This commit prevents the device from "hanging" when using lightsleep while
the WiFi chip is active.
Whenever the WiFi chip wants to interrupt the microcontroller to notify it
for a new package, it sets the CYW43_PIN_WL_HOST_WAKE pin to high,
triggering an IRQ. However, as polling the chip cannot happen in an
interrupt handler, it subsequently notifies the pendsv-service to do a poll
as soon as the interrupt handler ended. In order to prevent a new
interrupt from happening immediately afterwards, even before the poll has
run, the IRQ handler disables interrupts from the pin.
The first problem occurs, when a WiFi package arrives while the main loop
is in cyw43-code. In order to prevent concurrent access of the hardware,
the network code blocks pendsv from running again while entering lwIP code.
The same holds for direct cyw43 code (like changing the cyw43-gpios, i.e.
the LED on the Pico W). While the pendsv is disabled, interrupts can still
occur to schedule a poll (and disable further interrupts), but it will not
run. This can happen while the microcontroller is anywhere in rp2040 code.
In order to preserve power while waiting for cyw43 responses,
cyw43_configport.h defines CYW43_DO_IOCTL_WAIT and
CYW43_SDPCM_SEND_COMMON_WAIT to __WFI(). While this might work in most
cases, there are 2 edge cases where it fails:
- When an interrupt has already been received by the cyw43 stack, for
example due to an incoming ethernet packet.
- When the interrupt from the cyw43 response comes before the
microcontroller entered the __WFI() instruction.
When that happens, wfi will just block forever as no further interrupts are
received. The only way to safely use wfi to wake up from an interrupt is
inside a critical section, as this delays interrupts until the wfi is
entered, possibly resuming immediately until interrupts are reenabled and
the interrupt handler is run. Additionally this critical section needs to
check whether the interrupt has already been disabled and pendsv was
triggered, as in such a case, wfi can never be woken up, and needs to be
skipped, because there is already a package from the network chip waiting.
Note that this turns cyw43_yield into a nop (and thereby the cyw43-loops
into busy waits) from the second time onwards, as after the first call, a
pendsv request will definitely be pending. More logic could be added, to
explicitly enable the interrupt in this case.
Regarding lightsleep, this code has a similar problem. When an interrupt
occurs during lightsleep, the IRQ and pendsv handler and thereby poll are
run immediately, with the clocks still disabled, causing the SPI transfers
to fail. If we don't want to add complex logic inside the IRQ handler we
need to protect the whole lightsleep procedure form interrupts with a
critical section, exiting out early if an interrupt is pending for whatever
reason. Only then we can start to shut down clocks and only enable
interrupts when the system is ready again. Other interrupt handlers might
also be happy, that they are only run when the system is fully operational.
Tested on a Pico W, calling machine.lightsleep() within an endless loop and
pinging from the outside.
This includes:
- Configuration file for the cyw43-driver.
- Integration of cyw43-driver into the build, using lwIP.
- Enhancements to machine.Pin to support extension IO pins provided by the
CYW43xx.
- More mp-hal pin helper functions.
- mp_hal_get_mac_ascii MAC address helper function.
- Addition of rp2.country() function to set the country code.
A board can enable this driver by setting MICROPY_PY_NETWORK_CYW43 in their
cmake snippet.
Work done in collaboration with Graham Sanderson and Peter Harper.
Signed-off-by: Damien George <damien@micropython.org>
It is more reliable and scales better when more components need it.
Work done in collaboration with Graham Sanderson and Peter Harper.
Signed-off-by: Damien George <damien@micropython.org>
The callback passed to add_alarm_in_ms must return microseconds, even
though the initial delay is in milliseconds. Fix this use, and to avoid
further confusion use the add_alarm_in_us function instead.
Signed-off-by: Damien George <damien@micropython.org>
