diff --git a/CHANGELOG b/CHANGELOG --- a/CHANGELOG +++ b/CHANGELOG @@ -1,5 +1,17 @@ # Changelog +- 0.2.5 (2026-10-10) + * Further defends MAC comparison timing against "helpful" LLVM + optimizations. + * Detects support for the ARM SHA512 extensions at runtime, rather + than assuming it on all aarch64 builds, which crashed on aarch64 + CPUs lacking them (e.g. Graviton2, Raspberry Pi). + * The portable implementation now lives in an internal + 'Crypto.Hash.SHA512.Pure' module, and + 'Crypto.Hash.SHA512.Arm' is no longer exposed. Both are hidden + internal modules; the public API is unchanged. + * Improves 'hash_lazy' and 'hmac_lazy' performance. + - 0.2.4 (2026-07-06) * Reverts the unrolled MAC comparison introduced in 0.2.3, which was found to introduce timing variation on both aarch64 and x86-64 when diff --git a/cbits/sha512_arm.c b/cbits/sha512_arm.c --- a/cbits/sha512_arm.c +++ b/cbits/sha512_arm.c @@ -1,10 +1,19 @@ #include #include -#if defined(__aarch64__) && defined(__ARM_FEATURE_SHA512) +#if defined(__aarch64__) #include +#if defined(__APPLE__) +#include +#elif defined(__linux__) +#include +#ifndef HWCAP_SHA512 +#define HWCAP_SHA512 (1UL << 21) +#endif +#endif + static const uint64_t K[80] = { 0x428a2f98d728ae22ULL, 0x7137449123ef65cdULL, 0xb5c0fbcfec4d3b2fULL, 0xe9b5dba58189dbbcULL, @@ -55,7 +64,11 @@ * block: pointer to 16 uint64_t words (already native endian) * * The state is updated in place. + * + * Compiled for the SHA512 extension only; callers must first check + * 'sha512_arm_available'. */ +__attribute__((target("+sha3"))) void sha512_block_arm(uint64_t *state, const uint64_t *block) { /* Load current hash state */ uint64x2_t ab = vld1q_u64(&state[0]); @@ -446,14 +459,34 @@ vst1q_u64(&state[6], gh); } -/* Return 1 if ARM SHA512 is available, 0 otherwise */ +#if defined(__APPLE__) +static int sysctl_flag(const char *name) { + int val = 0; + size_t len = sizeof(val); + if (sysctlbyname(name, &val, &len, NULL, 0) != 0) return 0; + return val != 0; +} +#endif + +/* + * Return 1 if the CPU implements the ARM SHA512 extension, 0 + * otherwise. We only know how to ask Darwin and Linux, and report 0 + * (selecting the pure Haskell fallback) on any other OS. + */ int sha512_arm_available(void) { - return 1; +#if defined(__APPLE__) + return sysctl_flag("hw.optional.arm.FEAT_SHA512") + || sysctl_flag("hw.optional.armv8_2_sha512"); +#elif defined(__linux__) + return (getauxval(AT_HWCAP) & HWCAP_SHA512) != 0; +#else + return 0; +#endif } #else -/* Stub implementations when ARM SHA512 is not available */ +/* Stub implementations for non-aarch64 builds */ void sha512_block_arm(uint64_t *state, const uint64_t *block) { (void)state; (void)block; diff --git a/lib/Crypto/Hash/SHA512.hs b/lib/Crypto/Hash/SHA512.hs --- a/lib/Crypto/Hash/SHA512.hs +++ b/lib/Crypto/Hash/SHA512.hs @@ -1,9 +1,4 @@ {-# OPTIONS_HADDOCK prune #-} -{-# LANGUAGE BangPatterns #-} -{-# LANGUAGE MagicHash #-} -{-# LANGUAGE PatternSynonyms #-} -{-# LANGUAGE UnboxedTuples #-} -{-# LANGUAGE UnliftedNewtypes #-} -- | -- Module: Crypto.Hash.SHA512 @@ -36,20 +31,12 @@ ) where import qualified Data.ByteString as BS -import qualified Data.ByteString.Internal as BI -import qualified Data.ByteString.Unsafe as BU import Data.Word (Word8, Word64) import Foreign.Ptr (Ptr) -import qualified GHC.Exts as Exts import qualified Crypto.Hash.SHA512.Arm as Arm -import Crypto.Hash.SHA512.Internal +import Crypto.Hash.SHA512.Internal (MAC(..), Registers) import qualified Crypto.Hash.SHA512.Lazy as Lazy - --- utilities ------------------------------------------------------------------ - -fi :: (Integral a, Num b) => a -> b -fi = fromIntegral -{-# INLINE fi #-} +import qualified Crypto.Hash.SHA512.Pure as Pure -- hash ----------------------------------------------------------------------- @@ -63,39 +50,9 @@ hash :: BS.ByteString -> BS.ByteString hash m | Arm.sha512_arm_available = Arm.hash m - | otherwise = cat (_hash 0 (iv ()) m) + | otherwise = Pure.hash m {-# INLINABLE hash #-} -_hash - :: Word64 -- ^ extra prefix length for padding calculations - -> Registers -- ^ register state - -> BS.ByteString -- ^ input - -> Registers -_hash el rs m@(BI.PS _ _ l) = do - let !state = _hash_blocks rs m - !fin@(BI.PS _ _ ll) = BU.unsafeDrop (l - l `rem` 128) m - !total = el + fi l - if ll < 112 - then - let !ult = parse_pad1 fin total - in update state ult - else - let !(# pen, ult #) = parse_pad2 fin total - in update (update state pen) ult -{-# INLINABLE _hash #-} - -_hash_blocks - :: Registers -- ^ state - -> BS.ByteString -- ^ input - -> Registers -_hash_blocks rs m@(BI.PS _ _ l) = loop rs 0 where - loop !acc !j - | j + 128 > l = acc - | otherwise = - let !nacc = update acc (parse m j) - in loop nacc (j + 128) -{-# INLINABLE _hash_blocks #-} - -- hmac ---------------------------------------------------------------------- -- | Produce a message authentication code for a strict bytestring, @@ -111,26 +68,9 @@ hmac :: BS.ByteString -> BS.ByteString -> MAC hmac k m | Arm.sha512_arm_available = MAC (Arm.hmac k m) - | otherwise = MAC (cat (_hmac (prep_key k) m)) + | otherwise = MAC (Pure.hmac k m) {-# INLINABLE hmac #-} -prep_key :: BS.ByteString -> Block -prep_key k@(BI.PS _ _ l) - | l > 128 = parse_key (hash k) - | otherwise = parse_key k -{-# INLINABLE prep_key #-} - -_hmac - :: Block -- ^ padded key - -> BS.ByteString -- ^ message - -> Registers -_hmac k m = - let !rs0 = update (iv ()) (xor k (Exts.wordToWord64# 0x3636363636363636##)) - !block = pad_registers_with_length (_hash 128 rs0 m) - !rs1 = update (iv ()) (xor k (Exts.wordToWord64# 0x5C5C5C5C5C5C5C5C##)) - in update rs1 block -{-# INLINABLE _hmac #-} - -- the following functions are useful when we want to avoid allocating certain -- components of the HMAC key and message on the heap. @@ -145,25 +85,9 @@ -> IO () _hmac_rr rp bp k m | Arm.sha512_arm_available = Arm._hmac_rr rp bp k m - | otherwise = do - let !key = pad_registers k - !block = pad_registers_with_length m - !rs = _hmac_bb key block - poke_registers rp rs + | otherwise = Pure._hmac_rr rp bp k m {-# INLINABLE _hmac_rr #-} -_hmac_bb - :: Block -- ^ key - -> Block -- ^ message - -> Registers -_hmac_bb k m = - let !rs0 = update (iv ()) (xor k (Exts.wordToWord64# 0x3636363636363636##)) - !rs1 = update rs0 m - !inner = pad_registers_with_length rs1 - !rs2 = update (iv ()) (xor k (Exts.wordToWord64# 0x5C5C5C5C5C5C5C5C##)) - in update rs2 inner -{-# INLINABLE _hmac_bb #-} - -- Calculate hmac(k, m) where m is the concatenation of v (registers), a -- separator byte, and a ByteString. This avoids allocating 'v' on the -- heap. @@ -179,46 +103,5 @@ -> IO () _hmac_rsb rp bp k v sep dat | Arm.sha512_arm_available = Arm._hmac_rsb rp bp k v sep dat - | otherwise = do - let !key = pad_registers k - !rs0 = update (iv ()) (xor key (Exts.wordToWord64# 0x3636363636363636##)) - !inner = _hash_vsb 128 rs0 v sep dat - !block = pad_registers_with_length inner - !rs1 = update (iv ()) (xor key (Exts.wordToWord64# 0x5C5C5C5C5C5C5C5C##)) - !rs = update rs1 block - poke_registers rp rs + | otherwise = Pure._hmac_rsb rp bp k v sep dat {-# INLINABLE _hmac_rsb #-} - --- hash(v || sep || dat) with a custom initial state and extra --- prefix length. used for producing a more specialized hmac. -_hash_vsb - :: Word64 -- ^ extra prefix length - -> Registers -- ^ initial state - -> Registers -- ^ v - -> Word8 -- ^ sep - -> BS.ByteString -- ^ dat - -> Registers -_hash_vsb el rs0 v sep dat@(BI.PS _ _ l) - | l >= 63 = - -- first block is complete - let !b0 = parse_vsb v sep dat - !rs1 = update rs0 b0 - !rest = BU.unsafeDrop 63 dat - !rlen = l - 63 - !rs2 = _hash_blocks rs1 rest - !flen = rlen `rem` 128 - !fin = BU.unsafeDrop (rlen - flen) rest - !total = el + 65 + fi l - in if flen < 112 - then update rs2 (parse_pad1 fin total) - else let !(# pen, ult #) = parse_pad2 fin total - in update (update rs2 pen) ult - | otherwise = - -- message < 128 bytes, goes straight to padding - let !total = el + 65 + fi l - in if 65 + l < 112 - then update rs0 (parse_pad1_vsb v sep dat total) - else let !(# pen, ult #) = parse_pad2_vsb v sep dat total - in update (update rs0 pen) ult -{-# INLINABLE _hash_vsb #-} - diff --git a/lib/Crypto/Hash/SHA512/Arm.hs b/lib/Crypto/Hash/SHA512/Arm.hs --- a/lib/Crypto/Hash/SHA512/Arm.hs +++ b/lib/Crypto/Hash/SHA512/Arm.hs @@ -24,6 +24,7 @@ import qualified Data.ByteString.Internal as BI import qualified Data.ByteString.Unsafe as BU import Data.Word (Word8, Word64) +import Foreign.C.Types (CInt(..)) import Foreign.Marshal.Alloc (allocaBytes) import Foreign.Ptr (Ptr) import qualified GHC.Exts as Exts @@ -38,7 +39,7 @@ c_sha512_block :: Ptr Word64 -> Ptr Word64 -> IO () foreign import ccall unsafe "sha512_arm_available" - c_sha512_arm_available :: IO Int + c_sha512_arm_available :: IO CInt -- utilities ------------------------------------------------------------------ diff --git a/lib/Crypto/Hash/SHA512/Internal.hs b/lib/Crypto/Hash/SHA512/Internal.hs --- a/lib/Crypto/Hash/SHA512/Internal.hs +++ b/lib/Crypto/Hash/SHA512/Internal.hs @@ -50,6 +50,7 @@ , poke_registers ) where +import Data.Barrier (barrier) import qualified Data.Bits as B import qualified Data.ByteString as BS import qualified Data.ByteString.Internal as BI @@ -89,10 +90,13 @@ -- fused fold: OR the bytewise XORs into an accumulator -- directly, rather than via packZipWith, so no intermediate -- ByteString holding the (secret-derived) difference bytes - -- is ever materialised on the heap. + -- is ever materialised on the heap. The accumulator is routed + -- through 'barrier' before the zero-test so the LLVM backend + -- cannot recover the array-equality idiom and short-circuit on + -- the first mismatch (see "Data.Barrier"). go :: Word8 -> Int -> Bool go !acc !i - | i == la = acc == 0 + | i == la = barrier acc == 0 | otherwise = let !x = BU.unsafeIndex a i !y = BU.unsafeIndex b i diff --git a/lib/Crypto/Hash/SHA512/Lazy.hs b/lib/Crypto/Hash/SHA512/Lazy.hs --- a/lib/Crypto/Hash/SHA512/Lazy.hs +++ b/lib/Crypto/Hash/SHA512/Lazy.hs @@ -22,10 +22,10 @@ ) where import Crypto.Hash.SHA512.Internal +import qualified Crypto.Hash.SHA512.Pure as Pure import qualified Data.Bits as B import qualified Data.ByteString as BS import qualified Data.ByteString.Builder as BSB -import qualified Data.ByteString.Builder.Extra as BE import qualified Data.ByteString.Internal as BI import qualified Data.ByteString.Lazy as BL import qualified Data.ByteString.Lazy.Internal as BLI @@ -77,26 +77,20 @@ -- k such that (l + 1 + k) mod 128 = 112 sol :: Word64 -> Word64 -sol l = - let r = 112 - fi l `rem` 128 - 1 :: Integer -- fi prevents underflow - in fi (if r < 0 then r + 128 else r) +sol l = (111 - l) B..&. 127 -- wraps mod 2^64, a multiple of 128 -- RFC 6234 4.1 (lazy) pad_lazy :: BL.ByteString -> BL.ByteString pad_lazy (BL.toChunks -> m) = BL.fromChunks (walk 0 m) where walk !l bs = case bs of (c:cs) -> c : walk (l + fi (BS.length c)) cs - [] -> padding l (sol l) (BSB.word8 0x80) + [] -> [padding l] - padding l k bs - | k == 0 = - pure - . to_strict - -- more efficient for small builder - $ bs <> BSB.word64BE 0x00 <> BSB.word64BE (l * 8) - | otherwise = - let nacc = bs <> BSB.word8 0x00 - in padding l (pred k) nacc + padding l = to_strict $ + BSB.word8 0x80 + <> BSB.byteString (BS.replicate (fi (sol l)) 0x00) + <> BSB.word64BE 0x00 + <> BSB.word64BE (l * 8) -- | Compute a condensed representation of a lazy bytestring via -- SHA-512. @@ -137,37 +131,6 @@ step4 = hash_lazy step3 step5 = BS.map (B.xor 0x5C) step1 step6 = step5 <> step4 - in MAC (hash step6) + in MAC (Pure.hash step6) where - hash bs = cat (go (iv ()) (pad bs)) where - go :: Registers -> BS.ByteString -> Registers - go !acc b - | BS.null b = acc - | otherwise = case unsafe_splitAt 128 b of - SSPair c r -> go (update acc (parse c 0)) r - - pad m@(BI.PS _ _ (fi -> len)) - | len < 256 = to_strict_small padded - | otherwise = to_strict padded - where - padded = BSB.byteString m - <> fill (sol len) (BSB.word8 0x80) - <> BSB.word64BE 0x00 - <> BSB.word64BE (len * 8) - - to_strict_small = BL.toStrict . BE.toLazyByteStringWith - (BE.safeStrategy 256 BE.smallChunkSize) mempty - - fill j !acc - | j `rem` 8 == 0 = loop64 j acc - | otherwise = loop8 j acc - - loop64 j !acc - | j == 0 = acc - | otherwise = loop64 (j - 8) (acc <> BSB.word64BE 0x00) - - loop8 j !acc - | j == 0 = acc - | otherwise = loop8 (pred j) (acc <> BSB.word8 0x00) - - !(k, lk) = if l > 128 then (hash mk, 64) else (mk, l) + !(k, lk) = if l > 128 then (Pure.hash mk, 64) else (mk, l) diff --git a/lib/Crypto/Hash/SHA512/Pure.hs b/lib/Crypto/Hash/SHA512/Pure.hs new file mode 100644 --- /dev/null +++ b/lib/Crypto/Hash/SHA512/Pure.hs @@ -0,0 +1,189 @@ +{-# OPTIONS_HADDOCK hide #-} +{-# LANGUAGE BangPatterns #-} +{-# LANGUAGE MagicHash #-} +{-# LANGUAGE UnboxedTuples #-} + +-- | +-- Module: Crypto.Hash.SHA512.Pure +-- Copyright: (c) 2024 Jared Tobin +-- License: MIT +-- Maintainer: Jared Tobin +-- +-- Pure Haskell SHA-512 and HMAC-SHA512 implementations for strict +-- ByteStrings, used when the ARM cryptographic extensions are +-- unavailable. + +module Crypto.Hash.SHA512.Pure ( + hash + , hmac + , _hmac_rr + , _hmac_rsb + ) where + +import qualified Data.ByteString as BS +import qualified Data.ByteString.Internal as BI +import qualified Data.ByteString.Unsafe as BU +import Data.Word (Word8, Word64) +import Foreign.Ptr (Ptr) +import qualified GHC.Exts as Exts +import Crypto.Hash.SHA512.Internal + +-- utilities ------------------------------------------------------------------ + +fi :: (Integral a, Num b) => a -> b +fi = fromIntegral +{-# INLINE fi #-} + +-- hash ----------------------------------------------------------------------- + +-- | Compute a condensed representation of a strict bytestring via +-- SHA-512, using the pure Haskell implementation. +hash :: BS.ByteString -> BS.ByteString +hash m = cat (_hash 0 (iv ()) m) +{-# INLINABLE hash #-} + +_hash + :: Word64 -- ^ extra prefix length for padding calculations + -> Registers -- ^ register state + -> BS.ByteString -- ^ input + -> Registers +_hash el rs m@(BI.PS _ _ l) = do + let !state = _hash_blocks rs m + !fin@(BI.PS _ _ ll) = BU.unsafeDrop (l - l `rem` 128) m + !total = el + fi l + if ll < 112 + then + let !ult = parse_pad1 fin total + in update state ult + else + let !(# pen, ult #) = parse_pad2 fin total + in update (update state pen) ult +{-# INLINABLE _hash #-} + +_hash_blocks + :: Registers -- ^ state + -> BS.ByteString -- ^ input + -> Registers +_hash_blocks rs m@(BI.PS _ _ l) = loop rs 0 where + loop !acc !j + | j + 128 > l = acc + | otherwise = + let !nacc = update acc (parse m j) + in loop nacc (j + 128) +{-# INLINABLE _hash_blocks #-} + +-- hmac ---------------------------------------------------------------------- + +-- | Produce a message authentication code for a strict bytestring, +-- based on the provided (strict, bytestring) key, via HMAC-SHA512, +-- using the pure Haskell implementation. +hmac :: BS.ByteString -> BS.ByteString -> BS.ByteString +hmac k m = cat (_hmac (prep_key k) m) +{-# INLINABLE hmac #-} + +prep_key :: BS.ByteString -> Block +prep_key k@(BI.PS _ _ l) + | l > 128 = parse_key (hash k) + | otherwise = parse_key k +{-# INLINABLE prep_key #-} + +_hmac + :: Block -- ^ padded key + -> BS.ByteString -- ^ message + -> Registers +_hmac k m = + let !rs0 = update (iv ()) (xor k (Exts.wordToWord64# 0x3636363636363636##)) + !block = pad_registers_with_length (_hash 128 rs0 m) + !rs1 = update (iv ()) (xor k (Exts.wordToWord64# 0x5C5C5C5C5C5C5C5C##)) + in update rs1 block +{-# INLINABLE _hmac #-} + +-- the following functions are useful when we want to avoid allocating certain +-- components of the HMAC key and message on the heap. + +-- Computes hmac(k, v) when k and v are Registers. +-- +-- The 64-byte result is written to the destination pointer. +_hmac_rr + :: Ptr Word64 -- ^ destination (8 Word64s) + -> Ptr Word64 -- ^ scratch block buffer (unused) + -> Registers -- ^ key + -> Registers -- ^ message + -> IO () +_hmac_rr rp _ k m = do + let !key = pad_registers k + !block = pad_registers_with_length m + !rs = _hmac_bb key block + poke_registers rp rs +{-# INLINABLE _hmac_rr #-} + +_hmac_bb + :: Block -- ^ key + -> Block -- ^ message + -> Registers +_hmac_bb k m = + let !rs0 = update (iv ()) (xor k (Exts.wordToWord64# 0x3636363636363636##)) + !rs1 = update rs0 m + !inner = pad_registers_with_length rs1 + !rs2 = update (iv ()) (xor k (Exts.wordToWord64# 0x5C5C5C5C5C5C5C5C##)) + in update rs2 inner +{-# INLINABLE _hmac_bb #-} + +-- Calculate hmac(k, m) where m is the concatenation of v (registers), a +-- separator byte, and a ByteString. This avoids allocating 'v' on the +-- heap. +-- +-- The 64-byte result is written to the destination pointer. +_hmac_rsb + :: Ptr Word64 -- ^ destination pointer (8 x Word64) + -> Ptr Word64 -- ^ scratch block pointer (unused) + -> Registers -- ^ k + -> Registers -- ^ v + -> Word8 -- ^ separator byte + -> BS.ByteString -- ^ data + -> IO () +_hmac_rsb rp _ k v sep dat = do + let !key = pad_registers k + !rs0 = update (iv ()) + (xor key (Exts.wordToWord64# 0x3636363636363636##)) + !inner = _hash_vsb 128 rs0 v sep dat + !block = pad_registers_with_length inner + !rs1 = update (iv ()) + (xor key (Exts.wordToWord64# 0x5C5C5C5C5C5C5C5C##)) + !rs = update rs1 block + poke_registers rp rs +{-# INLINABLE _hmac_rsb #-} + +-- hash(v || sep || dat) with a custom initial state and extra +-- prefix length. used for producing a more specialized hmac. +_hash_vsb + :: Word64 -- ^ extra prefix length + -> Registers -- ^ initial state + -> Registers -- ^ v + -> Word8 -- ^ sep + -> BS.ByteString -- ^ dat + -> Registers +_hash_vsb el rs0 v sep dat@(BI.PS _ _ l) + | l >= 63 = + -- first block is complete + let !b0 = parse_vsb v sep dat + !rs1 = update rs0 b0 + !rest = BU.unsafeDrop 63 dat + !rlen = l - 63 + !rs2 = _hash_blocks rs1 rest + !flen = rlen `rem` 128 + !fin = BU.unsafeDrop (rlen - flen) rest + !total = el + 65 + fi l + in if flen < 112 + then update rs2 (parse_pad1 fin total) + else let !(# pen, ult #) = parse_pad2 fin total + in update (update rs2 pen) ult + | otherwise = + -- message < 128 bytes, goes straight to padding + let !total = el + 65 + fi l + in if 65 + l < 112 + then update rs0 (parse_pad1_vsb v sep dat total) + else let !(# pen, ult #) = parse_pad2_vsb v sep dat total + in update (update rs0 pen) ult +{-# INLINABLE _hash_vsb #-} + diff --git a/lib/Data/Barrier.hs b/lib/Data/Barrier.hs new file mode 100644 --- /dev/null +++ b/lib/Data/Barrier.hs @@ -0,0 +1,34 @@ +{-# OPTIONS_HADDOCK hide #-} + +-- | +-- Module: Data.Barrier +-- Copyright: (c) 2026 Jared Tobin +-- License: MIT +-- Maintainer: Jared Tobin +-- +-- An optimisation barrier for constant-time code. + +module Data.Barrier ( + barrier + ) where + +import Data.Word (Word8) + +-- | Identity on 'Word8', but opaque to the optimiser. A constant-time +-- accumulate-then-compare (e.g. an OR-fold of bytewise XORs, tested +-- against zero) routes its accumulator through this before the +-- zero-test, so the compiler cannot recognise it as an array-equality +-- test and lower it to a short-circuiting byte comparison (which would +-- leak the mismatch position). +-- +-- Both properties are required and must not be \"tidied\" away: +-- +-- * @NOINLINE@ -- if GHC inlines it, the LLVM backend regains the +-- accumulator's definition and short-circuits again. +-- * a /separate/ module -- a caller then compiles @barrier@ to an +-- external call it cannot see through. Inline it into the caller +-- and the barrier is gone. +-- +barrier :: Word8 -> Word8 +barrier x = x +{-# NOINLINE barrier #-} diff --git a/ppad-sha512.cabal b/ppad-sha512.cabal --- a/ppad-sha512.cabal +++ b/ppad-sha512.cabal @@ -1,6 +1,6 @@ cabal-version: 3.0 name: ppad-sha512 -version: 0.2.4 +version: 0.2.5 synopsis: The SHA-512 and HMAC-SHA512 algorithms license: MIT license-file: LICENSE @@ -37,16 +37,17 @@ ghc-options: -fllvm -O2 exposed-modules: Crypto.Hash.SHA512 - Crypto.Hash.SHA512.Arm Crypto.Hash.SHA512.Internal Crypto.Hash.SHA512.Lazy + Crypto.Hash.SHA512.Pure + other-modules: + Crypto.Hash.SHA512.Arm + Data.Barrier build-depends: base >= 4.9 && < 5 , bytestring >= 0.9 && < 0.13 c-sources: cbits/sha512_arm.c - if arch(aarch64) - cc-options: -march=armv8.2-a+sha3 if flag(sanitize) cc-options: -fsanitize=address,undefined -fno-omit-frame-pointer ghc-options: -optl=-fsanitize=address,undefined diff --git a/test/Property.hs b/test/Property.hs --- a/test/Property.hs +++ b/test/Property.hs @@ -1,13 +1,26 @@ {-# OPTIONS_GHC -fno-warn-orphans #-} +{-# LANGUAGE MagicHash #-} module Property ( properties ) where +import Data.Bits ((.|.)) +import qualified Data.Bits as B import qualified Data.ByteString as BS +import qualified Data.ByteString.Builder as BSB import qualified Data.ByteString.Lazy as BL +import Data.Word (Word8, Word64) +import Foreign.Marshal.Alloc (allocaBytes) +import Foreign.Marshal.Array (peekArray) +import Foreign.Ptr (Ptr) +import qualified GHC.Exts as Exts +import qualified GHC.Word import Crypto.Hash.SHA512 +import Crypto.Hash.SHA512.Internal (Registers(..)) +import qualified Crypto.Hash.SHA512.Pure as Pure import Test.Tasty +import qualified Test.Tasty.HUnit as H import qualified Test.Tasty.QuickCheck as Q import Test.QuickCheck.Instances.ByteString () @@ -41,6 +54,102 @@ lazy_chunked = BL.fromChunks (chunk_by (map Q.getPositive chunks) m) in hmac_lazy k lazy_single == hmac_lazy k lazy_chunked +-- dispatching vs. pure implementations ------------------------------------ +-- +-- on hosts with the ARM SHA512 extension, the public functions use it, so +-- these compare it against the pure Haskell implementation. + +-- messages spanning several blocks +newtype Msg = Msg BS.ByteString + deriving Show + +instance Q.Arbitrary Msg where + arbitrary = do + l <- Q.chooseInt (0, 8 * 128) + Msg . BS.pack <$> Q.vectorOf l Q.arbitrary + +-- register-sized strings +newtype Bytes64 = Bytes64 BS.ByteString + deriving Show + +instance Q.Arbitrary Bytes64 where + arbitrary = Bytes64 . BS.pack <$> Q.vectorOf 64 Q.arbitrary + +-- big-endian words of a 64-byte string, as registers +to_registers :: BS.ByteString -> Registers +to_registers bs = R (w 0) (w 1) (w 2) (w 3) (w 4) (w 5) (w 6) (w 7) where + w :: Int -> Exts.Word64# + w i = case word64 (BS.take 8 (BS.drop (8 * i) bs)) of + GHC.Word.W64# x -> x + word64 :: BS.ByteString -> Word64 + word64 = BS.foldl' (\acc b -> acc `B.shiftL` 8 .|. fromIntegral b) 0 + +-- run a destination-pointer HMAC, returning the big-endian digest +run_ptr :: (Ptr Word64 -> Ptr Word64 -> IO ()) -> IO BS.ByteString +run_ptr act = allocaBytes 64 $ \rp -> allocaBytes 128 $ \bp -> do + act rp bp + ws <- peekArray 8 rp + pure (BL.toStrict (BSB.toLazyByteString (foldMap BSB.word64BE ws))) + +unmac :: MAC -> BS.ByteString +unmac (MAC m) = m + +-- deterministic input of the given length +msg :: Int -> BS.ByteString +msg l = BS.pack (fmap fromIntegral (take l [(7 :: Int), 20 ..])) + +hash_matches_pure :: Msg -> Bool +hash_matches_pure (Msg m) = hash m == Pure.hash m + +hmac_matches_pure :: Msg -> Msg -> Bool +hmac_matches_pure (Msg k) (Msg m) = unmac (hmac k m) == Pure.hmac k m + +hmac_rr_matches_hmac :: Bytes64 -> Bytes64 -> Q.Property +hmac_rr_matches_hmac (Bytes64 k) (Bytes64 m) = Q.ioProperty $ do + let expected = unmac (hmac k m) + a <- run_ptr (\rp bp -> _hmac_rr rp bp (to_registers k) (to_registers m)) + b <- run_ptr (\rp bp -> + Pure._hmac_rr rp bp (to_registers k) (to_registers m)) + pure (a == expected && b == expected) + +hmac_rsb_matches_hmac :: Bytes64 -> Bytes64 -> Word8 -> Msg -> Q.Property +hmac_rsb_matches_hmac (Bytes64 k) (Bytes64 v) sep (Msg dat) = + Q.ioProperty (rsb_matches k v sep dat) + +rsb_matches + :: BS.ByteString -> BS.ByteString -> Word8 -> BS.ByteString -> IO Bool +rsb_matches k v sep dat = do + let expected = unmac (hmac k (v <> BS.singleton sep <> dat)) + a <- run_ptr (\rp bp -> + _hmac_rsb rp bp (to_registers k) (to_registers v) sep dat) + b <- run_ptr (\rp bp -> + Pure._hmac_rsb rp bp (to_registers k) (to_registers v) sep dat) + pure (a == expected && b == expected) + +-- every input length through three blocks, hitting each padding case +hash_all_lengths :: TestTree +hash_all_lengths = H.testCase "hash ~ Pure.hash (all lengths)" $ + H.assertBool mempty $ + all (\l -> hash (msg l) == Pure.hash (msg l)) [0 .. 3 * 128] + +hmac_all_lengths :: TestTree +hmac_all_lengths = H.testCase "hmac ~ Pure.hmac (all lengths)" $ + H.assertBool mempty $ + all (\l -> unmac (hmac (msg 64) (msg l)) == Pure.hmac (msg 64) (msg l)) + [0 .. 3 * 128] + +hmac_all_key_lengths :: TestTree +hmac_all_key_lengths = H.testCase "hmac ~ Pure.hmac (all key lengths)" $ + H.assertBool mempty $ + all (\l -> unmac (hmac (msg l) (msg 3)) == Pure.hmac (msg l) (msg 3)) + [0 .. 2 * 128 + 1] + +hmac_rsb_all_lengths :: TestTree +hmac_rsb_all_lengths = H.testCase "_hmac_rsb ~ hmac (all lengths)" $ do + oks <- traverse (rsb_matches (msg 64) (BS.reverse (msg 64)) 0x01 . msg) + [0 .. 3 * 128] + H.assertBool mempty (and oks) + chunk_by :: [Int] -> BS.ByteString -> [BS.ByteString] chunk_by _ bs | BS.null bs = [] chunk_by [] bs = [bs] @@ -62,4 +171,18 @@ Q.withMaxSuccess 1000 hash_chunking , Q.testProperty "hmac_lazy chunking-invariant" $ Q.withMaxSuccess 1000 hmac_chunking + , testGroup "dispatch ~ pure" [ + Q.testProperty "hash ~ Pure.hash" $ + Q.withMaxSuccess 1000 hash_matches_pure + , Q.testProperty "hmac ~ Pure.hmac" $ + Q.withMaxSuccess 1000 hmac_matches_pure + , Q.testProperty "_hmac_rr k m ~ hmac (cat k) (cat m)" $ + Q.withMaxSuccess 1000 hmac_rr_matches_hmac + , Q.testProperty "_hmac_rsb k v s d ~ hmac (cat k) (cat v <> s <> d)" $ + Q.withMaxSuccess 1000 hmac_rsb_matches_hmac + , hash_all_lengths + , hmac_all_lengths + , hmac_all_key_lengths + , hmac_rsb_all_lengths + ] ]