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