stats->ext_seq has been the SRTP receive index since srtp_index moved into
shm (08c7f5853d). The transcode sequencer overwrote it after every
emitted packet with its own roc<<16|seq, which lags the received packets
and, whenever the sequencer emits a late or replayed packet with a lower
sequence number, carries a roc that was incremented without a real
wrap-around. packet_index() then estimated the wrong index, and once the
sender's real roc reached 1 none of the fallback guesses matched, so every
packet of the stream failed authentication until the call ended.
packet_index() already maintains the index for every received packet, so
drop the write. Also make the sequencer's own roc increment only on a
forward wrap-around, never on a backwards step.
Closes#2221
Change-Id: Ia38bf8cbc63dd3435717927f1adfd81ac94ea347
(cherry picked from commit 956d7add09)
(cherry picked from commit ff56b32f81)
test6: lost / received / lost with forced emissions; the second forced
emission must return the lower queued packet, not the one that arrived
last. test7: a packet exactly one ring size ahead across the 65535->0
wrap-around must be handled as a seq reset.
Both tests fail on the previous sequencer code.
Change-Id: I47c5d0b961856255cb204fbcf2e656cf01b49d7d
(cherry picked from commit eaa2b75c42)
(cherry picked from commit 953ac7cb63)
Treat a packet exactly one ring size ahead across the 65535->0
wrap-around as a seq reset, like the non-wrapped case, instead of
accepting it into the head slot.
Fixes#2220
Change-Id: I94e2fdd3d50d5c8241cfed5da06829097566356e
(cherry picked from commit fc703ce91c)
(cherry picked from commit fa598a070c)
Ref #2211
cli_handle() takes a str, but most "set" handlers parsed their value
with strtol(), strtod() or atoi() on instr->s and printed it with "%s",
reading up to the next NUL. The ng "cli" command passes a slice of the
request: "set maxsessions 1000" set 10007, "set rediscmdtimeout 0" set
7 ms, and error messages quoted the rest of the request.
Parse with the str_to_X helpers, plus a new str_to_d for maxcpu and
maxload, and print with STR_FORMAT. As with the str_to_X family, a
value with trailing characters ("10x") is refused, and so is a
non-number in the seven handlers that read it as 0. "terminate" now
matches all, own and foreign exactly (str_cmp()), not as a prefix.
Tested by auto-daemon-tests-redis-errors.pl and the new t/test-str.c.
Change-Id: I92ed33a4229688aba513c3d0410b9d77a8bcd516
(cherry picked from commit f54da0bcec)
(cherry picked from commit 56c61e88d1)
Use epoll's u64 field
Change-Id: If56e980b2d65cc0e32736c58a12cad1000628f2a
(cherry picked from commit c6508f4fd9)
(cherry picked from commit 11ed687e35)
Currently unused and transparent
Change-Id: I35604c9d5f724ae6a3bce714d708bed0fbed563f
(cherry picked from commit b2cf472a20)
(cherry picked from commit ded8f0033f)
Ref #2211
When json_restore_call() fails, it sends "DEL <call-id>" twice: on the
write connection, in its own DB, and on the notification connection, in
whatever DB the last keyspace notification selected. For a foreign call
(a keyspace notification, or the restore of a subscribed keyspace at
startup) that key belongs to the instance that owns the call, which
keeps writing it: a standby that fails to restore one record deletes the
owner's copy, and the call cannot be taken over any more. For an own
call, the DEL on the notification connection can hit a peer's DB.
Only delete an unrestorable record when the call is our own, and only on
the write connection, in our own DB. A foreign record that can never be
restored now stays until its owner rewrites or deletes it, or until it
expires (--redis-expires). Seen in a lab with Redis Sentinel, where a
subscriber's DEL was refused only because its node had just become a
replica.
Tests, in auto-daemon-tests-redis-subscribe.pl: a keyspace notification
for a foreign call whose record cannot be parsed is followed by no
command on either connection, and an own record that cannot be parsed at
startup is deleted once, on the write connection only. Without this
change both get a DEL on both connections.
Change-Id: I5d8a53f3515596cf3fc2c58ea97847476385b876
(cherry picked from commit 6b3bf8299b)
(cherry picked from commit bb1901ef0a)
Ref #2211
redis_set_timeout() receives the timeout in microseconds (the
--redis-cmd-timeout value times 1000) and logged it unconverted, so
--redis-cmd-timeout=1000 was reported as:
Setting timeout for Redis commands to 1000000 milliseconds
Log the value in milliseconds, as the message says. The value passed to
hiredis is unchanged.
No functional change. Seen in the log of every connection made with a
command timeout set; there is no test, as no test reads this line.
Change-Id: Ie4d81e4f8baab1969210a589c98245407b088561
(cherry picked from commit 1c447b509b)
(cherry picked from commit 6e8ffd4499)
Ref #2211
redis_connect() finds "role:" in the INFO reply and compares what
follows with memcmp(): 11 bytes for "role:master", 19 for
"role:active-replica" and 10 for "role:slave". memcmp() may read all n
bytes, so a reply that ends less than 19 bytes after "role:" (for
example a replica's "role:slave\r\n" with nothing after it) is read
past its end. Real Redis and KeyDB servers send many more lines after
the role, so this does not show in practice, but AddressSanitizer
aborts on it (heap-buffer-overflow in the memcmp interceptor) with a
minimal reply, such as one from a test server or a proxy.
Use strncmp(): hiredis NUL-terminates the reply string, so the compare
stops at its end.
Tested by auto-daemon-tests-redis-errors.pl, added later in this
series, which answers a reconnect's INFO with just "role:slave": an
ASan build aborts there without this change. The change is early in
the series so that it can be taken on its own.
Change-Id: Icdf3d3a23464d861f57406436328d0b166794362
(cherry picked from commit e5ff3ea8af)
(cherry picked from commit bb58e6a8f3)
Ref #2211
"set rediscmdtimeout 0" reconnects the Redis connections, but each of
the three calls is guarded by rtpe_redis instead of its own pointer:
- with --redis and no notification connection (no --redis-subscribe and
no subscribed keyspace), redis_reconnect(rtpe_redis_notify) locks a
NULL mutex and rtpengine crashes;
- with --redis-write and no --redis, the write connection is never
reconnected, so its command timeout is never cleared.
Guard each call with its own pointer, as the non-zero branch already
does.
Tested by auto-daemon-tests-redis-errors.pl, added later in this
series ("set rediscmdtimeout 0" over the ng "cli" command with --redis
only): without this change rtpengine crashes and never replies. The
change is first in the series so that it can be taken on its own.
Change-Id: Ia458a6eb0b70a6906ab410fd91087ad5af581f3b
(cherry picked from commit 20bcdce73c)
(cherry picked from commit b789cab2c4)
monologue_publish() never copied the offered label into the media,
so sdp_create() omitted a=label from the publish answer and from
subscribe requests for that publisher. Add the missing
media_update_label() call.
Closes#2200
Change-Id: Id9aaf4ab09c21f98d3dcaf78632413b99f4a97ff
(cherry picked from commit b66c59bed6)
The kernel tracks lost packets per SSRC through the `lost_bits` window
in the shared ssrc_stats, which was left zeroed, while userspace starts
its own window as all-received. As userspace already advances `ext_seq`
before a stream is kernelised, the kernel then counted up to 32 packets
it did receive as lost, both per SSRC and in the interface and global
packets_lost counters. Start the shared window as all-received too.
Change-Id: I3d6cd85a1b764a6bb3296daf7489091767449554
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 3b73432826)
ssrc_collect_metrics() converted the running RFC 3550 jitter estimate
to milliseconds in place on every call timer run with --measure-rtp,
dividing it by clock_rate/1000 each time. The estimator is kept in
RTP timestamp units and also feeds generated RTCP RRs, which therefore
reported a jitter decaying towards zero. Convert a copy instead.
The RTCP test expected the resulting zero jitter; it now gets the
value actually measured from its back-to-back packets.
Change-Id: Ice29029322764fd76374eb3f7dda517570f7f75b
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 4dd1826ffa)
Copy create-answer and subscribe-answer SDP into call-owned storage before parsing, matching offer/answer and publish. Retained stream parameters otherwise reference released NG request buffers and later subscriptions can lose their media type and codecs.
Initialize created-answer monologue session state so generated subscriptions retain a valid SDP origin and session name. Add request-buffer-churn and decoded-audio regressions for created and subscribed endpoints.
Validation on standalone Sipwise master: daemon build, all 89 retained-SDP assertions, all 652 create assertions, and all 932 publish/subscribe assertions pass. The cumulative 26.2.1 release also has isolated browser Opus decoding, privacy, recording, and mixer regression coverage.
Closes#2194
Change-Id: Id8ca0e0ff691356d905ee94c0f41090f973fc70f
(cherry picked from commit 12ecdbc2cf)
Update media player, send timer, jitter buffer objects.
Set final memory arena pointer as the releasing the last reference might
reset it to null.
Exclude media_player stuff from no-transcoding builds and adjust ifdef
guards accordingly.
Closes#2186
Change-Id: I42dc5a449816d765bc4d1da2b35315e0b6f7c29b
(cherry picked from commit 2366392202)
Share SSRC dump and restore helpers between ingress and egress, keeping
ssrc_out_table as the outbound list name. Scope each hash lock to its
dump rather than holding both locks throughout media serialization.
Restore records without an outbound list through the original combined
in/out path, including records written before the SSRC hash split.
Adapt the takeover regression test from kbinetruy's PR #2188 to exercise
native and JSON records, sequence rollover, encryption after restart,
and both legacy record layouts. Follow the existing Redis helper,
initializer, iteration, and daemon-test patterns.
Ref #2185
Change-Id: I0ac7e6b351fd3e2bcab38d0318d1787e5332c442
Co-authored-by: Kevin Binetruy <kevin.binetruy@ovhcloud.com>
(cherry picked from commit afb98f00d3)
Since the SSRC contexts were split into separate ingress and egress
entries (cf22d829), only the ingress entries are written to the SSRC
table, and the lines for the outbound index are commented out. The
restore still looks for the out_* keys in the ingress entries, where
they no longer are.
A node that takes over a call therefore starts encrypting from index 0
again. A receiver estimates the ROC from its own state, so if the
sequence number of the call has already wrapped, it rejects every
packet after the takeover.
Write the egress entries to a list of their own, ssrc_out_table, and
restore them from there. Records without the list restore as before.
Fixes#2183Closes#2185
Change-Id: Id28dc86c658f7e94dd87443fc0f09225beaf484e
(cherry picked from commit ef16e3f1e7)
After a takeover, FOREIGN_MEDIA holds off the media timeout until the
call receives media. The flag is cleared by the first packet that
arrives on any of the call's sockets, before the packet is checked, so
a packet that is then dropped, for example because of strict-source,
ends the grace period without refreshing the stream's last packet time.
A call that was restored longer than the timeout ago is then closed on
the next timer run, before its real media has had a chance to arrive.
With the peer address confirmation now restored from Redis, this is
what happens to an adopted strict-source call when a stray packet
arrives ahead of the peer's first one.
Clear the flag only once a packet has passed the address check. The
timer still clears it for kernel-forwarded media, and the grace period
stays bounded by the timeout after the takeover.
Ref #2184
Change-Id: Ib06a170d271e8aea709abc57cbfd3bcc59196ab5
(cherry picked from commit 1d783ba1b9)
A call restored from Redis starts with every sfd unconfirmed and every
stream's el_flags at zero, which is EL_DELAYED, regardless of what the
offer asked for or what --endpoint-learning is set to. After a
keyspace-notification failover, all adopted calls therefore re-learn
their peer at the same time, and strict-source has nothing to compare
against until they have.
Write the sfd's confirmed flag and the stream's el_flags and
learned_endpoint to the record, and read them back on restore. The two
stream fields were so far only part of the rollback snapshot. Move them
into the shared stream encoder and decoder, so that both use the same
definition. Records without the new keys restore as before.
Fixes#2182Closes#2184
Change-Id: I6228e55796e7a680b05797df985573e8ab54c653
(cherry picked from commit bac92fbb5e)
Both mean that the ice towards the offerer gets rejected.
When a certain peer who previously offered ice, sends SDP without ice
(e.g. T.38 re-invite), or the ice was requested to be removed
(when processing answer) with `ICE=remove` then the ice agent is left alive
with its learned ufrag, candidates and succeeded pairs.
The stream endpoint is then taken from that SDP (non-ice path).
If this peer later offers ice again with the same credentials (e.g.
falling back to audio after an error response like 488 for that T.38),
then this stale agent is reused:
`__fill_stream()` ignores the new endpoint because the ice data
considered as unchanged, and then the `ice_update()` drops these candidates as duplicates.
Therefore no connectivity checks run and then the endpoint stays at the old non-iced address.
Media gets bound to the wrong port.
Shutdown the ice agent whenever the peer's own sdp no longer carries ice,
or we get the `ICE=remove`.
Then the offered ice again later starts from scratch.
Change-Id: I5ff3b6454262099022bf4f61c84b978835130dcc
(cherry picked from commit 39fd49b6c4)
Update the sockaddr_t to an endpoint_t and change the generic pointer to
a more type-safe argument
Closes#2176Closes#2178
Change-Id: Ib734079051a2a8916eeae9942df3ea53a1637985