Skip to content

Commit

Permalink
Fix NTP sync functionality
Browse files Browse the repository at this point in the history 
The existing NTP synchronization has a few issues. It only waits 3
milliseconds for a response from the iOS device, which is too short of a
delay. Also, if it fails to send the NTP request for some reason, or it
fails to receive the NTP response in time, it stops attempting NTP
synchronization altogether.

This commit attempts to address these problems. The NTP response time
has been increased to 300 ms, and if we fail to send or receive an NTP
packet, we will try again next time. The NTP socket is flushed prior to
each request to ensure we don't read an older delayed response.
  • Loading branch information
@dougg3 @FD-
dougg3 authored and FD- committed Sep 13, 2020
1 parent 576c5a6 commit 5900542
Showing 1 changed file with 73 additions and 54 deletions.
127 changes: 73 additions & 54 deletions lib/raop_ntp.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -202,7 +202,7 @@ raop_ntp_init_socket(raop_ntp_t *raop_ntp, int use_ipv6)
// We're calling recvfrom without knowing whether there is any data, so we need a timeout
struct timeval tv;
tv.tv_sec = 0;
tv.tv_usec = 3000;
tv.tv_usec = 300000;
if (setsockopt(tsock, SOL_SOCKET, SO_RCVTIMEO, &tv, sizeof(tv)) < 0) {
goto sockets_cleanup;
}
Expand All @@ -219,6 +219,23 @@ raop_ntp_init_socket(raop_ntp_t *raop_ntp, int use_ipv6)
return -1;
}

static void
raop_ntp_flush_socket(int fd)
{
int bytes_available = 0;
while (ioctl(fd, FIONREAD, &bytes_available) == 0 && bytes_available > 0)
{
// We are guaranteed that we won't block, because bytes are available.
// Read 1 byte. Extra bytes in the datagram will be discarded.
char c;
int result = recvfrom(fd, &c, sizeof(c), 0, NULL, NULL);
if (result < 0)
{
break;
}
}
}

static THREAD_RETVAL
raop_ntp_thread(void *arg)
{
Expand All @@ -240,6 +257,9 @@ raop_ntp_thread(void *arg)
}
MUTEX_UNLOCK(raop_ntp->run_mutex);

// Flush the socket in case a super delayed response arrived or something
raop_ntp_flush_socket(raop_ntp->tsock);

// Send request
uint64_t send_time = raop_ntp_get_local_time(raop_ntp);
byteutils_put_ntp_timestamp(request, 24, send_time);
Expand All @@ -248,60 +268,59 @@ raop_ntp_thread(void *arg)
logger_log(raop_ntp->logger, LOGGER_DEBUG, "raop_ntp send_len = %d", send_len);
if (send_len < 0) {
logger_log(raop_ntp->logger, LOGGER_ERR, "raop_ntp error sending request");
break;
} else {
// Read response
response_len = recvfrom(raop_ntp->tsock, (char *)response, sizeof(response), 0,
(struct sockaddr *) &raop_ntp->remote_saddr, &raop_ntp->remote_saddr_len);
if (response_len < 0) {
logger_log(raop_ntp->logger, LOGGER_ERR, "raop_ntp receive timeout");
} else {
logger_log(raop_ntp->logger, LOGGER_DEBUG, "raop_ntp receive time type_t packetlen = %d", response_len);

int64_t t3 = (int64_t) raop_ntp_get_local_time(raop_ntp);
// Local time of the client when the NTP request packet leaves the client
int64_t t0 = (int64_t) byteutils_get_ntp_timestamp(response, 8);
// Local time of the server when the NTP request packet arrives at the server
int64_t t1 = (int64_t) byteutils_get_ntp_timestamp(response, 16);
// Local time of the server when the response message leaves the server
int64_t t2 = (int64_t) byteutils_get_ntp_timestamp(response, 24);

// The iOS device sends its time in micro seconds relative to an arbitrary Epoch (the last boot).
// For a little bonus confusion, they add SECONDS_FROM_1900_TO_1970 * 1000000 us.
// This means we have to expect some rather huge offset, but its growth or shrink over time should be small.

raop_ntp->data_index = (raop_ntp->data_index + 1) % RAOP_NTP_DATA_COUNT;
raop_ntp->data[raop_ntp->data_index].time = t3;
raop_ntp->data[raop_ntp->data_index].offset = ((t1 - t0) + (t2 - t3)) / 2;
raop_ntp->data[raop_ntp->data_index].delay = ((t3 - t0) - (t2 - t1));
raop_ntp->data[raop_ntp->data_index].dispersion = RAOP_NTP_R_RHO + RAOP_NTP_S_RHO + (t3 - t0) * RAOP_NTP_PHI_PPM / 1000000u;

// Sort by delay
memcpy(data_sorted, raop_ntp->data, sizeof(data_sorted));
qsort(data_sorted, RAOP_NTP_DATA_COUNT, sizeof(data_sorted[0]), raop_ntp_compare);

uint64_t dispersion = 0ull;
int64_t offset = data_sorted[0].offset;
int64_t delay = data_sorted[RAOP_NTP_DATA_COUNT - 1].delay;

// Calculate dispersion
for(int i = 0; i < RAOP_NTP_DATA_COUNT; ++i) {
unsigned long long disp = raop_ntp->data[i].dispersion + (t3 - raop_ntp->data[i].time) * RAOP_NTP_PHI_PPM / 1000000u;
dispersion += disp / two_pow_n[i];
}

MUTEX_LOCK(raop_ntp->sync_params_mutex);

int64_t correction = offset - raop_ntp->sync_offset;
raop_ntp->sync_offset = offset;
raop_ntp->sync_dispersion = dispersion;
raop_ntp->sync_delay = delay;
MUTEX_UNLOCK(raop_ntp->sync_params_mutex);

logger_log(raop_ntp->logger, LOGGER_DEBUG, "raop_ntp sync correction = %lld", correction);
}
}

// Read response
response_len = recvfrom(raop_ntp->tsock, (char *)response, sizeof(response), 0,
(struct sockaddr *) &raop_ntp->remote_saddr, &raop_ntp->remote_saddr_len);
if (response_len < 0) {
logger_log(raop_ntp->logger, LOGGER_ERR, "raop_ntp receive timeout");
break;
}
logger_log(raop_ntp->logger, LOGGER_DEBUG, "raop_ntp receive time type_t packetlen = %d", response_len);

int64_t t3 = (int64_t) raop_ntp_get_local_time(raop_ntp);
// Local time of the client when the NTP request packet leaves the client
int64_t t0 = (int64_t) byteutils_get_ntp_timestamp(response, 8);
// Local time of the server when the NTP request packet arrives at the server
int64_t t1 = (int64_t) byteutils_get_ntp_timestamp(response, 16);
// Local time of the server when the response message leaves the server
int64_t t2 = (int64_t) byteutils_get_ntp_timestamp(response, 24);

// The iOS device sends its time in micro seconds relative to an arbitrary Epoch (the last boot).
// For a little bonus confusion, they add SECONDS_FROM_1900_TO_1970 * 1000000 us.
// This means we have to expect some rather huge offset, but its growth or shrink over time should be small.

raop_ntp->data_index = (raop_ntp->data_index + 1) % RAOP_NTP_DATA_COUNT;
raop_ntp->data[raop_ntp->data_index].time = t3;
raop_ntp->data[raop_ntp->data_index].offset = ((t1 - t0) + (t2 - t3)) / 2;
raop_ntp->data[raop_ntp->data_index].delay = ((t3 - t0) - (t2 - t1));
raop_ntp->data[raop_ntp->data_index].dispersion = RAOP_NTP_R_RHO + RAOP_NTP_S_RHO + (t3 - t0) * RAOP_NTP_PHI_PPM / 1000000u;

// Sort by delay
memcpy(data_sorted, raop_ntp->data, sizeof(data_sorted));
qsort(data_sorted, RAOP_NTP_DATA_COUNT, sizeof(data_sorted[0]), raop_ntp_compare);

uint64_t dispersion = 0ull;
int64_t offset = data_sorted[0].offset;
int64_t delay = data_sorted[RAOP_NTP_DATA_COUNT - 1].delay;

// Calculate dispersion
for(int i = 0; i < RAOP_NTP_DATA_COUNT; ++i) {
unsigned long long disp = raop_ntp->data[i].dispersion + (t3 - raop_ntp->data[i].time) * RAOP_NTP_PHI_PPM / 1000000u;
dispersion += disp / two_pow_n[i];
}

MUTEX_LOCK(raop_ntp->sync_params_mutex);

int64_t correction = offset - raop_ntp->sync_offset;
raop_ntp->sync_offset = offset;
raop_ntp->sync_dispersion = dispersion;
raop_ntp->sync_delay = delay;
MUTEX_UNLOCK(raop_ntp->sync_params_mutex);

logger_log(raop_ntp->logger, LOGGER_DEBUG, "raop_ntp sync correction = %lld", correction);

// Sleep for 3 seconds
struct timeval now;
struct timespec wait_time;
Expand Down Expand Up @@ -443,4 +462,4 @@ uint64_t raop_ntp_convert_local_time(raop_ntp_t *raop_ntp, uint64_t local_time)
uint64_t offset = raop_ntp->sync_offset;
MUTEX_UNLOCK(raop_ntp->sync_params_mutex);
return (uint64_t) ((int64_t) local_time) + ((int64_t) offset);
}
}

0 comments on commit 5900542

Please sign in to comment.