On Thu, Jan 28, 2016 at 01:47:00PM +0100, Thierry Reding wrote:
> On Thu, Jan 28, 2016 at 11:12:10AM +0100, Peter Zijlstra wrote:
> > On Tue, Jan 26, 2016 at 06:38:43PM +0100, Thierry Reding wrote:
> > > > Task or work item involved in memory reclaim trying to flush a
> > > > non-WQ_MEM_RECLAIM workqueue or one of its work items can lead to
> > > > deadlock.  Trigger WARN_ONCE() if such conditions are detected.
> > > I've started noticing the following during boot on some of the devices I
> > > work with:
> > > 
> > > [    4.723705] WARNING: CPU: 0 PID: 6 at kernel/workqueue.c:2361 check_flush_dependency+0x138/0x144()
> > > [    4.736818] workqueue: WQ_MEM_RECLAIM deferwq:deferred_probe_work_func is flushing !WQ_MEM_RECLAIM events:lru_add_drain_per_cpu
> > > [    4.748099] Modules linked in:
> > > [    4.751342] CPU: 0 PID: 6 Comm: kworker/u8:0 Not tainted 4.5.0-rc1-00018-g420fc292d9c7 #1
> > > [    4.759504] Hardware name: NVIDIA Tegra SoC (Flattened Device Tree)
> > > [    4.765762] Workqueue: deferwq deferred_probe_work_func
> > > [    4.771004] [<c0017acc>] (unwind_backtrace) from [<c0013134>] (show_stack+0x10/0x14)
> > > [    4.778746] [<c0013134>] (show_stack) from [<c0245f18>] (dump_stack+0x94/0xd4)
> > > [    4.785966] [<c0245f18>] (dump_stack) from [<c0026f9c>] (warn_slowpath_common+0x80/0xb0)
> > > [    4.794048] [<c0026f9c>] (warn_slowpath_common) from [<c0026ffc>] (warn_slowpath_fmt+0x30/0x40)
> > > [    4.802736] [<c0026ffc>] (warn_slowpath_fmt) from [<c00390b8>] (check_flush_dependency+0x138/0x144)
> > > [    4.811769] [<c00390b8>] (check_flush_dependency) from [<c0039ca0>] (flush_work+0x50/0x15c)
> > > [    4.820112] [<c0039ca0>] (flush_work) from [<c00c51b0>] (lru_add_drain_all+0x130/0x180)
> > > [    4.828110] [<c00c51b0>] (lru_add_drain_all) from [<c00f728c>] (migrate_prep+0x8/0x10)
> > 
> > Right, also, I think it makes sense to do lru_add_drain_all() from a
> > WQ_MEM_RECLAIM workqueue, it is, after all, aiding in getting memory
> > freed.
> > 
> > Does something like the below cure things?
> > 
> > TJ does this make sense to you?
> > 
> > ---
> >  mm/swap.c | 11 ++++++++++-
> >  1 file changed, 10 insertions(+), 1 deletion(-)
> > 
> > diff --git a/mm/swap.c b/mm/swap.c
> > index 09fe5e97714a..a3de016b2a9d 100644
> > --- a/mm/swap.c
> > +++ b/mm/swap.c
> > @@ -666,6 +666,15 @@ static void lru_add_drain_per_cpu(struct work_struct *dummy)
> >  
> >  static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
> >  
> > +static struct workqueue_struct *lru_wq;
> > +
> > +static int __init lru_init(void)
> > +{
> > +	lru_wq = create_workqueue("lru");
> > +	return 0;
> > +}
> > +early_initcall(lru_init);
> > +
> >  void lru_add_drain_all(void)
> >  {
> >  	static DEFINE_MUTEX(lock);
> > @@ -685,7 +694,7 @@ void lru_add_drain_all(void)
> >  		    pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
> >  		    need_activate_page_drain(cpu)) {
> >  			INIT_WORK(work, lru_add_drain_per_cpu);
> > -			schedule_work_on(cpu, work);
> > +			queue_work_on(cpu, &lru_wq, work);
>                                            ^
> 
> This ampersand is too much here and causes a compile-time warning.
> Removing it and booting the resulting kernel doesn't trigger the
> WQ_MEM_RECLAIM warning anymore, though.
> 
> Tested on top of next-20160128.

This implies that if you want to turn this into a proper patch:

Tested-by: Thierry Reding <treding@nvidia.com>

Alternatively, if you come up with a different way to fix things,
please let me know and I'll be happy to test again.

Thierry