Date:	Fri, 13 Jan 2012 00:26:06 -0800
From:	ebiederm@...ssion.com (Eric W. Biederman)
To:	Venki Pallipadi <venki@...gle.com>
Cc:	Andi Kleen <andi@...stfloor.org>, Ingo Molnar <mingo@...e.hu>,
	"H. Peter Anvin" <hpa@...or.com>,
	Thomas Gleixner <tglx@...utronix.de>,
	Len Brown <lenb@...nel.org>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Aaron Durbin <adurbin@...gle.com>, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] Kexec reboot and stale/inconsistent system state

Venki Pallipadi <venki@...gle.com> writes:

> On Thu, Jan 12, 2012 at 2:42 PM, Eric W. Biederman
> <ebiederm@...ssion.com> wrote:
>> Venkatesh Pallipadi <venki@...gle.com> writes:
>>
>>> Of course, kexec reboots has a lot of advantantages. There is one
>>> blind spot though, in the way we handle shutdown of the old kernel in kexec.
>>> The problem: there are drivers in the kernel that change
>>> system state either during init or during normal operation. During a
>>> hardware reboot these system states would get reset to default.
>>> But, with kexec reboot this system state can stay persistent with the
>>> new kernel and can lead to some inconsistencies.
>>>
>>> More concrete x86 example:
>>> * Kernel_v0 has a driver which sets a particular MSR.
>>> * Kernel_v1 has a different version of the driver which does not touch the
>>> MSR and assumes it to be at BIOS init state.
>>> * Now we have a kexec reboot from Kernel_v0 to Kernel_v1 and the MSR will
>>> stay set leading to potential inconsistent state.
>>>
>>> One such driver I was looking at is drivers/idle/intel_idle.c, which sets
>>> the MSR during init and forgets about it. There are other drivers which can
>>> potentially have similar problems - like MTRR, etc. There can be other
>>> state in the system like chipset state and device state etc, though
>>> I expect most of it won't be a problem based on the way drivers init.
>>>
>>> Seemingly simple solution is to do the needed cleanup in driver's module_exit
>>> code path. That is not enough as module_exit for all builtin drivers are
>>> all discarded.
>>>
>>> What if we retain all the module_exits in CONFIG_KEXEC case and do a
>>> cleanup before launching a new kernel. This should work, but most of the
>>> module_exit code assumes that they will be only called when corresponding
>>> init is successful. I couldn't find a clean and easy way of associating
>>> module's init function with its exit function and to save the result of
>>> init.
>>>
>>> Alternate quick solution is to create a new module interface to register
>>> kexec cleanup function as in the patch below. But, this is not generic enough
>>> yet and does not handle the case where we have done the cleanup and kexec
>>> fails after that and we have to do some setup again to run with old
>>> kernel.
>>>
>>> Before spending more time on this, just wanted to run this RFC along to see:
>>> * How big/serious this problem really is?
>>> * Has this been looked at before and any other potential solution for
>>> this?
>>
>> Use the drivers shutdown method.  The shutdown method is called
>> at reboot and at normal kexec.  If you care about the kexec on panic
>> case you need to write a paranoid driver initialization routine.
>>
>> Hmm.  Looking a little closer the intel idle driver is a weird thing.
>> I am not seeing any operations structures.  There are kobjects so
>> the intel idle driver might be properly hooked into the device tree.
>> If the intel idle driver isn't properly hooked into the device
>> tree you should be able to hook into the reboot notifier instead
>> of implementing a device shutdown method.  Although for a lot
>> of reasons implementing the standard device shutdown method is
>> preferred.
>>
>
> Yes. Doing in device shutdown is much cleaner. I need to figure out
> how to hook that into intel_idle.
>
> One other problem case I see is in mtrr. 'cat <something> >
> /proc/mtrr' gets carried over after a kexec reboot.
> So, we need some change to restore the boot time state there.

Assuming userspace changes /proc/mtrr and userspace calls reboot
that instance looks like a userspace problem.

Any time I have changed /proc/mtrr the change has happened because the
BIOS got it wrong and I needed to fix those values.

The other usecase for changning /proc/mtrr is to make video
frame-buffers write-combining, and with PAT support in the kernel
using /proc/mtrr for that should now be obsolete.  

So in practice if not in theory I think we should be good.

> Do you guys see any other place where we will need similar special
> handling for kexec reboots?

The rule is every driver needs to put the device back in a state where
the drivers init/probe routine can get the device out of that state.
Frequently people when the care chose to fix kdump.

Perhaps I am dense but I don't get the impression that you have
uncovered any new kinds of issues.  My impression is that you have
found another case where kexec isn't on the top of the list of the
driver developer and so a bug is found on the kexec path.

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

Powered by blists - more mailing lists