Re: [BUG] Xen vm kernel crash in get_free_entries.

[David Vrabel <david.vrabel@citrix.com>]

On 21/10/13 11:29, Matt Wilson wrote:
> On Sat, Oct 19, 2013 at 12:03:17PM +0100, Ian Campbell wrote:
>> On Sat, 2013-10-19 at 14:51 +0400, Astarta wrote:
>>> On 10/19/2013 03:14 AM, Sander Eikelenboom wrote:
>>
>>>> makes a HVM guest (qemu-xen-traditional) with xen_platform_pci=0 boot again using xl, haven't tested it with xend.
>>>>
>>> Great catch!
>>> I also confirm that 3.11.5 kernel boots just fine after reverting of 
>>> 'correctly initialize grant table version 1' patch.
>>
>> This could just be down to that patch adding some BUG_ONs to catch bad
>> things going on, e.g. the one in gnttab_expand which I think is being
>> hit here.
> 
> Indeed, the BUG_ON was added to ensure that the grant table system is
> initialized before we attempt to expand the grant table space.
> 
>> I have a feeling that it is still wrong (but just more benign) to be
>> hitting that call chain in a configuration where there is no platform
>> device driver running. IOW reverting that patch removes the obvious
>> symptom (blowing up) but not the root cause, i.e. the patch is doing its
>> job.
> 
> The initialization of the grant table is deferred when running on a
> HVM guest.
> 
> drivers/xen/grant-table.c:
> 
> static int __gnttab_init(void)
> {
>         /* Delay grant-table initialization in the PV on HVM case */
>         if (xen_hvm_domain())
>                 return 0;
> 
>         if (!xen_pv_domain())
>                 return -ENODEV;
> 
>         return gnttab_init();
> }
> 
> The Xen platform PCI driver initializes it when it binds to the PCI
> device:
> 
> drivers/xen/platform-pci.c:
> 
> static int platform_pci_init(struct pci_dev *pdev,
>                              const struct pci_device_id *ent)
> {
> ...
>         max_nr_gframes = gnttab_max_grant_frames();
>         xen_hvm_resume_frames = alloc_xen_mmio(PAGE_SIZE *
> 	max_nr_gframes);
>         ret = gnttab_init();
>         if (ret)
>                 goto out;
>         xenbus_probe(NULL);
>         return 0;
> ...
> 
> Lots of initialization depends on the presence of the Xen platform PCI
> device, I don't see how PV enlightenment can be expected to work if
> you don't enable the PCI device.

Yes, I think we need to defer the initialization of xenbus to
platform_pci_init() as well (and perhaps some other bits and pieces as
well?).

David