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 "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}
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}