Re: [PATCH v2 05/13] crypto: simd - Update `walksize` in simd skcipher

From: Jerry Shih
Date: Thu Nov 30 2023 - 21:10:03 EST


On Nov 29, 2023, at 01:22, Eric Biggers <ebiggers@xxxxxxxxxx> wrote:
> On Tue, Nov 28, 2023 at 01:38:29PM +0800, Jerry Shih wrote:
>> On Nov 28, 2023, at 11:58, Eric Biggers <ebiggers@xxxxxxxxxx> wrote:
>>> On Mon, Nov 27, 2023 at 03:06:55PM +0800, Jerry Shih wrote:
>>>> The `walksize` assignment is missed in simd skcipher.
>>>>
>>>> Signed-off-by: Jerry Shih <jerry.shih@xxxxxxxxxx>
>>>> ---
>>>> crypto/cryptd.c | 1 +
>>>> crypto/simd.c | 1 +
>>>> 2 files changed, 2 insertions(+)
>>>>
>>>> diff --git a/crypto/cryptd.c b/crypto/cryptd.c
>>>> index bbcc368b6a55..253d13504ccb 100644
>>>> --- a/crypto/cryptd.c
>>>> +++ b/crypto/cryptd.c
>>>> @@ -405,6 +405,7 @@ static int cryptd_create_skcipher(struct crypto_template *tmpl,
>>>> (alg->base.cra_flags & CRYPTO_ALG_INTERNAL);
>>>> inst->alg.ivsize = crypto_skcipher_alg_ivsize(alg);
>>>> inst->alg.chunksize = crypto_skcipher_alg_chunksize(alg);
>>>> + inst->alg.walksize = crypto_skcipher_alg_walksize(alg);
>>>> inst->alg.min_keysize = crypto_skcipher_alg_min_keysize(alg);
>>>> inst->alg.max_keysize = crypto_skcipher_alg_max_keysize(alg);
>>>>
>>>> diff --git a/crypto/simd.c b/crypto/simd.c
>>>> index edaa479a1ec5..ea0caabf90f1 100644
>>>> --- a/crypto/simd.c
>>>> +++ b/crypto/simd.c
>>>> @@ -181,6 +181,7 @@ struct simd_skcipher_alg *simd_skcipher_create_compat(const char *algname,
>>>>
>>>> alg->ivsize = ialg->ivsize;
>>>> alg->chunksize = ialg->chunksize;
>>>> + alg->walksize = ialg->walksize;
>>>> alg->min_keysize = ialg->min_keysize;
>>>> alg->max_keysize = ialg->max_keysize;
>>>
>>> What are the consequences of this bug? I wonder if it actually matters? The
>>> "inner" algorithm is the one that actually gets used for the "walk", right?
>>>
>>> - Eric
>>
>> Without this, we might still use chunksize or cra_blocksize as the walksize
>> even though we setup with the larger walksize.
>>
>> Here is the code for the walksize default value:
>> static int skcipher_prepare_alg(struct skcipher_alg *alg)
>> {
>> ...
>> if (!alg->chunksize)
>> alg->chunksize = base->cra_blocksize;
>> if (!alg->walksize)
>> alg->walksize = alg->chunksize;
>>
>> And we already have the bigger walksize for x86 aes-xts.
>> .base = {
>> .cra_name = "__xts(aes)",
>> ...
>> },
>> .walksize = 2 * AES_BLOCK_SIZE,
>>
>> The x86 aes-xts only uses one `walk` to handle the tail elements. It assumes
>> that the walksize contains 2 aes blocks. If walksize is not set correctly, maybe
>> some tail elements is not processed in simd-cipher mode for x86 aes-xts.
>
> With the SIMD helper there are three "algorithms": the underlying algorithm, the
> cryptd algorithm, and the simd algorithm. This patch makes the "walksize"
> property be propagated from the underlying algorithm to the cryptd and simd
> algorithms. I don't see how that actually makes a difference, since the only
> place the skcipher_walk happens is on the underlying algorithm. So it uses the
> "walksize" from the underlying algorithm, right?
>
> - Eric

Yes, you are right.
I re-check the cryptd and simd cipher flow. They use the underlying algorithms.
So, the actual `walksize` in the underlying algorithm is set by the user in
skcipher_alg def.
The x86 aes-xts works correctly for both cryptd and simd-cipher case.

This patch becomes fixing the `walksize` display error in `/proc/crypto`.

The aes-xts skcipher_alg def:
...
.ivsize = AES_BLOCK_SIZE,
.chunksize = AES_BLOCK_SIZE,
.walksize = AES_BLOCK_SIZE * 8,
.base = {
.cra_flags = CRYPTO_ALG_INTERNAL,
.cra_name = "__xts(aes)",
.cra_driver_name = "__xts-aes-riscv64-zvkned-zvbb-zvkg",
...
},


Without patch:
The original skcipher:
name : __xts(aes)
driver : __xts-aes-riscv64-zvkned-zvbb-zvkg
internal : yes
async : no
...
walksize : 128

The async skcipher registered by simd_register_skciphers_compat:
name : xts(aes)
driver : xts-aes-riscv64-zvkned-zvbb-zvkg
internal : no
async : yes
...
walksize : 16

...
name : __xts(aes)
driver : cryptd(__xts-aes-riscv64-zvkned-zvbb-zvkg)
internal : yes
async : yes
...
walksize : 16

With patch:
name : xts(aes)
driver : xts-aes-riscv64-zvkned-zvbb-zvkg
internal : no
async : yes
...
walksize : 128

...
name : __xts(aes)
driver : cryptd(__xts-aes-riscv64-zvkned-zvbb-zvkg)
internal : yes
async : yes
...
walksize : 128