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