direct-io.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Copyright (C) 2010 Red Hat, Inc.
  4 * Copyright (c) 2016-2025 Christoph Hellwig.
  5 */
  6#include <linux/fscrypt.h>
  7#include <linux/pagemap.h>
  8#include <linux/iomap.h>
  9#include <linux/task_io_accounting_ops.h>
 10#include "internal.h"
 11#include "trace.h"
 12
 13#include "../internal.h"
 14
 15/*
 16 * Private flags for iomap_dio, must not overlap with the public ones in
 17 * iomap.h:
 18 */
 19#define IOMAP_DIO_NO_INVALIDATE	(1U << 26)
 20#define IOMAP_DIO_COMP_WORK	(1U << 27)
 21#define IOMAP_DIO_WRITE_THROUGH	(1U << 28)
 22#define IOMAP_DIO_NEED_SYNC	(1U << 29)
 23#define IOMAP_DIO_WRITE		(1U << 30)
 24#define IOMAP_DIO_DIRTY		(1U << 31)
 25
 26struct iomap_dio {
 27	struct kiocb		*iocb;
 28	const struct iomap_dio_ops *dops;
 29	loff_t			i_size;
 30	loff_t			size;
 31	atomic_t		ref;
 32	unsigned		flags;
 33	int			error;
 34	size_t			done_before;
 35	bool			wait_for_completion;
 36
 37	union {
 38		/* used during submission and for synchronous completion: */
 39		struct {
 40			struct iov_iter		*iter;
 41			struct task_struct	*waiter;
 42		} submit;
 43
 44		/* used for aio completion: */
 45		struct {
 46			struct work_struct	work;
 47		} aio;
 48	};
 49};
 50
 51static struct bio *iomap_dio_alloc_bio(const struct iomap_iter *iter,
 52		struct iomap_dio *dio, unsigned short nr_vecs, blk_opf_t opf)
 53{
 54	if (dio->dops && dio->dops->bio_set)
 55		return bio_alloc_bioset(iter->iomap.bdev, nr_vecs, opf,
 56					GFP_KERNEL, dio->dops->bio_set);
 57	return bio_alloc(iter->iomap.bdev, nr_vecs, opf, GFP_KERNEL);
 58}
 59
 60static void iomap_dio_submit_bio(const struct iomap_iter *iter,
 61		struct iomap_dio *dio, struct bio *bio, loff_t pos)
 62{
 63	struct kiocb *iocb = dio->iocb;
 64
 65	atomic_inc(&dio->ref);
 66
 67	/* Sync dio can't be polled reliably */
 68	if ((iocb->ki_flags & IOCB_HIPRI) && !is_sync_kiocb(iocb)) {
 69		bio_set_polled(bio, iocb);
 70		WRITE_ONCE(iocb->private, bio);
 71	}
 72
 73	if (dio->dops && dio->dops->submit_io) {
 74		dio->dops->submit_io(iter, bio, pos);
 75	} else {
 76		WARN_ON_ONCE(iter->iomap.flags & IOMAP_F_ANON_WRITE);
 77		submit_bio(bio);
 78	}
 79}
 80
 81ssize_t iomap_dio_complete(struct iomap_dio *dio)
 82{
 83	const struct iomap_dio_ops *dops = dio->dops;
 84	struct kiocb *iocb = dio->iocb;
 85	loff_t offset = iocb->ki_pos;
 86	ssize_t ret = dio->error;
 87
 88	if (dops && dops->end_io)
 89		ret = dops->end_io(iocb, dio->size, ret, dio->flags);
 90
 91	if (likely(!ret)) {
 92		ret = dio->size;
 93		/* check for short read */
 94		if (offset + ret > dio->i_size &&
 95		    !(dio->flags & IOMAP_DIO_WRITE))
 96			ret = dio->i_size - offset;
 97	}
 98
 99	/*
100	 * Try again to invalidate clean pages which might have been cached by
101	 * non-direct readahead, or faulted in by get_user_pages() if the source
102	 * of the write was an mmap'ed region of the file we're writing.  Either
103	 * one is a pretty crazy thing to do, so we don't support it 100%.  If
104	 * this invalidation fails, tough, the write still worked...
105	 *
106	 * And this page cache invalidation has to be after ->end_io(), as some
107	 * filesystems convert unwritten extents to real allocations in
108	 * ->end_io() when necessary, otherwise a racing buffer read would cache
109	 * zeros from unwritten extents.
110	 */
111	if (!dio->error && dio->size && (dio->flags & IOMAP_DIO_WRITE) &&
112	    !(dio->flags & IOMAP_DIO_NO_INVALIDATE))
113		kiocb_invalidate_post_direct_write(iocb, dio->size);
114
115	inode_dio_end(file_inode(iocb->ki_filp));
116
117	if (ret > 0) {
118		iocb->ki_pos += ret;
119
120		/*
121		 * If this is a DSYNC write, make sure we push it to stable
122		 * storage now that we've written data.
123		 */
124		if (dio->flags & IOMAP_DIO_NEED_SYNC)
125			ret = generic_write_sync(iocb, ret);
126		if (ret > 0)
127			ret += dio->done_before;
128	}
129	trace_iomap_dio_complete(iocb, dio->error, ret);
130	kfree(dio);
131	return ret;
132}
133EXPORT_SYMBOL_GPL(iomap_dio_complete);
134
135static void iomap_dio_complete_work(struct work_struct *work)
136{
137	struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
138	struct kiocb *iocb = dio->iocb;
139
140	iocb->ki_complete(iocb, iomap_dio_complete(dio));
141}
142
143/*
144 * Set an error in the dio if none is set yet.  We have to use cmpxchg
145 * as the submission context and the completion context(s) can race to
146 * update the error.
147 */
148static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
149{
150	cmpxchg(&dio->error, 0, ret);
151}
152
153/*
154 * Called when dio->ref reaches zero from an I/O completion.
155 */
156static void iomap_dio_done(struct iomap_dio *dio)
157{
158	struct kiocb *iocb = dio->iocb;
159
160	if (dio->wait_for_completion) {
161		/*
162		 * Synchronous I/O, task itself will handle any completion work
163		 * that needs after IO. All we need to do is wake the task.
164		 */
165		struct task_struct *waiter = dio->submit.waiter;
166
167		WRITE_ONCE(dio->submit.waiter, NULL);
168		blk_wake_io_task(waiter);
169		return;
170	}
171
172	/*
173	 * Always run error completions in user context.  These are not
174	 * performance critical and some code relies on taking sleeping locks
175	 * for error handling.
176	 */
177	if (dio->error)
178		dio->flags |= IOMAP_DIO_COMP_WORK;
179
180	/*
181	 * Never invalidate pages from this context to avoid deadlocks with
182	 * buffered I/O completions when called from the ioend workqueue,
183	 * or avoid sleeping when called directly from ->bi_end_io.
184	 * Tough luck if you hit the tiny race with someone dirtying the range
185	 * right between this check and the actual completion.
186	 */
187	if ((dio->flags & IOMAP_DIO_WRITE) &&
188	    !(dio->flags & IOMAP_DIO_COMP_WORK)) {
189		if (dio->iocb->ki_filp->f_mapping->nrpages)
190			dio->flags |= IOMAP_DIO_COMP_WORK;
191		else
192			dio->flags |= IOMAP_DIO_NO_INVALIDATE;
193	}
194
195	if (dio->flags & IOMAP_DIO_COMP_WORK) {
196		struct inode *inode = file_inode(iocb->ki_filp);
197
198		/*
199		 * Async DIO completion that requires filesystem level
200		 * completion work gets punted to a work queue to complete as
201		 * the operation may require more IO to be issued to finalise
202		 * filesystem metadata changes or guarantee data integrity.
203		 */
204		INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
205		queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
206		return;
207	}
208
209	WRITE_ONCE(iocb->private, NULL);
210	iomap_dio_complete_work(&dio->aio.work);
211}
212
213void iomap_dio_bio_end_io(struct bio *bio)
214{
215	struct iomap_dio *dio = bio->bi_private;
216	bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
217
218	if (bio->bi_status)
219		iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
220
221	if (atomic_dec_and_test(&dio->ref))
222		iomap_dio_done(dio);
223
224	if (should_dirty) {
225		bio_check_pages_dirty(bio);
226	} else {
227		bio_release_pages(bio, false);
228		bio_put(bio);
229	}
230}
231EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
232
233u32 iomap_finish_ioend_direct(struct iomap_ioend *ioend)
234{
235	struct iomap_dio *dio = ioend->io_bio.bi_private;
236	bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
237	u32 vec_count = ioend->io_bio.bi_vcnt;
238
239	if (ioend->io_error)
240		iomap_dio_set_error(dio, ioend->io_error);
241
242	if (atomic_dec_and_test(&dio->ref)) {
243		/*
244		 * Try to avoid another context switch for the completion given
245		 * that we are already called from the ioend completion
246		 * workqueue.
247		 */
248		dio->flags &= ~IOMAP_DIO_COMP_WORK;
249		iomap_dio_done(dio);
250	}
251
252	if (should_dirty) {
253		bio_check_pages_dirty(&ioend->io_bio);
254	} else {
255		bio_release_pages(&ioend->io_bio, false);
256		bio_put(&ioend->io_bio);
257	}
258
259	/*
260	 * Return the number of bvecs completed as even direct I/O completions
261	 * do significant per-folio work and we'll still want to give up the
262	 * CPU after a lot of completions.
263	 */
264	return vec_count;
265}
266
267static int iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
268		loff_t pos, unsigned len)
269{
270	struct inode *inode = file_inode(dio->iocb->ki_filp);
271	struct bio *bio;
272	struct folio *zero_folio = largest_zero_folio();
273	int nr_vecs = max(1, i_blocksize(inode) / folio_size(zero_folio));
274
275	if (!len)
276		return 0;
277
278	/*
279	 * This limit shall never be reached as most filesystems have a
280	 * maximum blocksize of 64k.
281	 */
282	if (WARN_ON_ONCE(nr_vecs > BIO_MAX_VECS))
283		return -EINVAL;
284
285	bio = iomap_dio_alloc_bio(iter, dio, nr_vecs,
286				  REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
287	fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
288				  GFP_KERNEL);
289	bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
290	bio->bi_private = dio;
291	bio->bi_end_io = iomap_dio_bio_end_io;
292
293	while (len > 0) {
294		unsigned int io_len = min(len, folio_size(zero_folio));
295
296		bio_add_folio_nofail(bio, zero_folio, io_len, 0);
297		len -= io_len;
298	}
299	iomap_dio_submit_bio(iter, dio, bio, pos);
300
301	return 0;
302}
303
304static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
305{
306	const struct iomap *iomap = &iter->iomap;
307	struct inode *inode = iter->inode;
308	unsigned int fs_block_size = i_blocksize(inode), pad;
309	const loff_t length = iomap_length(iter);
310	loff_t pos = iter->pos;
311	blk_opf_t bio_opf = REQ_SYNC | REQ_IDLE;
312	struct bio *bio;
313	bool need_zeroout = false;
314	int nr_pages, ret = 0;
315	u64 copied = 0;
316	size_t orig_count;
317	unsigned int alignment;
318
319	/*
320	 * File systems that write out of place and always allocate new blocks
321	 * need each bio to be block aligned as that's the unit of allocation.
322	 */
323	if (dio->flags & IOMAP_DIO_FSBLOCK_ALIGNED)
324		alignment = fs_block_size;
325	else
326		alignment = bdev_logical_block_size(iomap->bdev);
327
328	if ((pos | length) & (alignment - 1))
329		return -EINVAL;
330
331	if (dio->flags & IOMAP_DIO_WRITE) {
332		bool need_completion_work = true;
333
334		switch (iomap->type) {
335		case IOMAP_MAPPED:
336			/*
337			 * Directly mapped I/O does not inherently need to do
338			 * work at I/O completion time.  But there are various
339			 * cases below where this will get set again.
340			 */
341			need_completion_work = false;
342			break;
343		case IOMAP_UNWRITTEN:
344			dio->flags |= IOMAP_DIO_UNWRITTEN;
345			need_zeroout = true;
346			break;
347		default:
348			break;
349		}
350
351		if (iomap->flags & IOMAP_F_ATOMIC_BIO) {
352			/*
353			 * Ensure that the mapping covers the full write
354			 * length, otherwise it won't be submitted as a single
355			 * bio, which is required to use hardware atomics.
356			 */
357			if (length != iter->len)
358				return -EINVAL;
359			bio_opf |= REQ_ATOMIC;
360		}
361
362		if (iomap->flags & IOMAP_F_SHARED) {
363			/*
364			 * Unsharing of needs to update metadata at I/O
365			 * completion time.
366			 */
367			need_completion_work = true;
368			dio->flags |= IOMAP_DIO_COW;
369		}
370
371		if (iomap->flags & IOMAP_F_NEW) {
372			/*
373			 * Newly allocated blocks might need recording in
374			 * metadata at I/O completion time.
375			 */
376			need_completion_work = true;
377			need_zeroout = true;
378		}
379
380		/*
381		 * Use a FUA write if we need datasync semantics and this is a
382		 * pure overwrite that doesn't require any metadata updates.
383		 *
384		 * This allows us to avoid cache flushes on I/O completion.
385		 */
386		if (dio->flags & IOMAP_DIO_WRITE_THROUGH) {
387			if (!need_completion_work &&
388			    !(iomap->flags & IOMAP_F_DIRTY) &&
389			    (!bdev_write_cache(iomap->bdev) ||
390			     bdev_fua(iomap->bdev)))
391				bio_opf |= REQ_FUA;
392			else
393				dio->flags &= ~IOMAP_DIO_WRITE_THROUGH;
394		}
395
396		/*
397		 * We can only do inline completion for pure overwrites that
398		 * don't require additional I/O at completion time.
399		 *
400		 * This rules out writes that need zeroing or metdata updates to
401		 * convert unwritten or shared extents.
402		 *
403		 * Writes that extend i_size are also not supported, but this is
404		 * handled in __iomap_dio_rw().
405		 */
406		if (need_completion_work)
407			dio->flags |= IOMAP_DIO_COMP_WORK;
408
409		bio_opf |= REQ_OP_WRITE;
410	} else {
411		bio_opf |= REQ_OP_READ;
412	}
413
414	/*
415	 * Save the original count and trim the iter to just the extent we
416	 * are operating on right now.  The iter will be re-expanded once
417	 * we are done.
418	 */
419	orig_count = iov_iter_count(dio->submit.iter);
420	iov_iter_truncate(dio->submit.iter, length);
421
422	if (!iov_iter_count(dio->submit.iter))
423		goto out;
424
425	/*
426	 * The rules for polled IO completions follow the guidelines as the
427	 * ones we set for inline and deferred completions. If none of those
428	 * are available for this IO, clear the polled flag.
429	 */
430	if (dio->flags & IOMAP_DIO_COMP_WORK)
431		dio->iocb->ki_flags &= ~IOCB_HIPRI;
432
433	if (need_zeroout) {
434		/* zero out from the start of the block to the write offset */
435		pad = pos & (fs_block_size - 1);
436
437		ret = iomap_dio_zero(iter, dio, pos - pad, pad);
438		if (ret)
439			goto out;
440	}
441
442	nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
443	do {
444		size_t n;
445
446		/*
447		 * If completions already occurred and reported errors, give up now and
448		 * don't bother submitting more bios.
449		 */
450		if (unlikely(data_race(dio->error))) {
451			ret = 0;
452			goto out;
453		}
454
455		bio = iomap_dio_alloc_bio(iter, dio, nr_pages, bio_opf);
456		fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
457					  GFP_KERNEL);
458		bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
459		bio->bi_write_hint = inode->i_write_hint;
460		bio->bi_ioprio = dio->iocb->ki_ioprio;
461		bio->bi_private = dio;
462		bio->bi_end_io = iomap_dio_bio_end_io;
463
464		ret = bio_iov_iter_get_pages(bio, dio->submit.iter,
465					     alignment - 1);
466		if (unlikely(ret)) {
467			/*
468			 * We have to stop part way through an IO. We must fall
469			 * through to the sub-block tail zeroing here, otherwise
470			 * this short IO may expose stale data in the tail of
471			 * the block we haven't written data to.
472			 */
473			bio_put(bio);
474			goto zero_tail;
475		}
476
477		n = bio->bi_iter.bi_size;
478		if (WARN_ON_ONCE((bio_opf & REQ_ATOMIC) && n != length)) {
479			/*
480			 * An atomic write bio must cover the complete length,
481			 * which it doesn't, so error. We may need to zero out
482			 * the tail (complete FS block), similar to when
483			 * bio_iov_iter_get_pages() returns an error, above.
484			 */
485			ret = -EINVAL;
486			bio_put(bio);
487			goto zero_tail;
488		}
489		if (dio->flags & IOMAP_DIO_WRITE)
490			task_io_account_write(n);
491		else if (dio->flags & IOMAP_DIO_DIRTY)
492			bio_set_pages_dirty(bio);
493
494		dio->size += n;
495		copied += n;
496
497		nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter,
498						 BIO_MAX_VECS);
499		/*
500		 * We can only poll for single bio I/Os.
501		 */
502		if (nr_pages)
503			dio->iocb->ki_flags &= ~IOCB_HIPRI;
504		iomap_dio_submit_bio(iter, dio, bio, pos);
505		pos += n;
506	} while (nr_pages);
507
508	/*
509	 * We need to zeroout the tail of a sub-block write if the extent type
510	 * requires zeroing or the write extends beyond EOF. If we don't zero
511	 * the block tail in the latter case, we can expose stale data via mmap
512	 * reads of the EOF block.
513	 */
514zero_tail:
515	if (need_zeroout ||
516	    ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
517		/* zero out from the end of the write to the end of the block */
518		pad = pos & (fs_block_size - 1);
519		if (pad)
520			ret = iomap_dio_zero(iter, dio, pos,
521					     fs_block_size - pad);
522	}
523out:
524	/* Undo iter limitation to current extent */
525	iov_iter_reexpand(dio->submit.iter, orig_count - copied);
526	if (copied)
527		return iomap_iter_advance(iter, copied);
528	return ret;
529}
530
531static int iomap_dio_hole_iter(struct iomap_iter *iter, struct iomap_dio *dio)
532{
533	loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter);
534
535	dio->size += length;
536	if (!length)
537		return -EFAULT;
538	return iomap_iter_advance(iter, length);
539}
540
541static int iomap_dio_inline_iter(struct iomap_iter *iomi, struct iomap_dio *dio)
542{
543	const struct iomap *iomap = &iomi->iomap;
544	struct iov_iter *iter = dio->submit.iter;
545	void *inline_data = iomap_inline_data(iomap, iomi->pos);
546	loff_t length = iomap_length(iomi);
547	loff_t pos = iomi->pos;
548	u64 copied;
549
550	if (WARN_ON_ONCE(!inline_data))
551		return -EIO;
552
553	if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap)))
554		return -EIO;
555
556	if (dio->flags & IOMAP_DIO_WRITE) {
557		loff_t size = iomi->inode->i_size;
558
559		if (pos > size)
560			memset(iomap_inline_data(iomap, size), 0, pos - size);
561		copied = copy_from_iter(inline_data, length, iter);
562		if (copied) {
563			if (pos + copied > size)
564				i_size_write(iomi->inode, pos + copied);
565			mark_inode_dirty(iomi->inode);
566		}
567	} else {
568		copied = copy_to_iter(inline_data, length, iter);
569	}
570	dio->size += copied;
571	if (!copied)
572		return -EFAULT;
573	return iomap_iter_advance(iomi, copied);
574}
575
576static int iomap_dio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
577{
578	switch (iter->iomap.type) {
579	case IOMAP_HOLE:
580		if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
581			return -EIO;
582		return iomap_dio_hole_iter(iter, dio);
583	case IOMAP_UNWRITTEN:
584		if (!(dio->flags & IOMAP_DIO_WRITE))
585			return iomap_dio_hole_iter(iter, dio);
586		return iomap_dio_bio_iter(iter, dio);
587	case IOMAP_MAPPED:
588		return iomap_dio_bio_iter(iter, dio);
589	case IOMAP_INLINE:
590		return iomap_dio_inline_iter(iter, dio);
591	case IOMAP_DELALLOC:
592		/*
593		 * DIO is not serialised against mmap() access at all, and so
594		 * if the page_mkwrite occurs between the writeback and the
595		 * iomap_iter() call in the DIO path, then it will see the
596		 * DELALLOC block that the page-mkwrite allocated.
597		 */
598		pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
599				    dio->iocb->ki_filp, current->comm);
600		return -EIO;
601	default:
602		WARN_ON_ONCE(1);
603		return -EIO;
604	}
605}
606
607/*
608 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
609 * is being issued as AIO or not.  This allows us to optimise pure data writes
610 * to use REQ_FUA rather than requiring generic_write_sync() to issue a
611 * REQ_FLUSH post write. This is slightly tricky because a single request here
612 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
613 * may be pure data writes. In that case, we still need to do a full data sync
614 * completion.
615 *
616 * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL,
617 * __iomap_dio_rw can return a partial result if it encounters a non-resident
618 * page in @iter after preparing a transfer.  In that case, the non-resident
619 * pages can be faulted in and the request resumed with @done_before set to the
620 * number of bytes previously transferred.  The request will then complete with
621 * the correct total number of bytes transferred; this is essential for
622 * completing partial requests asynchronously.
623 *
624 * Returns -ENOTBLK In case of a page invalidation invalidation failure for
625 * writes.  The callers needs to fall back to buffered I/O in this case.
626 */
627struct iomap_dio *
628__iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
629		const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
630		unsigned int dio_flags, void *private, size_t done_before)
631{
632	struct inode *inode = file_inode(iocb->ki_filp);
633	struct iomap_iter iomi = {
634		.inode		= inode,
635		.pos		= iocb->ki_pos,
636		.len		= iov_iter_count(iter),
637		.flags		= IOMAP_DIRECT,
638		.private	= private,
639	};
640	bool wait_for_completion =
641		is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
642	struct blk_plug plug;
643	struct iomap_dio *dio;
644	loff_t ret = 0;
645
646	trace_iomap_dio_rw_begin(iocb, iter, dio_flags, done_before);
647
648	if (!iomi.len)
649		return NULL;
650
651	dio = kmalloc(sizeof(*dio), GFP_KERNEL);
652	if (!dio)
653		return ERR_PTR(-ENOMEM);
654
655	dio->iocb = iocb;
656	atomic_set(&dio->ref, 1);
657	dio->size = 0;
658	dio->i_size = i_size_read(inode);
659	dio->dops = dops;
660	dio->error = 0;
661	dio->flags = 0;
662	dio->done_before = done_before;
663
664	dio->submit.iter = iter;
665	dio->submit.waiter = current;
666
667	if (iocb->ki_flags & IOCB_NOWAIT)
668		iomi.flags |= IOMAP_NOWAIT;
669
670	if (dio_flags & IOMAP_DIO_FSBLOCK_ALIGNED)
671		dio->flags |= IOMAP_DIO_FSBLOCK_ALIGNED;
672
673	if (iov_iter_rw(iter) == READ) {
674		if (iomi.pos >= dio->i_size)
675			goto out_free_dio;
676
677		if (user_backed_iter(iter))
678			dio->flags |= IOMAP_DIO_DIRTY;
679
680		ret = kiocb_write_and_wait(iocb, iomi.len);
681		if (ret)
682			goto out_free_dio;
683	} else {
684		iomi.flags |= IOMAP_WRITE;
685		dio->flags |= IOMAP_DIO_WRITE;
686
687		if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) {
688			ret = -EAGAIN;
689			if (iomi.pos >= dio->i_size ||
690			    iomi.pos + iomi.len > dio->i_size)
691				goto out_free_dio;
692			iomi.flags |= IOMAP_OVERWRITE_ONLY;
693		}
694
695		if (iocb->ki_flags & IOCB_ATOMIC)
696			iomi.flags |= IOMAP_ATOMIC;
697
698		/* for data sync or sync, we need sync completion processing */
699		if (iocb_is_dsync(iocb)) {
700			dio->flags |= IOMAP_DIO_NEED_SYNC;
701
702		       /*
703			* For datasync only writes, we optimistically try using
704			* WRITE_THROUGH for this IO. This flag requires either
705			* FUA writes through the device's write cache, or a
706			* normal write to a device without a volatile write
707			* cache. For the former, Any non-FUA write that occurs
708			* will clear this flag, hence we know before completion
709			* whether a cache flush is necessary.
710			*/
711			if (!(iocb->ki_flags & IOCB_SYNC))
712				dio->flags |= IOMAP_DIO_WRITE_THROUGH;
713		}
714
715		/*
716		 * i_size updates must to happen from process context.
717		 */
718		if (iomi.pos + iomi.len > dio->i_size)
719			dio->flags |= IOMAP_DIO_COMP_WORK;
720
721		/*
722		 * Try to invalidate cache pages for the range we are writing.
723		 * If this invalidation fails, let the caller fall back to
724		 * buffered I/O.
725		 */
726		ret = kiocb_invalidate_pages(iocb, iomi.len);
727		if (ret) {
728			if (ret != -EAGAIN) {
729				trace_iomap_dio_invalidate_fail(inode, iomi.pos,
730								iomi.len);
731				if (iocb->ki_flags & IOCB_ATOMIC) {
732					/*
733					 * folio invalidation failed, maybe
734					 * this is transient, unlock and see if
735					 * the caller tries again.
736					 */
737					ret = -EAGAIN;
738				} else {
739					/* fall back to buffered write */
740					ret = -ENOTBLK;
741				}
742			}
743			goto out_free_dio;
744		}
745	}
746
747	if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
748		ret = sb_init_dio_done_wq(inode->i_sb);
749		if (ret < 0)
750			goto out_free_dio;
751	}
752
753	inode_dio_begin(inode);
754
755	blk_start_plug(&plug);
756	while ((ret = iomap_iter(&iomi, ops)) > 0) {
757		iomi.status = iomap_dio_iter(&iomi, dio);
758
759		/*
760		 * We can only poll for single bio I/Os.
761		 */
762		iocb->ki_flags &= ~IOCB_HIPRI;
763	}
764
765	blk_finish_plug(&plug);
766
767	/*
768	 * We only report that we've read data up to i_size.
769	 * Revert iter to a state corresponding to that as some callers (such
770	 * as the splice code) rely on it.
771	 */
772	if (iov_iter_rw(iter) == READ && iomi.pos >= dio->i_size)
773		iov_iter_revert(iter, iomi.pos - dio->i_size);
774
775	if (ret == -EFAULT && dio->size && (dio_flags & IOMAP_DIO_PARTIAL)) {
776		if (!(iocb->ki_flags & IOCB_NOWAIT))
777			wait_for_completion = true;
778		ret = 0;
779	}
780
781	/* magic error code to fall back to buffered I/O */
782	if (ret == -ENOTBLK) {
783		wait_for_completion = true;
784		ret = 0;
785	}
786	if (ret < 0)
787		iomap_dio_set_error(dio, ret);
788
789	/*
790	 * If all the writes we issued were already written through to the
791	 * media, we don't need to flush the cache on IO completion. Clear the
792	 * sync flag for this case.
793	 *
794	 * Otherwise clear the inline completion flag if any sync work is
795	 * needed, as that needs to be performed from process context.
796	 */
797	if (dio->flags & IOMAP_DIO_WRITE_THROUGH)
798		dio->flags &= ~IOMAP_DIO_NEED_SYNC;
799	else if (dio->flags & IOMAP_DIO_NEED_SYNC)
800		dio->flags |= IOMAP_DIO_COMP_WORK;
801
802	/*
803	 * We are about to drop our additional submission reference, which
804	 * might be the last reference to the dio.  There are three different
805	 * ways we can progress here:
806	 *
807	 *  (a) If this is the last reference we will always complete and free
808	 *	the dio ourselves.
809	 *  (b) If this is not the last reference, and we serve an asynchronous
810	 *	iocb, we must never touch the dio after the decrement, the
811	 *	I/O completion handler will complete and free it.
812	 *  (c) If this is not the last reference, but we serve a synchronous
813	 *	iocb, the I/O completion handler will wake us up on the drop
814	 *	of the final reference, and we will complete and free it here
815	 *	after we got woken by the I/O completion handler.
816	 */
817	dio->wait_for_completion = wait_for_completion;
818	if (!atomic_dec_and_test(&dio->ref)) {
819		if (!wait_for_completion) {
820			trace_iomap_dio_rw_queued(inode, iomi.pos, iomi.len);
821			return ERR_PTR(-EIOCBQUEUED);
822		}
823
824		for (;;) {
825			set_current_state(TASK_UNINTERRUPTIBLE);
826			if (!READ_ONCE(dio->submit.waiter))
827				break;
828
829			blk_io_schedule();
830		}
831		__set_current_state(TASK_RUNNING);
832	}
833
834	return dio;
835
836out_free_dio:
837	kfree(dio);
838	if (ret)
839		return ERR_PTR(ret);
840	return NULL;
841}
842EXPORT_SYMBOL_GPL(__iomap_dio_rw);
843
844ssize_t
845iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
846		const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
847		unsigned int dio_flags, void *private, size_t done_before)
848{
849	struct iomap_dio *dio;
850
851	dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, private,
852			     done_before);
853	if (IS_ERR_OR_NULL(dio))
854		return PTR_ERR_OR_ZERO(dio);
855	return iomap_dio_complete(dio);
856}
857EXPORT_SYMBOL_GPL(iomap_dio_rw);