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) {
 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 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         seg_size = 0;
 317         nr_phys_segs = 0;
 318         for_each_bio(bio) {
 319                 bio_for_each_segment(bv, bio, iter) {
 320                         /*
 321                          * If SG merging is disabled, each bio vector is
 322                          * a segment
 323                          */
 324                         if (no_sg_merge)
 325                                 goto new_segment;
 326
 327                         if (prev) {
 328                                 if (seg_size + bv.bv_len
 329                                     > queue_max_segment_size(q))
 330                                         goto new_segment;
 331                                 if (!biovec_phys_mergeable(q, &bvprv, &bv))
 332                                         goto new_segment;
 333
 334                                 seg_size += bv.bv_len;
 335                                 bvprv = bv;
 336                                 continue;
 337                         }
 338 new_segment:
 339                         if (nr_phys_segs == 1 && seg_size >
 340                             fbio->bi_seg_front_size)
 341                                 fbio->bi_seg_front_size = seg_size;
 342
 343                         nr_phys_segs++;
 344                         bvprv = bv;
 345                         prev = 1;
 346                         seg_size = bv.bv_len;
 347                 }
 348                 bbio = bio;
 349         }
 350
 351         if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
 352                 fbio->bi_seg_front_size = seg_size;
 353         if (seg_size > bbio->bi_seg_back_size)
 354                 bbio->bi_seg_back_size = seg_size;
 355
 356         return nr_phys_segs;
 357 }
 358
 359 void blk_recalc_rq_segments(struct request *rq)
 360 {
 361         bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
 362                         &rq->q->queue_flags);
 363
 364         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
 365                         no_sg_merge);
 366 }
 367
 368 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 369 {
 370         unsigned short seg_cnt;
 371
 372         /* estimate segment number by bi_vcnt for non-cloned bio */
 373         if (bio_flagged(bio, BIO_CLONED))
 374                 seg_cnt = bio_segments(bio);
 375         else
 376                 seg_cnt = bio->bi_vcnt;
 377
 378         if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
 379                         (seg_cnt < queue_max_segments(q)))
 380                 bio->bi_phys_segments = seg_cnt;
 381         else {
 382                 struct bio *nxt = bio->bi_next;
 383
 384                 bio->bi_next = NULL;
 385                 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
 386                 bio->bi_next = nxt;
 387         }
 388
 389         bio_set_flag(bio, BIO_SEG_VALID);
 390 }
 391
 392 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 393                                    struct bio *nxt)
 394 {
 395         struct bio_vec end_bv = { NULL }, nxt_bv;
 396
 397         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 398             queue_max_segment_size(q))
 399                 return 0;
 400
 401         if (!bio_has_data(bio))
 402                 return 1;
 403
 404         bio_get_last_bvec(bio, &end_bv);
 405         bio_get_first_bvec(nxt, &nxt_bv);
 406
 407         return biovec_phys_mergeable(q, &end_bv, &nxt_bv);
 408 }
 409
 410 static inline void
 411 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
 412                      struct scatterlist *sglist, struct bio_vec *bvprv,
 413                      struct scatterlist **sg, int *nsegs)
 414 {
 415
 416         int nbytes = bvec->bv_len;
 417
 418         if (*sg) {
 419                 if ((*sg)->length + nbytes > queue_max_segment_size(q))
 420                         goto new_segment;
 421                 if (!biovec_phys_mergeable(q, bvprv, bvec))
 422                         goto new_segment;
 423
 424                 (*sg)->length += nbytes;
 425         } else {
 426 new_segment:
 427                 if (!*sg)
 428                         *sg = sglist;
 429                 else {
 430                         /*
 431                          * If the driver previously mapped a shorter
 432                          * list, we could see a termination bit
 433                          * prematurely unless it fully inits the sg
 434                          * table on each mapping. We KNOW that there
 435                          * must be more entries here or the driver
 436                          * would be buggy, so force clear the
 437                          * termination bit to avoid doing a full
 438                          * sg_init_table() in drivers for each command.
 439                          */
 440                         sg_unmark_end(*sg);
 441                         *sg = sg_next(*sg);
 442                 }
 443
 444                 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
 445                 (*nsegs)++;
 446         }
 447         *bvprv = *bvec;
 448 }
 449
 450 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
 451                 struct scatterlist *sglist, struct scatterlist **sg)
 452 {
 453         *sg = sglist;
 454         sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
 455         return 1;
 456 }
 457
 458 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 459                              struct scatterlist *sglist,
 460                              struct scatterlist **sg)
 461 {
 462         struct bio_vec bvec, bvprv = { NULL };
 463         struct bvec_iter iter;
 464         int nsegs = 0;
 465
 466         for_each_bio(bio)
 467                 bio_for_each_segment(bvec, bio, iter)
 468                         __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
 469                                              &nsegs);
 470
 471         return nsegs;
 472 }
 473
 474 /*
 475  * map a request to scatterlist, return number of sg entries setup. Caller
 476  * must make sure sg can hold rq->nr_phys_segments entries
 477  */
 478 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 479                   struct scatterlist *sglist)
 480 {
 481         struct scatterlist *sg = NULL;
 482         int nsegs = 0;
 483
 484         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
 485                 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
 486         else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
 487                 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
 488         else if (rq->bio)
 489                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
 490
 491         if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
 492             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 493                 unsigned int pad_len =
 494                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 495
 496                 sg->length += pad_len;
 497                 rq->extra_len += pad_len;
 498         }
 499
 500         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 501                 if (op_is_write(req_op(rq)))
 502                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 503
 504                 sg_unmark_end(sg);
 505                 sg = sg_next(sg);
 506                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 507                             q->dma_drain_size,
 508                             ((unsigned long)q->dma_drain_buffer) &
 509                             (PAGE_SIZE - 1));
 510                 nsegs++;
 511                 rq->extra_len += q->dma_drain_size;
 512         }
 513
 514         if (sg)
 515                 sg_mark_end(sg);
 516
 517         /*
 518          * Something must have been wrong if the figured number of
 519          * segment is bigger than number of req's physical segments
 520          */
 521         WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
 522
 523         return nsegs;
 524 }
 525 EXPORT_SYMBOL(blk_rq_map_sg);
 526
 527 static inline int ll_new_hw_segment(struct request_queue *q,
 528                                     struct request *req,
 529                                     struct bio *bio)
 530 {
 531         int nr_phys_segs = bio_phys_segments(q, bio);
 532
 533         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 534                 goto no_merge;
 535
 536         if (blk_integrity_merge_bio(q, req, bio) == false)
 537                 goto no_merge;
 538
 539         /*
 540          * This will form the start of a new hw segment.  Bump both
 541          * counters.
 542          */
 543         req->nr_phys_segments += nr_phys_segs;
 544         return 1;
 545
 546 no_merge:
 547         req_set_nomerge(q, req);
 548         return 0;
 549 }
 550
 551 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 552                      struct bio *bio)
 553 {
 554         if (req_gap_back_merge(req, bio))
 555                 return 0;
 556         if (blk_integrity_rq(req) &&
 557             integrity_req_gap_back_merge(req, bio))
 558                 return 0;
 559         if (blk_rq_sectors(req) + bio_sectors(bio) >
 560             blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
 561                 req_set_nomerge(q, req);
 562                 return 0;
 563         }
 564         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 565                 blk_recount_segments(q, req->biotail);
 566         if (!bio_flagged(bio, BIO_SEG_VALID))
 567                 blk_recount_segments(q, bio);
 568
 569         return ll_new_hw_segment(q, req, bio);
 570 }
 571
 572 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 573                       struct bio *bio)
 574 {
 575
 576         if (req_gap_front_merge(req, bio))
 577                 return 0;
 578         if (blk_integrity_rq(req) &&
 579             integrity_req_gap_front_merge(req, bio))
 580                 return 0;
 581         if (blk_rq_sectors(req) + bio_sectors(bio) >
 582             blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
 583                 req_set_nomerge(q, req);
 584                 return 0;
 585         }
 586         if (!bio_flagged(bio, BIO_SEG_VALID))
 587                 blk_recount_segments(q, bio);
 588         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 589                 blk_recount_segments(q, req->bio);
 590
 591         return ll_new_hw_segment(q, req, bio);
 592 }
 593
 594 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
 595                 struct request *next)
 596 {
 597         unsigned short segments = blk_rq_nr_discard_segments(req);
 598
 599         if (segments >= queue_max_discard_segments(q))
 600                 goto no_merge;
 601         if (blk_rq_sectors(req) + bio_sectors(next->bio) >
 602             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 603                 goto no_merge;
 604
 605         req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
 606         return true;
 607 no_merge:
 608         req_set_nomerge(q, req);
 609         return false;
 610 }
 611
 612 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 613                                 struct request *next)
 614 {
 615         int total_phys_segments;
 616         unsigned int seg_size =
 617                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 618
 619         if (req_gap_back_merge(req, next->bio))
 620                 return 0;
 621
 622         /*
 623          * Will it become too large?
 624          */
 625         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 626             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 627                 return 0;
 628
 629         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 630         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 631                 if (req->nr_phys_segments == 1)
 632                         req->bio->bi_seg_front_size = seg_size;
 633                 if (next->nr_phys_segments == 1)
 634                         next->biotail->bi_seg_back_size = seg_size;
 635                 total_phys_segments--;
 636         }
 637
 638         if (total_phys_segments > queue_max_segments(q))
 639                 return 0;
 640
 641         if (blk_integrity_merge_rq(q, req, next) == false)
 642                 return 0;
 643
 644         /* Merge is OK... */
 645         req->nr_phys_segments = total_phys_segments;
 646         return 1;
 647 }
 648
 649 /**
 650  * blk_rq_set_mixed_merge - mark a request as mixed merge
 651  * @rq: request to mark as mixed merge
 652  *
 653  * Description:
 654  *     @rq is about to be mixed merged.  Make sure the attributes
 655  *     which can be mixed are set in each bio and mark @rq as mixed
 656  *     merged.
 657  */
 658 void blk_rq_set_mixed_merge(struct request *rq)
 659 {
 660         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 661         struct bio *bio;
 662
 663         if (rq->rq_flags & RQF_MIXED_MERGE)
 664                 return;
 665
 666         /*
 667          * @rq will no longer represent mixable attributes for all the
 668          * contained bios.  It will just track those of the first one.
 669          * Distributes the attributs to each bio.
 670          */
 671         for (bio = rq->bio; bio; bio = bio->bi_next) {
 672                 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
 673                              (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
 674                 bio->bi_opf |= ff;
 675         }
 676         rq->rq_flags |= RQF_MIXED_MERGE;
 677 }
 678
 679 static void blk_account_io_merge(struct request *req)
 680 {
 681         if (blk_do_io_stat(req)) {
 682                 struct hd_struct *part;
 683
 684                 part_stat_lock();
 685                 part = req->part;
 686
 687                 part_dec_in_flight(req->q, part, rq_data_dir(req));
 688
 689                 hd_struct_put(part);
 690                 part_stat_unlock();
 691         }
 692 }
 693 /*
 694  * Two cases of handling DISCARD merge:
 695  * If max_discard_segments > 1, the driver takes every bio
 696  * as a range and send them to controller together. The ranges
 697  * needn't to be contiguous.
 698  * Otherwise, the bios/requests will be handled as same as
 699  * others which should be contiguous.
 700  */
 701 static inline bool blk_discard_mergable(struct request *req)
 702 {
 703         if (req_op(req) == REQ_OP_DISCARD &&
 704             queue_max_discard_segments(req->q) > 1)
 705                 return true;
 706         return false;
 707 }
 708
 709 static enum elv_merge blk_try_req_merge(struct request *req,
 710                                         struct request *next)
 711 {
 712         if (blk_discard_mergable(req))
 713                 return ELEVATOR_DISCARD_MERGE;
 714         else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
 715                 return ELEVATOR_BACK_MERGE;
 716
 717         return ELEVATOR_NO_MERGE;
 718 }
 719
 720 /*
 721  * For non-mq, this has to be called with the request spinlock acquired.
 722  * For mq with scheduling, the appropriate queue wide lock should be held.
 723  */
 724 static struct request *attempt_merge(struct request_queue *q,
 725                                      struct request *req, struct request *next)
 726 {
 727         if (!rq_mergeable(req) || !rq_mergeable(next))
 728                 return NULL;
 729
 730         if (req_op(req) != req_op(next))
 731                 return NULL;
 732
 733         if (rq_data_dir(req) != rq_data_dir(next)
 734             || req->rq_disk != next->rq_disk)
 735                 return NULL;
 736
 737         if (req_op(req) == REQ_OP_WRITE_SAME &&
 738             !blk_write_same_mergeable(req->bio, next->bio))
 739                 return NULL;
 740
 741         /*
 742          * Don't allow merge of different write hints, or for a hint with
 743          * non-hint IO.
 744          */
 745         if (req->write_hint != next->write_hint)
 746                 return NULL;
 747
 748         if (req->ioprio != next->ioprio)
 749                 return NULL;
 750
 751         /*
 752          * If we are allowed to merge, then append bio list
 753          * from next to rq and release next. merge_requests_fn
 754          * will have updated segment counts, update sector
 755          * counts here. Handle DISCARDs separately, as they
 756          * have separate settings.
 757          */
 758
 759         switch (blk_try_req_merge(req, next)) {
 760         case ELEVATOR_DISCARD_MERGE:
 761                 if (!req_attempt_discard_merge(q, req, next))
 762                         return NULL;
 763                 break;
 764         case ELEVATOR_BACK_MERGE:
 765                 if (!ll_merge_requests_fn(q, req, next))
 766                         return NULL;
 767                 break;
 768         default:
 769                 return NULL;
 770         }
 771
 772         /*
 773          * If failfast settings disagree or any of the two is already
 774          * a mixed merge, mark both as mixed before proceeding.  This
 775          * makes sure that all involved bios have mixable attributes
 776          * set properly.
 777          */
 778         if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
 779             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 780             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 781                 blk_rq_set_mixed_merge(req);
 782                 blk_rq_set_mixed_merge(next);
 783         }
 784
 785         /*
 786          * At this point we have either done a back merge or front merge. We
 787          * need the smaller start_time_ns of the merged requests to be the
 788          * current request for accounting purposes.
 789          */
 790         if (next->start_time_ns < req->start_time_ns)
 791                 req->start_time_ns = next->start_time_ns;
 792
 793         req->biotail->bi_next = next->bio;
 794         req->biotail = next->biotail;
 795
 796         req->__data_len += blk_rq_bytes(next);
 797
 798         if (!blk_discard_mergable(req))
 799                 elv_merge_requests(q, req, next);
 800
 801         /*
 802          * 'next' is going away, so update stats accordingly
 803          */
 804         blk_account_io_merge(next);
 805
 806         /*
 807          * ownership of bio passed from next to req, return 'next' for
 808          * the caller to free
 809          */
 810         next->bio = NULL;
 811         return next;
 812 }
 813
 814 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
 815 {
 816         struct request *next = elv_latter_request(q, rq);
 817
 818         if (next)
 819                 return attempt_merge(q, rq, next);
 820
 821         return NULL;
 822 }
 823
 824 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
 825 {
 826         struct request *prev = elv_former_request(q, rq);
 827
 828         if (prev)
 829                 return attempt_merge(q, prev, rq);
 830
 831         return NULL;
 832 }
 833
 834 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 835                           struct request *next)
 836 {
 837         struct request *free;
 838
 839         free = attempt_merge(q, rq, next);
 840         if (free) {
 841                 blk_put_request(free);
 842                 return 1;
 843         }
 844
 845         return 0;
 846 }
 847
 848 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 849 {
 850         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 851                 return false;
 852
 853         if (req_op(rq) != bio_op(bio))
 854                 return false;
 855
 856         /* different data direction or already started, don't merge */
 857         if (bio_data_dir(bio) != rq_data_dir(rq))
 858                 return false;
 859
 860         /* must be same device */
 861         if (rq->rq_disk != bio->bi_disk)
 862                 return false;
 863
 864         /* only merge integrity protected bio into ditto rq */
 865         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 866                 return false;
 867
 868         /* must be using the same buffer */
 869         if (req_op(rq) == REQ_OP_WRITE_SAME &&
 870             !blk_write_same_mergeable(rq->bio, bio))
 871                 return false;
 872
 873         /*
 874          * Don't allow merge of different write hints, or for a hint with
 875          * non-hint IO.
 876          */
 877         if (rq->write_hint != bio->bi_write_hint)
 878                 return false;
 879
 880         if (rq->ioprio != bio_prio(bio))
 881                 return false;
 882
 883         return true;
 884 }
 885
 886 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
 887 {
 888         if (blk_discard_mergable(rq))
 889                 return ELEVATOR_DISCARD_MERGE;
 890         else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 891                 return ELEVATOR_BACK_MERGE;
 892         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 893                 return ELEVATOR_FRONT_MERGE;
 894         return ELEVATOR_NO_MERGE;
 895 }