xref: /openbmc/linux/block/blk-merge.c (revision 7fc38225363dd8f19e667ad7c77b63bc4a5c065d)
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/scatterlist.h>
10 
11 #include <trace/events/block.h>
12 
13 #include "blk.h"
14 
15 /*
16  * Check if the two bvecs from two bios can be merged to one segment.  If yes,
17  * no need to check gap between the two bios since the 1st bio and the 1st bvec
18  * in the 2nd bio can be handled in one segment.
19  */
20 static inline bool bios_segs_mergeable(struct request_queue *q,
21 		struct bio *prev, struct bio_vec *prev_last_bv,
22 		struct bio_vec *next_first_bv)
23 {
24 	if (!biovec_phys_mergeable(q, prev_last_bv, next_first_bv))
25 		return false;
26 	if (prev->bi_seg_back_size + next_first_bv->bv_len >
27 			queue_max_segment_size(q))
28 		return false;
29 	return true;
30 }
31 
32 static inline bool bio_will_gap(struct request_queue *q,
33 		struct request *prev_rq, struct bio *prev, struct bio *next)
34 {
35 	struct bio_vec pb, nb;
36 
37 	if (!bio_has_data(prev) || !queue_virt_boundary(q))
38 		return false;
39 
40 	/*
41 	 * Don't merge if the 1st bio starts with non-zero offset, otherwise it
42 	 * is quite difficult to respect the sg gap limit.  We work hard to
43 	 * merge a huge number of small single bios in case of mkfs.
44 	 */
45 	if (prev_rq)
46 		bio_get_first_bvec(prev_rq->bio, &pb);
47 	else
48 		bio_get_first_bvec(prev, &pb);
49 	if (pb.bv_offset & queue_virt_boundary(q))
50 		return true;
51 
52 	/*
53 	 * We don't need to worry about the situation that the merged segment
54 	 * ends in unaligned virt boundary:
55 	 *
56 	 * - if 'pb' ends aligned, the merged segment ends aligned
57 	 * - if 'pb' ends unaligned, the next bio must include
58 	 *   one single bvec of 'nb', otherwise the 'nb' can't
59 	 *   merge with 'pb'
60 	 */
61 	bio_get_last_bvec(prev, &pb);
62 	bio_get_first_bvec(next, &nb);
63 	if (bios_segs_mergeable(q, prev, &pb, &nb))
64 		return false;
65 	return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
66 }
67 
68 static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
69 {
70 	return bio_will_gap(req->q, req, req->biotail, bio);
71 }
72 
73 static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
74 {
75 	return bio_will_gap(req->q, NULL, bio, req->bio);
76 }
77 
78 static struct bio *blk_bio_discard_split(struct request_queue *q,
79 					 struct bio *bio,
80 					 struct bio_set *bs,
81 					 unsigned *nsegs)
82 {
83 	unsigned int max_discard_sectors, granularity;
84 	int alignment;
85 	sector_t tmp;
86 	unsigned split_sectors;
87 
88 	*nsegs = 1;
89 
90 	/* Zero-sector (unknown) and one-sector granularities are the same.  */
91 	granularity = max(q->limits.discard_granularity >> 9, 1U);
92 
93 	max_discard_sectors = min(q->limits.max_discard_sectors,
94 			bio_allowed_max_sectors(q));
95 	max_discard_sectors -= max_discard_sectors % granularity;
96 
97 	if (unlikely(!max_discard_sectors)) {
98 		/* XXX: warn */
99 		return NULL;
100 	}
101 
102 	if (bio_sectors(bio) <= max_discard_sectors)
103 		return NULL;
104 
105 	split_sectors = max_discard_sectors;
106 
107 	/*
108 	 * If the next starting sector would be misaligned, stop the discard at
109 	 * the previous aligned sector.
110 	 */
111 	alignment = (q->limits.discard_alignment >> 9) % granularity;
112 
113 	tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
114 	tmp = sector_div(tmp, granularity);
115 
116 	if (split_sectors > tmp)
117 		split_sectors -= tmp;
118 
119 	return bio_split(bio, split_sectors, GFP_NOIO, bs);
120 }
121 
122 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
123 		struct bio *bio, struct bio_set *bs, unsigned *nsegs)
124 {
125 	*nsegs = 1;
126 
127 	if (!q->limits.max_write_zeroes_sectors)
128 		return NULL;
129 
130 	if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
131 		return NULL;
132 
133 	return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
134 }
135 
136 static struct bio *blk_bio_write_same_split(struct request_queue *q,
137 					    struct bio *bio,
138 					    struct bio_set *bs,
139 					    unsigned *nsegs)
140 {
141 	*nsegs = 1;
142 
143 	if (!q->limits.max_write_same_sectors)
144 		return NULL;
145 
146 	if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
147 		return NULL;
148 
149 	return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
150 }
151 
152 static inline unsigned get_max_io_size(struct request_queue *q,
153 				       struct bio *bio)
154 {
155 	unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
156 	unsigned mask = queue_logical_block_size(q) - 1;
157 
158 	/* aligned to logical block size */
159 	sectors &= ~(mask >> 9);
160 
161 	return sectors;
162 }
163 
164 static struct bio *blk_bio_segment_split(struct request_queue *q,
165 					 struct bio *bio,
166 					 struct bio_set *bs,
167 					 unsigned *segs)
168 {
169 	struct bio_vec bv, bvprv, *bvprvp = NULL;
170 	struct bvec_iter iter;
171 	unsigned seg_size = 0, nsegs = 0, sectors = 0;
172 	unsigned front_seg_size = bio->bi_seg_front_size;
173 	bool do_split = true;
174 	struct bio *new = NULL;
175 	const unsigned max_sectors = get_max_io_size(q, bio);
176 
177 	bio_for_each_segment(bv, bio, iter) {
178 		/*
179 		 * If the queue doesn't support SG gaps and adding this
180 		 * offset would create a gap, disallow it.
181 		 */
182 		if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
183 			goto split;
184 
185 		if (sectors + (bv.bv_len >> 9) > max_sectors) {
186 			/*
187 			 * Consider this a new segment if we're splitting in
188 			 * the middle of this vector.
189 			 */
190 			if (nsegs < queue_max_segments(q) &&
191 			    sectors < max_sectors) {
192 				nsegs++;
193 				sectors = max_sectors;
194 			}
195 			goto split;
196 		}
197 
198 		if (bvprvp) {
199 			if (seg_size + bv.bv_len > queue_max_segment_size(q))
200 				goto new_segment;
201 			if (!biovec_phys_mergeable(q, bvprvp, &bv))
202 				goto new_segment;
203 
204 			seg_size += bv.bv_len;
205 			bvprv = bv;
206 			bvprvp = &bvprv;
207 			sectors += bv.bv_len >> 9;
208 
209 			continue;
210 		}
211 new_segment:
212 		if (nsegs == queue_max_segments(q))
213 			goto split;
214 
215 		if (nsegs == 1 && seg_size > front_seg_size)
216 			front_seg_size = seg_size;
217 
218 		nsegs++;
219 		bvprv = bv;
220 		bvprvp = &bvprv;
221 		seg_size = bv.bv_len;
222 		sectors += bv.bv_len >> 9;
223 
224 	}
225 
226 	do_split = false;
227 split:
228 	*segs = nsegs;
229 
230 	if (do_split) {
231 		new = bio_split(bio, sectors, GFP_NOIO, bs);
232 		if (new)
233 			bio = new;
234 	}
235 
236 	if (nsegs == 1 && seg_size > front_seg_size)
237 		front_seg_size = seg_size;
238 	bio->bi_seg_front_size = front_seg_size;
239 	if (seg_size > bio->bi_seg_back_size)
240 		bio->bi_seg_back_size = seg_size;
241 
242 	return do_split ? new : NULL;
243 }
244 
245 void blk_queue_split(struct request_queue *q, struct bio **bio)
246 {
247 	struct bio *split, *res;
248 	unsigned nsegs;
249 
250 	switch (bio_op(*bio)) {
251 	case REQ_OP_DISCARD:
252 	case REQ_OP_SECURE_ERASE:
253 		split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
254 		break;
255 	case REQ_OP_WRITE_ZEROES:
256 		split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
257 		break;
258 	case REQ_OP_WRITE_SAME:
259 		split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
260 		break;
261 	default:
262 		split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
263 		break;
264 	}
265 
266 	/* physical segments can be figured out during splitting */
267 	res = split ? split : *bio;
268 	res->bi_phys_segments = nsegs;
269 	bio_set_flag(res, BIO_SEG_VALID);
270 
271 	if (split) {
272 		/* there isn't chance to merge the splitted bio */
273 		split->bi_opf |= REQ_NOMERGE;
274 
275 		bio_chain(split, *bio);
276 		trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
277 		generic_make_request(*bio);
278 		*bio = split;
279 	}
280 }
281 EXPORT_SYMBOL(blk_queue_split);
282 
283 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
284 					     struct bio *bio,
285 					     bool no_sg_merge)
286 {
287 	struct bio_vec bv, bvprv = { NULL };
288 	int prev = 0;
289 	unsigned int seg_size, nr_phys_segs;
290 	struct bio *fbio, *bbio;
291 	struct bvec_iter iter;
292 
293 	if (!bio)
294 		return 0;
295 
296 	switch (bio_op(bio)) {
297 	case REQ_OP_DISCARD:
298 	case REQ_OP_SECURE_ERASE:
299 	case REQ_OP_WRITE_ZEROES:
300 		return 0;
301 	case REQ_OP_WRITE_SAME:
302 		return 1;
303 	}
304 
305 	fbio = bio;
306 	seg_size = 0;
307 	nr_phys_segs = 0;
308 	for_each_bio(bio) {
309 		bio_for_each_segment(bv, bio, iter) {
310 			/*
311 			 * If SG merging is disabled, each bio vector is
312 			 * a segment
313 			 */
314 			if (no_sg_merge)
315 				goto new_segment;
316 
317 			if (prev) {
318 				if (seg_size + bv.bv_len
319 				    > queue_max_segment_size(q))
320 					goto new_segment;
321 				if (!biovec_phys_mergeable(q, &bvprv, &bv))
322 					goto new_segment;
323 
324 				seg_size += bv.bv_len;
325 				bvprv = bv;
326 				continue;
327 			}
328 new_segment:
329 			if (nr_phys_segs == 1 && seg_size >
330 			    fbio->bi_seg_front_size)
331 				fbio->bi_seg_front_size = seg_size;
332 
333 			nr_phys_segs++;
334 			bvprv = bv;
335 			prev = 1;
336 			seg_size = bv.bv_len;
337 		}
338 		bbio = bio;
339 	}
340 
341 	if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
342 		fbio->bi_seg_front_size = seg_size;
343 	if (seg_size > bbio->bi_seg_back_size)
344 		bbio->bi_seg_back_size = seg_size;
345 
346 	return nr_phys_segs;
347 }
348 
349 void blk_recalc_rq_segments(struct request *rq)
350 {
351 	bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
352 			&rq->q->queue_flags);
353 
354 	rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
355 			no_sg_merge);
356 }
357 
358 void blk_recount_segments(struct request_queue *q, struct bio *bio)
359 {
360 	unsigned short seg_cnt;
361 
362 	/* estimate segment number by bi_vcnt for non-cloned bio */
363 	if (bio_flagged(bio, BIO_CLONED))
364 		seg_cnt = bio_segments(bio);
365 	else
366 		seg_cnt = bio->bi_vcnt;
367 
368 	if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
369 			(seg_cnt < queue_max_segments(q)))
370 		bio->bi_phys_segments = seg_cnt;
371 	else {
372 		struct bio *nxt = bio->bi_next;
373 
374 		bio->bi_next = NULL;
375 		bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
376 		bio->bi_next = nxt;
377 	}
378 
379 	bio_set_flag(bio, BIO_SEG_VALID);
380 }
381 
382 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
383 				   struct bio *nxt)
384 {
385 	struct bio_vec end_bv = { NULL }, nxt_bv;
386 
387 	if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
388 	    queue_max_segment_size(q))
389 		return 0;
390 
391 	if (!bio_has_data(bio))
392 		return 1;
393 
394 	bio_get_last_bvec(bio, &end_bv);
395 	bio_get_first_bvec(nxt, &nxt_bv);
396 
397 	return biovec_phys_mergeable(q, &end_bv, &nxt_bv);
398 }
399 
400 static inline void
401 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
402 		     struct scatterlist *sglist, struct bio_vec *bvprv,
403 		     struct scatterlist **sg, int *nsegs)
404 {
405 
406 	int nbytes = bvec->bv_len;
407 
408 	if (*sg) {
409 		if ((*sg)->length + nbytes > queue_max_segment_size(q))
410 			goto new_segment;
411 		if (!biovec_phys_mergeable(q, bvprv, bvec))
412 			goto new_segment;
413 
414 		(*sg)->length += nbytes;
415 	} else {
416 new_segment:
417 		if (!*sg)
418 			*sg = sglist;
419 		else {
420 			/*
421 			 * If the driver previously mapped a shorter
422 			 * list, we could see a termination bit
423 			 * prematurely unless it fully inits the sg
424 			 * table on each mapping. We KNOW that there
425 			 * must be more entries here or the driver
426 			 * would be buggy, so force clear the
427 			 * termination bit to avoid doing a full
428 			 * sg_init_table() in drivers for each command.
429 			 */
430 			sg_unmark_end(*sg);
431 			*sg = sg_next(*sg);
432 		}
433 
434 		sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
435 		(*nsegs)++;
436 	}
437 	*bvprv = *bvec;
438 }
439 
440 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
441 		struct scatterlist *sglist, struct scatterlist **sg)
442 {
443 	*sg = sglist;
444 	sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
445 	return 1;
446 }
447 
448 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
449 			     struct scatterlist *sglist,
450 			     struct scatterlist **sg)
451 {
452 	struct bio_vec bvec, bvprv = { NULL };
453 	struct bvec_iter iter;
454 	int nsegs = 0;
455 
456 	for_each_bio(bio)
457 		bio_for_each_segment(bvec, bio, iter)
458 			__blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
459 					     &nsegs);
460 
461 	return nsegs;
462 }
463 
464 /*
465  * map a request to scatterlist, return number of sg entries setup. Caller
466  * must make sure sg can hold rq->nr_phys_segments entries
467  */
468 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
469 		  struct scatterlist *sglist)
470 {
471 	struct scatterlist *sg = NULL;
472 	int nsegs = 0;
473 
474 	if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
475 		nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
476 	else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
477 		nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
478 	else if (rq->bio)
479 		nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
480 
481 	if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
482 	    (blk_rq_bytes(rq) & q->dma_pad_mask)) {
483 		unsigned int pad_len =
484 			(q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
485 
486 		sg->length += pad_len;
487 		rq->extra_len += pad_len;
488 	}
489 
490 	if (q->dma_drain_size && q->dma_drain_needed(rq)) {
491 		if (op_is_write(req_op(rq)))
492 			memset(q->dma_drain_buffer, 0, q->dma_drain_size);
493 
494 		sg_unmark_end(sg);
495 		sg = sg_next(sg);
496 		sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
497 			    q->dma_drain_size,
498 			    ((unsigned long)q->dma_drain_buffer) &
499 			    (PAGE_SIZE - 1));
500 		nsegs++;
501 		rq->extra_len += q->dma_drain_size;
502 	}
503 
504 	if (sg)
505 		sg_mark_end(sg);
506 
507 	/*
508 	 * Something must have been wrong if the figured number of
509 	 * segment is bigger than number of req's physical segments
510 	 */
511 	WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
512 
513 	return nsegs;
514 }
515 EXPORT_SYMBOL(blk_rq_map_sg);
516 
517 static inline int ll_new_hw_segment(struct request_queue *q,
518 				    struct request *req,
519 				    struct bio *bio)
520 {
521 	int nr_phys_segs = bio_phys_segments(q, bio);
522 
523 	if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
524 		goto no_merge;
525 
526 	if (blk_integrity_merge_bio(q, req, bio) == false)
527 		goto no_merge;
528 
529 	/*
530 	 * This will form the start of a new hw segment.  Bump both
531 	 * counters.
532 	 */
533 	req->nr_phys_segments += nr_phys_segs;
534 	return 1;
535 
536 no_merge:
537 	req_set_nomerge(q, req);
538 	return 0;
539 }
540 
541 int ll_back_merge_fn(struct request_queue *q, struct request *req,
542 		     struct bio *bio)
543 {
544 	if (req_gap_back_merge(req, bio))
545 		return 0;
546 	if (blk_integrity_rq(req) &&
547 	    integrity_req_gap_back_merge(req, bio))
548 		return 0;
549 	if (blk_rq_sectors(req) + bio_sectors(bio) >
550 	    blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
551 		req_set_nomerge(q, req);
552 		return 0;
553 	}
554 	if (!bio_flagged(req->biotail, BIO_SEG_VALID))
555 		blk_recount_segments(q, req->biotail);
556 	if (!bio_flagged(bio, BIO_SEG_VALID))
557 		blk_recount_segments(q, bio);
558 
559 	return ll_new_hw_segment(q, req, bio);
560 }
561 
562 int ll_front_merge_fn(struct request_queue *q, struct request *req,
563 		      struct bio *bio)
564 {
565 
566 	if (req_gap_front_merge(req, bio))
567 		return 0;
568 	if (blk_integrity_rq(req) &&
569 	    integrity_req_gap_front_merge(req, bio))
570 		return 0;
571 	if (blk_rq_sectors(req) + bio_sectors(bio) >
572 	    blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
573 		req_set_nomerge(q, req);
574 		return 0;
575 	}
576 	if (!bio_flagged(bio, BIO_SEG_VALID))
577 		blk_recount_segments(q, bio);
578 	if (!bio_flagged(req->bio, BIO_SEG_VALID))
579 		blk_recount_segments(q, req->bio);
580 
581 	return ll_new_hw_segment(q, req, bio);
582 }
583 
584 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
585 		struct request *next)
586 {
587 	unsigned short segments = blk_rq_nr_discard_segments(req);
588 
589 	if (segments >= queue_max_discard_segments(q))
590 		goto no_merge;
591 	if (blk_rq_sectors(req) + bio_sectors(next->bio) >
592 	    blk_rq_get_max_sectors(req, blk_rq_pos(req)))
593 		goto no_merge;
594 
595 	req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
596 	return true;
597 no_merge:
598 	req_set_nomerge(q, req);
599 	return false;
600 }
601 
602 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
603 				struct request *next)
604 {
605 	int total_phys_segments;
606 	unsigned int seg_size =
607 		req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
608 
609 	if (req_gap_back_merge(req, next->bio))
610 		return 0;
611 
612 	/*
613 	 * Will it become too large?
614 	 */
615 	if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
616 	    blk_rq_get_max_sectors(req, blk_rq_pos(req)))
617 		return 0;
618 
619 	total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
620 	if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
621 		if (req->nr_phys_segments == 1)
622 			req->bio->bi_seg_front_size = seg_size;
623 		if (next->nr_phys_segments == 1)
624 			next->biotail->bi_seg_back_size = seg_size;
625 		total_phys_segments--;
626 	}
627 
628 	if (total_phys_segments > queue_max_segments(q))
629 		return 0;
630 
631 	if (blk_integrity_merge_rq(q, req, next) == false)
632 		return 0;
633 
634 	/* Merge is OK... */
635 	req->nr_phys_segments = total_phys_segments;
636 	return 1;
637 }
638 
639 /**
640  * blk_rq_set_mixed_merge - mark a request as mixed merge
641  * @rq: request to mark as mixed merge
642  *
643  * Description:
644  *     @rq is about to be mixed merged.  Make sure the attributes
645  *     which can be mixed are set in each bio and mark @rq as mixed
646  *     merged.
647  */
648 void blk_rq_set_mixed_merge(struct request *rq)
649 {
650 	unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
651 	struct bio *bio;
652 
653 	if (rq->rq_flags & RQF_MIXED_MERGE)
654 		return;
655 
656 	/*
657 	 * @rq will no longer represent mixable attributes for all the
658 	 * contained bios.  It will just track those of the first one.
659 	 * Distributes the attributs to each bio.
660 	 */
661 	for (bio = rq->bio; bio; bio = bio->bi_next) {
662 		WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
663 			     (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
664 		bio->bi_opf |= ff;
665 	}
666 	rq->rq_flags |= RQF_MIXED_MERGE;
667 }
668 
669 static void blk_account_io_merge(struct request *req)
670 {
671 	if (blk_do_io_stat(req)) {
672 		struct hd_struct *part;
673 
674 		part_stat_lock();
675 		part = req->part;
676 
677 		part_dec_in_flight(req->q, part, rq_data_dir(req));
678 
679 		hd_struct_put(part);
680 		part_stat_unlock();
681 	}
682 }
683 /*
684  * Two cases of handling DISCARD merge:
685  * If max_discard_segments > 1, the driver takes every bio
686  * as a range and send them to controller together. The ranges
687  * needn't to be contiguous.
688  * Otherwise, the bios/requests will be handled as same as
689  * others which should be contiguous.
690  */
691 static inline bool blk_discard_mergable(struct request *req)
692 {
693 	if (req_op(req) == REQ_OP_DISCARD &&
694 	    queue_max_discard_segments(req->q) > 1)
695 		return true;
696 	return false;
697 }
698 
699 static enum elv_merge blk_try_req_merge(struct request *req,
700 					struct request *next)
701 {
702 	if (blk_discard_mergable(req))
703 		return ELEVATOR_DISCARD_MERGE;
704 	else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
705 		return ELEVATOR_BACK_MERGE;
706 
707 	return ELEVATOR_NO_MERGE;
708 }
709 
710 /*
711  * For non-mq, this has to be called with the request spinlock acquired.
712  * For mq with scheduling, the appropriate queue wide lock should be held.
713  */
714 static struct request *attempt_merge(struct request_queue *q,
715 				     struct request *req, struct request *next)
716 {
717 	if (!rq_mergeable(req) || !rq_mergeable(next))
718 		return NULL;
719 
720 	if (req_op(req) != req_op(next))
721 		return NULL;
722 
723 	if (rq_data_dir(req) != rq_data_dir(next)
724 	    || req->rq_disk != next->rq_disk)
725 		return NULL;
726 
727 	if (req_op(req) == REQ_OP_WRITE_SAME &&
728 	    !blk_write_same_mergeable(req->bio, next->bio))
729 		return NULL;
730 
731 	/*
732 	 * Don't allow merge of different write hints, or for a hint with
733 	 * non-hint IO.
734 	 */
735 	if (req->write_hint != next->write_hint)
736 		return NULL;
737 
738 	if (req->ioprio != next->ioprio)
739 		return NULL;
740 
741 	/*
742 	 * If we are allowed to merge, then append bio list
743 	 * from next to rq and release next. merge_requests_fn
744 	 * will have updated segment counts, update sector
745 	 * counts here. Handle DISCARDs separately, as they
746 	 * have separate settings.
747 	 */
748 
749 	switch (blk_try_req_merge(req, next)) {
750 	case ELEVATOR_DISCARD_MERGE:
751 		if (!req_attempt_discard_merge(q, req, next))
752 			return NULL;
753 		break;
754 	case ELEVATOR_BACK_MERGE:
755 		if (!ll_merge_requests_fn(q, req, next))
756 			return NULL;
757 		break;
758 	default:
759 		return NULL;
760 	}
761 
762 	/*
763 	 * If failfast settings disagree or any of the two is already
764 	 * a mixed merge, mark both as mixed before proceeding.  This
765 	 * makes sure that all involved bios have mixable attributes
766 	 * set properly.
767 	 */
768 	if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
769 	    (req->cmd_flags & REQ_FAILFAST_MASK) !=
770 	    (next->cmd_flags & REQ_FAILFAST_MASK)) {
771 		blk_rq_set_mixed_merge(req);
772 		blk_rq_set_mixed_merge(next);
773 	}
774 
775 	/*
776 	 * At this point we have either done a back merge or front merge. We
777 	 * need the smaller start_time_ns of the merged requests to be the
778 	 * current request for accounting purposes.
779 	 */
780 	if (next->start_time_ns < req->start_time_ns)
781 		req->start_time_ns = next->start_time_ns;
782 
783 	req->biotail->bi_next = next->bio;
784 	req->biotail = next->biotail;
785 
786 	req->__data_len += blk_rq_bytes(next);
787 
788 	if (!blk_discard_mergable(req))
789 		elv_merge_requests(q, req, next);
790 
791 	/*
792 	 * 'next' is going away, so update stats accordingly
793 	 */
794 	blk_account_io_merge(next);
795 
796 	/*
797 	 * ownership of bio passed from next to req, return 'next' for
798 	 * the caller to free
799 	 */
800 	next->bio = NULL;
801 	return next;
802 }
803 
804 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
805 {
806 	struct request *next = elv_latter_request(q, rq);
807 
808 	if (next)
809 		return attempt_merge(q, rq, next);
810 
811 	return NULL;
812 }
813 
814 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
815 {
816 	struct request *prev = elv_former_request(q, rq);
817 
818 	if (prev)
819 		return attempt_merge(q, prev, rq);
820 
821 	return NULL;
822 }
823 
824 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
825 			  struct request *next)
826 {
827 	struct request *free;
828 
829 	free = attempt_merge(q, rq, next);
830 	if (free) {
831 		blk_put_request(free);
832 		return 1;
833 	}
834 
835 	return 0;
836 }
837 
838 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
839 {
840 	if (!rq_mergeable(rq) || !bio_mergeable(bio))
841 		return false;
842 
843 	if (req_op(rq) != bio_op(bio))
844 		return false;
845 
846 	/* different data direction or already started, don't merge */
847 	if (bio_data_dir(bio) != rq_data_dir(rq))
848 		return false;
849 
850 	/* must be same device */
851 	if (rq->rq_disk != bio->bi_disk)
852 		return false;
853 
854 	/* only merge integrity protected bio into ditto rq */
855 	if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
856 		return false;
857 
858 	/* must be using the same buffer */
859 	if (req_op(rq) == REQ_OP_WRITE_SAME &&
860 	    !blk_write_same_mergeable(rq->bio, bio))
861 		return false;
862 
863 	/*
864 	 * Don't allow merge of different write hints, or for a hint with
865 	 * non-hint IO.
866 	 */
867 	if (rq->write_hint != bio->bi_write_hint)
868 		return false;
869 
870 	if (rq->ioprio != bio_prio(bio))
871 		return false;
872 
873 	return true;
874 }
875 
876 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
877 {
878 	if (blk_discard_mergable(rq))
879 		return ELEVATOR_DISCARD_MERGE;
880 	else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
881 		return ELEVATOR_BACK_MERGE;
882 	else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
883 		return ELEVATOR_FRONT_MERGE;
884 	return ELEVATOR_NO_MERGE;
885 }
886