Message ID | 20210826012626.1163705-2-isabellabdoamaral@usp.br |
---|---|
State | New |
Headers | show |
Series | test_hash.c: refactor into KUnit | expand |
On Thu, Aug 26, 2021 at 9:26 AM Isabella Basso <isabellabdoamaral@usp.br> wrote: > > The HAVE_ARCH_HASH_32 (single underscore) define hasn't been used for > any known supported architectures that have their own hash function > implementation (i.e. m68k, Microblaze, H8/300, pa-risc) since George's > patch [1], which introduced it. > > The supported 32-bit architectures from the list above have only been > making use of the (more general) HAVE_ARCH__HASH_32 define, which only > lacks the right shift operator, that wasn't targeted for optimizations > so far. > > [1] https://lore.kernel.org/lkml/20160525073311.5600.qmail@ns.sciencehorizons.net/ > > Co-developed-by: Augusto Durães Camargo <augusto.duraes33@gmail.com> > Signed-off-by: Augusto Durães Camargo <augusto.duraes33@gmail.com> > Co-developed-by: Enzo Ferreira <ferreiraenzoa@gmail.com> > Signed-off-by: Enzo Ferreira <ferreiraenzoa@gmail.com> > Signed-off-by: Isabella Basso <isabellabdoamaral@usp.br> > --- I'm not familiar with the hash functions here, so take this with the appropriate heap of salt, but it took me a little while to understand exactly what this is doing. As I understand it: - There are separate __hash_32() and hash_32() functions. - Both of these have generic implementations, which can optionally be overridden by an architecture-specific optimised version. - There aren't any architectures which provide an optimised hash_32() implementation. - This patch therefore removes support for architecture-specific hash_32() implementations, and leaves only the generic implementation. - This generic implementation of hash_32() itself relies on __hash_32(), which may still be optimised. Could the commit description be updated to make this a bit clearer? While we are getting rid of the HAVE_ARCH_HASH_32 #define, that seems to be a side-effect/implementation detail of removing support for architecture-specific hash_32() implementations... The other wild, out-there option would be to remove __hash_32() entirely and make everything use hash_32(), which then could have architecture-specific implementations. A quick grep reveals that there's only one use of __hash_32() outside of the hashing code itself (in fs/namei.c). This would be much more consistent with what hash_64() does, but also would be significantly more work, and potentially could have some implication (full_name_hash() performance maybe?) which I'm not aware of. So it's possibly not worth it. Cheers, -- David > include/linux/hash.h | 5 +---- > lib/test_hash.c | 24 +----------------------- > tools/include/linux/hash.h | 5 +---- > 3 files changed, 3 insertions(+), 31 deletions(-) > > diff --git a/include/linux/hash.h b/include/linux/hash.h > index ad6fa21d977b..38edaa08f862 100644 > --- a/include/linux/hash.h > +++ b/include/linux/hash.h > @@ -62,10 +62,7 @@ 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) > +static inline u32 hash_32(u32 val, unsigned int bits) > { > /* High bits are more random, so use them. */ > return __hash_32(val) >> (32 - bits); > diff --git a/lib/test_hash.c b/lib/test_hash.c > index 0ee40b4a56dd..d4b0cfdb0377 100644 > --- a/lib/test_hash.c > +++ b/lib/test_hash.c > @@ -94,22 +94,7 @@ test_int_hash(unsigned long long h64, u32 hash_or[2][33]) > 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) { > @@ -227,13 +212,6 @@ test_hash_init(void) > #else > pr_info("__hash_32() has no arch implementation to test."); > #endif > -#ifdef HAVE_ARCH_HASH_32 > -#if HAVE_ARCH_HASH_32 != 1 > - pr_info("hash_32() is arch-specific; not compared to generic."); > -#endif > -#else > - pr_info("hash_32() has no arch implementation to test."); > -#endif > #ifdef HAVE_ARCH_HASH_64 > #if HAVE_ARCH_HASH_64 != 1 > pr_info("hash_64() is arch-specific; not compared to generic."); > diff --git a/tools/include/linux/hash.h b/tools/include/linux/hash.h > index ad6fa21d977b..38edaa08f862 100644 > --- a/tools/include/linux/hash.h > +++ b/tools/include/linux/hash.h > @@ -62,10 +62,7 @@ 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) > +static inline u32 hash_32(u32 val, unsigned int bits) > { > /* High bits are more random, so use them. */ > return __hash_32(val) >> (32 - bits); > -- > 2.33.0 >
Hi, David, On Thu, Aug 26, 2021 at 1:01 AM David Gow <davidgow@google.com> wrote: > > On Thu, Aug 26, 2021 at 9:26 AM Isabella Basso <isabellabdoamaral@usp.br> wrote: > > > > The HAVE_ARCH_HASH_32 (single underscore) define hasn't been used for > > any known supported architectures that have their own hash function > > implementation (i.e. m68k, Microblaze, H8/300, pa-risc) since George's > > patch [1], which introduced it. > > > > The supported 32-bit architectures from the list above have only been > > making use of the (more general) HAVE_ARCH__HASH_32 define, which only > > lacks the right shift operator, that wasn't targeted for optimizations > > so far. > > > > [1] https://lore.kernel.org/lkml/20160525073311.5600.qmail@ns.sciencehorizons.net/ > > > > Co-developed-by: Augusto Durães Camargo <augusto.duraes33@gmail.com> > > Signed-off-by: Augusto Durães Camargo <augusto.duraes33@gmail.com> > > Co-developed-by: Enzo Ferreira <ferreiraenzoa@gmail.com> > > Signed-off-by: Enzo Ferreira <ferreiraenzoa@gmail.com> > > Signed-off-by: Isabella Basso <isabellabdoamaral@usp.br> > > --- > > I'm not familiar with the hash functions here, so take this with the > appropriate heap of salt, but it took me a little while to understand > exactly what this is doing. > > As I understand it: > - There are separate __hash_32() and hash_32() functions. > - Both of these have generic implementations, which can optionally be > overridden by an architecture-specific optimised version. > - There aren't any architectures which provide an optimised hash_32() > implementation. > - This patch therefore removes support for architecture-specific > hash_32() implementations, and leaves only the generic implementation. > - This generic implementation of hash_32() itself relies on > __hash_32(), which may still be optimised. > > Could the commit description be updated to make this a bit clearer? Sure, that makes perfect sense! Thank you very much for the feedback! Writing those descriptions was quite a challenge for me, as I wasn't perfectly sure of what the appropriate reasoning should be for the message. I'm also glad I was able to get a grasp similar to yours. :) > While we are getting rid of the HAVE_ARCH_HASH_32 #define, that seems > to be a side-effect/implementation detail of removing support for > architecture-specific hash_32() implementations... > > The other wild, out-there option would be to remove __hash_32() > entirely and make everything use hash_32(), which then could have > architecture-specific implementations. A quick grep reveals that > there's only one use of __hash_32() outside of the hashing code itself > (in fs/namei.c). This would be much more consistent with what > hash_64() does, but also would be significantly more work, and > potentially could have some implication (full_name_hash() performance > maybe?) which I'm not aware of. So it's possibly not worth it. I do agree with you that it seems a bit over the top, as I'm also really not aware of the performance implications of such a change (and that seemed to be what motivated most of the patch series that introduced the __hash_32() to fs/namei.c), so I'd rather not mess with fs/namei.c based on consistency reasons alone. Thanks, -- Isabella Basso
diff --git a/include/linux/hash.h b/include/linux/hash.h index ad6fa21d977b..38edaa08f862 100644 --- a/include/linux/hash.h +++ b/include/linux/hash.h @@ -62,10 +62,7 @@ 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) +static inline u32 hash_32(u32 val, unsigned int bits) { /* High bits are more random, so use them. */ return __hash_32(val) >> (32 - bits); diff --git a/lib/test_hash.c b/lib/test_hash.c index 0ee40b4a56dd..d4b0cfdb0377 100644 --- a/lib/test_hash.c +++ b/lib/test_hash.c @@ -94,22 +94,7 @@ test_int_hash(unsigned long long h64, u32 hash_or[2][33]) 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) { @@ -227,13 +212,6 @@ test_hash_init(void) #else pr_info("__hash_32() has no arch implementation to test."); #endif -#ifdef HAVE_ARCH_HASH_32 -#if HAVE_ARCH_HASH_32 != 1 - pr_info("hash_32() is arch-specific; not compared to generic."); -#endif -#else - pr_info("hash_32() has no arch implementation to test."); -#endif #ifdef HAVE_ARCH_HASH_64 #if HAVE_ARCH_HASH_64 != 1 pr_info("hash_64() is arch-specific; not compared to generic."); diff --git a/tools/include/linux/hash.h b/tools/include/linux/hash.h index ad6fa21d977b..38edaa08f862 100644 --- a/tools/include/linux/hash.h +++ b/tools/include/linux/hash.h @@ -62,10 +62,7 @@ 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) +static inline u32 hash_32(u32 val, unsigned int bits) { /* High bits are more random, so use them. */ return __hash_32(val) >> (32 - bits);