block/blk-merge.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Functions related to segment and merge handling
   4  */
   5 #include <linux/kernel.h>
   6 #include <linux/module.h>
   7 #include <linux/bio.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/scatterlist.h>
  10
  11 #include <trace/events/block.h>
  12
  13 #include "blk.h"
  14
  15 static inline bool bio_will_gap(struct request_queue *q,
  16                 struct request *prev_rq, struct bio *prev, struct bio *next)
  17 {
  18         struct bio_vec pb, nb;
  19
  20         if (!bio_has_data(prev) || !queue_virt_boundary(q))
  21                 return false;
  22
  23         /*
  24          * Don't merge if the 1st bio starts with non-zero offset, otherwise it
  25          * is quite difficult to respect the sg gap limit.  We work hard to
  26          * merge a huge number of small single bios in case of mkfs.
  27          */
  28         if (prev_rq)
  29                 bio_get_first_bvec(prev_rq->bio, &pb);
  30         else
  31                 bio_get_first_bvec(prev, &pb);
  32         if (pb.bv_offset & queue_virt_boundary(q))
  33                 return true;
  34
  35         /*
  36          * We don't need to worry about the situation that the merged segment
  37          * ends in unaligned virt boundary:
  38          *
  39          * - if 'pb' ends aligned, the merged segment ends aligned
  40          * - if 'pb' ends unaligned, the next bio must include
  41          *   one single bvec of 'nb', otherwise the 'nb' can't
  42          *   merge with 'pb'
  43          */
  44         bio_get_last_bvec(prev, &pb);
  45         bio_get_first_bvec(next, &nb);
  46         if (biovec_phys_mergeable(q, &pb, &nb))
  47                 return false;
  48         return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
  49 }
  50
  51 static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
  52 {
  53         return bio_will_gap(req->q, req, req->biotail, bio);
  54 }
  55
  56 static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
  57 {
  58         return bio_will_gap(req->q, NULL, bio, req->bio);
  59 }
  60
  61 static struct bio *blk_bio_discard_split(struct request_queue *q,
  62                                          struct bio *bio,
  63                                          struct bio_set *bs,
  64                                          unsigned *nsegs)
  65 {
  66         unsigned int max_discard_sectors, granularity;
  67         int alignment;
  68         sector_t tmp;
  69         unsigned split_sectors;
  70
  71         *nsegs = 1;
  72
  73         /* Zero-sector (unknown) and one-sector granularities are the same.  */
  74         granularity = max(q->limits.discard_granularity >> 9, 1U);
  75
  76         max_discard_sectors = min(q->limits.max_discard_sectors,
  77                         bio_allowed_max_sectors(q));
  78         max_discard_sectors -= max_discard_sectors % granularity;
  79
  80         if (unlikely(!max_discard_sectors)) {
  81                 /* XXX: warn */
  82                 return NULL;
  83         }
  84
  85         if (bio_sectors(bio) <= max_discard_sectors)
  86                 return NULL;
  87
  88         split_sectors = max_discard_sectors;
  89
  90         /*
  91          * If the next starting sector would be misaligned, stop the discard at
  92          * the previous aligned sector.
  93          */
  94         alignment = (q->limits.discard_alignment >> 9) % granularity;
  95
  96         tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
  97         tmp = sector_div(tmp, granularity);
  98
  99         if (split_sectors > tmp)
 100                 split_sectors -= tmp;
 101
 102         return bio_split(bio, split_sectors, GFP_NOIO, bs);
 103 }
 104
 105 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
 106                 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
 107 {
 108         *nsegs = 0;
 109
 110         if (!q->limits.max_write_zeroes_sectors)
 111                 return NULL;
 112
 113         if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
 114                 return NULL;
 115
 116         return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
 117 }
 118
 119 static struct bio *blk_bio_write_same_split(struct request_queue *q,
 120                                             struct bio *bio,
 121                                             struct bio_set *bs,
 122                                             unsigned *nsegs)
 123 {
 124         *nsegs = 1;
 125
 126         if (!q->limits.max_write_same_sectors)
 127                 return NULL;
 128
 129         if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
 130                 return NULL;
 131
 132         return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
 133 }
 134
 135 /*
 136  * Return the maximum number of sectors from the start of a bio that may be
 137  * submitted as a single request to a block device. If enough sectors remain,
 138  * align the end to the physical block size. Otherwise align the end to the
 139  * logical block size. This approach minimizes the number of non-aligned
 140  * requests that are submitted to a block device if the start of a bio is not
 141  * aligned to a physical block boundary.
 142  */
 143 static inline unsigned get_max_io_size(struct request_queue *q,
 144                                        struct bio *bio)
 145 {
 146         unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
 147         unsigned max_sectors = sectors;
 148         unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT;
 149         unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT;
 150         unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1);
 151
 152         max_sectors += start_offset;
 153         max_sectors &= ~(pbs - 1);
 154         if (max_sectors > start_offset)
 155                 return max_sectors - start_offset;
 156
 157         return sectors & (lbs - 1);
 158 }
 159
 160 static inline unsigned get_max_segment_size(const struct request_queue *q,
 161                                             struct page *start_page,
 162                                             unsigned long offset)
 163 {
 164         unsigned long mask = queue_segment_boundary(q);
 165
 166         offset = mask & (page_to_phys(start_page) + offset);
 167
 168         /*
 169          * overflow may be triggered in case of zero page physical address
 170          * on 32bit arch, use queue's max segment size when that happens.
 171          */
 172         return min_not_zero(mask - offset + 1,
 173                         (unsigned long)queue_max_segment_size(q));
 174 }
 175
 176 /**
 177  * bvec_split_segs - verify whether or not a bvec should be split in the middle
 178  * @q:        [in] request queue associated with the bio associated with @bv
 179  * @bv:       [in] bvec to examine
 180  * @nsegs:    [in,out] Number of segments in the bio being built. Incremented
 181  *            by the number of segments from @bv that may be appended to that
 182  *            bio without exceeding @max_segs
 183  * @sectors:  [in,out] Number of sectors in the bio being built. Incremented
 184  *            by the number of sectors from @bv that may be appended to that
 185  *            bio without exceeding @max_sectors
 186  * @max_segs: [in] upper bound for *@nsegs
 187  * @max_sectors: [in] upper bound for *@sectors
 188  *
 189  * When splitting a bio, it can happen that a bvec is encountered that is too
 190  * big to fit in a single segment and hence that it has to be split in the
 191  * middle. This function verifies whether or not that should happen. The value
 192  * %true is returned if and only if appending the entire @bv to a bio with
 193  * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
 194  * the block driver.
 195  */
 196 static bool bvec_split_segs(const struct request_queue *q,
 197                             const struct bio_vec *bv, unsigned *nsegs,
 198                             unsigned *sectors, unsigned max_segs,
 199                             unsigned max_sectors)
 200 {
 201         unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9;
 202         unsigned len = min(bv->bv_len, max_len);
 203         unsigned total_len = 0;
 204         unsigned seg_size = 0;
 205
 206         while (len && *nsegs < max_segs) {
 207                 seg_size = get_max_segment_size(q, bv->bv_page,
 208                                                 bv->bv_offset + total_len);
 209                 seg_size = min(seg_size, len);
 210
 211                 (*nsegs)++;
 212                 total_len += seg_size;
 213                 len -= seg_size;
 214
 215                 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
 216                         break;
 217         }
 218
 219         *sectors += total_len >> 9;
 220
 221         /* tell the caller to split the bvec if it is too big to fit */
 222         return len > 0 || bv->bv_len > max_len;
 223 }
 224
 225 /**
 226  * blk_bio_segment_split - split a bio in two bios
 227  * @q:    [in] request queue pointer
 228  * @bio:  [in] bio to be split
 229  * @bs:   [in] bio set to allocate the clone from
 230  * @segs: [out] number of segments in the bio with the first half of the sectors
 231  *
 232  * Clone @bio, update the bi_iter of the clone to represent the first sectors
 233  * of @bio and update @bio->bi_iter to represent the remaining sectors. The
 234  * following is guaranteed for the cloned bio:
 235  * - That it has at most get_max_io_size(@q, @bio) sectors.
 236  * - That it has at most queue_max_segments(@q) segments.
 237  *
 238  * Except for discard requests the cloned bio will point at the bi_io_vec of
 239  * the original bio. It is the responsibility of the caller to ensure that the
 240  * original bio is not freed before the cloned bio. The caller is also
 241  * responsible for ensuring that @bs is only destroyed after processing of the
 242  * split bio has finished.
 243  */
 244 static struct bio *blk_bio_segment_split(struct request_queue *q,
 245                                          struct bio *bio,
 246                                          struct bio_set *bs,
 247                                          unsigned *segs)
 248 {
 249         struct bio_vec bv, bvprv, *bvprvp = NULL;
 250         struct bvec_iter iter;
 251         unsigned nsegs = 0, sectors = 0;
 252         const unsigned max_sectors = get_max_io_size(q, bio);
 253         const unsigned max_segs = queue_max_segments(q);
 254
 255         bio_for_each_bvec(bv, bio, iter) {
 256                 /*
 257                  * If the queue doesn't support SG gaps and adding this
 258                  * offset would create a gap, disallow it.
 259                  */
 260                 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
 261                         goto split;
 262
 263                 if (nsegs < max_segs &&
 264                     sectors + (bv.bv_len >> 9) <= max_sectors &&
 265                     bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
 266                         nsegs++;
 267                         sectors += bv.bv_len >> 9;
 268                 } else if (bvec_split_segs(q, &bv, &nsegs, &sectors, max_segs,
 269                                          max_sectors)) {
 270                         goto split;
 271                 }
 272
 273                 bvprv = bv;
 274                 bvprvp = &bvprv;
 275         }
 276
 277         *segs = nsegs;
 278         return NULL;
 279 split:
 280         *segs = nsegs;
 281         return bio_split(bio, sectors, GFP_NOIO, bs);
 282 }
 283
 284 /**
 285  * __blk_queue_split - split a bio and submit the second half
 286  * @q:       [in] request queue pointer
 287  * @bio:     [in, out] bio to be split
 288  * @nr_segs: [out] number of segments in the first bio
 289  *
 290  * Split a bio into two bios, chain the two bios, submit the second half and
 291  * store a pointer to the first half in *@bio. If the second bio is still too
 292  * big it will be split by a recursive call to this function. Since this
 293  * function may allocate a new bio from @q->bio_split, it is the responsibility
 294  * of the caller to ensure that @q is only released after processing of the
 295  * split bio has finished.
 296  */
 297 void __blk_queue_split(struct request_queue *q, struct bio **bio,
 298                 unsigned int *nr_segs)
 299 {
 300         struct bio *split = NULL;
 301
 302         switch (bio_op(*bio)) {
 303         case REQ_OP_DISCARD:
 304         case REQ_OP_SECURE_ERASE:
 305                 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs);
 306                 break;
 307         case REQ_OP_WRITE_ZEROES:
 308                 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split,
 309                                 nr_segs);
 310                 break;
 311         case REQ_OP_WRITE_SAME:
 312                 split = blk_bio_write_same_split(q, *bio, &q->bio_split,
 313                                 nr_segs);
 314                 break;
 315         default:
 316                 /*
 317                  * All drivers must accept single-segments bios that are <=
 318                  * PAGE_SIZE.  This is a quick and dirty check that relies on
 319                  * the fact that bi_io_vec[0] is always valid if a bio has data.
 320                  * The check might lead to occasional false negatives when bios
 321                  * are cloned, but compared to the performance impact of cloned
 322                  * bios themselves the loop below doesn't matter anyway.
 323                  */
 324                 if (!q->limits.chunk_sectors &&
 325                     (*bio)->bi_vcnt == 1 &&
 326                     ((*bio)->bi_io_vec[0].bv_len +
 327                      (*bio)->bi_io_vec[0].bv_offset) <= PAGE_SIZE) {
 328                         *nr_segs = 1;
 329                         break;
 330                 }
 331                 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs);
 332                 break;
 333         }
 334
 335         if (split) {
 336                 /* there isn't chance to merge the splitted bio */
 337                 split->bi_opf |= REQ_NOMERGE;
 338
 339                 /*
 340                  * Since we're recursing into make_request here, ensure
 341                  * that we mark this bio as already having entered the queue.
 342                  * If not, and the queue is going away, we can get stuck
 343                  * forever on waiting for the queue reference to drop. But
 344                  * that will never happen, as we're already holding a
 345                  * reference to it.
 346                  */
 347                 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
 348
 349                 bio_chain(split, *bio);
 350                 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
 351                 generic_make_request(*bio);
 352                 *bio = split;
 353         }
 354 }
 355
 356 /**
 357  * blk_queue_split - split a bio and submit the second half
 358  * @q:   [in] request queue pointer
 359  * @bio: [in, out] bio to be split
 360  *
 361  * Split a bio into two bios, chains the two bios, submit the second half and
 362  * store a pointer to the first half in *@bio. Since this function may allocate
 363  * a new bio from @q->bio_split, it is the responsibility of the caller to
 364  * ensure that @q is only released after processing of the split bio has
 365  * finished.
 366  */
 367 void blk_queue_split(struct request_queue *q, struct bio **bio)
 368 {
 369         unsigned int nr_segs;
 370
 371         __blk_queue_split(q, bio, &nr_segs);
 372 }
 373 EXPORT_SYMBOL(blk_queue_split);
 374
 375 unsigned int blk_recalc_rq_segments(struct request *rq)
 376 {
 377         unsigned int nr_phys_segs = 0;
 378         unsigned int nr_sectors = 0;
 379         struct req_iterator iter;
 380         struct bio_vec bv;
 381
 382         if (!rq->bio)
 383                 return 0;
 384
 385         switch (bio_op(rq->bio)) {
 386         case REQ_OP_DISCARD:
 387         case REQ_OP_SECURE_ERASE:
 388         case REQ_OP_WRITE_ZEROES:
 389                 return 0;
 390         case REQ_OP_WRITE_SAME:
 391                 return 1;
 392         }
 393
 394         rq_for_each_bvec(bv, rq, iter)
 395                 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors,
 396                                 UINT_MAX, UINT_MAX);
 397         return nr_phys_segs;
 398 }
 399
 400 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
 401                 struct scatterlist *sglist)
 402 {
 403         if (!*sg)
 404                 return sglist;
 405
 406         /*
 407          * If the driver previously mapped a shorter list, we could see a
 408          * termination bit prematurely unless it fully inits the sg table
 409          * on each mapping. We KNOW that there must be more entries here
 410          * or the driver would be buggy, so force clear the termination bit
 411          * to avoid doing a full sg_init_table() in drivers for each command.
 412          */
 413         sg_unmark_end(*sg);
 414         return sg_next(*sg);
 415 }
 416
 417 static unsigned blk_bvec_map_sg(struct request_queue *q,
 418                 struct bio_vec *bvec, struct scatterlist *sglist,
 419                 struct scatterlist **sg)
 420 {
 421         unsigned nbytes = bvec->bv_len;
 422         unsigned nsegs = 0, total = 0;
 423
 424         while (nbytes > 0) {
 425                 unsigned offset = bvec->bv_offset + total;
 426                 unsigned len = min(get_max_segment_size(q, bvec->bv_page,
 427                                         offset), nbytes);
 428                 struct page *page = bvec->bv_page;
 429
 430                 /*
 431                  * Unfortunately a fair number of drivers barf on scatterlists
 432                  * that have an offset larger than PAGE_SIZE, despite other
 433                  * subsystems dealing with that invariant just fine.  For now
 434                  * stick to the legacy format where we never present those from
 435                  * the block layer, but the code below should be removed once
 436                  * these offenders (mostly MMC/SD drivers) are fixed.
 437                  */
 438                 page += (offset >> PAGE_SHIFT);
 439                 offset &= ~PAGE_MASK;
 440
 441                 *sg = blk_next_sg(sg, sglist);
 442                 sg_set_page(*sg, page, len, offset);
 443
 444                 total += len;
 445                 nbytes -= len;
 446                 nsegs++;
 447         }
 448
 449         return nsegs;
 450 }
 451
 452 static inline int __blk_bvec_map_sg(struct bio_vec bv,
 453                 struct scatterlist *sglist, struct scatterlist **sg)
 454 {
 455         *sg = blk_next_sg(sg, sglist);
 456         sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
 457         return 1;
 458 }
 459
 460 /* only try to merge bvecs into one sg if they are from two bios */
 461 static inline bool
 462 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
 463                            struct bio_vec *bvprv, struct scatterlist **sg)
 464 {
 465
 466         int nbytes = bvec->bv_len;
 467
 468         if (!*sg)
 469                 return false;
 470
 471         if ((*sg)->length + nbytes > queue_max_segment_size(q))
 472                 return false;
 473
 474         if (!biovec_phys_mergeable(q, bvprv, bvec))
 475                 return false;
 476
 477         (*sg)->length += nbytes;
 478
 479         return true;
 480 }
 481
 482 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 483                              struct scatterlist *sglist,
 484                              struct scatterlist **sg)
 485 {
 486         struct bio_vec uninitialized_var(bvec), bvprv = { NULL };
 487         struct bvec_iter iter;
 488         int nsegs = 0;
 489         bool new_bio = false;
 490
 491         for_each_bio(bio) {
 492                 bio_for_each_bvec(bvec, bio, iter) {
 493                         /*
 494                          * Only try to merge bvecs from two bios given we
 495                          * have done bio internal merge when adding pages
 496                          * to bio
 497                          */
 498                         if (new_bio &&
 499                             __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
 500                                 goto next_bvec;
 501
 502                         if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
 503                                 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
 504                         else
 505                                 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
 506  next_bvec:
 507                         new_bio = false;
 508                 }
 509                 if (likely(bio->bi_iter.bi_size)) {
 510                         bvprv = bvec;
 511                         new_bio = true;
 512                 }
 513         }
 514
 515         return nsegs;
 516 }
 517
 518 /*
 519  * map a request to scatterlist, return number of sg entries setup. Caller
 520  * must make sure sg can hold rq->nr_phys_segments entries
 521  */
 522 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 523                   struct scatterlist *sglist)
 524 {
 525         struct scatterlist *sg = NULL;
 526         int nsegs = 0;
 527
 528         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
 529                 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, &sg);
 530         else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
 531                 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, &sg);
 532         else if (rq->bio)
 533                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
 534
 535         if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
 536             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 537                 unsigned int pad_len =
 538                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 539
 540                 sg->length += pad_len;
 541                 rq->extra_len += pad_len;
 542         }
 543
 544         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 545                 if (op_is_write(req_op(rq)))
 546                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 547
 548                 sg_unmark_end(sg);
 549                 sg = sg_next(sg);
 550                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 551                             q->dma_drain_size,
 552                             ((unsigned long)q->dma_drain_buffer) &
 553                             (PAGE_SIZE - 1));
 554                 nsegs++;
 555                 rq->extra_len += q->dma_drain_size;
 556         }
 557
 558         if (sg)
 559                 sg_mark_end(sg);
 560
 561         /*
 562          * Something must have been wrong if the figured number of
 563          * segment is bigger than number of req's physical segments
 564          */
 565         WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
 566
 567         return nsegs;
 568 }
 569 EXPORT_SYMBOL(blk_rq_map_sg);
 570
 571 static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
 572                 unsigned int nr_phys_segs)
 573 {
 574         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(req->q))
 575                 goto no_merge;
 576
 577         if (blk_integrity_merge_bio(req->q, req, bio) == false)
 578                 goto no_merge;
 579
 580         /*
 581          * This will form the start of a new hw segment.  Bump both
 582          * counters.
 583          */
 584         req->nr_phys_segments += nr_phys_segs;
 585         return 1;
 586
 587 no_merge:
 588         req_set_nomerge(req->q, req);
 589         return 0;
 590 }
 591
 592 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
 593 {
 594         if (req_gap_back_merge(req, bio))
 595                 return 0;
 596         if (blk_integrity_rq(req) &&
 597             integrity_req_gap_back_merge(req, bio))
 598                 return 0;
 599         if (blk_rq_sectors(req) + bio_sectors(bio) >
 600             blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
 601                 req_set_nomerge(req->q, req);
 602                 return 0;
 603         }
 604
 605         return ll_new_hw_segment(req, bio, nr_segs);
 606 }
 607
 608 int ll_front_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
 609 {
 610         if (req_gap_front_merge(req, bio))
 611                 return 0;
 612         if (blk_integrity_rq(req) &&
 613             integrity_req_gap_front_merge(req, bio))
 614                 return 0;
 615         if (blk_rq_sectors(req) + bio_sectors(bio) >
 616             blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
 617                 req_set_nomerge(req->q, req);
 618                 return 0;
 619         }
 620
 621         return ll_new_hw_segment(req, bio, nr_segs);
 622 }
 623
 624 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
 625                 struct request *next)
 626 {
 627         unsigned short segments = blk_rq_nr_discard_segments(req);
 628
 629         if (segments >= queue_max_discard_segments(q))
 630                 goto no_merge;
 631         if (blk_rq_sectors(req) + bio_sectors(next->bio) >
 632             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 633                 goto no_merge;
 634
 635         req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
 636         return true;
 637 no_merge:
 638         req_set_nomerge(q, req);
 639         return false;
 640 }
 641
 642 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 643                                 struct request *next)
 644 {
 645         int total_phys_segments;
 646
 647         if (req_gap_back_merge(req, next->bio))
 648                 return 0;
 649
 650         /*
 651          * Will it become too large?
 652          */
 653         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 654             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 655                 return 0;
 656
 657         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 658         if (total_phys_segments > queue_max_segments(q))
 659                 return 0;
 660
 661         if (blk_integrity_merge_rq(q, req, next) == false)
 662                 return 0;
 663
 664         /* Merge is OK... */
 665         req->nr_phys_segments = total_phys_segments;
 666         return 1;
 667 }
 668
 669 /**
 670  * blk_rq_set_mixed_merge - mark a request as mixed merge
 671  * @rq: request to mark as mixed merge
 672  *
 673  * Description:
 674  *     @rq is about to be mixed merged.  Make sure the attributes
 675  *     which can be mixed are set in each bio and mark @rq as mixed
 676  *     merged.
 677  */
 678 void blk_rq_set_mixed_merge(struct request *rq)
 679 {
 680         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 681         struct bio *bio;
 682
 683         if (rq->rq_flags & RQF_MIXED_MERGE)
 684                 return;
 685
 686         /*
 687          * @rq will no longer represent mixable attributes for all the
 688          * contained bios.  It will just track those of the first one.
 689          * Distributes the attributs to each bio.
 690          */
 691         for (bio = rq->bio; bio; bio = bio->bi_next) {
 692                 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
 693                              (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
 694                 bio->bi_opf |= ff;
 695         }
 696         rq->rq_flags |= RQF_MIXED_MERGE;
 697 }
 698
 699 static void blk_account_io_merge(struct request *req)
 700 {
 701         if (blk_do_io_stat(req)) {
 702                 struct hd_struct *part;
 703
 704                 part_stat_lock();
 705                 part = req->part;
 706
 707                 part_dec_in_flight(req->q, part, rq_data_dir(req));
 708
 709                 hd_struct_put(part);
 710                 part_stat_unlock();
 711         }
 712 }
 713 /*
 714  * Two cases of handling DISCARD merge:
 715  * If max_discard_segments > 1, the driver takes every bio
 716  * as a range and send them to controller together. The ranges
 717  * needn't to be contiguous.
 718  * Otherwise, the bios/requests will be handled as same as
 719  * others which should be contiguous.
 720  */
 721 static inline bool blk_discard_mergable(struct request *req)
 722 {
 723         if (req_op(req) == REQ_OP_DISCARD &&
 724             queue_max_discard_segments(req->q) > 1)
 725                 return true;
 726         return false;
 727 }
 728
 729 static enum elv_merge blk_try_req_merge(struct request *req,
 730                                         struct request *next)
 731 {
 732         if (blk_discard_mergable(req))
 733                 return ELEVATOR_DISCARD_MERGE;
 734         else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
 735                 return ELEVATOR_BACK_MERGE;
 736
 737         return ELEVATOR_NO_MERGE;
 738 }
 739
 740 /*
 741  * For non-mq, this has to be called with the request spinlock acquired.
 742  * For mq with scheduling, the appropriate queue wide lock should be held.
 743  */
 744 static struct request *attempt_merge(struct request_queue *q,
 745                                      struct request *req, struct request *next)
 746 {
 747         if (!rq_mergeable(req) || !rq_mergeable(next))
 748                 return NULL;
 749
 750         if (req_op(req) != req_op(next))
 751                 return NULL;
 752
 753         if (rq_data_dir(req) != rq_data_dir(next)
 754             || req->rq_disk != next->rq_disk)
 755                 return NULL;
 756
 757         if (req_op(req) == REQ_OP_WRITE_SAME &&
 758             !blk_write_same_mergeable(req->bio, next->bio))
 759                 return NULL;
 760
 761         /*
 762          * Don't allow merge of different write hints, or for a hint with
 763          * non-hint IO.
 764          */
 765         if (req->write_hint != next->write_hint)
 766                 return NULL;
 767
 768         if (req->ioprio != next->ioprio)
 769                 return NULL;
 770
 771         /*
 772          * If we are allowed to merge, then append bio list
 773          * from next to rq and release next. merge_requests_fn
 774          * will have updated segment counts, update sector
 775          * counts here. Handle DISCARDs separately, as they
 776          * have separate settings.
 777          */
 778
 779         switch (blk_try_req_merge(req, next)) {
 780         case ELEVATOR_DISCARD_MERGE:
 781                 if (!req_attempt_discard_merge(q, req, next))
 782                         return NULL;
 783                 break;
 784         case ELEVATOR_BACK_MERGE:
 785                 if (!ll_merge_requests_fn(q, req, next))
 786                         return NULL;
 787                 break;
 788         default:
 789                 return NULL;
 790         }
 791
 792         /*
 793          * If failfast settings disagree or any of the two is already
 794          * a mixed merge, mark both as mixed before proceeding.  This
 795          * makes sure that all involved bios have mixable attributes
 796          * set properly.
 797          */
 798         if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
 799             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 800             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 801                 blk_rq_set_mixed_merge(req);
 802                 blk_rq_set_mixed_merge(next);
 803         }
 804
 805         /*
 806          * At this point we have either done a back merge or front merge. We
 807          * need the smaller start_time_ns of the merged requests to be the
 808          * current request for accounting purposes.
 809          */
 810         if (next->start_time_ns < req->start_time_ns)
 811                 req->start_time_ns = next->start_time_ns;
 812
 813         req->biotail->bi_next = next->bio;
 814         req->biotail = next->biotail;
 815
 816         req->__data_len += blk_rq_bytes(next);
 817
 818         if (!blk_discard_mergable(req))
 819                 elv_merge_requests(q, req, next);
 820
 821         /*
 822          * 'next' is going away, so update stats accordingly
 823          */
 824         blk_account_io_merge(next);
 825
 826         /*
 827          * ownership of bio passed from next to req, return 'next' for
 828          * the caller to free
 829          */
 830         next->bio = NULL;
 831         return next;
 832 }
 833
 834 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
 835 {
 836         struct request *next = elv_latter_request(q, rq);
 837
 838         if (next)
 839                 return attempt_merge(q, rq, next);
 840
 841         return NULL;
 842 }
 843
 844 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
 845 {
 846         struct request *prev = elv_former_request(q, rq);
 847
 848         if (prev)
 849                 return attempt_merge(q, prev, rq);
 850
 851         return NULL;
 852 }
 853
 854 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 855                           struct request *next)
 856 {
 857         struct request *free;
 858
 859         free = attempt_merge(q, rq, next);
 860         if (free) {
 861                 blk_put_request(free);
 862                 return 1;
 863         }
 864
 865         return 0;
 866 }
 867
 868 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 869 {
 870         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 871                 return false;
 872
 873         if (req_op(rq) != bio_op(bio))
 874                 return false;
 875
 876         /* different data direction or already started, don't merge */
 877         if (bio_data_dir(bio) != rq_data_dir(rq))
 878                 return false;
 879
 880         /* must be same device */
 881         if (rq->rq_disk != bio->bi_disk)
 882                 return false;
 883
 884         /* only merge integrity protected bio into ditto rq */
 885         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 886                 return false;
 887
 888         /* must be using the same buffer */
 889         if (req_op(rq) == REQ_OP_WRITE_SAME &&
 890             !blk_write_same_mergeable(rq->bio, bio))
 891                 return false;
 892
 893         /*
 894          * Don't allow merge of different write hints, or for a hint with
 895          * non-hint IO.
 896          */
 897         if (rq->write_hint != bio->bi_write_hint)
 898                 return false;
 899
 900         if (rq->ioprio != bio_prio(bio))
 901                 return false;
 902
 903         return true;
 904 }
 905
 906 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
 907 {
 908         if (blk_discard_mergable(rq))
 909                 return ELEVATOR_DISCARD_MERGE;
 910         else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 911                 return ELEVATOR_BACK_MERGE;
 912         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 913                 return ELEVATOR_FRONT_MERGE;
 914         return ELEVATOR_NO_MERGE;
 915 }