block/blk-merge.c

   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
  14 static 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
  57 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
  58                 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
  59 {
  60         *nsegs = 1;
  61
  62         if (!q->limits.max_write_zeroes_sectors)
  63                 return NULL;
  64
  65         if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
  66                 return NULL;
  67
  68         return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
  69 }
  70
  71 static struct bio *blk_bio_write_same_split(struct request_queue *q,
  72                                             struct bio *bio,
  73                                             struct bio_set *bs,
  74                                             unsigned *nsegs)
  75 {
  76         *nsegs = 1;
  77
  78         if (!q->limits.max_write_same_sectors)
  79                 return NULL;
  80
  81         if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
  82                 return NULL;
  83
  84         return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
  85 }
  86
  87 static inline unsigned get_max_io_size(struct request_queue *q,
  88                                        struct bio *bio)
  89 {
  90         unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
  91         unsigned mask = queue_logical_block_size(q) - 1;
  92
  93         /* aligned to logical block size */
  94         sectors &= ~(mask >> 9);
  95
  96         return sectors;
  97 }
  98
  99 static struct bio *blk_bio_segment_split(struct request_queue *q,
 100                                          struct bio *bio,
 101                                          struct bio_set *bs,
 102                                          unsigned *segs)
 103 {
 104         struct bio_vec bv, bvprv, *bvprvp = NULL;
 105         struct bvec_iter iter;
 106         unsigned seg_size = 0, nsegs = 0, sectors = 0;
 107         unsigned front_seg_size = bio->bi_seg_front_size;
 108         bool do_split = true;
 109         struct bio *new = NULL;
 110         const unsigned max_sectors = get_max_io_size(q, bio);
 111         unsigned bvecs = 0;
 112
 113         bio_for_each_segment(bv, bio, iter) {
 114                 /*
 115                  * With arbitrary bio size, the incoming bio may be very
 116                  * big. We have to split the bio into small bios so that
 117                  * each holds at most BIO_MAX_PAGES bvecs because
 118                  * bio_clone() can fail to allocate big bvecs.
 119                  *
 120                  * It should have been better to apply the limit per
 121                  * request queue in which bio_clone() is involved,
 122                  * instead of globally. The biggest blocker is the
 123                  * bio_clone() in bio bounce.
 124                  *
 125                  * If bio is splitted by this reason, we should have
 126                  * allowed to continue bios merging, but don't do
 127                  * that now for making the change simple.
 128                  *
 129                  * TODO: deal with bio bounce's bio_clone() gracefully
 130                  * and convert the global limit into per-queue limit.
 131                  */
 132                 if (bvecs++ >= BIO_MAX_PAGES)
 133                         goto split;
 134
 135                 /*
 136                  * If the queue doesn't support SG gaps and adding this
 137                  * offset would create a gap, disallow it.
 138                  */
 139                 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
 140                         goto split;
 141
 142                 if (sectors + (bv.bv_len >> 9) > max_sectors) {
 143                         /*
 144                          * Consider this a new segment if we're splitting in
 145                          * the middle of this vector.
 146                          */
 147                         if (nsegs < queue_max_segments(q) &&
 148                             sectors < max_sectors) {
 149                                 nsegs++;
 150                                 sectors = max_sectors;
 151                         }
 152                         if (sectors)
 153                                 goto split;
 154                         /* Make this single bvec as the 1st segment */
 155                 }
 156
 157                 if (bvprvp && blk_queue_cluster(q)) {
 158                         if (seg_size + bv.bv_len > queue_max_segment_size(q))
 159                                 goto new_segment;
 160                         if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
 161                                 goto new_segment;
 162                         if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
 163                                 goto new_segment;
 164
 165                         seg_size += bv.bv_len;
 166                         bvprv = bv;
 167                         bvprvp = &bvprv;
 168                         sectors += bv.bv_len >> 9;
 169
 170                         if (nsegs == 1 && seg_size > front_seg_size)
 171                                 front_seg_size = seg_size;
 172                         continue;
 173                 }
 174 new_segment:
 175                 if (nsegs == queue_max_segments(q))
 176                         goto split;
 177
 178                 nsegs++;
 179                 bvprv = bv;
 180                 bvprvp = &bvprv;
 181                 seg_size = bv.bv_len;
 182                 sectors += bv.bv_len >> 9;
 183
 184                 if (nsegs == 1 && seg_size > front_seg_size)
 185                         front_seg_size = seg_size;
 186         }
 187
 188         do_split = false;
 189 split:
 190         *segs = nsegs;
 191
 192         if (do_split) {
 193                 new = bio_split(bio, sectors, GFP_NOIO, bs);
 194                 if (new)
 195                         bio = new;
 196         }
 197
 198         bio->bi_seg_front_size = front_seg_size;
 199         if (seg_size > bio->bi_seg_back_size)
 200                 bio->bi_seg_back_size = seg_size;
 201
 202         return do_split ? new : NULL;
 203 }
 204
 205 void blk_queue_split(struct request_queue *q, struct bio **bio,
 206                      struct bio_set *bs)
 207 {
 208         struct bio *split, *res;
 209         unsigned nsegs;
 210
 211         switch (bio_op(*bio)) {
 212         case REQ_OP_DISCARD:
 213         case REQ_OP_SECURE_ERASE:
 214                 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
 215                 break;
 216         case REQ_OP_WRITE_ZEROES:
 217                 split = blk_bio_write_zeroes_split(q, *bio, bs, &nsegs);
 218                 break;
 219         case REQ_OP_WRITE_SAME:
 220                 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
 221                 break;
 222         default:
 223                 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
 224                 break;
 225         }
 226
 227         /* physical segments can be figured out during splitting */
 228         res = split ? split : *bio;
 229         res->bi_phys_segments = nsegs;
 230         bio_set_flag(res, BIO_SEG_VALID);
 231
 232         if (split) {
 233                 /* there isn't chance to merge the splitted bio */
 234                 split->bi_opf |= REQ_NOMERGE;
 235
 236                 bio_chain(split, *bio);
 237                 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
 238                 generic_make_request(*bio);
 239                 *bio = split;
 240         }
 241 }
 242 EXPORT_SYMBOL(blk_queue_split);
 243
 244 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
 245                                              struct bio *bio,
 246                                              bool no_sg_merge)
 247 {
 248         struct bio_vec bv, bvprv = { NULL };
 249         int cluster, prev = 0;
 250         unsigned int seg_size, nr_phys_segs;
 251         struct bio *fbio, *bbio;
 252         struct bvec_iter iter;
 253
 254         if (!bio)
 255                 return 0;
 256
 257         switch (bio_op(bio)) {
 258         case REQ_OP_DISCARD:
 259         case REQ_OP_SECURE_ERASE:
 260         case REQ_OP_WRITE_ZEROES:
 261                 return 0;
 262         case REQ_OP_WRITE_SAME:
 263                 return 1;
 264         }
 265
 266         fbio = bio;
 267         cluster = blk_queue_cluster(q);
 268         seg_size = 0;
 269         nr_phys_segs = 0;
 270         for_each_bio(bio) {
 271                 bio_for_each_segment(bv, bio, iter) {
 272                         /*
 273                          * If SG merging is disabled, each bio vector is
 274                          * a segment
 275                          */
 276                         if (no_sg_merge)
 277                                 goto new_segment;
 278
 279                         if (prev && cluster) {
 280                                 if (seg_size + bv.bv_len
 281                                     > queue_max_segment_size(q))
 282                                         goto new_segment;
 283                                 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
 284                                         goto new_segment;
 285                                 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
 286                                         goto new_segment;
 287
 288                                 seg_size += bv.bv_len;
 289                                 bvprv = bv;
 290                                 continue;
 291                         }
 292 new_segment:
 293                         if (nr_phys_segs == 1 && seg_size >
 294                             fbio->bi_seg_front_size)
 295                                 fbio->bi_seg_front_size = seg_size;
 296
 297                         nr_phys_segs++;
 298                         bvprv = bv;
 299                         prev = 1;
 300                         seg_size = bv.bv_len;
 301                 }
 302                 bbio = bio;
 303         }
 304
 305         if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
 306                 fbio->bi_seg_front_size = seg_size;
 307         if (seg_size > bbio->bi_seg_back_size)
 308                 bbio->bi_seg_back_size = seg_size;
 309
 310         return nr_phys_segs;
 311 }
 312
 313 void blk_recalc_rq_segments(struct request *rq)
 314 {
 315         bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
 316                         &rq->q->queue_flags);
 317
 318         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
 319                         no_sg_merge);
 320 }
 321
 322 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 323 {
 324         unsigned short seg_cnt;
 325
 326         /* estimate segment number by bi_vcnt for non-cloned bio */
 327         if (bio_flagged(bio, BIO_CLONED))
 328                 seg_cnt = bio_segments(bio);
 329         else
 330                 seg_cnt = bio->bi_vcnt;
 331
 332         if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
 333                         (seg_cnt < queue_max_segments(q)))
 334                 bio->bi_phys_segments = seg_cnt;
 335         else {
 336                 struct bio *nxt = bio->bi_next;
 337
 338                 bio->bi_next = NULL;
 339                 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
 340                 bio->bi_next = nxt;
 341         }
 342
 343         bio_set_flag(bio, BIO_SEG_VALID);
 344 }
 345 EXPORT_SYMBOL(blk_recount_segments);
 346
 347 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 348                                    struct bio *nxt)
 349 {
 350         struct bio_vec end_bv = { NULL }, nxt_bv;
 351
 352         if (!blk_queue_cluster(q))
 353                 return 0;
 354
 355         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 356             queue_max_segment_size(q))
 357                 return 0;
 358
 359         if (!bio_has_data(bio))
 360                 return 1;
 361
 362         bio_get_last_bvec(bio, &end_bv);
 363         bio_get_first_bvec(nxt, &nxt_bv);
 364
 365         if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
 366                 return 0;
 367
 368         /*
 369          * bio and nxt are contiguous in memory; check if the queue allows
 370          * these two to be merged into one
 371          */
 372         if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
 373                 return 1;
 374
 375         return 0;
 376 }
 377
 378 static inline void
 379 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
 380                      struct scatterlist *sglist, struct bio_vec *bvprv,
 381                      struct scatterlist **sg, int *nsegs, int *cluster)
 382 {
 383
 384         int nbytes = bvec->bv_len;
 385
 386         if (*sg && *cluster) {
 387                 if ((*sg)->length + nbytes > queue_max_segment_size(q))
 388                         goto new_segment;
 389
 390                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 391                         goto new_segment;
 392                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 393                         goto new_segment;
 394
 395                 (*sg)->length += nbytes;
 396         } else {
 397 new_segment:
 398                 if (!*sg)
 399                         *sg = sglist;
 400                 else {
 401                         /*
 402                          * If the driver previously mapped a shorter
 403                          * list, we could see a termination bit
 404                          * prematurely unless it fully inits the sg
 405                          * table on each mapping. We KNOW that there
 406                          * must be more entries here or the driver
 407                          * would be buggy, so force clear the
 408                          * termination bit to avoid doing a full
 409                          * sg_init_table() in drivers for each command.
 410                          */
 411                         sg_unmark_end(*sg);
 412                         *sg = sg_next(*sg);
 413                 }
 414
 415                 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
 416                 (*nsegs)++;
 417         }
 418         *bvprv = *bvec;
 419 }
 420
 421 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
 422                 struct scatterlist *sglist, struct scatterlist **sg)
 423 {
 424         *sg = sglist;
 425         sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
 426         return 1;
 427 }
 428
 429 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 430                              struct scatterlist *sglist,
 431                              struct scatterlist **sg)
 432 {
 433         struct bio_vec bvec, bvprv = { NULL };
 434         struct bvec_iter iter;
 435         int cluster = blk_queue_cluster(q), nsegs = 0;
 436
 437         for_each_bio(bio)
 438                 bio_for_each_segment(bvec, bio, iter)
 439                         __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
 440                                              &nsegs, &cluster);
 441
 442         return nsegs;
 443 }
 444
 445 /*
 446  * map a request to scatterlist, return number of sg entries setup. Caller
 447  * must make sure sg can hold rq->nr_phys_segments entries
 448  */
 449 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 450                   struct scatterlist *sglist)
 451 {
 452         struct scatterlist *sg = NULL;
 453         int nsegs = 0;
 454
 455         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
 456                 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
 457         else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
 458                 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
 459         else if (rq->bio)
 460                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
 461
 462         if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
 463             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 464                 unsigned int pad_len =
 465                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 466
 467                 sg->length += pad_len;
 468                 rq->extra_len += pad_len;
 469         }
 470
 471         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 472                 if (op_is_write(req_op(rq)))
 473                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 474
 475                 sg_unmark_end(sg);
 476                 sg = sg_next(sg);
 477                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 478                             q->dma_drain_size,
 479                             ((unsigned long)q->dma_drain_buffer) &
 480                             (PAGE_SIZE - 1));
 481                 nsegs++;
 482                 rq->extra_len += q->dma_drain_size;
 483         }
 484
 485         if (sg)
 486                 sg_mark_end(sg);
 487
 488         /*
 489          * Something must have been wrong if the figured number of
 490          * segment is bigger than number of req's physical segments
 491          */
 492         WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
 493
 494         return nsegs;
 495 }
 496 EXPORT_SYMBOL(blk_rq_map_sg);
 497
 498 static inline int ll_new_hw_segment(struct request_queue *q,
 499                                     struct request *req,
 500                                     struct bio *bio)
 501 {
 502         int nr_phys_segs = bio_phys_segments(q, bio);
 503
 504         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 505                 goto no_merge;
 506
 507         if (blk_integrity_merge_bio(q, req, bio) == false)
 508                 goto no_merge;
 509
 510         /*
 511          * This will form the start of a new hw segment.  Bump both
 512          * counters.
 513          */
 514         req->nr_phys_segments += nr_phys_segs;
 515         return 1;
 516
 517 no_merge:
 518         req_set_nomerge(q, req);
 519         return 0;
 520 }
 521
 522 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 523                      struct bio *bio)
 524 {
 525         if (req_gap_back_merge(req, bio))
 526                 return 0;
 527         if (blk_integrity_rq(req) &&
 528             integrity_req_gap_back_merge(req, bio))
 529                 return 0;
 530         if (blk_rq_sectors(req) + bio_sectors(bio) >
 531             blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
 532                 req_set_nomerge(q, req);
 533                 return 0;
 534         }
 535         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 536                 blk_recount_segments(q, req->biotail);
 537         if (!bio_flagged(bio, BIO_SEG_VALID))
 538                 blk_recount_segments(q, bio);
 539
 540         return ll_new_hw_segment(q, req, bio);
 541 }
 542
 543 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 544                       struct bio *bio)
 545 {
 546
 547         if (req_gap_front_merge(req, bio))
 548                 return 0;
 549         if (blk_integrity_rq(req) &&
 550             integrity_req_gap_front_merge(req, bio))
 551                 return 0;
 552         if (blk_rq_sectors(req) + bio_sectors(bio) >
 553             blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
 554                 req_set_nomerge(q, req);
 555                 return 0;
 556         }
 557         if (!bio_flagged(bio, BIO_SEG_VALID))
 558                 blk_recount_segments(q, bio);
 559         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 560                 blk_recount_segments(q, req->bio);
 561
 562         return ll_new_hw_segment(q, req, bio);
 563 }
 564
 565 /*
 566  * blk-mq uses req->special to carry normal driver per-request payload, it
 567  * does not indicate a prepared command that we cannot merge with.
 568  */
 569 static bool req_no_special_merge(struct request *req)
 570 {
 571         struct request_queue *q = req->q;
 572
 573         return !q->mq_ops && req->special;
 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(cpu, part);
 660                 part_dec_in_flight(part, rq_data_dir(req));
 661
 662                 hd_struct_put(part);
 663                 part_stat_unlock();
 664         }
 665 }
 666
 667 /*
 668  * For non-mq, this has to be called with the request spinlock acquired.
 669  * For mq with scheduling, the appropriate queue wide lock should be held.
 670  */
 671 static struct request *attempt_merge(struct request_queue *q,
 672                                      struct request *req, struct request *next)
 673 {
 674         if (!rq_mergeable(req) || !rq_mergeable(next))
 675                 return NULL;
 676
 677         if (req_op(req) != req_op(next))
 678                 return NULL;
 679
 680         /*
 681          * not contiguous
 682          */
 683         if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
 684                 return NULL;
 685
 686         if (rq_data_dir(req) != rq_data_dir(next)
 687             || req->rq_disk != next->rq_disk
 688             || req_no_special_merge(next))
 689                 return NULL;
 690
 691         if (req_op(req) == REQ_OP_WRITE_SAME &&
 692             !blk_write_same_mergeable(req->bio, next->bio))
 693                 return NULL;
 694
 695         /*
 696          * If we are allowed to merge, then append bio list
 697          * from next to rq and release next. merge_requests_fn
 698          * will have updated segment counts, update sector
 699          * counts here.
 700          */
 701         if (!ll_merge_requests_fn(q, req, next))
 702                 return NULL;
 703
 704         /*
 705          * If failfast settings disagree or any of the two is already
 706          * a mixed merge, mark both as mixed before proceeding.  This
 707          * makes sure that all involved bios have mixable attributes
 708          * set properly.
 709          */
 710         if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
 711             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 712             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 713                 blk_rq_set_mixed_merge(req);
 714                 blk_rq_set_mixed_merge(next);
 715         }
 716
 717         /*
 718          * At this point we have either done a back merge
 719          * or front merge. We need the smaller start_time of
 720          * the merged requests to be the current request
 721          * for accounting purposes.
 722          */
 723         if (time_after(req->start_time, next->start_time))
 724                 req->start_time = next->start_time;
 725
 726         req->biotail->bi_next = next->bio;
 727         req->biotail = next->biotail;
 728
 729         req->__data_len += blk_rq_bytes(next);
 730
 731         elv_merge_requests(q, req, next);
 732
 733         /*
 734          * 'next' is going away, so update stats accordingly
 735          */
 736         blk_account_io_merge(next);
 737
 738         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 739         if (blk_rq_cpu_valid(next))
 740                 req->cpu = next->cpu;
 741
 742         /*
 743          * ownership of bio passed from next to req, return 'next' for
 744          * the caller to free
 745          */
 746         next->bio = NULL;
 747         return next;
 748 }
 749
 750 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
 751 {
 752         struct request *next = elv_latter_request(q, rq);
 753
 754         if (next)
 755                 return attempt_merge(q, rq, next);
 756
 757         return NULL;
 758 }
 759
 760 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
 761 {
 762         struct request *prev = elv_former_request(q, rq);
 763
 764         if (prev)
 765                 return attempt_merge(q, prev, rq);
 766
 767         return NULL;
 768 }
 769
 770 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 771                           struct request *next)
 772 {
 773         struct elevator_queue *e = q->elevator;
 774         struct request *free;
 775
 776         if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn)
 777                 if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next))
 778                         return 0;
 779
 780         free = attempt_merge(q, rq, next);
 781         if (free) {
 782                 __blk_put_request(q, free);
 783                 return 1;
 784         }
 785
 786         return 0;
 787 }
 788
 789 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 790 {
 791         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 792                 return false;
 793
 794         if (req_op(rq) != bio_op(bio))
 795                 return false;
 796
 797         /* different data direction or already started, don't merge */
 798         if (bio_data_dir(bio) != rq_data_dir(rq))
 799                 return false;
 800
 801         /* must be same device and not a special request */
 802         if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
 803                 return false;
 804
 805         /* only merge integrity protected bio into ditto rq */
 806         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 807                 return false;
 808
 809         /* must be using the same buffer */
 810         if (req_op(rq) == REQ_OP_WRITE_SAME &&
 811             !blk_write_same_mergeable(rq->bio, bio))
 812                 return false;
 813
 814         return true;
 815 }
 816
 817 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
 818 {
 819         if (req_op(rq) == REQ_OP_DISCARD &&
 820             queue_max_discard_segments(rq->q) > 1)
 821                 return ELEVATOR_DISCARD_MERGE;
 822         else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 823                 return ELEVATOR_BACK_MERGE;
 824         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 825                 return ELEVATOR_FRONT_MERGE;
 826         return ELEVATOR_NO_MERGE;
 827 }