static void rand_get_seed(struct rand_state *state,
uint8_t seed[CTR_DRBG_ENTROPY_LEN]) {
if (!state->last_block_valid) {
if (!hwrand(state->last_block, sizeof(state->last_block))) {
CRYPTO_sysrand(state->last_block, sizeof(state->last_block));
}
state->last_block_valid = 1;
}
// We overread from /dev/urandom or RDRAND by a factor of 10 and XOR to
// whiten.
#define FIPS_OVERREAD 10
uint8_t entropy[CTR_DRBG_ENTROPY_LEN * FIPS_OVERREAD];
if (!hwrand(entropy, sizeof(entropy))) {
CRYPTO_sysrand(entropy, sizeof(entropy));
}
// See FIPS 140-2, section 4.9.2. This is the “continuous random number
// generator test” which causes the program to randomly abort. Hopefully the
// rate of failure is small enough not to be a problem in practice.
if (CRYPTO_memcmp(state->last_block, entropy, CRNGT_BLOCK_SIZE) == 0) {
fprintf(stderr, "CRNGT failed.\n");
BORINGSSL_FIPS_abort();
}
for (size_t i = CRNGT_BLOCK_SIZE; i < sizeof(entropy);
i += CRNGT_BLOCK_SIZE) {
if (CRYPTO_memcmp(entropy + i - CRNGT_BLOCK_SIZE, entropy + i,
CRNGT_BLOCK_SIZE) == 0) {
fprintf(stderr, "CRNGT failed.\n");
BORINGSSL_FIPS_abort();
}
}
OPENSSL_memcpy(state->last_block,
entropy + sizeof(entropy) - CRNGT_BLOCK_SIZE,
CRNGT_BLOCK_SIZE);
OPENSSL_memcpy(seed, entropy, CTR_DRBG_ENTROPY_LEN);
for (size_t i = 1; i < FIPS_OVERREAD; i++) {
for (size_t j = 0; j < CTR_DRBG_ENTROPY_LEN; j++) {
seed[j] ^= entropy[CTR_DRBG_ENTROPY_LEN * i + j];
}
}
}
void RAND_bytes_with_additional_data(uint8_t *out, size_t out_len,
const uint8_t user_additional_data[32]) {
if (out_len == 0) {
return;
}
// Additional data is mixed into every CTR-DRBG call to protect, as best we
// can, against forks & VM clones. We do not over-read this information and
// don't reseed with it so, from the point of view of FIPS, this doesn't
// provide “prediction resistance”. But, in practice, it does.
uint8_t additional_data[32];
if (!hwrand(additional_data, sizeof(additional_data))) {
// Without a hardware RNG to save us from address-space duplication, the OS
// entropy is used. This can be expensive (one read per |RAND_bytes| call)
// and so can be disabled by applications that we have ensured don't fork
// and aren't at risk of VM cloning.
if (!rand_fork_unsafe_buffering_enabled()) {
CRYPTO_sysrand(additional_data, sizeof(additional_data));
} else {
OPENSSL_memset(additional_data, 0, sizeof(additional_data));
}
}
for (size_t i = 0; i < sizeof(additional_data); i++) {
additional_data[i] ^= user_additional_data[i];
}
struct rand_state stack_state;
struct rand_state *state = rand_state_get();
if (state == NULL) {
// If the system is out of memory, use an ephemeral state on the
// stack.
state = &stack_state;
rand_state_init(state);
}
if (state->calls >= kReseedInterval) {
uint8_t seed[CTR_DRBG_ENTROPY_LEN];
rand_get_seed(state, seed);
#if defined(BORINGSSL_FIPS)
// Take a read lock around accesses to |state->drbg|. This is needed to
// avoid returning bad entropy if we race with
// |rand_state_clear_all|.
//
// This lock must be taken after any calls to |CRYPTO_sysrand| to avoid a
// bug on ppc64le. glibc may implement pthread locks by wrapping user code
// in a hardware transaction, but, on some older versions of glibc and the
// kernel, syscalls made with |syscall| did not abort the transaction.
CRYPTO_STATIC_MUTEX_lock_read(rand_drbg_lock_bss_get());
#endif
if (!CTR_DRBG_reseed(&state->drbg, seed, NULL, 0)) {
abort();
}
state->calls = 0;
} else {
#if defined(BORINGSSL_FIPS)
CRYPTO_STATIC_MUTEX_lock_read(rand_drbg_lock_bss_get());
#endif
}
int first_call = 1;
while (out_len > 0) {
size_t todo = out_len;
if (todo > CTR_DRBG_MAX_GENERATE_LENGTH) {
todo = CTR_DRBG_MAX_GENERATE_LENGTH;
}
if (!CTR_DRBG_generate(&state->drbg, out, todo, additional_data,
first_call ? sizeof(additional_data) : 0)) {
abort();
}
out += todo;
out_len -= todo;
state->calls++;
first_call = 0;
}
if (state == &stack_state) {
CTR_DRBG_clear(&state->drbg);
}
#if defined(BORINGSSL_FIPS)
CRYPTO_STATIC_MUTEX_unlock_read(rand_drbg_lock_bss_get());
#endif
if (state != &stack_state) {
rand_state_put(state);
}
}
static void rand_get_seed(struct rand_state *state,
uint8_t seed[CTR_DRBG_ENTROPY_LEN]) {
// If not in FIPS mode, we don't overread from the system entropy source and
// we don't depend only on the hardware RDRAND.
CRYPTO_sysrand(seed, CTR_DRBG_ENTROPY_LEN);
}
int RAND_bytes(uint8_t *buf, size_t len) {
if (len == 0) {
return 1;
}
if (!CRYPTO_have_hwrand() ||
!CRYPTO_hwrand(buf, len)) {
/* Without a hardware RNG to save us from address-space duplication, the OS
* entropy is used directly. */
CRYPTO_sysrand(buf, len);
return 1;
}
struct rand_thread_state *state =
CRYPTO_get_thread_local(OPENSSL_THREAD_LOCAL_RAND);
if (state == NULL) {
state = OPENSSL_malloc(sizeof(struct rand_thread_state));
if (state == NULL ||
!CRYPTO_set_thread_local(OPENSSL_THREAD_LOCAL_RAND, state,
rand_thread_state_free)) {
CRYPTO_sysrand(buf, len);
return 1;
}
memset(state->partial_block, 0, sizeof(state->partial_block));
state->calls_used = kMaxCallsPerRefresh;
}
if (state->calls_used >= kMaxCallsPerRefresh ||
state->bytes_used >= kMaxBytesPerRefresh) {
CRYPTO_sysrand(state->key, sizeof(state->key));
state->calls_used = 0;
state->bytes_used = 0;
state->partial_block_used = sizeof(state->partial_block);
}
if (len >= sizeof(state->partial_block)) {
size_t remaining = len;
while (remaining > 0) {
// kMaxBytesPerCall is only 2GB, while ChaCha can handle 256GB. But this
// is sufficient and easier on 32-bit.
static const size_t kMaxBytesPerCall = 0x80000000;
size_t todo = remaining;
if (todo > kMaxBytesPerCall) {
todo = kMaxBytesPerCall;
}
CRYPTO_chacha_20(buf, buf, todo, state->key,
(uint8_t *)&state->calls_used, 0);
buf += todo;
remaining -= todo;
state->calls_used++;
}
} else {
if (sizeof(state->partial_block) - state->partial_block_used < len) {
CRYPTO_chacha_20(state->partial_block, state->partial_block,
sizeof(state->partial_block), state->key,
(uint8_t *)&state->calls_used, 0);
state->partial_block_used = 0;
}
unsigned i;
for (i = 0; i < len; i++) {
buf[i] ^= state->partial_block[state->partial_block_used++];
}
state->calls_used++;
}
state->bytes_used += len;
return 1;
}
