Re: Soft lockups during reading /proc/PID/smaps

[Aleksei Besogonov <alex.besogonov@gmail.com>]

On 02 Aug 2014, at 11:19, Oleg Nesterov <oleg@redhat.com> wrote:
> On 07/31, Aleksei Besogonov wrote:
>> 
>> On 31 Jul 2014, at 00:43, David Rientjes <rientjes@google.com> wrote:
>> 
>>> The while_each_thread() in vm_is_stack() looks suspicious since the task
>>> isn't current and rcu won't protect the iteration, and we also don't hold
>>> sighand lock or a readlock on tasklist_lock.
>>> I think Oleg will know how to proceed, cc'd.
>> I’m attaching a minimal test case that can reproduce the issue. Works in 100% cases on any system I’ve tried.
> Thanks. I think David is right and we need the simple patch below.
> This reminds me I should kill while_each_thread :/
> Any chance you can test it? If not, I will do this later and send
> the patch if it helps.
Thanks, it works on the rc7 kernel.

I can make a backported version for earlier kernels if nobody else is interested.

> 
> Oleg.
> 
> --- x/mm/util.c
> +++ x/mm/util.c
> @@ -277,17 +277,14 @@ pid_t vm_is_stack(struct task_struct *ta
> 
> 	if (in_group) {
> 		struct task_struct *t;
> -		rcu_read_lock();
> -		if (!pid_alive(task))
> -			goto done;
> 
> -		t = task;
> -		do {
> +		rcu_read_lock();
> +		for_each_thread(task, t) {
> 			if (vm_is_stack_for_task(t, vma)) {
> 				ret = t->pid;
> 				goto done;
> 			}
> -		} while_each_thread(task, t);
> +		}
> done:
> 		rcu_read_unlock();
> 	}
>