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) || blk_discard_rq(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 (bio->bi_size + nxt->bi_size > q->max_segment_size)
 135                 return 0;
 136
 137         if (!bio_has_data(bio))
 138                 return 1;
 139
 140         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
 141                 return 0;
 142
 143         /*
 144          * bio and nxt are contiguous in memory; check if the queue allows
 145          * these two to be merged into one
 146          */
 147         if (BIO_SEG_BOUNDARY(q, bio, nxt))
 148                 return 1;
 149
 150         return 0;
 151 }
 152
 153 static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
 154                                  struct bio *nxt)
 155 {
 156         if (!bio_flagged(bio, BIO_SEG_VALID))
 157                 blk_recount_segments(q, bio);
 158         if (!bio_flagged(nxt, BIO_SEG_VALID))
 159                 blk_recount_segments(q, nxt);
 160         if (bio_has_data(bio) &&
 161             (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
 162              BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size)))
 163                 return 0;
 164         if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
 165                 return 0;
 166
 167         return 1;
 168 }
 169
 170 /*
 171  * map a request to scatterlist, return number of sg entries setup. Caller
 172  * must make sure sg can hold rq->nr_phys_segments entries
 173  */
 174 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 175                   struct scatterlist *sglist)
 176 {
 177         struct bio_vec *bvec, *bvprv;
 178         struct req_iterator iter;
 179         struct scatterlist *sg;
 180         int nsegs, cluster;
 181
 182         nsegs = 0;
 183         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
 184
 185         /*
 186          * for each bio in rq
 187          */
 188         bvprv = NULL;
 189         sg = NULL;
 190         rq_for_each_segment(bvec, rq, iter) {
 191                 int nbytes = bvec->bv_len;
 192
 193                 if (bvprv && cluster) {
 194                         if (sg->length + nbytes > q->max_segment_size)
 195                                 goto new_segment;
 196
 197                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 198                                 goto new_segment;
 199                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 200                                 goto new_segment;
 201
 202                         sg->length += nbytes;
 203                 } else {
 204 new_segment:
 205                         if (!sg)
 206                                 sg = sglist;
 207                         else {
 208                                 /*
 209                                  * If the driver previously mapped a shorter
 210                                  * list, we could see a termination bit
 211                                  * prematurely unless it fully inits the sg
 212                                  * table on each mapping. We KNOW that there
 213                                  * must be more entries here or the driver
 214                                  * would be buggy, so force clear the
 215                                  * termination bit to avoid doing a full
 216                                  * sg_init_table() in drivers for each command.
 217                                  */
 218                                 sg->page_link &= ~0x02;
 219                                 sg = sg_next(sg);
 220                         }
 221
 222                         sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
 223                         nsegs++;
 224                 }
 225                 bvprv = bvec;
 226         } /* segments in rq */
 227
 228
 229         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 230             (rq->data_len & q->dma_pad_mask)) {
 231                 unsigned int pad_len = (q->dma_pad_mask & ~rq->data_len) + 1;
 232
 233                 sg->length += pad_len;
 234                 rq->extra_len += pad_len;
 235         }
 236
 237         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 238                 if (rq->cmd_flags & REQ_RW)
 239                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 240
 241                 sg->page_link &= ~0x02;
 242                 sg = sg_next(sg);
 243                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 244                             q->dma_drain_size,
 245                             ((unsigned long)q->dma_drain_buffer) &
 246                             (PAGE_SIZE - 1));
 247                 nsegs++;
 248                 rq->extra_len += q->dma_drain_size;
 249         }
 250
 251         if (sg)
 252                 sg_mark_end(sg);
 253
 254         return nsegs;
 255 }
 256 EXPORT_SYMBOL(blk_rq_map_sg);
 257
 258 static inline int ll_new_mergeable(struct request_queue *q,
 259                                    struct request *req,
 260                                    struct bio *bio)
 261 {
 262         int nr_phys_segs = bio_phys_segments(q, bio);
 263
 264         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 265                 req->cmd_flags |= REQ_NOMERGE;
 266                 if (req == q->last_merge)
 267                         q->last_merge = NULL;
 268                 return 0;
 269         }
 270
 271         /*
 272          * A hw segment is just getting larger, bump just the phys
 273          * counter.
 274          */
 275         req->nr_phys_segments += nr_phys_segs;
 276         return 1;
 277 }
 278
 279 static inline int ll_new_hw_segment(struct request_queue *q,
 280                                     struct request *req,
 281                                     struct bio *bio)
 282 {
 283         int nr_hw_segs = bio_hw_segments(q, bio);
 284         int nr_phys_segs = bio_phys_segments(q, bio);
 285
 286         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
 287             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 288                 req->cmd_flags |= REQ_NOMERGE;
 289                 if (req == q->last_merge)
 290                         q->last_merge = NULL;
 291                 return 0;
 292         }
 293
 294         /*
 295          * This will form the start of a new hw segment.  Bump both
 296          * counters.
 297          */
 298         req->nr_hw_segments += nr_hw_segs;
 299         req->nr_phys_segments += nr_phys_segs;
 300         return 1;
 301 }
 302
 303 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 304                      struct bio *bio)
 305 {
 306         unsigned short max_sectors;
 307         int len;
 308
 309         if (unlikely(blk_pc_request(req)))
 310                 max_sectors = q->max_hw_sectors;
 311         else
 312                 max_sectors = q->max_sectors;
 313
 314         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 315                 req->cmd_flags |= REQ_NOMERGE;
 316                 if (req == q->last_merge)
 317                         q->last_merge = NULL;
 318                 return 0;
 319         }
 320         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 321                 blk_recount_segments(q, req->biotail);
 322         if (!bio_flagged(bio, BIO_SEG_VALID))
 323                 blk_recount_segments(q, bio);
 324         len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
 325         if (!bio_has_data(bio) ||
 326             (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio))
 327              && !BIOVEC_VIRT_OVERSIZE(len))) {
 328                 int mergeable =  ll_new_mergeable(q, req, bio);
 329
 330                 if (mergeable) {
 331                         if (req->nr_hw_segments == 1)
 332                                 req->bio->bi_hw_front_size = len;
 333                         if (bio->bi_hw_segments == 1)
 334                                 bio->bi_hw_back_size = len;
 335                 }
 336                 return mergeable;
 337         }
 338
 339         return ll_new_hw_segment(q, req, bio);
 340 }
 341
 342 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 343                       struct bio *bio)
 344 {
 345         unsigned short max_sectors;
 346         int len;
 347
 348         if (unlikely(blk_pc_request(req)))
 349                 max_sectors = q->max_hw_sectors;
 350         else
 351                 max_sectors = q->max_sectors;
 352
 353
 354         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 355                 req->cmd_flags |= REQ_NOMERGE;
 356                 if (req == q->last_merge)
 357                         q->last_merge = NULL;
 358                 return 0;
 359         }
 360         len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
 361         if (!bio_flagged(bio, BIO_SEG_VALID))
 362                 blk_recount_segments(q, bio);
 363         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 364                 blk_recount_segments(q, req->bio);
 365         if (!bio_has_data(bio) ||
 366             (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
 367              !BIOVEC_VIRT_OVERSIZE(len))) {
 368                 int mergeable =  ll_new_mergeable(q, req, bio);
 369
 370                 if (mergeable) {
 371                         if (bio->bi_hw_segments == 1)
 372                                 bio->bi_hw_front_size = len;
 373                         if (req->nr_hw_segments == 1)
 374                                 req->biotail->bi_hw_back_size = len;
 375                 }
 376                 return mergeable;
 377         }
 378
 379         return ll_new_hw_segment(q, req, bio);
 380 }
 381
 382 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 383                                 struct request *next)
 384 {
 385         int total_phys_segments;
 386         int total_hw_segments;
 387
 388         /*
 389          * First check if the either of the requests are re-queued
 390          * requests.  Can't merge them if they are.
 391          */
 392         if (req->special || next->special)
 393                 return 0;
 394
 395         /*
 396          * Will it become too large?
 397          */
 398         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
 399                 return 0;
 400
 401         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 402         if (blk_phys_contig_segment(q, req->biotail, next->bio))
 403                 total_phys_segments--;
 404
 405         if (total_phys_segments > q->max_phys_segments)
 406                 return 0;
 407
 408         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
 409         if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
 410                 int len = req->biotail->bi_hw_back_size +
 411                                 next->bio->bi_hw_front_size;
 412                 /*
 413                  * propagate the combined length to the end of the requests
 414                  */
 415                 if (req->nr_hw_segments == 1)
 416                         req->bio->bi_hw_front_size = len;
 417                 if (next->nr_hw_segments == 1)
 418                         next->biotail->bi_hw_back_size = len;
 419                 total_hw_segments--;
 420         }
 421
 422         if (total_hw_segments > q->max_hw_segments)
 423                 return 0;
 424
 425         /* Merge is OK... */
 426         req->nr_phys_segments = total_phys_segments;
 427         req->nr_hw_segments = total_hw_segments;
 428         return 1;
 429 }
 430
 431 /*
 432  * Has to be called with the request spinlock acquired
 433  */
 434 static int attempt_merge(struct request_queue *q, struct request *req,
 435                           struct request *next)
 436 {
 437         if (!rq_mergeable(req) || !rq_mergeable(next))
 438                 return 0;
 439
 440         /*
 441          * not contiguous
 442          */
 443         if (req->sector + req->nr_sectors != next->sector)
 444                 return 0;
 445
 446         if (rq_data_dir(req) != rq_data_dir(next)
 447             || req->rq_disk != next->rq_disk
 448             || next->special)
 449                 return 0;
 450
 451         if (blk_integrity_rq(req) != blk_integrity_rq(next))
 452                 return 0;
 453
 454         /*
 455          * If we are allowed to merge, then append bio list
 456          * from next to rq and release next. merge_requests_fn
 457          * will have updated segment counts, update sector
 458          * counts here.
 459          */
 460         if (!ll_merge_requests_fn(q, req, next))
 461                 return 0;
 462
 463         /*
 464          * At this point we have either done a back merge
 465          * or front merge. We need the smaller start_time of
 466          * the merged requests to be the current request
 467          * for accounting purposes.
 468          */
 469         if (time_after(req->start_time, next->start_time))
 470                 req->start_time = next->start_time;
 471
 472         req->biotail->bi_next = next->bio;
 473         req->biotail = next->biotail;
 474
 475         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
 476
 477         elv_merge_requests(q, req, next);
 478
 479         if (req->rq_disk) {
 480                 struct hd_struct *part
 481                         = get_part(req->rq_disk, req->sector);
 482                 disk_round_stats(req->rq_disk);
 483                 req->rq_disk->in_flight--;
 484                 if (part) {
 485                         part_round_stats(part);
 486                         part->in_flight--;
 487                 }
 488         }
 489
 490         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 491
 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 }