iomap: optionally use ioends for direct I/O

struct iomap_ioend currently tracks outstanding buffered writes and has
some really nice code in core iomap and XFS to merge contiguous I/Os
an defer them to userspace for completion in a very efficient way.

For zoned writes we'll also need a per-bio user context completion to
record the written blocks, and the infrastructure for that would look
basically like the ioend handling for buffered I/O.

So instead of reinventing the wheel, reuse the existing infrastructure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20250206064035.2323428-8-hch@lst.de
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>

4 files changed, 55 insertions, 6 deletions

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index 277ece243770..138d246ec29d 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 /*
  * Copyright (C) 2010 Red Hat, Inc.
- * Copyright (c) 2016-2021 Christoph Hellwig.
+ * Copyright (c) 2016-2025 Christoph Hellwig.
  */
 #include <linux/module.h>
 #include <linux/compiler.h>
@@ -12,6 +12,7 @@
 #include <linux/backing-dev.h>
 #include <linux/uio.h>
 #include <linux/task_io_accounting_ops.h>
+#include "internal.h"
 #include "trace.h"
 
 #include "../internal.h"
@@ -20,6 +21,7 @@
  * Private flags for iomap_dio, must not overlap with the public ones in
  * iomap.h:
  */
+#define IOMAP_DIO_NO_INVALIDATE	(1U << 25)
 #define IOMAP_DIO_CALLER_COMP	(1U << 26)
 #define IOMAP_DIO_INLINE_COMP	(1U << 27)
 #define IOMAP_DIO_WRITE_THROUGH	(1U << 28)
@@ -119,7 +121,8 @@ ssize_t iomap_dio_complete(struct iomap_dio *dio)
 	 * ->end_io() when necessary, otherwise a racing buffer read would cache
 	 * zeros from unwritten extents.
 	 */
-	if (!dio->error && dio->size && (dio->flags & IOMAP_DIO_WRITE))
+	if (!dio->error && dio->size && (dio->flags & IOMAP_DIO_WRITE) &&
+	    !(dio->flags & IOMAP_DIO_NO_INVALIDATE))
 		kiocb_invalidate_post_direct_write(iocb, dio->size);
 
 	inode_dio_end(file_inode(iocb->ki_filp));
@@ -241,6 +244,47 @@ void iomap_dio_bio_end_io(struct bio *bio)
 }
 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
 
+u32 iomap_finish_ioend_direct(struct iomap_ioend *ioend)
+{
+	struct iomap_dio *dio = ioend->io_bio.bi_private;
+	bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
+	u32 vec_count = ioend->io_bio.bi_vcnt;
+
+	if (ioend->io_error)
+		iomap_dio_set_error(dio, ioend->io_error);
+
+	if (atomic_dec_and_test(&dio->ref)) {
+		/*
+		 * Try to avoid another context switch for the completion given
+		 * that we are already called from the ioend completion
+		 * workqueue, but never invalidate pages from this thread to
+		 * avoid deadlocks with buffered I/O completions.  Tough luck if
+		 * you hit the tiny race with someone dirtying the range now
+		 * between this check and the actual completion.
+		 */
+		if (!dio->iocb->ki_filp->f_mapping->nrpages) {
+			dio->flags |= IOMAP_DIO_INLINE_COMP;
+			dio->flags |= IOMAP_DIO_NO_INVALIDATE;
+		}
+		dio->flags &= ~IOMAP_DIO_CALLER_COMP;
+		iomap_dio_done(dio);
+	}
+
+	if (should_dirty) {
+		bio_check_pages_dirty(&ioend->io_bio);
+	} else {
+		bio_release_pages(&ioend->io_bio, false);
+		bio_put(&ioend->io_bio);
+	}
+
+	/*
+	 * Return the number of bvecs completed as even direct I/O completions
+	 * do significant per-folio work and we'll still want to give up the
+	 * CPU after a lot of completions.
+	 */
+	return vec_count;
+}
+
 static int iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
 		loff_t pos, unsigned len)
 {
diff --git a/fs/iomap/internal.h b/fs/iomap/internal.h
index 36d5c56e073e..f6992a3bf66a 100644
--- a/fs/iomap/internal.h
+++ b/fs/iomap/internal.h
@@ -5,5 +5,6 @@
 #define IOEND_BATCH_SIZE	4096
 
 u32 iomap_finish_ioend_buffered(struct iomap_ioend *ioend);
+u32 iomap_finish_ioend_direct(struct iomap_ioend *ioend);
 
 #endif /* _IOMAP_INTERNAL_H */
diff --git a/fs/iomap/ioend.c b/fs/iomap/ioend.c
index 97d43c50cdf7..44f254ecab55 100644
--- a/fs/iomap/ioend.c
+++ b/fs/iomap/ioend.c
@@ -41,6 +41,8 @@ static u32 iomap_finish_ioend(struct iomap_ioend *ioend, int error)
 
 	if (!atomic_dec_and_test(&ioend->io_remaining))
 		return 0;
+	if (ioend->io_flags & IOMAP_IOEND_DIRECT)
+		return iomap_finish_ioend_direct(ioend);
 	return iomap_finish_ioend_buffered(ioend);
 }
 
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 90c27875e39d..5768b9f2a1cc 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -343,20 +343,22 @@ sector_t iomap_bmap(struct address_space *mapping, sector_t bno,
 #define IOMAP_IOEND_UNWRITTEN		(1U << 1)
 /* don't merge into previous ioend */
 #define IOMAP_IOEND_BOUNDARY		(1U << 2)
+/* is direct I/O */
+#define IOMAP_IOEND_DIRECT		(1U << 3)
 
 /*
  * Flags that if set on either ioend prevent the merge of two ioends.
  * (IOMAP_IOEND_BOUNDARY also prevents merges, but only one-way)
  */
 #define IOMAP_IOEND_NOMERGE_FLAGS \
-	(IOMAP_IOEND_SHARED | IOMAP_IOEND_UNWRITTEN)
+	(IOMAP_IOEND_SHARED | IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_DIRECT)
 
 /*
  * Structure for writeback I/O completions.
  *
- * File systems implementing ->submit_ioend can split a bio generated
- * by iomap.  In that case the parent ioend it was split from is recorded
- * in ioend->io_parent.
+ * File systems implementing ->submit_ioend (for buffered I/O) or ->submit_io
+ * for direct I/O) can split a bio generated by iomap.  In that case the parent
+ * ioend it was split from is recorded in ioend->io_parent.
  */
 struct iomap_ioend {
 	struct list_head	io_list;	/* next ioend in chain */