Little-endian (x86) PBKDF2 (SHA1) which relies on the AF_ALG user-space interface to the Linux Kernel Cryptography system

pbkdf2.c

#include <sys/socket.h>
#include <linux/if_alg.h>
#include <string.h>
#include <unistd.h>
#include <byteswap.h>

static struct sockaddr_alg sa = {
        .salg_family = AF_ALG
,       .salg_type = "hash"
,       .salg_name = "hmac(sha1)"
};

#define BLOCKSIZE 20

void pbkdf2(
                const char *password, const int password_len,
                const char *salt, const int salt_len,
                const int iter,
                unsigned char *key_buffer, const int key_length
) {
        unsigned char last[BLOCKSIZE], xorb[BLOCKSIZE];
        unsigned char count_buffer[4];
        int block_count = (key_length + BLOCKSIZE - 1) / BLOCKSIZE;
        int block_rem = ((key_length + BLOCKSIZE - 1) % 20) + 1;
        int block, cur_iter, i, control, hmac, block_bswap;

        control = socket(AF_ALG, SOCK_SEQPACKET, 0);
        bind(control, (struct sockaddr *)&sa, sizeof(sa));
        setsockopt(control, SOL_ALG, ALG_SET_KEY, password, password_len);

        for (block = 1; block <= block_count; block++) {
                block_bswap = bswap_32(block);

                hmac = accept(control, NULL, 0);
                sendto(hmac, salt, salt_len, MSG_MORE, NULL, 0);
                sendto(hmac, &block_bswap, 4, 0, NULL, 0);
                recvfrom(hmac, last, BLOCKSIZE, 0, NULL, NULL);
                close(hmac);

                memcpy(xorb, last, BLOCKSIZE);

                for (cur_iter = 1; cur_iter < iter; cur_iter++) {
                        hmac = accept(control, NULL, 0);
                        sendto(hmac, last, BLOCKSIZE, 0, NULL, 0);
                        recvfrom(hmac, last, BLOCKSIZE, 0, NULL, NULL);
                        close(hmac);
                        for (i = 0; i < BLOCKSIZE; i++) {
                                xorb[i] ^= last[i];
                        }
                }

                memcpy(key_buffer + (block * BLOCKSIZE) - BLOCKSIZE, xorb, (block < block_count) ? BLOCKSIZE : block_rem);
        }

        close(control);

        return;
}

As a note, this code fragment is specifically avoiding any libc calls with the exception of memcpy which generally optimizes out to an inline macro regardless, and is targeting x86 explicitly thus the explicit byte-swap to big-endian which the PBKDF2 standard requires.

Including the sockaddr data structure this compiles down to <600 bytes total program space taken.