Re: [RFC PATCH] fs: Use a seperate wq for do_sync_work() to avoid a potential deadlock

[Dave Chinner <david@fromorbit.com>]

On Wed, Sep 17, 2014 at 09:46:35PM +0100, Aaron Tomlin wrote:
> On Wed, Sep 17, 2014 at 08:22:02PM +0200, Oleg Nesterov wrote:
> > On 09/17, Aaron Tomlin wrote:
> > >
> > > Since do_sync_work() is a deferred function it can block indefinitely by
> > > design. At present do_sync_work() is added to the global system_wq.
> > > As such a deadlock is theoretically possible between sys_unmount() and
> > > sync_filesystems():
> > >
> > >   * The current work fn on the system_wq (do_sync_work()) is blocked
> > >     waiting to aquire a sb's s_umount for reading.
> > >
> > >   * The "umount" task is the current owner of the s_umount in
> > >     question but is waiting for do_sync_work() to continue.
> > >     Thus we hit a deadlock situation.
> > >
> > I can't comment the patches in this area, but I am just curious...
> > 
> > Could you explain this deadlock in more details? I simply can't understand
> > what "waiting for do_sync_work()" actually means.
> 
> Hopefully this helps:
> 
> 	           "umount"                                      "events/1"
> 
> sys_umount					    sysrq_handle_sync
>   deactivate_super(sb)				      emergency_sync
>   {						    	schedule_work(work)
>     ...						    	  queue_work(system_wq, work)
>     down_write(&s->s_umount)			    	    do_sync_work(work)
>     ...						      	      sync_filesystems(0)
>     kill_block_super(s)				    		...
>       generic_shutdown_super(sb)		    		down_read(&sb->s_umount)
>       // sop->put_super(sb)
>       ext4_put_super(sb)
> 	invalidate_bdev(sb->s_bdev)
> 	  lru_add_drain_all()
> 	    for_each_online_cpu(cpu) {
> 	      schedule_work_on(cpu, work)
> 		queue_work_on(cpu, system_wq, work)
> 		...
> 	    }
>   }
> 
>   - Both lru_add_drain and do_sync_work work items are added to
>     the same global system_wq
> 
>   - The current work fn on the system_wq is do_sync_work and is
>     blocked waiting to aquire an sb's s_umount for reading
> 
>   - The umount task is the current owner of the s_umount in
>     question but is waiting for do_sync_work to continue.
>     Thus we hit a deadlock situation.

What kernel did you see this deadlock on?

I don't see a deadlock here on a mainline kernel. The emergency sync
work blocks, the new work gets queued, and the workqueue
infrastructure simply pulls another kworker thread from the pool and
runs the new work. IOWs, I can't see how this would deadlock unless
the system_wq kworker pool has been fully depleted it's defined
per-cpu concurrency depth. If the kworker thread pool is depleted
then you have bigger problems than emergency sync not
deadlocking....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com