Files
eepp/docs
Martín Lucas Golini 00999e26c3 Queue HTTP connection waits, add a socket budget and cancel on shutdown
Requests that hit a connection limit used to wait inside their worker,
polling every 10 ms, and an async request's timeout included that wait.
On a page with many resources from one origin, later requests timed out
while queued, and waiting requests tied up shared thread-pool workers
that other origins needed.

Connection queue:
- Serve blocking and async requests from one first-come, first-served
  queue per pool. Requests waiting on one origin's limit never block
  requests to other origins.
- In shared thread-pool mode, an async request runs on a worker only
  after it has been granted a connection. Grants are handed to the
  executor from a service thread, never from the thread releasing a
  connection, which could be inside the executor's own lock.
- Requests hold their executor only weakly, so a request can never be
  the last owner of the thread pool running it.
- Start an async request's timeout once it holds a connection. A
  blocking call's timeout still covers the wait plus the transfer, and
  a resumed download keeps its deadline across the HEAD probe.
- Wait on a condition variable instead of polling. Cancelling by ID,
  shutdown() and Pool::close() release queued requests immediately.
- Let running requests and blocking callers take over a slot reserved
  for queued work that has not started; that work queues again, first
  in line. Without this, a redirect, retry or resumed download on a
  busy executor could wait forever on slots reserved behind it.
- When queued work starts, switch to a connection that went idle while
  it waited, instead of opening a new one.
- Replace an idle connection the server has closed in place, keeping
  its slot.

Process-wide limit and idle expiry:
- Limit open sockets across all pools with
  Http::setMaxOpenConnections() (default 256). At the limit, close the
  least recently used idle socket in any pool, then a socket held for
  queued work, and only wait while every socket is in active use.
- Wake waiters on the limit when a socket goes idle, and recheck a
  change counter so a wake-up just before waiting is never missed.
- Only inspect a socket for reclaim once its connection is known to be
  idle or reserved, never while a request is using it.
- Close idle connections from a background timer when idleTimeout
  elapses, instead of only when a connection is next requested in the
  same pool.
- The budget and its service thread are intentionally never destroyed,
  so the thread is never joined during static destruction, which would
  deadlock under the Windows loader lock.

Shutdown:
- Add Http::cancelAllRequests(), which cancels every queued and running
  request in every pool and closes idle connections. A per-pool
  generation counter makes running requests stop at their next check,
  and the pools stay usable afterwards.
- Add Http::waitForConnectionsToClose().
- Engine cancels all requests as it starts shutting down. Once its
  windows are closed, it waits up to 100 ms for sockets to close before
  ending TLS.
- Check for cancellation every 25 ms instead of 100 ms while reading, so
  requests stop within that wait.

TLS:
- SSLSocket::end() now waits while TLS sockets exist: the last socket
  destroyed releases the backend, and a socket created in the meantime
  keeps it.
- Set up the backend and register a new socket a
  lock, so a concurrent end() cannot release the backend in between.
- Add SSLSocket::isInitialized() and a testing h
  socket construction.
2026-10-15 00:36:56 -03:00
..