Implementation of alloca() is very much architecture and compiler
dependent. Avoid the case in which it could return a non-aligned
pointer, which would mean Thread::_poison_stack() would do the wrong
thing.
../../libraries/AP_HAL_Linux/Thread.cpp: In member function ‘void Linux::Thread::_poison_stack()’:
../../libraries/AP_HAL_Linux/Thread.cpp:63:20: warning: declaration of ‘start’ shadows a member of 'this' [-Wshadow]
uint8_t *end, *start, *curr;
^
Sort include alphabetically and make them in order:
Main header
system headers
library headers
local headers
While reordering, change a include of endian.h to our sparse-endian.h
which is more reliant to toolchain changes.
Running the vehicles we check the stack size doesn't grow too much by
enabling the DEBUG_STACK in the scheduler. Even on 64bit boards the
stack is consistent around 4k. Just to be a little conservative, let it
be a little bit more that that: 256kB.
Since we have RT prio and we call mlock(), the memory for the stack of
each thread is locked in memory. This means we are effectively taking
that much memory. The default stack size varies per distro, but it's
common to have 8MB for 64 bit boards and 4MB for 32 bit boards. Here is
the output of ps -L -o 'comm,rtprio,rss $(pidof arducopter-quad)', showing the
RSS of arducopter-quad before and after this change:
Before:
COMMAND RTPRIO RSS
arducopter-quad 12 46960
sched-timer 15 46960
sched-uart 14 46960
sched-rcin 13 46960
sched-tonealarm 11 46960
sched-io 10 46960
After:
COMMAND RTPRIO RSS
arducopter-quad 12 7320
sched-timer 15 7320
sched-uart 14 7320
sched-rcin 13 7320
sched-tonealarm 11 7320
sched-io 10 7320
We don't need all the comments in the array declaration and we can
inline its declaration in the function call. This makes it easier to
copy it to other places.
This bug led to issues for us so it may help others to resolve it.
Currently, the AP_HAL_Linux RCInput::read(uint16_t*,uint8_t) function
only returns the first x nonzero channels. Once it hits a channel that
is set to zero, it stops and all remaining channels are returned as
zero, even if they are set. This causes discrepancies between the raw RC
input sent to the GCS and the RC input that is actually used on the
vehicle.
The fixes this issue and makes it behave exactly as it does on the
PX4_HAL code. We ran into this issue when sending rc_override messages
in which there were some channels set to zero.
0-length arrays are supported in C but forbidden in C++. GCC allows it
but clang is more strict:
../../libraries/AP_HAL_Linux/SPIDriver.cpp:75:35: fatal error: no matching constructor for initialization of 'Linux::SPIDeviceDriver [0]'
SPIDeviceDriver SPIDeviceManager::_device[0];
^
../../libraries/AP_HAL_Linux/SPIDriver.h:20:7: note: candidate constructor (the implicit move constructor) not viable: requires 1 argument, but 0 were provided
class SPIDeviceDriver : public AP_HAL::SPIDeviceDriver {
^
../../libraries/AP_HAL_Linux/SPIDriver.h:20:7: note: candidate constructor (the implicit copy constructor) not viable: requires 1 argument, but 0 were provided
../../libraries/AP_HAL_Linux/SPIDriver.h:25:5: note: candidate constructor not viable: requires 9 arguments, but 0 were provided
SPIDeviceDriver(const char *name, uint16_t bus, uint16_t subdev, enum AP_HAL::SPIDeviceType type, uint8_t mode, uint8_t bitsPerWord, int16_t cs_pin, uint32_t lowspeed, uint32_t highspeed);
^
1 error generated.
This allows us to re-use SPIDevice from SPIDeviceDriver (the
to-become-SPIDeviceProperties) while the drivers are
converted. We create a fake device by calling the temporary
SPIDeviceManager::get_device() method passing the descriptor. The
transfer and assert logic is still using the old code.
Now we can interoperate SPIDeviceDriver with the ones based in
SPIDevice since they are going to use the same semaphore and bus.
The way this code is structured is a little bit different from the
SPIDriver implementation:
- We only open the bus once, no matter how many devices we have in it
- There's a single transfer() method which uses half-duplex mode
instead of full duplex. The reason is that for all cases in the
codebase we are using half-duplex transfers using the full-duplex
API, i.e. a single SPI msg with both tx and rx buffers. This is
cumbersome because the buffers need to be of the same size and the
receive buffer using an offset of the same length as the actux data
being written. This means the high level APIs need to copy buffers
around.
If later we have uses for a real full duplex case it's just a matter
of adding another transfer_fullduplex() method or something like
this.
- The methods are implemented in the SPIDevice class instead of having
proxy methods to SPIDeviceManager as is the case of SPIDriver
Also from now on we refer to the SPIDriver objects as "descriptors"
because they have the parameters of each device in the
SPIDeviceManager::devices[] table. When SPIDeviceDriver is completely
replaced we can rename them to SPIDeviceProperties.
Save in the manager the number of devices so it can be used in other
places like the SPIDevice implementation. This is a temporary storage
while we migrate to SPIDevice.
While at it use protected rather than private.
