Sidenote: Linux does support a less awful way to change the kernel's view
of argv these days, using prctl(PR_SET_MM, PR_SET_MM_ARG_START (or _END),
addr, 0, 0). Sadly, it only allows root (CAP_SYS_RESOURCE) to use it. I'm
not sure why, perhaps that restriction could be relaxed for future
kernels...

See
https://github.com/systemd/systemd/blob/87305b0fbfc0e40a948cf0a683bcf9d47b8a41a3/src/basic/process-util.c#L256
for an example of use (including ugly workaround for the API being silly
and setting START/END with separate syscalls, but requiring START <= END at
all times)

On Wed, Aug 17, 2022 at 6:05 AM Elvis Pranskevichus <elvis@edgedb.com>
wrote:

> There is no guarantee that the environment block will remain intact.
> For example, PostgreSQL clobbers argv/environ area to implement its
> "setproctitle" emulation on non-BSD [1], and there is a popular Python
> library inspired by it [2].  As a result, setting `LD_LIBRARY_PATH`
> or `LD_PRELOAD` has no effect on Postgres subprocesses when linking
> against musl.
>
> Protect against this by making a copies instead of storing the
> original pointers directly.
>
> (please CC me, I'm not subscribed to the list)
>
> ---
>  ldso/dynlink.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/ldso/dynlink.c b/ldso/dynlink.c
> index cc677952..703342b8 100644
> --- a/ldso/dynlink.c
> +++ b/ldso/dynlink.c
> @@ -1756,8 +1756,8 @@ void __dls3(size_t *sp, size_t *auxv)
>
>         /* Only trust user/env if kernel says we're not suid/sgid */
>         if (!libc.secure) {
> -               env_path = getenv("LD_LIBRARY_PATH");
> -               env_preload = getenv("LD_PRELOAD");
> +               env_path = strdup(getenv("LD_LIBRARY_PATH"));
> +               env_preload = strdup(getenv("LD_PRELOAD"));
>         }
>
>         /* Activate error handler function */
>
>
>
>
>