The main thread would always be blocked on the semaphore to read the
data from accelerometer and gyroscope. Especially if we have a slow
update of these values in _accumulate() due to the I2C transfer function
taking too much time: the timer thread would never give up the CPU,
causing starvation on the main thread.
This fixes the issue by reducing the critical region using a flip-buffer
so _accumulate() can work on its own copy of the data. Now that the
critical region is smaller, also avoid the semaphore and use a spinlock
instead.
All threads share the same address space and have the same pages locked
into memory so it's not necessary to call mlockall() for each of them.
Grepping /proc/<tid>/status gives the same VmLck for all of them, even
when only the main thread locks the memory:
# for i in `seq 477 482`; do \
name=$(cat /proc/$i/comm); \
vm=$(cat /proc/$i/status |grep VmLck); \
echo -e "$name\t$vm"; \
done
ArduCopter.elf VmLck: 57868 kB
sched-timer VmLck: 57868 kB
sched-uart VmLck: 57868 kB
sched-rcin VmLck: 57868 kB
sched-tonealarm VmLck: 57868 kB
sched-io VmLck: 57868 kB
Use pthread_setname_np() to set thread name so it's easier to debug
what't going on with each of them. This is the example output of the
relevant par of "ps -Leo class,rtprio,wchan,comm":
FF 12 futex_ ArduCopter.elf
FF 15 usleep sched-timer
FF 14 hrtime sched-uart
FF 13 poll_s sched-rcin
FF 11 hrtime sched-tonealarm
FF 10 hrtime sched-io
It's undefined behavior to pass the callback to pthread to a class
member like we were doing. Refactor the code so the callbacks are static
members.
This fixes the following warnings:
libraries/AP_HAL_Linux/Scheduler.cpp: In member function 'virtual void Linux::LinuxScheduler::init(void*)':
/home/lucas/p/dronecode/ardupilot/libraries/AP_HAL_Linux/Scheduler.cpp:61:76: warning: converting from 'void* (Linux::LinuxScheduler::*)()' to 'Linux::LinuxScheduler::pthread_startroutine_t {aka void* (*)(void*)}' [-Wpmf-conversions]
(pthread_startroutine_t)&Linux::LinuxScheduler::_timer_thread);
^
libraries/AP_HAL_Linux/Scheduler.cpp:65:76: warning: converting from 'void* (Linux::LinuxScheduler::*)()' to 'Linux::LinuxScheduler::pthread_startroutine_t {aka void* (*)(void*)}' [-Wpmf-conversions]
(pthread_startroutine_t)&Linux::LinuxScheduler::_uart_thread);
^
libraries/AP_HAL_Linux/Scheduler.cpp:69:76: warning: converting from 'void* (Linux::LinuxScheduler::*)()' to 'Linux::LinuxScheduler::pthread_startroutine_t {aka void* (*)(void*)}' [-Wpmf-conversions]
(pthread_startroutine_t)&Linux::LinuxScheduler::_rcin_thread);
^
libraries/AP_HAL_Linux/Scheduler.cpp:73:76: warning: converting from 'void* (Linux::LinuxScheduler::*)()' to 'Linux::LinuxScheduler::pthread_startroutine_t {aka void* (*)(void*)}' [-Wpmf-conversions]
(pthread_startroutine_t)&Linux::LinuxScheduler::_tonealarm_thread);
^
libraries/AP_HAL_Linux/Scheduler.cpp:77:76: warning: converting from 'void* (Linux::LinuxScheduler::*)()' to 'Linux::LinuxScheduler::pthread_startroutine_t {aka void* (*)(void*)}' [-Wpmf-conversions]
(pthread_startroutine_t)&Linux::LinuxScheduler::_io_thread);
LinuxScheduler::init() was not really working as it should. This was the
result of "ps -Leo class,rtprio,wchan,comm | grep ArduCopter":
FF 12 futex_ ArduCopter.elf
FF 12 usleep ArduCopter.elf
FF 12 hrtime ArduCopter.elf
FF 12 poll_s ArduCopter.elf
FF 12 hrtime ArduCopter.elf
FF 12 hrtime ArduCopter.elf
As can be seen all the threads run with the same priority, the one of the main
thread. There were basically 2 mistakes:
1) pthread_attr_setschedpolicy() needs to be called before
pthread_attr_setschedparam(). Otherwise the latter will just return
an error and not set the priority
2) pthread_create() defaults to ignore the priority and inherit the
it from the parent thread. pthread_attr_setinheritsched() needs to
be called to change the behavior to PTHREAD_EXPLICIT_SCHED. See
pthread_attr_setinheritsched(3) for an example program to test the
behaviors.
Also, it's undefined behavior to call pthread_attr_init() several times on the
same pthread_attr_t. Although we could reutilize the same attribute without
calling pthread_attr_init() again, lets refactor the code a little bit, so all
the pthread calls are in a single place. Then also call pthread_attr_destroy()
when we are done.
In Linux the default stack size is always greater than 32k, either 2MB
or 8MB depending on the architecture. There's no point in creating a
function to lock 32k.
We are failing to execute arm-none-eabi-gcc due to missing 32 bits
libraries in travis. Install gcc-multilib so we get all of them.
This commit also calls this version of gcc in the end of
install-travis-env.sh so if it's to fail, force it to fail early on.
Staroselskii Georgii
Mikhail Avkhimenia
Sean O\'Brien
Lucas De Marchi
Andrew Tridgell
Gustavo Jose de Sousa
Randy Mackay
Jaime Machuca
Paul Riseborough