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 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
  42                                              struct bio *bio)
  43 {
  44         unsigned int phys_size;
  45         struct bio_vec *bv, *bvprv = NULL;
  46         int cluster, i, high, highprv = 1;
  47         unsigned int seg_size, nr_phys_segs;
  48         struct bio *fbio, *bbio;
  49
  50         if (!bio)
  51                 return 0;
  52
  53         fbio = bio;
  54         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
  55         seg_size = 0;
  56         phys_size = nr_phys_segs = 0;
  57         for_each_bio(bio) {
  58                 bio_for_each_segment(bv, bio, i) {
  59                         /*
  60                          * the trick here is making sure that a high page is
  61                          * never considered part of another segment, since that
  62                          * might 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 >
  81                             fbio->bi_seg_front_size)
  82                                 fbio->bi_seg_front_size = seg_size;
  83
  84                         nr_phys_segs++;
  85                         bvprv = bv;
  86                         seg_size = bv->bv_len;
  87                         highprv = high;
  88                 }
  89                 bbio = bio;
  90         }
  91
  92         if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
  93                 fbio->bi_seg_front_size = seg_size;
  94         if (seg_size > bbio->bi_seg_back_size)
  95                 bbio->bi_seg_back_size = seg_size;
  96
  97         return nr_phys_segs;
  98 }
  99
 100 void blk_recalc_rq_segments(struct request *rq)
 101 {
 102         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
 103 }
 104
 105 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 106 {
 107         struct bio *nxt = bio->bi_next;
 108
 109         bio->bi_next = NULL;
 110         bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
 111         bio->bi_next = nxt;
 112         bio->bi_flags |= (1 << BIO_SEG_VALID);
 113 }
 114 EXPORT_SYMBOL(blk_recount_segments);
 115
 116 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 117                                    struct bio *nxt)
 118 {
 119         if (!test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags))
 120                 return 0;
 121
 122         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 123             q->max_segment_size)
 124                 return 0;
 125
 126         if (!bio_has_data(bio))
 127                 return 1;
 128
 129         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
 130                 return 0;
 131
 132         /*
 133          * bio and nxt are contiguous in memory; check if the queue allows
 134          * these two to be merged into one
 135          */
 136         if (BIO_SEG_BOUNDARY(q, bio, nxt))
 137                 return 1;
 138
 139         return 0;
 140 }
 141
 142 /*
 143  * map a request to scatterlist, return number of sg entries setup. Caller
 144  * must make sure sg can hold rq->nr_phys_segments entries
 145  */
 146 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 147                   struct scatterlist *sglist)
 148 {
 149         struct bio_vec *bvec, *bvprv;
 150         struct req_iterator iter;
 151         struct scatterlist *sg;
 152         int nsegs, cluster;
 153
 154         nsegs = 0;
 155         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
 156
 157         /*
 158          * for each bio in rq
 159          */
 160         bvprv = NULL;
 161         sg = NULL;
 162         rq_for_each_segment(bvec, rq, iter) {
 163                 int nbytes = bvec->bv_len;
 164
 165                 if (bvprv && cluster) {
 166                         if (sg->length + nbytes > q->max_segment_size)
 167                                 goto new_segment;
 168
 169                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 170                                 goto new_segment;
 171                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 172                                 goto new_segment;
 173
 174                         sg->length += nbytes;
 175                 } else {
 176 new_segment:
 177                         if (!sg)
 178                                 sg = sglist;
 179                         else {
 180                                 /*
 181                                  * If the driver previously mapped a shorter
 182                                  * list, we could see a termination bit
 183                                  * prematurely unless it fully inits the sg
 184                                  * table on each mapping. We KNOW that there
 185                                  * must be more entries here or the driver
 186                                  * would be buggy, so force clear the
 187                                  * termination bit to avoid doing a full
 188                                  * sg_init_table() in drivers for each command.
 189                                  */
 190                                 sg->page_link &= ~0x02;
 191                                 sg = sg_next(sg);
 192                         }
 193
 194                         sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
 195                         nsegs++;
 196                 }
 197                 bvprv = bvec;
 198         } /* segments in rq */
 199
 200
 201         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 202             (rq->data_len & q->dma_pad_mask)) {
 203                 unsigned int pad_len = (q->dma_pad_mask & ~rq->data_len) + 1;
 204
 205                 sg->length += pad_len;
 206                 rq->extra_len += pad_len;
 207         }
 208
 209         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 210                 if (rq->cmd_flags & REQ_RW)
 211                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 212
 213                 sg->page_link &= ~0x02;
 214                 sg = sg_next(sg);
 215                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 216                             q->dma_drain_size,
 217                             ((unsigned long)q->dma_drain_buffer) &
 218                             (PAGE_SIZE - 1));
 219                 nsegs++;
 220                 rq->extra_len += q->dma_drain_size;
 221         }
 222
 223         if (sg)
 224                 sg_mark_end(sg);
 225
 226         return nsegs;
 227 }
 228 EXPORT_SYMBOL(blk_rq_map_sg);
 229
 230 static inline int ll_new_hw_segment(struct request_queue *q,
 231                                     struct request *req,
 232                                     struct bio *bio)
 233 {
 234         int nr_phys_segs = bio_phys_segments(q, bio);
 235
 236         if (req->nr_phys_segments + nr_phys_segs > q->max_hw_segments
 237             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 238                 req->cmd_flags |= REQ_NOMERGE;
 239                 if (req == q->last_merge)
 240                         q->last_merge = NULL;
 241                 return 0;
 242         }
 243
 244         /*
 245          * This will form the start of a new hw segment.  Bump both
 246          * counters.
 247          */
 248         req->nr_phys_segments += nr_phys_segs;
 249         return 1;
 250 }
 251
 252 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 253                      struct bio *bio)
 254 {
 255         unsigned short max_sectors;
 256
 257         if (unlikely(blk_pc_request(req)))
 258                 max_sectors = q->max_hw_sectors;
 259         else
 260                 max_sectors = q->max_sectors;
 261
 262         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 263                 req->cmd_flags |= REQ_NOMERGE;
 264                 if (req == q->last_merge)
 265                         q->last_merge = NULL;
 266                 return 0;
 267         }
 268         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 269                 blk_recount_segments(q, req->biotail);
 270         if (!bio_flagged(bio, BIO_SEG_VALID))
 271                 blk_recount_segments(q, bio);
 272
 273         return ll_new_hw_segment(q, req, bio);
 274 }
 275
 276 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 277                       struct bio *bio)
 278 {
 279         unsigned short max_sectors;
 280
 281         if (unlikely(blk_pc_request(req)))
 282                 max_sectors = q->max_hw_sectors;
 283         else
 284                 max_sectors = q->max_sectors;
 285
 286
 287         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
 288                 req->cmd_flags |= REQ_NOMERGE;
 289                 if (req == q->last_merge)
 290                         q->last_merge = NULL;
 291                 return 0;
 292         }
 293         if (!bio_flagged(bio, BIO_SEG_VALID))
 294                 blk_recount_segments(q, bio);
 295         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 296                 blk_recount_segments(q, req->bio);
 297
 298         return ll_new_hw_segment(q, req, bio);
 299 }
 300
 301 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 302                                 struct request *next)
 303 {
 304         int total_phys_segments;
 305         unsigned int seg_size =
 306                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 307
 308         /*
 309          * First check if the either of the requests are re-queued
 310          * requests.  Can't merge them if they are.
 311          */
 312         if (req->special || next->special)
 313                 return 0;
 314
 315         /*
 316          * Will it become too large?
 317          */
 318         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
 319                 return 0;
 320
 321         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 322         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 323                 if (req->nr_phys_segments == 1)
 324                         req->bio->bi_seg_front_size = seg_size;
 325                 if (next->nr_phys_segments == 1)
 326                         next->biotail->bi_seg_back_size = seg_size;
 327                 total_phys_segments--;
 328         }
 329
 330         if (total_phys_segments > q->max_phys_segments)
 331                 return 0;
 332
 333         if (total_phys_segments > q->max_hw_segments)
 334                 return 0;
 335
 336         /* Merge is OK... */
 337         req->nr_phys_segments = total_phys_segments;
 338         return 1;
 339 }
 340
 341 static void blk_account_io_merge(struct request *req)
 342 {
 343         if (blk_do_io_stat(req)) {
 344                 struct hd_struct *part;
 345                 int cpu;
 346
 347                 cpu = part_stat_lock();
 348                 part = disk_map_sector_rcu(req->rq_disk, req->sector);
 349
 350                 part_round_stats(cpu, part);
 351                 part_dec_in_flight(part);
 352
 353                 part_stat_unlock();
 354         }
 355 }
 356
 357 /*
 358  * Has to be called with the request spinlock acquired
 359  */
 360 static int attempt_merge(struct request_queue *q, struct request *req,
 361                           struct request *next)
 362 {
 363         if (!rq_mergeable(req) || !rq_mergeable(next))
 364                 return 0;
 365
 366         /*
 367          * not contiguous
 368          */
 369         if (req->sector + req->nr_sectors != next->sector)
 370                 return 0;
 371
 372         if (rq_data_dir(req) != rq_data_dir(next)
 373             || req->rq_disk != next->rq_disk
 374             || next->special)
 375                 return 0;
 376
 377         if (blk_integrity_rq(req) != blk_integrity_rq(next))
 378                 return 0;
 379
 380         /*
 381          * If we are allowed to merge, then append bio list
 382          * from next to rq and release next. merge_requests_fn
 383          * will have updated segment counts, update sector
 384          * counts here.
 385          */
 386         if (!ll_merge_requests_fn(q, req, next))
 387                 return 0;
 388
 389         /*
 390          * At this point we have either done a back merge
 391          * or front merge. We need the smaller start_time of
 392          * the merged requests to be the current request
 393          * for accounting purposes.
 394          */
 395         if (time_after(req->start_time, next->start_time))
 396                 req->start_time = next->start_time;
 397
 398         req->biotail->bi_next = next->bio;
 399         req->biotail = next->biotail;
 400
 401         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
 402
 403         elv_merge_requests(q, req, next);
 404
 405         /*
 406          * 'next' is going away, so update stats accordingly
 407          */
 408         blk_account_io_merge(next);
 409
 410         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 411         if (blk_rq_cpu_valid(next))
 412                 req->cpu = next->cpu;
 413
 414         /* owner-ship of bio passed from next to req */
 415         next->bio = NULL;
 416         __blk_put_request(q, next);
 417         return 1;
 418 }
 419
 420 int attempt_back_merge(struct request_queue *q, struct request *rq)
 421 {
 422         struct request *next = elv_latter_request(q, rq);
 423
 424         if (next)
 425                 return attempt_merge(q, rq, next);
 426
 427         return 0;
 428 }
 429
 430 int attempt_front_merge(struct request_queue *q, struct request *rq)
 431 {
 432         struct request *prev = elv_former_request(q, rq);
 433
 434         if (prev)
 435                 return attempt_merge(q, prev, rq);
 436
 437         return 0;
 438 }