- 28 Dec, 2017 11 commits
-
-
Horia Geantă authored
Offload split key generation in CAAM engine, using DKP. DKP is supported starting with Era 6. Note that the way assoclen is transmitted from the job descriptor to the shared descriptor changes - DPOVRD register is used instead of MATH3 (where available), since DKP protocol thrashes the MATH registers. The replacement of MDHA split key generation with DKP has the side effect of the crypto engine writing the authentication key, and thus the DMA mapping direction for the buffer holding the key has to change from DMA_TO_DEVICE to DMA_BIDIRECTIONAL. There are two cases: -key is inlined in descriptor - descriptor buffer mapping changes -key is referenced - key buffer mapping changes Signed-off-by:
Horia Geantă <horia.geanta@nxp.com> Signed-off-by:
Herbert Xu <herbert@gondor.apana.org.au>
-
Horia Geantă authored
Save Era in driver's private data for further usage, like deciding whether an erratum applies or a feature is available based on its value. Signed-off-by:
-
Horia Geantă authored
ablkcipher shared descriptors are relatively small, thus there is enough space for the key to be inlined. Accordingly, there is no need to copy the key in ctx->key. Signed-off-by:
-
Horia Geantă authored
Key data is not modified, it is copied in the shared descriptor. Signed-off-by:
-
Eric Biggers authored
Using %rbp as a temporary register breaks frame pointer convention and breaks stack traces when unwinding from an interrupt in the crypto code. In twofish-3way, we can't simply replace %rbp with another register because there are none available. Instead, we use the stack to hold the values that %rbp, %r11, and %r12 were holding previously. Each of these values represents the half of the output from the previous Feistel round that is being passed on unchanged to the following round. They are only used once per round, when they are exchanged with %rax, %rbx, and %rcx. As a result, we free up 3 registers (one per block) and can reassign them so that %rbp is not used, and additionally %r14 and %r15 are not used so they do not need to be saved/restored. There may be a small overhead caused by replacing 'xchg REG, REG' with the needed sequence 'mov MEM, REG; mov REG, MEM; mov REG, REG' once per round. But, counterintuitively, when I tested "ctr-twofish-3way" on a Haswell processor, the new version was actually about 2% faster. (Perhaps 'xchg' is not as well optimized as plain moves.) Reported-by:
syzbot <syzkaller@googlegroups.com> Signed-off-by:
Eric Biggers <ebiggers@google.com> Reviewed-by:
Josh Poimboeuf <jpoimboe@redhat.com> Signed-off-by:
-
Gilad Ben-Yossef authored
The performance of some aead tfm providers is affected by the amount of parallelism possible with the processing. Introduce an async aead concurrent multiple buffer processing speed test to be able to test performance of such tfm providers. Signed-off-by:
Gilad Ben-Yossef <gilad@benyossef.com> Signed-off-by:
-
Gilad Ben-Yossef authored
The performance of some skcipher tfm providers is affected by the amount of parallelism possible with the processing. Introduce an async skcipher concurrent multiple buffer processing speed test to be able to test performance of such tfm providers. Signed-off-by:
-
Gilad Ben-Yossef authored
The multi buffer concurrent requests ahash speed test only supported the cycles mode. Add support for the so called jiffies mode that test performance of bytes/sec. We only add support for digest mode at the moment. Signed-off-by:
-
Gilad Ben-Yossef authored
For multiple buffers speed tests, the number of buffers, or requests, used actually sets the level of parallelism a tfm provider may utilize to hide latency. The existing number (of 8) is good for some software based providers but not enough for many HW providers with deep FIFOs. Add a module parameter that allows setting the number of multiple buffers/requests used, leaving the default at 8. Signed-off-by:
-
Gilad Ben-Yossef authored
The AEAD speed test pretended to support decryption, however that support was broken as decryption requires a valid auth field which the test did not provide. Fix this by running the encryption path once with inout/output sgls switched to calculate the auth field prior to performing decryption speed tests. Signed-off-by:
-
Gilad Ben-Yossef authored
The multi buffer ahash speed test was allocating multiple buffers for use with the multiple outstanding requests it was starting but never actually using them (except to free them), instead using a different single statically allocated buffer for all requests. Fix this by actually using the allocated buffers for the test. It is noted that it may seem tempting to instead remove the allocation and free of the multiple buffers and leave things as they are since this is a hash test where the input is read only. However, after consideration I believe that multiple buffers better reflect real life scenario with regard to data cache and TLB behaviours etc. Signed-off-by:
-
- 22 Dec, 2017 29 commits
-
-
Gary R Hook authored
Commit 142a27f0 added support for a "best" RNG, and in doing so introduced a hang from rmmod/modprobe -r when the last RNG on the list was unloaded. When the hwrng list is depleted, return the global variables to their original state and decrement all references to the object. Fixes: 142a27f0 ("hwrng: core - Reset user selected rng by writing "" to rng_current") Signed-off-by:
Gary R Hook <gary.hook@amd.com> Reviewed-by:
PrasannaKumar Muralidharan <prasannatsmkumar@gmail.com> Signed-off-by:
-
Antoine Ténart authored
The Inside Secure SafeXcel driver was firstly designed to support the EIP197 cryptographic engine which is an evolution (with much more feature, better performances) of the EIP97 cryptographic engine. This patch convert the Inside Secure SafeXcel driver to support both engines (EIP97 + EIP197). The main differences are the register offsets and the context invalidation process which is EIP197 specific. This patch adds an indirection on the register offsets and adds checks not to send any invalidation request when driving the EIP97. A new compatible is added as well to bind the driver from device trees. Signed-off-by:
Antoine Tenart <antoine.tenart@free-electrons.com> Signed-off-by:
-
Antoine Ténart authored
The dequeueing function was putting back a request in the crypto queue on failure (when not enough resources are available) which is not perfect as the request will be handled much later. This patch updates this logic by keeping a reference on the failed request to try proceeding it later when enough resources are available. Signed-off-by:
-
Antoine Ténart authored
This patch modifies the result handling logic to continue handling results when the completed requests counter is full and not showing the actual number of requests to handle. Suggested-by:
Ofer Heifetz <oferh@marvell.com> Signed-off-by:
-
Antoine Ténart authored
This patches moves the result request acknowledgment from a per request process to acknowledging all the result requests handled at once. Suggested-by:
-
Antoine Ténart authored
Increase the ring size to handle more requests in parallel, while keeping the batch size (for interrupt coalescing) to its previous value. The ring size and batch size are now unlinked. Suggested-by:
-
Antoine Ténart authored
This patch updates the dequeueing logic to dequeue all requests at once. Since we can have many requests in the queue, the interrupt coalescing is kept so that the ring interrupt fires every EIP197_MAX_BATCH_SZ at most. To allow dequeueing all requests at once while still using reasonable settings for the interrupt coalescing, the result handling function was updated to setup the threshold interrupt when needed (i.e. when more requests than EIP197_MAX_BATCH_SZ are in the queue). When using this capability the ring is marked as busy so that the dequeue function enqueue new requests without setting the threshold interrupt. Suggested-by:
-
Antoine Ténart authored
This patch moves the result handling from an IRQ handler to a threaded IRQ handler, to improve the number of complete requests being handled at once. Suggested-by:
-
Antoine Ténart authored
This patch moves the request dequeueing into a workqueue to improve the coalescing of interrupts when sending requests to the engine; as the engine is capable of having one single interrupt for n requests sent. Using a workqueue allows to send more request at once. Suggested-by:
-
Ofer Heifetz authored
The SafeXcel context isn't used in the cache invalidation function. This cosmetic patch removes it (as well as from the function prototype in the header file and when the function is called). Signed-off-by:
-
Ofer Heifetz authored
The cipher direction can be different for requests within the same transformation context. This patch moves the direction flag from the context to the request scope. Signed-off-by:
-
Ofer Heifetz authored
When initializing the IVs crypto_ahash_update() is called, which at some point will call crypto_enqueue_request(). This function can return -EBUSY when no resource is available and the request is queued. Since this is a valid case, -EBUSY shouldn't be treated as an error. Signed-off-by:
-
Ofer Heifetz authored
The check to know if an invalidation is needed (i.e. when the context changes) is done even if the context does not exist yet. This happens when first setting a key for ciphers and/or hmac operations. This commits adds a check in the _setkey functions to only check if an invalidation is needed when a context exists, as there is no need to perform this check otherwise. Signed-off-by:
-
Antoine Ténart authored
Cosmetic patch adding a few comments to the ahash caching function to understand easily what calculations are made in the functions; and how the function is working. Signed-off-by:
-
Antoine Ténart authored
This patch removes an useless memset in the ahash_export function, as the zeroed buffer will be entirely overridden the next line. Suggested-by:
-
Antoine Ténart authored
Cosmetic patch fixing one typo in one of the driver's comments. Signed-off-by:
-
Antoine Ténart authored
Cosmetic patch removing an extra empty line between header inclusions. Signed-off-by:
-
git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6Herbert Xu authored
Merge the crypto tree to pick up inside-secure fixes.
-
Sabrina Dubroca authored
When I added generic-gcm-aes I didn't add a wrapper like the one provided for rfc4106(gcm(aes)). We need to add a cryptd wrapper to fall back on in case the FPU is not available, otherwise we might corrupt the FPU state. Fixes: cce2ea8d ("crypto: aesni - add generic gcm(aes)") Cc: <stable@vger.kernel.org> Reported-by:
Ilya Lesokhin <ilyal@mellanox.com> Signed-off-by:
Sabrina Dubroca <sd@queasysnail.net> Reviewed-by:
Stefano Brivio <sbrivio@redhat.com> Signed-off-by:
-
Sabrina Dubroca authored
generic_gcmaes_decrypt needs to use generic_gcmaes_ctx, not aesni_rfc4106_gcm_ctx. This is actually harmless because the fields in struct generic_gcmaes_ctx share the layout of the same fields in aesni_rfc4106_gcm_ctx. Fixes: cce2ea8d ("crypto: aesni - add generic gcm(aes)") Cc: <stable@vger.kernel.org> Signed-off-by:
-
Corentin Labbe authored
This patch remove two unused variable and some dead "code" using it. Fixes: 92932d03 ("crypto: seqiv - Remove AEAD compatibility code") Signed-off-by:
Corentin Labbe <clabbe@baylibre.com> Signed-off-by:
-
Corentin Labbe authored
This patch remove two unused variable and some dead "code" using it. Fixes: 66008d42 ("crypto: echainiv - Remove AEAD compatibility code") Signed-off-by:
-
Łukasz Stelmach authored
Hardware operations like reading random numbers and setting a seed need to be conducted in a single thread. Therefore a mutex is required to prevent multiple threads (processes) from accessing the hardware at the same time. The sequence of mutex_lock() and mutex_unlock() in the exynos_rng_reseed() function enables switching between different threads waiting for the driver to generate random numbers for them. Signed-off-by:
Łukasz Stelmach <l.stelmach@samsung.com> Reviewed-by:
Krzysztof Kozlowski <krzk@kernel.org> Signed-off-by:
-
Łukasz Stelmach authored
Reseed PRNG after reading 65 kB of randomness. Although this may reduce performance, in most cases the loss is not noticeable. Also the time based threshold for reseeding is changed to one second. Reseeding is performed whenever either limit is exceeded. Reseeding of a PRNG does not increase entropy, but it helps preventing backtracking the internal state of the device from its output sequence, and hence, prevents potential attacker from predicting numbers to be generated. Signed-off-by:
Stephan Mueller <smueller@chronox.de> Reviewed-by:
-
Łukasz Stelmach authored
Use memcpy_fromio() instead of custom exynos_rng_copy_random() function to retrieve generated numbers from the registers of PRNG. Signed-off-by:
-
Łukasz Stelmach authored
Add support for PRNG in Exynos5250+ SoCs. Signed-off-by:
-
Lars Persson authored
The IV size should not include the 32 bit counter. Because we had the IV size set as 16 the transform only worked when the IV input was zero padded. Fixes: a21eb94f ("crypto: axis - add ARTPEC-6/7 crypto accelerator driver") Signed-off-by:
Lars Persson <larper@axis.com> Signed-off-by:
-
Eric Biggers authored
The comment in gf128mul_x8_ble() was copy-and-pasted from gf128mul.h and makes no sense in the new context. Remove it. Cc: Harsh Jain <harsh@chelsio.com> Signed-off-by:
-
Arnd Bergmann authored
Passing the register value by reference here leads a large amount of stack being used when CONFIG_KASAN is enabled: drivers/crypto/qat/qat_common/qat_hal.c: In function 'qat_hal_exec_micro_inst.constprop': drivers/crypto/qat/qat_common/qat_hal.c:963:1: error: the frame size of 1792 bytes is larger than 1536 bytes [-Werror=frame-larger-than=] Changing the register-read function to return the value instead reduces the stack size to around 800 bytes, most of which is for the 'savuwords' array. The function now no longer returns an error code, but nothing ever evaluated that anyway. Signed-off-by:
Arnd Bergmann <arnd@arndb.de> Signed-off-by:
-
