[PATCH] arm64: mte: Use PAGE_KERNEL_TAGGED in arch_add_memory

[Catalin Marinas]

Hi Patrick,

Thanks for this. Adding a few more people and linux-mm.

On Tue, Mar 02, 2021 at 08:21:03PM -0800, Patrick Daly wrote:
> In a system which supports MTE, the linear kernel region must allow
> reading/writing allocation tags. For memory present at boot this
> is already being done in map_mem(). Add the same in arch_add_memory().
> 
> Signed-off-by: Patrick Daly <pdaly at codeaurora.org>
> ---
>  arch/arm64/mm/mmu.c | 12 +++++++++++-
>  1 file changed, 11 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 9b25d60b..0fcfe90 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -1463,6 +1463,7 @@ int arch_add_memory(int nid, u64 start, u64 size,
>  		    struct mhp_params *params)
>  {
>  	int ret, flags = 0;
> +	pgprot_t pgprot;
>  
>  	if (!inside_linear_region(start, size)) {
>  		pr_err("[%llx %llx] is outside linear mapping region\n", start, start + size);
> @@ -1477,8 +1478,17 @@ int arch_add_memory(int nid, u64 start, u64 size,
>  	    IS_ENABLED(CONFIG_KFENCE))
>  		flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
>  
> +	/*
> +	 * The linear map must allow allocation tags reading/writing
> +	 * if MTE is present. Otherwise, it has the same attributes as
> +	 * PAGE_KERNEL.
> +	 */
> +	pgprot = params->pgprot;
> +	if (pgprot_val(pgprot) == pgprot_val(PAGE_KERNEL))
> +		pgprot = PAGE_KERNEL_TAGGED;
> +
>  	__create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
> -			     size, params->pgprot, __pgd_pgtable_alloc,
> +			     size, pgprot, __pgd_pgtable_alloc,
>  			     flags);

We'll need a similar pattern for vmalloc() once we have the khwasan
support in place. So we could add a pgprot_tagged() function (similar to
pgprot_writecombine() etc.) which does the above check and returns
PAGE_KERNEL_TAGGED, maybe only checking the PTE_ATTRINDX_MASK bits
rather than the whole prot bits.

However, the bigger problem is that arch_add_memory() is also called for
ZONE_DEVICE mappings and we can't always guarantee that such range
supports tagging (most likely it doesn't, e.g. persistent memory),
leading to potential external aborts.

One option is to expand mhp_params to pass additional information so
that the arch code can make the right decision. Another option is to
make PAGE_KERNEL_TAGGED global which is just PAGE_KERNEL for all the
other architectures and use it in the core code.

Yet another option which we haven't fully explored with MTE is to have
PAGE_KERNEL always tagged but add a new PAGE_KERNEL_DEVICE (or
_UNTAGGED) for specific cases like ZONE_DEVICE. We need to make sure
that PAGE_KERNEL doesn't end up in places where the backing memory does
not support tags.

I'll give the last option a quick try and see if it falls apart (just
changing PAGE_KERNEL to tagged). In terms of tag cache usage, it
probably won't have much of an impact since the whole of the linear map
is tagged already.

-- 
Catalin