This enables flow control on CDCACM for the NuttX boards which fixes a
problem where HITL would stall.
The stall could happen if the hardware would be a bit too slow in
keeping up with the incoming messages. Often, this happened on arming
because the logger would take some time to log all parameters right at
the beginning.
The stall would then not recover due to NuttX bug where the rx interrupt
would not be restored correctly and instead only a slower watchdog would
release the next read. This watchdog takes 200ms which means it's hard
to impossible to get out of the situation without restarting sim and/or
PX4. For more information about the issue, see:
apache/incubator-nuttx#3633
As a workaround, until that bug is fixed, and because it makes sense
anyway, I propose to enable FLOWCONTROL for the serial via USB.
- cmake NuttX build wrapper compile in place instead of copying source tree to build directory
- slightly faster skipping necessary copying (depending on system)
- allows debugging in place
- easier to work directly in NuttX following official documentation
- simplifies overall build which should make it easier to resolve any remaining NuttX dependency issues in the build system
- the downside is switching back and forth between different builds always require rebuilding NuttX, but I think this is worth the improved developer experience
- also no longer builds px4io and bootloader in every single build, for most users these rarely change and we're wasting a lot of build time
- increase stack sizes to run cleanly under stackcheck
- this is likely overkill for most boards, but using stackcheck to set our minimum ensures we have a very safe margin on regular builds and it's something we can currently afford
- remove holybro_durandal-v1_stackcheck from test rack (there's only one unit)
- update to NuttX with stm32f4 and stm32f7 SPI DMA internal buffers
- remove explicit DMA buffer allocations from new IMU drivers
- restore original BOARD_DMA_ALLOC_POOL_SIZE
- decrease SPI DMA thresholds based on fmu-v2/v3/v4/v5 bench testing
On more complicated setups it's still possible to exceed 32 tasks. For example fmu-v5 with mavlink on every telem (+ USB), external spi usage (pmw3901), gimbal (vmount), multiple i2c sensors, and camera feedback is 35 tasks (with top running). This is a fairly extreme case, so I'm only going to increase CONFIG_MAX_TASKS on newer F7 boards.
But only on the first 4 FMU outputs, as the next ones conflict with px4io
serial dma (UART8_RX)
RX DMA is disabled on the GPS port as well (conflicts with TIM1).
The cause of the stack detection fault is because of the
level of nesting in the start up script. We need to
determine the worst case configuration and set the
bar there.
This fault occurred some 42 calls deep due to script
calling script (repeat).
The HW stack check requires as a margin of 204 bytes. That is
ISR HW stacking of CPU(8) FPU(18) registers and SW stacking of
CPU(11) and FPU(16) registers. Total CPU(19) registers is
68 bytes and the total FPU(34) registers is 136 bytes. On
a system with a separate ISR stack This only needs to be 104
so there is 100 bytes of headroom. But as coded the detection
will give a false positive detection and fault. This does not
mean that the stack will be corrupted.
Adjustments to that stack can have no effect due to rounding.
A stack size of 2608 and 2616 can yield the exact same size stack.
So even when the failure is due to a 4 byte overflow, it can take
greater than a 16 bytes increase to fix it. Because the final
stack size is calculated with an 8 byte alignment after a 4 byte
decrease. So 2624 becomes 2620 at runtime and will boot
with SYS_AUTOSTART=4001.
Daniel Agar