diff options
author | Shawn Landden <shawn@git.icu> | 2019-05-11 14:19:51 -0300 |
---|---|---|
committer | Shawn Landden <shawn@git.icu> | 2019-05-13 16:25:28 -0500 |
commit | cc6513fd7d0e049c0c0dae0c6ef62eb6ad11afe1 (patch) | |
tree | 8ea0fe17d01d9287d090a05305f1210cf64a555d /src/crypto/zinc/poly1305/poly1305.c | |
parent | [Zinc] Add PowerPC chacha20 implementation from openssl/cryptograms (diff) | |
download | wireguard-monolithic-historical-cc6513fd7d0e049c0c0dae0c6ef62eb6ad11afe1.tar.xz wireguard-monolithic-historical-cc6513fd7d0e049c0c0dae0c6ef62eb6ad11afe1.zip |
[zinc] Add PowerPC accelerated poly1305 from openssl/cryptogramssl/ppc
Unfortunately I am not seeing a speed up with this patch,
but it does decrease CPU usage.
Only (currently) runs on the outbound path, as the in-bound path is in
an interrupt, but that can be fixed in Linux.
v2: - Do not include the FPU version, as +10% performance on POWER8
(admittedly better on really old CPUs, like old world macs) is not
worth it, especially when there is a fast VSX version available.
- Honor CONFIG_VSX.
Signed-off-by: Shawn Landden <shawn@git.icu>
Diffstat (limited to 'src/crypto/zinc/poly1305/poly1305.c')
-rw-r--r-- | src/crypto/zinc/poly1305/poly1305.c | 69 |
1 files changed, 69 insertions, 0 deletions
diff --git a/src/crypto/zinc/poly1305/poly1305.c b/src/crypto/zinc/poly1305/poly1305.c index 7d373b9..dd2e1a3 100644 --- a/src/crypto/zinc/poly1305/poly1305.c +++ b/src/crypto/zinc/poly1305/poly1305.c @@ -16,12 +16,81 @@ #include <linux/module.h> #include <linux/init.h> +#if defined(CONFIG_ZINC_ARCH_ARM) || defined(CONFIG_ZINC_ARCH_ARM64) || defined(CONFIG_ZINC_ARCH_PPC32) || defined(CONFIG_ZINC_ARCH_PPC64) +#if defined(CONFIG_ZINC_ARCH_ARM64) || defined(CONFIG_ZINC_ARCH_PPC64) +struct poly1305_arch_internal { + union { + u32 h[5]; + struct { + u64 h0, h1, h2; + }; + }; + u64 is_base2_26; + u64 r[2]; +}; +#elif defined(CONFIG_ZINC_ARCH_ARM) || defined(CONFIG_ZINC_ARCH_PPC32) +struct poly1305_arch_internal { + union { + u32 h[5]; + struct { + u64 h0, h1; + u32 h2; + } __packed; + }; + u32 r[4]; + u32 is_base2_26; +}; +#endif +/* The NEON and AVX code uses base 2^26, while the scalar code uses base 2^64 on 64-bit + * and base 2^32 on 32-bit. If we hit the unfortunate situation of using NEON or AVX + * and then having to go back to scalar -- because the user is silly and has + * called the update function from two separate contexts -- then we need to + * convert back to the original base before proceeding. The below function is + * written for 64-bit integers, and so we have to swap words at the end on + * big-endian 32-bit. It is possible to reason that the initial reduction below + * is sufficient given the implementation invariants. However, for an avoidance + * of doubt and because this is not performance critical, we do the full + * reduction anyway. + */ +static void convert_to_base2_64(void *ctx) +{ + struct poly1305_arch_internal *state = ctx; + u32 cy; + + if (!(IS_ENABLED(CONFIG_KERNEL_MODE_NEON) || IS_ENABLED(CONFIG_AVX)) || !state->is_base2_26) + return; + + cy = state->h[0] >> 26; state->h[0] &= 0x3ffffff; state->h[1] += cy; + cy = state->h[1] >> 26; state->h[1] &= 0x3ffffff; state->h[2] += cy; + cy = state->h[2] >> 26; state->h[2] &= 0x3ffffff; state->h[3] += cy; + cy = state->h[3] >> 26; state->h[3] &= 0x3ffffff; state->h[4] += cy; + state->h0 = ((u64)state->h[2] << 52) | ((u64)state->h[1] << 26) | state->h[0]; + state->h1 = ((u64)state->h[4] << 40) | ((u64)state->h[3] << 14) | (state->h[2] >> 12); + state->h2 = state->h[4] >> 24; + if ((IS_ENABLED(CONFIG_ZINC_ARCH_ARM) || IS_ENABLED(CONFIG_ZINC_ARCH_PPC32)) && + IS_ENABLED(CONFIG_CPU_BIG_ENDIAN)) { + state->h0 = rol64(state->h0, 32); + state->h1 = rol64(state->h1, 32); + } +#define ULT(a, b) ((a ^ ((a ^ b) | ((a - b) ^ b))) >> (sizeof(a) * 8 - 1)) + cy = (state->h2 >> 2) + (state->h2 & ~3ULL); + state->h2 &= 3; + state->h0 += cy; + state->h1 += (cy = ULT(state->h0, cy)); + state->h2 += ULT(state->h1, cy); +#undef ULT + state->is_base2_26 = 0; +} +#endif + #if defined(CONFIG_ZINC_ARCH_X86_64) #include "poly1305-x86_64-glue.c" #elif defined(CONFIG_ZINC_ARCH_ARM) || defined(CONFIG_ZINC_ARCH_ARM64) #include "poly1305-arm-glue.c" #elif defined(CONFIG_ZINC_ARCH_MIPS) || defined(CONFIG_ZINC_ARCH_MIPS64) #include "poly1305-mips-glue.c" +#elif defined(CONFIG_ZINC_ARCH_PPC32) || defined(CONFIG_ZINC_ARCH_PPC64) +#include "poly1305-ppc-glue.c" #else static inline bool poly1305_init_arch(void *ctx, const u8 key[POLY1305_KEY_SIZE]) |