Re: linux-next ppc64: RCU mods cause __might_sleep BUGs

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Apr 30, 2012 at 10:10:06PM -0700, Hugh Dickins wrote:
> On Tue, 1 May 2012, Benjamin Herrenschmidt wrote:
> > On Mon, 2012-04-30 at 15:37 -0700, Hugh Dickins wrote:
> > > 
> > > BUG: sleeping function called from invalid context at include/linux/pagemap.h:354
> > > in_atomic(): 0, irqs_disabled(): 0, pid: 6886, name: cc1
> > 
> > Hrm ... in_atomic and irqs_disabled are both 0 ... so yeah it smells
> > like a preempt count problem... odd.
> > 
> > Did you get a specific bisect target yet ?
> 
> Oh, I went as far as we need, I think, but I didn't bother quite to
> complete it because, once in that area, we know the schedule_tail()
> omission would muddy the waters: the tail of my bisect log was

Agreed, your bisect is close enough for our purposes.

> # bad: [e798cf3385d3aa7c84afa65677eb92e0c0876dfd] rcu: Add exports for per-CPU variables used for inlining
> git bisect bad e798cf3385d3aa7c84afa65677eb92e0c0876dfd
> # good: [90aec3b06194393c909e3e5a47b6ed99bb8caba5] rcu: Make exit_rcu() more precise and consolidate
> git bisect good 90aec3b06194393c909e3e5a47b6ed99bb8caba5
> 
> from which I concluded that the patch responsible is
> 
> commit ab8fc41a8545d40a4b58d745876c125af72a8a5c
> Author: Paul E. McKenney <paul.mckenney@linaro.org>
> Date:   Fri Apr 13 14:32:01 2012 -0700
> 
>     rcu: Move __rcu_read_lock() and __rcu_read_unlock() to per-CPU variables
>     
>     This commit is another step towards inlinable __rcu_read_lock() and
>     __rcu_read_unlock() functions for preemptible RCU.  This keeps these two
>     functions out of line, but switches them to use the per-CPU variables
>     that are required to export their definitions without requiring that
>     all RCU users include sched.h.  These per-CPU variables are saved and
>     restored at context-switch time.
> 
> > 
> > Cheers,
> > Ben.
> > 
> > > Call Trace:
> > > [c0000001a99f78e0] [c00000000000f34c] .show_stack+0x6c/0x16c (unreliable)
> > > [c0000001a99f7990] [c000000000077b40] .__might_sleep+0x11c/0x134
> > > [c0000001a99f7a10] [c0000000000c6228] .filemap_fault+0x1fc/0x494
> > > [c0000001a99f7af0] [c0000000000e7c9c] .__do_fault+0x120/0x684
> > > [c0000001a99f7c00] [c000000000025790] .do_page_fault+0x458/0x664
> > > [c0000001a99f7e30] [c000000000005868] handle_page_fault+0x10/0x30
> > > 
> > > I've plenty more examples, most of them from page faults or from kswapd;
> > > but I don't think there's any more useful information in them.
> > > 
> > > Anything I can try later on? 
> 
> I'd forgotten about CONFIG_PROVE_RCU (and hadn't been using PROVE_LOCKING
> on that machine), but following Paul's suggestion have now turned them on.
> 
> But not much light shed, I'm afraid.  Within minutes it showed a trace
> exactly like the one above, but the only thing PROVE_LOCKING and PROVE_RCU
> had to say was that we're holding mmap_sem at that point, which is no
> surprise and not a problem, just something lockdep is right to note.
> 
> That was an isolated occurrence, it continued quietly for maybe 20 minutes,
> then output lots to the console screen - but garbled in a way I've not
> seen before - the 0s came out just right (or perhaps all the hex digits
> were being shown as 0s), but most everything else was grayly unreadable.
> Then after a few minutes, spontaneously rebooted.
> 
> Perhaps I should remind myself of netdump; but getting the trace above
> without complaint from PROVE_RCU tells me that it is not helping.

My guess is that the following happened:

1.	Task A is running, with its state in RCU's per-CPU variables.

2.	Task A creates Task B and switches to it, but without invoking
	schedule_tail() or schedule().  Task B is now running, with
	Task A's state in RCU's per-CPU variables.

3.	Task B switches context, saving Task A's per-CPU RCU variables
	(with modifications by Task B, just for fun).

4.	Task A starts running again, and loads obsolete versions of its
	per-CPU RCU variables.  This can cause rcu_read_unlock_do_special()
	to be invoked at inappropriate times, which could cause
	pretty arbitrary misbehavior.

5.	Mismatched values for the RCU read-side nesting could cause
	the read-side critical section to complete prematurely, which
	could cause all manner of mischief.  However, I would expect
	this to trigger the WARN_ON_ONCE() in __rcu_read_unlock().

Hmmm...

							Thanx, Paul