Module inotify

Expand description

A single inotify instance which watches the memory events files of all containers.

Structs§

SharedWatcher 🔒: A single inotify instance shared by all OOM watchers.
Subscription 🔒: A subscription to the SharedWatcher which unsubscribes on drop.
Watch 🔒: A watch of a file and its subscribers.
WatchEvent 🔒: A watch event as delivered to subscribers.
Watches 🔒: The watches by descriptor. A watch belongs to a file rather than to its path: a file which got recreated gets a watch of its own, and the events of the old one, like its removal, must not reach the subscribers of the new one.

MIN_DISPATCH_INTERVAL 🔒: How often the events get dispatched at most. The kernel notifies a change of the memory events of a cgroup up to every 10 ms, and one of the hierarchical ones of every ancestor, which alternate in the queue, so that inotify does not merge them. Under memory pressure, like at memory.high or memory.max, a single container would wake the server up to a hundred times a second otherwise, and every one of its watchers would read its counters as often. The queue of the kernel holds 16384 events by default, which takes more than 1600 watched files changing in one interval, far more than the containers of a pod. An overflow gets dispatched to every watcher, see Watches::dispatch.
WATCH_FLAGS 🔒: What a watch reports. cgroupfs notifies a change of the memory events with IN_MODIFY and the removal of a cgroup with IN_IGNORED only, which every watch gets anyway.

SHARED_WATCHER 🔒: The shared watcher is kept alive by its subscriptions. It gets created on the first subscription and dropped with the last one, which frees the inotify instance while no container is watched. A failed creation is retried on the next subscription instead of being cached, and so is a watcher whose reader stopped. The server has a single runtime, which reads the events.