[RFC PATCH v3 00/11] Add buffered write-through support to iomap & xfs

[RFC PATCH v3 00/11] Add buffered write-through support to iomap & xfs

[Ojaswin Mujoo]

This is the next revision of writethrough series. I'll quote part of the
original cover:

    Hi all,

    This patchset implements an early design prototype of buffered I/O
    write-through semantics in linux.

    This idea mainly picked up traction to enable RWF_ATOMIC buffered IO,
    however write-through path can have many use cases beyond atomic writes,
    - such as enabling truly async AIO buffered I/O when issued with O_DSYNC
    - better scalability for buffered I/O

==============================================
** Changes since rfc v2 ** [1]
==============================================

1. Refined the error handling semantics:
----------------------------------------
Writethrough has 2 steps - memcpy to folio & IO to disk. Only the part of write
completely submitted to disk is considered successful. With that in mind we have
2 types of errors. Suppose we have 16K write and we encounter an error after 4k:
 a) If have only memcpy'd the 4k and we are able to submit it for IO, then we
    can return 4k as a short write, since everything is consistent.

 b) If we have memcpy'd the whole 16k but faced and error (like EIO) after the
    first 4k write, then we return the error to user. Since the page cache is in
    inconsistent state, the user must treat this as a buffered IO fsync failure.


2. Introduce RWF_NOSERIAL to allow parallel writes
--------------------------------------------------
As per our discussions in LSFMM 2026, we noticed that writethrough (v2
design) suffered from a big regression (~65%) in workloads with multiple
writers writing to a single file. This is because writethrough submits
IO within the inode lock and since buffered IO has an exclusive lock in
write, this hurt massively. In LSFMM, we discussed 3 approaches:
  a) Defer IO submission outside inode lock
  b) Avoid folio dirty - clear cycle
  c) Shared lock for writes

We tried a) by only staging prepared folios in a list under inode lock and then
submitting them outside. However, the initial implementation still showed
around ~30% regression even though the code complexity was significantly
higher. As the ROI was not worth it, we dropped this idea (if there is
interest I can share a github link for these patches). So we finally
decided to go with b) and c).

With b), we can avoid cycling folio through dirty and clear as we immediately
submit the IO. We also don't use the heavy folio_start/end_writeback() and
instead only use the writeback master bit on folio. This cuts back
significantly on xa_lock contention.

With c), the NOSERIAL flag allows us to do writes under a shared lock if
possible. This violates guarantees that XFS has historically provided however
its expected to be an advanced feature that should be used by applications who
know what they are doing. If used correctly, it can give a very good
performance boost to parallel workloads. This is something that was also
discussed at LSFMM 2026 [3] and since WRITETHROUGH is a new path it
seems like a good time to introduce such a semantic change . For now we
auto apply NOSERIAL for writethrough and disallow users to pass it till
we finalize the semantics.

With b & c, we are able to see a good performance improvement in almost all of
the cases that we regressing. More details and performance numbers specific to
the NOSERIAL flag can be found in the respective patches.


3. Use REQ_SYNC | REQ_IDLE like dio
-----------------------------------
Writethrough doesn't go via the writeback mechanism and it's IO characteristics
are similar to dio. Hence we pass REQ_SYNC | REQ_IDLE in the bio, just like
dio, which allows us to bypass writeback throttling.


4. Move inode i_size update from completion path to write path
--------------------------------------------------------------
As per Dave's suggestion we originally wanted to update isize in completion
like dio however this resulted in a big regression in extending IO because we
end up holding the exclusive lock throughout the IO. To avoid this, we can just
take the buffered IO approach of updating i_size in write path so that we can
safely drop the inode lock and allow completion to finish outside the lock.
This brings back the append IO performance in par with buffered IO.

5. There's a deadock in v2 that is fixed in the last patch. If needed
this can be squashed in but I've kept it separate for now for easier
review.

6. Added a patch to make stable writes more compatible to our usage
model. Check patch 1 for details. Thanks to Darrick for suggesting
this!.

7. Addressed reviews from Jan and Sashiko (thanks).

===================================================================
Performance Comparison Tables (with fio snippet)
===================================================================

Table 1: Extending writes using (libaio + O_DSYNC) - All writers on single file
+----------+-------------------+---------------------+
| numjobs  | Buffered IO       | Writethrough IO     |
+----------+-------------------+---------------------+
|    1     | 133 MiB/s         | 133 MiB/s (+0.0%)   |
|    2     | 179 MiB/s         | 243 MiB/s (+35.8%)  |
|    4     | 253 MiB/s         | 358 MiB/s (+41.5%)  |
|    8     | 366 MiB/s         | 376 MiB/s (+2.7%)   |
|   16     | 474 MiB/s         | 449 MiB/s (-5.3%)   |
+----------+-------------------+---------------------+
(fio --ioengine=libaio --writethrough=0/1 --bs=4k --rw=write \
     --iodepth=32 --sync=dsync--file_append=1)


Table 2: Random pure overwrites (libaio + O_DSYNC) - All writers on single file
+----------+-------------------+---------------------+
| numjobs  | Buffered IO       | Writethrough IO     |
+----------+-------------------+---------------------+
|    1     | 131 MiB/s         | 391 MiB/s (+198.5%) |
|    2     | 377 MiB/s         | 791 MiB/s (+109.8%) |
|    4     | 695 MiB/s         | 1591 MiB/s (+128.9%)|
|    8     | 1217 MiB/s        | 1846 MiB/s (+51.7%) |
|   16     | 1197 MiB/s        | 1844 MiB/s (+54.1%) |
+----------+-------------------+---------------------+
(fio --ioengine=libaio --bs=4k --size=2G --sync=dsync --writethrough=0/1 \
    --overwrite=1--rw=randwrite --iodepth=32)


Table 3: Random pure overwrites (libaio + O_DSYNC) - Each write writes own file
+----------+-------------------+---------------------+
| numjobs  | Buffered IO       | Writethrough IO     |
+----------+-------------------+---------------------+
|    1     | 187 MiB/s         | 389 MiB/s (+108.0%) |
|    2     | 373 MiB/s         | 781 MiB/s (+109.4%) |
|    4     | 706 MiB/s         | 1568 MiB/s (+122.1%)|
|    8     | 1222 MiB/s        | 1796 MiB/s (+47.0%) |
+----------+-------------------+---------------------+
(fio --ioengine=libaio --bs=4k --size=2G --filename_format=.test_file.\$jobnum \
     --sync=dsync --writethrough=0/1 --overwrite=1 --rw=randwrite --iodepth=32)


Table 4: Random 16KB overwrites with sync_file_range:16 - Single file
(Roughly mimics postgresql IO pattern)
+----------+-------------------+---------------------+
| numjobs  | Buffered IO       | Writethrough IO     |
+----------+-------------------+---------------------+
|    1     | 1323 MiB/s        | 1275 MiB/s (-3.6%)  |
|    2     | 1779 MiB/s        | 2434 MiB/s (+36.8%) |
|    4     | 2092 MiB/s        | 2519 MiB/s (+20.4%) |
|    8     | 2272 MiB/s        | 2517 MiB/s (+10.8%) |
|   16     | 2328 MiB/s        | 2525 MiB/s (+8.5%)  |
+----------+-------------------+---------------------+
(fio --ioengine=libaio --bs=16k --size=5G --writethrough=0/1 --overwrite=1\
     --sync_file_range=wait_before,write:16 --rw=randwrite --iodepth=32)

* Environment details *

CPU        : IBM Power 11 LPAR
Memory     : 62Gi
Storage    : Samsung PM173-series Enterprise NVMe SSD
Kernel     : Linux 7.2-rc1
Filesystem : XFS (4k block size)


[1] https://lore.kernel.org/linux-xfs/cover.1775658795.git.ojaswin@linux.ibm.com/
[2] https://github.com/OjaswinM/xfstests/tree/iomap-buf-writethrough2
[3] https://lwn.net/Articles/1072019

As usual, thoughts and suggestions are welcome :)

Regards,
ojaswin

Ojaswin Mujoo (11):
  fs: Add counter to track inflight writes that need stable pages
  mm: Refactor folio_clear_dirty_for_io()
  iomap: Add helper to revert iomap iter
  iomap: Add initial support for buffered RWF_WRITETHROUGH
  xfs: Add RWF_WRITETHROUGH support to xfs
  iomap: Add aio support to RWF_WRITETHROUGH
  iomap: Add DSYNC support to RWF_WRITETHROUGH
  fs: Introduce RWF_NOSERIAL flag to indicate parallel reads/writes
  xfs: Implement RWF_NOSERIAL to parallelize RWF_WRITETHROUGH writes
  iomap: Avoid folio dirtying in case of RWF_WRITETHROUGH
  iomap: Handle deadlock due to repeating folios in RWF_WRITETHROUGH

 fs/iomap/buffered-io.c  | 570 +++++++++++++++++++++++++++++++++++++++-
 fs/iomap/iter.c         |   9 +
 fs/xfs/xfs_file.c       | 135 +++++++++-
 include/linux/fs.h      |  29 ++
 include/linux/iomap.h   |  50 ++++
 include/linux/pagemap.h |  16 +-
 include/uapi/linux/fs.h |   9 +-
 mm/filemap.c            |  22 ++
 mm/page-writeback.c     |  49 +++-
 9 files changed, 861 insertions(+), 28 deletions(-)

--
2.55.0

[RFC PATCH v3 01/11] fs: Add counter to track inflight writes that need stable pages

[Ojaswin Mujoo]

The current flag-style stable write implementation uses idempotent set
and clear functions. This is okay because the users for the most part
just want to set or clear it once based on factors like underlying
device support.

However, this scheme doesn't play well when we have parallel users
wanting to temporarily set and unset stable writes. For example, the
upcoming RWF_WRITETHROUGH patches need stable writes to be enabled for
the duration of the IO. The current scheme can lead to bugs like:

RWF_WRITETHROUGH write 1             RWF_WRITETHROUGH write 2

enable stable write		     enable stable write
submit IO
disable stable write   <---- WRONG

                                     submit IO
                                     disable stable write

The 2nd write loses the stable write guarantee midway which is not
correct. Fix this by introducing a new inflight_stable_write counter
which can be used by parallel users safely.

Unfortunately, due to the way the current users are designed, we cannot
directly migrate them to the counter approach hence for now we will
have to keep both methods till all the users adapt to the counters.

Suggested-by: "Darrick J. Wong" <djwong@kernel.org>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 include/linux/fs.h      |  1 +
 include/linux/pagemap.h | 14 +++++++++++++-
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/include/linux/fs.h b/include/linux/fs.h
index 8e9bc9dda0cb..5f17aa0ed4c7 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -483,6 +483,7 @@ struct address_space {
 	errseq_t		wb_err;
 	spinlock_t		i_private_lock;
 	struct rw_semaphore	i_mmap_rwsem;
+	atomic_t		inflight_stable_writes_count;
 } __attribute__((aligned(sizeof(long)))) __randomize_layout;
 	/*
 	 * On most architectures that alignment is already the case; but
diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
index 2c3718d592d6..eb8b7e292478 100644
--- a/include/linux/pagemap.h
+++ b/include/linux/pagemap.h
@@ -306,7 +306,8 @@ static inline void mapping_clear_release_always(struct address_space *mapping)
 
 static inline bool mapping_stable_writes(const struct address_space *mapping)
 {
-	return test_bit(AS_STABLE_WRITES, &mapping->flags);
+	return test_bit(AS_STABLE_WRITES, &mapping->flags) ||
+	       atomic_read(&mapping->inflight_stable_writes_count) > 0;
 }
 
 static inline void mapping_set_stable_writes(struct address_space *mapping)
@@ -319,6 +320,17 @@ static inline void mapping_clear_stable_writes(struct address_space *mapping)
 	clear_bit(AS_STABLE_WRITES, &mapping->flags);
 }
 
+static inline void mapping_inc_inflight_stable_writes(struct address_space *mapping)
+{
+	atomic_inc(&mapping->inflight_stable_writes_count);
+}
+
+static inline void mapping_dec_inflight_stable_writes(struct address_space *mapping)
+{
+	WARN_ON_ONCE(atomic_read(&mapping->inflight_stable_writes_count) == 0);
+	atomic_dec_if_positive(&mapping->inflight_stable_writes_count);
+}
+
 static inline void mapping_set_inaccessible(struct address_space *mapping)
 {
 	/*
-- 
2.55.0

[RFC PATCH v3 02/11] mm: Refactor folio_clear_dirty_for_io()

[Ojaswin Mujoo]

Add a new __folio_clear_dirty_for_io() helper which takes an extra
parameter to indicate folio_mkclean() is needed. This is in preparation
of buffered writethrough support where we already do folio_mkclean()
before calling into this function.

Co-developed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 mm/page-writeback.c | 39 +++++++++++++++++++++++++--------------
 1 file changed, 25 insertions(+), 14 deletions(-)

diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index e98748112d1e..3d184ca316a8 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -2856,20 +2856,12 @@ void __folio_cancel_dirty(struct folio *folio)
 EXPORT_SYMBOL(__folio_cancel_dirty);
 
 /*
- * Clear a folio's dirty flag, while caring for dirty memory accounting.
- * Returns true if the folio was previously dirty.
- *
- * This is for preparing to put the folio under writeout.  We leave
- * the folio tagged as dirty in the xarray so that a concurrent
- * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
- * The ->writepage implementation will run either folio_start_writeback()
- * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
- * and xarray dirty tag back into sync.
- *
- * This incoherency between the folio's dirty flag and xarray tag is
- * unfortunate, but it only exists while the folio is locked.
+ * Internal helper to take care of clearing dirty bit on a folio in preparation
+ * of an IO. For some cases we might not want to do mkclean, eg, if we've
+ * already taken care of it, hence pass the should_mkclean flag to indicate if
+ * its needed.
  */
-bool folio_clear_dirty_for_io(struct folio *folio)
+static bool __folio_clear_dirty_for_io(struct folio *folio, bool should_mkclean)
 {
 	struct address_space *mapping = folio_mapping(folio);
 	bool ret = false;
@@ -2906,7 +2898,7 @@ bool folio_clear_dirty_for_io(struct folio *folio)
 		 * as a serialization point for all the different
 		 * threads doing their things.
 		 */
-		if (folio_mkclean(folio))
+		if (should_mkclean && folio_mkclean(folio))
 			folio_mark_dirty(folio);
 		/*
 		 * We carefully synchronise fault handlers against
@@ -2931,6 +2923,25 @@ bool folio_clear_dirty_for_io(struct folio *folio)
 	}
 	return folio_test_clear_dirty(folio);
 }
+
+/*
+ * Clear a folio's dirty flag, while caring for dirty memory accounting.
+ * Returns true if the folio was previously dirty.
+ *
+ * This is for preparing to put the folio under writeout.  We leave
+ * the folio tagged as dirty in the xarray so that a concurrent
+ * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
+ * The ->writepage implementation will run either folio_start_writeback()
+ * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
+ * and xarray dirty tag back into sync.
+ *
+ * This incoherency between the folio's dirty flag and xarray tag is
+ * unfortunate, but it only exists while the folio is locked.
+ */
+bool folio_clear_dirty_for_io(struct folio *folio)
+{
+	return __folio_clear_dirty_for_io(folio, true);
+}
 EXPORT_SYMBOL(folio_clear_dirty_for_io);
 
 static void wb_inode_writeback_start(struct bdi_writeback *wb)
-- 
2.55.0

[RFC PATCH v3 03/11] iomap: Add helper to revert iomap iter

[Ojaswin Mujoo]

Add helper to revert back the iomap iter by count bytes. This will be
used by upcoming RWF_WRITETHROUGH feature to revert the state of the
iomap in case an error is encountered.

Writethrough can process and park folios so that they can be batched
together for IO later, this also advances the iomap iter. However if we
encounter any error, we would want to cancel the IO on those processed
folios. In this case, we want to revert the iter back so the upper
layers can know how much we were actually able to write before failure.

Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 fs/iomap/iter.c       | 9 +++++++++
 include/linux/iomap.h | 1 +
 2 files changed, 10 insertions(+)

diff --git a/fs/iomap/iter.c b/fs/iomap/iter.c
index c445a38b6285..8ea849e7ccef 100644
--- a/fs/iomap/iter.c
+++ b/fs/iomap/iter.c
@@ -32,6 +32,15 @@ int iomap_iter_advance(struct iomap_iter *iter, u64 count)
 	return 0;
 }
 
+int iomap_iter_revert(struct iomap_iter *iter, u64 count)
+{
+	if (WARN_ON_ONCE(count > iter->pos))
+		return -EIO;
+	iter->pos -= count;
+	iter->len += count;
+	return 0;
+}
+
 static inline void iomap_iter_done(struct iomap_iter *iter)
 {
 	WARN_ON_ONCE(iter->iomap.offset > iter->pos);
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 8c754eb974fb..1c8f13948f81 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -275,6 +275,7 @@ struct iomap_iter {
 
 int iomap_iter(struct iomap_iter *iter, const struct iomap_ops *ops);
 int iomap_iter_advance(struct iomap_iter *iter, u64 count);
+int iomap_iter_revert(struct iomap_iter *iter, u64 count);
 
 /**
  * iomap_length_trim - trimmed length of the current iomap iteration
-- 
2.55.0

[RFC PATCH v3 04/11] iomap: Add initial support for buffered RWF_WRITETHROUGH

[Ojaswin Mujoo]

This adds initial support for performing buffered non-aio
RWF_WRITETHROUGH write. The rough flow for a writethrough write is as
follows:

1. Acquire inode lock
2. initialize writethrough context (wt_ctx) and mark
   mapping as stable.
3. Start the iomap_iter() loop. For each iomap:
  3.1. Acquire folio and folio_lock.
  3.2. perform memcpy from user buffer to the folio and mark it
       dirty
  3.3. Wait for any current writeback to complete and then call
       folio_mkclean() to prevent mmap writes from changing it.
  3.4. Start writeback on the folio
  3.5. Add the folio range under write to wt_ctx->bvec and folio_unlock()
  3.6. If bvec is full, submit the current bvecs for IO.
  3.7. Repeat 3.2 to 3.6 till the whole iomap is processed. Submit
       the final set of bvecs for IO.
4. Repeat step 3 till we have no more data to write.
5. Finally, sleep in the syscall thread till all the IOs are
   completed (refcount == 0). Once that happens, the end io handler will
   wake us up.
6. Upon waking up, call fs ->end_io() callback (which updates inode
   size), record any errors and return.
7. inode_unlock()

This design gives buffered writethrough the same semantics as dio and
any error in the IO is directly returned to the caller. However, the
users should note that an error should be treated equivalent of a
buffered IO fsync error, since we can't always guarantee that the page
cache is consistent.

The design has deliberately open coded the IO submission and completion
flow (inspired by dio) rather than reusing the dio functions as
accommodating buffered writethrough logic in dio code was polluting it
with too many if else conditionals and special cases.

Suggested-by: Jan Kara <jack@suse.cz>
Suggested-by: Dave Chinner <dgc@kernel.org>
Co-developed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 fs/iomap/buffered-io.c  | 437 ++++++++++++++++++++++++++++++++++++++++
 include/linux/fs.h      |  11 +
 include/linux/iomap.h   |  42 ++++
 include/linux/pagemap.h |   1 +
 include/uapi/linux/fs.h |   5 +-
 mm/page-writeback.c     |  10 +
 6 files changed, 505 insertions(+), 1 deletion(-)

diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index 0a5ebfda90f1..3178e8c0fa13 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -10,6 +10,7 @@
 #include <linux/migrate.h>
 #include <linux/fserror.h>
 #include <linux/fsverity.h>
+#include <linux/rmap.h>
 #include "internal.h"
 #include "trace.h"
 
@@ -1185,6 +1186,360 @@ static bool iomap_write_end(struct iomap_iter *iter, size_t len, size_t copied,
 	return __iomap_write_end(iter->inode, pos, len, copied, folio);
 }
 
+static ssize_t iomap_writethrough_complete(struct iomap_writethrough_ctx *wt_ctx)
+{
+	struct kiocb *iocb	= wt_ctx->iocb;
+	struct inode *inode	= wt_ctx->inode;
+	ssize_t ret		= wt_ctx->error;
+
+	if (wt_ctx->end_io) {
+		int err = wt_ctx->end_io(wt_ctx, wt_ctx->written,
+					 wt_ctx->error,
+					 wt_ctx->flags);
+		if (err)
+			ret = err;
+	}
+
+	mapping_dec_inflight_stable_writes(inode->i_mapping);
+
+	if (!ret) {
+		ret = wt_ctx->written;
+		iocb->ki_pos += ret;
+	}
+
+	kfree(wt_ctx);
+	return ret;
+}
+
+static void iomap_writethrough_done(struct iomap_writethrough_ctx *wt_ctx)
+{
+	struct task_struct *waiter = wt_ctx->waiter;
+
+	WRITE_ONCE(wt_ctx->waiter, NULL);
+	blk_wake_io_task(waiter);
+}
+
+static void iomap_writethrough_bio_end_io(struct bio *bio)
+{
+	struct iomap_writethrough_ctx *wt_ctx = bio->bi_private;
+	struct folio_iter fi;
+
+	if (bio->bi_status)
+		cmpxchg(&wt_ctx->error, 0,
+			blk_status_to_errno(bio->bi_status));
+	bio_for_each_folio_all(fi, bio)
+		folio_end_writeback(fi.folio);
+
+	bio_put(bio);
+	if (atomic_dec_and_test(&wt_ctx->ref))
+		iomap_writethrough_done(wt_ctx);
+}
+
+static int
+iomap_writethrough_submit_bio(struct iomap_writethrough_ctx *wt_ctx,
+			      struct iomap *iomap,
+			      const struct iomap_writethrough_ops *wt_ops, int error)
+{
+	struct bio *bio;
+	unsigned int i;
+	u64 len = 0;
+	blk_opf_t opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
+
+	if (!wt_ctx->nr_bvecs)
+		goto exit;
+
+	for (i = 0; i < wt_ctx->nr_bvecs; i++)
+		len += wt_ctx->bvec[i].bv_len;
+
+	bio = bio_alloc(iomap->bdev, wt_ctx->nr_bvecs, opf, GFP_NOFS);
+	bio->bi_iter.bi_sector	= iomap_sector(iomap, wt_ctx->bio_pos);
+	bio->bi_end_io		= iomap_writethrough_bio_end_io;
+	bio->bi_private		= wt_ctx;
+
+	for (i = 0; i < wt_ctx->nr_bvecs; i++)
+		__bio_add_page(bio, wt_ctx->bvec[i].bv_page,
+				wt_ctx->bvec[i].bv_len,
+				wt_ctx->bvec[i].bv_offset);
+
+	if (!error && wt_ops->writethrough_submit)
+		error = wt_ops->writethrough_submit(wt_ctx->inode, iomap,
+						    wt_ctx->bio_pos, len);
+
+
+	atomic_inc(&wt_ctx->ref);
+
+	/*
+	 * In case of error we still need the I/O completion to run so we can
+	 * release references and end writeback on the folios.
+	 */
+	if (error) {
+		bio->bi_status = errno_to_blk_status(error);
+		bio_endio(bio);
+		return error;
+	}
+
+	submit_bio(bio);
+	wt_ctx->nr_bvecs = 0;
+
+exit:
+	return 0;
+}
+
+/*
+ * Submit any pending bvecs as a bio and account the written bytes.
+ * On failure, iomap_writethrough_submit_bio() has already called the
+ * endio completion handler to record the error.
+ */
+static int
+iomap_writethrough_try_submit(struct iomap_writethrough_ctx *wt_ctx,
+			      struct iomap *iomap,
+			      const struct iomap_writethrough_ops *wt_ops,
+			      ssize_t *pending)
+{
+	int ret = iomap_writethrough_submit_bio(wt_ctx, iomap, wt_ops, 0);
+
+	if (ret < 0)
+		return ret;
+	wt_ctx->written += *pending;
+	*pending = 0;
+	return 0;
+}
+
+/**
+ * iomap_writethrough_begin - prepare the various structures for writethrough
+ * @folio: folio to prepare for writethrough
+ * @off: offset of write within folio
+ * @len: len of write within folio
+ *
+ * This function does the major preparation work needed before starting the
+ * writethrough. The main task is to prepare folio for writeththrough by blocking
+ * mmap writes and setting writeback on it. Further, we must clear the write range
+ * to non-dirty. If this results in the complete folio becoming non-dirty, then we
+ * need to clear the master dirty bit.
+ */
+static void iomap_folio_prepare_writethrough(struct folio *folio, size_t off,
+					     size_t len)
+{
+	bool fully_written;
+	u64 zero = 0;
+
+	if (folio_test_writeback(folio))
+		folio_wait_writeback(folio);
+
+	if (folio_mkclean(folio))
+		folio_mark_dirty(folio);
+
+	/*
+	 * We might either write through the complete folio or a partial folio
+	 * writethrough might result in all blocks becoming non-dirty, so we need to
+	 * check and mark the folio clean if that is the case.
+	 */
+	fully_written = (off == 0 && len == folio_size(folio));
+	iomap_clear_range_dirty(folio, off, len);
+	if (fully_written ||
+	    !iomap_find_dirty_range(folio, &zero, folio_size(folio)))
+		folio_clear_dirty_for_writethrough(folio);
+
+	folio_start_writeback(folio);
+}
+
+/**
+ * iomap_writethrough_iter - perform RWF_WRITETHROUGH buffered write
+ * @wt_ctx: writethrough context
+ * @iter: iomap iter holding mapping information
+ * @i: iov_iter for write
+ * @wt_ops: the fs callbacks needed for writethrough
+ *
+ * This function copies the user buffer to folio similar to usual buffered
+ * IO path, with the difference that we immediately issue the IO. For this we
+ * utilize IO submission and completion mechanism that is inspired by dio.
+ *
+ * Folio handling note: We might be writing through a partial folio so we need
+ * to be careful to not clear the folio dirty bit unless there are no dirty blocks
+ * in the folio after the writethrough.
+ */
+static int iomap_writethrough_iter(struct iomap_writethrough_ctx *wt_ctx,
+				   struct iomap_iter *iter, struct iov_iter *i,
+				   const struct iomap_writethrough_ops *wt_ops)
+
+{
+	ssize_t total_written = 0, pending = 0;
+	loff_t submit_start_pos;
+	int status = 0;
+	struct address_space *mapping = iter->inode->i_mapping;
+	size_t chunk = mapping_max_folio_size(mapping);
+	unsigned int bdp_flags = (iter->flags & IOMAP_NOWAIT) ? BDP_ASYNC : 0;
+	unsigned int bs = i_blocksize(iter->inode);
+
+	/* copied over based on how DIO handles these flags */
+	if (iter->iomap.type == IOMAP_UNWRITTEN)
+		wt_ctx->flags |= IOMAP_DIO_UNWRITTEN;
+	if (iter->iomap.flags & IOMAP_F_SHARED)
+		wt_ctx->flags |= IOMAP_DIO_COW;
+
+	if (!(iter->flags & IOMAP_WRITETHROUGH))
+		return -EINVAL;
+
+	/*
+	 * IOMAP_INLINE mappings have NULL bdev and would cause
+	 * iomap_sector() to dereference invalid memory. Reject them.
+	 */
+	if (iter->iomap.type == IOMAP_INLINE)
+		return -EINVAL;
+
+	do {
+		struct folio *folio;
+		size_t offset;		/* Offset into folio */
+		loff_t old_size;
+		u64 bytes;		/* Bytes to write to folio */
+		size_t copied;		/* Bytes copied from user */
+		u64 written;		/* Bytes have been written */
+		loff_t pos;
+		size_t off_aligned, len_aligned;
+
+		bytes = iov_iter_count(i);
+retry:
+		offset = iter->pos & (chunk - 1);
+		bytes = min(chunk - offset, bytes);
+		status = balance_dirty_pages_ratelimited_flags(mapping,
+							       bdp_flags);
+		if (unlikely(status))
+			break;
+
+		/*
+		 * If completions already occurred and reported errors, give up
+		 * now and don't bother submitting more bios.
+		 */
+		status = data_race(wt_ctx->error);
+		if (unlikely(status)) {
+			wt_ctx->nr_bvecs = 0;
+			break;
+		}
+
+		if (bytes > iomap_length(iter))
+			bytes = iomap_length(iter);
+
+		/*
+		 * Bring in the user page that we'll copy from _first_.
+		 * Otherwise there's a nasty deadlock on copying from the
+		 * same page as we're writing to, without it being marked
+		 * up-to-date.
+		 *
+		 * For async buffered writes the assumption is that the user
+		 * page has already been faulted in. This can be optimized by
+		 * faulting the user page.
+		 */
+		if (unlikely(fault_in_iov_iter_readable(i, bytes) == bytes)) {
+			status = -EFAULT;
+			break;
+		}
+
+		status = iomap_write_begin(iter, wt_ops->write_ops, &folio,
+					   &offset, &bytes);
+		if (unlikely(status)) {
+			iomap_write_failed(iter->inode, iter->pos, bytes);
+			break;
+		}
+		if (iter->iomap.flags & IOMAP_F_STALE)
+			break;
+
+		pos = iter->pos;
+
+		if (mapping_writably_mapped(mapping))
+			flush_dcache_folio(folio);
+
+		copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
+		written = iomap_write_end(iter, bytes, copied, folio) ?
+			  copied : 0;
+
+		old_size = iter->inode->i_size;
+		if (pos + written > old_size)
+			i_size_write(iter->inode, pos + written);
+
+		if (!written)
+			goto put_folio;
+
+		off_aligned = round_down(offset, bs);
+		len_aligned = round_up(offset + written, bs) - off_aligned;
+
+		iomap_folio_prepare_writethrough(folio, off_aligned,
+						 len_aligned);
+
+		if (!wt_ctx->nr_bvecs) {
+			wt_ctx->bio_pos = round_down(pos, bs);
+			submit_start_pos = pos;
+		}
+
+		bvec_set_folio(&wt_ctx->bvec[wt_ctx->nr_bvecs], folio,
+			       len_aligned, off_aligned);
+		wt_ctx->nr_bvecs++;
+
+put_folio:
+		__iomap_put_folio(iter, wt_ops->write_ops, written, folio);
+
+		if (old_size < pos)
+			pagecache_isize_extended(iter->inode, old_size, pos);
+
+		cond_resched();
+		if (unlikely(written == 0)) {
+			iomap_write_failed(iter->inode, pos, bytes);
+			iov_iter_revert(i, copied);
+
+			if (chunk > PAGE_SIZE)
+				chunk /= 2;
+			if (copied) {
+				bytes = copied;
+				goto retry;
+			}
+		} else {
+			total_written += written;
+			pending += written;
+			iomap_iter_advance(iter, written);
+		}
+
+		/*
+		 * If we fail to submit the bio, we immediately call the
+		 * IO completion handler that records the error. We
+		 * shall not retry anymore cause this could lead to
+		 * infinite loops in case of non-transient errors.
+		 */
+		if (wt_ctx->nr_bvecs == wt_ctx->max_bvecs) {
+			status = iomap_writethrough_try_submit(wt_ctx,
+					&iter->iomap, wt_ops, &pending);
+			if (status)
+				goto submit_failed;
+		}
+
+	} while (iov_iter_count(i) && iomap_length(iter));
+
+	if (wt_ctx->nr_bvecs) {
+		status = iomap_writethrough_try_submit(wt_ctx,
+				&iter->iomap, wt_ops, &pending);
+		if (status)
+			goto submit_failed;
+	}
+
+	/*
+	 * In case of an error, we only consider the bytes we were actually able
+	 * to submit IO for as valid data and revert the iters accordingly
+	 */
+	if (status) {
+		/*
+		 * we still need to run the endio completion for cleanup work
+		 * hence call the below helper to take care of it, if we haven't
+		 * already done so. We can ignore the return value here.
+		 */
+		iomap_writethrough_submit_bio(wt_ctx, &iter->iomap, wt_ops, status);
+
+submit_failed:
+		iomap_write_failed(iter->inode, submit_start_pos, pending);
+		iomap_iter_revert(iter, pending);
+		iov_iter_revert(i, pending);
+	}
+
+	return status;
+}
+
 static int iomap_write_iter(struct iomap_iter *iter, struct iov_iter *i,
 		const struct iomap_write_ops *write_ops)
 {
@@ -1345,6 +1700,88 @@ int iomap_fsverity_write(struct file *file, loff_t pos, size_t length,
 }
 EXPORT_SYMBOL_GPL(iomap_fsverity_write);
 
+ssize_t iomap_file_writethrough_write(struct kiocb *iocb, struct iov_iter *i,
+				      const struct iomap_writethrough_ops *wt_ops,
+				      void *private)
+{
+	struct inode *inode = iocb->ki_filp->f_mapping->host;
+	struct iomap_iter iter = {
+		.inode		= inode,
+		.pos		= iocb->ki_pos,
+		.len		= iov_iter_count(i),
+		.flags		= IOMAP_WRITE | IOMAP_WRITETHROUGH,
+		.private	= private,
+	};
+	struct iomap_writethrough_ctx *wt_ctx;
+	unsigned int max_bvecs;
+	ssize_t ret;
+	struct blk_plug plug;
+	size_t min_folio_bytes = PAGE_SIZE
+				 << mapping_min_folio_order(inode->i_mapping);
+
+	/*
+	 * For now we don't support any other flag with WRITETHROUGH
+	 */
+	if (!(iocb->ki_flags & IOCB_WRITETHROUGH))
+		return -EINVAL;
+	if (iocb->ki_flags & (IOCB_DONTCACHE))
+		return -EINVAL;
+	if (iocb_is_dsync(iocb))
+		/* D_SYNC support not implemented yet */
+		return -EOPNOTSUPP;
+	if (!is_sync_kiocb(iocb))
+		/* aio support not implemented yet */
+		return -EOPNOTSUPP;
+
+	/*
+	 * +1 to max bvecs to account for unaligned write spanning multiple
+	 * folios. Guard against overflow since iov_iter_count() returns size_t.
+	 */
+	max_bvecs = (unsigned int)min_t(
+		size_t, DIV_ROUND_UP(iov_iter_count(i), min_folio_bytes) + 1,
+		BIO_MAX_VECS);
+
+	wt_ctx = kzalloc(struct_size(wt_ctx, bvec, max_bvecs), GFP_NOFS);
+	if (!wt_ctx)
+		return -ENOMEM;
+
+	wt_ctx->iocb = iocb;
+	wt_ctx->inode = inode;
+	wt_ctx->end_io = wt_ops->end_io;
+	wt_ctx->old_i_size = i_size_read(inode);
+	wt_ctx->max_bvecs = max_bvecs;
+	atomic_set(&wt_ctx->ref, 1);
+	wt_ctx->waiter = current;
+
+	mapping_inc_inflight_stable_writes(inode->i_mapping);
+
+	blk_start_plug(&plug);
+
+	while ((ret = iomap_iter(&iter, wt_ops->ops)) > 0) {
+		WARN_ON(iter.iomap.type != IOMAP_UNWRITTEN &&
+			iter.iomap.type != IOMAP_MAPPED);
+		iter.status = iomap_writethrough_iter(wt_ctx, &iter, i, wt_ops);
+	}
+
+	blk_finish_plug(&plug);
+
+	if (ret < 0)
+		cmpxchg(&wt_ctx->error, 0, ret);
+
+	if (!atomic_dec_and_test(&wt_ctx->ref)) {
+		for (;;) {
+			set_current_state(TASK_UNINTERRUPTIBLE);
+			if (!READ_ONCE(wt_ctx->waiter))
+				break;
+			blk_io_schedule();
+		}
+		__set_current_state(TASK_RUNNING);
+	}
+
+	return iomap_writethrough_complete(wt_ctx);
+}
+EXPORT_SYMBOL_GPL(iomap_file_writethrough_write);
+
 static void iomap_write_delalloc_ifs_punch(struct inode *inode,
 		struct folio *folio, loff_t start_byte, loff_t end_byte,
 		struct iomap *iomap, iomap_punch_t punch)
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 5f17aa0ed4c7..bff09a5f90e8 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -344,6 +344,7 @@ struct readahead_control;
 #define IOCB_ATOMIC		(__force int) RWF_ATOMIC
 #define IOCB_DONTCACHE		(__force int) RWF_DONTCACHE
 #define IOCB_NOSIGNAL		(__force int) RWF_NOSIGNAL
+#define IOCB_WRITETHROUGH	(__force int) RWF_WRITETHROUGH
 
 /* non-RWF related bits - start at 16 */
 #define IOCB_EVENTFD		(1 << 16)
@@ -1980,6 +1981,8 @@ struct file_operations {
 #define FOP_ASYNC_LOCK		((__force fop_flags_t)(1 << 6))
 /* File system supports uncached read/write buffered IO */
 #define FOP_DONTCACHE		((__force fop_flags_t)(1 << 7))
+/* File system supports write through buffered IO */
+#define FOP_WRITETHROUGH	((__force fop_flags_t)(1 << 8))
 
 /* Wrap a directory iterator that needs exclusive inode access */
 int wrap_directory_iterator(struct file *, struct dir_context *,
@@ -3464,6 +3467,14 @@ static inline int kiocb_set_rw_flags(struct kiocb *ki, rwf_t flags,
 		if (IS_DAX(ki->ki_filp->f_mapping->host))
 			return -EOPNOTSUPP;
 	}
+	if (flags & RWF_WRITETHROUGH) {
+		/* file system must support it */
+		if (!(ki->ki_filp->f_op->fop_flags & FOP_WRITETHROUGH))
+			return -EOPNOTSUPP;
+		/* DAX mappings not supported */
+		if (IS_DAX(ki->ki_filp->f_mapping->host))
+			return -EOPNOTSUPP;
+	}
 	kiocb_flags |= (__force int) (flags & RWF_SUPPORTED);
 	if (flags & RWF_SYNC)
 		kiocb_flags |= IOCB_DSYNC;
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 1c8f13948f81..427a2763221c 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -212,6 +212,7 @@ struct iomap_write_ops {
 #endif /* CONFIG_FS_DAX */
 #define IOMAP_ATOMIC		(1 << 9) /* torn-write protection */
 #define IOMAP_DONTCACHE		(1 << 10)
+#define IOMAP_WRITETHROUGH	(1 << 11)
 
 /*
  * Return the existing mapping at pos, or reserve space starting at pos for up
@@ -561,6 +562,29 @@ struct iomap_writepage_ctx {
 	void			*wb_ctx;	/* pending writeback context */
 };
 
+struct iomap_writethrough_ctx {
+	struct kiocb		*iocb;
+	struct inode		*inode;
+	loff_t			old_i_size;
+	loff_t			new_i_size;
+	loff_t			pos;
+	size_t			written;
+	atomic_t		ref;
+	unsigned int		flags;
+	int			error;
+
+	/* used during submission and for non-aio completion */
+	struct task_struct	*waiter;
+	int (*end_io)(struct iomap_writethrough_ctx *wt_ctx, ssize_t size,
+		      int error, unsigned int flags);
+
+	loff_t			bio_pos;
+	unsigned int		nr_bvecs;
+	unsigned int		max_bvecs;
+	struct bio_vec		bvec[];
+
+};
+
 struct iomap_ioend *iomap_init_ioend(struct inode *inode, struct bio *bio,
 		loff_t file_offset, u16 ioend_flags);
 struct iomap_ioend *iomap_split_ioend(struct iomap_ioend *ioend,
@@ -751,6 +775,24 @@ static __always_inline ssize_t iomap_dio_read_simple(struct kiocb *iocb,
 	return __iomap_dio_read_simple(iocb, iter, &iomi);
 }
 
+/*
+ * In writethrough, we copy user data to folio first and then send the folio
+ * to writeback via dio path. To achieve this, we need callbacks from iomap_ops
+ * and iomap_write_ops. This struct packs them together.
+ */
+struct iomap_writethrough_ops {
+	const struct iomap_ops *ops;
+	const struct iomap_write_ops *write_ops;
+	int (*writethrough_submit)(struct inode *inode, struct iomap *iomap,
+				   loff_t offset, u64 len);
+	int (*end_io)(struct iomap_writethrough_ctx *iocb, ssize_t size,
+		      int error, unsigned int flags);
+};
+
+ssize_t iomap_file_writethrough_write(struct kiocb *iocb, struct iov_iter *i,
+				      const struct iomap_writethrough_ops *wt_ops,
+				      void *private);
+
 #ifdef CONFIG_SWAP
 struct file;
 struct swap_info_struct;
diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
index eb8b7e292478..b20e38cc0fa0 100644
--- a/include/linux/pagemap.h
+++ b/include/linux/pagemap.h
@@ -1270,6 +1270,7 @@ static inline void folio_cancel_dirty(struct folio *folio)
 		__folio_cancel_dirty(folio);
 }
 bool folio_clear_dirty_for_io(struct folio *folio);
+bool folio_clear_dirty_for_writethrough(struct folio *folio);
 bool clear_page_dirty_for_io(struct page *page);
 void folio_invalidate(struct folio *folio, size_t offset, size_t length);
 bool noop_dirty_folio(struct address_space *mapping, struct folio *folio);
diff --git a/include/uapi/linux/fs.h b/include/uapi/linux/fs.h
index bd87262f2e34..9c8d91b926a7 100644
--- a/include/uapi/linux/fs.h
+++ b/include/uapi/linux/fs.h
@@ -451,10 +451,13 @@ typedef int __bitwise __kernel_rwf_t;
 /* prevent pipe and socket writes from raising SIGPIPE */
 #define RWF_NOSIGNAL	((__force __kernel_rwf_t)0x00000100)
 
+/* buffered IO that is asynchronously written through to disk after write */
+#define RWF_WRITETHROUGH	((__force __kernel_rwf_t)0x00000200)
+
 /* mask of flags supported by the kernel */
 #define RWF_SUPPORTED	(RWF_HIPRI | RWF_DSYNC | RWF_SYNC | RWF_NOWAIT |\
 			 RWF_APPEND | RWF_NOAPPEND | RWF_ATOMIC |\
-			 RWF_DONTCACHE | RWF_NOSIGNAL)
+			 RWF_DONTCACHE | RWF_NOSIGNAL | RWF_WRITETHROUGH)
 
 #define PROCFS_IOCTL_MAGIC 'f'
 
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index 3d184ca316a8..d4f2bb60cb47 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -2944,6 +2944,16 @@ bool folio_clear_dirty_for_io(struct folio *folio)
 }
 EXPORT_SYMBOL(folio_clear_dirty_for_io);
 
+/*
+ * Clear folio dirty in preparation of writethrough. Note that for writethrough
+ * we have already done folkio_mkclean so we avoid it here
+ */
+bool folio_clear_dirty_for_writethrough(struct folio *folio)
+{
+	return __folio_clear_dirty_for_io(folio, false);
+}
+EXPORT_SYMBOL(folio_clear_dirty_for_writethrough);
+
 static void wb_inode_writeback_start(struct bdi_writeback *wb)
 {
 	atomic_inc(&wb->writeback_inodes);
-- 
2.55.0

[RFC PATCH v3 05/11] xfs: Add RWF_WRITETHROUGH support to xfs

[Ojaswin Mujoo]

Add the boilerplate needed to start supporting RWF_WRITETHROUGH in XFS.
We use the direct write ->iomap_begin() functions to ensure the range
under write through always has a real non-delalloc extent. We reuse the xfs
dio's end IO function to perform extent conversion and i_size handling
for us.

*Note on COW extent over DATA hole case*

In case of an unmapped COW extent over a DATA hole
(due to COW preallocations), leave the extent unmapped until we are just
about to send IO. At that time, use the ->writethrough_submit() call
back to convert the COW extent to written.

We initially tried converting during iomap begin() time (like dio does)
but that results in a stale data exposure as follows:

1. iomap begin() - converts COW extent over DATA hole to written and
marks IOMAP_F_NEW to handle zeroing.
2. During iomap_write_begin() -> realise extent is stale and return back
without zeroing.
3. iomap begin() - Again sees the same COW extent but it's written
this time so we don't mark IOMAP_F_NEW
4. Since IOMAP_F_NEW is unmarked, we never zeroout and hence expose
stale data.

To avoid the above, take the buffered IO approach of converting the
extent just before IO, when we are sure to have zeroed out the folio.

Co-developed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 fs/xfs/xfs_file.c | 83 +++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 77 insertions(+), 6 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index 768cabf6250b..4b45ceacd461 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -702,6 +702,36 @@ static const struct iomap_dio_ops xfs_dio_write_ops = {
 	.end_io		= xfs_dio_write_end_io,
 };
 
+static int
+xfs_writethrough_end_io(
+	struct iomap_writethrough_ctx	*wt_ctx,
+	ssize_t				size,
+	int				error,
+	unsigned int			flags)
+{
+	struct xfs_inode *ip = XFS_I(wt_ctx->inode);
+	xfs_off_t offset = wt_ctx->iocb->ki_pos;
+
+	if (unlikely(error)) {
+		if (wt_ctx->flags & IOMAP_DIO_COW)
+			xfs_reflink_cancel_cow_range(ip, offset, size, true);
+
+		return error;
+	}
+
+	/*
+	 * writethrough completions are handled same as dio with the exception
+	 * that we need to explicitly change the i_disk_size. This is because
+	 * unlike dio, we have already updated the i_size and hence the
+	 * (i_disk_size < i_size) check will fail in dio code
+	 */
+	xfs_dio_write_end_io(wt_ctx->iocb, size, error, flags);
+	if (offset + size > ip->i_disk_size)
+		return xfs_setfilesize(ip, offset, size);
+
+	return 0;
+}
+
 static void
 xfs_dio_zoned_submit_io(
 	const struct iomap_iter	*iter,
@@ -1033,6 +1063,39 @@ xfs_file_dax_write(
 	return ret;
 }
 
+static int
+xfs_writethrough_submit(
+	struct inode		*inode,
+	struct iomap		*iomap,
+	loff_t			offset,
+	u64			count)
+{
+	int error = 0;
+	unsigned int		nofs_flag;
+
+	/*
+	 * Convert CoW extents to regular.
+	 *
+	 * We are under writethrough context with folio lock possibly held. To
+	 * avoid memory allocation deadlocks, set the task-wide nofs context.
+	 */
+	if (iomap->flags & IOMAP_F_SHARED) {
+		nofs_flag = memalloc_nofs_save();
+		error = xfs_reflink_convert_cow(XFS_I(inode), offset, count);
+		memalloc_nofs_restore(nofs_flag);
+	}
+
+	return error;
+}
+
+const struct iomap_writethrough_ops xfs_writethrough_ops = {
+	.ops			= &xfs_direct_write_iomap_ops,
+	.write_ops		= &xfs_iomap_write_ops,
+	.end_io		= xfs_writethrough_end_io,
+	.writethrough_submit	= &xfs_writethrough_submit
+};
+
+
 STATIC ssize_t
 xfs_file_buffered_write(
 	struct kiocb		*iocb,
@@ -1055,9 +1118,13 @@ xfs_file_buffered_write(
 		goto out;
 
 	trace_xfs_file_buffered_write(iocb, from);
-	ret = iomap_file_buffered_write(iocb, from,
-			&xfs_buffered_write_iomap_ops, &xfs_iomap_write_ops,
-			NULL);
+	if (iocb->ki_flags & IOCB_WRITETHROUGH) {
+		ret = iomap_file_writethrough_write(iocb, from,
+						    &xfs_writethrough_ops, NULL);
+	} else
+		ret = iomap_file_buffered_write(iocb, from,
+						&xfs_buffered_write_iomap_ops,
+						&xfs_iomap_write_ops, NULL);
 
 	/*
 	 * If we hit a space limit, try to free up some lingering preallocated
@@ -1092,8 +1159,12 @@ xfs_file_buffered_write(
 
 	if (ret > 0) {
 		XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
-		/* Handle various SYNC-type writes */
-		ret = generic_write_sync(iocb, ret);
+		/*
+		 * Handle various SYNC-type writes.
+		 * For writethrough, we handle sync during completion.
+		 */
+		if (!(iocb->ki_flags & IOCB_WRITETHROUGH))
+			ret = generic_write_sync(iocb, ret);
 	}
 	return ret;
 }
@@ -2104,7 +2175,7 @@ const struct file_operations xfs_file_operations = {
 	.remap_file_range = xfs_file_remap_range,
 	.fop_flags	= FOP_MMAP_SYNC | FOP_BUFFER_RASYNC |
 			  FOP_BUFFER_WASYNC | FOP_DIO_PARALLEL_WRITE |
-			  FOP_DONTCACHE,
+			  FOP_DONTCACHE | FOP_WRITETHROUGH,
 	.setlease	= generic_setlease,
 };
 
-- 
2.55.0

[RFC PATCH v3 06/11] iomap: Add aio support to RWF_WRITETHROUGH

[Ojaswin Mujoo]

With aio the only thing we need to be careful of is that writethrough
can be in progress even after dropping inode and folio lock. Due to
this, we need a way to synchronise with other paths where stable write
is not enough, example:

1. Truncate to 0 in xfs sets i_size = 0 before waiting for writeback to
   complete. In case of writethrough, the end io completion can again
   push the i_size to a non-zero value.
2. Dio reads might race with aio writethrough ->end_io() and read 0s if
   unwritten conversion is yet to happen.

Hence use the dio begin/end as it gives us the required guarantees.

Co-developed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 fs/iomap/buffered-io.c | 54 ++++++++++++++++++++++++++++++++++++------
 include/linux/iomap.h  | 11 +++++++--
 2 files changed, 56 insertions(+), 9 deletions(-)

diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index 3178e8c0fa13..22e4252dff4d 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -1202,6 +1202,9 @@ static ssize_t iomap_writethrough_complete(struct iomap_writethrough_ctx *wt_ctx
 
 	mapping_dec_inflight_stable_writes(inode->i_mapping);
 
+	if (wt_ctx->is_aio)
+		inode_dio_end(inode);
+
 	if (!ret) {
 		ret = wt_ctx->written;
 		iocb->ki_pos += ret;
@@ -1211,12 +1214,27 @@ static ssize_t iomap_writethrough_complete(struct iomap_writethrough_ctx *wt_ctx
 	return ret;
 }
 
+static void iomap_writethrough_complete_work(struct work_struct *work)
+{
+	struct iomap_writethrough_ctx *wt_ctx =
+		container_of(work, struct iomap_writethrough_ctx, aio_work);
+	struct kiocb *iocb = wt_ctx->iocb;
+
+	iocb->ki_complete(iocb, iomap_writethrough_complete(wt_ctx));
+}
+
 static void iomap_writethrough_done(struct iomap_writethrough_ctx *wt_ctx)
 {
-	struct task_struct *waiter = wt_ctx->waiter;
+	if (!wt_ctx->is_aio) {
+		struct task_struct *waiter = wt_ctx->waiter;
 
-	WRITE_ONCE(wt_ctx->waiter, NULL);
-	blk_wake_io_task(waiter);
+		WRITE_ONCE(wt_ctx->waiter, NULL);
+		blk_wake_io_task(waiter);
+		return;
+	}
+
+	INIT_WORK(&wt_ctx->aio_work, iomap_writethrough_complete_work);
+	queue_work(wt_ctx->inode->i_sb->s_dio_done_wq, &wt_ctx->aio_work);
 }
 
 static void iomap_writethrough_bio_end_io(struct bio *bio)
@@ -1729,9 +1747,6 @@ ssize_t iomap_file_writethrough_write(struct kiocb *iocb, struct iov_iter *i,
 	if (iocb_is_dsync(iocb))
 		/* D_SYNC support not implemented yet */
 		return -EOPNOTSUPP;
-	if (!is_sync_kiocb(iocb))
-		/* aio support not implemented yet */
-		return -EOPNOTSUPP;
 
 	/*
 	 * +1 to max bvecs to account for unaligned write spanning multiple
@@ -1750,11 +1765,33 @@ ssize_t iomap_file_writethrough_write(struct kiocb *iocb, struct iov_iter *i,
 	wt_ctx->end_io = wt_ops->end_io;
 	wt_ctx->old_i_size = i_size_read(inode);
 	wt_ctx->max_bvecs = max_bvecs;
+	wt_ctx->is_aio = !is_sync_kiocb(iocb);
 	atomic_set(&wt_ctx->ref, 1);
-	wt_ctx->waiter = current;
+
+	if (!wt_ctx->is_aio)
+		wt_ctx->waiter = current;
+	else
+		/*
+		 * With aio, writethrough can be in progress even after dropping
+		 * inode and folio lock. Due to this, we need a way to
+		 * synchronise with other paths where stable write is not enough
+		 * (example truncate). Hence use the dio begin/end as it gives
+		 * us the required guarantees.
+		 */
+		inode_dio_begin(inode);
 
 	mapping_inc_inflight_stable_writes(inode->i_mapping);
 
+	if (wt_ctx->is_aio && !inode->i_sb->s_dio_done_wq) {
+		ret = sb_init_dio_done_wq(inode->i_sb);
+		if (ret < 0) {
+			mapping_dec_inflight_stable_writes(inode->i_mapping);
+			inode_dio_end(inode);
+			kfree(wt_ctx);
+			return ret;
+		}
+	}
+
 	blk_start_plug(&plug);
 
 	while ((ret = iomap_iter(&iter, wt_ops->ops)) > 0) {
@@ -1769,6 +1806,9 @@ ssize_t iomap_file_writethrough_write(struct kiocb *iocb, struct iov_iter *i,
 		cmpxchg(&wt_ctx->error, 0, ret);
 
 	if (!atomic_dec_and_test(&wt_ctx->ref)) {
+		if (wt_ctx->is_aio)
+			return -EIOCBQUEUED;
+
 		for (;;) {
 			set_current_state(TASK_UNINTERRUPTIBLE);
 			if (!READ_ONCE(wt_ctx->waiter))
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 427a2763221c..7203c4d92170 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -572,9 +572,16 @@ struct iomap_writethrough_ctx {
 	atomic_t		ref;
 	unsigned int		flags;
 	int			error;
+	bool			is_aio;
+
+	union {
+		/* used during submission and for non-aio completion */
+		struct task_struct	*waiter;
+
+		/* used during aio completion */
+		struct work_struct	aio_work;
+	};
 
-	/* used during submission and for non-aio completion */
-	struct task_struct	*waiter;
 	int (*end_io)(struct iomap_writethrough_ctx *wt_ctx, ssize_t size,
 		      int error, unsigned int flags);
 
-- 
2.55.0

[RFC PATCH v3 07/11] iomap: Add DSYNC support to RWF_WRITETHROUGH

[Ojaswin Mujoo]

Add DSYNC support to writethrough buffered writes. Unlike the usual
buffered writes where we call generic_write_sync() inline during the
syscall path, for writethrough we instead sync the data during IO
completion path, just like dio.

This allows aio writethrough to be truly async where the syscall can
return after IO submission and the sync can then be done asynchronously
during IO completion time.

Further, just like dio, we utilize the FUA optimization, if available,
to avoid syncing the data for DSYNC operations.

Suggested-by: Dave Chinner <dgc@kernel.org>
Co-developed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 fs/iomap/buffered-io.c | 34 +++++++++++++++++++++++++++++++---
 include/linux/iomap.h  |  1 +
 2 files changed, 32 insertions(+), 3 deletions(-)

diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index 22e4252dff4d..d16694dd9995 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -1208,6 +1208,14 @@ static ssize_t iomap_writethrough_complete(struct iomap_writethrough_ctx *wt_ctx
 	if (!ret) {
 		ret = wt_ctx->written;
 		iocb->ki_pos += ret;
+
+		/*
+		 * If this is a DSYNC write and we couldn't optimize it, make
+		 * sure we push it to stable storage now that we've written
+		 * data.
+		 */
+		if (iocb_is_dsync(wt_ctx->iocb) && !wt_ctx->use_fua)
+			ret = generic_write_sync(iocb, ret);
 	}
 
 	kfree(wt_ctx);
@@ -1269,6 +1277,9 @@ iomap_writethrough_submit_bio(struct iomap_writethrough_ctx *wt_ctx,
 	for (i = 0; i < wt_ctx->nr_bvecs; i++)
 		len += wt_ctx->bvec[i].bv_len;
 
+	if (wt_ctx->use_fua)
+		opf |= REQ_FUA;
+
 	bio = bio_alloc(iomap->bdev, wt_ctx->nr_bvecs, opf, GFP_NOFS);
 	bio->bi_iter.bi_sector	= iomap_sector(iomap, wt_ctx->bio_pos);
 	bio->bi_end_io		= iomap_writethrough_bio_end_io;
@@ -1405,6 +1416,19 @@ static int iomap_writethrough_iter(struct iomap_writethrough_ctx *wt_ctx,
 	if (iter->iomap.type == IOMAP_INLINE)
 		return -EINVAL;
 
+	/*
+	 * If we realise that cache flush is necessary (eg FUA is not present
+	 * or we need metadata updates) then we turn off the optimization.
+	 */
+	if (wt_ctx->use_fua) {
+		if (iter->iomap.type != IOMAP_MAPPED ||
+		    (iter->iomap.flags &
+		     (IOMAP_F_NEW | IOMAP_F_SHARED | IOMAP_F_DIRTY)) ||
+		    (bdev_write_cache(iter->iomap.bdev) &&
+		     !bdev_fua(iter->iomap.bdev)))
+			wt_ctx->use_fua = false;
+	}
+
 	do {
 		struct folio *folio;
 		size_t offset;		/* Offset into folio */
@@ -1744,9 +1768,6 @@ ssize_t iomap_file_writethrough_write(struct kiocb *iocb, struct iov_iter *i,
 		return -EINVAL;
 	if (iocb->ki_flags & (IOCB_DONTCACHE))
 		return -EINVAL;
-	if (iocb_is_dsync(iocb))
-		/* D_SYNC support not implemented yet */
-		return -EOPNOTSUPP;
 
 	/*
 	 * +1 to max bvecs to account for unaligned write spanning multiple
@@ -1768,6 +1789,13 @@ ssize_t iomap_file_writethrough_write(struct kiocb *iocb, struct iov_iter *i,
 	wt_ctx->is_aio = !is_sync_kiocb(iocb);
 	atomic_set(&wt_ctx->ref, 1);
 
+	/*
+	 * Similar to dio, we optimistically set use_fua=true to avoid explicit
+	 * sync. In case we later realise cache flush is needed we set it back
+	 * to false.
+	 */
+	wt_ctx->use_fua = iocb_is_dsync(iocb) && !(iocb->ki_flags & IOCB_SYNC);
+
 	if (!wt_ctx->is_aio)
 		wt_ctx->waiter = current;
 	else
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 7203c4d92170..ba510d02c508 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -573,6 +573,7 @@ struct iomap_writethrough_ctx {
 	unsigned int		flags;
 	int			error;
 	bool			is_aio;
+	bool			use_fua;
 
 	union {
 		/* used during submission and for non-aio completion */
-- 
2.55.0

[RFC PATCH v3 08/11] fs: Introduce RWF_NOSERIAL flag to indicate parallel reads/writes

[Ojaswin Mujoo]

Introduce RWF_NOSERIAL flag to indicate that the application is
okay with its reads and writes going in parallel to other read/writes.
This flag will allow writes and read to go in parallel (for eg, under a
shared lock) increasing performance at the cost of losing the (loosely
implemented) POSIX guarantees wrt to R/W serialization that various
filesystems provide.

In this patch we just introduce the flag and in upcoming patches we will
use it to implement parallel writes to increase performance of
RWF_WRITETHROUGH

Co-developed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 include/linux/fs.h      | 10 ++++++++++
 include/uapi/linux/fs.h |  6 +++++-
 2 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/include/linux/fs.h b/include/linux/fs.h
index bff09a5f90e8..685ffe8da6ea 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -345,6 +345,7 @@ struct readahead_control;
 #define IOCB_DONTCACHE		(__force int) RWF_DONTCACHE
 #define IOCB_NOSIGNAL		(__force int) RWF_NOSIGNAL
 #define IOCB_WRITETHROUGH	(__force int) RWF_WRITETHROUGH
+#define IOCB_NOSERIAL		(__force int) RWF_NOSERIAL
 
 /* non-RWF related bits - start at 16 */
 #define IOCB_EVENTFD		(1 << 16)
@@ -3475,6 +3476,15 @@ static inline int kiocb_set_rw_flags(struct kiocb *ki, rwf_t flags,
 		if (IS_DAX(ki->ki_filp->f_mapping->host))
 			return -EOPNOTSUPP;
 	}
+
+	/*
+	 * For now we don't allow users to pass the NOSERIAL flag
+	 * TODO: Once the semantics of this flag are finalized we can lift this
+	 * restriction.
+	 */
+	if (flags & IOCB_NOSERIAL)
+		return -EOPNOTSUPP;
+
 	kiocb_flags |= (__force int) (flags & RWF_SUPPORTED);
 	if (flags & RWF_SYNC)
 		kiocb_flags |= IOCB_DSYNC;
diff --git a/include/uapi/linux/fs.h b/include/uapi/linux/fs.h
index 9c8d91b926a7..cde8c9492da6 100644
--- a/include/uapi/linux/fs.h
+++ b/include/uapi/linux/fs.h
@@ -454,10 +454,14 @@ typedef int __bitwise __kernel_rwf_t;
 /* buffered IO that is asynchronously written through to disk after write */
 #define RWF_WRITETHROUGH	((__force __kernel_rwf_t)0x00000200)
 
+/* buffered IO writes that are non sequential because they use a shared lock */
+#define RWF_NOSERIAL	((__force __kernel_rwf_t)0x00000400)
+
 /* mask of flags supported by the kernel */
 #define RWF_SUPPORTED	(RWF_HIPRI | RWF_DSYNC | RWF_SYNC | RWF_NOWAIT |\
 			 RWF_APPEND | RWF_NOAPPEND | RWF_ATOMIC |\
-			 RWF_DONTCACHE | RWF_NOSIGNAL | RWF_WRITETHROUGH)
+			 RWF_DONTCACHE | RWF_NOSIGNAL | RWF_WRITETHROUGH |\
+			 RWF_NOSERIAL)
 
 #define PROCFS_IOCTL_MAGIC 'f'
 
-- 
2.55.0

[RFC PATCH v3 09/11] xfs: Implement RWF_NOSERIAL to parallelize RWF_WRITETHROUGH writes

[Ojaswin Mujoo]

In xfs, buffered writethrough writes take an exclusive inode lock
similar to regular buffered write path. However, since writethrough also
submits the write under the inode lock, the increased critical section
really hurts performance when we have parallel writers writing to the
same file.

To mitigate this, implement RWF_NOSERIAL flag which allows us to perform
the write under a shared inode lock, instead of an exclusive lock. This
gives significant performance boost to single file, multiple writer
workloads at the cost of losing write-write serialization and read-write
serialization guarantees which XFS has historically provided.

Let's look at each of the guarantees and what changes with the NOSERIAL
flag:

Write-write guarantee:

Image 2 writers trying to write the same 3 folios. One is write As to
all 3 folios (denoted by AAA) and the other is writing BBB. Then under
exclusive lock the final state of the 3 folios could only be either AAA
or BBB.

However, with NOSERIAL writes, we could end up with mixed data in the
3 folios, like AAB, ABA, BAA etc. Note that this mixing will always
happen at the inter folio level, data contained within the same folio
will not be mixed as it is protected by the folio lock.

Read-write guarantee:

In XFS, reads also take a shared lock, however they don't take a folio
lock when copying from the folio to the user buffer. Due to the shared
lock of read and exclusive lock of write, we are able to guarantee that
the read always reads either completely old data or completely new data.

However, with the NOSERIAL flag, writethrough will take a shared lock for
writes, ie a read can race with a write which is still in middle of
copying data to the folio, hence the read can read a mix of old and new
data.

Despite the above changes in behavior, there might be applications
who would be okay to lose the guarantees because of the nature of their
workloads for example, if they never have multiple readers/writes doing
IO to the same range in the file. Such applications would benefit
significantly by using NOSERIAL writes.

Below are some fio performance numbers of RWF_WRITETHROUGH with and without
NOSERIAL writes.

** Multiple writes, single file (Pure overwrites, DSYNC) **
  Fio Workload:
    libaio, buffered randwrite, bs=4k, size=2G (pre written), O_DSYNC
    iodepth=32

  numjobs  baseline BW  RWF_NOSERIAL BW  Δ%
     1        350           392         +12.0%
     2        526           798         +51.7%
     4        597          1591        +166.5%
     8        630          1839        +191.9%
    16        570          1836        +222.1%

** Multiple writes, single file (Preallocated file, sync_file_range) **
  Fio Workload:
    libaio, buffered randwrite, bs=16k, size=5G (fallocated)
    sync_file_range=wait_before,write:16

  numjobs  baseline BW  RWF_NOSERIAL BW  Δ%
     1       1099          1279        +16.4%
     2       1482          2456        +65.7%
     4       2065          2479        +20.1%
     8       1778          2500        +40.6%
    16       1787          2508        +40.3%

* Multiple writes, single file (Truncated file, DSYNC) *
  Fio Workload:
    libaio, buffered randwrite, bs=4k, size=6.5G (truncated), O_DSYNC
    iodepth=32

  numjobs  baseline BW  RWF_NOSERIAL BW  Δ%
     1        78            80         +2.6%
     2        97            88         -9.3%
     4       100            96         -4.0%
     8       109            97        -11.0%
    16       106           107         +0.9%

Co-developed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 fs/xfs/xfs_file.c  | 54 ++++++++++++++++++++++++++++++++++++++++------
 include/linux/fs.h |  7 ++++++
 2 files changed, 55 insertions(+), 6 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index 4b45ceacd461..b79076c15c5f 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -520,6 +520,42 @@ xfs_file_write_checks(
 	return kiocb_modified(iocb);
 }
 
+STATIC ssize_t
+xfs_file_writethrough_checks(
+	struct kiocb		*iocb,
+	struct iov_iter		*from,
+	unsigned int		*iolock,
+	struct xfs_zone_alloc_ctx *ac)
+{
+	struct inode		*inode = iocb->ki_filp->f_mapping->host;
+	size_t			isize = i_size_read(inode);
+	size_t			count = iov_iter_count(from);
+	ssize_t			error;
+
+	error =  xfs_file_write_checks(iocb, from, iolock, ac);
+	if (error < 0)
+		return error;
+
+	if (*iolock == XFS_IOLOCK_EXCL)
+		return 0;
+
+	/*
+	 * Extending IO needs exclusive lock for i_size change
+	 */
+	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize)
+		goto upgrade_excl;
+
+	return 0;
+
+upgrade_excl:
+	xfs_iunlock(XFS_I(inode), *iolock);
+	*iolock = XFS_IOLOCK_EXCL;
+	error = xfs_ilock_iocb(iocb, *iolock);
+	if (error)
+		*iolock = 0;
+	return error;
+}
+
 static ssize_t
 xfs_zoned_write_space_reserve(
 	struct xfs_mount		*mp,
@@ -1108,23 +1144,29 @@ xfs_file_buffered_write(
 	unsigned int		iolock;
 
 write_retry:
-	iolock = XFS_IOLOCK_EXCL;
+	if (iocb->ki_flags & IOCB_NOSERIAL)
+		iolock = XFS_IOLOCK_SHARED;
+	else
+		iolock = XFS_IOLOCK_EXCL;
 	ret = xfs_ilock_iocb(iocb, iolock);
 	if (ret)
 		return ret;
 
-	ret = xfs_file_write_checks(iocb, from, &iolock, NULL);
-	if (ret)
-		goto out;
-
 	trace_xfs_file_buffered_write(iocb, from);
 	if (iocb->ki_flags & IOCB_WRITETHROUGH) {
+		ret = xfs_file_writethrough_checks(iocb, from, &iolock, NULL);
+		if (ret)
+			goto out;
 		ret = iomap_file_writethrough_write(iocb, from,
 						    &xfs_writethrough_ops, NULL);
-	} else
+	} else {
+		ret = xfs_file_write_checks(iocb, from, &iolock, NULL);
+		if (ret)
+			goto out;
 		ret = iomap_file_buffered_write(iocb, from,
 						&xfs_buffered_write_iomap_ops,
 						&xfs_iomap_write_ops, NULL);
+	}
 
 	/*
 	 * If we hit a space limit, try to free up some lingering preallocated
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 685ffe8da6ea..ed5144512835 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -3495,6 +3495,13 @@ static inline int kiocb_set_rw_flags(struct kiocb *ki, rwf_t flags,
 		ki->ki_flags &= ~IOCB_APPEND;
 	}
 
+	/*
+	 * Writethrough implies non-serial writes ie writes can go
+	 * parallelly.
+	 */
+	if (flags & RWF_WRITETHROUGH)
+		kiocb_flags |= IOCB_NOSERIAL;
+
 	ki->ki_flags |= kiocb_flags;
 	return 0;
 }
-- 
2.55.0

[RFC PATCH v3 10/11] iomap: Avoid folio dirtying in case of RWF_WRITETHROUGH

[Ojaswin Mujoo]

RWF_WRITETHROUGH dirties the folio only to send it for IO immediately,
in the same context. Due to this, we can optimize away the folio
dirtying and clearing step usually seen in buffered IO. Althrough we can
do away with most of the accounting there are a couple of counters we
need to take care of which we do during IO submission/completion.

Co-developed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 fs/iomap/buffered-io.c  | 64 +++++++++++++++++++++++++++++++++--------
 include/linux/pagemap.h |  1 +
 mm/filemap.c            | 21 ++++++++++++++
 3 files changed, 74 insertions(+), 12 deletions(-)

diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index d16694dd9995..0844361fe0f1 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -11,6 +11,9 @@
 #include <linux/fserror.h>
 #include <linux/fsverity.h>
 #include <linux/rmap.h>
+#include <linux/task_io_accounting_ops.h>
+#include "linux/pagemap.h"
+#include "linux/page-flags.h"
 #include "internal.h"
 #include "trace.h"
 
@@ -1162,6 +1165,34 @@ static bool iomap_write_end_inline(const struct iomap_iter *iter,
 	return true;
 }
 
+/*
+ * __iomap_writethrough_end() is almost same as __iomap_write_end() but with the difference
+ * that we don't mark folio dirty since we are about to issue it for IO anyways.
+ * Consequently, most of the accounting is skipped.
+ */
+static bool __iomap_writethrough_end(struct inode *inode, loff_t pos, size_t len,
+		size_t copied, struct folio *folio)
+{
+	flush_dcache_folio(folio);
+
+	/*
+	 * The blocks that were entirely written will now be up-to-date, so we
+	 * don't have to worry about a read_folio reading them and overwriting a
+	 * partial write.  However, if we've encountered a short write and only
+	 * partially written into a block, it will not be marked up-to-date, so a
+	 * read_folio might come in and destroy our partial write.
+	 *
+	 * Do the simplest thing and just treat any short write to a
+	 * non-uptodate page as a zero-length write, and force the caller to
+	 * redo the whole thing.
+	 */
+	if (unlikely(copied < len && !folio_test_uptodate(folio)))
+		return false;
+	iomap_set_range_uptodate(folio, offset_in_folio(folio, pos), len);
+	return true;
+}
+
+
 /*
  * Returns true if all copied bytes have been written to the pagecache,
  * otherwise return false.
@@ -1183,7 +1214,10 @@ static bool iomap_write_end(struct iomap_iter *iter, size_t len, size_t copied,
 		return bh_written == copied;
 	}
 
-	return __iomap_write_end(iter->inode, pos, len, copied, folio);
+	if (iter->flags & IOMAP_WRITETHROUGH)
+		return __iomap_writethrough_end(iter->inode, pos, len, copied, folio);
+	else
+		return __iomap_write_end(iter->inode, pos, len, copied, folio);
 }
 
 static ssize_t iomap_writethrough_complete(struct iomap_writethrough_ctx *wt_ctx)
@@ -1254,7 +1288,7 @@ static void iomap_writethrough_bio_end_io(struct bio *bio)
 		cmpxchg(&wt_ctx->error, 0,
 			blk_status_to_errno(bio->bi_status));
 	bio_for_each_folio_all(fi, bio)
-		folio_end_writeback(fi.folio);
+		folio_end_writethrough(fi.folio, wt_ctx->error);
 
 	bio_put(bio);
 	if (atomic_dec_and_test(&wt_ctx->ref))
@@ -1299,9 +1333,11 @@ iomap_writethrough_submit_bio(struct iomap_writethrough_ctx *wt_ctx,
 
 	/*
 	 * In case of error we still need the I/O completion to run so we can
-	 * release references and end writeback on the folios.
+	 * release references, handle accounting and end writeback on the
+	 * folios.
 	 */
 	if (error) {
+		task_io_account_cancelled_write(len);
 		bio->bi_status = errno_to_blk_status(error);
 		bio_endio(bio);
 		return error;
@@ -1351,6 +1387,7 @@ static void iomap_folio_prepare_writethrough(struct folio *folio, size_t off,
 {
 	bool fully_written;
 	u64 zero = 0;
+	u64 tmp_off = off;
 
 	if (folio_test_writeback(folio))
 		folio_wait_writeback(folio);
@@ -1359,17 +1396,20 @@ static void iomap_folio_prepare_writethrough(struct folio *folio, size_t off,
 		folio_mark_dirty(folio);
 
 	/*
-	 * We might either write through the complete folio or a partial folio
-	 * writethrough might result in all blocks becoming non-dirty, so we need to
-	 * check and mark the folio clean if that is the case.
+	 * For writethrough, we don't mark the write range dirty but we still
+	 * need clear the dirty range if someone else has dirtied it before.
+	 * Further, if the clearing results in folio becoming completely clean,
+	 * then we need to take care of accounting.
 	 */
-	fully_written = (off == 0 && len == folio_size(folio));
-	iomap_clear_range_dirty(folio, off, len);
-	if (fully_written ||
-	    !iomap_find_dirty_range(folio, &zero, folio_size(folio)))
-		folio_clear_dirty_for_writethrough(folio);
+	if (iomap_find_dirty_range(folio, &tmp_off, tmp_off + len)) {
+		iomap_clear_range_dirty(folio, off, len);
 
-	folio_start_writeback(folio);
+		if (!iomap_find_dirty_range(folio, &zero, folio_size(folio)))
+			folio_clear_dirty_for_writethrough(folio);
+	}
+
+	task_io_account_write(folio_nr_pages(folio) * PAGE_SIZE);
+	folio_test_set_writeback(folio);
 }
 
 /**
diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
index b20e38cc0fa0..774a2e57a9e0 100644
--- a/include/linux/pagemap.h
+++ b/include/linux/pagemap.h
@@ -1257,6 +1257,7 @@ void folio_wait_writeback(struct folio *folio);
 int folio_wait_writeback_killable(struct folio *folio);
 void end_page_writeback(struct page *page);
 void folio_end_writeback(struct folio *folio);
+void folio_end_writethrough(struct folio *folio, bool error);
 void folio_end_writeback_no_dropbehind(struct folio *folio);
 void folio_end_dropbehind(struct folio *folio);
 void folio_wait_stable(struct folio *folio);
diff --git a/mm/filemap.c b/mm/filemap.c
index 58eb9d240643..a1a5f8837e03 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1695,6 +1695,27 @@ void folio_end_writeback(struct folio *folio)
 }
 EXPORT_SYMBOL(folio_end_writeback);
 
+/**
+ * folio_end_writethrough - End writethrough against a folio
+ * @folio: The folio.
+ * error: Was there an error in IO.
+ *
+ * Context: May be called from process or interrupt context.
+ */
+void folio_end_writethrough(struct folio *folio, bool error)
+{
+	long nr = folio_nr_pages(folio);
+
+	VM_BUG_ON_FOLIO(!folio_test_writeback(folio), folio);
+
+	if (!error)
+		node_stat_mod_folio(folio, NR_WRITTEN, nr);
+
+	if (folio_xor_flags_has_waiters(folio, 1 << PG_writeback))
+		folio_wake_bit(folio, PG_writeback);
+}
+EXPORT_SYMBOL(folio_end_writethrough);
+
 /**
  * __folio_lock - Get a lock on the folio, assuming we need to sleep to get it.
  * @folio: The folio to lock
-- 
2.55.0

[RFC PATCH v3 11/11] iomap: Handle deadlock due to repeating folios in RWF_WRITETHROUGH

[Ojaswin Mujoo]

In iomap_writethrough_iter() we might encounter repeating folios across
multiple iterations. Repeating folios can occur if, example,
copy_folio_from_iter_atomic() does a short copy due to userspace pages
not faulted in. This is an issue because a previous loop might have
started writeback on them but not yet issued the IO. In the next
iteration trying to get the same folio with FGP_STABLE will result in a
deadlock. Since repeating folios will always be encountered back to
back, we can just use a simple cur != prev check to detect them.

Use this to avoid waiting for writeback or starting writeback on folios
we have already processed. Note that in ->endio() we might end up
calling folio_end_writethrough() twice on the same folio which can cause
issues with folio_xor_flags_has_waiters(). For simplicity, just change
the folio_xor_flags_has_waiters() call to an idempotent variant.

Reported-by: Pankaj Raghav <pankaj.raghav@linux.dev>
Co-developed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
---
 fs/iomap/buffered-io.c | 37 ++++++++++++++++++++++++++++++-------
 mm/filemap.c           |  3 ++-
 2 files changed, 32 insertions(+), 8 deletions(-)

diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index 0844361fe0f1..70c1565e5416 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -804,6 +804,13 @@ struct folio *iomap_get_folio(struct iomap_iter *iter, loff_t pos, size_t len)
 {
 	fgf_t fgp = FGP_WRITEBEGIN;
 
+	/*
+	 * For writethrough, we open code the FGP_STABLE logic directly in
+	 * iomap_writhrethrough_iter() so disable it here.. See
+	 * iomap_writethrough_iter() for details.
+	 */
+	if (iter->flags & IOMAP_WRITETHROUGH)
+		fgp &= ~FGP_STABLE;
 	if (iter->flags & IOMAP_NOWAIT)
 		fgp |= FGP_NOWAIT;
 	if (iter->flags & IOMAP_DONTCACHE)
@@ -1383,15 +1390,12 @@ iomap_writethrough_try_submit(struct iomap_writethrough_ctx *wt_ctx,
  * need to clear the master dirty bit.
  */
 static void iomap_folio_prepare_writethrough(struct folio *folio, size_t off,
-					     size_t len)
+					     size_t len, bool already_prepared)
 {
 	bool fully_written;
 	u64 zero = 0;
 	u64 tmp_off = off;
 
-	if (folio_test_writeback(folio))
-		folio_wait_writeback(folio);
-
 	if (folio_mkclean(folio))
 		folio_mark_dirty(folio);
 
@@ -1409,7 +1413,8 @@ static void iomap_folio_prepare_writethrough(struct folio *folio, size_t off,
 	}
 
 	task_io_account_write(folio_nr_pages(folio) * PAGE_SIZE);
-	folio_test_set_writeback(folio);
+	if (!already_prepared)
+		folio_test_set_writeback(folio);
 }
 
 /**
@@ -1426,6 +1431,17 @@ static void iomap_folio_prepare_writethrough(struct folio *folio, size_t off,
  * Folio handling note: We might be writing through a partial folio so we need
  * to be careful to not clear the folio dirty bit unless there are no dirty blocks
  * in the folio after the writethrough.
+ *
+ * **A corner case to be careful about**
+ *
+ * For writethrough, we open code the stable write behavior to handle the case
+ * where we encounter a folio that we already started writeback on but have not
+ * yet submitted. In that case we must not wait for writeback again to avoid
+ * deadlocking. Repeating folios can occur if, example,
+ * copy_folio_from_iter_atomic() does a short copy due to userspace pages not
+ * faulted in. Also, repeating folios will always be encountered back to back so
+ * we can just use a simple cur != prev check to detect them.
+
  */
 static int iomap_writethrough_iter(struct iomap_writethrough_ctx *wt_ctx,
 				   struct iomap_iter *iter, struct iov_iter *i,
@@ -1439,6 +1455,7 @@ static int iomap_writethrough_iter(struct iomap_writethrough_ctx *wt_ctx,
 	size_t chunk = mapping_max_folio_size(mapping);
 	unsigned int bdp_flags = (iter->flags & IOMAP_NOWAIT) ? BDP_ASYNC : 0;
 	unsigned int bs = i_blocksize(iter->inode);
+	struct folio *prev_folio = NULL;
 
 	/* copied over based on how DIO handles these flags */
 	if (iter->iomap.type == IOMAP_UNWRITTEN)
@@ -1530,6 +1547,10 @@ static int iomap_writethrough_iter(struct iomap_writethrough_ctx *wt_ctx,
 		if (mapping_writably_mapped(mapping))
 			flush_dcache_folio(folio);
 
+		 /* Open coding stable write behavior, see comment on top. */
+		if (prev_folio != folio)
+			folio_wait_writeback(folio);
+
 		copied = copy_folio_from_iter_atomic(folio, offset, bytes, i);
 		written = iomap_write_end(iter, bytes, copied, folio) ?
 			  copied : 0;
@@ -1544,8 +1565,10 @@ static int iomap_writethrough_iter(struct iomap_writethrough_ctx *wt_ctx,
 		off_aligned = round_down(offset, bs);
 		len_aligned = round_up(offset + written, bs) - off_aligned;
 
-		iomap_folio_prepare_writethrough(folio, off_aligned,
-						 len_aligned);
+		iomap_folio_prepare_writethrough(
+			folio, off_aligned, len_aligned, prev_folio == folio);
+
+		prev_folio = folio;
 
 		if (!wt_ctx->nr_bvecs) {
 			wt_ctx->bio_pos = round_down(pos, bs);
diff --git a/mm/filemap.c b/mm/filemap.c
index a1a5f8837e03..fc3ed0619838 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1711,7 +1711,8 @@ void folio_end_writethrough(struct folio *folio, bool error)
 	if (!error)
 		node_stat_mod_folio(folio, NR_WRITTEN, nr);
 
-	if (folio_xor_flags_has_waiters(folio, 1 << PG_writeback))
+	folio_test_clear_writeback(folio);
+	if (folio_test_waiters(folio))
 		folio_wake_bit(folio, PG_writeback);
 }
 EXPORT_SYMBOL(folio_end_writethrough);
-- 
2.55.0

Re: [RFC PATCH v3 00/11] Add buffered write-through support to iomap & xfs

[Ojaswin Mujoo]

On Wed, Aug 05, 2026 at 11:58:06AM +0530, Ojaswin Mujoo wrote:
> This is the next revision of writethrough series. I'll quote part of the

(+cc Andres)