The rollup 487-502 over v1.0-build486. No server release: the server stays
srtp-build478.
-
A relay that answers TURN and does not carry the handshake to the server
(contributed in #110, with follow-ups here).
Behind a peer-handshake timeout one connection tries another relay; when
the server answers there, the first relay is marked for the whole pool.
A mark needs that evidence from another relay, leaves live sessions
alone, and an attempt cut by a suspension is not judged. Recovery checks
return the relay (1 to 30 minutes apart, taken at a re-dial); a path
change clears every mark. The bootstrap's connection is not handed a
recovery check while another relay can seat it; a cleared mark is said
in the log; a run of short failures ends at a session that came up. -
A cold start has a second relay. The seed carries every relay VK named
with its credential, and in cookie mode the seed's credential is a slot
on each relay of its call. With the first relay dead or not forwarding
the tunnel comes up through the second without asking VK anything first. -
A relay that REFUSES the permission for our peer (an own relay whose
peer list lacks the server) is marked at once, on its own answer - never
on a 400 or a 437, which mean a changed mapping. Every transport tells
it: the explicit permission, the one asked inside the first write
(WRAP-A), and csqtt of both generations. The mark is cleared by a
session that came up through the relay. Before, such a tunnel did not
come up at all. -
A session whose TURN client has stopped reading is ended at once. The
client's read loop ends on any read error - a datagram refused while a
relay process restarts, a TCP connection the relay closed - and never
reads again; such a session stood deaf until a liveness rule found it,
and WRAP-A had none of its own. What follows for an OWN relay restarted
under the tunnel: over TCP its sessions end in that second, their
re-dials are refused while the relay is down, the relay is marked, and
the tunnel moves to the other relays of the list until a path change.
Over UDP that happens only if a datagram falls into the gap (WRAP-A);
otherwise the sessions end on the relay's 400 at their next permission
refresh and return to it - which wants a relay that answers such a
request: tools/turn-relay on the server repository's branch does, the
one of srtp-build478 stays silent. -
The inner traffic line: a switch in Settings > Advanced > Logging, off
by default. Once a minute the tunnel says what the phone's own traffic
meets inside it: TCP connections asked, answered, refused and silent
(the silent ones by host count, address family and port class);
established connections that sent data, heard, stalled 10 s or more,
re-sent; DNS; UDP flows never answered. Counts from packet headers
only - no address, no name. -
csqtt: the start says what the server's name resolved to and how long
the lookup took; the 2.5 engine says once a minute how the datagrams on
the wire were spaced.
Measured on the phone. First relay blackholed, a cold cache: forty
connections up 12.6 s behind the start. First relay accepting TURN and not
forwarding: about 33 s (the handshake's deadline, the bootstrap's pause,
and a second wave of connections that still meet the first relay). An own
relay refusing the permission: up through another relay 12 s behind the
start; its list mended, one recovery check clears the mark and the next
re-dials are back on it. An own relay restarted under the tunnel: native
over TCP - all forty on another relay 1 to 7 s later; native and csqtt 2.5
over UDP - back on the own relay 40 to 50 s later, at the permission
refresh; csqtt 2.1.9 over UDP - 75 s later, by its deafness rule.