block/elevator.c

   1 /*
   2  *  Block device elevator/IO-scheduler.
   3  *
   4  *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
   5  *
   6  * 30042000 Jens Axboe <axboe@kernel.dk> :
   7  *
   8  * Split the elevator a bit so that it is possible to choose a different
   9  * one or even write a new "plug in". There are three pieces:
  10  * - elevator_fn, inserts a new request in the queue list
  11  * - elevator_merge_fn, decides whether a new buffer can be merged with
  12  *   an existing request
  13  * - elevator_dequeue_fn, called when a request is taken off the active list
  14  *
  15  * 20082000 Dave Jones <davej@suse.de> :
  16  * Removed tests for max-bomb-segments, which was breaking elvtune
  17  *  when run without -bN
  18  *
  19  * Jens:
  20  * - Rework again to work with bio instead of buffer_heads
  21  * - loose bi_dev comparisons, partition handling is right now
  22  * - completely modularize elevator setup and teardown
  23  *
  24  */
  25 #include <linux/kernel.h>
  26 #include <linux/fs.h>
  27 #include <linux/blkdev.h>
  28 #include <linux/elevator.h>
  29 #include <linux/bio.h>
  30 #include <linux/module.h>
  31 #include <linux/slab.h>
  32 #include <linux/init.h>
  33 #include <linux/compiler.h>
  34 #include <linux/delay.h>
  35 #include <linux/blktrace_api.h>
  36 #include <linux/hash.h>
  37
  38 #include <asm/uaccess.h>
  39
  40 static DEFINE_SPINLOCK(elv_list_lock);
  41 static LIST_HEAD(elv_list);
  42
  43 /*
  44  * Merge hash stuff.
  45  */
  46 static const int elv_hash_shift = 6;
  47 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  48 #define ELV_HASH_FN(sec)        \
  49                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  50 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  51 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  52 #define ELV_ON_HASH(rq)         (!hlist_unhashed(&(rq)->hash))
  53
  54 /*
  55  * Query io scheduler to see if the current process issuing bio may be
  56  * merged with rq.
  57  */
  58 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  59 {
  60         struct request_queue *q = rq->q;
  61         elevator_t *e = q->elevator;
  62
  63         if (e->ops->elevator_allow_merge_fn)
  64                 return e->ops->elevator_allow_merge_fn(q, rq, bio);
  65
  66         return 1;
  67 }
  68
  69 /*
  70  * can we safely merge with this request?
  71  */
  72 int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  73 {
  74         if (!rq_mergeable(rq))
  75                 return 0;
  76
  77         /*
  78          * different data direction or already started, don't merge
  79          */
  80         if (bio_data_dir(bio) != rq_data_dir(rq))
  81                 return 0;
  82
  83         /*
  84          * must be same device and not a special request
  85          */
  86         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
  87                 return 0;
  88
  89         if (!elv_iosched_allow_merge(rq, bio))
  90                 return 0;
  91
  92         return 1;
  93 }
  94 EXPORT_SYMBOL(elv_rq_merge_ok);
  95
  96 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
  97 {
  98         int ret = ELEVATOR_NO_MERGE;
  99
 100         /*
 101          * we can merge and sequence is ok, check if it's possible
 102          */
 103         if (elv_rq_merge_ok(__rq, bio)) {
 104                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 105                         ret = ELEVATOR_BACK_MERGE;
 106                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 107                         ret = ELEVATOR_FRONT_MERGE;
 108         }
 109
 110         return ret;
 111 }
 112
 113 static struct elevator_type *elevator_find(const char *name)
 114 {
 115         struct elevator_type *e;
 116
 117         list_for_each_entry(e, &elv_list, list) {
 118                 if (!strcmp(e->elevator_name, name))
 119                         return e;
 120         }
 121
 122         return NULL;
 123 }
 124
 125 static void elevator_put(struct elevator_type *e)
 126 {
 127         module_put(e->elevator_owner);
 128 }
 129
 130 static struct elevator_type *elevator_get(const char *name)
 131 {
 132         struct elevator_type *e;
 133
 134         spin_lock(&elv_list_lock);
 135
 136         e = elevator_find(name);
 137         if (!e) {
 138                 char elv[ELV_NAME_MAX + strlen("-iosched")];
 139
 140                 spin_unlock(&elv_list_lock);
 141
 142                 if (!strcmp(name, "anticipatory"))
 143                         sprintf(elv, "as-iosched");
 144                 else
 145                         sprintf(elv, "%s-iosched", name);
 146
 147                 request_module(elv);
 148                 spin_lock(&elv_list_lock);
 149                 e = elevator_find(name);
 150         }
 151
 152         if (e && !try_module_get(e->elevator_owner))
 153                 e = NULL;
 154
 155         spin_unlock(&elv_list_lock);
 156
 157         return e;
 158 }
 159
 160 static void *elevator_init_queue(struct request_queue *q,
 161                                  struct elevator_queue *eq)
 162 {
 163         return eq->ops->elevator_init_fn(q);
 164 }
 165
 166 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 167                            void *data)
 168 {
 169         q->elevator = eq;
 170         eq->elevator_data = data;
 171 }
 172
 173 static char chosen_elevator[16];
 174
 175 static int __init elevator_setup(char *str)
 176 {
 177         /*
 178          * Be backwards-compatible with previous kernels, so users
 179          * won't get the wrong elevator.
 180          */
 181         if (!strcmp(str, "as"))
 182                 strcpy(chosen_elevator, "anticipatory");
 183         else
 184                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 185         return 1;
 186 }
 187
 188 __setup("elevator=", elevator_setup);
 189
 190 static struct kobj_type elv_ktype;
 191
 192 static elevator_t *elevator_alloc(struct request_queue *q,
 193                                   struct elevator_type *e)
 194 {
 195         elevator_t *eq;
 196         int i;
 197
 198         eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
 199         if (unlikely(!eq))
 200                 goto err;
 201
 202         eq->ops = &e->ops;
 203         eq->elevator_type = e;
 204         kobject_init(&eq->kobj, &elv_ktype);
 205         mutex_init(&eq->sysfs_lock);
 206
 207         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 208                                         GFP_KERNEL, q->node);
 209         if (!eq->hash)
 210                 goto err;
 211
 212         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 213                 INIT_HLIST_HEAD(&eq->hash[i]);
 214
 215         return eq;
 216 err:
 217         kfree(eq);
 218         elevator_put(e);
 219         return NULL;
 220 }
 221
 222 static void elevator_release(struct kobject *kobj)
 223 {
 224         elevator_t *e = container_of(kobj, elevator_t, kobj);
 225
 226         elevator_put(e->elevator_type);
 227         kfree(e->hash);
 228         kfree(e);
 229 }
 230
 231 int elevator_init(struct request_queue *q, char *name)
 232 {
 233         struct elevator_type *e = NULL;
 234         struct elevator_queue *eq;
 235         int ret = 0;
 236         void *data;
 237
 238         INIT_LIST_HEAD(&q->queue_head);
 239         q->last_merge = NULL;
 240         q->end_sector = 0;
 241         q->boundary_rq = NULL;
 242
 243         if (name) {
 244                 e = elevator_get(name);
 245                 if (!e)
 246                         return -EINVAL;
 247         }
 248
 249         if (!e && *chosen_elevator) {
 250                 e = elevator_get(chosen_elevator);
 251                 if (!e)
 252                         printk(KERN_ERR "I/O scheduler %s not found\n",
 253                                                         chosen_elevator);
 254         }
 255
 256         if (!e) {
 257                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 258                 if (!e) {
 259                         printk(KERN_ERR
 260                                 "Default I/O scheduler not found. " \
 261                                 "Using noop.\n");
 262                         e = elevator_get("noop");
 263                 }
 264         }
 265
 266         eq = elevator_alloc(q, e);
 267         if (!eq)
 268                 return -ENOMEM;
 269
 270         data = elevator_init_queue(q, eq);
 271         if (!data) {
 272                 kobject_put(&eq->kobj);
 273                 return -ENOMEM;
 274         }
 275
 276         elevator_attach(q, eq, data);
 277         return ret;
 278 }
 279 EXPORT_SYMBOL(elevator_init);
 280
 281 void elevator_exit(elevator_t *e)
 282 {
 283         mutex_lock(&e->sysfs_lock);
 284         if (e->ops->elevator_exit_fn)
 285                 e->ops->elevator_exit_fn(e);
 286         e->ops = NULL;
 287         mutex_unlock(&e->sysfs_lock);
 288
 289         kobject_put(&e->kobj);
 290 }
 291 EXPORT_SYMBOL(elevator_exit);
 292
 293 static void elv_activate_rq(struct request_queue *q, struct request *rq)
 294 {
 295         elevator_t *e = q->elevator;
 296
 297         if (e->ops->elevator_activate_req_fn)
 298                 e->ops->elevator_activate_req_fn(q, rq);
 299 }
 300
 301 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
 302 {
 303         elevator_t *e = q->elevator;
 304
 305         if (e->ops->elevator_deactivate_req_fn)
 306                 e->ops->elevator_deactivate_req_fn(q, rq);
 307 }
 308
 309 static inline void __elv_rqhash_del(struct request *rq)
 310 {
 311         hlist_del_init(&rq->hash);
 312 }
 313
 314 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 315 {
 316         if (ELV_ON_HASH(rq))
 317                 __elv_rqhash_del(rq);
 318 }
 319
 320 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 321 {
 322         elevator_t *e = q->elevator;
 323
 324         BUG_ON(ELV_ON_HASH(rq));
 325         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 326 }
 327
 328 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 329 {
 330         __elv_rqhash_del(rq);
 331         elv_rqhash_add(q, rq);
 332 }
 333
 334 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 335 {
 336         elevator_t *e = q->elevator;
 337         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 338         struct hlist_node *entry, *next;
 339         struct request *rq;
 340
 341         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 342                 BUG_ON(!ELV_ON_HASH(rq));
 343
 344                 if (unlikely(!rq_mergeable(rq))) {
 345                         __elv_rqhash_del(rq);
 346                         continue;
 347                 }
 348
 349                 if (rq_hash_key(rq) == offset)
 350                         return rq;
 351         }
 352
 353         return NULL;
 354 }
 355
 356 /*
 357  * RB-tree support functions for inserting/lookup/removal of requests
 358  * in a sorted RB tree.
 359  */
 360 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 361 {
 362         struct rb_node **p = &root->rb_node;
 363         struct rb_node *parent = NULL;
 364         struct request *__rq;
 365
 366         while (*p) {
 367                 parent = *p;
 368                 __rq = rb_entry(parent, struct request, rb_node);
 369
 370                 if (rq->sector < __rq->sector)
 371                         p = &(*p)->rb_left;
 372                 else if (rq->sector > __rq->sector)
 373                         p = &(*p)->rb_right;
 374                 else
 375                         return __rq;
 376         }
 377
 378         rb_link_node(&rq->rb_node, parent, p);
 379         rb_insert_color(&rq->rb_node, root);
 380         return NULL;
 381 }
 382 EXPORT_SYMBOL(elv_rb_add);
 383
 384 void elv_rb_del(struct rb_root *root, struct request *rq)
 385 {
 386         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 387         rb_erase(&rq->rb_node, root);
 388         RB_CLEAR_NODE(&rq->rb_node);
 389 }
 390 EXPORT_SYMBOL(elv_rb_del);
 391
 392 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 393 {
 394         struct rb_node *n = root->rb_node;
 395         struct request *rq;
 396
 397         while (n) {
 398                 rq = rb_entry(n, struct request, rb_node);
 399
 400                 if (sector < rq->sector)
 401                         n = n->rb_left;
 402                 else if (sector > rq->sector)
 403                         n = n->rb_right;
 404                 else
 405                         return rq;
 406         }
 407
 408         return NULL;
 409 }
 410 EXPORT_SYMBOL(elv_rb_find);
 411
 412 /*
 413  * Insert rq into dispatch queue of q.  Queue lock must be held on
 414  * entry.  rq is sort instead into the dispatch queue. To be used by
 415  * specific elevators.
 416  */
 417 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 418 {
 419         sector_t boundary;
 420         struct list_head *entry;
 421         int stop_flags;
 422
 423         if (q->last_merge == rq)
 424                 q->last_merge = NULL;
 425
 426         elv_rqhash_del(q, rq);
 427
 428         q->nr_sorted--;
 429
 430         boundary = q->end_sector;
 431         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 432         list_for_each_prev(entry, &q->queue_head) {
 433                 struct request *pos = list_entry_rq(entry);
 434
 435                 if (rq_data_dir(rq) != rq_data_dir(pos))
 436                         break;
 437                 if (pos->cmd_flags & stop_flags)
 438                         break;
 439                 if (rq->sector >= boundary) {
 440                         if (pos->sector < boundary)
 441                                 continue;
 442                 } else {
 443                         if (pos->sector >= boundary)
 444                                 break;
 445                 }
 446                 if (rq->sector >= pos->sector)
 447                         break;
 448         }
 449
 450         list_add(&rq->queuelist, entry);
 451 }
 452 EXPORT_SYMBOL(elv_dispatch_sort);
 453
 454 /*
 455  * Insert rq into dispatch queue of q.  Queue lock must be held on
 456  * entry.  rq is added to the back of the dispatch queue. To be used by
 457  * specific elevators.
 458  */
 459 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 460 {
 461         if (q->last_merge == rq)
 462                 q->last_merge = NULL;
 463
 464         elv_rqhash_del(q, rq);
 465
 466         q->nr_sorted--;
 467
 468         q->end_sector = rq_end_sector(rq);
 469         q->boundary_rq = rq;
 470         list_add_tail(&rq->queuelist, &q->queue_head);
 471 }
 472 EXPORT_SYMBOL(elv_dispatch_add_tail);
 473
 474 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 475 {
 476         elevator_t *e = q->elevator;
 477         struct request *__rq;
 478         int ret;
 479
 480         /*
 481          * First try one-hit cache.
 482          */
 483         if (q->last_merge) {
 484                 ret = elv_try_merge(q->last_merge, bio);
 485                 if (ret != ELEVATOR_NO_MERGE) {
 486                         *req = q->last_merge;
 487                         return ret;
 488                 }
 489         }
 490
 491         if (blk_queue_nomerges(q))
 492                 return ELEVATOR_NO_MERGE;
 493
 494         /*
 495          * See if our hash lookup can find a potential backmerge.
 496          */
 497         __rq = elv_rqhash_find(q, bio->bi_sector);
 498         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 499                 *req = __rq;
 500                 return ELEVATOR_BACK_MERGE;
 501         }
 502
 503         if (e->ops->elevator_merge_fn)
 504                 return e->ops->elevator_merge_fn(q, req, bio);
 505
 506         return ELEVATOR_NO_MERGE;
 507 }
 508
 509 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 510 {
 511         elevator_t *e = q->elevator;
 512
 513         if (e->ops->elevator_merged_fn)
 514                 e->ops->elevator_merged_fn(q, rq, type);
 515
 516         if (type == ELEVATOR_BACK_MERGE)
 517                 elv_rqhash_reposition(q, rq);
 518
 519         q->last_merge = rq;
 520 }
 521
 522 void elv_merge_requests(struct request_queue *q, struct request *rq,
 523                              struct request *next)
 524 {
 525         elevator_t *e = q->elevator;
 526
 527         if (e->ops->elevator_merge_req_fn)
 528                 e->ops->elevator_merge_req_fn(q, rq, next);
 529
 530         elv_rqhash_reposition(q, rq);
 531         elv_rqhash_del(q, next);
 532
 533         q->nr_sorted--;
 534         q->last_merge = rq;
 535 }
 536
 537 void elv_requeue_request(struct request_queue *q, struct request *rq)
 538 {
 539         /*
 540          * it already went through dequeue, we need to decrement the
 541          * in_flight count again
 542          */
 543         if (blk_account_rq(rq)) {
 544                 q->in_flight--;
 545                 if (blk_sorted_rq(rq))
 546                         elv_deactivate_rq(q, rq);
 547         }
 548
 549         rq->cmd_flags &= ~REQ_STARTED;
 550
 551         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 552 }
 553
 554 static void elv_drain_elevator(struct request_queue *q)
 555 {
 556         static int printed;
 557         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 558                 ;
 559         if (q->nr_sorted == 0)
 560                 return;
 561         if (printed++ < 10) {
 562                 printk(KERN_ERR "%s: forced dispatching is broken "
 563                        "(nr_sorted=%u), please report this\n",
 564                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 565         }
 566 }
 567
 568 void elv_insert(struct request_queue *q, struct request *rq, int where)
 569 {
 570         struct list_head *pos;
 571         unsigned ordseq;
 572         int unplug_it = 1;
 573
 574         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 575
 576         rq->q = q;
 577
 578         switch (where) {
 579         case ELEVATOR_INSERT_FRONT:
 580                 rq->cmd_flags |= REQ_SOFTBARRIER;
 581
 582                 list_add(&rq->queuelist, &q->queue_head);
 583                 break;
 584
 585         case ELEVATOR_INSERT_BACK:
 586                 rq->cmd_flags |= REQ_SOFTBARRIER;
 587                 elv_drain_elevator(q);
 588                 list_add_tail(&rq->queuelist, &q->queue_head);
 589                 /*
 590                  * We kick the queue here for the following reasons.
 591                  * - The elevator might have returned NULL previously
 592                  *   to delay requests and returned them now.  As the
 593                  *   queue wasn't empty before this request, ll_rw_blk
 594                  *   won't run the queue on return, resulting in hang.
 595                  * - Usually, back inserted requests won't be merged
 596                  *   with anything.  There's no point in delaying queue
 597                  *   processing.
 598                  */
 599                 blk_remove_plug(q);
 600                 q->request_fn(q);
 601                 break;
 602
 603         case ELEVATOR_INSERT_SORT:
 604                 BUG_ON(!blk_fs_request(rq));
 605                 rq->cmd_flags |= REQ_SORTED;
 606                 q->nr_sorted++;
 607                 if (rq_mergeable(rq)) {
 608                         elv_rqhash_add(q, rq);
 609                         if (!q->last_merge)
 610                                 q->last_merge = rq;
 611                 }
 612
 613                 /*
 614                  * Some ioscheds (cfq) run q->request_fn directly, so
 615                  * rq cannot be accessed after calling
 616                  * elevator_add_req_fn.
 617                  */
 618                 q->elevator->ops->elevator_add_req_fn(q, rq);
 619                 break;
 620
 621         case ELEVATOR_INSERT_REQUEUE:
 622                 /*
 623                  * If ordered flush isn't in progress, we do front
 624                  * insertion; otherwise, requests should be requeued
 625                  * in ordseq order.
 626                  */
 627                 rq->cmd_flags |= REQ_SOFTBARRIER;
 628
 629                 /*
 630                  * Most requeues happen because of a busy condition,
 631                  * don't force unplug of the queue for that case.
 632                  */
 633                 unplug_it = 0;
 634
 635                 if (q->ordseq == 0) {
 636                         list_add(&rq->queuelist, &q->queue_head);
 637                         break;
 638                 }
 639
 640                 ordseq = blk_ordered_req_seq(rq);
 641
 642                 list_for_each(pos, &q->queue_head) {
 643                         struct request *pos_rq = list_entry_rq(pos);
 644                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 645                                 break;
 646                 }
 647
 648                 list_add_tail(&rq->queuelist, pos);
 649                 break;
 650
 651         default:
 652                 printk(KERN_ERR "%s: bad insertion point %d\n",
 653                        __func__, where);
 654                 BUG();
 655         }
 656
 657         if (unplug_it && blk_queue_plugged(q)) {
 658                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 659                         - q->in_flight;
 660
 661                 if (nrq >= q->unplug_thresh)
 662                         __generic_unplug_device(q);
 663         }
 664 }
 665
 666 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 667                        int plug)
 668 {
 669         if (q->ordcolor)
 670                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 671
 672         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 673                 /*
 674                  * toggle ordered color
 675                  */
 676                 if (blk_barrier_rq(rq))
 677                         q->ordcolor ^= 1;
 678
 679                 /*
 680                  * barriers implicitly indicate back insertion
 681                  */
 682                 if (where == ELEVATOR_INSERT_SORT)
 683                         where = ELEVATOR_INSERT_BACK;
 684
 685                 /*
 686                  * this request is scheduling boundary, update
 687                  * end_sector
 688                  */
 689                 if (blk_fs_request(rq)) {
 690                         q->end_sector = rq_end_sector(rq);
 691                         q->boundary_rq = rq;
 692                 }
 693         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 694                     where == ELEVATOR_INSERT_SORT)
 695                 where = ELEVATOR_INSERT_BACK;
 696
 697         if (plug)
 698                 blk_plug_device(q);
 699
 700         elv_insert(q, rq, where);
 701 }
 702 EXPORT_SYMBOL(__elv_add_request);
 703
 704 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 705                      int plug)
 706 {
 707         unsigned long flags;
 708
 709         spin_lock_irqsave(q->queue_lock, flags);
 710         __elv_add_request(q, rq, where, plug);
 711         spin_unlock_irqrestore(q->queue_lock, flags);
 712 }
 713 EXPORT_SYMBOL(elv_add_request);
 714
 715 static inline struct request *__elv_next_request(struct request_queue *q)
 716 {
 717         struct request *rq;
 718
 719         while (1) {
 720                 while (!list_empty(&q->queue_head)) {
 721                         rq = list_entry_rq(q->queue_head.next);
 722                         if (blk_do_ordered(q, &rq))
 723                                 return rq;
 724                 }
 725
 726                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 727                         return NULL;
 728         }
 729 }
 730
 731 struct request *elv_next_request(struct request_queue *q)
 732 {
 733         struct request *rq;
 734         int ret;
 735
 736         while ((rq = __elv_next_request(q)) != NULL) {
 737                 /*
 738                  * Kill the empty barrier place holder, the driver must
 739                  * not ever see it.
 740                  */
 741                 if (blk_empty_barrier(rq)) {
 742                         end_queued_request(rq, 1);
 743                         continue;
 744                 }
 745                 if (!(rq->cmd_flags & REQ_STARTED)) {
 746                         /*
 747                          * This is the first time the device driver
 748                          * sees this request (possibly after
 749                          * requeueing).  Notify IO scheduler.
 750                          */
 751                         if (blk_sorted_rq(rq))
 752                                 elv_activate_rq(q, rq);
 753
 754                         /*
 755                          * just mark as started even if we don't start
 756                          * it, a request that has been delayed should
 757                          * not be passed by new incoming requests
 758                          */
 759                         rq->cmd_flags |= REQ_STARTED;
 760                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 761                 }
 762
 763                 if (!q->boundary_rq || q->boundary_rq == rq) {
 764                         q->end_sector = rq_end_sector(rq);
 765                         q->boundary_rq = NULL;
 766                 }
 767
 768                 if (rq->cmd_flags & REQ_DONTPREP)
 769                         break;
 770
 771                 if (q->dma_drain_size && rq->data_len) {
 772                         /*
 773                          * make sure space for the drain appears we
 774                          * know we can do this because max_hw_segments
 775                          * has been adjusted to be one fewer than the
 776                          * device can handle
 777                          */
 778                         rq->nr_phys_segments++;
 779                         rq->nr_hw_segments++;
 780                 }
 781
 782                 if (!q->prep_rq_fn)
 783                         break;
 784
 785                 ret = q->prep_rq_fn(q, rq);
 786                 if (ret == BLKPREP_OK) {
 787                         break;
 788                 } else if (ret == BLKPREP_DEFER) {
 789                         /*
 790                          * the request may have been (partially) prepped.
 791                          * we need to keep this request in the front to
 792                          * avoid resource deadlock.  REQ_STARTED will
 793                          * prevent other fs requests from passing this one.
 794                          */
 795                         if (q->dma_drain_size && rq->data_len &&
 796                             !(rq->cmd_flags & REQ_DONTPREP)) {
 797                                 /*
 798                                  * remove the space for the drain we added
 799                                  * so that we don't add it again
 800                                  */
 801                                 --rq->nr_phys_segments;
 802                                 --rq->nr_hw_segments;
 803                         }
 804
 805                         rq = NULL;
 806                         break;
 807                 } else if (ret == BLKPREP_KILL) {
 808                         rq->cmd_flags |= REQ_QUIET;
 809                         end_queued_request(rq, 0);
 810                 } else {
 811                         printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
 812                         break;
 813                 }
 814         }
 815
 816         return rq;
 817 }
 818 EXPORT_SYMBOL(elv_next_request);
 819
 820 void elv_dequeue_request(struct request_queue *q, struct request *rq)
 821 {
 822         BUG_ON(list_empty(&rq->queuelist));
 823         BUG_ON(ELV_ON_HASH(rq));
 824
 825         list_del_init(&rq->queuelist);
 826
 827         /*
 828          * the time frame between a request being removed from the lists
 829          * and to it is freed is accounted as io that is in progress at
 830          * the driver side.
 831          */
 832         if (blk_account_rq(rq))
 833                 q->in_flight++;
 834 }
 835 EXPORT_SYMBOL(elv_dequeue_request);
 836
 837 int elv_queue_empty(struct request_queue *q)
 838 {
 839         elevator_t *e = q->elevator;
 840
 841         if (!list_empty(&q->queue_head))
 842                 return 0;
 843
 844         if (e->ops->elevator_queue_empty_fn)
 845                 return e->ops->elevator_queue_empty_fn(q);
 846
 847         return 1;
 848 }
 849 EXPORT_SYMBOL(elv_queue_empty);
 850
 851 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 852 {
 853         elevator_t *e = q->elevator;
 854
 855         if (e->ops->elevator_latter_req_fn)
 856                 return e->ops->elevator_latter_req_fn(q, rq);
 857         return NULL;
 858 }
 859
 860 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 861 {
 862         elevator_t *e = q->elevator;
 863
 864         if (e->ops->elevator_former_req_fn)
 865                 return e->ops->elevator_former_req_fn(q, rq);
 866         return NULL;
 867 }
 868
 869 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 870 {
 871         elevator_t *e = q->elevator;
 872
 873         if (e->ops->elevator_set_req_fn)
 874                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 875
 876         rq->elevator_private = NULL;
 877         return 0;
 878 }
 879
 880 void elv_put_request(struct request_queue *q, struct request *rq)
 881 {
 882         elevator_t *e = q->elevator;
 883
 884         if (e->ops->elevator_put_req_fn)
 885                 e->ops->elevator_put_req_fn(rq);
 886 }
 887
 888 int elv_may_queue(struct request_queue *q, int rw)
 889 {
 890         elevator_t *e = q->elevator;
 891
 892         if (e->ops->elevator_may_queue_fn)
 893                 return e->ops->elevator_may_queue_fn(q, rw);
 894
 895         return ELV_MQUEUE_MAY;
 896 }
 897
 898 void elv_completed_request(struct request_queue *q, struct request *rq)
 899 {
 900         elevator_t *e = q->elevator;
 901
 902         /*
 903          * request is released from the driver, io must be done
 904          */
 905         if (blk_account_rq(rq)) {
 906                 q->in_flight--;
 907                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 908                         e->ops->elevator_completed_req_fn(q, rq);
 909         }
 910
 911         /*
 912          * Check if the queue is waiting for fs requests to be
 913          * drained for flush sequence.
 914          */
 915         if (unlikely(q->ordseq)) {
 916                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 917                 if (q->in_flight == 0 &&
 918                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 919                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 920                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 921                         q->request_fn(q);
 922                 }
 923         }
 924 }
 925
 926 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 927
 928 static ssize_t
 929 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 930 {
 931         elevator_t *e = container_of(kobj, elevator_t, kobj);
 932         struct elv_fs_entry *entry = to_elv(attr);
 933         ssize_t error;
 934
 935         if (!entry->show)
 936                 return -EIO;
 937
 938         mutex_lock(&e->sysfs_lock);
 939         error = e->ops ? entry->show(e, page) : -ENOENT;
 940         mutex_unlock(&e->sysfs_lock);
 941         return error;
 942 }
 943
 944 static ssize_t
 945 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 946                const char *page, size_t length)
 947 {
 948         elevator_t *e = container_of(kobj, elevator_t, kobj);
 949         struct elv_fs_entry *entry = to_elv(attr);
 950         ssize_t error;
 951
 952         if (!entry->store)
 953                 return -EIO;
 954
 955         mutex_lock(&e->sysfs_lock);
 956         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 957         mutex_unlock(&e->sysfs_lock);
 958         return error;
 959 }
 960
 961 static struct sysfs_ops elv_sysfs_ops = {
 962         .show   = elv_attr_show,
 963         .store  = elv_attr_store,
 964 };
 965
 966 static struct kobj_type elv_ktype = {
 967         .sysfs_ops      = &elv_sysfs_ops,
 968         .release        = elevator_release,
 969 };
 970
 971 int elv_register_queue(struct request_queue *q)
 972 {
 973         elevator_t *e = q->elevator;
 974         int error;
 975
 976         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 977         if (!error) {
 978                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 979                 if (attr) {
 980                         while (attr->attr.name) {
 981                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 982                                         break;
 983                                 attr++;
 984                         }
 985                 }
 986                 kobject_uevent(&e->kobj, KOBJ_ADD);
 987         }
 988         return error;
 989 }
 990
 991 static void __elv_unregister_queue(elevator_t *e)
 992 {
 993         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 994         kobject_del(&e->kobj);
 995 }
 996
 997 void elv_unregister_queue(struct request_queue *q)
 998 {
 999         if (q)
1000                 __elv_unregister_queue(q->elevator);
1001 }
1002
1003 void elv_register(struct elevator_type *e)
1004 {
1005         char *def = "";
1006
1007         spin_lock(&elv_list_lock);
1008         BUG_ON(elevator_find(e->elevator_name));
1009         list_add_tail(&e->list, &elv_list);
1010         spin_unlock(&elv_list_lock);
1011
1012         if (!strcmp(e->elevator_name, chosen_elevator) ||
1013                         (!*chosen_elevator &&
1014                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
1015                                 def = " (default)";
1016
1017         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
1018                                                                 def);
1019 }
1020 EXPORT_SYMBOL_GPL(elv_register);
1021
1022 void elv_unregister(struct elevator_type *e)
1023 {
1024         struct task_struct *g, *p;
1025
1026         /*
1027          * Iterate every thread in the process to remove the io contexts.
1028          */
1029         if (e->ops.trim) {
1030                 read_lock(&tasklist_lock);
1031                 do_each_thread(g, p) {
1032                         task_lock(p);
1033                         if (p->io_context)
1034                                 e->ops.trim(p->io_context);
1035                         task_unlock(p);
1036                 } while_each_thread(g, p);
1037                 read_unlock(&tasklist_lock);
1038         }
1039
1040         spin_lock(&elv_list_lock);
1041         list_del_init(&e->list);
1042         spin_unlock(&elv_list_lock);
1043 }
1044 EXPORT_SYMBOL_GPL(elv_unregister);
1045
1046 /*
1047  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1048  * we don't free the old io scheduler, before we have allocated what we
1049  * need for the new one. this way we have a chance of going back to the old
1050  * one, if the new one fails init for some reason.
1051  */
1052 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1053 {
1054         elevator_t *old_elevator, *e;
1055         void *data;
1056
1057         /*
1058          * Allocate new elevator
1059          */
1060         e = elevator_alloc(q, new_e);
1061         if (!e)
1062                 return 0;
1063
1064         data = elevator_init_queue(q, e);
1065         if (!data) {
1066                 kobject_put(&e->kobj);
1067                 return 0;
1068         }
1069
1070         /*
1071          * Turn on BYPASS and drain all requests w/ elevator private data
1072          */
1073         spin_lock_irq(q->queue_lock);
1074
1075         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
1076
1077         elv_drain_elevator(q);
1078
1079         while (q->rq.elvpriv) {
1080                 blk_remove_plug(q);
1081                 q->request_fn(q);
1082                 spin_unlock_irq(q->queue_lock);
1083                 msleep(10);
1084                 spin_lock_irq(q->queue_lock);
1085                 elv_drain_elevator(q);
1086         }
1087
1088         /*
1089          * Remember old elevator.
1090          */
1091         old_elevator = q->elevator;
1092
1093         /*
1094          * attach and start new elevator
1095          */
1096         elevator_attach(q, e, data);
1097
1098         spin_unlock_irq(q->queue_lock);
1099
1100         __elv_unregister_queue(old_elevator);
1101
1102         if (elv_register_queue(q))
1103                 goto fail_register;
1104
1105         /*
1106          * finally exit old elevator and turn off BYPASS.
1107          */
1108         elevator_exit(old_elevator);
1109         spin_lock_irq(q->queue_lock);
1110         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1111         spin_unlock_irq(q->queue_lock);
1112
1113         return 1;
1114
1115 fail_register:
1116         /*
1117          * switch failed, exit the new io scheduler and reattach the old
1118          * one again (along with re-adding the sysfs dir)
1119          */
1120         elevator_exit(e);
1121         q->elevator = old_elevator;
1122         elv_register_queue(q);
1123
1124         spin_lock_irq(q->queue_lock);
1125         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1126         spin_unlock_irq(q->queue_lock);
1127
1128         return 0;
1129 }
1130
1131 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1132                           size_t count)
1133 {
1134         char elevator_name[ELV_NAME_MAX];
1135         size_t len;
1136         struct elevator_type *e;
1137
1138         elevator_name[sizeof(elevator_name) - 1] = '\0';
1139         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
1140         len = strlen(elevator_name);
1141
1142         if (len && elevator_name[len - 1] == '\n')
1143                 elevator_name[len - 1] = '\0';
1144
1145         e = elevator_get(elevator_name);
1146         if (!e) {
1147                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1148                 return -EINVAL;
1149         }
1150
1151         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1152                 elevator_put(e);
1153                 return count;
1154         }
1155
1156         if (!elevator_switch(q, e))
1157                 printk(KERN_ERR "elevator: switch to %s failed\n",
1158                                                         elevator_name);
1159         return count;
1160 }
1161
1162 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1163 {
1164         elevator_t *e = q->elevator;
1165         struct elevator_type *elv = e->elevator_type;
1166         struct elevator_type *__e;
1167         int len = 0;
1168
1169         spin_lock(&elv_list_lock);
1170         list_for_each_entry(__e, &elv_list, list) {
1171                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1172                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1173                 else
1174                         len += sprintf(name+len, "%s ", __e->elevator_name);
1175         }
1176         spin_unlock(&elv_list_lock);
1177
1178         len += sprintf(len+name, "\n");
1179         return len;
1180 }
1181
1182 struct request *elv_rb_former_request(struct request_queue *q,
1183                                       struct request *rq)
1184 {
1185         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1186
1187         if (rbprev)
1188                 return rb_entry_rq(rbprev);
1189
1190         return NULL;
1191 }
1192 EXPORT_SYMBOL(elv_rb_former_request);
1193
1194 struct request *elv_rb_latter_request(struct request_queue *q,
1195                                       struct request *rq)
1196 {
1197         struct rb_node *rbnext = rb_next(&rq->rb_node);
1198
1199         if (rbnext)
1200                 return rb_entry_rq(rbnext);
1201
1202         return NULL;
1203 }
1204 EXPORT_SYMBOL(elv_rb_latter_request);