Query about ARM64 virt_to_phys and vice versa implementation

Query about ARM64 virt_to_phys and vice versa implementation

[Pratyush Anand]

Hi,

Would like to discuss virt_to_phys() and phys_to_virt() conversion in latest
kernel. I can see that if VA < PAGE_OFFSET, then we just subtract kimage_voffset
from VA and we get the PA. However, there is no such condition for
phys_to_virt(). phys_to_virt is always done by (PA - PHYS_OFFSET + PAGE_OFFSET).
So, how does phys_to_virt() work for all the cases? I must be missing something.

-- I am using a platform with VA bits = 42.
   Therefore PAGE_OFFSET on my platform is 0xfffffe0000000000
   PHYS_OFFSET is 0x8000000000
   kimage_voffset is 0xfffffb8006000000

Now lets find physical address of log_buf:
# cat /proc/kallsyms | grep -w "d log_buf"
fffffc0008ca6b40 d log_buf

Since VA < PAGE_OFFSET (fffffc0008ca6b40 < fffffe0000000000)
Therefore as per definition of virt_to_phys() from asm/memory.h:
PA = VA - kimage_voffset = fffffc0008ca6b40 - fffffb8006000000 = 8002CA6B40

Now convert PA=8002CA6B40 to VA.
VA = PA - PHYS_OFFSET + PAGE_OFFSET = 8002CA6B40 - 8000000000 + fffffe0000000000
= fffffe0002CA6B40 which not equal to fffffc0008ca6b40.

So, do we have dual mapping for all physical addresses? If not, how does it
work? What I am missing. 

If its dual mapping, then probably we will need to fix couple of things in
kexec-tools.

~Pratyush

Query about ARM64 virt_to_phys and vice versa implementation

[Catalin Marinas]

On Wed, Aug 10, 2016 at 01:58:08PM +0530, Pratyush Anand wrote:
> Would like to discuss virt_to_phys() and phys_to_virt() conversion in latest
> kernel. I can see that if VA < PAGE_OFFSET, then we just subtract kimage_voffset
> from VA and we get the PA. However, there is no such condition for
> phys_to_virt(). phys_to_virt is always done by (PA - PHYS_OFFSET + PAGE_OFFSET).
> So, how does phys_to_virt() work for all the cases? I must be missing something.

Because we recently relaxed where the kernel can be loaded in RAM, we
decoupled the actual RAM linear mapping from the kernel image virtual
address. The latter, when KASLR is disabled, is fixed but the
corresponding PA is not. So for various reasons, we need to map retrieve
the PA of a kernel image address/symbol, hence the virt_to_phys() needs
to take kimage_offset into account. The corresponding phys_to_virt()
would return an address in the linear mapping rather than the kernel
image VA but we didn't find a reason where we need phys_to_virt() to
return the latter.

> -- I am using a platform with VA bits = 42.
>    Therefore PAGE_OFFSET on my platform is 0xfffffe0000000000
>    PHYS_OFFSET is 0x8000000000
>    kimage_voffset is 0xfffffb8006000000
> 
> Now lets find physical address of log_buf:
> # cat /proc/kallsyms | grep -w "d log_buf"
> fffffc0008ca6b40 d log_buf
> 
> Since VA < PAGE_OFFSET (fffffc0008ca6b40 < fffffe0000000000)
> Therefore as per definition of virt_to_phys() from asm/memory.h:
> PA = VA - kimage_voffset = fffffc0008ca6b40 - fffffb8006000000 = 8002CA6B40
> 
> Now convert PA=8002CA6B40 to VA.
> VA = PA - PHYS_OFFSET + PAGE_OFFSET = 8002CA6B40 - 8000000000 + fffffe0000000000
> = fffffe0002CA6B40 which not equal to fffffc0008ca6b40.
> 
> So, do we have dual mapping for all physical addresses? If not, how does it
> work? What I am missing. 

Yes, we have dual mapping for the memory containing the kernel image. I
think the kernel memory layout print during boot should give you an idea
of how things look like from a VA perspective.

-- 
Catalin