Re: [GIT PULL rcu/urgent] Fix two more 4.3 regressions

From: Paul E. McKenney
Date: Sun Sep 27 2015 - 12:42:46 EST


On Sun, Sep 27, 2015 at 05:55:43PM +0200, Sedat Dilek wrote:
> On Sun, Sep 27, 2015 at 5:49 PM, Paul E. McKenney
> <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> > On Sun, Sep 27, 2015 at 09:37:05AM +0200, Sedat Dilek wrote:
> >> On Sun, Sep 27, 2015 at 9:32 AM, Paul E. McKenney
> >> <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> >> > On Sun, Sep 27, 2015 at 08:28:39AM +0200, Sedat Dilek wrote:

[ . . . ]

> Hi Paul :-),
>
> Here is the stack trace.
>
> [ 23.045871] BUG: sleeping function called from invalid context at
> kernel/workqueue.c:2678
> [ 23.045982] in_atomic(): 0, irqs_disabled(): 1, pid: 1399, name: acpid
> [ 23.046064] 3 locks held by acpid/1399:
> [ 23.046066] #0: (&evdev->mutex){+.+...}, at: [<ffffffff8174ac7c>]
> evdev_release+0xbc/0xf0
> [ 23.046081] #1: (&dev->mutex#2){+.+...}, at: [<ffffffff81742397>]
> input_close_device+0x27/0x70
> [ 23.046093] #2: (hid_open_mut){+.+...}, at: [<ffffffffa0056388>]
> usbhid_close+0x28/0xb0 [usbhid]
> [ 23.046106] irq event stamp: 3306
> [ 23.046109] hardirqs last enabled at (3305): [<ffffffff8192ae32>]
> _raw_spin_unlock_irq+0x32/0x60
> [ 23.046115] hardirqs last disabled at (3306): [<ffffffff81121017>]
> del_timer_sync+0x37/0x110
> [ 23.046122] softirqs last enabled at (2704): [<ffffffff818b12c9>]
> local_bh_enable+0x9/0x20
> [ 23.046128] softirqs last disabled at (2702): [<ffffffff818b12a9>]
> local_bh_disable+0x9/0x20
> [ 23.046136] CPU: 2 PID: 1399 Comm: acpid Not tainted
> 4.3.0-rc3-3-llvmlinux-amd64 #1
> [ 23.046139] Hardware name: SAMSUNG ELECTRONICS CO., LTD.
> 530U3BI/530U4BI/530U4BH/530U3BI/530U4BI/530U4BH, BIOS 13XK 03/28/2013
> [ 23.046143] ffff8800d36ee948 0000000000000092 0000000000000000
> ffff8800bbacfae8
> [ 23.046151] ffffffff8149adad ffff8800bbacfb18 ffffffff810cd1ea
> ffffffff81c56f0a
> [ 23.046158] ffff8800c22dc400 0000000000000000 0000000000000a76
> ffff8800bbacfb58
> [ 23.046165] Call Trace:
> [ 23.046172] [<ffffffff8149adad>] dump_stack+0x7d/0xa0
> [ 23.046177] [<ffffffff810cd1ea>] ___might_sleep+0x28a/0x2a0
> [ 23.046182] [<ffffffff810cceef>] __might_sleep+0x4f/0xc0
> [ 23.046187] [<ffffffff810afbff>] start_flush_work+0x2f/0x290
> [ 23.046192] [<ffffffff810afbac>] flush_work+0x5c/0x80
> [ 23.046195] [<ffffffff810afb6a>] ? flush_work+0x1a/0x80
> [ 23.046202] [<ffffffff810eed0d>] ? trace_hardirqs_off+0xd/0x10
> [ 23.046206] [<ffffffff810aecc8>] ? try_to_grab_pending+0x48/0x360
> [ 23.046211] [<ffffffff8192ac53>] ? _raw_spin_lock_irqsave+0x73/0x80
> [ 23.046216] [<ffffffff810afff9>] __cancel_work_timer+0x179/0x260
> [ 23.046221] [<ffffffff8192add2>] ? _raw_spin_unlock_irqrestore+0x52/0x80
> [ 23.046226] [<ffffffff81120fcd>] ? try_to_del_timer_sync+0xad/0xc0
> [ 23.046230] [<ffffffff810afe78>] cancel_work_sync+0x18/0x20
> [ 23.046237] [<ffffffffa00563d5>] usbhid_close+0x75/0xb0 [usbhid]
> [ 23.046245] [<ffffffffa00394d1>] hidinput_close+0x31/0x40 [hid]
> [ 23.046251] [<ffffffffa00394a0>] ? hidinput_open+0x40/0x40 [hid]
> [ 23.046256] [<ffffffff817423b8>] input_close_device+0x48/0x70
> [ 23.046261] [<ffffffff8174ac96>] evdev_release+0xd6/0xf0
> [ 23.046267] [<ffffffff812728c7>] __fput+0x107/0x240
> [ 23.046271] [<ffffffff81272756>] ____fput+0x16/0x20
> [ 23.046276] [<ffffffff810b945c>] task_work_run+0x6c/0xe0
> [ 23.046282] [<ffffffff81003aa7>] prepare_exit_to_usermode+0x117/0x120
> [ 23.046287] [<ffffffff81003ce1>] syscall_return_slowpath+0x231/0x2a0
> [ 23.046292] [<ffffffff8126efa5>] ? filp_close+0x65/0x90
> [ 23.046298] [<ffffffff810ef1c9>] ? trace_hardirqs_on_caller+0x19/0x290
> [ 23.046303] [<ffffffff81003017>] ? trace_hardirqs_on_thunk+0x17/0x19
> [ 23.046308] [<ffffffff8192bb62>] int_ret_from_sys_call+0x25/0x9f
>
> Can you give help on how to "debug" this?
>
> I switched from full-dynticks to simple cpu-accounting which did not help.
> But this was only a suspicion as Jiri pointed to the possibility
> del_timer_sync() could have get some mis-optimization.
>
> So, more empty head here.

I am not familiar with the "hardirqs last enabled" debug output, but
I am guessing that hardirqs are disabled because the "last disabled"
number is greater than the "last enabled" number. So if I understand
correctly, something somewhere up the call stack has irqs disabled,
but is calling a function that needs them enabled.

A quick look at usbhid_close() shows no problem, but then again, I am
looking at the v4.2 version rather than whatever you are testing.
But v4.3-rc2 looks the same.

Looks like you are already in contact with Jiri, which is good.

Another thing to try would be to insert might_sleep() calls, after
usbhid_close(), since it enables irqs explicitly, but before the call
to flush_work(). Jiri's comment can be interpreted as a suggestion
to insert a might_sleep() in try_to_del_timer_sync(), though
__cancel_work_timer() wouldn't be a bad place to try as well,
just before its call to flush_work().

Thanx, Paul

> - Sedat -
>
> >> $ grep LOCKDEP /boot/config-4.3.0-rc2-438.1-llvmlinux-amd64
> >> CONFIG_LOCKDEP_SUPPORT=y
> >> CONFIG_LOCKDEP=y
> >> # CONFIG_DEBUG_LOCKDEP is not set
> >>
> >> So, CLANG is a bit different compiler :-).
> >
> > True enough. Good to have more than one, though.
> >
> > Thanx, Paul
> >
> >> Thanks for taking care.
> >>
> >> Regards,
> >> - Sedat -
> >>
> >> >>
> >> >> [1] http://marc.info/?l=linux-kernel&m=144332758307978&w=2
> >> >> [2] http://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/patch/?id=66e8c57da6bf6b847a48a5a6fda59512f733ed78
> >> >> [3] http://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/patch/?id=19a5ecde086a6a5287978b12ae948fa691b197b7
> >> >>
> >> >
> >>
> >
>

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/