Re: BUG: KASAN: slab-out-of-bounds in perf_callchain_user+0x494/0x530

[Peter Zijlstra]

On Tue, Dec 05, 2017 at 11:47:18PM +0900, Namhyung Kim wrote:
> Sure, I mean the following code:
> 
> 	mutex_lock(&callchain_mutex);
> 
> 	count = atomic_inc_return(&nr_callchain_events);
> 	if (WARN_ON_ONCE(count < 1)) {
> 		err = -EINVAL;
> 		goto exit;
> 	}
> 
> 	if (count > 1) {
> 		/* If the allocation failed, give up */
> 		if (!callchain_cpus_entries)
> 			err = -ENOMEM;
> 
> 		goto exit;
> 	}
> 
> 	err = alloc_callchain_buffers();
> exit:
> 	if (err)
> 		atomic_dec(&nr_callchain_events);
> 
> 	mutex_unlock(&callchain_mutex);
> 
> 
> The callchain_cpus_entries is allocated in alloc_callchain_buffers()
> only when the count is 1.  But if it failed to allocate, it decrease
> the count so next event would try to allocate it again.  Thus it seems
> not possible to see the callchain_cpus_entries being NULL in the
> 'if (count > 1)' block.  If you want to make next event give up, it'd
> need to take an additional count IMHO.

There's also a race against put_callchain_buffers() there, consider:


	get_callchain_buffers()		put_callchain_buffers()
	  mutex_lock();
	  inc()
					  dec_and_test() // false

	  dec() // 0


And the buffers leak.