Loading...
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 <trace/events/block.h>
11
12#include "blk.h"
13
14static struct bio *blk_bio_discard_split(struct request_queue *q,
15 struct bio *bio,
16 struct bio_set *bs,
17 unsigned *nsegs)
18{
19 unsigned int max_discard_sectors, granularity;
20 int alignment;
21 sector_t tmp;
22 unsigned split_sectors;
23
24 *nsegs = 1;
25
26 /* Zero-sector (unknown) and one-sector granularities are the same. */
27 granularity = max(q->limits.discard_granularity >> 9, 1U);
28
29 max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
30 max_discard_sectors -= max_discard_sectors % granularity;
31
32 if (unlikely(!max_discard_sectors)) {
33 /* XXX: warn */
34 return NULL;
35 }
36
37 if (bio_sectors(bio) <= max_discard_sectors)
38 return NULL;
39
40 split_sectors = max_discard_sectors;
41
42 /*
43 * If the next starting sector would be misaligned, stop the discard at
44 * the previous aligned sector.
45 */
46 alignment = (q->limits.discard_alignment >> 9) % granularity;
47
48 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
49 tmp = sector_div(tmp, granularity);
50
51 if (split_sectors > tmp)
52 split_sectors -= tmp;
53
54 return bio_split(bio, split_sectors, GFP_NOIO, bs);
55}
56
57static struct bio *blk_bio_write_same_split(struct request_queue *q,
58 struct bio *bio,
59 struct bio_set *bs,
60 unsigned *nsegs)
61{
62 *nsegs = 1;
63
64 if (!q->limits.max_write_same_sectors)
65 return NULL;
66
67 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
68 return NULL;
69
70 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
71}
72
73static inline unsigned get_max_io_size(struct request_queue *q,
74 struct bio *bio)
75{
76 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
77 unsigned mask = queue_logical_block_size(q) - 1;
78
79 /* aligned to logical block size */
80 sectors &= ~(mask >> 9);
81
82 return sectors;
83}
84
85static struct bio *blk_bio_segment_split(struct request_queue *q,
86 struct bio *bio,
87 struct bio_set *bs,
88 unsigned *segs)
89{
90 struct bio_vec bv, bvprv, *bvprvp = NULL;
91 struct bvec_iter iter;
92 unsigned seg_size = 0, nsegs = 0, sectors = 0;
93 unsigned front_seg_size = bio->bi_seg_front_size;
94 bool do_split = true;
95 struct bio *new = NULL;
96 const unsigned max_sectors = get_max_io_size(q, bio);
97
98 bio_for_each_segment(bv, bio, iter) {
99 /*
100 * If the queue doesn't support SG gaps and adding this
101 * offset would create a gap, disallow it.
102 */
103 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
104 goto split;
105
106 if (sectors + (bv.bv_len >> 9) > max_sectors) {
107 /*
108 * Consider this a new segment if we're splitting in
109 * the middle of this vector.
110 */
111 if (nsegs < queue_max_segments(q) &&
112 sectors < max_sectors) {
113 nsegs++;
114 sectors = max_sectors;
115 }
116 if (sectors)
117 goto split;
118 /* Make this single bvec as the 1st segment */
119 }
120
121 if (bvprvp && blk_queue_cluster(q)) {
122 if (seg_size + bv.bv_len > queue_max_segment_size(q))
123 goto new_segment;
124 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
125 goto new_segment;
126 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
127 goto new_segment;
128
129 seg_size += bv.bv_len;
130 bvprv = bv;
131 bvprvp = &bvprv;
132 sectors += bv.bv_len >> 9;
133
134 if (nsegs == 1 && seg_size > front_seg_size)
135 front_seg_size = seg_size;
136 continue;
137 }
138new_segment:
139 if (nsegs == queue_max_segments(q))
140 goto split;
141
142 nsegs++;
143 bvprv = bv;
144 bvprvp = &bvprv;
145 seg_size = bv.bv_len;
146 sectors += bv.bv_len >> 9;
147
148 if (nsegs == 1 && seg_size > front_seg_size)
149 front_seg_size = seg_size;
150 }
151
152 do_split = false;
153split:
154 *segs = nsegs;
155
156 if (do_split) {
157 new = bio_split(bio, sectors, GFP_NOIO, bs);
158 if (new)
159 bio = new;
160 }
161
162 bio->bi_seg_front_size = front_seg_size;
163 if (seg_size > bio->bi_seg_back_size)
164 bio->bi_seg_back_size = seg_size;
165
166 return do_split ? new : NULL;
167}
168
169void blk_queue_split(struct request_queue *q, struct bio **bio,
170 struct bio_set *bs)
171{
172 struct bio *split, *res;
173 unsigned nsegs;
174
175 if ((*bio)->bi_rw & REQ_DISCARD)
176 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
177 else if ((*bio)->bi_rw & REQ_WRITE_SAME)
178 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
179 else
180 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
181
182 /* physical segments can be figured out during splitting */
183 res = split ? split : *bio;
184 res->bi_phys_segments = nsegs;
185 bio_set_flag(res, BIO_SEG_VALID);
186
187 if (split) {
188 /* there isn't chance to merge the splitted bio */
189 split->bi_rw |= REQ_NOMERGE;
190
191 bio_chain(split, *bio);
192 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
193 generic_make_request(*bio);
194 *bio = split;
195 }
196}
197EXPORT_SYMBOL(blk_queue_split);
198
199static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
200 struct bio *bio,
201 bool no_sg_merge)
202{
203 struct bio_vec bv, bvprv = { NULL };
204 int cluster, prev = 0;
205 unsigned int seg_size, nr_phys_segs;
206 struct bio *fbio, *bbio;
207 struct bvec_iter iter;
208
209 if (!bio)
210 return 0;
211
212 /*
213 * This should probably be returning 0, but blk_add_request_payload()
214 * (Christoph!!!!)
215 */
216 if (bio->bi_rw & REQ_DISCARD)
217 return 1;
218
219 if (bio->bi_rw & REQ_WRITE_SAME)
220 return 1;
221
222 fbio = bio;
223 cluster = blk_queue_cluster(q);
224 seg_size = 0;
225 nr_phys_segs = 0;
226 for_each_bio(bio) {
227 bio_for_each_segment(bv, bio, iter) {
228 /*
229 * If SG merging is disabled, each bio vector is
230 * a segment
231 */
232 if (no_sg_merge)
233 goto new_segment;
234
235 if (prev && cluster) {
236 if (seg_size + bv.bv_len
237 > queue_max_segment_size(q))
238 goto new_segment;
239 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
240 goto new_segment;
241 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
242 goto new_segment;
243
244 seg_size += bv.bv_len;
245 bvprv = bv;
246 continue;
247 }
248new_segment:
249 if (nr_phys_segs == 1 && seg_size >
250 fbio->bi_seg_front_size)
251 fbio->bi_seg_front_size = seg_size;
252
253 nr_phys_segs++;
254 bvprv = bv;
255 prev = 1;
256 seg_size = bv.bv_len;
257 }
258 bbio = bio;
259 }
260
261 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
262 fbio->bi_seg_front_size = seg_size;
263 if (seg_size > bbio->bi_seg_back_size)
264 bbio->bi_seg_back_size = seg_size;
265
266 return nr_phys_segs;
267}
268
269void blk_recalc_rq_segments(struct request *rq)
270{
271 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
272 &rq->q->queue_flags);
273
274 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
275 no_sg_merge);
276}
277
278void blk_recount_segments(struct request_queue *q, struct bio *bio)
279{
280 unsigned short seg_cnt;
281
282 /* estimate segment number by bi_vcnt for non-cloned bio */
283 if (bio_flagged(bio, BIO_CLONED))
284 seg_cnt = bio_segments(bio);
285 else
286 seg_cnt = bio->bi_vcnt;
287
288 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
289 (seg_cnt < queue_max_segments(q)))
290 bio->bi_phys_segments = seg_cnt;
291 else {
292 struct bio *nxt = bio->bi_next;
293
294 bio->bi_next = NULL;
295 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
296 bio->bi_next = nxt;
297 }
298
299 bio_set_flag(bio, BIO_SEG_VALID);
300}
301EXPORT_SYMBOL(blk_recount_segments);
302
303static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
304 struct bio *nxt)
305{
306 struct bio_vec end_bv = { NULL }, nxt_bv;
307
308 if (!blk_queue_cluster(q))
309 return 0;
310
311 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
312 queue_max_segment_size(q))
313 return 0;
314
315 if (!bio_has_data(bio))
316 return 1;
317
318 bio_get_last_bvec(bio, &end_bv);
319 bio_get_first_bvec(nxt, &nxt_bv);
320
321 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
322 return 0;
323
324 /*
325 * bio and nxt are contiguous in memory; check if the queue allows
326 * these two to be merged into one
327 */
328 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
329 return 1;
330
331 return 0;
332}
333
334static inline void
335__blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
336 struct scatterlist *sglist, struct bio_vec *bvprv,
337 struct scatterlist **sg, int *nsegs, int *cluster)
338{
339
340 int nbytes = bvec->bv_len;
341
342 if (*sg && *cluster) {
343 if ((*sg)->length + nbytes > queue_max_segment_size(q))
344 goto new_segment;
345
346 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
347 goto new_segment;
348 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
349 goto new_segment;
350
351 (*sg)->length += nbytes;
352 } else {
353new_segment:
354 if (!*sg)
355 *sg = sglist;
356 else {
357 /*
358 * If the driver previously mapped a shorter
359 * list, we could see a termination bit
360 * prematurely unless it fully inits the sg
361 * table on each mapping. We KNOW that there
362 * must be more entries here or the driver
363 * would be buggy, so force clear the
364 * termination bit to avoid doing a full
365 * sg_init_table() in drivers for each command.
366 */
367 sg_unmark_end(*sg);
368 *sg = sg_next(*sg);
369 }
370
371 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
372 (*nsegs)++;
373 }
374 *bvprv = *bvec;
375}
376
377static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
378 struct scatterlist *sglist,
379 struct scatterlist **sg)
380{
381 struct bio_vec bvec, bvprv = { NULL };
382 struct bvec_iter iter;
383 int nsegs, cluster;
384
385 nsegs = 0;
386 cluster = blk_queue_cluster(q);
387
388 if (bio->bi_rw & REQ_DISCARD) {
389 /*
390 * This is a hack - drivers should be neither modifying the
391 * biovec, nor relying on bi_vcnt - but because of
392 * blk_add_request_payload(), a discard bio may or may not have
393 * a payload we need to set up here (thank you Christoph) and
394 * bi_vcnt is really the only way of telling if we need to.
395 */
396
397 if (bio->bi_vcnt)
398 goto single_segment;
399
400 return 0;
401 }
402
403 if (bio->bi_rw & REQ_WRITE_SAME) {
404single_segment:
405 *sg = sglist;
406 bvec = bio_iovec(bio);
407 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
408 return 1;
409 }
410
411 for_each_bio(bio)
412 bio_for_each_segment(bvec, bio, iter)
413 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
414 &nsegs, &cluster);
415
416 return nsegs;
417}
418
419/*
420 * map a request to scatterlist, return number of sg entries setup. Caller
421 * must make sure sg can hold rq->nr_phys_segments entries
422 */
423int blk_rq_map_sg(struct request_queue *q, struct request *rq,
424 struct scatterlist *sglist)
425{
426 struct scatterlist *sg = NULL;
427 int nsegs = 0;
428
429 if (rq->bio)
430 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
431
432 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
433 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
434 unsigned int pad_len =
435 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
436
437 sg->length += pad_len;
438 rq->extra_len += pad_len;
439 }
440
441 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
442 if (rq->cmd_flags & REQ_WRITE)
443 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
444
445 sg_unmark_end(sg);
446 sg = sg_next(sg);
447 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
448 q->dma_drain_size,
449 ((unsigned long)q->dma_drain_buffer) &
450 (PAGE_SIZE - 1));
451 nsegs++;
452 rq->extra_len += q->dma_drain_size;
453 }
454
455 if (sg)
456 sg_mark_end(sg);
457
458 /*
459 * Something must have been wrong if the figured number of
460 * segment is bigger than number of req's physical segments
461 */
462 WARN_ON(nsegs > rq->nr_phys_segments);
463
464 return nsegs;
465}
466EXPORT_SYMBOL(blk_rq_map_sg);
467
468static inline int ll_new_hw_segment(struct request_queue *q,
469 struct request *req,
470 struct bio *bio)
471{
472 int nr_phys_segs = bio_phys_segments(q, bio);
473
474 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
475 goto no_merge;
476
477 if (blk_integrity_merge_bio(q, req, bio) == false)
478 goto no_merge;
479
480 /*
481 * This will form the start of a new hw segment. Bump both
482 * counters.
483 */
484 req->nr_phys_segments += nr_phys_segs;
485 return 1;
486
487no_merge:
488 req->cmd_flags |= REQ_NOMERGE;
489 if (req == q->last_merge)
490 q->last_merge = NULL;
491 return 0;
492}
493
494int 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)) {
504 req->cmd_flags |= REQ_NOMERGE;
505 if (req == q->last_merge)
506 q->last_merge = NULL;
507 return 0;
508 }
509 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
510 blk_recount_segments(q, req->biotail);
511 if (!bio_flagged(bio, BIO_SEG_VALID))
512 blk_recount_segments(q, bio);
513
514 return ll_new_hw_segment(q, req, bio);
515}
516
517int ll_front_merge_fn(struct request_queue *q, struct request *req,
518 struct bio *bio)
519{
520
521 if (req_gap_front_merge(req, bio))
522 return 0;
523 if (blk_integrity_rq(req) &&
524 integrity_req_gap_front_merge(req, bio))
525 return 0;
526 if (blk_rq_sectors(req) + bio_sectors(bio) >
527 blk_rq_get_max_sectors(req)) {
528 req->cmd_flags |= REQ_NOMERGE;
529 if (req == q->last_merge)
530 q->last_merge = NULL;
531 return 0;
532 }
533 if (!bio_flagged(bio, BIO_SEG_VALID))
534 blk_recount_segments(q, bio);
535 if (!bio_flagged(req->bio, BIO_SEG_VALID))
536 blk_recount_segments(q, req->bio);
537
538 return ll_new_hw_segment(q, req, bio);
539}
540
541/*
542 * blk-mq uses req->special to carry normal driver per-request payload, it
543 * does not indicate a prepared command that we cannot merge with.
544 */
545static bool req_no_special_merge(struct request *req)
546{
547 struct request_queue *q = req->q;
548
549 return !q->mq_ops && req->special;
550}
551
552static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
553 struct request *next)
554{
555 int total_phys_segments;
556 unsigned int seg_size =
557 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
558
559 /*
560 * First check if the either of the requests are re-queued
561 * requests. Can't merge them if they are.
562 */
563 if (req_no_special_merge(req) || req_no_special_merge(next))
564 return 0;
565
566 if (req_gap_back_merge(req, next->bio))
567 return 0;
568
569 /*
570 * Will it become too large?
571 */
572 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
573 blk_rq_get_max_sectors(req))
574 return 0;
575
576 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
577 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
578 if (req->nr_phys_segments == 1)
579 req->bio->bi_seg_front_size = seg_size;
580 if (next->nr_phys_segments == 1)
581 next->biotail->bi_seg_back_size = seg_size;
582 total_phys_segments--;
583 }
584
585 if (total_phys_segments > queue_max_segments(q))
586 return 0;
587
588 if (blk_integrity_merge_rq(q, req, next) == false)
589 return 0;
590
591 /* Merge is OK... */
592 req->nr_phys_segments = total_phys_segments;
593 return 1;
594}
595
596/**
597 * blk_rq_set_mixed_merge - mark a request as mixed merge
598 * @rq: request to mark as mixed merge
599 *
600 * Description:
601 * @rq is about to be mixed merged. Make sure the attributes
602 * which can be mixed are set in each bio and mark @rq as mixed
603 * merged.
604 */
605void blk_rq_set_mixed_merge(struct request *rq)
606{
607 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
608 struct bio *bio;
609
610 if (rq->cmd_flags & REQ_MIXED_MERGE)
611 return;
612
613 /*
614 * @rq will no longer represent mixable attributes for all the
615 * contained bios. It will just track those of the first one.
616 * Distributes the attributs to each bio.
617 */
618 for (bio = rq->bio; bio; bio = bio->bi_next) {
619 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
620 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
621 bio->bi_rw |= ff;
622 }
623 rq->cmd_flags |= REQ_MIXED_MERGE;
624}
625
626static void blk_account_io_merge(struct request *req)
627{
628 if (blk_do_io_stat(req)) {
629 struct hd_struct *part;
630 int cpu;
631
632 cpu = part_stat_lock();
633 part = req->part;
634
635 part_round_stats(cpu, part);
636 part_dec_in_flight(part, rq_data_dir(req));
637
638 hd_struct_put(part);
639 part_stat_unlock();
640 }
641}
642
643/*
644 * Has to be called with the request spinlock acquired
645 */
646static int attempt_merge(struct request_queue *q, struct request *req,
647 struct request *next)
648{
649 if (!rq_mergeable(req) || !rq_mergeable(next))
650 return 0;
651
652 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
653 return 0;
654
655 /*
656 * not contiguous
657 */
658 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
659 return 0;
660
661 if (rq_data_dir(req) != rq_data_dir(next)
662 || req->rq_disk != next->rq_disk
663 || req_no_special_merge(next))
664 return 0;
665
666 if (req->cmd_flags & REQ_WRITE_SAME &&
667 !blk_write_same_mergeable(req->bio, next->bio))
668 return 0;
669
670 /*
671 * If we are allowed to merge, then append bio list
672 * from next to rq and release next. merge_requests_fn
673 * will have updated segment counts, update sector
674 * counts here.
675 */
676 if (!ll_merge_requests_fn(q, req, next))
677 return 0;
678
679 /*
680 * If failfast settings disagree or any of the two is already
681 * a mixed merge, mark both as mixed before proceeding. This
682 * makes sure that all involved bios have mixable attributes
683 * set properly.
684 */
685 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
686 (req->cmd_flags & REQ_FAILFAST_MASK) !=
687 (next->cmd_flags & REQ_FAILFAST_MASK)) {
688 blk_rq_set_mixed_merge(req);
689 blk_rq_set_mixed_merge(next);
690 }
691
692 /*
693 * At this point we have either done a back merge
694 * or front merge. We need the smaller start_time of
695 * the merged requests to be the current request
696 * for accounting purposes.
697 */
698 if (time_after(req->start_time, next->start_time))
699 req->start_time = next->start_time;
700
701 req->biotail->bi_next = next->bio;
702 req->biotail = next->biotail;
703
704 req->__data_len += blk_rq_bytes(next);
705
706 elv_merge_requests(q, req, next);
707
708 /*
709 * 'next' is going away, so update stats accordingly
710 */
711 blk_account_io_merge(next);
712
713 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
714 if (blk_rq_cpu_valid(next))
715 req->cpu = next->cpu;
716
717 /* owner-ship of bio passed from next to req */
718 next->bio = NULL;
719 __blk_put_request(q, next);
720 return 1;
721}
722
723int attempt_back_merge(struct request_queue *q, struct request *rq)
724{
725 struct request *next = elv_latter_request(q, rq);
726
727 if (next)
728 return attempt_merge(q, rq, next);
729
730 return 0;
731}
732
733int attempt_front_merge(struct request_queue *q, struct request *rq)
734{
735 struct request *prev = elv_former_request(q, rq);
736
737 if (prev)
738 return attempt_merge(q, prev, rq);
739
740 return 0;
741}
742
743int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
744 struct request *next)
745{
746 return attempt_merge(q, rq, next);
747}
748
749bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
750{
751 if (!rq_mergeable(rq) || !bio_mergeable(bio))
752 return false;
753
754 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
755 return false;
756
757 /* different data direction or already started, don't merge */
758 if (bio_data_dir(bio) != rq_data_dir(rq))
759 return false;
760
761 /* must be same device and not a special request */
762 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
763 return false;
764
765 /* only merge integrity protected bio into ditto rq */
766 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
767 return false;
768
769 /* must be using the same buffer */
770 if (rq->cmd_flags & REQ_WRITE_SAME &&
771 !blk_write_same_mergeable(rq->bio, bio))
772 return false;
773
774 return true;
775}
776
777int blk_try_merge(struct request *rq, struct bio *bio)
778{
779 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
780 return ELEVATOR_BACK_MERGE;
781 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
782 return ELEVATOR_FRONT_MERGE;
783 return ELEVATOR_NO_MERGE;
784}
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/blk-integrity.h>
10#include <linux/scatterlist.h>
11#include <linux/part_stat.h>
12#include <linux/blk-cgroup.h>
13
14#include <trace/events/block.h>
15
16#include "blk.h"
17#include "blk-mq-sched.h"
18#include "blk-rq-qos.h"
19#include "blk-throttle.h"
20
21static inline void bio_get_first_bvec(struct bio *bio, struct bio_vec *bv)
22{
23 *bv = mp_bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
24}
25
26static inline void bio_get_last_bvec(struct bio *bio, struct bio_vec *bv)
27{
28 struct bvec_iter iter = bio->bi_iter;
29 int idx;
30
31 bio_get_first_bvec(bio, bv);
32 if (bv->bv_len == bio->bi_iter.bi_size)
33 return; /* this bio only has a single bvec */
34
35 bio_advance_iter(bio, &iter, iter.bi_size);
36
37 if (!iter.bi_bvec_done)
38 idx = iter.bi_idx - 1;
39 else /* in the middle of bvec */
40 idx = iter.bi_idx;
41
42 *bv = bio->bi_io_vec[idx];
43
44 /*
45 * iter.bi_bvec_done records actual length of the last bvec
46 * if this bio ends in the middle of one io vector
47 */
48 if (iter.bi_bvec_done)
49 bv->bv_len = iter.bi_bvec_done;
50}
51
52static inline bool bio_will_gap(struct request_queue *q,
53 struct request *prev_rq, struct bio *prev, struct bio *next)
54{
55 struct bio_vec pb, nb;
56
57 if (!bio_has_data(prev) || !queue_virt_boundary(q))
58 return false;
59
60 /*
61 * Don't merge if the 1st bio starts with non-zero offset, otherwise it
62 * is quite difficult to respect the sg gap limit. We work hard to
63 * merge a huge number of small single bios in case of mkfs.
64 */
65 if (prev_rq)
66 bio_get_first_bvec(prev_rq->bio, &pb);
67 else
68 bio_get_first_bvec(prev, &pb);
69 if (pb.bv_offset & queue_virt_boundary(q))
70 return true;
71
72 /*
73 * We don't need to worry about the situation that the merged segment
74 * ends in unaligned virt boundary:
75 *
76 * - if 'pb' ends aligned, the merged segment ends aligned
77 * - if 'pb' ends unaligned, the next bio must include
78 * one single bvec of 'nb', otherwise the 'nb' can't
79 * merge with 'pb'
80 */
81 bio_get_last_bvec(prev, &pb);
82 bio_get_first_bvec(next, &nb);
83 if (biovec_phys_mergeable(q, &pb, &nb))
84 return false;
85 return __bvec_gap_to_prev(&q->limits, &pb, nb.bv_offset);
86}
87
88static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
89{
90 return bio_will_gap(req->q, req, req->biotail, bio);
91}
92
93static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
94{
95 return bio_will_gap(req->q, NULL, bio, req->bio);
96}
97
98/*
99 * The max size one bio can handle is UINT_MAX becasue bvec_iter.bi_size
100 * is defined as 'unsigned int', meantime it has to be aligned to with the
101 * logical block size, which is the minimum accepted unit by hardware.
102 */
103static unsigned int bio_allowed_max_sectors(const struct queue_limits *lim)
104{
105 return round_down(UINT_MAX, lim->logical_block_size) >> SECTOR_SHIFT;
106}
107
108static struct bio *bio_submit_split(struct bio *bio, int split_sectors)
109{
110 if (unlikely(split_sectors < 0))
111 goto error;
112
113 if (split_sectors) {
114 struct bio *split;
115
116 split = bio_split(bio, split_sectors, GFP_NOIO,
117 &bio->bi_bdev->bd_disk->bio_split);
118 if (IS_ERR(split)) {
119 split_sectors = PTR_ERR(split);
120 goto error;
121 }
122 split->bi_opf |= REQ_NOMERGE;
123 blkcg_bio_issue_init(split);
124 bio_chain(split, bio);
125 trace_block_split(split, bio->bi_iter.bi_sector);
126 WARN_ON_ONCE(bio_zone_write_plugging(bio));
127 submit_bio_noacct(bio);
128 return split;
129 }
130
131 return bio;
132error:
133 bio->bi_status = errno_to_blk_status(split_sectors);
134 bio_endio(bio);
135 return NULL;
136}
137
138struct bio *bio_split_discard(struct bio *bio, const struct queue_limits *lim,
139 unsigned *nsegs)
140{
141 unsigned int max_discard_sectors, granularity;
142 sector_t tmp;
143 unsigned split_sectors;
144
145 *nsegs = 1;
146
147 granularity = max(lim->discard_granularity >> 9, 1U);
148
149 max_discard_sectors =
150 min(lim->max_discard_sectors, bio_allowed_max_sectors(lim));
151 max_discard_sectors -= max_discard_sectors % granularity;
152 if (unlikely(!max_discard_sectors))
153 return bio;
154
155 if (bio_sectors(bio) <= max_discard_sectors)
156 return bio;
157
158 split_sectors = max_discard_sectors;
159
160 /*
161 * If the next starting sector would be misaligned, stop the discard at
162 * the previous aligned sector.
163 */
164 tmp = bio->bi_iter.bi_sector + split_sectors -
165 ((lim->discard_alignment >> 9) % granularity);
166 tmp = sector_div(tmp, granularity);
167
168 if (split_sectors > tmp)
169 split_sectors -= tmp;
170
171 return bio_submit_split(bio, split_sectors);
172}
173
174static inline unsigned int blk_boundary_sectors(const struct queue_limits *lim,
175 bool is_atomic)
176{
177 /*
178 * chunk_sectors must be a multiple of atomic_write_boundary_sectors if
179 * both non-zero.
180 */
181 if (is_atomic && lim->atomic_write_boundary_sectors)
182 return lim->atomic_write_boundary_sectors;
183
184 return lim->chunk_sectors;
185}
186
187/*
188 * Return the maximum number of sectors from the start of a bio that may be
189 * submitted as a single request to a block device. If enough sectors remain,
190 * align the end to the physical block size. Otherwise align the end to the
191 * logical block size. This approach minimizes the number of non-aligned
192 * requests that are submitted to a block device if the start of a bio is not
193 * aligned to a physical block boundary.
194 */
195static inline unsigned get_max_io_size(struct bio *bio,
196 const struct queue_limits *lim)
197{
198 unsigned pbs = lim->physical_block_size >> SECTOR_SHIFT;
199 unsigned lbs = lim->logical_block_size >> SECTOR_SHIFT;
200 bool is_atomic = bio->bi_opf & REQ_ATOMIC;
201 unsigned boundary_sectors = blk_boundary_sectors(lim, is_atomic);
202 unsigned max_sectors, start, end;
203
204 /*
205 * We ignore lim->max_sectors for atomic writes because it may less
206 * than the actual bio size, which we cannot tolerate.
207 */
208 if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
209 max_sectors = lim->max_write_zeroes_sectors;
210 else if (is_atomic)
211 max_sectors = lim->atomic_write_max_sectors;
212 else
213 max_sectors = lim->max_sectors;
214
215 if (boundary_sectors) {
216 max_sectors = min(max_sectors,
217 blk_boundary_sectors_left(bio->bi_iter.bi_sector,
218 boundary_sectors));
219 }
220
221 start = bio->bi_iter.bi_sector & (pbs - 1);
222 end = (start + max_sectors) & ~(pbs - 1);
223 if (end > start)
224 return end - start;
225 return max_sectors & ~(lbs - 1);
226}
227
228/**
229 * get_max_segment_size() - maximum number of bytes to add as a single segment
230 * @lim: Request queue limits.
231 * @paddr: address of the range to add
232 * @len: maximum length available to add at @paddr
233 *
234 * Returns the maximum number of bytes of the range starting at @paddr that can
235 * be added to a single segment.
236 */
237static inline unsigned get_max_segment_size(const struct queue_limits *lim,
238 phys_addr_t paddr, unsigned int len)
239{
240 /*
241 * Prevent an overflow if mask = ULONG_MAX and offset = 0 by adding 1
242 * after having calculated the minimum.
243 */
244 return min_t(unsigned long, len,
245 min(lim->seg_boundary_mask - (lim->seg_boundary_mask & paddr),
246 (unsigned long)lim->max_segment_size - 1) + 1);
247}
248
249/**
250 * bvec_split_segs - verify whether or not a bvec should be split in the middle
251 * @lim: [in] queue limits to split based on
252 * @bv: [in] bvec to examine
253 * @nsegs: [in,out] Number of segments in the bio being built. Incremented
254 * by the number of segments from @bv that may be appended to that
255 * bio without exceeding @max_segs
256 * @bytes: [in,out] Number of bytes in the bio being built. Incremented
257 * by the number of bytes from @bv that may be appended to that
258 * bio without exceeding @max_bytes
259 * @max_segs: [in] upper bound for *@nsegs
260 * @max_bytes: [in] upper bound for *@bytes
261 *
262 * When splitting a bio, it can happen that a bvec is encountered that is too
263 * big to fit in a single segment and hence that it has to be split in the
264 * middle. This function verifies whether or not that should happen. The value
265 * %true is returned if and only if appending the entire @bv to a bio with
266 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
267 * the block driver.
268 */
269static bool bvec_split_segs(const struct queue_limits *lim,
270 const struct bio_vec *bv, unsigned *nsegs, unsigned *bytes,
271 unsigned max_segs, unsigned max_bytes)
272{
273 unsigned max_len = min(max_bytes, UINT_MAX) - *bytes;
274 unsigned len = min(bv->bv_len, max_len);
275 unsigned total_len = 0;
276 unsigned seg_size = 0;
277
278 while (len && *nsegs < max_segs) {
279 seg_size = get_max_segment_size(lim, bvec_phys(bv) + total_len, len);
280
281 (*nsegs)++;
282 total_len += seg_size;
283 len -= seg_size;
284
285 if ((bv->bv_offset + total_len) & lim->virt_boundary_mask)
286 break;
287 }
288
289 *bytes += total_len;
290
291 /* tell the caller to split the bvec if it is too big to fit */
292 return len > 0 || bv->bv_len > max_len;
293}
294
295static unsigned int bio_split_alignment(struct bio *bio,
296 const struct queue_limits *lim)
297{
298 if (op_is_write(bio_op(bio)) && lim->zone_write_granularity)
299 return lim->zone_write_granularity;
300 return lim->logical_block_size;
301}
302
303/**
304 * bio_split_rw_at - check if and where to split a read/write bio
305 * @bio: [in] bio to be split
306 * @lim: [in] queue limits to split based on
307 * @segs: [out] number of segments in the bio with the first half of the sectors
308 * @max_bytes: [in] maximum number of bytes per bio
309 *
310 * Find out if @bio needs to be split to fit the queue limits in @lim and a
311 * maximum size of @max_bytes. Returns a negative error number if @bio can't be
312 * split, 0 if the bio doesn't have to be split, or a positive sector offset if
313 * @bio needs to be split.
314 */
315int bio_split_rw_at(struct bio *bio, const struct queue_limits *lim,
316 unsigned *segs, unsigned max_bytes)
317{
318 struct bio_vec bv, bvprv, *bvprvp = NULL;
319 struct bvec_iter iter;
320 unsigned nsegs = 0, bytes = 0;
321
322 bio_for_each_bvec(bv, bio, iter) {
323 /*
324 * If the queue doesn't support SG gaps and adding this
325 * offset would create a gap, disallow it.
326 */
327 if (bvprvp && bvec_gap_to_prev(lim, bvprvp, bv.bv_offset))
328 goto split;
329
330 if (nsegs < lim->max_segments &&
331 bytes + bv.bv_len <= max_bytes &&
332 bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
333 nsegs++;
334 bytes += bv.bv_len;
335 } else {
336 if (bvec_split_segs(lim, &bv, &nsegs, &bytes,
337 lim->max_segments, max_bytes))
338 goto split;
339 }
340
341 bvprv = bv;
342 bvprvp = &bvprv;
343 }
344
345 *segs = nsegs;
346 return 0;
347split:
348 if (bio->bi_opf & REQ_ATOMIC)
349 return -EINVAL;
350
351 /*
352 * We can't sanely support splitting for a REQ_NOWAIT bio. End it
353 * with EAGAIN if splitting is required and return an error pointer.
354 */
355 if (bio->bi_opf & REQ_NOWAIT)
356 return -EAGAIN;
357
358 *segs = nsegs;
359
360 /*
361 * Individual bvecs might not be logical block aligned. Round down the
362 * split size so that each bio is properly block size aligned, even if
363 * we do not use the full hardware limits.
364 */
365 bytes = ALIGN_DOWN(bytes, bio_split_alignment(bio, lim));
366
367 /*
368 * Bio splitting may cause subtle trouble such as hang when doing sync
369 * iopoll in direct IO routine. Given performance gain of iopoll for
370 * big IO can be trival, disable iopoll when split needed.
371 */
372 bio_clear_polled(bio);
373 return bytes >> SECTOR_SHIFT;
374}
375EXPORT_SYMBOL_GPL(bio_split_rw_at);
376
377struct bio *bio_split_rw(struct bio *bio, const struct queue_limits *lim,
378 unsigned *nr_segs)
379{
380 return bio_submit_split(bio,
381 bio_split_rw_at(bio, lim, nr_segs,
382 get_max_io_size(bio, lim) << SECTOR_SHIFT));
383}
384
385/*
386 * REQ_OP_ZONE_APPEND bios must never be split by the block layer.
387 *
388 * But we want the nr_segs calculation provided by bio_split_rw_at, and having
389 * a good sanity check that the submitter built the bio correctly is nice to
390 * have as well.
391 */
392struct bio *bio_split_zone_append(struct bio *bio,
393 const struct queue_limits *lim, unsigned *nr_segs)
394{
395 int split_sectors;
396
397 split_sectors = bio_split_rw_at(bio, lim, nr_segs,
398 lim->max_zone_append_sectors << SECTOR_SHIFT);
399 if (WARN_ON_ONCE(split_sectors > 0))
400 split_sectors = -EINVAL;
401 return bio_submit_split(bio, split_sectors);
402}
403
404struct bio *bio_split_write_zeroes(struct bio *bio,
405 const struct queue_limits *lim, unsigned *nsegs)
406{
407 unsigned int max_sectors = get_max_io_size(bio, lim);
408
409 *nsegs = 0;
410
411 /*
412 * An unset limit should normally not happen, as bio submission is keyed
413 * off having a non-zero limit. But SCSI can clear the limit in the
414 * I/O completion handler, and we can race and see this. Splitting to a
415 * zero limit obviously doesn't make sense, so band-aid it here.
416 */
417 if (!max_sectors)
418 return bio;
419 if (bio_sectors(bio) <= max_sectors)
420 return bio;
421 return bio_submit_split(bio, max_sectors);
422}
423
424/**
425 * bio_split_to_limits - split a bio to fit the queue limits
426 * @bio: bio to be split
427 *
428 * Check if @bio needs splitting based on the queue limits of @bio->bi_bdev, and
429 * if so split off a bio fitting the limits from the beginning of @bio and
430 * return it. @bio is shortened to the remainder and re-submitted.
431 *
432 * The split bio is allocated from @q->bio_split, which is provided by the
433 * block layer.
434 */
435struct bio *bio_split_to_limits(struct bio *bio)
436{
437 unsigned int nr_segs;
438
439 return __bio_split_to_limits(bio, bdev_limits(bio->bi_bdev), &nr_segs);
440}
441EXPORT_SYMBOL(bio_split_to_limits);
442
443unsigned int blk_recalc_rq_segments(struct request *rq)
444{
445 unsigned int nr_phys_segs = 0;
446 unsigned int bytes = 0;
447 struct req_iterator iter;
448 struct bio_vec bv;
449
450 if (!rq->bio)
451 return 0;
452
453 switch (bio_op(rq->bio)) {
454 case REQ_OP_DISCARD:
455 case REQ_OP_SECURE_ERASE:
456 if (queue_max_discard_segments(rq->q) > 1) {
457 struct bio *bio = rq->bio;
458
459 for_each_bio(bio)
460 nr_phys_segs++;
461 return nr_phys_segs;
462 }
463 return 1;
464 case REQ_OP_WRITE_ZEROES:
465 return 0;
466 default:
467 break;
468 }
469
470 rq_for_each_bvec(bv, rq, iter)
471 bvec_split_segs(&rq->q->limits, &bv, &nr_phys_segs, &bytes,
472 UINT_MAX, UINT_MAX);
473 return nr_phys_segs;
474}
475
476static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
477 struct scatterlist *sglist)
478{
479 if (!*sg)
480 return sglist;
481
482 /*
483 * If the driver previously mapped a shorter list, we could see a
484 * termination bit prematurely unless it fully inits the sg table
485 * on each mapping. We KNOW that there must be more entries here
486 * or the driver would be buggy, so force clear the termination bit
487 * to avoid doing a full sg_init_table() in drivers for each command.
488 */
489 sg_unmark_end(*sg);
490 return sg_next(*sg);
491}
492
493static unsigned blk_bvec_map_sg(struct request_queue *q,
494 struct bio_vec *bvec, struct scatterlist *sglist,
495 struct scatterlist **sg)
496{
497 unsigned nbytes = bvec->bv_len;
498 unsigned nsegs = 0, total = 0;
499
500 while (nbytes > 0) {
501 unsigned offset = bvec->bv_offset + total;
502 unsigned len = get_max_segment_size(&q->limits,
503 bvec_phys(bvec) + total, nbytes);
504 struct page *page = bvec->bv_page;
505
506 /*
507 * Unfortunately a fair number of drivers barf on scatterlists
508 * that have an offset larger than PAGE_SIZE, despite other
509 * subsystems dealing with that invariant just fine. For now
510 * stick to the legacy format where we never present those from
511 * the block layer, but the code below should be removed once
512 * these offenders (mostly MMC/SD drivers) are fixed.
513 */
514 page += (offset >> PAGE_SHIFT);
515 offset &= ~PAGE_MASK;
516
517 *sg = blk_next_sg(sg, sglist);
518 sg_set_page(*sg, page, len, offset);
519
520 total += len;
521 nbytes -= len;
522 nsegs++;
523 }
524
525 return nsegs;
526}
527
528static inline int __blk_bvec_map_sg(struct bio_vec bv,
529 struct scatterlist *sglist, struct scatterlist **sg)
530{
531 *sg = blk_next_sg(sg, sglist);
532 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
533 return 1;
534}
535
536/* only try to merge bvecs into one sg if they are from two bios */
537static inline bool
538__blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
539 struct bio_vec *bvprv, struct scatterlist **sg)
540{
541
542 int nbytes = bvec->bv_len;
543
544 if (!*sg)
545 return false;
546
547 if ((*sg)->length + nbytes > queue_max_segment_size(q))
548 return false;
549
550 if (!biovec_phys_mergeable(q, bvprv, bvec))
551 return false;
552
553 (*sg)->length += nbytes;
554
555 return true;
556}
557
558static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
559 struct scatterlist *sglist,
560 struct scatterlist **sg)
561{
562 struct bio_vec bvec, bvprv = { NULL };
563 struct bvec_iter iter;
564 int nsegs = 0;
565 bool new_bio = false;
566
567 for_each_bio(bio) {
568 bio_for_each_bvec(bvec, bio, iter) {
569 /*
570 * Only try to merge bvecs from two bios given we
571 * have done bio internal merge when adding pages
572 * to bio
573 */
574 if (new_bio &&
575 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
576 goto next_bvec;
577
578 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
579 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
580 else
581 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
582 next_bvec:
583 new_bio = false;
584 }
585 if (likely(bio->bi_iter.bi_size)) {
586 bvprv = bvec;
587 new_bio = true;
588 }
589 }
590
591 return nsegs;
592}
593
594/*
595 * map a request to scatterlist, return number of sg entries setup. Caller
596 * must make sure sg can hold rq->nr_phys_segments entries
597 */
598int __blk_rq_map_sg(struct request_queue *q, struct request *rq,
599 struct scatterlist *sglist, struct scatterlist **last_sg)
600{
601 int nsegs = 0;
602
603 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
604 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg);
605 else if (rq->bio)
606 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg);
607
608 if (*last_sg)
609 sg_mark_end(*last_sg);
610
611 /*
612 * Something must have been wrong if the figured number of
613 * segment is bigger than number of req's physical segments
614 */
615 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
616
617 return nsegs;
618}
619EXPORT_SYMBOL(__blk_rq_map_sg);
620
621static inline unsigned int blk_rq_get_max_sectors(struct request *rq,
622 sector_t offset)
623{
624 struct request_queue *q = rq->q;
625 struct queue_limits *lim = &q->limits;
626 unsigned int max_sectors, boundary_sectors;
627 bool is_atomic = rq->cmd_flags & REQ_ATOMIC;
628
629 if (blk_rq_is_passthrough(rq))
630 return q->limits.max_hw_sectors;
631
632 boundary_sectors = blk_boundary_sectors(lim, is_atomic);
633 max_sectors = blk_queue_get_max_sectors(rq);
634
635 if (!boundary_sectors ||
636 req_op(rq) == REQ_OP_DISCARD ||
637 req_op(rq) == REQ_OP_SECURE_ERASE)
638 return max_sectors;
639 return min(max_sectors,
640 blk_boundary_sectors_left(offset, boundary_sectors));
641}
642
643static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
644 unsigned int nr_phys_segs)
645{
646 if (!blk_cgroup_mergeable(req, bio))
647 goto no_merge;
648
649 if (blk_integrity_merge_bio(req->q, req, bio) == false)
650 goto no_merge;
651
652 /* discard request merge won't add new segment */
653 if (req_op(req) == REQ_OP_DISCARD)
654 return 1;
655
656 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req))
657 goto no_merge;
658
659 /*
660 * This will form the start of a new hw segment. Bump both
661 * counters.
662 */
663 req->nr_phys_segments += nr_phys_segs;
664 if (bio_integrity(bio))
665 req->nr_integrity_segments += blk_rq_count_integrity_sg(req->q,
666 bio);
667 return 1;
668
669no_merge:
670 req_set_nomerge(req->q, req);
671 return 0;
672}
673
674int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
675{
676 if (req_gap_back_merge(req, bio))
677 return 0;
678 if (blk_integrity_rq(req) &&
679 integrity_req_gap_back_merge(req, bio))
680 return 0;
681 if (!bio_crypt_ctx_back_mergeable(req, bio))
682 return 0;
683 if (blk_rq_sectors(req) + bio_sectors(bio) >
684 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
685 req_set_nomerge(req->q, req);
686 return 0;
687 }
688
689 return ll_new_hw_segment(req, bio, nr_segs);
690}
691
692static int ll_front_merge_fn(struct request *req, struct bio *bio,
693 unsigned int nr_segs)
694{
695 if (req_gap_front_merge(req, bio))
696 return 0;
697 if (blk_integrity_rq(req) &&
698 integrity_req_gap_front_merge(req, bio))
699 return 0;
700 if (!bio_crypt_ctx_front_mergeable(req, bio))
701 return 0;
702 if (blk_rq_sectors(req) + bio_sectors(bio) >
703 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
704 req_set_nomerge(req->q, req);
705 return 0;
706 }
707
708 return ll_new_hw_segment(req, bio, nr_segs);
709}
710
711static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
712 struct request *next)
713{
714 unsigned short segments = blk_rq_nr_discard_segments(req);
715
716 if (segments >= queue_max_discard_segments(q))
717 goto no_merge;
718 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
719 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
720 goto no_merge;
721
722 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
723 return true;
724no_merge:
725 req_set_nomerge(q, req);
726 return false;
727}
728
729static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
730 struct request *next)
731{
732 int total_phys_segments;
733
734 if (req_gap_back_merge(req, next->bio))
735 return 0;
736
737 /*
738 * Will it become too large?
739 */
740 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
741 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
742 return 0;
743
744 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
745 if (total_phys_segments > blk_rq_get_max_segments(req))
746 return 0;
747
748 if (!blk_cgroup_mergeable(req, next->bio))
749 return 0;
750
751 if (blk_integrity_merge_rq(q, req, next) == false)
752 return 0;
753
754 if (!bio_crypt_ctx_merge_rq(req, next))
755 return 0;
756
757 /* Merge is OK... */
758 req->nr_phys_segments = total_phys_segments;
759 req->nr_integrity_segments += next->nr_integrity_segments;
760 return 1;
761}
762
763/**
764 * blk_rq_set_mixed_merge - mark a request as mixed merge
765 * @rq: request to mark as mixed merge
766 *
767 * Description:
768 * @rq is about to be mixed merged. Make sure the attributes
769 * which can be mixed are set in each bio and mark @rq as mixed
770 * merged.
771 */
772static void blk_rq_set_mixed_merge(struct request *rq)
773{
774 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK;
775 struct bio *bio;
776
777 if (rq->rq_flags & RQF_MIXED_MERGE)
778 return;
779
780 /*
781 * @rq will no longer represent mixable attributes for all the
782 * contained bios. It will just track those of the first one.
783 * Distributes the attributs to each bio.
784 */
785 for (bio = rq->bio; bio; bio = bio->bi_next) {
786 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
787 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
788 bio->bi_opf |= ff;
789 }
790 rq->rq_flags |= RQF_MIXED_MERGE;
791}
792
793static inline blk_opf_t bio_failfast(const struct bio *bio)
794{
795 if (bio->bi_opf & REQ_RAHEAD)
796 return REQ_FAILFAST_MASK;
797
798 return bio->bi_opf & REQ_FAILFAST_MASK;
799}
800
801/*
802 * After we are marked as MIXED_MERGE, any new RA bio has to be updated
803 * as failfast, and request's failfast has to be updated in case of
804 * front merge.
805 */
806static inline void blk_update_mixed_merge(struct request *req,
807 struct bio *bio, bool front_merge)
808{
809 if (req->rq_flags & RQF_MIXED_MERGE) {
810 if (bio->bi_opf & REQ_RAHEAD)
811 bio->bi_opf |= REQ_FAILFAST_MASK;
812
813 if (front_merge) {
814 req->cmd_flags &= ~REQ_FAILFAST_MASK;
815 req->cmd_flags |= bio->bi_opf & REQ_FAILFAST_MASK;
816 }
817 }
818}
819
820static void blk_account_io_merge_request(struct request *req)
821{
822 if (req->rq_flags & RQF_IO_STAT) {
823 part_stat_lock();
824 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
825 part_stat_local_dec(req->part,
826 in_flight[op_is_write(req_op(req))]);
827 part_stat_unlock();
828 }
829}
830
831static enum elv_merge blk_try_req_merge(struct request *req,
832 struct request *next)
833{
834 if (blk_discard_mergable(req))
835 return ELEVATOR_DISCARD_MERGE;
836 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
837 return ELEVATOR_BACK_MERGE;
838
839 return ELEVATOR_NO_MERGE;
840}
841
842static bool blk_atomic_write_mergeable_rq_bio(struct request *rq,
843 struct bio *bio)
844{
845 return (rq->cmd_flags & REQ_ATOMIC) == (bio->bi_opf & REQ_ATOMIC);
846}
847
848static bool blk_atomic_write_mergeable_rqs(struct request *rq,
849 struct request *next)
850{
851 return (rq->cmd_flags & REQ_ATOMIC) == (next->cmd_flags & REQ_ATOMIC);
852}
853
854/*
855 * For non-mq, this has to be called with the request spinlock acquired.
856 * For mq with scheduling, the appropriate queue wide lock should be held.
857 */
858static struct request *attempt_merge(struct request_queue *q,
859 struct request *req, struct request *next)
860{
861 if (!rq_mergeable(req) || !rq_mergeable(next))
862 return NULL;
863
864 if (req_op(req) != req_op(next))
865 return NULL;
866
867 if (req->bio->bi_write_hint != next->bio->bi_write_hint)
868 return NULL;
869 if (req->bio->bi_ioprio != next->bio->bi_ioprio)
870 return NULL;
871 if (!blk_atomic_write_mergeable_rqs(req, next))
872 return NULL;
873
874 /*
875 * If we are allowed to merge, then append bio list
876 * from next to rq and release next. merge_requests_fn
877 * will have updated segment counts, update sector
878 * counts here. Handle DISCARDs separately, as they
879 * have separate settings.
880 */
881
882 switch (blk_try_req_merge(req, next)) {
883 case ELEVATOR_DISCARD_MERGE:
884 if (!req_attempt_discard_merge(q, req, next))
885 return NULL;
886 break;
887 case ELEVATOR_BACK_MERGE:
888 if (!ll_merge_requests_fn(q, req, next))
889 return NULL;
890 break;
891 default:
892 return NULL;
893 }
894
895 /*
896 * If failfast settings disagree or any of the two is already
897 * a mixed merge, mark both as mixed before proceeding. This
898 * makes sure that all involved bios have mixable attributes
899 * set properly.
900 */
901 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
902 (req->cmd_flags & REQ_FAILFAST_MASK) !=
903 (next->cmd_flags & REQ_FAILFAST_MASK)) {
904 blk_rq_set_mixed_merge(req);
905 blk_rq_set_mixed_merge(next);
906 }
907
908 /*
909 * At this point we have either done a back merge or front merge. We
910 * need the smaller start_time_ns of the merged requests to be the
911 * current request for accounting purposes.
912 */
913 if (next->start_time_ns < req->start_time_ns)
914 req->start_time_ns = next->start_time_ns;
915
916 req->biotail->bi_next = next->bio;
917 req->biotail = next->biotail;
918
919 req->__data_len += blk_rq_bytes(next);
920
921 if (!blk_discard_mergable(req))
922 elv_merge_requests(q, req, next);
923
924 blk_crypto_rq_put_keyslot(next);
925
926 /*
927 * 'next' is going away, so update stats accordingly
928 */
929 blk_account_io_merge_request(next);
930
931 trace_block_rq_merge(next);
932
933 /*
934 * ownership of bio passed from next to req, return 'next' for
935 * the caller to free
936 */
937 next->bio = NULL;
938 return next;
939}
940
941static struct request *attempt_back_merge(struct request_queue *q,
942 struct request *rq)
943{
944 struct request *next = elv_latter_request(q, rq);
945
946 if (next)
947 return attempt_merge(q, rq, next);
948
949 return NULL;
950}
951
952static struct request *attempt_front_merge(struct request_queue *q,
953 struct request *rq)
954{
955 struct request *prev = elv_former_request(q, rq);
956
957 if (prev)
958 return attempt_merge(q, prev, rq);
959
960 return NULL;
961}
962
963/*
964 * Try to merge 'next' into 'rq'. Return true if the merge happened, false
965 * otherwise. The caller is responsible for freeing 'next' if the merge
966 * happened.
967 */
968bool blk_attempt_req_merge(struct request_queue *q, struct request *rq,
969 struct request *next)
970{
971 return attempt_merge(q, rq, next);
972}
973
974bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
975{
976 if (!rq_mergeable(rq) || !bio_mergeable(bio))
977 return false;
978
979 if (req_op(rq) != bio_op(bio))
980 return false;
981
982 if (!blk_cgroup_mergeable(rq, bio))
983 return false;
984 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
985 return false;
986 if (!bio_crypt_rq_ctx_compatible(rq, bio))
987 return false;
988 if (rq->bio->bi_write_hint != bio->bi_write_hint)
989 return false;
990 if (rq->bio->bi_ioprio != bio->bi_ioprio)
991 return false;
992 if (blk_atomic_write_mergeable_rq_bio(rq, bio) == false)
993 return false;
994
995 return true;
996}
997
998enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
999{
1000 if (blk_discard_mergable(rq))
1001 return ELEVATOR_DISCARD_MERGE;
1002 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
1003 return ELEVATOR_BACK_MERGE;
1004 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
1005 return ELEVATOR_FRONT_MERGE;
1006 return ELEVATOR_NO_MERGE;
1007}
1008
1009static void blk_account_io_merge_bio(struct request *req)
1010{
1011 if (req->rq_flags & RQF_IO_STAT) {
1012 part_stat_lock();
1013 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
1014 part_stat_unlock();
1015 }
1016}
1017
1018enum bio_merge_status bio_attempt_back_merge(struct request *req,
1019 struct bio *bio, unsigned int nr_segs)
1020{
1021 const blk_opf_t ff = bio_failfast(bio);
1022
1023 if (!ll_back_merge_fn(req, bio, nr_segs))
1024 return BIO_MERGE_FAILED;
1025
1026 trace_block_bio_backmerge(bio);
1027 rq_qos_merge(req->q, req, bio);
1028
1029 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1030 blk_rq_set_mixed_merge(req);
1031
1032 blk_update_mixed_merge(req, bio, false);
1033
1034 if (req->rq_flags & RQF_ZONE_WRITE_PLUGGING)
1035 blk_zone_write_plug_bio_merged(bio);
1036
1037 req->biotail->bi_next = bio;
1038 req->biotail = bio;
1039 req->__data_len += bio->bi_iter.bi_size;
1040
1041 bio_crypt_free_ctx(bio);
1042
1043 blk_account_io_merge_bio(req);
1044 return BIO_MERGE_OK;
1045}
1046
1047static enum bio_merge_status bio_attempt_front_merge(struct request *req,
1048 struct bio *bio, unsigned int nr_segs)
1049{
1050 const blk_opf_t ff = bio_failfast(bio);
1051
1052 /*
1053 * A front merge for writes to sequential zones of a zoned block device
1054 * can happen only if the user submitted writes out of order. Do not
1055 * merge such write to let it fail.
1056 */
1057 if (req->rq_flags & RQF_ZONE_WRITE_PLUGGING)
1058 return BIO_MERGE_FAILED;
1059
1060 if (!ll_front_merge_fn(req, bio, nr_segs))
1061 return BIO_MERGE_FAILED;
1062
1063 trace_block_bio_frontmerge(bio);
1064 rq_qos_merge(req->q, req, bio);
1065
1066 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1067 blk_rq_set_mixed_merge(req);
1068
1069 blk_update_mixed_merge(req, bio, true);
1070
1071 bio->bi_next = req->bio;
1072 req->bio = bio;
1073
1074 req->__sector = bio->bi_iter.bi_sector;
1075 req->__data_len += bio->bi_iter.bi_size;
1076
1077 bio_crypt_do_front_merge(req, bio);
1078
1079 blk_account_io_merge_bio(req);
1080 return BIO_MERGE_OK;
1081}
1082
1083static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q,
1084 struct request *req, struct bio *bio)
1085{
1086 unsigned short segments = blk_rq_nr_discard_segments(req);
1087
1088 if (segments >= queue_max_discard_segments(q))
1089 goto no_merge;
1090 if (blk_rq_sectors(req) + bio_sectors(bio) >
1091 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
1092 goto no_merge;
1093
1094 rq_qos_merge(q, req, bio);
1095
1096 req->biotail->bi_next = bio;
1097 req->biotail = bio;
1098 req->__data_len += bio->bi_iter.bi_size;
1099 req->nr_phys_segments = segments + 1;
1100
1101 blk_account_io_merge_bio(req);
1102 return BIO_MERGE_OK;
1103no_merge:
1104 req_set_nomerge(q, req);
1105 return BIO_MERGE_FAILED;
1106}
1107
1108static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q,
1109 struct request *rq,
1110 struct bio *bio,
1111 unsigned int nr_segs,
1112 bool sched_allow_merge)
1113{
1114 if (!blk_rq_merge_ok(rq, bio))
1115 return BIO_MERGE_NONE;
1116
1117 switch (blk_try_merge(rq, bio)) {
1118 case ELEVATOR_BACK_MERGE:
1119 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1120 return bio_attempt_back_merge(rq, bio, nr_segs);
1121 break;
1122 case ELEVATOR_FRONT_MERGE:
1123 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1124 return bio_attempt_front_merge(rq, bio, nr_segs);
1125 break;
1126 case ELEVATOR_DISCARD_MERGE:
1127 return bio_attempt_discard_merge(q, rq, bio);
1128 default:
1129 return BIO_MERGE_NONE;
1130 }
1131
1132 return BIO_MERGE_FAILED;
1133}
1134
1135/**
1136 * blk_attempt_plug_merge - try to merge with %current's plugged list
1137 * @q: request_queue new bio is being queued at
1138 * @bio: new bio being queued
1139 * @nr_segs: number of segments in @bio
1140 * from the passed in @q already in the plug list
1141 *
1142 * Determine whether @bio being queued on @q can be merged with the previous
1143 * request on %current's plugged list. Returns %true if merge was successful,
1144 * otherwise %false.
1145 *
1146 * Plugging coalesces IOs from the same issuer for the same purpose without
1147 * going through @q->queue_lock. As such it's more of an issuing mechanism
1148 * than scheduling, and the request, while may have elvpriv data, is not
1149 * added on the elevator at this point. In addition, we don't have
1150 * reliable access to the elevator outside queue lock. Only check basic
1151 * merging parameters without querying the elevator.
1152 *
1153 * Caller must ensure !blk_queue_nomerges(q) beforehand.
1154 */
1155bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio,
1156 unsigned int nr_segs)
1157{
1158 struct blk_plug *plug = current->plug;
1159 struct request *rq;
1160
1161 if (!plug || rq_list_empty(&plug->mq_list))
1162 return false;
1163
1164 rq_list_for_each(&plug->mq_list, rq) {
1165 if (rq->q == q) {
1166 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) ==
1167 BIO_MERGE_OK)
1168 return true;
1169 break;
1170 }
1171
1172 /*
1173 * Only keep iterating plug list for merges if we have multiple
1174 * queues
1175 */
1176 if (!plug->multiple_queues)
1177 break;
1178 }
1179 return false;
1180}
1181
1182/*
1183 * Iterate list of requests and see if we can merge this bio with any
1184 * of them.
1185 */
1186bool blk_bio_list_merge(struct request_queue *q, struct list_head *list,
1187 struct bio *bio, unsigned int nr_segs)
1188{
1189 struct request *rq;
1190 int checked = 8;
1191
1192 list_for_each_entry_reverse(rq, list, queuelist) {
1193 if (!checked--)
1194 break;
1195
1196 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) {
1197 case BIO_MERGE_NONE:
1198 continue;
1199 case BIO_MERGE_OK:
1200 return true;
1201 case BIO_MERGE_FAILED:
1202 return false;
1203 }
1204
1205 }
1206
1207 return false;
1208}
1209EXPORT_SYMBOL_GPL(blk_bio_list_merge);
1210
1211bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
1212 unsigned int nr_segs, struct request **merged_request)
1213{
1214 struct request *rq;
1215
1216 switch (elv_merge(q, &rq, bio)) {
1217 case ELEVATOR_BACK_MERGE:
1218 if (!blk_mq_sched_allow_merge(q, rq, bio))
1219 return false;
1220 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1221 return false;
1222 *merged_request = attempt_back_merge(q, rq);
1223 if (!*merged_request)
1224 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
1225 return true;
1226 case ELEVATOR_FRONT_MERGE:
1227 if (!blk_mq_sched_allow_merge(q, rq, bio))
1228 return false;
1229 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1230 return false;
1231 *merged_request = attempt_front_merge(q, rq);
1232 if (!*merged_request)
1233 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
1234 return true;
1235 case ELEVATOR_DISCARD_MERGE:
1236 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK;
1237 default:
1238 return false;
1239 }
1240}
1241EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);