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