Re: [Xen-devel] [PATCH 06/11] x86/xen: Add i386 kexec/kdump implementation

[Daniel Kiper <daniel.kiper@xxxxxxxxxx>]

On Mon, Oct 01, 2012 at 02:55:01PM +0100, Jan Beulich wrote:
> >>> On 01.10.12 at 14:52, Daniel Kiper <daniel.kiper@xxxxxxxxxx> wrote:
> > On Fri, Sep 28, 2012 at 09:11:47AM +0100, Jan Beulich wrote:
> >> Finally, as noticed in an earlier patch already, you appear to
> >> re-introduce stuff long dropped from the kernel - the forward
> >> ported kernels get away with just setting PA_CONTROL_PAGE,
> >> PA_PGD, and PA_SWAP_PAGE in the page list. Since the number
> >> and purpose of the pages is established entirely by the guest
> >> kernel, all you need to obey is that the hypervisor expects
> >> alternating PA_/VA_ pairs (where the VA_ ones can be left
> >> unpopulated). Perhaps taking a look at a recent SLES kernel
> >> would help...
> >
> > I have got 
> > ftp://ftp.suse.com/pub/projects/kernel/kotd/SLE11-SP2/src/kernel-source-3.0.43-6.1.src.rpm.
> > Does kexec/kdump work in your environment? In my it does not.
>
> While I never ran it myself, I know kdump has been working on
> SLE for quite a long while (leaving aside hardware or firmware
> quirks requiring extra workarounds).

It should work on baremetal without any issue. However,
I am almost sure that it does not work on Xen dom0
in any way. But maybe I missed something.

> > At least there is wrong assumption that
> > vaddr = (unsigned long)relocate_kernel
> > gets virtual address of relocate_kernel in Xen
>
> Where did you spot that? Afaics the only thing done with
> relocate_kernel is that it gets copied into the control page.

arch/x86/kernel/machine_kexec_64.c:270
There is also similar thing in i386 code.

> > (I have tested only x86_64 implementation but
> > as I saw i386 has similar problem). In real it is
> > fix mapped in hypervisor which is completely
> > different than address calculated in dom0 kernel.
> > Virtual address of control page (and others) is
> > only known by hypervisor kexec/kdump functions.
> > It means that transition page table could be
> > established by relocate_kernel code only.
> > If you would like to do optimistation as you
> > mentioned above you must reintroduce code
> > for page table establishment into generic
> > relocate_kernel_??.S. However, another
> > problem arises. New generic code utilizes
> > additional arguments such as swap page
> > (and potentially could use others in the future).
> > As I saw it is not possible to pass extra addresses
> > through page_list[] in struct xen_kexec_image
> > because its has insufficient size (I mean
> > x86_64 because i386 is a bit different story).
>
> No - there's no meaning assigned by Xen to any of the slots,
> except for said assumption about them representing alternating

I know that.

> PA_/VA_ pairs. Hence, as long as old entries get removed, new
> slots can easily be added (and the number of slots currently
> needed is far lower than what was used originally [2.6.18]).

As I said in other email I will try to optimize page table code.
We will see what could be done.

Daniel

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel