Re: suspend race -mm regression [Was: Power: fix suspend vt regression]

[Jiri Slaby <jirislaby@gmail.com>]

CCs reduced.

On 09/04/2009 01:49 PM, Jiri Slaby wrote:
> On 08/31/2009 09:32 PM, Rafael J. Wysocki wrote:
>> On Monday 31 August 2009, Jiri Slaby wrote:
>>> On 08/11/2009 11:19 PM, Jiri Slaby wrote:
>>>> However there is still a race or something. Sometimes the suspend goes
>>>> through, sometimes it doesn't. I will investigate this further.
>>>
>>> Hmm, this took a loong time to track down a bit. Code instrumentation by
>>> outb(XX, 0x80) usually caused the issue to disappear.
>>>
>>> However I found out that it's caused by might_sleep() calls in
>>> flush_workqueue() and flush_cpu_workqueue(). I.e. it looks like there is
>>> a task which deadlocks/spins forever. If we won't reschedule to it,
>>> suspend proceeds.
>>>
>>> I replaced the latter might_sleep() by show_state() and removed
>>> refrigerated tasks afterwards. The thing is that I don't know if the
>>> prank task is there. I need a scheduler to store "next" task pid or
>>> whatever to see what it picked as "next" and so what will run due to
>>> might_sched(). I can then show it on port 80 display and read it when
>>> the hangup occurs.
>>>
>>> Depending on which might_sleep(), either flush_workqueue() never (well,
>>> at least in next 5 minutes) proceeds to for_each_cpu() or
>>> wait_for_completion() in flush_cpu_workqueue() never returns.
>>>
>>> It's a regression against some -rc1 based -next tree. Bisection
>>> impossible, suspend needs to be run even 7 times before it occurs. Maybe
>>> a s/might_sleep/yield/ could make it happen earlier (going to try)?
>>
>> If /sys/class/rtc/rtc0/wakealarm works on this box, you can use it to trigger
>> resume in a loop.
>>
>> Basically, you can do
>>
>> # echo 0 > /sys/class/rtc/rtc0/wakealarm
>> # date +%s -d "+60 seconds" > /sys/class/rtc/rtc0/wakealarm
>>
>> then go to suspend and it will resume the box in ~1 minute.
> 
> Thanks, in the end I found it manually. Goddammit! It's an -mm thing:
> cpu_hotplug-dont-affect-current-tasks-affinity.patch

BTW. when I reverted it, during suspend I got a warning:
SMP alternatives: switching to UP code
------------[ cut here ]------------
WARNING: at kernel/smp.c:124
__generic_smp_call_function_interrupt+0xfd/0x110()
Hardware name: To Be Filled By O.E.M.
Modules linked in: nfs lockd auth_rpcgss sunrpc ath5k ath
Pid: 3423, comm: pm-suspend Not tainted 2.6.31-rc8-mm1_64 #762
Call Trace:
 [<ffffffff8103fc48>] warn_slowpath_common+0x78/0xb0
 [<ffffffff8103fc8f>] warn_slowpath_null+0xf/0x20
 [<ffffffff8106950d>] __generic_smp_call_function_interrupt+0xfd/0x110
 [<ffffffff8106956a>] hotplug_cfd+0x4a/0xa0
 [<ffffffff81434e47>] notifier_call_chain+0x47/0x90
 [<ffffffff8105b311>] raw_notifier_call_chain+0x11/0x20
 [<ffffffff8141ece0>] _cpu_down+0x150/0x2d0
 [<ffffffff8104169b>] disable_nonboot_cpus+0xab/0x130
 [<ffffffff8106ee3d>] suspend_devices_and_enter+0xad/0x1a0
 [<ffffffff8106f00b>] enter_state+0xdb/0xf0
 [<ffffffff8106e741>] state_store+0x91/0x100
 [<ffffffff8116c157>] kobj_attr_store+0x17/0x20
 [<ffffffff8111c6a0>] sysfs_write_file+0xe0/0x160
 [<ffffffff810c3ce8>] vfs_write+0xb8/0x1b0
 [<ffffffff81434c35>] ? do_page_fault+0x185/0x350
 [<ffffffff810c434c>] sys_write+0x4c/0x80
 [<ffffffff8100be2b>] system_call_fastpath+0x16/0x1b
---[ end trace 73264e95657dec65 ]---
CPU1 is down

> Well, I don't know why, but when the kthread overthere runs under
> suspend conditions and gets rescheduled (e.g. by the might_sleep()
> inside) it never returns. pick_next_task always returns the idle task
> from the idle queue. State of the thread is TASK_RUNNING.
> 
> Why is it not enqueued into some queue? I tried also
> sched_setscheduler(current, FIFO, 99) in the thread itself. Unless I did
> it wrong, it seems like a global scheduler problem?
> 
> Ingo, any ideas?