| Age | Commit message (Collapse) | Author |
|---|
|
The previous weak-symbol workaround let the module load on musl (which
treats RTLD_LAZY as RTLD_NOW) but introduced a silent failure: weak
undefined references are bound to NULL at our own dlopen time and are
not updated when a later RTLD_GLOBAL dlopen brings the protocol module
in. In any [modules] ordering where metrics comes before pppoe / l2tp /
pptp / sstp / ipoe, the function pointers stay NULL and per-protocol
session metrics silently disappear from both Prometheus and JSON
output, with no log to indicate why.
Replace the weak declarations with a small table and resolve each
protocol's stat_starting/stat_active pair via dlsym(RTLD_DEFAULT, ...)
the first time we render after the module is seen as loaded. dlsym
walks the live global scope at call time, so it picks up symbols
regardless of dlopen order; the resolved pointers are cached so
subsequent scrapes do not re-walk the loader. The five near-identical
render blocks in render_prometheus() and render_json() collapse into
table-driven loops.
libdl is already a transitive dependency of accel-pppd via triton, so
no build-system changes are needed.
Signed-off-by: Denys Fedoryshchenko <denys.f@collabora.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
|
write_all() previously did a blocking-style loop on a O_NONBLOCK
socket and bailed on the first EAGAIN. With a slow scrape client or
a small kernel send buffer that meant the response was truncated and
the connection dropped mid-flight.
Allocate one contiguous xmit_buf per response holding header + body,
then drain it in xmit_flush():
* full write → mark the client for disconnect on the next event
loop tick;
* EAGAIN/EWOULDBLOCK → enable MD_MODE_WRITE so cln_write() resumes
the drain when the socket becomes writable;
* hard error → mark for disconnect, caller tears down.
cln_read() now stops reading once a response is queued (read events
during the response phase are uninteresting since we'll close on
flush), and cln_write() finishes the drain and disconnects when the
last byte is out. The existing per-client read timer doubles as a
write deadline, so a peer that opens the connection and never reads
still gets cleaned up after read_timeout seconds.
Smoke-tested with a python client that uses SO_RCVBUF=256 and
sleep(0.05) between recv()s — it now reads the entire ~2.8 KiB body
across many short reads. Five concurrent slow readers plus a fast
scrape all complete successfully.
Signed-off-by: Denys Fedoryshchenko <denys.f@collabora.com>
|
|
Previously the only path out of serv_read()'s accept loop was an
EAGAIN/EWOULDBLOCK return; every other failure logged once and fell
back into `continue`. With a level-readable listening fd, that means
EMFILE/ENFILE/ENOBUFS/ENOMEM pin the worker thread spinning on
accept() and saturate the log.
Detect that class of error and pause the listener: disable
MD_MODE_READ on serv_hnd, arm a one-shot triton timer for one
second, and on expiry re-enable the handler. EINTR and ECONNABORTED
are kept as transient retries — those are normal and short-lived.
Verified by running the daemon under `prlimit --nofile=24` and
opening enough slow connections to exhaust the limit. The first
accept failure logs
metrics: accept failed: Too many open files; backing off 1s
then the daemon idles at 0% CPU instead of spinning; once fds free
up it resumes accepting.
Signed-off-by: Denys Fedoryshchenko <denys.f@collabora.com>
|
|
A scrape client that opens a TCP connection and never sends a full
request line+headers used to keep its accel-pppd-side fd registered
indefinitely. Combined with the default `allowed_ips` (= allow all),
a single peer could exhaust the daemon's file descriptors
slowloris-style.
Give every accepted connection a triton timer armed for
`read_timeout` seconds (default 5). On expiry, disconnect_client()
tears down the fd, the timer, and the buffer. The timer is canceled
implicitly when the client is disconnected for any other reason
because disconnect_client() now deletes the timer before freeing the
client.
Also cap the number of in-flight clients at `max_clients`
(default 64). Excess connections are accepted and immediately
closed so the kernel listen backlog still drains.
Both knobs accept 0 to disable. The default values are documented in
accel-ppp.conf(5) alongside the existing [metrics] options.
Smoke-tested:
* a connection that sends nothing is dropped from the daemon's fd
table when read_timeout elapses; subsequent scrapes still
succeed;
* with max_clients=3 and five concurrent silent connections, the
daemon holds exactly three ESTAB sockets, the others are closed.
Signed-off-by: Denys Fedoryshchenko <denys.f@collabora.com>
|
|
Replace the previous `{}` stub with a structured JSON document carrying
the same fields as the Prometheus exposition: build info, uptime,
CPU%, RSS/virtual memory, the triton core counters, session counts by
state, and a `protocols` object whose keys are only present for
modules that are actually loaded.
Strings are emitted through a small helper that escapes the JSON
control characters (\b, \f, \n, \r, \t, \", \\) and falls back to
\u00XX for other bytes below 0x20, so the version string and any
future textual labels survive without producing invalid JSON. The
output passes `python3 -m json.tool` against a running daemon.
Content-Type is already set to application/json by content_type(), so
no transport changes are needed.
Signed-off-by: Denys Fedoryshchenko <denys.f@collabora.com>
|
|
Fill in the body that GET /metrics returns when format=prometheus.
The exposed series mirror what `show stat` prints over the CLI today:
* accel_ppp_build_info{version="..."} 1
* accel_ppp_uptime_seconds
* accel_ppp_cpu_percent
* accel_ppp_memory_{rss,virt}_bytes (read from /proc/<pid>/statm)
* accel_ppp_core_mempool_{allocated,available}_bytes
* accel_ppp_core_threads{,_active}
* accel_ppp_core_contexts{,_sleeping,_pending}
* accel_ppp_core_md_handlers{,_pending}
* accel_ppp_core_timers{,_pending}
* accel_ppp_sessions{state="starting|active|finishing"}
* accel_ppp_protocol_sessions{protocol=...,state=...} for every
protocol module that is currently loaded — pppoe, l2tp, pptp,
sstp, ipoe — gated by triton_module_loaded() so we never call a
stat helper from a module that wasn't loaded.
A small growing strbuf helper coalesces the rendering into a single
buffer that is passed to send_response() in one shot. Per-protocol
forward declarations rely on the existing RTLD_LAZY|RTLD_GLOBAL load
behaviour used by net-snmp, so the metrics .so does not need to link
against pppoe.so, l2tp.so, etc.
json format remains a stubbed `{}` body — the next commit replaces
that with a real renderer.
Signed-off-by: Denys Fedoryshchenko <denys.f@collabora.com>
|
|
Parse the `allowed_ips` option in [metrics] as a comma-separated list
of IPv4 CIDR entries. Both the bracketed form
allowed_ips = ["1.2.3.4/32", "5.6.7.0/24"]
and a bare comma-separated form are accepted; surrounding whitespace,
matched quotes, and the optional [ ] are stripped. A bare address
without a prefix is treated as /32.
When the list is empty (option missing or empty value), all peers are
allowed and behavior is unchanged. Otherwise serv_read() rejects any
peer that doesn't match a configured CIDR by closing the freshly
accepted socket before allocating client state, so scanners get
nothing more than a TCP reset.
Signed-off-by: Denys Fedoryshchenko <denys.f@collabora.com>
|
|
Stand up a minimal HTTP/1.1 server on top of triton's md handler, using
the same pattern as cli/tcp.c. The listener:
* binds the address configured in [metrics]/address (IPv4 only for
now);
* accepts non-blocking connections, reads up to the first
"\r\n\r\n" into a fixed-size buffer (2 KiB), then dispatches one
request and closes the connection;
* routes GET /metrics to a placeholder 200 response (body is empty
until the metrics rendering lands in a later commit);
* returns 404 for other paths, 405 for non-GET, 413 if the request
headers do not fit, and 400 for an unparseable request line.
Content-Type is selected from the configured format (prometheus or
json) so the next commits can plug in real bodies without touching the
transport.
Signed-off-by: Denys Fedoryshchenko <denys.f@collabora.com>
|
|
Add a new extra module `metrics` that will eventually expose statistics
over HTTP. This first commit only:
* creates accel-pppd/extra/metrics.c with an init() that parses the
[metrics] section options `format` (prometheus|json) and
`address` (host:port);
* wires the new shared library into the extras CMakeLists.
No listener, no metrics rendering yet — those land in follow-up
commits. With this commit alone, loading the `metrics` module just
logs the configured listen address and format and is otherwise inert.
Signed-off-by: Denys Fedoryshchenko <denys.f@collabora.com>
|
