Re: !PageLocked BUG_ON hit in clear_page_dirty_for_io

[Dave Jones <davej@codemonkey.org.uk>]

On Thu, Dec 10, 2015 at 04:30:24PM -0500, Chris Mason wrote:
 > On Thu, Dec 10, 2015 at 02:35:55PM -0500, Dave Jones wrote:
 > > On Thu, Dec 10, 2015 at 02:02:20PM -0500, Chris Mason wrote:
 > >  > On Tue, Dec 08, 2015 at 11:25:28PM -0500, Dave Jones wrote:
 > >  > > Not sure if I've already reported this one, but I've been seeing this
 > >  > > a lot this last couple days.
 > >  > > 
 > >  > > kernel BUG at mm/page-writeback.c:2654!
 > >  > > invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC KASAN
 > >  > > CPU: 1 PID: 2566 Comm: trinity-c1 Tainted: G        W       4.4.0-rc4-think+ #14
 > >  > > task: ffff880462811b80 ti: ffff8800cd808000 task.ti: ffff8800cd808000
 > >  > > RIP: 0010:[<ffffffff8826ff40>]  [<ffffffff8826ff40>] clear_page_dirty_for_io+0x180/0x1d0
 > >  > 
 > >  > Huh, are you able to reproduce at will?  From this code path it should
 > >  > mean somebody else is unlocking a page they don't own.
 > > 
 > > pretty easily yeah. I hit it maybe a couple dozen times yesterday.
 > > So if you've got some idea of printk's to spray anywhere I can give
 > > that a shot.
 > 
 > I'd rather try to trigger it here.  Going to have to add some way to
 > record which stack trace last unlocked and/or freed the page.

I'm using..

trinity -q -l off -C8 -a64 -x fsync -x fdatasync -x syncfs -x sync --enable-fds=testfile,pseudo

interestingly, if I just use 'testfile' by itself, I can't reproduce it.
(That means "create a bunch a few files in current dir and use their fds")
the "pseudo" bit means "also use fds from /proc, /sys and /dev".

strange.

(also, using trinity.git rather than the last version released, though
 I doubt it makes a difference in this case)

	Dave