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/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_split_discard(struct bio *bio,
109 const struct queue_limits *lim,
110 unsigned *nsegs, struct bio_set *bs)
111{
112 unsigned int max_discard_sectors, granularity;
113 sector_t tmp;
114 unsigned split_sectors;
115
116 *nsegs = 1;
117
118 /* Zero-sector (unknown) and one-sector granularities are the same. */
119 granularity = max(lim->discard_granularity >> 9, 1U);
120
121 max_discard_sectors =
122 min(lim->max_discard_sectors, bio_allowed_max_sectors(lim));
123 max_discard_sectors -= max_discard_sectors % granularity;
124
125 if (unlikely(!max_discard_sectors)) {
126 /* XXX: warn */
127 return NULL;
128 }
129
130 if (bio_sectors(bio) <= max_discard_sectors)
131 return NULL;
132
133 split_sectors = max_discard_sectors;
134
135 /*
136 * If the next starting sector would be misaligned, stop the discard at
137 * the previous aligned sector.
138 */
139 tmp = bio->bi_iter.bi_sector + split_sectors -
140 ((lim->discard_alignment >> 9) % granularity);
141 tmp = sector_div(tmp, granularity);
142
143 if (split_sectors > tmp)
144 split_sectors -= tmp;
145
146 return bio_split(bio, split_sectors, GFP_NOIO, bs);
147}
148
149static struct bio *bio_split_write_zeroes(struct bio *bio,
150 const struct queue_limits *lim,
151 unsigned *nsegs, struct bio_set *bs)
152{
153 *nsegs = 0;
154 if (!lim->max_write_zeroes_sectors)
155 return NULL;
156 if (bio_sectors(bio) <= lim->max_write_zeroes_sectors)
157 return NULL;
158 return bio_split(bio, lim->max_write_zeroes_sectors, GFP_NOIO, bs);
159}
160
161/*
162 * Return the maximum number of sectors from the start of a bio that may be
163 * submitted as a single request to a block device. If enough sectors remain,
164 * align the end to the physical block size. Otherwise align the end to the
165 * logical block size. This approach minimizes the number of non-aligned
166 * requests that are submitted to a block device if the start of a bio is not
167 * aligned to a physical block boundary.
168 */
169static inline unsigned get_max_io_size(struct bio *bio,
170 const struct queue_limits *lim)
171{
172 unsigned pbs = lim->physical_block_size >> SECTOR_SHIFT;
173 unsigned lbs = lim->logical_block_size >> SECTOR_SHIFT;
174 unsigned max_sectors = lim->max_sectors, start, end;
175
176 if (lim->chunk_sectors) {
177 max_sectors = min(max_sectors,
178 blk_chunk_sectors_left(bio->bi_iter.bi_sector,
179 lim->chunk_sectors));
180 }
181
182 start = bio->bi_iter.bi_sector & (pbs - 1);
183 end = (start + max_sectors) & ~(pbs - 1);
184 if (end > start)
185 return end - start;
186 return max_sectors & ~(lbs - 1);
187}
188
189/**
190 * get_max_segment_size() - maximum number of bytes to add as a single segment
191 * @lim: Request queue limits.
192 * @start_page: See below.
193 * @offset: Offset from @start_page where to add a segment.
194 *
195 * Returns the maximum number of bytes that can be added as a single segment.
196 */
197static inline unsigned get_max_segment_size(const struct queue_limits *lim,
198 struct page *start_page, unsigned long offset)
199{
200 unsigned long mask = lim->seg_boundary_mask;
201
202 offset = mask & (page_to_phys(start_page) + offset);
203
204 /*
205 * Prevent an overflow if mask = ULONG_MAX and offset = 0 by adding 1
206 * after having calculated the minimum.
207 */
208 return min(mask - offset, (unsigned long)lim->max_segment_size - 1) + 1;
209}
210
211/**
212 * bvec_split_segs - verify whether or not a bvec should be split in the middle
213 * @lim: [in] queue limits to split based on
214 * @bv: [in] bvec to examine
215 * @nsegs: [in,out] Number of segments in the bio being built. Incremented
216 * by the number of segments from @bv that may be appended to that
217 * bio without exceeding @max_segs
218 * @bytes: [in,out] Number of bytes in the bio being built. Incremented
219 * by the number of bytes from @bv that may be appended to that
220 * bio without exceeding @max_bytes
221 * @max_segs: [in] upper bound for *@nsegs
222 * @max_bytes: [in] upper bound for *@bytes
223 *
224 * When splitting a bio, it can happen that a bvec is encountered that is too
225 * big to fit in a single segment and hence that it has to be split in the
226 * middle. This function verifies whether or not that should happen. The value
227 * %true is returned if and only if appending the entire @bv to a bio with
228 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
229 * the block driver.
230 */
231static bool bvec_split_segs(const struct queue_limits *lim,
232 const struct bio_vec *bv, unsigned *nsegs, unsigned *bytes,
233 unsigned max_segs, unsigned max_bytes)
234{
235 unsigned max_len = min(max_bytes, UINT_MAX) - *bytes;
236 unsigned len = min(bv->bv_len, max_len);
237 unsigned total_len = 0;
238 unsigned seg_size = 0;
239
240 while (len && *nsegs < max_segs) {
241 seg_size = get_max_segment_size(lim, bv->bv_page,
242 bv->bv_offset + total_len);
243 seg_size = min(seg_size, len);
244
245 (*nsegs)++;
246 total_len += seg_size;
247 len -= seg_size;
248
249 if ((bv->bv_offset + total_len) & lim->virt_boundary_mask)
250 break;
251 }
252
253 *bytes += total_len;
254
255 /* tell the caller to split the bvec if it is too big to fit */
256 return len > 0 || bv->bv_len > max_len;
257}
258
259/**
260 * bio_split_rw - split a bio in two bios
261 * @bio: [in] bio to be split
262 * @lim: [in] queue limits to split based on
263 * @segs: [out] number of segments in the bio with the first half of the sectors
264 * @bs: [in] bio set to allocate the clone from
265 * @max_bytes: [in] maximum number of bytes per bio
266 *
267 * Clone @bio, update the bi_iter of the clone to represent the first sectors
268 * of @bio and update @bio->bi_iter to represent the remaining sectors. The
269 * following is guaranteed for the cloned bio:
270 * - That it has at most @max_bytes worth of data
271 * - That it has at most queue_max_segments(@q) segments.
272 *
273 * Except for discard requests the cloned bio will point at the bi_io_vec of
274 * the original bio. It is the responsibility of the caller to ensure that the
275 * original bio is not freed before the cloned bio. The caller is also
276 * responsible for ensuring that @bs is only destroyed after processing of the
277 * split bio has finished.
278 */
279static struct bio *bio_split_rw(struct bio *bio, const struct queue_limits *lim,
280 unsigned *segs, struct bio_set *bs, unsigned max_bytes)
281{
282 struct bio_vec bv, bvprv, *bvprvp = NULL;
283 struct bvec_iter iter;
284 unsigned nsegs = 0, bytes = 0;
285
286 bio_for_each_bvec(bv, bio, iter) {
287 /*
288 * If the queue doesn't support SG gaps and adding this
289 * offset would create a gap, disallow it.
290 */
291 if (bvprvp && bvec_gap_to_prev(lim, bvprvp, bv.bv_offset))
292 goto split;
293
294 if (nsegs < lim->max_segments &&
295 bytes + bv.bv_len <= max_bytes &&
296 bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
297 nsegs++;
298 bytes += bv.bv_len;
299 } else {
300 if (bvec_split_segs(lim, &bv, &nsegs, &bytes,
301 lim->max_segments, max_bytes))
302 goto split;
303 }
304
305 bvprv = bv;
306 bvprvp = &bvprv;
307 }
308
309 *segs = nsegs;
310 return NULL;
311split:
312 /*
313 * We can't sanely support splitting for a REQ_NOWAIT bio. End it
314 * with EAGAIN if splitting is required and return an error pointer.
315 */
316 if (bio->bi_opf & REQ_NOWAIT) {
317 bio->bi_status = BLK_STS_AGAIN;
318 bio_endio(bio);
319 return ERR_PTR(-EAGAIN);
320 }
321
322 *segs = nsegs;
323
324 /*
325 * Individual bvecs might not be logical block aligned. Round down the
326 * split size so that each bio is properly block size aligned, even if
327 * we do not use the full hardware limits.
328 */
329 bytes = ALIGN_DOWN(bytes, lim->logical_block_size);
330
331 /*
332 * Bio splitting may cause subtle trouble such as hang when doing sync
333 * iopoll in direct IO routine. Given performance gain of iopoll for
334 * big IO can be trival, disable iopoll when split needed.
335 */
336 bio_clear_polled(bio);
337 return bio_split(bio, bytes >> SECTOR_SHIFT, GFP_NOIO, bs);
338}
339
340/**
341 * __bio_split_to_limits - split a bio to fit the queue limits
342 * @bio: bio to be split
343 * @lim: queue limits to split based on
344 * @nr_segs: returns the number of segments in the returned bio
345 *
346 * Check if @bio needs splitting based on the queue limits, and if so split off
347 * a bio fitting the limits from the beginning of @bio and return it. @bio is
348 * shortened to the remainder and re-submitted.
349 *
350 * The split bio is allocated from @q->bio_split, which is provided by the
351 * block layer.
352 */
353struct bio *__bio_split_to_limits(struct bio *bio,
354 const struct queue_limits *lim,
355 unsigned int *nr_segs)
356{
357 struct bio_set *bs = &bio->bi_bdev->bd_disk->bio_split;
358 struct bio *split;
359
360 switch (bio_op(bio)) {
361 case REQ_OP_DISCARD:
362 case REQ_OP_SECURE_ERASE:
363 split = bio_split_discard(bio, lim, nr_segs, bs);
364 break;
365 case REQ_OP_WRITE_ZEROES:
366 split = bio_split_write_zeroes(bio, lim, nr_segs, bs);
367 break;
368 default:
369 split = bio_split_rw(bio, lim, nr_segs, bs,
370 get_max_io_size(bio, lim) << SECTOR_SHIFT);
371 if (IS_ERR(split))
372 return NULL;
373 break;
374 }
375
376 if (split) {
377 /* there isn't chance to merge the split bio */
378 split->bi_opf |= REQ_NOMERGE;
379
380 blkcg_bio_issue_init(split);
381 bio_chain(split, bio);
382 trace_block_split(split, bio->bi_iter.bi_sector);
383 submit_bio_noacct(bio);
384 return split;
385 }
386 return bio;
387}
388
389/**
390 * bio_split_to_limits - split a bio to fit the queue limits
391 * @bio: bio to be split
392 *
393 * Check if @bio needs splitting based on the queue limits of @bio->bi_bdev, and
394 * if so split off a bio fitting the limits from the beginning of @bio and
395 * return it. @bio is shortened to the remainder and re-submitted.
396 *
397 * The split bio is allocated from @q->bio_split, which is provided by the
398 * block layer.
399 */
400struct bio *bio_split_to_limits(struct bio *bio)
401{
402 const struct queue_limits *lim = &bdev_get_queue(bio->bi_bdev)->limits;
403 unsigned int nr_segs;
404
405 if (bio_may_exceed_limits(bio, lim))
406 return __bio_split_to_limits(bio, lim, &nr_segs);
407 return bio;
408}
409EXPORT_SYMBOL(bio_split_to_limits);
410
411unsigned int blk_recalc_rq_segments(struct request *rq)
412{
413 unsigned int nr_phys_segs = 0;
414 unsigned int bytes = 0;
415 struct req_iterator iter;
416 struct bio_vec bv;
417
418 if (!rq->bio)
419 return 0;
420
421 switch (bio_op(rq->bio)) {
422 case REQ_OP_DISCARD:
423 case REQ_OP_SECURE_ERASE:
424 if (queue_max_discard_segments(rq->q) > 1) {
425 struct bio *bio = rq->bio;
426
427 for_each_bio(bio)
428 nr_phys_segs++;
429 return nr_phys_segs;
430 }
431 return 1;
432 case REQ_OP_WRITE_ZEROES:
433 return 0;
434 default:
435 break;
436 }
437
438 rq_for_each_bvec(bv, rq, iter)
439 bvec_split_segs(&rq->q->limits, &bv, &nr_phys_segs, &bytes,
440 UINT_MAX, UINT_MAX);
441 return nr_phys_segs;
442}
443
444static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
445 struct scatterlist *sglist)
446{
447 if (!*sg)
448 return sglist;
449
450 /*
451 * If the driver previously mapped a shorter list, we could see a
452 * termination bit prematurely unless it fully inits the sg table
453 * on each mapping. We KNOW that there must be more entries here
454 * or the driver would be buggy, so force clear the termination bit
455 * to avoid doing a full sg_init_table() in drivers for each command.
456 */
457 sg_unmark_end(*sg);
458 return sg_next(*sg);
459}
460
461static unsigned blk_bvec_map_sg(struct request_queue *q,
462 struct bio_vec *bvec, struct scatterlist *sglist,
463 struct scatterlist **sg)
464{
465 unsigned nbytes = bvec->bv_len;
466 unsigned nsegs = 0, total = 0;
467
468 while (nbytes > 0) {
469 unsigned offset = bvec->bv_offset + total;
470 unsigned len = min(get_max_segment_size(&q->limits,
471 bvec->bv_page, offset), nbytes);
472 struct page *page = bvec->bv_page;
473
474 /*
475 * Unfortunately a fair number of drivers barf on scatterlists
476 * that have an offset larger than PAGE_SIZE, despite other
477 * subsystems dealing with that invariant just fine. For now
478 * stick to the legacy format where we never present those from
479 * the block layer, but the code below should be removed once
480 * these offenders (mostly MMC/SD drivers) are fixed.
481 */
482 page += (offset >> PAGE_SHIFT);
483 offset &= ~PAGE_MASK;
484
485 *sg = blk_next_sg(sg, sglist);
486 sg_set_page(*sg, page, len, offset);
487
488 total += len;
489 nbytes -= len;
490 nsegs++;
491 }
492
493 return nsegs;
494}
495
496static inline int __blk_bvec_map_sg(struct bio_vec bv,
497 struct scatterlist *sglist, struct scatterlist **sg)
498{
499 *sg = blk_next_sg(sg, sglist);
500 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
501 return 1;
502}
503
504/* only try to merge bvecs into one sg if they are from two bios */
505static inline bool
506__blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
507 struct bio_vec *bvprv, struct scatterlist **sg)
508{
509
510 int nbytes = bvec->bv_len;
511
512 if (!*sg)
513 return false;
514
515 if ((*sg)->length + nbytes > queue_max_segment_size(q))
516 return false;
517
518 if (!biovec_phys_mergeable(q, bvprv, bvec))
519 return false;
520
521 (*sg)->length += nbytes;
522
523 return true;
524}
525
526static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
527 struct scatterlist *sglist,
528 struct scatterlist **sg)
529{
530 struct bio_vec bvec, bvprv = { NULL };
531 struct bvec_iter iter;
532 int nsegs = 0;
533 bool new_bio = false;
534
535 for_each_bio(bio) {
536 bio_for_each_bvec(bvec, bio, iter) {
537 /*
538 * Only try to merge bvecs from two bios given we
539 * have done bio internal merge when adding pages
540 * to bio
541 */
542 if (new_bio &&
543 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
544 goto next_bvec;
545
546 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
547 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
548 else
549 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
550 next_bvec:
551 new_bio = false;
552 }
553 if (likely(bio->bi_iter.bi_size)) {
554 bvprv = bvec;
555 new_bio = true;
556 }
557 }
558
559 return nsegs;
560}
561
562/*
563 * map a request to scatterlist, return number of sg entries setup. Caller
564 * must make sure sg can hold rq->nr_phys_segments entries
565 */
566int __blk_rq_map_sg(struct request_queue *q, struct request *rq,
567 struct scatterlist *sglist, struct scatterlist **last_sg)
568{
569 int nsegs = 0;
570
571 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
572 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg);
573 else if (rq->bio)
574 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg);
575
576 if (*last_sg)
577 sg_mark_end(*last_sg);
578
579 /*
580 * Something must have been wrong if the figured number of
581 * segment is bigger than number of req's physical segments
582 */
583 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
584
585 return nsegs;
586}
587EXPORT_SYMBOL(__blk_rq_map_sg);
588
589static inline unsigned int blk_rq_get_max_segments(struct request *rq)
590{
591 if (req_op(rq) == REQ_OP_DISCARD)
592 return queue_max_discard_segments(rq->q);
593 return queue_max_segments(rq->q);
594}
595
596static inline unsigned int blk_rq_get_max_sectors(struct request *rq,
597 sector_t offset)
598{
599 struct request_queue *q = rq->q;
600 unsigned int max_sectors;
601
602 if (blk_rq_is_passthrough(rq))
603 return q->limits.max_hw_sectors;
604
605 max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
606 if (!q->limits.chunk_sectors ||
607 req_op(rq) == REQ_OP_DISCARD ||
608 req_op(rq) == REQ_OP_SECURE_ERASE)
609 return max_sectors;
610 return min(max_sectors,
611 blk_chunk_sectors_left(offset, q->limits.chunk_sectors));
612}
613
614static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
615 unsigned int nr_phys_segs)
616{
617 if (!blk_cgroup_mergeable(req, bio))
618 goto no_merge;
619
620 if (blk_integrity_merge_bio(req->q, req, bio) == false)
621 goto no_merge;
622
623 /* discard request merge won't add new segment */
624 if (req_op(req) == REQ_OP_DISCARD)
625 return 1;
626
627 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req))
628 goto no_merge;
629
630 /*
631 * This will form the start of a new hw segment. Bump both
632 * counters.
633 */
634 req->nr_phys_segments += nr_phys_segs;
635 return 1;
636
637no_merge:
638 req_set_nomerge(req->q, req);
639 return 0;
640}
641
642int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
643{
644 if (req_gap_back_merge(req, bio))
645 return 0;
646 if (blk_integrity_rq(req) &&
647 integrity_req_gap_back_merge(req, bio))
648 return 0;
649 if (!bio_crypt_ctx_back_mergeable(req, bio))
650 return 0;
651 if (blk_rq_sectors(req) + bio_sectors(bio) >
652 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
653 req_set_nomerge(req->q, req);
654 return 0;
655 }
656
657 return ll_new_hw_segment(req, bio, nr_segs);
658}
659
660static int ll_front_merge_fn(struct request *req, struct bio *bio,
661 unsigned int nr_segs)
662{
663 if (req_gap_front_merge(req, bio))
664 return 0;
665 if (blk_integrity_rq(req) &&
666 integrity_req_gap_front_merge(req, bio))
667 return 0;
668 if (!bio_crypt_ctx_front_mergeable(req, bio))
669 return 0;
670 if (blk_rq_sectors(req) + bio_sectors(bio) >
671 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
672 req_set_nomerge(req->q, req);
673 return 0;
674 }
675
676 return ll_new_hw_segment(req, bio, nr_segs);
677}
678
679static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
680 struct request *next)
681{
682 unsigned short segments = blk_rq_nr_discard_segments(req);
683
684 if (segments >= queue_max_discard_segments(q))
685 goto no_merge;
686 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
687 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
688 goto no_merge;
689
690 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
691 return true;
692no_merge:
693 req_set_nomerge(q, req);
694 return false;
695}
696
697static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
698 struct request *next)
699{
700 int total_phys_segments;
701
702 if (req_gap_back_merge(req, next->bio))
703 return 0;
704
705 /*
706 * Will it become too large?
707 */
708 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
709 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
710 return 0;
711
712 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
713 if (total_phys_segments > blk_rq_get_max_segments(req))
714 return 0;
715
716 if (!blk_cgroup_mergeable(req, next->bio))
717 return 0;
718
719 if (blk_integrity_merge_rq(q, req, next) == false)
720 return 0;
721
722 if (!bio_crypt_ctx_merge_rq(req, next))
723 return 0;
724
725 /* Merge is OK... */
726 req->nr_phys_segments = total_phys_segments;
727 return 1;
728}
729
730/**
731 * blk_rq_set_mixed_merge - mark a request as mixed merge
732 * @rq: request to mark as mixed merge
733 *
734 * Description:
735 * @rq is about to be mixed merged. Make sure the attributes
736 * which can be mixed are set in each bio and mark @rq as mixed
737 * merged.
738 */
739void blk_rq_set_mixed_merge(struct request *rq)
740{
741 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK;
742 struct bio *bio;
743
744 if (rq->rq_flags & RQF_MIXED_MERGE)
745 return;
746
747 /*
748 * @rq will no longer represent mixable attributes for all the
749 * contained bios. It will just track those of the first one.
750 * Distributes the attributs to each bio.
751 */
752 for (bio = rq->bio; bio; bio = bio->bi_next) {
753 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
754 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
755 bio->bi_opf |= ff;
756 }
757 rq->rq_flags |= RQF_MIXED_MERGE;
758}
759
760static void blk_account_io_merge_request(struct request *req)
761{
762 if (blk_do_io_stat(req)) {
763 part_stat_lock();
764 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
765 part_stat_unlock();
766 }
767}
768
769static enum elv_merge blk_try_req_merge(struct request *req,
770 struct request *next)
771{
772 if (blk_discard_mergable(req))
773 return ELEVATOR_DISCARD_MERGE;
774 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
775 return ELEVATOR_BACK_MERGE;
776
777 return ELEVATOR_NO_MERGE;
778}
779
780/*
781 * For non-mq, this has to be called with the request spinlock acquired.
782 * For mq with scheduling, the appropriate queue wide lock should be held.
783 */
784static struct request *attempt_merge(struct request_queue *q,
785 struct request *req, struct request *next)
786{
787 if (!rq_mergeable(req) || !rq_mergeable(next))
788 return NULL;
789
790 if (req_op(req) != req_op(next))
791 return NULL;
792
793 if (rq_data_dir(req) != rq_data_dir(next))
794 return NULL;
795
796 if (req->ioprio != next->ioprio)
797 return NULL;
798
799 /*
800 * If we are allowed to merge, then append bio list
801 * from next to rq and release next. merge_requests_fn
802 * will have updated segment counts, update sector
803 * counts here. Handle DISCARDs separately, as they
804 * have separate settings.
805 */
806
807 switch (blk_try_req_merge(req, next)) {
808 case ELEVATOR_DISCARD_MERGE:
809 if (!req_attempt_discard_merge(q, req, next))
810 return NULL;
811 break;
812 case ELEVATOR_BACK_MERGE:
813 if (!ll_merge_requests_fn(q, req, next))
814 return NULL;
815 break;
816 default:
817 return NULL;
818 }
819
820 /*
821 * If failfast settings disagree or any of the two is already
822 * a mixed merge, mark both as mixed before proceeding. This
823 * makes sure that all involved bios have mixable attributes
824 * set properly.
825 */
826 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
827 (req->cmd_flags & REQ_FAILFAST_MASK) !=
828 (next->cmd_flags & REQ_FAILFAST_MASK)) {
829 blk_rq_set_mixed_merge(req);
830 blk_rq_set_mixed_merge(next);
831 }
832
833 /*
834 * At this point we have either done a back merge or front merge. We
835 * need the smaller start_time_ns of the merged requests to be the
836 * current request for accounting purposes.
837 */
838 if (next->start_time_ns < req->start_time_ns)
839 req->start_time_ns = next->start_time_ns;
840
841 req->biotail->bi_next = next->bio;
842 req->biotail = next->biotail;
843
844 req->__data_len += blk_rq_bytes(next);
845
846 if (!blk_discard_mergable(req))
847 elv_merge_requests(q, req, next);
848
849 /*
850 * 'next' is going away, so update stats accordingly
851 */
852 blk_account_io_merge_request(next);
853
854 trace_block_rq_merge(next);
855
856 /*
857 * ownership of bio passed from next to req, return 'next' for
858 * the caller to free
859 */
860 next->bio = NULL;
861 return next;
862}
863
864static struct request *attempt_back_merge(struct request_queue *q,
865 struct request *rq)
866{
867 struct request *next = elv_latter_request(q, rq);
868
869 if (next)
870 return attempt_merge(q, rq, next);
871
872 return NULL;
873}
874
875static struct request *attempt_front_merge(struct request_queue *q,
876 struct request *rq)
877{
878 struct request *prev = elv_former_request(q, rq);
879
880 if (prev)
881 return attempt_merge(q, prev, rq);
882
883 return NULL;
884}
885
886/*
887 * Try to merge 'next' into 'rq'. Return true if the merge happened, false
888 * otherwise. The caller is responsible for freeing 'next' if the merge
889 * happened.
890 */
891bool blk_attempt_req_merge(struct request_queue *q, struct request *rq,
892 struct request *next)
893{
894 return attempt_merge(q, rq, next);
895}
896
897bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
898{
899 if (!rq_mergeable(rq) || !bio_mergeable(bio))
900 return false;
901
902 if (req_op(rq) != bio_op(bio))
903 return false;
904
905 /* different data direction or already started, don't merge */
906 if (bio_data_dir(bio) != rq_data_dir(rq))
907 return false;
908
909 /* don't merge across cgroup boundaries */
910 if (!blk_cgroup_mergeable(rq, bio))
911 return false;
912
913 /* only merge integrity protected bio into ditto rq */
914 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
915 return false;
916
917 /* Only merge if the crypt contexts are compatible */
918 if (!bio_crypt_rq_ctx_compatible(rq, bio))
919 return false;
920
921 if (rq->ioprio != bio_prio(bio))
922 return false;
923
924 return true;
925}
926
927enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
928{
929 if (blk_discard_mergable(rq))
930 return ELEVATOR_DISCARD_MERGE;
931 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
932 return ELEVATOR_BACK_MERGE;
933 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
934 return ELEVATOR_FRONT_MERGE;
935 return ELEVATOR_NO_MERGE;
936}
937
938static void blk_account_io_merge_bio(struct request *req)
939{
940 if (!blk_do_io_stat(req))
941 return;
942
943 part_stat_lock();
944 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
945 part_stat_unlock();
946}
947
948enum bio_merge_status {
949 BIO_MERGE_OK,
950 BIO_MERGE_NONE,
951 BIO_MERGE_FAILED,
952};
953
954static enum bio_merge_status bio_attempt_back_merge(struct request *req,
955 struct bio *bio, unsigned int nr_segs)
956{
957 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK;
958
959 if (!ll_back_merge_fn(req, bio, nr_segs))
960 return BIO_MERGE_FAILED;
961
962 trace_block_bio_backmerge(bio);
963 rq_qos_merge(req->q, req, bio);
964
965 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
966 blk_rq_set_mixed_merge(req);
967
968 req->biotail->bi_next = bio;
969 req->biotail = bio;
970 req->__data_len += bio->bi_iter.bi_size;
971
972 bio_crypt_free_ctx(bio);
973
974 blk_account_io_merge_bio(req);
975 return BIO_MERGE_OK;
976}
977
978static enum bio_merge_status bio_attempt_front_merge(struct request *req,
979 struct bio *bio, unsigned int nr_segs)
980{
981 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK;
982
983 if (!ll_front_merge_fn(req, bio, nr_segs))
984 return BIO_MERGE_FAILED;
985
986 trace_block_bio_frontmerge(bio);
987 rq_qos_merge(req->q, req, bio);
988
989 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
990 blk_rq_set_mixed_merge(req);
991
992 bio->bi_next = req->bio;
993 req->bio = bio;
994
995 req->__sector = bio->bi_iter.bi_sector;
996 req->__data_len += bio->bi_iter.bi_size;
997
998 bio_crypt_do_front_merge(req, bio);
999
1000 blk_account_io_merge_bio(req);
1001 return BIO_MERGE_OK;
1002}
1003
1004static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q,
1005 struct request *req, struct bio *bio)
1006{
1007 unsigned short segments = blk_rq_nr_discard_segments(req);
1008
1009 if (segments >= queue_max_discard_segments(q))
1010 goto no_merge;
1011 if (blk_rq_sectors(req) + bio_sectors(bio) >
1012 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
1013 goto no_merge;
1014
1015 rq_qos_merge(q, req, bio);
1016
1017 req->biotail->bi_next = bio;
1018 req->biotail = bio;
1019 req->__data_len += bio->bi_iter.bi_size;
1020 req->nr_phys_segments = segments + 1;
1021
1022 blk_account_io_merge_bio(req);
1023 return BIO_MERGE_OK;
1024no_merge:
1025 req_set_nomerge(q, req);
1026 return BIO_MERGE_FAILED;
1027}
1028
1029static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q,
1030 struct request *rq,
1031 struct bio *bio,
1032 unsigned int nr_segs,
1033 bool sched_allow_merge)
1034{
1035 if (!blk_rq_merge_ok(rq, bio))
1036 return BIO_MERGE_NONE;
1037
1038 switch (blk_try_merge(rq, bio)) {
1039 case ELEVATOR_BACK_MERGE:
1040 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1041 return bio_attempt_back_merge(rq, bio, nr_segs);
1042 break;
1043 case ELEVATOR_FRONT_MERGE:
1044 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1045 return bio_attempt_front_merge(rq, bio, nr_segs);
1046 break;
1047 case ELEVATOR_DISCARD_MERGE:
1048 return bio_attempt_discard_merge(q, rq, bio);
1049 default:
1050 return BIO_MERGE_NONE;
1051 }
1052
1053 return BIO_MERGE_FAILED;
1054}
1055
1056/**
1057 * blk_attempt_plug_merge - try to merge with %current's plugged list
1058 * @q: request_queue new bio is being queued at
1059 * @bio: new bio being queued
1060 * @nr_segs: number of segments in @bio
1061 * from the passed in @q already in the plug list
1062 *
1063 * Determine whether @bio being queued on @q can be merged with the previous
1064 * request on %current's plugged list. Returns %true if merge was successful,
1065 * otherwise %false.
1066 *
1067 * Plugging coalesces IOs from the same issuer for the same purpose without
1068 * going through @q->queue_lock. As such it's more of an issuing mechanism
1069 * than scheduling, and the request, while may have elvpriv data, is not
1070 * added on the elevator at this point. In addition, we don't have
1071 * reliable access to the elevator outside queue lock. Only check basic
1072 * merging parameters without querying the elevator.
1073 *
1074 * Caller must ensure !blk_queue_nomerges(q) beforehand.
1075 */
1076bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio,
1077 unsigned int nr_segs)
1078{
1079 struct blk_plug *plug;
1080 struct request *rq;
1081
1082 plug = blk_mq_plug(bio);
1083 if (!plug || rq_list_empty(plug->mq_list))
1084 return false;
1085
1086 rq_list_for_each(&plug->mq_list, rq) {
1087 if (rq->q == q) {
1088 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) ==
1089 BIO_MERGE_OK)
1090 return true;
1091 break;
1092 }
1093
1094 /*
1095 * Only keep iterating plug list for merges if we have multiple
1096 * queues
1097 */
1098 if (!plug->multiple_queues)
1099 break;
1100 }
1101 return false;
1102}
1103
1104/*
1105 * Iterate list of requests and see if we can merge this bio with any
1106 * of them.
1107 */
1108bool blk_bio_list_merge(struct request_queue *q, struct list_head *list,
1109 struct bio *bio, unsigned int nr_segs)
1110{
1111 struct request *rq;
1112 int checked = 8;
1113
1114 list_for_each_entry_reverse(rq, list, queuelist) {
1115 if (!checked--)
1116 break;
1117
1118 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) {
1119 case BIO_MERGE_NONE:
1120 continue;
1121 case BIO_MERGE_OK:
1122 return true;
1123 case BIO_MERGE_FAILED:
1124 return false;
1125 }
1126
1127 }
1128
1129 return false;
1130}
1131EXPORT_SYMBOL_GPL(blk_bio_list_merge);
1132
1133bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
1134 unsigned int nr_segs, struct request **merged_request)
1135{
1136 struct request *rq;
1137
1138 switch (elv_merge(q, &rq, bio)) {
1139 case ELEVATOR_BACK_MERGE:
1140 if (!blk_mq_sched_allow_merge(q, rq, bio))
1141 return false;
1142 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1143 return false;
1144 *merged_request = attempt_back_merge(q, rq);
1145 if (!*merged_request)
1146 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
1147 return true;
1148 case ELEVATOR_FRONT_MERGE:
1149 if (!blk_mq_sched_allow_merge(q, rq, bio))
1150 return false;
1151 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1152 return false;
1153 *merged_request = attempt_front_merge(q, rq);
1154 if (!*merged_request)
1155 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
1156 return true;
1157 case ELEVATOR_DISCARD_MERGE:
1158 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK;
1159 default:
1160 return false;
1161 }
1162}
1163EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
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}