Re: [PATCH] Fix race in process_vm_rw_core

[Oleg Nesterov <oleg@redhat.com>]

On 01/13, Christopher Yeoh wrote:
>
> Hi Linus,
>
> Below is a patch which fixes the race in process_vm_core found by
> Oleg (http://article.gmane.org/gmane.linux.kernel/1235667/).
> It consolidates some code with mm_for_maps since what they do is almost
> identical.
>
> Oleg - I've kept the breakout of ptrace_may_attach and get_task_mm to
> preserve only having to take the task lock once. I see some performance
> difference with a microbenchmark but haven't had a chance to test with
> some HPC benchmarks yet so for the moment I'd like to leave it in.

I still think we should avoid the copy-and-paste code, we can do this
without the extra unlock+lock if it hurts.

However,

> At
> this stage I think its more important to get the race fixed and I'm at
> Linux.conf.au all next week. I'll send a patch out for the
> rw_copy_check_uvector cleanup after I get back from LCA.

OK, lets fix the bug first.

>  struct mm_struct *mm_for_maps(struct task_struct *task)
>  {
> -	struct mm_struct *mm;
> -	int err;
> -
> -	err =  mutex_lock_killable(&task->signal->cred_guard_mutex);
> -	if (err)
> -		return ERR_PTR(err);
> -
> -	mm = get_task_mm(task);
> -	if (mm && mm != current->mm &&
> -			!ptrace_may_access(task, PTRACE_MODE_READ)) {
> -		mmput(mm);
> -		mm = ERR_PTR(-EACCES);
> -	}
> -	mutex_unlock(&task->signal->cred_guard_mutex);
> -
> -	return mm;
> +	return get_check_task_mm(task, PTRACE_MODE_READ);
>  }
> ...
> +struct mm_struct *get_check_task_mm(struct task_struct *task, unsigned int mode)
> +{
> +	struct mm_struct *mm;
> +	int err;
> +
> +	err =  mutex_lock_killable(&task->signal->cred_guard_mutex);
> +	if (err)
> +		return ERR_PTR(err);
> +
> +	task_lock(task);
> +	if (__ptrace_may_access(task, mode)) {
> +		mm = ERR_PTR(-EACCES);
> +		goto out;
> +	}

Probably you should check "mm != current->mm" before __ptrace_may_access(),
otherwise this changes the rules for, say, /proc/pid/maps.

> @@ -298,23 +298,15 @@ static ssize_t process_vm_rw_core(pid_t pid, const struct iovec *lvec,
>  		goto free_proc_pages;
>  	}
>
> -	task_lock(task);
> -	if (__ptrace_may_access(task, PTRACE_MODE_ATTACH)) {
> -		task_unlock(task);
> -		rc = -EPERM;
> -		goto put_task_struct;
> -	}
> -	mm = task->mm;
> -
> -	if (!mm || (task->flags & PF_KTHREAD)) {
> -		task_unlock(task);
> -		rc = -EINVAL;
> +	mm = get_check_task_mm(task, PTRACE_MODE_ATTACH);
> +	if (!mm || IS_ERR(mm)) {
> +		if (!mm)
> +			rc = -EINVAL;
> +		else
> +			rc = -EPERM;

Cosmetic nit. I won't insist, but why -EPERM is better than -EACCES
returned by get_check_task_mm()? IOW, why not rc = PTR_ERR() ?

Note that get_check_task_mm() can return -EINTR, in this case -EPERM
looks confusing even if this doesn't really materr (the killed task
can't return to usermode).

Oleg.