[RFC PATCH 1/2] arm64: mm: Make virt_addr_valid to check for pfn_valid again

[Catalin Marinas]

Adding Mike and Anshuman,

On Wed, Jul 28, 2021 at 07:44:15PM +0300, Oleksandr Tyshchenko wrote:
> From: Oleksandr Tyshchenko <oleksandr_tyshchenko at epam.com>
> 
> The problem is that Arm's implementation of virt_addr_valid()
> leads to memblock_is_map_memory() check, which will fail for
> ZONE_DEVICE based addresses. But, the pfn_valid() check in turn
> is able to cope with ZONE_DEVICE based memory.
> 
> You can find a good explanation of that problem at:
> https://lore.kernel.org/lkml/1614921898-4099-2-git-send-email-anshuman.khandual@arm.com
> 
> Signed-off-by: Oleksandr Tyshchenko <oleksandr_tyshchenko at epam.com>
> ---
> I am not quite sure whether it is a "correct" place and
> the change itself, I just partially restored a behaviour before:
> https://lore.kernel.org/lkml/20210511100550.28178-4-rppt@kernel.org
> So, the target of this patch is to get a feedback how to resolve
> this properly if, of course, this really needs to be resolved
> (I might miss important bits here).
> 
> It is worth mentioning that patch doesn't fix the current code base
> (if I am not mistaken, no one calls virt_addr_valid() on Arm64 for
> ZONE_DEVICE based addresses at the moment, so it seems that nothing
> is broken), the fix is intended for the subsequent patch in this
> series that will try to enable Xen's "unpopulated-alloc" usage
> on Arm (it was enabled on x86 so far).
> Please see:
> [RFC PATCH 2/2] xen/unpopulated-alloc: Query hypervisor to provide
> unallocated space
> 
> The subsequent patch will enable the code where virt_addr_valid()
> is used in drivers/xen/unpopulated-alloc.c:fill_list() to check that
> a virtual address returned by memremap_pages() is valid.

I wonder what the point of calling virt_addr_valid() in fill_list() is?
If memremap_pages() succeeded, the pages were mapped at the returned
vaddr, there's no need for an additional virt_addr_valid() check.

> ---
>  arch/arm64/include/asm/memory.h | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/include/asm/memory.h b/arch/arm64/include/asm/memory.h
> index 824a365..1a35a44 100644
> --- a/arch/arm64/include/asm/memory.h
> +++ b/arch/arm64/include/asm/memory.h
> @@ -351,7 +351,7 @@ static inline void *phys_to_virt(phys_addr_t x)
>  
>  #define virt_addr_valid(addr)	({					\
>  	__typeof__(addr) __addr = __tag_reset(addr);			\
> -	__is_lm_address(__addr) && pfn_is_map_memory(virt_to_pfn(__addr));	\
> +	__is_lm_address(__addr) && pfn_valid(virt_to_pfn(__addr));	\
>  })

pfn_valid() only guarantees the presence of a struct page but not
necessarily that the virtual address is accessible (valid). So this
change would break the NOMAP ranges case.

-- 
Catalin