Skip to contentSkip to conversation

iommufd: Fix offset handling in folio subroutines

PATCH In review Alexey Kardashevskiy 2 participants 3 messages Apr 2, 2025 – Apr 7, 2025

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 iommufd: Fix offset handling in folio subroutines

The @offset from memfd_pin_folios() is pgoff_t and is bytes, not a number of pages as batch_from_folios() treats it.

Convert byte offset to pfn offset and use that.

Fixes ed9178fbfd4e ("iommufd: Folio subroutines")

---

Found this when tried mapping a bunch of folios (>64) with an offset and the first folio was used for all the mappings in the batch.

---
 drivers/iommu/iommufd/pages.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/iommu/iommufd/pages.c +3 -2
index 3427749bc5ce..6075d11691a6 100644
@@ -646,8 +646,9 @@ static int batch_from_folios(struct pfn_batch *batch, struct folio ***folios_p,
 
 	while (npages) {
 		struct folio *folio = *folios;
-		unsigned long nr = folio_nr_pages(folio) - offset;
-		unsigned long pfn = page_to_pfn(folio_page(folio, offset));
+		unsigned long offset_nr = offset >> PAGE_SHIFT;
+		unsigned long nr = folio_nr_pages(folio) - offset_nr;
+		unsigned long pfn = page_to_pfn(folio_page(folio, offset_nr));
 
 		nr = min(nr, npages);
 		npages -= nr;
Signature · 1 line
2.47.1
maintainer
#1 lore raw
Quoted 25 lines

From: Alexey Kardashevskiy <[email protected]>
Sent: Wednesday, April 2, 2025 9:44 AM

The @offset from memfd_pin_folios() is pgoff_t and is bytes, not a number of pages as batch_from_folios() treats it.

Convert byte offset to pfn offset and use that.

Fixes ed9178fbfd4e ("iommufd: Folio subroutines")

---

Found this when tried mapping a bunch of folios (>64) with an offset and the first folio was used for all the mappings in the batch.

---
 drivers/iommu/iommufd/pages.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/iommu/iommufd/pages.c +0 -0
index 3427749bc5ce..6075d11691a6 100644

drivers/iommu/iommufd/pages.c static int batch_from_folios(struct pfn_batch *batch, · line 652

Quoted 9 lines
drivers/iommu/iommufd/pages.c +3 -2
@@ -646,8 +646,9 @@ static int batch_from_folios(struct pfn_batch *batch,
 struct folio ***folios_p,
 
 	while (npages) {
 		struct folio *folio = *folios;
-		unsigned long nr = folio_nr_pages(folio) - offset;
-		unsigned long pfn = page_to_pfn(folio_page(folio, offset));
+		unsigned long offset_nr = offset >> PAGE_SHIFT;
+		unsigned long nr = folio_nr_pages(folio) - offset_nr;
+		unsigned long pfn = page_to_pfn(folio_page(folio, offset_nr));

pin_memfd_pages():

	nfolios = memfd_pin_folios(user->file, start, end, user->ufolios,
				      nfolios, &offset);
	offset >>= PAGE_SHIFT;
	user->ufolios_next = user->ufolios;
	user->ufolios_offset = offset;

pfn_reader_fill_span():

	if (!user->file)
		batch_from_pages(&pfns->batch, user->upages + start_index,
				    npages);
	else
		rc = batch_from_folios(&pfns->batch, &user->ufolios_next,
					&user->ufolios_offset, npages); 

So the offset is already about a page number as batch_from_folios() expects.

Quoted 5 lines
drivers/iommu/iommufd/pages.c +0 -1
 		nr = min(nr, npages);
 		npages -= nr;
--

2.47.1

#2 lore raw
Quoted 36 lines · Kevin
Quoted 25 lines

From: Alexey Kardashevskiy <[email protected]>
Sent: Wednesday, April 2, 2025 9:44 AM

The @offset from memfd_pin_folios() is pgoff_t and is bytes, not a number of pages as batch_from_folios() treats it.

Convert byte offset to pfn offset and use that.

Fixes ed9178fbfd4e ("iommufd: Folio subroutines")

---

Found this when tried mapping a bunch of folios (>64) with an offset and the first folio was used for all the mappings in the batch.

---
  drivers/iommu/iommufd/pages.c | 5 +++--
  1 file changed, 3 insertions(+), 2 deletions(-)
drivers/iommu/iommufd/pages.c +0 -0
index 3427749bc5ce..6075d11691a6 100644

drivers/iommu/iommufd/pages.c static int batch_from_folios(struct pfn_batch *batch, · line 652

Quoted 9 lines
drivers/iommu/iommufd/pages.c +3 -2
@@ -646,8 +646,9 @@ static int batch_from_folios(struct pfn_batch *batch,
 struct folio ***folios_p,
 
  	while (npages) {
  		struct folio *folio = *folios;
-		unsigned long nr = folio_nr_pages(folio) - offset;
-		unsigned long pfn = page_to_pfn(folio_page(folio, offset));
+		unsigned long offset_nr = offset >> PAGE_SHIFT;
+		unsigned long nr = folio_nr_pages(folio) - offset_nr;
+		unsigned long pfn = page_to_pfn(folio_page(folio, offset_nr));

pin_memfd_pages():

Quoted 3 lines
	nfolios = memfd_pin_folios(user->file, start, end, user->ufolios,
				      nfolios, &offset);
	offset >>= PAGE_SHIFT;

Aaaaahhh missed that so there is a bug in my variant of memfd_pin_folios() for gmemfd and sorry for the noise. Thanks,

Quoted 19 lines
	user->ufolios_next = user->ufolios;
	user->ufolios_offset = offset;

pfn_reader_fill_span():

	if (!user->file)
		batch_from_pages(&pfns->batch, user->upages + start_index,
				    npages);
	else
		rc = batch_from_folios(&pfns->batch, &user->ufolios_next,
					&user->ufolios_offset, npages);

So the offset is already about a page number as batch_from_folios() expects.

Quoted 5 lines
drivers/iommu/iommufd/pages.c +0 -1
  		nr = min(nr, npages);
  		npages -= nr;
--

2.47.1

Signature · 1 line
Alexey