quic/decrypt_crypto: Reuse OpenSSL context objects

[awelzel]

It is not necessary to allocate and free the context objects used for HKDF and AES all the time, they can be re-used. The main assumption here is no cross-thread usage, but this should be guaranteed even with the fibers: QUIC_decrypt_crypto_payload() always runs to completion.

A pcap with ~12k QUIC connections had ~15% samples in QUIC_decrypt_crypto_payload. After this change it is down to 5% of samples. The improvement in runtime is ~16%, 12.2 seconds to 10.2 seconds.

---

This is lots of copy paste currently, I have the skills to do #define, but wonder if there might be a nicer approch.

[bbannier]

This code clearly works today, but it will fail in very subtle ways (in production) should we ever process analyzers in parallel, e.g., during DPD. I feel this needs a very very fat warning on the top explaining why this is not thread safe at all.

[bbannier]

Just as a note, in multi-threaded code this would not be safe since ctx might be set once you reach this assignment. You'd either need to use double-checked locking, or use a static stack variable which is guaranteed by the language to be initialized exactly once on first use, e.g.,

EVP_CIPHER_CTX* get_aes_128_ecb()
	{
	static struct Ctx
		{
		Ctx() : val(EVP_CIPHER_CTX_new())
			{
			EVP_CipherInit_ex(val, EVP_aes_128_ecb(), NULL, NULL, NULL, 1);
			}
		EVP_CIPHER_CTX* val;
		} ctx;

	return ctx.val;
	}

Using such a struct would actually give you the option of implementing a destructor which calls EVP_PKEY_CTX_free so you do not leak the context anymore (which might work if the order in which static objects are destructed works out right).