[RFC] arm64: properly define SOFT_DIRTY for arm64

[Anshuman Khandual]

Hi Nico,

On 7/3/23 19:25, Nico Pache wrote:
> ARM64 has a soft-dirty bit (software dirty) but never properly defines
> CONFIG_ARCH_HAS_SOFT_DIRTY or its necessary functions. This patch
> introduces the ability to set/clear the soft dirty bit in a similar
> manner as the other arches that utilize it.
> 
> However, we must be careful... there are cases where the DBM bit is not
> available and the software dirty bit plays a essential role in determining
> whether or not a page is dirty. In these cases we must not allow the
> user to clear the software dirty bit. We can test for these cases by
> utilizing the arch_faults_on_old_pte() function which test the availability
> of DBM.

The current soft-dirty bit is a SW PTE bit i.e PTE_DIRTY, tracking PTE
dirtiness in absence of HW DBM support, although both these tracking
methods are very much intertwined. Current pte helpers like pte_dirty(),
pte_mkdirty(), pte_mkclean(), and pte_wrrotect() etc operate both on
HW and SW dirty tracking bits irrespective of whether DBM is supported
or not.

For soft dirty to work, we need a software PTE bit which sticks around
on the PTE entry until user space (only) clears it and above PTE_DIRTY
bit cannot be used for that purpose as it could be cleared in the kernel.

static inline pte_t pte_mkclean(pte_t pte)
{
        pte = clear_pte_bit(pte, __pgprot(PTE_DIRTY));
        pte = set_pte_bit(pte, __pgprot(PTE_RDONLY));

        return pte;
}

BTW arch_faults_on_old_pte() is no longer available mainline.

- Anshuman

> 
> Cc: Andrew Morton <akpm at linux-foundation.org>
> Cc: Anshuman Khandual <anshuman.khandual at arm.com>
> Cc: Catalin Marinas <catalin.marinas at arm.com>
> Cc: David Hildenbrand <david at redhat.com>
> Cc: Gerald Schaefer <gerald.schaefer at linux.ibm.com>
> Cc: Liu Shixin <liushixin2 at huawei.com>
> Cc: Will Deacon <will at kernel.org>
> Cc: Yu Zhao <yuzhao at google.com>
> Signed-off-by: Nico Pache <npache at redhat.com>
> ---
>  arch/arm64/Kconfig               |  1 +
>  arch/arm64/include/asm/pgtable.h | 77 +++++++++++++++++++++++++++++++-
>  2 files changed, 76 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> index 891ab530a665..4de491627f49 100644
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -173,6 +173,7 @@ config ARM64
>  	select HAVE_ARCH_PREL32_RELOCATIONS
>  	select HAVE_ARCH_RANDOMIZE_KSTACK_OFFSET
>  	select HAVE_ARCH_SECCOMP_FILTER
> +	select HAVE_ARCH_SOFT_DIRTY
>  	select HAVE_ARCH_STACKLEAK
>  	select HAVE_ARCH_THREAD_STRUCT_WHITELIST
>  	select HAVE_ARCH_TRACEHOOK
> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
> index 0bd18de9fd97..a0a15ffa2417 100644
> --- a/arch/arm64/include/asm/pgtable.h
> +++ b/arch/arm64/include/asm/pgtable.h
> @@ -121,8 +121,9 @@ static inline pteval_t __phys_to_pte_val(phys_addr_t phys)
>  })
>  
>  #define pte_hw_dirty(pte)	(pte_write(pte) && !(pte_val(pte) & PTE_RDONLY))
> -#define pte_sw_dirty(pte)	(!!(pte_val(pte) & PTE_DIRTY))
> -#define pte_dirty(pte)		(pte_sw_dirty(pte) || pte_hw_dirty(pte))
> +#define pte_soft_dirty(pte)	(!!(pte_val(pte) & PTE_DIRTY))
> +#define pte_dirty(pte)		(pte_soft_dirty(pte) || pte_hw_dirty(pte))
> +#define pte_swp_soft_dirty(pte)	pte_soft_dirty(pte)
>  
>  #define pte_valid(pte)		(!!(pte_val(pte) & PTE_VALID))
>  /*
> @@ -1096,6 +1097,78 @@ static inline bool pud_sect_supported(void)
>  	return PAGE_SIZE == SZ_4K;
>  }
>  
> +#ifdef CONFIG_ARM64_HW_AFDBM
> +/*
> + * if we have the DBM bit we can utilize the software dirty bit as
> + * a mechanism to introduce the soft_dirty functionality; however, without
> + * it this bit is crucial to determining if a entry is dirty and we cannot
> + * clear it via software. DBM can also be disabled or broken on some early
> + * armv8 devices, so check its availability before modifying it.
> + */
> +static inline pte_t pte_clear_soft_dirty(pte_t pte)
> +{
> +	if (arch_faults_on_old_pte())
> +		return pte;
> +
> +	return clear_pte_bit(pte, __pgprot(PTE_DIRTY));
> +}
> +
> +static inline pte_t pte_mksoft_dirty(pte_t pte)
> +{
> +	if (arch_faults_on_old_pte())
> +		return pte;
> +
> +	return set_pte_bit(pte, __pgprot(PTE_DIRTY));
> +}
> +
> +static inline pte_t pte_swp_clear_soft_dirty(pte_t pte)
> +{
> +	if (arch_faults_on_old_pte())
> +		return pte;
> +
> +	return clear_pte_bit(pte, __pgprot(PTE_DIRTY));
> +}
> +
> +static inline pte_t pte_swp_mksoft_dirty(pte_t pte)
> +{
> +	if (arch_faults_on_old_pte())
> +		return pte;
> +
> +	return set_pte_bit(pte, __pgprot(PTE_DIRTY));
> +}
> +
> +static inline int pmd_soft_dirty(pmd_t pmd)
> +{
> +	return pte_soft_dirty(pmd_pte(pmd));
> +}
> +
> +static inline pmd_t pmd_clear_soft_dirty(pmd_t pmd)
> +{
> +	return pte_pmd(pte_clear_soft_dirty(pmd_pte(pmd)));
> +}
> +
> +static inline pmd_t pmd_mksoft_dirty(pmd_t pmd)
> +{
> +	return pte_pmd(pte_mksoft_dirty(pmd_pte(pmd)));
> +}
> +
> +#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
> +static inline int pmd_swp_soft_dirty(pmd_t pmd)
> +{
> +	return pmd_soft_dirty(pmd);
> +}
> +
> +static inline pmd_t pmd_swp_clear_soft_dirty(pmd_t pmd)
> +{
> +	return pmd_clear_soft_dirty(pmd);
> +}
> +
> +static inline pmd_t pmd_swp_mksoft_dirty(pmd_t pmd)
> +{
> +	return pmd_mksoft_dirty(pmd);
> +}
> +#endif /* CONFIG_ARCH_ENABLE_THP_MIGRATION */
> +#endif /* CONFIG_ARM64_HW_AFDBM */
>  
>  #define __HAVE_ARCH_PTEP_MODIFY_PROT_TRANSACTION
>  #define ptep_modify_prot_start ptep_modify_prot_start