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/blk-crypto.h>
7 #include <linux/fscrypt.h>
8 #include <linux/pagemap.h>
9 #include <linux/iomap.h>
10 #include <linux/task_io_accounting_ops.h>
11 #include <linux/fserror.h>
12 #include "internal.h"
13 #include "trace.h"
14
15 #include "../internal.h"
16
17 /*
18 * Private flags for iomap_dio, must not overlap with the public ones in
19 * iomap.h:
20 */
21 #define IOMAP_DIO_NO_INVALIDATE (1U << 26)
22 #define IOMAP_DIO_COMP_WORK (1U << 27)
23 #define IOMAP_DIO_WRITE_THROUGH (1U << 28)
24 #define IOMAP_DIO_NEED_SYNC (1U << 29)
25 #define IOMAP_DIO_WRITE (1U << 30)
26 #define IOMAP_DIO_USER_BACKED (1U << 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 } submit;
45
46 /* used for aio completion: */
47 struct {
48 struct work_struct work;
49 } aio;
50 };
51 };
52
iomap_dio_alloc_bio(const struct iomap_iter * iter,struct iomap_dio * dio,unsigned short nr_vecs,blk_opf_t opf)53 static struct bio *iomap_dio_alloc_bio(const struct iomap_iter *iter,
54 struct iomap_dio *dio, unsigned short nr_vecs, blk_opf_t opf)
55 {
56 if (dio->dops && dio->dops->bio_set)
57 return bio_alloc_bioset(iter->iomap.bdev, nr_vecs, opf,
58 GFP_KERNEL, dio->dops->bio_set);
59 return bio_alloc(iter->iomap.bdev, nr_vecs, opf, GFP_KERNEL);
60 }
61
iomap_dio_submit_bio(const struct iomap_iter * iter,struct iomap_dio * dio,struct bio * bio,loff_t pos)62 static void iomap_dio_submit_bio(const struct iomap_iter *iter,
63 struct iomap_dio *dio, struct bio *bio, loff_t pos)
64 {
65 struct kiocb *iocb = dio->iocb;
66
67 atomic_inc(&dio->ref);
68
69 /* Sync dio can't be polled reliably */
70 if ((iocb->ki_flags & IOCB_HIPRI) && !is_sync_kiocb(iocb)) {
71 bio_set_polled(bio, iocb);
72 WRITE_ONCE(iocb->private, bio);
73 }
74
75 if (dio->dops && dio->dops->submit_io) {
76 dio->dops->submit_io(iter, bio, pos);
77 } else {
78 WARN_ON_ONCE(iter->iomap.flags & IOMAP_F_ANON_WRITE);
79 blk_crypto_submit_bio(bio);
80 }
81 }
82
iomap_dio_err_type(const struct iomap_dio * dio)83 static inline enum fserror_type iomap_dio_err_type(const struct iomap_dio *dio)
84 {
85 if (dio->flags & IOMAP_DIO_WRITE)
86 return FSERR_DIRECTIO_WRITE;
87 return FSERR_DIRECTIO_READ;
88 }
89
should_report_dio_fserror(const struct iomap_dio * dio)90 static inline bool should_report_dio_fserror(const struct iomap_dio *dio)
91 {
92 switch (dio->error) {
93 case 0:
94 case -EAGAIN:
95 case -ENOTBLK:
96 /* don't send fsnotify for success or magic retry codes */
97 return false;
98 default:
99 return true;
100 }
101 }
102
iomap_dio_complete(struct iomap_dio * dio)103 ssize_t iomap_dio_complete(struct iomap_dio *dio)
104 {
105 const struct iomap_dio_ops *dops = dio->dops;
106 struct kiocb *iocb = dio->iocb;
107 loff_t offset = iocb->ki_pos;
108 ssize_t ret = dio->error;
109
110 if (dops && dops->end_io)
111 ret = dops->end_io(iocb, dio->size, ret, dio->flags);
112 if (should_report_dio_fserror(dio))
113 fserror_report_io(file_inode(iocb->ki_filp),
114 iomap_dio_err_type(dio), offset, dio->size,
115 dio->error, GFP_NOFS);
116
117 if (likely(!ret)) {
118 ret = dio->size;
119 /* check for short read */
120 if (offset + ret > dio->i_size &&
121 !(dio->flags & IOMAP_DIO_WRITE))
122 ret = dio->i_size - offset;
123 }
124
125 /*
126 * Try again to invalidate clean pages which might have been cached by
127 * non-direct readahead, or faulted in by get_user_pages() if the source
128 * of the write was an mmap'ed region of the file we're writing. Either
129 * one is a pretty crazy thing to do, so we don't support it 100%. If
130 * this invalidation fails, tough, the write still worked...
131 *
132 * And this page cache invalidation has to be after ->end_io(), as some
133 * filesystems convert unwritten extents to real allocations in
134 * ->end_io() when necessary, otherwise a racing buffer read would cache
135 * zeros from unwritten extents.
136 */
137 if (!dio->error && dio->size && (dio->flags & IOMAP_DIO_WRITE) &&
138 !(dio->flags & IOMAP_DIO_NO_INVALIDATE))
139 kiocb_invalidate_post_direct_write(iocb, dio->size);
140
141 inode_dio_end(file_inode(iocb->ki_filp));
142
143 if (ret > 0) {
144 iocb->ki_pos += ret;
145
146 /*
147 * If this is a DSYNC write, make sure we push it to stable
148 * storage now that we've written data.
149 */
150 if (dio->flags & IOMAP_DIO_NEED_SYNC)
151 ret = generic_write_sync(iocb, ret);
152 if (ret > 0)
153 ret += dio->done_before;
154 }
155 trace_iomap_dio_complete(iocb, dio->error, ret);
156 kfree(dio);
157 return ret;
158 }
159 EXPORT_SYMBOL_GPL(iomap_dio_complete);
160
iomap_dio_complete_work(struct work_struct * work)161 static void iomap_dio_complete_work(struct work_struct *work)
162 {
163 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
164 struct kiocb *iocb = dio->iocb;
165
166 iocb->ki_complete(iocb, iomap_dio_complete(dio));
167 }
168
169 /*
170 * Set an error in the dio if none is set yet. We have to use cmpxchg
171 * as the submission context and the completion context(s) can race to
172 * update the error.
173 */
iomap_dio_set_error(struct iomap_dio * dio,int ret)174 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
175 {
176 cmpxchg(&dio->error, 0, ret);
177 }
178
179 /*
180 * Called when dio->ref reaches zero from an I/O completion.
181 */
iomap_dio_done(struct iomap_dio * dio)182 static void iomap_dio_done(struct iomap_dio *dio)
183 {
184 struct kiocb *iocb = dio->iocb;
185
186 if (dio->wait_for_completion) {
187 /*
188 * Synchronous I/O, task itself will handle any completion work
189 * that needs after IO. All we need to do is wake the task.
190 */
191 struct task_struct *waiter = dio->submit.waiter;
192
193 WRITE_ONCE(dio->submit.waiter, NULL);
194 blk_wake_io_task(waiter);
195 return;
196 }
197
198 /*
199 * Always run error completions in user context. These are not
200 * performance critical and some code relies on taking sleeping locks
201 * for error handling.
202 */
203 if (dio->error)
204 dio->flags |= IOMAP_DIO_COMP_WORK;
205
206 /*
207 * Never invalidate pages from this context to avoid deadlocks with
208 * buffered I/O completions when called from the ioend workqueue,
209 * or avoid sleeping when called directly from ->bi_end_io.
210 * Tough luck if you hit the tiny race with someone dirtying the range
211 * right between this check and the actual completion.
212 */
213 if ((dio->flags & IOMAP_DIO_WRITE) &&
214 !(dio->flags & IOMAP_DIO_COMP_WORK)) {
215 if (dio->iocb->ki_filp->f_mapping->nrpages)
216 dio->flags |= IOMAP_DIO_COMP_WORK;
217 else
218 dio->flags |= IOMAP_DIO_NO_INVALIDATE;
219 }
220
221 if (dio->flags & IOMAP_DIO_COMP_WORK) {
222 struct inode *inode = file_inode(iocb->ki_filp);
223
224 /*
225 * Async DIO completion that requires filesystem level
226 * completion work gets punted to a work queue to complete as
227 * the operation may require more IO to be issued to finalise
228 * filesystem metadata changes or guarantee data integrity.
229 */
230 INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
231 queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
232 return;
233 }
234
235 WRITE_ONCE(iocb->private, NULL);
236 iomap_dio_complete_work(&dio->aio.work);
237 }
238
__iomap_dio_bio_end_io(struct bio * bio,bool inline_completion)239 static void __iomap_dio_bio_end_io(struct bio *bio, bool inline_completion)
240 {
241 struct iomap_dio *dio = bio->bi_private;
242
243 if (dio->flags & IOMAP_DIO_BOUNCE) {
244 bio_iov_iter_unbounce(bio, !!dio->error,
245 dio->flags & IOMAP_DIO_USER_BACKED);
246 bio_put(bio);
247 } else if (dio->flags & IOMAP_DIO_USER_BACKED) {
248 bio_check_pages_dirty(bio);
249 } else {
250 bio_release_pages(bio, false);
251 bio_put(bio);
252 }
253
254 /* Do not touch bio below, we just gave up our reference. */
255
256 if (atomic_dec_and_test(&dio->ref)) {
257 /*
258 * Avoid another context switch for the completion when already
259 * called from the ioend completion workqueue.
260 */
261 if (inline_completion)
262 dio->flags &= ~IOMAP_DIO_COMP_WORK;
263 iomap_dio_done(dio);
264 }
265 }
266
iomap_dio_bio_end_io(struct bio * bio)267 void iomap_dio_bio_end_io(struct bio *bio)
268 {
269 struct iomap_dio *dio = bio->bi_private;
270
271 if (bio->bi_status)
272 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
273 __iomap_dio_bio_end_io(bio, false);
274 }
275 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
276
iomap_finish_ioend_direct(struct iomap_ioend * ioend)277 u32 iomap_finish_ioend_direct(struct iomap_ioend *ioend)
278 {
279 struct iomap_dio *dio = ioend->io_bio.bi_private;
280 u32 vec_count = ioend->io_bio.bi_vcnt;
281
282 if (ioend->io_error)
283 iomap_dio_set_error(dio, ioend->io_error);
284 __iomap_dio_bio_end_io(&ioend->io_bio, true);
285
286 /*
287 * Return the number of bvecs completed as even direct I/O completions
288 * do significant per-folio work and we'll still want to give up the
289 * CPU after a lot of completions.
290 */
291 return vec_count;
292 }
293
iomap_dio_zero(const struct iomap_iter * iter,struct iomap_dio * dio,loff_t pos,unsigned len)294 static int iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
295 loff_t pos, unsigned len)
296 {
297 struct inode *inode = file_inode(dio->iocb->ki_filp);
298 struct bio *bio;
299 struct folio *zero_folio = largest_zero_folio();
300 int nr_vecs = max(1, i_blocksize(inode) / folio_size(zero_folio));
301
302 if (!len)
303 return 0;
304
305 /*
306 * This limit shall never be reached as most filesystems have a
307 * maximum blocksize of 64k.
308 */
309 if (WARN_ON_ONCE(nr_vecs > BIO_MAX_VECS))
310 return -EINVAL;
311
312 bio = iomap_dio_alloc_bio(iter, dio, nr_vecs,
313 REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
314 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
315 GFP_KERNEL);
316 bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
317 bio->bi_private = dio;
318 bio->bi_end_io = iomap_dio_bio_end_io;
319
320 while (len > 0) {
321 unsigned int io_len = min(len, folio_size(zero_folio));
322
323 bio_add_folio_nofail(bio, zero_folio, io_len, 0);
324 len -= io_len;
325 }
326 iomap_dio_submit_bio(iter, dio, bio, pos);
327
328 return 0;
329 }
330
iomap_dio_bio_iter_one(struct iomap_iter * iter,struct iomap_dio * dio,loff_t pos,unsigned int alignment,blk_opf_t op)331 static ssize_t iomap_dio_bio_iter_one(struct iomap_iter *iter,
332 struct iomap_dio *dio, loff_t pos, unsigned int alignment,
333 blk_opf_t op)
334 {
335 unsigned int nr_vecs;
336 struct bio *bio;
337 ssize_t ret;
338
339 if (dio->flags & IOMAP_DIO_BOUNCE)
340 nr_vecs = bio_iov_bounce_nr_vecs(dio->submit.iter, op);
341 else
342 nr_vecs = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
343
344 bio = iomap_dio_alloc_bio(iter, dio, nr_vecs, op);
345 fscrypt_set_bio_crypt_ctx(bio, iter->inode,
346 pos >> iter->inode->i_blkbits, GFP_KERNEL);
347 bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
348 bio->bi_write_hint = iter->inode->i_write_hint;
349 bio->bi_ioprio = dio->iocb->ki_ioprio;
350 bio->bi_private = dio;
351 bio->bi_end_io = iomap_dio_bio_end_io;
352
353 if (dio->flags & IOMAP_DIO_BOUNCE)
354 ret = bio_iov_iter_bounce(bio, dio->submit.iter);
355 else
356 ret = bio_iov_iter_get_pages(bio, dio->submit.iter,
357 alignment - 1);
358 if (unlikely(ret))
359 goto out_put_bio;
360 ret = bio->bi_iter.bi_size;
361
362 /*
363 * An atomic write bio must cover the complete length. If it doesn't,
364 * error out.
365 */
366 if ((op & REQ_ATOMIC) && WARN_ON_ONCE(ret != iomap_length(iter))) {
367 ret = -EINVAL;
368 goto out_put_bio;
369 }
370
371 if (dio->flags & IOMAP_DIO_WRITE)
372 task_io_account_write(ret);
373 else if ((dio->flags & IOMAP_DIO_USER_BACKED) &&
374 !(dio->flags & IOMAP_DIO_BOUNCE))
375 bio_set_pages_dirty(bio);
376
377 /*
378 * We can only poll for single bio I/Os.
379 */
380 if (iov_iter_count(dio->submit.iter))
381 dio->iocb->ki_flags &= ~IOCB_HIPRI;
382 iomap_dio_submit_bio(iter, dio, bio, pos);
383 return ret;
384
385 out_put_bio:
386 bio_put(bio);
387 return ret;
388 }
389
iomap_dio_bio_iter(struct iomap_iter * iter,struct iomap_dio * dio)390 static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
391 {
392 const struct iomap *iomap = &iter->iomap;
393 struct inode *inode = iter->inode;
394 unsigned int fs_block_size = i_blocksize(inode), pad;
395 const loff_t length = iomap_length(iter);
396 loff_t pos = iter->pos;
397 blk_opf_t bio_opf = REQ_SYNC | REQ_IDLE;
398 bool need_zeroout = false;
399 u64 copied = 0;
400 size_t orig_count;
401 unsigned int alignment;
402 ssize_t ret = 0;
403
404 /*
405 * File systems that write out of place and always allocate new blocks
406 * need each bio to be block aligned as that's the unit of allocation.
407 */
408 if (dio->flags & IOMAP_DIO_FSBLOCK_ALIGNED)
409 alignment = fs_block_size;
410 else
411 alignment = bdev_logical_block_size(iomap->bdev);
412
413 if ((pos | length) & (alignment - 1))
414 return -EINVAL;
415
416 if (dio->flags & IOMAP_DIO_WRITE) {
417 bool need_completion_work = true;
418
419 switch (iomap->type) {
420 case IOMAP_MAPPED:
421 /*
422 * Directly mapped I/O does not inherently need to do
423 * work at I/O completion time. But there are various
424 * cases below where this will get set again.
425 */
426 need_completion_work = false;
427 break;
428 case IOMAP_UNWRITTEN:
429 dio->flags |= IOMAP_DIO_UNWRITTEN;
430 need_zeroout = true;
431 break;
432 default:
433 break;
434 }
435
436 if (iomap->flags & IOMAP_F_ATOMIC_BIO) {
437 /*
438 * Ensure that the mapping covers the full write
439 * length, otherwise it won't be submitted as a single
440 * bio, which is required to use hardware atomics.
441 */
442 if (length != iter->len)
443 return -EINVAL;
444 bio_opf |= REQ_ATOMIC;
445 }
446
447 if (iomap->flags & IOMAP_F_SHARED) {
448 /*
449 * Unsharing of needs to update metadata at I/O
450 * completion time.
451 */
452 need_completion_work = true;
453 dio->flags |= IOMAP_DIO_COW;
454 }
455
456 if (iomap->flags & IOMAP_F_NEW) {
457 /*
458 * Newly allocated blocks might need recording in
459 * metadata at I/O completion time.
460 */
461 need_completion_work = true;
462 need_zeroout = true;
463 }
464
465 /*
466 * Use a FUA write if we need datasync semantics and this is a
467 * pure overwrite that doesn't require any metadata updates.
468 *
469 * This allows us to avoid cache flushes on I/O completion.
470 */
471 if (dio->flags & IOMAP_DIO_WRITE_THROUGH) {
472 if (!need_completion_work &&
473 !(iomap->flags & IOMAP_F_DIRTY) &&
474 (!bdev_write_cache(iomap->bdev) ||
475 bdev_fua(iomap->bdev)))
476 bio_opf |= REQ_FUA;
477 else
478 dio->flags &= ~IOMAP_DIO_WRITE_THROUGH;
479 }
480
481 /*
482 * We can only do inline completion for pure overwrites that
483 * don't require additional I/O at completion time.
484 *
485 * This rules out writes that need zeroing or metdata updates to
486 * convert unwritten or shared extents.
487 *
488 * Writes that extend i_size are also not supported, but this is
489 * handled in __iomap_dio_rw().
490 */
491 if (need_completion_work)
492 dio->flags |= IOMAP_DIO_COMP_WORK;
493
494 bio_opf |= REQ_OP_WRITE;
495 } else {
496 bio_opf |= REQ_OP_READ;
497 }
498
499 /*
500 * Save the original count and trim the iter to just the extent we
501 * are operating on right now. The iter will be re-expanded once
502 * we are done.
503 */
504 orig_count = iov_iter_count(dio->submit.iter);
505 iov_iter_truncate(dio->submit.iter, length);
506
507 if (!iov_iter_count(dio->submit.iter))
508 goto out;
509
510 /*
511 * The rules for polled IO completions follow the guidelines as the
512 * ones we set for inline and deferred completions. If none of those
513 * are available for this IO, clear the polled flag.
514 */
515 if (dio->flags & IOMAP_DIO_COMP_WORK)
516 dio->iocb->ki_flags &= ~IOCB_HIPRI;
517
518 if (need_zeroout) {
519 /* zero out from the start of the block to the write offset */
520 pad = pos & (fs_block_size - 1);
521
522 ret = iomap_dio_zero(iter, dio, pos - pad, pad);
523 if (ret)
524 goto out;
525 }
526
527 do {
528 /*
529 * If completions already occurred and reported errors, give up now and
530 * don't bother submitting more bios.
531 */
532 if (unlikely(data_race(dio->error)))
533 goto out;
534
535 ret = iomap_dio_bio_iter_one(iter, dio, pos, alignment, bio_opf);
536 if (unlikely(ret < 0)) {
537 /*
538 * We have to stop part way through an IO. We must fall
539 * through to the sub-block tail zeroing here, otherwise
540 * this short IO may expose stale data in the tail of
541 * the block we haven't written data to.
542 */
543 break;
544 }
545 dio->size += ret;
546 copied += ret;
547 pos += ret;
548 ret = 0;
549 } while (iov_iter_count(dio->submit.iter));
550
551 /*
552 * We need to zeroout the tail of a sub-block write if the extent type
553 * requires zeroing or the write extends beyond EOF. If we don't zero
554 * the block tail in the latter case, we can expose stale data via mmap
555 * reads of the EOF block.
556 */
557 if (need_zeroout ||
558 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
559 /* zero out from the end of the write to the end of the block */
560 pad = pos & (fs_block_size - 1);
561 if (pad)
562 ret = iomap_dio_zero(iter, dio, pos,
563 fs_block_size - pad);
564 }
565 out:
566 /* Undo iter limitation to current extent */
567 iov_iter_reexpand(dio->submit.iter, orig_count - copied);
568 if (copied)
569 return iomap_iter_advance(iter, copied);
570 return ret;
571 }
572
iomap_dio_hole_iter(struct iomap_iter * iter,struct iomap_dio * dio)573 static int iomap_dio_hole_iter(struct iomap_iter *iter, struct iomap_dio *dio)
574 {
575 loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter);
576
577 dio->size += length;
578 if (!length)
579 return -EFAULT;
580 return iomap_iter_advance(iter, length);
581 }
582
iomap_dio_inline_iter(struct iomap_iter * iomi,struct iomap_dio * dio)583 static int iomap_dio_inline_iter(struct iomap_iter *iomi, struct iomap_dio *dio)
584 {
585 const struct iomap *iomap = &iomi->iomap;
586 struct iov_iter *iter = dio->submit.iter;
587 void *inline_data = iomap_inline_data(iomap, iomi->pos);
588 loff_t length = iomap_length(iomi);
589 loff_t pos = iomi->pos;
590 u64 copied;
591
592 if (WARN_ON_ONCE(!inline_data))
593 return -EIO;
594
595 if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap)))
596 return -EIO;
597
598 if (dio->flags & IOMAP_DIO_WRITE) {
599 loff_t size = iomi->inode->i_size;
600
601 if (pos > size)
602 memset(iomap_inline_data(iomap, size), 0, pos - size);
603 copied = copy_from_iter(inline_data, length, iter);
604 if (copied) {
605 if (pos + copied > size)
606 i_size_write(iomi->inode, pos + copied);
607 mark_inode_dirty(iomi->inode);
608 }
609 } else {
610 copied = copy_to_iter(inline_data, length, iter);
611 }
612 dio->size += copied;
613 if (!copied)
614 return -EFAULT;
615 return iomap_iter_advance(iomi, copied);
616 }
617
iomap_dio_iter(struct iomap_iter * iter,struct iomap_dio * dio)618 static int iomap_dio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
619 {
620 switch (iter->iomap.type) {
621 case IOMAP_HOLE:
622 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
623 return -EIO;
624 return iomap_dio_hole_iter(iter, dio);
625 case IOMAP_UNWRITTEN:
626 if (!(dio->flags & IOMAP_DIO_WRITE))
627 return iomap_dio_hole_iter(iter, dio);
628 return iomap_dio_bio_iter(iter, dio);
629 case IOMAP_MAPPED:
630 return iomap_dio_bio_iter(iter, dio);
631 case IOMAP_INLINE:
632 return iomap_dio_inline_iter(iter, dio);
633 case IOMAP_DELALLOC:
634 /*
635 * DIO is not serialised against mmap() access at all, and so
636 * if the page_mkwrite occurs between the writeback and the
637 * iomap_iter() call in the DIO path, then it will see the
638 * DELALLOC block that the page-mkwrite allocated.
639 */
640 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
641 dio->iocb->ki_filp, current->comm);
642 return -EIO;
643 default:
644 WARN_ON_ONCE(1);
645 return -EIO;
646 }
647 }
648
649 /*
650 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
651 * is being issued as AIO or not. This allows us to optimise pure data writes
652 * to use REQ_FUA rather than requiring generic_write_sync() to issue a
653 * REQ_FLUSH post write. This is slightly tricky because a single request here
654 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
655 * may be pure data writes. In that case, we still need to do a full data sync
656 * completion.
657 *
658 * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL,
659 * __iomap_dio_rw can return a partial result if it encounters a non-resident
660 * page in @iter after preparing a transfer. In that case, the non-resident
661 * pages can be faulted in and the request resumed with @done_before set to the
662 * number of bytes previously transferred. The request will then complete with
663 * the correct total number of bytes transferred; this is essential for
664 * completing partial requests asynchronously.
665 *
666 * Returns -ENOTBLK In case of a page invalidation invalidation failure for
667 * writes. The callers needs to fall back to buffered I/O in this case.
668 */
669 struct iomap_dio *
__iomap_dio_rw(struct kiocb * iocb,struct iov_iter * iter,const struct iomap_ops * ops,const struct iomap_dio_ops * dops,unsigned int dio_flags,void * private,size_t done_before)670 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
671 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
672 unsigned int dio_flags, void *private, size_t done_before)
673 {
674 struct inode *inode = file_inode(iocb->ki_filp);
675 struct iomap_iter iomi = {
676 .inode = inode,
677 .pos = iocb->ki_pos,
678 .len = iov_iter_count(iter),
679 .flags = IOMAP_DIRECT,
680 .private = private,
681 };
682 bool wait_for_completion =
683 is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
684 struct blk_plug plug;
685 struct iomap_dio *dio;
686 loff_t ret = 0;
687
688 trace_iomap_dio_rw_begin(iocb, iter, dio_flags, done_before);
689
690 if (!iomi.len)
691 return NULL;
692
693 dio = kmalloc_obj(*dio);
694 if (!dio)
695 return ERR_PTR(-ENOMEM);
696
697 dio->iocb = iocb;
698 atomic_set(&dio->ref, 1);
699 dio->size = 0;
700 dio->i_size = i_size_read(inode);
701 dio->dops = dops;
702 dio->error = 0;
703 dio->flags = dio_flags & (IOMAP_DIO_FSBLOCK_ALIGNED | IOMAP_DIO_BOUNCE);
704 dio->done_before = done_before;
705
706 dio->submit.iter = iter;
707 dio->submit.waiter = current;
708
709 if (iocb->ki_flags & IOCB_NOWAIT)
710 iomi.flags |= IOMAP_NOWAIT;
711
712 if (iov_iter_rw(iter) == READ) {
713 if (iomi.pos >= dio->i_size)
714 goto out_free_dio;
715
716 if (user_backed_iter(iter))
717 dio->flags |= IOMAP_DIO_USER_BACKED;
718
719 ret = kiocb_write_and_wait(iocb, iomi.len);
720 if (ret)
721 goto out_free_dio;
722 } else {
723 iomi.flags |= IOMAP_WRITE;
724 dio->flags |= IOMAP_DIO_WRITE;
725
726 if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) {
727 ret = -EAGAIN;
728 if (iomi.pos >= dio->i_size ||
729 iomi.pos + iomi.len > dio->i_size)
730 goto out_free_dio;
731 iomi.flags |= IOMAP_OVERWRITE_ONLY;
732 }
733
734 if (iocb->ki_flags & IOCB_ATOMIC)
735 iomi.flags |= IOMAP_ATOMIC;
736
737 /* for data sync or sync, we need sync completion processing */
738 if (iocb_is_dsync(iocb)) {
739 dio->flags |= IOMAP_DIO_NEED_SYNC;
740
741 /*
742 * For datasync only writes, we optimistically try using
743 * WRITE_THROUGH for this IO. This flag requires either
744 * FUA writes through the device's write cache, or a
745 * normal write to a device without a volatile write
746 * cache. For the former, Any non-FUA write that occurs
747 * will clear this flag, hence we know before completion
748 * whether a cache flush is necessary.
749 */
750 if (!(iocb->ki_flags & IOCB_SYNC))
751 dio->flags |= IOMAP_DIO_WRITE_THROUGH;
752 }
753
754 /*
755 * i_size updates must to happen from process context.
756 */
757 if (iomi.pos + iomi.len > dio->i_size)
758 dio->flags |= IOMAP_DIO_COMP_WORK;
759
760 /*
761 * Try to invalidate cache pages for the range we are writing.
762 * If this invalidation fails, let the caller fall back to
763 * buffered I/O.
764 */
765 ret = kiocb_invalidate_pages(iocb, iomi.len);
766 if (ret) {
767 if (ret != -EAGAIN) {
768 trace_iomap_dio_invalidate_fail(inode, iomi.pos,
769 iomi.len);
770 if (iocb->ki_flags & IOCB_ATOMIC) {
771 /*
772 * folio invalidation failed, maybe
773 * this is transient, unlock and see if
774 * the caller tries again.
775 */
776 ret = -EAGAIN;
777 } else {
778 /* fall back to buffered write */
779 ret = -ENOTBLK;
780 }
781 }
782 goto out_free_dio;
783 }
784 }
785
786 if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
787 ret = sb_init_dio_done_wq(inode->i_sb);
788 if (ret < 0)
789 goto out_free_dio;
790 }
791
792 inode_dio_begin(inode);
793
794 blk_start_plug(&plug);
795 while ((ret = iomap_iter(&iomi, ops)) > 0) {
796 iomi.status = iomap_dio_iter(&iomi, dio);
797
798 /*
799 * We can only poll for single bio I/Os.
800 */
801 iocb->ki_flags &= ~IOCB_HIPRI;
802 }
803
804 blk_finish_plug(&plug);
805
806 /*
807 * We only report that we've read data up to i_size.
808 * Revert iter to a state corresponding to that as some callers (such
809 * as the splice code) rely on it.
810 */
811 if (iov_iter_rw(iter) == READ && iomi.pos >= dio->i_size)
812 iov_iter_revert(iter, iomi.pos - dio->i_size);
813
814 if (ret == -EFAULT && dio->size && (dio_flags & IOMAP_DIO_PARTIAL)) {
815 if (!(iocb->ki_flags & IOCB_NOWAIT))
816 wait_for_completion = true;
817 ret = 0;
818 }
819
820 /* magic error code to fall back to buffered I/O */
821 if (ret == -ENOTBLK) {
822 wait_for_completion = true;
823 ret = 0;
824 }
825 if (ret < 0)
826 iomap_dio_set_error(dio, ret);
827
828 /*
829 * If all the writes we issued were already written through to the
830 * media, we don't need to flush the cache on IO completion. Clear the
831 * sync flag for this case.
832 *
833 * Otherwise clear the inline completion flag if any sync work is
834 * needed, as that needs to be performed from process context.
835 */
836 if (dio->flags & IOMAP_DIO_WRITE_THROUGH)
837 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
838 else if (dio->flags & IOMAP_DIO_NEED_SYNC)
839 dio->flags |= IOMAP_DIO_COMP_WORK;
840
841 /*
842 * We are about to drop our additional submission reference, which
843 * might be the last reference to the dio. There are three different
844 * ways we can progress here:
845 *
846 * (a) If this is the last reference we will always complete and free
847 * the dio ourselves.
848 * (b) If this is not the last reference, and we serve an asynchronous
849 * iocb, we must never touch the dio after the decrement, the
850 * I/O completion handler will complete and free it.
851 * (c) If this is not the last reference, but we serve a synchronous
852 * iocb, the I/O completion handler will wake us up on the drop
853 * of the final reference, and we will complete and free it here
854 * after we got woken by the I/O completion handler.
855 */
856 dio->wait_for_completion = wait_for_completion;
857 if (!atomic_dec_and_test(&dio->ref)) {
858 if (!wait_for_completion) {
859 trace_iomap_dio_rw_queued(inode, iomi.pos, iomi.len);
860 return ERR_PTR(-EIOCBQUEUED);
861 }
862
863 for (;;) {
864 set_current_state(TASK_UNINTERRUPTIBLE);
865 if (!READ_ONCE(dio->submit.waiter))
866 break;
867
868 blk_io_schedule();
869 }
870 __set_current_state(TASK_RUNNING);
871 }
872
873 return dio;
874
875 out_free_dio:
876 kfree(dio);
877 if (ret)
878 return ERR_PTR(ret);
879 return NULL;
880 }
881 EXPORT_SYMBOL_GPL(__iomap_dio_rw);
882
883 ssize_t
iomap_dio_rw(struct kiocb * iocb,struct iov_iter * iter,const struct iomap_ops * ops,const struct iomap_dio_ops * dops,unsigned int dio_flags,void * private,size_t done_before)884 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
885 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
886 unsigned int dio_flags, void *private, size_t done_before)
887 {
888 struct iomap_dio *dio;
889
890 dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, private,
891 done_before);
892 if (IS_ERR_OR_NULL(dio))
893 return PTR_ERR_OR_ZERO(dio);
894 return iomap_dio_complete(dio);
895 }
896 EXPORT_SYMBOL_GPL(iomap_dio_rw);
897