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         struct bio_vec *bv, *bvprv = NULL;
  16         int cluster, i, high, highprv = 1;
  17         unsigned int seg_size, nr_phys_segs;
  18         struct bio *fbio, *bbio;
  19
  20         if (!bio)
  21                 return 0;
  22
  23         fbio = bio;
  24         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
  25         seg_size = 0;
  26         nr_phys_segs = 0;
  27         for_each_bio(bio) {
  28                 bio_for_each_segment(bv, bio, i) {
  29                         /*
  30                          * the trick here is making sure that a high page is
  31                          * never considered part of another segment, since that
  32                          * might change with the bounce page.
  33                          */
  34                         high = page_to_pfn(bv->bv_page) > queue_bounce_pfn(q);
  35                         if (high || highprv)
  36                                 goto new_segment;
  37                         if (cluster) {
  38                                 if (seg_size + bv->bv_len
  39                                     > queue_max_segment_size(q))
  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             queue_max_segment_size(q))
  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 > queue_max_segment_size(q))
 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_WRITE)
 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 > queue_max_segments(q)) {
 209                 req->cmd_flags |= REQ_NOMERGE;
 210                 if (req == q->last_merge)
 211                         q->last_merge = NULL;
 212                 return 0;
 213         }
 214
 215         /*
 216          * This will form the start of a new hw segment.  Bump both
 217          * counters.
 218          */
 219         req->nr_phys_segments += nr_phys_segs;
 220         return 1;
 221 }
 222
 223 int ll_back_merge_fn(struct request_queue *q, struct request *req,
 224                      struct bio *bio)
 225 {
 226         unsigned short max_sectors;
 227
 228         if (unlikely(req->cmd_type == REQ_TYPE_BLOCK_PC))
 229                 max_sectors = queue_max_hw_sectors(q);
 230         else
 231                 max_sectors = queue_max_sectors(q);
 232
 233         if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
 234                 req->cmd_flags |= REQ_NOMERGE;
 235                 if (req == q->last_merge)
 236                         q->last_merge = NULL;
 237                 return 0;
 238         }
 239         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
 240                 blk_recount_segments(q, req->biotail);
 241         if (!bio_flagged(bio, BIO_SEG_VALID))
 242                 blk_recount_segments(q, bio);
 243
 244         return ll_new_hw_segment(q, req, bio);
 245 }
 246
 247 int ll_front_merge_fn(struct request_queue *q, struct request *req,
 248                       struct bio *bio)
 249 {
 250         unsigned short max_sectors;
 251
 252         if (unlikely(req->cmd_type == REQ_TYPE_BLOCK_PC))
 253                 max_sectors = queue_max_hw_sectors(q);
 254         else
 255                 max_sectors = queue_max_sectors(q);
 256
 257
 258         if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
 259                 req->cmd_flags |= REQ_NOMERGE;
 260                 if (req == q->last_merge)
 261                         q->last_merge = NULL;
 262                 return 0;
 263         }
 264         if (!bio_flagged(bio, BIO_SEG_VALID))
 265                 blk_recount_segments(q, bio);
 266         if (!bio_flagged(req->bio, BIO_SEG_VALID))
 267                 blk_recount_segments(q, req->bio);
 268
 269         return ll_new_hw_segment(q, req, bio);
 270 }
 271
 272 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 273                                 struct request *next)
 274 {
 275         int total_phys_segments;
 276         unsigned int seg_size =
 277                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
 278
 279         /*
 280          * First check if the either of the requests are re-queued
 281          * requests.  Can't merge them if they are.
 282          */
 283         if (req->special || next->special)
 284                 return 0;
 285
 286         /*
 287          * Will it become too large?
 288          */
 289         if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > queue_max_sectors(q))
 290                 return 0;
 291
 292         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
 293         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
 294                 if (req->nr_phys_segments == 1)
 295                         req->bio->bi_seg_front_size = seg_size;
 296                 if (next->nr_phys_segments == 1)
 297                         next->biotail->bi_seg_back_size = seg_size;
 298                 total_phys_segments--;
 299         }
 300
 301         if (total_phys_segments > queue_max_segments(q))
 302                 return 0;
 303
 304         /* Merge is OK... */
 305         req->nr_phys_segments = total_phys_segments;
 306         return 1;
 307 }
 308
 309 /**
 310  * blk_rq_set_mixed_merge - mark a request as mixed merge
 311  * @rq: request to mark as mixed merge
 312  *
 313  * Description:
 314  *     @rq is about to be mixed merged.  Make sure the attributes
 315  *     which can be mixed are set in each bio and mark @rq as mixed
 316  *     merged.
 317  */
 318 void blk_rq_set_mixed_merge(struct request *rq)
 319 {
 320         unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
 321         struct bio *bio;
 322
 323         if (rq->cmd_flags & REQ_MIXED_MERGE)
 324                 return;
 325
 326         /*
 327          * @rq will no longer represent mixable attributes for all the
 328          * contained bios.  It will just track those of the first one.
 329          * Distributes the attributs to each bio.
 330          */
 331         for (bio = rq->bio; bio; bio = bio->bi_next) {
 332                 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
 333                              (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
 334                 bio->bi_rw |= ff;
 335         }
 336         rq->cmd_flags |= REQ_MIXED_MERGE;
 337 }
 338
 339 static void blk_account_io_merge(struct request *req)
 340 {
 341         if (blk_do_io_stat(req)) {
 342                 struct hd_struct *part;
 343                 int cpu;
 344
 345                 cpu = part_stat_lock();
 346                 part = disk_map_sector_rcu(req->rq_disk, blk_rq_pos(req));
 347
 348                 part_round_stats(cpu, part);
 349                 part_dec_in_flight(part, rq_data_dir(req));
 350
 351                 part_stat_unlock();
 352         }
 353 }
 354
 355 /*
 356  * Has to be called with the request spinlock acquired
 357  */
 358 static int attempt_merge(struct request_queue *q, struct request *req,
 359                           struct request *next)
 360 {
 361         if (!rq_mergeable(req) || !rq_mergeable(next))
 362                 return 0;
 363
 364         /*
 365          * Don't merge file system requests and discard requests
 366          */
 367         if ((req->cmd_flags & REQ_DISCARD) != (next->cmd_flags & REQ_DISCARD))
 368                 return 0;
 369
 370         /*
 371          * Don't merge discard requests and secure discard requests
 372          */
 373         if ((req->cmd_flags & REQ_SECURE) != (next->cmd_flags & REQ_SECURE))
 374                 return 0;
 375
 376         /*
 377          * not contiguous
 378          */
 379         if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
 380                 return 0;
 381
 382         if (rq_data_dir(req) != rq_data_dir(next)
 383             || req->rq_disk != next->rq_disk
 384             || next->special)
 385                 return 0;
 386
 387         if (blk_integrity_rq(req) != blk_integrity_rq(next))
 388                 return 0;
 389
 390         /*
 391          * If we are allowed to merge, then append bio list
 392          * from next to rq and release next. merge_requests_fn
 393          * will have updated segment counts, update sector
 394          * counts here.
 395          */
 396         if (!ll_merge_requests_fn(q, req, next))
 397                 return 0;
 398
 399         /*
 400          * If failfast settings disagree or any of the two is already
 401          * a mixed merge, mark both as mixed before proceeding.  This
 402          * makes sure that all involved bios have mixable attributes
 403          * set properly.
 404          */
 405         if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
 406             (req->cmd_flags & REQ_FAILFAST_MASK) !=
 407             (next->cmd_flags & REQ_FAILFAST_MASK)) {
 408                 blk_rq_set_mixed_merge(req);
 409                 blk_rq_set_mixed_merge(next);
 410         }
 411
 412         /*
 413          * At this point we have either done a back merge
 414          * or front merge. We need the smaller start_time of
 415          * the merged requests to be the current request
 416          * for accounting purposes.
 417          */
 418         if (time_after(req->start_time, next->start_time))
 419                 req->start_time = next->start_time;
 420
 421         req->biotail->bi_next = next->bio;
 422         req->biotail = next->biotail;
 423
 424         req->__data_len += blk_rq_bytes(next);
 425
 426         elv_merge_requests(q, req, next);
 427
 428         /*
 429          * 'next' is going away, so update stats accordingly
 430          */
 431         blk_account_io_merge(next);
 432
 433         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
 434         if (blk_rq_cpu_valid(next))
 435                 req->cpu = next->cpu;
 436
 437         /* owner-ship of bio passed from next to req */
 438         next->bio = NULL;
 439         __blk_put_request(q, next);
 440         return 1;
 441 }
 442
 443 int attempt_back_merge(struct request_queue *q, struct request *rq)
 444 {
 445         struct request *next = elv_latter_request(q, rq);
 446
 447         if (next)
 448                 return attempt_merge(q, rq, next);
 449
 450         return 0;
 451 }
 452
 453 int attempt_front_merge(struct request_queue *q, struct request *rq)
 454 {
 455         struct request *prev = elv_former_request(q, rq);
 456
 457         if (prev)
 458                 return attempt_merge(q, prev, rq);
 459
 460         return 0;
 461 }