Hugetlb calculates addresses and page offsets differently from the rest of
mm. In order to pass struct vm_fault through the fault pathway we will let
hugetlb_fault() and __handle_mm_fault() set those variables themselves
instead.

Signed-off-by: Vishal Moola (Oracle) <vishal.moola@gmail.com>
---
 include/linux/mm.h | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/include/linux/mm.h b/include/linux/mm.h
index f5a97dec5169..c6874aa7b7f0 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -507,10 +507,11 @@ struct vm_fault {
 	const struct {
 		struct vm_area_struct *vma;	/* Target VMA */
 		gfp_t gfp_mask;			/* gfp mask to be used for allocations */
-		pgoff_t pgoff;			/* Logical page offset based on vma */
-		unsigned long address;		/* Faulting virtual address - masked */
 		unsigned long real_address;	/* Faulting virtual address - unmasked */
 	};
+	unsigned long address;		/* Faulting virtual address - masked */
+	pgoff_t pgoff;			/* Logical page offset based on vma */
+
 	enum fault_flag flags;		/* FAULT_FLAG_xxx flags
 					 * XXX: should really be 'const' */
 	pmd_t *pmd;			/* Pointer to pmd entry matching
-- 
2.43.0

On Mon, Mar 25, 2024 at 03:33:38PM -0700, Vishal Moola (Oracle) wrote:
> Hugetlb calculates addresses and page offsets differently from the rest of
> mm. In order to pass struct vm_fault through the fault pathway we will let
> hugetlb_fault() and __handle_mm_fault() set those variables themselves
> instead.

I don't think this is a great idea.  I'd rather not do patch 5 than do
patch 4+5.  If you look at the history, commits 742d33729a0df11 and
5857c9209ce58f show that drivers got into the bad habit of changing
address & pgoff, so they got made const to prevent that.

So can we make hugetlbfs OK with using addresses & pgoffsets that aren't
aligned to HPAGE boundaries?  Worth playing with for a bit to see how
deep that assumption runs.

On Mon, Mar 25, 2024 at 7:38 PM Matthew Wilcox <willy@infradead.org> wrote:
>
> On Mon, Mar 25, 2024 at 03:33:38PM -0700, Vishal Moola (Oracle) wrote:
> > Hugetlb calculates addresses and page offsets differently from the rest of
> > mm. In order to pass struct vm_fault through the fault pathway we will let
> > hugetlb_fault() and __handle_mm_fault() set those variables themselves
> > instead.
>
> I don't think this is a great idea.  I'd rather not do patch 5 than do
> patch 4+5.  If you look at the history, commits 742d33729a0df11 and
> 5857c9209ce58f show that drivers got into the bad habit of changing
> address & pgoff, so they got made const to prevent that.
>
> So can we make hugetlbfs OK with using addresses & pgoffsets that aren't
> aligned to HPAGE boundaries?  Worth playing with for a bit to see how
> deep that assumption runs.

Hmmm, I'll take a look. I don't think there should be too many issues
with that.