start_command/
cgroup_memory.rs

1//! cgroup v2 memory counters of a detached docker container (issue #182).
2//!
3//! Docker's `State.OOMKilled` is container-wide and sticky (moby/moby#43564):
4//! it says that *some* process in the container was OOM-killed at some point,
5//! but not how many, not whether the container hit its own `--memory` limit or
6//! the whole host ran out, and not how close the execution came to its limit.
7//! The kernel provides raw counters per cgroup
8//! (<https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files>):
9//!
10//! ```text
11//! memory.events `oom`       allocation events reaching the memory limit
12//! memory.events `oom_kill`  processes killed here by *any* OOM killer
13//! memory.peak / memory.max  peak usage (Linux 5.19+) and limit (`max` when
14//!                           unlimited)
15//! ```
16//! These counters have different units and are hierarchical. A group OOM can
17//! kill several processes for one allocation event; earlier allocation events
18//! can also coexist with a later host OOM. Their comparison cannot establish
19//! container, parent or host scope (issue #185).
20//!
21//! The cgroup is removed when the container stops, so the values cannot be read
22//! afterwards. The detached completion watcher therefore samples them while the
23//! container runs and hands the last sample to the post-mortem, the finalizer
24//! and the recovery step. Best effort by design: no cgroup v2, a remote docker
25//! daemon or a hidden `/proc` leave the counters unknown, and a kill in the last
26//! sampling interval can be missed when the cgroup is gone before the final
27//! read.
28
29use serde::{Deserialize, Deserializer, Serialize};
30use serde_json::Value;
31
32use crate::isolation::isolation_log::shell_quote;
33
34/// Default seconds between samples; the watcher environment may override it.
35pub const CGROUP_SAMPLE_INTERVAL_SECONDS: u64 = 5;
36
37/// Shell variables used by the sampler snippets.
38pub mod cgroup_shell_vars {
39    /// Last sample: `<memory.max> <memory.peak> <oom> <oom_kill>`.
40    pub const SAMPLE: &str = "__start_command_cgroup";
41    pub const SAMPLER: &str = "__start_command_cgroup_sampler";
42    pub const FILE: &str = "__start_command_cgroup_file";
43}
44
45/// Where the OOM killer that killed processes in the container came from.
46#[derive(Debug, Clone, Copy, PartialEq, Eq)]
47pub enum OomScope {
48    /// Raw counters do not establish container, parent or host scope.
49    Unknown,
50    /// Retained scope name; raw counters alone cannot establish it.
51    ContainerLimit,
52    /// Retained scope name; raw counters alone cannot establish it.
53    HostOrParent,
54}
55
56impl OomScope {
57    pub fn as_str(&self) -> &'static str {
58        match self {
59            OomScope::Unknown => "unknown",
60            OomScope::ContainerLimit => "container-limit",
61            OomScope::HostOrParent => "host-or-parent",
62        }
63    }
64
65    pub fn note(&self) -> &'static str {
66        match self {
67            OomScope::Unknown => "OOM kill scope unknown",
68            OomScope::ContainerLimit => "the container hit its own memory limit",
69            OomScope::HostOrParent => "a host-wide or parent cgroup OOM killed processes here",
70        }
71    }
72}
73
74/// The counters of one sample. `limit_bytes` is `None` without a limit.
75#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize)]
76#[serde(rename_all = "camelCase")]
77pub struct CgroupMemory {
78    pub limit_bytes: Option<u64>,
79    pub peak_bytes: Option<u64>,
80    pub oom_events: Option<u64>,
81    pub oom_kills: Option<u64>,
82}
83
84/// Shell functions shared by the sampler and its final read:
85///
86/// ```text
87/// __start_command_cgroup_dir NAME  print the container's cgroup v2 directory
88/// __start_command_cgroup_read DIR  print `<max> <peak> <oom> <oom_kill> DIR`
89/// ```
90///
91/// The directory comes from `/proc/<.State.Pid>/cgroup` (`0::<path>`), which
92/// covers the systemd and cgroupfs drivers, `--cgroup-parent` and rootless
93/// docker alike; the two default layouts are fallbacks. A candidate must name
94/// the container ID, so a PID reported by a remote daemon can never select an
95/// unrelated local cgroup. `START_COMMAND_CGROUP_ROOT` and
96/// `START_COMMAND_PROC_ROOT` replace `/sys/fs/cgroup` and `/proc` (tests).
97pub fn build_cgroup_functions_snippet() -> String {
98    r###"__start_command_cgroup_dir() { __scd_id=$(docker inspect -f '{{.Id}}' "$1" 2>/dev/null); [ -n "$__scd_id" ] || return 1; __scd_pid=$(docker inspect -f '{{.State.Pid}}' "$1" 2>/dev/null); __scd_rel=''; case "$__scd_pid" in ''|0|*[!0-9]*) ;; *) __scd_rel=$(sed -n 's/^0:://p' "${START_COMMAND_PROC_ROOT:-/proc}/$__scd_pid/cgroup" 2>/dev/null);; esac; __scd_root=${START_COMMAND_CGROUP_ROOT:-/sys/fs/cgroup}; for __scd_dir in ${__scd_rel:+"$__scd_root$__scd_rel"} "$__scd_root/system.slice/docker-$__scd_id.scope" "$__scd_root/docker/$__scd_id"; do case "$__scd_dir" in *"$__scd_id"*) if [ -r "$__scd_dir/memory.events" ]; then printf '%s' "$__scd_dir"; return 0; fi;; esac; done; return 1; }; __start_command_cgroup_read() { __scr_events=$(cat "$1/memory.events" 2>/dev/null) || return 1; __scr_oom=$(printf '%s\n' "$__scr_events" | sed -n 's/^oom //p'); __scr_kill=$(printf '%s\n' "$__scr_events" | sed -n 's/^oom_kill //p'); case "$__scr_oom:$__scr_kill" in *[!0-9:]*|:*|*:) return 1;; esac; __scr_max=$(cat "$1/memory.max" 2>/dev/null); __scr_peak=$(cat "$1/memory.peak" 2>/dev/null); printf '%s %s %s %s %s\n' "${__scr_max:--}" "${__scr_peak:--}" "${__scr_oom:--}" "${__scr_kill:--}" "$1"; }"###.to_string()
99}
100
101/// In-container loop of the long-lived sampler exec (issue #242): `$1` is the
102/// cgroup directory. The host paces it with one newline per interval on stdin
103/// (`docker exec -i`), so it runs only shell builtins: the task's own CPU
104/// limit pays for no `cat`, `sed` or `sleep` processes. Stdin reaches EOF when
105/// the exec's client is gone, so the loop never outlives the sampler; dockerd
106/// would accept its output even then. Exit 3: unreadable counters.
107pub fn build_cgroup_stream_script() -> String {
108    r###"while read -r __scx_tick; do __scx_oom=; __scx_kill=; __scx_max=-; __scx_peak=-; [ -r "$1/memory.events" ] || exit 3; while read -r __scx_key __scx_value; do case "$__scx_key" in oom) __scx_oom=$__scx_value;; oom_kill) __scx_kill=$__scx_value;; esac; done < "$1/memory.events"; case "$__scx_oom:$__scx_kill" in *[!0-9:]*|:*|*:) exit 3;; esac; [ ! -r "$1/memory.max" ] || read -r __scx_max < "$1/memory.max"; [ ! -r "$1/memory.peak" ] || read -r __scx_peak < "$1/memory.peak"; printf '%s %s %s %s %s\n' "${__scx_max:--}" "${__scx_peak:--}" "$__scx_oom" "$__scx_kill" "$1" || exit 0; done"###.to_string()
109}
110
111/// `$0` of the long-lived exec, visible in `ps` inside the task container.
112pub const CGROUP_STREAM_NAME: &str = "start-command-cgroup-stream";
113
114/// Interval value that turns the sampler off (issue #242).
115pub const CGROUP_SAMPLING_OFF_REASON: &str =
116    "sampling disabled by START_COMMAND_CGROUP_SAMPLE_INTERVAL_SECONDS=off";
117
118/// Shell fragment starting the background sampler. It writes each sample to a
119/// temp file with `mv`, so the watcher never reads a half-written line, and
120/// ends by itself once the container is gone.
121///
122/// Where the cgroup is readable locally it reads it directly. Otherwise, with
123/// a private cgroup namespace (a mounted docker.sock), one long-lived
124/// `docker exec -i` answers one host tick per interval (issue #242); a
125/// successful read proves the container runs, so Docker is asked for liveness
126/// only after a failed round or when the stream ends. Without a usable stream it falls back
127/// to one `docker exec` per round; without `sh` in the image (exit 126/127) or
128/// without any readable cgroup it records the reason and stops.
129/// `START_COMMAND_CGROUP_SAMPLE_INTERVAL_SECONDS=off` disables sampling.
130pub fn build_cgroup_sampler_start_snippet(container_name: &str) -> String {
131    let functions = build_cgroup_functions_snippet();
132    let remote = format!("{}; __start_command_cgroup_read /sys/fs/cgroup", functions);
133    format!(
134        "{}; {}",
135        functions,
136        r###"__start_command_cgroup_file="${TMPDIR:-/tmp}/start-command-cgroup.$$"; rm -f "$__start_command_cgroup_file" "$__start_command_cgroup_file.tmp" "$__start_command_cgroup_file.reason"; __start_command_memory_limit=$(docker inspect -f '{{.HostConfig.Memory}}' '__SC_NAME__' 2>/dev/null); __start_command_sample_name='__SC_NAME__'; ( export START_COMMAND_CGROUP_SAMPLING=1; __scs_interval=${START_COMMAND_CGROUP_SAMPLE_INTERVAL_SECONDS:-__SC_INTERVAL__}; case "$__scs_interval" in [Oo][Ff][Ff]) printf '%s\n' __SC_OFF__ > "$__start_command_cgroup_file.reason"; exit 0;; *[!0-9]*|'') __scs_interval=__SC_INTERVAL__;; esac; [ "$__scs_interval" -gt 0 ] 2>/dev/null || __scs_interval=__SC_INTERVAL__; __scs_record() { set -f; set -- $1; set +f; case "$3:$4" in *[!0-9:]*|:*|*:) return 1;; esac; __scs_now=$(date +%s); __scs_previous=''; __scs_changed=0; __scs_previous_at=0; [ ! -r "$__start_command_cgroup_file" ] || read -r __scs_p __scs_p __scs_p __scs_previous __scs_changed __scs_previous_at __scs_p < "$__start_command_cgroup_file"; if [ -n "$__scs_previous" ] && [ "$4" -gt "$__scs_previous" ] 2>/dev/null && [ "$((__scs_now-__scs_previous_at))" -le 3 ]; then __scs_changed=$__scs_now; fi; printf '%s %s %s %s %s %s %s\n' "$1" "$2" "$3" "$4" "$__scs_changed" "$__scs_now" "$__scs_dir" > "$__start_command_cgroup_file.tmp" && mv -f "$__start_command_cgroup_file.tmp" "$__start_command_cgroup_file"; }; __scs_dir=$(__start_command_cgroup_dir '__SC_NAME__'); __scs_ns=$(docker inspect -f '{{.HostConfig.CgroupnsMode}}' '__SC_NAME__' 2>/dev/null); __scs_stream=1; __scs_reader=''; while :; do __scs_ok=''; if [ -n "$__scs_dir" ]; then __scs_line=$(__start_command_cgroup_read "$__scs_dir") && __scs_record "$__scs_line" && __scs_ok=1; elif [ "$__scs_ns" != private ]; then printf '%s\n' "local cgroup unavailable; container cgroup namespace=${__scs_ns:-unavailable}, private required" > "$__start_command_cgroup_file.reason"; break; elif [ "$__scs_stream" = 1 ]; then trap 'kill "$__scs_reader" 2>/dev/null; wait "$__scs_reader" 2>/dev/null; exit 0' TERM; { { while :; do echo || exit 0; sleep "$__scs_interval" || exit 0; done 2>/dev/null & } | docker exec -i '__SC_NAME__' sh -c __SC_STREAM__ __SC_STREAM_NAME__ /sys/fs/cgroup 2>&1; printf '__scs_exit %s\n' "$?"; } | { __scs_got=''; __scs_exit=''; __scs_detail=''; while IFS= read -r __scs_line; do case "$__scs_line" in '__scs_exit '*) __scs_exit=${__scs_line#__scs_exit };; *) if __scs_record "$__scs_line"; then __scs_got=1; elif [ -z "$__scs_detail" ]; then __scs_detail=$(printf '%s' "$__scs_line" | cut -c 1-240); fi;; esac; done; [ -z "$__scs_got" ] || exit 0; printf '%s\n' "local cgroup unavailable; Docker exec exit=$__scs_exit: ${__scs_detail:-private cgroup v2 counters unreadable}" > "$__start_command_cgroup_file.reason"; case "$__scs_exit" in ''|0|*[!0-9]*) exit 1;; esac; exit "$__scs_exit"; } & __scs_reader=$!; wait "$__scs_reader"; __scs_exit=$?; trap - TERM; __scs_reader=''; case "$__scs_exit" in 0) ;; 126|127) break;; *) __scs_stream=0;; esac; else __scs_remote=$(docker exec '__SC_NAME__' sh -c __SC_REMOTE__ 2>&1); __scs_exit=$?; if [ "$__scs_exit" = 0 ] && __scs_record "$__scs_remote"; then __scs_ok=1; else __scs_detail=$(printf "%s" "$__scs_remote" | sed -n "1p" | cut -c 1-240); printf '%s\n' "local cgroup unavailable; Docker exec exit=$__scs_exit: ${__scs_detail:-private cgroup v2 counters unreadable}" > "$__start_command_cgroup_file.reason"; fi; fi; if [ -z "$__scs_ok" ] && [ "$(docker inspect -f '{{.State.Running}}' '__SC_NAME__' 2>/dev/null)" = false ]; then break; fi; sleep "$__scs_interval"; done ) >/dev/null 2>&1 & __start_command_cgroup_sampler=$!"###
137            .replace("'__SC_NAME__'", &shell_quote(container_name))
138            .replace("__SC_INTERVAL__", &CGROUP_SAMPLE_INTERVAL_SECONDS.to_string())
139            .replace("__SC_STREAM_NAME__", CGROUP_STREAM_NAME)
140            .replace("__SC_STREAM__", &shell_quote(&build_cgroup_stream_script()))
141            .replace("__SC_REMOTE__", &shell_quote(&remote))
142            .replace("__SC_OFF__", &shell_quote(CGROUP_SAMPLING_OFF_REASON))
143    )
144}
145
146/// Shell fragment stopping the sampler once the container has exited and
147/// leaving the last sample (`<max> <peak> <oom> <oom_kill>`, or empty) in
148/// `$__start_command_cgroup`. The cgroup can outlive the main process for a
149/// moment, so it is read one last time when it still exists.
150pub fn build_cgroup_sampler_stop_snippet() -> String {
151    r###"kill "$__start_command_cgroup_sampler" 2>/dev/null; wait "$__start_command_cgroup_sampler" 2>/dev/null; __start_command_cgroup=$(cat "$__start_command_cgroup_file" 2>/dev/null); __start_command_memory_unavailable=$(cat "$__start_command_cgroup_file.reason" 2>/dev/null); __scs_dir=$(printf '%s' "$__start_command_cgroup" | cut -s -d ' ' -f 7-); if [ -n "$__scs_dir" ] && __scs_line=$(__start_command_cgroup_read "$__scs_dir"); then __scs_old=$(printf '%s' "$__start_command_cgroup" | cut -d ' ' -f 4); __scs_at=$(printf '%s' "$__start_command_cgroup" | cut -d ' ' -f 6); __scs_changed=$(printf '%s' "$__start_command_cgroup" | cut -d ' ' -f 5); __scs_now=$(date +%s); __scs_new=$(printf '%s' "$__scs_line" | cut -d ' ' -f 4); if [ "$__scs_new" -gt "$__scs_old" ] 2>/dev/null && [ "$((__scs_now-__scs_at))" -le 3 ]; then __scs_changed=$__scs_now; fi; __start_command_cgroup="$(printf '%s' "$__scs_line" | cut -d ' ' -f 1-4) $__scs_changed $__scs_now"; fi; rm -f "$__start_command_cgroup_file" "$__start_command_cgroup_file.tmp" "$__start_command_cgroup_file.reason"; __start_command_cgroup=$(printf '%s' "$__start_command_cgroup" | cut -d ' ' -f 1-6)"###.to_string()
152}
153
154/// Shell fragment appending the `Memory:` line of the post-mortem when a
155/// sample exists. `quoted_log_path` must already be shell-quoted.
156pub fn build_cgroup_memory_log_snippet(quoted_log_path: &str) -> String {
157    r###"__start_command_cgroup_log() { __scm_note=''; if [ "$4" -gt 0 ] 2>/dev/null; then __scm_note=' (OOM kill scope unknown)'; fi; printf 'Memory:     memory.max=%s memory.peak=%s oom=%s oom_kill=%s%s\n' "$1" "$2" "$3" "$4" "$__scm_note"; }; if [ -n "$__start_command_cgroup" ]; then __start_command_cgroup_log $__start_command_cgroup >> __SC_LOG__; else printf 'Memory:     unavailable (%s) memory.limit=%s (HostConfig)\n' "${__start_command_memory_unavailable:-sampler stopped before a reading; local cgroup unavailable}" "${__start_command_memory_limit:-unknown}" >> __SC_LOG__; fi"###.replace("__SC_LOG__", quoted_log_path)
158}
159
160fn parse_counter(text: &str) -> Option<u64> {
161    if !text.is_empty() && text.chars().all(|c| c.is_ascii_digit()) {
162        text.parse().ok()
163    } else {
164        None
165    }
166}
167
168/// Parse a sample written by the watcher
169/// (`<memory.max> <memory.peak> <oom> <oom_kill>`); `None` without a sample.
170pub fn parse_cgroup_memory_sample(text: &str) -> Option<CgroupMemory> {
171    let fields: Vec<&str> = text.split_whitespace().collect();
172    if fields.len() < 4 {
173        return None;
174    }
175    let counters = CgroupMemory {
176        limit_bytes: parse_counter(fields[0]),
177        peak_bytes: parse_counter(fields[1]),
178        oom_events: parse_counter(fields[2]),
179        oom_kills: parse_counter(fields[3]),
180    };
181    if counters.oom_events.is_none() && counters.oom_kills.is_none() {
182        // memory.events was unreadable: the line carries no information.
183        return None;
184    }
185    Some(counters)
186}
187
188/// The `Memory:` line for a sample, exactly as the watcher's shell writes it.
189/// Used where Rust writes the post-mortem itself (the recovery step).
190pub fn format_cgroup_memory_log_line(text: &str) -> Option<String> {
191    let counters = parse_cgroup_memory_sample(text)?;
192    let fields: Vec<&str> = text.split_whitespace().collect();
193    let note = describe_cgroup_oom_scope(&counters)
194        .map(|scope| format!(" ({})", scope.note()))
195        .unwrap_or_default();
196    Some(format!(
197        "Memory:     memory.max={} memory.peak={} oom={} oom_kill={}{}",
198        fields[0], fields[1], fields[2], fields[3], note
199    ))
200}
201
202/// `Unknown` for observed kills, `None` when no kills were observed.
203///
204/// Separately attributed kernel/cgroup evidence would be needed to establish
205/// container, parent or host scope. An unknown `oom_events` remains unknown.
206pub fn describe_cgroup_oom_scope(memory: &CgroupMemory) -> Option<OomScope> {
207    memory.oom_kills.filter(|kills| *kills > 0)?;
208    Some(OomScope::Unknown)
209}
210
211fn format_bytes(bytes: Option<u64>) -> String {
212    let Some(bytes) = bytes else {
213        return "unknown".to_string();
214    };
215    const UNITS: [&str; 5] = ["B", "KiB", "MiB", "GiB", "TiB"];
216    let mut value = bytes as f64;
217    let mut unit = 0;
218    while value >= 1024.0 && unit < UNITS.len() - 1 {
219        value /= 1024.0;
220        unit += 1;
221    }
222    if unit == 0 {
223        format!("{} B", bytes)
224    } else {
225        format!("{:.1} {}", value, UNITS[unit])
226    }
227}
228
229/// Human-readable counters for `--status`, e.g.
230/// `peak 255.9 MiB of 256.0 MiB limit, oom 0, oom_kill 3 (...)`.
231pub fn format_cgroup_memory(memory: &CgroupMemory) -> String {
232    let limit = match memory.limit_bytes {
233        Some(_) => format!("{} limit", format_bytes(memory.limit_bytes)),
234        None => "no limit".to_string(),
235    };
236    let count = |value: Option<u64>| value.map_or_else(|| "unknown".to_string(), |v| v.to_string());
237    let note = describe_cgroup_oom_scope(memory)
238        .map(|scope| format!(" ({})", scope.note()))
239        .unwrap_or_default();
240    format!(
241        "peak {} of {}, oom {}, oom_kill {}{}",
242        format_bytes(memory.peak_bytes),
243        limit,
244        count(memory.oom_events),
245        count(memory.oom_kills),
246        note
247    )
248}
249
250fn counter_from_value(value: Option<&Value>) -> Option<u64> {
251    match value? {
252        Value::Number(number) => number.as_u64().or_else(|| {
253            number
254                .as_f64()
255                .filter(|v| v.fract() == 0.0 && *v >= 0.0 && *v <= 9_007_199_254_740_991.0)
256                .map(|v| v as u64)
257        }),
258        Value::String(text) => parse_counter(text.trim()),
259        _ => None,
260    }
261}
262
263/// Normalize counters read back from an execution record. Lenient, like the
264/// JavaScript reader: a malformed value never makes the record unreadable.
265pub fn normalize_cgroup_memory(value: &Value) -> Option<CgroupMemory> {
266    let object = value.as_object()?;
267    Some(CgroupMemory {
268        limit_bytes: counter_from_value(object.get("limitBytes")),
269        peak_bytes: counter_from_value(object.get("peakBytes")),
270        oom_events: counter_from_value(object.get("oomEvents")),
271        oom_kills: counter_from_value(object.get("oomKills")),
272    })
273}
274
275/// `deserialize_with` helper for `ExecutionRecord::cgroup_memory`.
276pub fn deserialize_cgroup_memory<'de, D>(deserializer: D) -> Result<Option<CgroupMemory>, D::Error>
277where
278    D: Deserializer<'de>,
279{
280    let value = Option::<Value>::deserialize(deserializer)?;
281    Ok(value.as_ref().and_then(normalize_cgroup_memory))
282}
283
284#[cfg(test)]
285mod tests {
286    use super::*;
287
288    #[test]
289    fn parses_counters_and_treats_max_as_no_limit() {
290        assert_eq!(
291            parse_cgroup_memory_sample("max 1024 0 2"),
292            Some(CgroupMemory {
293                limit_bytes: None,
294                peak_bytes: Some(1024),
295                oom_events: Some(0),
296                oom_kills: Some(2),
297            })
298        );
299        assert_eq!(parse_cgroup_memory_sample("max - - -"), None);
300        assert_eq!(parse_cgroup_memory_sample("1 2 3"), None);
301    }
302
303    #[test]
304    fn normalizes_floats_strings_and_nulls() {
305        let value = serde_json::json!({
306            "limitBytes": 268435456.0,
307            "peakBytes": "1024",
308            "oomEvents": null,
309            "oomKills": -1
310        });
311        assert_eq!(
312            normalize_cgroup_memory(&value),
313            Some(CgroupMemory {
314                limit_bytes: Some(268_435_456),
315                peak_bytes: Some(1024),
316                oom_events: None,
317                oom_kills: None,
318            })
319        );
320        assert_eq!(normalize_cgroup_memory(&Value::Null), None);
321    }
322}
start_command/cgroup_memory.rs

start_command/
cgroup_memory.rs
