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