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 #include "blk-rq-qos.h"
  15
  16 static inline bool bio_will_gap(struct request_queue *q,
  17                 struct request *prev_rq, struct bio *prev, struct bio *next)
  18 {
  19         struct bio_vec pb, nb;
  20
  21         if (!bio_has_data(prev) || !queue_virt_boundary(q))
  22                 return false;
  23
  24         /*
  25          * Don't merge if the 1st bio starts with non-zero offset, otherwise it
  26          * is quite difficult to respect the sg gap limit.  We work hard to
  27          * merge a huge number of small single bios in case of mkfs.
  28          */
  29         if (prev_rq)
  30                 bio_get_first_bvec(prev_rq->bio, &pb);
  31         else
  32                 bio_get_first_bvec(prev, &pb);
  33         if (pb.bv_offset & queue_virt_boundary(q))
  34                 return true;
  35
  36         /*
  37          * We don't need to worry about the situation that the merged segment
  38          * ends in unaligned virt boundary:
  39          *
  40          * - if 'pb' ends aligned, the merged segment ends aligned
  41          * - if 'pb' ends unaligned, the next bio must include
  42          *   one single bvec of 'nb', otherwise the 'nb' can't
  43          *   merge with 'pb'
  44          */
  45         bio_get_last_bvec(prev, &pb);
  46         bio_get_first_bvec(next, &nb);
  47         if (biovec_phys_mergeable(q, &pb, &nb))
  48                 return false;
  49         return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
  50 }
  51
  52 static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
  53 {
  54         return bio_will_gap(req->q, req, req->biotail, bio);
  55 }
  56
  57 static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
  58 {
  59         return bio_will_gap(req->q, NULL, bio, req->bio);
  60 }
  61
  62 static struct bio *blk_bio_discard_split(struct request_queue *q,
  63                                          struct bio *bio,
  64                                          struct bio_set *bs,
  65                                          unsigned *nsegs)
  66 {
  67         unsigned int max_discard_sectors, granularity;
  68         int alignment;
  69         sector_t tmp;
  70         unsigned split_sectors;
  71
  72         *nsegs = 1;
  73
  74         /* Zero-sector (unknown) and one-sector granularities are the same.  */
  75         granularity = max(q->limits.discard_granularity >> 9, 1U);
  76
  77         max_discard_sectors = min(q->limits.max_discard_sectors,
  78                         bio_allowed_max_sectors(q));
  79         max_discard_sectors -= max_discard_sectors % granularity;
  80
  81         if (unlikely(!max_discard_sectors)) {
  82                 /* XXX: warn */
  83                 return NULL;
  84         }
  85
  86         if (bio_sectors(bio) <= max_discard_sectors)
  87                 return NULL;
  88
  89         split_sectors = max_discard_sectors;
  90
  91         /*
  92          * If the next starting sector would be misaligned, stop the discard at
  93          * the previous aligned sector.
  94          */
  95         alignment = (q->limits.discard_alignment >> 9) % granularity;
  96
  97         tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
  98         tmp = sector_div(tmp, granularity);
  99
 100         if (split_sectors > tmp)
 101                 split_sectors -= tmp;
 102
 103         return bio_split(bio, split_sectors, GFP_NOIO, bs);
 104 }
 105
 106 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
 107                 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
 108 {
 109         *nsegs = 0;
 110
 111         if (!q->limits.max_write_zeroes_sectors)
 112                 return NULL;
 113
 114         if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
 115                 return NULL;
 116
 117         return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
 118 }
 119
 120 static struct bio *blk_bio_write_same_split(struct request_queue *q,
 121                                             struct bio *bio,
 122                                             struct bio_set *bs,
 123                                             unsigned *nsegs)
 124 {
 125         *nsegs = 1;
 126
 127         if (!q->limits.max_write_same_sectors)
 128                 return NULL;
 129
 130         if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
 131                 return NULL;
 132
 133         return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
 134 }
 135
 136 /*
 137  * Return the maximum number of sectors from the start of a bio that may be
 138  * submitted as a single request to a block device. If enough sectors remain,
 139  * align the end to the physical block size. Otherwise align the end to the
 140  * logical block size. This approach minimizes the number of non-aligned
 141  * requests that are submitted to a block device if the start of a bio is not
 142  * aligned to a physical block boundary.
 143  */
 144 static inline unsigned get_max_io_size(struct request_queue *q,
 145                                        struct bio *bio)
 146 {
 147         unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector, 0);
 148         unsigned max_sectors = sectors;
 149         unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT;
 150         unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT;
 151         unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1);
 152
 153         max_sectors += start_offset;
 154         max_sectors &= ~(pbs - 1);
 155         if (max_sectors > start_offset)
 156                 return max_sectors - start_offset;
 157
 158         return sectors & ~(lbs - 1);
 159 }
 160
 161 static inline unsigned get_max_segment_size(const struct request_queue *q,
 162                                             struct page *start_page,
 163                                             unsigned long offset)
 164 {
 165         unsigned long mask = queue_segment_boundary(q);
 166
 167         offset = mask & (page_to_phys(start_page) + offset);
 168
 169         /*
 170          * overflow may be triggered in case of zero page physical address
 171          * on 32bit arch, use queue's max segment size when that happens.
 172          */
 173         return min_not_zero(mask - offset + 1,
 174                         (unsigned long)queue_max_segment_size(q));
 175 }
 176
 177 /**
 178  * bvec_split_segs - verify whether or not a bvec should be split in the middle
 179  * @q:        [in] request queue associated with the bio associated with @bv
 180  * @bv:       [in] bvec to examine
 181  * @nsegs:    [in,out] Number of segments in the bio being built. Incremented
 182  *            by the number of segments from @bv that may be appended to that
 183  *            bio without exceeding @max_segs
 184  * @sectors:  [in,out] Number of sectors in the bio being built. Incremented
 185  *            by the number of sectors from @bv that may be appended to that
 186  *            bio without exceeding @max_sectors
 187  * @max_segs: [in] upper bound for *@nsegs
 188  * @max_sectors: [in] upper bound for *@sectors
 189  *
 190  * When splitting a bio, it can happen that a bvec is encountered that is too
 191  * big to fit in a single segment and hence that it has to be split in the
 192  * middle. This function verifies whether or not that should happen. The value
 193  * %true is returned if and only if appending the entire @bv to a bio with
 194  * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
 195  * the block driver.
 196  */
 197 static bool bvec_split_segs(const struct request_queue *q,
 198                             const struct bio_vec *bv, unsigned *nsegs,
 199                             unsigned *sectors, unsigned max_segs,
 200                             unsigned max_sectors)
 201 {
 202         unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9;
 203         unsigned len = min(bv->bv_len, max_len);
 204         unsigned total_len = 0;
 205         unsigned seg_size = 0;
 206
 207         while (len && *nsegs < max_segs) {
 208                 seg_size = get_max_segment_size(q, bv->bv_page,
 209                                                 bv->bv_offset + total_len);
 210                 seg_size = min(seg_size, len);
 211
 212                 (*nsegs)++;
 213                 total_len += seg_size;
 214                 len -= seg_size;
 215
 216                 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
 217                         break;
 218         }
 219
 220         *sectors += total_len >> 9;
 221
 222         /* tell the caller to split the bvec if it is too big to fit */
 223         return len > 0 || bv->bv_len > max_len;
 224 }
 225
 226 /**
 227  * blk_bio_segment_split - split a bio in two bios
 228  * @q:    [in] request queue pointer
 229  * @bio:  [in] bio to be split
 230  * @bs:   [in] bio set to allocate the clone from
 231  * @segs: [out] number of segments in the bio with the first half of the sectors
 232  *
 233  * Clone @bio, update the bi_iter of the clone to represent the first sectors
 234  * of @bio and update @bio->bi_iter to represent the remaining sectors. The
 235  * following is guaranteed for the cloned bio:
 236  * - That it has at most get_max_io_size(@q, @bio) sectors.
 237  * - That it has at most queue_max_segments(@q) segments.
 238  *
 239  * Except for discard requests the cloned bio will point at the bi_io_vec of
 240  * the original bio. It is the responsibility of the caller to ensure that the
 241  * original bio is not freed before the cloned bio. The caller is also
 242  * responsible for ensuring that @bs is only destroyed after processing of the
 243  * split bio has finished.
 244  */
 245 static struct bio *blk_bio_segment_split(struct request_queue *q,
 246                                          struct bio *bio,
 247                                          struct bio_set *bs,
 248                                          unsigned *segs)
 249 {
 250         struct bio_vec bv, bvprv, *bvprvp = NULL;
 251         struct bvec_iter iter;
 252         unsigned nsegs = 0, sectors = 0;
 253         const unsigned max_sectors = get_max_io_size(q, bio);
 254         const unsigned max_segs = queue_max_segments(q);
 255
 256         bio_for_each_bvec(bv, bio, iter) {
 257                 /*
 258                  * If the queue doesn't support SG gaps and adding this
 259                  * offset would create a gap, disallow it.
 260                  */
 261                 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
 262                         goto split;
 263
 264                 if (nsegs < max_segs &&
 265                     sectors + (bv.bv_len >> 9) <= max_sectors &&
 266                     bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
 267                         nsegs++;
 268                         sectors += bv.bv_len >> 9;
 269                 } else if (bvec_split_segs(q, &bv, &nsegs, &sectors, max_segs,
 270                                          max_sectors)) {
 271                         goto split;
 272                 }
 273
 274                 bvprv = bv;
 275                 bvprvp = &bvprv;
 276         }
 277
 278         *segs = nsegs;
 279         return NULL;
 280 split:
 281         *segs = nsegs;
 282
 283         /*
 284          * Bio splitting may cause subtle trouble such as hang when doing sync
 285          * iopoll in direct IO routine. Given performance gain of iopoll for
 286          * big IO can be trival, disable iopoll when split needed.
 287          */
 288         bio->bi_opf &= ~REQ_HIPRI;
 289
 290         return bio_split(bio, sectors, GFP_NOIO, bs);
 291 }
 292
 293 /**
 294  * __blk_queue_split - split a bio and submit the second half
 295  * @bio:     [in, out] bio to be split
 296  * @nr_segs: [out] number of segments in the first bio
 297  *
 298  * Split a bio into two bios, chain the two bios, submit the second half and
 299  * store a pointer to the first half in *@bio. If the second bio is still too
 300  * big it will be split by a recursive call to this function. Since this
 301  * function may allocate a new bio from @bio->bi_disk->queue->bio_split, it is
 302  * the responsibility of the caller to ensure that
 303  * @bio->bi_disk->queue->bio_split is only released after processing of the
 304  * split bio has finished.
 305  */
 306 void __blk_queue_split(struct bio **bio, unsigned int *nr_segs)
 307 {
 308         struct request_queue *q = (*bio)->bi_disk->queue;
 309         struct bio *split = NULL;
 310
 311         switch (bio_op(*bio)) {
 312         case REQ_OP_DISCARD:
 313         case REQ_OP_SECURE_ERASE:
 314                 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs);
 315                 break;
 316         case REQ_OP_WRITE_ZEROES:
 317                 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split,
 318                                 nr_segs);
 319                 break;
 320         case REQ_OP_WRITE_SAME:
 321                 split = blk_bio_write_same_split(q, *bio, &q->bio_split,
 322                                 nr_segs);
 323                 break;
 324         default:
 325                 /*
 326                  * All drivers must accept single-segments bios that are <=
 327                  * PAGE_SIZE.  This is a quick and dirty check that relies on
 328                  * the fact that bi_io_vec[0] is always valid if a bio has data.
 329                  * The check might lead to occasional false negatives when bios
 330                  * are cloned, but compared to the performance impact of cloned
 331                  * bios themselves the loop below doesn't matter anyway.
 332                  */
 333                 if (!q->limits.chunk_sectors &&
 334                     (*bio)->bi_vcnt == 1 &&
 335                     ((*bio)->bi_io_vec[0].bv_len +
 336                      (*bio)->bi_io_vec[0].bv_offset) <= PAGE_SIZE) {
 337                         *nr_segs = 1;
 338                         break;
 339                 }
 340                 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs);
 341                 break;
 342         }
 343
 344         if (split) {
 345                 /* there isn't chance to merge the splitted bio */
 346                 split->bi_opf |= REQ_NOMERGE;
 347
 348                 bio_chain(split, *bio);
 349                 trace_block_split(split, (*bio)->bi_iter.bi_sector);
 350                 submit_bio_noacct(*bio);
 351                 *bio = split;
 352         }
 353 }
 354
 355 /**
 356  * blk_queue_split - split a bio and submit the second half
 357  * @bio: [in, out] bio to be split
 358  *
 359  * Split a bio into two bios, chains the two bios, submit the second half and
 360  * store a pointer to the first half in *@bio. Since this function may allocate
 361  * a new bio from @bio->bi_disk->queue->bio_split, it is the responsibility of
 362  * the caller to ensure that @bio->bi_disk->queue->bio_split is only released
 363  * after processing of the split bio has finished.
 364  */
 365 void blk_queue_split(struct bio **bio)
 366 {
 367         unsigned int nr_segs;
 368
 369         __blk_queue_split(bio, &nr_segs);
 370 }
 371 EXPORT_SYMBOL(blk_queue_split);
 372
 373 unsigned int blk_recalc_rq_segments(struct request *rq)
 374 {
 375         unsigned int nr_phys_segs = 0;
 376         unsigned int nr_sectors = 0;
 377         struct req_iterator iter;
 378         struct bio_vec bv;
 379
 380         if (!rq->bio)
 381                 return 0;
 382
 383         switch (bio_op(rq->bio)) {
 384         case REQ_OP_DISCARD:
 385         case REQ_OP_SECURE_ERASE:
 386         case REQ_OP_WRITE_ZEROES:
 387                 return 0;
 388         case REQ_OP_WRITE_SAME:
 389                 return 1;
 390         }
 391
 392         rq_for_each_bvec(bv, rq, iter)
 393                 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors,
 394                                 UINT_MAX, UINT_MAX);
 395         return nr_phys_segs;
 396 }
 397
 398 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
 399                 struct scatterlist *sglist)
 400 {
 401         if (!*sg)
 402                 return sglist;
 403
 404         /*
 405          * If the driver previously mapped a shorter list, we could see a
 406          * termination bit prematurely unless it fully inits the sg table
 407          * on each mapping. We KNOW that there must be more entries here
 408          * or the driver would be buggy, so force clear the termination bit
 409          * to avoid doing a full sg_init_table() in drivers for each command.
 410          */
 411         sg_unmark_end(*sg);
 412         return sg_next(*sg);
 413 }
 414
 415 static unsigned blk_bvec_map_sg(struct request_queue *q,
 416                 struct bio_vec *bvec, struct scatterlist *sglist,
 417                 struct scatterlist **sg)
 418 {
 419         unsigned nbytes = bvec->bv_len;
 420         unsigned nsegs = 0, total = 0;
 421
 422         while (nbytes > 0) {
 423                 unsigned offset = bvec->bv_offset + total;
 424                 unsigned len = min(get_max_segment_size(q, bvec->bv_page,
 425                                         offset), nbytes);
 426                 struct page *page = bvec->bv_page;
 427
 428                 /*
 429                  * Unfortunately a fair number of drivers barf on scatterlists
 430                  * that have an offset larger than PAGE_SIZE, despite other
 431                  * subsystems dealing with that invariant just fine.  For now
 432                  * stick to the legacy format where we never present those from
 433                  * the block layer, but the code below should be removed once
 434                  * these offenders (mostly MMC/SD drivers) are fixed.
 435                  */
 436                 page += (offset >> PAGE_SHIFT);
 437                 offset &= ~PAGE_MASK;
 438
 439                 *sg = blk_next_sg(sg, sglist);
 440                 sg_set_page(*sg, page, len, offset);
 441
 442                 total += len;
 443                 nbytes -= len;
 444                 nsegs++;
 445         }
 446
 447         return nsegs;
 448 }
 449
 450 static inline int __blk_bvec_map_sg(struct bio_vec bv,
 451                 struct scatterlist *sglist, struct scatterlist **sg)
 452 {
 453         *sg = blk_next_sg(sg, sglist);
 454         sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
 455         return 1;
 456 }
 457
 458 /* only try to merge bvecs into one sg if they are from two bios */
 459 static inline bool
 460 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
 461                            struct bio_vec *bvprv, struct scatterlist **sg)
 462 {
 463
 464         int nbytes = bvec->bv_len;
 465
 466         if (!*sg)
 467                 return false;
 468
 469         if ((*sg)->length + nbytes > queue_max_segment_size(q))
 470                 return false;
 471
 472         if (!biovec_phys_mergeable(q, bvprv, bvec))
 473                 return false;
 474
 475         (*sg)->length += nbytes;
 476
 477         return true;
 478 }
 479
 480 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 481                              struct scatterlist *sglist,
 482                              struct scatterlist **sg)
 483 {
 484         struct bio_vec bvec, bvprv = { NULL };
 485         struct bvec_iter iter;
 486         int nsegs = 0;
 487         bool new_bio = false;
 488
 489         for_each_bio(bio) {
 490                 bio_for_each_bvec(bvec, bio, iter) {
 491                         /*
 492                          * Only try to merge bvecs from two bios given we
 493                          * have done bio internal merge when adding pages
 494                          * to bio
 495                          */
 496                         if (new_bio &&
 497                             __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
 498                                 goto next_bvec;
 499
 500                         if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
 501                                 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
 502                         else
 503                                 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
 504  next_bvec:
 505                         new_bio = false;
 506                 }
 507                 if (likely(bio->bi_iter.bi_size)) {
 508                         bvprv = bvec;
 509                         new_bio = true;
 510                 }
 511         }
 512
 513         return nsegs;
 514 }
 515
 516 /*
 517  * map a request to scatterlist, return number of sg entries setup. Caller
 518  * must make sure sg can hold rq->nr_phys_segments entries
 519  */
 520 int __blk_rq_map_sg(struct request_queue *q, struct request *rq,
 521                 struct scatterlist *sglist, struct scatterlist **last_sg)
 522 {
 523         int nsegs = 0;
 524
 525         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
 526                 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg);
 527         else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
 528                 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, last_sg);
 529         else if (rq->bio)
 530                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg);
 531
 532         if (*last_sg)
 533                 sg_mark_end(*last_sg);
 534
 535         /*
 536          * Something must have been wrong if the figured number of
 537          * segment is bigger than number of req's physical segments
 538          */
 539         WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
 540
 541         return nsegs;
 542 }
 543 EXPORT_SYMBOL(__blk_rq_map_sg);
 544
 545 static inline unsigned int blk_rq_get_max_segments(struct request *rq)
 546 {
 547         if (req_op(rq) == REQ_OP_DISCARD)
 548                 return queue_max_discard_segments(rq->q);
 549         return queue_max_segments(rq->q);
 550 }
 551
 552 static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
 553                 unsigned int nr_phys_segs)
 554 {
 555         if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req))
 556                 goto no_merge;
 557
 558         if (blk_integrity_merge_bio(req->q, req, bio) == false)
 559                 goto no_merge;
 560
 561         /*
 562          * This will form the start of a new hw segment.  Bump both
 563          * counters.
 564          */
 565         req->nr_phys_segments += nr_phys_segs;
 566         return 1;
 567
 568 no_merge:
 569         req_set_nomerge(req->q, req);
 570         return 0;
 571 }
 572
 573 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
 574 {
 575         if (req_gap_back_merge(req, bio))
 576                 return 0;
 577         if (blk_integrity_rq(req) &&
 578             integrity_req_gap_back_merge(req, bio))
 579                 return 0;
 580         if (!bio_crypt_ctx_back_mergeable(req, bio))
 581                 return 0;
 582         if (blk_rq_sectors(req) + bio_sectors(bio) >
 583             blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
 584                 req_set_nomerge(req->q, req);
 585                 return 0;
 586         }
 587
 588         return ll_new_hw_segment(req, bio, nr_segs);
 589 }
 590
 591 static int ll_front_merge_fn(struct request *req, struct bio *bio,
 592                 unsigned int nr_segs)
 593 {
 594         if (req_gap_front_merge(req, bio))
 595                 return 0;
 596         if (blk_integrity_rq(req) &&
 597             integrity_req_gap_front_merge(req, bio))
 598                 return 0;
 599         if (!bio_crypt_ctx_front_mergeable(req, bio))
 600                 return 0;
 601         if (blk_rq_sectors(req) + bio_sectors(bio) >
 602             blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
 603                 req_set_nomerge(req->q, req);
 604                 return 0;
 605         }
 606
 607         return ll_new_hw_segment(req, bio, nr_segs);
 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
 633         if (req_gap_back_merge(req, next->bio))
 634                 return 0;
 635
 636         /*
 637          * Will it become too large?
 638          */
 639         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 640             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 641                 return 0;
 642
 643         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 644         if (total_phys_segments > blk_rq_get_max_segments(req))
 645                 return 0;
 646
 647         if (blk_integrity_merge_rq(q, req, next) == false)
 648                 return 0;
 649
 650         if (!bio_crypt_ctx_merge_rq(req, next))
 651                 return 0;
 652
 653         /* Merge is OK... */
 654         req->nr_phys_segments = total_phys_segments;
 655         return 1;
 656 }
 657
 658 /**
 659  * blk_rq_set_mixed_merge - mark a request as mixed merge
 660  * @rq: request to mark as mixed merge
 661  *
 662  * Description:
 663  *     @rq is about to be mixed merged.  Make sure the attributes
 664  *     which can be mixed are set in each bio and mark @rq as mixed
 665  *     merged.
 666  */
 667 void blk_rq_set_mixed_merge(struct request *rq)
 668 {
 669         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 670         struct bio *bio;
 671
 672         if (rq->rq_flags & RQF_MIXED_MERGE)
 673                 return;
 674
 675         /*
 676          * @rq will no longer represent mixable attributes for all the
 677          * contained bios.  It will just track those of the first one.
 678          * Distributes the attributs to each bio.
 679          */
 680         for (bio = rq->bio; bio; bio = bio->bi_next) {
 681                 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
 682                              (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
 683                 bio->bi_opf |= ff;
 684         }
 685         rq->rq_flags |= RQF_MIXED_MERGE;
 686 }
 687
 688 static void blk_account_io_merge_request(struct request *req)
 689 {
 690         if (blk_do_io_stat(req)) {
 691                 part_stat_lock();
 692                 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
 693                 part_stat_unlock();
 694         }
 695 }
 696
 697 /*
 698  * Two cases of handling DISCARD merge:
 699  * If max_discard_segments > 1, the driver takes every bio
 700  * as a range and send them to controller together. The ranges
 701  * needn't to be contiguous.
 702  * Otherwise, the bios/requests will be handled as same as
 703  * others which should be contiguous.
 704  */
 705 static inline bool blk_discard_mergable(struct request *req)
 706 {
 707         if (req_op(req) == REQ_OP_DISCARD &&
 708             queue_max_discard_segments(req->q) > 1)
 709                 return true;
 710         return false;
 711 }
 712
 713 static enum elv_merge blk_try_req_merge(struct request *req,
 714                                         struct request *next)
 715 {
 716         if (blk_discard_mergable(req))
 717                 return ELEVATOR_DISCARD_MERGE;
 718         else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
 719                 return ELEVATOR_BACK_MERGE;
 720
 721         return ELEVATOR_NO_MERGE;
 722 }
 723
 724 /*
 725  * For non-mq, this has to be called with the request spinlock acquired.
 726  * For mq with scheduling, the appropriate queue wide lock should be held.
 727  */
 728 static struct request *attempt_merge(struct request_queue *q,
 729                                      struct request *req, struct request *next)
 730 {
 731         if (!rq_mergeable(req) || !rq_mergeable(next))
 732                 return NULL;
 733
 734         if (req_op(req) != req_op(next))
 735                 return NULL;
 736
 737         if (rq_data_dir(req) != rq_data_dir(next)
 738             || req->rq_disk != next->rq_disk)
 739                 return NULL;
 740
 741         if (req_op(req) == REQ_OP_WRITE_SAME &&
 742             !blk_write_same_mergeable(req->bio, next->bio))
 743                 return NULL;
 744
 745         /*
 746          * Don't allow merge of different write hints, or for a hint with
 747          * non-hint IO.
 748          */
 749         if (req->write_hint != next->write_hint)
 750                 return NULL;
 751
 752         if (req->ioprio != next->ioprio)
 753                 return NULL;
 754
 755         /*
 756          * If we are allowed to merge, then append bio list
 757          * from next to rq and release next. merge_requests_fn
 758          * will have updated segment counts, update sector
 759          * counts here. Handle DISCARDs separately, as they
 760          * have separate settings.
 761          */
 762
 763         switch (blk_try_req_merge(req, next)) {
 764         case ELEVATOR_DISCARD_MERGE:
 765                 if (!req_attempt_discard_merge(q, req, next))
 766                         return NULL;
 767                 break;
 768         case ELEVATOR_BACK_MERGE:
 769                 if (!ll_merge_requests_fn(q, req, next))
 770                         return NULL;
 771                 break;
 772         default:
 773                 return NULL;
 774         }
 775
 776         /*
 777          * If failfast settings disagree or any of the two is already
 778          * a mixed merge, mark both as mixed before proceeding.  This
 779          * makes sure that all involved bios have mixable attributes
 780          * set properly.
 781          */
 782         if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
 783             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 784             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 785                 blk_rq_set_mixed_merge(req);
 786                 blk_rq_set_mixed_merge(next);
 787         }
 788
 789         /*
 790          * At this point we have either done a back merge or front merge. We
 791          * need the smaller start_time_ns of the merged requests to be the
 792          * current request for accounting purposes.
 793          */
 794         if (next->start_time_ns < req->start_time_ns)
 795                 req->start_time_ns = next->start_time_ns;
 796
 797         req->biotail->bi_next = next->bio;
 798         req->biotail = next->biotail;
 799
 800         req->__data_len += blk_rq_bytes(next);
 801
 802         if (!blk_discard_mergable(req))
 803                 elv_merge_requests(q, req, next);
 804
 805         /*
 806          * 'next' is going away, so update stats accordingly
 807          */
 808         blk_account_io_merge_request(next);
 809
 810         trace_block_rq_merge(next);
 811
 812         /*
 813          * ownership of bio passed from next to req, return 'next' for
 814          * the caller to free
 815          */
 816         next->bio = NULL;
 817         return next;
 818 }
 819
 820 static struct request *attempt_back_merge(struct request_queue *q,
 821                 struct request *rq)
 822 {
 823         struct request *next = elv_latter_request(q, rq);
 824
 825         if (next)
 826                 return attempt_merge(q, rq, next);
 827
 828         return NULL;
 829 }
 830
 831 static struct request *attempt_front_merge(struct request_queue *q,
 832                 struct request *rq)
 833 {
 834         struct request *prev = elv_former_request(q, rq);
 835
 836         if (prev)
 837                 return attempt_merge(q, prev, rq);
 838
 839         return NULL;
 840 }
 841
 842 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 843                           struct request *next)
 844 {
 845         struct request *free;
 846
 847         free = attempt_merge(q, rq, next);
 848         if (free) {
 849                 blk_put_request(free);
 850                 return 1;
 851         }
 852
 853         return 0;
 854 }
 855
 856 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 857 {
 858         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 859                 return false;
 860
 861         if (req_op(rq) != bio_op(bio))
 862                 return false;
 863
 864         /* different data direction or already started, don't merge */
 865         if (bio_data_dir(bio) != rq_data_dir(rq))
 866                 return false;
 867
 868         /* must be same device */
 869         if (rq->rq_disk != bio->bi_disk)
 870                 return false;
 871
 872         /* only merge integrity protected bio into ditto rq */
 873         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 874                 return false;
 875
 876         /* Only merge if the crypt contexts are compatible */
 877         if (!bio_crypt_rq_ctx_compatible(rq, bio))
 878                 return false;
 879
 880         /* must be using the same buffer */
 881         if (req_op(rq) == REQ_OP_WRITE_SAME &&
 882             !blk_write_same_mergeable(rq->bio, bio))
 883                 return false;
 884
 885         /*
 886          * Don't allow merge of different write hints, or for a hint with
 887          * non-hint IO.
 888          */
 889         if (rq->write_hint != bio->bi_write_hint)
 890                 return false;
 891
 892         if (rq->ioprio != bio_prio(bio))
 893                 return false;
 894
 895         return true;
 896 }
 897
 898 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
 899 {
 900         if (blk_discard_mergable(rq))
 901                 return ELEVATOR_DISCARD_MERGE;
 902         else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 903                 return ELEVATOR_BACK_MERGE;
 904         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 905                 return ELEVATOR_FRONT_MERGE;
 906         return ELEVATOR_NO_MERGE;
 907 }
 908
 909 static void blk_account_io_merge_bio(struct request *req)
 910 {
 911         if (!blk_do_io_stat(req))
 912                 return;
 913
 914         part_stat_lock();
 915         part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
 916         part_stat_unlock();
 917 }
 918
 919 enum bio_merge_status {
 920         BIO_MERGE_OK,
 921         BIO_MERGE_NONE,
 922         BIO_MERGE_FAILED,
 923 };
 924
 925 static enum bio_merge_status bio_attempt_back_merge(struct request *req,
 926                 struct bio *bio, unsigned int nr_segs)
 927 {
 928         const int ff = bio->bi_opf & REQ_FAILFAST_MASK;
 929
 930         if (!ll_back_merge_fn(req, bio, nr_segs))
 931                 return BIO_MERGE_FAILED;
 932
 933         trace_block_bio_backmerge(bio);
 934         rq_qos_merge(req->q, req, bio);
 935
 936         if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
 937                 blk_rq_set_mixed_merge(req);
 938
 939         req->biotail->bi_next = bio;
 940         req->biotail = bio;
 941         req->__data_len += bio->bi_iter.bi_size;
 942
 943         bio_crypt_free_ctx(bio);
 944
 945         blk_account_io_merge_bio(req);
 946         return BIO_MERGE_OK;
 947 }
 948
 949 static enum bio_merge_status bio_attempt_front_merge(struct request *req,
 950                 struct bio *bio, unsigned int nr_segs)
 951 {
 952         const int ff = bio->bi_opf & REQ_FAILFAST_MASK;
 953
 954         if (!ll_front_merge_fn(req, bio, nr_segs))
 955                 return BIO_MERGE_FAILED;
 956
 957         trace_block_bio_frontmerge(bio);
 958         rq_qos_merge(req->q, req, bio);
 959
 960         if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
 961                 blk_rq_set_mixed_merge(req);
 962
 963         bio->bi_next = req->bio;
 964         req->bio = bio;
 965
 966         req->__sector = bio->bi_iter.bi_sector;
 967         req->__data_len += bio->bi_iter.bi_size;
 968
 969         bio_crypt_do_front_merge(req, bio);
 970
 971         blk_account_io_merge_bio(req);
 972         return BIO_MERGE_OK;
 973 }
 974
 975 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q,
 976                 struct request *req, struct bio *bio)
 977 {
 978         unsigned short segments = blk_rq_nr_discard_segments(req);
 979
 980         if (segments >= queue_max_discard_segments(q))
 981                 goto no_merge;
 982         if (blk_rq_sectors(req) + bio_sectors(bio) >
 983             blk_rq_get_max_sectors(req, blk_rq_pos(req)))
 984                 goto no_merge;
 985
 986         rq_qos_merge(q, req, bio);
 987
 988         req->biotail->bi_next = bio;
 989         req->biotail = bio;
 990         req->__data_len += bio->bi_iter.bi_size;
 991         req->nr_phys_segments = segments + 1;
 992
 993         blk_account_io_merge_bio(req);
 994         return BIO_MERGE_OK;
 995 no_merge:
 996         req_set_nomerge(q, req);
 997         return BIO_MERGE_FAILED;
 998 }
 999
1000 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q,
1001                                                    struct request *rq,
1002                                                    struct bio *bio,
1003                                                    unsigned int nr_segs,
1004                                                    bool sched_allow_merge)
1005 {
1006         if (!blk_rq_merge_ok(rq, bio))
1007                 return BIO_MERGE_NONE;
1008
1009         switch (blk_try_merge(rq, bio)) {
1010         case ELEVATOR_BACK_MERGE:
1011                 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1012                         return bio_attempt_back_merge(rq, bio, nr_segs);
1013                 break;
1014         case ELEVATOR_FRONT_MERGE:
1015                 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1016                         return bio_attempt_front_merge(rq, bio, nr_segs);
1017                 break;
1018         case ELEVATOR_DISCARD_MERGE:
1019                 return bio_attempt_discard_merge(q, rq, bio);
1020         default:
1021                 return BIO_MERGE_NONE;
1022         }
1023
1024         return BIO_MERGE_FAILED;
1025 }
1026
1027 /**
1028  * blk_attempt_plug_merge - try to merge with %current's plugged list
1029  * @q: request_queue new bio is being queued at
1030  * @bio: new bio being queued
1031  * @nr_segs: number of segments in @bio
1032  * @same_queue_rq: pointer to &struct request that gets filled in when
1033  * another request associated with @q is found on the plug list
1034  * (optional, may be %NULL)
1035  *
1036  * Determine whether @bio being queued on @q can be merged with a request
1037  * on %current's plugged list.  Returns %true if merge was successful,
1038  * otherwise %false.
1039  *
1040  * Plugging coalesces IOs from the same issuer for the same purpose without
1041  * going through @q->queue_lock.  As such it's more of an issuing mechanism
1042  * than scheduling, and the request, while may have elvpriv data, is not
1043  * added on the elevator at this point.  In addition, we don't have
1044  * reliable access to the elevator outside queue lock.  Only check basic
1045  * merging parameters without querying the elevator.
1046  *
1047  * Caller must ensure !blk_queue_nomerges(q) beforehand.
1048  */
1049 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio,
1050                 unsigned int nr_segs, struct request **same_queue_rq)
1051 {
1052         struct blk_plug *plug;
1053         struct request *rq;
1054         struct list_head *plug_list;
1055
1056         plug = blk_mq_plug(q, bio);
1057         if (!plug)
1058                 return false;
1059
1060         plug_list = &plug->mq_list;
1061
1062         list_for_each_entry_reverse(rq, plug_list, queuelist) {
1063                 if (rq->q == q && same_queue_rq) {
1064                         /*
1065                          * Only blk-mq multiple hardware queues case checks the
1066                          * rq in the same queue, there should be only one such
1067                          * rq in a queue
1068                          **/
1069                         *same_queue_rq = rq;
1070                 }
1071
1072                 if (rq->q != q)
1073                         continue;
1074
1075                 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) ==
1076                     BIO_MERGE_OK)
1077                         return true;
1078         }
1079
1080         return false;
1081 }
1082
1083 /*
1084  * Iterate list of requests and see if we can merge this bio with any
1085  * of them.
1086  */
1087 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list,
1088                         struct bio *bio, unsigned int nr_segs)
1089 {
1090         struct request *rq;
1091         int checked = 8;
1092
1093         list_for_each_entry_reverse(rq, list, queuelist) {
1094                 if (!checked--)
1095                         break;
1096
1097                 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) {
1098                 case BIO_MERGE_NONE:
1099                         continue;
1100                 case BIO_MERGE_OK:
1101                         return true;
1102                 case BIO_MERGE_FAILED:
1103                         return false;
1104                 }
1105
1106         }
1107
1108         return false;
1109 }
1110 EXPORT_SYMBOL_GPL(blk_bio_list_merge);
1111
1112 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
1113                 unsigned int nr_segs, struct request **merged_request)
1114 {
1115         struct request *rq;
1116
1117         switch (elv_merge(q, &rq, bio)) {
1118         case ELEVATOR_BACK_MERGE:
1119                 if (!blk_mq_sched_allow_merge(q, rq, bio))
1120                         return false;
1121                 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1122                         return false;
1123                 *merged_request = attempt_back_merge(q, rq);
1124                 if (!*merged_request)
1125                         elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
1126                 return true;
1127         case ELEVATOR_FRONT_MERGE:
1128                 if (!blk_mq_sched_allow_merge(q, rq, bio))
1129                         return false;
1130                 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1131                         return false;
1132                 *merged_request = attempt_front_merge(q, rq);
1133                 if (!*merged_request)
1134                         elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
1135                 return true;
1136         case ELEVATOR_DISCARD_MERGE:
1137                 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK;
1138         default:
1139                 return false;
1140         }
1141 }
1142 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);