Files
crosspoint-reader-mod/docs/activity-manager.md

473 lines
21 KiB
Markdown
Raw Permalink Normal View History

# Activity & ActivityManager Migration Guide
This document explains the refactoring from the original per-activity render task model to the centralized `ActivityManager` introduced in [PR #1016](https://github.com/crosspoint-reader/crosspoint-reader/pull/1016). It covers the architectural differences, what changed for activity authors, and the FreeRTOS task and locking model that underpins the system.
## Overview of Changes
| Aspect | Old Model | New Model |
|--------|-----------|-----------|
| Render task | One per activity (8KB stack each) | Single shared task in `ActivityManager` |
| Render mutex | Per-activity `renderingMutex` | Single global mutex in `ActivityManager` |
| `RenderLock` | Inner class of `Activity` | Standalone class, acquires global mutex |
| Subactivities | `ActivityWithSubactivity` base class | Activity stack managed by `ActivityManager` |
| Navigation | Free functions in `main.cpp` | `activityManager.goHome()`, `goToReader()`, etc. |
| Subactivity results | Callback lambdas stored in parent | `startActivityForResult()` / `setResult()` / `finish()` |
| `requestUpdate()` | Notifies activity's own render task | Delegates to `ActivityManager` (immediate or deferred) |
## Architecture
### Old Model: Per-Activity Render Tasks
Each activity created its own FreeRTOS render task on entry and destroyed it on exit:
```text
┌─────────────────────────────────────────────────────────┐
│ Main Task (Arduino loop) │
│ ┌───────────────────────────────────────────────────┐ │
│ │ currentActivity->loop() │ │
│ │ ├── handle input │ │
│ │ ├── update state (under RenderLock) │ │
│ │ └── requestUpdate() ──notify──► Render Task │ │
│ │ (per-activity)│ │
│ │ 8KB stack │ │
│ │ owns mutex │ │
│ └───────────────────────────────────────────────────┘ │
│ │
│ ActivityWithSubactivity: │
│ ┌──────────────┐ ┌──────────────┐ │
│ │ Parent │────►│ SubActivity │ │
│ │ (has render │ │ (has own │ │
│ │ task) │ │ render task) │ │
│ └──────────────┘ └──────────────┘ │
└─────────────────────────────────────────────────────────┘
```
Problems with this approach:
- **8KB per render task**: Each activity allocated an 8KB FreeRTOS stack for its render task, even though only one renders at a time
- **Dangerous deletion patterns**: `exitActivity()` + `enterNewActivity()` in callbacks led to `delete this` situations where the caller was destroyed while its code was still on the stack
- **Subactivity coupling**: Parents stored callbacks to child results, creating tight coupling and lifetime hazards
### New Model: Centralized ActivityManager
A single `ActivityManager` owns the render task and manages an activity stack:
```text
┌──────────────────────────────────────────────────────────┐
│ Main Task (Arduino loop) │
│ │
│ activityManager.loop() │
│ │ │
│ ├── currentActivity->loop() │
│ │ ├── handle input │
│ │ ├── update state (under RenderLock) │
│ │ └── requestUpdate() │
│ │ │
│ ├── process pending actions (Push / Pop / Replace) │
│ │ │
│ └── if requestedUpdate: ──notify──► Render Task │
│ (single, shared) │
│ 8KB stack │
│ global mutex │
│ │
│ Activity Stack: │
│ ┌──────────┬──────────┬──────────┐ ┌──────────┐ │
│ │ Home │ Settings │ Wifi │ │ Keyboard │ │
│ │ (stack) │ (stack) │ (stack) │ │ (current)│ │
│ └──────────┴──────────┴──────────┘ └──────────┘ │
│ stackActivities[] currentActivity │
└──────────────────────────────────────────────────────────┘
```
## Migration Checklist
### 1. Change Base Class
If your activity extended `ActivityWithSubactivity`, change it to extend `Activity`:
```cpp
// BEFORE
class MyActivity final : public ActivityWithSubactivity {
MyActivity(GfxRenderer& r, MappedInputManager& m, std::function<void()> goBack)
: ActivityWithSubactivity("MyActivity", r, m), goBack(goBack) {}
};
// AFTER
class MyActivity final : public Activity {
MyActivity(GfxRenderer& r, MappedInputManager& m)
: Activity("MyActivity", r, m) {}
};
```
Note that navigation callbacks like `goBack` are no longer stored — use `finish()` or `activityManager.goHome()` instead.
### 2. Replace Navigation Functions
The free functions `exitActivity()` / `enterNewActivity()` in `main.cpp` are gone. Use `ActivityManager` methods:
```cpp
// BEFORE (in main.cpp or via stored callbacks)
exitActivity();
enterNewActivity(new SettingsActivity(renderer, mappedInput, onGoHome));
// AFTER (from any Activity method)
activityManager.goToSettings();
// or for arbitrary navigation:
activityManager.replaceActivity(std::make_unique<MyActivity>(renderer, mappedInput));
```
`replaceActivity()` destroys the current activity and clears the stack. Use it for top-level navigation (home, reader, settings, etc.).
### 3. Replace Subactivity Pattern
The `enterNewActivity()` / `exitActivity()` subactivity pattern is replaced by a stack with typed results:
```cpp
// BEFORE
void MyActivity::launchWifi() {
enterNewActivity(new WifiSelectionActivity(renderer, mappedInput,
[this](bool connected) { onWifiDone(connected); }));
}
// Child calls: onComplete(true); // triggers callback, which may call exitActivity()
// AFTER
void MyActivity::launchWifi() {
startActivityForResult(
std::make_unique<WifiSelectionActivity>(renderer, mappedInput),
[this](const ActivityResult& result) {
if (result.isCancelled) return;
auto& wifi = std::get<WifiResult>(result.data);
onWifiDone(wifi.connected);
});
}
// Child calls:
// setResult(WifiResult{.connected = true, .ssid = ssid});
// finish();
```
Key differences:
- **`startActivityForResult()`** pushes the current activity onto the stack and launches the child
- **`setResult()`** stores a typed result on the child activity
- **`finish()`** signals the manager to pop the child, call the result handler, and resume the parent
- The parent is never deleted during this process — it's safely stored on the stack
### 4. Update `render()` Signature
The `RenderLock` type changed from `Activity::RenderLock` (inner class) to standalone `RenderLock`:
```cpp
// BEFORE
void render(Activity::RenderLock&&) override;
// AFTER
void render(RenderLock&&) override;
```
Include `RenderLock.h` if not transitively included via `Activity.h`.
### 5. Update `onEnter()` / `onExit()`
Activities no longer create or destroy render tasks:
```cpp
// BEFORE
void MyActivity::onEnter() {
Activity::onEnter(); // created render task + logged
// ... allocate resources
requestUpdate();
}
void MyActivity::onExit() {
// ... free resources
Activity::onExit(); // acquired RenderLock, deleted render task
}
// AFTER
void MyActivity::onEnter() {
Activity::onEnter(); // just logs
// ... allocate resources
requestUpdate();
}
void MyActivity::onExit() {
// ... free resources
Activity::onExit(); // just logs
}
```
The render task lifecycle is handled entirely by `ActivityManager::begin()`.
### 6. Update `requestUpdate()` Calls
The signature changed to accept an `immediate` flag:
```cpp
// BEFORE
void requestUpdate(); // always immediate notification to per-activity render task
// AFTER
void requestUpdate(bool immediate = false);
// immediate=false (default): deferred until end of current loop iteration
// immediate=true: sends notification to render task right away
```
**When to use `immediate`**: Almost never. Deferred updates are batched — if `loop()` triggers multiple state changes that each call `requestUpdate()`, only one render happens. Use `immediate` only when you need the render to start before the current function returns (e.g., before a blocking network call).
**`requestUpdateAndWait()`**: Blocks the calling task until the render completes. Use sparingly — it's designed for cases where you need the screen to reflect new state before proceeding (e.g., showing "Checking for update..." before calling a network API).
### 7. Remove Stored Navigation Callbacks
Old activities often stored `std::function` callbacks for navigation:
```cpp
// BEFORE
class SettingsActivity : public ActivityWithSubactivity {
const std::function<void()> goBack; // stored callback
const std::function<void()> goHome; // stored callback
public:
SettingsActivity(GfxRenderer& r, MappedInputManager& m,
std::function<void()> goBack, std::function<void()> goHome)
: ActivityWithSubactivity("Settings", r, m), goBack(goBack), goHome(goHome) {}
};
// AFTER
class SettingsActivity : public Activity {
public:
SettingsActivity(GfxRenderer& r, MappedInputManager& m)
: Activity("Settings", r, m) {}
// Use finish() to go back, activityManager.goHome() to go home
};
```
This removes `std::function` overhead (~2-4KB per unique signature) and eliminates lifetime risks from captured `this` pointers.
## Technical Details
### FreeRTOS Task Model
The firmware runs on an ESP32-C3, a single-core RISC-V microcontroller. FreeRTOS provides cooperative and preemptive multitasking on this single core — only one task executes at any moment, and the scheduler switches between tasks at yield points (blocking calls, `vTaskDelay`, `taskYIELD`) or when a tick interrupt promotes a higher-priority task.
There are two tasks relevant to the activity system:
```text
┌──────────────────────┐ ┌──────────────────────────┐
│ Main Task │ │ Render Task │
│ (Arduino loop) │ │ (ActivityManager-owned) │
│ Priority: 1 │ │ Priority: 1 │
│ │ │ │
│ Runs: │ │ Runs: │
│ - gpio.update() │ │ - ulTaskNotifyTake() │
│ - activity->loop() │ │ (blocks until notified)│
│ - pending actions │ │ - RenderLock (mutex) │
│ - sleep/power mgmt │ │ - activity->render() │
│ - requestUpdate →────┼─────┼─► xTaskNotify() │
│ (end of loop) │ │ │
└──────────────────────┘ └──────────────────────────┘
```
Both tasks run at priority 1. Since the ESP32-C3 is single-core, they alternate execution: the main task runs `loop()`, then at the end of the loop iteration, notifies the render task if an update was requested. The render task wakes, acquires the mutex, calls `render()`, releases the mutex, and blocks again.
Do not use `xTaskCreate` inside activities. If you have a use case that seems to require a background task, open a discussion to propose a lifecycle-aware `Worker` abstraction first.
### The Render Mutex and RenderLock
A single FreeRTOS mutex (`renderingMutex`) protects shared state between `loop()` and `render()`. Since these run on different tasks, any state read by `render()` and written by `loop()` must be guarded.
`RenderLock` is an RAII wrapper:
```cpp
// Standalone class (not tied to any specific activity)
class RenderLock {
bool isLocked = false;
public:
explicit RenderLock(); // acquires activityManager.renderingMutex
explicit RenderLock(Activity&); // same — Activity& param kept for compatibility
~RenderLock(); // releases mutex if still held
void unlock(); // early release
};
```
**Usage patterns:**
```cpp
// In loop(): protect state mutations that render() reads
void MyActivity::loop() {
if (somethingChanged) {
RenderLock lock;
state = newState; // safe — render() can't run while lock is held
}
requestUpdate(); // trigger render after lock is released
}
// In render(): lock is passed in, held for duration of render
void MyActivity::render(RenderLock&&) {
// Lock is held — safe to read shared state
renderer.clearScreen();
renderer.drawText(..., stateString, ...);
renderer.displayBuffer();
// Lock released when RenderLock destructor runs
}
```
**Critical rule**: Never call `requestUpdateAndWait()` while holding a `RenderLock`. The render task needs the mutex to call `render()`, so holding it while waiting for the render to complete is a deadlock:
```text
Main Task Render Task
────────── ───────────
RenderLock lock; (blocked on mutex)
requestUpdateAndWait();
→ notify render task
→ block waiting for
render to complete → wakes up
→ tries to acquire mutex
→ DEADLOCK: main holds mutex,
waits for render; render
waits for mutex
```
### requestUpdate() vs requestUpdateAndWait()
```text
requestUpdate(false) requestUpdate(true)
───────────────── ─────────────────
Sets flag only. Notifies render task
Render happens after immediately.
loop() returns and Render may start
ActivityManager checks before the calling
the flag. function returns.
(Does NOT wait for
render to complete.)
requestUpdateAndWait()
──────────────────────
Notifies render task AND
blocks calling task until
render is done. Uses
FreeRTOS direct-to-task
notification on the
caller's task handle.
```
`requestUpdateAndWait()` flow in detail:
```text
Calling Task Render Task
──────────── ───────────
requestUpdateAndWait()
├─ assert: not render task
├─ assert: not holding RenderLock
├─ store waitingTaskHandle
├─ xTaskNotify(renderTask) → wakes render task
└─ ulTaskNotifyTake() ─┐
(blocked) │ RenderLock lock;
│ activity->render();
│ // render complete
│ taskENTER_CRITICAL
│ waiter = waitingTaskHandle
│ waitingTaskHandle = nullptr
│ taskEXIT_CRITICAL
│ xTaskNotify(waiter) ───┐
│ │
┌──────────────────────┘ │
│ (woken by notification) ◄────────────────────────┘
└─ return
```
### Activity Lifecycle Under ActivityManager
```text
activityManager.replaceActivity(make_unique<MyActivity>(...))
╔═══════════════════════════════════════════════════╗
║ pendingAction = Replace ║
║ pendingActivity = MyActivity ║
╚═══════════════════════════════════════════════════╝
▼ (next loop iteration)
ActivityManager::loop()
├── currentActivity->loop() // old activity's last loop
├── process pending action:
│ ├── RenderLock lock;
│ ├── oldActivity->onExit() // cleanup under lock
│ ├── delete oldActivity
│ ├── clear stack
│ ├── currentActivity = MyActivity
│ ├── lock.unlock()
│ └── MyActivity->onEnter() // init new activity
└── if requestedUpdate:
└── notify render task
```
For push/pop (subactivity) navigation:
```text
Parent calls: startActivityForResult(make_unique<Child>(...), handler)
╔══════════════════════════════════════╗
║ pendingAction = Push ║
║ pendingActivity = Child ║
║ parent->resultHandler = handler ║
╚══════════════════════════════════════╝
▼ (next loop iteration)
├── Parent moved to stackActivities[]
├── currentActivity = Child
└── Child->onEnter()
... child runs ...
Child calls: setResult(MyResult{...}); finish();
╔══════════════════════════════════════╗
║ pendingAction = Pop ║
║ child->result = MyResult{...} ║
╚══════════════════════════════════════╝
▼ (next loop iteration)
├── result = child->result
├── Child->onExit(); delete Child
├── currentActivity = Parent (popped from stack)
├── Parent->resultHandler(result)
└── requestUpdate() // automatic re-render for parent
```
### Common Pitfalls
**Calling `finish()` and continuing to access `this`**: `finish()` sets `pendingAction = Pop` but does not immediately destroy the activity. The activity is destroyed on the next `ActivityManager::loop()` iteration. It's safe to access member variables after `finish()` within the same function, but don't rely on the activity surviving past the current `loop()` call.
**Modifying shared state without `RenderLock`**: If `render()` reads a variable and `loop()` writes it, the write must be under a `RenderLock`. Without it, `render()` could see a half-written value (e.g., a partially updated string or struct).
**Creating background tasks that outlive the activity**: Any FreeRTOS task created in `onEnter()` must be deleted in `onExit()` before the activity is destroyed. The `ActivityManager` does not track or clean up background tasks.
**Holding `RenderLock` across blocking calls**: The render task is blocked on the mutex while you hold the lock. Keep critical sections short — acquire, mutate state, release, then do blocking work.
```cpp
// WRONG — blocks render for the entire network call
void MyActivity::doNetworkStuff() {
RenderLock lock;
state = LOADING;
auto result = http.get(url); // blocks for seconds with lock held
state = DONE;
}
// CORRECT — release lock before blocking
void MyActivity::doNetworkStuff() {
{
RenderLock lock;
state = LOADING;
}
requestUpdate(true); // render "Loading..." immediately, before we block
auto result = http.get(url); // lock is not held
{
RenderLock lock;
state = DONE;
}
requestUpdate();
}
```