2024-06-11 14:19:07

by Andrew Bresticker

[permalink] [raw]
Subject: [PATCH] mm/memory: Pass head page to do_set_pmd()

The requirement that the head page be passed to do_set_pmd() was added
in commit ef37b2ea08ac ("mm/memory: page_add_file_rmap() ->
folio_add_file_rmap_[pte|pmd]()") and prevents pmd-mapping in the
finish_fault() path if vmf->page is anything but the head page for an
otherwise suitable vma and pmd-sized page. Have finish_fault() pass in
the head page instead.

Fixes: ef37b2ea08ac ("mm/memory: page_add_file_rmap() -> folio_add_file_rmap_[pte|pmd]()")
Signed-off-by: Andrew Bresticker <[email protected]>
---
mm/memory.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/memory.c b/mm/memory.c
index 0f47a533014e..f13b953b507c 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -4764,7 +4764,7 @@ vm_fault_t finish_fault(struct vm_fault *vmf)

if (pmd_none(*vmf->pmd)) {
if (PageTransCompound(page)) {
- ret = do_set_pmd(vmf, page);
+ ret = do_set_pmd(vmf, compound_head(page));
if (ret != VM_FAULT_FALLBACK)
return ret;
}
--
2.34.1



2024-06-11 14:46:32

by David Hildenbrand

[permalink] [raw]
Subject: Re: [PATCH] mm/memory: Pass head page to do_set_pmd()

On 11.06.24 16:18, Andrew Bresticker wrote:
> The requirement that the head page be passed to do_set_pmd() was added
> in commit ef37b2ea08ac ("mm/memory: page_add_file_rmap() ->
> folio_add_file_rmap_[pte|pmd]()") and prevents pmd-mapping in the
> finish_fault() path if vmf->page is anything but the head page for an
> otherwise suitable vma and pmd-sized page. Have finish_fault() pass in
> the head page instead.
>
> Fixes: ef37b2ea08ac ("mm/memory: page_add_file_rmap() -> folio_add_file_rmap_[pte|pmd]()")
> Signed-off-by: Andrew Bresticker <[email protected]>
> ---
> mm/memory.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/mm/memory.c b/mm/memory.c
> index 0f47a533014e..f13b953b507c 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -4764,7 +4764,7 @@ vm_fault_t finish_fault(struct vm_fault *vmf)
>
> if (pmd_none(*vmf->pmd)) {
> if (PageTransCompound(page)) {
> - ret = do_set_pmd(vmf, page);
> + ret = do_set_pmd(vmf, compound_head(page));
> if (ret != VM_FAULT_FALLBACK)
> return ret;
> }

That certainly makes the "page != &folio->page" check happy.

It is *likely* incorrect if we would ever have folios > PMD size (which
we don't have on that path yet).

I assume that the thp_vma_suitable_order() check would detect any kind
of "different placement of the folio in virtual address space", where we
could mess up.

Question is: should we instead drop the "page != &folio->page" check
that I added?

I think I added that check because I saw the "compound_order(page)"
check and assumed it would return 0 for tail pages, but missed that we
get the compound head first.

--
Cheers,

David / dhildenb


2024-06-11 14:48:41

by David Hildenbrand

[permalink] [raw]
Subject: Re: [PATCH] mm/memory: Pass head page to do_set_pmd()

On 11.06.24 16:46, David Hildenbrand wrote:
> On 11.06.24 16:18, Andrew Bresticker wrote:
>> The requirement that the head page be passed to do_set_pmd() was added
>> in commit ef37b2ea08ac ("mm/memory: page_add_file_rmap() ->
>> folio_add_file_rmap_[pte|pmd]()") and prevents pmd-mapping in the
>> finish_fault() path if vmf->page is anything but the head page for an
>> otherwise suitable vma and pmd-sized page. Have finish_fault() pass in
>> the head page instead.
>>
>> Fixes: ef37b2ea08ac ("mm/memory: page_add_file_rmap() -> folio_add_file_rmap_[pte|pmd]()")
>> Signed-off-by: Andrew Bresticker <[email protected]>
>> ---
>> mm/memory.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/mm/memory.c b/mm/memory.c
>> index 0f47a533014e..f13b953b507c 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -4764,7 +4764,7 @@ vm_fault_t finish_fault(struct vm_fault *vmf)
>>
>> if (pmd_none(*vmf->pmd)) {
>> if (PageTransCompound(page)) {
>> - ret = do_set_pmd(vmf, page);
>> + ret = do_set_pmd(vmf, compound_head(page));
>> if (ret != VM_FAULT_FALLBACK)
>> return ret;
>> }
>
> That certainly makes the "page != &folio->page" check happy.
>
> It is *likely* incorrect if we would ever have folios > PMD size (which
> we don't have on that path yet).
>
> I assume that the thp_vma_suitable_order() check would detect any kind
> of "different placement of the folio in virtual address space", where we
> could mess up.
>
> Question is: should we instead drop the "page != &folio->page" check
> that I added?

To be precise, the something like following:

diff --git a/mm/memory.c b/mm/memory.c
index 5e633c9c27a8b..76f48dc5899cd 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -4699,8 +4699,9 @@ vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
return ret;

- if (page != &folio->page || folio_order(folio) != HPAGE_PMD_ORDER)
+ if (folio_order(folio) != HPAGE_PMD_ORDER)
return ret;
+ page = &folio->page;

/*
* Just backoff if any subpage of a THP is corrupted otherwise


--
Cheers,

David / dhildenb


2024-06-11 15:28:31

by Andrew Bresticker

[permalink] [raw]
Subject: Re: [PATCH] mm/memory: Pass head page to do_set_pmd()

On Tue, Jun 11, 2024 at 10:46 AM David Hildenbrand <[email protected]> wrote:
>
> On 11.06.24 16:18, Andrew Bresticker wrote:
> > The requirement that the head page be passed to do_set_pmd() was added
> > in commit ef37b2ea08ac ("mm/memory: page_add_file_rmap() ->
> > folio_add_file_rmap_[pte|pmd]()") and prevents pmd-mapping in the
> > finish_fault() path if vmf->page is anything but the head page for an
> > otherwise suitable vma and pmd-sized page. Have finish_fault() pass in
> > the head page instead.
> >
> > Fixes: ef37b2ea08ac ("mm/memory: page_add_file_rmap() -> folio_add_file_rmap_[pte|pmd]()")
> > Signed-off-by: Andrew Bresticker <[email protected]>
> > ---
> > mm/memory.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/mm/memory.c b/mm/memory.c
> > index 0f47a533014e..f13b953b507c 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -4764,7 +4764,7 @@ vm_fault_t finish_fault(struct vm_fault *vmf)
> >
> > if (pmd_none(*vmf->pmd)) {
> > if (PageTransCompound(page)) {
> > - ret = do_set_pmd(vmf, page);
> > + ret = do_set_pmd(vmf, compound_head(page));
> > if (ret != VM_FAULT_FALLBACK)
> > return ret;
> > }
>
> That certainly makes the "page != &folio->page" check happy.
>
> It is *likely* incorrect if we would ever have folios > PMD size (which
> we don't have on that path yet).
>
> I assume that the thp_vma_suitable_order() check would detect any kind
> of "different placement of the folio in virtual address space", where we
> could mess up.

thp_vma_suitable_order() checks that the folio fits in the vma, if
that's what you mean.

> Question is: should we instead drop the "page != &folio->page" check
> that I added?

Indeed that's probably better as I'm now noticing that
filemap_map_pmd() has the same issue. Will send a v2.

Thanks.
-Andrew

>
> I think I added that check because I saw the "compound_order(page)"
> check and assumed it would return 0 for tail pages, but missed that we
> get the compound head first.
>
> --
> Cheers,
>
> David / dhildenb
>