| Message ID | 20170401151755.11875-1-omosnacek@gmail.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Herbert Xu |
| Headers | show |
Oops, sorry, wrong prefix... 2017-04-01 17:17 GMT+02:00 Ondrej Mosnacek <omosnacek@gmail.com>: > The gf128mul_x_ble function is currently defined in gf128mul.c, because > it depends on the gf128mul_table_be multiplication table. > > However, since the function is very small and only uses two values from > the table, it is better for it to be defined as inline function in > gf128mul.h. That way, the function can be inlined by the compiler for > better performance. > > For consistency, the other gf128mul_x_* functions are also moved to the > header file. In addition, the code is rewritten to be constant-time. > > After this change, the speed of the generic 'xts(aes)' implementation > increased from ~225 MiB/s to ~235 MiB/s (measured using 'cryptsetup > benchmark -c aes-xts-plain64' on an Intel system with CRYPTO_AES_X86_64 > and CRYPTO_AES_NI_INTEL disabled). > > Signed-off-by: Ondrej Mosnacek <omosnacek@gmail.com> > Cc: Eric Biggers <ebiggers@google.com> > --- > v3 -> v4: a faster version of gf128mul_x_lle > v2 -> v3: constant-time implementation > v1 -> v2: move all _x_ functions to the header, not just gf128mul_x_ble > > crypto/gf128mul.c | 33 +--------------------------- > include/crypto/gf128mul.h | 55 +++++++++++++++++++++++++++++++++++++++++++++-- > 2 files changed, 54 insertions(+), 34 deletions(-) > > diff --git a/crypto/gf128mul.c b/crypto/gf128mul.c > index 04facc0..dc01212 100644 > --- a/crypto/gf128mul.c > +++ b/crypto/gf128mul.c > @@ -130,43 +130,12 @@ static const u16 gf128mul_table_le[256] = gf128mul_dat(xda_le); > static const u16 gf128mul_table_be[256] = gf128mul_dat(xda_be); > > /* > - * The following functions multiply a field element by x or by x^8 in > + * The following functions multiply a field element by x^8 in > * the polynomial field representation. They use 64-bit word operations > * to gain speed but compensate for machine endianness and hence work > * correctly on both styles of machine. > */ > > -static void gf128mul_x_lle(be128 *r, const be128 *x) > -{ > - u64 a = be64_to_cpu(x->a); > - u64 b = be64_to_cpu(x->b); > - u64 _tt = gf128mul_table_le[(b << 7) & 0xff]; > - > - r->b = cpu_to_be64((b >> 1) | (a << 63)); > - r->a = cpu_to_be64((a >> 1) ^ (_tt << 48)); > -} > - > -static void gf128mul_x_bbe(be128 *r, const be128 *x) > -{ > - u64 a = be64_to_cpu(x->a); > - u64 b = be64_to_cpu(x->b); > - u64 _tt = gf128mul_table_be[a >> 63]; > - > - r->a = cpu_to_be64((a << 1) | (b >> 63)); > - r->b = cpu_to_be64((b << 1) ^ _tt); > -} > - > -void gf128mul_x_ble(be128 *r, const be128 *x) > -{ > - u64 a = le64_to_cpu(x->a); > - u64 b = le64_to_cpu(x->b); > - u64 _tt = gf128mul_table_be[b >> 63]; > - > - r->a = cpu_to_le64((a << 1) ^ _tt); > - r->b = cpu_to_le64((b << 1) | (a >> 63)); > -} > -EXPORT_SYMBOL(gf128mul_x_ble); > - > static void gf128mul_x8_lle(be128 *x) > { > u64 a = be64_to_cpu(x->a); > diff --git a/include/crypto/gf128mul.h b/include/crypto/gf128mul.h > index 0bc9b5f..35ced9d 100644 > --- a/include/crypto/gf128mul.h > +++ b/include/crypto/gf128mul.h > @@ -49,6 +49,7 @@ > #ifndef _CRYPTO_GF128MUL_H > #define _CRYPTO_GF128MUL_H > > +#include <asm/byteorder.h> > #include <crypto/b128ops.h> > #include <linux/slab.h> > > @@ -163,8 +164,58 @@ void gf128mul_lle(be128 *a, const be128 *b); > > void gf128mul_bbe(be128 *a, const be128 *b); > > -/* multiply by x in ble format, needed by XTS */ > -void gf128mul_x_ble(be128 *a, const be128 *b); > +/* > + * The following functions multiply a field element by x in > + * the polynomial field representation. They use 64-bit word operations > + * to gain speed but compensate for machine endianness and hence work > + * correctly on both styles of machine. > + * > + * They are defined here for performance. > + */ > + > +static inline u64 gf128mul_mask_from_bit(u64 x, int which) > +{ > + /* a constant-time version of 'x & ((u64)1 << which) ? (u64)-1 : 0' */ > + return ((s64)(x << (63 - which)) >> 63); > +} > + > +static inline void gf128mul_x_lle(be128 *r, const be128 *x) > +{ > + u64 a = be64_to_cpu(x->a); > + u64 b = be64_to_cpu(x->b); > + > + /* equivalent to gf128mul_table_le[(b << 7) & 0xff] << 48 > + * (see crypto/gf128mul.c): */ > + u64 _tt = gf128mul_mask_from_bit(b, 0) & ((u64)0xe1 << 56); > + > + r->b = cpu_to_be64((b >> 1) | (a << 63)); > + r->a = cpu_to_be64((a >> 1) ^ _tt); > +} > + > +static inline void gf128mul_x_bbe(be128 *r, const be128 *x) > +{ > + u64 a = be64_to_cpu(x->a); > + u64 b = be64_to_cpu(x->b); > + > + /* equivalent to gf128mul_table_be[a >> 63] (see crypto/gf128mul.c): */ > + u64 _tt = gf128mul_mask_from_bit(a, 63) & 0x87; > + > + r->a = cpu_to_be64((a << 1) | (b >> 63)); > + r->b = cpu_to_be64((b << 1) ^ _tt); > +} > + > +/* needed by XTS */ > +static inline void gf128mul_x_ble(be128 *r, const be128 *x) > +{ > + u64 a = le64_to_cpu(x->a); > + u64 b = le64_to_cpu(x->b); > + > + /* equivalent to gf128mul_table_be[b >> 63] (see crypto/gf128mul.c): */ > + u64 _tt = gf128mul_mask_from_bit(b, 63) & 0x87; > + > + r->a = cpu_to_le64((a << 1) ^ _tt); > + r->b = cpu_to_le64((b << 1) | (a >> 63)); > +} > > /* 4k table optimization */ > > -- > 2.9.3 >
Never mind, Gmail is confusing me... there is indeed "v4" in the subject :) O.M. 2017-04-01 17:19 GMT+02:00 Ondrej Mosnáček <omosnacek@gmail.com>: > Oops, sorry, wrong prefix... > > 2017-04-01 17:17 GMT+02:00 Ondrej Mosnacek <omosnacek@gmail.com>: >> The gf128mul_x_ble function is currently defined in gf128mul.c, because >> it depends on the gf128mul_table_be multiplication table. >> >> However, since the function is very small and only uses two values from >> the table, it is better for it to be defined as inline function in >> gf128mul.h. That way, the function can be inlined by the compiler for >> better performance. >> >> For consistency, the other gf128mul_x_* functions are also moved to the >> header file. In addition, the code is rewritten to be constant-time. >> >> After this change, the speed of the generic 'xts(aes)' implementation >> increased from ~225 MiB/s to ~235 MiB/s (measured using 'cryptsetup >> benchmark -c aes-xts-plain64' on an Intel system with CRYPTO_AES_X86_64 >> and CRYPTO_AES_NI_INTEL disabled). >> >> Signed-off-by: Ondrej Mosnacek <omosnacek@gmail.com> >> Cc: Eric Biggers <ebiggers@google.com> >> --- >> v3 -> v4: a faster version of gf128mul_x_lle >> v2 -> v3: constant-time implementation >> v1 -> v2: move all _x_ functions to the header, not just gf128mul_x_ble >> >> crypto/gf128mul.c | 33 +--------------------------- >> include/crypto/gf128mul.h | 55 +++++++++++++++++++++++++++++++++++++++++++++-- >> 2 files changed, 54 insertions(+), 34 deletions(-) >> >> diff --git a/crypto/gf128mul.c b/crypto/gf128mul.c >> index 04facc0..dc01212 100644 >> --- a/crypto/gf128mul.c >> +++ b/crypto/gf128mul.c >> @@ -130,43 +130,12 @@ static const u16 gf128mul_table_le[256] = gf128mul_dat(xda_le); >> static const u16 gf128mul_table_be[256] = gf128mul_dat(xda_be); >> >> /* >> - * The following functions multiply a field element by x or by x^8 in >> + * The following functions multiply a field element by x^8 in >> * the polynomial field representation. They use 64-bit word operations >> * to gain speed but compensate for machine endianness and hence work >> * correctly on both styles of machine. >> */ >> >> -static void gf128mul_x_lle(be128 *r, const be128 *x) >> -{ >> - u64 a = be64_to_cpu(x->a); >> - u64 b = be64_to_cpu(x->b); >> - u64 _tt = gf128mul_table_le[(b << 7) & 0xff]; >> - >> - r->b = cpu_to_be64((b >> 1) | (a << 63)); >> - r->a = cpu_to_be64((a >> 1) ^ (_tt << 48)); >> -} >> - >> -static void gf128mul_x_bbe(be128 *r, const be128 *x) >> -{ >> - u64 a = be64_to_cpu(x->a); >> - u64 b = be64_to_cpu(x->b); >> - u64 _tt = gf128mul_table_be[a >> 63]; >> - >> - r->a = cpu_to_be64((a << 1) | (b >> 63)); >> - r->b = cpu_to_be64((b << 1) ^ _tt); >> -} >> - >> -void gf128mul_x_ble(be128 *r, const be128 *x) >> -{ >> - u64 a = le64_to_cpu(x->a); >> - u64 b = le64_to_cpu(x->b); >> - u64 _tt = gf128mul_table_be[b >> 63]; >> - >> - r->a = cpu_to_le64((a << 1) ^ _tt); >> - r->b = cpu_to_le64((b << 1) | (a >> 63)); >> -} >> -EXPORT_SYMBOL(gf128mul_x_ble); >> - >> static void gf128mul_x8_lle(be128 *x) >> { >> u64 a = be64_to_cpu(x->a); >> diff --git a/include/crypto/gf128mul.h b/include/crypto/gf128mul.h >> index 0bc9b5f..35ced9d 100644 >> --- a/include/crypto/gf128mul.h >> +++ b/include/crypto/gf128mul.h >> @@ -49,6 +49,7 @@ >> #ifndef _CRYPTO_GF128MUL_H >> #define _CRYPTO_GF128MUL_H >> >> +#include <asm/byteorder.h> >> #include <crypto/b128ops.h> >> #include <linux/slab.h> >> >> @@ -163,8 +164,58 @@ void gf128mul_lle(be128 *a, const be128 *b); >> >> void gf128mul_bbe(be128 *a, const be128 *b); >> >> -/* multiply by x in ble format, needed by XTS */ >> -void gf128mul_x_ble(be128 *a, const be128 *b); >> +/* >> + * The following functions multiply a field element by x in >> + * the polynomial field representation. They use 64-bit word operations >> + * to gain speed but compensate for machine endianness and hence work >> + * correctly on both styles of machine. >> + * >> + * They are defined here for performance. >> + */ >> + >> +static inline u64 gf128mul_mask_from_bit(u64 x, int which) >> +{ >> + /* a constant-time version of 'x & ((u64)1 << which) ? (u64)-1 : 0' */ >> + return ((s64)(x << (63 - which)) >> 63); >> +} >> + >> +static inline void gf128mul_x_lle(be128 *r, const be128 *x) >> +{ >> + u64 a = be64_to_cpu(x->a); >> + u64 b = be64_to_cpu(x->b); >> + >> + /* equivalent to gf128mul_table_le[(b << 7) & 0xff] << 48 >> + * (see crypto/gf128mul.c): */ >> + u64 _tt = gf128mul_mask_from_bit(b, 0) & ((u64)0xe1 << 56); >> + >> + r->b = cpu_to_be64((b >> 1) | (a << 63)); >> + r->a = cpu_to_be64((a >> 1) ^ _tt); >> +} >> + >> +static inline void gf128mul_x_bbe(be128 *r, const be128 *x) >> +{ >> + u64 a = be64_to_cpu(x->a); >> + u64 b = be64_to_cpu(x->b); >> + >> + /* equivalent to gf128mul_table_be[a >> 63] (see crypto/gf128mul.c): */ >> + u64 _tt = gf128mul_mask_from_bit(a, 63) & 0x87; >> + >> + r->a = cpu_to_be64((a << 1) | (b >> 63)); >> + r->b = cpu_to_be64((b << 1) ^ _tt); >> +} >> + >> +/* needed by XTS */ >> +static inline void gf128mul_x_ble(be128 *r, const be128 *x) >> +{ >> + u64 a = le64_to_cpu(x->a); >> + u64 b = le64_to_cpu(x->b); >> + >> + /* equivalent to gf128mul_table_be[b >> 63] (see crypto/gf128mul.c): */ >> + u64 _tt = gf128mul_mask_from_bit(b, 63) & 0x87; >> + >> + r->a = cpu_to_le64((a << 1) ^ _tt); >> + r->b = cpu_to_le64((b << 1) | (a >> 63)); >> +} >> >> /* 4k table optimization */ >> >> -- >> 2.9.3 >>
diff --git a/crypto/gf128mul.c b/crypto/gf128mul.c index 04facc0..dc01212 100644 --- a/crypto/gf128mul.c +++ b/crypto/gf128mul.c @@ -130,43 +130,12 @@ static const u16 gf128mul_table_le[256] = gf128mul_dat(xda_le); static const u16 gf128mul_table_be[256] = gf128mul_dat(xda_be); /* - * The following functions multiply a field element by x or by x^8 in + * The following functions multiply a field element by x^8 in * the polynomial field representation. They use 64-bit word operations * to gain speed but compensate for machine endianness and hence work * correctly on both styles of machine. */ -static void gf128mul_x_lle(be128 *r, const be128 *x) -{ - u64 a = be64_to_cpu(x->a); - u64 b = be64_to_cpu(x->b); - u64 _tt = gf128mul_table_le[(b << 7) & 0xff]; - - r->b = cpu_to_be64((b >> 1) | (a << 63)); - r->a = cpu_to_be64((a >> 1) ^ (_tt << 48)); -} - -static void gf128mul_x_bbe(be128 *r, const be128 *x) -{ - u64 a = be64_to_cpu(x->a); - u64 b = be64_to_cpu(x->b); - u64 _tt = gf128mul_table_be[a >> 63]; - - r->a = cpu_to_be64((a << 1) | (b >> 63)); - r->b = cpu_to_be64((b << 1) ^ _tt); -} - -void gf128mul_x_ble(be128 *r, const be128 *x) -{ - u64 a = le64_to_cpu(x->a); - u64 b = le64_to_cpu(x->b); - u64 _tt = gf128mul_table_be[b >> 63]; - - r->a = cpu_to_le64((a << 1) ^ _tt); - r->b = cpu_to_le64((b << 1) | (a >> 63)); -} -EXPORT_SYMBOL(gf128mul_x_ble); - static void gf128mul_x8_lle(be128 *x) { u64 a = be64_to_cpu(x->a); diff --git a/include/crypto/gf128mul.h b/include/crypto/gf128mul.h index 0bc9b5f..35ced9d 100644 --- a/include/crypto/gf128mul.h +++ b/include/crypto/gf128mul.h @@ -49,6 +49,7 @@ #ifndef _CRYPTO_GF128MUL_H #define _CRYPTO_GF128MUL_H +#include <asm/byteorder.h> #include <crypto/b128ops.h> #include <linux/slab.h> @@ -163,8 +164,58 @@ void gf128mul_lle(be128 *a, const be128 *b); void gf128mul_bbe(be128 *a, const be128 *b); -/* multiply by x in ble format, needed by XTS */ -void gf128mul_x_ble(be128 *a, const be128 *b); +/* + * The following functions multiply a field element by x in + * the polynomial field representation. They use 64-bit word operations + * to gain speed but compensate for machine endianness and hence work + * correctly on both styles of machine. + * + * They are defined here for performance. + */ + +static inline u64 gf128mul_mask_from_bit(u64 x, int which) +{ + /* a constant-time version of 'x & ((u64)1 << which) ? (u64)-1 : 0' */ + return ((s64)(x << (63 - which)) >> 63); +} + +static inline void gf128mul_x_lle(be128 *r, const be128 *x) +{ + u64 a = be64_to_cpu(x->a); + u64 b = be64_to_cpu(x->b); + + /* equivalent to gf128mul_table_le[(b << 7) & 0xff] << 48 + * (see crypto/gf128mul.c): */ + u64 _tt = gf128mul_mask_from_bit(b, 0) & ((u64)0xe1 << 56); + + r->b = cpu_to_be64((b >> 1) | (a << 63)); + r->a = cpu_to_be64((a >> 1) ^ _tt); +} + +static inline void gf128mul_x_bbe(be128 *r, const be128 *x) +{ + u64 a = be64_to_cpu(x->a); + u64 b = be64_to_cpu(x->b); + + /* equivalent to gf128mul_table_be[a >> 63] (see crypto/gf128mul.c): */ + u64 _tt = gf128mul_mask_from_bit(a, 63) & 0x87; + + r->a = cpu_to_be64((a << 1) | (b >> 63)); + r->b = cpu_to_be64((b << 1) ^ _tt); +} + +/* needed by XTS */ +static inline void gf128mul_x_ble(be128 *r, const be128 *x) +{ + u64 a = le64_to_cpu(x->a); + u64 b = le64_to_cpu(x->b); + + /* equivalent to gf128mul_table_be[b >> 63] (see crypto/gf128mul.c): */ + u64 _tt = gf128mul_mask_from_bit(b, 63) & 0x87; + + r->a = cpu_to_le64((a << 1) ^ _tt); + r->b = cpu_to_le64((b << 1) | (a >> 63)); +} /* 4k table optimization */
The gf128mul_x_ble function is currently defined in gf128mul.c, because it depends on the gf128mul_table_be multiplication table. However, since the function is very small and only uses two values from the table, it is better for it to be defined as inline function in gf128mul.h. That way, the function can be inlined by the compiler for better performance. For consistency, the other gf128mul_x_* functions are also moved to the header file. In addition, the code is rewritten to be constant-time. After this change, the speed of the generic 'xts(aes)' implementation increased from ~225 MiB/s to ~235 MiB/s (measured using 'cryptsetup benchmark -c aes-xts-plain64' on an Intel system with CRYPTO_AES_X86_64 and CRYPTO_AES_NI_INTEL disabled). Signed-off-by: Ondrej Mosnacek <omosnacek@gmail.com> Cc: Eric Biggers <ebiggers@google.com> --- v3 -> v4: a faster version of gf128mul_x_lle v2 -> v3: constant-time implementation v1 -> v2: move all _x_ functions to the header, not just gf128mul_x_ble crypto/gf128mul.c | 33 +--------------------------- include/crypto/gf128mul.h | 55 +++++++++++++++++++++++++++++++++++++++++++++-- 2 files changed, 54 insertions(+), 34 deletions(-)