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         unsigned int phys_size;
  45         struct bio_vec *bv, *bvprv = NULL;
  46         int seg_size;
  47         int cluster;
  48         struct req_iterator iter;
  49         int high, highprv = 1;
  50         struct request_queue *q = rq->q;
  51
  52         if (!rq->bio)
  53                 return;
  54
  55         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
  56         seg_size = 0;
  57         phys_size = nr_phys_segs = 0;
  58         rq_for_each_segment(bv, rq, iter) {
  59                 /*
  60                  * the trick here is making sure that a high page is never
  61                  * considered part of another segment, since that might
  62                  * change with the bounce page.
  63                  */
  64                 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
  65                 if (high || highprv)
  66                         goto new_segment;
  67                 if (cluster) {
  68                         if (seg_size + bv->bv_len > q->max_segment_size)
  69                                 goto new_segment;
  70                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
  71                                 goto new_segment;
  72                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
  73                                 goto new_segment;
  74
  75                         seg_size += bv->bv_len;
  76                         bvprv = bv;
  77                         continue;
  78                 }
  79 new_segment:
  80                 if (nr_phys_segs == 1 && seg_size > rq->bio->bi_seg_front_size)
  81                         rq->bio->bi_seg_front_size = seg_size;
  82
  83                 nr_phys_segs++;
  84                 bvprv = bv;
  85                 seg_size = bv->bv_len;
  86                 highprv = high;
  87         }
  88
  89         if (nr_phys_segs == 1 && seg_size > rq->bio->bi_seg_front_size)
  90                 rq->bio->bi_seg_front_size = seg_size;
  91         if (seg_size > rq->biotail->bi_seg_back_size)
  92                 rq->biotail->bi_seg_back_size = seg_size;
  93
  94         rq->nr_phys_segments = nr_phys_segs;
  95 }
  96
  97 void blk_recount_segments(struct request_queue *q, struct bio *bio)
  98 {
  99         struct request rq;
 100         struct bio *nxt = bio->bi_next;
 101         rq.q = q;
 102         rq.bio = rq.biotail = bio;
 103         bio->bi_next = NULL;
 104         blk_recalc_rq_segments(&rq);
 105         bio->bi_next = nxt;
 106         bio->bi_phys_segments = rq.nr_phys_segments;
 107         bio->bi_flags |= (1 << BIO_SEG_VALID);
 108 }
 109 EXPORT_SYMBOL(blk_recount_segments);
 110
 111 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 112                                    struct bio *nxt)
 113 {
 114         if (!test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags))
 115                 return 0;
 116
 117         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 118             q->max_segment_size)
 119                 return 0;
 120
 121         if (!bio_has_data(bio))
 122                 return 1;
 123
 124         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
 125                 return 0;
 126
 127         /*
 128          * bio and nxt are contiguous in memory; check if the queue allows
 129          * these two to be merged into one
 130          */
 131         if (BIO_SEG_BOUNDARY(q, bio, nxt))
 132                 return 1;
 133
 134         return 0;
 135 }
 136
 137 /*
 138  * map a request to scatterlist, return number of sg entries setup. Caller
 139  * must make sure sg can hold rq->nr_phys_segments entries
 140  */
 141 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 142                   struct scatterlist *sglist)
 143 {
 144         struct bio_vec *bvec, *bvprv;
 145         struct req_iterator iter;
 146         struct scatterlist *sg;
 147         int nsegs, cluster;
 148
 149         nsegs = 0;
 150         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
 151
 152         /*
 153          * for each bio in rq
 154          */
 155         bvprv = NULL;
 156         sg = NULL;
 157         rq_for_each_segment(bvec, rq, iter) {
 158                 int nbytes = bvec->bv_len;
 159
 160                 if (bvprv && cluster) {
 161                         if (sg->length + nbytes > q->max_segment_size)
 162                                 goto new_segment;
 163
 164                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 165                                 goto new_segment;
 166                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 167                                 goto new_segment;
 168
 169                         sg->length += nbytes;
 170                 } else {
 171 new_segment:
 172                         if (!sg)
 173                                 sg = sglist;
 174                         else {
 175                                 /*
 176                                  * If the driver previously mapped a shorter
 177                                  * list, we could see a termination bit
 178                                  * prematurely unless it fully inits the sg
 179                                  * table on each mapping. We KNOW that there
 180                                  * must be more entries here or the driver
 181                                  * would be buggy, so force clear the
 182                                  * termination bit to avoid doing a full
 183                                  * sg_init_table() in drivers for each command.
 184                                  */
 185                                 sg->page_link &= ~0x02;
 186                                 sg = sg_next(sg);
 187                         }
 188
 189                         sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
 190                         nsegs++;
 191                 }
 192                 bvprv = bvec;
 193         } /* segments in rq */
 194
 195
 196         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 197             (rq->data_len & q->dma_pad_mask)) {
 198                 unsigned int pad_len = (q->dma_pad_mask & ~rq->data_len) + 1;
 199
 200                 sg->length += pad_len;
 201                 rq->extra_len += pad_len;
 202         }
 203
 204         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 205                 if (rq->cmd_flags & REQ_RW)
 206                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 207
 208                 sg->page_link &= ~0x02;
 209                 sg = sg_next(sg);
 210                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 211                             q->dma_drain_size,
 212                             ((unsigned long)q->dma_drain_buffer) &
 213                             (PAGE_SIZE - 1));
 214                 nsegs++;
 215                 rq->extra_len += q->dma_drain_size;
 216         }
 217
 218         if (sg)
 219                 sg_mark_end(sg);
 220
 221         return nsegs;
 222 }
 223 EXPORT_SYMBOL(blk_rq_map_sg);
 224
 225 static inline int ll_new_hw_segment(struct request_queue *q,
 226                                     struct request *req,
 227                                     struct bio *bio)
 228 {
 229         int nr_phys_segs = bio_phys_segments(q, bio);
 230
 231         if (req->nr_phys_segments + nr_phys_segs > q->max_hw_segments
 232             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 233                 req->cmd_flags |= REQ_NOMERGE;
 234                 if (req == q->last_merge)
 235                         q->last_merge = NULL;
 236                 return 0;
 237         }
 238
 239         /*
 240          * This will form the start of a new hw segment.  Bump both
 241          * counters.
 242          */
 243         req->nr_phys_segments += nr_phys_segs;
 244         return 1;
 245 }
 246
 247 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 248                      struct bio *bio)
 249 {
 250         unsigned short max_sectors;
 251
 252         if (unlikely(blk_pc_request(req)))
 253                 max_sectors = q->max_hw_sectors;
 254         else
 255                 max_sectors = q->max_sectors;
 256
 257         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 258                 req->cmd_flags |= REQ_NOMERGE;
 259                 if (req == q->last_merge)
 260                         q->last_merge = NULL;
 261                 return 0;
 262         }
 263         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 264                 blk_recount_segments(q, req->biotail);
 265         if (!bio_flagged(bio, BIO_SEG_VALID))
 266                 blk_recount_segments(q, bio);
 267
 268         return ll_new_hw_segment(q, req, bio);
 269 }
 270
 271 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 272                       struct bio *bio)
 273 {
 274         unsigned short max_sectors;
 275
 276         if (unlikely(blk_pc_request(req)))
 277                 max_sectors = q->max_hw_sectors;
 278         else
 279                 max_sectors = q->max_sectors;
 280
 281
 282         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 283                 req->cmd_flags |= REQ_NOMERGE;
 284                 if (req == q->last_merge)
 285                         q->last_merge = NULL;
 286                 return 0;
 287         }
 288         if (!bio_flagged(bio, BIO_SEG_VALID))
 289                 blk_recount_segments(q, bio);
 290         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 291                 blk_recount_segments(q, req->bio);
 292
 293         return ll_new_hw_segment(q, req, bio);
 294 }
 295
 296 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 297                                 struct request *next)
 298 {
 299         int total_phys_segments;
 300         unsigned int seg_size =
 301                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 302
 303         /*
 304          * First check if the either of the requests are re-queued
 305          * requests.  Can't merge them if they are.
 306          */
 307         if (req->special || next->special)
 308                 return 0;
 309
 310         /*
 311          * Will it become too large?
 312          */
 313         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
 314                 return 0;
 315
 316         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 317         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 318                 if (req->nr_phys_segments == 1)
 319                         req->bio->bi_seg_front_size = seg_size;
 320                 if (next->nr_phys_segments == 1)
 321                         next->biotail->bi_seg_back_size = seg_size;
 322                 total_phys_segments--;
 323         }
 324
 325         if (total_phys_segments > q->max_phys_segments)
 326                 return 0;
 327
 328         if (total_phys_segments > q->max_hw_segments)
 329                 return 0;
 330
 331         /* Merge is OK... */
 332         req->nr_phys_segments = total_phys_segments;
 333         return 1;
 334 }
 335
 336 /*
 337  * Has to be called with the request spinlock acquired
 338  */
 339 static int attempt_merge(struct request_queue *q, struct request *req,
 340                           struct request *next)
 341 {
 342         if (!rq_mergeable(req) || !rq_mergeable(next))
 343                 return 0;
 344
 345         /*
 346          * not contiguous
 347          */
 348         if (req->sector + req->nr_sectors != next->sector)
 349                 return 0;
 350
 351         if (rq_data_dir(req) != rq_data_dir(next)
 352             || req->rq_disk != next->rq_disk
 353             || next->special)
 354                 return 0;
 355
 356         if (blk_integrity_rq(req) != blk_integrity_rq(next))
 357                 return 0;
 358
 359         /*
 360          * If we are allowed to merge, then append bio list
 361          * from next to rq and release next. merge_requests_fn
 362          * will have updated segment counts, update sector
 363          * counts here.
 364          */
 365         if (!ll_merge_requests_fn(q, req, next))
 366                 return 0;
 367
 368         /*
 369          * At this point we have either done a back merge
 370          * or front merge. We need the smaller start_time of
 371          * the merged requests to be the current request
 372          * for accounting purposes.
 373          */
 374         if (time_after(req->start_time, next->start_time))
 375                 req->start_time = next->start_time;
 376
 377         req->biotail->bi_next = next->bio;
 378         req->biotail = next->biotail;
 379
 380         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
 381
 382         elv_merge_requests(q, req, next);
 383
 384         if (req->rq_disk) {
 385                 struct hd_struct *part;
 386                 int cpu;
 387
 388                 cpu = part_stat_lock();
 389                 part = disk_map_sector_rcu(req->rq_disk, req->sector);
 390
 391                 part_round_stats(cpu, part);
 392                 part_dec_in_flight(part);
 393
 394                 part_stat_unlock();
 395         }
 396
 397         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 398         if (blk_rq_cpu_valid(next))
 399                 req->cpu = next->cpu;
 400
 401         __blk_put_request(q, next);
 402         return 1;
 403 }
 404
 405 int attempt_back_merge(struct request_queue *q, struct request *rq)
 406 {
 407         struct request *next = elv_latter_request(q, rq);
 408
 409         if (next)
 410                 return attempt_merge(q, rq, next);
 411
 412         return 0;
 413 }
 414
 415 int attempt_front_merge(struct request_queue *q, struct request *rq)
 416 {
 417         struct request *prev = elv_former_request(q, rq);
 418
 419         if (prev)
 420                 return attempt_merge(q, prev, rq);
 421
 422         return 0;
 423 }