Merge branch 'hash' of git://ftp.sciencehorizons.net/linux

Pull string hash improvements from George Spelvin:
 "This series does several related things:

   - Makes the dcache hash (fs/namei.c) useful for general kernel use.

     (Thanks to Bruce for noticing the zero-length corner case)

   - Converts the string hashes in <linux/sunrpc/svcauth.h> to use the
     above.

   - Avoids 64-bit multiplies in hash_64() on 32-bit platforms.  Two
     32-bit multiplies will do well enough.

   - Rids the world of the bad hash multipliers in hash_32.

     This finishes the job started in commit 689de1d6ca95 ("Minimal
     fix-up of bad hashing behavior of hash_64()")

     The vast majority of Linux architectures have hardware support for
     32x32-bit multiply and so derive no benefit from "simplified"
     multipliers.

     The few processors that do not (68000, h8/300 and some models of
     Microblaze) have arch-specific implementations added.  Those
     patches are last in the series.

   - Overhauls the dcache hash mixing.

     The patch in commit 0fed3ac866ea ("namei: Improve hash mixing if
     CONFIG_DCACHE_WORD_ACCESS") was an off-the-cuff suggestion.
     Replaced with a much more careful design that's simultaneously
     faster and better.  (My own invention, as there was noting suitable
     in the literature I could find.  Comments welcome!)

   - Modify the hash_name() loop to skip the initial HASH_MIX().  This
     would let us salt the hash if we ever wanted to.

   - Sort out partial_name_hash().

     The hash function is declared as using a long state, even though
     it's truncated to 32 bits at the end and the extra internal state
     contributes nothing to the result.  And some callers do odd things:

      - fs/hfs/string.c only allocates 32 bits of state
      - fs/hfsplus/unicode.c uses it to hash 16-bit unicode symbols not bytes

   - Modify bytemask_from_count to handle inputs of 1..sizeof(long)
     rather than 0..sizeof(long)-1.  This would simplify users other
     than full_name_hash"

  Special thanks to Bruce Fields for testing and finding bugs in v1.  (I
  learned some humbling lessons about "obviously correct" code.)

  On the arch-specific front, the m68k assembly has been tested in a
  standalone test harness, I've been in contact with the Microblaze
  maintainers who mostly don't care, as the hardware multiplier is never
  omitted in real-world applications, and I haven't heard anything from
  the H8/300 world"

* 'hash' of git://ftp.sciencehorizons.net/linux:
  h8300: Add <asm/hash.h>
  microblaze: Add <asm/hash.h>
  m68k: Add <asm/hash.h>
  <linux/hash.h>: Add support for architecture-specific functions
  fs/namei.c: Improve dcache hash function
  Eliminate bad hash multipliers from hash_32() and  hash_64()
  Change hash_64() return value to 32 bits
  <linux/sunrpc/svcauth.h>: Define hash_str() in terms of hashlen_string()
  fs/namei.c: Add hashlen_string() function
  Pull out string hash to <linux/stringhash.h>

17 files changed, 734 insertions, 150 deletions

diff --git a/arch/Kconfig b/arch/Kconfig
index b16e74e..d794384 100644
--- a/arch/Kconfig
+++ b/arch/Kconfig
@@ -598,6 +598,14 @@ config HAVE_STACK_VALIDATION
 	  Architecture supports the 'objtool check' host tool command, which
 	  performs compile-time stack metadata validation.
 
+config HAVE_ARCH_HASH
+	bool
+	default n
+	help
+	  If this is set, the architecture provides an <asm/hash.h>
+	  file which provides platform-specific implementations of some
+	  functions in <linux/hash.h> or fs/namei.c.
+
 #
 # ABI hall of shame
 #
diff --git a/arch/h8300/Kconfig b/arch/h8300/Kconfig
index aa232de..3ae8525 100644
--- a/arch/h8300/Kconfig
+++ b/arch/h8300/Kconfig
@@ -20,6 +20,7 @@ config H8300
 	select HAVE_KERNEL_GZIP
 	select HAVE_KERNEL_LZO
 	select HAVE_ARCH_KGDB
+	select HAVE_ARCH_HASH
 	select CPU_NO_EFFICIENT_FFS
 
 config RWSEM_GENERIC_SPINLOCK
diff --git a/arch/h8300/include/asm/hash.h b/arch/h8300/include/asm/hash.h
new file mode 100644
index 0000000..04cfbd2
--- /dev/null
+++ b/arch/h8300/include/asm/hash.h
@@ -0,0 +1,53 @@
+#ifndef _ASM_HASH_H
+#define _ASM_HASH_H
+
+/*
+ * The later H8SX models have a 32x32-bit multiply, but the H8/300H
+ * and H8S have only 16x16->32.  Since it's tolerably compact, this is
+ * basically an inlined version of the __mulsi3 code.  Since the inputs
+ * are not expected to be small, it's also simplfied by skipping the
+ * early-out checks.
+ *
+ * (Since neither CPU has any multi-bit shift instructions, a
+ * shift-and-add version is a non-starter.)
+ *
+ * TODO: come up with an arch-specific version of the hashing in fs/namei.c,
+ * since that is heavily dependent on rotates.  Which, as mentioned, suck
+ * horribly on H8.
+ */
+
+#if defined(CONFIG_CPU_H300H) || defined(CONFIG_CPU_H8S)
+
+#define HAVE_ARCH__HASH_32 1
+
+/*
+ * Multiply by k = 0x61C88647.  Fitting this into three registers requires
+ * one extra instruction, but reducing register pressure will probably
+ * make that back and then some.
+ *
+ * GCC asm note: %e1 is the high half of operand %1, while %f1 is the
+ * low half.  So if %1 is er4, then %e1 is e4 and %f1 is r4.
+ *
+ * This has been designed to modify x in place, since that's the most
+ * common usage, but preserve k, since hash_64() makes two calls in
+ * quick succession.
+ */
+static inline u32 __attribute_const__ __hash_32(u32 x)
+{
+	u32 temp;
+
+	asm(   "mov.w	%e1,%f0"
+	"\n	mulxu.w	%f2,%0"		/* klow * xhigh */
+	"\n	mov.w	%f0,%e1"	/* The extra instruction */
+	"\n	mov.w	%f1,%f0"
+	"\n	mulxu.w	%e2,%0"		/* khigh * xlow */
+	"\n	add.w	%e1,%f0"
+	"\n	mulxu.w	%f2,%1"		/* klow * xlow */
+	"\n	add.w	%f0,%e1"
+	: "=&r" (temp), "=r" (x)
+	: "%r" (GOLDEN_RATIO_32), "1" (x));
+	return x;
+}
+
+#endif
+#endif /* _ASM_HASH_H */
diff --git a/arch/m68k/Kconfig.cpu b/arch/m68k/Kconfig.cpu
index 8ace920..967260f 100644
--- a/arch/m68k/Kconfig.cpu
+++ b/arch/m68k/Kconfig.cpu
@@ -41,6 +41,7 @@ config M68000
 	select CPU_HAS_NO_UNALIGNED
 	select GENERIC_CSUM
 	select CPU_NO_EFFICIENT_FFS
+	select HAVE_ARCH_HASH
 	help
 	  The Freescale (was Motorola) 68000 CPU is the first generation of
 	  the well known M68K family of processors. The CPU core as well as
diff --git a/arch/m68k/include/asm/hash.h b/arch/m68k/include/asm/hash.h
new file mode 100644
index 0000000..6407af8
--- /dev/null
+++ b/arch/m68k/include/asm/hash.h
@@ -0,0 +1,59 @@
+#ifndef _ASM_HASH_H
+#define _ASM_HASH_H
+
+/*
+ * If CONFIG_M68000=y (original mc68000/010), this file is #included
+ * to work around the lack of a MULU.L instruction.
+ */
+
+#define HAVE_ARCH__HASH_32 1
+/*
+ * While it would be legal to substitute a different hash operation
+ * entirely, let's keep it simple and just use an optimized multiply
+ * by GOLDEN_RATIO_32 = 0x61C88647.
+ *
+ * The best way to do that appears to be to multiply by 0x8647 with
+ * shifts and adds, and use mulu.w to multiply the high half by 0x61C8.
+ *
+ * Because the 68000 has multi-cycle shifts, this addition chain is
+ * chosen to minimise the shift distances.
+ *
+ * Despite every attempt to spoon-feed it simple operations, GCC
+ * 6.1.1 doggedly insists on doing annoying things like converting
+ * "lsl.l #2,<reg>" (12 cycles) to two adds (8+8 cycles).
+ *
+ * It also likes to notice two shifts in a row, like "a = x << 2" and
+ * "a <<= 7", and convert that to "a = x << 9".  But shifts longer
+ * than 8 bits are extra-slow on m68k, so that's a lose.
+ *
+ * Since the 68000 is a very simple in-order processor with no
+ * instruction scheduling effects on execution time, we can safely
+ * take it out of GCC's hands and write one big asm() block.
+ *
+ * Without calling overhead, this operation is 30 bytes (14 instructions
+ * plus one immediate constant) and 166 cycles.
+ *
+ * (Because %2 is fetched twice, it can't be postincrement, and thus it
+ * can't be a fully general "g" or "m".  Register is preferred, but
+ * offsettable memory or immediate will work.)
+ */
+static inline u32 __attribute_const__ __hash_32(u32 x)
+{
+	u32 a, b;
+
+	asm(   "move.l %2,%0"	/* a = x * 0x0001 */
+	"\n	lsl.l #2,%0"	/* a = x * 0x0004 */
+	"\n	move.l %0,%1"
+	"\n	lsl.l #7,%0"	/* a = x * 0x0200 */
+	"\n	add.l %2,%0"	/* a = x * 0x0201 */
+	"\n	add.l %0,%1"	/* b = x * 0x0205 */
+	"\n	add.l %0,%0"	/* a = x * 0x0402 */
+	"\n	add.l %0,%1"	/* b = x * 0x0607 */
+	"\n	lsl.l #5,%0"	/* a = x * 0x8040 */
+	: "=&d,d" (a), "=&r,r" (b)
+	: "r,roi?" (x));	/* a+b = x*0x8647 */
+
+	return ((u16)(x*0x61c8) << 16) + a + b;
+}
+
+#endif	/* _ASM_HASH_H */
diff --git a/arch/microblaze/Kconfig b/arch/microblaze/Kconfig
index f17c3a4..636e072 100644
--- a/arch/microblaze/Kconfig
+++ b/arch/microblaze/Kconfig
@@ -16,6 +16,7 @@ config MICROBLAZE
 	select GENERIC_IRQ_SHOW
 	select GENERIC_PCI_IOMAP
 	select GENERIC_SCHED_CLOCK
+	select HAVE_ARCH_HASH
 	select HAVE_ARCH_KGDB
 	select HAVE_DEBUG_KMEMLEAK
 	select HAVE_DMA_API_DEBUG
diff --git a/arch/microblaze/include/asm/hash.h b/arch/microblaze/include/asm/hash.h
new file mode 100644
index 0000000..753513a
--- /dev/null
+++ b/arch/microblaze/include/asm/hash.h
@@ -0,0 +1,81 @@
+#ifndef _ASM_HASH_H
+#define _ASM_HASH_H
+
+/*
+ * Fortunately, most people who want to run Linux on Microblaze enable
+ * both multiplier and barrel shifter, but omitting them is technically
+ * a supported configuration.
+ *
+ * With just a barrel shifter, we can implement an efficient constant
+ * multiply using shifts and adds.  GCC can find a 9-step solution, but
+ * this 6-step solution was found by Yevgen Voronenko's implementation
+ * of the Hcub algorithm at http://spiral.ece.cmu.edu/mcm/gen.html.
+ *
+ * That software is really not designed for a single multiplier this large,
+ * but if you run it enough times with different seeds, it'll find several
+ * 6-shift, 6-add sequences for computing x * 0x61C88647.  They are all
+ *	c = (x << 19) + x;
+ *	a = (x <<  9) + c;
+ *	b = (x << 23) + a;
+ *	return (a<<11) + (b<<6) + (c<<3) - b;
+ * with variations on the order of the final add.
+ *
+ * Without even a shifter, it's hopless; any hash function will suck.
+ */
+
+#if CONFIG_XILINX_MICROBLAZE0_USE_HW_MUL == 0
+
+#define HAVE_ARCH__HASH_32 1
+
+/* Multiply by GOLDEN_RATIO_32 = 0x61C88647 */
+static inline u32 __attribute_const__ __hash_32(u32 a)
+{
+#if CONFIG_XILINX_MICROBLAZE0_USE_BARREL
+	unsigned int b, c;
+
+	/* Phase 1: Compute three intermediate values */
+	b =  a << 23;
+	c = (a << 19) + a;
+	a = (a <<  9) + c;
+	b += a;
+
+	/* Phase 2: Compute (a << 11) + (b << 6) + (c << 3) - b */
+	a <<= 5;
+	a += b;		/* (a << 5) + b */
+	a <<= 3;
+	a += c;		/* (a << 8) + (b << 3) + c */
+	a <<= 3;
+	return a - b;	/* (a << 11) + (b << 6) + (c << 3) - b */
+#else
+	/*
+	 * "This is really going to hurt."
+	 *
+	 * Without a barrel shifter, left shifts are implemented as
+	 * repeated additions, and the best we can do is an optimal
+	 * addition-subtraction chain.  This one is not known to be
+	 * optimal, but at 37 steps, it's decent for a 31-bit multiplier.
+	 *
+	 * Question: given its size (37*4 = 148 bytes per instance),
+	 * and slowness, is this worth having inline?
+	 */
+	unsigned int b, c, d;
+
+	b = a << 4;	/* 4    */
+	c = b << 1;	/* 1  5 */
+	b += a;		/* 1  6 */
+	c += b;		/* 1  7 */
+	c <<= 3;	/* 3 10 */
+	c -= a;		/* 1 11 */
+	d = c << 7;	/* 7 18 */
+	d += b;		/* 1 19 */
+	d <<= 8;	/* 8 27 */
+	d += a;		/* 1 28 */
+	d <<= 1;	/* 1 29 */
+	d += b;		/* 1 30 */
+	d <<= 6;	/* 6 36 */
+	return d + c;	/* 1 37 total instructions*/
+#endif
+}
+
+#endif /* !CONFIG_XILINX_MICROBLAZE0_USE_HW_MUL */
+#endif /* _ASM_HASH_H */
diff --git a/drivers/media/usb/dvb-usb-v2/af9015.c b/drivers/media/usb/dvb-usb-v2/af9015.c
index 95a7388..09e0f58 100644
--- a/drivers/media/usb/dvb-usb-v2/af9015.c
+++ b/drivers/media/usb/dvb-usb-v2/af9015.c
@@ -398,6 +398,8 @@ error:
 }
 
 #define AF9015_EEPROM_SIZE 256
+/* 2^31 + 2^29 - 2^25 + 2^22 - 2^19 - 2^16 + 1 */
+#define GOLDEN_RATIO_PRIME_32 0x9e370001UL
 
 /* hash (and dump) eeprom */
 static int af9015_eeprom_hash(struct dvb_usb_device *d)
diff --git a/fs/dcache.c b/fs/dcache.c
index c622872..ad4a542 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -1670,8 +1670,7 @@ struct dentry *d_alloc_name(struct dentry *parent, const char *name)
 	struct qstr q;
 
 	q.name = name;
-	q.len = strlen(name);
-	q.hash = full_name_hash(q.name, q.len);
+	q.hash_len = hashlen_string(name);
 	return d_alloc(parent, &q);
 }
 EXPORT_SYMBOL(d_alloc_name);
diff --git a/fs/namei.c b/fs/namei.c
index 15b124c..e7bf99d 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -35,6 +35,7 @@
 #include <linux/fs_struct.h>
 #include <linux/posix_acl.h>
 #include <linux/hash.h>
+#include <linux/bitops.h>
 #include <asm/uaccess.h>
 
 #include "internal.h"
@@ -1797,74 +1798,144 @@ static int walk_component(struct nameidata *nd, int flags)
 
 #include <asm/word-at-a-time.h>
 
-#ifdef CONFIG_64BIT
+#ifdef HASH_MIX
 
-static inline unsigned int fold_hash(unsigned long hash)
-{
-	return hash_64(hash, 32);
-}
+/* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
 
+#elif defined(CONFIG_64BIT)
 /*
- * This is George Marsaglia's XORSHIFT generator.
- * It implements a maximum-period LFSR in only a few
- * instructions.  It also has the property (required
- * by hash_name()) that mix_hash(0) = 0.
+ * Register pressure in the mixing function is an issue, particularly
+ * on 32-bit x86, but almost any function requires one state value and
+ * one temporary.  Instead, use a function designed for two state values
+ * and no temporaries.
+ *
+ * This function cannot create a collision in only two iterations, so
+ * we have two iterations to achieve avalanche.  In those two iterations,
+ * we have six layers of mixing, which is enough to spread one bit's
+ * influence out to 2^6 = 64 state bits.
+ *
+ * Rotate constants are scored by considering either 64 one-bit input
+ * deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
+ * probability of that delta causing a change to each of the 128 output
+ * bits, using a sample of random initial states.
+ *
+ * The Shannon entropy of the computed probabilities is then summed
+ * to produce a score.  Ideally, any input change has a 50% chance of
+ * toggling any given output bit.
+ *
+ * Mixing scores (in bits) for (12,45):
+ * Input delta: 1-bit      2-bit
+ * 1 round:     713.3    42542.6
+ * 2 rounds:   2753.7   140389.8
+ * 3 rounds:   5954.1   233458.2
+ * 4 rounds:   7862.6   256672.2
+ * Perfect:    8192     258048
+ *            (64*128) (64*63/2 * 128)
  */
-static inline unsigned long mix_hash(unsigned long hash)
+#define HASH_MIX(x, y, a)	\
+	(	x ^= (a),	\
+	y ^= x,	x = rol64(x,12),\
+	x += y,	y = rol64(y,45),\
+	y *= 9			)
+
+/*
+ * Fold two longs into one 32-bit hash value.  This must be fast, but
+ * latency isn't quite as critical, as there is a fair bit of additional
+ * work done before the hash value is used.
+ */
+static inline unsigned int fold_hash(unsigned long x, unsigned long y)
 {
-	hash ^= hash << 13;
-	hash ^= hash >> 7;
-	hash ^= hash << 17;
-	return hash;
+	y ^= x * GOLDEN_RATIO_64;
+	y *= GOLDEN_RATIO_64;
+	return y >> 32;
 }
 
 #else	/* 32-bit case */
 
-#define fold_hash(x) (x)
+/*
+ * Mixing scores (in bits) for (7,20):
+ * Input delta: 1-bit      2-bit
+ * 1 round:     330.3     9201.6
+ * 2 rounds:   1246.4    25475.4
+ * 3 rounds:   1907.1    31295.1
+ * 4 rounds:   2042.3    31718.6
+ * Perfect:    2048      31744
+ *            (32*64)   (32*31/2 * 64)
+ */
+#define HASH_MIX(x, y, a)	\
+	(	x ^= (a),	\
+	y ^= x,	x = rol32(x, 7),\
+	x += y,	y = rol32(y,20),\
+	y *= 9			)
 
-static inline unsigned long mix_hash(unsigned long hash)
+static inline unsigned int fold_hash(unsigned long x, unsigned long y)
 {
-	hash ^= hash << 13;
-	hash ^= hash >> 17;
-	hash ^= hash << 5;
-	return hash;
+	/* Use arch-optimized multiply if one exists */
+	return __hash_32(y ^ __hash_32(x));
 }
 
 #endif
 
-unsigned int full_name_hash(const unsigned char *name, unsigned int len)
+/*
+ * Return the hash of a string of known length.  This is carfully
+ * designed to match hash_name(), which is the more critical function.
+ * In particular, we must end by hashing a final word containing 0..7
+ * payload bytes, to match the way that hash_name() iterates until it
+ * finds the delimiter after the name.
+ */
+unsigned int full_name_hash(const char *name, unsigned int len)
 {
-	unsigned long a, hash = 0;
+	unsigned long a, x = 0, y = 0;
 
 	for (;;) {
+		if (!len)
+			goto done;
 		a = load_unaligned_zeropad(name);
 		if (len < sizeof(unsigned long))
 			break;
-		hash = mix_hash(hash + a);
+		HASH_MIX(x, y, a);
 		name += sizeof(unsigned long);
 		len -= sizeof(unsigned long);
-		if (!len)
-			goto done;
 	}
-	hash += a & bytemask_from_count(len);
+	x ^= a & bytemask_from_count(len);
 done:
-	return fold_hash(hash);
+	return fold_hash(x, y);
 }
 EXPORT_SYMBOL(full_name_hash);
 
+/* Return the "hash_len" (hash and length) of a null-terminated string */
+u64 hashlen_string(const char *name)
+{
+	unsigned long a = 0, x = 0, y = 0, adata, mask, len;
+	const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
+
+	len = -sizeof(unsigned long);
+	do {
+		HASH_MIX(x, y, a);
+		len += sizeof(unsigned long);
+		a = load_unaligned_zeropad(name+len);
+	} while (!has_zero(a, &adata, &constants));
+
+	adata = prep_zero_mask(a, adata, &constants);
+	mask = create_zero_mask(adata);
+	x ^= a & zero_bytemask(mask);
+
+	return hashlen_create(fold_hash(x, y), len + find_zero(mask));
+}
+EXPORT_SYMBOL(hashlen_string);
+
 /*
  * Calculate the length and hash of the path component, and
  * return the "hash_len" as the result.
  */
 static inline u64 hash_name(const char *name)
 {
-	unsigned long a, b, adata, bdata, mask, hash, len;
+	unsigned long a = 0, b, x = 0, y = 0, adata, bdata, mask, len;
 	const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
 
-	hash = a = 0;
 	len = -sizeof(unsigned long);
 	do {
-		hash = mix_hash(hash + a);
+		HASH_MIX(x, y, a);
 		len += sizeof(unsigned long);
 		a = load_unaligned_zeropad(name+len);
 		b = a ^ REPEAT_BYTE('/');
@@ -1872,25 +1943,40 @@ static inline u64 hash_name(const char *name)
 
 	adata = prep_zero_mask(a, adata, &constants);
 	bdata = prep_zero_mask(b, bdata, &constants);
-
 	mask = create_zero_mask(adata | bdata);
+	x ^= a & zero_bytemask(mask);
 
-	hash += a & zero_bytemask(mask);
-	len += find_zero(mask);
-	return hashlen_create(fold_hash(hash), len);
+	return hashlen_create(fold_hash(x, y), len + find_zero(mask));
 }
 
-#else
+#else	/* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
 
-unsigned int full_name_hash(const unsigned char *name, unsigned int len)
+/* Return the hash of a string of known length */
+unsigned int full_name_hash(const char *name, unsigned int len)
 {
 	unsigned long hash = init_name_hash();
 	while (len--)
-		hash = partial_name_hash(*name++, hash);
+		hash = partial_name_hash((unsigned char)*name++, hash);
 	return end_name_hash(hash);
 }
 EXPORT_SYMBOL(full_name_hash);
 
+/* Return the "hash_len" (hash and length) of a null-terminated string */
+u64 hash_string(const char *name)
+{
+	unsigned long hash = init_name_hash();
+	unsigned long len = 0, c;
+
+	c = (unsigned char)*name;
+	do {
+		len++;
+		hash = partial_name_hash(c, hash);
+		c = (unsigned char)name[len];
+	} while (c);
+	return hashlen_create(end_name_hash(hash), len);
+}
+EXPORT_SYMBOL(hash_string);
+
 /*
  * We know there's a real path component here of at least
  * one character.
@@ -1934,7 +2020,7 @@ static int link_path_walk(const char *name, struct nameidata *nd)
 		int type;
 
 		err = may_lookup(nd);
- 		if (err)
+		if (err)
 			return err;
 
 		hash_len = hash_name(name);
diff --git a/include/linux/dcache.h b/include/linux/dcache.h
index f8506e8..484c879 100644
--- a/include/linux/dcache.h
+++ b/include/linux/dcache.h
@@ -10,6 +10,7 @@
 #include <linux/cache.h>
 #include <linux/rcupdate.h>
 #include <linux/lockref.h>
+#include <linux/stringhash.h>
 
 struct path;
 struct vfsmount;
@@ -52,9 +53,6 @@ struct qstr {
 };
 
 #define QSTR_INIT(n,l) { { { .len = l } }, .name = n }
-#define hashlen_hash(hashlen) ((u32) (hashlen))
-#define hashlen_len(hashlen)  ((u32)((hashlen) >> 32))
-#define hashlen_create(hash,len) (((u64)(len)<<32)|(u32)(hash))
 
 struct dentry_stat_t {
 	long nr_dentry;
@@ -65,29 +63,6 @@ struct dentry_stat_t {
 };
 extern struct dentry_stat_t dentry_stat;
 
-/* Name hashing routines. Initial hash value */
-/* Hash courtesy of the R5 hash in reiserfs modulo sign bits */
-#define init_name_hash()		0
-
-/* partial hash update function. Assume roughly 4 bits per character */
-static inline unsigned long
-partial_name_hash(unsigned long c, unsigned long prevhash)
-{
-	return (prevhash + (c << 4) + (c >> 4)) * 11;
-}
-
-/*
- * Finally: cut down the number of bits to a int value (and try to avoid
- * losing bits)
- */
-static inline unsigned long end_name_hash(unsigned long hash)
-{
-	return (unsigned int) hash;
-}
-
-/* Compute the hash for a name string. */
-extern unsigned int full_name_hash(const unsigned char *, unsigned int);
-
 /*
  * Try to keep struct dentry aligned on 64 byte cachelines (this will
  * give reasonable cacheline footprint with larger lines without the
diff --git a/include/linux/hash.h b/include/linux/hash.h
index 79c52fa..ad6fa21 100644
--- a/include/linux/hash.h
+++ b/include/linux/hash.h
@@ -3,92 +3,94 @@
 /* Fast hashing routine for ints,  longs and pointers.
    (C) 2002 Nadia Yvette Chambers, IBM */
 
-/*
- * Knuth recommends primes in approximately golden ratio to the maximum
- * integer representable by a machine word for multiplicative hashing.
- * Chuck Lever verified the effectiveness of this technique:
- * http://www.citi.umich.edu/techreports/reports/citi-tr-00-1.pdf
- *
- * These primes are chosen to be bit-sparse, that is operations on
- * them can use shifts and additions instead of multiplications for
- * machines where multiplications are slow.
- */
-
 #include <asm/types.h>
 #include <linux/compiler.h>
 
-/* 2^31 + 2^29 - 2^25 + 2^22 - 2^19 - 2^16 + 1 */
-#define GOLDEN_RATIO_PRIME_32 0x9e370001UL
-/*  2^63 + 2^61 - 2^57 + 2^54 - 2^51 - 2^18 + 1 */
-#define GOLDEN_RATIO_PRIME_64 0x9e37fffffffc0001UL
-
+/*
+ * The "GOLDEN_RATIO_PRIME" is used in ifs/btrfs/brtfs_inode.h and
+ * fs/inode.c.  It's not actually prime any more (the previous primes
+ * were actively bad for hashing), but the name remains.
+ */
 #if BITS_PER_LONG == 32
-#define GOLDEN_RATIO_PRIME GOLDEN_RATIO_PRIME_32
+#define GOLDEN_RATIO_PRIME GOLDEN_RATIO_32
 #define hash_long(val, bits) hash_32(val, bits)
 #elif BITS_PER_LONG == 64
 #define hash_long(val, bits) hash_64(val, bits)
-#define GOLDEN_RATIO_PRIME GOLDEN_RATIO_PRIME_64
+#define GOLDEN_RATIO_PRIME GOLDEN_RATIO_64
 #else
 #error Wordsize not 32 or 64
 #endif
 
 /*
- * The above primes are actively bad for hashing, since they are
- * too sparse. The 32-bit one is mostly ok, the 64-bit one causes
- * real problems. Besides, the "prime" part is pointless for the
- * multiplicative hash.
+ * This hash multiplies the input by a large odd number and takes the
+ * high bits.  Since multiplication propagates changes to the most
+ * significant end only, it is essential that the high bits of the
+ * product be used for the hash value.
+ *
+ * Chuck Lever verified the effectiveness of this technique:
+ * http://www.citi.umich.edu/techreports/reports/citi-tr-00-1.pdf
  *
  * Although a random odd number will do, it turns out that the golden
  * ratio phi = (sqrt(5)-1)/2, or its negative, has particularly nice
- * properties.
+ * properties.  (See Knuth vol 3, section 6.4, exercise 9.)
  *
- * These are the negative, (1 - phi) = (phi^2) = (3 - sqrt(5))/2.
- * (See Knuth vol 3, section 6.4, exercise 9.)
+ * These are the negative, (1 - phi) = phi**2 = (3 - sqrt(5))/2,
+ * which is very slightly easier to multiply by and makes no
+ * difference to the hash distribution.
  */
 #define GOLDEN_RATIO_32 0x61C88647
 #define GOLDEN_RATIO_64 0x61C8864680B583EBull
 
-static __always_inline u64 hash_64(u64 val, unsigned int bits)
-{
-	u64 hash = val;
+#ifdef CONFIG_HAVE_ARCH_HASH
+/* This header may use the GOLDEN_RATIO_xx constants */
+#include <asm/hash.h>
+#endif
 
-#if BITS_PER_LONG == 64
-	hash = hash * GOLDEN_RATIO_64;
-#else
-	/*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
-	u64 n = hash;
-	n <<= 18;
-	hash -= n;
-	n <<= 33;
-	hash -= n;
-	n <<= 3;
-	hash += n;
-	n <<= 3;
-	hash -= n;
-	n <<= 4;
-	hash += n;
-	n <<= 2;
-	hash += n;
+/*
+ * The _generic versions exist only so lib/test_hash.c can compare
+ * the arch-optimized versions with the generic.
+ *
+ * Note that if you change these, any <asm/hash.h> that aren't updated
+ * to match need to have their HAVE_ARCH_* define values updated so the
+ * self-test will not false-positive.
+ */
+#ifndef HAVE_ARCH__HASH_32
+#define __hash_32 __hash_32_generic
 #endif
+static inline u32 __hash_32_generic(u32 val)
+{
+	return val * GOLDEN_RATIO_32;
+}
 
+#ifndef HAVE_ARCH_HASH_32
+#define hash_32 hash_32_generic
+#endif
+static inline u32 hash_32_generic(u32 val, unsigned int bits)
+{
 	/* High bits are more random, so use them. */
-	return hash >> (64 - bits);
+	return __hash_32(val) >> (32 - bits);
 }
 
-static inline u32 hash_32(u32 val, unsigned int bits)
+#ifndef HAVE_ARCH_HASH_64
+#define hash_64 hash_64_generic
+#endif
+static __always_inline u32 hash_64_generic(u64 val, unsigned int bits)
 {
-	/* On some cpus multiply is faster, on others gcc will do shifts */
-	u32 hash = val * GOLDEN_RATIO_PRIME_32;
-
-	/* High bits are more random, so use them. */
-	return hash >> (32 - bits);
+#if BITS_PER_LONG == 64
+	/* 64x64-bit multiply is efficient on all 64-bit processors */
+	return val * GOLDEN_RATIO_64 >> (64 - bits);
+#else
+	/* Hash 64 bits using only 32x32-bit multiply. */
+	return hash_32((u32)val ^ __hash_32(val >> 32), bits);
+#endif
 }
 
-static inline unsigned long hash_ptr(const void *ptr, unsigned int bits)
+static inline u32 hash_ptr(const void *ptr, unsigned int bits)
 {
 	return hash_long((unsigned long)ptr, bits);
 }
 
+/* This really should be called fold32_ptr; it does no hashing to speak of. */
 static inline u32 hash32_ptr(const void *ptr)
 {
 	unsigned long val = (unsigned long)ptr;
diff --git a/include/linux/stringhash.h b/include/linux/stringhash.h
new file mode 100644
index 0000000..451771d
--- /dev/null
+++ b/include/linux/stringhash.h
@@ -0,0 +1,76 @@
+#ifndef __LINUX_STRINGHASH_H
+#define __LINUX_STRINGHASH_H
+
+#include <linux/compiler.h>	/* For __pure */
+#include <linux/types.h>	/* For u32, u64 */
+
+/*
+ * Routines for hashing strings of bytes to a 32-bit hash value.
+ *
+ * These hash functions are NOT GUARANTEED STABLE between kernel
+ * versions, architectures, or even repeated boots of the same kernel.
+ * (E.g. they may depend on boot-time hardware detection or be
+ * deliberately randomized.)
+ *
+ * They are also not intended to be secure against collisions caused by
+ * malicious inputs; much slower hash functions are required for that.
+ *
+ * They are optimized for pathname components, meaning short strings.
+ * Even if a majority of files have longer names, the dynamic profile of
+ * pathname components skews short due to short directory names.
+ * (E.g. /usr/lib/libsesquipedalianism.so.3.141.)
+ */
+
+/*
+ * Version 1: one byte at a time.  Example of use:
+ *
+ * unsigned long hash = init_name_hash;
+ * while (*p)
+ *	hash = partial_name_hash(tolower(*p++), hash);
+ * hash = end_name_hash(hash);
+ *
+ * Although this is designed for bytes, fs/hfsplus/unicode.c
+ * abuses it to hash 16-bit values.
+ */
+
+/* Hash courtesy of the R5 hash in reiserfs modulo sign bits */
+#define init_name_hash()		0
+
+/* partial hash update function. Assume roughly 4 bits per character */
+static inline unsigned long
+partial_name_hash(unsigned long c, unsigned long prevhash)
+{
+	return (prevhash + (c << 4) + (c >> 4)) * 11;
+}
+
+/*
+ * Finally: cut down the number of bits to a int value (and try to avoid
+ * losing bits)
+ */
+static inline unsigned long end_name_hash(unsigned long hash)
+{
+	return (unsigned int)hash;
+}
+
+/*
+ * Version 2: One word (32 or 64 bits) at a time.
+ * If CONFIG_DCACHE_WORD_ACCESS is defined (meaning <asm/word-at-a-time.h>
+ * exists, which describes major Linux platforms like x86 and ARM), then
+ * this computes a different hash function much faster.
+ *
+ * If not set, this falls back to a wrapper around the preceding.
+ */
+extern unsigned int __pure full_name_hash(const char *, unsigned int);
+
+/*
+ * A hash_len is a u64 with the hash of a string in the low
+ * half and the length in the high half.
+ */
+#define hashlen_hash(hashlen) ((u32)(hashlen))
+#define hashlen_len(hashlen)  ((u32)((hashlen) >> 32))
+#define hashlen_create(hash, len) ((u64)(len)<<32 | (u32)(hash))
+
+/* Return the "hash_len" (hash and length) of a null-terminated string */
+extern u64 __pure hashlen_string(const char *name);
+
+#endif	/* __LINUX_STRINGHASH_H */
diff --git a/include/linux/sunrpc/svcauth.h b/include/linux/sunrpc/svcauth.h
index c00f53a..91d5a5d 100644
--- a/include/linux/sunrpc/svcauth.h
+++ b/include/linux/sunrpc/svcauth.h
@@ -16,6 +16,7 @@
 #include <linux/sunrpc/cache.h>
 #include <linux/sunrpc/gss_api.h>
 #include <linux/hash.h>
+#include <linux/stringhash.h>
 #include <linux/cred.h>
 
 struct svc_cred {
@@ -165,41 +166,18 @@ extern int svcauth_unix_set_client(struct svc_rqst *rqstp);
 extern int unix_gid_cache_create(struct net *net);
 extern void unix_gid_cache_destroy(struct net *net);
 
-static inline unsigned long hash_str(char *name, int bits)
+/*
+ * The <stringhash.h> functions are good enough that we don't need to
+ * use hash_32() on them; just extracting the high bits is enough.
+ */
+static inline unsigned long hash_str(char const *name, int bits)
 {
-	unsigned long hash = 0;
-	unsigned long l = 0;
-	int len = 0;
-	unsigned char c;
-	do {
-		if (unlikely(!(c = *name++))) {
-			c = (char)len; len = -1;
-		}
-		l = (l << 8) | c;
-		len++;
-		if ((len & (BITS_PER_LONG/8-1))==0)
-			hash = hash_long(hash^l, BITS_PER_LONG);
-	} while (len);
-	return hash >> (BITS_PER_LONG - bits);
+	return hashlen_hash(hashlen_string(name)) >> (32 - bits);
 }
 
-static inline unsigned long hash_mem(char *buf, int length, int bits)
+static inline unsigned long hash_mem(char const *buf, int length, int bits)
 {
-	unsigned long hash = 0;
-	unsigned long l = 0;
-	int len = 0;
-	unsigned char c;
-	do {
-		if (len == length) {
-			c = (char)len; len = -1;
-		} else
-			c = *buf++;
-		l = (l << 8) | c;
-		len++;
-		if ((len & (BITS_PER_LONG/8-1))==0)
-			hash = hash_long(hash^l, BITS_PER_LONG);
-	} while (len);
-	return hash >> (BITS_PER_LONG - bits);
+	return full_name_hash(buf, length) >> (32 - bits);
 }
 
 #endif /* __KERNEL__ */
diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
index e707ab3..77d7d03 100644
--- a/lib/Kconfig.debug
+++ b/lib/Kconfig.debug
@@ -1849,6 +1849,17 @@ config TEST_RHASHTABLE
 
 	  If unsure, say N.
 
+config TEST_HASH
+	tristate "Perform selftest on hash functions"
+	default n
+	help
+	  Enable this option to test the kernel's integer (<linux/hash,h>)
+	  and string (<linux/stringhash.h>) hash functions on boot
+	  (or module load).
+
+	  This is intended to help people writing architecture-specific
+	  optimized versions.  If unsure, say N.
+
 endmenu # runtime tests
 
 config PROVIDE_OHCI1394_DMA_INIT
diff --git a/lib/Makefile b/lib/Makefile
index 42b6918..499fb35 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -48,6 +48,7 @@ obj-$(CONFIG_TEST_HEXDUMP) += test_hexdump.o
 obj-y += kstrtox.o
 obj-$(CONFIG_TEST_BPF) += test_bpf.o
 obj-$(CONFIG_TEST_FIRMWARE) += test_firmware.o
+obj-$(CONFIG_TEST_HASH) += test_hash.o
 obj-$(CONFIG_TEST_KASAN) += test_kasan.o
 obj-$(CONFIG_TEST_KSTRTOX) += test-kstrtox.o
 obj-$(CONFIG_TEST_LKM) += test_module.o
diff --git a/lib/test_hash.c b/lib/test_hash.c
new file mode 100644
index 0000000..c9549c8
--- /dev/null
+++ b/lib/test_hash.c
@@ -0,0 +1,250 @@
+/*
+ * Test cases for <linux/hash.h> and <linux/stringhash.h>
+ * This just verifies that various ways of computing a hash
+ * produce the same thing and, for cases where a k-bit hash
+ * value is requested, is of the requested size.
+ *
+ * We fill a buffer with a 255-byte null-terminated string,
+ * and use both full_name_hash() and hashlen_string() to hash the
+ * substrings from i to j, where 0 <= i < j < 256.
+ *
+ * The returned values are used to check that __hash_32() and
+ * __hash_32_generic() compute the same thing.  Likewise hash_32()
+ * and hash_64().
+ */
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt "\n"
+
+#include <linux/compiler.h>
+#include <linux/types.h>
+#include <linux/module.h>
+#include <linux/hash.h>
+#include <linux/stringhash.h>
+#include <linux/printk.h>
+
+/* 32-bit XORSHIFT generator.  Seed must not be zero. */
+static u32 __init __attribute_const__
+xorshift(u32 seed)
+{
+	seed ^= seed << 13;
+	seed ^= seed >> 17;
+	seed ^= seed << 5;
+	return seed;
+}
+
+/* Given a non-zero x, returns a non-zero byte. */
+static u8 __init __attribute_const__
+mod255(u32 x)
+{
+	x = (x & 0xffff) + (x >> 16);	/* 1 <= x <= 0x1fffe */
+	x = (x & 0xff) + (x >> 8);	/* 1 <= x <= 0x2fd */
+	x = (x & 0xff) + (x >> 8);	/* 1 <= x <= 0x100 */
+	x = (x & 0xff) + (x >> 8);	/* 1 <= x <= 0xff */
+	return x;
+}
+
+/* Fill the buffer with non-zero bytes. */
+static void __init
+fill_buf(char *buf, size_t len, u32 seed)
+{
+	size_t i;
+
+	for (i = 0; i < len; i++) {
+		seed = xorshift(seed);
+		buf[i] = mod255(seed);
+	}
+}
+
+/*
+ * Test the various integer hash functions.  h64 (or its low-order bits)
+ * is the integer to hash.  hash_or accumulates the OR of the hash values,
+ * which are later checked to see that they cover all the requested bits.
+ *
+ * Because these functions (as opposed to the string hashes) are all
+ * inline, the code being tested is actually in the module, and you can
+ * recompile and re-test the module without rebooting.
+ */
+static bool __init
+test_int_hash(unsigned long long h64, u32 hash_or[2][33])
+{
+	int k;
+	u32 h0 = (u32)h64, h1, h2;
+
+	/* Test __hash32 */
+	hash_or[0][0] |= h1 = __hash_32(h0);
+#ifdef HAVE_ARCH__HASH_32
+	hash_or[1][0] |= h2 = __hash_32_generic(h0);
+#if HAVE_ARCH__HASH_32 == 1
+	if (h1 != h2) {
+		pr_err("__hash_32(%#x) = %#x != __hash_32_generic() = %#x",
+			h0, h1, h2);
+		return false;
+	}
+#endif
+#endif
+
+	/* Test k = 1..32 bits */
+	for (k = 1; k <= 32; k++) {
+		u32 const m = ((u32)2 << (k-1)) - 1;	/* Low k bits set */
+
+		/* Test hash_32 */
+		hash_or[0][k] |= h1 = hash_32(h0, k);
+		if (h1 > m) {
+			pr_err("hash_32(%#x, %d) = %#x > %#x", h0, k, h1, m);
+			return false;
+		}
+#ifdef HAVE_ARCH_HASH_32
+		h2 = hash_32_generic(h0, k);
+#if HAVE_ARCH_HASH_32 == 1
+		if (h1 != h2) {
+			pr_err("hash_32(%#x, %d) = %#x != hash_32_generic() "
+				" = %#x", h0, k, h1, h2);
+			return false;
+		}
+#else
+		if (h2 > m) {
+			pr_err("hash_32_generic(%#x, %d) = %#x > %#x",
+				h0, k, h1, m);
+			return false;
+		}
+#endif
+#endif
+		/* Test hash_64 */
+		hash_or[1][k] |= h1 = hash_64(h64, k);
+		if (h1 > m) {
+			pr_err("hash_64(%#llx, %d) = %#x > %#x", h64, k, h1, m);
+			return false;
+		}
+#ifdef HAVE_ARCH_HASH_64
+		h2 = hash_64_generic(h64, k);
+#if HAVE_ARCH_HASH_64 == 1
+		if (h1 != h2) {
+			pr_err("hash_64(%#llx, %d) = %#x != hash_64_generic() "
+				"= %#x", h64, k, h1, h2);
+			return false;
+		}
+#else
+		if (h2 > m) {
+			pr_err("hash_64_generic(%#llx, %d) = %#x > %#x",
+				h64, k, h1, m);
+			return false;
+		}
+#endif
+#endif
+	}
+
+	(void)h2;	/* Suppress unused variable warning */
+	return true;
+}
+
+#define SIZE 256	/* Run time is cubic in SIZE */
+
+static int __init
+test_hash_init(void)
+{
+	char buf[SIZE+1];
+	u32 string_or = 0, hash_or[2][33] = { 0 };
+	unsigned tests = 0;
+	unsigned long long h64 = 0;
+	int i, j;
+
+	fill_buf(buf, SIZE, 1);
+
+	/* Test every possible non-empty substring in the buffer. */
+	for (j = SIZE; j > 0; --j) {
+		buf[j] = '\0';
+
+		for (i = 0; i <= j; i++) {
+			u64 hashlen = hashlen_string(buf+i);
+			u32 h0 = full_name_hash(buf+i, j-i);
+
+			/* Check that hashlen_string gets the length right */
+			if (hashlen_len(hashlen) != j-i) {
+				pr_err("hashlen_string(%d..%d) returned length"
+					" %u, expected %d",
+					i, j, hashlen_len(hashlen), j-i);
+				return -EINVAL;
+			}
+			/* Check that the hashes match */
+			if (hashlen_hash(hashlen) != h0) {
+				pr_err("hashlen_string(%d..%d) = %08x != "
+					"full_name_hash() = %08x",
+					i, j, hashlen_hash(hashlen), h0);
+				return -EINVAL;
+			}
+
+			string_or |= h0;
+			h64 = h64 << 32 | h0;	/* For use with hash_64 */
+			if (!test_int_hash(h64, hash_or))
+				return -EINVAL;
+			tests++;
+		} /* i */
+	} /* j */
+
+	/* The OR of all the hash values should cover all the bits */
+	if (~string_or) {
+		pr_err("OR of all string hash results = %#x != %#x",
+			string_or, -1u);
+		return -EINVAL;
+	}
+	if (~hash_or[0][0]) {
+		pr_err("OR of all __hash_32 results = %#x != %#x",
+			hash_or[0][0], -1u);
+		return -EINVAL;
+	}
+#ifdef HAVE_ARCH__HASH_32
+#if HAVE_ARCH__HASH_32 != 1	/* Test is pointless if results match */
+	if (~hash_or[1][0]) {
+		pr_err("OR of all __hash_32_generic results = %#x != %#x",
+			hash_or[1][0], -1u);
+		return -EINVAL;
+	}
+#endif
+#endif
+
+	/* Likewise for all the i-bit hash values */
+	for (i = 1; i <= 32; i++) {
+		u32 const m = ((u32)2 << (i-1)) - 1;	/* Low i bits set */
+
+		if (hash_or[0][i] != m) {
+			pr_err("OR of all hash_32(%d) results = %#x "
+				"(%#x expected)", i, hash_or[0][i], m);
+			return -EINVAL;
+		}
+		if (hash_or[1][i] != m) {
+			pr_err("OR of all hash_64(%d) results = %#x "
+				"(%#x expected)", i, hash_or[1][i], m);
+			return -EINVAL;
+		}
+	}
+
+	/* Issue notices about skipped tests. */
+#ifndef HAVE_ARCH__HASH_32
+	pr_info("__hash_32() has no arch implementation to test.");
+#elif HAVE_ARCH__HASH_32 != 1
+	pr_info("__hash_32() is arch-specific; not compared to generic.");
+#endif
+#ifndef HAVE_ARCH_HASH_32
+	pr_info("hash_32() has no arch implementation to test.");
+#elif HAVE_ARCH_HASH_32 != 1
+	pr_info("hash_32() is arch-specific; not compared to generic.");
+#endif
+#ifndef HAVE_ARCH_HASH_64
+	pr_info("hash_64() has no arch implementation to test.");
+#elif HAVE_ARCH_HASH_64 != 1
+	pr_info("hash_64() is arch-specific; not compared to generic.");
+#endif
+
+	pr_notice("%u tests passed.", tests);
+
+	return 0;
+}
+
+static void __exit test_hash_exit(void)
+{
+}
+
+module_init(test_hash_init);	/* Does everything */
+module_exit(test_hash_exit);	/* Does nothing */
+
+MODULE_LICENSE("GPL");