Hello!

Although I have no doubt, the current version works, several (substantial?) improvements can/should be made. I incorporate all of these in the G2O module for Varnish, which I wrote for my employer and so speak from some experience...

Multiple nonce-keys.

The process of changing the nonce/password with Akamai is not instantaneous. This is why Akamai are using the nonce concept to begin with -- to allow for multiple (at least -- two) passwords to remain valid in parallel. For some time after you change the Akamai configuration some of their servers will still be using the old nonce, so your back-end (origin) should recognize both. I implement this with a simple array of structures -- because the number of valid nonces is small (though greater than 1), linear search through the array is the fastest method :)

Conversion to string

You are wasting time for each hit converting the computed digest into a character string and then using strcmp to compare it with the supplied signature. Instead, it is faster to convert the supplied signature into numbers (of type uint64_t) and then comparing the two pairs of numbers:

    union qdigest {
        unsigned char   bytes[16];
        uint64_t    numbers[2];
    };
    ...
    union qdigest    expected, computed;
    ...
    b64decode(signature, expected.bytes);
    ...
    HMAC_Final(tcontexts + count, computed.bytes, &len);
    ...
    if (len != 16 ||
        expected.numbers[0] != computed.numbers[0] ||
        expected.numbers[1] != computed.numbers[1]) {
        /* Report mismatch */

Recomputing the digest anew

Some of the HMAC computations need only be done once (for each key), yet the current code recomputes the digest anew for each hit. Each thread can keep the HMAC-context for each key and keep reusing it. These two calls need only be done once for each key (in each thread):

    HMAC_CTX_init(context); /* Is this even necessary? */
    HMAC_Init_ex(context, key, len, EVP_md5(), NULL);

After this, for each hit only the following needs to happen:

    HMAC_Update(context, (const unsigned char *)d, strlen(d));
    HMAC_Update(context, (const unsigned char *)path, strlen(path));
    HMAC_Final(context, computed.bytes, &len);
    HMAC_Init_ex(context, NULL, 0, NULL, NULL);

The last call resets the context back to just the key - by passing the zeros, the caller indicates, he wants to reuse the already-supplied key. Presumably, this is faster, than full reinitialization.

I can not (easily) offer a working patch, because I don't have an nginx installation readily available. But you should be able to incorporate (some of) these ideas easily yourself.

There's no licence file checked-in with the code, and the source code has (c)opyright tags at the top.

Is the source able to be copied? or open source? Is it possible to specify the licence, either here in this issue, or ideally in the code. See https://help.github.com/articles/open-source-licensing/