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 struct bio *blk_bio_segment_split(struct request_queue *q,
 165                                          struct bio *bio,
 166                                          struct bio_set *bs,
 167                                          unsigned *segs)
 168 {
 169         struct bio_vec bv, bvprv, *bvprvp = NULL;
 170         struct bvec_iter iter;
 171         unsigned seg_size = 0, nsegs = 0, sectors = 0;
 172         unsigned front_seg_size = bio->bi_seg_front_size;
 173         bool do_split = true;
 174         struct bio *new = NULL;
 175         const unsigned max_sectors = get_max_io_size(q, bio);
 176
 177         bio_for_each_segment(bv, bio, iter) {
 178                 /*
 179                  * If the queue doesn't support SG gaps and adding this
 180                  * offset would create a gap, disallow it.
 181                  */
 182                 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
 183                         goto split;
 184
 185                 if (sectors + (bv.bv_len >> 9) > max_sectors) {
 186                         /*
 187                          * Consider this a new segment if we're splitting in
 188                          * the middle of this vector.
 189                          */
 190                         if (nsegs < queue_max_segments(q) &&
 191                             sectors < max_sectors) {
 192                                 nsegs++;
 193                                 sectors = max_sectors;
 194                         }
 195                         goto split;
 196                 }
 197
 198                 if (bvprvp && blk_queue_cluster(q)) {
 199                         if (seg_size + bv.bv_len > queue_max_segment_size(q))
 200                                 goto new_segment;
 201                         if (!biovec_phys_mergeable(q, bvprvp, &bv))
 202                                 goto new_segment;
 203
 204                         seg_size += bv.bv_len;
 205                         bvprv = bv;
 206                         bvprvp = &bvprv;
 207                         sectors += bv.bv_len >> 9;
 208
 209                         continue;
 210                 }
 211 new_segment:
 212                 if (nsegs == queue_max_segments(q))
 213                         goto split;
 214
 215                 if (nsegs == 1 && seg_size > front_seg_size)
 216                         front_seg_size = seg_size;
 217
 218                 nsegs++;
 219                 bvprv = bv;
 220                 bvprvp = &bvprv;
 221                 seg_size = bv.bv_len;
 222                 sectors += bv.bv_len >> 9;
 223
 224         }
 225
 226         do_split = false;
 227 split:
 228         *segs = nsegs;
 229
 230         if (do_split) {
 231                 new = bio_split(bio, sectors, GFP_NOIO, bs);
 232                 if (new)
 233                         bio = new;
 234         }
 235
 236         if (nsegs == 1 && seg_size > front_seg_size)
 237                 front_seg_size = seg_size;
 238         bio->bi_seg_front_size = front_seg_size;
 239         if (seg_size > bio->bi_seg_back_size)
 240                 bio->bi_seg_back_size = seg_size;
 241
 242         return do_split ? new : NULL;
 243 }
 244
 245 void blk_queue_split(struct request_queue *q, struct bio **bio)
 246 {
 247         struct bio *split, *res;
 248         unsigned nsegs;
 249
 250         switch (bio_op(*bio)) {
 251         case REQ_OP_DISCARD:
 252         case REQ_OP_SECURE_ERASE:
 253                 split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
 254                 break;
 255         case REQ_OP_WRITE_ZEROES:
 256                 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
 257                 break;
 258         case REQ_OP_WRITE_SAME:
 259                 split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
 260                 break;
 261         default:
 262                 split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
 263                 break;
 264         }
 265
 266         /* physical segments can be figured out during splitting */
 267         res = split ? split : *bio;
 268         res->bi_phys_segments = nsegs;
 269         bio_set_flag(res, BIO_SEG_VALID);
 270
 271         if (split) {
 272                 /* there isn't chance to merge the splitted bio */
 273                 split->bi_opf |= REQ_NOMERGE;
 274
 275                 /*
 276                  * Since we're recursing into make_request here, ensure
 277                  * that we mark this bio as already having entered the queue.
 278                  * If not, and the queue is going away, we can get stuck
 279                  * forever on waiting for the queue reference to drop. But
 280                  * that will never happen, as we're already holding a
 281                  * reference to it.
 282                  */
 283                 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
 284
 285                 bio_chain(split, *bio);
 286                 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
 287                 generic_make_request(*bio);
 288                 *bio = split;
 289         }
 290 }
 291 EXPORT_SYMBOL(blk_queue_split);
 292
 293 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
 294                                              struct bio *bio,
 295                                              bool no_sg_merge)
 296 {
 297         struct bio_vec bv, bvprv = { NULL };
 298         int cluster, prev = 0;
 299         unsigned int seg_size, nr_phys_segs;
 300         struct bio *fbio, *bbio;
 301         struct bvec_iter iter;
 302
 303         if (!bio)
 304                 return 0;
 305
 306         switch (bio_op(bio)) {
 307         case REQ_OP_DISCARD:
 308         case REQ_OP_SECURE_ERASE:
 309         case REQ_OP_WRITE_ZEROES:
 310                 return 0;
 311         case REQ_OP_WRITE_SAME:
 312                 return 1;
 313         }
 314
 315         fbio = bio;
 316         cluster = blk_queue_cluster(q);
 317         seg_size = 0;
 318         nr_phys_segs = 0;
 319         for_each_bio(bio) {
 320                 bio_for_each_segment(bv, bio, iter) {
 321                         /*
 322                          * If SG merging is disabled, each bio vector is
 323                          * a segment
 324                          */
 325                         if (no_sg_merge)
 326                                 goto new_segment;
 327
 328                         if (prev && cluster) {
 329                                 if (seg_size + bv.bv_len
 330                                     > queue_max_segment_size(q))
 331                                         goto new_segment;
 332                                 if (!biovec_phys_mergeable(q, &bvprv, &bv))
 333                                         goto new_segment;
 334
 335                                 seg_size += bv.bv_len;
 336                                 bvprv = bv;
 337                                 continue;
 338                         }
 339 new_segment:
 340                         if (nr_phys_segs == 1 && seg_size >
 341                             fbio->bi_seg_front_size)
 342                                 fbio->bi_seg_front_size = seg_size;
 343
 344                         nr_phys_segs++;
 345                         bvprv = bv;
 346                         prev = 1;
 347                         seg_size = bv.bv_len;
 348                 }
 349                 bbio = bio;
 350         }
 351
 352         if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
 353                 fbio->bi_seg_front_size = seg_size;
 354         if (seg_size > bbio->bi_seg_back_size)
 355                 bbio->bi_seg_back_size = seg_size;
 356
 357         return nr_phys_segs;
 358 }
 359
 360 void blk_recalc_rq_segments(struct request *rq)
 361 {
 362         bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
 363                         &rq->q->queue_flags);
 364
 365         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
 366                         no_sg_merge);
 367 }
 368
 369 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 370 {
 371         unsigned short seg_cnt;
 372
 373         /* estimate segment number by bi_vcnt for non-cloned bio */
 374         if (bio_flagged(bio, BIO_CLONED))
 375                 seg_cnt = bio_segments(bio);
 376         else
 377                 seg_cnt = bio->bi_vcnt;
 378
 379         if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
 380                         (seg_cnt < queue_max_segments(q)))
 381                 bio->bi_phys_segments = seg_cnt;
 382         else {
 383                 struct bio *nxt = bio->bi_next;
 384
 385                 bio->bi_next = NULL;
 386                 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
 387                 bio->bi_next = nxt;
 388         }
 389
 390         bio_set_flag(bio, BIO_SEG_VALID);
 391 }
 392 EXPORT_SYMBOL(blk_recount_segments);
 393
 394 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 395                                    struct bio *nxt)
 396 {
 397         struct bio_vec end_bv = { NULL }, nxt_bv;
 398
 399         if (!blk_queue_cluster(q))
 400                 return 0;
 401
 402         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 403             queue_max_segment_size(q))
 404                 return 0;
 405
 406         if (!bio_has_data(bio))
 407                 return 1;
 408
 409         bio_get_last_bvec(bio, &end_bv);
 410         bio_get_first_bvec(nxt, &nxt_bv);
 411
 412         return biovec_phys_mergeable(q, &end_bv, &nxt_bv);
 413 }
 414
 415 static inline void
 416 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
 417                      struct scatterlist *sglist, struct bio_vec *bvprv,
 418                      struct scatterlist **sg, int *nsegs, int *cluster)
 419 {
 420
 421         int nbytes = bvec->bv_len;
 422
 423         if (*sg && *cluster) {
 424                 if ((*sg)->length + nbytes > queue_max_segment_size(q))
 425                         goto new_segment;
 426                 if (!biovec_phys_mergeable(q, bvprv, bvec))
 427                         goto new_segment;
 428
 429                 (*sg)->length += nbytes;
 430         } else {
 431 new_segment:
 432                 if (!*sg)
 433                         *sg = sglist;
 434                 else {
 435                         /*
 436                          * If the driver previously mapped a shorter
 437                          * list, we could see a termination bit
 438                          * prematurely unless it fully inits the sg
 439                          * table on each mapping. We KNOW that there
 440                          * must be more entries here or the driver
 441                          * would be buggy, so force clear the
 442                          * termination bit to avoid doing a full
 443                          * sg_init_table() in drivers for each command.
 444                          */
 445                         sg_unmark_end(*sg);
 446                         *sg = sg_next(*sg);
 447                 }
 448
 449                 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
 450                 (*nsegs)++;
 451         }
 452         *bvprv = *bvec;
 453 }
 454
 455 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
 456                 struct scatterlist *sglist, struct scatterlist **sg)
 457 {
 458         *sg = sglist;
 459         sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
 460         return 1;
 461 }
 462
 463 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 464                              struct scatterlist *sglist,
 465                              struct scatterlist **sg)
 466 {
 467         struct bio_vec bvec, bvprv = { NULL };
 468         struct bvec_iter iter;
 469         int cluster = blk_queue_cluster(q), nsegs = 0;
 470
 471         for_each_bio(bio)
 472                 bio_for_each_segment(bvec, bio, iter)
 473                         __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
 474                                              &nsegs, &cluster);
 475
 476         return nsegs;
 477 }
 478
 479 /*
 480  * map a request to scatterlist, return number of sg entries setup. Caller
 481  * must make sure sg can hold rq->nr_phys_segments entries
 482  */
 483 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 484                   struct scatterlist *sglist)
 485 {
 486         struct scatterlist *sg = NULL;
 487         int nsegs = 0;
 488
 489         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
 490                 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
 491         else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
 492                 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
 493         else if (rq->bio)
 494                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
 495
 496         if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
 497             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 498                 unsigned int pad_len =
 499                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 500
 501                 sg->length += pad_len;
 502                 rq->extra_len += pad_len;
 503         }
 504
 505         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 506                 if (op_is_write(req_op(rq)))
 507                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 508
 509                 sg_unmark_end(sg);
 510                 sg = sg_next(sg);
 511                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 512                             q->dma_drain_size,
 513                             ((unsigned long)q->dma_drain_buffer) &
 514                             (PAGE_SIZE - 1));
 515                 nsegs++;
 516                 rq->extra_len += q->dma_drain_size;
 517         }
 518
 519         if (sg)
 520                 sg_mark_end(sg);
 521
 522         /*
 523          * Something must have been wrong if the figured number of
 524          * segment is bigger than number of req's physical segments
 525          */
 526         WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
 527
 528         return nsegs;
 529 }
 530 EXPORT_SYMBOL(blk_rq_map_sg);
 531
 532 static inline int ll_new_hw_segment(struct request_queue *q,
 533                                     struct request *req,
 534                                     struct bio *bio)
 535 {
 536         int nr_phys_segs = bio_phys_segments(q, bio);
 537
 538         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 539                 goto no_merge;
 540
 541         if (blk_integrity_merge_bio(q, req, bio) == false)
 542                 goto no_merge;
 543
 544         /*
 545          * This will form the start of a new hw segment.  Bump both
 546          * counters.
 547          */
 548         req->nr_phys_segments += nr_phys_segs;
 549         return 1;
 550
 551 no_merge:
 552         req_set_nomerge(q, req);
 553         return 0;
 554 }
 555
 556 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 557                      struct bio *bio)
 558 {
 559         if (req_gap_back_merge(req, bio))
 560                 return 0;
 561         if (blk_integrity_rq(req) &&
 562             integrity_req_gap_back_merge(req, bio))
 563                 return 0;
 564         if (blk_rq_sectors(req) + bio_sectors(bio) >
 565             blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
 566                 req_set_nomerge(q, req);
 567                 return 0;
 568         }
 569         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 570                 blk_recount_segments(q, req->biotail);
 571         if (!bio_flagged(bio, BIO_SEG_VALID))
 572                 blk_recount_segments(q, bio);
 573
 574         return ll_new_hw_segment(q, req, bio);
 575 }
 576
 577 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 578                       struct bio *bio)
 579 {
 580
 581         if (req_gap_front_merge(req, bio))
 582                 return 0;
 583         if (blk_integrity_rq(req) &&
 584             integrity_req_gap_front_merge(req, bio))
 585                 return 0;
 586         if (blk_rq_sectors(req) + bio_sectors(bio) >
 587             blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
 588                 req_set_nomerge(q, req);
 589                 return 0;
 590         }
 591         if (!bio_flagged(bio, BIO_SEG_VALID))
 592                 blk_recount_segments(q, bio);
 593         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 594                 blk_recount_segments(q, req->bio);
 595
 596         return ll_new_hw_segment(q, req, bio);
 597 }
 598
 599 /*
 600  * blk-mq uses req->special to carry normal driver per-request payload, it
 601  * does not indicate a prepared command that we cannot merge with.
 602  */
 603 static bool req_no_special_merge(struct request *req)
 604 {
 605         struct request_queue *q = req->q;
 606
 607         return !q->mq_ops && req->special;
 608 }
 609
 610 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
 611                 struct request *next)
 612 {
 613         unsigned short segments = blk_rq_nr_discard_segments(req);
 614
 615         if (segments >= queue_max_discard_segments(q))
 616                 goto no_merge;
 617         if (blk_rq_sectors(req) + bio_sectors(next->bio) >
 618             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 619                 goto no_merge;
 620
 621         req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
 622         return true;
 623 no_merge:
 624         req_set_nomerge(q, req);
 625         return false;
 626 }
 627
 628 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 629                                 struct request *next)
 630 {
 631         int total_phys_segments;
 632         unsigned int seg_size =
 633                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 634
 635         /*
 636          * First check if the either of the requests are re-queued
 637          * requests.  Can't merge them if they are.
 638          */
 639         if (req_no_special_merge(req) || req_no_special_merge(next))
 640                 return 0;
 641
 642         if (req_gap_back_merge(req, next->bio))
 643                 return 0;
 644
 645         /*
 646          * Will it become too large?
 647          */
 648         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 649             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 650                 return 0;
 651
 652         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 653         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 654                 if (req->nr_phys_segments == 1)
 655                         req->bio->bi_seg_front_size = seg_size;
 656                 if (next->nr_phys_segments == 1)
 657                         next->biotail->bi_seg_back_size = seg_size;
 658                 total_phys_segments--;
 659         }
 660
 661         if (total_phys_segments > queue_max_segments(q))
 662                 return 0;
 663
 664         if (blk_integrity_merge_rq(q, req, next) == false)
 665                 return 0;
 666
 667         /* Merge is OK... */
 668         req->nr_phys_segments = total_phys_segments;
 669         return 1;
 670 }
 671
 672 /**
 673  * blk_rq_set_mixed_merge - mark a request as mixed merge
 674  * @rq: request to mark as mixed merge
 675  *
 676  * Description:
 677  *     @rq is about to be mixed merged.  Make sure the attributes
 678  *     which can be mixed are set in each bio and mark @rq as mixed
 679  *     merged.
 680  */
 681 void blk_rq_set_mixed_merge(struct request *rq)
 682 {
 683         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 684         struct bio *bio;
 685
 686         if (rq->rq_flags & RQF_MIXED_MERGE)
 687                 return;
 688
 689         /*
 690          * @rq will no longer represent mixable attributes for all the
 691          * contained bios.  It will just track those of the first one.
 692          * Distributes the attributs to each bio.
 693          */
 694         for (bio = rq->bio; bio; bio = bio->bi_next) {
 695                 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
 696                              (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
 697                 bio->bi_opf |= ff;
 698         }
 699         rq->rq_flags |= RQF_MIXED_MERGE;
 700 }
 701
 702 static void blk_account_io_merge(struct request *req)
 703 {
 704         if (blk_do_io_stat(req)) {
 705                 struct hd_struct *part;
 706                 int cpu;
 707
 708                 cpu = part_stat_lock();
 709                 part = req->part;
 710
 711                 part_round_stats(req->q, cpu, part);
 712                 part_dec_in_flight(req->q, part, rq_data_dir(req));
 713
 714                 hd_struct_put(part);
 715                 part_stat_unlock();
 716         }
 717 }
 718 /*
 719  * Two cases of handling DISCARD merge:
 720  * If max_discard_segments > 1, the driver takes every bio
 721  * as a range and send them to controller together. The ranges
 722  * needn't to be contiguous.
 723  * Otherwise, the bios/requests will be handled as same as
 724  * others which should be contiguous.
 725  */
 726 static inline bool blk_discard_mergable(struct request *req)
 727 {
 728         if (req_op(req) == REQ_OP_DISCARD &&
 729             queue_max_discard_segments(req->q) > 1)
 730                 return true;
 731         return false;
 732 }
 733
 734 enum elv_merge blk_try_req_merge(struct request *req, struct request *next)
 735 {
 736         if (blk_discard_mergable(req))
 737                 return ELEVATOR_DISCARD_MERGE;
 738         else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
 739                 return ELEVATOR_BACK_MERGE;
 740
 741         return ELEVATOR_NO_MERGE;
 742 }
 743
 744 /*
 745  * For non-mq, this has to be called with the request spinlock acquired.
 746  * For mq with scheduling, the appropriate queue wide lock should be held.
 747  */
 748 static struct request *attempt_merge(struct request_queue *q,
 749                                      struct request *req, struct request *next)
 750 {
 751         if (!q->mq_ops)
 752                 lockdep_assert_held(q->queue_lock);
 753
 754         if (!rq_mergeable(req) || !rq_mergeable(next))
 755                 return NULL;
 756
 757         if (req_op(req) != req_op(next))
 758                 return NULL;
 759
 760         if (rq_data_dir(req) != rq_data_dir(next)
 761             || req->rq_disk != next->rq_disk
 762             || req_no_special_merge(next))
 763                 return NULL;
 764
 765         if (req_op(req) == REQ_OP_WRITE_SAME &&
 766             !blk_write_same_mergeable(req->bio, next->bio))
 767                 return NULL;
 768
 769         /*
 770          * Don't allow merge of different write hints, or for a hint with
 771          * non-hint IO.
 772          */
 773         if (req->write_hint != next->write_hint)
 774                 return NULL;
 775
 776         /*
 777          * If we are allowed to merge, then append bio list
 778          * from next to rq and release next. merge_requests_fn
 779          * will have updated segment counts, update sector
 780          * counts here. Handle DISCARDs separately, as they
 781          * have separate settings.
 782          */
 783
 784         switch (blk_try_req_merge(req, next)) {
 785         case ELEVATOR_DISCARD_MERGE:
 786                 if (!req_attempt_discard_merge(q, req, next))
 787                         return NULL;
 788                 break;
 789         case ELEVATOR_BACK_MERGE:
 790                 if (!ll_merge_requests_fn(q, req, next))
 791                         return NULL;
 792                 break;
 793         default:
 794                 return NULL;
 795         }
 796
 797         /*
 798          * If failfast settings disagree or any of the two is already
 799          * a mixed merge, mark both as mixed before proceeding.  This
 800          * makes sure that all involved bios have mixable attributes
 801          * set properly.
 802          */
 803         if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
 804             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 805             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 806                 blk_rq_set_mixed_merge(req);
 807                 blk_rq_set_mixed_merge(next);
 808         }
 809
 810         /*
 811          * At this point we have either done a back merge or front merge. We
 812          * need the smaller start_time_ns of the merged requests to be the
 813          * current request for accounting purposes.
 814          */
 815         if (next->start_time_ns < req->start_time_ns)
 816                 req->start_time_ns = next->start_time_ns;
 817
 818         req->biotail->bi_next = next->bio;
 819         req->biotail = next->biotail;
 820
 821         req->__data_len += blk_rq_bytes(next);
 822
 823         if (!blk_discard_mergable(req))
 824                 elv_merge_requests(q, req, next);
 825
 826         /*
 827          * 'next' is going away, so update stats accordingly
 828          */
 829         blk_account_io_merge(next);
 830
 831         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 832         if (blk_rq_cpu_valid(next))
 833                 req->cpu = next->cpu;
 834
 835         /*
 836          * ownership of bio passed from next to req, return 'next' for
 837          * the caller to free
 838          */
 839         next->bio = NULL;
 840         return next;
 841 }
 842
 843 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
 844 {
 845         struct request *next = elv_latter_request(q, rq);
 846
 847         if (next)
 848                 return attempt_merge(q, rq, next);
 849
 850         return NULL;
 851 }
 852
 853 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
 854 {
 855         struct request *prev = elv_former_request(q, rq);
 856
 857         if (prev)
 858                 return attempt_merge(q, prev, rq);
 859
 860         return NULL;
 861 }
 862
 863 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 864                           struct request *next)
 865 {
 866         struct elevator_queue *e = q->elevator;
 867         struct request *free;
 868
 869         if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn)
 870                 if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next))
 871                         return 0;
 872
 873         free = attempt_merge(q, rq, next);
 874         if (free) {
 875                 __blk_put_request(q, free);
 876                 return 1;
 877         }
 878
 879         return 0;
 880 }
 881
 882 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 883 {
 884         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 885                 return false;
 886
 887         if (req_op(rq) != bio_op(bio))
 888                 return false;
 889
 890         /* different data direction or already started, don't merge */
 891         if (bio_data_dir(bio) != rq_data_dir(rq))
 892                 return false;
 893
 894         /* must be same device and not a special request */
 895         if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq))
 896                 return false;
 897
 898         /* only merge integrity protected bio into ditto rq */
 899         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 900                 return false;
 901
 902         /* must be using the same buffer */
 903         if (req_op(rq) == REQ_OP_WRITE_SAME &&
 904             !blk_write_same_mergeable(rq->bio, bio))
 905                 return false;
 906
 907         /*
 908          * Don't allow merge of different write hints, or for a hint with
 909          * non-hint IO.
 910          */
 911         if (rq->write_hint != bio->bi_write_hint)
 912                 return false;
 913
 914         return true;
 915 }
 916
 917 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
 918 {
 919         if (blk_discard_mergable(rq))
 920                 return ELEVATOR_DISCARD_MERGE;
 921         else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 922                 return ELEVATOR_BACK_MERGE;
 923         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 924                 return ELEVATOR_FRONT_MERGE;
 925         return ELEVATOR_NO_MERGE;
 926 }