| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Instead of yielding from the bowels of the asm routine if a reschedule
is needed, divide up the input into 4 KB chunks in the C glue. This
simplifies the code substantially, and avoids scheduling out the task
with the asm routine on the call stack, which is undesirable from a
CFI/instrumentation point of view.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
There is no need for elaborate yield handling in the bit-sliced NEON
implementation of AES, given that skciphers are naturally bounded by the
size of the chunks returned by the skcipher_walk API. So remove the
yield calls from the asm code.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Instead of calling into kernel_neon_end() and kernel_neon_begin() (and
potentially into schedule()) from the assembler code when running in
task mode and a reschedule is pending, perform only the preempt count
check in assembler, but simply return early in this case, and let the C
code deal with the consequences.
This reverts commit 6caf7adc5e458f77f550b6c6ca8effa152d61b4a.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Instead of calling into kernel_neon_end() and kernel_neon_begin() (and
potentially into schedule()) from the assembler code when running in
task mode and a reschedule is pending, perform only the preempt count
check in assembler, but simply return early in this case, and let the C
code deal with the consequences.
This reverts commit 7edc86cb1c18b4c274672232117586ea2bef1d9a.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Instead of calling into kernel_neon_end() and kernel_neon_begin() (and
potentially into schedule()) from the assembler code when running in
task mode and a reschedule is pending, perform only the preempt count
check in assembler, but simply return early in this case, and let the C
code deal with the consequences.
This reverts commit d82f37ab5e2426287013eba38b1212e8b71e5be3.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Instead of calling into kernel_neon_end() and kernel_neon_begin() (and
potentially into schedule()) from the assembler code when running in
task mode and a reschedule is pending, perform only the preempt count
check in assembler, but simply return early in this case, and let the C
code deal with the consequences.
This reverts commit 7df8d164753e6e6f229b72767595072bc6a71f48.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Eliminate the following coccicheck warning:
./arch/powerpc/crypto/sha256-spe-glue.c:132:2-3: Unneeded
semicolon
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Pull change from arm64 tree that's needed for crypto arm changes.
|
|
Add a macro cond_yield that branches to a specified label when called if
the TIF_NEED_RESCHED flag is set and decreasing the preempt count would
make the task preemptible again, resulting in a schedule to occur. This
can be used by kernel mode SIMD code that keeps a lot of state in SIMD
registers, which would make chunking the input in order to perform the
cond_resched() check from C code disproportionately costly.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20210203113626.220151-2-ardb@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>
|
|
Taking ownership of the FPU in kernel mode disables preemption, and
this may result in excessive scheduling blackouts if the size of the
data being processed on the FPU is unbounded.
Given that taking and releasing the FPU is cheap these days on x86, we
can limit the impact of this issue easily for skcipher implementations,
by moving the FPU begin/end calls inside the skcipher walk processing
loop. Considering that skcipher walks operate on at most one page at a
time, doing so fully mitigates this issue.
This also permits the skcipher walk logic to use non-atomic kmalloc()
calls etc so we can change the 'atomic' bool argument in the calls to
skcipher_walk_virt() to false as well.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Indirect calls are very expensive on x86, so use a static call to set
the system-wide AES-NI/CTR asm helper.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The accelerated, instruction based implementations of SHA1, SHA2 and
SHA3 are autoloaded based on CPU capabilities, given that the code is
modest in size, and widely used, which means that resolving the algo
name, loading all compatible modules and picking the one with the
highest priority is taken to be suboptimal.
However, if these algorithms are requested before this CPU feature
based matching and autoloading occurs, these modules are not even
considered, and we end up with suboptimal performance.
So add the missing module aliases for the various SHA implementations.
Cc: <stable@vger.kernel.org>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The Camellia, Serpent and Twofish related header files only contain
declarations that are shared between different implementations of the
respective algorithms residing under arch/x86/crypto, and none of their
contents should be used elsewhere. So move the header files into the
same location, and use local #includes instead.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
All dependencies on the x86 glue helper module have been replaced by
local instantiations of the new ECB/CBC preprocessor helper macros, so
the glue helper module can be retired.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Replace the glue helper dependency with implementations of ECB and CBC
based on the new CPP macros, which avoid the need for indirect calls.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Replace the glue helper dependency with implementations of ECB and CBC
based on the new CPP macros, which avoid the need for indirect calls.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Replace the glue helper dependency with implementations of ECB and CBC
based on the new CPP macros, which avoid the need for indirect calls.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Replace the glue helper dependency with implementations of ECB and CBC
based on the new CPP macros, which avoid the need for indirect calls.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Replace the glue helper dependency with implementations of ECB and CBC
based on the new CPP macros, which avoid the need for indirect calls.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The x86 glue helper module is starting to show its age:
- It relies heavily on function pointers to invoke asm helper functions that
operate on fixed input sizes that are relatively small. This means the
performance is severely impacted by retpolines.
- It goes to great lengths to amortize the cost of kernel_fpu_begin()/end()
over as much work as possible, which is no longer necessary now that FPU
save/restore is done lazily, and doing so may cause unbounded scheduling
blackouts due to the fact that enabling the FPU in kernel mode disables
preemption.
- The CBC mode decryption helper makes backward strides through the input, in
order to avoid a single block size memcpy() between chunks. Consuming the
input in this manner is highly likely to defeat any hardware prefetchers,
so it is better to go through the data linearly, and perform the extra
memcpy() where needed (which is turned into direct loads and stores by the
compiler anyway). Note that benchmarks won't show this effect, given that
the memory they use is always cache hot.
- It implements blockwise XOR in terms of le128 pointers, which imply an
alignment that is not guaranteed by the API, violating the C standard.
GCC does not seem to be smart enough to elide the indirect calls when the
function pointers are passed as arguments to static inline helper routines
modeled after the existing ones. So instead, let's create some CPP macros
that encapsulate the core of the ECB and CBC processing, so we can wire
them up for existing users of the glue helper module, i.e., Camellia,
Serpent, Twofish and CAST6.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Blowfish in counter mode is never used in the kernel, so there
is no point in keeping an accelerated implementation around.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
DES or Triple DES in counter mode is never used in the kernel, so there
is no point in keeping an accelerated implementation around.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The glue helper's CTR routines are no longer used, so drop them.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Twofish in CTR mode is never used by the kernel directly, and is highly
unlikely to be relied upon by dm-crypt or algif_skcipher. So let's drop
the accelerated CTR mode implementation, and instead, rely on the CTR
template and the bare cipher.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
CAST6 in CTR mode is never used by the kernel directly, and is highly
unlikely to be relied upon by dm-crypt or algif_skcipher. So let's drop
the accelerated CTR mode implementation, and instead, rely on the CTR
template and the bare cipher.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
CAST5 in CTR mode is never used by the kernel directly, and is highly
unlikely to be relied upon by dm-crypt or algif_skcipher. So let's drop
the accelerated CTR mode implementation, and instead, rely on the CTR
template and the bare cipher.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Serpent in CTR mode is never used by the kernel directly, and is highly
unlikely to be relied upon by dm-crypt or algif_skcipher. So let's drop
the accelerated CTR mode implementation, and instead, rely on the CTR
template and the bare cipher.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Camellia in CTR mode is never used by the kernel directly, and is highly
unlikely to be relied upon by dm-crypt or algif_skcipher. So let's drop
the accelerated CTR mode implementation, and instead, rely on the CTR
template and the bare cipher.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The glue helper's XTS routines are no longer used, so drop them.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Now that the XTS template can wrap accelerated ECB modes, it can be
used to implement Twofish in XTS mode as well, which turns out to
be at least as fast, and sometimes even faster
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Now that the XTS template can wrap accelerated ECB modes, it can be
used to implement Serpent in XTS mode as well, which turns out to
be at least as fast, and sometimes even faster
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Now that the XTS template can wrap accelerated ECB modes, it can be
used to implement CAST6 in XTS mode as well, which turns out to
be at least as fast, and sometimes even faster
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Now that the XTS template can wrap accelerated ECB modes, it can be
used to implement Camellia in XTS mode as well, which turns out to
be at least as fast, and sometimes even faster.
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Replace the function pointers in the GCM implementation with static branches,
which are based on code patching, which occurs only at module load time.
This avoids the severe performance penalty caused by the use of retpolines.
In order to retain the ability to switch between different versions of the
implementation based on the input size on cores that support AVX and AVX2,
use static branches instead of static calls.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Currently, the gcm(aes-ni) driver open codes the scatterlist handling
that is encapsulated by the skcipher walk API. So let's switch to that
instead.
Also, move the handling at the end of gcmaes_crypt_by_sg() that is
dependent on whether we are encrypting or decrypting into the callers,
which always do one or the other.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The gcm(aes-ni) driver is only built for x86_64, which does not make
use of highmem. So testing for PageHighMem is pointless and can be
omitted.
While at it, replace GFP_ATOMIC with the appropriate runtime decided
value based on the context.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Drop some prototypes that are declared but never called.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The GCM mode driver uses 16 byte aligned buffers on the stack to pass
the IV to the asm helpers, but unfortunately, the x86 port does not
guarantee that the stack pointer is 16 byte aligned upon entry in the
first place. Since the compiler is not aware of this, it will not emit
the additional stack realignment sequence that is needed, and so the
alignment is not guaranteed to be more than 8 bytes.
So instead, allocate some padding on the stack, and realign the IV
pointer by hand.
Cc: <stable@vger.kernel.org>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The AES-NI driver implements XTS via the glue helper, which consumes
a struct with sets of function pointers which are invoked on chunks
of input data of the appropriate size, as annotated in the struct.
Let's get rid of this indirection, so that we can perform direct calls
to the assembler helpers. Instead, let's adopt the arm64 strategy, i.e.,
provide a helper which can consume inputs of any size, provided that the
penultimate, full block is passed via the last call if ciphertext stealing
needs to be applied.
This also allows us to enable the XTS mode for i386.
Tested-by: Eric Biggers <ebiggers@google.com> # x86_64
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Reported-by: kernel test robot <lkp@intel.com>
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The XTS asm helper arrangement is a bit odd: the 8-way stride helper
consists of back-to-back calls to the 4-way core transforms, which
are called indirectly, based on a boolean that indicates whether we
are performing encryption or decryption.
Given how costly indirect calls are on x86, let's switch to direct
calls, and given how the 8-way stride doesn't really add anything
substantial, use a 4-way stride instead, and make the asm core
routine deal with any multiple of 4 blocks. Since 512 byte sectors
or 4 KB blocks are the typical quantities XTS operates on, increase
the stride exported to the glue helper to 512 bytes as well.
As a result, the number of indirect calls is reduced from 3 per 64 bytes
of in/output to 1 per 512 bytes of in/output, which produces a 65% speedup
when operating on 1 KB blocks (measured on a Intel(R) Core(TM) i7-8650U CPU)
Fixes: 9697fa39efd3f ("x86/retpoline/crypto: Convert crypto assembler indirect jumps")
Tested-by: Eric Biggers <ebiggers@google.com> # x86_64
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Add a NEON-accelerated implementation of BLAKE2b.
On Cortex-A7 (which these days is the most common ARM processor that
doesn't have the ARMv8 Crypto Extensions), this is over twice as fast as
SHA-256, and slightly faster than SHA-1. It is also almost three times
as fast as the generic implementation of BLAKE2b:
Algorithm Cycles per byte (on 4096-byte messages)
=================== =======================================
blake2b-256-neon 14.0
sha1-neon 16.3
blake2s-256-arm 18.8
sha1-asm 20.8
blake2s-256-generic 26.0
sha256-neon 28.9
sha256-asm 32.0
blake2b-256-generic 38.9
This implementation isn't directly based on any other implementation,
but it borrows some ideas from previous NEON code I've written as well
as from chacha-neon-core.S. At least on Cortex-A7, it is faster than
the other NEON implementations of BLAKE2b I'm aware of (the
implementation in the BLAKE2 official repository using intrinsics, and
Andrew Moon's implementation which can be found in SUPERCOP). It does
only one block at a time, so it performs well on short messages too.
NEON-accelerated BLAKE2b is useful because there is interest in using
BLAKE2b-256 for dm-verity on low-end Android devices (specifically,
devices that lack the ARMv8 Crypto Extensions) to replace SHA-1. On
these devices, the performance cost of upgrading to SHA-256 may be
unacceptable, whereas BLAKE2b-256 would actually improve performance.
Although BLAKE2b is intended for 64-bit platforms (unlike BLAKE2s which
is intended for 32-bit platforms), on 32-bit ARM processors with NEON,
BLAKE2b is actually faster than BLAKE2s. This is because NEON supports
64-bit operations, and because BLAKE2s's block size is too small for
NEON to be helpful for it. The best I've been able to do with BLAKE2s
on Cortex-A7 is 18.8 cpb with an optimized scalar implementation.
(I didn't try BLAKE2sp and BLAKE3, which in theory would be faster, but
they're more complex as they require running multiple hashes at once.
Note that BLAKE2b already uses all the NEON bandwidth on the Cortex-A7,
so I expect that any speedup from BLAKE2sp or BLAKE3 would come only
from the smaller number of rounds, not from the extra parallelism.)
For now this BLAKE2b implementation is only wired up to the shash API,
since there is no library API for BLAKE2b yet. However, I've tried to
keep things consistent with BLAKE2s, e.g. by defining
blake2b_compress_arch() which is analogous to blake2s_compress_arch()
and could be exported for use by the library API later if needed.
Acked-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Tested-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Add an ARM scalar optimized implementation of BLAKE2s.
NEON isn't very useful for BLAKE2s because the BLAKE2s block size is too
small for NEON to help. Each NEON instruction would depend on the
previous one, resulting in poor performance.
With scalar instructions, on the other hand, we can take advantage of
ARM's "free" rotations (like I did in chacha-scalar-core.S) to get an
implementation get runs much faster than the C implementation.
Performance results on Cortex-A7 in cycles per byte using the shash API:
4096-byte messages:
blake2s-256-arm: 18.8
blake2s-256-generic: 26.0
500-byte messages:
blake2s-256-arm: 20.3
blake2s-256-generic: 27.9
100-byte messages:
blake2s-256-arm: 29.7
blake2s-256-generic: 39.2
32-byte messages:
blake2s-256-arm: 50.6
blake2s-256-generic: 66.2
Except on very short messages, this is still slower than the NEON
implementation of BLAKE2b which I've written; that is 14.0, 16.4, 25.8,
and 76.1 cpb on 4096, 500, 100, and 32-byte messages, respectively.
However, optimized BLAKE2s is useful for cases where BLAKE2s is used
instead of BLAKE2b, such as WireGuard.
This new implementation is added in the form of a new module
blake2s-arm.ko, which is analogous to blake2s-x86_64.ko in that it
provides blake2s_compress_arch() for use by the library API as well as
optionally register the algorithms with the shash API.
Acked-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Tested-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Add helper functions for shash implementations of BLAKE2s to
include/crypto/internal/blake2s.h, taking advantage of
__blake2s_update() and __blake2s_final() that were added by the previous
patch to share more code between the library and shash implementations.
crypto_blake2s_setkey() and crypto_blake2s_init() are usable as
shash_alg::setkey and shash_alg::init directly, while
crypto_blake2s_update() and crypto_blake2s_final() take an extra
'blake2s_compress_t' function pointer parameter. This allows the
implementation of the compression function to be overridden, which is
the only part that optimized implementations really care about.
The new functions are inline functions (similar to those in sha1_base.h,
sha256_base.h, and sm3_base.h) because this avoids needing to add a new
module blake2s_helpers.ko, they aren't *too* long, and this avoids
indirect calls which are expensive these days. Note that they can't go
in blake2s_generic.ko, as that would require selecting CRYPTO_BLAKE2S
from CRYPTO_BLAKE2S_X86, which would cause a recursive dependency.
Finally, use these new helper functions in the x86 implementation of
BLAKE2s. (This part should be a separate patch, but unfortunately the
x86 implementation used the exact same function names like
"crypto_blake2s_update()", so it had to be updated at the same time.)
Signed-off-by: Eric Biggers <ebiggers@google.com>
Acked-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The shash_alg structs for the four variants of BLAKE2s are identical
except for the algorithm name, driver name, and digest size. So, avoid
code duplication by using a macro to define these structs.
Acked-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Counter mode is a stream cipher chaining mode that is typically used
with inputs that are of arbitrarily length, and so a tail block which
is smaller than a full AES block is rule rather than exception.
The current ctr(aes) implementation for arm64 always makes a separate
call into the assembler routine to process this tail block, which is
suboptimal, given that it requires reloading of the AES round keys,
and prevents us from handling this tail block using the 5-way stride
that we use for better performance on deep pipelines.
So let's update the assembler routine so it can handle any input size,
and uses NEON permutation instructions and overlapping loads and stores
to handle the tail block. This results in a ~16% speedup for 1420 byte
blocks on cores with deep pipelines such as ThunderX2.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Commit 69b6f2e817e5b ("crypto: arm64/aes-neon - limit exposed routines if
faster driver is enabled") intended to hide modes from the plain NEON
driver that are also implemented by the faster bit sliced NEON one if
both are enabled. However, the defined() CPP function does not detect
if the bit sliced NEON driver is enabled as a module. So instead, let's
use IS_ENABLED() here.
Fixes: 69b6f2e817e5b ("crypto: arm64/aes-neon - limit exposed routines if ...")
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The cipher routines in the crypto API are mostly intended for templates
implementing skcipher modes generically in software, and shouldn't be
used outside of the crypto subsystem. So move the prototypes and all
related definitions to a new header file under include/crypto/internal.
Also, let's use the new module namespace feature to move the symbol
exports into a new namespace CRYPTO_INTERNAL.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Acked-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Follow the same approach as the arm64 driver for implementing a version
of AES-NI in CBC mode that supports ciphertext stealing. This results in
a ~2x speed increase for relatively short inputs (less than 256 bytes),
which is relevant given that AES-CBC with ciphertext stealing is used
for filename encryption in the fscrypt layer. For larger inputs, the
speedup is still significant (~25% on decryption, ~6% on encryption)
Tested-by: Eric Biggers <ebiggers@google.com> # x86_64
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Commit 86cd97ec4b943af3 ("crypto: arm/chacha-neon - optimize for non-block
size multiples") refactored the chacha block handling in the glue code in
a way that may result in the counter increment to be omitted when calling
chacha_block_xor_neon() to process a full block. This violates the skcipher
API, which requires that the output IV is suitable for handling more input
as long as the preceding input has been presented in round multiples of the
block size. Also, the same code is exposed via the chacha library interface
whose callers may actually rely on this increment to occur even for final
blocks that are smaller than the chacha block size.
So increment the counter after calling chacha_block_xor_neon().
Fixes: 86cd97ec4b943af3 ("crypto: arm/chacha-neon - optimize for non-block size multiples")
Reported-by: Eric Biggers <ebiggers@kernel.org>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|