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