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 = bio_segments(bio);
 313
 314         if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
 315                         (seg_cnt < queue_max_segments(q)))
 316                 bio->bi_phys_segments = seg_cnt;
 317         else {
 318                 struct bio *nxt = bio->bi_next;
 319
 320                 bio->bi_next = NULL;
 321                 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
 322                 bio->bi_next = nxt;
 323         }
 324
 325         bio_set_flag(bio, BIO_SEG_VALID);
 326 }
 327 EXPORT_SYMBOL(blk_recount_segments);
 328
 329 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 330                                    struct bio *nxt)
 331 {
 332         struct bio_vec end_bv = { NULL }, nxt_bv;
 333
 334         if (!blk_queue_cluster(q))
 335                 return 0;
 336
 337         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 338             queue_max_segment_size(q))
 339                 return 0;
 340
 341         if (!bio_has_data(bio))
 342                 return 1;
 343
 344         bio_get_last_bvec(bio, &end_bv);
 345         bio_get_first_bvec(nxt, &nxt_bv);
 346
 347         if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
 348                 return 0;
 349
 350         /*
 351          * bio and nxt are contiguous in memory; check if the queue allows
 352          * these two to be merged into one
 353          */
 354         if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
 355                 return 1;
 356
 357         return 0;
 358 }
 359
 360 static inline void
 361 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
 362                      struct scatterlist *sglist, struct bio_vec *bvprv,
 363                      struct scatterlist **sg, int *nsegs, int *cluster)
 364 {
 365
 366         int nbytes = bvec->bv_len;
 367
 368         if (*sg && *cluster) {
 369                 if ((*sg)->length + nbytes > queue_max_segment_size(q))
 370                         goto new_segment;
 371
 372                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 373                         goto new_segment;
 374                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 375                         goto new_segment;
 376
 377                 (*sg)->length += nbytes;
 378         } else {
 379 new_segment:
 380                 if (!*sg)
 381                         *sg = sglist;
 382                 else {
 383                         /*
 384                          * If the driver previously mapped a shorter
 385                          * list, we could see a termination bit
 386                          * prematurely unless it fully inits the sg
 387                          * table on each mapping. We KNOW that there
 388                          * must be more entries here or the driver
 389                          * would be buggy, so force clear the
 390                          * termination bit to avoid doing a full
 391                          * sg_init_table() in drivers for each command.
 392                          */
 393                         sg_unmark_end(*sg);
 394                         *sg = sg_next(*sg);
 395                 }
 396
 397                 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
 398                 (*nsegs)++;
 399         }
 400         *bvprv = *bvec;
 401 }
 402
 403 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
 404                 struct scatterlist *sglist, struct scatterlist **sg)
 405 {
 406         *sg = sglist;
 407         sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
 408         return 1;
 409 }
 410
 411 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 412                              struct scatterlist *sglist,
 413                              struct scatterlist **sg)
 414 {
 415         struct bio_vec bvec, bvprv = { NULL };
 416         struct bvec_iter iter;
 417         int cluster = blk_queue_cluster(q), nsegs = 0;
 418
 419         for_each_bio(bio)
 420                 bio_for_each_segment(bvec, bio, iter)
 421                         __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
 422                                              &nsegs, &cluster);
 423
 424         return nsegs;
 425 }
 426
 427 /*
 428  * map a request to scatterlist, return number of sg entries setup. Caller
 429  * must make sure sg can hold rq->nr_phys_segments entries
 430  */
 431 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 432                   struct scatterlist *sglist)
 433 {
 434         struct scatterlist *sg = NULL;
 435         int nsegs = 0;
 436
 437         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
 438                 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
 439         else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
 440                 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
 441         else if (rq->bio)
 442                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
 443
 444         if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
 445             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 446                 unsigned int pad_len =
 447                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 448
 449                 sg->length += pad_len;
 450                 rq->extra_len += pad_len;
 451         }
 452
 453         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 454                 if (op_is_write(req_op(rq)))
 455                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 456
 457                 sg_unmark_end(sg);
 458                 sg = sg_next(sg);
 459                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 460                             q->dma_drain_size,
 461                             ((unsigned long)q->dma_drain_buffer) &
 462                             (PAGE_SIZE - 1));
 463                 nsegs++;
 464                 rq->extra_len += q->dma_drain_size;
 465         }
 466
 467         if (sg)
 468                 sg_mark_end(sg);
 469
 470         /*
 471          * Something must have been wrong if the figured number of
 472          * segment is bigger than number of req's physical segments
 473          */
 474         WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
 475
 476         return nsegs;
 477 }
 478 EXPORT_SYMBOL(blk_rq_map_sg);
 479
 480 static inline int ll_new_hw_segment(struct request_queue *q,
 481                                     struct request *req,
 482                                     struct bio *bio)
 483 {
 484         int nr_phys_segs = bio_phys_segments(q, bio);
 485
 486         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 487                 goto no_merge;
 488
 489         if (blk_integrity_merge_bio(q, req, bio) == false)
 490                 goto no_merge;
 491
 492         /*
 493          * This will form the start of a new hw segment.  Bump both
 494          * counters.
 495          */
 496         req->nr_phys_segments += nr_phys_segs;
 497         return 1;
 498
 499 no_merge:
 500         req_set_nomerge(q, req);
 501         return 0;
 502 }
 503
 504 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 505                      struct bio *bio)
 506 {
 507         if (req_gap_back_merge(req, bio))
 508                 return 0;
 509         if (blk_integrity_rq(req) &&
 510             integrity_req_gap_back_merge(req, bio))
 511                 return 0;
 512         if (blk_rq_sectors(req) + bio_sectors(bio) >
 513             blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
 514                 req_set_nomerge(q, req);
 515                 return 0;
 516         }
 517         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 518                 blk_recount_segments(q, req->biotail);
 519         if (!bio_flagged(bio, BIO_SEG_VALID))
 520                 blk_recount_segments(q, bio);
 521
 522         return ll_new_hw_segment(q, req, bio);
 523 }
 524
 525 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 526                       struct bio *bio)
 527 {
 528
 529         if (req_gap_front_merge(req, bio))
 530                 return 0;
 531         if (blk_integrity_rq(req) &&
 532             integrity_req_gap_front_merge(req, bio))
 533                 return 0;
 534         if (blk_rq_sectors(req) + bio_sectors(bio) >
 535             blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
 536                 req_set_nomerge(q, req);
 537                 return 0;
 538         }
 539         if (!bio_flagged(bio, BIO_SEG_VALID))
 540                 blk_recount_segments(q, bio);
 541         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 542                 blk_recount_segments(q, req->bio);
 543
 544         return ll_new_hw_segment(q, req, bio);
 545 }
 546
 547 /*
 548  * blk-mq uses req->special to carry normal driver per-request payload, it
 549  * does not indicate a prepared command that we cannot merge with.
 550  */
 551 static bool req_no_special_merge(struct request *req)
 552 {
 553         struct request_queue *q = req->q;
 554
 555         return !q->mq_ops && req->special;
 556 }
 557
 558 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
 559                 struct request *next)
 560 {
 561         unsigned short segments = blk_rq_nr_discard_segments(req);
 562
 563         if (segments >= queue_max_discard_segments(q))
 564                 goto no_merge;
 565         if (blk_rq_sectors(req) + bio_sectors(next->bio) >
 566             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 567                 goto no_merge;
 568
 569         req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
 570         return true;
 571 no_merge:
 572         req_set_nomerge(q, req);
 573         return false;
 574 }
 575
 576 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 577                                 struct request *next)
 578 {
 579         int total_phys_segments;
 580         unsigned int seg_size =
 581                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 582
 583         /*
 584          * First check if the either of the requests are re-queued
 585          * requests.  Can't merge them if they are.
 586          */
 587         if (req_no_special_merge(req) || req_no_special_merge(next))
 588                 return 0;
 589
 590         if (req_gap_back_merge(req, next->bio))
 591                 return 0;
 592
 593         /*
 594          * Will it become too large?
 595          */
 596         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 597             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 598                 return 0;
 599
 600         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 601         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 602                 if (req->nr_phys_segments == 1)
 603                         req->bio->bi_seg_front_size = seg_size;
 604                 if (next->nr_phys_segments == 1)
 605                         next->biotail->bi_seg_back_size = seg_size;
 606                 total_phys_segments--;
 607         }
 608
 609         if (total_phys_segments > queue_max_segments(q))
 610                 return 0;
 611
 612         if (blk_integrity_merge_rq(q, req, next) == false)
 613                 return 0;
 614
 615         /* Merge is OK... */
 616         req->nr_phys_segments = total_phys_segments;
 617         return 1;
 618 }
 619
 620 /**
 621  * blk_rq_set_mixed_merge - mark a request as mixed merge
 622  * @rq: request to mark as mixed merge
 623  *
 624  * Description:
 625  *     @rq is about to be mixed merged.  Make sure the attributes
 626  *     which can be mixed are set in each bio and mark @rq as mixed
 627  *     merged.
 628  */
 629 void blk_rq_set_mixed_merge(struct request *rq)
 630 {
 631         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 632         struct bio *bio;
 633
 634         if (rq->rq_flags & RQF_MIXED_MERGE)
 635                 return;
 636
 637         /*
 638          * @rq will no longer represent mixable attributes for all the
 639          * contained bios.  It will just track those of the first one.
 640          * Distributes the attributs to each bio.
 641          */
 642         for (bio = rq->bio; bio; bio = bio->bi_next) {
 643                 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
 644                              (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
 645                 bio->bi_opf |= ff;
 646         }
 647         rq->rq_flags |= RQF_MIXED_MERGE;
 648 }
 649
 650 static void blk_account_io_merge(struct request *req)
 651 {
 652         if (blk_do_io_stat(req)) {
 653                 struct hd_struct *part;
 654                 int cpu;
 655
 656                 cpu = part_stat_lock();
 657                 part = req->part;
 658
 659                 part_round_stats(req->q, cpu, part);
 660                 part_dec_in_flight(req->q, part, rq_data_dir(req));
 661
 662                 hd_struct_put(part);
 663                 part_stat_unlock();
 664         }
 665 }
 666 /*
 667  * Two cases of handling DISCARD merge:
 668  * If max_discard_segments > 1, the driver takes every bio
 669  * as a range and send them to controller together. The ranges
 670  * needn't to be contiguous.
 671  * Otherwise, the bios/requests will be handled as same as
 672  * others which should be contiguous.
 673  */
 674 static inline bool blk_discard_mergable(struct request *req)
 675 {
 676         if (req_op(req) == REQ_OP_DISCARD &&
 677             queue_max_discard_segments(req->q) > 1)
 678                 return true;
 679         return false;
 680 }
 681
 682 enum elv_merge blk_try_req_merge(struct request *req, struct request *next)
 683 {
 684         if (blk_discard_mergable(req))
 685                 return ELEVATOR_DISCARD_MERGE;
 686         else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
 687                 return ELEVATOR_BACK_MERGE;
 688
 689         return ELEVATOR_NO_MERGE;
 690 }
 691
 692 /*
 693  * For non-mq, this has to be called with the request spinlock acquired.
 694  * For mq with scheduling, the appropriate queue wide lock should be held.
 695  */
 696 static struct request *attempt_merge(struct request_queue *q,
 697                                      struct request *req, struct request *next)
 698 {
 699         if (!q->mq_ops)
 700                 lockdep_assert_held(q->queue_lock);
 701
 702         if (!rq_mergeable(req) || !rq_mergeable(next))
 703                 return NULL;
 704
 705         if (req_op(req) != req_op(next))
 706                 return NULL;
 707
 708         if (rq_data_dir(req) != rq_data_dir(next)
 709             || req->rq_disk != next->rq_disk
 710             || req_no_special_merge(next))
 711                 return NULL;
 712
 713         if (req_op(req) == REQ_OP_WRITE_SAME &&
 714             !blk_write_same_mergeable(req->bio, next->bio))
 715                 return NULL;
 716
 717         /*
 718          * Don't allow merge of different write hints, or for a hint with
 719          * non-hint IO.
 720          */
 721         if (req->write_hint != next->write_hint)
 722                 return NULL;
 723
 724         /*
 725          * If we are allowed to merge, then append bio list
 726          * from next to rq and release next. merge_requests_fn
 727          * will have updated segment counts, update sector
 728          * counts here. Handle DISCARDs separately, as they
 729          * have separate settings.
 730          */
 731
 732         switch (blk_try_req_merge(req, next)) {
 733         case ELEVATOR_DISCARD_MERGE:
 734                 if (!req_attempt_discard_merge(q, req, next))
 735                         return NULL;
 736                 break;
 737         case ELEVATOR_BACK_MERGE:
 738                 if (!ll_merge_requests_fn(q, req, next))
 739                         return NULL;
 740                 break;
 741         default:
 742                 return NULL;
 743         }
 744
 745         /*
 746          * If failfast settings disagree or any of the two is already
 747          * a mixed merge, mark both as mixed before proceeding.  This
 748          * makes sure that all involved bios have mixable attributes
 749          * set properly.
 750          */
 751         if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
 752             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 753             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 754                 blk_rq_set_mixed_merge(req);
 755                 blk_rq_set_mixed_merge(next);
 756         }
 757
 758         /*
 759          * At this point we have either done a back merge or front merge. We
 760          * need the smaller start_time_ns of the merged requests to be the
 761          * current request for accounting purposes.
 762          */
 763         if (next->start_time_ns < req->start_time_ns)
 764                 req->start_time_ns = next->start_time_ns;
 765
 766         req->biotail->bi_next = next->bio;
 767         req->biotail = next->biotail;
 768
 769         req->__data_len += blk_rq_bytes(next);
 770
 771         if (!blk_discard_mergable(req))
 772                 elv_merge_requests(q, req, next);
 773
 774         /*
 775          * 'next' is going away, so update stats accordingly
 776          */
 777         blk_account_io_merge(next);
 778
 779         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 780         if (blk_rq_cpu_valid(next))
 781                 req->cpu = next->cpu;
 782
 783         /*
 784          * ownership of bio passed from next to req, return 'next' for
 785          * the caller to free
 786          */
 787         next->bio = NULL;
 788         return next;
 789 }
 790
 791 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
 792 {
 793         struct request *next = elv_latter_request(q, rq);
 794
 795         if (next)
 796                 return attempt_merge(q, rq, next);
 797
 798         return NULL;
 799 }
 800
 801 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
 802 {
 803         struct request *prev = elv_former_request(q, rq);
 804
 805         if (prev)
 806                 return attempt_merge(q, prev, rq);
 807
 808         return NULL;
 809 }
 810
 811 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 812                           struct request *next)
 813 {
 814         struct elevator_queue *e = q->elevator;
 815         struct request *free;
 816
 817         if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn)
 818                 if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next))
 819                         return 0;
 820
 821         free = attempt_merge(q, rq, next);
 822         if (free) {
 823                 __blk_put_request(q, free);
 824                 return 1;
 825         }
 826
 827         return 0;
 828 }
 829
 830 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 831 {
 832         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 833                 return false;
 834
 835         if (req_op(rq) != bio_op(bio))
 836                 return false;
 837
 838         /* different data direction or already started, don't merge */
 839         if (bio_data_dir(bio) != rq_data_dir(rq))
 840                 return false;
 841
 842         /* must be same device and not a special request */
 843         if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq))
 844                 return false;
 845
 846         /* only merge integrity protected bio into ditto rq */
 847         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 848                 return false;
 849
 850         /* must be using the same buffer */
 851         if (req_op(rq) == REQ_OP_WRITE_SAME &&
 852             !blk_write_same_mergeable(rq->bio, bio))
 853                 return false;
 854
 855         /*
 856          * Don't allow merge of different write hints, or for a hint with
 857          * non-hint IO.
 858          */
 859         if (rq->write_hint != bio->bi_write_hint)
 860                 return false;
 861
 862         return true;
 863 }
 864
 865 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
 866 {
 867         if (blk_discard_mergable(rq))
 868                 return ELEVATOR_DISCARD_MERGE;
 869         else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 870                 return ELEVATOR_BACK_MERGE;
 871         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 872                 return ELEVATOR_FRONT_MERGE;
 873         return ELEVATOR_NO_MERGE;
 874 }