Re: [PATCH v2] perf/core: Fix data race between perf_event_set_output and perf_mmap_close

[Yang Jihong]

Hello,

On 2022/7/6 20:29, Yang Jihong wrote:
> Hello,
>
> On 2022/7/5 21:07, Peter Zijlstra wrote:
> > On Mon, Jul 04, 2022 at 05:26:04PM +0200, Peter Zijlstra wrote:
> > > On Mon, Jul 04, 2022 at 08:00:06PM +0800, Yang Jihong wrote:
> > > > Data race exists between perf_event_set_output and perf_mmap_close.
> > > > The scenario is as follows:
> > > >
> > > >                    
> > > > CPU1                                                       CPU2
> > > >                                                                      
> > > > perf_mmap_close(event2)
> > > >                                                                        
> > > > if (atomic_dec_and_test(&event2->rb->mmap_count)  // mmap_count 1 -> 0
> > > >                                                                          
> > > > detach_rest = true;
> > > > ioctl(event1, PERF_EVENT_IOC_SET_OUTPUT, event2)
> > > >    perf_event_set_output(event1, event2)
> > > >                                                                        
> > > > if (!detach_rest)
> > > >                                                                          
> > > > goto out_put;
> > > >                                                                        
> > > > list_for_each_entry_rcu(event, &event2->rb->event_list, rb_entry)
> > > >                                                                          
> > > > ring_buffer_attach(event, NULL)
> > > >                                                                        
> > > > // because event1 has not been added to event2->rb->event_list,
> > > >                                                                        
> > > > // event1->rb is not set to NULL in these loops
> > > >
> > > >      ring_buffer_attach(event1, event2->rb)
> > > >        list_add_rcu(&event1->rb_entry, &event2->rb->event_list)
> > > >
> > > > The above data race causes a problem, that is, event1->rb is not 
> > > > NULL, but event1->rb->mmap_count is 0.
> > > > If the perf_mmap interface is invoked for the fd of event1, the 
> > > > kernel keeps in the perf_mmap infinite loop:
> > > >
> > > > again:
> > > >          mutex_lock(&event->mmap_mutex);
> > > >          if (event->rb) {
> > > > <SNIP>
> > > >                  if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
> > > >                          /*
> > > >                           * Raced against perf_mmap_close() through
> > > >                           * perf_event_set_output(). Try again, hope 
> > > > for better
> > > >                           * luck.
> > > >                           */
> > > >                          mutex_unlock(&event->mmap_mutex);
> > > >                          goto again;
> > > >                  }
> > > > <SNIP>
> > >
> > > Too tired, must look again tomorrow, little feeback below.
> >
> > With brain more awake I ended up with the below. Does that work?
I have verified that this patch can solve the problem.

Do I submit this patch? Or do you submit it?

Thanks,
Yang

> Yes, I apply the patch on kernel versions 5.10 and mainline,
> and it could fixed the problem.
>
> Tested-by: Yang Jihong <yangjihong1@xxxxxxxxxx>
>
> Thanks,
> Yang
> .