[Asterisk-code-review] res rtp asterisk: RTT miscalculation in RTCP (asterisk[master])
Richard Mudgett
asteriskteam at digium.com
Wed Nov 23 10:15:43 CST 2016
Richard Mudgett has uploaded a new change for review. ( https://gerrit.asterisk.org/4491 )
Change subject: res_rtp_asterisk: RTT miscalculation in RTCP
......................................................................
res_rtp_asterisk: RTT miscalculation in RTCP
When retrieving RTCP stats for PJSIP channels, RTT values are unreliable.
RTT calculation is correct, but the data representation isn't. RTT is
represented by a 32-bit fixed-point number with the integer part in the
first 16 bits and the fractional part in the last 16 bits. In order to
get the RTT value, the fractional part is miscalculated, there is an
unnecessary 16 bit shift that causes overflow. Besides this there is
another mistake, when transforming the integer value to the fixed point
fractional part via bitwise operation, that loses precision.
* RTT fractional part is no longer shifted, avoiding overflow.
* RTT fractional part is transformed to its fixed-point value more
precisely.
* Fixed timeval2ntp() and ntp2timeval() second fraction conversions.
* Fixed NTP timestamp report logging. The usec was inexplicably
multiplied by 4096.
ASTERISK-26566 #close
Reported by Hector Royo Concepcion
Change-Id: Ie09bdabfee75afb3f1b8ddfd963e5219ada3b96f
---
M main/rtp_engine.c
M res/res_rtp_asterisk.c
2 files changed, 42 insertions(+), 9 deletions(-)
git pull ssh://gerrit.asterisk.org:29418/asterisk refs/changes/91/4491/1
diff --git a/main/rtp_engine.c b/main/rtp_engine.c
index 4fc1414..3ef2d87 100644
--- a/main/rtp_engine.c
+++ b/main/rtp_engine.c
@@ -2493,7 +2493,7 @@
if (type == AST_RTP_RTCP_SR) {
ast_str_append(&packet_string, 0, "SentNTP: %lu.%06lu\r\n",
(unsigned long)payload->report->sender_information.ntp_timestamp.tv_sec,
- (unsigned long)payload->report->sender_information.ntp_timestamp.tv_usec * 4096);
+ (unsigned long)payload->report->sender_information.ntp_timestamp.tv_usec);
ast_str_append(&packet_string, 0, "SentRTP: %u\r\n",
payload->report->sender_information.rtp_timestamp);
ast_str_append(&packet_string, 0, "SentPackets: %u\r\n",
diff --git a/res/res_rtp_asterisk.c b/res/res_rtp_asterisk.c
index a882033..4136f7e 100644
--- a/res/res_rtp_asterisk.c
+++ b/res/res_rtp_asterisk.c
@@ -3084,7 +3084,26 @@
unsigned int sec, usec, frac;
sec = tv.tv_sec + 2208988800u; /* Sec between 1900 and 1970 */
usec = tv.tv_usec;
- frac = (usec << 12) + (usec << 8) - ((usec * 3650) >> 6);
+ /*
+ * Convert usec to 0.32 bit fixed point without overflow.
+ *
+ * = usec * 2^32 / 10^6
+ * = usec * 2^32 / (2^6 * 5^6)
+ * = usec * 2^26 / 5^6
+ *
+ * The usec value needs 20 bits to represent 999999 usec. So
+ * splitting the 2^26 to get the most precision using 32 bit
+ * values gives:
+ *
+ * = ((usec * 2^12) / 5^6) * 2^14
+ *
+ * Splitting the division into two stages preserves all the
+ * available significant bits of usec over doing the division
+ * all at once.
+ *
+ * = ((((usec * 2^12) / 5^3) * 2^7) / 5^3) * 2^7
+ */
+ frac = ((((usec << 12) / 125) << 7) / 125) << 7;
*msw = sec;
*lsw = frac;
}
@@ -3092,7 +3111,8 @@
static void ntp2timeval(unsigned int msw, unsigned int lsw, struct timeval *tv)
{
tv->tv_sec = msw - 2208988800u;
- tv->tv_usec = ((lsw << 6) / 3650) - (lsw >> 12) - (lsw >> 8);
+ /* Reverse the sequence in timeval2ntp() */
+ tv->tv_usec = ((((lsw >> 7) * 125) >> 7) * 125) >> 12;
}
static void calculate_lost_packet_statistics(struct ast_rtp *rtp,
@@ -3277,9 +3297,9 @@
ast_sockaddr_stringify(&remote_address), ice ? " (via ICE)" : "");
ast_verbose(" Our SSRC: %u\n", rtcp_report->ssrc);
if (sr) {
- ast_verbose(" Sent(NTP): %u.%010u\n",
+ ast_verbose(" Sent(NTP): %u.%06u\n",
(unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_sec,
- (unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_usec * 4096);
+ (unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_usec);
ast_verbose(" Sent(RTP): %u\n", rtcp_report->sender_information.rtp_timestamp);
ast_verbose(" Sent packets: %u\n", rtcp_report->sender_information.packet_count);
ast_verbose(" Sent octets: %u\n", rtcp_report->sender_information.octet_count);
@@ -4016,9 +4036,22 @@
lsr_a = ((msw & 0x0000ffff) << 16) | ((lsw & 0xffff0000) >> 16);
rtt = lsr_a - lsr - dlsr;
rtt_msw = (rtt & 0xffff0000) >> 16;
- rtt_lsw = (rtt & 0x0000ffff) << 16;
+ rtt_lsw = (rtt & 0x0000ffff);
rtt_tv.tv_sec = rtt_msw;
- rtt_tv.tv_usec = ((rtt_lsw << 6) / 3650) - (rtt_lsw >> 12) - (rtt_lsw >> 8);
+ /*
+ * Convert 16.16 fixed point rtt_lsw to usec without
+ * overflow.
+ *
+ * = rtt_lsw * 10^6 / 2^16
+ * = rtt_lsw * (2^6 * 5^6) / 2^16
+ * = rtt_lsw * 5^6 / 2^10
+ *
+ * The rtt_lsw value is in 16.16 fixed point format and 5^6
+ * requires 14 bits to represent. We have enough space to
+ * directly do the conversion because there is no integer
+ * component in rtt_lsw.
+ */
+ rtt_tv.tv_usec = (rtt_lsw * 15625) >> 10;
rtp->rtcp->rtt = (double)rtt_tv.tv_sec + ((double)rtt_tv.tv_usec / 1000000);
if (lsr_a - dlsr < lsr) {
return 1;
@@ -4214,9 +4247,9 @@
&rtcp_report->sender_information.ntp_timestamp);
rtcp_report->sender_information.rtp_timestamp = ntohl(rtcpheader[i + 2]);
if (rtcp_debug_test_addr(&addr)) {
- ast_verbose("NTP timestamp: %u.%010u\n",
+ ast_verbose("NTP timestamp: %u.%06u\n",
(unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_sec,
- (unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_usec * 4096);
+ (unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_usec);
ast_verbose("RTP timestamp: %u\n", rtcp_report->sender_information.rtp_timestamp);
ast_verbose("SPC: %u\tSOC: %u\n",
rtcp_report->sender_information.packet_count,
--
To view, visit https://gerrit.asterisk.org/4491
To unsubscribe, visit https://gerrit.asterisk.org/settings
Gerrit-MessageType: newchange
Gerrit-Change-Id: Ie09bdabfee75afb3f1b8ddfd963e5219ada3b96f
Gerrit-PatchSet: 1
Gerrit-Project: asterisk
Gerrit-Branch: master
Gerrit-Owner: Richard Mudgett <rmudgett at digium.com>
Gerrit-Reviewer: Hector Royo Concepcion <hectorroyo92 at gmail.com>
More information about the asterisk-code-review
mailing list