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