Merge pull request #1224 from DeusData/fix/daemon-app-tsan-wait-budget

test: size the daemon-application wait budget for TSan's slowdown tail

scripts/run-tests-parallel.sh

@@ -48,16 +48,32 @@ mkdir -p "$LOGDIR"
 # setup. Two idempotent steps: protect the DIRECTORY (inheritance flags are
 # directory-only — a /T re-root leaves files with empty deny-all DACLs),
 # then /reset the children to re-inherit the clean set.
-case "$(uname -s 2>/dev/null)" in
-MINGW* | MSYS*)
+stamp_windows_build_dir() {
+    local when="$1"
+    case "$(uname -s 2>/dev/null)" in
+    MINGW* | MSYS*) ;;
+    *) return 0 ;;
+    esac
+    local runner_dir_w me stamp_out reset_out
     runner_dir_w="$(cygpath -w "$(dirname "$RUNNER")")"
     me="$(whoami | tr -d '\r')"
-    MSYS2_ARG_CONV_EXCL='*' icacls "$runner_dir_w" /inheritance:r \
+    stamp_out=$(MSYS2_ARG_CONV_EXCL='*' icacls "$runner_dir_w" /inheritance:r \
         /grant:r "${me}:(OI)(CI)F" '*S-1-5-18:(OI)(CI)F' '*S-1-5-32-544:(OI)(CI)F' \
-        /Q >/dev/null 2>&1 || true
-    MSYS2_ARG_CONV_EXCL='*' icacls "${runner_dir_w}\\*" /reset /T /C /Q >/dev/null 2>&1 || true
-    ;;
-esac
+        /Q 2>&1) || echo "WARN: build-dir DACL stamp ($when) failed (user=$me dir=$runner_dir_w): $stamp_out"
+    reset_out=$(MSYS2_ARG_CONV_EXCL='*' icacls "${runner_dir_w}\\*" /reset /T /C /Q 2>&1) ||
+        echo "WARN: build-dir child DACL reset ($when) failed: $(printf '%s' "$reset_out" | tail -2)"
+    # The stamp is load-bearing for the install-flow suites: verify it and say
+    # so, in either direction — a silent stamp once cost a full CI round to
+    # even see WHETHER it had run.
+    if MSYS2_ARG_CONV_EXCL='*' icacls "$runner_dir_w" 2>/dev/null |
+        grep -qE 'Authenticated Users|CREATOR OWNER'; then
+        echo "WARN: build-dir DACL still grants cross-account mutation after $when stamp:"
+        MSYS2_ARG_CONV_EXCL='*' icacls "$runner_dir_w" 2>&1 | head -8
+    else
+        echo "build-dir DACL stamped clean ($when, $runner_dir_w, user=$me)"
+    fi
+}
+stamp_windows_build_dir pre-wave
 
 SUITES_FILE="$LOGDIR/suites.txt"
 RESULTS_FILE="$LOGDIR/results.txt"
@@ -232,6 +248,13 @@ while IFS= read -r sname; do
         echo "$sname" >> "$FLEX_FILE"
     fi
 done < "$SER_FILE"
+# Wave suites spawn Cygwin-family tooling that can rewrite the build
+# directory's DACL behind the first stamp (observed: an arm shard whose
+# tail held the install-flow suites failed the source-directory policy
+# minutes after a clean pre-wave stamp, while its sibling shard passed).
+# Re-stamp at the tail boundary so the deadline-sensitive tail — which
+# hosts those suites — always starts from the verified-clean shape.
+stamp_windows_build_dir pre-tail
 xargs -P "${CBM_TAIL_JOBS:-2}" -I{} bash -c 'run_one "$@"' _ {} < "$FLEX_FILE"
 while IFS= read -r sname; do
     run_one "$sname"

src/daemon/service.c

@@ -856,12 +856,26 @@ static bool windows_private_file_prepare(const char *directory, const char *base
      * to the ~16 ms timer granularity, and a permanently obstructed path
      * (e.g. a directory squatting on it) must fail fast — a rejected client
      * is waiting behind this on the hello path with its own timeout. */
-    ULONGLONG deadline = GetTickCount64() + 250;
+    ULONGLONG obstruction_deadline = GetTickCount64() + 250;
+    /* Genuine share collisions get a wider budget than obstruction: with
+     * many concurrent appenders on a loaded, sanitized CI runner the
+     * last-in-line writer can legitimately spend more than 250 ms behind
+     * its peers' brief exclusive windows, and the no-drop contract for
+     * conflict events outranks latency there — while a permanently
+     * obstructed path (a directory squatting on the name) still fails
+     * inside 250 ms, because a rejected client is waiting behind this on
+     * the hello path with its own timeout. */
+    ULONGLONG contention_deadline = GetTickCount64() + 2000;
     for (;;) {
+        SetLastError(ERROR_SUCCESS);
         FILE *file = cbm_daemon_ipc_private_log_open(directory, base, SIZE_MAX);
         if (file) {
             return fclose(file) == 0;
         }
+        DWORD open_error = GetLastError();
+        bool transient_collision =
+            open_error == ERROR_SHARING_VIOLATION || open_error == ERROR_LOCK_VIOLATION;
+        ULONGLONG deadline = transient_collision ? contention_deadline : obstruction_deadline;
         if (GetTickCount64() >= deadline) {
             return false;
         }

tests/test_daemon_application.c

@@ -27,7 +27,12 @@
 #include <stdlib.h>
 #include <string.h>
 
-enum { APP_TEST_TIMEOUT_MS = 2000, APP_TEST_PATH_CAP = 1024 };
+/* Observation-wait budget. This is a HANG DETECTOR, not a latency
+ * assertion: every waiter returns the moment its condition holds, so green
+ * runs never pay it. It must absorb ThreadSanitizer's 5-20x slowdown on a
+ * loaded 3-core CI runner — at 2000 ms the cancel-delivery wait lost the
+ * tail of that distribution once in seven otherwise-green TSan rounds. */
+enum { APP_TEST_TIMEOUT_MS = 10000, APP_TEST_PATH_CAP = 1024 };
 
 typedef struct {
     char runtime_parent[APP_TEST_PATH_CAP];