xref: /linux/fs/iomap/ioend.c (revision 0b3bb205808195159be633a8cefb602670e856fb)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2016-2025 Christoph Hellwig.
4  */
5 #include <linux/iomap.h>
6 #include <linux/list_sort.h>
7 #include <linux/pagemap.h>
8 #include <linux/writeback.h>
9 #include <linux/fserror.h>
10 #include "internal.h"
11 #include "trace.h"
12 
13 struct bio_set iomap_ioend_bioset;
14 EXPORT_SYMBOL_GPL(iomap_ioend_bioset);
15 
iomap_init_ioend(struct inode * inode,struct bio * bio,loff_t file_offset,u16 ioend_flags)16 struct iomap_ioend *iomap_init_ioend(struct inode *inode,
17 		struct bio *bio, loff_t file_offset, u16 ioend_flags)
18 {
19 	struct iomap_ioend *ioend = iomap_ioend_from_bio(bio);
20 
21 	atomic_set(&ioend->io_remaining, 1);
22 	ioend->io_error = 0;
23 	ioend->io_parent = NULL;
24 	INIT_LIST_HEAD(&ioend->io_list);
25 	ioend->io_flags = ioend_flags;
26 	ioend->io_inode = inode;
27 	ioend->io_offset = file_offset;
28 	ioend->io_size = bio->bi_iter.bi_size;
29 	ioend->io_sector = bio->bi_iter.bi_sector;
30 	ioend->io_private = NULL;
31 	return ioend;
32 }
33 EXPORT_SYMBOL_GPL(iomap_init_ioend);
34 
35 /*
36  * We're now finished for good with this ioend structure.  Update the folio
37  * state, release holds on bios, and finally free up memory.  Do not use the
38  * ioend after this.
39  */
iomap_finish_ioend_buffered(struct iomap_ioend * ioend)40 static u32 iomap_finish_ioend_buffered(struct iomap_ioend *ioend)
41 {
42 	struct inode *inode = ioend->io_inode;
43 	struct bio *bio = &ioend->io_bio;
44 	struct folio_iter fi;
45 	u32 folio_count = 0;
46 
47 	if (ioend->io_error) {
48 		mapping_set_error(inode->i_mapping, ioend->io_error);
49 		if (!bio_flagged(bio, BIO_QUIET)) {
50 			pr_err_ratelimited(
51 "%s: writeback error on inode %lu, offset %lld, sector %llu",
52 				inode->i_sb->s_id, inode->i_ino,
53 				ioend->io_offset, ioend->io_sector);
54 		}
55 	}
56 
57 	/* walk all folios in bio, ending page IO on them */
58 	bio_for_each_folio_all(fi, bio) {
59 		if (ioend->io_error)
60 			fserror_report_io(inode, FSERR_BUFFERED_WRITE,
61 					  folio_pos(fi.folio) + fi.offset,
62 					  fi.length, ioend->io_error,
63 					  GFP_ATOMIC);
64 		iomap_finish_folio_write(inode, fi.folio, fi.length);
65 		folio_count++;
66 	}
67 
68 	bio_put(bio);	/* frees the ioend */
69 	return folio_count;
70 }
71 
72 static DEFINE_SPINLOCK(failed_ioend_lock);
73 static LIST_HEAD(failed_ioend_list);
74 
75 static void
iomap_fail_ioends(struct work_struct * work)76 iomap_fail_ioends(
77 	struct work_struct	*work)
78 {
79 	struct iomap_ioend	*ioend;
80 	struct list_head	tmp;
81 	unsigned long		flags;
82 
83 	spin_lock_irqsave(&failed_ioend_lock, flags);
84 	list_replace_init(&failed_ioend_list, &tmp);
85 	spin_unlock_irqrestore(&failed_ioend_lock, flags);
86 
87 	while ((ioend = list_first_entry_or_null(&tmp, struct iomap_ioend,
88 			io_list))) {
89 		list_del_init(&ioend->io_list);
90 		iomap_finish_ioend_buffered(ioend);
91 		cond_resched();
92 	}
93 }
94 
95 static DECLARE_WORK(failed_ioend_work, iomap_fail_ioends);
96 
iomap_fail_ioend_buffered(struct iomap_ioend * ioend)97 static void iomap_fail_ioend_buffered(struct iomap_ioend *ioend)
98 {
99 	unsigned long flags;
100 
101 	/*
102 	 * Bounce I/O errors to a workqueue to avoid nested i_lock acquisitions
103 	 * in the fserror code.  The caller no longer owns the ioend reference
104 	 * after the spinlock drops.
105 	 */
106 	spin_lock_irqsave(&failed_ioend_lock, flags);
107 	if (list_empty(&failed_ioend_list))
108 		WARN_ON_ONCE(!schedule_work(&failed_ioend_work));
109 	list_add_tail(&ioend->io_list, &failed_ioend_list);
110 	spin_unlock_irqrestore(&failed_ioend_lock, flags);
111 }
112 
ioend_writeback_end_bio(struct bio * bio)113 static void ioend_writeback_end_bio(struct bio *bio)
114 {
115 	struct iomap_ioend *ioend = iomap_ioend_from_bio(bio);
116 
117 	ioend->io_error = blk_status_to_errno(bio->bi_status);
118 	if (ioend->io_error) {
119 		iomap_fail_ioend_buffered(ioend);
120 		return;
121 	}
122 
123 	iomap_finish_ioend_buffered(ioend);
124 }
125 
126 /*
127  * We cannot cancel the ioend directly in case of an error, so call the bio end
128  * I/O handler with the error status here to run the normal I/O completion
129  * handler.
130  */
iomap_ioend_writeback_submit(struct iomap_writepage_ctx * wpc,int error)131 int iomap_ioend_writeback_submit(struct iomap_writepage_ctx *wpc, int error)
132 {
133 	struct iomap_ioend *ioend = wpc->wb_ctx;
134 
135 	if (!ioend->io_bio.bi_end_io)
136 		ioend->io_bio.bi_end_io = ioend_writeback_end_bio;
137 
138 	if (WARN_ON_ONCE(wpc->iomap.flags & IOMAP_F_ANON_WRITE))
139 		error = -EIO;
140 
141 	if (error) {
142 		ioend->io_bio.bi_status = errno_to_blk_status(error);
143 		bio_endio(&ioend->io_bio);
144 		return error;
145 	}
146 
147 	submit_bio(&ioend->io_bio);
148 	return 0;
149 }
150 EXPORT_SYMBOL_GPL(iomap_ioend_writeback_submit);
151 
iomap_alloc_ioend(struct iomap_writepage_ctx * wpc,loff_t pos,u16 ioend_flags)152 static struct iomap_ioend *iomap_alloc_ioend(struct iomap_writepage_ctx *wpc,
153 		loff_t pos, u16 ioend_flags)
154 {
155 	struct bio *bio;
156 
157 	bio = bio_alloc_bioset(wpc->iomap.bdev, BIO_MAX_VECS,
158 			       REQ_OP_WRITE | wbc_to_write_flags(wpc->wbc),
159 			       GFP_NOFS, &iomap_ioend_bioset);
160 	bio->bi_iter.bi_sector = iomap_sector(&wpc->iomap, pos);
161 	bio->bi_write_hint = wpc->inode->i_write_hint;
162 	wbc_init_bio(wpc->wbc, bio);
163 	wpc->nr_folios = 0;
164 	return iomap_init_ioend(wpc->inode, bio, pos, ioend_flags);
165 }
166 
iomap_can_add_to_ioend(struct iomap_writepage_ctx * wpc,loff_t pos,u16 ioend_flags)167 static bool iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t pos,
168 		u16 ioend_flags)
169 {
170 	struct iomap_ioend *ioend = wpc->wb_ctx;
171 
172 	if (ioend_flags & IOMAP_IOEND_BOUNDARY)
173 		return false;
174 	if ((ioend_flags & IOMAP_IOEND_NOMERGE_FLAGS) !=
175 	    (ioend->io_flags & IOMAP_IOEND_NOMERGE_FLAGS))
176 		return false;
177 	if (pos != ioend->io_offset + ioend->io_size)
178 		return false;
179 	if (!(wpc->iomap.flags & IOMAP_F_ANON_WRITE) &&
180 	    iomap_sector(&wpc->iomap, pos) != bio_end_sector(&ioend->io_bio))
181 		return false;
182 	/*
183 	 * Limit ioend bio chain lengths to minimise IO completion latency. This
184 	 * also prevents long tight loops ending page writeback on all the
185 	 * folios in the ioend.
186 	 */
187 	if (wpc->nr_folios >= IOEND_BATCH_SIZE)
188 		return false;
189 	return true;
190 }
191 
192 /*
193  * Test to see if we have an existing ioend structure that we could append to
194  * first; otherwise finish off the current ioend and start another.
195  *
196  * If a new ioend is created and cached, the old ioend is submitted to the block
197  * layer instantly.  Batching optimisations are provided by higher level block
198  * plugging.
199  *
200  * At the end of a writeback pass, there will be a cached ioend remaining on the
201  * writepage context that the caller will need to submit.
202  */
iomap_add_to_ioend(struct iomap_writepage_ctx * wpc,struct folio * folio,loff_t pos,loff_t end_pos,unsigned int dirty_len)203 ssize_t iomap_add_to_ioend(struct iomap_writepage_ctx *wpc, struct folio *folio,
204 		loff_t pos, loff_t end_pos, unsigned int dirty_len)
205 {
206 	struct iomap_ioend *ioend = wpc->wb_ctx;
207 	size_t poff = offset_in_folio(folio, pos);
208 	unsigned int ioend_flags = 0;
209 	unsigned int map_len = min_t(u64, dirty_len,
210 		wpc->iomap.offset + wpc->iomap.length - pos);
211 	int error;
212 
213 	trace_iomap_add_to_ioend(wpc->inode, pos, dirty_len, &wpc->iomap);
214 
215 	WARN_ON_ONCE(!folio->private && map_len < dirty_len);
216 
217 	switch (wpc->iomap.type) {
218 	case IOMAP_UNWRITTEN:
219 		ioend_flags |= IOMAP_IOEND_UNWRITTEN;
220 		break;
221 	case IOMAP_MAPPED:
222 		break;
223 	case IOMAP_HOLE:
224 		return map_len;
225 	default:
226 		WARN_ON_ONCE(1);
227 		return -EIO;
228 	}
229 
230 	if (wpc->iomap.flags & IOMAP_F_SHARED)
231 		ioend_flags |= IOMAP_IOEND_SHARED;
232 	if (folio_test_dropbehind(folio))
233 		ioend_flags |= IOMAP_IOEND_DONTCACHE;
234 	if (pos == wpc->iomap.offset && (wpc->iomap.flags & IOMAP_F_BOUNDARY))
235 		ioend_flags |= IOMAP_IOEND_BOUNDARY;
236 
237 	if (!ioend || !iomap_can_add_to_ioend(wpc, pos, ioend_flags)) {
238 new_ioend:
239 		if (ioend) {
240 			error = wpc->ops->writeback_submit(wpc, 0);
241 			if (error)
242 				return error;
243 		}
244 		wpc->wb_ctx = ioend = iomap_alloc_ioend(wpc, pos, ioend_flags);
245 	}
246 
247 	if (!bio_add_folio(&ioend->io_bio, folio, map_len, poff))
248 		goto new_ioend;
249 
250 	/*
251 	 * Clamp io_offset and io_size to the incore EOF so that ondisk
252 	 * file size updates in the ioend completion are byte-accurate.
253 	 * This avoids recovering files with zeroed tail regions when
254 	 * writeback races with appending writes:
255 	 *
256 	 *    Thread 1:                  Thread 2:
257 	 *    ------------               -----------
258 	 *    write [A, A+B]
259 	 *    update inode size to A+B
260 	 *    submit I/O [A, A+BS]
261 	 *                               write [A+B, A+B+C]
262 	 *                               update inode size to A+B+C
263 	 *    <I/O completes, updates disk size to min(A+B+C, A+BS)>
264 	 *    <power failure>
265 	 *
266 	 *  After reboot:
267 	 *    1) with A+B+C < A+BS, the file has zero padding in range
268 	 *       [A+B, A+B+C]
269 	 *
270 	 *    |<     Block Size (BS)   >|
271 	 *    |DDDDDDDDDDDD0000000000000|
272 	 *    ^           ^        ^
273 	 *    A          A+B     A+B+C
274 	 *                       (EOF)
275 	 *
276 	 *    2) with A+B+C > A+BS, the file has zero padding in range
277 	 *       [A+B, A+BS]
278 	 *
279 	 *    |<     Block Size (BS)   >|<     Block Size (BS)    >|
280 	 *    |DDDDDDDDDDDD0000000000000|00000000000000000000000000|
281 	 *    ^           ^             ^           ^
282 	 *    A          A+B           A+BS       A+B+C
283 	 *                             (EOF)
284 	 *
285 	 *    D = Valid Data
286 	 *    0 = Zero Padding
287 	 *
288 	 * Note that this defeats the ability to chain the ioends of
289 	 * appending writes.
290 	 */
291 	ioend->io_size += map_len;
292 	if (ioend->io_offset + ioend->io_size > end_pos)
293 		ioend->io_size = end_pos - ioend->io_offset;
294 
295 	wbc_account_cgroup_owner(wpc->wbc, folio, map_len);
296 	return map_len;
297 }
298 EXPORT_SYMBOL_GPL(iomap_add_to_ioend);
299 
iomap_finish_ioend(struct iomap_ioend * ioend,int error)300 static u32 iomap_finish_ioend(struct iomap_ioend *ioend, int error)
301 {
302 	if (ioend->io_parent) {
303 		struct bio *bio = &ioend->io_bio;
304 
305 		ioend = ioend->io_parent;
306 		bio_put(bio);
307 	}
308 
309 	if (error)
310 		cmpxchg(&ioend->io_error, 0, error);
311 
312 	if (!atomic_dec_and_test(&ioend->io_remaining))
313 		return 0;
314 	if (ioend->io_flags & IOMAP_IOEND_DIRECT)
315 		return iomap_finish_ioend_direct(ioend);
316 	return iomap_finish_ioend_buffered(ioend);
317 }
318 
319 /*
320  * Ioend completion routine for merged bios. This can only be called from task
321  * contexts as merged ioends can be of unbound length. Hence we have to break up
322  * the writeback completions into manageable chunks to avoid long scheduler
323  * holdoffs. We aim to keep scheduler holdoffs down below 10ms so that we get
324  * good batch processing throughput without creating adverse scheduler latency
325  * conditions.
326  */
iomap_finish_ioends(struct iomap_ioend * ioend,int error)327 void iomap_finish_ioends(struct iomap_ioend *ioend, int error)
328 {
329 	struct list_head tmp;
330 	u32 completions;
331 
332 	might_sleep();
333 
334 	list_replace_init(&ioend->io_list, &tmp);
335 	completions = iomap_finish_ioend(ioend, error);
336 
337 	while (!list_empty(&tmp)) {
338 		if (completions > IOEND_BATCH_SIZE * 8) {
339 			cond_resched();
340 			completions = 0;
341 		}
342 		ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
343 		list_del_init(&ioend->io_list);
344 		completions += iomap_finish_ioend(ioend, error);
345 	}
346 }
347 EXPORT_SYMBOL_GPL(iomap_finish_ioends);
348 
349 /*
350  * We can merge two adjacent ioends if they have the same set of work to do.
351  */
iomap_ioend_can_merge(struct iomap_ioend * ioend,struct iomap_ioend * next)352 static bool iomap_ioend_can_merge(struct iomap_ioend *ioend,
353 		struct iomap_ioend *next)
354 {
355 	/*
356 	 * There is no point in merging reads as there is no completion
357 	 * processing that can be easily batched up for them.
358 	 */
359 	if (bio_op(&ioend->io_bio) == REQ_OP_READ ||
360 	    bio_op(&next->io_bio) == REQ_OP_READ)
361 		return false;
362 
363 	if (ioend->io_bio.bi_status != next->io_bio.bi_status)
364 		return false;
365 	if (next->io_flags & IOMAP_IOEND_BOUNDARY)
366 		return false;
367 	if ((ioend->io_flags & IOMAP_IOEND_NOMERGE_FLAGS) !=
368 	    (next->io_flags & IOMAP_IOEND_NOMERGE_FLAGS))
369 		return false;
370 	if (ioend->io_offset + ioend->io_size != next->io_offset)
371 		return false;
372 	/*
373 	 * Do not merge physically discontiguous ioends. The filesystem
374 	 * completion functions will have to iterate the physical
375 	 * discontiguities even if we merge the ioends at a logical level, so
376 	 * we don't gain anything by merging physical discontiguities here.
377 	 *
378 	 * We cannot use bio->bi_iter.bi_sector here as it is modified during
379 	 * submission so does not point to the start sector of the bio at
380 	 * completion.
381 	 */
382 	if (ioend->io_sector + (ioend->io_size >> SECTOR_SHIFT) !=
383 	    next->io_sector)
384 		return false;
385 	return true;
386 }
387 
iomap_ioend_try_merge(struct iomap_ioend * ioend,struct list_head * more_ioends)388 void iomap_ioend_try_merge(struct iomap_ioend *ioend,
389 		struct list_head *more_ioends)
390 {
391 	struct iomap_ioend *next;
392 
393 	INIT_LIST_HEAD(&ioend->io_list);
394 
395 	while ((next = list_first_entry_or_null(more_ioends, struct iomap_ioend,
396 			io_list))) {
397 		if (!iomap_ioend_can_merge(ioend, next))
398 			break;
399 		list_move_tail(&next->io_list, &ioend->io_list);
400 		ioend->io_size += next->io_size;
401 	}
402 }
403 EXPORT_SYMBOL_GPL(iomap_ioend_try_merge);
404 
iomap_ioend_compare(void * priv,const struct list_head * a,const struct list_head * b)405 static int iomap_ioend_compare(void *priv, const struct list_head *a,
406 		const struct list_head *b)
407 {
408 	struct iomap_ioend *ia = container_of(a, struct iomap_ioend, io_list);
409 	struct iomap_ioend *ib = container_of(b, struct iomap_ioend, io_list);
410 
411 	if (ia->io_offset < ib->io_offset)
412 		return -1;
413 	if (ia->io_offset > ib->io_offset)
414 		return 1;
415 	return 0;
416 }
417 
iomap_sort_ioends(struct list_head * ioend_list)418 void iomap_sort_ioends(struct list_head *ioend_list)
419 {
420 	list_sort(NULL, ioend_list, iomap_ioend_compare);
421 }
422 EXPORT_SYMBOL_GPL(iomap_sort_ioends);
423 
424 /*
425  * Split up to the first @max_len bytes from @ioend if the ioend covers more
426  * than @max_len bytes.
427  *
428  * If @is_append is set, the split will be based on the hardware limits for
429  * REQ_OP_ZONE_APPEND commands and can be less than @max_len if the hardware
430  * limits don't allow the entire @max_len length.
431  *
432  * The bio embedded into @ioend must be a REQ_OP_WRITE because the block layer
433  * does not allow splitting REQ_OP_ZONE_APPEND bios.  The file systems has to
434  * switch the operation after this call, but before submitting the bio.
435  */
iomap_split_ioend(struct iomap_ioend * ioend,unsigned int max_len,bool is_append)436 struct iomap_ioend *iomap_split_ioend(struct iomap_ioend *ioend,
437 		unsigned int max_len, bool is_append)
438 {
439 	struct bio *bio = &ioend->io_bio;
440 	struct iomap_ioend *split_ioend;
441 	unsigned int nr_segs;
442 	int sector_offset;
443 	struct bio *split;
444 
445 	if (is_append) {
446 		struct queue_limits *lim = bdev_limits(bio->bi_bdev);
447 
448 		max_len = min(max_len,
449 			      lim->max_zone_append_sectors << SECTOR_SHIFT);
450 
451 		sector_offset = bio_split_rw_at(bio, lim, &nr_segs, max_len);
452 		if (unlikely(sector_offset < 0))
453 			return ERR_PTR(sector_offset);
454 		if (!sector_offset)
455 			return NULL;
456 	} else {
457 		if (bio->bi_iter.bi_size <= max_len)
458 			return NULL;
459 		sector_offset = max_len >> SECTOR_SHIFT;
460 	}
461 
462 	/* ensure the split ioend is still block size aligned */
463 	sector_offset = ALIGN_DOWN(sector_offset << SECTOR_SHIFT,
464 			i_blocksize(ioend->io_inode)) >> SECTOR_SHIFT;
465 
466 	split = bio_split(bio, sector_offset, GFP_NOFS, &iomap_ioend_bioset);
467 	if (IS_ERR(split))
468 		return ERR_CAST(split);
469 	split->bi_private = bio->bi_private;
470 	split->bi_end_io = bio->bi_end_io;
471 
472 	split_ioend = iomap_init_ioend(ioend->io_inode, split, ioend->io_offset,
473 			ioend->io_flags);
474 	split_ioend->io_parent = ioend;
475 
476 	atomic_inc(&ioend->io_remaining);
477 	ioend->io_offset += split_ioend->io_size;
478 	ioend->io_size -= split_ioend->io_size;
479 
480 	split_ioend->io_sector = ioend->io_sector;
481 	if (!is_append)
482 		ioend->io_sector += (split_ioend->io_size >> SECTOR_SHIFT);
483 	return split_ioend;
484 }
485 EXPORT_SYMBOL_GPL(iomap_split_ioend);
486 
iomap_ioend_init(void)487 static int __init iomap_ioend_init(void)
488 {
489 	return bioset_init(&iomap_ioend_bioset, 4 * (PAGE_SIZE / SECTOR_SIZE),
490 			   offsetof(struct iomap_ioend, io_bio),
491 			   BIOSET_NEED_BVECS);
492 }
493 fs_initcall(iomap_ioend_init);
494