Re: [PATCH 2/2] trace_workqueue: add refcnt to struct cpu_workqueue_stats

[Li Zefan]

Frederic Weisbecker wrote:
> On Tue, Jul 07, 2009 at 02:07:35PM +0800, Li Zefan wrote:
>>> The stat entries can be freed when the stat file is being read.
>>> The worse is, the ptr can be freed immediately after it's returned
>>> from workqueue_stat_start/next().
>>>
>>> Add a refcnt to struct cpu_workqueue_stats to avoid use-after-free.
>>>
>>> Signed-off-by: Lai Jiangshan <laijs@xxxxxxxxxxxxxx>
>>> Signed-off-by: Li Zefan <lizf@xxxxxxxxxxxxxx>
>>> ---
>> ...
>>> @@ -175,11 +184,14 @@ static void *workqueue_stat_next(void *prev, int idx)
>>>  				return NULL;
>>>  		} while (!(ret = workqueue_stat_start_cpu(cpu)));
>>>  		return ret;
>>> +	} else {
>>> +		ret = list_entry(prev_cws->list.next,
>>> +				 struct cpu_workqueue_stats, list);
>> I just realized accessing prev_cws->list.next can be invalid!
>>
>> We can fix it by using list_del_init() to delete cws->list in
>> probe_workqueue_destruction(), but then if the race happened,
>> the next time stat_next() is called, NULL will be returned.
>> I guess this is Ok, since the race is rare.
> 
> 
> If you ensure the kref_get/put are under the
> workqueue_cpu_stat(cpu)->lock, it should be fine, right?
> 

Unfortunately no.

It's safe to dereference prev_cws, but not safe to retreive
prevw_cws->list.next.

Suppose: head->n1->n2

T1                      T2
---------------         -------------------
stat_start()
  -> return n1
                        list_del(n1)
                          -> n1->list->next = LIST_POISON1;
stat_next()
  -> prev = n1
    -> list_entry(prev->list.next) !!!

You see why it's not safe..

>  
>> (I never like the design of trace_stat..Fortunately we'll
>> probably switch to perfcounter for this kind of statistics
>> reporting)
> 
> 
> I don't like its design either. I wrote it specifically for
> the branch tracer and didn't think about free-able events :-/
> 

Yeah, for free-able events it's buggy to use trace_stat.
Similar bug exists in ksym_tracer.

Another way to fix it is not use trace_stat but use seq_file
directly. They don't need to be sorted anyway.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/