Re: [v2 PATCH 1/2] hugetlb: arm64: add mte support

[Yang Shi <yang@os.amperecomputing.com>]

On 9/3/24 2:35 PM, David Hildenbrand wrote:
> On 03.09.24 18:46, Yang Shi wrote:
>>
>>
>> On 9/2/24 7:33 AM, David Hildenbrand wrote:
>>> On 21.08.24 20:47, Yang Shi wrote:
>>>> Enable MTE support for hugetlb.
>>>>
>>>> The MTE page flags will be set on the head page only.  When copying
>>>> hugetlb folio, the tags for all tail pages will be copied when copying
>>>> head page.
>>>>
>>>> When freeing hugetlb folio, the MTE flags will be cleared.
>>>>
>>>> Signed-off-by: Yang Shi <yang@os.amperecomputing.com>
>>>> ---
>>>>    arch/arm64/include/asm/hugetlb.h | 11 ++++++++++-
>>>>    arch/arm64/include/asm/mman.h    |  3 ++-
>>>>    arch/arm64/kernel/hibernate.c    |  7 +++++++
>>>>    arch/arm64/kernel/mte.c          | 25 +++++++++++++++++++++++--
>>>>    arch/arm64/kvm/guest.c           | 16 +++++++++++++---
>>>>    arch/arm64/kvm/mmu.c             | 11 +++++++++++
>>>>    arch/arm64/mm/copypage.c         | 25 +++++++++++++++++++++++--
>>>>    fs/hugetlbfs/inode.c             |  2 +-
>>>>    8 files changed, 90 insertions(+), 10 deletions(-)
>>>>
>>>> v2: * Reimplemented the patch to fix the comments from Catalin.
>>>>       * Added test cases (patch #2) per Catalin.
>>>>
>>>> diff --git a/arch/arm64/include/asm/hugetlb.h
>>>> b/arch/arm64/include/asm/hugetlb.h
>>>> index 293f880865e8..00a1f75d40ee 100644
>>>> --- a/arch/arm64/include/asm/hugetlb.h
>>>> +++ b/arch/arm64/include/asm/hugetlb.h
>>>> @@ -11,6 +11,7 @@
>>>>    #define __ASM_HUGETLB_H
>>>>      #include <asm/cacheflush.h>
>>>> +#include <asm/mte.h>
>>>>    #include <asm/page.h>
>>>>      #ifdef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
>>>> @@ -20,7 +21,15 @@ extern bool
>>>> arch_hugetlb_migration_supported(struct hstate *h);
>>>>      static inline void arch_clear_hugetlb_flags(struct folio *folio)
>>>>    {
>>>> -    clear_bit(PG_dcache_clean, &folio->flags);
>>>> +    const unsigned long clear_flags = BIT(PG_dcache_clean) |
>>>> +        BIT(PG_mte_tagged) | BIT(PG_mte_lock);
>>>> +
>>>> +    if (!system_supports_mte()) {
>>>> +        clear_bit(PG_dcache_clean, &folio->flags);
>>>> +        return;
>>>> +    }
>>>> +
>>>> +    folio->flags &= ~clear_flags;
>>>>    }
>>>>    #define arch_clear_hugetlb_flags arch_clear_hugetlb_flags
>>>>    diff --git a/arch/arm64/include/asm/mman.h
>>>> b/arch/arm64/include/asm/mman.h
>>>> index 5966ee4a6154..304dfc499e68 100644
>>>> --- a/arch/arm64/include/asm/mman.h
>>>> +++ b/arch/arm64/include/asm/mman.h
>>>> @@ -28,7 +28,8 @@ static inline unsigned long
>>>> arch_calc_vm_flag_bits(unsigned long flags)
>>>>         * backed by tags-capable memory. The vm_flags may be
>>>> overridden by a
>>>>         * filesystem supporting MTE (RAM-based).
>>>>         */
>>>> -    if (system_supports_mte() && (flags & MAP_ANONYMOUS))
>>>> +    if (system_supports_mte() &&
>>>> +        (flags & (MAP_ANONYMOUS | MAP_HUGETLB)))
>>>>            return VM_MTE_ALLOWED;
>>>>          return 0;
>>>> diff --git a/arch/arm64/kernel/hibernate.c
>>>> b/arch/arm64/kernel/hibernate.c
>>>> index 02870beb271e..722e76f29141 100644
>>>> --- a/arch/arm64/kernel/hibernate.c
>>>> +++ b/arch/arm64/kernel/hibernate.c
>>>> @@ -266,10 +266,17 @@ static int swsusp_mte_save_tags(void)
>>>>            max_zone_pfn = zone_end_pfn(zone);
>>>>            for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; 
>>>> pfn++) {
>>>>                struct page *page = pfn_to_online_page(pfn);
>>>> +            struct folio *folio;
>>>>                  if (!page)
>>>>                    continue;
>>>>    +            folio = page_folio(page);
>>>> +
>>>> +            if (folio_test_hugetlb(folio) &&
>>>> +                !page_mte_tagged(&folio->page))
>>>> +                continue;
>>>
>>> Can we have folio_test_mte_tagged() whereby you make sure that only
>>> folio_test_hugetlb() uses it for now (VM_WARN_ON_ONCE) and then make
>>> sure that nobody uses page_mte_tagged() on hugetlb folios
>>> (VM_WARN_ON_ONCE)?
>>
>>
>> IIUC, you mean something like the below?
>>
>> bool folio_test_mte_tagged(struct folio *folio)
>> {
>>       VM_WARN_ON_ONCE(!folio_test_hugetlb(folio));
>>
>>       return test_bit(PG_mte_tagged, &folio->page->flags);
>
> folio->flags
>
>> }
>>
>> bool page_mte_tagged(struct page *page)
>> {
>>       struct folio *folio = page_folio(page);
>>
>>       VM_WARN_ON_ONCE(folio_test_hugetlb(folio));
>
> Yes, but better as
>
> VM_WARN_ON_ONCE(folio_test_hugetlb(page_folio(page)));

I see. But I think all the call sites for folio_test_mte_tagged() 
actually need have folio_test_hugetlb() before it. So the warn seems not 
very useful other than warning on some misuse.

>
>>
>>       return test_bit(PG_mte_tagged, &pge->flags);
>> }
>>
>>>
>>> Same for folio_set_mte_tagged() and other functions. We could slap a
>>> "hugetlb" into the function names, but maybe in the future we'll be
>>> able to use a single flag per folio (I know, it's complicated ...).
>>
>> A single flag per folio should be the future direction, but I haven't
>> done the research so can't tell how complicated it will be.
>
> There were some discussions on it, and it's tricky. So maybe we should 
> really just have folio_test_hugetlb_mte_tagged() etc. for now

Either w/ hugetlb or w/o hugetlb is fine, I don't have strong opinion on 
the naming.