Re: EIP is at device_shutdown+0x32/0x60

[Mark Lord <lkml@rtr.ca>]

Greg KH wrote:
> On Thu, Nov 15, 2007 at 01:29:04PM -0500, Mark Lord wrote:
>> Greg KH wrote:
>>> On Thu, Nov 15, 2007 at 12:07:48PM -0500, Mark Lord wrote:
>> ..
>>>> Greg, I don't know if this is relevant or not,
>>>> but x86 has bugs in the halt/reboot code for SMP.
>>>>
>>>> Specifically, in native_smp_send_stop() the code now uses
>>>> spin_trylock() to "lock" the shared call buffers,
>>>> but then ignores the result.
>>>>
>>>> This means that multiple CPUs can/will clobber each other
>>>> in that code.
>>>>
>>>> The second bug, is that this code does not wait for the
>>>> target CPUs to actually stop before it continues.
>>>>
>>>> This was the real cause of the failure-to-poweroff problems
>>>> I was having with 2.6.23, which we fixed by using CPU hotplug
>>>> to disable_nonboot_cpus() before the above code ever got run.
>>> I have noticed that the shutdown path is quite weird, shutting down
>>> sysdev devices differently depending on the type of shutdown, which is
>>> probably not good.
>>> But what change are you talking about for the poweroff problem?  I have
>>> a _lot_ of people reporting that 2.6.22 is not powering off for them and
>>> I can't seem to figure it out.  Do you have a changeset for something
>>> that went in to fix this issue?
>> ..
>>
>> Well, the real bugs that cause the problem are described by me above.
>> I don't have a fix for those, but the workaround is in 2.6.23
>> under git 4047727e5ae33f9b8d2b7766d1994ea6e5ec2991 Fix SMP poweroff hangs.
>>
>> With that workaround, there's no more hanging on halt,
>> though there could still be a hang on reboot.
>>
>> A problem with that workaround is that it has no effect unless
>> the CPU hotplug code is configured (CONFIG_PM_SLEEP_SMP and pals).
> 
> Hm, that's not going to be a fix for 2.6.22 as we already do a call to
> that function in that call when shutting down, something in the
> 2.6.23-rc series must have changed that logic to make it required again.
..

But remember, it has no effect unless the kernel has suspend/resume
configured in.  Many desktops don't.

And the real unfixed bugs are still there in native_smp_send_stop()
as described above.  You could prod whoever takes care of that code
to fix them.

Cheers