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 void blk_recalc_rq_sectors(struct request *rq, int nsect)
  13 {
  14         if (blk_fs_request(rq)) {
  15                 rq->hard_sector += nsect;
  16                 rq->hard_nr_sectors -= nsect;
  17
  18                 /*
  19                  * Move the I/O submission pointers ahead if required.
  20                  */
  21                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
  22                     (rq->sector <= rq->hard_sector)) {
  23                         rq->sector = rq->hard_sector;
  24                         rq->nr_sectors = rq->hard_nr_sectors;
  25                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
  26                         rq->current_nr_sectors = rq->hard_cur_sectors;
  27                         rq->buffer = bio_data(rq->bio);
  28                 }
  29
  30                 /*
  31                  * if total number of sectors is less than the first segment
  32                  * size, something has gone terribly wrong
  33                  */
  34                 if (rq->nr_sectors < rq->current_nr_sectors) {
  35                         printk(KERN_ERR "blk: request botched\n");
  36                         rq->nr_sectors = rq->current_nr_sectors;
  37                 }
  38         }
  39 }
  40
  41 void blk_recalc_rq_segments(struct request *rq)
  42 {
  43         int nr_phys_segs;
  44         int nr_hw_segs;
  45         unsigned int phys_size;
  46         unsigned int hw_size;
  47         struct bio_vec *bv, *bvprv = NULL;
  48         int seg_size;
  49         int hw_seg_size;
  50         int cluster;
  51         struct req_iterator iter;
  52         int high, highprv = 1;
  53         struct request_queue *q = rq->q;
  54
  55         if (!rq->bio)
  56                 return;
  57
  58         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
  59         hw_seg_size = seg_size = 0;
  60         phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
  61         rq_for_each_segment(bv, rq, iter) {
  62                 /*
  63                  * the trick here is making sure that a high page is never
  64                  * considered part of another segment, since that might
  65                  * change with the bounce page.
  66                  */
  67                 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
  68                 if (high || highprv)
  69                         goto new_hw_segment;
  70                 if (cluster) {
  71                         if (seg_size + bv->bv_len > q->max_segment_size)
  72                                 goto new_segment;
  73                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
  74                                 goto new_segment;
  75                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
  76                                 goto new_segment;
  77                         if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
  78                                 goto new_hw_segment;
  79
  80                         seg_size += bv->bv_len;
  81                         hw_seg_size += bv->bv_len;
  82                         bvprv = bv;
  83                         continue;
  84                 }
  85 new_segment:
  86                 if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
  87                     !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
  88                         hw_seg_size += bv->bv_len;
  89                 else {
  90 new_hw_segment:
  91                         if (nr_hw_segs == 1 &&
  92                             hw_seg_size > rq->bio->bi_hw_front_size)
  93                                 rq->bio->bi_hw_front_size = hw_seg_size;
  94                         hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
  95                         nr_hw_segs++;
  96                 }
  97
  98                 nr_phys_segs++;
  99                 bvprv = bv;
 100                 seg_size = bv->bv_len;
 101                 highprv = high;
 102         }
 103
 104         if (nr_hw_segs == 1 &&
 105             hw_seg_size > rq->bio->bi_hw_front_size)
 106                 rq->bio->bi_hw_front_size = hw_seg_size;
 107         if (hw_seg_size > rq->biotail->bi_hw_back_size)
 108                 rq->biotail->bi_hw_back_size = hw_seg_size;
 109         rq->nr_phys_segments = nr_phys_segs;
 110         rq->nr_hw_segments = nr_hw_segs;
 111 }
 112
 113 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 114 {
 115         struct request rq;
 116         struct bio *nxt = bio->bi_next;
 117         rq.q = q;
 118         rq.bio = rq.biotail = bio;
 119         bio->bi_next = NULL;
 120         blk_recalc_rq_segments(&rq);
 121         bio->bi_next = nxt;
 122         bio->bi_phys_segments = rq.nr_phys_segments;
 123         bio->bi_hw_segments = rq.nr_hw_segments;
 124         bio->bi_flags |= (1 << BIO_SEG_VALID);
 125 }
 126 EXPORT_SYMBOL(blk_recount_segments);
 127
 128 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 129                                    struct bio *nxt)
 130 {
 131         if (!test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags))
 132                 return 0;
 133
 134         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
 135                 return 0;
 136         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
 137                 return 0;
 138
 139         /*
 140          * bio and nxt are contigous in memory, check if the queue allows
 141          * these two to be merged into one
 142          */
 143         if (BIO_SEG_BOUNDARY(q, bio, nxt))
 144                 return 1;
 145
 146         return 0;
 147 }
 148
 149 static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
 150                                  struct bio *nxt)
 151 {
 152         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
 153                 blk_recount_segments(q, bio);
 154         if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
 155                 blk_recount_segments(q, nxt);
 156         if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
 157             BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size))
 158                 return 0;
 159         if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
 160                 return 0;
 161
 162         return 1;
 163 }
 164
 165 /*
 166  * map a request to scatterlist, return number of sg entries setup. Caller
 167  * must make sure sg can hold rq->nr_phys_segments entries
 168  */
 169 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 170                   struct scatterlist *sglist)
 171 {
 172         struct bio_vec *bvec, *bvprv;
 173         struct req_iterator iter;
 174         struct scatterlist *sg;
 175         int nsegs, cluster;
 176
 177         nsegs = 0;
 178         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
 179
 180         /*
 181          * for each bio in rq
 182          */
 183         bvprv = NULL;
 184         sg = NULL;
 185         rq_for_each_segment(bvec, rq, iter) {
 186                 int nbytes = bvec->bv_len;
 187
 188                 if (bvprv && cluster) {
 189                         if (sg->length + nbytes > q->max_segment_size)
 190                                 goto new_segment;
 191
 192                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 193                                 goto new_segment;
 194                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 195                                 goto new_segment;
 196
 197                         sg->length += nbytes;
 198                 } else {
 199 new_segment:
 200                         if (!sg)
 201                                 sg = sglist;
 202                         else {
 203                                 /*
 204                                  * If the driver previously mapped a shorter
 205                                  * list, we could see a termination bit
 206                                  * prematurely unless it fully inits the sg
 207                                  * table on each mapping. We KNOW that there
 208                                  * must be more entries here or the driver
 209                                  * would be buggy, so force clear the
 210                                  * termination bit to avoid doing a full
 211                                  * sg_init_table() in drivers for each command.
 212                                  */
 213                                 sg->page_link &= ~0x02;
 214                                 sg = sg_next(sg);
 215                         }
 216
 217                         sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
 218                         nsegs++;
 219                 }
 220                 bvprv = bvec;
 221         } /* segments in rq */
 222
 223
 224         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 225             (rq->data_len & q->dma_pad_mask)) {
 226                 unsigned int pad_len = (q->dma_pad_mask & ~rq->data_len) + 1;
 227
 228                 sg->length += pad_len;
 229                 rq->extra_len += pad_len;
 230         }
 231
 232         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 233                 if (rq->cmd_flags & REQ_RW)
 234                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 235
 236                 sg->page_link &= ~0x02;
 237                 sg = sg_next(sg);
 238                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 239                             q->dma_drain_size,
 240                             ((unsigned long)q->dma_drain_buffer) &
 241                             (PAGE_SIZE - 1));
 242                 nsegs++;
 243                 rq->extra_len += q->dma_drain_size;
 244         }
 245
 246         if (sg)
 247                 sg_mark_end(sg);
 248
 249         return nsegs;
 250 }
 251 EXPORT_SYMBOL(blk_rq_map_sg);
 252
 253 static inline int ll_new_mergeable(struct request_queue *q,
 254                                    struct request *req,
 255                                    struct bio *bio)
 256 {
 257         int nr_phys_segs = bio_phys_segments(q, bio);
 258
 259         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 260                 req->cmd_flags |= REQ_NOMERGE;
 261                 if (req == q->last_merge)
 262                         q->last_merge = NULL;
 263                 return 0;
 264         }
 265
 266         /*
 267          * A hw segment is just getting larger, bump just the phys
 268          * counter.
 269          */
 270         req->nr_phys_segments += nr_phys_segs;
 271         return 1;
 272 }
 273
 274 static inline int ll_new_hw_segment(struct request_queue *q,
 275                                     struct request *req,
 276                                     struct bio *bio)
 277 {
 278         int nr_hw_segs = bio_hw_segments(q, bio);
 279         int nr_phys_segs = bio_phys_segments(q, bio);
 280
 281         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
 282             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 283                 req->cmd_flags |= REQ_NOMERGE;
 284                 if (req == q->last_merge)
 285                         q->last_merge = NULL;
 286                 return 0;
 287         }
 288
 289         /*
 290          * This will form the start of a new hw segment.  Bump both
 291          * counters.
 292          */
 293         req->nr_hw_segments += nr_hw_segs;
 294         req->nr_phys_segments += nr_phys_segs;
 295         return 1;
 296 }
 297
 298 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 299                      struct bio *bio)
 300 {
 301         unsigned short max_sectors;
 302         int len;
 303
 304         if (unlikely(blk_pc_request(req)))
 305                 max_sectors = q->max_hw_sectors;
 306         else
 307                 max_sectors = q->max_sectors;
 308
 309         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 310                 req->cmd_flags |= REQ_NOMERGE;
 311                 if (req == q->last_merge)
 312                         q->last_merge = NULL;
 313                 return 0;
 314         }
 315         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
 316                 blk_recount_segments(q, req->biotail);
 317         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
 318                 blk_recount_segments(q, bio);
 319         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
 320         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio))
 321             && !BIOVEC_VIRT_OVERSIZE(len)) {
 322                 int mergeable =  ll_new_mergeable(q, req, bio);
 323
 324                 if (mergeable) {
 325                         if (req->nr_hw_segments == 1)
 326                                 req->bio->bi_hw_front_size = len;
 327                         if (bio->bi_hw_segments == 1)
 328                                 bio->bi_hw_back_size = len;
 329                 }
 330                 return mergeable;
 331         }
 332
 333         return ll_new_hw_segment(q, req, bio);
 334 }
 335
 336 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 337                       struct bio *bio)
 338 {
 339         unsigned short max_sectors;
 340         int len;
 341
 342         if (unlikely(blk_pc_request(req)))
 343                 max_sectors = q->max_hw_sectors;
 344         else
 345                 max_sectors = q->max_sectors;
 346
 347
 348         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 349                 req->cmd_flags |= REQ_NOMERGE;
 350                 if (req == q->last_merge)
 351                         q->last_merge = NULL;
 352                 return 0;
 353         }
 354         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
 355         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
 356                 blk_recount_segments(q, bio);
 357         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
 358                 blk_recount_segments(q, req->bio);
 359         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
 360             !BIOVEC_VIRT_OVERSIZE(len)) {
 361                 int mergeable =  ll_new_mergeable(q, req, bio);
 362
 363                 if (mergeable) {
 364                         if (bio->bi_hw_segments == 1)
 365                                 bio->bi_hw_front_size = len;
 366                         if (req->nr_hw_segments == 1)
 367                                 req->biotail->bi_hw_back_size = len;
 368                 }
 369                 return mergeable;
 370         }
 371
 372         return ll_new_hw_segment(q, req, bio);
 373 }
 374
 375 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 376                                 struct request *next)
 377 {
 378         int total_phys_segments;
 379         int total_hw_segments;
 380
 381         /*
 382          * First check if the either of the requests are re-queued
 383          * requests.  Can't merge them if they are.
 384          */
 385         if (req->special || next->special)
 386                 return 0;
 387
 388         /*
 389          * Will it become too large?
 390          */
 391         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
 392                 return 0;
 393
 394         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 395         if (blk_phys_contig_segment(q, req->biotail, next->bio))
 396                 total_phys_segments--;
 397
 398         if (total_phys_segments > q->max_phys_segments)
 399                 return 0;
 400
 401         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
 402         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
 403                 int len = req->biotail->bi_hw_back_size +
 404                                 next->bio->bi_hw_front_size;
 405                 /*
 406                  * propagate the combined length to the end of the requests
 407                  */
 408                 if (req->nr_hw_segments == 1)
 409                         req->bio->bi_hw_front_size = len;
 410                 if (next->nr_hw_segments == 1)
 411                         next->biotail->bi_hw_back_size = len;
 412                 total_hw_segments--;
 413         }
 414
 415         if (total_hw_segments > q->max_hw_segments)
 416                 return 0;
 417
 418         /* Merge is OK... */
 419         req->nr_phys_segments = total_phys_segments;
 420         req->nr_hw_segments = total_hw_segments;
 421         return 1;
 422 }
 423
 424 /*
 425  * Has to be called with the request spinlock acquired
 426  */
 427 static int attempt_merge(struct request_queue *q, struct request *req,
 428                           struct request *next)
 429 {
 430         if (!rq_mergeable(req) || !rq_mergeable(next))
 431                 return 0;
 432
 433         /*
 434          * not contiguous
 435          */
 436         if (req->sector + req->nr_sectors != next->sector)
 437                 return 0;
 438
 439         if (rq_data_dir(req) != rq_data_dir(next)
 440             || req->rq_disk != next->rq_disk
 441             || next->special)
 442                 return 0;
 443
 444         /*
 445          * If we are allowed to merge, then append bio list
 446          * from next to rq and release next. merge_requests_fn
 447          * will have updated segment counts, update sector
 448          * counts here.
 449          */
 450         if (!ll_merge_requests_fn(q, req, next))
 451                 return 0;
 452
 453         /*
 454          * At this point we have either done a back merge
 455          * or front merge. We need the smaller start_time of
 456          * the merged requests to be the current request
 457          * for accounting purposes.
 458          */
 459         if (time_after(req->start_time, next->start_time))
 460                 req->start_time = next->start_time;
 461
 462         req->biotail->bi_next = next->bio;
 463         req->biotail = next->biotail;
 464
 465         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
 466
 467         elv_merge_requests(q, req, next);
 468
 469         if (req->rq_disk) {
 470                 struct hd_struct *part
 471                         = get_part(req->rq_disk, req->sector);
 472                 disk_round_stats(req->rq_disk);
 473                 req->rq_disk->in_flight--;
 474                 if (part) {
 475                         part_round_stats(part);
 476                         part->in_flight--;
 477                 }
 478         }
 479
 480         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 481
 482         __blk_put_request(q, next);
 483         return 1;
 484 }
 485
 486 int attempt_back_merge(struct request_queue *q, struct request *rq)
 487 {
 488         struct request *next = elv_latter_request(q, rq);
 489
 490         if (next)
 491                 return attempt_merge(q, rq, next);
 492
 493         return 0;
 494 }
 495
 496 int attempt_front_merge(struct request_queue *q, struct request *rq)
 497 {
 498         struct request *prev = elv_former_request(q, rq);
 499
 500         if (prev)
 501                 return attempt_merge(q, prev, rq);
 502
 503         return 0;
 504 }