1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2010 Red Hat, Inc.
4 * Copyright (c) 2016-2018 Christoph Hellwig.
5 */
6 #include <linux/module.h>
7 #include <linux/compiler.h>
8 #include <linux/fs.h>
9 #include <linux/iomap.h>
10 #include <linux/backing-dev.h>
11 #include <linux/uio.h>
12 #include <linux/task_io_accounting_ops.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_WRITE_FUA (1 << 28)
22 #define IOMAP_DIO_NEED_SYNC (1 << 29)
23 #define IOMAP_DIO_WRITE (1 << 30)
24 #define IOMAP_DIO_DIRTY (1 << 31)
25
26 struct 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 bool wait_for_completion;
35
36 union {
37 /* used during submission and for synchronous completion: */
38 struct {
39 struct iov_iter *iter;
40 struct task_struct *waiter;
41 struct request_queue *last_queue;
42 blk_qc_t cookie;
43 } submit;
44
45 /* used for aio completion: */
46 struct {
47 struct work_struct work;
48 } aio;
49 };
50 };
51
iomap_dio_iopoll(struct kiocb * kiocb,bool spin)52 int iomap_dio_iopoll(struct kiocb *kiocb, bool spin)
53 {
54 struct request_queue *q = READ_ONCE(kiocb->private);
55
56 if (!q)
57 return 0;
58 return blk_poll(q, READ_ONCE(kiocb->ki_cookie), spin);
59 }
60 EXPORT_SYMBOL_GPL(iomap_dio_iopoll);
61
iomap_dio_submit_bio(struct iomap_dio * dio,struct iomap * iomap,struct bio * bio,loff_t pos)62 static void iomap_dio_submit_bio(struct iomap_dio *dio, struct iomap *iomap,
63 struct bio *bio, loff_t pos)
64 {
65 atomic_inc(&dio->ref);
66
67 if (dio->iocb->ki_flags & IOCB_HIPRI)
68 bio_set_polled(bio, dio->iocb);
69
70 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
71 if (dio->dops && dio->dops->submit_io)
72 dio->submit.cookie = dio->dops->submit_io(
73 file_inode(dio->iocb->ki_filp),
74 iomap, bio, pos);
75 else
76 dio->submit.cookie = submit_bio(bio);
77 }
78
iomap_dio_complete(struct iomap_dio * dio)79 ssize_t iomap_dio_complete(struct iomap_dio *dio)
80 {
81 const struct iomap_dio_ops *dops = dio->dops;
82 struct kiocb *iocb = dio->iocb;
83 struct inode *inode = file_inode(iocb->ki_filp);
84 loff_t offset = iocb->ki_pos;
85 ssize_t ret = dio->error;
86
87 if (dops && dops->end_io)
88 ret = dops->end_io(iocb, dio->size, ret, dio->flags);
89
90 if (likely(!ret)) {
91 ret = dio->size;
92 /* check for short read */
93 if (offset + ret > dio->i_size &&
94 !(dio->flags & IOMAP_DIO_WRITE))
95 ret = dio->i_size - offset;
96 iocb->ki_pos += ret;
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 &&
112 (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
113 int err;
114 err = invalidate_inode_pages2_range(inode->i_mapping,
115 offset >> PAGE_SHIFT,
116 (offset + dio->size - 1) >> PAGE_SHIFT);
117 if (err)
118 dio_warn_stale_pagecache(iocb->ki_filp);
119 }
120
121 inode_dio_end(file_inode(iocb->ki_filp));
122 /*
123 * If this is a DSYNC write, make sure we push it to stable storage now
124 * that we've written data.
125 */
126 if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
127 ret = generic_write_sync(iocb, ret);
128
129 kfree(dio);
130
131 return ret;
132 }
133 EXPORT_SYMBOL_GPL(iomap_dio_complete);
134
iomap_dio_complete_work(struct work_struct * work)135 static 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), 0);
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 */
iomap_dio_set_error(struct iomap_dio * dio,int ret)148 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
149 {
150 cmpxchg(&dio->error, 0, ret);
151 }
152
iomap_dio_bio_end_io(struct bio * bio)153 static void iomap_dio_bio_end_io(struct bio *bio)
154 {
155 struct iomap_dio *dio = bio->bi_private;
156 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
157
158 if (bio->bi_status)
159 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
160
161 if (atomic_dec_and_test(&dio->ref)) {
162 if (dio->wait_for_completion) {
163 struct task_struct *waiter = dio->submit.waiter;
164 WRITE_ONCE(dio->submit.waiter, NULL);
165 blk_wake_io_task(waiter);
166 } else if (dio->flags & IOMAP_DIO_WRITE) {
167 struct inode *inode = file_inode(dio->iocb->ki_filp);
168
169 INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
170 queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
171 } else {
172 iomap_dio_complete_work(&dio->aio.work);
173 }
174 }
175
176 if (should_dirty) {
177 bio_check_pages_dirty(bio);
178 } else {
179 bio_release_pages(bio, false);
180 bio_put(bio);
181 }
182 }
183
184 static void
iomap_dio_zero(struct iomap_dio * dio,struct iomap * iomap,loff_t pos,unsigned len)185 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
186 unsigned len)
187 {
188 struct page *page = ZERO_PAGE(0);
189 int flags = REQ_SYNC | REQ_IDLE;
190 struct bio *bio;
191
192 bio = bio_alloc(GFP_KERNEL, 1);
193 bio_set_dev(bio, iomap->bdev);
194 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
195 bio->bi_private = dio;
196 bio->bi_end_io = iomap_dio_bio_end_io;
197
198 get_page(page);
199 __bio_add_page(bio, page, len, 0);
200 bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
201 iomap_dio_submit_bio(dio, iomap, bio, pos);
202 }
203
204 static loff_t
iomap_dio_bio_actor(struct inode * inode,loff_t pos,loff_t length,struct iomap_dio * dio,struct iomap * iomap)205 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
206 struct iomap_dio *dio, struct iomap *iomap)
207 {
208 unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
209 unsigned int fs_block_size = i_blocksize(inode), pad;
210 unsigned int align = iov_iter_alignment(dio->submit.iter);
211 struct bio *bio;
212 bool need_zeroout = false;
213 bool use_fua = false;
214 int nr_pages, ret = 0;
215 size_t copied = 0;
216 size_t orig_count;
217
218 if ((pos | length | align) & ((1 << blkbits) - 1))
219 return -EINVAL;
220
221 if (iomap->type == IOMAP_UNWRITTEN) {
222 dio->flags |= IOMAP_DIO_UNWRITTEN;
223 need_zeroout = true;
224 }
225
226 if (iomap->flags & IOMAP_F_SHARED)
227 dio->flags |= IOMAP_DIO_COW;
228
229 if (iomap->flags & IOMAP_F_NEW) {
230 need_zeroout = true;
231 } else if (iomap->type == IOMAP_MAPPED) {
232 /*
233 * Use a FUA write if we need datasync semantics, this is a pure
234 * data IO that doesn't require any metadata updates (including
235 * after IO completion such as unwritten extent conversion) and
236 * the underlying device supports FUA. This allows us to avoid
237 * cache flushes on IO completion.
238 */
239 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
240 (dio->flags & IOMAP_DIO_WRITE_FUA) &&
241 blk_queue_fua(bdev_get_queue(iomap->bdev)))
242 use_fua = true;
243 }
244
245 /*
246 * Save the original count and trim the iter to just the extent we
247 * are operating on right now. The iter will be re-expanded once
248 * we are done.
249 */
250 orig_count = iov_iter_count(dio->submit.iter);
251 iov_iter_truncate(dio->submit.iter, length);
252
253 nr_pages = iov_iter_npages(dio->submit.iter, BIO_MAX_PAGES);
254 if (nr_pages <= 0) {
255 ret = nr_pages;
256 goto out;
257 }
258
259 if (need_zeroout) {
260 /* zero out from the start of the block to the write offset */
261 pad = pos & (fs_block_size - 1);
262 if (pad)
263 iomap_dio_zero(dio, iomap, pos - pad, pad);
264 }
265
266 do {
267 size_t n;
268 if (dio->error) {
269 iov_iter_revert(dio->submit.iter, copied);
270 copied = ret = 0;
271 goto out;
272 }
273
274 bio = bio_alloc(GFP_KERNEL, nr_pages);
275 bio_set_dev(bio, iomap->bdev);
276 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
277 bio->bi_write_hint = dio->iocb->ki_hint;
278 bio->bi_ioprio = dio->iocb->ki_ioprio;
279 bio->bi_private = dio;
280 bio->bi_end_io = iomap_dio_bio_end_io;
281
282 ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
283 if (unlikely(ret)) {
284 /*
285 * We have to stop part way through an IO. We must fall
286 * through to the sub-block tail zeroing here, otherwise
287 * this short IO may expose stale data in the tail of
288 * the block we haven't written data to.
289 */
290 bio_put(bio);
291 goto zero_tail;
292 }
293
294 n = bio->bi_iter.bi_size;
295 if (dio->flags & IOMAP_DIO_WRITE) {
296 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
297 if (use_fua)
298 bio->bi_opf |= REQ_FUA;
299 else
300 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
301 task_io_account_write(n);
302 } else {
303 bio->bi_opf = REQ_OP_READ;
304 if (dio->flags & IOMAP_DIO_DIRTY)
305 bio_set_pages_dirty(bio);
306 }
307
308 dio->size += n;
309 copied += n;
310
311 nr_pages = iov_iter_npages(dio->submit.iter, BIO_MAX_PAGES);
312 iomap_dio_submit_bio(dio, iomap, bio, pos);
313 pos += n;
314 } while (nr_pages);
315
316 /*
317 * We need to zeroout the tail of a sub-block write if the extent type
318 * requires zeroing or the write extends beyond EOF. If we don't zero
319 * the block tail in the latter case, we can expose stale data via mmap
320 * reads of the EOF block.
321 */
322 zero_tail:
323 if (need_zeroout ||
324 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
325 /* zero out from the end of the write to the end of the block */
326 pad = pos & (fs_block_size - 1);
327 if (pad)
328 iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
329 }
330 out:
331 /* Undo iter limitation to current extent */
332 iov_iter_reexpand(dio->submit.iter, orig_count - copied);
333 if (copied)
334 return copied;
335 return ret;
336 }
337
338 static loff_t
iomap_dio_hole_actor(loff_t length,struct iomap_dio * dio)339 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
340 {
341 length = iov_iter_zero(length, dio->submit.iter);
342 dio->size += length;
343 return length;
344 }
345
346 static loff_t
iomap_dio_inline_actor(struct inode * inode,loff_t pos,loff_t length,struct iomap_dio * dio,struct iomap * iomap)347 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
348 struct iomap_dio *dio, struct iomap *iomap)
349 {
350 struct iov_iter *iter = dio->submit.iter;
351 size_t copied;
352
353 BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
354
355 if (dio->flags & IOMAP_DIO_WRITE) {
356 loff_t size = inode->i_size;
357
358 if (pos > size)
359 memset(iomap->inline_data + size, 0, pos - size);
360 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
361 if (copied) {
362 if (pos + copied > size)
363 i_size_write(inode, pos + copied);
364 mark_inode_dirty(inode);
365 }
366 } else {
367 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
368 }
369 dio->size += copied;
370 return copied;
371 }
372
373 static loff_t
iomap_dio_actor(struct inode * inode,loff_t pos,loff_t length,void * data,struct iomap * iomap,struct iomap * srcmap)374 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
375 void *data, struct iomap *iomap, struct iomap *srcmap)
376 {
377 struct iomap_dio *dio = data;
378
379 switch (iomap->type) {
380 case IOMAP_HOLE:
381 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
382 return -EIO;
383 return iomap_dio_hole_actor(length, dio);
384 case IOMAP_UNWRITTEN:
385 if (!(dio->flags & IOMAP_DIO_WRITE))
386 return iomap_dio_hole_actor(length, dio);
387 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
388 case IOMAP_MAPPED:
389 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
390 case IOMAP_INLINE:
391 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
392 case IOMAP_DELALLOC:
393 /*
394 * DIO is not serialised against mmap() access at all, and so
395 * if the page_mkwrite occurs between the writeback and the
396 * iomap_apply() call in the DIO path, then it will see the
397 * DELALLOC block that the page-mkwrite allocated.
398 */
399 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
400 dio->iocb->ki_filp, current->comm);
401 return -EIO;
402 default:
403 WARN_ON_ONCE(1);
404 return -EIO;
405 }
406 }
407
408 /*
409 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
410 * is being issued as AIO or not. This allows us to optimise pure data writes
411 * to use REQ_FUA rather than requiring generic_write_sync() to issue a
412 * REQ_FLUSH post write. This is slightly tricky because a single request here
413 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
414 * may be pure data writes. In that case, we still need to do a full data sync
415 * completion.
416 *
417 * Returns -ENOTBLK In case of a page invalidation invalidation failure for
418 * writes. The callers needs to fall back to buffered I/O in this case.
419 */
420 struct iomap_dio *
__iomap_dio_rw(struct kiocb * iocb,struct iov_iter * iter,const struct iomap_ops * ops,const struct iomap_dio_ops * dops,bool wait_for_completion)421 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
422 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
423 bool wait_for_completion)
424 {
425 struct address_space *mapping = iocb->ki_filp->f_mapping;
426 struct inode *inode = file_inode(iocb->ki_filp);
427 size_t count = iov_iter_count(iter);
428 loff_t pos = iocb->ki_pos;
429 loff_t end = iocb->ki_pos + count - 1, ret = 0;
430 unsigned int flags = IOMAP_DIRECT;
431 struct blk_plug plug;
432 struct iomap_dio *dio;
433
434 if (!count)
435 return NULL;
436
437 if (WARN_ON(is_sync_kiocb(iocb) && !wait_for_completion))
438 return ERR_PTR(-EIO);
439
440 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
441 if (!dio)
442 return ERR_PTR(-ENOMEM);
443
444 dio->iocb = iocb;
445 atomic_set(&dio->ref, 1);
446 dio->size = 0;
447 dio->i_size = i_size_read(inode);
448 dio->dops = dops;
449 dio->error = 0;
450 dio->flags = 0;
451
452 dio->submit.iter = iter;
453 dio->submit.waiter = current;
454 dio->submit.cookie = BLK_QC_T_NONE;
455 dio->submit.last_queue = NULL;
456
457 if (iov_iter_rw(iter) == READ) {
458 if (pos >= dio->i_size)
459 goto out_free_dio;
460
461 if (iter_is_iovec(iter))
462 dio->flags |= IOMAP_DIO_DIRTY;
463 } else {
464 flags |= IOMAP_WRITE;
465 dio->flags |= IOMAP_DIO_WRITE;
466
467 /* for data sync or sync, we need sync completion processing */
468 if (iocb->ki_flags & IOCB_DSYNC)
469 dio->flags |= IOMAP_DIO_NEED_SYNC;
470
471 /*
472 * For datasync only writes, we optimistically try using FUA for
473 * this IO. Any non-FUA write that occurs will clear this flag,
474 * hence we know before completion whether a cache flush is
475 * necessary.
476 */
477 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
478 dio->flags |= IOMAP_DIO_WRITE_FUA;
479 }
480
481 if (iocb->ki_flags & IOCB_NOWAIT) {
482 if (filemap_range_has_page(mapping, pos, end)) {
483 ret = -EAGAIN;
484 goto out_free_dio;
485 }
486 flags |= IOMAP_NOWAIT;
487 }
488
489 ret = filemap_write_and_wait_range(mapping, pos, end);
490 if (ret)
491 goto out_free_dio;
492
493 if (iov_iter_rw(iter) == WRITE) {
494 /*
495 * Try to invalidate cache pages for the range we are writing.
496 * If this invalidation fails, let the caller fall back to
497 * buffered I/O.
498 */
499 if (invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT,
500 end >> PAGE_SHIFT)) {
501 trace_iomap_dio_invalidate_fail(inode, pos, count);
502 ret = -ENOTBLK;
503 goto out_free_dio;
504 }
505
506 if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
507 ret = sb_init_dio_done_wq(inode->i_sb);
508 if (ret < 0)
509 goto out_free_dio;
510 }
511 }
512
513 inode_dio_begin(inode);
514
515 blk_start_plug(&plug);
516 do {
517 ret = iomap_apply(inode, pos, count, flags, ops, dio,
518 iomap_dio_actor);
519 if (ret <= 0) {
520 /* magic error code to fall back to buffered I/O */
521 if (ret == -ENOTBLK) {
522 wait_for_completion = true;
523 ret = 0;
524 }
525 break;
526 }
527 pos += ret;
528
529 if (iov_iter_rw(iter) == READ && pos >= dio->i_size) {
530 /*
531 * We only report that we've read data up to i_size.
532 * Revert iter to a state corresponding to that as
533 * some callers (such as splice code) rely on it.
534 */
535 iov_iter_revert(iter, pos - dio->i_size);
536 break;
537 }
538 } while ((count = iov_iter_count(iter)) > 0);
539 blk_finish_plug(&plug);
540
541 if (ret < 0)
542 iomap_dio_set_error(dio, ret);
543
544 /*
545 * If all the writes we issued were FUA, we don't need to flush the
546 * cache on IO completion. Clear the sync flag for this case.
547 */
548 if (dio->flags & IOMAP_DIO_WRITE_FUA)
549 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
550
551 WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
552 WRITE_ONCE(iocb->private, dio->submit.last_queue);
553
554 /*
555 * We are about to drop our additional submission reference, which
556 * might be the last reference to the dio. There are three different
557 * ways we can progress here:
558 *
559 * (a) If this is the last reference we will always complete and free
560 * the dio ourselves.
561 * (b) If this is not the last reference, and we serve an asynchronous
562 * iocb, we must never touch the dio after the decrement, the
563 * I/O completion handler will complete and free it.
564 * (c) If this is not the last reference, but we serve a synchronous
565 * iocb, the I/O completion handler will wake us up on the drop
566 * of the final reference, and we will complete and free it here
567 * after we got woken by the I/O completion handler.
568 */
569 dio->wait_for_completion = wait_for_completion;
570 if (!atomic_dec_and_test(&dio->ref)) {
571 if (!wait_for_completion)
572 return ERR_PTR(-EIOCBQUEUED);
573
574 for (;;) {
575 set_current_state(TASK_UNINTERRUPTIBLE);
576 if (!READ_ONCE(dio->submit.waiter))
577 break;
578
579 if (!(iocb->ki_flags & IOCB_HIPRI) ||
580 !dio->submit.last_queue ||
581 !blk_poll(dio->submit.last_queue,
582 dio->submit.cookie, true))
583 blk_io_schedule();
584 }
585 __set_current_state(TASK_RUNNING);
586 }
587
588 return dio;
589
590 out_free_dio:
591 kfree(dio);
592 if (ret)
593 return ERR_PTR(ret);
594 return NULL;
595 }
596 EXPORT_SYMBOL_GPL(__iomap_dio_rw);
597
598 ssize_t
iomap_dio_rw(struct kiocb * iocb,struct iov_iter * iter,const struct iomap_ops * ops,const struct iomap_dio_ops * dops,bool wait_for_completion)599 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
600 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
601 bool wait_for_completion)
602 {
603 struct iomap_dio *dio;
604
605 dio = __iomap_dio_rw(iocb, iter, ops, dops, wait_for_completion);
606 if (IS_ERR_OR_NULL(dio))
607 return PTR_ERR_OR_ZERO(dio);
608 return iomap_dio_complete(dio);
609 }
610 EXPORT_SYMBOL_GPL(iomap_dio_rw);
611