[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH 2/2] Xen/vMCE: bugfix to remove problematic is_vmce_ready check

On 03.05.13 17:51, Liu, Jinsong wrote:
> Jan Beulich wrote:
>>>>> On 03.05.13 at 16:16, "Liu, Jinsong" <jinsong.liu@xxxxxxxxx> wrote:
>>> Jan Beulich wrote:
>>>>>>> On 03.05.13 at 10:41, "Liu, Jinsong" <jinsong.liu@xxxxxxxxx>
>>>>>>> wrote: 
>>>>> Jan Beulich wrote:
>>>>>>>>> On 27.04.13 at 10:38, "Liu, Jinsong" <jinsong.liu@xxxxxxxxx>
>>>>>>>>> wrote:
>>>>>>> From 9098666db640183f894b9aec09599dd32dddb7fa Mon Sep 17 00:00:00
>>>>>>> 2001 From: Liu Jinsong <jinsong.liu@xxxxxxxxx>
>>>>>>> Date: Sat, 27 Apr 2013 22:37:35 +0800
>>>>>>> Subject: [PATCH 2/2] Xen/vMCE: bugfix to remove problematic
>>>>>>> is_vmce_ready check 
>>>>>>> is_vmce_ready() is problematic:
>>>>>>> * For dom0, it checks if virq bind to dom0 mcelog driver. If not,
>>>>>>> it results dom0 crash. However, it's problematic and overkilled
>>>>>>> since mcelog as a dom0 feature could be enabled/disabled per dom0
>>>>>>> option: (XEN) MCE: This error page is ownded by DOM 0 (XEN) DOM0
>>>>>>> not ready for vMCE (XEN) domain_crash called from mcaction.c:133
>>>>>>> (XEN) Domain 0 reported crashed by domain 32767 on cpu#31:
>>>>>>> (XEN) Domain 0 crashed: rebooting machine in 5 seconds.
>>>>>>> (XEN) Resetting with ACPI MEMORY or I/O RESET_REG.
>>>>>>> * For dom0, if really need check, it should check whether vMCE
>>>>>>> injection for dom0 ready (say, exception trap bounce check, which
>>>>>>> has been done at inject_vmce()), not check dom0 mcelog ready
>>>>>>> (which has been done at mce_softirq() before send global virq to
>>>>>>> dom0). 
>>>>>> Following the argumentation above, I wonder which of the other
>>>>>> "goto vmce_failed" are really appropriate, i.e. whether the patch
>>>>>> shouldn't be extended (at least for the Dom0 case).
>>>>> You mean other 'goto vmce_failed' are also not appropriate (I'm not
>>>>> quite clear your point)?
>>>> Yes.
>>>>> Would you please point out which point you think not appropriate?
>>>> I question whether it is correct/necessary to crash the domain in
>>>> any of those failure cases. Perhaps when we fail to unmap the
>>>> page it is, but failure of fill_vmsr_data() and inject_vmce() don't
>>>> appear to be valid reasons once the is_vmce_ready() path is being
>>>> dropped.
>>> For fill_vmsr_data(), it failed only when MCG_STATUS_MCIP bit still
>>> set when next vMCE# occur, means the 2nd vMCE# occur when the 1st
>>> vMCE# not handled yet. Per SDM it should shutdown.
>>> For inject_vmce(), it failed when
>>> 1). vcpu is still mce_pending, or
>>> 2). pv not register trap callback
>>> Maybe it's some overkilled for dom0 (for other guest, it's ok to
>>> kill them), but any graceful way to quit?
>> Just exit and do nothing (except perhaps log a rate limited
>> message)?
> One concern of quiet exit is, the error will be totally ignored by guest --> 
> it didn't get preperly handled, and may recursively occur to make worse error 
> --> it's better to kill guest under such case.
>>> or, considering it rarely happens, how about keep current way (kill
>>> guest no matter dom0 or not)?
>> Possibly - I was merely asking why this one condition was found to
>> be too strict, while the others are being left as is.
>> Jan
> Ah, the reason of removing is_vmce_ready check is, it's
> problematic (check mcelog driver, not vmce tap callback),
> and overkilled (since defaultly dom0 will not start mcelog driver,
> under which case system will crash whenever vmce inject to dom0)
> --> So patch 2/2 is not too strict for dom0.

Please keep in mind the mcelog userland/kernel interface is not designed
with xen in mind. mcelog cannot report which guest is impacted for
example, although xen reports that to dom0.
I object 'fixing' the hypervisor to come over with mcelog drawbacks.
I prefer fixing Dom0 instead.

>From the design perspective, the virq for Dom0 is for logging purpose
only and the trap handler has equal purpose for both Dom0 and DomU.

BTW: I heard from Andre Pryzwara, he is working with Borislav Petkov
together on a complete new machine check interface in Linux
that supports ARM and x86. We should have an eye on this that
they do not repeat the mcelog design mistakes.


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.