xref: /openbmc/linux/fs/iomap/direct-io.c (revision 29c37341)
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 
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 
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 
79 static 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 &&
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 	/*
122 	 * If this is a DSYNC write, make sure we push it to stable storage now
123 	 * that we've written data.
124 	 */
125 	if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
126 		ret = generic_write_sync(iocb, ret);
127 
128 	inode_dio_end(file_inode(iocb->ki_filp));
129 	kfree(dio);
130 
131 	return ret;
132 }
133 
134 static void iomap_dio_complete_work(struct work_struct *work)
135 {
136 	struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
137 	struct kiocb *iocb = dio->iocb;
138 
139 	iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
140 }
141 
142 /*
143  * Set an error in the dio if none is set yet.  We have to use cmpxchg
144  * as the submission context and the completion context(s) can race to
145  * update the error.
146  */
147 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
148 {
149 	cmpxchg(&dio->error, 0, ret);
150 }
151 
152 static void iomap_dio_bio_end_io(struct bio *bio)
153 {
154 	struct iomap_dio *dio = bio->bi_private;
155 	bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
156 
157 	if (bio->bi_status)
158 		iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
159 
160 	if (atomic_dec_and_test(&dio->ref)) {
161 		if (dio->wait_for_completion) {
162 			struct task_struct *waiter = dio->submit.waiter;
163 			WRITE_ONCE(dio->submit.waiter, NULL);
164 			blk_wake_io_task(waiter);
165 		} else if (dio->flags & IOMAP_DIO_WRITE) {
166 			struct inode *inode = file_inode(dio->iocb->ki_filp);
167 
168 			INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
169 			queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
170 		} else {
171 			iomap_dio_complete_work(&dio->aio.work);
172 		}
173 	}
174 
175 	if (should_dirty) {
176 		bio_check_pages_dirty(bio);
177 	} else {
178 		bio_release_pages(bio, false);
179 		bio_put(bio);
180 	}
181 }
182 
183 static void
184 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
185 		unsigned len)
186 {
187 	struct page *page = ZERO_PAGE(0);
188 	int flags = REQ_SYNC | REQ_IDLE;
189 	struct bio *bio;
190 
191 	bio = bio_alloc(GFP_KERNEL, 1);
192 	bio_set_dev(bio, iomap->bdev);
193 	bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
194 	bio->bi_private = dio;
195 	bio->bi_end_io = iomap_dio_bio_end_io;
196 
197 	get_page(page);
198 	__bio_add_page(bio, page, len, 0);
199 	bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
200 	iomap_dio_submit_bio(dio, iomap, bio, pos);
201 }
202 
203 static loff_t
204 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
205 		struct iomap_dio *dio, struct iomap *iomap)
206 {
207 	unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
208 	unsigned int fs_block_size = i_blocksize(inode), pad;
209 	unsigned int align = iov_iter_alignment(dio->submit.iter);
210 	struct bio *bio;
211 	bool need_zeroout = false;
212 	bool use_fua = false;
213 	int nr_pages, ret = 0;
214 	size_t copied = 0;
215 	size_t orig_count;
216 
217 	if ((pos | length | align) & ((1 << blkbits) - 1))
218 		return -EINVAL;
219 
220 	if (iomap->type == IOMAP_UNWRITTEN) {
221 		dio->flags |= IOMAP_DIO_UNWRITTEN;
222 		need_zeroout = true;
223 	}
224 
225 	if (iomap->flags & IOMAP_F_SHARED)
226 		dio->flags |= IOMAP_DIO_COW;
227 
228 	if (iomap->flags & IOMAP_F_NEW) {
229 		need_zeroout = true;
230 	} else if (iomap->type == IOMAP_MAPPED) {
231 		/*
232 		 * Use a FUA write if we need datasync semantics, this is a pure
233 		 * data IO that doesn't require any metadata updates (including
234 		 * after IO completion such as unwritten extent conversion) and
235 		 * the underlying device supports FUA. This allows us to avoid
236 		 * cache flushes on IO completion.
237 		 */
238 		if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
239 		    (dio->flags & IOMAP_DIO_WRITE_FUA) &&
240 		    blk_queue_fua(bdev_get_queue(iomap->bdev)))
241 			use_fua = true;
242 	}
243 
244 	/*
245 	 * Save the original count and trim the iter to just the extent we
246 	 * are operating on right now.  The iter will be re-expanded once
247 	 * we are done.
248 	 */
249 	orig_count = iov_iter_count(dio->submit.iter);
250 	iov_iter_truncate(dio->submit.iter, length);
251 
252 	nr_pages = iov_iter_npages(dio->submit.iter, BIO_MAX_PAGES);
253 	if (nr_pages <= 0) {
254 		ret = nr_pages;
255 		goto out;
256 	}
257 
258 	if (need_zeroout) {
259 		/* zero out from the start of the block to the write offset */
260 		pad = pos & (fs_block_size - 1);
261 		if (pad)
262 			iomap_dio_zero(dio, iomap, pos - pad, pad);
263 	}
264 
265 	do {
266 		size_t n;
267 		if (dio->error) {
268 			iov_iter_revert(dio->submit.iter, copied);
269 			copied = ret = 0;
270 			goto out;
271 		}
272 
273 		bio = bio_alloc(GFP_KERNEL, nr_pages);
274 		bio_set_dev(bio, iomap->bdev);
275 		bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
276 		bio->bi_write_hint = dio->iocb->ki_hint;
277 		bio->bi_ioprio = dio->iocb->ki_ioprio;
278 		bio->bi_private = dio;
279 		bio->bi_end_io = iomap_dio_bio_end_io;
280 
281 		ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
282 		if (unlikely(ret)) {
283 			/*
284 			 * We have to stop part way through an IO. We must fall
285 			 * through to the sub-block tail zeroing here, otherwise
286 			 * this short IO may expose stale data in the tail of
287 			 * the block we haven't written data to.
288 			 */
289 			bio_put(bio);
290 			goto zero_tail;
291 		}
292 
293 		n = bio->bi_iter.bi_size;
294 		if (dio->flags & IOMAP_DIO_WRITE) {
295 			bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
296 			if (use_fua)
297 				bio->bi_opf |= REQ_FUA;
298 			else
299 				dio->flags &= ~IOMAP_DIO_WRITE_FUA;
300 			task_io_account_write(n);
301 		} else {
302 			bio->bi_opf = REQ_OP_READ;
303 			if (dio->flags & IOMAP_DIO_DIRTY)
304 				bio_set_pages_dirty(bio);
305 		}
306 
307 		dio->size += n;
308 		copied += n;
309 
310 		nr_pages = iov_iter_npages(dio->submit.iter, BIO_MAX_PAGES);
311 		iomap_dio_submit_bio(dio, iomap, bio, pos);
312 		pos += n;
313 	} while (nr_pages);
314 
315 	/*
316 	 * We need to zeroout the tail of a sub-block write if the extent type
317 	 * requires zeroing or the write extends beyond EOF. If we don't zero
318 	 * the block tail in the latter case, we can expose stale data via mmap
319 	 * reads of the EOF block.
320 	 */
321 zero_tail:
322 	if (need_zeroout ||
323 	    ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
324 		/* zero out from the end of the write to the end of the block */
325 		pad = pos & (fs_block_size - 1);
326 		if (pad)
327 			iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
328 	}
329 out:
330 	/* Undo iter limitation to current extent */
331 	iov_iter_reexpand(dio->submit.iter, orig_count - copied);
332 	if (copied)
333 		return copied;
334 	return ret;
335 }
336 
337 static loff_t
338 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
339 {
340 	length = iov_iter_zero(length, dio->submit.iter);
341 	dio->size += length;
342 	return length;
343 }
344 
345 static loff_t
346 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
347 		struct iomap_dio *dio, struct iomap *iomap)
348 {
349 	struct iov_iter *iter = dio->submit.iter;
350 	size_t copied;
351 
352 	BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
353 
354 	if (dio->flags & IOMAP_DIO_WRITE) {
355 		loff_t size = inode->i_size;
356 
357 		if (pos > size)
358 			memset(iomap->inline_data + size, 0, pos - size);
359 		copied = copy_from_iter(iomap->inline_data + pos, length, iter);
360 		if (copied) {
361 			if (pos + copied > size)
362 				i_size_write(inode, pos + copied);
363 			mark_inode_dirty(inode);
364 		}
365 	} else {
366 		copied = copy_to_iter(iomap->inline_data + pos, length, iter);
367 	}
368 	dio->size += copied;
369 	return copied;
370 }
371 
372 static loff_t
373 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
374 		void *data, struct iomap *iomap, struct iomap *srcmap)
375 {
376 	struct iomap_dio *dio = data;
377 
378 	switch (iomap->type) {
379 	case IOMAP_HOLE:
380 		if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
381 			return -EIO;
382 		return iomap_dio_hole_actor(length, dio);
383 	case IOMAP_UNWRITTEN:
384 		if (!(dio->flags & IOMAP_DIO_WRITE))
385 			return iomap_dio_hole_actor(length, dio);
386 		return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
387 	case IOMAP_MAPPED:
388 		return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
389 	case IOMAP_INLINE:
390 		return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
391 	default:
392 		WARN_ON_ONCE(1);
393 		return -EIO;
394 	}
395 }
396 
397 /*
398  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
399  * is being issued as AIO or not.  This allows us to optimise pure data writes
400  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
401  * REQ_FLUSH post write. This is slightly tricky because a single request here
402  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
403  * may be pure data writes. In that case, we still need to do a full data sync
404  * completion.
405  *
406  * Returns -ENOTBLK In case of a page invalidation invalidation failure for
407  * writes.  The callers needs to fall back to buffered I/O in this case.
408  */
409 ssize_t
410 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
411 		const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
412 		bool wait_for_completion)
413 {
414 	struct address_space *mapping = iocb->ki_filp->f_mapping;
415 	struct inode *inode = file_inode(iocb->ki_filp);
416 	size_t count = iov_iter_count(iter);
417 	loff_t pos = iocb->ki_pos;
418 	loff_t end = iocb->ki_pos + count - 1, ret = 0;
419 	unsigned int flags = IOMAP_DIRECT;
420 	struct blk_plug plug;
421 	struct iomap_dio *dio;
422 
423 	if (!count)
424 		return 0;
425 
426 	if (WARN_ON(is_sync_kiocb(iocb) && !wait_for_completion))
427 		return -EIO;
428 
429 	dio = kmalloc(sizeof(*dio), GFP_KERNEL);
430 	if (!dio)
431 		return -ENOMEM;
432 
433 	dio->iocb = iocb;
434 	atomic_set(&dio->ref, 1);
435 	dio->size = 0;
436 	dio->i_size = i_size_read(inode);
437 	dio->dops = dops;
438 	dio->error = 0;
439 	dio->flags = 0;
440 
441 	dio->submit.iter = iter;
442 	dio->submit.waiter = current;
443 	dio->submit.cookie = BLK_QC_T_NONE;
444 	dio->submit.last_queue = NULL;
445 
446 	if (iov_iter_rw(iter) == READ) {
447 		if (pos >= dio->i_size)
448 			goto out_free_dio;
449 
450 		if (iter_is_iovec(iter))
451 			dio->flags |= IOMAP_DIO_DIRTY;
452 	} else {
453 		flags |= IOMAP_WRITE;
454 		dio->flags |= IOMAP_DIO_WRITE;
455 
456 		/* for data sync or sync, we need sync completion processing */
457 		if (iocb->ki_flags & IOCB_DSYNC)
458 			dio->flags |= IOMAP_DIO_NEED_SYNC;
459 
460 		/*
461 		 * For datasync only writes, we optimistically try using FUA for
462 		 * this IO.  Any non-FUA write that occurs will clear this flag,
463 		 * hence we know before completion whether a cache flush is
464 		 * necessary.
465 		 */
466 		if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
467 			dio->flags |= IOMAP_DIO_WRITE_FUA;
468 	}
469 
470 	if (iocb->ki_flags & IOCB_NOWAIT) {
471 		if (filemap_range_has_page(mapping, pos, end)) {
472 			ret = -EAGAIN;
473 			goto out_free_dio;
474 		}
475 		flags |= IOMAP_NOWAIT;
476 	}
477 
478 	ret = filemap_write_and_wait_range(mapping, pos, end);
479 	if (ret)
480 		goto out_free_dio;
481 
482 	if (iov_iter_rw(iter) == WRITE) {
483 		/*
484 		 * Try to invalidate cache pages for the range we are writing.
485 		 * If this invalidation fails, let the caller fall back to
486 		 * buffered I/O.
487 		 */
488 		if (invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT,
489 				end >> PAGE_SHIFT)) {
490 			trace_iomap_dio_invalidate_fail(inode, pos, count);
491 			ret = -ENOTBLK;
492 			goto out_free_dio;
493 		}
494 
495 		if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
496 			ret = sb_init_dio_done_wq(inode->i_sb);
497 			if (ret < 0)
498 				goto out_free_dio;
499 		}
500 	}
501 
502 	inode_dio_begin(inode);
503 
504 	blk_start_plug(&plug);
505 	do {
506 		ret = iomap_apply(inode, pos, count, flags, ops, dio,
507 				iomap_dio_actor);
508 		if (ret <= 0) {
509 			/* magic error code to fall back to buffered I/O */
510 			if (ret == -ENOTBLK) {
511 				wait_for_completion = true;
512 				ret = 0;
513 			}
514 			break;
515 		}
516 		pos += ret;
517 
518 		if (iov_iter_rw(iter) == READ && pos >= dio->i_size) {
519 			/*
520 			 * We only report that we've read data up to i_size.
521 			 * Revert iter to a state corresponding to that as
522 			 * some callers (such as splice code) rely on it.
523 			 */
524 			iov_iter_revert(iter, pos - dio->i_size);
525 			break;
526 		}
527 	} while ((count = iov_iter_count(iter)) > 0);
528 	blk_finish_plug(&plug);
529 
530 	if (ret < 0)
531 		iomap_dio_set_error(dio, ret);
532 
533 	/*
534 	 * If all the writes we issued were FUA, we don't need to flush the
535 	 * cache on IO completion. Clear the sync flag for this case.
536 	 */
537 	if (dio->flags & IOMAP_DIO_WRITE_FUA)
538 		dio->flags &= ~IOMAP_DIO_NEED_SYNC;
539 
540 	WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
541 	WRITE_ONCE(iocb->private, dio->submit.last_queue);
542 
543 	/*
544 	 * We are about to drop our additional submission reference, which
545 	 * might be the last reference to the dio.  There are three different
546 	 * ways we can progress here:
547 	 *
548 	 *  (a) If this is the last reference we will always complete and free
549 	 *	the dio ourselves.
550 	 *  (b) If this is not the last reference, and we serve an asynchronous
551 	 *	iocb, we must never touch the dio after the decrement, the
552 	 *	I/O completion handler will complete and free it.
553 	 *  (c) If this is not the last reference, but we serve a synchronous
554 	 *	iocb, the I/O completion handler will wake us up on the drop
555 	 *	of the final reference, and we will complete and free it here
556 	 *	after we got woken by the I/O completion handler.
557 	 */
558 	dio->wait_for_completion = wait_for_completion;
559 	if (!atomic_dec_and_test(&dio->ref)) {
560 		if (!wait_for_completion)
561 			return -EIOCBQUEUED;
562 
563 		for (;;) {
564 			set_current_state(TASK_UNINTERRUPTIBLE);
565 			if (!READ_ONCE(dio->submit.waiter))
566 				break;
567 
568 			if (!(iocb->ki_flags & IOCB_HIPRI) ||
569 			    !dio->submit.last_queue ||
570 			    !blk_poll(dio->submit.last_queue,
571 					 dio->submit.cookie, true))
572 				blk_io_schedule();
573 		}
574 		__set_current_state(TASK_RUNNING);
575 	}
576 
577 	return iomap_dio_complete(dio);
578 
579 out_free_dio:
580 	kfree(dio);
581 	return ret;
582 }
583 EXPORT_SYMBOL_GPL(iomap_dio_rw);
584