Re: [PATCHv2 1/3] x86, ptdump: Add section for EFI runtime services
From: Mathias Krause
Date: Wed Oct 08 2014 - 17:58:29 EST
On 8 October 2014 17:17, Borislav Petkov <bp@xxxxxxxxx> wrote:
> On Tue, Oct 07, 2014 at 07:07:48PM +0200, Mathias Krause wrote:
>> What you can see here are actually the EFI runtime service mappings, not
>> the ESP fix area. Check the addresses and compare them. You should find
>> similarities ;) And, in fact, the EFI mappings are incomplete in the
>> second dump, i.e. the vanilla kernel one, because of the enforced limit
>> for the ESP fix area.
>> So, in your examples are actually *no* ESP fix area mappings as those
>> would be r/o. In fact, I think, the above dumps are the result of a
>> CONFIG_EFI_PGT_DUMP enabled kernel that dumps the page table after
>> setting up the EFI mappings. There are no ESP fix mappings in this dump
>> because those are only set up after the EFI runtime service mappings.
> Ok, I think I know what the deal is:
> So, the ptdump we do to dmesg very early at boot is the EFI pagetable which
> shouldn't have espfix mappings...
>> See the following code in init/main:
>> #ifdef CONFIG_X86
>> if (efi_enabled(EFI_RUNTIME_SERVICES))
>> #ifdef CONFIG_X86_ESPFIX64
>> /* Should be run before the first non-init thread is created */
> ... exactly because of this: we're setting up the EFI mappings in
> the EFI page table before we do the espfix mappings in the *kernel*
> pagetable which is a separate one.
Well, that is only partly correct. The call chain in efi_map_regions()
[ -> efi_map_region() -> __map_region() -> kernel_map_pages_in_pgd()
-> ..."magic"... ] does not only map the EFI regions in
trampoline_pgd, but also in kernel page table, i.e. init_level4_pgt.
That can easily be shown by looking at the kernel_page_tables debugfs
file on a running system. You'll notice large RWX portions covering
the "phys" mappings in the "Low Kernel Mapping" area and the "virt"
mappings in the "EFI Runtime Services" area. Now reboot with "noefi"
and see those be gone.
> So, if we have to be really correct about it, the first dump to dmesg
> which comes down the efi_enter_virtual_mode() path shouldn't contain the
> espfix area at all.
Correct -- because init_espfix_bsp() hadn't been called by then. Nice,
we agree on this, at least ;)
> Later dumps from debugfs cannot select the EFI pagetable so they should
> not be dumping the EFI runtime services.
Well, beside the debugfs file is always using init_level4_pgt, reality
shows the EFI mappings are visible there, too. So why omit them?
> I don't have a good idea about how to do that right now though, maybe
> the address markers should have flags or so...
Well, maybe I got it all wrong and there should be no EFI mappings in
the kernel page table at all? If so, how about fixing
kernel_map_pages_in_pgd() to not do so? It's you're code after all...
> Sent from a fat crate under my desk. Formatting is fine.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/