algorithm test */
118 raid6_select_algo();
119
120 if (err)
121 printf("\n*** ERRORS FOUND ***\n");
122
123 return err;
124}
diff --git a/lib/raid6/raid6x86.h b/lib/raid6/raid6x86.h
new file mode 100644
index 000000000000..4c22c1568558
--- /dev/null
+++ b/lib/raid6/raid6x86.h
@@ -0,0 +1,61 @@
1/* ----------------------------------------------------------------------- *
2 *
3 * Copyright 2002-2004 H. Peter Anvin - All Rights Reserved
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
8 * Boston MA 02111-1307, USA; either version 2 of the License, or
9 * (at your option) any later version; incorporated herein by reference.
10 *
11 * ----------------------------------------------------------------------- */
12
13/*
14 * raid6x86.h
15 *
16 * Definitions common to x86 and x86-64 RAID-6 code only
17 */
18
19#ifndef LINUX_RAID_RAID6X86_H
20#define LINUX_RAID_RAID6X86_H
21
22#if (defined(__i386__) || defined(__x86_64__)) && !defined(__arch_um__)
23
24#ifdef __KERNEL__ /* Real code */
25
26#include <asm/i387.h>
27
28#else /* Dummy code for user space testing */
29
30static inline void kernel_fpu_begin(void)
31{
32}
33
34static inline void kernel_fpu_end(void)
35{
36}
37
38#define X86_FEATURE_MMX (0*32+23) /* Multimedia Extensions */
39#define X86_FEATURE_FXSR (0*32+24) /* FXSAVE and FXRSTOR instructions
40 * (fast save and restore) */
41#define X86_FEATURE_XMM (0*32+25) /* Streaming SIMD Extensions */
42#define X86_FEATURE_XMM2 (0*32+26) /* Streaming SIMD Extensions-2 */
43#define X86_FEATURE_MMXEXT (1*32+22) /* AMD MMX extensions */
44
45/* Should work well enough on modern CPUs for testing */
46static inline int boot_cpu_has(int flag)
47{
48 u32 eax = (flag >> 5) ? 0x80000001 : 1;
49 u32 edx;
50
51 asm volatile("cpuid"
52 : "+a" (eax), "=d" (edx)
53 : : "ecx", "ebx");
54
55 return (edx >> (flag & 31)) & 1;
56}
57
58#endif /* ndef __KERNEL__ */
59
60#endif
61#endif
diff --git a/lib/raid6/unroll.awk b/lib/raid6/unroll.awk
new file mode 100644
index 000000000000..c6aa03631df8
--- /dev/null
+++ b/lib/raid6/unroll.awk
@@ -0,0 +1,20 @@
1
2# This filter requires one command line option of form -vN=n
3# where n must be a decimal number.
4#
5# Repeat each input line containing $$ n times, replacing $$ with 0...n-1.
6# Replace each $# with n, and each $* with a single $.
7
8BEGIN {
9 n = N + 0
10}
11{
12 if (/\$\$/) { rep = n } else { rep = 1 }
13 for (i = 0; i < rep; ++i) {
14 tmp = $0
15 gsub(/\$\$/, i, tmp)
16 gsub(/\$\#/, n, tmp)
17 gsub(/\$\*/, "$", tmp)
18 print tmp
19 }
20}
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-15 04:54:13 -0500