Skip to contentSkip to conversation

iomap: fix offset truncation in iomap_fill_dirty_folios() on 32-bit

PATCH Applied Shashank Mohan Jain 3 participants 4 messages Nov 17 – Nov 18

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

PATCH 1/1 iomap: fix offset truncation in iomap_fill_dirty_folios() on 32-bit

iomap_fill_dirty_folios() turns the page index at which the dirty folio lookup ended into the file offset it returns to the caller:

	*start = (pstart << PAGE_SHIFT);

pstart is a pgoff_t. On a 32-bit kernel the shift is done in 32 bits and the upper bits are lost for every offset at or above 4 GiB.

XFS uses the returned offset in xfs_buffered_write_iomap_begin() to trim an unwritten mapping to the range covered by the lookup. With the truncated offset the end of that range is below its start and xfs_trim_extent() leaves a mapping of zero length. iomap_iter_done() warns about it,

  WARNING: fs/iomap/iter.c:38 at iomap_iter+0x33c/0x36c
  WARNING: fs/iomap/iter.c:39 at iomap_iter+0x34c/0x36c

and iomap_zero_range() does not terminate: a mapping of zero length is what the iterator takes for the start of an iteration, so it asks for the next mapping again and again without checking how much of the range is left.

On i386 with XFS (4k blocks) this is enough to trigger it:

	fd = open("f", O_RDWR | O_CREAT, 0644);
	fallocate(fd, 0, 5ULL << 30, 1 << 20);
	ftruncate(fd, (5ULL << 30) + 1000);

ftruncate() never returns. The task spins in iomap_zero_range() with the inode locked and does not react to SIGKILL. A write behind an unaligned EOF which is in an unwritten block above 4 GiB fails with EFSCORRUPTED instead, after "XFS (vda): Access to block zero in inode 132 start_block: 0 start_off: 140002 blkcnt: 0 extent-state: 0":

	fallocate(fd, 0, 5ULL << 30, 1000);
	pwrite(fd, buf, 100, (5ULL << 30) + 8192);

The same calls work below 4 GiB, and v6.19 behaves like v7.3-rc6.

Cast the page index to loff_t before the shift.

Tested with an i386 kernel in qemu and XFS on a virtio disk: the two sequences above hang and fail without this change and succeed with it.

Fixes 395ed1ef0012 ("iomap: optional zero range dirty folio processing")
Assisted-by LLM

---
Prepared with Claude Code (Anthropic), model Claude Opus 5.5 (claude-opus-5-5).

First sent privately to [email protected] on 11 October; Matthew Wilcox asked for it to be posted as a normal patch. Sorry for the delay. The code is unchanged in mainline, so this is the same fix, rebased on 666539eecdf4.

 fs/iomap/buffered-io.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/iomap/buffered-io.c +1 -1
index 6d843b1..846f87a 100644
@@ -1735,7 +1735,7 @@ iomap_fill_dirty_folios(
 	}
 
 	count = filemap_get_folios_dirty(mapping, &pstart, pend, iter->fbatch);
-	*start = (pstart << PAGE_SHIFT);
+	*start = (loff_t)pstart << PAGE_SHIFT;
 	*iomap_flags |= IOMAP_F_FOLIO_BATCH;
 	return count;
 }

base-commit: 666539eecdf4db623f5e3c3e55cf9f2dfad24aa4

Signature · 1 line
2.43.0
Reviewed-by on v1 1/1
#1 lore raw
Quoted 45 lines · Shashank Mohan Jain

iomap_fill_dirty_folios() turns the page index at which the dirty folio lookup ended into the file offset it returns to the caller:

	*start = (pstart << PAGE_SHIFT);

pstart is a pgoff_t. On a 32-bit kernel the shift is done in 32 bits and the upper bits are lost for every offset at or above 4 GiB.

XFS uses the returned offset in xfs_buffered_write_iomap_begin() to trim an unwritten mapping to the range covered by the lookup. With the truncated offset the end of that range is below its start and xfs_trim_extent() leaves a mapping of zero length. iomap_iter_done() warns about it,

  WARNING: fs/iomap/iter.c:38 at iomap_iter+0x33c/0x36c
  WARNING: fs/iomap/iter.c:39 at iomap_iter+0x34c/0x36c

and iomap_zero_range() does not terminate: a mapping of zero length is what the iterator takes for the start of an iteration, so it asks for the next mapping again and again without checking how much of the range is left.

On i386 with XFS (4k blocks) this is enough to trigger it:

	fd = open("f", O_RDWR | O_CREAT, 0644);
	fallocate(fd, 0, 5ULL << 30, 1 << 20);
	ftruncate(fd, (5ULL << 30) + 1000);

ftruncate() never returns. The task spins in iomap_zero_range() with the inode locked and does not react to SIGKILL. A write behind an unaligned EOF which is in an unwritten block above 4 GiB fails with EFSCORRUPTED instead, after "XFS (vda): Access to block zero in inode 132 start_block: 0 start_off: 140002 blkcnt: 0 extent-state: 0":

	fallocate(fd, 0, 5ULL << 30, 1000);
	pwrite(fd, buf, 100, (5ULL << 30) + 8192);

The same calls work below 4 GiB, and v6.19 behaves like v7.3-rc6.

Cast the page index to loff_t before the shift.

Tested with an i386 kernel in qemu and XFS on a virtio disk: the two sequences above hang and fail without this change and succeed with it.

Quoted 5 lines
Fixes 395ed1ef0012 ("iomap: optional zero range dirty folio processing")
Assisted-by LLM

---

Quoted 29 lines

Prepared with Claude Code (Anthropic), model Claude Opus 5.5 (claude-opus-5-5).

First sent privately to [email protected] on 11 October; Matthew Wilcox asked for it to be posted as a normal patch. Sorry for the delay. The code is unchanged in mainline, so this is the same fix, rebased on 666539eecdf4.

 fs/iomap/buffered-io.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/iomap/buffered-io.c +1 -1
index 6d843b1..846f87a 100644
@@ -1735,7 +1735,7 @@ iomap_fill_dirty_folios(
 	}
 
 	count = filemap_get_folios_dirty(mapping, &pstart, pend, iter->fbatch);
-	*start = (pstart << PAGE_SHIFT);
+	*start = (loff_t)pstart << PAGE_SHIFT;
 	*iomap_flags |= IOMAP_F_FOLIO_BATCH;
 	return count;
 }

base-commit: 666539eecdf4db623f5e3c3e55cf9f2dfad24aa4

Signature · 1 line
2.43.0
#2 lore raw
In reply to Shashank Mohan Jain

One correction to the note under the ---: it was Willy Tarreau on [email protected] who asked me to post this publicly, not Matthew Wilcox. Sorry for the noise, Matthew.

Shashank

Quoted 80 lines · Shashank Mohan Jain

iomap_fill_dirty_folios() turns the page index at which the dirty folio lookup ended into the file offset it returns to the caller:

        *start = (pstart << PAGE_SHIFT);

pstart is a pgoff_t. On a 32-bit kernel the shift is done in 32 bits and the upper bits are lost for every offset at or above 4 GiB.

XFS uses the returned offset in xfs_buffered_write_iomap_begin() to trim an unwritten mapping to the range covered by the lookup. With the truncated offset the end of that range is below its start and xfs_trim_extent() leaves a mapping of zero length. iomap_iter_done() warns about it,

  WARNING: fs/iomap/iter.c:38 at iomap_iter+0x33c/0x36c
  WARNING: fs/iomap/iter.c:39 at iomap_iter+0x34c/0x36c

and iomap_zero_range() does not terminate: a mapping of zero length is what the iterator takes for the start of an iteration, so it asks for the next mapping again and again without checking how much of the range is left.

On i386 with XFS (4k blocks) this is enough to trigger it:

        fd = open("f", O_RDWR | O_CREAT, 0644);
        fallocate(fd, 0, 5ULL << 30, 1 << 20);
        ftruncate(fd, (5ULL << 30) + 1000);

ftruncate() never returns. The task spins in iomap_zero_range() with the inode locked and does not react to SIGKILL. A write behind an unaligned EOF which is in an unwritten block above 4 GiB fails with EFSCORRUPTED instead, after "XFS (vda): Access to block zero in inode 132 start_block: 0 start_off: 140002 blkcnt: 0 extent-state: 0":

        fallocate(fd, 0, 5ULL << 30, 1000);
        pwrite(fd, buf, 100, (5ULL << 30) + 8192);

The same calls work below 4 GiB, and v6.19 behaves like v7.3-rc6.

Cast the page index to loff_t before the shift.

Tested with an i386 kernel in qemu and XFS on a virtio disk: the two sequences above hang and fail without this change and succeed with it.

Fixes 395ed1ef0012 ("iomap: optional zero range dirty folio processing")
Assisted-by LLM

---
Prepared with Claude Code (Anthropic), model Claude Opus 5.5 (claude-opus-5-5).

First sent privately to [email protected] on 11 October; Matthew Wilcox asked for it to be posted as a normal patch. Sorry for the delay. The code is unchanged in mainline, so this is the same fix, rebased on 666539eecdf4.

 fs/iomap/buffered-io.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/iomap/buffered-io.c +1 -1
index 6d843b1..846f87a 100644
@@ -1735,7 +1735,7 @@ iomap_fill_dirty_folios(
        }
 
        count = filemap_get_folios_dirty(mapping, &pstart, pend, iter->fbatch);
-       *start = (pstart << PAGE_SHIFT);
+       *start = (loff_t)pstart << PAGE_SHIFT;
        *iomap_flags |= IOMAP_F_FOLIO_BATCH;
        return count;
 }

base-commit: 666539eecdf4db623f5e3c3e55cf9f2dfad24aa4 --
2.43.0

maintainer Applied
#3 lore raw
In reply to Shashank Mohan Jain
Quoted 9 lines · Shashank Mohan Jain

iomap_fill_dirty_folios() turns the page index at which the dirty folio lookup ended into the file offset it returns to the caller:

	*start = (pstart << PAGE_SHIFT);

pstart is a pgoff_t. On a 32-bit kernel the shift is done in 32 bits and the upper bits are lost for every offset at or above 4 GiB.

[...]

Applied to the vfs.fixes branch of the vfs/vfs.git tree. Patches in the vfs.fixes branch should appear in linux-next soon.

Please report any outstanding bugs that were missed during review in a new review to the original patch series allowing us to drop it.

It's encouraged to provide Acked-bys and Reviewed-bys even though the patch has now been applied. If possible patch trailers will be updated.

Note that commit hashes shown below are subject to change due to rebase, trailer updates or similar. If in doubt, please check the listed branch.

tree:   https://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs.git
branch: vfs.fixes

[1/1] iomap: fix offset truncation in iomap_fill_dirty_folios() on 32-bit

      https://git.kernel.org/vfs/vfs/c/4d841b11ce1d