Re: [PATCH bpf-next v3 1/2] bpf: add skcipher API support to TC/XDP programs

[Vadim Fedorenko <vadim.fedorenko@linux.dev>]

On 01.11.2023 22:59, Martin KaFai Lau wrote:
> On 11/1/23 3:50 PM, Vadim Fedorenko wrote:
>>>> +static int bpf_crypto_skcipher_crypt(struct crypto_sync_skcipher *tfm,
>>>> +                     const struct bpf_dynptr_kern *src,
>>>> +                     struct bpf_dynptr_kern *dst,
>>>> +                     const struct bpf_dynptr_kern *iv,
>>>> +                     bool decrypt)
>>>> +{
>>>> +    struct skcipher_request *req = NULL;
>>>> +    struct scatterlist sgin, sgout;
>>>> +    int err;
>>>> +
>>>> +    if (crypto_sync_skcipher_get_flags(tfm) & CRYPTO_TFM_NEED_KEY)
>>>> +        return -EINVAL;
>>>> +
>>>> +    if (__bpf_dynptr_is_rdonly(dst))
>>>> +        return -EINVAL;
>>>> +
>>>> +    if (!__bpf_dynptr_size(dst) || !__bpf_dynptr_size(src))
>>>> +        return -EINVAL;
>>>> +
>>>> +    if (__bpf_dynptr_size(iv) != crypto_sync_skcipher_ivsize(tfm))
>>>> +        return -EINVAL;
>>>> +
>>>> +    req = skcipher_request_alloc(&tfm->base, GFP_ATOMIC);
>>>
>>> Doing alloc per packet may kill performance. Is it possible to optimize it 
>>> somehow? What is the usual size of the req (e.g. the example in the selftest)?
>>>
>>
>> In ktls code aead_request is allocated every time encryption is invoked, see 
>> tls_decrypt_sg(), apparently per skb. Doesn't look like performance
>> killer. For selftest it's only sizeof(struct skcipher_request).
> 
> ktls is doing the en/decrypt on the userspace behalf to compensate the cost.
> 
> When this kfunc is used in xdp to decrypt a few bytes for each packet and then 
> XDP_TX out, this extra alloc will be quite noticeable. If the size is usually 
> small, can it be done in the stack memory?

Hmm... looks like SYNC_SKCIPHER_REQUEST_ON_STACK will help us. Ok, I'll change 
this part to use stack allocations.