[PATCH] RAID-6

From: "H. Peter Anvin" <hpa@zytor.com>

RAID6 implementation.  See Kconfig help for usage details.

The next release of `mdadm' has raid6 userspace support.

drivers/md/Kconfig

@@ -107,6 +107,39 @@ config MD_RAID5
 
 	  If unsure, say Y.
 
+config MD_RAID6
+	tristate "RAID-6 mode (EXPERIMENTAL)"
+	depends on BLK_DEV_MD && EXPERIMENTAL
+	---help---
+	  WARNING: RAID-6 is currently highly experimental.  If you
+	  use it, there is no guarantee whatsoever that it won't
+	  destroy your data, eat your disk drives, insult your mother,
+	  or re-appoint George W. Bush.
+
+	  A RAID-6 set of N drives with a capacity of C MB per drive
+	  provides the capacity of C * (N - 2) MB, and protects
+	  against a failure of any two drives. For a given sector
+	  (row) number, (N - 2) drives contain data sectors, and two
+	  drives contains two independent redundancy syndromes.  Like
+	  RAID-5, RAID-6 distributes the syndromes across the drives
+	  in one of the available parity distribution methods.
+
+	  RAID-6 currently requires a specially patched version of
+	  mdadm; the patch is available at:
+
+	  ftp://ftp.kernel.org/pub/linux/kernel/people/hpa/
+
+	  ... and the mdadm source code at ...
+
+	  ftp://ftp.kernel.org/pub/linux/utils/raid/mdadm/
+
+	  If you want to use such a RAID-6 set, say Y.  To compile
+	  this code as a module, choose M here: the module will be
+	  called raid6.
+
+
+	  If unsure, say N.
+
 config MD_MULTIPATH
 	tristate "Multipath I/O support"
 	depends on BLK_DEV_MD

drivers/md/Makefile

@@ -4,6 +4,11 @@
 
 dm-mod-objs	:= dm.o dm-table.o dm-target.o dm-linear.o dm-stripe.o \
 		   dm-ioctl.o
+raid6-objs	:= raid6main.o raid6algos.o raid6recov.o raid6tables.o \
+		   raid6int1.o raid6int2.o raid6int4.o \
+		   raid6int8.o raid6int16.o raid6int32.o \
+		   raid6mmx.o raid6sse1.o raid6sse2.o
+host-progs	:= mktables
 
 # Note: link order is important.  All raid personalities
 # and xor.o must come before md.o, as they each initialise 
@@ -14,6 +19,31 @@ obj-$(CONFIG_MD_LINEAR)		+= linear.o
 obj-$(CONFIG_MD_RAID0)		+= raid0.o
 obj-$(CONFIG_MD_RAID1)		+= raid1.o
 obj-$(CONFIG_MD_RAID5)		+= raid5.o xor.o
+obj-$(CONFIG_MD_RAID6)		+= raid6.o xor.o
 obj-$(CONFIG_MD_MULTIPATH)	+= multipath.o
 obj-$(CONFIG_BLK_DEV_MD)	+= md.o
 obj-$(CONFIG_BLK_DEV_DM)	+= dm-mod.o
+
+# Files generated that shall be removed upon make clean
+clean-files := raid6int*.c raid6tables.c mktables
+
+$(obj)/raid6int1.c:   $(src)/raid6int.uc $(src)/unroll.pl
+	$(PERL) $(src)/unroll.pl 1 < $< > $@ || ( rm -f $@ && exit 1 )
+
+$(obj)/raid6int2.c:   $(src)/raid6int.uc $(src)/unroll.pl
+	$(PERL) $(src)/unroll.pl 2 < $< > $@ || ( rm -f $@ && exit 1 )
+
+$(obj)/raid6int4.c:   $(src)/raid6int.uc $(src)/unroll.pl
+	$(PERL) $(src)/unroll.pl 4 < $< > $@ || ( rm -f $@ && exit 1 )
+
+$(obj)/raid6int8.c:   $(src)/raid6int.uc $(src)/unroll.pl
+	$(PERL) $(src)/unroll.pl 8 < $< > $@ || ( rm -f $@ && exit 1 )
+
+$(obj)/raid6int16.c:  $(src)/raid6int.uc $(src)/unroll.pl
+	$(PERL) $(src)/unroll.pl 16 < $< > $@ || ( rm -f $@ && exit 1 )
+
+$(obj)/raid6int32.c:  $(src)/raid6int.uc $(src)/unroll.pl
+	$(PERL) $(src)/unroll.pl 32 < $< > $@ || ( rm -f $@ && exit 1 )
+
+$(obj)/raid6tables.c: $(obj)/mktables
+	$(obj)/mktables > $@ || ( rm -f $@ && exit 1 )

drivers/md/md.c

@@ -7,6 +7,7 @@
    Changes:
 
    - RAID-1/RAID-5 extensions by Miguel de Icaza, Gadi Oxman, Ingo Molnar
+   - RAID-6 extensions by H. Peter Anvin <hpa@zytor.com>
    - boot support for linear and striped mode by Harald Hoyer <HarryH@Royal.Net>
    - kerneld support by Boris Tobotras <boris@xtalk.msk.su>
    - kmod support by: Cyrus Durgin
@@ -1435,8 +1436,9 @@ static int analyze_sbs(mddev_t * mddev)
 		goto abort;
 	}
 
-	if ((mddev->recovery_cp != MaxSector) && ((mddev->level == 1) ||
-			(mddev->level == 4) || (mddev->level == 5)))
+	if ((mddev->recovery_cp != MaxSector) &&
+	    ((mddev->level == 1) ||
+	     ((mddev->level >= 4) && (mddev->level <= 6))))
 		printk(KERN_ERR "md: md%d: raid array is not clean"
 		       " -- starting background reconstruction\n",
 		       mdidx(mddev));
@@ -3014,7 +3016,9 @@ static struct file_operations md_seq_fops = {
 int register_md_personality(int pnum, mdk_personality_t *p)
 {
 	if (pnum >= MAX_PERSONALITY) {
-		MD_BUG();
+		printk(KERN_ERR
+		       "md: tried to install personality %s as nr %d, but max is %lu\n",
+		       p->name, pnum, MAX_PERSONALITY-1);
 		return -EINVAL;
 	}
 

drivers/md/mktables.c

+#ident "$Id: mktables.c,v 1.2 2002/12/12 22:41:27 hpa Exp $"
+/* ----------------------------------------------------------------------- *
+ *
+ *   Copyright 2002 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+/*
+ * mktables.c
+ *
+ * Make RAID-6 tables.  This is a host user space program to be run at
+ * compile time.
+ */
+
+#include <stdio.h>
+#include <string.h>
+#include <inttypes.h>
+#include <stdlib.h>
+#include <time.h>
+
+static uint8_t gfmul(uint8_t a, uint8_t b)
+{
+  uint8_t v = 0;
+
+  while ( b ) {
+    if ( b & 1 ) v ^= a;
+    a = (a << 1) ^ (a & 0x80 ? 0x1d : 0);
+    b >>= 1;
+  }
+  return v;
+}
+
+static uint8_t gfpow(uint8_t a, int b)
+{
+  uint8_t v = 1;
+
+  b %= 255;
+  if ( b < 0 )
+    b += 255;
+
+  while ( b ) {
+    if ( b & 1 ) v = gfmul(v,a);
+    a = gfmul(a,a);
+    b >>= 1;
+  }
+  return v;
+}
+
+int main(int argc, char *argv[])
+{
+  int i, j, k;
+  uint8_t v;
+  uint8_t exptbl[256], invtbl[256];
+
+  printf("#include \"raid6.h\"\n");
+
+  /* Compute multiplication table */
+  printf("\nconst u8  __attribute__((aligned(256)))\n"
+	 "raid6_gfmul[256][256] =\n"
+	 "{\n");
+  for ( i = 0 ; i < 256 ; i++ ) {
+    printf("\t{\n");
+    for ( j = 0 ; j < 256 ; j += 8 ) {
+      printf("\t\t");
+      for ( k = 0 ; k < 8 ; k++ ) {
+	printf("0x%02x, ", gfmul(i,j+k));
+      }
+      printf("\n");
+    }
+    printf("\t},\n");
+  }
+  printf("};\n");
+
+  /* Compute power-of-2 table (exponent) */
+  v = 1;
+  printf("\nconst u8 __attribute__((aligned(256)))\n"
+	 "raid6_gfexp[256] =\n"
+	 "{\n");
+  for ( i = 0 ; i < 256 ; i += 8 ) {
+    printf("\t");
+    for ( j = 0 ; j < 8 ; j++ ) {
+      exptbl[i+j] = v;
+      printf("0x%02x, ", v);
+      v = gfmul(v,2);
+      if ( v == 1 ) v = 0;	/* For entry 255, not a real entry */
+    }
+    printf("\n");
+  }
+  printf("};\n");
+
+  /* Compute inverse table x^-1 == x^254 */
+  printf("\nconst u8 __attribute__((aligned(256)))\n"
+	 "raid6_gfinv[256] =\n"
+	 "{\n");
+  for ( i = 0 ; i < 256 ; i += 8 ) {
+    printf("\t");
+    for ( j = 0 ; j < 8 ; j++ ) {
+      invtbl[i+j] = v = gfpow(i+j,254);
+      printf("0x%02x, ", v);
+    }
+    printf("\n");
+  }
+  printf("};\n");
+
+  /* Compute inv(2^x + 1) (exponent-xor-inverse) table */
+  printf("\nconst u8 __attribute__((aligned(256)))\n"
+	 "raid6_gfexi[256] =\n"
+	 "{\n");
+  for ( i = 0 ; i < 256 ; i += 8 ) {
+    printf("\t");
+    for ( j = 0 ; j < 8 ; j++ ) {
+      printf("0x%02x, ", invtbl[exptbl[i+j]^1]);
+    }
+    printf("\n");
+  }
+  printf("};\n\n");
+
+  return 0;
+}

drivers/md/raid6.h

+/* -*- linux-c -*- ------------------------------------------------------- *
+ *
+ *   Copyright 2003 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+#ifndef LINUX_RAID_RAID6_H
+#define LINUX_RAID_RAID6_H
+
+#ifdef __KERNEL__
+
+/* Set to 1 to use kernel-wide empty_zero_page */
+#define RAID6_USE_EMPTY_ZERO_PAGE 0
+
+#include <linux/module.h>
+#include <linux/stddef.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+#include <linux/errno.h>
+#include <linux/mempool.h>
+#include <linux/list.h>
+#include <linux/vmalloc.h>
+#include <linux/raid/md.h>
+#include <linux/raid/raid5.h>
+
+typedef raid5_conf_t raid6_conf_t; /* Same configuration */
+
+/* Additional compute_parity mode -- updates the parity w/o LOCKING */
+#define UPDATE_PARITY	4
+
+/* We need a pre-zeroed page... if we don't want to use the kernel-provided
+   one define it here */
+#if RAID6_USE_EMPTY_ZERO_PAGE
+# define raid6_empty_zero_page empty_zero_page
+#else
+extern const char raid6_empty_zero_page[PAGE_SIZE];
+#endif
+
+#else /* ! __KERNEL__ */
+/* Used for testing in user space */
+
+#include <stddef.h>
+#include <sys/types.h>
+#include <inttypes.h>
+#include <errno.h>
+#include <sys/mman.h>
+
+typedef uint8_t  u8;
+typedef uint16_t u16;
+typedef uint32_t u32;
+typedef uint64_t u64;
+
+#ifndef PAGE_SIZE
+# define PAGE_SIZE 4096
+#endif
+extern const char raid6_empty_zero_page[PAGE_SIZE];
+
+#define __init
+#define __exit
+
+#define preempt_enable()
+#define preempt_disable()
+
+#endif /* __KERNEL__ */
+
+/* Change this from BITS_PER_LONG if there is something better... */
+#if BITS_PER_LONG == 64
+# define NBYTES(x) ((x) * 0x0101010101010101UL)
+# define NSIZE  8
+# define NSHIFT 3
+# define NSTRING "64"
+typedef u64 unative_t;
+#else
+# define NBYTES(x) ((x) * 0x01010101U)
+# define NSIZE  4
+# define NSHIFT 2
+# define NSTRING "32"
+typedef u32 unative_t;
+#endif
+
+/* Routine choices */
+struct raid6_calls {
+	void (*gen_syndrome)(int, size_t, void **);
+	int  (*valid)(void);	/* Returns 1 if this routine set is usable */
+	const char *name;	/* Name of this routine set */
+	int prefer;		/* Has special performance attribute */
+};
+
+/* Selected algorithm */
+extern struct raid6_calls raid6_call;
+
+/* Algorithm list */
+extern const struct raid6_calls * const raid6_algos[];
+int raid6_select_algo(void);
+
+/* Return values from chk_syndrome */
+#define RAID6_OK	0
+#define RAID6_P_BAD	1
+#define RAID6_Q_BAD	2
+#define RAID6_PQ_BAD	3
+
+/* Galois field tables */
+extern const u8 raid6_gfmul[256][256] __attribute__((aligned(256)));
+extern const u8 raid6_gfexp[256]      __attribute__((aligned(256)));
+extern const u8 raid6_gfinv[256]      __attribute__((aligned(256)));
+extern const u8 raid6_gfexi[256]      __attribute__((aligned(256)));
+
+/* Recovery routines */
+void raid6_2data_recov(int disks, size_t bytes, int faila, int failb, void **ptrs);
+void raid6_datap_recov(int disks, size_t bytes, int faila, void **ptrs);
+void raid6_dual_recov(int disks, size_t bytes, int faila, int failb, void **ptrs);
+
+/* Some definitions to allow code to be compiled for testing in userspace */
+#ifndef __KERNEL__
+
+# define jiffies	raid6_jiffies()
+# define printk 	printf
+# define GFP_KERNEL	0
+# define __get_free_pages(x,y)	((unsigned long)mmap(NULL, PAGE_SIZE << (y), PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, 0, 0))
+# define free_pages(x,y)	munmap((void *)(x), (y)*PAGE_SIZE)
+
+static inline void cpu_relax(void)
+{
+	/* Nothing */
+}
+
+#undef  HZ
+#define HZ 1000
+static inline uint32_t raid6_jiffies(void)
+{
+	struct timeval tv;
+	gettimeofday(&tv, NULL);
+	return tv.tv_sec*1000 + tv.tv_usec/1000;
+}
+
+#endif /* ! __KERNEL__ */
+
+#endif /* LINUX_RAID_RAID6_H */

drivers/md/raid6algos.c

+/* -*- linux-c -*- ------------------------------------------------------- *
+ *
+ *   Copyright 2002 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+/*
+ * raid6algos.c
+ *
+ * Algorithm list and algorithm selection for RAID-6
+ */
+
+#include "raid6.h"
+#ifndef __KERNEL__
+#include <sys/mman.h>
+#endif
+
+struct raid6_calls raid6_call;
+
+/* Various routine sets */
+extern const struct raid6_calls raid6_intx1;
+extern const struct raid6_calls raid6_intx2;
+extern const struct raid6_calls raid6_intx4;
+extern const struct raid6_calls raid6_intx8;
+extern const struct raid6_calls raid6_intx16;
+extern const struct raid6_calls raid6_mmxx1;
+extern const struct raid6_calls raid6_mmxx2;
+extern const struct raid6_calls raid6_sse1x1;
+extern const struct raid6_calls raid6_sse1x2;
+extern const struct raid6_calls raid6_sse2x1;
+extern const struct raid6_calls raid6_sse2x2;
+extern const struct raid6_calls raid6_sse2x4;
+
+const struct raid6_calls * const raid6_algos[] = {
+	&raid6_intx1,
+	&raid6_intx2,
+	&raid6_intx4,
+	&raid6_intx8,
+#if defined(__ia64__)
+	&raid6_intx16,
+	&raid6_intx32,
+#endif
+#if defined(__i386__) || defined(__x86_64__)
+	&raid6_mmxx1,
+	&raid6_mmxx2,
+	&raid6_sse1x1,
+	&raid6_sse1x2,
+	&raid6_sse2x1,
+	&raid6_sse2x2,
+#endif
+#if defined(__x86_64__)
+	&raid6_sse2x4,
+#endif
+	NULL
+};
+
+#ifdef __KERNEL__
+#define RAID6_TIME_JIFFIES_LG2	4
+#else
+/* Need more time to be stable in userspace */
+#define RAID6_TIME_JIFFIES_LG2	9
+#endif
+
+/* Try to pick the best algorithm */
+/* This code uses the gfmul table as convenient data set to abuse */
+
+int __init raid6_select_algo(void)
+{
+	const struct raid6_calls * const * algo;
+	const struct raid6_calls * best;
+	char *syndromes;
+	void *dptrs[(65536/PAGE_SIZE)+2];
+	int i, disks;
+	unsigned long perf, bestperf;
+	int bestprefer;
+	unsigned long j0, j1;
+
+	disks = (65536/PAGE_SIZE)+2;
+	for ( i = 0 ; i < disks-2 ; i++ ) {
+		dptrs[i] = ((char *)raid6_gfmul) + PAGE_SIZE*i;
+	}
+
+	/* Normal code - use a 2-page allocation to avoid D$ conflict */
+	syndromes = (void *) __get_free_pages(GFP_KERNEL, 1);
+
+	if ( !syndromes ) {
+		printk("raid6: Yikes!  No memory available.\n");
+		return -ENOMEM;
+	}
+
+	dptrs[disks-2] = syndromes;
+	dptrs[disks-1] = syndromes + PAGE_SIZE;
+
+	bestperf = 0;  bestprefer = 0;  best = NULL;
+
+	for ( algo = raid6_algos ; *algo ; algo++ ) {
+		if ( !(*algo)->valid || (*algo)->valid() ) {
+			perf = 0;
+
+			preempt_disable();
+			j0 = jiffies;
+			while ( (j1 = jiffies) == j0 )
+				cpu_relax();
+			while ( (jiffies-j1) < (1 << RAID6_TIME_JIFFIES_LG2) ) {
+				(*algo)->gen_syndrome(disks, PAGE_SIZE, dptrs);
+				perf++;
+			}
+			preempt_enable();
+
+			if ( (*algo)->prefer > bestprefer ||
+			     ((*algo)->prefer == bestprefer &&
+			      perf > bestperf) ) {
+				best = *algo;
+				bestprefer = best->prefer;
+				bestperf = perf;
+			}
+			printk("raid6: %-8s %5ld MB/s\n", (*algo)->name,
+			       (perf*HZ) >> (20-16+RAID6_TIME_JIFFIES_LG2));
+		}
+	}
+
+	if ( best )
+		printk("raid6: using algorithm %s (%ld MB/s)\n",
+		       best->name,
+		       (bestperf*HZ) >> (20-16+RAID6_TIME_JIFFIES_LG2));
+	else
+		printk("raid6: Yikes!  No algorithm found!\n");
+
+	raid6_call = *best;
+
+	free_pages((unsigned long)syndromes, 1);
+
+	return best ? 0 : -EINVAL;
+}

drivers/md/raid6int.uc

+/* -*- linux-c -*- ------------------------------------------------------- *
+ *
+ *   Copyright 2002 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+/*
+ * raid6int$#.c
+ *
+ * $#-way unrolled portable integer math RAID-6 instruction set
+ *
+ * This file is postprocessed using unroller.pl
+ */
+
+#include "raid6.h"
+
+/*
+ * IA-64 wants insane amounts of unrolling.  On other architectures that
+ * is just a waste of space.
+ */
+
+#if ($# <= 8) || defined(_ia64__)
+
+static void raid6_int$#_gen_syndrome(int disks, size_t bytes, void **ptrs)
+{
+	u8 **dptr = (u8 **)ptrs;
+	u8 *p, *q;
+	int d, z, z0;
+
+	unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
+
+	z0 = disks - 3;		/* Highest data disk */
+	p = dptr[z0+1];		/* XOR parity */
+	q = dptr[z0+2];		/* RS syndrome */
+
+	for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
+		wq$$ = wp$$ = *(unative_t *)&dptr[z0][d+$$*NSIZE];
+		for ( z = z0-1 ; z >= 0 ; z-- ) {
+			wd$$ = *(unative_t *)&dptr[z][d+$$*NSIZE];
+			wp$$ ^= wd$$;
+			w2$$ = wq$$ & NBYTES(0x80);
+			w1$$ = (wq$$ << 1) & NBYTES(0xfe);
+			w2$$ = (w2$$ << 1) - (w2$$ >> 7);
+			w2$$ &= NBYTES(0x1d);
+			w1$$ ^= w2$$;
+			wq$$ = w1$$ ^ wd$$;
+		}
+		*(unative_t *)&p[d+NSIZE*$$] = wp$$;
+		*(unative_t *)&q[d+NSIZE*$$] = wq$$;
+	}
+}
+
+const struct raid6_calls raid6_intx$# = {
+	raid6_int$#_gen_syndrome,
+	NULL,		/* always valid */
+	"int" NSTRING "x$#",
+	0
+};
+
+#endif

drivers/md/raid6mmx.c

+/* -*- linux-c -*- ------------------------------------------------------- *
+ *
+ *   Copyright 2002 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+/*
+ * raid6mmx.c
+ *
+ * MMX implementation of RAID-6 syndrome functions
+ */
+
+#if defined(__i386__) || defined(__x86_64__)
+
+#include "raid6.h"
+#include "raid6x86.h"
+
+/* Shared with raid6sse1.c */
+const struct raid6_mmx_constants {
+	u64 x1d;
+} raid6_mmx_constants = {
+	0x1d1d1d1d1d1d1d1dULL,
+};
+
+static int raid6_have_mmx(void)
+{
+#ifdef __KERNEL__
+	/* Not really "boot_cpu" but "all_cpus" */
+	return boot_cpu_has(X86_FEATURE_MMX);
+#else
+	/* User space test code */
+	u32 features = cpuid_features();
+	return ( (features & (1<<23)) == (1<<23) );
+#endif
+}
+
+/*
+ * Plain MMX implementation
+ */
+static void raid6_mmx1_gen_syndrome(int disks, size_t bytes, void **ptrs)
+{
+	u8 **dptr = (u8 **)ptrs;
+	u8 *p, *q;
+	int d, z, z0;
+	raid6_mmx_save_t sa;
+
+	z0 = disks - 3;		/* Highest data disk */
+	p = dptr[z0+1];		/* XOR parity */
+	q = dptr[z0+2];		/* RS syndrome */
+
+	raid6_before_mmx(&sa);
+
+	asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
+	asm volatile("pxor %mm5,%mm5");	/* Zero temp */
+
+	for ( d = 0 ; d < bytes ; d += 8 ) {
+		asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
+		asm volatile("movq %mm2,%mm4");	/* Q[0] */
+		for ( z = z0-1 ; z >= 0 ; z-- ) {
+			asm volatile("movq %0,%%mm6" : : "m" (dptr[z][d]));
+			asm volatile("pcmpgtb %mm4,%mm5");
+			asm volatile("paddb %mm4,%mm4");
+			asm volatile("pand %mm0,%mm5");
+			asm volatile("pxor %mm5,%mm4");
+			asm volatile("pxor %mm5,%mm5");
+			asm volatile("pxor %mm6,%mm2");
+			asm volatile("pxor %mm6,%mm4");
+		}
+		asm volatile("movq %%mm2,%0" : "=m" (p[d]));
+		asm volatile("pxor %mm2,%mm2");
+		asm volatile("movq %%mm4,%0" : "=m" (q[d]));
+		asm volatile("pxor %mm4,%mm4");
+	}
+
+	raid6_after_mmx(&sa);
+}
+
+const struct raid6_calls raid6_mmxx1 = {
+	raid6_mmx1_gen_syndrome,
+	raid6_have_mmx,
+	"mmxx1",
+	0
+};
+
+/*
+ * Unrolled-by-2 MMX implementation
+ */
+static void raid6_mmx2_gen_syndrome(int disks, size_t bytes, void **ptrs)
+{
+	u8 **dptr = (u8 **)ptrs;
+	u8 *p, *q;
+	int d, z, z0;
+	raid6_mmx_save_t sa;
+
+	z0 = disks - 3;		/* Highest data disk */
+	p = dptr[z0+1];		/* XOR parity */
+	q = dptr[z0+2];		/* RS syndrome */
+
+	raid6_before_mmx(&sa);
+
+	asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
+	asm volatile("pxor %mm5,%mm5");	/* Zero temp */
+	asm volatile("pxor %mm7,%mm7"); /* Zero temp */
+
+	for ( d = 0 ; d < bytes ; d += 16 ) {
+		asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
+		asm volatile("movq %0,%%mm3" : : "m" (dptr[z0][d+8]));
+		asm volatile("movq %mm2,%mm4"); /* Q[0] */
+		asm volatile("movq %mm3,%mm6"); /* Q[1] */
+		for ( z = z0-1 ; z >= 0 ; z-- ) {
+			asm volatile("pcmpgtb %mm4,%mm5");
+			asm volatile("pcmpgtb %mm6,%mm7");
+			asm volatile("paddb %mm4,%mm4");
+			asm volatile("paddb %mm6,%mm6");
+			asm volatile("pand %mm0,%mm5");
+			asm volatile("pand %mm0,%mm7");
+			asm volatile("pxor %mm5,%mm4");
+			asm volatile("pxor %mm7,%mm6");
+			asm volatile("movq %0,%%mm5" : : "m" (dptr[z][d]));
+			asm volatile("movq %0,%%mm7" : : "m" (dptr[z][d+8]));
+			asm volatile("pxor %mm5,%mm2");
+			asm volatile("pxor %mm7,%mm3");
+			asm volatile("pxor %mm5,%mm4");
+			asm volatile("pxor %mm7,%mm6");
+			asm volatile("pxor %mm5,%mm5");
+			asm volatile("pxor %mm7,%mm7");
+		}
+		asm volatile("movq %%mm2,%0" : "=m" (p[d]));
+		asm volatile("movq %%mm3,%0" : "=m" (p[d+8]));
+		asm volatile("movq %%mm4,%0" : "=m" (q[d]));
+		asm volatile("movq %%mm6,%0" : "=m" (q[d+8]));
+	}
+
+	raid6_after_mmx(&sa);
+}
+
+const struct raid6_calls raid6_mmxx2 = {
+	raid6_mmx2_gen_syndrome,
+	raid6_have_mmx,
+	"mmxx2",
+	0
+};
+
+#endif

drivers/md/raid6recov.c

+/* -*- linux-c -*- ------------------------------------------------------- *
+ *
+ *   Copyright 2002 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+/*
+ * raid6recov.c
+ *
+ * RAID-6 data recovery in dual failure mode.  In single failure mode,
+ * use the RAID-5 algorithm (or, in the case of Q failure, just reconstruct
+ * the syndrome.)
+ */
+
+#include "raid6.h"
+
+/* Recover two failed data blocks. */
+void raid6_2data_recov(int disks, size_t bytes, int faila, int failb,
+		       void **ptrs)
+{
+	u8 *p, *q, *dp, *dq;
+	u8 px, qx, db;
+	const u8 *pbmul;	/* P multiplier table for B data */
+	const u8 *qmul;		/* Q multiplier table (for both) */
+
+	p = (u8 *)ptrs[disks-2];
+	q = (u8 *)ptrs[disks-1];
+
+	/* Compute syndrome with zero for the missing data pages
+	   Use the dead data pages as temporary storage for
+	   delta p and delta q */
+	dp = (u8 *)ptrs[faila];
+	ptrs[faila] = (void *)raid6_empty_zero_page;
+	ptrs[disks-2] = dp;
+	dq = (u8 *)ptrs[failb];
+	ptrs[failb] = (void *)raid6_empty_zero_page;
+	ptrs[disks-1] = dq;
+
+	raid6_call.gen_syndrome(disks, bytes, ptrs);
+
+	/* Restore pointer table */
+	ptrs[faila]   = dp;
+	ptrs[failb]   = dq;
+	ptrs[disks-2] = p;
+	ptrs[disks-1] = q;
+
+	/* Now, pick the proper data tables */
+	pbmul = raid6_gfmul[raid6_gfexi[failb-faila]];
+	qmul  = raid6_gfmul[raid6_gfinv[raid6_gfexp[faila]^raid6_gfexp[failb]]];
+
+	/* Now do it... */
+	while ( bytes-- ) {
+		px    = *p ^ *dp;
+		qx    = qmul[*q ^ *dq];
+		*dq++ = db = pbmul[px] ^ qx; /* Reconstructed B */
+		*dp++ = db ^ px; /* Reconstructed A */
+		p++; q++;
+	}
+}
+
+
+
+
+/* Recover failure of one data block plus the P block */
+void raid6_datap_recov(int disks, size_t bytes, int faila, void **ptrs)
+{
+	u8 *p, *q, *dq;
+	const u8 *qmul;		/* Q multiplier table */
+
+	p = (u8 *)ptrs[disks-2];
+	q = (u8 *)ptrs[disks-1];
+
+	/* Compute syndrome with zero for the missing data page
+	   Use the dead data page as temporary storage for delta q */
+	dq = (u8 *)ptrs[faila];
+	ptrs[faila] = (void *)raid6_empty_zero_page;
+	ptrs[disks-1] = dq;
+
+	raid6_call.gen_syndrome(disks, bytes, ptrs);
+
+	/* Restore pointer table */
+	ptrs[faila]   = dq;
+	ptrs[disks-1] = q;
+
+	/* Now, pick the proper data tables */
+	qmul  = raid6_gfmul[raid6_gfinv[raid6_gfexp[faila]]];
+
+	/* Now do it... */
+	while ( bytes-- ) {
+		*p++ ^= *dq = qmul[*q ^ *dq];
+		q++; dq++;
+	}
+}
+
+
+#ifndef __KERNEL__		/* Testing only */
+
+/* Recover two failed blocks. */
+void raid6_dual_recov(int disks, size_t bytes, int faila, int failb, void **ptrs)
+{
+	if ( faila > failb ) {
+		int tmp = faila;
+		faila = failb;
+		failb = tmp;
+	}
+
+	if ( failb == disks-1 ) {
+		if ( faila == disks-2 ) {
+			/* P+Q failure.  Just rebuild the syndrome. */
+			raid6_call.gen_syndrome(disks, bytes, ptrs);
+		} else {
+			/* data+Q failure.  Reconstruct data from P,
+			   then rebuild syndrome. */
+			/* FIX */
+		}
+	} else {
+		if ( failb == disks-2 ) {
+			/* data+P failure. */
+			raid6_datap_recov(disks, bytes, faila, ptrs);
+		} else {
+			/* data+data failure. */
+			raid6_2data_recov(disks, bytes, faila, failb, ptrs);
+		}
+	}
+}
+
+#endif

drivers/md/raid6sse1.c

+/* -*- linux-c -*- ------------------------------------------------------- *
+ *
+ *   Copyright 2002 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+/*
+ * raid6sse1.c
+ *
+ * SSE-1/MMXEXT implementation of RAID-6 syndrome functions
+ *
+ * This is really an MMX implementation, but it requires SSE-1 or
+ * AMD MMXEXT for prefetch support and a few other features.  The
+ * support for nontemporal memory accesses is enough to make this
+ * worthwhile as a separate implementation.
+ */
+
+#if defined(__i386__) || defined(__x86_64__)
+
+#include "raid6.h"
+#include "raid6x86.h"
+
+/* Defined in raid6mmx.c */
+extern const struct raid6_mmx_constants {
+	u64 x1d;
+} raid6_mmx_constants;
+
+static int raid6_have_sse1_or_mmxext(void)
+{
+#ifdef __KERNEL__
+	/* Not really boot_cpu but "all_cpus" */
+	return boot_cpu_has(X86_FEATURE_MMX) &&
+		(boot_cpu_has(X86_FEATURE_XMM) ||
+		 boot_cpu_has(X86_FEATURE_MMXEXT));
+#else
+	/* User space test code - this incorrectly breaks on some Athlons */
+	u32 features = cpuid_features();
+	return ( (features & (5<<23)) == (5<<23) );
+#endif
+}
+
+/*
+ * Plain SSE1 implementation
+ */
+static void raid6_sse11_gen_syndrome(int disks, size_t bytes, void **ptrs)
+{
+	u8 **dptr = (u8 **)ptrs;
+	u8 *p, *q;
+	int d, z, z0;
+	raid6_mmx_save_t sa;
+
+	z0 = disks - 3;		/* Highest data disk */
+	p = dptr[z0+1];		/* XOR parity */
+	q = dptr[z0+2];		/* RS syndrome */
+
+	/* This is really MMX code, not SSE */
+	raid6_before_mmx(&sa);
+
+	asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
+	asm volatile("pxor %mm5,%mm5");	/* Zero temp */
+
+	for ( d = 0 ; d < bytes ; d += 8 ) {
+		asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
+		asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
+		asm volatile("prefetchnta %0" : : "m" (dptr[z0-1][d]));
+		asm volatile("movq %mm2,%mm4");	/* Q[0] */
+		asm volatile("movq %0,%%mm6" : : "m" (dptr[z0-1][d]));
+		for ( z = z0-2 ; z >= 0 ; z-- ) {
+			asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
+			asm volatile("pcmpgtb %mm4,%mm5");
+			asm volatile("paddb %mm4,%mm4");
+			asm volatile("pand %mm0,%mm5");
+			asm volatile("pxor %mm5,%mm4");
+			asm volatile("pxor %mm5,%mm5");
+			asm volatile("pxor %mm6,%mm2");
+			asm volatile("pxor %mm6,%mm4");
+			asm volatile("movq %0,%%mm6" : : "m" (dptr[z][d]));
+		}
+		asm volatile("pcmpgtb %mm4,%mm5");
+		asm volatile("paddb %mm4,%mm4");
+		asm volatile("pand %mm0,%mm5");
+		asm volatile("pxor %mm5,%mm4");
+		asm volatile("pxor %mm5,%mm5");
+		asm volatile("pxor %mm6,%mm2");
+		asm volatile("pxor %mm6,%mm4");
+
+		asm volatile("movntq %%mm2,%0" : "=m" (p[d]));
+		asm volatile("movntq %%mm4,%0" : "=m" (q[d]));
+	}
+
+	raid6_after_mmx(&sa);
+	asm volatile("sfence" : : : "memory");
+}
+
+const struct raid6_calls raid6_sse1x1 = {
+	raid6_sse11_gen_syndrome,
+	raid6_have_sse1_or_mmxext,
+	"sse1x1",
+	1			/* Has cache hints */
+};
+
+/*
+ * Unrolled-by-2 SSE1 implementation
+ */
+static void raid6_sse12_gen_syndrome(int disks, size_t bytes, void **ptrs)
+{
+	u8 **dptr = (u8 **)ptrs;
+	u8 *p, *q;
+	int d, z, z0;
+	raid6_mmx_save_t sa;
+
+	z0 = disks - 3;		/* Highest data disk */
+	p = dptr[z0+1];		/* XOR parity */
+	q = dptr[z0+2];		/* RS syndrome */
+
+	raid6_before_mmx(&sa);
+
+	asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
+	asm volatile("pxor %mm5,%mm5");	/* Zero temp */
+	asm volatile("pxor %mm7,%mm7"); /* Zero temp */
+
+	/* We uniformly assume a single prefetch covers at least 16 bytes */
+	for ( d = 0 ; d < bytes ; d += 16 ) {
+		asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
+		asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
+		asm volatile("movq %0,%%mm3" : : "m" (dptr[z0][d+8])); /* P[1] */
+		asm volatile("movq %mm2,%mm4");	/* Q[0] */
+		asm volatile("movq %mm3,%mm6"); /* Q[1] */
+		for ( z = z0-1 ; z >= 0 ; z-- ) {
+			asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
+			asm volatile("pcmpgtb %mm4,%mm5");
+			asm volatile("pcmpgtb %mm6,%mm7");
+			asm volatile("paddb %mm4,%mm4");
+			asm volatile("paddb %mm6,%mm6");
+			asm volatile("pand %mm0,%mm5");
+			asm volatile("pand %mm0,%mm7");
+			asm volatile("pxor %mm5,%mm4");
+			asm volatile("pxor %mm7,%mm6");
+			asm volatile("movq %0,%%mm5" : : "m" (dptr[z][d]));
+			asm volatile("movq %0,%%mm7" : : "m" (dptr[z][d+8]));
+			asm volatile("pxor %mm5,%mm2");
+			asm volatile("pxor %mm7,%mm3");
+			asm volatile("pxor %mm5,%mm4");
+			asm volatile("pxor %mm7,%mm6");
+			asm volatile("pxor %mm5,%mm5");
+			asm volatile("pxor %mm7,%mm7");
+		}
+		asm volatile("movntq %%mm2,%0" : "=m" (p[d]));
+		asm volatile("movntq %%mm3,%0" : "=m" (p[d+8]));
+		asm volatile("movntq %%mm4,%0" : "=m" (q[d]));
+		asm volatile("movntq %%mm6,%0" : "=m" (q[d+8]));
+	}
+
+	raid6_after_mmx(&sa);
+	asm volatile("sfence" : :: "memory");
+}
+
+const struct raid6_calls raid6_sse1x2 = {
+	raid6_sse12_gen_syndrome,
+	raid6_have_sse1_or_mmxext,
+	"sse1x2",
+	1			/* Has cache hints */
+};
+
+#endif

drivers/md/raid6sse2.c

+/* -*- linux-c -*- ------------------------------------------------------- *
+ *
+ *   Copyright 2002 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+/*
+ * raid6sse2.c
+ *
+ * SSE-2 implementation of RAID-6 syndrome functions
+ *
+ */
+
+#if defined(__i386__) || defined(__x86_64__)
+
+#include "raid6.h"
+#include "raid6x86.h"
+
+static const struct raid6_sse_constants {
+	u64 x1d[2];
+} raid6_sse_constants  __attribute__((aligned(16))) = {
+	{ 0x1d1d1d1d1d1d1d1dULL, 0x1d1d1d1d1d1d1d1dULL },
+};
+
+static int raid6_have_sse2(void)
+{
+#ifdef __KERNEL__
+	/* Not really boot_cpu but "all_cpus" */
+	return boot_cpu_has(X86_FEATURE_MMX) &&
+		boot_cpu_has(X86_FEATURE_FXSR) &&
+		boot_cpu_has(X86_FEATURE_XMM) &&
+		boot_cpu_has(X86_FEATURE_XMM2);
+#else
+	/* User space test code */
+	u32 features = cpuid_features();
+	return ( (features & (15<<23)) == (15<<23) );
+#endif
+}
+
+/*
+ * Plain SSE2 implementation
+ */
+static void raid6_sse21_gen_syndrome(int disks, size_t bytes, void **ptrs)
+{
+	u8 **dptr = (u8 **)ptrs;
+	u8 *p, *q;
+	int d, z, z0;
+	raid6_sse_save_t sa;
+
+	z0 = disks - 3;		/* Highest data disk */
+	p = dptr[z0+1];		/* XOR parity */
+	q = dptr[z0+2];		/* RS syndrome */
+
+	raid6_before_sse2(&sa);
+
+	asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
+	asm volatile("pxor %xmm5,%xmm5");	/* Zero temp */
+
+	for ( d = 0 ; d < bytes ; d += 16 ) {
+		asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
+		asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d])); /* P[0] */
+		asm volatile("prefetchnta %0" : : "m" (dptr[z0-1][d]));
+		asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */
+		asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z0-1][d]));
+		for ( z = z0-2 ; z >= 0 ; z-- ) {
+			asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
+			asm volatile("pcmpgtb %xmm4,%xmm5");
+			asm volatile("paddb %xmm4,%xmm4");
+			asm volatile("pand %xmm0,%xmm5");
+			asm volatile("pxor %xmm5,%xmm4");
+			asm volatile("pxor %xmm5,%xmm5");
+			asm volatile("pxor %xmm6,%xmm2");
+			asm volatile("pxor %xmm6,%xmm4");
+			asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z][d]));
+		}
+		asm volatile("pcmpgtb %xmm4,%xmm5");
+		asm volatile("paddb %xmm4,%xmm4");
+		asm volatile("pand %xmm0,%xmm5");
+		asm volatile("pxor %xmm5,%xmm4");
+		asm volatile("pxor %xmm5,%xmm5");
+		asm volatile("pxor %xmm6,%xmm2");
+		asm volatile("pxor %xmm6,%xmm4");
+
+		asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
+		asm volatile("pxor %xmm2,%xmm2");
+		asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
+		asm volatile("pxor %xmm4,%xmm4");
+	}
+
+	raid6_after_sse2(&sa);
+	asm volatile("sfence" : : : "memory");
+}
+
+const struct raid6_calls raid6_sse2x1 = {
+	raid6_sse21_gen_syndrome,
+	raid6_have_sse2,
+	"sse2x1",
+	1			/* Has cache hints */
+};
+
+/*
+ * Unrolled-by-2 SSE2 implementation
+ */
+static void raid6_sse22_gen_syndrome(int disks, size_t bytes, void **ptrs)
+{
+	u8 **dptr = (u8 **)ptrs;
+	u8 *p, *q;
+	int d, z, z0;
+	raid6_sse_save_t sa;
+
+	z0 = disks - 3;		/* Highest data disk */
+	p = dptr[z0+1];		/* XOR parity */
+	q = dptr[z0+2];		/* RS syndrome */
+
+	raid6_before_sse2(&sa);
+
+	asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
+	asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */
+	asm volatile("pxor %xmm7,%xmm7"); /* Zero temp */
+
+	/* We uniformly assume a single prefetch covers at least 32 bytes */
+	for ( d = 0 ; d < bytes ; d += 32 ) {
+		asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
+		asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d]));    /* P[0] */
+		asm volatile("movdqa %0,%%xmm3" : : "m" (dptr[z0][d+16])); /* P[1] */
+		asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */
+		asm volatile("movdqa %xmm3,%xmm6"); /* Q[1] */
+		for ( z = z0-1 ; z >= 0 ; z-- ) {
+			asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
+			asm volatile("pcmpgtb %xmm4,%xmm5");
+			asm volatile("pcmpgtb %xmm6,%xmm7");
+			asm volatile("paddb %xmm4,%xmm4");
+			asm volatile("paddb %xmm6,%xmm6");
+			asm volatile("pand %xmm0,%xmm5");
+			asm volatile("pand %xmm0,%xmm7");
+			asm volatile("pxor %xmm5,%xmm4");
+			asm volatile("pxor %xmm7,%xmm6");
+			asm volatile("movdqa %0,%%xmm5" : : "m" (dptr[z][d]));
+			asm volatile("movdqa %0,%%xmm7" : : "m" (dptr[z][d+16]));
+			asm volatile("pxor %xmm5,%xmm2");
+			asm volatile("pxor %xmm7,%xmm3");
+			asm volatile("pxor %xmm5,%xmm4");
+			asm volatile("pxor %xmm7,%xmm6");
+			asm volatile("pxor %xmm5,%xmm5");
+			asm volatile("pxor %xmm7,%xmm7");
+		}
+		asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
+		asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
+		asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
+		asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
+	}
+
+	raid6_after_sse2(&sa);
+	asm volatile("sfence" : : : "memory");
+}
+
+const struct raid6_calls raid6_sse2x2 = {
+	raid6_sse22_gen_syndrome,
+	raid6_have_sse2,
+	"sse2x2",
+	1			/* Has cache hints */
+};
+
+#endif
+
+#ifdef __x86_64__
+
+/*
+ * Unrolled-by-4 SSE2 implementation
+ */
+static void raid6_sse24_gen_syndrome(int disks, size_t bytes, void **ptrs)
+{
+	u8 **dptr = (u8 **)ptrs;
+	u8 *p, *q;
+	int d, z, z0;
+	raid6_sse16_save_t sa;
+
+	z0 = disks - 3;		/* Highest data disk */
+	p = dptr[z0+1];		/* XOR parity */
+	q = dptr[z0+2];		/* RS syndrome */
+
+	raid6_before_sse16(&sa);
+
+	asm volatile("movdqa %0,%%xmm0" :: "m" (raid6_sse_constants.x1d[0]));
+	asm volatile("pxor %xmm2,%xmm2");	/* P[0] */
+	asm volatile("pxor %xmm3,%xmm3");	/* P[1] */
+	asm volatile("pxor %xmm4,%xmm4"); 	/* Q[0] */
+	asm volatile("pxor %xmm5,%xmm5");	/* Zero temp */
+	asm volatile("pxor %xmm6,%xmm6"); 	/* Q[1] */
+	asm volatile("pxor %xmm7,%xmm7"); 	/* Zero temp */
+	asm volatile("pxor %xmm10,%xmm10");	/* P[2] */
+	asm volatile("pxor %xmm11,%xmm11");	/* P[3] */
+	asm volatile("pxor %xmm12,%xmm12"); 	/* Q[2] */
+	asm volatile("pxor %xmm13,%xmm13");	/* Zero temp */
+	asm volatile("pxor %xmm14,%xmm14"); 	/* Q[3] */
+	asm volatile("pxor %xmm15,%xmm15"); 	/* Zero temp */
+
+	for ( d = 0 ; d < bytes ; d += 64 ) {
+		for ( z = z0 ; z >= 0 ; z-- ) {
+			/* The second prefetch seems to improve performance... */
+			asm volatile("prefetchnta %0" :: "m" (dptr[z][d]));
+			asm volatile("prefetchnta %0" :: "m" (dptr[z][d+32]));
+			asm volatile("pcmpgtb %xmm4,%xmm5");
+			asm volatile("pcmpgtb %xmm6,%xmm7");
+			asm volatile("pcmpgtb %xmm12,%xmm13");
+			asm volatile("pcmpgtb %xmm14,%xmm15");
+			asm volatile("paddb %xmm4,%xmm4");
+			asm volatile("paddb %xmm6,%xmm6");
+			asm volatile("paddb %xmm12,%xmm12");
+			asm volatile("paddb %xmm14,%xmm14");
+			asm volatile("pand %xmm0,%xmm5");
+			asm volatile("pand %xmm0,%xmm7");
+			asm volatile("pand %xmm0,%xmm13");
+			asm volatile("pand %xmm0,%xmm15");
+			asm volatile("pxor %xmm5,%xmm4");
+			asm volatile("pxor %xmm7,%xmm6");
+			asm volatile("pxor %xmm13,%xmm12");
+			asm volatile("pxor %xmm15,%xmm14");
+			asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
+			asm volatile("movdqa %0,%%xmm7" :: "m" (dptr[z][d+16]));
+			asm volatile("movdqa %0,%%xmm13" :: "m" (dptr[z][d+32]));
+			asm volatile("movdqa %0,%%xmm15" :: "m" (dptr[z][d+48]));
+			asm volatile("pxor %xmm5,%xmm2");
+			asm volatile("pxor %xmm7,%xmm3");
+			asm volatile("pxor %xmm13,%xmm10");
+			asm volatile("pxor %xmm15,%xmm11");
+			asm volatile("pxor %xmm5,%xmm4");
+			asm volatile("pxor %xmm7,%xmm6");
+			asm volatile("pxor %xmm13,%xmm12");
+			asm volatile("pxor %xmm15,%xmm14");
+			asm volatile("pxor %xmm5,%xmm5");
+			asm volatile("pxor %xmm7,%xmm7");
+			asm volatile("pxor %xmm13,%xmm13");
+			asm volatile("pxor %xmm15,%xmm15");
+		}
+		asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
+		asm volatile("pxor %xmm2,%xmm2");
+		asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
+		asm volatile("pxor %xmm3,%xmm3");
+		asm volatile("movntdq %%xmm10,%0" : "=m" (p[d+32]));
+		asm volatile("pxor %xmm10,%xmm10");
+		asm volatile("movntdq %%xmm11,%0" : "=m" (p[d+48]));
+		asm volatile("pxor %xmm11,%xmm11");
+		asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
+		asm volatile("pxor %xmm4,%xmm4");
+		asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
+		asm volatile("pxor %xmm6,%xmm6");
+		asm volatile("movntdq %%xmm12,%0" : "=m" (q[d+32]));
+		asm volatile("pxor %xmm12,%xmm12");
+		asm volatile("movntdq %%xmm14,%0" : "=m" (q[d+48]));
+		asm volatile("pxor %xmm14,%xmm14");
+	}
+	asm volatile("sfence" : : : "memory");
+	raid6_after_sse16(&sa);
+}
+
+const struct raid6_calls raid6_sse2x4 = {
+	raid6_sse24_gen_syndrome,
+	raid6_have_sse2,
+	"sse2x4",
+	1			/* Has cache hints */
+};
+
+#endif

drivers/md/raid6test/Makefile

+#
+# This is a simple Makefile to test some of the RAID-6 code
+# from userspace.
+#
+
+CC	= gcc
+CFLAGS	= -I.. -O2 -g -march=i686
+LD	= ld
+PERL	= perl
+
+.c.o:
+	$(CC) $(CFLAGS) -c -o $@ $<
+
+%.c: ../%.c
+	cp -f $< $@
+
+%.uc: ../%.uc
+	cp -f $< $@
+
+%.pl: ../%.pl
+	cp -f $< $@
+
+all:	raid6.o raid6test
+
+raid6.o: raid6int1.o raid6int2.o raid6int4.o raid6int8.o raid6int16.o \
+	 raid6mmx.o raid6sse1.o raid6sse2.o \
+	 raid6recov.o raid6algos.o \
+	 raid6tables.o
+	$(LD) -r -o $@ $^
+
+raid6test: raid6.o test.c
+	$(CC) $(CFLAGS) -o raid6test $^
+
+raid6int1.c: raid6int.uc unroller.pl
+	$(PERL) ./unroller.pl 1 < raid6int.uc > $@
+
+raid6int2.c: raid6int.uc unroller.pl
+	$(PERL) ./unroller.pl 2 < raid6int.uc > $@
+
+raid6int4.c: raid6int.uc unroller.pl
+	$(PERL) ./unroller.pl 4 < raid6int.uc > $@
+
+raid6int8.c: raid6int.uc unroller.pl
+	$(PERL) ./unroller.pl 8 < raid6int.uc > $@
+
+raid6int16.c: raid6int.uc unroller.pl
+	$(PERL) ./unroller.pl 16 < raid6int.uc > $@
+
+raid6tables.c: mktables
+	./mktables > raid6tables.c
+
+clean:
+	rm -f *.o mktables mktables.c raid6int.uc raid6*.c raid6test
+
+spotless: clean
+	rm -f *~

drivers/md/raid6test/test.c

+/* -*- linux-c -*- ------------------------------------------------------- *
+ *
+ *   Copyright 2002 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+/*
+ * raid6test.c
+ *
+ * Test RAID-6 recovery with various algorithms
+ */
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include "raid6.h"
+
+#define NDISKS		16	/* Including P and Q */
+
+const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
+struct raid6_calls raid6_call;
+
+char *dataptrs[NDISKS];
+char data[NDISKS][PAGE_SIZE];
+char recovi[PAGE_SIZE], recovj[PAGE_SIZE];
+
+void makedata(void)
+{
+	int i, j;
+
+	for (  i = 0 ; i < NDISKS ; i++ ) {
+		for ( j = 0 ; j < PAGE_SIZE ; j++ ) {
+			data[i][j] = rand();
+		}
+		dataptrs[i] = data[i];
+	}
+}
+
+int main(int argc, char *argv[])
+{
+	const struct raid6_calls * const * algo;
+	int i, j;
+	int erra, errb;
+
+	makedata();
+
+	for ( algo = raid6_algos ; *algo ; algo++ ) {
+		if ( !(*algo)->valid || (*algo)->valid() ) {
+			raid6_call = **algo;
+
+			/* Nuke syndromes */
+			memset(data[NDISKS-2], 0xee, 2*PAGE_SIZE);
+
+			/* Generate assumed good syndrome */
+			raid6_call.gen_syndrome(NDISKS, PAGE_SIZE, (void **)&dataptrs);
+
+			for ( i = 0 ; i < NDISKS-1 ; i++ ) {
+				for ( j = i+1 ; j < NDISKS ; j++ ) {
+					memset(recovi, 0xf0, PAGE_SIZE);
+					memset(recovj, 0xba, PAGE_SIZE);
+
+					dataptrs[i] = recovi;
+					dataptrs[j] = recovj;
+
+					raid6_dual_recov(NDISKS, PAGE_SIZE, i, j, (void **)&dataptrs);
+
+					erra = memcmp(data[i], recovi, PAGE_SIZE);
+					errb = memcmp(data[j], recovj, PAGE_SIZE);
+
+					printf("algo=%-8s  faila=%3d(%c)  failb=%3d(%c)  %s\n",
+					       raid6_call.name,
+					       i, (i==NDISKS-2)?'P':'D',
+					       j, (j==NDISKS-1)?'Q':(j==NDISKS-2)?'P':'D',
+					       (!erra && !errb) ? "OK" :
+					       !erra ? "ERRB" :
+					       !errb ? "ERRA" :
+					       "ERRAB");
+
+					dataptrs[i] = data[i];
+					dataptrs[j] = data[j];
+				}
+			}
+		}
+		printf("\n");
+	}
+
+	printf("\n");
+	/* Pick the best algorithm test */
+	raid6_select_algo();
+
+	return 0;
+}

drivers/md/raid6x86.h

+#ident "$Id: raid6x86.h,v 1.3 2002/12/12 22:41:27 hpa Exp $"
+/* ----------------------------------------------------------------------- *
+ *
+ *   Copyright 2002 H. Peter Anvin - All Rights Reserved
+ *
+ *   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, Inc., 53 Temple Place Ste 330,
+ *   Bostom MA 02111-1307, USA; either version 2 of the License, or
+ *   (at your option) any later version; incorporated herein by reference.
+ *
+ * ----------------------------------------------------------------------- */
+
+/*
+ * raid6x86.h
+ *
+ * Definitions common to x86 and x86-64 RAID-6 code only
+ */
+
+#ifndef LINUX_RAID_RAID6X86_H
+#define LINUX_RAID_RAID6X86_H
+
+#if defined(__i386__) || defined(__x86_64__)
+
+typedef struct {
+	unsigned int fsave[27];
+	unsigned int cr0;
+} raid6_mmx_save_t;
+
+/* N.B.: For SSE we only save %xmm0-%xmm7 even for x86-64, since
+   the code doesn't know about the additional x86-64 registers */
+/* The +3 is so we can make sure the area is aligned properly */
+typedef struct {
+	unsigned int sarea[8*4+3];
+	unsigned int cr0;
+} raid6_sse_save_t __attribute__((aligned(16)));
+
+#ifdef __x86_64__
+
+/* This is for x86-64-specific code which uses all 16 XMM registers */
+typedef struct {
+	unsigned int sarea[16*4+3];
+	unsigned int cr0;
+} raid6_sse16_save_t __attribute__((aligned(16)));
+
+#endif
+
+#ifdef __KERNEL__ /* Real code */
+
+ static inline u32 raid6_get_fpu(void)
+{
+	u32 cr0;
+
+	preempt_disable();
+	asm volatile("movl %%cr0,%0 ; clts" : "=r" (cr0));
+	return cr0;
+}
+
+static inline void raid6_put_fpu(u32 cr0)
+{
+	asm volatile("movl %0,%%cr0" : : "r" (cr0));
+	preempt_enable();
+}
+
+#else /* Dummy code for user space testing */
+
+static inline u32 raid6_get_fpu(void)
+{
+	return 0xf00ba6;
+}
+
+static inline void raid6_put_fpu(u32 cr0)
+{
+	(void)cr0;
+}
+
+#endif
+
+static inline void raid6_before_mmx(raid6_mmx_save_t *s)
+{
+	s->cr0 = raid6_get_fpu();
+	asm volatile("fsave %0 ; fwait" : "=m" (s->fsave[0]));
+}
+
+static inline void raid6_after_mmx(raid6_mmx_save_t *s)
+{
+	asm volatile("frstor %0" : : "m" (s->fsave[0]));
+	raid6_put_fpu(s->cr0);
+}
+
+static inline void raid6_before_sse(raid6_sse_save_t *s)
+{
+#ifdef __x86_64__
+	unsigned int *rsa = s->sarea;
+#else
+	/* On i386 the save area may not be aligned */
+	unsigned int *rsa =
+		(unsigned int *)((((unsigned long)&s->sarea)+15) & ~15);
+#endif
+	s->cr0 = raid6_get_fpu();
+
+	asm volatile("movaps %%xmm0,%0" : "=m" (rsa[0]));
+	asm volatile("movaps %%xmm1,%0" : "=m" (rsa[4]));
+	asm volatile("movaps %%xmm2,%0" : "=m" (rsa[8]));
+	asm volatile("movaps %%xmm3,%0" : "=m" (rsa[12]));
+	asm volatile("movaps %%xmm4,%0" : "=m" (rsa[16]));
+	asm volatile("movaps %%xmm5,%0" : "=m" (rsa[20]));
+	asm volatile("movaps %%xmm6,%0" : "=m" (rsa[24]));
+	asm volatile("movaps %%xmm7,%0" : "=m" (rsa[28]));
+}
+
+static inline void raid6_after_sse(raid6_sse_save_t *s)
+{
+#ifdef __x86_64__
+	unsigned int *rsa = s->sarea;
+#else
+	/* On i386 the save area may not be aligned */
+	unsigned int *rsa =
+		(unsigned int *)((((unsigned long)&s->sarea)+15) & ~15);
+#endif
+	asm volatile("movaps %0,%%xmm0" : : "m" (rsa[0]));
+	asm volatile("movaps %0,%%xmm1" : : "m" (rsa[4]));
+	asm volatile("movaps %0,%%xmm2" : : "m" (rsa[8]));
+	asm volatile("movaps %0,%%xmm3" : : "m" (rsa[12]));
+	asm volatile("movaps %0,%%xmm4" : : "m" (rsa[16]));
+	asm volatile("movaps %0,%%xmm5" : : "m" (rsa[20]));
+	asm volatile("movaps %0,%%xmm6" : : "m" (rsa[24]));
+	asm volatile("movaps %0,%%xmm7" : : "m" (rsa[28]));
+
+	raid6_put_fpu(s->cr0);
+}
+
+static inline void raid6_before_sse2(raid6_sse_save_t *s)
+{
+#ifdef __x86_64__
+	unsigned int *rsa = &s->sarea;
+#else
+	/* On i386 the save area may not be aligned */
+	unsigned int *rsa =
+		(unsigned int *)((((unsigned long)&s->sarea)+15) & ~15);
+#endif
+	s->cr0 = raid6_get_fpu();
+
+	asm volatile("movdqa %%xmm0,%0" : "=m" (rsa[0]));
+	asm volatile("movdqa %%xmm1,%0" : "=m" (rsa[4]));
+	asm volatile("movdqa %%xmm2,%0" : "=m" (rsa[8]));
+	asm volatile("movdqa %%xmm3,%0" : "=m" (rsa[12]));
+	asm volatile("movdqa %%xmm4,%0" : "=m" (rsa[16]));
+	asm volatile("movdqa %%xmm5,%0" : "=m" (rsa[20]));
+	asm volatile("movdqa %%xmm6,%0" : "=m" (rsa[24]));
+	asm volatile("movdqa %%xmm7,%0" : "=m" (rsa[28]));
+}
+
+static inline void raid6_after_sse2(raid6_sse_save_t *s)
+{
+#ifdef __x86_64__
+	unsigned int *rsa = s->sarea;
+#else
+	/* On i386 the save area may not be aligned */
+	unsigned int *rsa =
+		(unsigned int *)((((unsigned long)&s->sarea)+15) & ~15);
+#endif
+	asm volatile("movdqa %0,%%xmm0" : : "m" (rsa[0]));
+	asm volatile("movdqa %0,%%xmm1" : : "m" (rsa[4]));
+	asm volatile("movdqa %0,%%xmm2" : : "m" (rsa[8]));
+	asm volatile("movdqa %0,%%xmm3" : : "m" (rsa[12]));
+	asm volatile("movdqa %0,%%xmm4" : : "m" (rsa[16]));
+	asm volatile("movdqa %0,%%xmm5" : : "m" (rsa[20]));
+	asm volatile("movdqa %0,%%xmm6" : : "m" (rsa[24]));
+	asm volatile("movdqa %0,%%xmm7" : : "m" (rsa[28]));
+
+	raid6_put_fpu(s->cr0);
+}
+
+#ifdef __x86_64__
+
+static inline raid6_before_sse16(raid6_sse16_save_t *s)
+{
+	unsigned int *rsa = s->sarea;
+
+	s->cr0 = raid6_get_fpu();
+
+	asm volatile("movdqa %%xmm0,%0" : "=m" (rsa[0]));
+	asm volatile("movdqa %%xmm1,%0" : "=m" (rsa[4]));
+	asm volatile("movdqa %%xmm2,%0" : "=m" (rsa[8]));
+	asm volatile("movdqa %%xmm3,%0" : "=m" (rsa[12]));
+	asm volatile("movdqa %%xmm4,%0" : "=m" (rsa[16]));
+	asm volatile("movdqa %%xmm5,%0" : "=m" (rsa[20]));
+	asm volatile("movdqa %%xmm6,%0" : "=m" (rsa[24]));
+	asm volatile("movdqa %%xmm7,%0" : "=m" (rsa[28]));
+	asm volatile("movdqa %%xmm8,%0" : "=m" (rsa[32]));
+	asm volatile("movdqa %%xmm9,%0" : "=m" (rsa[36]));
+	asm volatile("movdqa %%xmm10,%0" : "=m" (rsa[40]));
+	asm volatile("movdqa %%xmm11,%0" : "=m" (rsa[44]));
+	asm volatile("movdqa %%xmm12,%0" : "=m" (rsa[48]));
+	asm volatile("movdqa %%xmm13,%0" : "=m" (rsa[52]));
+	asm volatile("movdqa %%xmm14,%0" : "=m" (rsa[56]));
+	asm volatile("movdqa %%xmm15,%0" : "=m" (rsa[60]));
+}
+
+static inline raid6_after_sse16(raid6_sse16_save_t *s)
+{
+	unsigned int *rsa = s->sarea;
+
+	asm volatile("movdqa %0,%%xmm0" : : "m" (rsa[0]));
+	asm volatile("movdqa %0,%%xmm1" : : "m" (rsa[4]));
+	asm volatile("movdqa %0,%%xmm2" : : "m" (rsa[8]));
+	asm volatile("movdqa %0,%%xmm3" : : "m" (rsa[12]));
+	asm volatile("movdqa %0,%%xmm4" : : "m" (rsa[16]));
+	asm volatile("movdqa %0,%%xmm5" : : "m" (rsa[20]));
+	asm volatile("movdqa %0,%%xmm6" : : "m" (rsa[24]));
+	asm volatile("movdqa %0,%%xmm7" : : "m" (rsa[28]));
+	asm volatile("movdqa %0,%%xmm8" : : "m" (rsa[32]));
+	asm volatile("movdqa %0,%%xmm9" : : "m" (rsa[36]));
+	asm volatile("movdqa %0,%%xmm10" : : "m" (rsa[40]));
+	asm volatile("movdqa %0,%%xmm11" : : "m" (rsa[44]));
+	asm volatile("movdqa %0,%%xmm12" : : "m" (rsa[48]));
+	asm volatile("movdqa %0,%%xmm13" : : "m" (rsa[52]));
+	asm volatile("movdqa %0,%%xmm14" : : "m" (rsa[56]));
+	asm volatile("movdqa %0,%%xmm15" : : "m" (rsa[60]));
+
+	raid6_put_fpu(s->cr0);
+}
+
+#endif /* __x86_64__ */
+
+/* User space test hack */
+#ifndef __KERNEL__
+static inline int cpuid_features(void)
+{
+	u32 eax = 1;
+	u32 ebx, ecx, edx;
+
+	asm volatile("cpuid" :
+		     "+a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx));
+
+	return edx;
+}
+#endif /* ndef __KERNEL__ */
+
+#endif
+#endif

drivers/md/unroll.pl

+#!/usr/bin/perl
+#
+# Take a piece of C code and for each line which contains the sequence $$
+# repeat n times with $ replaced by 0...n-1; the sequence $# is replaced
+# by the unrolling factor, and $* with a single $
+#
+
+($n) = @ARGV;
+$n += 0;
+
+while ( defined($line = <STDIN>) ) {
+    if ( $line =~ /\$\$/ ) {
+	$rep = $n;
+    } else {
+	$rep = 1;
+    }
+    for ( $i = 0 ; $i < $rep ; $i++ ) {
+	$tmp = $line;
+	$tmp =~ s/\$\$/$i/g;
+	$tmp =~ s/\$\#/$n/g;
+	$tmp =~ s/\$\*/\$/g;
+	print $tmp;
+    }
+}

include/linux/raid/md_k.h

@@ -23,7 +23,8 @@
 #define TRANSLUCENT       5UL
 #define HSM               6UL
 #define MULTIPATH         7UL
-#define MAX_PERSONALITY   8UL
+#define RAID6		  8UL
+#define MAX_PERSONALITY   9UL
 
 #define	LEVEL_MULTIPATH		(-4)
 #define	LEVEL_LINEAR		(-1)
@@ -41,6 +42,7 @@ static inline int pers_to_level (int pers)
 		case RAID0:		return 0;
 		case RAID1:		return 1;
 		case RAID5:		return 5;
+		case RAID6:		return 6;
 	}
 	BUG();
 	return MD_RESERVED;
@@ -57,6 +59,7 @@ static inline int level_to_pers (int level)
 		case 1: return RAID1;
 		case 4:
 		case 5: return RAID5;
+		case 6: return RAID6;
 	}
 	return MD_RESERVED;
 }