Re: [PATCH] hugetlbfs: add MTE support

[David Hildenbrand <david@redhat.com>]

On 03.07.24 02:20, Yang Shi wrote:
> 
> 
> On 7/2/24 6:09 AM, David Hildenbrand wrote:
>> On 02.07.24 14:34, Catalin Marinas wrote:
>>> On Tue, Jun 25, 2024 at 04:37:17PM -0700, Yang Shi wrote:
>>>> MTE can be supported on ram based filesystem. It is supported on tmpfs.
>>>> There is use case to use MTE on hugetlbfs as well, adding MTE support.
>>>>
>>>> Signed-off-by: Yang Shi <yang@os.amperecomputing.com>
>>>> ---
>>>>    fs/hugetlbfs/inode.c | 2 +-
>>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c
>>>> index ecad73a4f713..c34faef62daf 100644
>>>> --- a/fs/hugetlbfs/inode.c
>>>> +++ b/fs/hugetlbfs/inode.c
>>>> @@ -110,7 +110,7 @@ static int hugetlbfs_file_mmap(struct file
>>>> *file, struct vm_area_struct *vma)
>>>>         * way when do_mmap unwinds (may be important on powerpc
>>>>         * and ia64).
>>>>         */
>>>> -    vm_flags_set(vma, VM_HUGETLB | VM_DONTEXPAND);
>>>> +    vm_flags_set(vma, VM_HUGETLB | VM_DONTEXPAND | VM_MTE_ALLOWED);
>>>>        vma->vm_ops = &hugetlb_vm_ops;
>>>
>>> Last time I checked, about a year ago, this was not sufficient. One
>>> issue is that there's no arch_clear_hugetlb_flags() implemented by your
>>> patch, leaving PG_arch_{2,3} set on a page. The other issue was that I
>>> initially tried to do this only on the head page but this did not go
>>> well with the folio_copy() -> copy_highpage() which expects the
>>> PG_arch_* flags on each individual page. The alternative was for
>>> arch_clear_hugetlb_flags() to iterate over all the pages in a folio.
>>
>> This would likely also add a blocker for
>> ARCH_WANT_OPTIMIZE_HUGETLB_VMEMMAP on arm64 (no idea if there are now
>> ways to move forward with that now, or if we are still not sure if we
>> can actually add support), correct?
> 
> IIUC, it is not. We just need to guarantee each subpage has
> PG_mte_tagged flag and allocated tags. The HVO just maps the 7 vmemmap
> pages for sub pages to the first page, they still see the flag and the
> space for tag is not impacted, right? Did I miss something?

In the R/O vmemmap optimization we won't be able to modify the flags of 
the double-mapped vmemmap pages via the double mappings.

Of course, we could find HVO-specific ways to only modify the flags of 
the first vmemmap page, but it does sound wrong ...

Really, the question is if we can have a per-folio flag for hugetlb 
instead and avoid all that?

-- 
Cheers,

David / dhildenb