[PATCHv4 4/4] arm64: fix PAGE_OFFSET calc for flipped mm

[Philipp Rudo <prudo@redhat.com>]

Hi Pingfan,

On Tue, 18 Jan 2022 15:48:12 +0800
Pingfan Liu <piliu@redhat.com> wrote:

> From: Kairui Song <kasong@tencent.com>
> 
> Since kernel commit 14c127c957c1 ('arm64: mm: Flip kernel VA space'),
> the memory layout on arm64 have changed, and kexec-tools can no longer
> get the the right PAGE_OFFSET based on _text symbol.
> 
> Prior to that, the kimage (_text) lays above PAGE_END with this layout:
> 0               -> VA_START                 : Usespace
> VA_START        -> VA_START + 256M          : BPF JIT, Modules
> VA_START + 256M -> PAGE_OFFSET - (~GB misc) : Vmalloc (KERNEL _text HERE)
> PAGE_OFFSET     -> ...                      : * Linear map *
> 
> And here we have:
> VA_START    = -1UL << VA_BITS
> PAGE_OFFSET = -1UL << (VA_BITS - 1)
> _text < -1UL << (VA_BITS - 1)
> 
> Kernel image lays somewhere between VA_START and PAGE_OFFSET, so we just
> calc VA_BITS by getting the highest unset bit of _text symbol address,
> and shift one less bit of VA_BITS to get page offset. This works as long
> as KASLR don't put kernel in a too high location (which is commented inline).
> 
> And after that commit, kernel layout have changed:
> 0               -> PAGE_OFFSET              : Userspace
> PAGE_OFFSET     -> PAGE_END                 : * Linear map *
> PAGE_END        -> PAGE_END + 128M          : bpf jit region
> PAGE_END + 128M -> PAGE_END + 256MB         : modules
> PAGE_END + 256M -> ...                      : vmalloc (KERNEL _text HERE)
> 
> Here we have:
> PAGE_OFFSET = -1UL << VA_BITS
> PAGE_END    = -1UL << (VA_BITS - 1)
> _text > -1UL << (VA_BITS - 1)
> 
> Kernel image now lays above PAGE_END, so we have to shift one more bit to
> get the VA_BITS, and shift the exact VA_BITS for PAGE_OFFSET.
> 
> We can simply check if "_text > -1UL << (VA_BITS - 1)" is true to judge
> which layout is being used and shift the page offset occordingly.
> 
> Signed-off-by: Kairui Song <kasong@tencent.com>
> (rebased and stripped by Pingfan )
> Signed-off-by: Pingfan Liu <piliu@redhat.com>
> Cc: Simon Horman <horms@verge.net.au>
> Cc: Philipp Rudo <prudo@redhat.com>
> To: kexec at lists.infradead.org
> ---
>  kexec/arch/arm64/kexec-arm64.c | 14 +++++++++++++-
>  1 file changed, 13 insertions(+), 1 deletion(-)
> 
> diff --git a/kexec/arch/arm64/kexec-arm64.c b/kexec/arch/arm64/kexec-arm64.c
> index 793799b..ce7a5bb 100644
> --- a/kexec/arch/arm64/kexec-arm64.c
> +++ b/kexec/arch/arm64/kexec-arm64.c
> @@ -923,13 +923,25 @@ out:
>  
>  int get_page_offset(unsigned long *page_offset)
>  {
> +	unsigned long long text_sym_addr, kernel_va_mid;
>  	int ret;
>  
> +	text_sym_addr = get_kernel_sym("_text");
> +	if (text_sym_addr == 0) {
> +		fprintf(stderr, "Can't get the symbol of _text to calculate page_offset.\n");
> +		return -1;
> +	}
> +
>  	ret = get_va_bits();
>  	if (ret < 0)
>  		return ret;
>  
> -	if (va_bits < 52)
> +	/* Since kernel 5.4, kernel image is put above
> +	 * UINT64_MAX << (va_bits - 1)
> +	 */
> +	kernel_va_mid = UINT64_MAX << (va_bits - 1);
> +	/* older kernel */
> +	if (text_sym_addr < kernel_va_mid)
>  		*page_offset = UINT64_MAX << (va_bits - 1);
>  	else
>  		*page_offset = UINT64_MAX << va_bits;

I would drop the kernel_va_mid and simply use

	*page_offset = UINT64_MAX << (va_bits - 1)
	if (*page_offset > text_sym_addr > *page_offset)
		*page_offset = UINT64_MAX << va_bits

but that's more a matter of taste.

Reviewed-by: Philipp Rudo <prudo@redhat.com>