Re: possible deadlock in lru_add_drain_all

[Peter Zijlstra]

On Tue, Oct 31, 2017 at 02:13:33PM +0100, Michal Hocko wrote:
> On Mon 30-10-17 16:10:09, Peter Zijlstra wrote:

> > However, that splat translates like:
> > 
> > 	__cpuhp_setup_state()
> > #0	  cpus_read_lock()
> > 	  __cpuhp_setup_state_cpuslocked()
> > #1	    mutex_lock(&cpuhp_state_mutex)
> > 
> > 
> > 
> > 	__cpuhp_state_add_instance()
> > #2	  mutex_lock(&cpuhp_state_mutex)
> 
> this should be #1 right?

Yes

> > 	  cpuhp_issue_call()
> > 	    cpuhp_invoke_ap_callback()
> > #3	      wait_for_completion()
> > 
> > 						msr_device_create()
> > 						  ...
> > #4						    filename_create()
> > #3						complete()
> > 
> > 
> > 
> > 	do_splice()
> > #4	  file_start_write()
> > 	  do_splice_from()
> > 	    iter_file_splice_write()
> > #5	      pipe_lock()
> > 	      vfs_iter_write()
> > 	        ...
> > #6		  inode_lock()
> > 
> > 
> > 
> > 	sys_fcntl()
> > 	  do_fcntl()
> > 	    shmem_fcntl()
> > #5	      inode_lock()

And that #6

> > 	      shmem_wait_for_pins()
> > 	        if (!scan)
> > 		  lru_add_drain_all()
> > #0		    cpus_read_lock()
> > 
> > 
> > 
> > Which is an actual real deadlock, there is no mixing of up and down.
> 
> thanks a lot, this made it more clear to me. It took a while to
> actually see 0 -> 1 -> 3 -> 4 -> 5 -> 0 cycle. I have only focused
> on lru_add_drain_all while it was holding the cpus lock.

Yeah, these things are a pain to read, which is why I always construct
something like the above first.