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/// Seconds between two samples of a running container's cgroup.
35pub const CGROUP_SAMPLE_INTERVAL_SECONDS: u64 = 1;
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:--}" "$1"; }"###.to_string()
99}
100
101/// Shell fragment starting the background sampler. It writes each sample to a
102/// temp file with `mv`, so the watcher never reads a half-written line, and
103/// ends by itself once the cgroup is gone.
104pub fn build_cgroup_sampler_start_snippet(container_name: &str) -> String {
105    r###"__start_command_cgroup_dir() { __scd_id=$(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:--}" "$1"; }; __start_command_cgroup_file="${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=$(docker inspect -f '{{.HostConfig.Memory}}' '__SC_NAME__' 2>/dev/null); __start_command_sample_name='__SC_NAME__'; ( export START_COMMAND_CGROUP_SAMPLING=1; __scs_previous=''; __scs_previous_at=0; __scs_changed=0; __scs_dir=$(__start_command_cgroup_dir '__SC_NAME__'); while :; do __scs_line=""; if [ -n "$__scs_dir" ]; then __scs_line=$(__start_command_cgroup_read "$__scs_dir"); fi; if [ -z "$__scs_line" ]; then __scs_ns=$(docker inspect -f '{{.HostConfig.CgroupnsMode}}' '__SC_NAME__' 2>/dev/null); if [ "$__scs_ns" = private ]; then __scs_remote=$(docker exec '__SC_NAME__' sh -c '__start_command_cgroup_dir() { __scd_id=$(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:--}" "$1"; }; __start_command_cgroup_read /sys/fs/cgroup' 2>&1); __scs_exit=$?; if [ "$__scs_exit" = 0 ]; then __scs_line=$__scs_remote; else __scs_detail=$(printf "%s" "$__scs_remote" | sed -n "1p" | cut -c 1-240); printf '%s
106' "local cgroup unavailable; Docker exec exit=$__scs_exit: ${__scs_detail:-private cgroup v2 counters unreadable}" > "$__start_command_cgroup_file.reason"; fi; else printf '%s
107' "local cgroup unavailable; container cgroup namespace=${__scs_ns:-unavailable}, private required" > "$__start_command_cgroup_file.reason"; fi; fi; __scs_now=$(date +%s); if [ -n "$__scs_line" ]; then __scs_kill=$(printf "%s" "$__scs_line" | cut -d " " -f 4); if [ -n "$__scs_previous" ] && [ "$__scs_kill" -gt "$__scs_previous" ] 2>/dev/null && [ "$((__scs_now-__scs_previous_at))" -le 3 ]; then __scs_changed=$__scs_now; fi; __scs_previous=$__scs_kill; __scs_previous_at=$__scs_now; __scs_fields=$(printf "%s" "$__scs_line" | cut -d " " -f 1-4); printf '%s %s %s %s\n' "$__scs_fields" "$__scs_changed" "$__scs_now" "$__scs_dir" > "$__start_command_cgroup_file.tmp" && mv -f "$__start_command_cgroup_file.tmp" "$__start_command_cgroup_file"; fi; if [ "$(docker inspect -f '{{.State.Running}}' '__SC_NAME__' 2>/dev/null)" = false ]; then break; fi; sleep 1; done ) >/dev/null 2>&1 & __start_command_cgroup_sampler=$!"###.replace("'__SC_NAME__'", &shell_quote(container_name))
108}
109
110/// Shell fragment stopping the sampler once the container has exited and
111/// leaving the last sample (`<max> <peak> <oom> <oom_kill>`, or empty) in
112/// `$__start_command_cgroup`. The cgroup can outlive the main process for a
113/// moment, so it is read one last time when it still exists.
114pub fn build_cgroup_sampler_stop_snippet() -> String {
115    r###"kill "$__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()
116}
117
118/// Shell fragment appending the `Memory:` line of the post-mortem when a
119/// sample exists. `quoted_log_path` must already be shell-quoted.
120pub fn build_cgroup_memory_log_snippet(quoted_log_path: &str) -> String {
121    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)
122}
123
124fn parse_counter(text: &str) -> Option<u64> {
125    if !text.is_empty() && text.chars().all(|c| c.is_ascii_digit()) {
126        text.parse().ok()
127    } else {
128        None
129    }
130}
131
132/// Parse a sample written by the watcher
133/// (`<memory.max> <memory.peak> <oom> <oom_kill>`); `None` without a sample.
134pub fn parse_cgroup_memory_sample(text: &str) -> Option<CgroupMemory> {
135    let fields: Vec<&str> = text.split_whitespace().collect();
136    if fields.len() < 4 {
137        return None;
138    }
139    let counters = CgroupMemory {
140        limit_bytes: parse_counter(fields[0]),
141        peak_bytes: parse_counter(fields[1]),
142        oom_events: parse_counter(fields[2]),
143        oom_kills: parse_counter(fields[3]),
144    };
145    if counters.oom_events.is_none() && counters.oom_kills.is_none() {
146        // memory.events was unreadable: the line carries no information.
147        return None;
148    }
149    Some(counters)
150}
151
152/// The `Memory:` line for a sample, exactly as the watcher's shell writes it.
153/// Used where Rust writes the post-mortem itself (the recovery step).
154pub fn format_cgroup_memory_log_line(text: &str) -> Option<String> {
155    let counters = parse_cgroup_memory_sample(text)?;
156    let fields: Vec<&str> = text.split_whitespace().collect();
157    let note = describe_cgroup_oom_scope(&counters)
158        .map(|scope| format!(" ({})", scope.note()))
159        .unwrap_or_default();
160    Some(format!(
161        "Memory:     memory.max={} memory.peak={} oom={} oom_kill={}{}",
162        fields[0], fields[1], fields[2], fields[3], note
163    ))
164}
165
166/// `Unknown` for observed kills, `None` when no kills were observed.
167///
168/// Separately attributed kernel/cgroup evidence would be needed to establish
169/// container, parent or host scope. An unknown `oom_events` remains unknown.
170pub fn describe_cgroup_oom_scope(memory: &CgroupMemory) -> Option<OomScope> {
171    memory.oom_kills.filter(|kills| *kills > 0)?;
172    Some(OomScope::Unknown)
173}
174
175fn format_bytes(bytes: Option<u64>) -> String {
176    let Some(bytes) = bytes else {
177        return "unknown".to_string();
178    };
179    const UNITS: [&str; 5] = ["B", "KiB", "MiB", "GiB", "TiB"];
180    let mut value = bytes as f64;
181    let mut unit = 0;
182    while value >= 1024.0 && unit < UNITS.len() - 1 {
183        value /= 1024.0;
184        unit += 1;
185    }
186    if unit == 0 {
187        format!("{} B", bytes)
188    } else {
189        format!("{:.1} {}", value, UNITS[unit])
190    }
191}
192
193/// Human-readable counters for `--status`, e.g.
194/// `peak 255.9 MiB of 256.0 MiB limit, oom 0, oom_kill 3 (...)`.
195pub fn format_cgroup_memory(memory: &CgroupMemory) -> String {
196    let limit = match memory.limit_bytes {
197        Some(_) => format!("{} limit", format_bytes(memory.limit_bytes)),
198        None => "no limit".to_string(),
199    };
200    let count = |value: Option<u64>| value.map_or_else(|| "unknown".to_string(), |v| v.to_string());
201    let note = describe_cgroup_oom_scope(memory)
202        .map(|scope| format!(" ({})", scope.note()))
203        .unwrap_or_default();
204    format!(
205        "peak {} of {}, oom {}, oom_kill {}{}",
206        format_bytes(memory.peak_bytes),
207        limit,
208        count(memory.oom_events),
209        count(memory.oom_kills),
210        note
211    )
212}
213
214fn counter_from_value(value: Option<&Value>) -> Option<u64> {
215    match value? {
216        Value::Number(number) => number.as_u64().or_else(|| {
217            number
218                .as_f64()
219                .filter(|v| v.fract() == 0.0 && *v >= 0.0 && *v <= 9_007_199_254_740_991.0)
220                .map(|v| v as u64)
221        }),
222        Value::String(text) => parse_counter(text.trim()),
223        _ => None,
224    }
225}
226
227/// Normalize counters read back from an execution record. Lenient, like the
228/// JavaScript reader: a malformed value never makes the record unreadable.
229pub fn normalize_cgroup_memory(value: &Value) -> Option<CgroupMemory> {
230    let object = value.as_object()?;
231    Some(CgroupMemory {
232        limit_bytes: counter_from_value(object.get("limitBytes")),
233        peak_bytes: counter_from_value(object.get("peakBytes")),
234        oom_events: counter_from_value(object.get("oomEvents")),
235        oom_kills: counter_from_value(object.get("oomKills")),
236    })
237}
238
239/// `deserialize_with` helper for `ExecutionRecord::cgroup_memory`.
240pub fn deserialize_cgroup_memory<'de, D>(deserializer: D) -> Result<Option<CgroupMemory>, D::Error>
241where
242    D: Deserializer<'de>,
243{
244    let value = Option::<Value>::deserialize(deserializer)?;
245    Ok(value.as_ref().and_then(normalize_cgroup_memory))
246}
247
248#[cfg(test)]
249mod tests {
250    use super::*;
251
252    #[test]
253    fn parses_counters_and_treats_max_as_no_limit() {
254        assert_eq!(
255            parse_cgroup_memory_sample("max 1024 0 2"),
256            Some(CgroupMemory {
257                limit_bytes: None,
258                peak_bytes: Some(1024),
259                oom_events: Some(0),
260                oom_kills: Some(2),
261            })
262        );
263        assert_eq!(parse_cgroup_memory_sample("max - - -"), None);
264        assert_eq!(parse_cgroup_memory_sample("1 2 3"), None);
265    }
266
267    #[test]
268    fn normalizes_floats_strings_and_nulls() {
269        let value = serde_json::json!({
270            "limitBytes": 268435456.0,
271            "peakBytes": "1024",
272            "oomEvents": null,
273            "oomKills": -1
274        });
275        assert_eq!(
276            normalize_cgroup_memory(&value),
277            Some(CgroupMemory {
278                limit_bytes: Some(268_435_456),
279                peak_bytes: Some(1024),
280                oom_events: None,
281                oom_kills: None,
282            })
283        );
284        assert_eq!(normalize_cgroup_memory(&Value::Null), None);
285    }
286}
start_command/cgroup_memory.rs

start_command/
cgroup_memory.rs
