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 = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
  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 (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
 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 = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
 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         if (q->dma_drain_size) {
 224                 sg->page_link &= ~0x02;
 225                 sg = sg_next(sg);
 226                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 227                             q->dma_drain_size,
 228                             ((unsigned long)q->dma_drain_buffer) &
 229                             (PAGE_SIZE - 1));
 230                 nsegs++;
 231         }
 232
 233         if (sg)
 234                 sg_mark_end(sg);
 235
 236         return nsegs;
 237 }
 238 EXPORT_SYMBOL(blk_rq_map_sg);
 239
 240 static inline int ll_new_mergeable(struct request_queue *q,
 241                                    struct request *req,
 242                                    struct bio *bio)
 243 {
 244         int nr_phys_segs = bio_phys_segments(q, bio);
 245
 246         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 247                 req->cmd_flags |= REQ_NOMERGE;
 248                 if (req == q->last_merge)
 249                         q->last_merge = NULL;
 250                 return 0;
 251         }
 252
 253         /*
 254          * A hw segment is just getting larger, bump just the phys
 255          * counter.
 256          */
 257         req->nr_phys_segments += nr_phys_segs;
 258         return 1;
 259 }
 260
 261 static inline int ll_new_hw_segment(struct request_queue *q,
 262                                     struct request *req,
 263                                     struct bio *bio)
 264 {
 265         int nr_hw_segs = bio_hw_segments(q, bio);
 266         int nr_phys_segs = bio_phys_segments(q, bio);
 267
 268         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
 269             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 270                 req->cmd_flags |= REQ_NOMERGE;
 271                 if (req == q->last_merge)
 272                         q->last_merge = NULL;
 273                 return 0;
 274         }
 275
 276         /*
 277          * This will form the start of a new hw segment.  Bump both
 278          * counters.
 279          */
 280         req->nr_hw_segments += nr_hw_segs;
 281         req->nr_phys_segments += nr_phys_segs;
 282         return 1;
 283 }
 284
 285 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 286                      struct bio *bio)
 287 {
 288         unsigned short max_sectors;
 289         int len;
 290
 291         if (unlikely(blk_pc_request(req)))
 292                 max_sectors = q->max_hw_sectors;
 293         else
 294                 max_sectors = q->max_sectors;
 295
 296         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 297                 req->cmd_flags |= REQ_NOMERGE;
 298                 if (req == q->last_merge)
 299                         q->last_merge = NULL;
 300                 return 0;
 301         }
 302         if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
 303                 blk_recount_segments(q, req->biotail);
 304         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
 305                 blk_recount_segments(q, bio);
 306         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
 307         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio))
 308             && !BIOVEC_VIRT_OVERSIZE(len)) {
 309                 int mergeable =  ll_new_mergeable(q, req, bio);
 310
 311                 if (mergeable) {
 312                         if (req->nr_hw_segments == 1)
 313                                 req->bio->bi_hw_front_size = len;
 314                         if (bio->bi_hw_segments == 1)
 315                                 bio->bi_hw_back_size = len;
 316                 }
 317                 return mergeable;
 318         }
 319
 320         return ll_new_hw_segment(q, req, bio);
 321 }
 322
 323 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 324                       struct bio *bio)
 325 {
 326         unsigned short max_sectors;
 327         int len;
 328
 329         if (unlikely(blk_pc_request(req)))
 330                 max_sectors = q->max_hw_sectors;
 331         else
 332                 max_sectors = q->max_sectors;
 333
 334
 335         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 336                 req->cmd_flags |= REQ_NOMERGE;
 337                 if (req == q->last_merge)
 338                         q->last_merge = NULL;
 339                 return 0;
 340         }
 341         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
 342         if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
 343                 blk_recount_segments(q, bio);
 344         if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
 345                 blk_recount_segments(q, req->bio);
 346         if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
 347             !BIOVEC_VIRT_OVERSIZE(len)) {
 348                 int mergeable =  ll_new_mergeable(q, req, bio);
 349
 350                 if (mergeable) {
 351                         if (bio->bi_hw_segments == 1)
 352                                 bio->bi_hw_front_size = len;
 353                         if (req->nr_hw_segments == 1)
 354                                 req->biotail->bi_hw_back_size = len;
 355                 }
 356                 return mergeable;
 357         }
 358
 359         return ll_new_hw_segment(q, req, bio);
 360 }
 361
 362 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 363                                 struct request *next)
 364 {
 365         int total_phys_segments;
 366         int total_hw_segments;
 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->special || next->special)
 373                 return 0;
 374
 375         /*
 376          * Will it become too large?
 377          */
 378         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
 379                 return 0;
 380
 381         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 382         if (blk_phys_contig_segment(q, req->biotail, next->bio))
 383                 total_phys_segments--;
 384
 385         if (total_phys_segments > q->max_phys_segments)
 386                 return 0;
 387
 388         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
 389         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
 390                 int len = req->biotail->bi_hw_back_size +
 391                                 next->bio->bi_hw_front_size;
 392                 /*
 393                  * propagate the combined length to the end of the requests
 394                  */
 395                 if (req->nr_hw_segments == 1)
 396                         req->bio->bi_hw_front_size = len;
 397                 if (next->nr_hw_segments == 1)
 398                         next->biotail->bi_hw_back_size = len;
 399                 total_hw_segments--;
 400         }
 401
 402         if (total_hw_segments > q->max_hw_segments)
 403                 return 0;
 404
 405         /* Merge is OK... */
 406         req->nr_phys_segments = total_phys_segments;
 407         req->nr_hw_segments = total_hw_segments;
 408         return 1;
 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          * not contiguous
 422          */
 423         if (req->sector + req->nr_sectors != next->sector)
 424                 return 0;
 425
 426         if (rq_data_dir(req) != rq_data_dir(next)
 427             || req->rq_disk != next->rq_disk
 428             || next->special)
 429                 return 0;
 430
 431         /*
 432          * If we are allowed to merge, then append bio list
 433          * from next to rq and release next. merge_requests_fn
 434          * will have updated segment counts, update sector
 435          * counts here.
 436          */
 437         if (!ll_merge_requests_fn(q, req, next))
 438                 return 0;
 439
 440         /*
 441          * At this point we have either done a back merge
 442          * or front merge. We need the smaller start_time of
 443          * the merged requests to be the current request
 444          * for accounting purposes.
 445          */
 446         if (time_after(req->start_time, next->start_time))
 447                 req->start_time = next->start_time;
 448
 449         req->biotail->bi_next = next->bio;
 450         req->biotail = next->biotail;
 451
 452         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
 453
 454         elv_merge_requests(q, req, next);
 455
 456         if (req->rq_disk) {
 457                 struct hd_struct *part
 458                         = get_part(req->rq_disk, req->sector);
 459                 disk_round_stats(req->rq_disk);
 460                 req->rq_disk->in_flight--;
 461                 if (part) {
 462                         part_round_stats(part);
 463                         part->in_flight--;
 464                 }
 465         }
 466
 467         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 468
 469         __blk_put_request(q, next);
 470         return 1;
 471 }
 472
 473 int attempt_back_merge(struct request_queue *q, struct request *rq)
 474 {
 475         struct request *next = elv_latter_request(q, rq);
 476
 477         if (next)
 478                 return attempt_merge(q, rq, next);
 479
 480         return 0;
 481 }
 482
 483 int attempt_front_merge(struct request_queue *q, struct request *rq)
 484 {
 485         struct request *prev = elv_former_request(q, rq);
 486
 487         if (prev)
 488                 return attempt_merge(q, prev, rq);
 489
 490         return 0;
 491 }