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