Re: [PATCHSET] mempool, percpu, blkcg: fix percpu stat allocation and remove stats_lock

[Vivek Goyal <vgoyal@redhat.com>]

On Mon, Feb 27, 2012 at 09:25:29AM -0500, Vivek Goyal wrote:
> On Sun, Feb 26, 2012 at 07:21:13AM +0900, Tejun Heo wrote:
> > On Sun, Feb 26, 2012 at 06:46:41AM +0900, Tejun Heo wrote:
> > > Hello,
> > > 
> > > On Fri, Feb 24, 2012 at 10:44:32PM -0500, Vivek Goyal wrote:
> > > > Booting with blkcg-stacking branch and changing io scheduler from cfq to
> > > > deadline oopsed.
> > > > 
> > > > login: [   67.382768] general protection fault: 0000 [#1] SMP DEBUG_PAGEALLOC
> > > > [   67.383037] CPU 1 
> > > > [   67.383037] Modules linked in: floppy [last unloaded: scsi_wait_scan]
> > > > [   67.383037] 
> > > > [   67.383037] Pid: 4763, comm: bash Not tainted 3.3.0-rc3-tejun-misc+ #6 Hewlett-Packard HP xw6600 Workstation/0A9Ch
> > > > [   67.383037] RIP: 0010:[<ffffffff81311793>]  [<ffffffff81311793>] cfq_put_queue+0xb3/0x1d0
> > > 
> > > Hmmm... weird.  Looking into it.  I'm away from office for a week and
> > > will probably be slow.
> > 
> > It won't reproduce here.  Can you please explain how to trigger it?
> > Can you please also run addr2line on the oops address?
> 
> I have BLK_CGROUP enabled. CFQ is deafult scheduler. I boot the system and
> just change the scheduler to deadline on sda and crash happens. It is
> consistently reproducible on my machine.
> 
> add2line points to, blk-cgroup.h
> 
> blkg_put() {
> 	WARN_ON_ONCE(blkg->refcnt <= 0);
> }
> 
> I put more printk and we are putting down async queues when crash happens.
> 
> cfq_put_async_queues().
> 
> So looks like a group might have already been freed. May be it is a group
> refcount issue. I see 6b6b6b... pattern in RBX. Sounds like a use after
> free thing.

I think problem might be that we have destroyed policy data (cfqg also)
early and later we access it.

So we call following.

elevator_switch()
  blkg_destroy_all()
    update_root_blkg();

Here update_root_blkg() will free up the blkg->pd for cfq. 

And later we call.

elevator_exit()
  cfq_exit_queue()
    cfq_put_async_queues()
      cfq_put_queue()
        blkg_put(cfqg_to_blkg(cfqg)); <------ trying to reach blkg through
already freed policy data.

So we should not free up root group policy data till old elevator has
exited.

Thanks
Vivek

> 
> Thanks
> Vivek