8KB was empirically determined to work on musl. We can't go lower (musl
supports up to 2KB) because posix_spawn allocates a temporary stack for
vfork (via clone(2)) on its own stack.
This improves our resource consumption and increases the likeliness for
pthread_create to work even when there isn't a lot of free memory.
Using pthread_setattr_default_np isn't possible, because on musl it
doesn't allow us to go below the default thread stack size. This way is
also more portable.
We don't remove the guard pages entirely, because a shell prompt might
be active in attacker controlled directories, and it's a reasonable
protection against any stack overflow exploits that might appear.
Avoid unnecessary heap allocations (though we added new ones in main())
if possible by using a PATH_MAX buffer.
Check for PWD matching current directory ourselves, which allows us to
save a stat syscall, by re-using the stat(".") result to compare to
stat("$HOME"). Improve how we check if HOME is actually part of pwd -
previous would get confused with "${HOME}char", which should not get the
prompt to display '~'.
Use strtok_r to scan characters now. Writing directly to p() makes more
sense than writing chars to a buffer, so remove that.
I'm somewhat unhappy with the rpwd and mrpwd solution, since while
coding I forgot to set one or the other a couple of times. It would be
nice if I could "unconst" the variable at some point, though I greatly
dislike that pattern.
Member launched of struct threaded_task should be accessed after some
sort of synchronization happens (what we currently do is join the thread
that touches the struct), so it doesn't need to be an atomic variable.
Érico Nogueira