xref: /openbmc/linux/fs/iomap/direct-io.c (revision d6e0cbb1)
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 	iomap_dio_end_io_t	*end_io;
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 	struct kiocb *iocb = dio->iocb;
76 	struct inode *inode = file_inode(iocb->ki_filp);
77 	loff_t offset = iocb->ki_pos;
78 	ssize_t ret;
79 
80 	if (dio->end_io) {
81 		ret = dio->end_io(iocb,
82 				dio->error ? dio->error : dio->size,
83 				dio->flags);
84 	} else {
85 		ret = dio->error;
86 	}
87 
88 	if (likely(!ret)) {
89 		ret = dio->size;
90 		/* check for short read */
91 		if (offset + ret > dio->i_size &&
92 		    !(dio->flags & IOMAP_DIO_WRITE))
93 			ret = dio->i_size - offset;
94 		iocb->ki_pos += ret;
95 	}
96 
97 	/*
98 	 * Try again to invalidate clean pages which might have been cached by
99 	 * non-direct readahead, or faulted in by get_user_pages() if the source
100 	 * of the write was an mmap'ed region of the file we're writing.  Either
101 	 * one is a pretty crazy thing to do, so we don't support it 100%.  If
102 	 * this invalidation fails, tough, the write still worked...
103 	 *
104 	 * And this page cache invalidation has to be after dio->end_io(), as
105 	 * some filesystems convert unwritten extents to real allocations in
106 	 * end_io() when necessary, otherwise a racing buffer read would cache
107 	 * zeros from unwritten extents.
108 	 */
109 	if (!dio->error &&
110 	    (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
111 		int err;
112 		err = invalidate_inode_pages2_range(inode->i_mapping,
113 				offset >> PAGE_SHIFT,
114 				(offset + dio->size - 1) >> PAGE_SHIFT);
115 		if (err)
116 			dio_warn_stale_pagecache(iocb->ki_filp);
117 	}
118 
119 	/*
120 	 * If this is a DSYNC write, make sure we push it to stable storage now
121 	 * that we've written data.
122 	 */
123 	if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
124 		ret = generic_write_sync(iocb, ret);
125 
126 	inode_dio_end(file_inode(iocb->ki_filp));
127 	kfree(dio);
128 
129 	return ret;
130 }
131 
132 static void iomap_dio_complete_work(struct work_struct *work)
133 {
134 	struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
135 	struct kiocb *iocb = dio->iocb;
136 
137 	iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
138 }
139 
140 /*
141  * Set an error in the dio if none is set yet.  We have to use cmpxchg
142  * as the submission context and the completion context(s) can race to
143  * update the error.
144  */
145 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
146 {
147 	cmpxchg(&dio->error, 0, ret);
148 }
149 
150 static void iomap_dio_bio_end_io(struct bio *bio)
151 {
152 	struct iomap_dio *dio = bio->bi_private;
153 	bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
154 
155 	if (bio->bi_status)
156 		iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
157 
158 	if (atomic_dec_and_test(&dio->ref)) {
159 		if (dio->wait_for_completion) {
160 			struct task_struct *waiter = dio->submit.waiter;
161 			WRITE_ONCE(dio->submit.waiter, NULL);
162 			blk_wake_io_task(waiter);
163 		} else if (dio->flags & IOMAP_DIO_WRITE) {
164 			struct inode *inode = file_inode(dio->iocb->ki_filp);
165 
166 			INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
167 			queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
168 		} else {
169 			iomap_dio_complete_work(&dio->aio.work);
170 		}
171 	}
172 
173 	if (should_dirty) {
174 		bio_check_pages_dirty(bio);
175 	} else {
176 		bio_release_pages(bio, false);
177 		bio_put(bio);
178 	}
179 }
180 
181 static void
182 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
183 		unsigned len)
184 {
185 	struct page *page = ZERO_PAGE(0);
186 	int flags = REQ_SYNC | REQ_IDLE;
187 	struct bio *bio;
188 
189 	bio = bio_alloc(GFP_KERNEL, 1);
190 	bio_set_dev(bio, iomap->bdev);
191 	bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
192 	bio->bi_private = dio;
193 	bio->bi_end_io = iomap_dio_bio_end_io;
194 
195 	get_page(page);
196 	__bio_add_page(bio, page, len, 0);
197 	bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
198 	iomap_dio_submit_bio(dio, iomap, bio);
199 }
200 
201 static loff_t
202 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
203 		struct iomap_dio *dio, struct iomap *iomap)
204 {
205 	unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
206 	unsigned int fs_block_size = i_blocksize(inode), pad;
207 	unsigned int align = iov_iter_alignment(dio->submit.iter);
208 	struct iov_iter iter;
209 	struct bio *bio;
210 	bool need_zeroout = false;
211 	bool use_fua = false;
212 	int nr_pages, ret = 0;
213 	size_t copied = 0;
214 
215 	if ((pos | length | align) & ((1 << blkbits) - 1))
216 		return -EINVAL;
217 
218 	if (iomap->type == IOMAP_UNWRITTEN) {
219 		dio->flags |= IOMAP_DIO_UNWRITTEN;
220 		need_zeroout = true;
221 	}
222 
223 	if (iomap->flags & IOMAP_F_SHARED)
224 		dio->flags |= IOMAP_DIO_COW;
225 
226 	if (iomap->flags & IOMAP_F_NEW) {
227 		need_zeroout = true;
228 	} else if (iomap->type == IOMAP_MAPPED) {
229 		/*
230 		 * Use a FUA write if we need datasync semantics, this is a pure
231 		 * data IO that doesn't require any metadata updates (including
232 		 * after IO completion such as unwritten extent conversion) and
233 		 * the underlying device supports FUA. This allows us to avoid
234 		 * cache flushes on IO completion.
235 		 */
236 		if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
237 		    (dio->flags & IOMAP_DIO_WRITE_FUA) &&
238 		    blk_queue_fua(bdev_get_queue(iomap->bdev)))
239 			use_fua = true;
240 	}
241 
242 	/*
243 	 * Operate on a partial iter trimmed to the extent we were called for.
244 	 * We'll update the iter in the dio once we're done with this extent.
245 	 */
246 	iter = *dio->submit.iter;
247 	iov_iter_truncate(&iter, length);
248 
249 	nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
250 	if (nr_pages <= 0)
251 		return nr_pages;
252 
253 	if (need_zeroout) {
254 		/* zero out from the start of the block to the write offset */
255 		pad = pos & (fs_block_size - 1);
256 		if (pad)
257 			iomap_dio_zero(dio, iomap, pos - pad, pad);
258 	}
259 
260 	do {
261 		size_t n;
262 		if (dio->error) {
263 			iov_iter_revert(dio->submit.iter, copied);
264 			return 0;
265 		}
266 
267 		bio = bio_alloc(GFP_KERNEL, nr_pages);
268 		bio_set_dev(bio, iomap->bdev);
269 		bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
270 		bio->bi_write_hint = dio->iocb->ki_hint;
271 		bio->bi_ioprio = dio->iocb->ki_ioprio;
272 		bio->bi_private = dio;
273 		bio->bi_end_io = iomap_dio_bio_end_io;
274 
275 		ret = bio_iov_iter_get_pages(bio, &iter);
276 		if (unlikely(ret)) {
277 			/*
278 			 * We have to stop part way through an IO. We must fall
279 			 * through to the sub-block tail zeroing here, otherwise
280 			 * this short IO may expose stale data in the tail of
281 			 * the block we haven't written data to.
282 			 */
283 			bio_put(bio);
284 			goto zero_tail;
285 		}
286 
287 		n = bio->bi_iter.bi_size;
288 		if (dio->flags & IOMAP_DIO_WRITE) {
289 			bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
290 			if (use_fua)
291 				bio->bi_opf |= REQ_FUA;
292 			else
293 				dio->flags &= ~IOMAP_DIO_WRITE_FUA;
294 			task_io_account_write(n);
295 		} else {
296 			bio->bi_opf = REQ_OP_READ;
297 			if (dio->flags & IOMAP_DIO_DIRTY)
298 				bio_set_pages_dirty(bio);
299 		}
300 
301 		iov_iter_advance(dio->submit.iter, n);
302 
303 		dio->size += n;
304 		pos += n;
305 		copied += n;
306 
307 		nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
308 		iomap_dio_submit_bio(dio, iomap, bio);
309 	} while (nr_pages);
310 
311 	/*
312 	 * We need to zeroout the tail of a sub-block write if the extent type
313 	 * requires zeroing or the write extends beyond EOF. If we don't zero
314 	 * the block tail in the latter case, we can expose stale data via mmap
315 	 * reads of the EOF block.
316 	 */
317 zero_tail:
318 	if (need_zeroout ||
319 	    ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
320 		/* zero out from the end of the write to the end of the block */
321 		pad = pos & (fs_block_size - 1);
322 		if (pad)
323 			iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
324 	}
325 	return copied ? copied : ret;
326 }
327 
328 static loff_t
329 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
330 {
331 	length = iov_iter_zero(length, dio->submit.iter);
332 	dio->size += length;
333 	return length;
334 }
335 
336 static loff_t
337 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
338 		struct iomap_dio *dio, struct iomap *iomap)
339 {
340 	struct iov_iter *iter = dio->submit.iter;
341 	size_t copied;
342 
343 	BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
344 
345 	if (dio->flags & IOMAP_DIO_WRITE) {
346 		loff_t size = inode->i_size;
347 
348 		if (pos > size)
349 			memset(iomap->inline_data + size, 0, pos - size);
350 		copied = copy_from_iter(iomap->inline_data + pos, length, iter);
351 		if (copied) {
352 			if (pos + copied > size)
353 				i_size_write(inode, pos + copied);
354 			mark_inode_dirty(inode);
355 		}
356 	} else {
357 		copied = copy_to_iter(iomap->inline_data + pos, length, iter);
358 	}
359 	dio->size += copied;
360 	return copied;
361 }
362 
363 static loff_t
364 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
365 		void *data, struct iomap *iomap)
366 {
367 	struct iomap_dio *dio = data;
368 
369 	switch (iomap->type) {
370 	case IOMAP_HOLE:
371 		if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
372 			return -EIO;
373 		return iomap_dio_hole_actor(length, dio);
374 	case IOMAP_UNWRITTEN:
375 		if (!(dio->flags & IOMAP_DIO_WRITE))
376 			return iomap_dio_hole_actor(length, dio);
377 		return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
378 	case IOMAP_MAPPED:
379 		return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
380 	case IOMAP_INLINE:
381 		return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
382 	default:
383 		WARN_ON_ONCE(1);
384 		return -EIO;
385 	}
386 }
387 
388 /*
389  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
390  * is being issued as AIO or not.  This allows us to optimise pure data writes
391  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
392  * REQ_FLUSH post write. This is slightly tricky because a single request here
393  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
394  * may be pure data writes. In that case, we still need to do a full data sync
395  * completion.
396  */
397 ssize_t
398 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
399 		const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
400 {
401 	struct address_space *mapping = iocb->ki_filp->f_mapping;
402 	struct inode *inode = file_inode(iocb->ki_filp);
403 	size_t count = iov_iter_count(iter);
404 	loff_t pos = iocb->ki_pos, start = pos;
405 	loff_t end = iocb->ki_pos + count - 1, ret = 0;
406 	unsigned int flags = IOMAP_DIRECT;
407 	bool wait_for_completion = is_sync_kiocb(iocb);
408 	struct blk_plug plug;
409 	struct iomap_dio *dio;
410 
411 	lockdep_assert_held(&inode->i_rwsem);
412 
413 	if (!count)
414 		return 0;
415 
416 	dio = kmalloc(sizeof(*dio), GFP_KERNEL);
417 	if (!dio)
418 		return -ENOMEM;
419 
420 	dio->iocb = iocb;
421 	atomic_set(&dio->ref, 1);
422 	dio->size = 0;
423 	dio->i_size = i_size_read(inode);
424 	dio->end_io = end_io;
425 	dio->error = 0;
426 	dio->flags = 0;
427 
428 	dio->submit.iter = iter;
429 	dio->submit.waiter = current;
430 	dio->submit.cookie = BLK_QC_T_NONE;
431 	dio->submit.last_queue = NULL;
432 
433 	if (iov_iter_rw(iter) == READ) {
434 		if (pos >= dio->i_size)
435 			goto out_free_dio;
436 
437 		if (iter_is_iovec(iter) && iov_iter_rw(iter) == READ)
438 			dio->flags |= IOMAP_DIO_DIRTY;
439 	} else {
440 		flags |= IOMAP_WRITE;
441 		dio->flags |= IOMAP_DIO_WRITE;
442 
443 		/* for data sync or sync, we need sync completion processing */
444 		if (iocb->ki_flags & IOCB_DSYNC)
445 			dio->flags |= IOMAP_DIO_NEED_SYNC;
446 
447 		/*
448 		 * For datasync only writes, we optimistically try using FUA for
449 		 * this IO.  Any non-FUA write that occurs will clear this flag,
450 		 * hence we know before completion whether a cache flush is
451 		 * necessary.
452 		 */
453 		if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
454 			dio->flags |= IOMAP_DIO_WRITE_FUA;
455 	}
456 
457 	if (iocb->ki_flags & IOCB_NOWAIT) {
458 		if (filemap_range_has_page(mapping, start, end)) {
459 			ret = -EAGAIN;
460 			goto out_free_dio;
461 		}
462 		flags |= IOMAP_NOWAIT;
463 	}
464 
465 	ret = filemap_write_and_wait_range(mapping, start, end);
466 	if (ret)
467 		goto out_free_dio;
468 
469 	/*
470 	 * Try to invalidate cache pages for the range we're direct
471 	 * writing.  If this invalidation fails, tough, the write will
472 	 * still work, but racing two incompatible write paths is a
473 	 * pretty crazy thing to do, so we don't support it 100%.
474 	 */
475 	ret = invalidate_inode_pages2_range(mapping,
476 			start >> PAGE_SHIFT, end >> PAGE_SHIFT);
477 	if (ret)
478 		dio_warn_stale_pagecache(iocb->ki_filp);
479 	ret = 0;
480 
481 	if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
482 	    !inode->i_sb->s_dio_done_wq) {
483 		ret = sb_init_dio_done_wq(inode->i_sb);
484 		if (ret < 0)
485 			goto out_free_dio;
486 	}
487 
488 	inode_dio_begin(inode);
489 
490 	blk_start_plug(&plug);
491 	do {
492 		ret = iomap_apply(inode, pos, count, flags, ops, dio,
493 				iomap_dio_actor);
494 		if (ret <= 0) {
495 			/* magic error code to fall back to buffered I/O */
496 			if (ret == -ENOTBLK) {
497 				wait_for_completion = true;
498 				ret = 0;
499 			}
500 			break;
501 		}
502 		pos += ret;
503 
504 		if (iov_iter_rw(iter) == READ && pos >= dio->i_size)
505 			break;
506 	} while ((count = iov_iter_count(iter)) > 0);
507 	blk_finish_plug(&plug);
508 
509 	if (ret < 0)
510 		iomap_dio_set_error(dio, ret);
511 
512 	/*
513 	 * If all the writes we issued were FUA, we don't need to flush the
514 	 * cache on IO completion. Clear the sync flag for this case.
515 	 */
516 	if (dio->flags & IOMAP_DIO_WRITE_FUA)
517 		dio->flags &= ~IOMAP_DIO_NEED_SYNC;
518 
519 	WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
520 	WRITE_ONCE(iocb->private, dio->submit.last_queue);
521 
522 	/*
523 	 * We are about to drop our additional submission reference, which
524 	 * might be the last reference to the dio.  There are three three
525 	 * different ways we can progress here:
526 	 *
527 	 *  (a) If this is the last reference we will always complete and free
528 	 *	the dio ourselves.
529 	 *  (b) If this is not the last reference, and we serve an asynchronous
530 	 *	iocb, we must never touch the dio after the decrement, the
531 	 *	I/O completion handler will complete and free it.
532 	 *  (c) If this is not the last reference, but we serve a synchronous
533 	 *	iocb, the I/O completion handler will wake us up on the drop
534 	 *	of the final reference, and we will complete and free it here
535 	 *	after we got woken by the I/O completion handler.
536 	 */
537 	dio->wait_for_completion = wait_for_completion;
538 	if (!atomic_dec_and_test(&dio->ref)) {
539 		if (!wait_for_completion)
540 			return -EIOCBQUEUED;
541 
542 		for (;;) {
543 			set_current_state(TASK_UNINTERRUPTIBLE);
544 			if (!READ_ONCE(dio->submit.waiter))
545 				break;
546 
547 			if (!(iocb->ki_flags & IOCB_HIPRI) ||
548 			    !dio->submit.last_queue ||
549 			    !blk_poll(dio->submit.last_queue,
550 					 dio->submit.cookie, true))
551 				io_schedule();
552 		}
553 		__set_current_state(TASK_RUNNING);
554 	}
555 
556 	return iomap_dio_complete(dio);
557 
558 out_free_dio:
559 	kfree(dio);
560 	return ret;
561 }
562 EXPORT_SYMBOL_GPL(iomap_dio_rw);
563