io_uring: split provided buffers handling into its own file
[linux-2.6-microblaze.git] / fs / iomap / direct-io.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2010 Red Hat, Inc.
4  * Copyright (c) 2016-2021 Christoph Hellwig.
5  */
6 #include <linux/module.h>
7 #include <linux/compiler.h>
8 #include <linux/fs.h>
9 #include <linux/fscrypt.h>
10 #include <linux/pagemap.h>
11 #include <linux/iomap.h>
12 #include <linux/backing-dev.h>
13 #include <linux/uio.h>
14 #include <linux/task_io_accounting_ops.h>
15 #include "trace.h"
16
17 #include "../internal.h"
18
19 /*
20  * Private flags for iomap_dio, must not overlap with the public ones in
21  * iomap.h:
22  */
23 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
24 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
25 #define IOMAP_DIO_WRITE         (1 << 30)
26 #define IOMAP_DIO_DIRTY         (1 << 31)
27
28 struct iomap_dio {
29         struct kiocb            *iocb;
30         const struct iomap_dio_ops *dops;
31         loff_t                  i_size;
32         loff_t                  size;
33         atomic_t                ref;
34         unsigned                flags;
35         int                     error;
36         size_t                  done_before;
37         bool                    wait_for_completion;
38
39         union {
40                 /* used during submission and for synchronous completion: */
41                 struct {
42                         struct iov_iter         *iter;
43                         struct task_struct      *waiter;
44                         struct bio              *poll_bio;
45                 } submit;
46
47                 /* used for aio completion: */
48                 struct {
49                         struct work_struct      work;
50                 } aio;
51         };
52 };
53
54 static struct bio *iomap_dio_alloc_bio(const struct iomap_iter *iter,
55                 struct iomap_dio *dio, unsigned short nr_vecs, unsigned int opf)
56 {
57         if (dio->dops && dio->dops->bio_set)
58                 return bio_alloc_bioset(iter->iomap.bdev, nr_vecs, opf,
59                                         GFP_KERNEL, dio->dops->bio_set);
60         return bio_alloc(iter->iomap.bdev, nr_vecs, opf, GFP_KERNEL);
61 }
62
63 static void iomap_dio_submit_bio(const struct iomap_iter *iter,
64                 struct iomap_dio *dio, struct bio *bio, loff_t pos)
65 {
66         atomic_inc(&dio->ref);
67
68         /* Sync dio can't be polled reliably */
69         if ((dio->iocb->ki_flags & IOCB_HIPRI) && !is_sync_kiocb(dio->iocb)) {
70                 bio_set_polled(bio, dio->iocb);
71                 dio->submit.poll_bio = bio;
72         }
73
74         if (dio->dops && dio->dops->submit_io)
75                 dio->dops->submit_io(iter, bio, pos);
76         else
77                 submit_bio(bio);
78 }
79
80 ssize_t iomap_dio_complete(struct iomap_dio *dio)
81 {
82         const struct iomap_dio_ops *dops = dio->dops;
83         struct kiocb *iocb = dio->iocb;
84         struct inode *inode = file_inode(iocb->ki_filp);
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                 iocb->ki_pos += ret;
98         }
99
100         /*
101          * Try again to invalidate clean pages which might have been cached by
102          * non-direct readahead, or faulted in by get_user_pages() if the source
103          * of the write was an mmap'ed region of the file we're writing.  Either
104          * one is a pretty crazy thing to do, so we don't support it 100%.  If
105          * this invalidation fails, tough, the write still worked...
106          *
107          * And this page cache invalidation has to be after ->end_io(), as some
108          * filesystems convert unwritten extents to real allocations in
109          * ->end_io() when necessary, otherwise a racing buffer read would cache
110          * zeros from unwritten extents.
111          */
112         if (!dio->error && dio->size &&
113             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
114                 int err;
115                 err = invalidate_inode_pages2_range(inode->i_mapping,
116                                 offset >> PAGE_SHIFT,
117                                 (offset + dio->size - 1) >> PAGE_SHIFT);
118                 if (err)
119                         dio_warn_stale_pagecache(iocb->ki_filp);
120         }
121
122         inode_dio_end(file_inode(iocb->ki_filp));
123         /*
124          * If this is a DSYNC write, make sure we push it to stable storage now
125          * that we've written data.
126          */
127         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
128                 ret = generic_write_sync(iocb, ret);
129
130         if (ret > 0)
131                 ret += dio->done_before;
132
133         kfree(dio);
134
135         return ret;
136 }
137 EXPORT_SYMBOL_GPL(iomap_dio_complete);
138
139 static void iomap_dio_complete_work(struct work_struct *work)
140 {
141         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
142         struct kiocb *iocb = dio->iocb;
143
144         iocb->ki_complete(iocb, iomap_dio_complete(dio));
145 }
146
147 /*
148  * Set an error in the dio if none is set yet.  We have to use cmpxchg
149  * as the submission context and the completion context(s) can race to
150  * update the error.
151  */
152 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
153 {
154         cmpxchg(&dio->error, 0, ret);
155 }
156
157 void iomap_dio_bio_end_io(struct bio *bio)
158 {
159         struct iomap_dio *dio = bio->bi_private;
160         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
161
162         if (bio->bi_status)
163                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
164
165         if (atomic_dec_and_test(&dio->ref)) {
166                 if (dio->wait_for_completion) {
167                         struct task_struct *waiter = dio->submit.waiter;
168                         WRITE_ONCE(dio->submit.waiter, NULL);
169                         blk_wake_io_task(waiter);
170                 } else if (dio->flags & IOMAP_DIO_WRITE) {
171                         struct inode *inode = file_inode(dio->iocb->ki_filp);
172
173                         WRITE_ONCE(dio->iocb->private, NULL);
174                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
175                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
176                 } else {
177                         WRITE_ONCE(dio->iocb->private, NULL);
178                         iomap_dio_complete_work(&dio->aio.work);
179                 }
180         }
181
182         if (should_dirty) {
183                 bio_check_pages_dirty(bio);
184         } else {
185                 bio_release_pages(bio, false);
186                 bio_put(bio);
187         }
188 }
189 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
190
191 static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
192                 loff_t pos, unsigned len)
193 {
194         struct inode *inode = file_inode(dio->iocb->ki_filp);
195         struct page *page = ZERO_PAGE(0);
196         struct bio *bio;
197
198         bio = iomap_dio_alloc_bio(iter, dio, 1, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
199         fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
200                                   GFP_KERNEL);
201         bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
202         bio->bi_private = dio;
203         bio->bi_end_io = iomap_dio_bio_end_io;
204
205         get_page(page);
206         __bio_add_page(bio, page, len, 0);
207         iomap_dio_submit_bio(iter, dio, bio, pos);
208 }
209
210 /*
211  * Figure out the bio's operation flags from the dio request, the
212  * mapping, and whether or not we want FUA.  Note that we can end up
213  * clearing the WRITE_FUA flag in the dio request.
214  */
215 static inline unsigned int iomap_dio_bio_opflags(struct iomap_dio *dio,
216                 const struct iomap *iomap, bool use_fua)
217 {
218         unsigned int opflags = REQ_SYNC | REQ_IDLE;
219
220         if (!(dio->flags & IOMAP_DIO_WRITE)) {
221                 WARN_ON_ONCE(iomap->flags & IOMAP_F_ZONE_APPEND);
222                 return REQ_OP_READ;
223         }
224
225         if (iomap->flags & IOMAP_F_ZONE_APPEND)
226                 opflags |= REQ_OP_ZONE_APPEND;
227         else
228                 opflags |= REQ_OP_WRITE;
229
230         if (use_fua)
231                 opflags |= REQ_FUA;
232         else
233                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
234
235         return opflags;
236 }
237
238 static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
239                 struct iomap_dio *dio)
240 {
241         const struct iomap *iomap = &iter->iomap;
242         struct inode *inode = iter->inode;
243         unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
244         unsigned int fs_block_size = i_blocksize(inode), pad;
245         unsigned int align = iov_iter_alignment(dio->submit.iter);
246         loff_t length = iomap_length(iter);
247         loff_t pos = iter->pos;
248         unsigned int bio_opf;
249         struct bio *bio;
250         bool need_zeroout = false;
251         bool use_fua = false;
252         int nr_pages, ret = 0;
253         size_t copied = 0;
254         size_t orig_count;
255
256         if ((pos | length | align) & ((1 << blkbits) - 1))
257                 return -EINVAL;
258
259         if (iomap->type == IOMAP_UNWRITTEN) {
260                 dio->flags |= IOMAP_DIO_UNWRITTEN;
261                 need_zeroout = true;
262         }
263
264         if (iomap->flags & IOMAP_F_SHARED)
265                 dio->flags |= IOMAP_DIO_COW;
266
267         if (iomap->flags & IOMAP_F_NEW) {
268                 need_zeroout = true;
269         } else if (iomap->type == IOMAP_MAPPED) {
270                 /*
271                  * Use a FUA write if we need datasync semantics, this is a pure
272                  * data IO that doesn't require any metadata updates (including
273                  * after IO completion such as unwritten extent conversion) and
274                  * the underlying device supports FUA. This allows us to avoid
275                  * cache flushes on IO completion.
276                  */
277                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
278                     (dio->flags & IOMAP_DIO_WRITE_FUA) && bdev_fua(iomap->bdev))
279                         use_fua = true;
280         }
281
282         /*
283          * Save the original count and trim the iter to just the extent we
284          * are operating on right now.  The iter will be re-expanded once
285          * we are done.
286          */
287         orig_count = iov_iter_count(dio->submit.iter);
288         iov_iter_truncate(dio->submit.iter, length);
289
290         if (!iov_iter_count(dio->submit.iter))
291                 goto out;
292
293         /*
294          * We can only poll for single bio I/Os.
295          */
296         if (need_zeroout ||
297             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode)))
298                 dio->iocb->ki_flags &= ~IOCB_HIPRI;
299
300         if (need_zeroout) {
301                 /* zero out from the start of the block to the write offset */
302                 pad = pos & (fs_block_size - 1);
303                 if (pad)
304                         iomap_dio_zero(iter, dio, pos - pad, pad);
305         }
306
307         /*
308          * Set the operation flags early so that bio_iov_iter_get_pages
309          * can set up the page vector appropriately for a ZONE_APPEND
310          * operation.
311          */
312         bio_opf = iomap_dio_bio_opflags(dio, iomap, use_fua);
313
314         nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
315         do {
316                 size_t n;
317                 if (dio->error) {
318                         iov_iter_revert(dio->submit.iter, copied);
319                         copied = ret = 0;
320                         goto out;
321                 }
322
323                 bio = iomap_dio_alloc_bio(iter, dio, nr_pages, bio_opf);
324                 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
325                                           GFP_KERNEL);
326                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
327                 bio->bi_ioprio = dio->iocb->ki_ioprio;
328                 bio->bi_private = dio;
329                 bio->bi_end_io = iomap_dio_bio_end_io;
330
331                 ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
332                 if (unlikely(ret)) {
333                         /*
334                          * We have to stop part way through an IO. We must fall
335                          * through to the sub-block tail zeroing here, otherwise
336                          * this short IO may expose stale data in the tail of
337                          * the block we haven't written data to.
338                          */
339                         bio_put(bio);
340                         goto zero_tail;
341                 }
342
343                 n = bio->bi_iter.bi_size;
344                 if (dio->flags & IOMAP_DIO_WRITE) {
345                         task_io_account_write(n);
346                 } else {
347                         if (dio->flags & IOMAP_DIO_DIRTY)
348                                 bio_set_pages_dirty(bio);
349                 }
350
351                 dio->size += n;
352                 copied += n;
353
354                 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter,
355                                                  BIO_MAX_VECS);
356                 /*
357                  * We can only poll for single bio I/Os.
358                  */
359                 if (nr_pages)
360                         dio->iocb->ki_flags &= ~IOCB_HIPRI;
361                 iomap_dio_submit_bio(iter, dio, bio, pos);
362                 pos += n;
363         } while (nr_pages);
364
365         /*
366          * We need to zeroout the tail of a sub-block write if the extent type
367          * requires zeroing or the write extends beyond EOF. If we don't zero
368          * the block tail in the latter case, we can expose stale data via mmap
369          * reads of the EOF block.
370          */
371 zero_tail:
372         if (need_zeroout ||
373             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
374                 /* zero out from the end of the write to the end of the block */
375                 pad = pos & (fs_block_size - 1);
376                 if (pad)
377                         iomap_dio_zero(iter, dio, pos, fs_block_size - pad);
378         }
379 out:
380         /* Undo iter limitation to current extent */
381         iov_iter_reexpand(dio->submit.iter, orig_count - copied);
382         if (copied)
383                 return copied;
384         return ret;
385 }
386
387 static loff_t iomap_dio_hole_iter(const struct iomap_iter *iter,
388                 struct iomap_dio *dio)
389 {
390         loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter);
391
392         dio->size += length;
393         if (!length)
394                 return -EFAULT;
395         return length;
396 }
397
398 static loff_t iomap_dio_inline_iter(const struct iomap_iter *iomi,
399                 struct iomap_dio *dio)
400 {
401         const struct iomap *iomap = &iomi->iomap;
402         struct iov_iter *iter = dio->submit.iter;
403         void *inline_data = iomap_inline_data(iomap, iomi->pos);
404         loff_t length = iomap_length(iomi);
405         loff_t pos = iomi->pos;
406         size_t copied;
407
408         if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap)))
409                 return -EIO;
410
411         if (dio->flags & IOMAP_DIO_WRITE) {
412                 loff_t size = iomi->inode->i_size;
413
414                 if (pos > size)
415                         memset(iomap_inline_data(iomap, size), 0, pos - size);
416                 copied = copy_from_iter(inline_data, length, iter);
417                 if (copied) {
418                         if (pos + copied > size)
419                                 i_size_write(iomi->inode, pos + copied);
420                         mark_inode_dirty(iomi->inode);
421                 }
422         } else {
423                 copied = copy_to_iter(inline_data, length, iter);
424         }
425         dio->size += copied;
426         if (!copied)
427                 return -EFAULT;
428         return copied;
429 }
430
431 static loff_t iomap_dio_iter(const struct iomap_iter *iter,
432                 struct iomap_dio *dio)
433 {
434         switch (iter->iomap.type) {
435         case IOMAP_HOLE:
436                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
437                         return -EIO;
438                 return iomap_dio_hole_iter(iter, dio);
439         case IOMAP_UNWRITTEN:
440                 if (!(dio->flags & IOMAP_DIO_WRITE))
441                         return iomap_dio_hole_iter(iter, dio);
442                 return iomap_dio_bio_iter(iter, dio);
443         case IOMAP_MAPPED:
444                 return iomap_dio_bio_iter(iter, dio);
445         case IOMAP_INLINE:
446                 return iomap_dio_inline_iter(iter, dio);
447         case IOMAP_DELALLOC:
448                 /*
449                  * DIO is not serialised against mmap() access at all, and so
450                  * if the page_mkwrite occurs between the writeback and the
451                  * iomap_iter() call in the DIO path, then it will see the
452                  * DELALLOC block that the page-mkwrite allocated.
453                  */
454                 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
455                                     dio->iocb->ki_filp, current->comm);
456                 return -EIO;
457         default:
458                 WARN_ON_ONCE(1);
459                 return -EIO;
460         }
461 }
462
463 /*
464  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
465  * is being issued as AIO or not.  This allows us to optimise pure data writes
466  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
467  * REQ_FLUSH post write. This is slightly tricky because a single request here
468  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
469  * may be pure data writes. In that case, we still need to do a full data sync
470  * completion.
471  *
472  * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL,
473  * __iomap_dio_rw can return a partial result if it encounters a non-resident
474  * page in @iter after preparing a transfer.  In that case, the non-resident
475  * pages can be faulted in and the request resumed with @done_before set to the
476  * number of bytes previously transferred.  The request will then complete with
477  * the correct total number of bytes transferred; this is essential for
478  * completing partial requests asynchronously.
479  *
480  * Returns -ENOTBLK In case of a page invalidation invalidation failure for
481  * writes.  The callers needs to fall back to buffered I/O in this case.
482  */
483 struct iomap_dio *
484 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
485                 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
486                 unsigned int dio_flags, void *private, size_t done_before)
487 {
488         struct address_space *mapping = iocb->ki_filp->f_mapping;
489         struct inode *inode = file_inode(iocb->ki_filp);
490         struct iomap_iter iomi = {
491                 .inode          = inode,
492                 .pos            = iocb->ki_pos,
493                 .len            = iov_iter_count(iter),
494                 .flags          = IOMAP_DIRECT,
495                 .private        = private,
496         };
497         loff_t end = iomi.pos + iomi.len - 1, ret = 0;
498         bool wait_for_completion =
499                 is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
500         struct blk_plug plug;
501         struct iomap_dio *dio;
502
503         if (!iomi.len)
504                 return NULL;
505
506         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
507         if (!dio)
508                 return ERR_PTR(-ENOMEM);
509
510         dio->iocb = iocb;
511         atomic_set(&dio->ref, 1);
512         dio->size = 0;
513         dio->i_size = i_size_read(inode);
514         dio->dops = dops;
515         dio->error = 0;
516         dio->flags = 0;
517         dio->done_before = done_before;
518
519         dio->submit.iter = iter;
520         dio->submit.waiter = current;
521         dio->submit.poll_bio = NULL;
522
523         if (iov_iter_rw(iter) == READ) {
524                 if (iomi.pos >= dio->i_size)
525                         goto out_free_dio;
526
527                 if (iocb->ki_flags & IOCB_NOWAIT) {
528                         if (filemap_range_needs_writeback(mapping, iomi.pos,
529                                         end)) {
530                                 ret = -EAGAIN;
531                                 goto out_free_dio;
532                         }
533                         iomi.flags |= IOMAP_NOWAIT;
534                 }
535
536                 if (iter_is_iovec(iter))
537                         dio->flags |= IOMAP_DIO_DIRTY;
538         } else {
539                 iomi.flags |= IOMAP_WRITE;
540                 dio->flags |= IOMAP_DIO_WRITE;
541
542                 if (iocb->ki_flags & IOCB_NOWAIT) {
543                         if (filemap_range_has_page(mapping, iomi.pos, end)) {
544                                 ret = -EAGAIN;
545                                 goto out_free_dio;
546                         }
547                         iomi.flags |= IOMAP_NOWAIT;
548                 }
549
550                 /* for data sync or sync, we need sync completion processing */
551                 if (iocb->ki_flags & IOCB_DSYNC)
552                         dio->flags |= IOMAP_DIO_NEED_SYNC;
553
554                 /*
555                  * For datasync only writes, we optimistically try using FUA for
556                  * this IO.  Any non-FUA write that occurs will clear this flag,
557                  * hence we know before completion whether a cache flush is
558                  * necessary.
559                  */
560                 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
561                         dio->flags |= IOMAP_DIO_WRITE_FUA;
562         }
563
564         if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) {
565                 ret = -EAGAIN;
566                 if (iomi.pos >= dio->i_size ||
567                     iomi.pos + iomi.len > dio->i_size)
568                         goto out_free_dio;
569                 iomi.flags |= IOMAP_OVERWRITE_ONLY;
570         }
571
572         ret = filemap_write_and_wait_range(mapping, iomi.pos, end);
573         if (ret)
574                 goto out_free_dio;
575
576         if (iov_iter_rw(iter) == WRITE) {
577                 /*
578                  * Try to invalidate cache pages for the range we are writing.
579                  * If this invalidation fails, let the caller fall back to
580                  * buffered I/O.
581                  */
582                 if (invalidate_inode_pages2_range(mapping,
583                                 iomi.pos >> PAGE_SHIFT, end >> PAGE_SHIFT)) {
584                         trace_iomap_dio_invalidate_fail(inode, iomi.pos,
585                                                         iomi.len);
586                         ret = -ENOTBLK;
587                         goto out_free_dio;
588                 }
589
590                 if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
591                         ret = sb_init_dio_done_wq(inode->i_sb);
592                         if (ret < 0)
593                                 goto out_free_dio;
594                 }
595         }
596
597         inode_dio_begin(inode);
598
599         blk_start_plug(&plug);
600         while ((ret = iomap_iter(&iomi, ops)) > 0) {
601                 iomi.processed = iomap_dio_iter(&iomi, dio);
602
603                 /*
604                  * We can only poll for single bio I/Os.
605                  */
606                 iocb->ki_flags &= ~IOCB_HIPRI;
607         }
608
609         blk_finish_plug(&plug);
610
611         /*
612          * We only report that we've read data up to i_size.
613          * Revert iter to a state corresponding to that as some callers (such
614          * as the splice code) rely on it.
615          */
616         if (iov_iter_rw(iter) == READ && iomi.pos >= dio->i_size)
617                 iov_iter_revert(iter, iomi.pos - dio->i_size);
618
619         if (ret == -EFAULT && dio->size && (dio_flags & IOMAP_DIO_PARTIAL)) {
620                 if (!(iocb->ki_flags & IOCB_NOWAIT))
621                         wait_for_completion = true;
622                 ret = 0;
623         }
624
625         /* magic error code to fall back to buffered I/O */
626         if (ret == -ENOTBLK) {
627                 wait_for_completion = true;
628                 ret = 0;
629         }
630         if (ret < 0)
631                 iomap_dio_set_error(dio, ret);
632
633         /*
634          * If all the writes we issued were FUA, we don't need to flush the
635          * cache on IO completion. Clear the sync flag for this case.
636          */
637         if (dio->flags & IOMAP_DIO_WRITE_FUA)
638                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
639
640         WRITE_ONCE(iocb->private, dio->submit.poll_bio);
641
642         /*
643          * We are about to drop our additional submission reference, which
644          * might be the last reference to the dio.  There are three different
645          * ways we can progress here:
646          *
647          *  (a) If this is the last reference we will always complete and free
648          *      the dio ourselves.
649          *  (b) If this is not the last reference, and we serve an asynchronous
650          *      iocb, we must never touch the dio after the decrement, the
651          *      I/O completion handler will complete and free it.
652          *  (c) If this is not the last reference, but we serve a synchronous
653          *      iocb, the I/O completion handler will wake us up on the drop
654          *      of the final reference, and we will complete and free it here
655          *      after we got woken by the I/O completion handler.
656          */
657         dio->wait_for_completion = wait_for_completion;
658         if (!atomic_dec_and_test(&dio->ref)) {
659                 if (!wait_for_completion)
660                         return ERR_PTR(-EIOCBQUEUED);
661
662                 for (;;) {
663                         set_current_state(TASK_UNINTERRUPTIBLE);
664                         if (!READ_ONCE(dio->submit.waiter))
665                                 break;
666
667                         blk_io_schedule();
668                 }
669                 __set_current_state(TASK_RUNNING);
670         }
671
672         return dio;
673
674 out_free_dio:
675         kfree(dio);
676         if (ret)
677                 return ERR_PTR(ret);
678         return NULL;
679 }
680 EXPORT_SYMBOL_GPL(__iomap_dio_rw);
681
682 ssize_t
683 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
684                 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
685                 unsigned int dio_flags, void *private, size_t done_before)
686 {
687         struct iomap_dio *dio;
688
689         dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, private,
690                              done_before);
691         if (IS_ERR_OR_NULL(dio))
692                 return PTR_ERR_OR_ZERO(dio);
693         return iomap_dio_complete(dio);
694 }
695 EXPORT_SYMBOL_GPL(iomap_dio_rw);