1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Functions related to segment and merge handling
4 */
5 #include <linux/kernel.h>
6 #include <linux/module.h>
7 #include <linux/bio.h>
8 #include <linux/blkdev.h>
9 #include <linux/blk-integrity.h>
10 #include <linux/scatterlist.h>
11 #include <linux/part_stat.h>
12 #include <linux/blk-cgroup.h>
13
14 #include <trace/events/block.h>
15
16 #include "blk.h"
17 #include "blk-mq-sched.h"
18 #include "blk-rq-qos.h"
19 #include "blk-throttle.h"
20
21 static inline void bio_get_first_bvec(struct bio *bio, struct bio_vec *bv)
22 {
23 *bv = mp_bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
24 }
25
26 static inline void bio_get_last_bvec(struct bio *bio, struct bio_vec *bv)
27 {
28 struct bvec_iter iter = bio->bi_iter;
29 int idx;
30
31 bio_get_first_bvec(bio, bv);
32 if (bv->bv_len == bio->bi_iter.bi_size)
33 return; /* this bio only has a single bvec */
34
35 bio_advance_iter(bio, &iter, iter.bi_size);
36
37 if (!iter.bi_bvec_done)
38 idx = iter.bi_idx - 1;
39 else /* in the middle of bvec */
40 idx = iter.bi_idx;
41
42 *bv = bio->bi_io_vec[idx];
43
44 /*
45 * iter.bi_bvec_done records actual length of the last bvec
46 * if this bio ends in the middle of one io vector
47 */
48 if (iter.bi_bvec_done)
49 bv->bv_len = iter.bi_bvec_done;
50 }
51
52 static inline bool bio_will_gap(struct request_queue *q,
53 struct request *prev_rq, struct bio *prev, struct bio *next)
54 {
55 struct bio_vec pb, nb;
56
57 if (!bio_has_data(prev) || !queue_virt_boundary(q))
58 return false;
59
60 /*
61 * Don't merge if the 1st bio starts with non-zero offset, otherwise it
62 * is quite difficult to respect the sg gap limit. We work hard to
63 * merge a huge number of small single bios in case of mkfs.
64 */
65 if (prev_rq)
66 bio_get_first_bvec(prev_rq->bio, &pb);
67 else
68 bio_get_first_bvec(prev, &pb);
69 if (pb.bv_offset & queue_virt_boundary(q))
70 return true;
71
72 /*
73 * We don't need to worry about the situation that the merged segment
74 * ends in unaligned virt boundary:
75 *
76 * - if 'pb' ends aligned, the merged segment ends aligned
77 * - if 'pb' ends unaligned, the next bio must include
78 * one single bvec of 'nb', otherwise the 'nb' can't
79 * merge with 'pb'
80 */
81 bio_get_last_bvec(prev, &pb);
82 bio_get_first_bvec(next, &nb);
83 if (biovec_phys_mergeable(q, &pb, &nb))
84 return false;
85 return __bvec_gap_to_prev(&q->limits, &pb, nb.bv_offset);
86 }
87
88 static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
89 {
90 return bio_will_gap(req->q, req, req->biotail, bio);
91 }
92
93 static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
94 {
95 return bio_will_gap(req->q, NULL, bio, req->bio);
96 }
97
98 /*
99 * The max size one bio can handle is UINT_MAX becasue bvec_iter.bi_size
100 * is defined as 'unsigned int', meantime it has to be aligned to with the
101 * logical block size, which is the minimum accepted unit by hardware.
102 */
103 static unsigned int bio_allowed_max_sectors(const struct queue_limits *lim)
104 {
105 return round_down(UINT_MAX, lim->logical_block_size) >> SECTOR_SHIFT;
106 }
107
108 static struct bio *bio_split_discard(struct bio *bio,
109 const struct queue_limits *lim,
110 unsigned *nsegs, struct bio_set *bs)
111 {
112 unsigned int max_discard_sectors, granularity;
113 sector_t tmp;
114 unsigned split_sectors;
115
116 *nsegs = 1;
117
118 granularity = max(lim->discard_granularity >> 9, 1U);
119
120 max_discard_sectors =
121 min(lim->max_discard_sectors, bio_allowed_max_sectors(lim));
122 max_discard_sectors -= max_discard_sectors % granularity;
123 if (unlikely(!max_discard_sectors))
124 return NULL;
125
126 if (bio_sectors(bio) <= max_discard_sectors)
127 return NULL;
128
129 split_sectors = max_discard_sectors;
130
131 /*
132 * If the next starting sector would be misaligned, stop the discard at
133 * the previous aligned sector.
134 */
135 tmp = bio->bi_iter.bi_sector + split_sectors -
136 ((lim->discard_alignment >> 9) % granularity);
137 tmp = sector_div(tmp, granularity);
138
139 if (split_sectors > tmp)
140 split_sectors -= tmp;
141
142 return bio_split(bio, split_sectors, GFP_NOIO, bs);
143 }
144
145 static struct bio *bio_split_write_zeroes(struct bio *bio,
146 const struct queue_limits *lim,
147 unsigned *nsegs, struct bio_set *bs)
148 {
149 *nsegs = 0;
150 if (!lim->max_write_zeroes_sectors)
151 return NULL;
152 if (bio_sectors(bio) <= lim->max_write_zeroes_sectors)
153 return NULL;
154 return bio_split(bio, lim->max_write_zeroes_sectors, GFP_NOIO, bs);
155 }
156
157 static inline unsigned int blk_boundary_sectors(const struct queue_limits *lim,
158 bool is_atomic)
159 {
160 /*
161 * chunk_sectors must be a multiple of atomic_write_boundary_sectors if
162 * both non-zero.
163 */
164 if (is_atomic && lim->atomic_write_boundary_sectors)
165 return lim->atomic_write_boundary_sectors;
166
167 return lim->chunk_sectors;
168 }
169
170 /*
171 * Return the maximum number of sectors from the start of a bio that may be
172 * submitted as a single request to a block device. If enough sectors remain,
173 * align the end to the physical block size. Otherwise align the end to the
174 * logical block size. This approach minimizes the number of non-aligned
175 * requests that are submitted to a block device if the start of a bio is not
176 * aligned to a physical block boundary.
177 */
178 static inline unsigned get_max_io_size(struct bio *bio,
179 const struct queue_limits *lim)
180 {
181 unsigned pbs = lim->physical_block_size >> SECTOR_SHIFT;
182 unsigned lbs = lim->logical_block_size >> SECTOR_SHIFT;
183 bool is_atomic = bio->bi_opf & REQ_ATOMIC;
184 unsigned boundary_sectors = blk_boundary_sectors(lim, is_atomic);
185 unsigned max_sectors, start, end;
186
187 /*
188 * We ignore lim->max_sectors for atomic writes because it may less
189 * than the actual bio size, which we cannot tolerate.
190 */
191 if (is_atomic)
192 max_sectors = lim->atomic_write_max_sectors;
193 else
194 max_sectors = lim->max_sectors;
195
196 if (boundary_sectors) {
197 max_sectors = min(max_sectors,
198 blk_boundary_sectors_left(bio->bi_iter.bi_sector,
199 boundary_sectors));
200 }
201
202 start = bio->bi_iter.bi_sector & (pbs - 1);
203 end = (start + max_sectors) & ~(pbs - 1);
204 if (end > start)
205 return end - start;
206 return max_sectors & ~(lbs - 1);
207 }
208
209 /**
210 * get_max_segment_size() - maximum number of bytes to add as a single segment
211 * @lim: Request queue limits.
212 * @paddr: address of the range to add
213 * @len: maximum length available to add at @paddr
214 *
215 * Returns the maximum number of bytes of the range starting at @paddr that can
216 * be added to a single segment.
217 */
218 static inline unsigned get_max_segment_size(const struct queue_limits *lim,
219 phys_addr_t paddr, unsigned int len)
220 {
221 /*
222 * Prevent an overflow if mask = ULONG_MAX and offset = 0 by adding 1
223 * after having calculated the minimum.
224 */
225 return min_t(unsigned long, len,
226 min(lim->seg_boundary_mask - (lim->seg_boundary_mask & paddr),
227 (unsigned long)lim->max_segment_size - 1) + 1);
228 }
229
230 /**
231 * bvec_split_segs - verify whether or not a bvec should be split in the middle
232 * @lim: [in] queue limits to split based on
233 * @bv: [in] bvec to examine
234 * @nsegs: [in,out] Number of segments in the bio being built. Incremented
235 * by the number of segments from @bv that may be appended to that
236 * bio without exceeding @max_segs
237 * @bytes: [in,out] Number of bytes in the bio being built. Incremented
238 * by the number of bytes from @bv that may be appended to that
239 * bio without exceeding @max_bytes
240 * @max_segs: [in] upper bound for *@nsegs
241 * @max_bytes: [in] upper bound for *@bytes
242 *
243 * When splitting a bio, it can happen that a bvec is encountered that is too
244 * big to fit in a single segment and hence that it has to be split in the
245 * middle. This function verifies whether or not that should happen. The value
246 * %true is returned if and only if appending the entire @bv to a bio with
247 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
248 * the block driver.
249 */
250 static bool bvec_split_segs(const struct queue_limits *lim,
251 const struct bio_vec *bv, unsigned *nsegs, unsigned *bytes,
252 unsigned max_segs, unsigned max_bytes)
253 {
254 unsigned max_len = min(max_bytes, UINT_MAX) - *bytes;
255 unsigned len = min(bv->bv_len, max_len);
256 unsigned total_len = 0;
257 unsigned seg_size = 0;
258
259 while (len && *nsegs < max_segs) {
260 seg_size = get_max_segment_size(lim, bvec_phys(bv) + total_len, len);
261
262 (*nsegs)++;
263 total_len += seg_size;
264 len -= seg_size;
265
266 if ((bv->bv_offset + total_len) & lim->virt_boundary_mask)
267 break;
268 }
269
270 *bytes += total_len;
271
272 /* tell the caller to split the bvec if it is too big to fit */
273 return len > 0 || bv->bv_len > max_len;
274 }
275
276 /**
277 * bio_split_rw - split a bio in two bios
278 * @bio: [in] bio to be split
279 * @lim: [in] queue limits to split based on
280 * @segs: [out] number of segments in the bio with the first half of the sectors
281 * @bs: [in] bio set to allocate the clone from
282 * @max_bytes: [in] maximum number of bytes per bio
283 *
284 * Clone @bio, update the bi_iter of the clone to represent the first sectors
285 * of @bio and update @bio->bi_iter to represent the remaining sectors. The
286 * following is guaranteed for the cloned bio:
287 * - That it has at most @max_bytes worth of data
288 * - That it has at most queue_max_segments(@q) segments.
289 *
290 * Except for discard requests the cloned bio will point at the bi_io_vec of
291 * the original bio. It is the responsibility of the caller to ensure that the
292 * original bio is not freed before the cloned bio. The caller is also
293 * responsible for ensuring that @bs is only destroyed after processing of the
294 * split bio has finished.
295 */
296 struct bio *bio_split_rw(struct bio *bio, const struct queue_limits *lim,
297 unsigned *segs, struct bio_set *bs, unsigned max_bytes)
298 {
299 struct bio_vec bv, bvprv, *bvprvp = NULL;
300 struct bvec_iter iter;
301 unsigned nsegs = 0, bytes = 0;
302
303 bio_for_each_bvec(bv, bio, iter) {
304 /*
305 * If the queue doesn't support SG gaps and adding this
306 * offset would create a gap, disallow it.
307 */
308 if (bvprvp && bvec_gap_to_prev(lim, bvprvp, bv.bv_offset))
309 goto split;
310
311 if (nsegs < lim->max_segments &&
312 bytes + bv.bv_len <= max_bytes &&
313 bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
314 nsegs++;
315 bytes += bv.bv_len;
316 } else {
317 if (bvec_split_segs(lim, &bv, &nsegs, &bytes,
318 lim->max_segments, max_bytes))
319 goto split;
320 }
321
322 bvprv = bv;
323 bvprvp = &bvprv;
324 }
325
326 *segs = nsegs;
327 return NULL;
328 split:
329 if (bio->bi_opf & REQ_ATOMIC) {
330 bio->bi_status = BLK_STS_INVAL;
331 bio_endio(bio);
332 return ERR_PTR(-EINVAL);
333 }
334 /*
335 * We can't sanely support splitting for a REQ_NOWAIT bio. End it
336 * with EAGAIN if splitting is required and return an error pointer.
337 */
338 if (bio->bi_opf & REQ_NOWAIT) {
339 bio->bi_status = BLK_STS_AGAIN;
340 bio_endio(bio);
341 return ERR_PTR(-EAGAIN);
342 }
343
344 *segs = nsegs;
345
346 /*
347 * Individual bvecs might not be logical block aligned. Round down the
348 * split size so that each bio is properly block size aligned, even if
349 * we do not use the full hardware limits.
350 */
351 bytes = ALIGN_DOWN(bytes, lim->logical_block_size);
352
353 /*
354 * Bio splitting may cause subtle trouble such as hang when doing sync
355 * iopoll in direct IO routine. Given performance gain of iopoll for
356 * big IO can be trival, disable iopoll when split needed.
357 */
358 bio_clear_polled(bio);
359 return bio_split(bio, bytes >> SECTOR_SHIFT, GFP_NOIO, bs);
360 }
361 EXPORT_SYMBOL_GPL(bio_split_rw);
362
363 /**
364 * __bio_split_to_limits - split a bio to fit the queue limits
365 * @bio: bio to be split
366 * @lim: queue limits to split based on
367 * @nr_segs: returns the number of segments in the returned bio
368 *
369 * Check if @bio needs splitting based on the queue limits, and if so split off
370 * a bio fitting the limits from the beginning of @bio and return it. @bio is
371 * shortened to the remainder and re-submitted.
372 *
373 * The split bio is allocated from @q->bio_split, which is provided by the
374 * block layer.
375 */
376 struct bio *__bio_split_to_limits(struct bio *bio,
377 const struct queue_limits *lim,
378 unsigned int *nr_segs)
379 {
380 struct bio_set *bs = &bio->bi_bdev->bd_disk->bio_split;
381 struct bio *split;
382
383 switch (bio_op(bio)) {
384 case REQ_OP_DISCARD:
385 case REQ_OP_SECURE_ERASE:
386 split = bio_split_discard(bio, lim, nr_segs, bs);
387 break;
388 case REQ_OP_WRITE_ZEROES:
389 split = bio_split_write_zeroes(bio, lim, nr_segs, bs);
390 break;
391 default:
392 split = bio_split_rw(bio, lim, nr_segs, bs,
393 get_max_io_size(bio, lim) << SECTOR_SHIFT);
394 if (IS_ERR(split))
395 return NULL;
396 break;
397 }
398
399 if (split) {
400 /* there isn't chance to merge the split bio */
401 split->bi_opf |= REQ_NOMERGE;
402
403 blkcg_bio_issue_init(split);
404 bio_chain(split, bio);
405 trace_block_split(split, bio->bi_iter.bi_sector);
406 WARN_ON_ONCE(bio_zone_write_plugging(bio));
407 submit_bio_noacct(bio);
408 return split;
409 }
410 return bio;
411 }
412
413 /**
414 * bio_split_to_limits - split a bio to fit the queue limits
415 * @bio: bio to be split
416 *
417 * Check if @bio needs splitting based on the queue limits of @bio->bi_bdev, and
418 * if so split off a bio fitting the limits from the beginning of @bio and
419 * return it. @bio is shortened to the remainder and re-submitted.
420 *
421 * The split bio is allocated from @q->bio_split, which is provided by the
422 * block layer.
423 */
424 struct bio *bio_split_to_limits(struct bio *bio)
425 {
426 const struct queue_limits *lim = &bdev_get_queue(bio->bi_bdev)->limits;
427 unsigned int nr_segs;
428
429 if (bio_may_exceed_limits(bio, lim))
430 return __bio_split_to_limits(bio, lim, &nr_segs);
431 return bio;
432 }
433 EXPORT_SYMBOL(bio_split_to_limits);
434
435 unsigned int blk_recalc_rq_segments(struct request *rq)
436 {
437 unsigned int nr_phys_segs = 0;
438 unsigned int bytes = 0;
439 struct req_iterator iter;
440 struct bio_vec bv;
441
442 if (!rq->bio)
443 return 0;
444
445 switch (bio_op(rq->bio)) {
446 case REQ_OP_DISCARD:
447 case REQ_OP_SECURE_ERASE:
448 if (queue_max_discard_segments(rq->q) > 1) {
449 struct bio *bio = rq->bio;
450
451 for_each_bio(bio)
452 nr_phys_segs++;
453 return nr_phys_segs;
454 }
455 return 1;
456 case REQ_OP_WRITE_ZEROES:
457 return 0;
458 default:
459 break;
460 }
461
462 rq_for_each_bvec(bv, rq, iter)
463 bvec_split_segs(&rq->q->limits, &bv, &nr_phys_segs, &bytes,
464 UINT_MAX, UINT_MAX);
465 return nr_phys_segs;
466 }
467
468 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
469 struct scatterlist *sglist)
470 {
471 if (!*sg)
472 return sglist;
473
474 /*
475 * If the driver previously mapped a shorter list, we could see a
476 * termination bit prematurely unless it fully inits the sg table
477 * on each mapping. We KNOW that there must be more entries here
478 * or the driver would be buggy, so force clear the termination bit
479 * to avoid doing a full sg_init_table() in drivers for each command.
480 */
481 sg_unmark_end(*sg);
482 return sg_next(*sg);
483 }
484
485 static unsigned blk_bvec_map_sg(struct request_queue *q,
486 struct bio_vec *bvec, struct scatterlist *sglist,
487 struct scatterlist **sg)
488 {
489 unsigned nbytes = bvec->bv_len;
490 unsigned nsegs = 0, total = 0;
491
492 while (nbytes > 0) {
493 unsigned offset = bvec->bv_offset + total;
494 unsigned len = get_max_segment_size(&q->limits,
495 bvec_phys(bvec) + total, nbytes);
496 struct page *page = bvec->bv_page;
497
498 /*
499 * Unfortunately a fair number of drivers barf on scatterlists
500 * that have an offset larger than PAGE_SIZE, despite other
501 * subsystems dealing with that invariant just fine. For now
502 * stick to the legacy format where we never present those from
503 * the block layer, but the code below should be removed once
504 * these offenders (mostly MMC/SD drivers) are fixed.
505 */
506 page += (offset >> PAGE_SHIFT);
507 offset &= ~PAGE_MASK;
508
509 *sg = blk_next_sg(sg, sglist);
510 sg_set_page(*sg, page, len, offset);
511
512 total += len;
513 nbytes -= len;
514 nsegs++;
515 }
516
517 return nsegs;
518 }
519
520 static inline int __blk_bvec_map_sg(struct bio_vec bv,
521 struct scatterlist *sglist, struct scatterlist **sg)
522 {
523 *sg = blk_next_sg(sg, sglist);
524 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
525 return 1;
526 }
527
528 /* only try to merge bvecs into one sg if they are from two bios */
529 static inline bool
530 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
531 struct bio_vec *bvprv, struct scatterlist **sg)
532 {
533
534 int nbytes = bvec->bv_len;
535
536 if (!*sg)
537 return false;
538
539 if ((*sg)->length + nbytes > queue_max_segment_size(q))
540 return false;
541
542 if (!biovec_phys_mergeable(q, bvprv, bvec))
543 return false;
544
545 (*sg)->length += nbytes;
546
547 return true;
548 }
549
550 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
551 struct scatterlist *sglist,
552 struct scatterlist **sg)
553 {
554 struct bio_vec bvec, bvprv = { NULL };
555 struct bvec_iter iter;
556 int nsegs = 0;
557 bool new_bio = false;
558
559 for_each_bio(bio) {
560 bio_for_each_bvec(bvec, bio, iter) {
561 /*
562 * Only try to merge bvecs from two bios given we
563 * have done bio internal merge when adding pages
564 * to bio
565 */
566 if (new_bio &&
567 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
568 goto next_bvec;
569
570 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
571 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
572 else
573 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
574 next_bvec:
575 new_bio = false;
576 }
577 if (likely(bio->bi_iter.bi_size)) {
578 bvprv = bvec;
579 new_bio = true;
580 }
581 }
582
583 return nsegs;
584 }
585
586 /*
587 * map a request to scatterlist, return number of sg entries setup. Caller
588 * must make sure sg can hold rq->nr_phys_segments entries
589 */
590 int __blk_rq_map_sg(struct request_queue *q, struct request *rq,
591 struct scatterlist *sglist, struct scatterlist **last_sg)
592 {
593 int nsegs = 0;
594
595 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
596 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg);
597 else if (rq->bio)
598 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg);
599
600 if (*last_sg)
601 sg_mark_end(*last_sg);
602
603 /*
604 * Something must have been wrong if the figured number of
605 * segment is bigger than number of req's physical segments
606 */
607 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
608
609 return nsegs;
610 }
611 EXPORT_SYMBOL(__blk_rq_map_sg);
612
613 static inline unsigned int blk_rq_get_max_sectors(struct request *rq,
614 sector_t offset)
615 {
616 struct request_queue *q = rq->q;
617 struct queue_limits *lim = &q->limits;
618 unsigned int max_sectors, boundary_sectors;
619 bool is_atomic = rq->cmd_flags & REQ_ATOMIC;
620
621 if (blk_rq_is_passthrough(rq))
622 return q->limits.max_hw_sectors;
623
624 boundary_sectors = blk_boundary_sectors(lim, is_atomic);
625 max_sectors = blk_queue_get_max_sectors(rq);
626
627 if (!boundary_sectors ||
628 req_op(rq) == REQ_OP_DISCARD ||
629 req_op(rq) == REQ_OP_SECURE_ERASE)
630 return max_sectors;
631 return min(max_sectors,
632 blk_boundary_sectors_left(offset, boundary_sectors));
633 }
634
635 static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
636 unsigned int nr_phys_segs)
637 {
638 if (!blk_cgroup_mergeable(req, bio))
639 goto no_merge;
640
641 if (blk_integrity_merge_bio(req->q, req, bio) == false)
642 goto no_merge;
643
644 /* discard request merge won't add new segment */
645 if (req_op(req) == REQ_OP_DISCARD)
646 return 1;
647
648 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req))
649 goto no_merge;
650
651 /*
652 * This will form the start of a new hw segment. Bump both
653 * counters.
654 */
655 req->nr_phys_segments += nr_phys_segs;
656 return 1;
657
658 no_merge:
659 req_set_nomerge(req->q, req);
660 return 0;
661 }
662
663 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
664 {
665 if (req_gap_back_merge(req, bio))
666 return 0;
667 if (blk_integrity_rq(req) &&
668 integrity_req_gap_back_merge(req, bio))
669 return 0;
670 if (!bio_crypt_ctx_back_mergeable(req, bio))
671 return 0;
672 if (blk_rq_sectors(req) + bio_sectors(bio) >
673 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
674 req_set_nomerge(req->q, req);
675 return 0;
676 }
677
678 return ll_new_hw_segment(req, bio, nr_segs);
679 }
680
681 static int ll_front_merge_fn(struct request *req, struct bio *bio,
682 unsigned int nr_segs)
683 {
684 if (req_gap_front_merge(req, bio))
685 return 0;
686 if (blk_integrity_rq(req) &&
687 integrity_req_gap_front_merge(req, bio))
688 return 0;
689 if (!bio_crypt_ctx_front_mergeable(req, bio))
690 return 0;
691 if (blk_rq_sectors(req) + bio_sectors(bio) >
692 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
693 req_set_nomerge(req->q, req);
694 return 0;
695 }
696
697 return ll_new_hw_segment(req, bio, nr_segs);
698 }
699
700 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
701 struct request *next)
702 {
703 unsigned short segments = blk_rq_nr_discard_segments(req);
704
705 if (segments >= queue_max_discard_segments(q))
706 goto no_merge;
707 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
708 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
709 goto no_merge;
710
711 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
712 return true;
713 no_merge:
714 req_set_nomerge(q, req);
715 return false;
716 }
717
718 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
719 struct request *next)
720 {
721 int total_phys_segments;
722
723 if (req_gap_back_merge(req, next->bio))
724 return 0;
725
726 /*
727 * Will it become too large?
728 */
729 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
730 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
731 return 0;
732
733 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
734 if (total_phys_segments > blk_rq_get_max_segments(req))
735 return 0;
736
737 if (!blk_cgroup_mergeable(req, next->bio))
738 return 0;
739
740 if (blk_integrity_merge_rq(q, req, next) == false)
741 return 0;
742
743 if (!bio_crypt_ctx_merge_rq(req, next))
744 return 0;
745
746 /* Merge is OK... */
747 req->nr_phys_segments = total_phys_segments;
748 return 1;
749 }
750
751 /**
752 * blk_rq_set_mixed_merge - mark a request as mixed merge
753 * @rq: request to mark as mixed merge
754 *
755 * Description:
756 * @rq is about to be mixed merged. Make sure the attributes
757 * which can be mixed are set in each bio and mark @rq as mixed
758 * merged.
759 */
760 static void blk_rq_set_mixed_merge(struct request *rq)
761 {
762 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK;
763 struct bio *bio;
764
765 if (rq->rq_flags & RQF_MIXED_MERGE)
766 return;
767
768 /*
769 * @rq will no longer represent mixable attributes for all the
770 * contained bios. It will just track those of the first one.
771 * Distributes the attributs to each bio.
772 */
773 for (bio = rq->bio; bio; bio = bio->bi_next) {
774 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
775 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
776 bio->bi_opf |= ff;
777 }
778 rq->rq_flags |= RQF_MIXED_MERGE;
779 }
780
781 static inline blk_opf_t bio_failfast(const struct bio *bio)
782 {
783 if (bio->bi_opf & REQ_RAHEAD)
784 return REQ_FAILFAST_MASK;
785
786 return bio->bi_opf & REQ_FAILFAST_MASK;
787 }
788
789 /*
790 * After we are marked as MIXED_MERGE, any new RA bio has to be updated
791 * as failfast, and request's failfast has to be updated in case of
792 * front merge.
793 */
794 static inline void blk_update_mixed_merge(struct request *req,
795 struct bio *bio, bool front_merge)
796 {
797 if (req->rq_flags & RQF_MIXED_MERGE) {
798 if (bio->bi_opf & REQ_RAHEAD)
799 bio->bi_opf |= REQ_FAILFAST_MASK;
800
801 if (front_merge) {
802 req->cmd_flags &= ~REQ_FAILFAST_MASK;
803 req->cmd_flags |= bio->bi_opf & REQ_FAILFAST_MASK;
804 }
805 }
806 }
807
808 static void blk_account_io_merge_request(struct request *req)
809 {
810 if (blk_do_io_stat(req)) {
811 part_stat_lock();
812 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
813 part_stat_local_dec(req->part,
814 in_flight[op_is_write(req_op(req))]);
815 part_stat_unlock();
816 }
817 }
818
819 static enum elv_merge blk_try_req_merge(struct request *req,
820 struct request *next)
821 {
822 if (blk_discard_mergable(req))
823 return ELEVATOR_DISCARD_MERGE;
824 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
825 return ELEVATOR_BACK_MERGE;
826
827 return ELEVATOR_NO_MERGE;
828 }
829
830 static bool blk_atomic_write_mergeable_rq_bio(struct request *rq,
831 struct bio *bio)
832 {
833 return (rq->cmd_flags & REQ_ATOMIC) == (bio->bi_opf & REQ_ATOMIC);
834 }
835
836 static bool blk_atomic_write_mergeable_rqs(struct request *rq,
837 struct request *next)
838 {
839 return (rq->cmd_flags & REQ_ATOMIC) == (next->cmd_flags & REQ_ATOMIC);
840 }
841
842 /*
843 * For non-mq, this has to be called with the request spinlock acquired.
844 * For mq with scheduling, the appropriate queue wide lock should be held.
845 */
846 static struct request *attempt_merge(struct request_queue *q,
847 struct request *req, struct request *next)
848 {
849 if (!rq_mergeable(req) || !rq_mergeable(next))
850 return NULL;
851
852 if (req_op(req) != req_op(next))
853 return NULL;
854
855 if (rq_data_dir(req) != rq_data_dir(next))
856 return NULL;
857
858 /* Don't merge requests with different write hints. */
859 if (req->write_hint != next->write_hint)
860 return NULL;
861
862 if (req->ioprio != next->ioprio)
863 return NULL;
864
865 if (!blk_atomic_write_mergeable_rqs(req, next))
866 return NULL;
867
868 /*
869 * If we are allowed to merge, then append bio list
870 * from next to rq and release next. merge_requests_fn
871 * will have updated segment counts, update sector
872 * counts here. Handle DISCARDs separately, as they
873 * have separate settings.
874 */
875
876 switch (blk_try_req_merge(req, next)) {
877 case ELEVATOR_DISCARD_MERGE:
878 if (!req_attempt_discard_merge(q, req, next))
879 return NULL;
880 break;
881 case ELEVATOR_BACK_MERGE:
882 if (!ll_merge_requests_fn(q, req, next))
883 return NULL;
884 break;
885 default:
886 return NULL;
887 }
888
889 /*
890 * If failfast settings disagree or any of the two is already
891 * a mixed merge, mark both as mixed before proceeding. This
892 * makes sure that all involved bios have mixable attributes
893 * set properly.
894 */
895 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
896 (req->cmd_flags & REQ_FAILFAST_MASK) !=
897 (next->cmd_flags & REQ_FAILFAST_MASK)) {
898 blk_rq_set_mixed_merge(req);
899 blk_rq_set_mixed_merge(next);
900 }
901
902 /*
903 * At this point we have either done a back merge or front merge. We
904 * need the smaller start_time_ns of the merged requests to be the
905 * current request for accounting purposes.
906 */
907 if (next->start_time_ns < req->start_time_ns)
908 req->start_time_ns = next->start_time_ns;
909
910 req->biotail->bi_next = next->bio;
911 req->biotail = next->biotail;
912
913 req->__data_len += blk_rq_bytes(next);
914
915 if (!blk_discard_mergable(req))
916 elv_merge_requests(q, req, next);
917
918 blk_crypto_rq_put_keyslot(next);
919
920 /*
921 * 'next' is going away, so update stats accordingly
922 */
923 blk_account_io_merge_request(next);
924
925 trace_block_rq_merge(next);
926
927 /*
928 * ownership of bio passed from next to req, return 'next' for
929 * the caller to free
930 */
931 next->bio = NULL;
932 return next;
933 }
934
935 static struct request *attempt_back_merge(struct request_queue *q,
936 struct request *rq)
937 {
938 struct request *next = elv_latter_request(q, rq);
939
940 if (next)
941 return attempt_merge(q, rq, next);
942
943 return NULL;
944 }
945
946 static struct request *attempt_front_merge(struct request_queue *q,
947 struct request *rq)
948 {
949 struct request *prev = elv_former_request(q, rq);
950
951 if (prev)
952 return attempt_merge(q, prev, rq);
953
954 return NULL;
955 }
956
957 /*
958 * Try to merge 'next' into 'rq'. Return true if the merge happened, false
959 * otherwise. The caller is responsible for freeing 'next' if the merge
960 * happened.
961 */
962 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq,
963 struct request *next)
964 {
965 return attempt_merge(q, rq, next);
966 }
967
968 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
969 {
970 if (!rq_mergeable(rq) || !bio_mergeable(bio))
971 return false;
972
973 if (req_op(rq) != bio_op(bio))
974 return false;
975
976 /* different data direction or already started, don't merge */
977 if (bio_data_dir(bio) != rq_data_dir(rq))
978 return false;
979
980 /* don't merge across cgroup boundaries */
981 if (!blk_cgroup_mergeable(rq, bio))
982 return false;
983
984 /* only merge integrity protected bio into ditto rq */
985 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
986 return false;
987
988 /* Only merge if the crypt contexts are compatible */
989 if (!bio_crypt_rq_ctx_compatible(rq, bio))
990 return false;
991
992 /* Don't merge requests with different write hints. */
993 if (rq->write_hint != bio->bi_write_hint)
994 return false;
995
996 if (rq->ioprio != bio_prio(bio))
997 return false;
998
999 if (blk_atomic_write_mergeable_rq_bio(rq, bio) == false)
1000 return false;
1001
1002 return true;
1003 }
1004
1005 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
1006 {
1007 if (blk_discard_mergable(rq))
1008 return ELEVATOR_DISCARD_MERGE;
1009 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
1010 return ELEVATOR_BACK_MERGE;
1011 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
1012 return ELEVATOR_FRONT_MERGE;
1013 return ELEVATOR_NO_MERGE;
1014 }
1015
1016 static void blk_account_io_merge_bio(struct request *req)
1017 {
1018 if (!blk_do_io_stat(req))
1019 return;
1020
1021 part_stat_lock();
1022 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
1023 part_stat_unlock();
1024 }
1025
1026 enum bio_merge_status bio_attempt_back_merge(struct request *req,
1027 struct bio *bio, unsigned int nr_segs)
1028 {
1029 const blk_opf_t ff = bio_failfast(bio);
1030
1031 if (!ll_back_merge_fn(req, bio, nr_segs))
1032 return BIO_MERGE_FAILED;
1033
1034 trace_block_bio_backmerge(bio);
1035 rq_qos_merge(req->q, req, bio);
1036
1037 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1038 blk_rq_set_mixed_merge(req);
1039
1040 blk_update_mixed_merge(req, bio, false);
1041
1042 if (req->rq_flags & RQF_ZONE_WRITE_PLUGGING)
1043 blk_zone_write_plug_bio_merged(bio);
1044
1045 req->biotail->bi_next = bio;
1046 req->biotail = bio;
1047 req->__data_len += bio->bi_iter.bi_size;
1048
1049 bio_crypt_free_ctx(bio);
1050
1051 blk_account_io_merge_bio(req);
1052 return BIO_MERGE_OK;
1053 }
1054
1055 static enum bio_merge_status bio_attempt_front_merge(struct request *req,
1056 struct bio *bio, unsigned int nr_segs)
1057 {
1058 const blk_opf_t ff = bio_failfast(bio);
1059
1060 /*
1061 * A front merge for writes to sequential zones of a zoned block device
1062 * can happen only if the user submitted writes out of order. Do not
1063 * merge such write to let it fail.
1064 */
1065 if (req->rq_flags & RQF_ZONE_WRITE_PLUGGING)
1066 return BIO_MERGE_FAILED;
1067
1068 if (!ll_front_merge_fn(req, bio, nr_segs))
1069 return BIO_MERGE_FAILED;
1070
1071 trace_block_bio_frontmerge(bio);
1072 rq_qos_merge(req->q, req, bio);
1073
1074 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1075 blk_rq_set_mixed_merge(req);
1076
1077 blk_update_mixed_merge(req, bio, true);
1078
1079 bio->bi_next = req->bio;
1080 req->bio = bio;
1081
1082 req->__sector = bio->bi_iter.bi_sector;
1083 req->__data_len += bio->bi_iter.bi_size;
1084
1085 bio_crypt_do_front_merge(req, bio);
1086
1087 blk_account_io_merge_bio(req);
1088 return BIO_MERGE_OK;
1089 }
1090
1091 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q,
1092 struct request *req, struct bio *bio)
1093 {
1094 unsigned short segments = blk_rq_nr_discard_segments(req);
1095
1096 if (segments >= queue_max_discard_segments(q))
1097 goto no_merge;
1098 if (blk_rq_sectors(req) + bio_sectors(bio) >
1099 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
1100 goto no_merge;
1101
1102 rq_qos_merge(q, req, bio);
1103
1104 req->biotail->bi_next = bio;
1105 req->biotail = bio;
1106 req->__data_len += bio->bi_iter.bi_size;
1107 req->nr_phys_segments = segments + 1;
1108
1109 blk_account_io_merge_bio(req);
1110 return BIO_MERGE_OK;
1111 no_merge:
1112 req_set_nomerge(q, req);
1113 return BIO_MERGE_FAILED;
1114 }
1115
1116 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q,
1117 struct request *rq,
1118 struct bio *bio,
1119 unsigned int nr_segs,
1120 bool sched_allow_merge)
1121 {
1122 if (!blk_rq_merge_ok(rq, bio))
1123 return BIO_MERGE_NONE;
1124
1125 switch (blk_try_merge(rq, bio)) {
1126 case ELEVATOR_BACK_MERGE:
1127 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1128 return bio_attempt_back_merge(rq, bio, nr_segs);
1129 break;
1130 case ELEVATOR_FRONT_MERGE:
1131 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1132 return bio_attempt_front_merge(rq, bio, nr_segs);
1133 break;
1134 case ELEVATOR_DISCARD_MERGE:
1135 return bio_attempt_discard_merge(q, rq, bio);
1136 default:
1137 return BIO_MERGE_NONE;
1138 }
1139
1140 return BIO_MERGE_FAILED;
1141 }
1142
1143 /**
1144 * blk_attempt_plug_merge - try to merge with %current's plugged list
1145 * @q: request_queue new bio is being queued at
1146 * @bio: new bio being queued
1147 * @nr_segs: number of segments in @bio
1148 * from the passed in @q already in the plug list
1149 *
1150 * Determine whether @bio being queued on @q can be merged with the previous
1151 * request on %current's plugged list. Returns %true if merge was successful,
1152 * otherwise %false.
1153 *
1154 * Plugging coalesces IOs from the same issuer for the same purpose without
1155 * going through @q->queue_lock. As such it's more of an issuing mechanism
1156 * than scheduling, and the request, while may have elvpriv data, is not
1157 * added on the elevator at this point. In addition, we don't have
1158 * reliable access to the elevator outside queue lock. Only check basic
1159 * merging parameters without querying the elevator.
1160 *
1161 * Caller must ensure !blk_queue_nomerges(q) beforehand.
1162 */
1163 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio,
1164 unsigned int nr_segs)
1165 {
1166 struct blk_plug *plug = current->plug;
1167 struct request *rq;
1168
1169 if (!plug || rq_list_empty(plug->mq_list))
1170 return false;
1171
1172 rq_list_for_each(&plug->mq_list, rq) {
1173 if (rq->q == q) {
1174 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) ==
1175 BIO_MERGE_OK)
1176 return true;
1177 break;
1178 }
1179
1180 /*
1181 * Only keep iterating plug list for merges if we have multiple
1182 * queues
1183 */
1184 if (!plug->multiple_queues)
1185 break;
1186 }
1187 return false;
1188 }
1189
1190 /*
1191 * Iterate list of requests and see if we can merge this bio with any
1192 * of them.
1193 */
1194 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list,
1195 struct bio *bio, unsigned int nr_segs)
1196 {
1197 struct request *rq;
1198 int checked = 8;
1199
1200 list_for_each_entry_reverse(rq, list, queuelist) {
1201 if (!checked--)
1202 break;
1203
1204 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) {
1205 case BIO_MERGE_NONE:
1206 continue;
1207 case BIO_MERGE_OK:
1208 return true;
1209 case BIO_MERGE_FAILED:
1210 return false;
1211 }
1212
1213 }
1214
1215 return false;
1216 }
1217 EXPORT_SYMBOL_GPL(blk_bio_list_merge);
1218
1219 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
1220 unsigned int nr_segs, struct request **merged_request)
1221 {
1222 struct request *rq;
1223
1224 switch (elv_merge(q, &rq, bio)) {
1225 case ELEVATOR_BACK_MERGE:
1226 if (!blk_mq_sched_allow_merge(q, rq, bio))
1227 return false;
1228 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1229 return false;
1230 *merged_request = attempt_back_merge(q, rq);
1231 if (!*merged_request)
1232 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
1233 return true;
1234 case ELEVATOR_FRONT_MERGE:
1235 if (!blk_mq_sched_allow_merge(q, rq, bio))
1236 return false;
1237 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1238 return false;
1239 *merged_request = attempt_front_merge(q, rq);
1240 if (!*merged_request)
1241 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
1242 return true;
1243 case ELEVATOR_DISCARD_MERGE:
1244 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK;
1245 default:
1246 return false;
1247 }
1248 }
1249 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
1250
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.