block/blk-merge.c

   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
  15 static 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,
  31                         bio_allowed_max_sectors(q));
  32         max_discard_sectors -= max_discard_sectors % granularity;
  33
  34         if (unlikely(!max_discard_sectors)) {
  35                 /* XXX: warn */
  36                 return NULL;
  37         }
  38
  39         if (bio_sectors(bio) <= max_discard_sectors)
  40                 return NULL;
  41
  42         split_sectors = max_discard_sectors;
  43
  44         /*
  45          * If the next starting sector would be misaligned, stop the discard at
  46          * the previous aligned sector.
  47          */
  48         alignment = (q->limits.discard_alignment >> 9) % granularity;
  49
  50         tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
  51         tmp = sector_div(tmp, granularity);
  52
  53         if (split_sectors > tmp)
  54                 split_sectors -= tmp;
  55
  56         return bio_split(bio, split_sectors, GFP_NOIO, bs);
  57 }
  58
  59 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
  60                 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
  61 {
  62         *nsegs = 1;
  63
  64         if (!q->limits.max_write_zeroes_sectors)
  65                 return NULL;
  66
  67         if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
  68                 return NULL;
  69
  70         return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
  71 }
  72
  73 static struct bio *blk_bio_write_same_split(struct request_queue *q,
  74                                             struct bio *bio,
  75                                             struct bio_set *bs,
  76                                             unsigned *nsegs)
  77 {
  78         *nsegs = 1;
  79
  80         if (!q->limits.max_write_same_sectors)
  81                 return NULL;
  82
  83         if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
  84                 return NULL;
  85
  86         return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
  87 }
  88
  89 static inline unsigned get_max_io_size(struct request_queue *q,
  90                                        struct bio *bio)
  91 {
  92         unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
  93         unsigned mask = queue_logical_block_size(q) - 1;
  94
  95         /* aligned to logical block size */
  96         sectors &= ~(mask >> 9);
  97
  98         return sectors;
  99 }
 100
 101 static struct bio *blk_bio_segment_split(struct request_queue *q,
 102                                          struct bio *bio,
 103                                          struct bio_set *bs,
 104                                          unsigned *segs)
 105 {
 106         struct bio_vec bv, bvprv, *bvprvp = NULL;
 107         struct bvec_iter iter;
 108         unsigned seg_size = 0, nsegs = 0, sectors = 0;
 109         unsigned front_seg_size = bio->bi_seg_front_size;
 110         bool do_split = true;
 111         struct bio *new = NULL;
 112         const unsigned max_sectors = get_max_io_size(q, bio);
 113
 114         bio_for_each_segment(bv, bio, iter) {
 115                 /*
 116                  * If the queue doesn't support SG gaps and adding this
 117                  * offset would create a gap, disallow it.
 118                  */
 119                 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
 120                         goto split;
 121
 122                 if (sectors + (bv.bv_len >> 9) > max_sectors) {
 123                         /*
 124                          * Consider this a new segment if we're splitting in
 125                          * the middle of this vector.
 126                          */
 127                         if (nsegs < queue_max_segments(q) &&
 128                             sectors < max_sectors) {
 129                                 nsegs++;
 130                                 sectors = max_sectors;
 131                         }
 132                         goto split;
 133                 }
 134
 135                 if (bvprvp && blk_queue_cluster(q)) {
 136                         if (seg_size + bv.bv_len > queue_max_segment_size(q))
 137                                 goto new_segment;
 138                         if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
 139                                 goto new_segment;
 140                         if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
 141                                 goto new_segment;
 142
 143                         seg_size += bv.bv_len;
 144                         bvprv = bv;
 145                         bvprvp = &bvprv;
 146                         sectors += bv.bv_len >> 9;
 147
 148                         continue;
 149                 }
 150 new_segment:
 151                 if (nsegs == queue_max_segments(q))
 152                         goto split;
 153
 154                 if (nsegs == 1 && seg_size > front_seg_size)
 155                         front_seg_size = seg_size;
 156
 157                 nsegs++;
 158                 bvprv = bv;
 159                 bvprvp = &bvprv;
 160                 seg_size = bv.bv_len;
 161                 sectors += bv.bv_len >> 9;
 162
 163         }
 164
 165         do_split = false;
 166 split:
 167         *segs = nsegs;
 168
 169         if (do_split) {
 170                 new = bio_split(bio, sectors, GFP_NOIO, bs);
 171                 if (new)
 172                         bio = new;
 173         }
 174
 175         if (nsegs == 1 && seg_size > front_seg_size)
 176                 front_seg_size = seg_size;
 177         bio->bi_seg_front_size = front_seg_size;
 178         if (seg_size > bio->bi_seg_back_size)
 179                 bio->bi_seg_back_size = seg_size;
 180
 181         return do_split ? new : NULL;
 182 }
 183
 184 void blk_queue_split(struct request_queue *q, struct bio **bio)
 185 {
 186         struct bio *split, *res;
 187         unsigned nsegs;
 188
 189         switch (bio_op(*bio)) {
 190         case REQ_OP_DISCARD:
 191         case REQ_OP_SECURE_ERASE:
 192                 split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
 193                 break;
 194         case REQ_OP_WRITE_ZEROES:
 195                 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
 196                 break;
 197         case REQ_OP_WRITE_SAME:
 198                 split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
 199                 break;
 200         default:
 201                 split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
 202                 break;
 203         }
 204
 205         /* physical segments can be figured out during splitting */
 206         res = split ? split : *bio;
 207         res->bi_phys_segments = nsegs;
 208         bio_set_flag(res, BIO_SEG_VALID);
 209
 210         if (split) {
 211                 /* there isn't chance to merge the splitted bio */
 212                 split->bi_opf |= REQ_NOMERGE;
 213
 214                 /*
 215                  * Since we're recursing into make_request here, ensure
 216                  * that we mark this bio as already having entered the queue.
 217                  * If not, and the queue is going away, we can get stuck
 218                  * forever on waiting for the queue reference to drop. But
 219                  * that will never happen, as we're already holding a
 220                  * reference to it.
 221                  */
 222                 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
 223
 224                 bio_chain(split, *bio);
 225                 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
 226                 generic_make_request(*bio);
 227                 *bio = split;
 228         }
 229 }
 230 EXPORT_SYMBOL(blk_queue_split);
 231
 232 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
 233                                              struct bio *bio,
 234                                              bool no_sg_merge)
 235 {
 236         struct bio_vec bv, bvprv = { NULL };
 237         int cluster, prev = 0;
 238         unsigned int seg_size, nr_phys_segs;
 239         struct bio *fbio, *bbio;
 240         struct bvec_iter iter;
 241
 242         if (!bio)
 243                 return 0;
 244
 245         switch (bio_op(bio)) {
 246         case REQ_OP_DISCARD:
 247         case REQ_OP_SECURE_ERASE:
 248         case REQ_OP_WRITE_ZEROES:
 249                 return 0;
 250         case REQ_OP_WRITE_SAME:
 251                 return 1;
 252         }
 253
 254         fbio = bio;
 255         cluster = blk_queue_cluster(q);
 256         seg_size = 0;
 257         nr_phys_segs = 0;
 258         for_each_bio(bio) {
 259                 bio_for_each_segment(bv, bio, iter) {
 260                         /*
 261                          * If SG merging is disabled, each bio vector is
 262                          * a segment
 263                          */
 264                         if (no_sg_merge)
 265                                 goto new_segment;
 266
 267                         if (prev && cluster) {
 268                                 if (seg_size + bv.bv_len
 269                                     > queue_max_segment_size(q))
 270                                         goto new_segment;
 271                                 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
 272                                         goto new_segment;
 273                                 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
 274                                         goto new_segment;
 275
 276                                 seg_size += bv.bv_len;
 277                                 bvprv = bv;
 278                                 continue;
 279                         }
 280 new_segment:
 281                         if (nr_phys_segs == 1 && seg_size >
 282                             fbio->bi_seg_front_size)
 283                                 fbio->bi_seg_front_size = seg_size;
 284
 285                         nr_phys_segs++;
 286                         bvprv = bv;
 287                         prev = 1;
 288                         seg_size = bv.bv_len;
 289                 }
 290                 bbio = bio;
 291         }
 292
 293         if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
 294                 fbio->bi_seg_front_size = seg_size;
 295         if (seg_size > bbio->bi_seg_back_size)
 296                 bbio->bi_seg_back_size = seg_size;
 297
 298         return nr_phys_segs;
 299 }
 300
 301 void blk_recalc_rq_segments(struct request *rq)
 302 {
 303         bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
 304                         &rq->q->queue_flags);
 305
 306         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
 307                         no_sg_merge);
 308 }
 309
 310 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 311 {
 312         unsigned short seg_cnt;
 313
 314         /* estimate segment number by bi_vcnt for non-cloned bio */
 315         if (bio_flagged(bio, BIO_CLONED))
 316                 seg_cnt = bio_segments(bio);
 317         else
 318                 seg_cnt = bio->bi_vcnt;
 319
 320         if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
 321                         (seg_cnt < queue_max_segments(q)))
 322                 bio->bi_phys_segments = seg_cnt;
 323         else {
 324                 struct bio *nxt = bio->bi_next;
 325
 326                 bio->bi_next = NULL;
 327                 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
 328                 bio->bi_next = nxt;
 329         }
 330
 331         bio_set_flag(bio, BIO_SEG_VALID);
 332 }
 333 EXPORT_SYMBOL(blk_recount_segments);
 334
 335 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 336                                    struct bio *nxt)
 337 {
 338         struct bio_vec end_bv = { NULL }, nxt_bv;
 339
 340         if (!blk_queue_cluster(q))
 341                 return 0;
 342
 343         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 344             queue_max_segment_size(q))
 345                 return 0;
 346
 347         if (!bio_has_data(bio))
 348                 return 1;
 349
 350         bio_get_last_bvec(bio, &end_bv);
 351         bio_get_first_bvec(nxt, &nxt_bv);
 352
 353         if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
 354                 return 0;
 355
 356         /*
 357          * bio and nxt are contiguous in memory; check if the queue allows
 358          * these two to be merged into one
 359          */
 360         if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
 361                 return 1;
 362
 363         return 0;
 364 }
 365
 366 static inline void
 367 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
 368                      struct scatterlist *sglist, struct bio_vec *bvprv,
 369                      struct scatterlist **sg, int *nsegs, int *cluster)
 370 {
 371
 372         int nbytes = bvec->bv_len;
 373
 374         if (*sg && *cluster) {
 375                 if ((*sg)->length + nbytes > queue_max_segment_size(q))
 376                         goto new_segment;
 377
 378                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 379                         goto new_segment;
 380                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 381                         goto new_segment;
 382
 383                 (*sg)->length += nbytes;
 384         } else {
 385 new_segment:
 386                 if (!*sg)
 387                         *sg = sglist;
 388                 else {
 389                         /*
 390                          * If the driver previously mapped a shorter
 391                          * list, we could see a termination bit
 392                          * prematurely unless it fully inits the sg
 393                          * table on each mapping. We KNOW that there
 394                          * must be more entries here or the driver
 395                          * would be buggy, so force clear the
 396                          * termination bit to avoid doing a full
 397                          * sg_init_table() in drivers for each command.
 398                          */
 399                         sg_unmark_end(*sg);
 400                         *sg = sg_next(*sg);
 401                 }
 402
 403                 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
 404                 (*nsegs)++;
 405         }
 406         *bvprv = *bvec;
 407 }
 408
 409 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
 410                 struct scatterlist *sglist, struct scatterlist **sg)
 411 {
 412         *sg = sglist;
 413         sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
 414         return 1;
 415 }
 416
 417 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 418                              struct scatterlist *sglist,
 419                              struct scatterlist **sg)
 420 {
 421         struct bio_vec bvec, bvprv = { NULL };
 422         struct bvec_iter iter;
 423         int cluster = blk_queue_cluster(q), nsegs = 0;
 424
 425         for_each_bio(bio)
 426                 bio_for_each_segment(bvec, bio, iter)
 427                         __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
 428                                              &nsegs, &cluster);
 429
 430         return nsegs;
 431 }
 432
 433 /*
 434  * map a request to scatterlist, return number of sg entries setup. Caller
 435  * must make sure sg can hold rq->nr_phys_segments entries
 436  */
 437 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 438                   struct scatterlist *sglist)
 439 {
 440         struct scatterlist *sg = NULL;
 441         int nsegs = 0;
 442
 443         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
 444                 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
 445         else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
 446                 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
 447         else if (rq->bio)
 448                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
 449
 450         if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
 451             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 452                 unsigned int pad_len =
 453                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 454
 455                 sg->length += pad_len;
 456                 rq->extra_len += pad_len;
 457         }
 458
 459         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 460                 if (op_is_write(req_op(rq)))
 461                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 462
 463                 sg_unmark_end(sg);
 464                 sg = sg_next(sg);
 465                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 466                             q->dma_drain_size,
 467                             ((unsigned long)q->dma_drain_buffer) &
 468                             (PAGE_SIZE - 1));
 469                 nsegs++;
 470                 rq->extra_len += q->dma_drain_size;
 471         }
 472
 473         if (sg)
 474                 sg_mark_end(sg);
 475
 476         /*
 477          * Something must have been wrong if the figured number of
 478          * segment is bigger than number of req's physical segments
 479          */
 480         WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
 481
 482         return nsegs;
 483 }
 484 EXPORT_SYMBOL(blk_rq_map_sg);
 485
 486 static inline int ll_new_hw_segment(struct request_queue *q,
 487                                     struct request *req,
 488                                     struct bio *bio)
 489 {
 490         int nr_phys_segs = bio_phys_segments(q, bio);
 491
 492         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 493                 goto no_merge;
 494
 495         if (blk_integrity_merge_bio(q, req, bio) == false)
 496                 goto no_merge;
 497
 498         /*
 499          * This will form the start of a new hw segment.  Bump both
 500          * counters.
 501          */
 502         req->nr_phys_segments += nr_phys_segs;
 503         return 1;
 504
 505 no_merge:
 506         req_set_nomerge(q, req);
 507         return 0;
 508 }
 509
 510 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 511                      struct bio *bio)
 512 {
 513         if (req_gap_back_merge(req, bio))
 514                 return 0;
 515         if (blk_integrity_rq(req) &&
 516             integrity_req_gap_back_merge(req, bio))
 517                 return 0;
 518         if (blk_rq_sectors(req) + bio_sectors(bio) >
 519             blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
 520                 req_set_nomerge(q, req);
 521                 return 0;
 522         }
 523         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 524                 blk_recount_segments(q, req->biotail);
 525         if (!bio_flagged(bio, BIO_SEG_VALID))
 526                 blk_recount_segments(q, bio);
 527
 528         return ll_new_hw_segment(q, req, bio);
 529 }
 530
 531 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 532                       struct bio *bio)
 533 {
 534
 535         if (req_gap_front_merge(req, bio))
 536                 return 0;
 537         if (blk_integrity_rq(req) &&
 538             integrity_req_gap_front_merge(req, bio))
 539                 return 0;
 540         if (blk_rq_sectors(req) + bio_sectors(bio) >
 541             blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
 542                 req_set_nomerge(q, req);
 543                 return 0;
 544         }
 545         if (!bio_flagged(bio, BIO_SEG_VALID))
 546                 blk_recount_segments(q, bio);
 547         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 548                 blk_recount_segments(q, req->bio);
 549
 550         return ll_new_hw_segment(q, req, bio);
 551 }
 552
 553 /*
 554  * blk-mq uses req->special to carry normal driver per-request payload, it
 555  * does not indicate a prepared command that we cannot merge with.
 556  */
 557 static bool req_no_special_merge(struct request *req)
 558 {
 559         struct request_queue *q = req->q;
 560
 561         return !q->mq_ops && req->special;
 562 }
 563
 564 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
 565                 struct request *next)
 566 {
 567         unsigned short segments = blk_rq_nr_discard_segments(req);
 568
 569         if (segments >= queue_max_discard_segments(q))
 570                 goto no_merge;
 571         if (blk_rq_sectors(req) + bio_sectors(next->bio) >
 572             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 573                 goto no_merge;
 574
 575         req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
 576         return true;
 577 no_merge:
 578         req_set_nomerge(q, req);
 579         return false;
 580 }
 581
 582 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 583                                 struct request *next)
 584 {
 585         int total_phys_segments;
 586         unsigned int seg_size =
 587                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 588
 589         /*
 590          * First check if the either of the requests are re-queued
 591          * requests.  Can't merge them if they are.
 592          */
 593         if (req_no_special_merge(req) || req_no_special_merge(next))
 594                 return 0;
 595
 596         if (req_gap_back_merge(req, next->bio))
 597                 return 0;
 598
 599         /*
 600          * Will it become too large?
 601          */
 602         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 603             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 604                 return 0;
 605
 606         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 607         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 608                 if (req->nr_phys_segments == 1)
 609                         req->bio->bi_seg_front_size = seg_size;
 610                 if (next->nr_phys_segments == 1)
 611                         next->biotail->bi_seg_back_size = seg_size;
 612                 total_phys_segments--;
 613         }
 614
 615         if (total_phys_segments > queue_max_segments(q))
 616                 return 0;
 617
 618         if (blk_integrity_merge_rq(q, req, next) == false)
 619                 return 0;
 620
 621         /* Merge is OK... */
 622         req->nr_phys_segments = total_phys_segments;
 623         return 1;
 624 }
 625
 626 /**
 627  * blk_rq_set_mixed_merge - mark a request as mixed merge
 628  * @rq: request to mark as mixed merge
 629  *
 630  * Description:
 631  *     @rq is about to be mixed merged.  Make sure the attributes
 632  *     which can be mixed are set in each bio and mark @rq as mixed
 633  *     merged.
 634  */
 635 void blk_rq_set_mixed_merge(struct request *rq)
 636 {
 637         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 638         struct bio *bio;
 639
 640         if (rq->rq_flags & RQF_MIXED_MERGE)
 641                 return;
 642
 643         /*
 644          * @rq will no longer represent mixable attributes for all the
 645          * contained bios.  It will just track those of the first one.
 646          * Distributes the attributs to each bio.
 647          */
 648         for (bio = rq->bio; bio; bio = bio->bi_next) {
 649                 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
 650                              (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
 651                 bio->bi_opf |= ff;
 652         }
 653         rq->rq_flags |= RQF_MIXED_MERGE;
 654 }
 655
 656 static void blk_account_io_merge(struct request *req)
 657 {
 658         if (blk_do_io_stat(req)) {
 659                 struct hd_struct *part;
 660                 int cpu;
 661
 662                 cpu = part_stat_lock();
 663                 part = req->part;
 664
 665                 part_round_stats(req->q, cpu, part);
 666                 part_dec_in_flight(req->q, part, rq_data_dir(req));
 667
 668                 hd_struct_put(part);
 669                 part_stat_unlock();
 670         }
 671 }
 672 /*
 673  * Two cases of handling DISCARD merge:
 674  * If max_discard_segments > 1, the driver takes every bio
 675  * as a range and send them to controller together. The ranges
 676  * needn't to be contiguous.
 677  * Otherwise, the bios/requests will be handled as same as
 678  * others which should be contiguous.
 679  */
 680 static inline bool blk_discard_mergable(struct request *req)
 681 {
 682         if (req_op(req) == REQ_OP_DISCARD &&
 683             queue_max_discard_segments(req->q) > 1)
 684                 return true;
 685         return false;
 686 }
 687
 688 enum elv_merge blk_try_req_merge(struct request *req, struct request *next)
 689 {
 690         if (blk_discard_mergable(req))
 691                 return ELEVATOR_DISCARD_MERGE;
 692         else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
 693                 return ELEVATOR_BACK_MERGE;
 694
 695         return ELEVATOR_NO_MERGE;
 696 }
 697
 698 /*
 699  * For non-mq, this has to be called with the request spinlock acquired.
 700  * For mq with scheduling, the appropriate queue wide lock should be held.
 701  */
 702 static struct request *attempt_merge(struct request_queue *q,
 703                                      struct request *req, struct request *next)
 704 {
 705         if (!q->mq_ops)
 706                 lockdep_assert_held(q->queue_lock);
 707
 708         if (!rq_mergeable(req) || !rq_mergeable(next))
 709                 return NULL;
 710
 711         if (req_op(req) != req_op(next))
 712                 return NULL;
 713
 714         if (rq_data_dir(req) != rq_data_dir(next)
 715             || req->rq_disk != next->rq_disk
 716             || req_no_special_merge(next))
 717                 return NULL;
 718
 719         if (req_op(req) == REQ_OP_WRITE_SAME &&
 720             !blk_write_same_mergeable(req->bio, next->bio))
 721                 return NULL;
 722
 723         /*
 724          * Don't allow merge of different write hints, or for a hint with
 725          * non-hint IO.
 726          */
 727         if (req->write_hint != next->write_hint)
 728                 return NULL;
 729
 730         /*
 731          * If we are allowed to merge, then append bio list
 732          * from next to rq and release next. merge_requests_fn
 733          * will have updated segment counts, update sector
 734          * counts here. Handle DISCARDs separately, as they
 735          * have separate settings.
 736          */
 737
 738         switch (blk_try_req_merge(req, next)) {
 739         case ELEVATOR_DISCARD_MERGE:
 740                 if (!req_attempt_discard_merge(q, req, next))
 741                         return NULL;
 742                 break;
 743         case ELEVATOR_BACK_MERGE:
 744                 if (!ll_merge_requests_fn(q, req, next))
 745                         return NULL;
 746                 break;
 747         default:
 748                 return NULL;
 749         }
 750
 751         /*
 752          * If failfast settings disagree or any of the two is already
 753          * a mixed merge, mark both as mixed before proceeding.  This
 754          * makes sure that all involved bios have mixable attributes
 755          * set properly.
 756          */
 757         if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
 758             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 759             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 760                 blk_rq_set_mixed_merge(req);
 761                 blk_rq_set_mixed_merge(next);
 762         }
 763
 764         /*
 765          * At this point we have either done a back merge or front merge. We
 766          * need the smaller start_time_ns of the merged requests to be the
 767          * current request for accounting purposes.
 768          */
 769         if (next->start_time_ns < req->start_time_ns)
 770                 req->start_time_ns = next->start_time_ns;
 771
 772         req->biotail->bi_next = next->bio;
 773         req->biotail = next->biotail;
 774
 775         req->__data_len += blk_rq_bytes(next);
 776
 777         if (req_op(req) != REQ_OP_DISCARD)
 778                 elv_merge_requests(q, req, next);
 779
 780         /*
 781          * 'next' is going away, so update stats accordingly
 782          */
 783         blk_account_io_merge(next);
 784
 785         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 786         if (blk_rq_cpu_valid(next))
 787                 req->cpu = next->cpu;
 788
 789         /*
 790          * ownership of bio passed from next to req, return 'next' for
 791          * the caller to free
 792          */
 793         next->bio = NULL;
 794         return next;
 795 }
 796
 797 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
 798 {
 799         struct request *next = elv_latter_request(q, rq);
 800
 801         if (next)
 802                 return attempt_merge(q, rq, next);
 803
 804         return NULL;
 805 }
 806
 807 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
 808 {
 809         struct request *prev = elv_former_request(q, rq);
 810
 811         if (prev)
 812                 return attempt_merge(q, prev, rq);
 813
 814         return NULL;
 815 }
 816
 817 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 818                           struct request *next)
 819 {
 820         struct elevator_queue *e = q->elevator;
 821         struct request *free;
 822
 823         if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn)
 824                 if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next))
 825                         return 0;
 826
 827         free = attempt_merge(q, rq, next);
 828         if (free) {
 829                 __blk_put_request(q, free);
 830                 return 1;
 831         }
 832
 833         return 0;
 834 }
 835
 836 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 837 {
 838         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 839                 return false;
 840
 841         if (req_op(rq) != bio_op(bio))
 842                 return false;
 843
 844         /* different data direction or already started, don't merge */
 845         if (bio_data_dir(bio) != rq_data_dir(rq))
 846                 return false;
 847
 848         /* must be same device and not a special request */
 849         if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq))
 850                 return false;
 851
 852         /* only merge integrity protected bio into ditto rq */
 853         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 854                 return false;
 855
 856         /* must be using the same buffer */
 857         if (req_op(rq) == REQ_OP_WRITE_SAME &&
 858             !blk_write_same_mergeable(rq->bio, bio))
 859                 return false;
 860
 861         /*
 862          * Don't allow merge of different write hints, or for a hint with
 863          * non-hint IO.
 864          */
 865         if (rq->write_hint != bio->bi_write_hint)
 866                 return false;
 867
 868         return true;
 869 }
 870
 871 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
 872 {
 873         if (blk_discard_mergable(rq))
 874                 return ELEVATOR_DISCARD_MERGE;
 875         else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 876                 return ELEVATOR_BACK_MERGE;
 877         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 878                 return ELEVATOR_FRONT_MERGE;
 879         return ELEVATOR_NO_MERGE;
 880 }