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