Re: [PATCH v5 2/3] selftests/lam: Skip test if LAM is disabled

[Dave Hansen]

On 11/27/24 09:35, Maciej Wieczor-Retman wrote:
> +static inline int kernel_has_lam(void)
> +{
> +	unsigned long bits;
> +
> +	syscall(SYS_arch_prctl, ARCH_GET_MAX_TAG_BITS, &bits);
> +	return !!bits;
> +}

Generally, I'm less picky about selftest/ code than in-kernel code. But
people really do take selftest code and use it as a starting point for
production code.

I'd much rather have overly verbose, obviously correct code:

	err = syscall(SYS_arch_prctl, ARCH_GET_MAX_TAG_BITS, &bits);

	/* Handle syscall failure, like pre-LAM kernels: */
	if (err)
		return 0

	/* Tag bits are empty on non-LAM systems: */
	return !!bits;

Actually, I was going to argue for that^ just on style and writing good
code. But then I spotted a bug. What happens if the kernel has
CONFIG_ADDRESS_MASKING=n, either because it is config'd off or it's old?
The:

	put_user(0, (unsigned long __user *)arg2);

won't ever get run and 'bits' will be uninitialized.

So, I think this code was trying to be compact, fast and clever. But it
really just turns out to be buggy.