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
  10 #include "blk.h"
  11
  12 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
  13                                              struct bio *bio,
  14                                              bool no_sg_merge)
  15 {
  16         struct bio_vec bv, bvprv = { NULL };
  17         int cluster, high, highprv = 1;
  18         unsigned int seg_size, nr_phys_segs;
  19         struct bio *fbio, *bbio;
  20         struct bvec_iter iter;
  21
  22         if (!bio)
  23                 return 0;
  24
  25         /*
  26          * This should probably be returning 0, but blk_add_request_payload()
  27          * (Christoph!!!!)
  28          */
  29         if (bio->bi_rw & REQ_DISCARD)
  30                 return 1;
  31
  32         if (bio->bi_rw & REQ_WRITE_SAME)
  33                 return 1;
  34
  35         fbio = bio;
  36         cluster = blk_queue_cluster(q);
  37         seg_size = 0;
  38         nr_phys_segs = 0;
  39         high = 0;
  40         for_each_bio(bio) {
  41                 bio_for_each_segment(bv, bio, iter) {
  42                         /*
  43                          * If SG merging is disabled, each bio vector is
  44                          * a segment
  45                          */
  46                         if (no_sg_merge)
  47                                 goto new_segment;
  48
  49                         /*
  50                          * the trick here is making sure that a high page is
  51                          * never considered part of another segment, since
  52                          * that might change with the bounce page.
  53                          */
  54                         high = page_to_pfn(bv.bv_page) > queue_bounce_pfn(q);
  55                         if (!high && !highprv && cluster) {
  56                                 if (seg_size + bv.bv_len
  57                                     > queue_max_segment_size(q))
  58                                         goto new_segment;
  59                                 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
  60                                         goto new_segment;
  61                                 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
  62                                         goto new_segment;
  63
  64                                 seg_size += bv.bv_len;
  65                                 bvprv = bv;
  66                                 continue;
  67                         }
  68 new_segment:
  69                         if (nr_phys_segs == 1 && seg_size >
  70                             fbio->bi_seg_front_size)
  71                                 fbio->bi_seg_front_size = seg_size;
  72
  73                         nr_phys_segs++;
  74                         bvprv = bv;
  75                         seg_size = bv.bv_len;
  76                         highprv = high;
  77                 }
  78                 bbio = bio;
  79         }
  80
  81         if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
  82                 fbio->bi_seg_front_size = seg_size;
  83         if (seg_size > bbio->bi_seg_back_size)
  84                 bbio->bi_seg_back_size = seg_size;
  85
  86         return nr_phys_segs;
  87 }
  88
  89 void blk_recalc_rq_segments(struct request *rq)
  90 {
  91         bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
  92                         &rq->q->queue_flags);
  93
  94         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
  95                         no_sg_merge);
  96 }
  97
  98 void blk_recount_segments(struct request_queue *q, struct bio *bio)
  99 {
 100         bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
 101                         &q->queue_flags);
 102         bool merge_not_need = bio->bi_vcnt < queue_max_segments(q);
 103
 104         if (no_sg_merge && !bio_flagged(bio, BIO_CLONED) &&
 105                         merge_not_need)
 106                 bio->bi_phys_segments = bio->bi_vcnt;
 107         else {
 108                 struct bio *nxt = bio->bi_next;
 109
 110                 bio->bi_next = NULL;
 111                 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio,
 112                                 no_sg_merge && merge_not_need);
 113                 bio->bi_next = nxt;
 114         }
 115
 116         bio->bi_flags |= (1 << BIO_SEG_VALID);
 117 }
 118 EXPORT_SYMBOL(blk_recount_segments);
 119
 120 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 121                                    struct bio *nxt)
 122 {
 123         struct bio_vec end_bv = { NULL }, nxt_bv;
 124         struct bvec_iter iter;
 125
 126         if (!blk_queue_cluster(q))
 127                 return 0;
 128
 129         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 130             queue_max_segment_size(q))
 131                 return 0;
 132
 133         if (!bio_has_data(bio))
 134                 return 1;
 135
 136         bio_for_each_segment(end_bv, bio, iter)
 137                 if (end_bv.bv_len == iter.bi_size)
 138                         break;
 139
 140         nxt_bv = bio_iovec(nxt);
 141
 142         if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
 143                 return 0;
 144
 145         /*
 146          * bio and nxt are contiguous in memory; check if the queue allows
 147          * these two to be merged into one
 148          */
 149         if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
 150                 return 1;
 151
 152         return 0;
 153 }
 154
 155 static inline void
 156 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
 157                      struct scatterlist *sglist, struct bio_vec *bvprv,
 158                      struct scatterlist **sg, int *nsegs, int *cluster)
 159 {
 160
 161         int nbytes = bvec->bv_len;
 162
 163         if (*sg && *cluster) {
 164                 if ((*sg)->length + nbytes > queue_max_segment_size(q))
 165                         goto new_segment;
 166
 167                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 168                         goto new_segment;
 169                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 170                         goto new_segment;
 171
 172                 (*sg)->length += nbytes;
 173         } else {
 174 new_segment:
 175                 if (!*sg)
 176                         *sg = sglist;
 177                 else {
 178                         /*
 179                          * If the driver previously mapped a shorter
 180                          * list, we could see a termination bit
 181                          * prematurely unless it fully inits the sg
 182                          * table on each mapping. We KNOW that there
 183                          * must be more entries here or the driver
 184                          * would be buggy, so force clear the
 185                          * termination bit to avoid doing a full
 186                          * sg_init_table() in drivers for each command.
 187                          */
 188                         sg_unmark_end(*sg);
 189                         *sg = sg_next(*sg);
 190                 }
 191
 192                 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
 193                 (*nsegs)++;
 194         }
 195         *bvprv = *bvec;
 196 }
 197
 198 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
 199                              struct scatterlist *sglist,
 200                              struct scatterlist **sg)
 201 {
 202         struct bio_vec bvec, bvprv = { NULL };
 203         struct bvec_iter iter;
 204         int nsegs, cluster;
 205
 206         nsegs = 0;
 207         cluster = blk_queue_cluster(q);
 208
 209         if (bio->bi_rw & REQ_DISCARD) {
 210                 /*
 211                  * This is a hack - drivers should be neither modifying the
 212                  * biovec, nor relying on bi_vcnt - but because of
 213                  * blk_add_request_payload(), a discard bio may or may not have
 214                  * a payload we need to set up here (thank you Christoph) and
 215                  * bi_vcnt is really the only way of telling if we need to.
 216                  */
 217
 218                 if (bio->bi_vcnt)
 219                         goto single_segment;
 220
 221                 return 0;
 222         }
 223
 224         if (bio->bi_rw & REQ_WRITE_SAME) {
 225 single_segment:
 226                 *sg = sglist;
 227                 bvec = bio_iovec(bio);
 228                 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
 229                 return 1;
 230         }
 231
 232         for_each_bio(bio)
 233                 bio_for_each_segment(bvec, bio, iter)
 234                         __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
 235                                              &nsegs, &cluster);
 236
 237         return nsegs;
 238 }
 239
 240 /*
 241  * map a request to scatterlist, return number of sg entries setup. Caller
 242  * must make sure sg can hold rq->nr_phys_segments entries
 243  */
 244 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 245                   struct scatterlist *sglist)
 246 {
 247         struct scatterlist *sg = NULL;
 248         int nsegs = 0;
 249
 250         if (rq->bio)
 251                 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
 252
 253         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 254             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 255                 unsigned int pad_len =
 256                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 257
 258                 sg->length += pad_len;
 259                 rq->extra_len += pad_len;
 260         }
 261
 262         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 263                 if (rq->cmd_flags & REQ_WRITE)
 264                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 265
 266                 sg->page_link &= ~0x02;
 267                 sg = sg_next(sg);
 268                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 269                             q->dma_drain_size,
 270                             ((unsigned long)q->dma_drain_buffer) &
 271                             (PAGE_SIZE - 1));
 272                 nsegs++;
 273                 rq->extra_len += q->dma_drain_size;
 274         }
 275
 276         if (sg)
 277                 sg_mark_end(sg);
 278
 279         return nsegs;
 280 }
 281 EXPORT_SYMBOL(blk_rq_map_sg);
 282
 283 /**
 284  * blk_bio_map_sg - map a bio to a scatterlist
 285  * @q: request_queue in question
 286  * @bio: bio being mapped
 287  * @sglist: scatterlist being mapped
 288  *
 289  * Note:
 290  *    Caller must make sure sg can hold bio->bi_phys_segments entries
 291  *
 292  * Will return the number of sg entries setup
 293  */
 294 int blk_bio_map_sg(struct request_queue *q, struct bio *bio,
 295                    struct scatterlist *sglist)
 296 {
 297         struct scatterlist *sg = NULL;
 298         int nsegs;
 299         struct bio *next = bio->bi_next;
 300         bio->bi_next = NULL;
 301
 302         nsegs = __blk_bios_map_sg(q, bio, sglist, &sg);
 303         bio->bi_next = next;
 304         if (sg)
 305                 sg_mark_end(sg);
 306
 307         BUG_ON(bio->bi_phys_segments && nsegs > bio->bi_phys_segments);
 308         return nsegs;
 309 }
 310 EXPORT_SYMBOL(blk_bio_map_sg);
 311
 312 static inline int ll_new_hw_segment(struct request_queue *q,
 313                                     struct request *req,
 314                                     struct bio *bio)
 315 {
 316         int nr_phys_segs = bio_phys_segments(q, bio);
 317
 318         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 319                 goto no_merge;
 320
 321         if (blk_integrity_merge_bio(q, req, bio) == false)
 322                 goto no_merge;
 323
 324         /*
 325          * This will form the start of a new hw segment.  Bump both
 326          * counters.
 327          */
 328         req->nr_phys_segments += nr_phys_segs;
 329         return 1;
 330
 331 no_merge:
 332         req->cmd_flags |= REQ_NOMERGE;
 333         if (req == q->last_merge)
 334                 q->last_merge = NULL;
 335         return 0;
 336 }
 337
 338 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 339                      struct bio *bio)
 340 {
 341         if (blk_rq_sectors(req) + bio_sectors(bio) >
 342             blk_rq_get_max_sectors(req)) {
 343                 req->cmd_flags |= REQ_NOMERGE;
 344                 if (req == q->last_merge)
 345                         q->last_merge = NULL;
 346                 return 0;
 347         }
 348         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 349                 blk_recount_segments(q, req->biotail);
 350         if (!bio_flagged(bio, BIO_SEG_VALID))
 351                 blk_recount_segments(q, bio);
 352
 353         return ll_new_hw_segment(q, req, bio);
 354 }
 355
 356 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 357                       struct bio *bio)
 358 {
 359         if (blk_rq_sectors(req) + bio_sectors(bio) >
 360             blk_rq_get_max_sectors(req)) {
 361                 req->cmd_flags |= REQ_NOMERGE;
 362                 if (req == q->last_merge)
 363                         q->last_merge = NULL;
 364                 return 0;
 365         }
 366         if (!bio_flagged(bio, BIO_SEG_VALID))
 367                 blk_recount_segments(q, bio);
 368         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 369                 blk_recount_segments(q, req->bio);
 370
 371         return ll_new_hw_segment(q, req, bio);
 372 }
 373
 374 /*
 375  * blk-mq uses req->special to carry normal driver per-request payload, it
 376  * does not indicate a prepared command that we cannot merge with.
 377  */
 378 static bool req_no_special_merge(struct request *req)
 379 {
 380         struct request_queue *q = req->q;
 381
 382         return !q->mq_ops && req->special;
 383 }
 384
 385 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 386                                 struct request *next)
 387 {
 388         int total_phys_segments;
 389         unsigned int seg_size =
 390                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 391
 392         /*
 393          * First check if the either of the requests are re-queued
 394          * requests.  Can't merge them if they are.
 395          */
 396         if (req_no_special_merge(req) || req_no_special_merge(next))
 397                 return 0;
 398
 399         /*
 400          * Will it become too large?
 401          */
 402         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 403             blk_rq_get_max_sectors(req))
 404                 return 0;
 405
 406         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 407         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 408                 if (req->nr_phys_segments == 1)
 409                         req->bio->bi_seg_front_size = seg_size;
 410                 if (next->nr_phys_segments == 1)
 411                         next->biotail->bi_seg_back_size = seg_size;
 412                 total_phys_segments--;
 413         }
 414
 415         if (total_phys_segments > queue_max_segments(q))
 416                 return 0;
 417
 418         if (blk_integrity_merge_rq(q, req, next) == false)
 419                 return 0;
 420
 421         /* Merge is OK... */
 422         req->nr_phys_segments = total_phys_segments;
 423         return 1;
 424 }
 425
 426 /**
 427  * blk_rq_set_mixed_merge - mark a request as mixed merge
 428  * @rq: request to mark as mixed merge
 429  *
 430  * Description:
 431  *     @rq is about to be mixed merged.  Make sure the attributes
 432  *     which can be mixed are set in each bio and mark @rq as mixed
 433  *     merged.
 434  */
 435 void blk_rq_set_mixed_merge(struct request *rq)
 436 {
 437         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 438         struct bio *bio;
 439
 440         if (rq->cmd_flags & REQ_MIXED_MERGE)
 441                 return;
 442
 443         /*
 444          * @rq will no longer represent mixable attributes for all the
 445          * contained bios.  It will just track those of the first one.
 446          * Distributes the attributs to each bio.
 447          */
 448         for (bio = rq->bio; bio; bio = bio->bi_next) {
 449                 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
 450                              (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
 451                 bio->bi_rw |= ff;
 452         }
 453         rq->cmd_flags |= REQ_MIXED_MERGE;
 454 }
 455
 456 static void blk_account_io_merge(struct request *req)
 457 {
 458         if (blk_do_io_stat(req)) {
 459                 struct hd_struct *part;
 460                 int cpu;
 461
 462                 cpu = part_stat_lock();
 463                 part = req->part;
 464
 465                 part_round_stats(cpu, part);
 466                 part_dec_in_flight(part, rq_data_dir(req));
 467
 468                 hd_struct_put(part);
 469                 part_stat_unlock();
 470         }
 471 }
 472
 473 /*
 474  * Has to be called with the request spinlock acquired
 475  */
 476 static int attempt_merge(struct request_queue *q, struct request *req,
 477                           struct request *next)
 478 {
 479         if (!rq_mergeable(req) || !rq_mergeable(next))
 480                 return 0;
 481
 482         if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
 483                 return 0;
 484
 485         /*
 486          * not contiguous
 487          */
 488         if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
 489                 return 0;
 490
 491         if (rq_data_dir(req) != rq_data_dir(next)
 492             || req->rq_disk != next->rq_disk
 493             || req_no_special_merge(next))
 494                 return 0;
 495
 496         if (req->cmd_flags & REQ_WRITE_SAME &&
 497             !blk_write_same_mergeable(req->bio, next->bio))
 498                 return 0;
 499
 500         /*
 501          * If we are allowed to merge, then append bio list
 502          * from next to rq and release next. merge_requests_fn
 503          * will have updated segment counts, update sector
 504          * counts here.
 505          */
 506         if (!ll_merge_requests_fn(q, req, next))
 507                 return 0;
 508
 509         /*
 510          * If failfast settings disagree or any of the two is already
 511          * a mixed merge, mark both as mixed before proceeding.  This
 512          * makes sure that all involved bios have mixable attributes
 513          * set properly.
 514          */
 515         if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
 516             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 517             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 518                 blk_rq_set_mixed_merge(req);
 519                 blk_rq_set_mixed_merge(next);
 520         }
 521
 522         /*
 523          * At this point we have either done a back merge
 524          * or front merge. We need the smaller start_time of
 525          * the merged requests to be the current request
 526          * for accounting purposes.
 527          */
 528         if (time_after(req->start_time, next->start_time))
 529                 req->start_time = next->start_time;
 530
 531         req->biotail->bi_next = next->bio;
 532         req->biotail = next->biotail;
 533
 534         req->__data_len += blk_rq_bytes(next);
 535
 536         elv_merge_requests(q, req, next);
 537
 538         /*
 539          * 'next' is going away, so update stats accordingly
 540          */
 541         blk_account_io_merge(next);
 542
 543         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 544         if (blk_rq_cpu_valid(next))
 545                 req->cpu = next->cpu;
 546
 547         /* owner-ship of bio passed from next to req */
 548         next->bio = NULL;
 549         __blk_put_request(q, next);
 550         return 1;
 551 }
 552
 553 int attempt_back_merge(struct request_queue *q, struct request *rq)
 554 {
 555         struct request *next = elv_latter_request(q, rq);
 556
 557         if (next)
 558                 return attempt_merge(q, rq, next);
 559
 560         return 0;
 561 }
 562
 563 int attempt_front_merge(struct request_queue *q, struct request *rq)
 564 {
 565         struct request *prev = elv_former_request(q, rq);
 566
 567         if (prev)
 568                 return attempt_merge(q, prev, rq);
 569
 570         return 0;
 571 }
 572
 573 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
 574                           struct request *next)
 575 {
 576         return attempt_merge(q, rq, next);
 577 }
 578
 579 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 580 {
 581         struct request_queue *q = rq->q;
 582
 583         if (!rq_mergeable(rq) || !bio_mergeable(bio))
 584                 return false;
 585
 586         if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
 587                 return false;
 588
 589         /* different data direction or already started, don't merge */
 590         if (bio_data_dir(bio) != rq_data_dir(rq))
 591                 return false;
 592
 593         /* must be same device and not a special request */
 594         if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
 595                 return false;
 596
 597         /* only merge integrity protected bio into ditto rq */
 598         if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 599                 return false;
 600
 601         /* must be using the same buffer */
 602         if (rq->cmd_flags & REQ_WRITE_SAME &&
 603             !blk_write_same_mergeable(rq->bio, bio))
 604                 return false;
 605
 606         if (q->queue_flags & (1 << QUEUE_FLAG_SG_GAPS)) {
 607                 struct bio_vec *bprev;
 608
 609                 bprev = &rq->biotail->bi_io_vec[bio->bi_vcnt - 1];
 610                 if (bvec_gap_to_prev(bprev, bio->bi_io_vec[0].bv_offset))
 611                         return false;
 612         }
 613
 614         return true;
 615 }
 616
 617 int blk_try_merge(struct request *rq, struct bio *bio)
 618 {
 619         if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
 620                 return ELEVATOR_BACK_MERGE;
 621         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 622                 return ELEVATOR_FRONT_MERGE;
 623         return ELEVATOR_NO_MERGE;
 624 }