Re: [RFCv2 03/10] x86: Introduce userspace API to handle per-thread features

[Thomas Gleixner]

On Wed, May 11 2022 at 05:27, Kirill A. Shutemov wrote:
> Add three new arch_prctl() handles:
>  
> +static long thread_feature_prctl(struct task_struct *task, int option,
> +				 unsigned long features)

Bah. I really hate the task pointer on all these x86 prctls. @task must
always be current, so this @task argument is just confusion.

> +{
> +	const unsigned long known_features = 0;
> +
> +	if (features & ~known_features)
> +		return -EINVAL;

This implementation allows to task->read features_[locked] with
@features == 0. That should be documented somewhere.

> +	if (option == ARCH_THREAD_FEATURE_LOCK) {
> +		task->thread.features_locked |= features;
> +		return task->thread.features_locked;
> +	}


> +	/* Do not allow to change locked features */
> +	if (features & task->thread.features_locked)
> +		return -EPERM;
> +
> +	if (option == ARCH_THREAD_FEATURE_DISABLE) {
> +		task->thread.features &= ~features;
> +		goto out;
> +	}
> +
> +	/* Handle ARCH_THREAD_FEATURE_ENABLE */
> +
> +	task->thread.features |= features;
> +out:
> +	return task->thread.features;

Eyes bleed.

	if (option == ARCH_THREAD_FEATURE_ENABLE)
        	task->thread.features |= features;
        else
        	task->thread.features &= ~features;

	return task->thread.features;

It's bloody obvious from the code that the counterpart of enable is
disable, no? So neither the goto nor the 'comment the obvious' is useful
in any way.

Thanks,

        tglx