This solves a potential dead-lock when trying to shutdown: a call to exit()
stops all threads and calls all destructors for static objects.
The destructor of LockstepScheduler takes a lock. However this is not
safe, as the lock could already be taken (by any thread).
Some threads do not exit and are still running when
trying to exit SITL running under Windows in Cygwin.
This problem was introduced with:
- posix shell #10173 because of strating a child process
for the startup script and mixing up the signal handling
(only Ctrl+C broken)
- lockstep #10648 because of simulator threads not
stopping gracefully anymore with timing race conditions
(no graceful exit possible anymore)
I leave the SIGINT handler on its default implementation for
Cygwin which kills the process and all its threads when pressing
Ctrl+C.
This hotfix at least allows the usage of Ctrl+C again instead
of forcing the user to use the task manager to shut down
px4.exe going crazy on CPU load instead of exiting
everytime.
- use a linked-list instead of std::vector. Insertion and removal are now
O(1)
- avoid malloc and use a thread_local instance of TimedWait.
It gets destroyed when the thread exits, so we have to add protection
in case a thread exits too quickly. This in turn requires a fix to the
unit-tests.
The API of cond_timedwait was wrong. It used return -1 and set errno
instead of returning the error as specified for pthread_cond_timedwait
which it tries to mock.
@bkueng found that the old implementation was likely to wrap-around
given seconds is only a uint32_t. We now cast it directly to uint64_t
and therefore should fix this problem.
This uses the "fake" px4_sem based on mutex and condition_variable on
all POSIX system, not just macOS and Cygwin. This means that we can
change px4_sem_timedwait under the hood and inject the simulated time.
This integrates the lockstep_scheduler, so that the system time is set
by the mavlink HIL_SENSOR message.
This means that the speed factor is removed and the speed is entirely
given by the simulator.
These contains some rough changes trying to get SITL to speed up by a
SPEED_FACTOR.
This platform time code probably requires some more thought and refactor
but this gets a demo at 4x working.
The threads running commands for clients through the Posix daemon used
to write to a char buffer through snprintf (etc.) which was then written
directly to the file descriptor, whereas in the other case printf
(etc.) was used to write to stdout (FILE*). Both versions used some
macro's and repeated code to have the same output.
This change unifies these two cases by using a FILE* in both cases. The
(line) buffering is done by the standard C library's implementation
(just like with stdout), and px4_log.c now uses the same code in all
cases (using fprintf, etc.) for printing (colored) output.
Unlike pipes, unix sockets provide bi-directional
communication with each connected client.
- No need to generate a unique uuid per client anymore.
- The client doesn't have to create its own pipe anymore.
- Since there is no risk of multiple client's writes getting mixed up,
messages don't need to fit in a single write anymore, removing the
limit on command length.
- Since the server can detect a connection closing, the client no longer
needs to handle signals. When the client is killed, the connection is
automatically closed, which will cause the server to kill the related
px4 thread.
Since this does not rely on handling signals and the client sending an
additional message, this is much more reliable.
- Client is no longer a singleton.
- The protocol is simplified. Standard output is directly written to the
socket back to the client, without wrapping it in any protocol
message.
- Because of the simple protocol, one could now even use netcat to run a
px4 command:
$ echo hello | netcat -UN /tmp/px4-sock-0
Also removes a few race conditions.
Beat Küng
Daniel Agar
Matthias Grob
Lorenz Meier
karliss
Julian Oes
Kārlis Seņko
Mara Bos
dkang