block/blk-merge.c

   1 /*
   2  * Functions related to segment and merge handling
   3  */
   4 #include <linux/kernel.h>
   5 #include <linux/module.h>
   6 #include <linux/bio.h>
   7 #include <linux/blkdev.h>
   8 #include <linux/scatterlist.h>
   9 #include <linux/pfk.h>
  10 #include <linux/pft.h>
  11
  12 #include "blk.h"
  13
  14 static struct bio *blk_bio_discard_split(struct request_queue *q,
  15                                          struct bio *bio,
  16                                          struct bio_set *bs,
  17                                          unsigned *nsegs)
  18 {
  19         unsigned int max_discard_sectors, granularity;
  20         int alignment;
  21         sector_t tmp;
  22         unsigned split_sectors;
  23
  24         *nsegs = 1;
  25
  26         /* Zero-sector (unknown) and one-sector granularities are the same.  */
  27         granularity = max(q->limits.discard_granularity >> 9, 1U);
  28
  29         max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
  30         max_discard_sectors -= max_discard_sectors % granularity;
  31
  32         if (unlikely(!max_discard_sectors)) {
  33                 /* XXX: warn */
  34                 return NULL;
  35         }
  36
  37         if (bio_sectors(bio) <= max_discard_sectors)
  38                 return NULL;
  39
  40         split_sectors = max_discard_sectors;
  41
  42         /*
  43          * If the next starting sector would be misaligned, stop the discard at
  44          * the previous aligned sector.
  45          */
  46         alignment = (q->limits.discard_alignment >> 9) % granularity;
  47
  48         tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
  49         tmp = sector_div(tmp, granularity);
  50
  51         if (split_sectors > tmp)
  52                 split_sectors -= tmp;
  53
  54         return bio_split(bio, split_sectors, GFP_NOIO, bs);
  55 }
  56
  57 static struct bio *blk_bio_write_same_split(struct request_queue *q,
  58                                             struct bio *bio,
  59                                             struct bio_set *bs,
  60                                             unsigned *nsegs)
  61 {
  62         *nsegs = 1;
  63
  64         if (!q->limits.max_write_same_sectors)
  65                 return NULL;
  66
  67         if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
  68                 return NULL;
  69
  70         return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
  71 }
  72
  73 static inline unsigned get_max_io_size(struct request_queue *q,
  74                                        struct bio *bio)
  75 {
  76         unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
  77         unsigned mask = queue_logical_block_size(q) - 1;
  78
  79         /* aligned to logical block size */
  80         sectors &= ~(mask >> 9);
  81
  82         return sectors;
  83 }
  84
  85 static struct bio *blk_bio_segment_split(struct request_queue *q,
  86                                          struct bio *bio,
  87                                          struct bio_set *bs,
  88                                          unsigned *segs)
  89 {
  90         struct bio_vec bv, bvprv, *bvprvp = NULL;
  91         struct bvec_iter iter;
  92         unsigned seg_size = 0, nsegs = 0, sectors = 0;
  93         unsigned front_seg_size = bio->bi_seg_front_size;
  94         bool do_split = true;
  95         struct bio *new = NULL;
  96         const unsigned max_sectors = get_max_io_size(q, bio);
  97         unsigned bvecs = 0;
  98
  99         bio_for_each_segment(bv, bio, iter) {
 100                 /*
 101                  * With arbitrary bio size, the incoming bio may be very
 102                  * big. We have to split the bio into small bios so that
 103                  * each holds at most BIO_MAX_PAGES bvecs because
 104                  * bio_clone() can fail to allocate big bvecs.
 105                  *
 106                  * It should have been better to apply the limit per
 107                  * request queue in which bio_clone() is involved,
 108                  * instead of globally. The biggest blocker is the
 109                  * bio_clone() in bio bounce.
 110                  *
 111                  * If bio is splitted by this reason, we should have
 112                  * allowed to continue bios merging, but don't do
 113                  * that now for making the change simple.
 114                  *
 115                  * TODO: deal with bio bounce's bio_clone() gracefully
 116                  * and convert the global limit into per-queue limit.
 117                  */
 118                 if (bvecs++ >= BIO_MAX_PAGES)
 119                         goto split;
 120
 121                 /*
 122                  * If the queue doesn't support SG gaps and adding this
 123                  * offset would create a gap, disallow it.
 124                  */
 125                 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
 126                         goto split;
 127
 128                 if (sectors + (bv.bv_len >> 9) > max_sectors) {
 129                         /*
 130                          * Consider this a new segment if we're splitting in
 131                          * the middle of this vector.
 132                          */
 133                         if (nsegs < queue_max_segments(q) &&
 134                             sectors < max_sectors) {
 135                                 nsegs++;
 136                                 sectors = max_sectors;
 137                         }
 138                         if (sectors)
 139                                 goto split;
 140                         /* Make this single bvec as the 1st segment */
 141                 }
 142
 143                 if (bvprvp && blk_queue_cluster(q)) {
 144                         if (seg_size + bv.bv_len > queue_max_segment_size(q))
 145                                 goto new_segment;
 146                         if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
 147                                 goto new_segment;
 148                         if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
 149                                 goto new_segment;
 150
 151                         seg_size += bv.bv_len;
 152                         bvprv = bv;
 153                         bvprvp = &bvprv;
 154                         sectors += bv.bv_len >> 9;
 155
 156                         if (nsegs == 1 && seg_size > front_seg_size)
 157                                 front_seg_size = seg_size;
 158                         continue;
 159                 }
 160 new_segment:
 161                 if (nsegs == queue_max_segments(q))
 162                         goto split;
 163
 164                 nsegs++;
 165                 bvprv = bv;
 166                 bvprvp = &bvprv;
 167                 seg_size = bv.bv_len;
 168                 sectors += bv.bv_len >> 9;
 169
 170                 if (nsegs == 1 && seg_size > front_seg_size)
 171                         front_seg_size = seg_size;
 172         }
 173
 174         do_split = false;
 175 split:
 176         *segs = nsegs;
 177
 178         if (do_split) {
 179                 new = bio_split(bio, sectors, GFP_NOIO, bs);
 180                 if (new)
 181                         bio = new;
 182         }
 183
 184         bio->bi_seg_front_size = front_seg_size;
 185         if (seg_size > bio->bi_seg_back_size)
 186                 bio->bi_seg_back_size = seg_size;
 187
 188         return do_split ? new : NULL;
 189 }
 190
 191 void blk_queue_split(struct request_queue *q, struct bio **bio,
 192                      struct bio_set *bs)
 193 {
 194         struct bio *split, *res;
 195         unsigned nsegs;
 196
 197         if ((*bio)->bi_rw & REQ_DISCARD)
 198                 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
 199         else if ((*bio)->bi_rw & REQ_WRITE_SAME)
 200                 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
 201         else
 202                 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
 203
 204         /* physical segments can be figured out during splitting */
 205         res = split ? split : *bio;
 206         res->bi_phys_segments = nsegs;
 207         bio_set_flag(res, BIO_SEG_VALID);
 208
 209         if (split) {
 210                 /* there isn't chance to merge the splitted bio */
 211                 split->bi_rw |= REQ_NOMERGE;
 212
 213                 bio_chain(split, *bio);
 214                 generic_make_request(*bio);
 215                 *bio = split;
 216         }
 217 }
 218 EXPORT_SYMBOL(blk_queue_split);
 219
 220 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
 221                                              struct bio *bio,
 222                                              bool no_sg_merge)
 223 {
 224         struct bio_vec bv, bvprv = { NULL };
 225         int cluster, prev = 0;
 226         unsigned int seg_size, nr_phys_segs;
 227         struct bio *fbio, *bbio;
 228         struct bvec_iter iter;
 229
 230         if (!bio)
 231                 return 0;
 232
 233         /*
 234          * This should probably be returning 0, but blk_add_request_payload()
 235          * (Christoph!!!!)
 236          */
 237         if (bio->bi_rw & REQ_DISCARD)
 238                 return 1;
 239
 240         if (bio->bi_rw & REQ_WRITE_SAME)
 241                 return 1;
 242
 243         fbio = bio;
 244         cluster = blk_queue_cluster(q);
 245         seg_size = 0;
 246         nr_phys_segs = 0;
 247         for_each_bio(bio) {
 248                 bio_for_each_segment(bv, bio, iter) {
 249                         /*
 250                          * If SG merging is disabled, each bio vector is
 251                          * a segment
 252                          */
 253                         if (no_sg_merge)
 254                                 goto new_segment;
 255
 256                         if (prev && cluster) {
 257                                 if (seg_size + bv.bv_len
 258                                     > queue_max_segment_size(q))
 259                                         goto new_segment;
 260                                 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
 261                                         goto new_segment;
 262                                 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
 263                                         goto new_segment;
 264
 265                                 seg_size += bv.bv_len;
 266                                 bvprv = bv;
 267                                 continue;
 268                         }
 269 new_segment:
 270                         if (nr_phys_segs == 1 && seg_size >
 271                             fbio->bi_seg_front_size)
 272                                 fbio->bi_seg_front_size = seg_size;
 273
 274                         nr_phys_segs++;
 275                         bvprv = bv;
 276                         prev = 1;
 277                         seg_size = bv.bv_len;
 278                 }
 279                 bbio = bio;
 280         }
 281
 282         if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
 283                 fbio->bi_seg_front_size = seg_size;
 284         if (seg_size > bbio->bi_seg_back_size)
 285                 bbio->bi_seg_back_size = seg_size;
 286
 287         return nr_phys_segs;
 288 }
 289
 290 void blk_recalc_rq_segments(struct request *rq)
 291 {
 292         bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
 293                         &rq->q->queue_flags);
 294
 295         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
 296                         no_sg_merge);
 297 }
 298
 299 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 300 {
 301         unsigned short seg_cnt;
 302
 303         /* estimate segment number by bi_vcnt for non-cloned bio */
 304         if (bio_flagged(bio, BIO_CLONED))
 305                 seg_cnt = bio_segments(bio);
 306         else
 307                 seg_cnt = bio->bi_vcnt;
 308
 309         if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
 310                         (seg_cnt < queue_max_segments(q)))
 311                 bio->bi_phys_segments = seg_cnt;
 312         else {
 313                 struct bio *nxt = bio->bi_next;
 314
 315                 bio->bi_next = NULL;
 316                 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
 317                 bio->bi_next = nxt;
 318         }
 319
 320         bio_set_flag(bio, BIO_SEG_VALID);
 321 }
 322 EXPORT_SYMBOL(blk_recount_segments);
 323
 324 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 325                                    struct bio *nxt)
 326 {
 327         struct bio_vec end_bv = { NULL }, nxt_bv;
 328         struct bvec_iter iter;
 329
 330         if (!blk_queue_cluster(q))
 331                 return 0;
 332
 333         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 334             queue_max_segment_size(q))
 335                 return 0;
 336
 337         if (!bio_has_data(bio))
 338                 return 1;
 339
 340         bio_for_each_segment(end_bv, bio, iter)
 341                 if (end_bv.bv_len == iter.bi_size)
 342                         break;
 343
 344         nxt_bv = bio_iovec(nxt);
 345
 346         if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
 347                 return 0;
 348
 349         /*
 350          * bio and nxt are contiguous in memory; check if the queue allows
 351          * these two to be merged into one
 352          */
 353         if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
 354                 return 1;
 355
 356         return 0;
 357 }
 358
 359 static inline void
 360 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
 361                      struct scatterlist *sglist, struct bio_vec *bvprv,
 362                      struct scatterlist **sg, int *nsegs, int *cluster)
 363 {
 364
 365         int nbytes = bvec->bv_len;
 366
 367         if (*sg && *cluster) {
 368                 if ((*sg)->length + nbytes > queue_max_segment_size(q))
 369                         goto new_segment;
 370
 371                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 372                         goto new_segment;
 373                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 374                         goto new_segment;
 375
 376                 (*sg)->length += nbytes;
 377         } else {
 378 new_segment:
 379                 if (!*sg)
 380                         *sg = sglist;
 381                 else {
 382                         /*
 383                          * If the driver previously mapped a shorter
 384                          * list, we could see a termination bit
 385                          * prematurely unless it fully inits the sg
 386                          * table on each mapping. We KNOW that there
 387                          * must be more entries here or the driver
 388                          * would be buggy, so force clear the
 389                          * termination bit to avoid doing a full
 390                          * sg_init_table() in drivers for each command.
 391                          */
 392                         sg_unmark_end(*sg);
 393                         *sg = sg_next(*sg);
 394                 }
 395
 396                 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
 397                 (*nsegs)++;
 398         }
 399         *bvprv = *bvec;
 400 }
 401
 402 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 403                              struct scatterlist *sglist,
 404                              struct scatterlist **sg)
 405 {
 406         struct bio_vec bvec, bvprv = { NULL };
 407         struct bvec_iter iter;
 408         int nsegs, cluster;
 409
 410         nsegs = 0;
 411         cluster = blk_queue_cluster(q);
 412
 413         if (bio->bi_rw & REQ_DISCARD) {
 414                 /*
 415                  * This is a hack - drivers should be neither modifying the
 416                  * biovec, nor relying on bi_vcnt - but because of
 417                  * blk_add_request_payload(), a discard bio may or may not have
 418                  * a payload we need to set up here (thank you Christoph) and
 419                  * bi_vcnt is really the only way of telling if we need to.
 420                  */
 421
 422                 if (bio->bi_vcnt)
 423                         goto single_segment;
 424
 425                 return 0;
 426         }
 427
 428         if (bio->bi_rw & REQ_WRITE_SAME) {
 429 single_segment:
 430                 *sg = sglist;
 431                 bvec = bio_iovec(bio);
 432                 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
 433                 return 1;
 434         }
 435
 436         for_each_bio(bio)
 437                 bio_for_each_segment(bvec, bio, iter)
 438                         __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
 439                                              &nsegs, &cluster);
 440
 441         return nsegs;
 442 }
 443
 444 /*
 445  * map a request to scatterlist, return number of sg entries setup. Caller
 446  * must make sure sg can hold rq->nr_phys_segments entries
 447  */
 448 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 449                   struct scatterlist *sglist)
 450 {
 451         struct scatterlist *sg = NULL;
 452         int nsegs = 0;
 453
 454         if (rq->bio)
 455                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
 456
 457         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 458             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 459                 unsigned int pad_len =
 460                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 461
 462                 sg->length += pad_len;
 463                 rq->extra_len += pad_len;
 464         }
 465
 466         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 467                 if (rq->cmd_flags & REQ_WRITE)
 468                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 469
 470                 sg_unmark_end(sg);
 471                 sg = sg_next(sg);
 472                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 473                             q->dma_drain_size,
 474                             ((unsigned long)q->dma_drain_buffer) &
 475                             (PAGE_SIZE - 1));
 476                 nsegs++;
 477                 rq->extra_len += q->dma_drain_size;
 478         }
 479
 480         if (sg)
 481                 sg_mark_end(sg);
 482
 483         /*
 484          * Something must have been wrong if the figured number of
 485          * segment is bigger than number of req's physical segments
 486          */
 487         WARN_ON(nsegs > rq->nr_phys_segments);
 488
 489         return nsegs;
 490 }
 491 EXPORT_SYMBOL(blk_rq_map_sg);
 492
 493 /*
 494  * map a request to scatterlist without combining PHY CONT
 495  * blocks, return number of sg entries setup. Caller
 496  * must make sure sg can hold rq->nr_phys_segments entries
 497  */
 498 int blk_rq_map_sg_no_cluster(struct request_queue *q, struct request *rq,
 499                   struct scatterlist *sglist)
 500 {
 501         struct bio_vec bvec, bvprv = { NULL };
 502         struct req_iterator iter;
 503         struct scatterlist *sg;
 504         int nsegs, cluster = 0;
 505
 506         nsegs = 0;
 507
 508         /*
 509          * for each bio in rq
 510          */
 511         sg = NULL;
 512         rq_for_each_segment(bvec, rq, iter) {
 513                 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, &sg,
 514                                      &nsegs, &cluster);
 515         } /* segments in rq */
 516
 517
 518         if (!sg)
 519                 return nsegs;
 520
 521         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 522             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 523                 unsigned int pad_len =
 524                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 525
 526                 sg->length += pad_len;
 527                 rq->extra_len += pad_len;
 528         }
 529
 530         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 531                 if (rq->cmd_flags & REQ_WRITE)
 532                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 533
 534                 sg->page_link &= ~0x02;
 535                 sg = sg_next(sg);
 536                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 537                             q->dma_drain_size,
 538                             ((unsigned long)q->dma_drain_buffer) &
 539                             (PAGE_SIZE - 1));
 540                 nsegs++;
 541                 rq->extra_len += q->dma_drain_size;
 542         }
 543
 544         if (sg)
 545                 sg_mark_end(sg);
 546
 547         return nsegs;
 548 }
 549 EXPORT_SYMBOL(blk_rq_map_sg_no_cluster);
 550
 551 /**
 552  * blk_bio_map_sg - map a bio to a scatterlist
 553  * @q: request_queue in question
 554  * @bio: bio being mapped
 555  * @sglist: scatterlist being mapped
 556  *
 557  * Note:
 558  *    Caller must make sure sg can hold bio->bi_phys_segments entries
 559  *
 560  * Will return the number of sg entries setup
 561  */
 562 int blk_bio_map_sg(struct request_queue *q, struct bio *bio,
 563                    struct scatterlist *sglist)
 564 {
 565         struct scatterlist *sg = NULL;
 566         int nsegs;
 567         struct bio *next = bio->bi_next;
 568
 569         bio->bi_next = NULL;
 570         nsegs = __blk_bios_map_sg(q, bio, sglist, &sg);
 571         bio->bi_next = next;
 572         if (sg)
 573                 sg_mark_end(sg);
 574
 575         WARN_ON(bio->bi_phys_segments && nsegs > bio->bi_phys_segments);
 576         return nsegs;
 577 }
 578 EXPORT_SYMBOL(blk_bio_map_sg);
 579
 580 static inline int ll_new_hw_segment(struct request_queue *q,
 581                                     struct request *req,
 582                                     struct bio *bio)
 583 {
 584         int nr_phys_segs = bio_phys_segments(q, bio);
 585
 586         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 587                 goto no_merge;
 588
 589         if (blk_integrity_merge_bio(q, req, bio) == false)
 590                 goto no_merge;
 591
 592         /*
 593          * This will form the start of a new hw segment.  Bump both
 594          * counters.
 595          */
 596         req->nr_phys_segments += nr_phys_segs;
 597         return 1;
 598
 599 no_merge:
 600         req->cmd_flags |= REQ_NOMERGE;
 601         if (req == q->last_merge)
 602                 q->last_merge = NULL;
 603         return 0;
 604 }
 605
 606 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 607                      struct bio *bio)
 608 {
 609         if (req_gap_back_merge(req, bio))
 610                 return 0;
 611         if (blk_integrity_rq(req) &&
 612             integrity_req_gap_back_merge(req, bio))
 613                 return 0;
 614         if (blk_rq_sectors(req) + bio_sectors(bio) >
 615             blk_rq_get_max_sectors(req)) {
 616                 req->cmd_flags |= REQ_NOMERGE;
 617                 if (req == q->last_merge)
 618                         q->last_merge = NULL;
 619                 return 0;
 620         }
 621         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 622                 blk_recount_segments(q, req->biotail);
 623         if (!bio_flagged(bio, BIO_SEG_VALID))
 624                 blk_recount_segments(q, bio);
 625
 626         return ll_new_hw_segment(q, req, bio);
 627 }
 628
 629 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 630                       struct bio *bio)
 631 {
 632
 633         if (req_gap_front_merge(req, bio))
 634                 return 0;
 635         if (blk_integrity_rq(req) &&
 636             integrity_req_gap_front_merge(req, bio))
 637                 return 0;
 638         if (blk_rq_sectors(req) + bio_sectors(bio) >
 639             blk_rq_get_max_sectors(req)) {
 640                 req->cmd_flags |= REQ_NOMERGE;
 641                 if (req == q->last_merge)
 642                         q->last_merge = NULL;
 643                 return 0;
 644         }
 645         if (!bio_flagged(bio, BIO_SEG_VALID))
 646                 blk_recount_segments(q, bio);
 647         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 648                 blk_recount_segments(q, req->bio);
 649
 650         return ll_new_hw_segment(q, req, bio);
 651 }
 652
 653 /*
 654  * blk-mq uses req->special to carry normal driver per-request payload, it
 655  * does not indicate a prepared command that we cannot merge with.
 656  */
 657 static bool req_no_special_merge(struct request *req)
 658 {
 659         struct request_queue *q = req->q;
 660
 661         return !q->mq_ops && req->special;
 662 }
 663
 664 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 665                                 struct request *next)
 666 {
 667         int total_phys_segments;
 668         unsigned int seg_size =
 669                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 670
 671         /*
 672          * First check if the either of the requests are re-queued
 673          * requests.  Can't merge them if they are.
 674          */
 675         if (req_no_special_merge(req) || req_no_special_merge(next))
 676                 return 0;
 677
 678         if (req_gap_back_merge(req, next->bio))
 679                 return 0;
 680
 681         /*
 682          * Will it become too large?
 683          */
 684         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 685             blk_rq_get_max_sectors(req))
 686                 return 0;
 687
 688         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 689         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 690                 if (req->nr_phys_segments == 1)
 691                         req->bio->bi_seg_front_size = seg_size;
 692                 if (next->nr_phys_segments == 1)
 693                         next->biotail->bi_seg_back_size = seg_size;
 694                 total_phys_segments--;
 695         }
 696
 697         if (total_phys_segments > queue_max_segments(q))
 698                 return 0;
 699
 700         if (blk_integrity_merge_rq(q, req, next) == false)
 701                 return 0;
 702
 703         /* Merge is OK... */
 704         req->nr_phys_segments = total_phys_segments;
 705         return 1;
 706 }
 707
 708 /**
 709  * blk_rq_set_mixed_merge - mark a request as mixed merge
 710  * @rq: request to mark as mixed merge
 711  *
 712  * Description:
 713  *     @rq is about to be mixed merged.  Make sure the attributes
 714  *     which can be mixed are set in each bio and mark @rq as mixed
 715  *     merged.
 716  */
 717 void blk_rq_set_mixed_merge(struct request *rq)
 718 {
 719         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 720         struct bio *bio;
 721
 722         if (rq->cmd_flags & REQ_MIXED_MERGE)
 723                 return;
 724
 725         /*
 726          * @rq will no longer represent mixable attributes for all the
 727          * contained bios.  It will just track those of the first one.
 728          * Distributes the attributs to each bio.
 729          */
 730         for (bio = rq->bio; bio; bio = bio->bi_next) {
 731                 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
 732                              (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
 733                 bio->bi_rw |= ff;
 734         }
 735         rq->cmd_flags |= REQ_MIXED_MERGE;
 736 }
 737
 738 static void blk_account_io_merge(struct request *req)
 739 {
 740         if (blk_do_io_stat(req)) {
 741                 struct hd_struct *part;
 742                 int cpu;
 743
 744                 cpu = part_stat_lock();
 745                 part = req->part;
 746
 747                 part_round_stats(cpu, part);
 748                 part_dec_in_flight(part, rq_data_dir(req));
 749
 750                 hd_struct_put(part);
 751                 part_stat_unlock();
 752         }
 753 }
 754
 755 static bool crypto_not_mergeable(const struct bio *bio, const struct bio *nxt)
 756 {
 757         return (!pft_allow_merge_bio(bio, nxt) ||
 758                 !pfk_allow_merge_bio(bio, nxt));
 759 }
 760
 761 /*
 762  * Has to be called with the request spinlock acquired
 763  */
 764 static int attempt_merge(struct request_queue *q, struct request *req,
 765                           struct request *next)
 766 {
 767         if (!rq_mergeable(req) || !rq_mergeable(next))
 768                 return 0;
 769
 770         if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
 771                 return 0;
 772
 773         /*
 774          * not contiguous
 775          */
 776         if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
 777                 return 0;
 778
 779         if (rq_data_dir(req) != rq_data_dir(next)
 780             || req->rq_disk != next->rq_disk
 781             || req_no_special_merge(next))
 782                 return 0;
 783
 784         if (req->cmd_flags & REQ_WRITE_SAME &&
 785             !blk_write_same_mergeable(req->bio, next->bio))
 786                 return 0;
 787
 788         if (crypto_not_mergeable(req->bio, next->bio))
 789                 return 0;
 790
 791         /*
 792          * If we are allowed to merge, then append bio list
 793          * from next to rq and release next. merge_requests_fn
 794          * will have updated segment counts, update sector
 795          * counts here.
 796          */
 797         if (!ll_merge_requests_fn(q, req, next))
 798                 return 0;
 799
 800         /*
 801          * If failfast settings disagree or any of the two is already
 802          * a mixed merge, mark both as mixed before proceeding.  This
 803          * makes sure that all involved bios have mixable attributes
 804          * set properly.
 805          */
 806         if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
 807             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 808             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 809                 blk_rq_set_mixed_merge(req);
 810                 blk_rq_set_mixed_merge(next);
 811         }
 812
 813         /*
 814          * At this point we have either done a back merge
 815          * or front merge. We need the smaller start_time of
 816          * the merged requests to be the current request
 817          * for accounting purposes.
 818          */
 819         if (time_after(req->start_time, next->start_time))
 820                 req->start_time = next->start_time;
 821
 822         req->biotail->bi_next = next->bio;
 823         req->biotail = next->biotail;
 824
 825         req->__data_len += blk_rq_bytes(next);
 826
 827         elv_merge_requests(q, req, next);
 828
 829         /*
 830          * 'next' is going away, so update stats accordingly
 831          */
 832         blk_account_io_merge(next);
 833
 834         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 835         if (blk_rq_cpu_valid(next))
 836                 req->cpu = next->cpu;
 837
 838         /* owner-ship of bio passed from next to req */
 839         next->bio = NULL;
 840         __blk_put_request(q, next);
 841         return 1;
 842 }
 843
 844 int attempt_back_merge(struct request_queue *q, struct request *rq)
 845 {
 846         struct request *next = elv_latter_request(q, rq);
 847
 848         if (next)
 849                 return attempt_merge(q, rq, next);
 850
 851         return 0;
 852 }
 853
 854 int attempt_front_merge(struct request_queue *q, struct request *rq)
 855 {
 856         struct request *prev = elv_former_request(q, rq);
 857
 858         if (prev)
 859                 return attempt_merge(q, prev, rq);
 860
 861         return 0;
 862 }
 863
 864 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 865                           struct request *next)
 866 {
 867         return attempt_merge(q, rq, next);
 868 }
 869
 870 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 871 {
 872         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 873                 return false;
 874
 875         if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
 876                 return false;
 877
 878         /* different data direction or already started, don't merge */
 879         if (bio_data_dir(bio) != rq_data_dir(rq))
 880                 return false;
 881
 882         /* must be same device and not a special request */
 883         if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
 884                 return false;
 885
 886         /* only merge integrity protected bio into ditto rq */
 887         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 888                 return false;
 889
 890         /* must be using the same buffer */
 891         if (rq->cmd_flags & REQ_WRITE_SAME &&
 892             !blk_write_same_mergeable(rq->bio, bio))
 893                 return false;
 894
 895         if (crypto_not_mergeable(rq->bio, bio))
 896                 return false;
 897
 898         return true;
 899 }
 900
 901 int blk_try_merge(struct request *rq, struct bio *bio)
 902 {
 903         if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 904                 return ELEVATOR_BACK_MERGE;
 905         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 906                 return ELEVATOR_FRONT_MERGE;
 907         return ELEVATOR_NO_MERGE;
 908 }