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 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
  13                                              struct bio *bio)
  14 {
  15         unsigned int phys_size;
  16         struct bio_vec *bv, *bvprv = NULL;
  17         int cluster, i, high, highprv = 1;
  18         unsigned int seg_size, nr_phys_segs;
  19         struct bio *fbio, *bbio;
  20
  21         if (!bio)
  22                 return 0;
  23
  24         fbio = bio;
  25         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
  26         seg_size = 0;
  27         phys_size = nr_phys_segs = 0;
  28         for_each_bio(bio) {
  29                 bio_for_each_segment(bv, bio, i) {
  30                         /*
  31                          * the trick here is making sure that a high page is
  32                          * never considered part of another segment, since that
  33                          * might change with the bounce page.
  34                          */
  35                         high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
  36                         if (high || highprv)
  37                                 goto new_segment;
  38                         if (cluster) {
  39                                 if (seg_size + bv->bv_len > q->max_segment_size)
  40                                         goto new_segment;
  41                                 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
  42                                         goto new_segment;
  43                                 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
  44                                         goto new_segment;
  45
  46                                 seg_size += bv->bv_len;
  47                                 bvprv = bv;
  48                                 continue;
  49                         }
  50 new_segment:
  51                         if (nr_phys_segs == 1 && seg_size >
  52                             fbio->bi_seg_front_size)
  53                                 fbio->bi_seg_front_size = seg_size;
  54
  55                         nr_phys_segs++;
  56                         bvprv = bv;
  57                         seg_size = bv->bv_len;
  58                         highprv = high;
  59                 }
  60                 bbio = bio;
  61         }
  62
  63         if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
  64                 fbio->bi_seg_front_size = seg_size;
  65         if (seg_size > bbio->bi_seg_back_size)
  66                 bbio->bi_seg_back_size = seg_size;
  67
  68         return nr_phys_segs;
  69 }
  70
  71 void blk_recalc_rq_segments(struct request *rq)
  72 {
  73         rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
  74 }
  75
  76 void blk_recount_segments(struct request_queue *q, struct bio *bio)
  77 {
  78         struct bio *nxt = bio->bi_next;
  79
  80         bio->bi_next = NULL;
  81         bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
  82         bio->bi_next = nxt;
  83         bio->bi_flags |= (1 << BIO_SEG_VALID);
  84 }
  85 EXPORT_SYMBOL(blk_recount_segments);
  86
  87 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
  88                                    struct bio *nxt)
  89 {
  90         if (!test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags))
  91                 return 0;
  92
  93         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
  94             q->max_segment_size)
  95                 return 0;
  96
  97         if (!bio_has_data(bio))
  98                 return 1;
  99
 100         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
 101                 return 0;
 102
 103         /*
 104          * bio and nxt are contiguous in memory; check if the queue allows
 105          * these two to be merged into one
 106          */
 107         if (BIO_SEG_BOUNDARY(q, bio, nxt))
 108                 return 1;
 109
 110         return 0;
 111 }
 112
 113 /*
 114  * map a request to scatterlist, return number of sg entries setup. Caller
 115  * must make sure sg can hold rq->nr_phys_segments entries
 116  */
 117 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 118                   struct scatterlist *sglist)
 119 {
 120         struct bio_vec *bvec, *bvprv;
 121         struct req_iterator iter;
 122         struct scatterlist *sg;
 123         int nsegs, cluster;
 124
 125         nsegs = 0;
 126         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
 127
 128         /*
 129          * for each bio in rq
 130          */
 131         bvprv = NULL;
 132         sg = NULL;
 133         rq_for_each_segment(bvec, rq, iter) {
 134                 int nbytes = bvec->bv_len;
 135
 136                 if (bvprv && cluster) {
 137                         if (sg->length + nbytes > q->max_segment_size)
 138                                 goto new_segment;
 139
 140                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
 141                                 goto new_segment;
 142                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
 143                                 goto new_segment;
 144
 145                         sg->length += nbytes;
 146                 } else {
 147 new_segment:
 148                         if (!sg)
 149                                 sg = sglist;
 150                         else {
 151                                 /*
 152                                  * If the driver previously mapped a shorter
 153                                  * list, we could see a termination bit
 154                                  * prematurely unless it fully inits the sg
 155                                  * table on each mapping. We KNOW that there
 156                                  * must be more entries here or the driver
 157                                  * would be buggy, so force clear the
 158                                  * termination bit to avoid doing a full
 159                                  * sg_init_table() in drivers for each command.
 160                                  */
 161                                 sg->page_link &= ~0x02;
 162                                 sg = sg_next(sg);
 163                         }
 164
 165                         sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
 166                         nsegs++;
 167                 }
 168                 bvprv = bvec;
 169         } /* segments in rq */
 170
 171
 172         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
 173             (blk_rq_bytes(rq) & q->dma_pad_mask)) {
 174                 unsigned int pad_len =
 175                         (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
 176
 177                 sg->length += pad_len;
 178                 rq->extra_len += pad_len;
 179         }
 180
 181         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
 182                 if (rq->cmd_flags & REQ_RW)
 183                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
 184
 185                 sg->page_link &= ~0x02;
 186                 sg = sg_next(sg);
 187                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
 188                             q->dma_drain_size,
 189                             ((unsigned long)q->dma_drain_buffer) &
 190                             (PAGE_SIZE - 1));
 191                 nsegs++;
 192                 rq->extra_len += q->dma_drain_size;
 193         }
 194
 195         if (sg)
 196                 sg_mark_end(sg);
 197
 198         return nsegs;
 199 }
 200 EXPORT_SYMBOL(blk_rq_map_sg);
 201
 202 static inline int ll_new_hw_segment(struct request_queue *q,
 203                                     struct request *req,
 204                                     struct bio *bio)
 205 {
 206         int nr_phys_segs = bio_phys_segments(q, bio);
 207
 208         if (req->nr_phys_segments + nr_phys_segs > q->max_hw_segments
 209             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
 210                 req->cmd_flags |= REQ_NOMERGE;
 211                 if (req == q->last_merge)
 212                         q->last_merge = NULL;
 213                 return 0;
 214         }
 215
 216         /*
 217          * This will form the start of a new hw segment.  Bump both
 218          * counters.
 219          */
 220         req->nr_phys_segments += nr_phys_segs;
 221         return 1;
 222 }
 223
 224 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 225                      struct bio *bio)
 226 {
 227         unsigned short max_sectors;
 228
 229         if (unlikely(blk_pc_request(req)))
 230                 max_sectors = q->max_hw_sectors;
 231         else
 232                 max_sectors = q->max_sectors;
 233
 234         if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
 235                 req->cmd_flags |= REQ_NOMERGE;
 236                 if (req == q->last_merge)
 237                         q->last_merge = NULL;
 238                 return 0;
 239         }
 240         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 241                 blk_recount_segments(q, req->biotail);
 242         if (!bio_flagged(bio, BIO_SEG_VALID))
 243                 blk_recount_segments(q, bio);
 244
 245         return ll_new_hw_segment(q, req, bio);
 246 }
 247
 248 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 249                       struct bio *bio)
 250 {
 251         unsigned short max_sectors;
 252
 253         if (unlikely(blk_pc_request(req)))
 254                 max_sectors = q->max_hw_sectors;
 255         else
 256                 max_sectors = q->max_sectors;
 257
 258
 259         if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
 260                 req->cmd_flags |= REQ_NOMERGE;
 261                 if (req == q->last_merge)
 262                         q->last_merge = NULL;
 263                 return 0;
 264         }
 265         if (!bio_flagged(bio, BIO_SEG_VALID))
 266                 blk_recount_segments(q, bio);
 267         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 268                 blk_recount_segments(q, req->bio);
 269
 270         return ll_new_hw_segment(q, req, bio);
 271 }
 272
 273 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 274                                 struct request *next)
 275 {
 276         int total_phys_segments;
 277         unsigned int seg_size =
 278                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 279
 280         /*
 281          * First check if the either of the requests are re-queued
 282          * requests.  Can't merge them if they are.
 283          */
 284         if (req->special || next->special)
 285                 return 0;
 286
 287         /*
 288          * Will it become too large?
 289          */
 290         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > q->max_sectors)
 291                 return 0;
 292
 293         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 294         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 295                 if (req->nr_phys_segments == 1)
 296                         req->bio->bi_seg_front_size = seg_size;
 297                 if (next->nr_phys_segments == 1)
 298                         next->biotail->bi_seg_back_size = seg_size;
 299                 total_phys_segments--;
 300         }
 301
 302         if (total_phys_segments > q->max_phys_segments)
 303                 return 0;
 304
 305         if (total_phys_segments > q->max_hw_segments)
 306                 return 0;
 307
 308         /* Merge is OK... */
 309         req->nr_phys_segments = total_phys_segments;
 310         return 1;
 311 }
 312
 313 static void blk_account_io_merge(struct request *req)
 314 {
 315         if (blk_do_io_stat(req)) {
 316                 struct hd_struct *part;
 317                 int cpu;
 318
 319                 cpu = part_stat_lock();
 320                 part = disk_map_sector_rcu(req->rq_disk, blk_rq_pos(req));
 321
 322                 part_round_stats(cpu, part);
 323                 part_dec_in_flight(part);
 324
 325                 part_stat_unlock();
 326         }
 327 }
 328
 329 /*
 330  * Has to be called with the request spinlock acquired
 331  */
 332 static int attempt_merge(struct request_queue *q, struct request *req,
 333                           struct request *next)
 334 {
 335         if (!rq_mergeable(req) || !rq_mergeable(next))
 336                 return 0;
 337
 338         /*
 339          * not contiguous
 340          */
 341         if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
 342                 return 0;
 343
 344         if (rq_data_dir(req) != rq_data_dir(next)
 345             || req->rq_disk != next->rq_disk
 346             || next->special)
 347                 return 0;
 348
 349         if (blk_integrity_rq(req) != blk_integrity_rq(next))
 350                 return 0;
 351
 352         /*
 353          * If we are allowed to merge, then append bio list
 354          * from next to rq and release next. merge_requests_fn
 355          * will have updated segment counts, update sector
 356          * counts here.
 357          */
 358         if (!ll_merge_requests_fn(q, req, next))
 359                 return 0;
 360
 361         /*
 362          * At this point we have either done a back merge
 363          * or front merge. We need the smaller start_time of
 364          * the merged requests to be the current request
 365          * for accounting purposes.
 366          */
 367         if (time_after(req->start_time, next->start_time))
 368                 req->start_time = next->start_time;
 369
 370         req->biotail->bi_next = next->bio;
 371         req->biotail = next->biotail;
 372
 373         req->__data_len += blk_rq_bytes(next);
 374
 375         elv_merge_requests(q, req, next);
 376
 377         /*
 378          * 'next' is going away, so update stats accordingly
 379          */
 380         blk_account_io_merge(next);
 381
 382         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 383         if (blk_rq_cpu_valid(next))
 384                 req->cpu = next->cpu;
 385
 386         /* owner-ship of bio passed from next to req */
 387         next->bio = NULL;
 388         __blk_put_request(q, next);
 389         return 1;
 390 }
 391
 392 int attempt_back_merge(struct request_queue *q, struct request *rq)
 393 {
 394         struct request *next = elv_latter_request(q, rq);
 395
 396         if (next)
 397                 return attempt_merge(q, rq, next);
 398
 399         return 0;
 400 }
 401
 402 int attempt_front_merge(struct request_queue *q, struct request *rq)
 403 {
 404         struct request *prev = elv_former_request(q, rq);
 405
 406         if (prev)
 407                 return attempt_merge(q, prev, rq);
 408
 409         return 0;
 410 }