xref: /linux/fs/iomap/direct-io.c (revision 0889d07f3e4b171c453b2aaf2b257f9074cdf624)
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 
14 #include "../internal.h"
15 
16 /*
17  * Private flags for iomap_dio, must not overlap with the public ones in
18  * iomap.h:
19  */
20 #define IOMAP_DIO_WRITE_FUA	(1 << 28)
21 #define IOMAP_DIO_NEED_SYNC	(1 << 29)
22 #define IOMAP_DIO_WRITE		(1 << 30)
23 #define IOMAP_DIO_DIRTY		(1 << 31)
24 
25 struct iomap_dio {
26 	struct kiocb		*iocb;
27 	const struct iomap_dio_ops *dops;
28 	loff_t			i_size;
29 	loff_t			size;
30 	atomic_t		ref;
31 	unsigned		flags;
32 	int			error;
33 	bool			wait_for_completion;
34 
35 	union {
36 		/* used during submission and for synchronous completion: */
37 		struct {
38 			struct iov_iter		*iter;
39 			struct task_struct	*waiter;
40 			struct request_queue	*last_queue;
41 			blk_qc_t		cookie;
42 		} submit;
43 
44 		/* used for aio completion: */
45 		struct {
46 			struct work_struct	work;
47 		} aio;
48 	};
49 };
50 
51 int iomap_dio_iopoll(struct kiocb *kiocb, bool spin)
52 {
53 	struct request_queue *q = READ_ONCE(kiocb->private);
54 
55 	if (!q)
56 		return 0;
57 	return blk_poll(q, READ_ONCE(kiocb->ki_cookie), spin);
58 }
59 EXPORT_SYMBOL_GPL(iomap_dio_iopoll);
60 
61 static void iomap_dio_submit_bio(struct iomap_dio *dio, struct iomap *iomap,
62 		struct bio *bio)
63 {
64 	atomic_inc(&dio->ref);
65 
66 	if (dio->iocb->ki_flags & IOCB_HIPRI)
67 		bio_set_polled(bio, dio->iocb);
68 
69 	dio->submit.last_queue = bdev_get_queue(iomap->bdev);
70 	dio->submit.cookie = submit_bio(bio);
71 }
72 
73 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
74 {
75 	const struct iomap_dio_ops *dops = dio->dops;
76 	struct kiocb *iocb = dio->iocb;
77 	struct inode *inode = file_inode(iocb->ki_filp);
78 	loff_t offset = iocb->ki_pos;
79 	ssize_t ret = dio->error;
80 
81 	if (dops && dops->end_io)
82 		ret = dops->end_io(iocb, dio->size, ret, dio->flags);
83 
84 	if (likely(!ret)) {
85 		ret = dio->size;
86 		/* check for short read */
87 		if (offset + ret > dio->i_size &&
88 		    !(dio->flags & IOMAP_DIO_WRITE))
89 			ret = dio->i_size - offset;
90 		iocb->ki_pos += ret;
91 	}
92 
93 	/*
94 	 * Try again to invalidate clean pages which might have been cached by
95 	 * non-direct readahead, or faulted in by get_user_pages() if the source
96 	 * of the write was an mmap'ed region of the file we're writing.  Either
97 	 * one is a pretty crazy thing to do, so we don't support it 100%.  If
98 	 * this invalidation fails, tough, the write still worked...
99 	 *
100 	 * And this page cache invalidation has to be after ->end_io(), as some
101 	 * filesystems convert unwritten extents to real allocations in
102 	 * ->end_io() when necessary, otherwise a racing buffer read would cache
103 	 * zeros from unwritten extents.
104 	 */
105 	if (!dio->error &&
106 	    (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
107 		int err;
108 		err = invalidate_inode_pages2_range(inode->i_mapping,
109 				offset >> PAGE_SHIFT,
110 				(offset + dio->size - 1) >> PAGE_SHIFT);
111 		if (err)
112 			dio_warn_stale_pagecache(iocb->ki_filp);
113 	}
114 
115 	/*
116 	 * If this is a DSYNC write, make sure we push it to stable storage now
117 	 * that we've written data.
118 	 */
119 	if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
120 		ret = generic_write_sync(iocb, ret);
121 
122 	inode_dio_end(file_inode(iocb->ki_filp));
123 	kfree(dio);
124 
125 	return ret;
126 }
127 
128 static void iomap_dio_complete_work(struct work_struct *work)
129 {
130 	struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
131 	struct kiocb *iocb = dio->iocb;
132 
133 	iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
134 }
135 
136 /*
137  * Set an error in the dio if none is set yet.  We have to use cmpxchg
138  * as the submission context and the completion context(s) can race to
139  * update the error.
140  */
141 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
142 {
143 	cmpxchg(&dio->error, 0, ret);
144 }
145 
146 static void iomap_dio_bio_end_io(struct bio *bio)
147 {
148 	struct iomap_dio *dio = bio->bi_private;
149 	bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
150 
151 	if (bio->bi_status)
152 		iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
153 
154 	if (atomic_dec_and_test(&dio->ref)) {
155 		if (dio->wait_for_completion) {
156 			struct task_struct *waiter = dio->submit.waiter;
157 			WRITE_ONCE(dio->submit.waiter, NULL);
158 			blk_wake_io_task(waiter);
159 		} else if (dio->flags & IOMAP_DIO_WRITE) {
160 			struct inode *inode = file_inode(dio->iocb->ki_filp);
161 
162 			INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
163 			queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
164 		} else {
165 			iomap_dio_complete_work(&dio->aio.work);
166 		}
167 	}
168 
169 	if (should_dirty) {
170 		bio_check_pages_dirty(bio);
171 	} else {
172 		bio_release_pages(bio, false);
173 		bio_put(bio);
174 	}
175 }
176 
177 static void
178 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
179 		unsigned len)
180 {
181 	struct page *page = ZERO_PAGE(0);
182 	int flags = REQ_SYNC | REQ_IDLE;
183 	struct bio *bio;
184 
185 	bio = bio_alloc(GFP_KERNEL, 1);
186 	bio_set_dev(bio, iomap->bdev);
187 	bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
188 	bio->bi_private = dio;
189 	bio->bi_end_io = iomap_dio_bio_end_io;
190 
191 	get_page(page);
192 	__bio_add_page(bio, page, len, 0);
193 	bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
194 	iomap_dio_submit_bio(dio, iomap, bio);
195 }
196 
197 static loff_t
198 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
199 		struct iomap_dio *dio, struct iomap *iomap)
200 {
201 	unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
202 	unsigned int fs_block_size = i_blocksize(inode), pad;
203 	unsigned int align = iov_iter_alignment(dio->submit.iter);
204 	struct iov_iter iter;
205 	struct bio *bio;
206 	bool need_zeroout = false;
207 	bool use_fua = false;
208 	int nr_pages, ret = 0;
209 	size_t copied = 0;
210 
211 	if ((pos | length | align) & ((1 << blkbits) - 1))
212 		return -EINVAL;
213 
214 	if (iomap->type == IOMAP_UNWRITTEN) {
215 		dio->flags |= IOMAP_DIO_UNWRITTEN;
216 		need_zeroout = true;
217 	}
218 
219 	if (iomap->flags & IOMAP_F_SHARED)
220 		dio->flags |= IOMAP_DIO_COW;
221 
222 	if (iomap->flags & IOMAP_F_NEW) {
223 		need_zeroout = true;
224 	} else if (iomap->type == IOMAP_MAPPED) {
225 		/*
226 		 * Use a FUA write if we need datasync semantics, this is a pure
227 		 * data IO that doesn't require any metadata updates (including
228 		 * after IO completion such as unwritten extent conversion) and
229 		 * the underlying device supports FUA. This allows us to avoid
230 		 * cache flushes on IO completion.
231 		 */
232 		if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
233 		    (dio->flags & IOMAP_DIO_WRITE_FUA) &&
234 		    blk_queue_fua(bdev_get_queue(iomap->bdev)))
235 			use_fua = true;
236 	}
237 
238 	/*
239 	 * Operate on a partial iter trimmed to the extent we were called for.
240 	 * We'll update the iter in the dio once we're done with this extent.
241 	 */
242 	iter = *dio->submit.iter;
243 	iov_iter_truncate(&iter, length);
244 
245 	nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
246 	if (nr_pages <= 0)
247 		return nr_pages;
248 
249 	if (need_zeroout) {
250 		/* zero out from the start of the block to the write offset */
251 		pad = pos & (fs_block_size - 1);
252 		if (pad)
253 			iomap_dio_zero(dio, iomap, pos - pad, pad);
254 	}
255 
256 	do {
257 		size_t n;
258 		if (dio->error) {
259 			iov_iter_revert(dio->submit.iter, copied);
260 			return 0;
261 		}
262 
263 		bio = bio_alloc(GFP_KERNEL, nr_pages);
264 		bio_set_dev(bio, iomap->bdev);
265 		bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
266 		bio->bi_write_hint = dio->iocb->ki_hint;
267 		bio->bi_ioprio = dio->iocb->ki_ioprio;
268 		bio->bi_private = dio;
269 		bio->bi_end_io = iomap_dio_bio_end_io;
270 
271 		ret = bio_iov_iter_get_pages(bio, &iter);
272 		if (unlikely(ret)) {
273 			/*
274 			 * We have to stop part way through an IO. We must fall
275 			 * through to the sub-block tail zeroing here, otherwise
276 			 * this short IO may expose stale data in the tail of
277 			 * the block we haven't written data to.
278 			 */
279 			bio_put(bio);
280 			goto zero_tail;
281 		}
282 
283 		n = bio->bi_iter.bi_size;
284 		if (dio->flags & IOMAP_DIO_WRITE) {
285 			bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
286 			if (use_fua)
287 				bio->bi_opf |= REQ_FUA;
288 			else
289 				dio->flags &= ~IOMAP_DIO_WRITE_FUA;
290 			task_io_account_write(n);
291 		} else {
292 			bio->bi_opf = REQ_OP_READ;
293 			if (dio->flags & IOMAP_DIO_DIRTY)
294 				bio_set_pages_dirty(bio);
295 		}
296 
297 		iov_iter_advance(dio->submit.iter, n);
298 
299 		dio->size += n;
300 		pos += n;
301 		copied += n;
302 
303 		nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
304 		iomap_dio_submit_bio(dio, iomap, bio);
305 	} while (nr_pages);
306 
307 	/*
308 	 * We need to zeroout the tail of a sub-block write if the extent type
309 	 * requires zeroing or the write extends beyond EOF. If we don't zero
310 	 * the block tail in the latter case, we can expose stale data via mmap
311 	 * reads of the EOF block.
312 	 */
313 zero_tail:
314 	if (need_zeroout ||
315 	    ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
316 		/* zero out from the end of the write to the end of the block */
317 		pad = pos & (fs_block_size - 1);
318 		if (pad)
319 			iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
320 	}
321 	return copied ? copied : ret;
322 }
323 
324 static loff_t
325 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
326 {
327 	length = iov_iter_zero(length, dio->submit.iter);
328 	dio->size += length;
329 	return length;
330 }
331 
332 static loff_t
333 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
334 		struct iomap_dio *dio, struct iomap *iomap)
335 {
336 	struct iov_iter *iter = dio->submit.iter;
337 	size_t copied;
338 
339 	BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
340 
341 	if (dio->flags & IOMAP_DIO_WRITE) {
342 		loff_t size = inode->i_size;
343 
344 		if (pos > size)
345 			memset(iomap->inline_data + size, 0, pos - size);
346 		copied = copy_from_iter(iomap->inline_data + pos, length, iter);
347 		if (copied) {
348 			if (pos + copied > size)
349 				i_size_write(inode, pos + copied);
350 			mark_inode_dirty(inode);
351 		}
352 	} else {
353 		copied = copy_to_iter(iomap->inline_data + pos, length, iter);
354 	}
355 	dio->size += copied;
356 	return copied;
357 }
358 
359 static loff_t
360 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
361 		void *data, struct iomap *iomap)
362 {
363 	struct iomap_dio *dio = data;
364 
365 	switch (iomap->type) {
366 	case IOMAP_HOLE:
367 		if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
368 			return -EIO;
369 		return iomap_dio_hole_actor(length, dio);
370 	case IOMAP_UNWRITTEN:
371 		if (!(dio->flags & IOMAP_DIO_WRITE))
372 			return iomap_dio_hole_actor(length, dio);
373 		return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
374 	case IOMAP_MAPPED:
375 		return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
376 	case IOMAP_INLINE:
377 		return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
378 	default:
379 		WARN_ON_ONCE(1);
380 		return -EIO;
381 	}
382 }
383 
384 /*
385  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
386  * is being issued as AIO or not.  This allows us to optimise pure data writes
387  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
388  * REQ_FLUSH post write. This is slightly tricky because a single request here
389  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
390  * may be pure data writes. In that case, we still need to do a full data sync
391  * completion.
392  */
393 ssize_t
394 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
395 		const struct iomap_ops *ops, const struct iomap_dio_ops *dops)
396 {
397 	struct address_space *mapping = iocb->ki_filp->f_mapping;
398 	struct inode *inode = file_inode(iocb->ki_filp);
399 	size_t count = iov_iter_count(iter);
400 	loff_t pos = iocb->ki_pos, start = pos;
401 	loff_t end = iocb->ki_pos + count - 1, ret = 0;
402 	unsigned int flags = IOMAP_DIRECT;
403 	bool wait_for_completion = is_sync_kiocb(iocb);
404 	struct blk_plug plug;
405 	struct iomap_dio *dio;
406 
407 	lockdep_assert_held(&inode->i_rwsem);
408 
409 	if (!count)
410 		return 0;
411 
412 	dio = kmalloc(sizeof(*dio), GFP_KERNEL);
413 	if (!dio)
414 		return -ENOMEM;
415 
416 	dio->iocb = iocb;
417 	atomic_set(&dio->ref, 1);
418 	dio->size = 0;
419 	dio->i_size = i_size_read(inode);
420 	dio->dops = dops;
421 	dio->error = 0;
422 	dio->flags = 0;
423 
424 	dio->submit.iter = iter;
425 	dio->submit.waiter = current;
426 	dio->submit.cookie = BLK_QC_T_NONE;
427 	dio->submit.last_queue = NULL;
428 
429 	if (iov_iter_rw(iter) == READ) {
430 		if (pos >= dio->i_size)
431 			goto out_free_dio;
432 
433 		if (iter_is_iovec(iter) && iov_iter_rw(iter) == READ)
434 			dio->flags |= IOMAP_DIO_DIRTY;
435 	} else {
436 		flags |= IOMAP_WRITE;
437 		dio->flags |= IOMAP_DIO_WRITE;
438 
439 		/* for data sync or sync, we need sync completion processing */
440 		if (iocb->ki_flags & IOCB_DSYNC)
441 			dio->flags |= IOMAP_DIO_NEED_SYNC;
442 
443 		/*
444 		 * For datasync only writes, we optimistically try using FUA for
445 		 * this IO.  Any non-FUA write that occurs will clear this flag,
446 		 * hence we know before completion whether a cache flush is
447 		 * necessary.
448 		 */
449 		if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
450 			dio->flags |= IOMAP_DIO_WRITE_FUA;
451 	}
452 
453 	if (iocb->ki_flags & IOCB_NOWAIT) {
454 		if (filemap_range_has_page(mapping, start, end)) {
455 			ret = -EAGAIN;
456 			goto out_free_dio;
457 		}
458 		flags |= IOMAP_NOWAIT;
459 	}
460 
461 	ret = filemap_write_and_wait_range(mapping, start, end);
462 	if (ret)
463 		goto out_free_dio;
464 
465 	/*
466 	 * Try to invalidate cache pages for the range we're direct
467 	 * writing.  If this invalidation fails, tough, the write will
468 	 * still work, but racing two incompatible write paths is a
469 	 * pretty crazy thing to do, so we don't support it 100%.
470 	 */
471 	ret = invalidate_inode_pages2_range(mapping,
472 			start >> PAGE_SHIFT, end >> PAGE_SHIFT);
473 	if (ret)
474 		dio_warn_stale_pagecache(iocb->ki_filp);
475 	ret = 0;
476 
477 	if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
478 	    !inode->i_sb->s_dio_done_wq) {
479 		ret = sb_init_dio_done_wq(inode->i_sb);
480 		if (ret < 0)
481 			goto out_free_dio;
482 	}
483 
484 	inode_dio_begin(inode);
485 
486 	blk_start_plug(&plug);
487 	do {
488 		ret = iomap_apply(inode, pos, count, flags, ops, dio,
489 				iomap_dio_actor);
490 		if (ret <= 0) {
491 			/* magic error code to fall back to buffered I/O */
492 			if (ret == -ENOTBLK) {
493 				wait_for_completion = true;
494 				ret = 0;
495 			}
496 			break;
497 		}
498 		pos += ret;
499 
500 		if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
501 			break;
502 	} while ((count = iov_iter_count(iter)) > 0);
503 	blk_finish_plug(&plug);
504 
505 	if (ret < 0)
506 		iomap_dio_set_error(dio, ret);
507 
508 	/*
509 	 * If all the writes we issued were FUA, we don't need to flush the
510 	 * cache on IO completion. Clear the sync flag for this case.
511 	 */
512 	if (dio->flags & IOMAP_DIO_WRITE_FUA)
513 		dio->flags &= ~IOMAP_DIO_NEED_SYNC;
514 
515 	WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
516 	WRITE_ONCE(iocb->private, dio->submit.last_queue);
517 
518 	/*
519 	 * We are about to drop our additional submission reference, which
520 	 * might be the last reference to the dio.  There are three three
521 	 * different ways we can progress here:
522 	 *
523 	 *  (a) If this is the last reference we will always complete and free
524 	 *	the dio ourselves.
525 	 *  (b) If this is not the last reference, and we serve an asynchronous
526 	 *	iocb, we must never touch the dio after the decrement, the
527 	 *	I/O completion handler will complete and free it.
528 	 *  (c) If this is not the last reference, but we serve a synchronous
529 	 *	iocb, the I/O completion handler will wake us up on the drop
530 	 *	of the final reference, and we will complete and free it here
531 	 *	after we got woken by the I/O completion handler.
532 	 */
533 	dio->wait_for_completion = wait_for_completion;
534 	if (!atomic_dec_and_test(&dio->ref)) {
535 		if (!wait_for_completion)
536 			return -EIOCBQUEUED;
537 
538 		for (;;) {
539 			set_current_state(TASK_UNINTERRUPTIBLE);
540 			if (!READ_ONCE(dio->submit.waiter))
541 				break;
542 
543 			if (!(iocb->ki_flags & IOCB_HIPRI) ||
544 			    !dio->submit.last_queue ||
545 			    !blk_poll(dio->submit.last_queue,
546 					 dio->submit.cookie, true))
547 				io_schedule();
548 		}
549 		__set_current_state(TASK_RUNNING);
550 	}
551 
552 	return iomap_dio_complete(dio);
553 
554 out_free_dio:
555 	kfree(dio);
556 	return ret;
557 }
558 EXPORT_SYMBOL_GPL(iomap_dio_rw);
559