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         fbio = bio;
  25         cluster = blk_queue_cluster(q);
  26         seg_size = 0;
  27         nr_phys_segs = 0;
  28         for_each_bio(bio) {
  29                 bio_for_each_segment(bv, bio, iter) {
  30                         /*
  31                          * the trick here is making sure that a high page is
  32                          * never considered part of another segment, since that
  33                          * might change with the bounce page.
  34                          */
  35                         high = page_to_pfn(bv.bv_page) > queue_bounce_pfn(q);
  36                         if (!high && !highprv && cluster) {
  37                                 if (seg_size + bv.bv_len
  38                                     > queue_max_segment_size(q))
  39                                         goto new_segment;
  40                                 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
  41                                         goto new_segment;
  42                                 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
  43                                         goto new_segment;
  44
  45                                 seg_size += bv.bv_len;
  46                                 bvprv = bv;
  47                                 continue;
  48                         }
  49 new_segment:
  50                         if (nr_phys_segs == 1 && seg_size >
  51                             fbio->bi_seg_front_size)
  52                                 fbio->bi_seg_front_size = seg_size;
  53
  54                         nr_phys_segs++;
  55                         bvprv = bv;
  56                         seg_size = bv.bv_len;
  57                         highprv = high;
  58                 }
  59                 bbio = bio;
  60         }
  61
  62         if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
  63                 fbio->bi_seg_front_size = seg_size;
  64         if (seg_size > bbio->bi_seg_back_size)
  65                 bbio->bi_seg_back_size = seg_size;
  66
  67         return nr_phys_segs;
  68 }
  69
  70 void blk_recalc_rq_segments(struct request *rq)
  71 {
  72         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
  73 }
  74
  75 void blk_recount_segments(struct request_queue *q, struct bio *bio)
  76 {
  77         struct bio *nxt = bio->bi_next;
  78
  79         bio->bi_next = NULL;
  80         bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
  81         bio->bi_next = nxt;
  82         bio->bi_flags |= (1 << BIO_SEG_VALID);
  83 }
  84 EXPORT_SYMBOL(blk_recount_segments);
  85
  86 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
  87                                    struct bio *nxt)
  88 {
  89         struct bio_vec end_bv = { NULL }, nxt_bv;
  90         struct bvec_iter iter;
  91
  92         if (!blk_queue_cluster(q))
  93                 return 0;
  94
  95         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
  96             queue_max_segment_size(q))
  97                 return 0;
  98
  99         if (!bio_has_data(bio))
 100                 return 1;
 101
 102         bio_for_each_segment(end_bv, bio, iter)
 103                 if (end_bv.bv_len == iter.bi_size)
 104                         break;
 105
 106         nxt_bv = bio_iovec(nxt);
 107
 108         if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
 109                 return 0;
 110
 111         /*
 112          * bio and nxt are contiguous in memory; check if the queue allows
 113          * these two to be merged into one
 114          */
 115         if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
 116                 return 1;
 117
 118         return 0;
 119 }
 120
 121 static inline void
 122 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
 123                      struct scatterlist *sglist, struct bio_vec *bvprv,
 124                      struct scatterlist **sg, int *nsegs, int *cluster)
 125 {
 126
 127         int nbytes = bvec->bv_len;
 128
 129         if (*sg && *cluster) {
 130                 if ((*sg)->length + nbytes > queue_max_segment_size(q))
 131                         goto new_segment;
 132
 133                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 134                         goto new_segment;
 135                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 136                         goto new_segment;
 137
 138                 (*sg)->length += nbytes;
 139         } else {
 140 new_segment:
 141                 if (!*sg)
 142                         *sg = sglist;
 143                 else {
 144                         /*
 145                          * If the driver previously mapped a shorter
 146                          * list, we could see a termination bit
 147                          * prematurely unless it fully inits the sg
 148                          * table on each mapping. We KNOW that there
 149                          * must be more entries here or the driver
 150                          * would be buggy, so force clear the
 151                          * termination bit to avoid doing a full
 152                          * sg_init_table() in drivers for each command.
 153                          */
 154                         sg_unmark_end(*sg);
 155                         *sg = sg_next(*sg);
 156                 }
 157
 158                 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
 159                 (*nsegs)++;
 160         }
 161         *bvprv = *bvec;
 162 }
 163
 164 /*
 165  * map a request to scatterlist, return number of sg entries setup. Caller
 166  * must make sure sg can hold rq->nr_phys_segments entries
 167  */
 168 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 169                   struct scatterlist *sglist)
 170 {
 171         struct bio_vec bvec, bvprv = { NULL };
 172         struct req_iterator iter;
 173         struct scatterlist *sg;
 174         int nsegs, cluster;
 175
 176         nsegs = 0;
 177         cluster = blk_queue_cluster(q);
 178
 179         /*
 180          * for each bio in rq
 181          */
 182         sg = NULL;
 183         rq_for_each_segment(bvec, rq, iter) {
 184                 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, &sg,
 185                                      &nsegs, &cluster);
 186         } /* segments in rq */
 187
 188
 189         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 190             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 191                 unsigned int pad_len =
 192                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 193
 194                 sg->length += pad_len;
 195                 rq->extra_len += pad_len;
 196         }
 197
 198         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 199                 if (rq->cmd_flags & REQ_WRITE)
 200                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 201
 202                 sg->page_link &= ~0x02;
 203                 sg = sg_next(sg);
 204                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 205                             q->dma_drain_size,
 206                             ((unsigned long)q->dma_drain_buffer) &
 207                             (PAGE_SIZE - 1));
 208                 nsegs++;
 209                 rq->extra_len += q->dma_drain_size;
 210         }
 211
 212         if (sg)
 213                 sg_mark_end(sg);
 214
 215         return nsegs;
 216 }
 217 EXPORT_SYMBOL(blk_rq_map_sg);
 218
 219 /**
 220  * blk_bio_map_sg - map a bio to a scatterlist
 221  * @q: request_queue in question
 222  * @bio: bio being mapped
 223  * @sglist: scatterlist being mapped
 224  *
 225  * Note:
 226  *    Caller must make sure sg can hold bio->bi_phys_segments entries
 227  *
 228  * Will return the number of sg entries setup
 229  */
 230 int blk_bio_map_sg(struct request_queue *q, struct bio *bio,
 231                    struct scatterlist *sglist)
 232 {
 233         struct bio_vec bvec, bvprv = { NULL };
 234         struct scatterlist *sg;
 235         int nsegs, cluster;
 236         struct bvec_iter iter;
 237
 238         nsegs = 0;
 239         cluster = blk_queue_cluster(q);
 240
 241         sg = NULL;
 242         bio_for_each_segment(bvec, bio, iter) {
 243                 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, &sg,
 244                                      &nsegs, &cluster);
 245         } /* segments in bio */
 246
 247         if (sg)
 248                 sg_mark_end(sg);
 249
 250         BUG_ON(bio->bi_phys_segments && nsegs > bio->bi_phys_segments);
 251         return nsegs;
 252 }
 253 EXPORT_SYMBOL(blk_bio_map_sg);
 254
 255 static inline int ll_new_hw_segment(struct request_queue *q,
 256                                     struct request *req,
 257                                     struct bio *bio)
 258 {
 259         int nr_phys_segs = bio_phys_segments(q, bio);
 260
 261         if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 262                 goto no_merge;
 263
 264         if (bio_integrity(bio) && blk_integrity_merge_bio(q, req, bio))
 265                 goto no_merge;
 266
 267         /*
 268          * This will form the start of a new hw segment.  Bump both
 269          * counters.
 270          */
 271         req->nr_phys_segments += nr_phys_segs;
 272         return 1;
 273
 274 no_merge:
 275         req->cmd_flags |= REQ_NOMERGE;
 276         if (req == q->last_merge)
 277                 q->last_merge = NULL;
 278         return 0;
 279 }
 280
 281 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 282                      struct bio *bio)
 283 {
 284         if (blk_rq_sectors(req) + bio_sectors(bio) >
 285             blk_rq_get_max_sectors(req)) {
 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         if (blk_rq_sectors(req) + bio_sectors(bio) >
 303             blk_rq_get_max_sectors(req)) {
 304                 req->cmd_flags |= REQ_NOMERGE;
 305                 if (req == q->last_merge)
 306                         q->last_merge = NULL;
 307                 return 0;
 308         }
 309         if (!bio_flagged(bio, BIO_SEG_VALID))
 310                 blk_recount_segments(q, bio);
 311         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 312                 blk_recount_segments(q, req->bio);
 313
 314         return ll_new_hw_segment(q, req, bio);
 315 }
 316
 317 /*
 318  * blk-mq uses req->special to carry normal driver per-request payload, it
 319  * does not indicate a prepared command that we cannot merge with.
 320  */
 321 static bool req_no_special_merge(struct request *req)
 322 {
 323         struct request_queue *q = req->q;
 324
 325         return !q->mq_ops && req->special;
 326 }
 327
 328 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 329                                 struct request *next)
 330 {
 331         int total_phys_segments;
 332         unsigned int seg_size =
 333                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 334
 335         /*
 336          * First check if the either of the requests are re-queued
 337          * requests.  Can't merge them if they are.
 338          */
 339         if (req_no_special_merge(req) || req_no_special_merge(next))
 340                 return 0;
 341
 342         /*
 343          * Will it become too large?
 344          */
 345         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
 346             blk_rq_get_max_sectors(req))
 347                 return 0;
 348
 349         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 350         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 351                 if (req->nr_phys_segments == 1)
 352                         req->bio->bi_seg_front_size = seg_size;
 353                 if (next->nr_phys_segments == 1)
 354                         next->biotail->bi_seg_back_size = seg_size;
 355                 total_phys_segments--;
 356         }
 357
 358         if (total_phys_segments > queue_max_segments(q))
 359                 return 0;
 360
 361         if (blk_integrity_rq(req) && blk_integrity_merge_rq(q, req, next))
 362                 return 0;
 363
 364         /* Merge is OK... */
 365         req->nr_phys_segments = total_phys_segments;
 366         return 1;
 367 }
 368
 369 /**
 370  * blk_rq_set_mixed_merge - mark a request as mixed merge
 371  * @rq: request to mark as mixed merge
 372  *
 373  * Description:
 374  *     @rq is about to be mixed merged.  Make sure the attributes
 375  *     which can be mixed are set in each bio and mark @rq as mixed
 376  *     merged.
 377  */
 378 void blk_rq_set_mixed_merge(struct request *rq)
 379 {
 380         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 381         struct bio *bio;
 382
 383         if (rq->cmd_flags & REQ_MIXED_MERGE)
 384                 return;
 385
 386         /*
 387          * @rq will no longer represent mixable attributes for all the
 388          * contained bios.  It will just track those of the first one.
 389          * Distributes the attributs to each bio.
 390          */
 391         for (bio = rq->bio; bio; bio = bio->bi_next) {
 392                 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
 393                              (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
 394                 bio->bi_rw |= ff;
 395         }
 396         rq->cmd_flags |= REQ_MIXED_MERGE;
 397 }
 398
 399 static void blk_account_io_merge(struct request *req)
 400 {
 401         if (blk_do_io_stat(req)) {
 402                 struct hd_struct *part;
 403                 int cpu;
 404
 405                 cpu = part_stat_lock();
 406                 part = req->part;
 407
 408                 part_round_stats(cpu, part);
 409                 part_dec_in_flight(part, rq_data_dir(req));
 410
 411                 hd_struct_put(part);
 412                 part_stat_unlock();
 413         }
 414 }
 415
 416 /*
 417  * Has to be called with the request spinlock acquired
 418  */
 419 static int attempt_merge(struct request_queue *q, struct request *req,
 420                           struct request *next)
 421 {
 422         if (!rq_mergeable(req) || !rq_mergeable(next))
 423                 return 0;
 424
 425         if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
 426                 return 0;
 427
 428         /*
 429          * not contiguous
 430          */
 431         if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
 432                 return 0;
 433
 434         if (rq_data_dir(req) != rq_data_dir(next)
 435             || req->rq_disk != next->rq_disk
 436             || req_no_special_merge(next))
 437                 return 0;
 438
 439         if (req->cmd_flags & REQ_WRITE_SAME &&
 440             !blk_write_same_mergeable(req->bio, next->bio))
 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) || !bio_mergeable(bio))
 525                 return false;
 526
 527         if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
 528                 return false;
 529
 530         /* different data direction or already started, don't merge */
 531         if (bio_data_dir(bio) != rq_data_dir(rq))
 532                 return false;
 533
 534         /* must be same device and not a special request */
 535         if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
 536                 return false;
 537
 538         /* only merge integrity protected bio into ditto rq */
 539         if (bio_integrity(bio) != blk_integrity_rq(rq))
 540                 return false;
 541
 542         /* must be using the same buffer */
 543         if (rq->cmd_flags & REQ_WRITE_SAME &&
 544             !blk_write_same_mergeable(rq->bio, bio))
 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_iter.bi_sector)
 553                 return ELEVATOR_BACK_MERGE;
 554         else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
 555                 return ELEVATOR_FRONT_MERGE;
 556         return ELEVATOR_NO_MERGE;
 557 }