[PATCH 1/2] kexec: add phys_offset field

[Dave Anderson]

----- kexec-request at lists.infradead.org wrote:

> 
> Message: 1
> Date: Mon, 26 Apr 2010 12:03:30 +0300
> From: Mika Westerberg <ext-mika.1.westerberg at nokia.com>
> To: kexec at lists.infradead.org
> Cc: linux-arm-kernel at lists.infradead.org
> Subject: [PATCH 1/2] kexec: add phys_offset field
> Message-ID:
> 
> <e376969c3d29696ce11488af59605ada56fdccff.1272271391.git.ext-mika.1.westerberg at nokia.com>
> 	
> 
> Add new field 'phys_offset' to struct elf_info. This field is used to calculate
> virtual address of PT_LOAD segment on architectures where physical memory
> doesn't always start at address 0 (namely ARM).
> 
> Signed-off-by: Mika Westerberg <ext-mika.1.westerberg at nokia.com>
> ---
>  kexec/crashdump-elf.c |    9 ++++++++-
>  kexec/crashdump.h     |    1 +
>  2 files changed, 9 insertions(+), 1 deletions(-)
> 
> diff --git a/kexec/crashdump-elf.c b/kexec/crashdump-elf.c
> index 6bcef8d..1291174 100644
> --- a/kexec/crashdump-elf.c
> +++ b/kexec/crashdump-elf.c
> @@ -236,7 +236,14 @@ int FUNC(struct kexec_info *info,
>  		 * memory region.
>  		 */
>  		phdr->p_paddr = mstart;
> -		phdr->p_vaddr = mstart + elf_info->page_offset;
> +		/*
> +		 * In some architectures (namely ARM), physical memory doesn't
> +		 * always start at 0 so we need to calculate virtual address
> +		 * based on elf_info->phys_offset like:
> +		 *     virt = phys + page_offset - phys_offset
> +		 */
> +		phdr->p_vaddr = mstart + elf_info->page_offset -
> +				elf_info->phys_offset;
>  		phdr->p_filesz	= phdr->p_memsz	= mend - mstart + 1;
>  		/* Do we need any alignment of segments? */
>  		phdr->p_align	= 0;
> diff --git a/kexec/crashdump.h b/kexec/crashdump.h
> index 30d6f29..8c0ccda 100644
> --- a/kexec/crashdump.h
> +++ b/kexec/crashdump.h
> @@ -27,6 +27,7 @@ struct crash_elf_info {
>  	unsigned long backup_src_end;
>  
>  	unsigned long long page_offset;
> +	unsigned long long phys_offset;
>  	unsigned long lowmem_limit;
>  
>  	int (*get_note_info)(int cpu, uint64_t *addr, uint64_t *len);
> -- 

Can you can possibly leave the p_vaddr field as-is on the other
non-ARM architectures?

The p_vaddr field in the ELF header has always been presumed
to be a kernel unity-mapped virtual address.  As I understand
it, this will create a "hybrid" virtual address that is no longer
aligned with the kernel's concept of a unity-mapped address.

The crash utility for other architectures calculates the phys_offset,
or has the phys_base passed in the makedumpfile header.

Dave Anderson