[next,1/3] truncate,shmem: Fix data loss when hole punched in folio

Commit Message

Hugh Dickins Jan. 3, 2022, 1:32 a.m. UTC

As reported before in
https://lore.kernel.org/lkml/[e-mail]/
shmem_undo_range() mods sometimes caused good data to be zeroed when
running my tmpfs swapping loads.  Only the ext4-on-loop0-on-tmpfs mount
was seen to suffer, and that is mounted with "-o discard": which punches
holes in the underlying tmpfs file.

shmem_undo_range() partial_end handling was wrong: if (lend + 1) aligned
to page but not to folio, the second shmem_get_folio() could be skipped,
then the whole of that folio punched out instead of treated partially.

Rename same_page to same_folio (like in truncate.c), and rely on that
instead of partial_end: fewer variables, less confusion.  And considering
an off-by-one in setting same_folio initially, pointed to an off-by-one
in the second shmem_get_folio(): it should be on (lend >> PAGE_SHIFT) not
end - which had caused no data loss, but could split folio unnecessarily.

And apply these same fixes in truncate_inode_pages_range().

Fixes: 8842c9c23524 ("truncate,shmem: Handle truncates that split large folios")
Signed-off-by: Hugh Dickins <[e-mail]>
---

 mm/shmem.c    |   16 ++++++----------
 mm/truncate.c |   15 +++++++--------
 2 files changed, 13 insertions(+), 18 deletions(-)

Comments

Christoph Hellwig Jan. 3, 2022, 7:51 a.m. UTC | #1

On Sun, Jan 02, 2022 at 05:32:28PM -0800, Hugh Dickins wrote:
>  
> +	same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);

Should this move to the else branch?

Same for the other copy of this code.  Otherwise this looks sane.

Hugh Dickins Jan. 3, 2022, 8:17 p.m. UTC | #2

On Sun, 2 Jan 2022, Christoph Hellwig wrote:
> On Sun, Jan 02, 2022 at 05:32:28PM -0800, Hugh Dickins wrote:
> >  
> > +	same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
> 
> Should this move to the else branch?

We could add an else branch and move it there, yes.  I liked
Matthew's else-less style with partial_end, and followed that.

Whatever: since he posted that diff yesterday, I imagine he'll
just merge this into the fixed patch, in whatever style he prefers.

> 
> Same for the other copy of this code.  Otherwise this looks sane.

Thanks,
Hugh

Patch

--- next-20211224/mm/shmem.c
+++ hughd1/mm/shmem.c
@@ -908,10 +908,10 @@  static void shmem_undo_range(struct inod
 	struct folio_batch fbatch;
 	pgoff_t indices[PAGEVEC_SIZE];
 	struct folio *folio;
+	bool same_folio;
 	long nr_swaps_freed = 0;
 	pgoff_t index;
 	int i;
-	bool partial_end;
 
 	if (lend == -1)
 		end = -1;	/* unsigned, so actually very big */
@@ -947,18 +947,14 @@  static void shmem_undo_range(struct inod
 		index++;
 	}
 
-	partial_end = ((lend + 1) % PAGE_SIZE) > 0;
+	same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
 	shmem_get_folio(inode, lstart >> PAGE_SHIFT, &folio, SGP_READ);
 	if (folio) {
-		bool same_page;
-
-		same_page = lend < folio_pos(folio) + folio_size(folio);
-		if (same_page)
-			partial_end = false;
+		same_folio = lend < folio_pos(folio) + folio_size(folio);
 		folio_mark_dirty(folio);
 		if (!truncate_inode_partial_folio(folio, lstart, lend)) {
 			start = folio->index + folio_nr_pages(folio);
-			if (same_page)
+			if (same_folio)
 				end = folio->index;
 		}
 		folio_unlock(folio);
@@ -966,8 +962,8 @@  static void shmem_undo_range(struct inod
 		folio = NULL;
 	}
 
-	if (partial_end)
-		shmem_get_folio(inode, end, &folio, SGP_READ);
+	if (!same_folio)
+		shmem_get_folio(inode, lend >> PAGE_SHIFT, &folio, SGP_READ);
 	if (folio) {
 		folio_mark_dirty(folio);
 		if (!truncate_inode_partial_folio(folio, lstart, lend))
--- next-20211224/mm/truncate.c
+++ hughd1/mm/truncate.c
@@ -347,8 +347,8 @@  void truncate_inode_pages_range(struct a
 	pgoff_t		indices[PAGEVEC_SIZE];
 	pgoff_t		index;
 	int		i;
-	struct folio *	folio;
-	bool partial_end;
+	struct folio	*folio;
+	bool		same_folio;
 
 	if (mapping_empty(mapping))
 		goto out;
@@ -385,12 +385,10 @@  void truncate_inode_pages_range(struct a
 		cond_resched();
 	}
 
-	partial_end = ((lend + 1) % PAGE_SIZE) > 0;
+	same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
 	folio = __filemap_get_folio(mapping, lstart >> PAGE_SHIFT, FGP_LOCK, 0);
 	if (folio) {
-		bool same_folio = lend < folio_pos(folio) + folio_size(folio);
-		if (same_folio)
-			partial_end = false;
+		same_folio = lend < folio_pos(folio) + folio_size(folio);
 		if (!truncate_inode_partial_folio(folio, lstart, lend)) {
 			start = folio->index + folio_nr_pages(folio);
 			if (same_folio)
@@ -401,8 +399,9 @@  void truncate_inode_pages_range(struct a
 		folio = NULL;
 	}
 
-	if (partial_end)
-		folio = __filemap_get_folio(mapping, end, FGP_LOCK, 0);
+	if (!same_folio)
+		folio = __filemap_get_folio(mapping, lend >> PAGE_SHIFT,
+						FGP_LOCK, 0);
 	if (folio) {
 		if (!truncate_inode_partial_folio(folio, lstart, lend))
 			end = folio->index;