Re: [Xen-devel] Xen dom0 crash: "d0:v0: unhandled page fault (ec=0000)"

[Jeremy Fitzhardinge <jeremy@xxxxxxxx>]

 On 10/29/2010 12:15 PM, Konrad Rzeszutek Wilk wrote:
> On Fri, Oct 29, 2010 at 04:44:23PM +0100, Gianni Tedesco wrote:
>> On Wed, 2010-10-20 at 09:54 +0100, Gianni Tedesco wrote:
>>> On Wed, 2010-10-20 at 00:31 +0100, Andreas Kinzler wrote:
>>>> On 19.10.2010 17:45, Gianni Tedesco wrote:
>>>>> ditto, I suspected a known bug in my gcc version which broke xchg
>>>>> because when I compiled with -O2 instead of -Os... the problem went away
>>>>> but then something else bad happened later (I forget the details, and it
>>>>> was too many config tweaks ago to get back to last time I had it working
>>>>> that well)
>>>> Jeremy, one user earlier reported that he found out that for him there 
>>>> seems to be a relation between kernel size and crash status. He just 
>>>> added/removed some options that could never influence the "crash status" 
>>>> (like adding/removing netfilter modules). With all the experiences here, 
>>>> is may be useful to check for code paths related to kernel size.
>>>>
>>>> Regards Andreas
>> I have dmesg output from 2.6.32.18-ge6b9b2c and the current broken
>> version.
>>
>> http://pastebin.com/3m0DpDdW - 2.6.32.24-gd0054d6-dirty - broken
> Gianni pointed out to me that he spotted this:
>
> [    0.000000] last_pfn = 0x2d0699 max_arch_pfn = 0x400000000
> [    0.000000] x86 PAT enabled: cpu 0, old 0x50100070406, new 0x7010600070106
> [    0.000000] last_pfn = 0x2f000 max_arch_pfn = 0x400000000
>
> I am not sure why "last_pfn" is being printed twice, but it could be
> Gianni test-patch.
>
> It looks as if the initial E820 is created with a max_pfn of
> 0x2d0699, which rougly translates to 8G of memory instead of
> the 752MB.
>
> There were a bunch of changes in arch/x86/xen/setup.c and mmu.c
> code that figures out the max_pfn. Actually, there is one
> (git commit 6c8e75f5e712e596ab138597e65aac426ff03382):
>
>  HYPERVISOR_shared_info->arch.max_pfn = xen_max_p2m_pfn

That sets the extent that the toolstack will look at the P2M for
migration; it has no direct effect on the domain itself.

> Which would set the this to the highest PFN. But that number
> should not have been used by the E820 calculation which uses
> nr_pages entry to clamp the E820. Oh wait, it does not - it actually
> still parses the E820, but marks the area above the nr_pages
> as "XEN EXTRA" (git commit 8d0d6d6d275d4514780ba3d350e57d48e3b5b5e1)
> so they should not figure in the last_pfn calculation and instead
> lay unused. But the 'initial memory mapping' ignores that and
> still tries to setup mapping on _all_ E820_RAM regions, even
> if they are reserved from by the early memory allocator. This would
> imply that the page table is being actually put right in the
> area that is reserved by the early memory allocator.
>
> Hmm, so Gianni, I think if you shortcircuited the setup.c code
> to not parse the E820_RAM regions above the nr_pages that might
> do it. And also try to figure out who or what resets the last_pfn.
>
> Or in the code that sets the 'XEN EXTRA', make it set that region
> of pages as E820_RESERVED and see what happens then.

The way is this is supposed to work is:

   1. Xen gives the domain N pages
   2. There's an E820 which describes M pages (M > N)
   3. The kernel traverses the existing E820 and finds holes and adds
      the memory to a new E820_RAM region beyond M
   4. Set up P2M for pages up to N
   5. When the kernel maps all "RAM", the region from N-M is not
      present, and has no valid P2M mapping; in that case, xen_make_pte
      will return a non-present pte.

The important part of making XEN EXTRA E820_RAM is that the kernel will
allocate page structures for them, even if the pages are absent.  Making
it RESERVED will suppress that and make the exercise pointless.

    J

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