RE: [Xen-devel] [PATCH 0/2] MCA support with page offlining

["Wang, Shane" <shane.wang@xxxxxxxxx>]

For page offlining, It may not be a good way to put the page into an array but 
a list.

+       pg->count_info |= PGC_reserved;
+       page_offlining[num_page_offlining++] = pg;

I guess the code targets for offlining domain pages only, right? How about free 
pages and xen pages?
If so, no need to check in the following when allocating free pages, since the 
offlined pages will not be freed into heap()()().
If not, the following may have a bug.

+                if ( !list_empty(&heap(node, zone, j)) ) {
+                    pg = list_entry(heap(node, zone, j).next, struct 
page_info, list);
+                    if (!(pg->count_info & PGC_reserved))
+                        goto found;
+                    else
+                        printk(XENLOG_DEBUG "Page %p(%lx) is not to be 
allocated.\n",
+                               pg, page_to_maddr(pg));
+                }

If one free page (not pg) within pg and pg+(1U<<j) is offlined, the range 
pg~pg+(1U<<j) has the risk to be allocated with that page.

Shane

Jiang, Yunhong wrote:
> xen-devel-bounces@xxxxxxxxxxxxxxxxxxx <> wrote:
>> Hi all,
>> 
>> I had posted about MCA support for Intel64 before. It had only a
>> function to log the MCA error data received from hypervisor.
>> 
>> http://lists.xensource.com/archives/html/xen-devel/2008-09/msg0
>> 0876.html 
>> 
>> I attach patches that support not only error logging but also Page
>> Offlining function. The page where an MCA occurs will offline and not
>> reuse. A new flag 'PGC_reserved' was added in page count_info to
>> mark the impacted page. 
>> 
>> I know that it is better to implement the page offlining for general
>> purpose, but I implemented for MCA specialized in this first step.
> 
> Maybe the MCA page offline is a bit different to normal page offline
> requirement, so take it as first step maybe a good choice :) 
> 
> As for your current page_offlining, I'm not sure why the PGC_reserved
> page should not be freed? Also, for following code, will that make
> the heap(node, zone, j) can't be allocated anymore? Maybe we can
> creat a special list to hold all those pages and remove them from the
> heap list?    
> 
> +                if ( !list_empty(&heap(node, zone, j)) ) {
> +                    pg = list_entry(heap(node, zone, j).next, struct
> page_info, list); +                    if (!(pg->count_info &
> PGC_reserved)) +                        goto found;
> +                    else
> +                        printk(XENLOG_DEBUG "Page %p(%lx) is not to
> be allocated.\n", +                               pg,
> page_to_maddr(pg)); +
> 
> 
>> 
>> And I also implement the MCA handler of Dom0 which support to
>> shutdown the remote domain where a MCA occurred. If the MCA occurred
>> on a DomU, Dom0 notifies it to the DomU. When the notify is failed,
>> Dom0 calls SCHEDOP_remote_shutdown hypercall.
>> 
>> [1/2] xen part: mca-support-with-page-offlining-xen.patch
> 
> We are not sure we really need pass all #MC information to dom0
> firstly, and let dom0 to notify domU. Xen should knows about
> everything, so it may have knowledge to decide inject virtual #MC to
> guest or not. Of course, this does not impact your patch.   
> 
>> [2/2] linux/x86_64 part: mca-support-with-page-offlining-linux.patch
> 
> As for how to inject virtual #MC to guest (including dom0), I think
> we need consider following point: 
> 
> a) Benefit from reusing guest #MC handler's . #MC handler is well
> known difficult to test, and the native guest handler may have been
> tested more widely. Also #MC handler improves as time going-on, reuse
> guest's MCA handler share us those improvement.   
> b) Maintain the PV handler to different OS version may not so easy,
> especially as hardware improves, and kernel may have better support
> for error handling/containment.  
> c) #MC handler may need some model specific information to decide the
> action, while guest (not dom0) has virtualized CPUID information. 
> d) Guest's MCA handler may requires the physical information when the
> #MC hapen, like the CPU number the #MC happens. 
> e) For HVM domain, PV handler will be difficult (considering Windows
> guest). 
> 
> And we have several option to support virtual #MC to guest:
> 
> Option 1 is what currently implemented. A PV #MC handler is
> implemented in guest. This PV handler gets MCA information from Xen
> HV through hypercall, including MCA MSR value, also some additional
> information, like which physical CPU the MCA happened. Option 1 will
> help us on issue d), but we need main a PV handler, and can't get
> benifit from native handler. Also it does not resolve issue c) quite
> well.      
> 
> option 2, Xen will provide MCA MSR virtualization so that guest's
> native #MC handler can run without changes. It can benifit from guest
> #MC handler, but it will be difficult to get model specific
> information, and has no physical information.   
> 
> Option 3 uses a PV #MC handler for guest as option 1, but interface
> between Xen/guest is abstract event, like offline offending page,
> terminate current execution context etc. This should be straight
> forward for Linux, but may be difficult to Windows and other OS.   
> 
> Currently we are considering option 2 to provide MCA MSR
> virtualization to guest, and dom0 can also benifit from such support
> (if guest has different CPUID as native, we will either keep guest
> running, or kill guest based on error code). Of course, current
> mechanism of passing MCA information from xen to dom0 will still be
> useful, but that will be used for logging purpose or for Correcatable
> Error. How do you think about this?      
> 
> Thanks
> Yunhong Jiang
> 
>> Signed-off-by: Kazuhiro Suzuki <kaz@xxxxxxxxxxxxxx>
>> 
>> Thanks,
>> KAZ
>> 
>> _______________________________________________
>> Xen-devel mailing list
>> Xen-devel@xxxxxxxxxxxxxxxxxxx
>> http://lists.xensource.com/xen-devel
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@xxxxxxxxxxxxxxxxxxx
> http://lists.xensource.com/xen-devel


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel