Lua opens scripts to load them into memory, then the logger opens them
after to stream them into the dataflash log. When loading multiple large
Lua scripts from ROMFS, decompression takes a significant amount of
time. This creates the opportunity for the Lua interpreter and logging
threads to both be inside `AP_Filesystem_ROMFS::open()` decompressing a
file.
If this happens, the function can return the same `fd` for two different
calls as the `fd` is chosen before decompression starts, but only marked
as being used after that finishes. The read pointers then stomp on each
other, so Lua loads garbled scripts (usually resulting in a syntax
error) and the logger dumps garbled data.
Fix the issue by locking before searching for a free record (or marking
a record as free). Apply the same fix to directories as well. This
doesn't protect against using the same `fd`/`dirp` from multiple
threads, but that behavior is to be discouraged anyway and is not the
root cause here.
There are currently issues where the non-.exe-suffixed files can't be
copied into the `artifacts` folder; `cp` claims "File exists".
Previously this worked but the suffix was added by Cygwin so all files
in `artifacts` had a `.exe` suffix anyway.
This is evidently intended, though non-intuitive, behavior:
https://sourceware.org/legacy-ml/cygwin/2009-08/msg00293.html
> On Cygwin, you should avoid having a file "foo" and a file "foo.exe"
> in the same directory at all cost to avoid puzzeling POSIX borderline
> behaviour like this. What you do is essentially in the "not
> supported" class of problems.
> [...] Cygwin does not check for a file "foo", if the name of the file
> is explicitely given as "foo.exe".
Apparently something similar was addressed in PR #20926; the current
code installs files with both suffixes, but that fix contradicts the
info above and now has broken.
This PR changes the code to only install .exe-suffixed files, as opposed
to only non-.exe-suffixed files, which was the behavior before that PR.
There might be multiple copies of the flags, so remove all of them. Also
pull the setup into a helper function to avoid duplicating the logic.
On ChibiOS boards, this fixes SITL (which had issues with position
quantization) and affects the SBF GPS driver (which was probably fine
before but does change slightly).
