Pull crypto update from Herbert Xu:
"API:
- Add support for AEAD in simd
- Add fuzz testing to testmgr
- Add panic_on_fail module parameter to testmgr
- Use per-CPU struct instead multiple variables in scompress
- Change verify API for akcipher
Algorithms:
- Convert x86 AEAD algorithms over to simd
- Forbid 2-key 3DES in FIPS mode
- Add EC-RDSA (GOST 34.10) algorithm
Drivers:
- Set output IV with ctr-aes in crypto4xx
- Set output IV in rockchip
- Fix potential length overflow with hashing in sun4i-ss
- Fix computation error with ctr in vmx
- Add SM4 protected keys support in ccree
- Remove long-broken mxc-scc driver
- Add rfc4106(gcm(aes)) cipher support in cavium/nitrox"
* 'linus' of git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6: (179 commits)
crypto: ccree - use a proper le32 type for le32 val
crypto: ccree - remove set but not used variable 'du_size'
crypto: ccree - Make cc_sec_disable static
crypto: ccree - fix spelling mistake "protedcted" -> "protected"
crypto: caam/qi2 - generate hash keys in-place
crypto: caam/qi2 - fix DMA mapping of stack memory
crypto: caam/qi2 - fix zero-length buffer DMA mapping
crypto: stm32/cryp - update to return iv_out
crypto: stm32/cryp - remove request mutex protection
crypto: stm32/cryp - add weak key check for DES
crypto: atmel - remove set but not used variable 'alg_name'
crypto: picoxcell - Use dev_get_drvdata()
crypto: crypto4xx - get rid of redundant using_sd variable
crypto: crypto4xx - use sync skcipher for fallback
crypto: crypto4xx - fix cfb and ofb "overran dst buffer" issues
crypto: crypto4xx - fix ctr-aes missing output IV
crypto: ecrdsa - select ASN1 and OID_REGISTRY for EC-RDSA
crypto: ux500 - use ccflags-y instead of CFLAGS_<basename>.o
crypto: ccree - handle tee fips error during power management resume
crypto: ccree - add function to handle cryptocell tee fips error
...
The flags field in 'struct shash_desc' never actually does anything.
The only ostensibly supported flag is CRYPTO_TFM_REQ_MAY_SLEEP.
However, no shash algorithm ever sleeps, making this flag a no-op.
With this being the case, inevitably some users who can't sleep wrongly
pass MAY_SLEEP. These would all need to be fixed if any shash algorithm
actually started sleeping. For example, the shash_ahash_*() functions,
which wrap a shash algorithm with the ahash API, pass through MAY_SLEEP
from the ahash API to the shash API. However, the shash functions are
called under kmap_atomic(), so actually they're assumed to never sleep.
Even if it turns out that some users do need preemption points while
hashing large buffers, we could easily provide a helper function
crypto_shash_update_large() which divides the data into smaller chunks
and calls crypto_shash_update() and cond_resched() for each chunk. It's
not necessary to have a flag in 'struct shash_desc', nor is it necessary
to make individual shash algorithms aware of this at all.
Therefore, remove shash_desc::flags, and document that the
crypto_shash_*() functions can be called from any context.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The x86_64 implementation of Poly1305 produces the wrong result on some
inputs because poly1305_4block_avx2() incorrectly assumes that when
partially reducing the accumulator, the bits carried from limb 'd4' to
limb 'h0' fit in a 32-bit integer. This is true for poly1305-generic
which processes only one block at a time. However, it's not true for
the AVX2 implementation, which processes 4 blocks at a time and
therefore can produce intermediate limbs about 4x larger.
Fix it by making the relevant calculations use 64-bit arithmetic rather
than 32-bit. Note that most of the carries already used 64-bit
arithmetic, but the d4 -> h0 carry was different for some reason.
To be safe I also made the same change to the corresponding SSE2 code,
though that only operates on 1 or 2 blocks at a time. I don't think
it's really needed for poly1305_block_sse2(), but it doesn't hurt
because it's already x86_64 code. It *might* be needed for
poly1305_2block_sse2(), but overflows aren't easy to reproduce there.
This bug was originally detected by my patches that improve testmgr to
fuzz algorithms against their generic implementation. But also add a
test vector which reproduces it directly (in the AVX2 case).
Fixes: b1ccc8f4b6 ("crypto: poly1305 - Add a four block AVX2 variant for x86_64")
Fixes: c70f4abef0 ("crypto: poly1305 - Add a SSE2 SIMD variant for x86_64")
Cc: <stable@vger.kernel.org> # v4.3+
Cc: Martin Willi <martin@strongswan.org>
Cc: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The ->digest() method of crct10dif-pclmul reads the current CRC value
from the shash_desc context. But this value is uninitialized, causing
crypto_shash_digest() to compute the wrong result. Fix it.
Probably this wasn't noticed before because lib/crc-t10dif.c only uses
crypto_shash_update(), not crypto_shash_digest(). Likewise,
crypto_shash_digest() is not yet tested by the crypto self-tests because
those only test the ahash API which only uses shash init/update/final.
Fixes: 0b95a7f857 ("crypto: crct10dif - Glue code to cast accelerated CRCT10DIF assembly as a crypto transform")
Cc: <stable@vger.kernel.org> # v3.11+
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Replace all calls to irq_fpu_usable() in the x86 crypto code with
crypto_simd_usable(), in order to allow testing the no-SIMD code paths.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Convert the x86 implementations of MORUS-1280 to use the AEAD SIMD
helpers, rather than hand-rolling the same functionality. This
simplifies the code and also fixes the bug where the user-provided
aead_request is modified.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Convert the x86 implementation of MORUS-640 to use the AEAD SIMD
helpers, rather than hand-rolling the same functionality. This
simplifies the code and also fixes the bug where the user-provided
aead_request is modified.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Convert the x86 implementation of AEGIS-256 to use the AEAD SIMD
helpers, rather than hand-rolling the same functionality. This
simplifies the code and also fixes the bug where the user-provided
aead_request is modified.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Convert the x86 implementation of AEGIS-128L to use the AEAD SIMD
helpers, rather than hand-rolling the same functionality. This
simplifies the code and also fixes the bug where the user-provided
aead_request is modified.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Convert the x86 implementation of AEGIS-128 to use the AEAD SIMD
helpers, rather than hand-rolling the same functionality. This
simplifies the code and also fixes the bug where the user-provided
aead_request is modified.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Convert the AES-NI implementations of "gcm(aes)" and "rfc4106(gcm(aes))"
to use the AEAD SIMD helpers, rather than hand-rolling the same
functionality. This simplifies the code and also fixes the bug where
the user-provided aead_request is modified.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Convert the AES-NI glue code to use simd_register_skciphers_compat() to
create SIMD wrappers for all the internal skcipher algorithms at once,
rather than wrapping each one individually. This simplifies the code.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
1-block SSE2 variant of poly1305 stores variables s1..s4 containing key
material on the stack. This commit adds missing zeroing of the stack
memory. Benchmarks show negligible performance hit (tested on i7-3770).
Signed-off-by: Tommi Hirvola <tommi@hirvola.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
gcmaes_crypt_by_sg() dereferences the NULL pointer returned by
scatterwalk_ffwd() when encrypting an empty plaintext and the source
scatterlist ends immediately after the associated data.
Fix it by only fast-forwarding to the src/dst data scatterlists if the
data length is nonzero.
This bug is reproduced by the "rfc4543(gcm(aes))" test vectors when run
with the new AEAD test manager.
Fixes: e845520707 ("crypto: aesni - Update aesni-intel_glue to use scatter/gather")
Cc: <stable@vger.kernel.org> # v4.17+
Cc: Dave Watson <davejwatson@fb.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The x86 MORUS implementations all fail the improved AEAD tests because
they produce the wrong result with some data layouts. The issue is that
they assume that if the skcipher_walk API gives 'nbytes' not aligned to
the walksize (a.k.a. walk.stride), then it is the end of the data. In
fact, this can happen before the end.
Also, when the CRYPTO_TFM_REQ_MAY_SLEEP flag is given, they can
incorrectly sleep in the skcipher_walk_*() functions while preemption
has been disabled by kernel_fpu_begin().
Fix these bugs.
Fixes: 56e8e57fc3 ("crypto: morus - Add common SIMD glue code for MORUS")
Cc: <stable@vger.kernel.org> # v4.18+
Cc: Ondrej Mosnacek <omosnace@redhat.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Ondrej Mosnacek <omosnace@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The x86 AEGIS implementations all fail the improved AEAD tests because
they produce the wrong result with some data layouts. The issue is that
they assume that if the skcipher_walk API gives 'nbytes' not aligned to
the walksize (a.k.a. walk.stride), then it is the end of the data. In
fact, this can happen before the end.
Also, when the CRYPTO_TFM_REQ_MAY_SLEEP flag is given, they can
incorrectly sleep in the skcipher_walk_*() functions while preemption
has been disabled by kernel_fpu_begin().
Fix these bugs.
Fixes: 1d373d4e8e ("crypto: x86 - Add optimized AEGIS implementations")
Cc: <stable@vger.kernel.org> # v4.18+
Cc: Ondrej Mosnacek <omosnace@redhat.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Ondrej Mosnacek <omosnace@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The x86, arm, and arm64 asm implementations of crct10dif are very
difficult to understand partly because many of the comments, labels, and
macros are named incorrectly: the lengths mentioned are usually off by a
factor of two from the actual code. Many other things are unnecessarily
convoluted as well, e.g. there are many more fold constants than
actually needed and some aren't fully reduced.
This series therefore cleans up all these implementations to be much
more maintainable. I also made some small optimizations where I saw
opportunities, resulting in slightly better performance.
This patch cleans up the x86 version.
As part of this, I removed support for len < 16 from the x86 assembly;
now the glue code falls back to the generic table-based implementation
in this case. Due to the overhead of kernel_fpu_begin(), this actually
significantly improves performance on these lengths. (And even if
kernel_fpu_begin() were free, the generic code is still faster for about
len < 11.) This removal also eliminates error-prone special cases and
makes the x86, arm32, and arm64 ports of the code match more closely.
Acked-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add missing static keywords to fix the following sparse warnings:
arch/x86/crypto/aesni-intel_glue.c:197:24: warning: symbol 'aesni_gcm_tfm_sse' was not declared. Should it be static?
arch/x86/crypto/aesni-intel_glue.c:246:24: warning: symbol 'aesni_gcm_tfm_avx_gen2' was not declared. Should it be static?
arch/x86/crypto/aesni-intel_glue.c:291:24: warning: symbol 'aesni_gcm_tfm_avx_gen4' was not declared. Should it be static?
I also made the affected structures 'const', and adjusted the
indentation in the struct definition to not be insane.
Cc: Dave Watson <davejwatson@fb.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Pull crypto updates from Herbert Xu:
"API:
- Add 1472-byte test to tcrypt for IPsec
- Reintroduced crypto stats interface with numerous changes
- Support incremental algorithm dumps
Algorithms:
- Add xchacha12/20
- Add nhpoly1305
- Add adiantum
- Add streebog hash
- Mark cts(cbc(aes)) as FIPS allowed
Drivers:
- Improve performance of arm64/chacha20
- Improve performance of x86/chacha20
- Add NEON-accelerated nhpoly1305
- Add SSE2 accelerated nhpoly1305
- Add AVX2 accelerated nhpoly1305
- Add support for 192/256-bit keys in gcmaes AVX
- Add SG support in gcmaes AVX
- ESN for inline IPsec tx in chcr
- Add support for CryptoCell 703 in ccree
- Add support for CryptoCell 713 in ccree
- Add SM4 support in ccree
- Add SM3 support in ccree
- Add support for chacha20 in caam/qi2
- Add support for chacha20 + poly1305 in caam/jr
- Add support for chacha20 + poly1305 in caam/qi2
- Add AEAD cipher support in cavium/nitrox"
* 'linus' of git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6: (130 commits)
crypto: skcipher - remove remnants of internal IV generators
crypto: cavium/nitrox - Fix build with !CONFIG_DEBUG_FS
crypto: salsa20-generic - don't unnecessarily use atomic walk
crypto: skcipher - add might_sleep() to skcipher_walk_virt()
crypto: x86/chacha - avoid sleeping under kernel_fpu_begin()
crypto: cavium/nitrox - Added AEAD cipher support
crypto: mxc-scc - fix build warnings on ARM64
crypto: api - document missing stats member
crypto: user - remove unused dump functions
crypto: chelsio - Fix wrong error counter increments
crypto: chelsio - Reset counters on cxgb4 Detach
crypto: chelsio - Handle PCI shutdown event
crypto: chelsio - cleanup:send addr as value in function argument
crypto: chelsio - Use same value for both channel in single WR
crypto: chelsio - Swap location of AAD and IV sent in WR
crypto: chelsio - remove set but not used variable 'kctx_len'
crypto: ux500 - Use proper enum in hash_set_dma_transfer
crypto: ux500 - Use proper enum in cryp_set_dma_transfer
crypto: aesni - Add scatter/gather avx stubs, and use them in C
crypto: aesni - Introduce partial block macro
..
Passing atomic=true to skcipher_walk_virt() only makes the later
skcipher_walk_done() calls use atomic memory allocations, not
skcipher_walk_virt() itself. Thus, we have to move it outside of the
preemption-disabled region (kernel_fpu_begin()/kernel_fpu_end()).
(skcipher_walk_virt() only allocates memory for certain layouts of the
input scatterlist, hence why I didn't notice this earlier...)
Reported-by: syzbot+9bf843c33f782d73ae7d@syzkaller.appspotmail.com
Fixes: 4af7826187 ("crypto: x86/chacha20 - add XChaCha20 support")
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add the appropriate scatter/gather stubs to the avx asm.
In the C code, we can now always use crypt_by_sg, since both
sse and asm code now support scatter/gather.
Introduce a new struct, aesni_gcm_tfm, that is initialized on
startup to point to either the SSE, AVX, or AVX2 versions of the
four necessary encryption/decryption routines.
GENX_OPTSIZE is still checked at the start of crypt_by_sg. The
total size of the data is checked, since the additional overhead
is in the init function, calculating additional HashKeys.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Before this diff, multiple calls to GCM_ENC_DEC will
succeed, but only if all calls are a multiple of 16 bytes.
Handle partial blocks at the start of GCM_ENC_DEC, and update
aadhash as appropriate.
The data offset %r11 is also updated after the partial block.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Introduce READ_PARTIAL_BLOCK macro, and use it in the two existing
partial block cases: AAD and the end of ENC_DEC. In particular,
the ENC_DEC case should be faster, since we read by 8/4 bytes if
possible.
This macro will also be used to read partial blocks between
enc_update and dec_update calls.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Prepare to handle partial blocks between scatter/gather calls.
For the last partial block, we only want to calculate the aadhash
in GCM_COMPLETE, and a new partial block macro will handle both
aadhash update and encrypting partial blocks between calls.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Fill in aadhash, aadlen, pblocklen, curcount with appropriate values.
pblocklen, aadhash, and pblockenckey are also updated at the end
of each scatter/gather operation, to be carried over to the next
operation.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The precompute functions differ only by the sub-macros
they call, merge them to a single macro. Later diffs
add more code to fill in the gcm_context_data structure,
this allows changes in a single place.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
AAD hash only needs to be calculated once for each scatter/gather operation.
Move it to its own macro, and call it from GCM_INIT instead of
INITIAL_BLOCKS.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Merge encode and decode tag calculations in GCM_COMPLETE macro.
Scatter/gather routines will call this once at the end of encryption
or decryption.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add support for 192/256-bit keys using the avx gcm/aes routines.
The sse routines were previously updated in e31ac32d3b (Add support
for 192 & 256 bit keys to AESNI RFC4106).
Instead of adding an additional loop in the hotpath as in e31ac32d3b,
this diff instead generates separate versions of the code using macros,
and the entry routines choose which version once. This results
in a 5% performance improvement vs. adding a loop to the hot path.
This is the same strategy chosen by the intel isa-l_crypto library.
The key size checks are removed from the c code where appropriate.
Note that this diff depends on using gcm_context_data - 256 bit keys
require 16 HashKeys + 15 expanded keys, which is larger than
struct crypto_aes_ctx, so they are stored in struct gcm_context_data.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Macro-ify function save and restore. These will be used in new functions
added for scatter/gather update operations.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add the gcm_context_data structure to the avx asm routines.
This will be necessary to support both 256 bit keys and
scatter/gather.
The pre-computed HashKeys are now stored in the gcm_context_data
struct, which is expanded to hold the greater number of hashkeys
necessary for avx.
Loads and stores to the new struct are always done unlaligned to
avoid compiler issues, see e5b954e8 "Use unaligned loads from
gcm_context_data"
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The GCM_ENC_DEC routines for AVX and AVX2 are identical, except they
call separate sub-macros. Pass the macros as arguments, and merge them.
This facilitates additional refactoring, by requiring changes in only
one place.
The GCM_ENC_DEC macro was moved above the CONFIG_AS_AVX* ifdefs,
since it will be used by both AVX and AVX2.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
To improve responsiveness, yield the FPU (temporarily re-enabling
preemption) every 4 KiB encrypted/decrypted, rather than keeping
preemption disabled during the entire encryption/decryption operation.
Alternatively we could do this for every skcipher_walk step, but steps
may be small in some cases, and yielding the FPU is expensive on x86.
Suggested-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Now that the x86_64 SIMD implementations of ChaCha20 and XChaCha20 have
been refactored to support varying the number of rounds, add support for
XChaCha12. This is identical to XChaCha20 except for the number of
rounds, which is 12 instead of 20. This can be used by Adiantum.
Reviewed-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
In preparation for adding XChaCha12 support, rename/refactor the x86_64
SIMD implementations of ChaCha20 to support different numbers of rounds.
Reviewed-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add an XChaCha20 implementation that is hooked up to the x86_64 SIMD
implementations of ChaCha20. This can be used by Adiantum.
An SSSE3 implementation of single-block HChaCha20 is also added so that
XChaCha20 can use it rather than the generic implementation. This
required refactoring the ChaCha permutation into its own function.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add a 64-bit AVX2 implementation of NHPoly1305, an ε-almost-∆-universal
hash function used in the Adiantum encryption mode. For now, only the
NH portion is actually AVX2-accelerated; the Poly1305 part is less
performance-critical so is just implemented in C.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add a 64-bit SSE2 implementation of NHPoly1305, an ε-almost-∆-universal
hash function used in the Adiantum encryption mode. For now, only the
NH portion is actually SSE2-accelerated; the Poly1305 part is less
performance-critical so is just implemented in C.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Go over arch/x86/ and fix common typos in comments,
and a typo in an actual function argument name.
No change in functionality intended.
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This version uses the same principle as the AVX2 version by scheduling the
operations for two block pairs in parallel. It benefits from the AVX-512VL
rotate instructions and the more efficient partial block handling using
"vmovdqu8", resulting in a speedup of the raw block function of ~20%.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This version uses the same principle as the AVX2 version. It benefits
from the AVX-512VL rotate instructions and the more efficient partial
block handling using "vmovdqu8", resulting in a speedup of ~20%.
Unlike the AVX2 version, it is faster than the single block SSSE3 version
to process a single block. Hence we engage that function for (partial)
single block lengths as well.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This variant is similar to the AVX2 version, but benefits from the AVX-512
rotate instructions and the additional registers, so it can operate without
any data on the stack. It uses ymm registers only to avoid the massive core
throttling on Skylake-X platforms. Nontheless does it bring a ~30% speed
improvement compared to the AVX2 variant for random encryption lengths.
The AVX2 version uses "rep movsb" for partial block XORing via the stack.
With AVX-512, the new "vmovdqu8" can do this much more efficiently. The
associated "kmov" instructions to work with dynamic masks is not part of
the AVX-512VL instruction set, hence we depend on AVX-512BW as well. Given
that the major AVX-512VL architectures provide AVX-512BW and this extension
does not affect core clocking, this seems to be no problem at least for
now.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
In preparation for exposing a low-level Poly1305 API which implements
the ε-almost-∆-universal (εA∆U) hash function underlying the Poly1305
MAC and supports block-aligned inputs only, create structures
poly1305_key and poly1305_state which hold the limbs of the Poly1305
"r" key and accumulator, respectively.
These structures could actually have the same type (e.g. poly1305_val),
but different types are preferable, to prevent misuse.
Acked-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Acked-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
In preparation for adding XChaCha12 support, rename/refactor
chacha20-generic to support different numbers of rounds. The
justification for needing XChaCha12 support is explained in more detail
in the patch "crypto: chacha - add XChaCha12 support".
The only difference between ChaCha{8,12,20} are the number of rounds
itself; all other parts of the algorithm are the same. Therefore,
remove the "20" from all definitions, structures, functions, files, etc.
that will be shared by all ChaCha versions.
Also make ->setkey() store the round count in the chacha_ctx (previously
chacha20_ctx). The generic code then passes the round count through to
chacha_block(). There will be a ->setkey() function for each explicitly
allowed round count; the encrypt/decrypt functions will be the same. I
decided not to do it the opposite way (same ->setkey() function for all
round counts, with different encrypt/decrypt functions) because that
would have required more boilerplate code in architecture-specific
implementations of ChaCha and XChaCha.
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This variant builds upon the idea of the 2-block AVX2 variant that
shuffles words after each round. The shuffling has a rather high latency,
so the arithmetic units are not optimally used.
Given that we have plenty of registers in AVX, this version parallelizes
the 2-block variant to do four blocks. While the first two blocks are
shuffling, the CPU can do the XORing on the second two blocks and
vice-versa, which makes this version much faster than the SSSE3 variant
for four blocks. The latter is now mostly for systems that do not have
AVX2, but there it is the work-horse, so we keep it in place.
The partial XORing function trailer is very similar to the AVX2 2-block
variant. While it could be shared, that code segment is rather short;
profiling is also easier with the trailer integrated, so we keep it per
function.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This variant uses the same principle as the single block SSSE3 variant
by shuffling the state matrix after each round. With the wider AVX
registers, we can do two blocks in parallel, though.
This function can increase performance and efficiency significantly for
lengths that would otherwise require a 4-block function.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Now that all block functions support partial lengths, engage the wider
block sizes more aggressively. This prevents using smaller block
functions multiple times, where the next larger block function would
have been faster.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add a length argument to the eight block function for AVX2, so the
block function may XOR only a partial length of eight blocks.
To avoid unnecessary operations, we integrate XORing of the first four
blocks in the final lane interleaving; this also avoids some work in
the partial lengths path.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add a length argument to the quad block function for SSSE3, so the
block function may XOR only a partial length of four blocks.
As we already have the stack set up, the partial XORing does not need
to. This gives a slightly different function trailer, so we keep that
separate from the 1-block function.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add a length argument to the single block function for SSSE3, so the
block function may XOR only a partial length of the full block. Given
that the setup code is rather cheap, the function does not process more
than one block; this allows us to keep the block function selection in
the C glue code.
The required branching does not negatively affect performance for full
block sizes. The partial XORing uses simple "rep movsb" to copy the
data before and after doing XOR in SSE. This is rather efficient on
modern processors; movsw can be slightly faster, but the additional
complexity is probably not worth it.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
