test(benchmarks): add non-Asio controls for connection stalls

benchmarks/CMakeLists.txt

@@ -116,6 +116,16 @@ if(ARKNET_BUILD_TESTS)
     endif()
     arknet_add_check(connection_socket_operation_trace 10 arknet_connection_socket_trace_test)
 endif()
+if(UNIX)
+    # POSIX poll(2) or kqueue peer without Asio or strands; speaks the native control protocol.
+    add_executable(arknet_poll_connection_control poll_connections.cpp)
+    target_compile_features(arknet_poll_connection_control PRIVATE cxx_std_20)
+    target_link_libraries(arknet_poll_connection_control PRIVATE Threads::Threads)
+    if(ARKNET_BUILD_TESTS)
+        arknet_add_check(benchmark_poll_connections_help 10 arknet_poll_connection_control --help)
+        arknet_add_check(benchmark_poll_connections_self_test 20 arknet_poll_connection_control --self-test)
+    endif()
+endif()
 if(ARKNET_BUILD_TESTS AND UNIX)
     add_executable(arknet_asio_control_process_test ../tests/connection_load_native_control.cpp)
     add_dependencies(arknet_asio_control_process_test arknet_asio_connection_control arknet_connection_benchmark)

benchmarks/CONNECTIONS.md

@@ -191,6 +191,25 @@ HELLO, START, DATA, FIN and acknowledgement identity/sequence rules match the
 original benchmark. Payload is limited to 1–4023 bytes, preserving the original
 conservative 4096-byte receive-cap validation.
 
+Two diagnostic isolation switches change one mechanism each; their defaults keep
+the behavior above, and the final `result` records both. `--session-strands off`
+binds every session handler to the context executor instead of a per-session
+strand; with a single IO worker this still serializes all handlers.
+`--write-buffers joined` copies the length prefix and protocol frame into one
+buffer per write instead of passing two buffers; the extra copy is not included in
+the declared buffer budget. Pair both peers with the same switches when isolating a
+mechanism, and compare against interleaved default runs.
+
+`--io-histogram on` wraps each socket in a counting stream and adds `io_histogram`
+to the `result`: per direction, successful `async_read_some`/`async_write_some`
+calls, bytes, short completions and log2 histograms of requested and completed
+bytes. `async_read_until` issues zero-byte reads when a frame is already buffered;
+these appear as requested 0 and do not call the kernel. Readiness waits and
+would-block results are not visible through Asio and are reported as 0. The option
+cannot be combined with `--socket-operation-trace`, and the wrapper itself is part
+of the measured configuration. The poll control always reports the same object,
+including `EAGAIN` counts, waits and returned events.
+
 The default active capacity is 5000, connect parallelism is 32, data inflight
 capacity is 20000 and the send/data byte limit is 64 MiB. These are declared limits,
 not measured capacity. `--max-inflight` adds an explicit global DATA count limit;
@@ -456,3 +475,96 @@ not the older published IO snapshot time. Validate wall/steady offset stability
 across anchors; a clock step, large bracket or missing cancellation anchor makes
 precise cross-process timing inconclusive. Earlier output without these anchors
 cannot be precisely aligned to packet wall timestamps from its steady epoch alone.
+
+## POSIX poll control without Asio
+
+`arknet_poll_connection_control` is a second diagnostic TCP peer, built with the
+benchmark option on POSIX systems. It uses neither Asio nor strands: by default one
+thread drives non-blocking IPv4 sockets with `poll(2)`. It speaks the same
+length-prefixed HELLO, START, DATA, FIN and acknowledgement protocol as
+`arknet_asio_connection_control`, so either role can be paired with the native
+Asio control. Use the four pairings to separate the event loop of each peer from
+the transport underneath.
+
+```sh
+cmake --build build --target arknet_poll_connection_control --parallel 2
+build/benchmarks/arknet_poll_connection_control --self-test
+build/benchmarks/arknet_poll_connection_control --role server --listen 127.0.0.1 \
+  --port 21002 --run-id 71 --expected-connections 8 --window 4 \
+  --hold-seconds 0.1 --deadline-seconds 5
+# Separate terminal/process; the server may also be arknet_asio_connection_control
+# with --scenario traffic:
+build/benchmarks/arknet_poll_connection_control --role client --target 127.0.0.1 \
+  --port 21002 --run-id 71 --client-id 91 --connections 8 --window 4 \
+  --hold-seconds 0.1 --deadline-seconds 5
+```
+
+The client paces connects with `--connect-rate` and `--connect-parallel`, sends
+HELLO with `TCP_NODELAY`, starts traffic after every connection is ready and keeps
+`--window` DATA requests outstanding until `--hold-seconds` elapses. Each peer tries
+`send` immediately and waits for `POLLOUT` only for the unsent remainder; input
+uses the same 4096-byte bound and canonical length rules. Payload is limited to
+1–4023 bytes. The traffic scenario is the only scenario.
+
+On Apple and BSD systems `--events kqueue` replaces `poll(2)` with a loop that
+registers events the way the Boost.Asio kqueue reactor does, leaving the protocol,
+socket calls and diagnostics unchanged. Sends, receives, accepts and connects are
+attempted first; only one that must wait registers its descriptor with
+`EV_ADD | EV_CLEAR`, `EVFILT_READ` for reads and `EVFILT_READ` plus `EVFILT_WRITE`
+once a write or connect must wait. A write event that arrives with nothing left to
+write deletes the `EVFILT_WRITE` registration, and one `kevent` call returns at
+most 128 events. Other platforms reject `kqueue`; the `result` event records
+`events`. Pair both peers in the same mode to compare the two notification models.
+
+`--write-pacing serial` reproduces the native control's write cadence in either
+event mode: each session writes one frame at a time, trying `send` immediately
+when nothing is outstanding. A fully written frame posts its completion to the
+next loop pass, which then starts the next queued frame; posted completions make
+that pass's wait non-blocking. The default `burst` appends each frame and sends
+at once, so a window of requests usually leaves in consecutive calls. The `result`
+event records `write_pacing`.
+
+`--read-pacing serial` reproduces `async_read_until` on the native control's
+4096-byte stream buffer. A frame already buffered completes through a posted step
+without `recv`. Otherwise one `recv` is attempted and its completion is posted to a
+later pass, where framing is checked again; a would-block result waits for one
+readiness event, which performs one `recv`. Each completion handles one frame and
+then starts the next read; readiness that arrives while no read waits is dropped.
+The default `drain` reads until `EAGAIN` and handles every buffered frame at once.
+The `result` event records `read_pacing`.
+
+`--socket-setup asio` prepares sockets the way Boost.Asio's `socket_ops` does on
+Apple and FreeBSD: `SO_NOSIGPIPE` on the listener and on every created and
+accepted socket, and non-blocking mode through `ioctl(FIONBIO)`. The default sets
+`O_NONBLOCK` with `fcntl` and relies on an ignored `SIGPIPE`. Single-buffer reads
+and writes already use `recv` and `send` in both programs. The `result` event
+records `socket_setup`.
+
+`--send-queue-watch on` makes either role query every open socket's kernel send
+queue every 250 ms. A queue that stays non-empty and unchanged for one second has
+made no delivery progress; the peer writes a `send_queue_stalled` event with
+`phase` `detected`, then `sample` once per further second (at most 20 per
+connection and 64 connections), and `progressed` once the queue changes. Each
+event carries ports, sequence state, application and kernel queue bytes and
+`tcp_info` from Apple `TCP_CONNECTION_INFO` (state, flags, RTO, congestion window,
+slow-start threshold, peer window, transmitted, retransmitted and out-of-order
+bytes); other platforms report `"status":"unsupported"`. `stalled_connection`
+events include the same `tcp_info`. The scan adds one `getsockopt` per socket per
+interval and is part of the measured configuration; the `result` event records
+`send_queue_watch`.
+
+The first request older than `--request-timeout-ms` fails the client. It writes one
+`stalled_connection` event per affected connection with `phase` `at_failure`,
+then stops admitting requests. `--observe-seconds` keeps reading for that interval
+and reports the same connections again with `phase` `after_observation`. The server
+reports, at the first premature EOF, deadline or signal, unfinished connections
+whose reads stopped at least `--request-timeout-ms` before its latest read on any
+connection. Each event contains the local and peer ports, sequence state, bytes
+handed to and read from the kernel, bytes still queued in the application, and the
+kernel readable and send-queue byte counts (`FIONREAD`; `SO_NWRITE` on macOS,
+`SIOCOUTQ` on Linux, otherwise -1). Match peers by reversed ports.
+
+Stdout ends with one `result` event (`success`, `failure_scope`, counters). Exit
+status 0 means success, 1 a failed run and 2 an argument or startup error. This
+control has no operation trace, kernel TCP samples or resource events; it does not
+replace the arknet benchmark or the native Asio control.

benchmarks/CONNECTIONS_CN.md

@@ -185,6 +185,19 @@ build/benchmarks/arknet_asio_connection_control --role client --target 127.0.0.1
 序号规则与原 benchmark 一致。payload 限制为 1–4023 字节，保持原程序对 4096 字节
 接收上限的保守校验。
 
+两个诊断隔离开关各只改变一种机制；默认值保持上述行为，最终 `result` 会记录两者。
+`--session-strands off` 把每个会话的回调绑定到 context executor，而不是每会话 strand；
+只有一个 IO 工作线程时，所有回调仍然串行执行。`--write-buffers joined` 在每次写入时
+把长度前缀和协议帧复制到同一个 buffer，而不是传入两个 buffer；这份额外副本不计入声明的
+buffer 预算。隔离某种机制时，两端应使用相同开关，并与交替运行的默认配置对比。
+
+`--io-histogram on` 用一个计数 stream 包装每个 socket，并在 `result` 中增加 `io_histogram`：
+按读写方向记录成功的 `async_read_some`/`async_write_some` 调用次数、字节数、未完成全部请求的
+次数，以及请求字节数和完成字节数的 log2 直方图。帧已在缓冲区中时，`async_read_until` 会发起
+零字节读取；它们记为请求 0 字节，不会调用内核。Asio 不暴露就绪等待和 would-block 结果，
+这两项报告为 0。该选项不能与 `--socket-operation-trace` 同时使用，包装层本身也属于被测配置。
+poll 对照总是输出同样的对象，并包含 `EAGAIN` 次数、等待次数和返回的事件数。
+
 默认活动连接容量为 5000，最多同时建连 32 个，DATA 在途容量为 20000，发送和 DATA
 字节限制为 64 MiB。这些是配置边界，不是已验证容量。`--max-inflight` 增加显式全局
 DATA 数量限制；声明的连接窗口必须同时满足数量和字节限制。控制帧也占用发送字节额度。
@@ -374,3 +387,73 @@ OS 缺陷。首次发布仍依赖 IO worker，发布前 worker 已阻塞时不
 `[w+q-b, w+q-a]`，再纳入PCAP时间戳量化误差。应使用query区间，不用较旧的IO状态快照时间。
 须检查anchor间wall/steady偏移稳定性；时钟跳变、过宽区间或缺取消anchor时，精确跨进程
 对齐不可判定。旧输出仅有steady epoch，无法精确对齐包的wall时间戳。
+
+## 不使用 Asio 的 POSIX poll 对照
+
+`arknet_poll_connection_control` 是第二个诊断用 TCP 对端，在 POSIX 系统上随 benchmark
+选项构建。它不使用 Asio 或 strand：默认由一个线程用 `poll(2)` 驱动非阻塞
+IPv4 socket。它与 `arknet_asio_connection_control` 使用相同的长度前缀 HELLO、START、
+DATA、FIN 及确认协议，因此任一角色都可以与原生 Asio 对照配对。用四种配对区分两端各自的
+事件循环与其下的传输层。
+
+```sh
+cmake --build build --target arknet_poll_connection_control --parallel 2
+build/benchmarks/arknet_poll_connection_control --self-test
+build/benchmarks/arknet_poll_connection_control --role server --listen 127.0.0.1 \
+  --port 21002 --run-id 71 --expected-connections 8 --window 4 \
+  --hold-seconds 0.1 --deadline-seconds 5
+# 在另一个终端或进程中运行；服务端也可以换成带 --scenario traffic 的
+# arknet_asio_connection_control：
+build/benchmarks/arknet_poll_connection_control --role client --target 127.0.0.1 \
+  --port 21002 --run-id 71 --client-id 91 --connections 8 --window 4 \
+  --hold-seconds 0.1 --deadline-seconds 5
+```
+
+客户端按 `--connect-rate` 和 `--connect-parallel` 控制建连节奏，开启 `TCP_NODELAY` 后发送
+HELLO，所有连接就绪后开始流量，并在 `--hold-seconds` 内保持 `--window` 个 DATA 请求在途。
+两端都先立即调用 `send`，只对未发出的剩余部分等待 `POLLOUT`；输入使用相同的 4096 字节上限
+和规范长度规则。payload 限制为 1–4023 字节。只支持 traffic 场景。
+
+在 Apple 和 BSD 系统上，`--events kqueue` 用一个按 Boost.Asio kqueue reactor 方式注册事件
+的循环替换 `poll(2)`，协议、socket 调用和诊断输出保持不变。send、recv、accept 和 connect
+都先直接尝试；只有需要等待的操作才用 `EV_ADD | EV_CLEAR` 注册描述符：读取只注册
+`EVFILT_READ`，写入或建连需要等待时同时注册 `EVFILT_READ` 和 `EVFILT_WRITE`。写事件到达时
+若已无待写数据，就删除 `EVFILT_WRITE` 注册；每次 `kevent` 调用最多返回 128 个事件。其他
+平台会拒绝 `kqueue`；`result` 事件记录 `events`。比较两种通知模型时，两端应使用相同模式。
+
+`--write-pacing serial` 在两种事件模式下都能复现原生对照的写入节奏：每个会话同一时间只写一帧，
+没有在途写入时立即尝试 `send`。一帧完全写出后，把完成通知投递到下一轮循环，由那一轮再开始写
+队列中的下一帧；有待处理的完成时，那一轮的等待不会阻塞。默认的 `burst` 会追加每一帧并立即
+发送，因此同一窗口的请求通常在连续的调用中发出。`result` 事件记录 `write_pacing`。
+
+`--read-pacing serial` 复现原生对照在 4096 字节 stream buffer 上的 `async_read_until`：
+缓冲区里已有完整帧时，不调用 `recv`，通过投递的一步完成；否则尝试一次 `recv`，并把完成投递到
+下一轮，在那一轮重新检查帧边界；遇到 would-block 时等待一次就绪事件，由它执行一次 `recv`。
+每次完成只处理一帧，然后开始下一次读取；没有读取在等待时到达的就绪事件会被丢弃。默认的
+`drain` 一直读到 `EAGAIN`，并一次处理缓冲区里的全部帧。`result` 事件记录 `read_pacing`。
+
+`--socket-setup asio` 按 Boost.Asio 的 `socket_ops` 在 Apple 和 FreeBSD 上的做法准备 socket：
+监听 socket 以及每个创建和接受的 socket 都设置 `SO_NOSIGPIPE`，并用 `ioctl(FIONBIO)` 设置非阻塞。
+默认方式用 `fcntl` 设置 `O_NONBLOCK`，并依靠忽略 `SIGPIPE`。两个程序的单缓冲区读写本来就都使用
+`recv` 和 `send`。`result` 事件记录 `socket_setup`。
+
+`--send-queue-watch on` 让任一角色每 250 ms 查询一次每个打开 socket 的内核发送队列。发送队列
+连续一秒非空且没有变化，说明没有任何投递进展；此时输出 `phase` 为 `detected` 的
+`send_queue_stalled` 事件，之后每多一秒输出一次 `sample`（每个连接最多 20 次，最多跟踪 64 个
+连接），队列一旦变化就输出 `progressed`。每条事件包含端口、序号状态、应用层与内核队列字节数，
+以及来自 Apple `TCP_CONNECTION_INFO` 的 `tcp_info`（状态、标志、RTO、拥塞窗口、慢启动阈值、
+对端窗口、已发送、重传和乱序字节数）；其他平台报告 `"status":"unsupported"`。
+`stalled_connection` 事件也包含同样的 `tcp_info`。每个间隔对每个 socket 多一次 `getsockopt`，
+这属于被测配置的一部分；`result` 事件记录 `send_queue_watch`。
+
+第一个超过 `--request-timeout-ms` 的请求会使客户端失败。客户端为每个受影响的连接输出一条
+`phase` 为 `at_failure` 的 `stalled_connection` 事件，然后停止发出新请求。
+`--observe-seconds` 会在该时长内继续读取，并以 `phase` `after_observation` 再次报告同一批
+连接。服务端在第一次提前 EOF、到达截止时间或收到信号时，报告读取停止时间比任一连接最近一次
+读取早至少 `--request-timeout-ms` 的未完成连接。每条事件包含本端与对端端口、序号状态、交给
+内核及从内核读取的字节数、仍在应用层排队的字节数，以及内核可读字节数和发送队列字节数
+（`FIONREAD`；macOS 用 `SO_NWRITE`，Linux 用 `SIOCOUTQ`，其他平台为 -1）。按端口互换匹配两端。
+
+stdout 以一条 `result` 事件结束（`success`、`failure_scope` 及计数）。退出码 0 表示成功，
+1 表示运行失败，2 表示参数或启动错误。该对照不提供操作历史、内核 TCP 采样或资源事件，
+不能替代 arknet benchmark 或原生 Asio 对照。