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 #include <linux/blkdev.h>
11 #include <linux/blk-cgroup.h>
12 
13 #include <trace/events/block.h>
14 
15 #include "blk.h"
16 
blk_bio_discard_split(struct request_queue * q,struct bio * bio,struct bio_set * bs,unsigned * nsegs)17 static struct bio *blk_bio_discard_split(struct request_queue *q,
18 					 struct bio *bio,
19 					 struct bio_set *bs,
20 					 unsigned *nsegs)
21 {
22 	unsigned int max_discard_sectors, granularity;
23 	int alignment;
24 	sector_t tmp;
25 	unsigned split_sectors;
26 
27 	*nsegs = 1;
28 
29 	/* Zero-sector (unknown) and one-sector granularities are the same.  */
30 	granularity = max(q->limits.discard_granularity >> 9, 1U);
31 
32 	max_discard_sectors = min(q->limits.max_discard_sectors,
33 			bio_allowed_max_sectors(q));
34 	max_discard_sectors -= max_discard_sectors % granularity;
35 
36 	if (unlikely(!max_discard_sectors)) {
37 		/* XXX: warn */
38 		return NULL;
39 	}
40 
41 	if (bio_sectors(bio) <= max_discard_sectors)
42 		return NULL;
43 
44 	split_sectors = max_discard_sectors;
45 
46 	/*
47 	 * If the next starting sector would be misaligned, stop the discard at
48 	 * the previous aligned sector.
49 	 */
50 	alignment = (q->limits.discard_alignment >> 9) % granularity;
51 
52 	tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
53 	tmp = sector_div(tmp, granularity);
54 
55 	if (split_sectors > tmp)
56 		split_sectors -= tmp;
57 
58 	return bio_split(bio, split_sectors, GFP_NOIO, bs);
59 }
60 
blk_bio_write_zeroes_split(struct request_queue * q,struct bio * bio,struct bio_set * bs,unsigned * nsegs)61 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
62 		struct bio *bio, struct bio_set *bs, unsigned *nsegs)
63 {
64 	*nsegs = 1;
65 
66 	if (!q->limits.max_write_zeroes_sectors)
67 		return NULL;
68 
69 	if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
70 		return NULL;
71 
72 	return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
73 }
74 
blk_bio_write_same_split(struct request_queue * q,struct bio * bio,struct bio_set * bs,unsigned * nsegs)75 static struct bio *blk_bio_write_same_split(struct request_queue *q,
76 					    struct bio *bio,
77 					    struct bio_set *bs,
78 					    unsigned *nsegs)
79 {
80 	*nsegs = 1;
81 
82 	if (!q->limits.max_write_same_sectors)
83 		return NULL;
84 
85 	if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
86 		return NULL;
87 
88 	return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
89 }
90 
get_max_io_size(struct request_queue * q,struct bio * bio)91 static inline unsigned get_max_io_size(struct request_queue *q,
92 				       struct bio *bio)
93 {
94 	unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
95 	unsigned mask = queue_logical_block_size(q) - 1;
96 
97 	/* aligned to logical block size */
98 	sectors &= ~(mask >> 9);
99 
100 	return sectors;
101 }
102 
blk_bio_segment_split(struct request_queue * q,struct bio * bio,struct bio_set * bs,unsigned * segs)103 static struct bio *blk_bio_segment_split(struct request_queue *q,
104 					 struct bio *bio,
105 					 struct bio_set *bs,
106 					 unsigned *segs)
107 {
108 	struct bio_vec bv, bvprv, *bvprvp = NULL;
109 	struct bvec_iter iter;
110 	unsigned seg_size = 0, nsegs = 0, sectors = 0;
111 	unsigned front_seg_size = bio->bi_seg_front_size;
112 	bool do_split = true;
113 	struct bio *new = NULL;
114 	const unsigned max_sectors = get_max_io_size(q, bio);
115 
116 	bio_for_each_segment(bv, bio, iter) {
117 		/*
118 		 * If the queue doesn't support SG gaps and adding this
119 		 * offset would create a gap, disallow it.
120 		 */
121 		if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
122 			goto split;
123 
124 		if (sectors + (bv.bv_len >> 9) > max_sectors) {
125 			/*
126 			 * Consider this a new segment if we're splitting in
127 			 * the middle of this vector.
128 			 */
129 			if (nsegs < queue_max_segments(q) &&
130 			    sectors < max_sectors) {
131 				nsegs++;
132 				sectors = max_sectors;
133 			}
134 			goto split;
135 		}
136 
137 		if (bvprvp && blk_queue_cluster(q)) {
138 			if (seg_size + bv.bv_len > queue_max_segment_size(q))
139 				goto new_segment;
140 			if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
141 				goto new_segment;
142 			if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
143 				goto new_segment;
144 
145 			seg_size += bv.bv_len;
146 			bvprv = bv;
147 			bvprvp = &bvprv;
148 			sectors += bv.bv_len >> 9;
149 
150 			continue;
151 		}
152 new_segment:
153 		if (nsegs == queue_max_segments(q))
154 			goto split;
155 
156 		if (nsegs == 1 && seg_size > front_seg_size)
157 			front_seg_size = seg_size;
158 
159 		nsegs++;
160 		bvprv = bv;
161 		bvprvp = &bvprv;
162 		seg_size = bv.bv_len;
163 		sectors += bv.bv_len >> 9;
164 
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 	if (nsegs == 1 && seg_size > front_seg_size)
178 		front_seg_size = seg_size;
179 	bio->bi_seg_front_size = front_seg_size;
180 	if (seg_size > bio->bi_seg_back_size)
181 		bio->bi_seg_back_size = seg_size;
182 
183 	return do_split ? new : NULL;
184 }
185 
blk_queue_split(struct request_queue * q,struct bio ** bio)186 void blk_queue_split(struct request_queue *q, struct bio **bio)
187 {
188 	struct bio *split, *res;
189 	unsigned nsegs;
190 
191 	switch (bio_op(*bio)) {
192 	case REQ_OP_DISCARD:
193 	case REQ_OP_SECURE_ERASE:
194 		split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
195 		break;
196 	case REQ_OP_WRITE_ZEROES:
197 		split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
198 		break;
199 	case REQ_OP_WRITE_SAME:
200 		split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
201 		break;
202 	default:
203 		split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
204 		break;
205 	}
206 
207 	/* physical segments can be figured out during splitting */
208 	res = split ? split : *bio;
209 	res->bi_phys_segments = nsegs;
210 	bio_set_flag(res, BIO_SEG_VALID);
211 
212 	if (split) {
213 		/* there isn't chance to merge the splitted bio */
214 		split->bi_opf |= REQ_NOMERGE;
215 
216 		/*
217 		 * Since we're recursing into make_request here, ensure
218 		 * that we mark this bio as already having entered the queue.
219 		 * If not, and the queue is going away, we can get stuck
220 		 * forever on waiting for the queue reference to drop. But
221 		 * that will never happen, as we're already holding a
222 		 * reference to it.
223 		 */
224 		bio_set_flag(*bio, BIO_QUEUE_ENTERED);
225 
226 		bio_chain(split, *bio);
227 		trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
228 		generic_make_request(*bio);
229 		*bio = split;
230 	}
231 }
232 EXPORT_SYMBOL(blk_queue_split);
233 
__blk_recalc_rq_segments(struct request_queue * q,struct bio * bio,bool no_sg_merge)234 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
235 					     struct bio *bio,
236 					     bool no_sg_merge)
237 {
238 	struct bio_vec bv, bvprv = { NULL };
239 	int cluster, prev = 0;
240 	unsigned int seg_size, nr_phys_segs;
241 	struct bio *fbio, *bbio;
242 	struct bvec_iter iter;
243 
244 	if (!bio)
245 		return 0;
246 
247 	switch (bio_op(bio)) {
248 	case REQ_OP_DISCARD:
249 	case REQ_OP_SECURE_ERASE:
250 	case REQ_OP_WRITE_ZEROES:
251 		return 0;
252 	case REQ_OP_WRITE_SAME:
253 		return 1;
254 	}
255 
256 	fbio = bio;
257 	cluster = blk_queue_cluster(q);
258 	seg_size = 0;
259 	nr_phys_segs = 0;
260 	for_each_bio(bio) {
261 		bio_for_each_segment(bv, bio, iter) {
262 			/*
263 			 * If SG merging is disabled, each bio vector is
264 			 * a segment
265 			 */
266 			if (no_sg_merge)
267 				goto new_segment;
268 
269 			if (prev && cluster) {
270 				if (seg_size + bv.bv_len
271 				    > queue_max_segment_size(q))
272 					goto new_segment;
273 				if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
274 					goto new_segment;
275 				if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
276 					goto new_segment;
277 
278 				seg_size += bv.bv_len;
279 				bvprv = bv;
280 				continue;
281 			}
282 new_segment:
283 			if (nr_phys_segs == 1 && seg_size >
284 			    fbio->bi_seg_front_size)
285 				fbio->bi_seg_front_size = seg_size;
286 
287 			nr_phys_segs++;
288 			bvprv = bv;
289 			prev = 1;
290 			seg_size = bv.bv_len;
291 		}
292 		bbio = bio;
293 	}
294 
295 	if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
296 		fbio->bi_seg_front_size = seg_size;
297 	if (seg_size > bbio->bi_seg_back_size)
298 		bbio->bi_seg_back_size = seg_size;
299 
300 	return nr_phys_segs;
301 }
302 
blk_recalc_rq_segments(struct request * rq)303 void blk_recalc_rq_segments(struct request *rq)
304 {
305 	bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
306 			&rq->q->queue_flags);
307 
308 	rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
309 			no_sg_merge);
310 }
311 
blk_recount_segments(struct request_queue * q,struct bio * bio)312 void blk_recount_segments(struct request_queue *q, struct bio *bio)
313 {
314 	unsigned short seg_cnt = bio_segments(bio);
315 
316 	if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
317 			(seg_cnt < queue_max_segments(q)))
318 		bio->bi_phys_segments = seg_cnt;
319 	else {
320 		struct bio *nxt = bio->bi_next;
321 
322 		bio->bi_next = NULL;
323 		bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
324 		bio->bi_next = nxt;
325 	}
326 
327 	bio_set_flag(bio, BIO_SEG_VALID);
328 }
329 EXPORT_SYMBOL(blk_recount_segments);
330 
blk_phys_contig_segment(struct request_queue * q,struct bio * bio,struct bio * nxt)331 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
332 				   struct bio *nxt)
333 {
334 	struct bio_vec end_bv = { NULL }, nxt_bv;
335 
336 	if (!blk_queue_cluster(q))
337 		return 0;
338 
339 	if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
340 	    queue_max_segment_size(q))
341 		return 0;
342 
343 	if (!bio_has_data(bio))
344 		return 1;
345 
346 	bio_get_last_bvec(bio, &end_bv);
347 	bio_get_first_bvec(nxt, &nxt_bv);
348 
349 	if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
350 		return 0;
351 
352 	/*
353 	 * bio and nxt are contiguous in memory; check if the queue allows
354 	 * these two to be merged into one
355 	 */
356 	if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
357 		return 1;
358 
359 	return 0;
360 }
361 
362 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)363 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
364 		     struct scatterlist *sglist, struct bio_vec *bvprv,
365 		     struct scatterlist **sg, int *nsegs, int *cluster)
366 {
367 
368 	int nbytes = bvec->bv_len;
369 
370 	if (*sg && *cluster) {
371 		if ((*sg)->length + nbytes > queue_max_segment_size(q))
372 			goto new_segment;
373 
374 		if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
375 			goto new_segment;
376 		if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
377 			goto new_segment;
378 
379 		(*sg)->length += nbytes;
380 	} else {
381 new_segment:
382 		if (!*sg)
383 			*sg = sglist;
384 		else {
385 			/*
386 			 * If the driver previously mapped a shorter
387 			 * list, we could see a termination bit
388 			 * prematurely unless it fully inits the sg
389 			 * table on each mapping. We KNOW that there
390 			 * must be more entries here or the driver
391 			 * would be buggy, so force clear the
392 			 * termination bit to avoid doing a full
393 			 * sg_init_table() in drivers for each command.
394 			 */
395 			sg_unmark_end(*sg);
396 			*sg = sg_next(*sg);
397 		}
398 
399 		sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
400 		(*nsegs)++;
401 	}
402 	*bvprv = *bvec;
403 }
404 
__blk_bvec_map_sg(struct request_queue * q,struct bio_vec bv,struct scatterlist * sglist,struct scatterlist ** sg)405 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
406 		struct scatterlist *sglist, struct scatterlist **sg)
407 {
408 	*sg = sglist;
409 	sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
410 	return 1;
411 }
412 
__blk_bios_map_sg(struct request_queue * q,struct bio * bio,struct scatterlist * sglist,struct scatterlist ** sg)413 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
414 			     struct scatterlist *sglist,
415 			     struct scatterlist **sg)
416 {
417 	struct bio_vec bvec, bvprv = { NULL };
418 	struct bvec_iter iter;
419 	int cluster = blk_queue_cluster(q), nsegs = 0;
420 
421 	for_each_bio(bio)
422 		bio_for_each_segment(bvec, bio, iter)
423 			__blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
424 					     &nsegs, &cluster);
425 
426 	return nsegs;
427 }
428 
429 /*
430  * map a request to scatterlist, return number of sg entries setup. Caller
431  * must make sure sg can hold rq->nr_phys_segments entries
432  */
blk_rq_map_sg(struct request_queue * q,struct request * rq,struct scatterlist * sglist)433 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
434 		  struct scatterlist *sglist)
435 {
436 	struct scatterlist *sg = NULL;
437 	int nsegs = 0;
438 
439 	if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
440 		nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
441 	else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
442 		nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
443 	else if (rq->bio)
444 		nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
445 
446 	if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
447 	    (blk_rq_bytes(rq) & q->dma_pad_mask)) {
448 		unsigned int pad_len =
449 			(q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
450 
451 		sg->length += pad_len;
452 		rq->extra_len += pad_len;
453 	}
454 
455 	if (q->dma_drain_size && q->dma_drain_needed(rq)) {
456 		if (op_is_write(req_op(rq)))
457 			memset(q->dma_drain_buffer, 0, q->dma_drain_size);
458 
459 		sg_unmark_end(sg);
460 		sg = sg_next(sg);
461 		sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
462 			    q->dma_drain_size,
463 			    ((unsigned long)q->dma_drain_buffer) &
464 			    (PAGE_SIZE - 1));
465 		nsegs++;
466 		rq->extra_len += q->dma_drain_size;
467 	}
468 
469 	if (sg)
470 		sg_mark_end(sg);
471 
472 	/*
473 	 * Something must have been wrong if the figured number of
474 	 * segment is bigger than number of req's physical segments
475 	 */
476 	WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
477 
478 	return nsegs;
479 }
480 EXPORT_SYMBOL(blk_rq_map_sg);
481 
ll_new_hw_segment(struct request_queue * q,struct request * req,struct bio * bio)482 static inline int ll_new_hw_segment(struct request_queue *q,
483 				    struct request *req,
484 				    struct bio *bio)
485 {
486 	int nr_phys_segs = bio_phys_segments(q, bio);
487 
488 	if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
489 		goto no_merge;
490 
491 	if (!blk_cgroup_mergeable(req, bio))
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_cgroup_mergeable(req, next->bio))
618 		return 0;
619 
620 	if (blk_integrity_merge_rq(q, req, next) == false)
621 		return 0;
622 
623 	/* Merge is OK... */
624 	req->nr_phys_segments = total_phys_segments;
625 	return 1;
626 }
627 
628 /**
629  * blk_rq_set_mixed_merge - mark a request as mixed merge
630  * @rq: request to mark as mixed merge
631  *
632  * Description:
633  *     @rq is about to be mixed merged.  Make sure the attributes
634  *     which can be mixed are set in each bio and mark @rq as mixed
635  *     merged.
636  */
blk_rq_set_mixed_merge(struct request * rq)637 void blk_rq_set_mixed_merge(struct request *rq)
638 {
639 	unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
640 	struct bio *bio;
641 
642 	if (rq->rq_flags & RQF_MIXED_MERGE)
643 		return;
644 
645 	/*
646 	 * @rq will no longer represent mixable attributes for all the
647 	 * contained bios.  It will just track those of the first one.
648 	 * Distributes the attributs to each bio.
649 	 */
650 	for (bio = rq->bio; bio; bio = bio->bi_next) {
651 		WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
652 			     (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
653 		bio->bi_opf |= ff;
654 	}
655 	rq->rq_flags |= RQF_MIXED_MERGE;
656 }
657 
blk_account_io_merge(struct request * req)658 static void blk_account_io_merge(struct request *req)
659 {
660 	if (blk_do_io_stat(req)) {
661 		struct hd_struct *part;
662 		int cpu;
663 
664 		cpu = part_stat_lock();
665 		part = req->part;
666 
667 		part_round_stats(req->q, cpu, part);
668 		part_dec_in_flight(req->q, part, rq_data_dir(req));
669 
670 		hd_struct_put(part);
671 		part_stat_unlock();
672 	}
673 }
674 /*
675  * Two cases of handling DISCARD merge:
676  * If max_discard_segments > 1, the driver takes every bio
677  * as a range and send them to controller together. The ranges
678  * needn't to be contiguous.
679  * Otherwise, the bios/requests will be handled as same as
680  * others which should be contiguous.
681  */
blk_discard_mergable(struct request * req)682 static inline bool blk_discard_mergable(struct request *req)
683 {
684 	if (req_op(req) == REQ_OP_DISCARD &&
685 	    queue_max_discard_segments(req->q) > 1)
686 		return true;
687 	return false;
688 }
689 
blk_try_req_merge(struct request * req,struct request * next)690 enum elv_merge blk_try_req_merge(struct request *req, struct request *next)
691 {
692 	if (blk_discard_mergable(req))
693 		return ELEVATOR_DISCARD_MERGE;
694 	else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
695 		return ELEVATOR_BACK_MERGE;
696 
697 	return ELEVATOR_NO_MERGE;
698 }
699 
700 /*
701  * For non-mq, this has to be called with the request spinlock acquired.
702  * For mq with scheduling, the appropriate queue wide lock should be held.
703  */
attempt_merge(struct request_queue * q,struct request * req,struct request * next)704 static struct request *attempt_merge(struct request_queue *q,
705 				     struct request *req, struct request *next)
706 {
707 	if (!q->mq_ops)
708 		lockdep_assert_held(q->queue_lock);
709 
710 	if (!rq_mergeable(req) || !rq_mergeable(next))
711 		return NULL;
712 
713 	if (req_op(req) != req_op(next))
714 		return NULL;
715 
716 	if (rq_data_dir(req) != rq_data_dir(next)
717 	    || req->rq_disk != next->rq_disk
718 	    || req_no_special_merge(next))
719 		return NULL;
720 
721 	if (req_op(req) == REQ_OP_WRITE_SAME &&
722 	    !blk_write_same_mergeable(req->bio, next->bio))
723 		return NULL;
724 
725 	/*
726 	 * Don't allow merge of different write hints, or for a hint with
727 	 * non-hint IO.
728 	 */
729 	if (req->write_hint != next->write_hint)
730 		return NULL;
731 
732 	/*
733 	 * If we are allowed to merge, then append bio list
734 	 * from next to rq and release next. merge_requests_fn
735 	 * will have updated segment counts, update sector
736 	 * counts here. Handle DISCARDs separately, as they
737 	 * have separate settings.
738 	 */
739 
740 	switch (blk_try_req_merge(req, next)) {
741 	case ELEVATOR_DISCARD_MERGE:
742 		if (!req_attempt_discard_merge(q, req, next))
743 			return NULL;
744 		break;
745 	case ELEVATOR_BACK_MERGE:
746 		if (!ll_merge_requests_fn(q, req, next))
747 			return NULL;
748 		break;
749 	default:
750 		return NULL;
751 	}
752 
753 	/*
754 	 * If failfast settings disagree or any of the two is already
755 	 * a mixed merge, mark both as mixed before proceeding.  This
756 	 * makes sure that all involved bios have mixable attributes
757 	 * set properly.
758 	 */
759 	if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
760 	    (req->cmd_flags & REQ_FAILFAST_MASK) !=
761 	    (next->cmd_flags & REQ_FAILFAST_MASK)) {
762 		blk_rq_set_mixed_merge(req);
763 		blk_rq_set_mixed_merge(next);
764 	}
765 
766 	/*
767 	 * At this point we have either done a back merge or front merge. We
768 	 * need the smaller start_time_ns of the merged requests to be the
769 	 * current request for accounting purposes.
770 	 */
771 	if (next->start_time_ns < req->start_time_ns)
772 		req->start_time_ns = next->start_time_ns;
773 
774 	req->biotail->bi_next = next->bio;
775 	req->biotail = next->biotail;
776 
777 	req->__data_len += blk_rq_bytes(next);
778 
779 	if (!blk_discard_mergable(req))
780 		elv_merge_requests(q, req, next);
781 
782 	/*
783 	 * 'next' is going away, so update stats accordingly
784 	 */
785 	blk_account_io_merge(next);
786 
787 	req->ioprio = ioprio_best(req->ioprio, next->ioprio);
788 	if (blk_rq_cpu_valid(next))
789 		req->cpu = next->cpu;
790 
791 	/*
792 	 * ownership of bio passed from next to req, return 'next' for
793 	 * the caller to free
794 	 */
795 	next->bio = NULL;
796 	return next;
797 }
798 
attempt_back_merge(struct request_queue * q,struct request * rq)799 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
800 {
801 	struct request *next = elv_latter_request(q, rq);
802 
803 	if (next)
804 		return attempt_merge(q, rq, next);
805 
806 	return NULL;
807 }
808 
attempt_front_merge(struct request_queue * q,struct request * rq)809 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
810 {
811 	struct request *prev = elv_former_request(q, rq);
812 
813 	if (prev)
814 		return attempt_merge(q, prev, rq);
815 
816 	return NULL;
817 }
818 
blk_attempt_req_merge(struct request_queue * q,struct request * rq,struct request * next)819 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
820 			  struct request *next)
821 {
822 	struct elevator_queue *e = q->elevator;
823 	struct request *free;
824 
825 	if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn)
826 		if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next))
827 			return 0;
828 
829 	free = attempt_merge(q, rq, next);
830 	if (free) {
831 		__blk_put_request(q, free);
832 		return 1;
833 	}
834 
835 	return 0;
836 }
837 
blk_rq_merge_ok(struct request * rq,struct bio * bio)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 and not a special request */
851 	if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq))
852 		return false;
853 
854 	/* don't merge across cgroup boundaries */
855 	if (!blk_cgroup_mergeable(rq, bio))
856 		return false;
857 
858 	/* only merge integrity protected bio into ditto rq */
859 	if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
860 		return false;
861 
862 	/* must be using the same buffer */
863 	if (req_op(rq) == REQ_OP_WRITE_SAME &&
864 	    !blk_write_same_mergeable(rq->bio, bio))
865 		return false;
866 
867 	/*
868 	 * Don't allow merge of different write hints, or for a hint with
869 	 * non-hint IO.
870 	 */
871 	if (rq->write_hint != bio->bi_write_hint)
872 		return false;
873 
874 	return true;
875 }
876 
blk_try_merge(struct request * rq,struct bio * bio)877 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
878 {
879 	if (blk_discard_mergable(rq))
880 		return ELEVATOR_DISCARD_MERGE;
881 	else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
882 		return ELEVATOR_BACK_MERGE;
883 	else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
884 		return ELEVATOR_FRONT_MERGE;
885 	return ELEVATOR_NO_MERGE;
886 }
887