[PATCH] arm64: Fix for ARM64 3 level translation tables

[Pratyush Anand]

Hi Sameer,

Thanks for the patch.

On 02/12/2015:01:40:54 PM, Sameer Goel wrote:
> Add the implementation for PMD translation in case of 3 level page tables.
> Assign the kernel page size based on the page structure read from the kernel
> ELF file.
> 
> Signed-off-by: Sameer Goel <sgoel at codeaurora.org>
> ---
> This change targets the arm64_support branch of Pratush Anand's repository at:
> https://github.com/pratyushanand/makedumpfile.git
> 
>  arch/arm64.c   | 36 ++++++++++++++++++++++++++++++------
>  makedumpfile.c |  2 +-
>  2 files changed, 31 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/arm64.c b/arch/arm64.c
> index a94a4ba..ab20389 100644
> --- a/arch/arm64.c
> +++ b/arch/arm64.c
> @@ -36,7 +36,7 @@ typedef struct {
>  } pmd_t;
>  
>  #define pud_offset(pgd, vaddr) 	((pud_t *)pgd)
> -#define pmd_offset(pud, vaddr) 	((pmd_t *)pud)
> +
>  #define pgd_val(x)		((x).pgd)
>  #define pud_val(x)		(pgd_val((x).pgd))
>  #define pmd_val(x)		(pud_val((x).pud))
> @@ -68,6 +68,11 @@ typedef struct {
>  */
>  #define PHYS_MASK_SHIFT		48
>  #define PHYS_MASK		((1UL << PHYS_MASK_SHIFT) - 1)
> +/*
> + * Remove the highest order bits that are not a part of the
> + * physical address in a section
> + */
> +#define PMD_SECTION_MASK        ((1UL << 40) - 1)
>  
>  #define PMD_TYPE_MASK		3
>  #define PMD_TYPE_SECT		1
> @@ -83,10 +88,18 @@ typedef struct {
>  #define pmd_page_vaddr(pmd)		(__va(pmd_val(pmd) & PHYS_MASK & (int32_t)PAGE_MASK))
>  #define pte_offset(dir, vaddr) 		((pte_t*)pmd_page_vaddr((*dir)) + pte_index(vaddr))
>  
> +
> +#define pmd_offset_pgtbl_lvl_2(pud, vaddr) ((pmd_t *)pud)
> +
> +#define pmd_index(vaddr)		(((vaddr) >> PMD_SHIFT) & (PTRS_PER_PMD - 1))
> +#define pud_page_vaddr(pud)		(__va(pud_val(pud) & PHYS_MASK & (int32_t)PAGE_MASK))
> +#define pmd_offset_pgtbl_lvl_3(pud, vaddr) ((pmd_t *)pud_page_vaddr((*pud)) + pmd_index(vaddr))
> +
>  /* kernel struct page size can be kernel version dependent, currently
>   * keep it constant.
>   */
> -#define KERN_STRUCT_PAGE_SIZE		64
> +#define KERN_STRUCT_PAGE_SIZE		get_structure_size("page", DWARF_INFO_GET_STRUCT_SIZE)
> +
>  #define ALIGN(x, a) 			(((x) + (a) - 1) & ~((a) - 1))
>  #define PFN_DOWN(x)			((x) >> PAGE_SHIFT)
>  #define VMEMMAP_SIZE			ALIGN((1UL << (VA_BITS - PAGE_SHIFT)) * KERN_STRUCT_PAGE_SIZE, PUD_SIZE)
> @@ -97,6 +110,15 @@ static int pgtable_level;
>  static int va_bits;
>  static int page_shift;
>  
> +pmd_t *
> +pmd_offset(pud_t *pud, unsigned long vaddr)
> +{
> +	if (pgtable_level == 2) {
> +		return pmd_offset_pgtbl_lvl_2(pud, vaddr);
> +	} else {
> +		return pmd_offset_pgtbl_lvl_3(pud, vaddr);
> +	}
> +}
>  int
>  get_pgtable_level_arm64(void)
>  {
> @@ -256,7 +278,7 @@ vtop_arm64(unsigned long vaddr)
>  {
>  	unsigned long long paddr = NOT_PADDR;
>  	pgd_t	*pgda, pgdv;
> -	pud_t	*puda;
> +	pud_t	*puda, pudv;
>  	pmd_t	*pmda, pmdv;
>  	pte_t 	*ptea, ptev;
>  
> @@ -271,8 +293,9 @@ vtop_arm64(unsigned long vaddr)
>  		return NOT_PADDR;
>  	}
>  
> -	puda = pud_offset(pgda, vaddr);
> -	pmda = pmd_offset(puda, vaddr);
> +	pudv.pgd = pgdv;
> +
> +	pmda = pmd_offset(&pudv, vaddr);
>  	if (!readmem(VADDR, (unsigned long long)pmda, &pmdv, sizeof(pmdv))) {
>  		ERRMSG("Can't read pmd\n");
>  		return NOT_PADDR;
> @@ -298,7 +321,8 @@ vtop_arm64(unsigned long vaddr)
>  		break;
>  	case PMD_TYPE_SECT:
>  		/* 1GB section */
> -		paddr = (pmd_val(pmdv) & PMD_MASK) + (vaddr & (PMD_SIZE - 1));
> +		paddr = (pmd_val(pmdv) & (PMD_MASK & PMD_SECTION_MASK))
> +					+ (vaddr & (PMD_SIZE - 1));
>  		break;
>  	}
>  
> diff --git a/makedumpfile.c b/makedumpfile.c
> index cc71f20..df11f73 100644
> --- a/makedumpfile.c
> +++ b/makedumpfile.c
> @@ -5955,7 +5955,7 @@ create_dump_bitmap(void)
>  	ret = TRUE;
>  out:
>  	/* Should keep the buffer in the 1-cycle case. */
> -	if (info->flag_cyclic)
> +	if (info->flag_cyclic && (!info->flag_elf_dumpfile))

I think, this is a change which is not ARM64 specific, so should go to a
separate patch.
Moreover, probably you need:
if (info->flag_cyclic && !info->flag_elf_dumpfile && ret)

~Pratyush