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 /*
  16  * Check if the two bvecs from two bios can be merged to one segment.  If yes,
  17  * no need to check gap between the two bios since the 1st bio and the 1st bvec
  18  * in the 2nd bio can be handled in one segment.
  19  */
  20 static inline bool bios_segs_mergeable(struct request_queue *q,
  21                 struct bio *prev, struct bio_vec *prev_last_bv,
  22                 struct bio_vec *next_first_bv)
  23 {
  24         if (!biovec_phys_mergeable(q, prev_last_bv, next_first_bv))
  25                 return false;
  26         if (prev->bi_seg_back_size + next_first_bv->bv_len >
  27                         queue_max_segment_size(q))
  28                 return false;
  29         return true;
  30 }
  31
  32 static inline bool bio_will_gap(struct request_queue *q,
  33                 struct request *prev_rq, struct bio *prev, struct bio *next)
  34 {
  35         struct bio_vec pb, nb;
  36
  37         if (!bio_has_data(prev) || !queue_virt_boundary(q))
  38                 return false;
  39
  40         /*
  41          * Don't merge if the 1st bio starts with non-zero offset, otherwise it
  42          * is quite difficult to respect the sg gap limit.  We work hard to
  43          * merge a huge number of small single bios in case of mkfs.
  44          */
  45         if (prev_rq)
  46                 bio_get_first_bvec(prev_rq->bio, &pb);
  47         else
  48                 bio_get_first_bvec(prev, &pb);
  49         if (pb.bv_offset & queue_virt_boundary(q))
  50                 return true;
  51
  52         /*
  53          * We don't need to worry about the situation that the merged segment
  54          * ends in unaligned virt boundary:
  55          *
  56          * - if 'pb' ends aligned, the merged segment ends aligned
  57          * - if 'pb' ends unaligned, the next bio must include
  58          *   one single bvec of 'nb', otherwise the 'nb' can't
  59          *   merge with 'pb'
  60          */
  61         bio_get_last_bvec(prev, &pb);
  62         bio_get_first_bvec(next, &nb);
  63         if (bios_segs_mergeable(q, prev, &pb, &nb))
  64                 return false;
  65         return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
  66 }
  67
  68 static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
  69 {
  70         return bio_will_gap(req->q, req, req->biotail, bio);
  71 }
  72
  73 static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
  74 {
  75         return bio_will_gap(req->q, NULL, bio, req->bio);
  76 }
  77
  78 static struct bio *blk_bio_discard_split(struct request_queue *q,
  79                                          struct bio *bio,
  80                                          struct bio_set *bs,
  81                                          unsigned *nsegs)
  82 {
  83         unsigned int max_discard_sectors, granularity;
  84         int alignment;
  85         sector_t tmp;
  86         unsigned split_sectors;
  87
  88         *nsegs = 1;
  89
  90         /* Zero-sector (unknown) and one-sector granularities are the same.  */
  91         granularity = max(q->limits.discard_granularity >> 9, 1U);
  92
  93         max_discard_sectors = min(q->limits.max_discard_sectors,
  94                         bio_allowed_max_sectors(q));
  95         max_discard_sectors -= max_discard_sectors % granularity;
  96
  97         if (unlikely(!max_discard_sectors)) {
  98                 /* XXX: warn */
  99                 return NULL;
 100         }
 101
 102         if (bio_sectors(bio) <= max_discard_sectors)
 103                 return NULL;
 104
 105         split_sectors = max_discard_sectors;
 106
 107         /*
 108          * If the next starting sector would be misaligned, stop the discard at
 109          * the previous aligned sector.
 110          */
 111         alignment = (q->limits.discard_alignment >> 9) % granularity;
 112
 113         tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
 114         tmp = sector_div(tmp, granularity);
 115
 116         if (split_sectors > tmp)
 117                 split_sectors -= tmp;
 118
 119         return bio_split(bio, split_sectors, GFP_NOIO, bs);
 120 }
 121
 122 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
 123                 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
 124 {
 125         *nsegs = 1;
 126
 127         if (!q->limits.max_write_zeroes_sectors)
 128                 return NULL;
 129
 130         if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
 131                 return NULL;
 132
 133         return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
 134 }
 135
 136 static struct bio *blk_bio_write_same_split(struct request_queue *q,
 137                                             struct bio *bio,
 138                                             struct bio_set *bs,
 139                                             unsigned *nsegs)
 140 {
 141         *nsegs = 1;
 142
 143         if (!q->limits.max_write_same_sectors)
 144                 return NULL;
 145
 146         if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
 147                 return NULL;
 148
 149         return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
 150 }
 151
 152 static inline unsigned get_max_io_size(struct request_queue *q,
 153                                        struct bio *bio)
 154 {
 155         unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
 156         unsigned mask = queue_logical_block_size(q) - 1;
 157
 158         /* aligned to logical block size */
 159         sectors &= ~(mask >> 9);
 160
 161         return sectors;
 162 }
 163
 164 static unsigned get_max_segment_size(struct request_queue *q,
 165                                      unsigned offset)
 166 {
 167         unsigned long mask = queue_segment_boundary(q);
 168
 169         /* default segment boundary mask means no boundary limit */
 170         if (mask == BLK_SEG_BOUNDARY_MASK)
 171                 return queue_max_segment_size(q);
 172
 173         return min_t(unsigned long, mask - (mask & offset) + 1,
 174                      queue_max_segment_size(q));
 175 }
 176
 177 /*
 178  * Split the bvec @bv into segments, and update all kinds of
 179  * variables.
 180  */
 181 static bool bvec_split_segs(struct request_queue *q, struct bio_vec *bv,
 182                 unsigned *nsegs, unsigned *last_seg_size,
 183                 unsigned *front_seg_size, unsigned *sectors)
 184 {
 185         unsigned len = bv->bv_len;
 186         unsigned total_len = 0;
 187         unsigned new_nsegs = 0, seg_size = 0;
 188
 189         /*
 190          * Multi-page bvec may be too big to hold in one segment, so the
 191          * current bvec has to be splitted as multiple segments.
 192          */
 193         while (len && new_nsegs + *nsegs < queue_max_segments(q)) {
 194                 seg_size = get_max_segment_size(q, bv->bv_offset + total_len);
 195                 seg_size = min(seg_size, len);
 196
 197                 new_nsegs++;
 198                 total_len += seg_size;
 199                 len -= seg_size;
 200
 201                 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
 202                         break;
 203         }
 204
 205         if (!new_nsegs)
 206                 return !!len;
 207
 208         /* update front segment size */
 209         if (!*nsegs) {
 210                 unsigned first_seg_size;
 211
 212                 if (new_nsegs == 1)
 213                         first_seg_size = get_max_segment_size(q, bv->bv_offset);
 214                 else
 215                         first_seg_size = queue_max_segment_size(q);
 216
 217                 if (*front_seg_size < first_seg_size)
 218                         *front_seg_size = first_seg_size;
 219         }
 220
 221         /* update other varibles */
 222         *last_seg_size = seg_size;
 223         *nsegs += new_nsegs;
 224         if (sectors)
 225                 *sectors += total_len >> 9;
 226
 227         /* split in the middle of the bvec if len != 0 */
 228         return !!len;
 229 }
 230
 231 static struct bio *blk_bio_segment_split(struct request_queue *q,
 232                                          struct bio *bio,
 233                                          struct bio_set *bs,
 234                                          unsigned *segs)
 235 {
 236         struct bio_vec bv, bvprv, *bvprvp = NULL;
 237         struct bvec_iter iter;
 238         unsigned seg_size = 0, nsegs = 0, sectors = 0;
 239         unsigned front_seg_size = bio->bi_seg_front_size;
 240         bool do_split = true;
 241         struct bio *new = NULL;
 242         const unsigned max_sectors = get_max_io_size(q, bio);
 243
 244         bio_for_each_bvec(bv, bio, iter) {
 245                 /*
 246                  * If the queue doesn't support SG gaps and adding this
 247                  * offset would create a gap, disallow it.
 248                  */
 249                 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
 250                         goto split;
 251
 252                 if (sectors + (bv.bv_len >> 9) > max_sectors) {
 253                         /*
 254                          * Consider this a new segment if we're splitting in
 255                          * the middle of this vector.
 256                          */
 257                         if (nsegs < queue_max_segments(q) &&
 258                             sectors < max_sectors) {
 259                                 /* split in the middle of bvec */
 260                                 bv.bv_len = (max_sectors - sectors) << 9;
 261                                 bvec_split_segs(q, &bv, &nsegs,
 262                                                 &seg_size,
 263                                                 &front_seg_size,
 264                                                 &sectors);
 265                         }
 266                         goto split;
 267                 }
 268
 269                 if (bvprvp) {
 270                         if (seg_size + bv.bv_len > queue_max_segment_size(q))
 271                                 goto new_segment;
 272                         if (!biovec_phys_mergeable(q, bvprvp, &bv))
 273                                 goto new_segment;
 274
 275                         seg_size += bv.bv_len;
 276                         bvprv = bv;
 277                         bvprvp = &bvprv;
 278                         sectors += bv.bv_len >> 9;
 279
 280                         continue;
 281                 }
 282 new_segment:
 283                 if (nsegs == queue_max_segments(q))
 284                         goto split;
 285
 286                 bvprv = bv;
 287                 bvprvp = &bvprv;
 288
 289                 if (bvec_split_segs(q, &bv, &nsegs, &seg_size,
 290                                     &front_seg_size, &sectors))
 291                         goto split;
 292
 293         }
 294
 295         do_split = false;
 296 split:
 297         *segs = nsegs;
 298
 299         if (do_split) {
 300                 new = bio_split(bio, sectors, GFP_NOIO, bs);
 301                 if (new)
 302                         bio = new;
 303         }
 304
 305         bio->bi_seg_front_size = front_seg_size;
 306         if (seg_size > bio->bi_seg_back_size)
 307                 bio->bi_seg_back_size = seg_size;
 308
 309         return do_split ? new : NULL;
 310 }
 311
 312 void blk_queue_split(struct request_queue *q, struct bio **bio)
 313 {
 314         struct bio *split, *res;
 315         unsigned nsegs;
 316
 317         switch (bio_op(*bio)) {
 318         case REQ_OP_DISCARD:
 319         case REQ_OP_SECURE_ERASE:
 320                 split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
 321                 break;
 322         case REQ_OP_WRITE_ZEROES:
 323                 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
 324                 break;
 325         case REQ_OP_WRITE_SAME:
 326                 split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
 327                 break;
 328         default:
 329                 split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
 330                 break;
 331         }
 332
 333         /* physical segments can be figured out during splitting */
 334         res = split ? split : *bio;
 335         res->bi_phys_segments = nsegs;
 336         bio_set_flag(res, BIO_SEG_VALID);
 337
 338         if (split) {
 339                 /* there isn't chance to merge the splitted bio */
 340                 split->bi_opf |= REQ_NOMERGE;
 341
 342                 /*
 343                  * Since we're recursing into make_request here, ensure
 344                  * that we mark this bio as already having entered the queue.
 345                  * If not, and the queue is going away, we can get stuck
 346                  * forever on waiting for the queue reference to drop. But
 347                  * that will never happen, as we're already holding a
 348                  * reference to it.
 349                  */
 350                 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
 351
 352                 bio_chain(split, *bio);
 353                 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
 354                 generic_make_request(*bio);
 355                 *bio = split;
 356         }
 357 }
 358 EXPORT_SYMBOL(blk_queue_split);
 359
 360 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
 361                                              struct bio *bio)
 362 {
 363         struct bio_vec bv, bvprv = { NULL };
 364         int prev = 0;
 365         unsigned int seg_size, nr_phys_segs;
 366         unsigned front_seg_size;
 367         struct bio *fbio, *bbio;
 368         struct bvec_iter iter;
 369
 370         if (!bio)
 371                 return 0;
 372
 373         front_seg_size = bio->bi_seg_front_size;
 374
 375         switch (bio_op(bio)) {
 376         case REQ_OP_DISCARD:
 377         case REQ_OP_SECURE_ERASE:
 378         case REQ_OP_WRITE_ZEROES:
 379                 return 0;
 380         case REQ_OP_WRITE_SAME:
 381                 return 1;
 382         }
 383
 384         fbio = bio;
 385         seg_size = 0;
 386         nr_phys_segs = 0;
 387         for_each_bio(bio) {
 388                 bio_for_each_bvec(bv, bio, iter) {
 389                         if (prev) {
 390                                 if (seg_size + bv.bv_len
 391                                     > queue_max_segment_size(q))
 392                                         goto new_segment;
 393                                 if (!biovec_phys_mergeable(q, &bvprv, &bv))
 394                                         goto new_segment;
 395
 396                                 seg_size += bv.bv_len;
 397                                 bvprv = bv;
 398                                 continue;
 399                         }
 400 new_segment:
 401                         bvprv = bv;
 402                         prev = 1;
 403                         bvec_split_segs(q, &bv, &nr_phys_segs, &seg_size,
 404                                         &front_seg_size, NULL);
 405                 }
 406                 bbio = bio;
 407         }
 408
 409         fbio->bi_seg_front_size = front_seg_size;
 410         if (seg_size > bbio->bi_seg_back_size)
 411                 bbio->bi_seg_back_size = seg_size;
 412
 413         return nr_phys_segs;
 414 }
 415
 416 void blk_recalc_rq_segments(struct request *rq)
 417 {
 418         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
 419 }
 420
 421 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 422 {
 423         struct bio *nxt = bio->bi_next;
 424
 425         bio->bi_next = NULL;
 426         bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
 427         bio->bi_next = nxt;
 428
 429         bio_set_flag(bio, BIO_SEG_VALID);
 430 }
 431
 432 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 433                                    struct bio *nxt)
 434 {
 435         struct bio_vec end_bv = { NULL }, nxt_bv;
 436
 437         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 438             queue_max_segment_size(q))
 439                 return 0;
 440
 441         if (!bio_has_data(bio))
 442                 return 1;
 443
 444         bio_get_last_bvec(bio, &end_bv);
 445         bio_get_first_bvec(nxt, &nxt_bv);
 446
 447         return biovec_phys_mergeable(q, &end_bv, &nxt_bv);
 448 }
 449
 450 static struct scatterlist *blk_next_sg(struct scatterlist **sg,
 451                 struct scatterlist *sglist)
 452 {
 453         if (!*sg)
 454                 return sglist;
 455
 456         /*
 457          * If the driver previously mapped a shorter list, we could see a
 458          * termination bit prematurely unless it fully inits the sg table
 459          * on each mapping. We KNOW that there must be more entries here
 460          * or the driver would be buggy, so force clear the termination bit
 461          * to avoid doing a full sg_init_table() in drivers for each command.
 462          */
 463         sg_unmark_end(*sg);
 464         return sg_next(*sg);
 465 }
 466
 467 static unsigned blk_bvec_map_sg(struct request_queue *q,
 468                 struct bio_vec *bvec, struct scatterlist *sglist,
 469                 struct scatterlist **sg)
 470 {
 471         unsigned nbytes = bvec->bv_len;
 472         unsigned nsegs = 0, total = 0, offset = 0;
 473
 474         while (nbytes > 0) {
 475                 unsigned seg_size;
 476                 struct page *pg;
 477                 unsigned idx;
 478
 479                 *sg = blk_next_sg(sg, sglist);
 480
 481                 seg_size = get_max_segment_size(q, bvec->bv_offset + total);
 482                 seg_size = min(nbytes, seg_size);
 483
 484                 offset = (total + bvec->bv_offset) % PAGE_SIZE;
 485                 idx = (total + bvec->bv_offset) / PAGE_SIZE;
 486                 pg = nth_page(bvec->bv_page, idx);
 487
 488                 sg_set_page(*sg, pg, seg_size, offset);
 489
 490                 total += seg_size;
 491                 nbytes -= seg_size;
 492                 nsegs++;
 493         }
 494
 495         return nsegs;
 496 }
 497
 498 static inline void
 499 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
 500                      struct scatterlist *sglist, struct bio_vec *bvprv,
 501                      struct scatterlist **sg, int *nsegs)
 502 {
 503
 504         int nbytes = bvec->bv_len;
 505
 506         if (*sg) {
 507                 if ((*sg)->length + nbytes > queue_max_segment_size(q))
 508                         goto new_segment;
 509                 if (!biovec_phys_mergeable(q, bvprv, bvec))
 510                         goto new_segment;
 511
 512                 (*sg)->length += nbytes;
 513         } else {
 514 new_segment:
 515                 (*nsegs) += blk_bvec_map_sg(q, bvec, sglist, sg);
 516         }
 517         *bvprv = *bvec;
 518 }
 519
 520 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
 521                 struct scatterlist *sglist, struct scatterlist **sg)
 522 {
 523         *sg = sglist;
 524         sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
 525         return 1;
 526 }
 527
 528 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 529                              struct scatterlist *sglist,
 530                              struct scatterlist **sg)
 531 {
 532         struct bio_vec bvec, bvprv = { NULL };
 533         struct bvec_iter iter;
 534         int nsegs = 0;
 535
 536         for_each_bio(bio)
 537                 bio_for_each_bvec(bvec, bio, iter)
 538                         __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
 539                                              &nsegs);
 540
 541         return nsegs;
 542 }
 543
 544 /*
 545  * map a request to scatterlist, return number of sg entries setup. Caller
 546  * must make sure sg can hold rq->nr_phys_segments entries
 547  */
 548 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 549                   struct scatterlist *sglist)
 550 {
 551         struct scatterlist *sg = NULL;
 552         int nsegs = 0;
 553
 554         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
 555                 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
 556         else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
 557                 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
 558         else if (rq->bio)
 559                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
 560
 561         if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
 562             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 563                 unsigned int pad_len =
 564                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 565
 566                 sg->length += pad_len;
 567                 rq->extra_len += pad_len;
 568         }
 569
 570         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 571                 if (op_is_write(req_op(rq)))
 572                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 573
 574                 sg_unmark_end(sg);
 575                 sg = sg_next(sg);
 576                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 577                             q->dma_drain_size,
 578                             ((unsigned long)q->dma_drain_buffer) &
 579                             (PAGE_SIZE - 1));
 580                 nsegs++;
 581                 rq->extra_len += q->dma_drain_size;
 582         }
 583
 584         if (sg)
 585                 sg_mark_end(sg);
 586
 587         /*
 588          * Something must have been wrong if the figured number of
 589          * segment is bigger than number of req's physical segments
 590          */
 591         WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
 592
 593         return nsegs;
 594 }
 595 EXPORT_SYMBOL(blk_rq_map_sg);
 596
 597 static inline int ll_new_hw_segment(struct request_queue *q,
 598                                     struct request *req,
 599                                     struct bio *bio)
 600 {
 601         int nr_phys_segs = bio_phys_segments(q, bio);
 602
 603         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 604                 goto no_merge;
 605
 606         if (blk_integrity_merge_bio(q, req, bio) == false)
 607                 goto no_merge;
 608
 609         /*
 610          * This will form the start of a new hw segment.  Bump both
 611          * counters.
 612          */
 613         req->nr_phys_segments += nr_phys_segs;
 614         return 1;
 615
 616 no_merge:
 617         req_set_nomerge(q, req);
 618         return 0;
 619 }
 620
 621 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 622                      struct bio *bio)
 623 {
 624         if (req_gap_back_merge(req, bio))
 625                 return 0;
 626         if (blk_integrity_rq(req) &&
 627             integrity_req_gap_back_merge(req, bio))
 628                 return 0;
 629         if (blk_rq_sectors(req) + bio_sectors(bio) >
 630             blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
 631                 req_set_nomerge(q, req);
 632                 return 0;
 633         }
 634         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 635                 blk_recount_segments(q, req->biotail);
 636         if (!bio_flagged(bio, BIO_SEG_VALID))
 637                 blk_recount_segments(q, bio);
 638
 639         return ll_new_hw_segment(q, req, bio);
 640 }
 641
 642 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 643                       struct bio *bio)
 644 {
 645
 646         if (req_gap_front_merge(req, bio))
 647                 return 0;
 648         if (blk_integrity_rq(req) &&
 649             integrity_req_gap_front_merge(req, bio))
 650                 return 0;
 651         if (blk_rq_sectors(req) + bio_sectors(bio) >
 652             blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
 653                 req_set_nomerge(q, req);
 654                 return 0;
 655         }
 656         if (!bio_flagged(bio, BIO_SEG_VALID))
 657                 blk_recount_segments(q, bio);
 658         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 659                 blk_recount_segments(q, req->bio);
 660
 661         return ll_new_hw_segment(q, req, bio);
 662 }
 663
 664 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
 665                 struct request *next)
 666 {
 667         unsigned short segments = blk_rq_nr_discard_segments(req);
 668
 669         if (segments >= queue_max_discard_segments(q))
 670                 goto no_merge;
 671         if (blk_rq_sectors(req) + bio_sectors(next->bio) >
 672             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 673                 goto no_merge;
 674
 675         req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
 676         return true;
 677 no_merge:
 678         req_set_nomerge(q, req);
 679         return false;
 680 }
 681
 682 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 683                                 struct request *next)
 684 {
 685         int total_phys_segments;
 686         unsigned int seg_size =
 687                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 688
 689         if (req_gap_back_merge(req, next->bio))
 690                 return 0;
 691
 692         /*
 693          * Will it become too large?
 694          */
 695         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 696             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 697                 return 0;
 698
 699         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 700         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 701                 if (req->nr_phys_segments == 1)
 702                         req->bio->bi_seg_front_size = seg_size;
 703                 if (next->nr_phys_segments == 1)
 704                         next->biotail->bi_seg_back_size = seg_size;
 705                 total_phys_segments--;
 706         }
 707
 708         if (total_phys_segments > queue_max_segments(q))
 709                 return 0;
 710
 711         if (blk_integrity_merge_rq(q, req, next) == false)
 712                 return 0;
 713
 714         /* Merge is OK... */
 715         req->nr_phys_segments = total_phys_segments;
 716         return 1;
 717 }
 718
 719 /**
 720  * blk_rq_set_mixed_merge - mark a request as mixed merge
 721  * @rq: request to mark as mixed merge
 722  *
 723  * Description:
 724  *     @rq is about to be mixed merged.  Make sure the attributes
 725  *     which can be mixed are set in each bio and mark @rq as mixed
 726  *     merged.
 727  */
 728 void blk_rq_set_mixed_merge(struct request *rq)
 729 {
 730         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 731         struct bio *bio;
 732
 733         if (rq->rq_flags & RQF_MIXED_MERGE)
 734                 return;
 735
 736         /*
 737          * @rq will no longer represent mixable attributes for all the
 738          * contained bios.  It will just track those of the first one.
 739          * Distributes the attributs to each bio.
 740          */
 741         for (bio = rq->bio; bio; bio = bio->bi_next) {
 742                 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
 743                              (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
 744                 bio->bi_opf |= ff;
 745         }
 746         rq->rq_flags |= RQF_MIXED_MERGE;
 747 }
 748
 749 static void blk_account_io_merge(struct request *req)
 750 {
 751         if (blk_do_io_stat(req)) {
 752                 struct hd_struct *part;
 753
 754                 part_stat_lock();
 755                 part = req->part;
 756
 757                 part_dec_in_flight(req->q, part, rq_data_dir(req));
 758
 759                 hd_struct_put(part);
 760                 part_stat_unlock();
 761         }
 762 }
 763 /*
 764  * Two cases of handling DISCARD merge:
 765  * If max_discard_segments > 1, the driver takes every bio
 766  * as a range and send them to controller together. The ranges
 767  * needn't to be contiguous.
 768  * Otherwise, the bios/requests will be handled as same as
 769  * others which should be contiguous.
 770  */
 771 static inline bool blk_discard_mergable(struct request *req)
 772 {
 773         if (req_op(req) == REQ_OP_DISCARD &&
 774             queue_max_discard_segments(req->q) > 1)
 775                 return true;
 776         return false;
 777 }
 778
 779 static enum elv_merge blk_try_req_merge(struct request *req,
 780                                         struct request *next)
 781 {
 782         if (blk_discard_mergable(req))
 783                 return ELEVATOR_DISCARD_MERGE;
 784         else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
 785                 return ELEVATOR_BACK_MERGE;
 786
 787         return ELEVATOR_NO_MERGE;
 788 }
 789
 790 /*
 791  * For non-mq, this has to be called with the request spinlock acquired.
 792  * For mq with scheduling, the appropriate queue wide lock should be held.
 793  */
 794 static struct request *attempt_merge(struct request_queue *q,
 795                                      struct request *req, struct request *next)
 796 {
 797         if (!rq_mergeable(req) || !rq_mergeable(next))
 798                 return NULL;
 799
 800         if (req_op(req) != req_op(next))
 801                 return NULL;
 802
 803         if (rq_data_dir(req) != rq_data_dir(next)
 804             || req->rq_disk != next->rq_disk)
 805                 return NULL;
 806
 807         if (req_op(req) == REQ_OP_WRITE_SAME &&
 808             !blk_write_same_mergeable(req->bio, next->bio))
 809                 return NULL;
 810
 811         /*
 812          * Don't allow merge of different write hints, or for a hint with
 813          * non-hint IO.
 814          */
 815         if (req->write_hint != next->write_hint)
 816                 return NULL;
 817
 818         if (req->ioprio != next->ioprio)
 819                 return NULL;
 820
 821         /*
 822          * If we are allowed to merge, then append bio list
 823          * from next to rq and release next. merge_requests_fn
 824          * will have updated segment counts, update sector
 825          * counts here. Handle DISCARDs separately, as they
 826          * have separate settings.
 827          */
 828
 829         switch (blk_try_req_merge(req, next)) {
 830         case ELEVATOR_DISCARD_MERGE:
 831                 if (!req_attempt_discard_merge(q, req, next))
 832                         return NULL;
 833                 break;
 834         case ELEVATOR_BACK_MERGE:
 835                 if (!ll_merge_requests_fn(q, req, next))
 836                         return NULL;
 837                 break;
 838         default:
 839                 return NULL;
 840         }
 841
 842         /*
 843          * If failfast settings disagree or any of the two is already
 844          * a mixed merge, mark both as mixed before proceeding.  This
 845          * makes sure that all involved bios have mixable attributes
 846          * set properly.
 847          */
 848         if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
 849             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 850             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 851                 blk_rq_set_mixed_merge(req);
 852                 blk_rq_set_mixed_merge(next);
 853         }
 854
 855         /*
 856          * At this point we have either done a back merge or front merge. We
 857          * need the smaller start_time_ns of the merged requests to be the
 858          * current request for accounting purposes.
 859          */
 860         if (next->start_time_ns < req->start_time_ns)
 861                 req->start_time_ns = next->start_time_ns;
 862
 863         req->biotail->bi_next = next->bio;
 864         req->biotail = next->biotail;
 865
 866         req->__data_len += blk_rq_bytes(next);
 867
 868         if (!blk_discard_mergable(req))
 869                 elv_merge_requests(q, req, next);
 870
 871         /*
 872          * 'next' is going away, so update stats accordingly
 873          */
 874         blk_account_io_merge(next);
 875
 876         /*
 877          * ownership of bio passed from next to req, return 'next' for
 878          * the caller to free
 879          */
 880         next->bio = NULL;
 881         return next;
 882 }
 883
 884 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
 885 {
 886         struct request *next = elv_latter_request(q, rq);
 887
 888         if (next)
 889                 return attempt_merge(q, rq, next);
 890
 891         return NULL;
 892 }
 893
 894 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
 895 {
 896         struct request *prev = elv_former_request(q, rq);
 897
 898         if (prev)
 899                 return attempt_merge(q, prev, rq);
 900
 901         return NULL;
 902 }
 903
 904 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 905                           struct request *next)
 906 {
 907         struct request *free;
 908
 909         free = attempt_merge(q, rq, next);
 910         if (free) {
 911                 blk_put_request(free);
 912                 return 1;
 913         }
 914
 915         return 0;
 916 }
 917
 918 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 919 {
 920         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 921                 return false;
 922
 923         if (req_op(rq) != bio_op(bio))
 924                 return false;
 925
 926         /* different data direction or already started, don't merge */
 927         if (bio_data_dir(bio) != rq_data_dir(rq))
 928                 return false;
 929
 930         /* must be same device */
 931         if (rq->rq_disk != bio->bi_disk)
 932                 return false;
 933
 934         /* only merge integrity protected bio into ditto rq */
 935         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 936                 return false;
 937
 938         /* must be using the same buffer */
 939         if (req_op(rq) == REQ_OP_WRITE_SAME &&
 940             !blk_write_same_mergeable(rq->bio, bio))
 941                 return false;
 942
 943         /*
 944          * Don't allow merge of different write hints, or for a hint with
 945          * non-hint IO.
 946          */
 947         if (rq->write_hint != bio->bi_write_hint)
 948                 return false;
 949
 950         if (rq->ioprio != bio_prio(bio))
 951                 return false;
 952
 953         return true;
 954 }
 955
 956 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
 957 {
 958         if (blk_discard_mergable(rq))
 959                 return ELEVATOR_DISCARD_MERGE;
 960         else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 961                 return ELEVATOR_BACK_MERGE;
 962         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 963                 return ELEVATOR_FRONT_MERGE;
 964         return ELEVATOR_NO_MERGE;
 965 }