Re: [PATCH v3 1/6] mm: memory: extend finish_fault() to support large folio

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 2024/6/3 17:01, Barry Song wrote:
> On Mon, Jun 3, 2024 at 8:58 PM Barry Song <21cnbao@gmail.com> wrote:
>>
>> On Mon, Jun 3, 2024 at 8:29 PM Baolin Wang
>> <baolin.wang@linux.alibaba.com> wrote:
>>>
>>>
>>>
>>> On 2024/6/3 13:28, Barry Song wrote:
>>>> On Thu, May 30, 2024 at 2:04 PM Baolin Wang
>>>> <baolin.wang@linux.alibaba.com> wrote:
>>>>>
>>>>> Add large folio mapping establishment support for finish_fault() as a preparation,
>>>>> to support multi-size THP allocation of anonymous shmem pages in the following
>>>>> patches.
>>>>>
>>>>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
>>>>> ---
>>>>>    mm/memory.c | 58 ++++++++++++++++++++++++++++++++++++++++++++---------
>>>>>    1 file changed, 48 insertions(+), 10 deletions(-)
>>>>>
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index eef4e482c0c2..435187ff7ea4 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -4831,9 +4831,12 @@ vm_fault_t finish_fault(struct vm_fault *vmf)
>>>>>    {
>>>>>           struct vm_area_struct *vma = vmf->vma;
>>>>>           struct page *page;
>>>>> +       struct folio *folio;
>>>>>           vm_fault_t ret;
>>>>>           bool is_cow = (vmf->flags & FAULT_FLAG_WRITE) &&
>>>>>                         !(vma->vm_flags & VM_SHARED);
>>>>> +       int type, nr_pages, i;
>>>>> +       unsigned long addr = vmf->address;
>>>>>
>>>>>           /* Did we COW the page? */
>>>>>           if (is_cow)
>>>>> @@ -4864,24 +4867,59 @@ vm_fault_t finish_fault(struct vm_fault *vmf)
>>>>>                           return VM_FAULT_OOM;
>>>>>           }
>>>>>
>>>>> +       folio = page_folio(page);
>>>>> +       nr_pages = folio_nr_pages(folio);
>>>>> +
>>>>> +       /*
>>>>> +        * Using per-page fault to maintain the uffd semantics, and same
>>>>> +        * approach also applies to non-anonymous-shmem faults to avoid
>>>>> +        * inflating the RSS of the process.
>>>>
>>>> I don't feel the comment explains the root cause.
>>>> For non-shmem, anyway we have allocated the memory? Avoiding inflating
>>>> RSS seems not so useful as we have occupied the memory. the memory footprint
>>>
>>> This is also to keep the same behavior as before for non-anon-shmem, and
>>> will be discussed in the future.
>>
>> OK.
>>
>>>
>>>> is what we really care about. so we want to rely on read-ahead hints of subpage
>>>> to determine read-ahead size? that is why we don't map nr_pages for non-shmem
>>>> files though we can potentially reduce nr_pages - 1 page faults?
>>>
>>> IMHO, there is 2 cases for non-anon-shmem:
>>> (1) read mmap() faults: we can rely on the 'fault_around_bytes'
>>> interface to determin what size of mapping to build.
>>> (2) writable mmap() faults: I want to keep the same behavior as before
>>> (per-page fault), but we can talk about this when I send new patches to
>>> use mTHP to control large folio allocation for writable mmap().
>>
>> OK.
>>
>>>
>>>>> +        */
>>>>> +       if (!vma_is_anon_shmem(vma) || unlikely(userfaultfd_armed(vma))) {
>>>>> +               nr_pages = 1;
>>>>> +       } else if (nr_pages > 1) {
>>>>> +               pgoff_t idx = folio_page_idx(folio, page);
>>>>> +               /* The page offset of vmf->address within the VMA. */
>>>>> +               pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
>>>>> +
>>>>> +               /*
>>>>> +                * Fallback to per-page fault in case the folio size in page
>>>>> +                * cache beyond the VMA limits.
>>>>> +                */
>>>>> +               if (unlikely(vma_off < idx ||
>>>>> +                            vma_off + (nr_pages - idx) > vma_pages(vma))) {
>>>>> +                       nr_pages = 1;
>>>>> +               } else {
>>>>> +                       /* Now we can set mappings for the whole large folio. */
>>>>> +                       addr = vmf->address - idx * PAGE_SIZE;
>>>>> +                       page = &folio->page;
>>>>> +               }
>>>>> +       }
>>>>> +
>>>>>           vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
>>>>> -                                     vmf->address, &vmf->ptl);
>>>>> +                                      addr, &vmf->ptl);
>>>>>           if (!vmf->pte)
>>>>>                   return VM_FAULT_NOPAGE;
>>>>>
>>>>>           /* Re-check under ptl */
>>>>> -       if (likely(!vmf_pte_changed(vmf))) {
>>>>> -               struct folio *folio = page_folio(page);
>>>>> -               int type = is_cow ? MM_ANONPAGES : mm_counter_file(folio);
>>>>> -
>>>>> -               set_pte_range(vmf, folio, page, 1, vmf->address);
>>>>> -               add_mm_counter(vma->vm_mm, type, 1);
>>>>> -               ret = 0;
>>>>> -       } else {
>>>>> -               update_mmu_tlb(vma, vmf->address, vmf->pte);
>>>>> +       if (nr_pages == 1 && unlikely(vmf_pte_changed(vmf))) {
>>>>> +               update_mmu_tlb(vma, addr, vmf->pte);
>>>>>                   ret = VM_FAULT_NOPAGE;
>>>>> +               goto unlock;
>>>>> +       } else if (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages)) {
>>>>
>>>> In what case we can't use !pte_range_none(vmf->pte, 1) for nr_pages == 1
>>>> then unify the code for nr_pages==1 and nr_pages > 1?
>>>>
>>>> It seems this has been discussed before, but I forget the reason.
>>>
>>> IIUC, this is for uffd case, which is not a none pte entry.
>>
>> Is it possible to have a COW case for shmem? For example, if someone
>> maps a shmem
>> file as read-only and then writes to it, would that prevent the use of
>> pte_range_none?
> 
> sorry, i mean PRIVATE but not READ-ONLY.

Yes, I think so. Now CoW case still use per-page fault in do_cow_fault().

>> Furthermore, if we encounter a large folio in shmem while reading,
>> does it necessarily
>> mean we can map the entire folio? Is it possible for some processes to

Now this will depend on the 'fault_around_bytes' interface.

>> only map part
>> of large folios? For instance, if process A allocates large folios and
>> process B maps
>> only part of this shmem file or partially unmaps a large folio, how
>> would that be handled?

This is certainly possible.

For tmpfs:
(1) If 'fault_around_bytes' is enabled, filemap_map_pages() will handle 
partially mapping of the large folio for process B.

(2) If 'fault_around_bytes' is set to 0, finish_fault() will fallback to 
per-page fault.

For Anonomous shmem, process B should be the child of process A in your 
case, then:
(1) If 'fault_around_bytes' is enabled, behavior is same with tmpfs.

(2) If 'fault_around_bytes' is set to 0, finish_fault() will build the 
whole large folio mapping for process B. Since process B will copy the 
same shared VMA from parent process A, which means a mTHP mapping to share.

>> Apologies for not debugging this thoroughly, but these two corner
>> cases seem worth
>> considering. If these scenarios have already been addressed, please disregard my
>> comments.

No worries:) Thanks for your valuable input.