fix(ci): FE-GATE-013 provider 시험의 exit=125 묶음을 고친다 #58

Open
donghyeon.kang wants to merge 4 commits from donghyeon.kang/ci/provider-exit-125 AGit into main
12 changed files with 397 additions and 31 deletions
+1
View File
@@ -50,6 +50,7 @@ Gitea runner에서 merge 단계를 초록으로 만들고, worker를 늘리지
| HTTP 실행기의 deadline 분류 시험이 가짜 시계로 대기 지점에서 deadline을 낸다. 실제 30ms deadline이 부하에서 backoff 전에 지나 TIMEOUT 대신 전송 실패로 끝났다 | `tests/unit/http-execution-v3.test.ts` | `37a85004` |
| 위험 커버리지가 `"use strict"` 밖의 지시문(`"use client"` 등)을 문장으로 센다. V8 coverage가 그렇게 센다(`check:v8-coverage-counter-semantics` fixture에 지시문 모듈 둘) | `scripts/lib/risk-coverage.ts`, `scripts/lib/v8-coverage-counter-semantics.ts` | `0c3c9d33` |
| e2e fixture build를 브라우저 project마다가 아니라 run에 한 번 만든다. global setup이 공유 폴더를 만들고 처음 닿은 worker만 build한다. 자기 spec에서 `prepareServiceWorkerBuilds`를 부르면 `key`를 준다 | `playwright.config.ts`, `tests/e2e/support/shared-fixture-builds.ts`, `tests/e2e/support/fixture-builds.ts`, `tests/e2e/support/service-worker-builds.ts` | `c6d7f43a` |
| FE-GATE-013의 시스템 컨테이너가 systemd 전에 root cgroup 제어기를 켠다. 부팅 중 `docker exec`가 root cgroup에 남으면 systemd가 user manager에 제어기를 넘기지 못해 provider 시험 18개가 exit 125로 떨어졌다(desktop 4회 중 1회). 사전 점검이 user manager의 cpu·memory·pids 제어기를 보고, provider 실패 메시지에 우리 단계가 쓴 고정 이유 줄을 남긴다 | `scripts/run-in-system-container.sh`, `scripts/check-system-test-prerequisites.ts`, `scripts/lib/provider-sandbox-diagnostics.ts`, `scripts/run-and-validate-provider.ts`, `scripts/lib/provider-scope-wrapper.ts` | `6d6fd61a`, `4e1a385f`, `13b40934` |
## 미출시 — `8327e221` → 하네스와 작업 매뉴얼
+1
View File
@@ -6,3 +6,4 @@
{"schema": 1, "kind": "actual", "task": "w2-fe-ci-perf", "recorded_at": "2026-11-12T22:25+09:00", "transcript": "00d79702-1dea-4e11-b79a-f021d0d9f4bb", "minutes": 1463, "active_minutes": 308, "calls": 447, "output_tokens": 413297, "cache_read_tokens": 246773496, "max_context": 943246, "within": "p30", "result": "partial", "commits": ["ddfcb6f6", "79a225b6", "0d8e4f5a", "b4210753", "66138440", "1d3eb901", "92abfd20", "1f09c1cc", "27876ffc", "35b60a04", "f90ffb6c", "f6e2560e", "ba72b865"], "why": "작업 중 시간은 p30 안이지만 일이 끝나지 않고 인계됐다(PR #19 Draft). 벽시계의 대부분은 runner 대기열(백엔드 정기 lane 뒤)과 desktop heavy 잠금 대기다. 남은 것: runner 초록 확인, concurrency, firefox 흔들림, coverage 경로, 느린 시험 둘. 세션 기록으로 다시 셌고 인계 절을 쓴 22:10 KST 뒤까지 들어 있다. 기록은 이어받은 세션이 남긴다"}
{"schema": 1, "kind": "estimate", "task": "w2-fe2-takeover", "recorded_at": "2026-11-12T22:25+09:00", "transcript": "fa9a12aa-7abc-46ca-a5fd-ee00cab98618", "started_at": "2026-11-12T21:05+09:00", "base": "ba72b865", "task_kind": "improve+operate", "size": "large", "minutes": {"p30": 180, "p50": 300, "p80": 480}, "calls": {"p30": 250, "p50": 400, "p80": 650}, "output_tokens": {"p30": 150000, "p50": 250000, "p80": 400000}, "basis": "PR #19 이어받기(인계 절 2026-11-12 22:10). concurrency·coverage 경로·느린 시험 둘은 각 30~60분(작음, 선례 W2-BE concurrency). firefox 008 흔들림은 원인 미확정이라 1~4시간. runner 1대를 백엔드와 같이 써 PR run 하나가 대기 포함 1시간 이상(W2-BE 기록: PR run이 2~4시간 뒤로 밀림), 초록 3회와 실패-고침 1~2번을 p80에 넣었다. 첫 5분 뒤에 적었다(인계 대기 중 시작)"}
{"schema": 1, "kind": "estimate", "task": "r9-fe-group-A", "recorded_at": "2026-11-13T18:50+09:00", "transcript": "66f9f89d-a13f-4cee-a2bf-c3c528c6bfb2", "started_at": "2026-11-13T17:57+09:00", "base": "ee95573f", "task_kind": "improve", "size": "large", "minutes": {"p30": 300, "p50": 420, "p80": 600}, "calls": {"p30": 350, "p50": 500, "p80": 750}, "output_tokens": {"p30": 200000, "p50": 300000, "p80": 450000}, "basis": "round9 그룹 A 발견 22건(화면·토큰·story·검사·접근성 기록)을 PR 셋으로. 바꿀 파일 약 60개, CSS 선언 118개 토큰화와 story 37개가 가장 크다. desktop heavy 잠금을 다른 세션 넷과 같이 써 실행마다 대기 10~30분, merge 단계 셋(각 약 21분)과 골든 갱신·변이 실험, runner run 셋(각 약 23분)을 p80에 넣었다. 선례 r9-fe F4a·F5(묶음당 2~4시간). 착수 약 50분 뒤 발견 원문을 다 읽고 PR1 고침을 쓰던 중에 적었다"}
{"schema": 1, "kind": "estimate", "task": "fe-ci-fix3", "recorded_at": "2026-11-14T06:12+09:00", "transcript": "b28bbd6f-b8e2-4b03-998f-95cc55c1ef86", "started_at": "2026-11-14T04:50+09:00", "base": "e98d8ee9", "task_kind": "improve", "size": "medium", "minutes": {"p30": 150, "p50": 210, "p80": 300}, "calls": {"p30": 180, "p50": 260, "p80": 380}, "output_tokens": {"p30": 70000, "p50": 110000, "p80": 170000}, "basis": "FE-GATE-013 exit=125 묶음(18개) 진단·원인·고침 PR 하나. 원인 미확정, 실패율 약 25%라 desktop 파일 전체 반복(회당 약 2.5분) 16회를 고치기 전·뒤 두 번, merge 단계(약 25분) 한두 번, 새 눈 리뷰와 확인 리뷰, runner run 하나(약 25분, 1대 공유 대기)를 넣었다. 선례 fe-ci-fix2(작음, 예측 p50 150분). 착수 때 적지 못하고 원인을 확정한 뒤(착수 약 80분) 적어 이미 쓴 시간에 기댄다"}
+2
View File
@@ -8,3 +8,5 @@
{"date": "2026-11-13", "test": "tests/browser-capabilities/indexeddb-runtime.spec.ts:256", "category": "resource-io", "first_run": "/tmp/claude-gates/ci-pr-gate-selection-merge/gate-artifacts/FE-GATE-008.txt", "verdict": "[firefox] desktop merge(트리 148e68f4, #47 없음)에서 30초 timeout 1회. 같은 게이트가 Firefox 프로필을 RAM에 둔 #47 트리에서는 69개 전부 통과. Firefox IndexedDB·OPFS가 프로필 디스크를 기다린 것으로 봄", "quarantined_until": "", "fixed_by": "프론트 #47"}
{"date": "2026-11-13", "test": "tests/browser-capabilities/local-draft-composition.spec.ts:10", "category": "resource-io", "first_run": "/tmp/claude-gates/ci-pr-gate-selection-merge/gate-artifacts/FE-GATE-008.txt", "verdict": "[firefox] desktop merge(트리 148e68f4, #47 없음)에서 30초 timeout 1회. 같은 게이트가 Firefox 프로필을 RAM에 둔 #47 트리에서는 69개 전부 통과. Firefox IndexedDB·OPFS가 프로필 디스크를 기다린 것으로 봄", "quarantined_until": "", "fixed_by": "프론트 #47"}
{"date": "2026-11-13", "test": "tests/browser-capabilities/opfs-runtime.spec.ts:338", "category": "resource-io", "first_run": "/tmp/claude-gates/ci-pr-gate-selection-merge/gate-artifacts/FE-GATE-008.txt", "verdict": "[firefox] desktop merge(트리 148e68f4, #47 없음)에서 30초 timeout 1회. 같은 게이트가 Firefox 프로필을 RAM에 둔 #47 트리에서는 69개 전부 통과. Firefox IndexedDB·OPFS가 프로필 디스크를 기다린 것으로 봄", "quarantined_until": "", "fixed_by": "프론트 #47"}
{"date": "2026-11-13", "test": "tests/system/ci-runner/ci-artifact-contract.test.ts:671", "category": "concurrency-order", "first_run": "/tmp/claude-gates/w2fe2-t08m/sys-2.json", "verdict": "원인 확정(2026-11-14). desktop 시스템 컨테이너 파일 전체 4회 중 1회 79번째(그 파일에서 처음 provider를 띄우는 시험)부터 18개가 exit=125, 진단을 넣은 뒤 메시지는 `provider cgroup enforcement failed: cannot read memory.max`. scripts/run-in-system-container.sh가 부팅 중에 띄운 docker exec가 systemd가 root cgroup에 제어기를 켜기 전에 root cgroup에 들어가, cgroup v2가 그 cgroup에 제어기 켜기를 막아 cpu·memory·pids가 user manager까지 넘어가지 않았다. 부팅 직후 exec 15개 탐침이 6/6 결정적으로 재현. 고친 트리는 탐침 6/6 정상, 파일 전체 16회 0/16", "quarantined_until": "", "fixed_by": "프론트 #58"}
{"date": "2026-11-14", "test": "tests/system/ci-runner/ci-artifact-contract.test.ts:983", "category": "concurrency-order", "first_run": "/tmp/claude-gates/fe125-loop1-runs/run-1.log", "verdict": "desktop 시스템 컨테이너 파일 전체 4회 중 1회 실패(시작 load 15.34). 기대한 `provider guardian failed|SIGKILL` 대신 supervisor가 guardian commit 쓰기(scripts/lib/provider-guardian-client.ts:286)의 처리되지 않은 `write EPIPE`로 죽었다. 원인 조사 전", "quarantined_until": "", "fixed_by": ""}
+13 -1
View File
@@ -239,10 +239,22 @@ runner의 Docker proxy가 안쪽 dind daemon에 컨테이너를 띄운다.
release 단계 provider job은 아직 job 컨테이너에서 돈다.
- runner가 privileged 컨테이너를 막는 설정(rootless dind 등)으로 바뀌면 이
길이 깨진다. 그때는 위 조건을 갖춘 전용 runner(VM 등)로 옮긴다.
- 컨테이너의 첫 프로세스는 systemd를 띄우기 전에 root cgroup의 프로세스를
`init.scope`로 옮기고 root에 cgroup 제어기를 켠다. 그러지 않으면 부팅 중에
닿은 `docker exec`가 root cgroup에 들어가 systemd가 제어기를 넘기지 못하고,
provider scope에 `memory.max`가 없어 아래 확인이 provider를 exit 125로
멈춘다(2026-11-14 desktop 4회 중 1회). `test:system`의 사전 점검
(`scripts/check-system-test-prerequisites.ts`)이 user manager에 cpu·memory·pids
제어기가 넘어왔는지 보고, 없으면 빠진 이름을 대며 실패한다.
각 provider는 고유한 collected user scope에서 실행된다. Supervisor는 실행 전에
실제 cgroup membership과 `memory.max=1073741824`, `memory.swap.max=0`,
`pids.max=64`, `cpu.max="100000 100000"`을 확인한다. 내부 process에는 core 0,
`pids.max=64`, `cpu.max="100000 100000"`을 확인한다. 확인이나 scope 준비가
provider 전에 실패하면 supervisor 오류 끝에 이유가 붙는다
(`sandbox reported: provider cgroup enforcement failed: ...`). 남기는 줄은 8줄까지,
우리 단계가 쓰는 고정 문구와 sandbox 도구의 접두 줄(`bwrap:` 등)뿐이다. provider
출력은 자격 증명을 실을 수 있어 그 밖의 줄은 남기지 않는다
(`scripts/lib/provider-sandbox-diagnostics.ts`). 내부 process에는 core 0,
file-size 8,388,607 bytes, open FD 64, CPU 1,200초 상한도 적용된다.
같은 UID 전체에 합산되는 `RLIMIT_NPROC`로 provider별 32개를 보장한다고 주장하지
않으며 aggregate PID authority는 cgroup `TasksMax=64`다.
@@ -71,10 +71,13 @@ export async function findSystemTestPrerequisiteFailures(
}
}
let cgroupV2 = false;
try {
const controllers = await host.readFile("/sys/fs/cgroup/cgroup.controllers");
if (controllers.trim().length === 0) {
failures.push("cgroup v2 controllers are unavailable");
} else {
cgroupV2 = true;
}
} catch {
failures.push("cgroup v2 is required at /sys/fs/cgroup/cgroup.controllers");
@@ -87,6 +90,10 @@ export async function findSystemTestPrerequisiteFailures(
}
}
// The user manager's cgroup path below is a cgroup v2 path; without v2 it would name the wrong cause.
const controllerFailure = cgroupV2 ? await findProviderScopeControllerFailure(host) : undefined;
if (controllerFailure !== undefined) failures.push(controllerFailure);
if (executable.get(BWRAP)) {
const failure = await findProviderSandboxNamespaceFailure(host);
if (failure !== undefined) failures.push(failure);
@@ -141,6 +148,35 @@ async function namespaceFailure(
].join("\n");
}
/** The controllers every provider scope is limited by (scripts/lib/provider-cgroup.ts). */
const PROVIDER_SCOPE_CONTROLLERS = Object.freeze(["cpu", "memory", "pids"] as const);
/**
* Whether the systemd user manager that provider scopes run under was handed
* the cgroup controllers their limits need. A manager without them passed
* every check above, and then each provider stopped with `provider cgroup
* enforcement failed: cannot read memory.max` (2026-11-14: a `docker exec`
* sat in the system container's root cgroup while systemd started, so systemd
* delegated nothing).
*/
async function findProviderScopeControllerFailure(host: SystemTestHost): Promise<string | undefined> {
const status = await host.readFile("/proc/self/status").catch(() => "");
const uid = /^Uid:\s+(\d+)\s/mu.exec(status)?.[1];
if (uid === undefined) return "this process's uid is unreadable at /proc/self/status";
const manager = `/sys/fs/cgroup/user.slice/user-${uid}.slice/user@${uid}.service`;
let delegated: string[];
try {
delegated = (await host.readFile(`${manager}/cgroup.controllers`)).trim().split(/\s+/u);
} catch {
return `a systemd user manager is required at ${manager}`;
}
const missing = PROVIDER_SCOPE_CONTROLLERS.filter((controller) => !delegated.includes(controller));
return missing.length === 0
? undefined
: `the systemd user manager at ${manager} has no ${missing.join(", ")} cgroup controller ` +
`for provider scopes (it has: ${delegated.join(" ") || "none"})`;
}
export function formatSystemTestPrerequisiteReport(failures: readonly string[]): string {
return [
"System/CI-runner test prerequisites are unavailable:",
@@ -0,0 +1,93 @@
/**
* Lines the sandbox tooling itself emits, kept so a launch failure can say
* why. Everything else the child writes is provider output and may carry
* credentials, so it is counted and discarded as before.
*
* Without this a sandbox that never started reported only `exit=1`, and the
* actual cause — `bwrap: loopback: Failed RTM_NEWADDR: Operation not
* permitted` on a host with `kernel.apparmor_restrict_unprivileged_userns=1`
* — was invisible. That turned a host restriction into an unexplained
* product failure.
*/
const SANDBOX_TOOL = /^(?:bwrap|prlimit|systemd-run|systemctl):\s.*$/u;
/**
* Our own launch stages exit 125 before the provider runs: the cgroup gate
* (provider-cgroup.ts ENFORCEMENT_GATE) and the scope wrapper
* (provider-scope-wrapper.ts). Their lines are matched as whole templates,
* not by prefix, so provider output that merely starts the same way is still
* discarded. Only the liveness line carries free text, Node's own error for
* descriptor 0.
*/
/** An errno or Node error name, the only free part of the scope wrapper's stop line. */
const ERROR_CODE = String.raw`(?:E[A-Z0-9]{1,15}|ERR_[A-Z0-9_]{1,40})`;
const CGROUP_FILE = String.raw`(?:memory\.max|memory\.swap\.max|pids\.max|cpu\.max)`;
const CGROUP_VALUE = String.raw`(?:max|[0-9]{1,20})(?: [0-9]{1,20})?`;
const LAUNCH_STAGE = [
new RegExp(
String.raw`^provider cgroup enforcement failed: (?:unified cgroup v2 membership is required|` +
String.raw`unit membership is invalid|cannot read ${CGROUP_FILE}|` +
String.raw`${CGROUP_FILE} is (?:${CGROUP_VALUE})?, expected ${CGROUP_VALUE})$`,
"u",
),
new RegExp(
String.raw`^provider scope stopped before launch: (?:supervisor channel ended|` +
String.raw`supervisor channel error (?:${ERROR_CODE}|UNKNOWN)|protocol failure)$`,
"u",
),
/^provider scope (?:received trailing protocol bytes|frame length is invalid|frame has trailing bytes)$/u,
/^provider scope liveness channel is unavailable: [\x20-\x7e]{1,200}$/u,
] as const;
const MAX_LINES = 8;
/** A diagnostic line is short; anything longer is provider output. */
const MAX_LINE_CHARACTERS = 512;
export type ProviderSandboxDiagnostics = Readonly<{
consume(stream: "stdout" | "stderr", chunk: Buffer | string): void;
lines(): readonly string[];
}>;
/** Whether the scope wrapper may name this error code in its stop line. */
export function isProviderScopeErrorCode(code: string): boolean {
return new RegExp(`^${ERROR_CODE}$`, "u").test(code);
}
export function isProviderSandboxDiagnostic(line: string): boolean {
return SANDBOX_TOOL.test(line) || LAUNCH_STAGE.some((pattern) => pattern.test(line));
}
/**
* Reassembles each stream into whole lines before matching, so a line split
* across two reads is neither missed nor kept in a truncated form. A line is
* only judged once its newline arrives.
*/
export function createProviderSandboxDiagnostics(): ProviderSandboxDiagnostics {
const partial = { stdout: "", stderr: "" };
const overlong = { stdout: false, stderr: false };
const kept: string[] = [];
const keep = (line: string): void => {
if (
kept.length < MAX_LINES && line.length <= MAX_LINE_CHARACTERS &&
!kept.includes(line) && isProviderSandboxDiagnostic(line)
) {
kept.push(line);
}
};
return Object.freeze({
consume: (stream: "stdout" | "stderr", chunk: Buffer | string) => {
const pieces = String(chunk).split("\n");
const tail = pieces.pop() ?? "";
for (const piece of pieces) {
if (!overlong[stream]) keep(partial[stream] + piece);
partial[stream] = "";
overlong[stream] = false;
}
if (overlong[stream]) return;
partial[stream] += tail;
if (partial[stream].length > MAX_LINE_CHARACTERS) {
partial[stream] = "";
overlong[stream] = true;
}
},
lines: () => Object.freeze([...kept]),
});
}
+16 -4
View File
@@ -4,6 +4,7 @@ import { Socket } from "node:net";
import { traceOwnProviderProcess, traceSpawnedProviderProcess } from "./provider-process-trace.ts";
import { cleanupOwnedProviderReport } from "./provider-raw-cleanup.ts";
import { isProviderScopeErrorCode } from "./provider-sandbox-diagnostics.ts";
const MAX_FRAME_BYTES = 16_777_216;
const reportIdentity = parseReportIdentity(process.argv.slice(2));
@@ -13,6 +14,8 @@ let expectedBytes: number | undefined;
let provider: ReturnType<typeof spawn> | undefined;
let providerClosed = false;
let livenessLost = false;
/** Why the scope stopped before it launched the provider, for its exit-125 line. */
let stopReason = "supervisor channel ended";
/**
* The supervisor keeps this pipe open for the scope's whole life — that is how
* parent loss is observed — and only ever writes one frame into it.
@@ -47,8 +50,8 @@ liveness.on("data", (chunk: Buffer | string) => {
}
});
liveness.once("end", () => terminateForParentLoss());
liveness.once("error", () => terminateForParentLoss());
liveness.once("end", () => terminateForParentLoss("supervisor channel ended"));
liveness.once("error", (error) => terminateForParentLoss(`supervisor channel error ${errorCode(error)}`));
function launchProvider(payload: Buffer): void {
const frame = parseFrame(payload);
@@ -96,9 +99,10 @@ async function finishProvider(
process.exit(code ?? 1);
}
function terminateForParentLoss(): void {
function terminateForParentLoss(reason: string): void {
if (livenessLost) return;
livenessLost = true;
stopReason = reason;
if (!provider || providerClosed) {
void cleanupAfterParentLossAndExit();
return;
@@ -117,12 +121,15 @@ async function cleanupAfterParentLossAndExit(): Promise<void> {
writeSync(2, `${error instanceof Error ? error.message : String(error)}\n`);
}
closeLivenessInput();
// The supervisor keeps only fixed launch-stage lines (provider-sandbox-diagnostics.ts);
// without this one an exit 125 here was indistinguishable from the cgroup gate's.
if (!provider) writeSync(2, `provider scope stopped before launch: ${stopReason}\n`);
process.exit(125);
}
function terminateForProtocolFailure(message: string): void {
writeSync(2, `${message}\n`);
terminateForParentLoss();
terminateForParentLoss("protocol failure");
}
function openLivenessChannel(): Socket {
@@ -212,6 +219,11 @@ function signalNumber(signal: NodeJS.Signals): number {
return signal === "SIGKILL" ? 9 : signal === "SIGXCPU" ? 24 : 1;
}
function errorCode(error: unknown): string {
const code = error && typeof error === "object" && "code" in error ? String(error.code) : "";
return isProviderScopeErrorCode(code) ? code : "UNKNOWN";
}
function hasErrorCode(error: unknown, code: string): boolean {
return Boolean(error && typeof error === "object" && "code" in error && error.code === code);
}
+9 -23
View File
@@ -30,6 +30,7 @@ import {
type ProviderGuardianLease,
} from "./lib/provider-guardian-client.ts";
import { createProviderOutputLimiter } from "./lib/provider-output-limiter.ts";
import { createProviderSandboxDiagnostics } from "./lib/provider-sandbox-diagnostics.ts";
import { findProviderSandboxNamespaceFailure } from "./check-system-test-prerequisites.ts";
import { PROVIDER_SANDBOX_NAMESPACE_ARGUMENTS } from "./lib/provider-sandbox-namespaces.ts";
import {
@@ -400,30 +401,15 @@ async function waitForProvider(
PROVIDER_MAX_OUTPUT_BYTES,
() => terminate("output"),
);
/**
* Lines the sandbox tooling itself emits, kept so a launch failure can say
* why. Everything else the child writes is provider output and may carry
* credentials, so it is counted and discarded as before.
*
* Without this a sandbox that never started reported only `exit=1`, and the
* actual cause — `bwrap: loopback: Failed RTM_NEWADDR: Operation not
* permitted` on a host with `kernel.apparmor_restrict_unprivileged_userns=1`
* — was invisible. That turned a host restriction into an unexplained
* product failure.
*/
const SANDBOX_DIAGNOSTIC = /^(?:bwrap|prlimit|systemd-run|systemctl):\s.*$/gmu;
const sandboxDiagnostics: string[] = [];
const capture = (chunk: Buffer | string): void => {
for (const line of String(chunk).matchAll(SANDBOX_DIAGNOSTIC)) {
if (sandboxDiagnostics.length < 8 && !sandboxDiagnostics.includes(line[0])) {
sandboxDiagnostics.push(line[0]);
}
}
// Only lines our own launch stages write survive (lib/provider-sandbox-diagnostics.ts).
const sandboxDiagnostics = createProviderSandboxDiagnostics();
const capture = (stream: "stdout" | "stderr") => (chunk: Buffer | string): void => {
sandboxDiagnostics.consume(stream, chunk);
if (termination) return;
outputLimiter.consume(chunk);
};
child.stdout?.on("data", capture);
child.stderr?.on("data", capture);
child.stdout?.on("data", capture("stdout"));
child.stderr?.on("data", capture("stderr"));
const timeout = setTimeout(() => terminate("timeout"), timeoutMs);
try {
const first = await Promise.race([
@@ -475,8 +461,8 @@ async function waitForProvider(
throw new Error(
`sandboxed external provider failed: exit=${result.code ?? "none"}, ` +
`signal=${result.signal ?? "none"}` +
(sandboxDiagnostics.length > 0
? `; sandbox reported: ${sandboxDiagnostics.join("; ")}`
(sandboxDiagnostics.lines().length > 0
? `; sandbox reported: ${sandboxDiagnostics.lines().join("; ")}`
: ""),
);
}
+24 -1
View File
@@ -42,9 +42,32 @@ if ! $docker image inspect "$tag" > /dev/null 2>&1; then
say "building $tag"
$docker build -q -t "$tag" -f "$dockerfile" config/ci > /dev/null || fail "image build failed"
fi
cid=$($docker run -d --privileged --cgroupns=private --tmpfs /run --tmpfs /run/lock "$tag") \
# systemd hands the cgroup controllers the provider scope needs (cpu, memory, pids) down from the container's
# root cgroup. A `docker exec` that arrives while systemd starts, before it has enabled controllers there,
# joins that root cgroup, and while any process sits in it the kernel refuses to enable controllers on it
# (cgroup v2: a cgroup that hands controllers to children holds no processes). systemd then delegates
# nothing, every later exec joins the root cgroup too, and each provider scope lacks memory.max, so the
# provider cgroup gate exits 125 (1 run in 4 on desktop, 2026-11-14). So the container's first process
# moves every process out of the root cgroup and enables its controllers before it execs systemd, as
# moby's hack/dind does. From then on an exec cannot join the root cgroup and lands in PID 1's.
boot='cg=/sys/fs/cgroup
controllers=$(cat $cg/cgroup.controllers 2> /dev/null) || controllers=
if [ -n "$controllers" ]; then
mkdir -p $cg/init.scope
enable=$(printf "+%s " $controllers)
tries=0
until { xargs -rn1 < $cg/cgroup.procs > $cg/init.scope/cgroup.procs; echo "$enable" > $cg/cgroup.subtree_control; } 2> /dev/null; do
tries=$((tries + 1))
[ $tries -lt 50 ] || { echo "system-container: cannot enable the root cgroup controllers" >&2; break; }
sleep 0.1
done
fi
exec /sbin/init'
cid=$($docker run -d --privileged --cgroupns=private --tmpfs /run --tmpfs /run/lock "$tag" /bin/sh -c "$boot") \
|| fail "container did not start"
cleanup() {
# The boot step writes to the container log only when it gave up; say so whatever ends the run.
$docker logs "$cid" 2>&1 | grep '^system-container:' >&2 || :
if [ "${SYSTEM_CONTAINER_KEEP:-0}" = 1 ]; then say "kept container $cid"; else $docker rm -f "$cid" > /dev/null 2>&1 || :; fi
}
trap cleanup EXIT
@@ -0,0 +1,147 @@
import { spawnSync } from "node:child_process";
import { mkdtemp, rm, writeFile } from "node:fs/promises";
import { tmpdir } from "node:os";
import path from "node:path";
import { describe, expect, it } from "vitest";
import {
formatProviderCgroupUnitName,
systemdRunProviderArguments,
} from "../../scripts/lib/provider-cgroup.ts";
import { createProviderSandboxDiagnostics } from "../../scripts/lib/provider-sandbox-diagnostics.ts";
import { childProcessTimeout } from "../helpers/child-process-timeout.ts";
const CREDENTIAL = "provider-credential-must-not-reach-ci";
const GATE_LINE = "provider cgroup enforcement failed: ";
/** The command systemd-run starts in the scope: `/bin/sh -eu -c <gate> <unit> …`. */
function gateShell(): string[] {
const unit = formatProviderCgroupUnitName("vulnerability", 4242, "0".repeat(24));
const systemdRun = systemdRunProviderArguments(unit, 1_000, 1, "/bin/node", "/wrapper.ts", "/r", 1, 1);
return systemdRun.slice(systemdRun.indexOf("--") + 1);
}
describe("provider sandbox diagnostics", () => {
it("keeps the line the real cgroup gate writes when it refuses a scope", () => {
// Outside a provider scope the gate's own membership check fails, which is
// the same shell, text, and exit code a broken scope produces.
const shell = gateShell();
const gate = spawnSync(shell[0]!, shell.slice(1), { encoding: "utf8" });
expect(gate.status).toBe(125);
expect(gate.stderr).toMatch(/^provider cgroup enforcement failed: \S/u);
const diagnostics = createProviderSandboxDiagnostics();
diagnostics.consume("stderr", gate.stderr);
expect(diagnostics.lines()).toEqual([gate.stderr.trimEnd()]);
}, childProcessTimeout(4));
it("keeps the line the real gate check writes for every unreadable or wrong limit", async () => {
// The gate's own limit check and its own calls, run against a stand-in
// cgroup directory: 2026-11-14's exit 125 was `cannot read memory.max`.
const gate = gateShell()[3]!;
const start = gate.indexOf("require_cgroup_value() {");
const definition = gate.slice(start, gate.indexOf("\n}\n", start) + 2);
const calls = gate.split("\n").filter((line) => line.startsWith("require_cgroup_value "));
expect(calls).toHaveLength(4);
const root = await mkdtemp(path.join(tmpdir(), "provider-cgroup-limits-"));
try {
for (const content of [undefined, "max\n"]) {
for (const call of calls) {
const file = path.join(root, call.split(" ")[1]!);
if (content === undefined) await rm(file, { force: true });
else await writeFile(file, content);
const check = spawnSync("/bin/sh", ["-eu", "-c", `cgroup_root=$1\n${definition}\n${call}`, "check", root], {
encoding: "utf8",
});
expect(check.status, call).toBe(125);
const written = check.stderr.split("\n").filter((line) => line.startsWith(GATE_LINE));
expect(written, call).toHaveLength(1);
const diagnostics = createProviderSandboxDiagnostics();
diagnostics.consume("stderr", check.stderr);
expect(diagnostics.lines(), call).toEqual(written);
}
}
} finally {
await rm(root, { recursive: true, force: true });
}
}, childProcessTimeout(40));
it.each([
"provider cgroup enforcement failed: unified cgroup v2 membership is required",
"provider cgroup enforcement failed: unit membership is invalid",
"provider cgroup enforcement failed: cannot read memory.max",
"provider cgroup enforcement failed: memory.max is max, expected 1073741824",
"provider cgroup enforcement failed: memory.swap.max is , expected 0",
"provider cgroup enforcement failed: cpu.max is max 100000, expected 100000 100000",
"provider scope stopped before launch: supervisor channel ended",
"provider scope stopped before launch: supervisor channel error EPIPE",
"provider scope stopped before launch: supervisor channel error ERR_STREAM_DESTROYED",
"provider scope stopped before launch: supervisor channel error UNKNOWN",
"provider scope stopped before launch: protocol failure",
"provider scope frame length is invalid",
"provider scope liveness channel is unavailable: Unsupported fd type: FILE",
"bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted",
])("keeps the launch-stage line %j", (line) => {
const diagnostics = createProviderSandboxDiagnostics();
diagnostics.consume("stderr", `${line}\n`);
expect(diagnostics.lines()).toEqual([line]);
});
it.each([
CREDENTIAL,
`provider cgroup enforcement failed: memory.max is ${CREDENTIAL}, expected 1073741824`,
`provider cgroup enforcement failed: cannot read ${CREDENTIAL}`,
`provider scope stopped before launch: ${CREDENTIAL}`,
`provider scope stopped before launch: supervisor channel error ${CREDENTIAL}`,
// Upper case alone is not enough: a key-like value is not an errno or Node error name.
"provider scope stopped before launch: supervisor channel error ABCDEFGHIJKLMNOPQRST",
` provider cgroup enforcement failed: unit membership is invalid`,
// A whole template followed by more text is provider output, not ours.
`provider cgroup enforcement failed: cannot read memory.max ${CREDENTIAL}`,
`provider cgroup enforcement failed: memory.max is max, expected 1073741824 ${CREDENTIAL}`,
`provider scope stopped before launch: supervisor channel ended ${CREDENTIAL}`,
`provider scope frame length is invalid ${CREDENTIAL}`,
`provider scope liveness channel is unavailable: ${"x".repeat(201)}`,
`bwrap:${CREDENTIAL}`,
])("drops provider output that only resembles a launch-stage line: %j", (line) => {
const diagnostics = createProviderSandboxDiagnostics();
diagnostics.consume("stdout", `${line}\n`);
diagnostics.consume("stderr", `${line}\n`);
expect(diagnostics.lines()).toEqual([]);
expect(JSON.stringify(diagnostics.lines())).not.toContain(CREDENTIAL);
});
it("judges a line only once it is whole, per stream", () => {
const diagnostics = createProviderSandboxDiagnostics();
diagnostics.consume("stderr", "provider cgroup enforcement failed: cannot ");
diagnostics.consume("stdout", "provider output\n");
expect(diagnostics.lines()).toEqual([]);
diagnostics.consume("stderr", "read pids.max\nprovider cgroup enforcement failed: unit");
expect(diagnostics.lines()).toEqual(["provider cgroup enforcement failed: cannot read pids.max"]);
});
it("drops the rest of an overlong line even when that rest arrives as its own write", () => {
const diagnostics = createProviderSandboxDiagnostics();
diagnostics.consume("stderr", "x".repeat(600));
diagnostics.consume("stderr", `bwrap: ${CREDENTIAL}\n`);
expect(diagnostics.lines()).toEqual([]);
});
it("drops an overlong line even when it starts like a diagnostic", () => {
const diagnostics = createProviderSandboxDiagnostics();
diagnostics.consume("stderr", `bwrap: ${"x".repeat(600)}`);
diagnostics.consume("stderr", `${CREDENTIAL}\nbwrap: ${"y".repeat(600)}\n`);
expect(diagnostics.lines()).toEqual([]);
});
it("keeps at most eight distinct lines", () => {
const diagnostics = createProviderSandboxDiagnostics();
for (let index = 0; index < 12; index += 1) {
diagnostics.consume("stderr", `bwrap: failure ${index % 10}\nbwrap: failure ${index % 10}\n`);
}
expect(diagnostics.lines()).toEqual(
Array.from({ length: 8 }, (_unused, index) => `bwrap: failure ${index}`),
);
});
});
+47 -2
View File
@@ -15,6 +15,7 @@ type SpawnCall = Readonly<{ command: string; args: readonly string[] }>;
type SpawnResult = ReturnType<SystemTestHost["spawn"]>;
const RTM_NEWADDR = "bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted";
const USER_MANAGER_CONTROLLERS = "/sys/fs/cgroup/user.slice/user-1000.slice/user@1000.service/cgroup.controllers";
/** A host that has every prerequisite, except what `bwrap` and `files` say. */
function fakeHost(
@@ -25,8 +26,12 @@ function fakeHost(
const host: SystemTestHost = {
access: async () => undefined,
readFile: async (target) => {
if (target === "/sys/fs/cgroup/cgroup.controllers") return "cpuset cpu io memory pids\n";
const content = files[target];
const content = {
"/sys/fs/cgroup/cgroup.controllers": "cpuset cpu io memory pids\n",
"/proc/self/status": "Name:\tnode\nUid:\t1000\t1000\t1000\t1000\n",
[USER_MANAGER_CONTROLLERS]: "cpu memory pids\n",
...files,
}[target];
if (content === undefined) throw new Error(`ENOENT: ${target}`);
return content;
},
@@ -125,6 +130,46 @@ describe("system test prerequisites (1105/12-01)", () => {
});
});
// 2026-11-14. In one system container in four, systemd delegated no cgroup
// controller to the root user manager; this check printed PASS and then every
// provider stopped with `provider cgroup enforcement failed: cannot read
// memory.max`.
describe("system test prerequisites: provider scope controllers", () => {
it.each([
["none", "\n", "cpu, memory, pids", "none"],
["memory", "cpu pids\n", "memory", "cpu pids"],
])("names the controllers the user manager lacks (%s missing)", async (_label, delegated, missing, has) => {
const { host } = fakeHost({ status: 0, stderr: "" }, { [USER_MANAGER_CONTROLLERS]: delegated });
expect(await findSystemTestPrerequisiteFailures(host)).toEqual([
"the systemd user manager at /sys/fs/cgroup/user.slice/user-1000.slice/user@1000.service " +
`has no ${missing} cgroup controller for provider scopes (it has: ${has})`,
]);
});
it("reads the user manager only where cgroup v2 is there", async () => {
const { host } = fakeHost({ status: 0, stderr: "" }, { "/sys/fs/cgroup/cgroup.controllers": "\n", [USER_MANAGER_CONTROLLERS]: "\n" });
expect(await findSystemTestPrerequisiteFailures(host)).toEqual(["cgroup v2 controllers are unavailable"]);
});
it("names an unreadable uid", async () => {
const { host } = fakeHost({ status: 0, stderr: "" }, { "/proc/self/status": "Name:\tnode\n" });
expect(await findSystemTestPrerequisiteFailures(host)).toEqual([
"this process's uid is unreadable at /proc/self/status",
]);
});
it("requires a user manager for this process's own uid", async () => {
const { host } = fakeHost({ status: 0, stderr: "" }, { "/proc/self/status": "Uid:\t0\t0\t0\t0\n" });
expect(await findSystemTestPrerequisiteFailures(host)).toEqual([
"a systemd user manager is required at /sys/fs/cgroup/user.slice/user-0.slice/user@0.service",
]);
});
});
// 1101/13-03. The release provider jobs run the same sandbox on whatever
// runner picks them up. The supervisor checked only that the executables
// exist, so a host that refuses the namespaces failed later with bwrap's raw
@@ -10,6 +10,7 @@ import { afterEach, describe, expect, it } from "vitest";
import { loadCiGateContract, parseCiGateContract } from "../../scripts/contracts/ci-gates.ts";
import { generateCiWorkflow } from "../../scripts/generate-ci-workflow.ts";
import { validatePackageScriptGraph } from "../../scripts/lib/package-script-graph.ts";
import { createProviderSandboxDiagnostics } from "../../scripts/lib/provider-sandbox-diagnostics.ts";
import { childProcessTimeout } from "../helpers/child-process-timeout.ts";
const roots: string[] = [];
@@ -167,6 +168,9 @@ describe("selective Task 3 contract closure", () => {
String(identity.ino),
]);
const completion = waitForChildResult(child);
let stderr = "";
child.stderr!.setEncoding("utf8");
child.stderr!.on("data", (chunk: string) => { stderr += chunk; });
await waitForScopeWrapperLoaded(child);
// Time to exit on its own, which it must not do before EOF.
await new Promise<void>((resolve) => setTimeout(resolve, 75));
@@ -175,6 +179,10 @@ describe("selective Task 3 contract closure", () => {
child.stdin!.end();
const result = await within(completion, PROCESS_WATCHDOG_MS, "provider scope EOF close");
expect(result).toEqual({ code: 125, signal: null });
// The supervisor keeps this fixed line, so its exit-125 report names the stage.
const diagnostics = createProviderSandboxDiagnostics();
diagnostics.consume("stderr", stderr);
expect(diagnostics.lines()).toEqual(["provider scope stopped before launch: supervisor channel ended"]);
await expect(lstat(reportPath)).rejects.toMatchObject({ code: "ENOENT" });
await expect(readdir(root)).resolves.toEqual([]);
}, childProcessTimeout(433));