Re: [RFC] c/r: prctl: Add ability to set new mm_struct::exe_file

[Oleg Nesterov <oleg@redhat.com>]

On 02/29, Cyrill Gorcunov wrote:
>
> +static int prctl_set_mm_exe_file(struct mm_struct *mm,
> +				 const void __user *path,
> +				 size_t size)
> +{
> +	struct file *new_exe_file;
> +	char *pathbuf;
> +	int ret = 0;
> +
> +	if (size >= PATH_MAX)
> +		return -EINVAL;
> +
> +	/*
> +	 * We allow to change only those exe's which
> +	 * are not mapped several times. This one
> +	 * is early test while mmap_sem is taken.
> +	 */
> +	if (mm->num_exe_file_vmas > 1)
> +		return -EBUSY;

I don't really understand this check, but it is racy. Another thread
can change ->num_exe_file_vmas right after the check.

> +       up_read(&mm->mmap_sem);

up? I do not see down...

> +	new_exe_file = open_exec(pathbuf);
> +	kfree(pathbuf);
> +
> +	down_read(&mm->mmap_sem);

probably you meant "up" here. OK, I am ignoring ->mmap_sem, I can't
understand what did you really mean ;)

> +	if (IS_ERR(new_exe_file))
> +		return PTR_ERR(new_exe_file);
> +
> +	/*
> +	 * We allow to change only those exe's which
> +	 * are not mapped several times.
> +	 */
> +	if (mm->num_exe_file_vmas < 2) {
> +		set_mm_exe_file(mm, new_exe_file);
> +		ret = 0;
> +	} else
> +		ret = -EBUSY;
> +
> +	return ret;

Both success/EBUSY leak new_exe_file. And I agree with Pavel,
prctl_set_mm_exe_file() should take fd, not filename.

I simply can't understand why set_mm_exe_file() is safe. What
if we race with another thread doing set_mm_exe_file() too?
Or it can race with added_exe_file_vma/removed_exe_file_vma.

And. set_mm_exe_file() sets ->num_exe_file_vmas = 0, this is
simply wrong? It should match the number of VM_EXECUTABLE
vmas.

In short, I do not understand the patch at all. It seems, you
only need to replace mm->exe_file under down_write(mmap_sem)
and nothing else.

Oleg.