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_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
  78                         seg_size += bv->bv_len;
  79                         hw_seg_size += bv->bv_len;
  80                         bvprv = bv;
  81                         continue;
  82                 }
  83 new_segment:
  84                 if (nr_hw_segs == 1 &&
  85                     hw_seg_size > rq->bio->bi_hw_front_size)
  86                         rq->bio->bi_hw_front_size = hw_seg_size;
  87                 hw_seg_size = bv->bv_len;
  88                 nr_hw_segs++;
  89
  90                 nr_phys_segs++;
  91                 bvprv = bv;
  92                 seg_size = bv->bv_len;
  93                 highprv = high;
  94         }
  95
  96         if (nr_hw_segs == 1 &&
  97             hw_seg_size > rq->bio->bi_hw_front_size)
  98                 rq->bio->bi_hw_front_size = hw_seg_size;
  99         if (hw_seg_size > rq->biotail->bi_hw_back_size)
 100                 rq->biotail->bi_hw_back_size = hw_seg_size;
 101         rq->nr_phys_segments = nr_phys_segs;
 102         rq->nr_hw_segments = nr_hw_segs;
 103 }
 104
 105 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 106 {
 107         struct request rq;
 108         struct bio *nxt = bio->bi_next;
 109         rq.q = q;
 110         rq.bio = rq.biotail = bio;
 111         bio->bi_next = NULL;
 112         blk_recalc_rq_segments(&rq);
 113         bio->bi_next = nxt;
 114         bio->bi_phys_segments = rq.nr_phys_segments;
 115         bio->bi_hw_segments = rq.nr_hw_segments;
 116         bio->bi_flags |= (1 << BIO_SEG_VALID);
 117 }
 118 EXPORT_SYMBOL(blk_recount_segments);
 119
 120 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 121                                    struct bio *nxt)
 122 {
 123         if (!test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags))
 124                 return 0;
 125
 126         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
 127                 return 0;
 128
 129         if (!bio_has_data(bio))
 130                 return 1;
 131
 132         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
 133                 return 0;
 134
 135         /*
 136          * bio and nxt are contiguous in memory; check if the queue allows
 137          * these two to be merged into one
 138          */
 139         if (BIO_SEG_BOUNDARY(q, bio, nxt))
 140                 return 1;
 141
 142         return 0;
 143 }
 144
 145 /*
 146  * map a request to scatterlist, return number of sg entries setup. Caller
 147  * must make sure sg can hold rq->nr_phys_segments entries
 148  */
 149 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 150                   struct scatterlist *sglist)
 151 {
 152         struct bio_vec *bvec, *bvprv;
 153         struct req_iterator iter;
 154         struct scatterlist *sg;
 155         int nsegs, cluster;
 156
 157         nsegs = 0;
 158         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
 159
 160         /*
 161          * for each bio in rq
 162          */
 163         bvprv = NULL;
 164         sg = NULL;
 165         rq_for_each_segment(bvec, rq, iter) {
 166                 int nbytes = bvec->bv_len;
 167
 168                 if (bvprv && cluster) {
 169                         if (sg->length + nbytes > q->max_segment_size)
 170                                 goto new_segment;
 171
 172                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 173                                 goto new_segment;
 174                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 175                                 goto new_segment;
 176
 177                         sg->length += nbytes;
 178                 } else {
 179 new_segment:
 180                         if (!sg)
 181                                 sg = sglist;
 182                         else {
 183                                 /*
 184                                  * If the driver previously mapped a shorter
 185                                  * list, we could see a termination bit
 186                                  * prematurely unless it fully inits the sg
 187                                  * table on each mapping. We KNOW that there
 188                                  * must be more entries here or the driver
 189                                  * would be buggy, so force clear the
 190                                  * termination bit to avoid doing a full
 191                                  * sg_init_table() in drivers for each command.
 192                                  */
 193                                 sg->page_link &= ~0x02;
 194                                 sg = sg_next(sg);
 195                         }
 196
 197                         sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
 198                         nsegs++;
 199                 }
 200                 bvprv = bvec;
 201         } /* segments in rq */
 202
 203
 204         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 205             (rq->data_len & q->dma_pad_mask)) {
 206                 unsigned int pad_len = (q->dma_pad_mask & ~rq->data_len) + 1;
 207
 208                 sg->length += pad_len;
 209                 rq->extra_len += pad_len;
 210         }
 211
 212         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 213                 if (rq->cmd_flags & REQ_RW)
 214                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 215
 216                 sg->page_link &= ~0x02;
 217                 sg = sg_next(sg);
 218                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 219                             q->dma_drain_size,
 220                             ((unsigned long)q->dma_drain_buffer) &
 221                             (PAGE_SIZE - 1));
 222                 nsegs++;
 223                 rq->extra_len += q->dma_drain_size;
 224         }
 225
 226         if (sg)
 227                 sg_mark_end(sg);
 228
 229         return nsegs;
 230 }
 231 EXPORT_SYMBOL(blk_rq_map_sg);
 232
 233 static inline int ll_new_mergeable(struct request_queue *q,
 234                                    struct request *req,
 235                                    struct bio *bio)
 236 {
 237         int nr_phys_segs = bio_phys_segments(q, bio);
 238
 239         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 240                 req->cmd_flags |= REQ_NOMERGE;
 241                 if (req == q->last_merge)
 242                         q->last_merge = NULL;
 243                 return 0;
 244         }
 245
 246         /*
 247          * A hw segment is just getting larger, bump just the phys
 248          * counter.
 249          */
 250         req->nr_phys_segments += nr_phys_segs;
 251         return 1;
 252 }
 253
 254 static inline int ll_new_hw_segment(struct request_queue *q,
 255                                     struct request *req,
 256                                     struct bio *bio)
 257 {
 258         int nr_hw_segs = bio_hw_segments(q, bio);
 259         int nr_phys_segs = bio_phys_segments(q, bio);
 260
 261         if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
 262             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 263                 req->cmd_flags |= REQ_NOMERGE;
 264                 if (req == q->last_merge)
 265                         q->last_merge = NULL;
 266                 return 0;
 267         }
 268
 269         /*
 270          * This will form the start of a new hw segment.  Bump both
 271          * counters.
 272          */
 273         req->nr_hw_segments += nr_hw_segs;
 274         req->nr_phys_segments += nr_phys_segs;
 275         return 1;
 276 }
 277
 278 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 279                      struct bio *bio)
 280 {
 281         unsigned short max_sectors;
 282
 283         if (unlikely(blk_pc_request(req)))
 284                 max_sectors = q->max_hw_sectors;
 285         else
 286                 max_sectors = q->max_sectors;
 287
 288         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 289                 req->cmd_flags |= REQ_NOMERGE;
 290                 if (req == q->last_merge)
 291                         q->last_merge = NULL;
 292                 return 0;
 293         }
 294         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 295                 blk_recount_segments(q, req->biotail);
 296         if (!bio_flagged(bio, BIO_SEG_VALID))
 297                 blk_recount_segments(q, bio);
 298
 299         return ll_new_hw_segment(q, req, bio);
 300 }
 301
 302 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 303                       struct bio *bio)
 304 {
 305         unsigned short max_sectors;
 306
 307         if (unlikely(blk_pc_request(req)))
 308                 max_sectors = q->max_hw_sectors;
 309         else
 310                 max_sectors = q->max_sectors;
 311
 312
 313         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 314                 req->cmd_flags |= REQ_NOMERGE;
 315                 if (req == q->last_merge)
 316                         q->last_merge = NULL;
 317                 return 0;
 318         }
 319         if (!bio_flagged(bio, BIO_SEG_VALID))
 320                 blk_recount_segments(q, bio);
 321         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 322                 blk_recount_segments(q, req->bio);
 323
 324         return ll_new_hw_segment(q, req, bio);
 325 }
 326
 327 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 328                                 struct request *next)
 329 {
 330         int total_phys_segments;
 331         int total_hw_segments;
 332
 333         /*
 334          * First check if the either of the requests are re-queued
 335          * requests.  Can't merge them if they are.
 336          */
 337         if (req->special || next->special)
 338                 return 0;
 339
 340         /*
 341          * Will it become too large?
 342          */
 343         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
 344                 return 0;
 345
 346         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 347         if (blk_phys_contig_segment(q, req->biotail, next->bio))
 348                 total_phys_segments--;
 349
 350         if (total_phys_segments > q->max_phys_segments)
 351                 return 0;
 352
 353         total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
 354
 355         if (total_hw_segments > q->max_hw_segments)
 356                 return 0;
 357
 358         /* Merge is OK... */
 359         req->nr_phys_segments = total_phys_segments;
 360         req->nr_hw_segments = total_hw_segments;
 361         return 1;
 362 }
 363
 364 /*
 365  * Has to be called with the request spinlock acquired
 366  */
 367 static int attempt_merge(struct request_queue *q, struct request *req,
 368                           struct request *next)
 369 {
 370         if (!rq_mergeable(req) || !rq_mergeable(next))
 371                 return 0;
 372
 373         /*
 374          * not contiguous
 375          */
 376         if (req->sector + req->nr_sectors != next->sector)
 377                 return 0;
 378
 379         if (rq_data_dir(req) != rq_data_dir(next)
 380             || req->rq_disk != next->rq_disk
 381             || next->special)
 382                 return 0;
 383
 384         if (blk_integrity_rq(req) != blk_integrity_rq(next))
 385                 return 0;
 386
 387         /*
 388          * If we are allowed to merge, then append bio list
 389          * from next to rq and release next. merge_requests_fn
 390          * will have updated segment counts, update sector
 391          * counts here.
 392          */
 393         if (!ll_merge_requests_fn(q, req, next))
 394                 return 0;
 395
 396         /*
 397          * At this point we have either done a back merge
 398          * or front merge. We need the smaller start_time of
 399          * the merged requests to be the current request
 400          * for accounting purposes.
 401          */
 402         if (time_after(req->start_time, next->start_time))
 403                 req->start_time = next->start_time;
 404
 405         req->biotail->bi_next = next->bio;
 406         req->biotail = next->biotail;
 407
 408         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
 409
 410         elv_merge_requests(q, req, next);
 411
 412         if (req->rq_disk) {
 413                 struct hd_struct *part
 414                         = get_part(req->rq_disk, req->sector);
 415                 disk_round_stats(req->rq_disk);
 416                 req->rq_disk->in_flight--;
 417                 if (part) {
 418                         part_round_stats(part);
 419                         part->in_flight--;
 420                 }
 421         }
 422
 423         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 424
 425         __blk_put_request(q, next);
 426         return 1;
 427 }
 428
 429 int attempt_back_merge(struct request_queue *q, struct request *rq)
 430 {
 431         struct request *next = elv_latter_request(q, rq);
 432
 433         if (next)
 434                 return attempt_merge(q, rq, next);
 435
 436         return 0;
 437 }
 438
 439 int attempt_front_merge(struct request_queue *q, struct request *rq)
 440 {
 441         struct request *prev = elv_former_request(q, rq);
 442
 443         if (prev)
 444                 return attempt_merge(q, prev, rq);
 445
 446         return 0;
 447 }