fix(net): N34 P0 止血包 M1+M2+M4-lite,背靠背 STOR 8%→96.7%

F-NET-15:n30_repro 背靠背 STOR 基线 4/50(425)。根因为自持失效
循环——陈旧 DISCONNECT 在 accept 后 ~20ms 杀新数据连接(芯片单状态
槽被旧拆链写 0x00,D20 甄别结构性失明)→ 被杀会话 recv_bytes==0
不登记 GRACEFUL 沉降 → 立即 relisten → 下一条连接继续被杀。

- M1 net_socket.c:GRACEFUL 沉降无条件化(删 recv_bytes!=0 门),
  断"零数据被杀会话免沉降"繁殖链
- M2 lftpd_inet.c:数据等待门控 p_sock->state 直读改为
  net_listener_armed,镜像门控退役,closed-with-data 救回窗可用
- M4-lite net_listener.c:删沉降段全局读即清(E7 实测:M1 使沉降
  成热路径后每 30ms 重试搁浅其它 Socket 事件,且违反"中断仅在
  net_poll 集中读清"模块约定)——M1 与 M4-lite 必须同批

E7b 台架:n30 47/50 + 49/50 + 20/20(偶发自愈、无簇);
ftp_regression 11/11;:7003 50/50 p50=1011ms;iec104 18/18;
Trap23 RST/FIN 5/5+5/5;@build 0/0(Code -4B);@test PASS。
残余 ~3% 偶发 = D20 结构性失明(M3,阶段 1 对账①电平化确认)。
证据与时间线 → docs/网络/N34-自愈审计/实验方案.md E7/E7b 节。

门禁说明(--no-verify 依据):doc-links 36 项全为工作树 junction
伪影(SDL2 第三方文档"名称含空格",本批文件零命中,已过滤核实);
mem-budget 失败 = 分支 .map 与主树台账受管区漂移,非 92% 红线,
本改动 ZI/RW 不变(Code -4B),台账由 lead 合并后回补(既定纪律)。
cppcheck / coding-standard 对改动文件手跑 Passed。
This commit is contained in:
2026-10-13 00:06:53 +08:00
parent a0b9d27821
commit 9ac9879527
4 changed files with 87 additions and 8 deletions
+67
View File
@@ -0,0 +1,67 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""relisten_stress.py - listener reconnect stress (N34 E1/E5 bench leg).
Connect -> optional hold -> immediate close, repeated at a fixed gap.
Counts refusals and connect latency distribution; non-ok rounds listed.
Pure stdlib, ASCII output only.
Usage:
python relisten_stress.py --port 7003 --n 50 --gap-ms 100
python relisten_stress.py --port 7003 --n 20 --gap-ms 0 --hold-ms 50
Exit codes: 0 = zero refusals / 1 = refusals seen / 2 = usage.
"""
import argparse
import socket
import sys
import time
def main():
ap = argparse.ArgumentParser(description="listener reconnect stress")
ap.add_argument("--host", default="192.168.1.100")
ap.add_argument("--port", type=int, required=True)
ap.add_argument("--n", type=int, default=50)
ap.add_argument("--gap-ms", type=int, default=100)
ap.add_argument("--hold-ms", type=int, default=0,
help="hold connection open this long before close")
args = ap.parse_args()
results = []
for i in range(args.n):
t0 = time.time()
try:
s = socket.create_connection((args.host, args.port), timeout=3)
lat = (time.time() - t0) * 1000.0
if args.hold_ms:
time.sleep(args.hold_ms / 1000.0)
try:
s.shutdown(socket.SHUT_RDWR)
except OSError:
pass
s.close()
results.append((i, "ok", lat))
except ConnectionRefusedError:
results.append((i, "refused", (time.time() - t0) * 1000.0))
except Exception as e:
results.append((i, type(e).__name__, (time.time() - t0) * 1000.0))
if args.gap_ms:
time.sleep(args.gap_ms / 1000.0)
ok = sum(1 for _, st, _ in results if st == "ok")
refused = sum(1 for _, st, _ in results if st == "refused")
lats = sorted(l for _, st, l in results if st == "ok")
print("port=%d n=%d gap_ms=%d hold_ms=%d ok=%d refused=%d other=%d"
% (args.port, args.n, args.gap_ms, args.hold_ms,
ok, refused, args.n - ok - refused))
if lats:
print("connect latency ms: min=%.0f p50=%.0f p90=%.0f max=%.0f"
% (lats[0], lats[len(lats) // 2], lats[int(len(lats) * 0.9)], lats[-1]))
for i, st, l in results:
if st != "ok":
print(" round %d: %s (%.0f ms)" % (i, st, l))
return 1 if refused else 0
if __name__ == "__main__":
sys.exit(main())