xref: /linux/block/blk-merge.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
1 /*
2  * Functions related to segment and merge handling
3  */
4 #include <linux/kernel.h>
5 #include <linux/module.h>
6 #include <linux/bio.h>
7 #include <linux/blkdev.h>
8 #include <linux/scatterlist.h>
9 
10 #include "blk.h"
11 
12 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
13 					     struct bio *bio,
14 					     bool no_sg_merge)
15 {
16 	struct bio_vec bv, bvprv = { NULL };
17 	int cluster, high, highprv = 1;
18 	unsigned int seg_size, nr_phys_segs;
19 	struct bio *fbio, *bbio;
20 	struct bvec_iter iter;
21 
22 	if (!bio)
23 		return 0;
24 
25 	/*
26 	 * This should probably be returning 0, but blk_add_request_payload()
27 	 * (Christoph!!!!)
28 	 */
29 	if (bio->bi_rw & REQ_DISCARD)
30 		return 1;
31 
32 	if (bio->bi_rw & REQ_WRITE_SAME)
33 		return 1;
34 
35 	fbio = bio;
36 	cluster = blk_queue_cluster(q);
37 	seg_size = 0;
38 	nr_phys_segs = 0;
39 	high = 0;
40 	for_each_bio(bio) {
41 		bio_for_each_segment(bv, bio, iter) {
42 			/*
43 			 * If SG merging is disabled, each bio vector is
44 			 * a segment
45 			 */
46 			if (no_sg_merge)
47 				goto new_segment;
48 
49 			/*
50 			 * the trick here is making sure that a high page is
51 			 * never considered part of another segment, since
52 			 * that might change with the bounce page.
53 			 */
54 			high = page_to_pfn(bv.bv_page) > queue_bounce_pfn(q);
55 			if (!high && !highprv && cluster) {
56 				if (seg_size + bv.bv_len
57 				    > queue_max_segment_size(q))
58 					goto new_segment;
59 				if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
60 					goto new_segment;
61 				if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
62 					goto new_segment;
63 
64 				seg_size += bv.bv_len;
65 				bvprv = bv;
66 				continue;
67 			}
68 new_segment:
69 			if (nr_phys_segs == 1 && seg_size >
70 			    fbio->bi_seg_front_size)
71 				fbio->bi_seg_front_size = seg_size;
72 
73 			nr_phys_segs++;
74 			bvprv = bv;
75 			seg_size = bv.bv_len;
76 			highprv = high;
77 		}
78 		bbio = bio;
79 	}
80 
81 	if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
82 		fbio->bi_seg_front_size = seg_size;
83 	if (seg_size > bbio->bi_seg_back_size)
84 		bbio->bi_seg_back_size = seg_size;
85 
86 	return nr_phys_segs;
87 }
88 
89 void blk_recalc_rq_segments(struct request *rq)
90 {
91 	bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
92 			&rq->q->queue_flags);
93 
94 	rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
95 			no_sg_merge);
96 }
97 
98 void blk_recount_segments(struct request_queue *q, struct bio *bio)
99 {
100 	if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
101 			bio->bi_vcnt < queue_max_segments(q))
102 		bio->bi_phys_segments = bio->bi_vcnt;
103 	else {
104 		struct bio *nxt = bio->bi_next;
105 
106 		bio->bi_next = NULL;
107 		bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
108 		bio->bi_next = nxt;
109 	}
110 
111 	bio->bi_flags |= (1 << BIO_SEG_VALID);
112 }
113 EXPORT_SYMBOL(blk_recount_segments);
114 
115 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
116 				   struct bio *nxt)
117 {
118 	struct bio_vec end_bv = { NULL }, nxt_bv;
119 	struct bvec_iter iter;
120 
121 	if (!blk_queue_cluster(q))
122 		return 0;
123 
124 	if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
125 	    queue_max_segment_size(q))
126 		return 0;
127 
128 	if (!bio_has_data(bio))
129 		return 1;
130 
131 	bio_for_each_segment(end_bv, bio, iter)
132 		if (end_bv.bv_len == iter.bi_size)
133 			break;
134 
135 	nxt_bv = bio_iovec(nxt);
136 
137 	if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
138 		return 0;
139 
140 	/*
141 	 * bio and nxt are contiguous in memory; check if the queue allows
142 	 * these two to be merged into one
143 	 */
144 	if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
145 		return 1;
146 
147 	return 0;
148 }
149 
150 static inline void
151 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
152 		     struct scatterlist *sglist, struct bio_vec *bvprv,
153 		     struct scatterlist **sg, int *nsegs, int *cluster)
154 {
155 
156 	int nbytes = bvec->bv_len;
157 
158 	if (*sg && *cluster) {
159 		if ((*sg)->length + nbytes > queue_max_segment_size(q))
160 			goto new_segment;
161 
162 		if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
163 			goto new_segment;
164 		if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
165 			goto new_segment;
166 
167 		(*sg)->length += nbytes;
168 	} else {
169 new_segment:
170 		if (!*sg)
171 			*sg = sglist;
172 		else {
173 			/*
174 			 * If the driver previously mapped a shorter
175 			 * list, we could see a termination bit
176 			 * prematurely unless it fully inits the sg
177 			 * table on each mapping. We KNOW that there
178 			 * must be more entries here or the driver
179 			 * would be buggy, so force clear the
180 			 * termination bit to avoid doing a full
181 			 * sg_init_table() in drivers for each command.
182 			 */
183 			sg_unmark_end(*sg);
184 			*sg = sg_next(*sg);
185 		}
186 
187 		sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
188 		(*nsegs)++;
189 	}
190 	*bvprv = *bvec;
191 }
192 
193 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
194 			     struct scatterlist *sglist,
195 			     struct scatterlist **sg)
196 {
197 	struct bio_vec bvec, bvprv = { NULL };
198 	struct bvec_iter iter;
199 	int nsegs, cluster;
200 
201 	nsegs = 0;
202 	cluster = blk_queue_cluster(q);
203 
204 	if (bio->bi_rw & REQ_DISCARD) {
205 		/*
206 		 * This is a hack - drivers should be neither modifying the
207 		 * biovec, nor relying on bi_vcnt - but because of
208 		 * blk_add_request_payload(), a discard bio may or may not have
209 		 * a payload we need to set up here (thank you Christoph) and
210 		 * bi_vcnt is really the only way of telling if we need to.
211 		 */
212 
213 		if (bio->bi_vcnt)
214 			goto single_segment;
215 
216 		return 0;
217 	}
218 
219 	if (bio->bi_rw & REQ_WRITE_SAME) {
220 single_segment:
221 		*sg = sglist;
222 		bvec = bio_iovec(bio);
223 		sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
224 		return 1;
225 	}
226 
227 	for_each_bio(bio)
228 		bio_for_each_segment(bvec, bio, iter)
229 			__blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
230 					     &nsegs, &cluster);
231 
232 	return nsegs;
233 }
234 
235 /*
236  * map a request to scatterlist, return number of sg entries setup. Caller
237  * must make sure sg can hold rq->nr_phys_segments entries
238  */
239 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
240 		  struct scatterlist *sglist)
241 {
242 	struct scatterlist *sg = NULL;
243 	int nsegs = 0;
244 
245 	if (rq->bio)
246 		nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
247 
248 	if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
249 	    (blk_rq_bytes(rq) & q->dma_pad_mask)) {
250 		unsigned int pad_len =
251 			(q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
252 
253 		sg->length += pad_len;
254 		rq->extra_len += pad_len;
255 	}
256 
257 	if (q->dma_drain_size && q->dma_drain_needed(rq)) {
258 		if (rq->cmd_flags & REQ_WRITE)
259 			memset(q->dma_drain_buffer, 0, q->dma_drain_size);
260 
261 		sg->page_link &= ~0x02;
262 		sg = sg_next(sg);
263 		sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
264 			    q->dma_drain_size,
265 			    ((unsigned long)q->dma_drain_buffer) &
266 			    (PAGE_SIZE - 1));
267 		nsegs++;
268 		rq->extra_len += q->dma_drain_size;
269 	}
270 
271 	if (sg)
272 		sg_mark_end(sg);
273 
274 	return nsegs;
275 }
276 EXPORT_SYMBOL(blk_rq_map_sg);
277 
278 /**
279  * blk_bio_map_sg - map a bio to a scatterlist
280  * @q: request_queue in question
281  * @bio: bio being mapped
282  * @sglist: scatterlist being mapped
283  *
284  * Note:
285  *    Caller must make sure sg can hold bio->bi_phys_segments entries
286  *
287  * Will return the number of sg entries setup
288  */
289 int blk_bio_map_sg(struct request_queue *q, struct bio *bio,
290 		   struct scatterlist *sglist)
291 {
292 	struct scatterlist *sg = NULL;
293 	int nsegs;
294 	struct bio *next = bio->bi_next;
295 	bio->bi_next = NULL;
296 
297 	nsegs = __blk_bios_map_sg(q, bio, sglist, &sg);
298 	bio->bi_next = next;
299 	if (sg)
300 		sg_mark_end(sg);
301 
302 	BUG_ON(bio->bi_phys_segments && nsegs > bio->bi_phys_segments);
303 	return nsegs;
304 }
305 EXPORT_SYMBOL(blk_bio_map_sg);
306 
307 static inline int ll_new_hw_segment(struct request_queue *q,
308 				    struct request *req,
309 				    struct bio *bio)
310 {
311 	int nr_phys_segs = bio_phys_segments(q, bio);
312 
313 	if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
314 		goto no_merge;
315 
316 	if (bio_integrity(bio) && blk_integrity_merge_bio(q, req, bio))
317 		goto no_merge;
318 
319 	/*
320 	 * This will form the start of a new hw segment.  Bump both
321 	 * counters.
322 	 */
323 	req->nr_phys_segments += nr_phys_segs;
324 	return 1;
325 
326 no_merge:
327 	req->cmd_flags |= REQ_NOMERGE;
328 	if (req == q->last_merge)
329 		q->last_merge = NULL;
330 	return 0;
331 }
332 
333 int ll_back_merge_fn(struct request_queue *q, struct request *req,
334 		     struct bio *bio)
335 {
336 	if (blk_rq_sectors(req) + bio_sectors(bio) >
337 	    blk_rq_get_max_sectors(req)) {
338 		req->cmd_flags |= REQ_NOMERGE;
339 		if (req == q->last_merge)
340 			q->last_merge = NULL;
341 		return 0;
342 	}
343 	if (!bio_flagged(req->biotail, BIO_SEG_VALID))
344 		blk_recount_segments(q, req->biotail);
345 	if (!bio_flagged(bio, BIO_SEG_VALID))
346 		blk_recount_segments(q, bio);
347 
348 	return ll_new_hw_segment(q, req, bio);
349 }
350 
351 int ll_front_merge_fn(struct request_queue *q, struct request *req,
352 		      struct bio *bio)
353 {
354 	if (blk_rq_sectors(req) + bio_sectors(bio) >
355 	    blk_rq_get_max_sectors(req)) {
356 		req->cmd_flags |= REQ_NOMERGE;
357 		if (req == q->last_merge)
358 			q->last_merge = NULL;
359 		return 0;
360 	}
361 	if (!bio_flagged(bio, BIO_SEG_VALID))
362 		blk_recount_segments(q, bio);
363 	if (!bio_flagged(req->bio, BIO_SEG_VALID))
364 		blk_recount_segments(q, req->bio);
365 
366 	return ll_new_hw_segment(q, req, bio);
367 }
368 
369 /*
370  * blk-mq uses req->special to carry normal driver per-request payload, it
371  * does not indicate a prepared command that we cannot merge with.
372  */
373 static bool req_no_special_merge(struct request *req)
374 {
375 	struct request_queue *q = req->q;
376 
377 	return !q->mq_ops && req->special;
378 }
379 
380 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
381 				struct request *next)
382 {
383 	int total_phys_segments;
384 	unsigned int seg_size =
385 		req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
386 
387 	/*
388 	 * First check if the either of the requests are re-queued
389 	 * requests.  Can't merge them if they are.
390 	 */
391 	if (req_no_special_merge(req) || req_no_special_merge(next))
392 		return 0;
393 
394 	/*
395 	 * Will it become too large?
396 	 */
397 	if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
398 	    blk_rq_get_max_sectors(req))
399 		return 0;
400 
401 	total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
402 	if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
403 		if (req->nr_phys_segments == 1)
404 			req->bio->bi_seg_front_size = seg_size;
405 		if (next->nr_phys_segments == 1)
406 			next->biotail->bi_seg_back_size = seg_size;
407 		total_phys_segments--;
408 	}
409 
410 	if (total_phys_segments > queue_max_segments(q))
411 		return 0;
412 
413 	if (blk_integrity_rq(req) && blk_integrity_merge_rq(q, req, next))
414 		return 0;
415 
416 	/* Merge is OK... */
417 	req->nr_phys_segments = total_phys_segments;
418 	return 1;
419 }
420 
421 /**
422  * blk_rq_set_mixed_merge - mark a request as mixed merge
423  * @rq: request to mark as mixed merge
424  *
425  * Description:
426  *     @rq is about to be mixed merged.  Make sure the attributes
427  *     which can be mixed are set in each bio and mark @rq as mixed
428  *     merged.
429  */
430 void blk_rq_set_mixed_merge(struct request *rq)
431 {
432 	unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
433 	struct bio *bio;
434 
435 	if (rq->cmd_flags & REQ_MIXED_MERGE)
436 		return;
437 
438 	/*
439 	 * @rq will no longer represent mixable attributes for all the
440 	 * contained bios.  It will just track those of the first one.
441 	 * Distributes the attributs to each bio.
442 	 */
443 	for (bio = rq->bio; bio; bio = bio->bi_next) {
444 		WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
445 			     (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
446 		bio->bi_rw |= ff;
447 	}
448 	rq->cmd_flags |= REQ_MIXED_MERGE;
449 }
450 
451 static void blk_account_io_merge(struct request *req)
452 {
453 	if (blk_do_io_stat(req)) {
454 		struct hd_struct *part;
455 		int cpu;
456 
457 		cpu = part_stat_lock();
458 		part = req->part;
459 
460 		part_round_stats(cpu, part);
461 		part_dec_in_flight(part, rq_data_dir(req));
462 
463 		hd_struct_put(part);
464 		part_stat_unlock();
465 	}
466 }
467 
468 /*
469  * Has to be called with the request spinlock acquired
470  */
471 static int attempt_merge(struct request_queue *q, struct request *req,
472 			  struct request *next)
473 {
474 	if (!rq_mergeable(req) || !rq_mergeable(next))
475 		return 0;
476 
477 	if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
478 		return 0;
479 
480 	/*
481 	 * not contiguous
482 	 */
483 	if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
484 		return 0;
485 
486 	if (rq_data_dir(req) != rq_data_dir(next)
487 	    || req->rq_disk != next->rq_disk
488 	    || req_no_special_merge(next))
489 		return 0;
490 
491 	if (req->cmd_flags & REQ_WRITE_SAME &&
492 	    !blk_write_same_mergeable(req->bio, next->bio))
493 		return 0;
494 
495 	/*
496 	 * If we are allowed to merge, then append bio list
497 	 * from next to rq and release next. merge_requests_fn
498 	 * will have updated segment counts, update sector
499 	 * counts here.
500 	 */
501 	if (!ll_merge_requests_fn(q, req, next))
502 		return 0;
503 
504 	/*
505 	 * If failfast settings disagree or any of the two is already
506 	 * a mixed merge, mark both as mixed before proceeding.  This
507 	 * makes sure that all involved bios have mixable attributes
508 	 * set properly.
509 	 */
510 	if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
511 	    (req->cmd_flags & REQ_FAILFAST_MASK) !=
512 	    (next->cmd_flags & REQ_FAILFAST_MASK)) {
513 		blk_rq_set_mixed_merge(req);
514 		blk_rq_set_mixed_merge(next);
515 	}
516 
517 	/*
518 	 * At this point we have either done a back merge
519 	 * or front merge. We need the smaller start_time of
520 	 * the merged requests to be the current request
521 	 * for accounting purposes.
522 	 */
523 	if (time_after(req->start_time, next->start_time))
524 		req->start_time = next->start_time;
525 
526 	req->biotail->bi_next = next->bio;
527 	req->biotail = next->biotail;
528 
529 	req->__data_len += blk_rq_bytes(next);
530 
531 	elv_merge_requests(q, req, next);
532 
533 	/*
534 	 * 'next' is going away, so update stats accordingly
535 	 */
536 	blk_account_io_merge(next);
537 
538 	req->ioprio = ioprio_best(req->ioprio, next->ioprio);
539 	if (blk_rq_cpu_valid(next))
540 		req->cpu = next->cpu;
541 
542 	/* owner-ship of bio passed from next to req */
543 	next->bio = NULL;
544 	__blk_put_request(q, next);
545 	return 1;
546 }
547 
548 int attempt_back_merge(struct request_queue *q, struct request *rq)
549 {
550 	struct request *next = elv_latter_request(q, rq);
551 
552 	if (next)
553 		return attempt_merge(q, rq, next);
554 
555 	return 0;
556 }
557 
558 int attempt_front_merge(struct request_queue *q, struct request *rq)
559 {
560 	struct request *prev = elv_former_request(q, rq);
561 
562 	if (prev)
563 		return attempt_merge(q, prev, rq);
564 
565 	return 0;
566 }
567 
568 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
569 			  struct request *next)
570 {
571 	return attempt_merge(q, rq, next);
572 }
573 
574 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
575 {
576 	struct request_queue *q = rq->q;
577 
578 	if (!rq_mergeable(rq) || !bio_mergeable(bio))
579 		return false;
580 
581 	if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
582 		return false;
583 
584 	/* different data direction or already started, don't merge */
585 	if (bio_data_dir(bio) != rq_data_dir(rq))
586 		return false;
587 
588 	/* must be same device and not a special request */
589 	if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
590 		return false;
591 
592 	/* only merge integrity protected bio into ditto rq */
593 	if (bio_integrity(bio) != blk_integrity_rq(rq))
594 		return false;
595 
596 	/* must be using the same buffer */
597 	if (rq->cmd_flags & REQ_WRITE_SAME &&
598 	    !blk_write_same_mergeable(rq->bio, bio))
599 		return false;
600 
601 	if (q->queue_flags & (1 << QUEUE_FLAG_SG_GAPS)) {
602 		struct bio_vec *bprev;
603 
604 		bprev = &rq->biotail->bi_io_vec[bio->bi_vcnt - 1];
605 		if (bvec_gap_to_prev(bprev, bio->bi_io_vec[0].bv_offset))
606 			return false;
607 	}
608 
609 	return true;
610 }
611 
612 int blk_try_merge(struct request *rq, struct bio *bio)
613 {
614 	if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
615 		return ELEVATOR_BACK_MERGE;
616 	else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
617 		return ELEVATOR_FRONT_MERGE;
618 	return ELEVATOR_NO_MERGE;
619 }
620