mirror of
https://github.com/trezor/trezor-firmware.git
synced 2024-12-22 14:28:07 +00:00
docs: document trezor-core event loop
This commit is contained in:
parent
4a889f34cd
commit
c16f54ad04
195
docs/core/src/event-loop.md
Normal file
195
docs/core/src/event-loop.md
Normal file
@ -0,0 +1,195 @@
|
||||
# Trezor Core event loop
|
||||
|
||||
The event loop is implemented in `src/trezor/loop.py` and forms the core of the
|
||||
processing. At boot time, default tasks are started and inserted into an event queue.
|
||||
Such task will usually run in an endless loop: wait for event, process event, loop back.
|
||||
|
||||
Application code is written with `async/await` constructs. Low level of the event queue
|
||||
processes running coroutines via `coroutine.send()` and `coroutine.throw()` calls.
|
||||
|
||||
### MicroPython details
|
||||
|
||||
MicroPython does not distinguish between coroutines, awaitables, and generators. Some
|
||||
low-level constructs are using `yield` and `yield from` constructions.
|
||||
|
||||
`async def` definition marks the function as a generator, even if it does not contain
|
||||
`await` or `yield` expressions. It is thus possible to see `async def __iter__`, which
|
||||
indicates that the function is a generator.
|
||||
|
||||
For type-checking purposes, objects usually define an `__await__` method that delegates
|
||||
to `__iter__`. The `__await__` method is never executed, however.
|
||||
|
||||
|
||||
## Low-level API
|
||||
|
||||
### Function summary
|
||||
|
||||
`loop.run()` starts the event loop. The call only returns when there are no further
|
||||
waiting tasks -- so, in usual conditions, never.
|
||||
|
||||
`loop.schedule(task, value, deadline, finalizer)` schedules an awaitable to be run
|
||||
either as soon as possible, or at a specified time (given as a `deadline` in
|
||||
microseconds since system bootup.)
|
||||
|
||||
In addition, when the task finishes processing or is closed externally, the `finalizer`
|
||||
callback will be executed, with the task and the return value (or the raised exception)
|
||||
as a parameter.
|
||||
|
||||
`loop.close(task)` removes a previously scheduled task from the list of waiting tasks
|
||||
and calls its finalizer.
|
||||
|
||||
`loop.pause(task, interface)` sets the task as waiting for a particular _interface_:
|
||||
either reading from or writing to one of the USB interfaces, or waiting for a touch
|
||||
event.
|
||||
|
||||
### Implementation details
|
||||
|
||||
Trezor Core runs coroutine-based cooperative multitasking, i.e., there is no preemption.
|
||||
|
||||
Every _task_ is a coroutine, which means that it runs uninterrupted until it yields a
|
||||
value (or, in async terms, until it awaits something). In every processing step, the
|
||||
currently selected coroutine is resumed by sending a value to it (which is returned as a
|
||||
result of the yield/await, or raised as an exception if it is an instance of
|
||||
`BaseException`). The tasks then runs uninterrupted again, until it yields or exits.
|
||||
|
||||
A loop in `loop.run()` spins for as long as any tasks are waiting. Two lists of waiting
|
||||
tasks exist:
|
||||
|
||||
- `_queue` is a priority queue where the ordering is defined by real-time deadlines.
|
||||
In most cases, tasks are scheduled for "now", which makes them run one after another
|
||||
in FIFO order. It is also possible to schedule a task to run in the future.
|
||||
|
||||
- `_paused` is a collection of tasks grouped by the interface for which they are
|
||||
waiting.
|
||||
|
||||
In each run of the loop, `io.poll` is called to query I/O events. If an event arrives on
|
||||
an interface, _all_ tasks waiting on that interface are resumed one after another. No
|
||||
scheduled tasks in `_queue` can execute until the waiting tasks yield again.
|
||||
|
||||
At most one I/O event is processsed in this phase.
|
||||
|
||||
When the I/O phase is done, a task with the highest priority is popped from `_queue` and
|
||||
resumed.
|
||||
|
||||
### I/O wait
|
||||
|
||||
When no tasks are paused on a given interface, events on that interface remain in queue.
|
||||
|
||||
When multiple tasks are paused on the same interface, all of them receive every event.
|
||||
However, a waiting task receives at most one event. To receive more, it must pause
|
||||
itself again. Event processing is usually done in an endless loop with a pause call.
|
||||
|
||||
If two tasks are attempting to read from the same interface, and one of them re-pauses
|
||||
itself immediately while the other doesn't (possibly due to use of `loop.race`, which
|
||||
introduces scheduling gaps), the other task might lose some events.
|
||||
|
||||
For this reason, you should avoid waiting on the same interface from multiple tasks.
|
||||
|
||||
### Syscalls
|
||||
|
||||
Syscalls bridge the gap between `await`-based application code and the coroutine-based
|
||||
low-level implementation.
|
||||
|
||||
Every sequence of `await`s will at some point boil down to yielding a `Syscall`
|
||||
instance. (Yielding anything else is an error.) When that happens, control returns to
|
||||
the event loop.
|
||||
|
||||
The `handle(task)` method is called on the result. This way the syscall gets hold of the
|
||||
task object, and can `schedule()` or `pause()` it as appropriate.
|
||||
|
||||
As an example, consider pausing on an input event. A running task has no way to call
|
||||
`pause()` _on itself_. It would need to pass a separate function as a callback.
|
||||
|
||||
The `wait` syscall can be implemented as a simple wrapper around the `pause()` low-level
|
||||
call:
|
||||
|
||||
```python
|
||||
class wait(Syscall):
|
||||
def __init__(self, msg_iface: int) -> None:
|
||||
self.msg_iface = msg_iface
|
||||
|
||||
def handle(self, task: Task) -> None:
|
||||
pause(task, self.msg_iface)
|
||||
```
|
||||
|
||||
The `__init__()` method takes all the arguments of the "call", and `handle()` pauses the
|
||||
task on the given interface.
|
||||
|
||||
Calling code will look like this:
|
||||
```python
|
||||
event = await loop.wait(io.TOUCH)
|
||||
```
|
||||
The `loop.wait(io.TOUCH)` expression instantiates a new `Syscall` object. The argument
|
||||
is passed to the constructor, and stored on the instance. The rest boils down to
|
||||
```python
|
||||
event = await some_syscall_instance
|
||||
```
|
||||
which is equivalent to
|
||||
```python
|
||||
event = yield from some_syscall_instance.__iter__()
|
||||
```
|
||||
The `Syscall.__iter__()` method yields `self`, returning control to the event loop. The
|
||||
event loop invokes `some_syscall_instance.handle(task_object)`. The `task_object` is
|
||||
then set to resume when a touch event arrives.
|
||||
|
||||
A side-effect of this design is that it is possible to store and reuse syscall
|
||||
instances. That can be advantageous for avoiding unnecessary allocations.
|
||||
```python
|
||||
while True:
|
||||
# every run of the loop allocates a new object
|
||||
event = await loop.wait(io.TOUCH)
|
||||
process_event(event)
|
||||
|
||||
touch_source = loop.wait(io.TOUCH)
|
||||
while True:
|
||||
# same instance is reused
|
||||
event = await touch_source
|
||||
process_event(event)
|
||||
```
|
||||
|
||||
## High-level API
|
||||
|
||||
Application code should not be using any of the above low-level functions. Awaiting
|
||||
syscalls is the preferred method of writing code.
|
||||
|
||||
The following syscalls and constructs are available:
|
||||
|
||||
**`loop.sleep(delay_us: int)`**: Suspend execution until the given delay (in
|
||||
microseconds) elapses. Sub-millisecond precision is not guaranteed, however.
|
||||
Return value is the planned deadline in microseconds since system start.
|
||||
|
||||
Calling `await loop.sleep(0)` yields execution to other tasks, and schedules the current
|
||||
task for the next tick.
|
||||
|
||||
**`loop.wait(interface)`**: Wait indefinitely for an event on the given interface.
|
||||
Return value is the event.
|
||||
|
||||
_Upcoming code modification adds a timeout parameter to `loop.wait`._
|
||||
|
||||
**`loop.race(*children)`**: Schedule each argument to run, and suspend execution until
|
||||
the first of them finishes.
|
||||
|
||||
It is possible to specify wait timeout for `loop.wait` by using `loop.race`:
|
||||
```python
|
||||
result = await loop.race(loop.wait(io.TOUCH), loop.sleep(1000 * 1000))
|
||||
```
|
||||
This introduces scheduling gaps: every child is treated as a task and scheduled
|
||||
to run. This means that if the child is a syscall, as in the above example, its action
|
||||
is not done immediately. Instead, the `wait` begins on the next tick (or whenever the
|
||||
newly created coroutine runs) and the `sleep` in the tick afterwards. When nesting
|
||||
multiple `race`s, the child `races` also run later.
|
||||
|
||||
Also, when a child task is done, another scheduling gap happens, and the parent task
|
||||
is scheduled to run on the next tick.
|
||||
|
||||
_Upcoming changes may solve this in relevant cases, by inlining syscall operations._
|
||||
|
||||
**`loop.chan()`** is a unidirectional communication channel that actually implements two
|
||||
syscalls:
|
||||
|
||||
* **`chan.put()`** sends a value to the channel, and waits until it is picked up
|
||||
by a taker task.
|
||||
* **`chan.take()`** waits until a value is sent to the channel and then returns it.
|
||||
|
||||
It is possible to put in a value without waiting for a taker, by calling
|
||||
`chan.publish()`. It is not possible to take a value without waiting.
|
Loading…
Reference in New Issue
Block a user