Re: possible deadlock in lru_add_drain_all

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Oct 30, 2017 at 04:10:09PM +0100, Peter Zijlstra wrote:
> I can indeed confirm it's running old code; cpuhp_state is no more.
> 
> 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)
> 	  cpuhp_issue_call()
> 	    cpuhp_invoke_ap_callback()
> #3	      wait_for_completion()
> 
> 						msr_device_create()
> 						  ...
> #4						    filename_create()
> #3						complete()
> 


So all this you can get in a single callchain when you do something
shiny like:

	modprobe msr


> 	do_splice()
> #4	  file_start_write()
> 	  do_splice_from()
> 	    iter_file_splice_write()
> #5	      pipe_lock()
> 	      vfs_iter_write()
> 	        ...
> #6		  inode_lock()
> 
> 

This is a splice into a devtmpfs file


> 	sys_fcntl()
> 	  do_fcntl()
> 	    shmem_fcntl()
> #5	      inode_lock()

#6 (obviously)

> 	      shmem_wait_for_pins()
> 	        if (!scan)
> 		  lru_add_drain_all()
> #0		    cpus_read_lock()
> 

Is the right fcntl()


So 3 different callchains, and *splat*..