diff --git a/scripts/redeploy-fleetd.sh b/scripts/redeploy-fleetd.sh index 5f94df8..4be9e6a 100755 --- a/scripts/redeploy-fleetd.sh +++ b/scripts/redeploy-fleetd.sh @@ -169,7 +169,37 @@ hash256() { # on PATH). "absent" must never be the answer for a file that exists — that conflation, on Linux, # was the whole defect this ticket fixes. jar_id() { local f="${1:-$JAR}"; [ -f "$f" ] && hash256 "$f" || echo "absent"; } -running_pid() { pgrep -f "$PATTERN" || true; } + +# fleetd #593 — `pgrep -f "$PATTERN"` matches ANY process whose full command line CONTAINS the +# pattern text, and that is not the same thing as "is the daemon". A shell that merely embeds the +# pattern as literal text — a human typing this exact investigation by hand, an ssh-shaped +# `sh -c '...; ...'`, a pipeline, or any other non-exec'ing shell that never replaced itself with +# the pattern-holding command — still shows up in that match, and it is the INSTRUMENT, not the +# daemon. Measured live on this Mac: `sh -c 'echo "target/fleetd.jar" >/dev/null; sleep 30' &` +# leaves a real `sh` process alive (it forks for the `sleep`, it does not exec into it) whose own +# `ps -o args` is `sh -c echo "target/fleetd.jar" >/dev/null; sleep 30` — `pgrep -f "$PATTERN"` +# matches that line right alongside the real `java -jar target/fleetd.jar` process. `pgrep -c` +# (an in-one-call count) does not exist on BSD/macOS at all, so this cannot be fixed by switching +# pgrep flags — it has to filter what pgrep already found, after the fact, in a way that still +# runs on BSD. The daemon is always started as `java -jar target/fleetd.jar`, so its process name +# (comm) is always `java`; a self-matching wrapper of this shape is always a shell. So: keep pgrep +# to FIND candidates (that part is already portable), then drop any candidate whose `comm` names a +# shell — the wrapper is excluded, a genuine second daemon-shaped process (java, or in a test, +# anything that is not itself a shell) still counts. `comm` can carry a login shell's leading '-' +# (e.g. "-bash") or, on some `ps` builds, a full path — both are stripped before the comparison. +running_pid() { + local pid comm out='' + for pid in $(pgrep -f "$PATTERN" 2>/dev/null || true); do + comm="$(ps -o comm= -p "$pid" 2>/dev/null || true)" + comm="${comm##*/}" + comm="${comm#-}" + case "$comm" in + sh|bash|zsh|dash|ksh) continue ;; + esac + out="$out$pid"$'\n' + done + printf '%s' "$out" +} # fleetd #493 — three small, independently testable pieces of "never build into the path a # running process holds": @@ -507,8 +537,11 @@ assert_single_daemon() { die "more than one fleetd process is running after this restart (pids: $(printf '%s' "$pids" | tr '\n' ' ')). This is the exact failure a racing supervisor produces: the OLD jar was revived by its supervisor while this script started a NEW copy. Two daemons on one herdr session kill - each other's members. Investigate with 'pgrep -f \"$PATTERN\"' and stop the wrong one by - hand — do not assume either pid is the one you want." + each other's members. Investigate with 'ps -eo pid,comm,args | grep -F \"$PATTERN\"' and + check the COMM column of each hit yourself before acting — a bare 'pgrep -f \"$PATTERN\"' + (fleetd #593) can match the very shell you type it into, not just the daemon, so it is not + safe remediation advice on its own. Stop the wrong one by hand — do not assume either pid + is the one you want." fi } diff --git a/scripts/test-redeploy-fleetd.sh b/scripts/test-redeploy-fleetd.sh index 439e831..d4dd9d6 100755 --- a/scripts/test-redeploy-fleetd.sh +++ b/scripts/test-redeploy-fleetd.sh @@ -433,6 +433,67 @@ test_assert_single_daemon_rejects_two_pids() { printf '%s' "$output" | grep -qF '4343' || fail "refusal message does not list the pids it found" } +# fleetd #593 instance 2 — `running_pid()` used to be a bare `pgrep -f "$PATTERN"`, which matches +# ANY process whose full command line contains the pattern TEXT, including a shell that merely +# embeds it as literal text rather than being the daemon. Measured live on this Mac: `pgrep -c` +# (a one-call count) does not exist on BSD at all, and `bash -c ""` execs in place +# so no parent shell survives to hold the pattern — which is exactly why the defect did not +# reproduce from a plain script and needs a wrapper shaped like this instead. A `sh -c '...; ...'` +# with MORE THAN ONE statement does not get that exec-in-place treatment: the shell forks a child +# for the second statement and stays alive itself, holding the whole `-c` string — pattern text +# included — in its own `ps -o args`, for as long as it runs. That is the same shape an +# `ssh host "…; …"` wrapper or a hand-typed pipeline leaves behind. Before the fix this test would +# have found the wrapper's pid in running_pid()'s output; it must not. +test_running_pid_excludes_self_matching_wrapper_shell() { + local before after wrapper_pid + before="$(running_pid)" + sh -c 'echo "target/fleetd.jar" >/dev/null; sleep 20' & + wrapper_pid=$! + sleep 0.3 + after="$(running_pid)" + kill "$wrapper_pid" 2>/dev/null || true + wait "$wrapper_pid" 2>/dev/null || true + [ "$after" = "$before" ] \ + || fail "running_pid() counted a self-matching wrapper shell (pid $wrapper_pid, holding the pattern as literal text in its own argv, not the daemon): before=[$before] after=[$after]" +} + +# fleetd #593 instance 2, the other half — the fix above must not exclude too much. A genuine +# daemon-shaped process (comm is not a shell) whose own argv holds the pattern must still be +# found. `exec -a` gives a harmless `sleep` an argv[0] containing the pattern without starting a +# real JVM: copying a system binary into a scratch path and executing it from there was tried +# first and the OS killed it outright (SIGKILL, exit 137 — almost certainly a code-signing check), +# so this uses `exec -a` instead, which needs no binary of its own and nothing under a scratch +# directory. +test_running_pid_still_finds_a_real_second_process() { + local before after standin_pid + before="$(running_pid)" + ( exec -a "fleetd-593-test-standin-target/fleetd.jar" sleep 20 ) & + standin_pid=$! + sleep 0.3 + after="$(running_pid)" + kill "$standin_pid" 2>/dev/null || true + wait "$standin_pid" 2>/dev/null || true + printf '%s\n' "$after" | grep -qxF "$standin_pid" \ + || fail "running_pid() did not find a real second daemon-shaped process (pid $standin_pid) whose own argv holds the pattern: before=[$before] after=[$after]" +} + +# fleetd #593 instance 3 — assert_single_daemon's refusal message used to tell the operator to +# "Investigate with 'pgrep -f \"\$PATTERN\"'", which — typed by hand or over ssh — is precisely the +# self-matching invocation instance 2 above fixes. A source-text check, the same technique +# test_no_error_lines_message_gated_by_drain_state uses: this is prose inside a die() call, never +# reached by sourcing (the SOURCED guard stops before the main flow, and this text only prints +# from inside a call assert_single_daemon makes when it is already refusing). +test_die_message_does_not_recommend_bare_pgrep_as_remediation() { + local src="$ROOT/scripts/redeploy-fleetd.sh" block bad + block="$(grep -A6 -F 'racing supervisor produces' "$src" || true)" + [ -n "$block" ] || fail "could not find the assert_single_daemon refusal message in redeploy-fleetd.sh" + bad="$(printf '%s' "$block" | grep -F "Investigate with 'pgrep -f" || true)" + [ -z "$bad" ] \ + || fail "assert_single_daemon's die message still hands the operator a bare 'pgrep -f \"\$PATTERN\"' as remediation (fleetd #593) — that is exactly the self-matching invocation" + printf '%s' "$block" | grep -qF 'fleetd #593' \ + || fail "assert_single_daemon's die message does not say in words that a pattern can match the caller (fleetd #593)" +} + # fleetd #511 — jar_id()'s no-argument default was unpinned by any test: nothing proved it reports # $JAR (the live path) rather than $JAR_STAGED. Both halves matter, so this pins both: the bare call # must hash the live jar, and an explicit path argument must hash THAT file, not fall back to $JAR. @@ -1894,6 +1955,9 @@ test_require_drivable_supervisor_accepts_known_kinds test_count_daemon_pids test_assert_single_daemon_accepts_one_pid test_assert_single_daemon_rejects_two_pids +test_running_pid_excludes_self_matching_wrapper_shell +test_running_pid_still_finds_a_real_second_process +test_die_message_does_not_recommend_bare_pgrep_as_remediation test_jar_id_defaults_to_live_and_reports_explicit_path test_hash256_computes_a_real_sha256 test_jar_id_reports_absent_for_missing_file