btrfs: fix generic/362 failures with nodatasum

Changelog
v3:
- Change the timing of isize revert to avoid race with buffered read Normally we need to take extent lock to update isize, but in this particular case, the ordered extent we're holding acts like an extent lock, thus we are safe to modify the isize as long as the OE is not finished.

- Add a comment on why it's safe to revert the isize

- Explain that the isize revert is not perfect
We should hold the extent lock until endio/iomap_end, and only update the isize for a successful submission.

But that will require quite a lot of extra work, and should be done later by a dedicated series other than a short fix.

v2:
- Add a new patch to allow retry after no byte is submitted for direct write

- Add a more clean callchain to explain why __iomap_dio_rw() doesn't return -EFAULT/-ENOTBLK directly.

- Add a code comment explaining the OE split sitation and why we can truncate all the remaining OE after a short write

- Fix the isize revert when part of the direct write succeeded

- Dig deeper into the original cause
Which is very old, dates back to v5.15 LTS at least, and add a reason why no specific fixes tag is provided.

The test case generic/362 is pretty well hidden by several factors:

- "nodatasum" will not take effect due to the bad design of the test case
If the target file already exists, the test case will reuse it, meanwhile "nodatasum" mount option only affects new files, meaning if the test case is executed before with data checksum, it will never go through the nodatasum path.

There is already an update on the test case to make the failure reliably reproducible:
https://lore.kernel.org/linux-btrfs/[email protected]/

- Btrfs always falls back to buffered IO if the inode has csum Thus we do not exercise the zero-copy path, hide the failure for the default mount option.

With that said, the test case failure needs to be fixed, and there are several bugs in the direct write path:

- Treat any short write as an error
This is especially common as we have disabled page faulting for reading from the @from iov_iter.

This is fixed by the first patch.

- No isize rollback after a short write
The isize is increased at btrfs_get_blocks_direct_write() but when a short write happened, the isize is not propoerly rolledback, causing later buffered fallback to write at the new isize.

This is fixed by the second patch.

- No page fault-in retry after a zero-submitted short write Unlike the previous two, this is only a minor problem which reduces the chance to do zero-copy direct IO.

This is fixed by the third patch.

Qu Wenruo (3):
btrfs: fix false IO failure after falling back to buffered write
btrfs: fix incorrect buffered IO fallback for append direct writes
btrfs: retry faulting in the pages after a zero sized short direct
write

 fs/btrfs/direct-io.c    | 62 +++++++++++++++++++++++++++++++++++++----
 fs/btrfs/inode.c        |  6 +---
 fs/btrfs/ordered-data.c | 12 ++++++++
 fs/btrfs/ordered-data.h |  2 ++
 4 files changed, 72 insertions(+), 10 deletions(-)

2.54.0

[BUG]
The test case generic/362 will fail with "nodatasum" mount option (*):

MOUNT_OPTIONS -- -o nodatasum /dev/mapper/test-scratch1 /mnt/scratch

generic/362 0s ... - output mismatch (see /home/adam/xfstests/results//generic/362.out.bad)

    --- tests/generic/362.out	2024-08-24 15:31:37.200000000 +0930
    +++ /home/adam/xfstests/results//generic/362.out.bad	2026-05-27 10:21:17.574771567 +0930
    @@ -1,2 +1,3 @@
     QA output created by 362
    +First write failed: Input/output error
     Silence is golden
    ...

*: If the test case has been executed before with default data checksum, the failure will not reproduce. Need the following fix to make it reliably reproducible:
https://lore.kernel.org/linux-btrfs/[email protected]/

[CAUSE]
Inside __iomap_dio_rw(), the -EFAULT/-ENOTBLK error is not directly returned. Thus we never got an error pointer from __iomap_dio_rw().

The call chain looks like this:

btrfs_direct_write()

 |- btrfs_dio_write()
 |-  __iomap_dio_rw()
 |  |- iomap_iter()
 |  |  |- btrfs_dio_iomap_begin()
 |  |     Now an ordered extent is allocated for the 4K write.
 |  |
 |  |- iomi.status = iomap_dio_iter()
 |  |  Where iomap_dio_iter() returned -EFAULT.
 |  |
 |  |- ret = iomap_iter()
 |  |  |- btrfs_dio_iomap_end()
 |  |  |  |- btrfs_finish_ordered_extent(uptodate = false)
 |  |  |  |  |- can_finish_ordered_extent()
 |  |  |  |     |- btrfs_mark_ordered_extent_error()
 |  |  |  |        |- mapping_set_error()
 |  |  |  |           Now the address space is marked error.
 |  |  |  | return -ENOTBLK
 |  |  |- return -ENOTBLK
 |  |- if (ret == -ENOTBLK) { ret = 0; }
 |     Now the return value is reset to 0.
 |     Thus no error pointer will be returned.
 |
 |- ret = iomap_dio_complete()
 |  Since no byte is submitted, @ret is 0.
 |
 |- Fallback to buffered IO
 |  And the buffered write finished without error
 |
 |- filemap_fdatawait_range()

|- filemap_check_errors()
The previous error is recorded, thus an error is returned

However the buffered write is properly submitted and finished, the error is from the btrfs_finish_ordered_extent() call with @uptodate = false.

[FIX]
When a short dio write happened, any range that is submitted will have btrfs_extract_ordered_extent() to be called, thus the submitted range will always have an OE just covering the submitted range.

The remaining OE range is never submitted, thus they should be treated as truncated, not an error. So that we can properly reclaim and not insert an unnecessary file extent item, without marking the mapping as error.

Extract a helper, btrfs_mark_ordered_extent_truncated(), and utilize that helper to mark the direct IO ordered extent as truncated, so it won't cause failure for the later buffered fallback.

[REASON FOR NO FIXES TAG]
The bug itself is pretty old, at commit f85781fb505e ("btrfs: switch to iomap for direct IO") we're already passing @uptodate=false finishing the OE.
But at that time OE with IOERR won't call mapping_set_error(), so it's not exposed.
Later commit d61bec08b904 ("btrfs: mark ordered extent and inode with error if we fail to finish") finally exposed the bug, but that commit is doing a correct job, not the root cause.

Anyway the bug is very old, dating back to 5.1x days, thus only CC to stable.

---
 fs/btrfs/direct-io.c    | 17 ++++++++++++++---
 fs/btrfs/inode.c        |  6 +-----
 fs/btrfs/ordered-data.c | 12 ++++++++++++
 fs/btrfs/ordered-data.h |  2 ++
 4 files changed, 29 insertions(+), 8 deletions(-)

index 57167d56dc72..88cb2e82a507 100644
@@ -624,12 +624,23 @@ static int btrfs_dio_iomap_end(struct inode *inode, loff_t pos, loff_t length,
 	if (submitted < length) {
 		pos += submitted;
 		length -= submitted;
-		if (write)
+		if (write) {
+			/*
+			 * We have a short write, if there is any range
+			 * that is submitted properly, that part will have
+			 * its own OE split from the original one.
+			 *
+			 * So for the OE at dio_data->ordered, it's the part
+			 * that is not submitted, and should be marked
+			 * as fully truncated.
+			 */
+			btrfs_mark_ordered_extent_truncated(dio_data->ordered, 0);
 			btrfs_finish_ordered_extent(dio_data->ordered,
-						    pos, length, false);
-		else
+						    pos, length, true);
+		} else {
 			btrfs_unlock_dio_extent(&BTRFS_I(inode)->io_tree, pos,
 						pos + length - 1, NULL);
+		}
 		ret = -ENOTBLK;
 	}
 	if (write) {

index 973a89301baa..2c0131452754 100644
@@ -7590,11 +7590,7 @@ static void btrfs_invalidate_folio(struct folio *folio, size_t offset,
 					       EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
 					       EXTENT_DEFRAG, &cached_state);
 
-		spin_lock(&inode->ordered_tree_lock);
-		set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
-		ordered->truncated_len = min(ordered->truncated_len,
-					     cur - ordered->file_offset);
-		spin_unlock(&inode->ordered_tree_lock);
+		btrfs_mark_ordered_extent_truncated(ordered, cur - ordered->file_offset);
 
 		/*
 		 * If the ordered extent has finished, we're safe to delete all

index f5f77c33cf59..b32d4eabe0ab 100644
@@ -358,6 +358,18 @@ void btrfs_mark_ordered_extent_error(struct btrfs_ordered_extent *ordered)
 		mapping_set_error(ordered->inode->vfs_inode.i_mapping, -EIO);
 }
 
+void btrfs_mark_ordered_extent_truncated(struct btrfs_ordered_extent *ordered,
+					 u64 truncate_len)
+{
+	struct btrfs_inode *inode = ordered->inode;
+
+	ASSERT(truncate_len <= ordered->num_bytes);
+	spin_lock(&inode->ordered_tree_lock);
+	set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
+	ordered->truncated_len = min(ordered->truncated_len, truncate_len);
+	spin_unlock(&inode->ordered_tree_lock);
+}
+
 static void finish_ordered_fn(struct btrfs_work *work)
 {
 	struct btrfs_ordered_extent *ordered_extent;

index 03e12380a2fd..8d5d5ba1e02f 100644
@@ -226,6 +226,8 @@ bool btrfs_try_lock_ordered_range(struct btrfs_inode *inode, u64 start, u64 end,
 struct btrfs_ordered_extent *btrfs_split_ordered_extent(
 			struct btrfs_ordered_extent *ordered, u64 len);
 void btrfs_mark_ordered_extent_error(struct btrfs_ordered_extent *ordered);
+void btrfs_mark_ordered_extent_truncated(struct btrfs_ordered_extent *ordered,
+					 u64 truncate_len);
 int __init ordered_data_init(void);
 void __cold ordered_data_exit(void);
 

2.54.0

[BUG]
With the previous bug of short direct writes fixed, test case generic/362 (*) still fails with the following error with nodatasum mount option:

generic/362 0s ... - output mismatch (see /home/adam/xfstests/results//generic/362.out.bad)
- output mismatch (see /home/adam/xfstests/results//generic/362.out.bad)

    --- tests/generic/362.out	2024-08-24 15:31:37.200000000 +0930
    +++ /home/adam/xfstests/results//generic/362.out.bad	2026-05-27 10:13:09.072485767 +0930
    @@ -1,2 +1,3 @@
     QA output created by 362
    +Wrong file size after first write, got 8192 expected 4096
     Silence is golden
    ...

*: If the test case has been executed before with default data checksum, the failure will not reproduce. Need the following fix to make it reliably reproducible:
https://lore.kernel.org/linux-btrfs/[email protected]/

[CAUSE]
Inside btrfs_dio_iomap_begin() for a direct write, we increase the isize if it's beyond the current isize.

But if the direct io finished short, we do not revert the isize to the previous value nor to the short write end.

Then if we need to fall back to buffered writes, and the write has IOCB_APPEND flag, then the buffered write will be positioned at the incorrect isize.

The call chain looks like this:

btrfs_direct_write(pos=0, length=4K)

 |- __iomap_dio_rw()
 |  |- iomap_iter()
 |  |  |- btrfs_dio_iomap_begin()
 |  |     |- btrfs_get_blocks_direct_write()
 |  |        |- i_size_write()
 |  |           Which updates the isize to the write end (4K).
 |  |
 |  |- iomap_dio_iter()
 |  |  Failed with -EFAULT on the first page.
 |  |
 |  |- iomap_iter()
 |  |  |- btrfs_dio_iomap_end()
 |  |     Detects a short write, return -ENOTBLK
 |  |- if (ret == -ENOTBLK) { ret = 0;}
 |     Which resets the return value.
 |
 |- ret = iomap_dio_complet()
 |  Which returns 0.
 |
 |- btrfs_buffered_write(iocb, from);

|- generic_write_checks()
|- iocb->ki_pos = i_size_read()
Which is still the new size (4K), other than the original isize 0.

[FIX]
Introduce the following btrfs_dio_data members:

- old_isize

- updated_isize
If the direct write has enlarged the isize.

Then if we got a short write, and btrfs_dio_data::updated_isize is set, revert to the correct isize based on old_isize and current file position.

And here we call i_size_write() without holding an extent lock, which is a very special case that we're safe to do:

- Only a single writer can be enlarging isize Enlarging isize will take the exclusive inode lock.

- Buffered readers need to wait for the OE we're holding Buffered readers will lock extent and wait for OE of the folio range. Sometimes we can skip the OE wait, but since all page cache is invalidated, the OE wait can not be skipped.

But I do not think this is the most elegant solution, nor covers all cases. E.g. if the bio is submitted but IO failed, we are unable to do the revert.

I believe the more elegant one would be extend the EXTENT_DIO_LOCKED lifespan for direct writes, so that we can update the isize when a write beyond EOF finished successfully.

However that change is too huge for a small bug fix. So only implement the minimal partial fix for now.

[REASON FOR NO FIXES TAG]
The bug is again very old, before commit f85781fb505e ("btrfs: switch to iomap for direct IO") we are already increasing isize without a proper rollback for short writes.

Thus only a CC to stable.

---
 fs/btrfs/direct-io.c | 43 ++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 42 insertions(+), 1 deletion(-)

index 88cb2e82a507..412309825d6f 100644
@@ -15,10 +15,12 @@
 
 struct btrfs_dio_data {
 	ssize_t submitted;
+	loff_t old_isize;
 	struct extent_changeset *data_reserved;
 	struct btrfs_ordered_extent *ordered;
 	bool data_space_reserved;
 	bool nocow_done;
+	bool updated_isize;
 };
 
 struct btrfs_dio_private {

@@ -228,6 +230,7 @@ static int btrfs_get_blocks_direct_write(struct extent_map **map,
 	bool space_reserved = false;
 	u64 len = *lenp;
 	u64 prev_len;
+	loff_t old_isize;
 	int ret = 0;
 
 	/*

@@ -341,8 +344,14 @@ static int btrfs_get_blocks_direct_write(struct extent_map **map,
 	 * Need to update the i_size under the extent lock so buffered
 	 * readers will get the updated i_size when we unlock.
 	 */
-	if (start + len > i_size_read(inode))
+	old_isize = i_size_read(inode);
+	if (start + len > old_isize) {
+		if (!dio_data->updated_isize) {
+			dio_data->old_isize = old_isize;
+			dio_data->updated_isize = true;
+		}
 		i_size_write(inode, start + len);
+	}
 out:
 	if (ret && space_reserved) {
 		btrfs_delalloc_release_extents(BTRFS_I(inode), len);

@@ -625,6 +634,38 @@ static int btrfs_dio_iomap_end(struct inode *inode, loff_t pos, loff_t length,
 		pos += submitted;
 		length -= submitted;
 		if (write) {
+			/*
+			 * Got a short write and have updated the isize, need to
+			 * revert the isize change.
+			 *
+			 * Normally we need to update isize with extent lock hold,
+			 * but we're safe due to the following factors:
+			 *
+			 * - Only a single writer can be enlarging isize
+			 *   Enlarging isize will take the exclusive inode lock.
+			 *
+			 * - Buffered readers need to wait for the OE we're holding
+			 *   Buffered readers will lock extent and wait for OE
+			 *   of the folio range, and since page cache is invalidated
+			 *   the OE wait can not be skipped.
+			 *
+			 * So here we are safe to revert the isize before
+			 * finishing the OE, and no reader of the remaining range
+			 * can see the enlarged size.
+			 *
+			 * TODO: Extend the DIO_LOCKED lifespan for direct writes,
+			 * and only enlarge isize after a successful write.
+			 */
+			if (dio_data->updated_isize) {
+				u64 new_isize;
+
+				if (submitted == 0)
+					new_isize = dio_data->old_isize;
+				else
+					new_isize = max(dio_data->old_isize, pos);
+				i_size_write(inode, new_isize);
+				dio_data->updated_isize = false;
+			}
 			/*
 			 * We have a short write, if there is any range
 			 * that is submitted properly, that part will have

2.54.0

Looks good to me, now, thanks!

Currently btrfs_direct_write() will not try to fault in the pages, but directly fall back to buffered writes, if the first page of the buffer can not be faulted in.

For example, during generic/362 with nodatasum mount option, there is a write at file offset 0, length PAGE_SIZE, and the page is not faulted in. Then we go the following callchain and directly fall back to buffered IO:

btrfs_direct_write()

 |- btrfs_dio_write()
 |-  __iomap_dio_rw()
 |  |- iomap_iter()
 |  |  |- btrfs_dio_iomap_begin()
 |  |     Now an ordered extent is allocated for the 4K write.
 |  |
 |  |- iomi.status = iomap_dio_iter()
 |  |  Where iomap_dio_iter() returned -EFAULT.
 |  |
 |  |- ret = iomap_iter()
 |  |  |- btrfs_dio_iomap_end()
 |  |  |  | return -ENOTBLK
 |  |  |- return -ENOTBLK
 |  |- if (ret == -ENOTBLK) { ret = 0; }
 |     Now the return value is reset to 0.
 |
 |- ret = iomap_dio_complete()
 |  Since no byte is submitted, @ret is now zero.
 |
 |- if (iov_iter_count() > 0 && (ret == -EFAULT || ret > 0))
 |  @ret is zero, thus not meeting the above retry condition
 |
 |- Fallback to buffered

Just slightly loosen the condition to allow retry faulting in pages after a zero sized short write.

Unlike the previous two bug fixes, this one is not really cause any real bug, but only reducing the chance to do zero-copy direct IO. Thus it doesn't really require stable-CC nor fixes-tag.

---
 fs/btrfs/direct-io.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index 412309825d6f..460326d34143 100644
@@ -978,7 +978,7 @@ ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
 	if (ret > 0)
 		written = ret;
 
-	if (iov_iter_count(from) > 0 && (ret == -EFAULT || ret > 0)) {
+	if (iov_iter_count(from) > 0 && (ret == -EFAULT || ret >= 0)) {
 		const size_t left = iov_iter_count(from);
 		/*
 		 * We have more data left to write. Try to fault in as many as

2.54.0