Rework peripheral reset and always use finalizers

[tannewt]

While investigating #8726 I realized the crash was due to a conflict between "bulk" reset such as "pulsein_reset()" and "deinit()" based reset that is triggered by deiniting the heap. With early ports, we relied heavily on "bulk" reset so that we knew exactly the state of peripherals at the start of every run. However, we also implemented deinit() so that user code can deinit peripherals at a precise time. (del and finalizers are called at collect time or vm end.) If deinit() wasn't called but the object was GCed, then the resources would "leak" until the next "bulk" reset. More and more objects are now allocated with deinit() as the finalizer so that when the object is GCed, the resources aren't (hopefully) temporarily leaked. The original bug was caused when PulseIn was switched to using finalizers but not removing/changing the "bulk" reset.

We should refactor the CircuitPython API to always use finalizers and remove the bulk resets. That way we'll only have one way that resets occur and it doesn't require state outside the CP objects to achieve it.

The main external data we have is "never reset" information that keeps resources that are actually still in use from being reset by "bulk" reset (but not deinit() reset). This too is a source for bugs. Once we remove bulk resets, we can also remove never_reset because there is no danger for mis-deinit. It is tied to explicit deinit() calls. Any class that needs to live longer than the VM will need to be allocated outside the VM heap so that its finalizer isn't called at VM heap end. It must be explicitly managed instead.

(@dhalbert edit):
May be able to use finalizers; now use *_reset() or reset_*(). List gathered from reset_port() in all ports.

• - audioio.AudioOut (switch atmel-samd AudioOut to use finalizer #9165)
• audiobusio.PDMIn
• audiobusio.I2SOut
• i2s
• frequencyio.FrequencyIn
• touchio.TouchIn
• analogio.AnalogIn
• analogio.AnalogOut
• watchdog
• pew
• busio.I2C
• busio.SPI
• busio.UART
• rtc (? reset is in shared-bindings)
• PWMOut: obsolete reference in broadcom
• espcamera
• ssl
• dualbank
• espnow
• espulp
• ps2io
• sdioio
• socketpool
• touchio
• neopixel_write
• countio
• rp2pio
• bleio

[tannewt]

We can wait until micropython#13682 is released to make this simpler.

[timchinowsky]

Is the resolution to this issue exactly to make sure that all of these classes call deinit() in their __del__ method? Or is more required? Looking at the classes above, it looks like some do this and some do not. And, as far as documentation, am I right that explicit calls to deinit() should be de-emphasized? You indicate that sometimes explicit calls can be useful to control the timing of the deinit(), but this is not usually necessary, right?

[dhalbert]

The resolution is to make all these objects have finalizers and have __del__() call deinit(). In addition we can then remove the bulk reset operations like pulsein_reset() (already removed).

You can also use a context manager to do the deinit on context exit.

The only previous reason to call deinit() was if you wanted free up a piece of hardware that was being held by a native class. That is still true. All the hardware was also supposed to be freed up when the VM shut down and then restarted, but that was less reliable than we would have liked. Instead of doing it manually, we can depend on the garbage collector to call the finalizers to do it. And a hard reset will of course reset everything.

There are long-lifetime things that exist across VM boundaries, such as display objects, those don't get gc'd, so their finalizer isn't called.

We have a long list of these to do, but it's not a huge task.