Struct OOMWatcher 

pub(crate) struct OOMWatcher;

Watches the cgroup of a process for OOM kills, see OOMWatcher::watch.

Implementations

impl OOMWatcher

pub(super) fn process_cgroup_subsystem_path_cgroup_v2( pid: Pid, ) -> Result<Option<PathBuf>>

The cgroup v2 path of a process, none if the process is gone.

pub(super) fn unified_cgroup_path(cgroup_file: impl BufRead) -> Result<PathBuf>

The cgroupfs path of the unified hierarchy from a /proc/<pid>/cgroup file. Named cgroup v1 hierarchies, like the one of systemd, are listed as well and may come first.

pub(super) fn cgroup_v2_path(line: &str) -> Result<PathBuf>

The cgroupfs path of the unified hierarchy line 0::cgroup-path of /proc/<pid>/cgroup. The path may contain colons itself.

fn strip_cgroup_root(path: Option<&str>) -> Result<&str>

The cgroup path relative to the root of its hierarchy.

fn try_open_cgroup_path(pid: Pid) -> Result<Option<StdFile>>

impl OOMWatcher

const MAX_WATCHER_RESTARTS: u32 = 3

How often an OOM watcher watches its files again on a new shared watcher, after the reader of the events failed.

pub(super) async fn oom_handling_cgroup_v2( token: &CancellationToken, pid: Pid, ) -> Result<Option<OOMEvent>>

Watch the cgroup of a process with cgroup v2, see OOMWatcher::watch.

async fn watch_cgroup_v2(token: &CancellationToken, leaf: &Path) -> OOMEvent

Wait for an OOM of the cgroup v2 leaf until token gets cancelled and return what was found.

An OOM counts once the counters of the leaf went up. The leaf may be gone at the exit, the systemd cgroup manager removes the scope of a container once its processes exited. Then the hierarchical oom_kill counter of the nearest ancestor which outlives it tells whether a kill happened meanwhile, see OOMEvent::possible_oom. A removal is no OOM on its own. cgroupfs reports it with IN_IGNORED only, which ends the watch without an event.

fn watch_failed(message: &str, e: &Error)

Log a failure to watch memory events. A file which is gone, like the one of a cgroup which got removed already, is expected. Others, like running out of inotify instances or watches, get a warning, only the first one until a watch works again, see WATCH_FAILING.

fn is_not_found(e: &Error) -> bool

Whether watching or reading failed because the file does not exist. cgroupfs fails with ENODEV instead for a file whose cgroup got removed while it got opened or read.

fn initial_oom_counters(path: &Path) -> OomCounters

Read the current OOM counters of a memory events file to establish a baseline, so that only OOMs during this container’s lifetime count.

fn check_for_oom( memory_events_file_path: &Path, last: OomCounters, ) -> Result<(OomCounters, bool)>

fn read_oom_counters(path: &Path) -> Result<OomCounters>

Read the oom and oom_kill counters of a memory events file.

pub(super) fn parse_oom_counters(content: &str) -> Result<OomCounters>

Parse the oom and oom_kill counters of the content of a memory events file.

fn file_not_found(f: impl AsRef<Path>) -> bool

Checks if a file does not exist on disk.

impl OOMWatcher

pub(crate) async fn watch( token: &CancellationToken, pid: Pid, exit_paths: &[PathBuf], ) -> Option<OOMEvent>

Watch the cgroup of the process for an OOM until token gets cancelled, which the exit of the process does, and return what was found. The OOM files get written for an OOM found on its own. None if the cgroup could not be watched, like without cgroup v2, or the process was gone already.

fn supported(cgroup_v2: bool, warned: &AtomicBool) -> bool

Whether the OOM kills of the host can be watched, which takes the unified hierarchy of cgroup v2 at the cgroup root. Without it, like with cgroup v1 or a hybrid setup, the containers run without the OOM detection, which a warning tells once per server, as noted in warned, instead of with every container.

fn counters( content: &str, ) -> impl Iterator<Item = (&str, Result<u64, ParseIntError>)>

The counters of a cgroup file of key value lines, like memory.events, by their key.

pub(crate) async fn write_oom_files(exit_paths: &[PathBuf])

Create the empty OOM files, like conmon does. A failure is only logged, the exit gets reported either way, only without the OOM. A symlink at the path does not get followed, like for the log files.