On Wed, 2018-08-08 at 17:56 +0800, Herbert Xu wrote: > On Thu, Jul 26, 2018 at 05:25:07PM -0700, Megha Dey wrote: > > > > 1. On the existing algorithms covered in aesni_intel-glue.c (eg: > > __cbc-aes-aesni), 3 algorithms are registered in /proc/crypto: > > > > __cbc(aes) > > cryptd(__cbc-aes-aesni)--> registered via cryptd_create_skcipher > > > > cbc(aes) > > cbc-aes-aesni --> registered via simd_skcipher_create_compat > > > > __cbc(aes) > > __cbc-aes-aesni --> registered as the internal algorithm > > > > I would want to know why do we need the cryptd(__cbc-aes-aesni) > > algorithm at all. I do not see any of the associated setkey, encrypt or > > decrypt functions getting called during the selftest or while running > > tcrypt. I just see the simd_(setkey, encrypt, decrypt) functions > > directly called the inner algorithms. However, if I remove the cryptd > > algorithm, none of the algorithms are registered. > > The simd functions are the fast path where you are running in a > context where SIMD can be used directly. cryptd is the slow path > where we defer the work to a work queue. Hi Herbert, Thank you for the clarification. I seem to have gotten things to work (i.e remove mcryptd layer). I have tried this with the skcipher on top of my previously posted patches for the aes-cbc-mb multibuffer algorithm since the simd wrappers already exist for it. I am working on extending to hashes, sorry for the confusion. I would like to get your approval first on the changes I have made in the cryptd layer: 1. @@ -495,7 +534,10 @@ static void cryptd_skcipher_encrypt(struct crypto_async_request *base, skcipher_request_set_crypt(subreq, req->src, req->dst, req->cryptlen, req->iv); - err = crypto_skcipher_encrypt(subreq); + subreq->base.data = req->base.data; + subreq->base.complete = rctx->complete; + rctx->desc = *subreq; + err = crypto_skcipher_encrypt(&rctx->desc); skcipher_request_zero(subreq); This change is necessary because for the multibuffer algorithms, the inner algorithm needs a pointer to the original request. In the slow path, since we allocate a skcipher_request on the stack, there is no easy way to retrieve the request. In the mcryptd_layer, we had extra logic to store this pointer. 2. Currently, -struct cryptd_skcipher_request_ctx { - crypto_completion_t complete; -}; - For multibuffer algorithms, we need more structure members: +struct cryptd_skcipher_request_ctx { + struct list_head waiter; + crypto_completion_t complete; + struct cryptd_tag tag; + struct skcipher_walk walk; + u8 flag; + int nbytes; + int error; + struct skcipher_request desc; + void *job; + u128 seq_iv; I am not sure if adding the member to the original structure definition is acceptable or I should introduce a new structure. Lastly, for hashes, we have struct cryptd_hash_request_ctx { crypto_completion_t complete; struct shash_desc desc; }; If we were to use this(with the added fields for multibuffer), we should update the shash_desc to ahash_request since we are an async algorithm right? > > > > What you need to do is create an actual simd wrapper with cryptd > > > > This simd wrapper is already present for skcipher right(in simd.c)? > > Assuming we only have ciphers and no hash algorithms, are any changes > > required in these wrappers? > > For skcipher yes they already exist. But this thread was about > hashes. > > Cheers,