It's a fallback that is typically never attempted on linux:

arc4_seed(void)
{
	int ok = 0;
	/* We try every method that might work, and don't give up even if one
	 * does seem to work.  There's no real harm in over-seeding, and if
	 * one of these sources turns out to be broken, that would be bad. */
#ifdef TRY_SEED_WIN32
	if (0 == arc4_seed_win32())
		ok = 1;
#endif
#ifdef TRY_SEED_URANDOM
	if (0 == arc4_seed_urandom())
		ok = 1;
#endif
#ifdef TRY_SEED_PROC_SYS_KERNEL_RANDOM_UUID
	if (arc4random_urandom_filename == NULL &&
	    0 == arc4_seed_proc_sys_kernel_random_uuid())
		ok = 1;
#endif
#ifdef TRY_SEED_SYSCTL_LINUX
	/* Apparently Linux is deprecating sysctl, and spewing warning
	 * messages when you try to use it. */
	if (!ok && 0 == arc4_seed_sysctl_linux())
		ok = 1;
#endif

So if /dev/urandom is accessible, or if /proc/sys/kernel/random/uuid is accessible, it doesn't actually try the sysctl libc call. So if you can ignore the compiler warning, it's fine :)

Anyway, that sysctl strategy should probably be replaced by getrandom() if available.

Replaced with getrandom, @ploxiln thanks for the suggestion and @Jingzhao123 thanks for the report!

very nice, thanks @azat

By the way, OpenBSD and MacOS have a getentropy() syscall that is very similar, and I just noticed that glibc also offers this function, for which it uses getrandom() underneath: http://manpages.ubuntu.com/manpages/bionic/man3/getentropy.3.html

But arc4_seed() is probably fine as-is if the bsd sysctl method still works fine on those OS anyway.

MacOS have a getentropy()

There are some issues on osx sierra at least (see #398), so let's keep it as-is for now, to avoid introducing yet another build switch.