On 12/14/2010 10:01 AM, Takao Indoh wrote:
> On Tue, 14 Dec 2010 12:43:58 +0900, Kenji Kaneshige wrote:
>
>> Hi,
>>
>> I tested this patch on the system that has large amount of memory (1TB),
>> and I encountered the immediate system reset problem that happens every
>> time I modify the EFI boot entry using efibootmgr command. It seems that
>> triple fault happens due to the incorrect page table setup.
>>
>>> +void __init efi_pagetable_init(void)
>>> +{
>> (snip.)
>>> + pgd = efi_pgd + pgd_index(PAGE_OFFSET);
>>> + set_pgd(pgd, *pgd_offset_k(PAGE_OFFSET));
>>> + pgd = efi_pgd + pgd_index(__START_KERNEL_map);
>>> + set_pgd(pgd, *pgd_offset_k(__START_KERNEL_map));
>>> +}
>>
>> Maybe we need to map whole kernel address space. The problem doesn't
>> happen by modifying as follows.
>>
>> clone_pgd_range(efi_pgd + KERNEL_PGD_BOUNDARY,
>> swapper_pg_dir + KERNEL_PGD_BOUNDARY,
>> KERNEL_PGD_PTRS);
>
> Agree, I'll merge this to the patch. And I got another bug report. This
> should be also merged.
>
>
> @@ -177,8 +177,8 @@ void __init efi_pagetable_init(void)
> memset(efi_pgd, 0, sizeof(efi_pgd));
> for (p = memmap.map; p < memmap.map_end; p += memmap.desc_size) {
> md = p;
> - if (!(md->type & EFI_RUNTIME_SERVICES_CODE) &&
> - !(md->type & EFI_RUNTIME_SERVICES_DATA))
> + if ((md->type != EFI_RUNTIME_SERVICES_CODE) &&
> + (md->type != EFI_RUNTIME_SERVICES_DATA))
> continue;
>
> start_pfn = md->phys_addr >> PAGE_SHIFT;
> @@ -191,7 +191,7 @@ void __init efi_pagetable_init(void)
> pud = fill_pud(pgd, vaddr);
> pmd = fill_pmd(pud, vaddr);
> pte = fill_pte(pmd, vaddr);
> - if (md->type & EFI_RUNTIME_SERVICES_CODE)
> + if (md->type == EFI_RUNTIME_SERVICES_CODE)
> set_pte(pte, pfn_pte(pfn, PAGE_KERNEL_EXEC));
> else
> set_pte(pte, pfn_pte(pfn, PAGE_KERNEL));
>
Note: the current kernel has initial_page_table around for 32 bits; we
should do the equivalent for 64 bits and always keep a 1:1 page table
set around instead of having a special page table. As such, I would
prefer to not merge a patchset that adds yet another set of ad hoc page
tables if at all possible.
-hpa
