blob: e5132fc4d7fe2811a0c33623a9af78cf829c172a [file] [log] [blame]
// Copyright 2020 The Fuchsia Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
library fuchsia.memorypressure;
/// Indicates the memory pressure level.
type Level = strict enum {
/// The memory pressure level is healthy.
/// Registered clients are free to hold on to caches and allocate memory
/// unrestricted.
/// However, clients should take care to not proactively re-create caches on a
/// transition back to the NORMAL level, causing a memory spike that immediately
/// pushes the level over to WARNING again.
/// The memory pressure level is somewhat constrained, and might cross over to
/// the critical pressure range if left unchecked.
/// Registered clients are expected to optimize their operation to limit memory
/// usage, rather than for best performance, for example, by reducing cache sizes
/// and non-essential memory allocations.
/// Clients must take care to regulate the amount of work they undertake in
/// order to reclaim memory, and ensure that it does not cause visible
/// performance degradation. There exists some memory pressure, but not enough
/// to justify trading off user responsiveness to reclaim memory.
/// The memory pressure level is very constrained.
/// Registered clients are expected to drop all non-essential memory, and refrain
/// from allocating more memory. Failing to do so might result in the job
/// getting terminated, or the system being rebooted in the case of global
/// memory pressure.
/// Clients may undertake expensive work to reclaim memory if required, since
/// failing to do so might result in termination. The client might decide that a
/// performance hit is a fair tradeoff in this case.
/// Registration protocol
protocol Provider {
/// Used to register for memory pressure level changes.
/// `watcher`: memory pressure `Watcher` channel that the `Provider` will use to send
/// level change messages to the client.
/// The current memory pressure level is immediately sent to the watcher
/// when this method is called.
/// It is recommended that the root job in a component tree register for changes,
/// rather than having individual jobs further down the tree register individually.
/// A low client count will help minimize system churn due to a large number of
/// memory pressure messages in transit at the same time.
/// Also, the more context a job has, the better equipped it will be to react to
/// memory pressure by controlling the behavior of children jobs in its tree.
RegisterWatcher(resource struct {
watcher client_end:Watcher;
/// Watcher protocol
/// To be implemented by clients who wish to be notified on memory pressure level changes.
protocol Watcher {
/// Sent to the registered client when the memory pressure level changes.
/// `level`: indicates the current memory pressure level.
/// Will also be invoked on initial connection via `RegisterWatcher`, so that a newly
/// registered client can discover the current memory pressure level.
/// The watcher must immediately reply with a message to acknowledge that it has
/// received the level change notification, and has initiated required actions as a
/// result. It may then continue to reclaim memory asynchronously after sending
/// the acknowledgement.
/// Some helpful guidelines for clients:
/// 1. The watcher will be notified of new pressure level changes only after a reply
/// corresponding to the previous message has been received by the provider.
/// If multiple level transitions occur during that time, the watcher will be
/// notified of the latest pressure level.
/// 2. The level changes are edge-triggered, and clients are expected to maintain
/// local state to track the current pressure level, if required. For example,
/// a job might be notified of a CRITICAL level and drop all its caches as a result.
/// Some time after this, it might want to trigger an activity that causes a
/// fair amount of memory to be allocated. At this point, the job is expected to
/// remember that the last pressure level it saw was CRITICAL, and refrain from
/// triggering the memory-intensive activity.
/// 3. As a performance optimization, the provider may decide to skip sending
/// messages for some pressure level changes. For example, when oscillating across
/// the NORMAL / WARNING boundary, it might not be worth notifying clients of every
/// single transition. The provider might rate-limit messages in this case.
/// On a similar note, the provider may decide to send repeated messages at the
/// same pressure level, particularly CRITICAL, to indicate that further action
/// needs to be taken.
OnLevelChanged(struct {
level Level;
}) -> ();