We had several cases of code needing a strong alias, so we switch those
to use the new macro. This covers systems that support the alias
attribute and others such as macOS where we need to use assembler
directives to add the alias as the attribute is not supported.
The environ variable is supposed to be defined by the code using it, but
on glibc-based systems it will get defined if we request it, by including
<unistd.h> and defining _GNU_SOURCE.
The two arrays might not reference contiguous memory, and assuming they
are does break at least now on GNU/Hurd, which contains an unmapped
memory block between the memory used by the two arrays.
Just check that each element is strictly after the previous one, so that
we know there are no unmapped memory blocks inbetween.
The automatic initialization cannot be part of the main shared library,
because there is no thread-safe way to change the environ global
variable. This is not a problem if the initializaion happens just at
program load time, but becomes one if the shared library is directly or
indirectly dlopen()ed during the execution of the program, which could
have either kept references to the old environ or could change it in
some other thread. This has been observed for example on systems using
Samba NSS modules.
To avoid any other possible fallout, the constructor is split into a
new static library that needs to be linked explicitly into programs
using setproctitle(). As an additional safety measure the pkg-config
linker flags will mark the program as not allowing to be dlopen()ed
so that we avoid the problem described above.
Reported-by: Jan Alexander Steffens (heftig) <jan.steffens@gmail.com>
Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=66679
Because clearenv() or setenv() might free the environ array of pointers,
we should make sure to copy it so that we can access it later on when
doing the deep copy via setenv().
Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=65470
The GNU .init_array support is an extension over the standard System V
ABI .init_array support, which passes the main() arguments to the init
function.
This support comes in three parts. First the dynamic linker (from glibc)
needs to support it. Then function pointers need to be placed in the
section, for example by using __attribute__((constructor)), that the
compiler (gcc or clang for example) might place in section .ctors and
the linker (from binutils) will move to .init_array on the output
object, or by placing them directly into .init_array by the compiler
when compiling. If this does not happen and the function pointers end
up in .ctors, then they will not get passed the main() arguments, which
we do really need in this case.
But this relies on recent binutils or gcc having native .init_array
support, and not having it disabled through --disable-initfini-array.
To guarantee we get the correct behaviour, let's just place the function
pointer in the .init_array section directly, so we only require a recent
enough glibc.
Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=65029
Make the 0.5 version the default, so that code wanting the actual
implemented version can get a proper versioned depdendency. For code
linked against the old version, make it available as an alias.
Use local getprogname()/setprogname() instead of reimplementing them
locally. Use clearenv() if available, not just on glibc. Use bool
instead of _Bool. Use paranthesis on sizeof. Fold the SPT_MIN macro
into spt_min(). Make spt_init() static. Avoid unnecessary gotos.
