Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Functions related to segment and merge handling
4 */
5#include <linux/kernel.h>
6#include <linux/module.h>
7#include <linux/bio.h>
8#include <linux/blkdev.h>
9#include <linux/scatterlist.h>
10
11#include <trace/events/block.h>
12
13#include "blk.h"
14
15static struct bio *blk_bio_discard_split(struct request_queue *q,
16 struct bio *bio,
17 struct bio_set *bs,
18 unsigned *nsegs)
19{
20 unsigned int max_discard_sectors, granularity;
21 int alignment;
22 sector_t tmp;
23 unsigned split_sectors;
24
25 *nsegs = 1;
26
27 /* Zero-sector (unknown) and one-sector granularities are the same. */
28 granularity = max(q->limits.discard_granularity >> 9, 1U);
29
30 max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
31 max_discard_sectors -= max_discard_sectors % granularity;
32
33 if (unlikely(!max_discard_sectors)) {
34 /* XXX: warn */
35 return NULL;
36 }
37
38 if (bio_sectors(bio) <= max_discard_sectors)
39 return NULL;
40
41 split_sectors = max_discard_sectors;
42
43 /*
44 * If the next starting sector would be misaligned, stop the discard at
45 * the previous aligned sector.
46 */
47 alignment = (q->limits.discard_alignment >> 9) % granularity;
48
49 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
50 tmp = sector_div(tmp, granularity);
51
52 if (split_sectors > tmp)
53 split_sectors -= tmp;
54
55 return bio_split(bio, split_sectors, GFP_NOIO, bs);
56}
57
58static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
59 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
60{
61 *nsegs = 1;
62
63 if (!q->limits.max_write_zeroes_sectors)
64 return NULL;
65
66 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
67 return NULL;
68
69 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
70}
71
72static struct bio *blk_bio_write_same_split(struct request_queue *q,
73 struct bio *bio,
74 struct bio_set *bs,
75 unsigned *nsegs)
76{
77 *nsegs = 1;
78
79 if (!q->limits.max_write_same_sectors)
80 return NULL;
81
82 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
83 return NULL;
84
85 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
86}
87
88static inline unsigned get_max_io_size(struct request_queue *q,
89 struct bio *bio)
90{
91 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
92 unsigned mask = queue_logical_block_size(q) - 1;
93
94 /* aligned to logical block size */
95 sectors &= ~(mask >> 9);
96
97 return sectors;
98}
99
100static struct bio *blk_bio_segment_split(struct request_queue *q,
101 struct bio *bio,
102 struct bio_set *bs,
103 unsigned *segs)
104{
105 struct bio_vec bv, bvprv, *bvprvp = NULL;
106 struct bvec_iter iter;
107 unsigned seg_size = 0, nsegs = 0, sectors = 0;
108 unsigned front_seg_size = bio->bi_seg_front_size;
109 bool do_split = true;
110 struct bio *new = NULL;
111 const unsigned max_sectors = get_max_io_size(q, bio);
112
113 bio_for_each_segment(bv, bio, iter) {
114 /*
115 * If the queue doesn't support SG gaps and adding this
116 * offset would create a gap, disallow it.
117 */
118 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
119 goto split;
120
121 if (sectors + (bv.bv_len >> 9) > max_sectors) {
122 /*
123 * Consider this a new segment if we're splitting in
124 * the middle of this vector.
125 */
126 if (nsegs < queue_max_segments(q) &&
127 sectors < max_sectors) {
128 nsegs++;
129 sectors = max_sectors;
130 }
131 goto split;
132 }
133
134 if (bvprvp && blk_queue_cluster(q)) {
135 if (seg_size + bv.bv_len > queue_max_segment_size(q))
136 goto new_segment;
137 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
138 goto new_segment;
139 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
140 goto new_segment;
141
142 seg_size += bv.bv_len;
143 bvprv = bv;
144 bvprvp = &bvprv;
145 sectors += bv.bv_len >> 9;
146
147 continue;
148 }
149new_segment:
150 if (nsegs == queue_max_segments(q))
151 goto split;
152
153 if (nsegs == 1 && seg_size > front_seg_size)
154 front_seg_size = seg_size;
155
156 nsegs++;
157 bvprv = bv;
158 bvprvp = &bvprv;
159 seg_size = bv.bv_len;
160 sectors += bv.bv_len >> 9;
161
162 }
163
164 do_split = false;
165split:
166 *segs = nsegs;
167
168 if (do_split) {
169 new = bio_split(bio, sectors, GFP_NOIO, bs);
170 if (new)
171 bio = new;
172 }
173
174 if (nsegs == 1 && seg_size > front_seg_size)
175 front_seg_size = seg_size;
176 bio->bi_seg_front_size = front_seg_size;
177 if (seg_size > bio->bi_seg_back_size)
178 bio->bi_seg_back_size = seg_size;
179
180 return do_split ? new : NULL;
181}
182
183void blk_queue_split(struct request_queue *q, struct bio **bio)
184{
185 struct bio *split, *res;
186 unsigned nsegs;
187
188 switch (bio_op(*bio)) {
189 case REQ_OP_DISCARD:
190 case REQ_OP_SECURE_ERASE:
191 split = blk_bio_discard_split(q, *bio, q->bio_split, &nsegs);
192 break;
193 case REQ_OP_WRITE_ZEROES:
194 split = blk_bio_write_zeroes_split(q, *bio, q->bio_split, &nsegs);
195 break;
196 case REQ_OP_WRITE_SAME:
197 split = blk_bio_write_same_split(q, *bio, q->bio_split, &nsegs);
198 break;
199 default:
200 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
201 break;
202 }
203
204 /* physical segments can be figured out during splitting */
205 res = split ? split : *bio;
206 res->bi_phys_segments = nsegs;
207 bio_set_flag(res, BIO_SEG_VALID);
208
209 if (split) {
210 /* there isn't chance to merge the splitted bio */
211 split->bi_opf |= REQ_NOMERGE;
212
213 bio_chain(split, *bio);
214 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
215 generic_make_request(*bio);
216 *bio = split;
217 }
218}
219EXPORT_SYMBOL(blk_queue_split);
220
221static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
222 struct bio *bio,
223 bool no_sg_merge)
224{
225 struct bio_vec bv, bvprv = { NULL };
226 int cluster, prev = 0;
227 unsigned int seg_size, nr_phys_segs;
228 struct bio *fbio, *bbio;
229 struct bvec_iter iter;
230
231 if (!bio)
232 return 0;
233
234 switch (bio_op(bio)) {
235 case REQ_OP_DISCARD:
236 case REQ_OP_SECURE_ERASE:
237 case REQ_OP_WRITE_ZEROES:
238 return 0;
239 case REQ_OP_WRITE_SAME:
240 return 1;
241 }
242
243 fbio = bio;
244 cluster = blk_queue_cluster(q);
245 seg_size = 0;
246 nr_phys_segs = 0;
247 for_each_bio(bio) {
248 bio_for_each_segment(bv, bio, iter) {
249 /*
250 * If SG merging is disabled, each bio vector is
251 * a segment
252 */
253 if (no_sg_merge)
254 goto new_segment;
255
256 if (prev && cluster) {
257 if (seg_size + bv.bv_len
258 > queue_max_segment_size(q))
259 goto new_segment;
260 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
261 goto new_segment;
262 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
263 goto new_segment;
264
265 seg_size += bv.bv_len;
266 bvprv = bv;
267 continue;
268 }
269new_segment:
270 if (nr_phys_segs == 1 && seg_size >
271 fbio->bi_seg_front_size)
272 fbio->bi_seg_front_size = seg_size;
273
274 nr_phys_segs++;
275 bvprv = bv;
276 prev = 1;
277 seg_size = bv.bv_len;
278 }
279 bbio = bio;
280 }
281
282 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
283 fbio->bi_seg_front_size = seg_size;
284 if (seg_size > bbio->bi_seg_back_size)
285 bbio->bi_seg_back_size = seg_size;
286
287 return nr_phys_segs;
288}
289
290void blk_recalc_rq_segments(struct request *rq)
291{
292 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
293 &rq->q->queue_flags);
294
295 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
296 no_sg_merge);
297}
298
299void blk_recount_segments(struct request_queue *q, struct bio *bio)
300{
301 unsigned short seg_cnt;
302
303 /* estimate segment number by bi_vcnt for non-cloned bio */
304 if (bio_flagged(bio, BIO_CLONED))
305 seg_cnt = bio_segments(bio);
306 else
307 seg_cnt = bio->bi_vcnt;
308
309 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
310 (seg_cnt < queue_max_segments(q)))
311 bio->bi_phys_segments = seg_cnt;
312 else {
313 struct bio *nxt = bio->bi_next;
314
315 bio->bi_next = NULL;
316 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
317 bio->bi_next = nxt;
318 }
319
320 bio_set_flag(bio, BIO_SEG_VALID);
321}
322EXPORT_SYMBOL(blk_recount_segments);
323
324static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
325 struct bio *nxt)
326{
327 struct bio_vec end_bv = { NULL }, nxt_bv;
328
329 if (!blk_queue_cluster(q))
330 return 0;
331
332 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
333 queue_max_segment_size(q))
334 return 0;
335
336 if (!bio_has_data(bio))
337 return 1;
338
339 bio_get_last_bvec(bio, &end_bv);
340 bio_get_first_bvec(nxt, &nxt_bv);
341
342 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
343 return 0;
344
345 /*
346 * bio and nxt are contiguous in memory; check if the queue allows
347 * these two to be merged into one
348 */
349 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
350 return 1;
351
352 return 0;
353}
354
355static inline void
356__blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
357 struct scatterlist *sglist, struct bio_vec *bvprv,
358 struct scatterlist **sg, int *nsegs, int *cluster)
359{
360
361 int nbytes = bvec->bv_len;
362
363 if (*sg && *cluster) {
364 if ((*sg)->length + nbytes > queue_max_segment_size(q))
365 goto new_segment;
366
367 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
368 goto new_segment;
369 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
370 goto new_segment;
371
372 (*sg)->length += nbytes;
373 } else {
374new_segment:
375 if (!*sg)
376 *sg = sglist;
377 else {
378 /*
379 * If the driver previously mapped a shorter
380 * list, we could see a termination bit
381 * prematurely unless it fully inits the sg
382 * table on each mapping. We KNOW that there
383 * must be more entries here or the driver
384 * would be buggy, so force clear the
385 * termination bit to avoid doing a full
386 * sg_init_table() in drivers for each command.
387 */
388 sg_unmark_end(*sg);
389 *sg = sg_next(*sg);
390 }
391
392 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
393 (*nsegs)++;
394 }
395 *bvprv = *bvec;
396}
397
398static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
399 struct scatterlist *sglist, struct scatterlist **sg)
400{
401 *sg = sglist;
402 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
403 return 1;
404}
405
406static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
407 struct scatterlist *sglist,
408 struct scatterlist **sg)
409{
410 struct bio_vec bvec, bvprv = { NULL };
411 struct bvec_iter iter;
412 int cluster = blk_queue_cluster(q), nsegs = 0;
413
414 for_each_bio(bio)
415 bio_for_each_segment(bvec, bio, iter)
416 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
417 &nsegs, &cluster);
418
419 return nsegs;
420}
421
422/*
423 * map a request to scatterlist, return number of sg entries setup. Caller
424 * must make sure sg can hold rq->nr_phys_segments entries
425 */
426int blk_rq_map_sg(struct request_queue *q, struct request *rq,
427 struct scatterlist *sglist)
428{
429 struct scatterlist *sg = NULL;
430 int nsegs = 0;
431
432 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
433 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
434 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
435 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
436 else if (rq->bio)
437 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
438
439 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
440 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
441 unsigned int pad_len =
442 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
443
444 sg->length += pad_len;
445 rq->extra_len += pad_len;
446 }
447
448 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
449 if (op_is_write(req_op(rq)))
450 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
451
452 sg_unmark_end(sg);
453 sg = sg_next(sg);
454 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
455 q->dma_drain_size,
456 ((unsigned long)q->dma_drain_buffer) &
457 (PAGE_SIZE - 1));
458 nsegs++;
459 rq->extra_len += q->dma_drain_size;
460 }
461
462 if (sg)
463 sg_mark_end(sg);
464
465 /*
466 * Something must have been wrong if the figured number of
467 * segment is bigger than number of req's physical segments
468 */
469 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
470
471 return nsegs;
472}
473EXPORT_SYMBOL(blk_rq_map_sg);
474
475static inline int ll_new_hw_segment(struct request_queue *q,
476 struct request *req,
477 struct bio *bio)
478{
479 int nr_phys_segs = bio_phys_segments(q, bio);
480
481 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
482 goto no_merge;
483
484 if (blk_integrity_merge_bio(q, req, bio) == false)
485 goto no_merge;
486
487 /*
488 * This will form the start of a new hw segment. Bump both
489 * counters.
490 */
491 req->nr_phys_segments += nr_phys_segs;
492 return 1;
493
494no_merge:
495 req_set_nomerge(q, req);
496 return 0;
497}
498
499int ll_back_merge_fn(struct request_queue *q, struct request *req,
500 struct bio *bio)
501{
502 if (req_gap_back_merge(req, bio))
503 return 0;
504 if (blk_integrity_rq(req) &&
505 integrity_req_gap_back_merge(req, bio))
506 return 0;
507 if (blk_rq_sectors(req) + bio_sectors(bio) >
508 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
509 req_set_nomerge(q, req);
510 return 0;
511 }
512 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
513 blk_recount_segments(q, req->biotail);
514 if (!bio_flagged(bio, BIO_SEG_VALID))
515 blk_recount_segments(q, bio);
516
517 return ll_new_hw_segment(q, req, bio);
518}
519
520int ll_front_merge_fn(struct request_queue *q, struct request *req,
521 struct bio *bio)
522{
523
524 if (req_gap_front_merge(req, bio))
525 return 0;
526 if (blk_integrity_rq(req) &&
527 integrity_req_gap_front_merge(req, bio))
528 return 0;
529 if (blk_rq_sectors(req) + bio_sectors(bio) >
530 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
531 req_set_nomerge(q, req);
532 return 0;
533 }
534 if (!bio_flagged(bio, BIO_SEG_VALID))
535 blk_recount_segments(q, bio);
536 if (!bio_flagged(req->bio, BIO_SEG_VALID))
537 blk_recount_segments(q, req->bio);
538
539 return ll_new_hw_segment(q, req, bio);
540}
541
542/*
543 * blk-mq uses req->special to carry normal driver per-request payload, it
544 * does not indicate a prepared command that we cannot merge with.
545 */
546static bool req_no_special_merge(struct request *req)
547{
548 struct request_queue *q = req->q;
549
550 return !q->mq_ops && req->special;
551}
552
553static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
554 struct request *next)
555{
556 unsigned short segments = blk_rq_nr_discard_segments(req);
557
558 if (segments >= queue_max_discard_segments(q))
559 goto no_merge;
560 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
561 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
562 goto no_merge;
563
564 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
565 return true;
566no_merge:
567 req_set_nomerge(q, req);
568 return false;
569}
570
571static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
572 struct request *next)
573{
574 int total_phys_segments;
575 unsigned int seg_size =
576 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
577
578 /*
579 * First check if the either of the requests are re-queued
580 * requests. Can't merge them if they are.
581 */
582 if (req_no_special_merge(req) || req_no_special_merge(next))
583 return 0;
584
585 if (req_gap_back_merge(req, next->bio))
586 return 0;
587
588 /*
589 * Will it become too large?
590 */
591 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
592 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
593 return 0;
594
595 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
596 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
597 if (req->nr_phys_segments == 1)
598 req->bio->bi_seg_front_size = seg_size;
599 if (next->nr_phys_segments == 1)
600 next->biotail->bi_seg_back_size = seg_size;
601 total_phys_segments--;
602 }
603
604 if (total_phys_segments > queue_max_segments(q))
605 return 0;
606
607 if (blk_integrity_merge_rq(q, req, next) == false)
608 return 0;
609
610 /* Merge is OK... */
611 req->nr_phys_segments = total_phys_segments;
612 return 1;
613}
614
615/**
616 * blk_rq_set_mixed_merge - mark a request as mixed merge
617 * @rq: request to mark as mixed merge
618 *
619 * Description:
620 * @rq is about to be mixed merged. Make sure the attributes
621 * which can be mixed are set in each bio and mark @rq as mixed
622 * merged.
623 */
624void blk_rq_set_mixed_merge(struct request *rq)
625{
626 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
627 struct bio *bio;
628
629 if (rq->rq_flags & RQF_MIXED_MERGE)
630 return;
631
632 /*
633 * @rq will no longer represent mixable attributes for all the
634 * contained bios. It will just track those of the first one.
635 * Distributes the attributs to each bio.
636 */
637 for (bio = rq->bio; bio; bio = bio->bi_next) {
638 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
639 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
640 bio->bi_opf |= ff;
641 }
642 rq->rq_flags |= RQF_MIXED_MERGE;
643}
644
645static void blk_account_io_merge(struct request *req)
646{
647 if (blk_do_io_stat(req)) {
648 struct hd_struct *part;
649 int cpu;
650
651 cpu = part_stat_lock();
652 part = req->part;
653
654 part_round_stats(req->q, cpu, part);
655 part_dec_in_flight(req->q, part, rq_data_dir(req));
656
657 hd_struct_put(part);
658 part_stat_unlock();
659 }
660}
661
662/*
663 * For non-mq, this has to be called with the request spinlock acquired.
664 * For mq with scheduling, the appropriate queue wide lock should be held.
665 */
666static struct request *attempt_merge(struct request_queue *q,
667 struct request *req, struct request *next)
668{
669 if (!q->mq_ops)
670 lockdep_assert_held(q->queue_lock);
671
672 if (!rq_mergeable(req) || !rq_mergeable(next))
673 return NULL;
674
675 if (req_op(req) != req_op(next))
676 return NULL;
677
678 /*
679 * not contiguous
680 */
681 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
682 return NULL;
683
684 if (rq_data_dir(req) != rq_data_dir(next)
685 || req->rq_disk != next->rq_disk
686 || req_no_special_merge(next))
687 return NULL;
688
689 if (req_op(req) == REQ_OP_WRITE_SAME &&
690 !blk_write_same_mergeable(req->bio, next->bio))
691 return NULL;
692
693 /*
694 * Don't allow merge of different write hints, or for a hint with
695 * non-hint IO.
696 */
697 if (req->write_hint != next->write_hint)
698 return NULL;
699
700 /*
701 * If we are allowed to merge, then append bio list
702 * from next to rq and release next. merge_requests_fn
703 * will have updated segment counts, update sector
704 * counts here. Handle DISCARDs separately, as they
705 * have separate settings.
706 */
707 if (req_op(req) == REQ_OP_DISCARD) {
708 if (!req_attempt_discard_merge(q, req, next))
709 return NULL;
710 } else if (!ll_merge_requests_fn(q, req, next))
711 return NULL;
712
713 /*
714 * If failfast settings disagree or any of the two is already
715 * a mixed merge, mark both as mixed before proceeding. This
716 * makes sure that all involved bios have mixable attributes
717 * set properly.
718 */
719 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
720 (req->cmd_flags & REQ_FAILFAST_MASK) !=
721 (next->cmd_flags & REQ_FAILFAST_MASK)) {
722 blk_rq_set_mixed_merge(req);
723 blk_rq_set_mixed_merge(next);
724 }
725
726 /*
727 * At this point we have either done a back merge
728 * or front merge. We need the smaller start_time of
729 * the merged requests to be the current request
730 * for accounting purposes.
731 */
732 if (time_after(req->start_time, next->start_time))
733 req->start_time = next->start_time;
734
735 req->biotail->bi_next = next->bio;
736 req->biotail = next->biotail;
737
738 req->__data_len += blk_rq_bytes(next);
739
740 if (req_op(req) != REQ_OP_DISCARD)
741 elv_merge_requests(q, req, next);
742
743 /*
744 * 'next' is going away, so update stats accordingly
745 */
746 blk_account_io_merge(next);
747
748 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
749 if (blk_rq_cpu_valid(next))
750 req->cpu = next->cpu;
751
752 /*
753 * ownership of bio passed from next to req, return 'next' for
754 * the caller to free
755 */
756 next->bio = NULL;
757 return next;
758}
759
760struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
761{
762 struct request *next = elv_latter_request(q, rq);
763
764 if (next)
765 return attempt_merge(q, rq, next);
766
767 return NULL;
768}
769
770struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
771{
772 struct request *prev = elv_former_request(q, rq);
773
774 if (prev)
775 return attempt_merge(q, prev, rq);
776
777 return NULL;
778}
779
780int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
781 struct request *next)
782{
783 struct elevator_queue *e = q->elevator;
784 struct request *free;
785
786 if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn)
787 if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next))
788 return 0;
789
790 free = attempt_merge(q, rq, next);
791 if (free) {
792 __blk_put_request(q, free);
793 return 1;
794 }
795
796 return 0;
797}
798
799bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
800{
801 if (!rq_mergeable(rq) || !bio_mergeable(bio))
802 return false;
803
804 if (req_op(rq) != bio_op(bio))
805 return false;
806
807 /* different data direction or already started, don't merge */
808 if (bio_data_dir(bio) != rq_data_dir(rq))
809 return false;
810
811 /* must be same device and not a special request */
812 if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq))
813 return false;
814
815 /* only merge integrity protected bio into ditto rq */
816 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
817 return false;
818
819 /* must be using the same buffer */
820 if (req_op(rq) == REQ_OP_WRITE_SAME &&
821 !blk_write_same_mergeable(rq->bio, bio))
822 return false;
823
824 /*
825 * Don't allow merge of different write hints, or for a hint with
826 * non-hint IO.
827 */
828 if (rq->write_hint != bio->bi_write_hint)
829 return false;
830
831 return true;
832}
833
834enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
835{
836 if (req_op(rq) == REQ_OP_DISCARD &&
837 queue_max_discard_segments(rq->q) > 1)
838 return ELEVATOR_DISCARD_MERGE;
839 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
840 return ELEVATOR_BACK_MERGE;
841 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
842 return ELEVATOR_FRONT_MERGE;
843 return ELEVATOR_NO_MERGE;
844}
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
12static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
13 struct bio *bio)
14{
15 struct bio_vec bv, bvprv = { NULL };
16 int cluster, high, highprv = 1;
17 unsigned int seg_size, nr_phys_segs;
18 struct bio *fbio, *bbio;
19 struct bvec_iter iter;
20
21 if (!bio)
22 return 0;
23
24 /*
25 * This should probably be returning 0, but blk_add_request_payload()
26 * (Christoph!!!!)
27 */
28 if (bio->bi_rw & REQ_DISCARD)
29 return 1;
30
31 if (bio->bi_rw & REQ_WRITE_SAME)
32 return 1;
33
34 fbio = bio;
35 cluster = blk_queue_cluster(q);
36 seg_size = 0;
37 nr_phys_segs = 0;
38 for_each_bio(bio) {
39 bio_for_each_segment(bv, bio, iter) {
40 /*
41 * the trick here is making sure that a high page is
42 * never considered part of another segment, since that
43 * might change with the bounce page.
44 */
45 high = page_to_pfn(bv.bv_page) > queue_bounce_pfn(q);
46 if (!high && !highprv && cluster) {
47 if (seg_size + bv.bv_len
48 > queue_max_segment_size(q))
49 goto new_segment;
50 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
51 goto new_segment;
52 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
53 goto new_segment;
54
55 seg_size += bv.bv_len;
56 bvprv = bv;
57 continue;
58 }
59new_segment:
60 if (nr_phys_segs == 1 && seg_size >
61 fbio->bi_seg_front_size)
62 fbio->bi_seg_front_size = seg_size;
63
64 nr_phys_segs++;
65 bvprv = bv;
66 seg_size = bv.bv_len;
67 highprv = high;
68 }
69 bbio = bio;
70 }
71
72 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
73 fbio->bi_seg_front_size = seg_size;
74 if (seg_size > bbio->bi_seg_back_size)
75 bbio->bi_seg_back_size = seg_size;
76
77 return nr_phys_segs;
78}
79
80void blk_recalc_rq_segments(struct request *rq)
81{
82 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
83}
84
85void blk_recount_segments(struct request_queue *q, struct bio *bio)
86{
87 struct bio *nxt = bio->bi_next;
88
89 bio->bi_next = NULL;
90 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
91 bio->bi_next = nxt;
92 bio->bi_flags |= (1 << BIO_SEG_VALID);
93}
94EXPORT_SYMBOL(blk_recount_segments);
95
96static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
97 struct bio *nxt)
98{
99 struct bio_vec end_bv = { NULL }, nxt_bv;
100 struct bvec_iter iter;
101
102 if (!blk_queue_cluster(q))
103 return 0;
104
105 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
106 queue_max_segment_size(q))
107 return 0;
108
109 if (!bio_has_data(bio))
110 return 1;
111
112 bio_for_each_segment(end_bv, bio, iter)
113 if (end_bv.bv_len == iter.bi_size)
114 break;
115
116 nxt_bv = bio_iovec(nxt);
117
118 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
119 return 0;
120
121 /*
122 * bio and nxt are contiguous in memory; check if the queue allows
123 * these two to be merged into one
124 */
125 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
126 return 1;
127
128 return 0;
129}
130
131static inline void
132__blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
133 struct scatterlist *sglist, struct bio_vec *bvprv,
134 struct scatterlist **sg, int *nsegs, int *cluster)
135{
136
137 int nbytes = bvec->bv_len;
138
139 if (*sg && *cluster) {
140 if ((*sg)->length + nbytes > queue_max_segment_size(q))
141 goto new_segment;
142
143 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
144 goto new_segment;
145 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
146 goto new_segment;
147
148 (*sg)->length += nbytes;
149 } else {
150new_segment:
151 if (!*sg)
152 *sg = sglist;
153 else {
154 /*
155 * If the driver previously mapped a shorter
156 * list, we could see a termination bit
157 * prematurely unless it fully inits the sg
158 * table on each mapping. We KNOW that there
159 * must be more entries here or the driver
160 * would be buggy, so force clear the
161 * termination bit to avoid doing a full
162 * sg_init_table() in drivers for each command.
163 */
164 sg_unmark_end(*sg);
165 *sg = sg_next(*sg);
166 }
167
168 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
169 (*nsegs)++;
170 }
171 *bvprv = *bvec;
172}
173
174static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
175 struct scatterlist *sglist,
176 struct scatterlist **sg)
177{
178 struct bio_vec bvec, bvprv = { NULL };
179 struct bvec_iter iter;
180 int nsegs, cluster;
181
182 nsegs = 0;
183 cluster = blk_queue_cluster(q);
184
185 if (bio->bi_rw & REQ_DISCARD) {
186 /*
187 * This is a hack - drivers should be neither modifying the
188 * biovec, nor relying on bi_vcnt - but because of
189 * blk_add_request_payload(), a discard bio may or may not have
190 * a payload we need to set up here (thank you Christoph) and
191 * bi_vcnt is really the only way of telling if we need to.
192 */
193
194 if (bio->bi_vcnt)
195 goto single_segment;
196
197 return 0;
198 }
199
200 if (bio->bi_rw & REQ_WRITE_SAME) {
201single_segment:
202 *sg = sglist;
203 bvec = bio_iovec(bio);
204 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
205 return 1;
206 }
207
208 for_each_bio(bio)
209 bio_for_each_segment(bvec, bio, iter)
210 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
211 &nsegs, &cluster);
212
213 return nsegs;
214}
215
216/*
217 * map a request to scatterlist, return number of sg entries setup. Caller
218 * must make sure sg can hold rq->nr_phys_segments entries
219 */
220int blk_rq_map_sg(struct request_queue *q, struct request *rq,
221 struct scatterlist *sglist)
222{
223 struct scatterlist *sg = NULL;
224 int nsegs = 0;
225
226 if (rq->bio)
227 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
228
229 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
230 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
231 unsigned int pad_len =
232 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
233
234 sg->length += pad_len;
235 rq->extra_len += pad_len;
236 }
237
238 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
239 if (rq->cmd_flags & REQ_WRITE)
240 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
241
242 sg->page_link &= ~0x02;
243 sg = sg_next(sg);
244 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
245 q->dma_drain_size,
246 ((unsigned long)q->dma_drain_buffer) &
247 (PAGE_SIZE - 1));
248 nsegs++;
249 rq->extra_len += q->dma_drain_size;
250 }
251
252 if (sg)
253 sg_mark_end(sg);
254
255 return nsegs;
256}
257EXPORT_SYMBOL(blk_rq_map_sg);
258
259/**
260 * blk_bio_map_sg - map a bio to a scatterlist
261 * @q: request_queue in question
262 * @bio: bio being mapped
263 * @sglist: scatterlist being mapped
264 *
265 * Note:
266 * Caller must make sure sg can hold bio->bi_phys_segments entries
267 *
268 * Will return the number of sg entries setup
269 */
270int blk_bio_map_sg(struct request_queue *q, struct bio *bio,
271 struct scatterlist *sglist)
272{
273 struct scatterlist *sg = NULL;
274 int nsegs;
275 struct bio *next = bio->bi_next;
276 bio->bi_next = NULL;
277
278 nsegs = __blk_bios_map_sg(q, bio, sglist, &sg);
279 bio->bi_next = next;
280 if (sg)
281 sg_mark_end(sg);
282
283 BUG_ON(bio->bi_phys_segments && nsegs > bio->bi_phys_segments);
284 return nsegs;
285}
286EXPORT_SYMBOL(blk_bio_map_sg);
287
288static inline int ll_new_hw_segment(struct request_queue *q,
289 struct request *req,
290 struct bio *bio)
291{
292 int nr_phys_segs = bio_phys_segments(q, bio);
293
294 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
295 goto no_merge;
296
297 if (bio_integrity(bio) && blk_integrity_merge_bio(q, req, bio))
298 goto no_merge;
299
300 /*
301 * This will form the start of a new hw segment. Bump both
302 * counters.
303 */
304 req->nr_phys_segments += nr_phys_segs;
305 return 1;
306
307no_merge:
308 req->cmd_flags |= REQ_NOMERGE;
309 if (req == q->last_merge)
310 q->last_merge = NULL;
311 return 0;
312}
313
314int ll_back_merge_fn(struct request_queue *q, struct request *req,
315 struct bio *bio)
316{
317 if (blk_rq_sectors(req) + bio_sectors(bio) >
318 blk_rq_get_max_sectors(req)) {
319 req->cmd_flags |= REQ_NOMERGE;
320 if (req == q->last_merge)
321 q->last_merge = NULL;
322 return 0;
323 }
324 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
325 blk_recount_segments(q, req->biotail);
326 if (!bio_flagged(bio, BIO_SEG_VALID))
327 blk_recount_segments(q, bio);
328
329 return ll_new_hw_segment(q, req, bio);
330}
331
332int ll_front_merge_fn(struct request_queue *q, struct request *req,
333 struct bio *bio)
334{
335 if (blk_rq_sectors(req) + bio_sectors(bio) >
336 blk_rq_get_max_sectors(req)) {
337 req->cmd_flags |= REQ_NOMERGE;
338 if (req == q->last_merge)
339 q->last_merge = NULL;
340 return 0;
341 }
342 if (!bio_flagged(bio, BIO_SEG_VALID))
343 blk_recount_segments(q, bio);
344 if (!bio_flagged(req->bio, BIO_SEG_VALID))
345 blk_recount_segments(q, req->bio);
346
347 return ll_new_hw_segment(q, req, bio);
348}
349
350/*
351 * blk-mq uses req->special to carry normal driver per-request payload, it
352 * does not indicate a prepared command that we cannot merge with.
353 */
354static bool req_no_special_merge(struct request *req)
355{
356 struct request_queue *q = req->q;
357
358 return !q->mq_ops && req->special;
359}
360
361static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
362 struct request *next)
363{
364 int total_phys_segments;
365 unsigned int seg_size =
366 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
367
368 /*
369 * First check if the either of the requests are re-queued
370 * requests. Can't merge them if they are.
371 */
372 if (req_no_special_merge(req) || req_no_special_merge(next))
373 return 0;
374
375 /*
376 * Will it become too large?
377 */
378 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
379 blk_rq_get_max_sectors(req))
380 return 0;
381
382 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
383 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
384 if (req->nr_phys_segments == 1)
385 req->bio->bi_seg_front_size = seg_size;
386 if (next->nr_phys_segments == 1)
387 next->biotail->bi_seg_back_size = seg_size;
388 total_phys_segments--;
389 }
390
391 if (total_phys_segments > queue_max_segments(q))
392 return 0;
393
394 if (blk_integrity_rq(req) && blk_integrity_merge_rq(q, req, next))
395 return 0;
396
397 /* Merge is OK... */
398 req->nr_phys_segments = total_phys_segments;
399 return 1;
400}
401
402/**
403 * blk_rq_set_mixed_merge - mark a request as mixed merge
404 * @rq: request to mark as mixed merge
405 *
406 * Description:
407 * @rq is about to be mixed merged. Make sure the attributes
408 * which can be mixed are set in each bio and mark @rq as mixed
409 * merged.
410 */
411void blk_rq_set_mixed_merge(struct request *rq)
412{
413 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
414 struct bio *bio;
415
416 if (rq->cmd_flags & REQ_MIXED_MERGE)
417 return;
418
419 /*
420 * @rq will no longer represent mixable attributes for all the
421 * contained bios. It will just track those of the first one.
422 * Distributes the attributs to each bio.
423 */
424 for (bio = rq->bio; bio; bio = bio->bi_next) {
425 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
426 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
427 bio->bi_rw |= ff;
428 }
429 rq->cmd_flags |= REQ_MIXED_MERGE;
430}
431
432static void blk_account_io_merge(struct request *req)
433{
434 if (blk_do_io_stat(req)) {
435 struct hd_struct *part;
436 int cpu;
437
438 cpu = part_stat_lock();
439 part = req->part;
440
441 part_round_stats(cpu, part);
442 part_dec_in_flight(part, rq_data_dir(req));
443
444 hd_struct_put(part);
445 part_stat_unlock();
446 }
447}
448
449/*
450 * Has to be called with the request spinlock acquired
451 */
452static int attempt_merge(struct request_queue *q, struct request *req,
453 struct request *next)
454{
455 if (!rq_mergeable(req) || !rq_mergeable(next))
456 return 0;
457
458 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
459 return 0;
460
461 /*
462 * not contiguous
463 */
464 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
465 return 0;
466
467 if (rq_data_dir(req) != rq_data_dir(next)
468 || req->rq_disk != next->rq_disk
469 || req_no_special_merge(next))
470 return 0;
471
472 if (req->cmd_flags & REQ_WRITE_SAME &&
473 !blk_write_same_mergeable(req->bio, next->bio))
474 return 0;
475
476 /*
477 * If we are allowed to merge, then append bio list
478 * from next to rq and release next. merge_requests_fn
479 * will have updated segment counts, update sector
480 * counts here.
481 */
482 if (!ll_merge_requests_fn(q, req, next))
483 return 0;
484
485 /*
486 * If failfast settings disagree or any of the two is already
487 * a mixed merge, mark both as mixed before proceeding. This
488 * makes sure that all involved bios have mixable attributes
489 * set properly.
490 */
491 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
492 (req->cmd_flags & REQ_FAILFAST_MASK) !=
493 (next->cmd_flags & REQ_FAILFAST_MASK)) {
494 blk_rq_set_mixed_merge(req);
495 blk_rq_set_mixed_merge(next);
496 }
497
498 /*
499 * At this point we have either done a back merge
500 * or front merge. We need the smaller start_time of
501 * the merged requests to be the current request
502 * for accounting purposes.
503 */
504 if (time_after(req->start_time, next->start_time))
505 req->start_time = next->start_time;
506
507 req->biotail->bi_next = next->bio;
508 req->biotail = next->biotail;
509
510 req->__data_len += blk_rq_bytes(next);
511
512 elv_merge_requests(q, req, next);
513
514 /*
515 * 'next' is going away, so update stats accordingly
516 */
517 blk_account_io_merge(next);
518
519 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
520 if (blk_rq_cpu_valid(next))
521 req->cpu = next->cpu;
522
523 /* owner-ship of bio passed from next to req */
524 next->bio = NULL;
525 __blk_put_request(q, next);
526 return 1;
527}
528
529int attempt_back_merge(struct request_queue *q, struct request *rq)
530{
531 struct request *next = elv_latter_request(q, rq);
532
533 if (next)
534 return attempt_merge(q, rq, next);
535
536 return 0;
537}
538
539int attempt_front_merge(struct request_queue *q, struct request *rq)
540{
541 struct request *prev = elv_former_request(q, rq);
542
543 if (prev)
544 return attempt_merge(q, prev, rq);
545
546 return 0;
547}
548
549int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
550 struct request *next)
551{
552 return attempt_merge(q, rq, next);
553}
554
555bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
556{
557 if (!rq_mergeable(rq) || !bio_mergeable(bio))
558 return false;
559
560 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
561 return false;
562
563 /* different data direction or already started, don't merge */
564 if (bio_data_dir(bio) != rq_data_dir(rq))
565 return false;
566
567 /* must be same device and not a special request */
568 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
569 return false;
570
571 /* only merge integrity protected bio into ditto rq */
572 if (bio_integrity(bio) != blk_integrity_rq(rq))
573 return false;
574
575 /* must be using the same buffer */
576 if (rq->cmd_flags & REQ_WRITE_SAME &&
577 !blk_write_same_mergeable(rq->bio, bio))
578 return false;
579
580 return true;
581}
582
583int blk_try_merge(struct request *rq, struct bio *bio)
584{
585 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
586 return ELEVATOR_BACK_MERGE;
587 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
588 return ELEVATOR_FRONT_MERGE;
589 return ELEVATOR_NO_MERGE;
590}