## Summary
Flips the Linux default for `--udp-recvmmsg` from **off** to **on**.
Operators opt out with `--udp-recvmmsg=false` (or `=0`).
> **Stacked on #1929.** This depends on the recvmmsg-scoping change in
#1929 and is based on that branch, so the diff shows only the default-on
change. GitHub will auto-retarget the base to `master` once #1929
merges. Merge #1929 first.
## Why this is now safe
The original objection to default-on (recorded in
`docs/PerformanceIterationLog.md`) was the **per-session-relay-socket
prealloc tax**: `--udp-recvmmsg` applied the 16-buffer batch path to
every connected relay socket, which only ever carries one flow, so the
churn ate the listener-side win.
#1929 scoped recvmmsg to **shared fan-in sockets only**
(`udp_recvmmsg_eligible`: the client listener, plus the per-thread
shared relay socket under `--multiplex-peer`). Per-session relay sockets
now stay on the single-recv path regardless of the flag, so that tax is
gone. The one socket touched by default — the client listener — is a
genuine fan-in point:
- batches whenever client concurrency is non-trivial (measured
`avg_batch ≈ 16` under load), and
- costs little when idle (few packets ⇒ few prealloc cycles).
## What changed
- `mainrelay.c`: `turn_params.udp_recvmmsg` default `false → true`
(Linux only).
- Removed the now-dead `--multiplex-peer` auto-enable block and the
`udp_recvmmsg_set_explicitly` tracking it relied on; multiplex-peer gets
its recvmmsg window from the default. The opt-out flows through the
normal `get_bool_value` path.
- Help text, `man/man1/turnserver.1`, `examples/etc/turnserver.conf`,
`CLAUDE.md`, and `docs/PerformanceIterationLog.md` updated for the new
default + opt-out.
Per-session relay sockets and DTLS session sockets are unchanged.
## Validation
- **Format:** clang-format 15.0.7 clean.
- **macOS:** build + ctest 6/6 + `run_tests.sh` pass.
- **Linux (Docker, clean build):** ctest 5/5; `run_tests.sh`,
`run_tests_conf.sh`, `run_tests_multiplex_peer.sh` all pass (no FAIL).
- **Runtime proof (loopback, `--udp-recvmmsg-log`):**
- Default, no flag: recvmmsg active, `calls=13714 packets=219306
avg_batch=15.99`.
- `--udp-recvmmsg=false`: zero recvmmsg activity — opt-out confirmed.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Summary
Adds a Linux-only `--udp-sendmmsg-log` flag (mirroring
`--udp-recvmmsg-log`) that logs per-relay-thread **egress** batch
statistics every 10 s: flush count, total datagrams, average batch
occupancy, UDP-GSO engagement
(`gso_flushes`/`gso_datagrams`/`gso_frac`), and a per-flush occupancy
histogram.
## Motivation
`--multiplex-peer` enables `sendmmsg`/UDP-GSO coalescing on the egress
path, but there was no way to see whether it actually coalesces
anything. While investigating this I confirmed a non-obvious property
worth documenting:
- Per-session UDP **client** sockets are children of the listener
(`parent_s`), and `udp_send_fd()` returns the shared **listener fd** for
all of them. Combined with the `recvmmsg`-driven batch window,
relay→client **downlink** sends to *different clients* already coalesce
into one `sendmmsg` on the listener fd — the listener fd is effectively
a shared client-facing send socket. (In non-multiplex mode each
allocation has its own relay socket, so a `recvmmsg` drain spans only
one session and the downlink batch is a singleton — cross-client
batching genuinely requires `--multiplex-peer`.)
- UDP-GSO only engages when destination **and** segment size match
across a batch, so at low per-flow packet rates (VoIP-style, tens of pps
per flow) it rarely fires.
This flag makes both effects measurable instead of assumed.
## Example
Captured under a `--multiplex-peer --udp-gso` load:
```
udp-sendmmsg stats: flushes=21 datagrams=27 avg_batch=1.29 gso_flushes=0 \
gso_datagrams=0 gso_frac=0.000 hist_1=17 hist_2=2 hist_3_4=2 hist_5_8=0 hist_9_16=0 hist_17_32=0
```
- `avg_batch` — mean datagrams per flush (1.0 = no coalescing)
- `gso_frac` — fraction of datagrams sent via UDP-GSO (~0 = GSO not
earning its keep)
- `hist_*` — per-flush occupancy histogram
Both rise with aggregate pps and sit near the floor on lightly loaded
servers.
## Implementation
- Counters on `ioa_engine` (behind `#if defined(__linux__)`), bumped
once per flush in `udp_sendmmsg_flush()` and once per `recvmmsg` call —
no per-datagram cost.
- New `--udp-sendmmsg-log` CLI flag + 10 s periodic logger in the engine
timer, mirroring the existing recvmmsg stats path.
- Docs: a new "Egress Batching (sendmmsg / UDP-GSO) and Observability"
section in `docs/multiplex-peer.md`.
## Testing
- macOS: build + `ctest` (6/6) + `run_tests.sh` pass (instrumentation is
`#if __linux__`, validating flag plumbing + the non-Linux stub path).
- Linux (clean Ubuntu 24.04 Docker): build + `ctest` (4/4) +
`run_tests.sh` / `run_tests_conf.sh` / `run_tests_multiplex_peer.sh`
pass, and a `--multiplex-peer --udp-gso --udp-sendmmsg-log` load run
emitted the stats line above.
## Summary
Adds **`--multiplex-peer`**, a non-standard relay mode that replaces the
per-allocation peer-side port bind with **one shared IPv4+IPv6 UDP
socket pair per relay thread**. Sessions are demultiplexed by exact peer
IP:port in a per-thread `mp_table`. This lifts the ~16 k allocation cap
that the default 49152-65535 relay port range imposes, and dramatically
reduces kernel-level UDP receive-buffer drops under high pps.
Design and trade-offs: [docs/multiplex-peer.md](docs/multiplex-peer.md).
## What changes
### Server (turnserver)
- **`--multiplex-peer`** (cross-platform) — enable the shared per-thread
relay sockets. Replaces the per-session port bind. Implies sendmmsg
batching on Linux and default-enables `--udp-recvmmsg` (override with
`--udp-recvmmsg=0`). Incompatible with EVEN-PORT — those Allocates are
rejected with 400.
- **`--multiplex-peer-port <port>`** (cross-platform, default 3480) —
base port; thread `i` binds `<base>+2i` (IPv4) and `<base>+2i+1` (IPv6).
A 4-thread server consumes 8 ports.
- **`--udp-gso`** (Linux-only CLI) — UDP-GSO (`UDP_SEGMENT` cmsg) on the
relay send path. Requires `--multiplex-peer` (which is what enables the
sendmmsg batching GSO piggybacks on); passing `--udp-gso` alone is a
silent no-op.
- **CLI surface tightened**: `--udp-recvmmsg`, `--udp-recvmmsg-log`,
`--udp-gso` and their fields are now `#if defined(__linux__)` — absent
from `--help`, rejected with `unrecognized option`, and the code paths
compile out on macOS/Windows.
- **Windows portability**: `SO_REUSEPORT` in `mp_open_socket` wrapped in
`#ifdef` (MSVC's Winsock doesn't define it; REUSEPORT was defensive
anyway because the per-thread port layout is unique by construction).
- **`--sock-buf-size` honoured at startup**: the shared multiplex-peer
relay socket now calls `set_ioa_socket_buf_size` in `mp_open_socket` so
the configured rcvbuf is in effect from the moment the socket exists,
not deferred to the first Allocate.
### turnutils_uclient (loadgen)
- **`--no-even-port`** — force `ep = -1` on Allocate. The default path
randomly attaches EVEN-PORT (with no-R bit) even under `-c`, which
`--multiplex-peer` strictly rejects with 400; this flag makes
alloc-flood runs against multiplex-peer deterministic.
- **Legacy `timer_handler` now wraps the per-tick send batch with
`uclient_send_batch_begin/_end`** — without this, runs with
`--sender-threads 0` (the default for `-m < 4`) silently fell through
every send to plain `send(2)`. strace A/B: 205 k `sendto` → 61 k
`sendmsg` (GSO) + 4 k `sendmmsg` + small `sendto` residual for control.
## Measured impact (3-droplet DigitalOcean, c-4 / 4 vCPU, 8 concurrent
UDP streams, 45 s)
| | baseline | `--udp-recvmmsg` | `--multiplex-peer` | `--multiplex-peer
--udp-gso` |
|---|---:|---:|---:|---:|
| Server NIC rx pps (UDP relay both legs) | 350 k | 334 k | 326 k | 294
k |
| Server `UdpInDatagrams` pps | 279 k | 292 k | 300 k | 294 k |
| **Server `UdpRcvbufErrors` pps** | **71 k** | 42 k | 26 k | **0.3 k
(−99.6 %)** |
| **`turnserver` process CPU** | **387 %** | 205 % | 283 % | **133 %
(−65 %)** |
| Server host idle | 22 % | 49 % | 41 % | **68 %** |
Same loadgen-side packet rate (~2 M pps reported by uclient `send_pps`
after the legacy-path batching fix). Iteration log:
[docs/PerformanceIterationLog.md](docs/PerformanceIterationLog.md).
## Test plan
- [x] `ctest --test-dir build` — 3/3 pass (test_ioaddr, test_stun_msg,
test_http_server) on macOS + Linux.
- [x] `examples/run_tests.sh` — 4 protocols + 4 threaded + load-gen
smoke on Linux; 4 protocols on macOS.
- [x] `examples/run_tests_conf.sh` — same coverage, conf-driven.
- [x] `examples/run_tests_multiplex_peer.sh` — UDP/TCP/TLS/DTLS via
`--multiplex-peer --multiplex-peer-port=35000` on macOS + Linux.
- [x] Flag matrix smoke on macOS: `--multiplex-peer`,
`--multiplex-peer-port=42000`, `--multiplex-peer --udp-gso` (no-op),
`uclient --no-even-port`, `uclient --listener-threads N --sender-threads
M` — all pass; `--udp-recvmmsg` / `--udp-gso` correctly rejected with
`unrecognized option`.
- [x] Flag matrix smoke on Linux (Docker): same + `--udp-recvmmsg`
accepted, `--multiplex-peer` auto-enables `--udp-recvmmsg`,
`--udp-recvmmsg=0` overrides the auto-enable.
- [x] Windows compile fix verified — `SO_REUSEPORT` no longer referenced
unconditionally.
- [x] 3-droplet perf matrix completed; per-hop UDP counters captured.
## Docs updated
- New: [docs/multiplex-peer.md](docs/multiplex-peer.md)
- [README.turnserver](README.turnserver): full entries for
`--multiplex-peer`, `--multiplex-peer-port`, `--udp-gso`; clarified
`--udp-recvmmsg` auto-enable semantics.
- [README.turnutils](README.turnutils): added `--no-even-port`, plus
previously-undocumented `--listener-threads` / `--sender-threads`
loadgen pool flags.
- [examples/etc/turnserver.conf](examples/etc/turnserver.conf):
commented `udp-recvmmsg`, `udp-recvmmsg-log`, `udp-gso`,
`multiplex-peer`, `multiplex-peer-port` keys with one-paragraph
descriptions and pointer to `docs/multiplex-peer.md`.
- Man pages regenerated via `./make-man.sh`.
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
## Summary
- New `--udp-gso` flag (Linux, requires `--udp-sendmmsg`) collapses
same-destination, same-size sendmmsg batches into a single `sendmsg`
with a `UDP_SEGMENT` cmsg, so the kernel allocates one super-skb that
traverses the network stack once and is segmented at egress instead of
running `udp_sendmsg → ip_finish_output → __dev_queue_xmit` per
datagram.
- Also wraps the relay-side `recvmmsg` callback loop in
`udp_sendmmsg_batch_begin/end` so peer→client sends triggered inside a
recv batch can also coalesce — without that wrapping the relay path
issues one `sendto` per delivered datagram.
- Sticky-disable on `EINVAL/ENOPROTOOPT` for older kernels/NICs that
lack UDP-GSO; one warning logged, then transparent fallback to the
existing `sendmmsg` and `udp_send` paths.
## Why
The `--udp-recvmmsg` and `--udp-sendmmsg` follow-ups confirmed (see
[docs/PerformanceIterationLog.md](docs/PerformanceIterationLog.md)) that
on the relay flood workload the dominant cost is the per-datagram kernel
TX path. mmsg-style batching reduces only the syscall entry/exit, not
the per-skb stack traversal — UDP-GSO collapses both.
## Result
DigitalOcean nyc1 c-4, 30 s alternating A/B, `-Y packet -m 1`, eth1 TX
as the authoritative server forwarding metric:
| Variant | eth1 RX | eth1 TX | sys CPU | idle CPU |
|---|---:|---:|---:|---:|
| baseline (no flags) | 322,091 | 127,445 | 22.9 % | 67.5 % |
| `--udp-recvmmsg --udp-sendmmsg --udp-gso` | 266,068 | **257,996** |
15.0 % | 78.7 % |
| baseline (no flags) | 309,475 | 125,573 | 20.9 % | 70.7 % |
| `--udp-recvmmsg --udp-sendmmsg --udp-gso` | 275,992 | **225,366** |
14.9 % | 74.3 % |
Mean server forwarding rate: **126.5 k → 241.7 k pps (+91 %, 1.91×)**,
mean system CPU **21.9 % → 14.9 %** — about **2.8× CPU efficiency** (TX
pps per system-CPU-%). Full perf-children comparison and methodology in
the new section of
[docs/PerformanceIterationLog.md](docs/PerformanceIterationLog.md).
## Notes for reviewers
- `--udp-gso` is opt-in and requires `--udp-sendmmsg` (the help text
states the dependency). Without `--udp-sendmmsg` the batch state never
accumulates and GSO has nothing to flush.
- GSO eligibility resets on every `_begin/_end`. Mixed-destination,
mixed-size, or oversize batches transparently fall back through
`sendmmsg` / `udp_send`.
- Rebased onto current `master`; the recvmmsg dependency is already
merged via #1906.
## Test plan
- [x] `cmake --build build --target turnserver` (RelWithDebInfo + ASan
local builds clean)
- [x] `ctest --test-dir build --output-on-failure` — 3/3 unit tests pass
- [x] `examples/run_tests.sh` — TCP/TLS/UDP pass; DTLS pre-existing
failure on macOS environment, unrelated to this change
- [x] DigitalOcean A/B perf validation captured above
- [ ] Reviewer to confirm CI green on Linux build/test/CodeQL
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
Extends the existing Linux-only `--udp-recvmmsg` flag from the UDP
listener socket to also cover **connected per-session UDP relay
sockets**, so steady-state client→relay and peer→relay traffic on plain
UDP is read in batches of up to 16 datagrams per `recvmmsg(2)` instead
of one `recvmsg` per packet. DTLS sessions still go through the SSL read
path and are unchanged.
The flag stays **opt-in**: receive-side batching works correctly, but on
the current `m=1` / `m=100` benchmarks throughput is flat to slightly
negative — the bottleneck has moved past receive (see results below).
## What's in the change
- **Shared receive helpers** (`src/apps/relay/ns_ioalib_engine_impl.c`,
`src/apps/relay/ns_ioalib_impl.h`):
- `ioa_parse_udp_recvmsg_cmsg()` — single TTL/TOS/`IP_RECVERR` cmsg
parser used by both `udp_recvfrom()` and the new batch path. Replaces
the duplicated parser previously inlined in `dtls_listener.c` and
`udp_recvfrom()`.
- `ioa_init_recvmmsg_hdr()` — single initializer for
`mmsghdr`/`iovec`/cmsg/source-address fields, also used by the listener.
- New `IOA_UDP_RECVMMSG_MAX_BATCH = 16` constant; both listener and
relay paths now share it.
- **Connected relay batch read** (`socket_udp_read_batch_recvmmsg` in
`ns_ioalib_engine_impl.c`): called from `socket_input_worker` for
non-SSL UDP sockets when `--udp-recvmmsg` is on. Allocates per-message
`stun_buffer_list_elem`s, calls `recvmmsg(MSG_DONTWAIT)`, dispatches
each datagram through the existing `read_cb` path, and falls back
cleanly on `ENOSYS`/`EINVAL`/`EOPNOTSUPP` (auto-disables the flag) and
on `EAGAIN`/short-batch (releases unused buffers).
- **Per-engine scratch state**: the `mmsghdr[16]` / `iovec[16]` / cmsg /
src-addr arrays live on `ioa_engine`, not on every socket — keeps memory
flat at thousands of allocations.
- **TTL/TOS-sized cmsg buffers** in the listener: the listener
previously over-allocated `64 KiB` per slot; it now uses the same
TTL+TOS sizing as the relay path.
- **Opt-in occupancy stats** behind a new `--udp-recvmmsg-log` flag:
every 10 s the relay logs `udp-recvmmsg stats: calls=… packets=…
avg_batch=… wouldblock=… unavailable=… no_buffer=… hist_1=… hist_2=…
hist_3_4=… hist_5_8=… hist_9_16=…`. Counters are always tracked (cheap);
the periodic log is gated by the new flag so default operation is
silent.
- **CLI plumbing**: `--udp-recvmmsg-log` long option in
`mainrelay.c`/`mainrelay.h`, `cli_print_flag` entry in
`turn_admin_server.c`, doc updates in `README.turnserver`.
- **Docs**: `docs/PerformanceIterationLog.md` records the iteration
steps, validation, and two rounds of DigitalOcean A/B numbers.
`CLAUDE.md` load-test instructions updated to mention the new flag and
the `tot_recv_msgs` / `tot_recv_bytes` workaround.
Remove the two engine implementations (NEV_UDP_SOCKET_PER_SESSION and
NEV_UDP_SOCKET_PER_ENDPOINT) and all the dispatch/selection logic around
them. NEV_UDP_SOCKET_PER_THREAD is now the sole, unconditional
implementation.
- mainrelay.h: removed _NET_ENG_VERSION enum, typedef, and
net_engine_version / net_engine_version_txt struct fields
- mainrelay.c: removed NE_TYPE_OPT CLI option, set_network_engine(),
per-endpoint branch in print_features(), and all remaining
net_engine_version references
- netengine.c: removed run_udp_listener_thread(),
setup_socket_per_endpoint_udp_listener_servers() (~190 lines),
setup_socket_per_session_udp_listener_servers() (~90 lines); simplified
setup_barriers(), setup_relay_server(), run_general_relay_thread(),
setup_general_relay_servers(), and setup_server() by eliminating all
engine-type conditionals
- turn_admin_server.c: replaced dynamic engine version lookups with
hardcoded values (3 / "UDP thread per CPU core") in CLI and HTTPS status
handlers
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
A rewrite and reformat of the PostgreSQL.md file:
- Applies some minor formatting to the document.
- Applies consistency between the 'turn' user and 'coturn' database
names.
- Adds some warnings about issues I encountered during my own
installation process (public schema error)
- Adds a minor example of restarting a service with SystemD
Add basic diagram with modules of coturn inside Developer.md.
I also wanted to move the existing Flow chart to mermaid so that our doc
doesn't depend on external tools but i couldn't find the time to do it
properly (horizontal arrows are hard to represent with mermaid) so i
postponed that part.
This is the result of the mermaid diagram:
The flowchart will be kept updated:
- Use [drawio](https://app.diagrams.net/) to edit.
1. Open [drawio](https://app.diagrams.net/) in brower
2. Menu → File → Open from ... → Device:
Select [FlowChart.html](FlowChart.html)
3. Edit flow chart
4. Export to svg:
Menu → File → Export as... → SVG...
This is not a long term documentation works - this is a temp proposal to
modernize existing documentation
- Create docs folder
- Move existing documentation files into docs
- Split one massive INSTALL files into smaller files (no text editing
was done except small markdown header formatting)
- Rename files to be markdowns for nicer rendering
- Bad md rendering is expected in many places
Pavel PunskyandClaude Opus 4.8
Captain Picarl
Gustavo Garcia
dominiquefournier
Kang Lin
vuittont60