drivers/block/elevator.c

   1 /*
   2  *  linux/drivers/block/elevator.c
   3  *
   4  *  Block device elevator/IO-scheduler.
   5  *
   6  *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
   7  *
   8  * 30042000 Jens Axboe <axboe@suse.de> :
   9  *
  10  * Split the elevator a bit so that it is possible to choose a different
  11  * one or even write a new "plug in". There are three pieces:
  12  * - elevator_fn, inserts a new request in the queue list
  13  * - elevator_merge_fn, decides whether a new buffer can be merged with
  14  *   an existing request
  15  * - elevator_dequeue_fn, called when a request is taken off the active list
  16  *
  17  * 20082000 Dave Jones <davej@suse.de> :
  18  * Removed tests for max-bomb-segments, which was breaking elvtune
  19  *  when run without -bN
  20  *
  21  * Jens:
  22  * - Rework again to work with bio instead of buffer_heads
  23  * - loose bi_dev comparisons, partition handling is right now
  24  * - completely modularize elevator setup and teardown
  25  *
  26  */
  27 #include <linux/kernel.h>
  28 #include <linux/fs.h>
  29 #include <linux/blkdev.h>
  30 #include <linux/elevator.h>
  31 #include <linux/bio.h>
  32 #include <linux/config.h>
  33 #include <linux/module.h>
  34 #include <linux/slab.h>
  35 #include <linux/init.h>
  36 #include <linux/compiler.h>
  37
  38 #include <asm/uaccess.h>
  39
  40 /*
  41  * can we safely merge with this request?
  42  */
  43 inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  44 {
  45         if (!rq_mergeable(rq))
  46                 return 0;
  47
  48         /*
  49          * different data direction or already started, don't merge
  50          */
  51         if (bio_data_dir(bio) != rq_data_dir(rq))
  52                 return 0;
  53
  54         /*
  55          * same device and no special stuff set, merge is ok
  56          */
  57         if (rq->rq_disk == bio->bi_bdev->bd_disk &&
  58             !rq->waiting && !rq->special)
  59                 return 1;
  60
  61         return 0;
  62 }
  63
  64 inline int elv_try_merge(struct request *__rq, struct bio *bio)
  65 {
  66         int ret = ELEVATOR_NO_MERGE;
  67
  68         /*
  69          * we can merge and sequence is ok, check if it's possible
  70          */
  71         if (elv_rq_merge_ok(__rq, bio)) {
  72                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
  73                         ret = ELEVATOR_BACK_MERGE;
  74                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
  75                         ret = ELEVATOR_FRONT_MERGE;
  76         }
  77
  78         return ret;
  79 }
  80
  81 inline int elv_try_last_merge(request_queue_t *q, struct bio *bio)
  82 {
  83         if (q->last_merge)
  84                 return elv_try_merge(q->last_merge, bio);
  85
  86         return ELEVATOR_NO_MERGE;
  87 }
  88
  89 /*
  90  * general block -> elevator interface starts here
  91  */
  92 int elevator_init(request_queue_t *q, elevator_t *type)
  93 {
  94         elevator_t *e = &q->elevator;
  95
  96         memcpy(e, type, sizeof(*e));
  97
  98         INIT_LIST_HEAD(&q->queue_head);
  99         q->last_merge = NULL;
 100
 101         if (e->elevator_init_fn)
 102                 return e->elevator_init_fn(q, e);
 103
 104         return 0;
 105 }
 106
 107 void elevator_exit(request_queue_t *q)
 108 {
 109         elevator_t *e = &q->elevator;
 110
 111         if (e->elevator_exit_fn)
 112                 e->elevator_exit_fn(q, e);
 113 }
 114
 115 int elevator_global_init(void)
 116 {
 117         return 0;
 118 }
 119
 120 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
 121 {
 122         elevator_t *e = &q->elevator;
 123
 124         if (e->elevator_merge_fn)
 125                 return e->elevator_merge_fn(q, req, bio);
 126
 127         return ELEVATOR_NO_MERGE;
 128 }
 129
 130 void elv_merged_request(request_queue_t *q, struct request *rq)
 131 {
 132         elevator_t *e = &q->elevator;
 133
 134         if (e->elevator_merged_fn)
 135                 e->elevator_merged_fn(q, rq);
 136 }
 137
 138 void elv_merge_requests(request_queue_t *q, struct request *rq,
 139                              struct request *next)
 140 {
 141         elevator_t *e = &q->elevator;
 142
 143         if (q->last_merge == next)
 144                 q->last_merge = NULL;
 145
 146         if (e->elevator_merge_req_fn)
 147                 e->elevator_merge_req_fn(q, rq, next);
 148 }
 149
 150 void elv_requeue_request(request_queue_t *q, struct request *rq)
 151 {
 152         /*
 153          * it already went through dequeue, we need to decrement the
 154          * in_flight count again
 155          */
 156         if (blk_account_rq(rq))
 157                 q->in_flight--;
 158
 159         /*
 160          * if iosched has an explicit requeue hook, then use that. otherwise
 161          * just put the request at the front of the queue
 162          */
 163         if (q->elevator.elevator_requeue_req_fn)
 164                 q->elevator.elevator_requeue_req_fn(q, rq);
 165         else
 166                 __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
 167 }
 168
 169 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
 170                        int plug)
 171 {
 172         /*
 173          * barriers implicitly indicate back insertion
 174          */
 175         if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER) &&
 176             where == ELEVATOR_INSERT_SORT)
 177                 where = ELEVATOR_INSERT_BACK;
 178
 179         if (plug)
 180                 blk_plug_device(q);
 181
 182         rq->q = q;
 183         q->elevator.elevator_add_req_fn(q, rq, where);
 184
 185         if (blk_queue_plugged(q)) {
 186                 int nrq = q->rq.count[READ] + q->rq.count[WRITE] - q->in_flight;
 187
 188                 if (nrq == q->unplug_thresh)
 189                         __generic_unplug_device(q);
 190         }
 191
 192 }
 193
 194 void elv_add_request(request_queue_t *q, struct request *rq, int where,
 195                      int plug)
 196 {
 197         unsigned long flags;
 198
 199         spin_lock_irqsave(q->queue_lock, flags);
 200         __elv_add_request(q, rq, where, plug);
 201         spin_unlock_irqrestore(q->queue_lock, flags);
 202 }
 203
 204 static inline struct request *__elv_next_request(request_queue_t *q)
 205 {
 206         return q->elevator.elevator_next_req_fn(q);
 207 }
 208
 209 struct request *elv_next_request(request_queue_t *q)
 210 {
 211         struct request *rq;
 212         int ret;
 213
 214         while ((rq = __elv_next_request(q)) != NULL) {
 215                 /*
 216                  * just mark as started even if we don't start it, a request
 217                  * that has been delayed should not be passed by new incoming
 218                  * requests
 219                  */
 220                 rq->flags |= REQ_STARTED;
 221
 222                 if (rq == q->last_merge)
 223                         q->last_merge = NULL;
 224
 225                 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
 226                         break;
 227
 228                 ret = q->prep_rq_fn(q, rq);
 229                 if (ret == BLKPREP_OK) {
 230                         break;
 231                 } else if (ret == BLKPREP_DEFER) {
 232                         rq = NULL;
 233                         break;
 234                 } else if (ret == BLKPREP_KILL) {
 235                         int nr_bytes = rq->hard_nr_sectors << 9;
 236
 237                         if (!nr_bytes)
 238                                 nr_bytes = rq->data_len;
 239
 240                         blkdev_dequeue_request(rq);
 241                         rq->flags |= REQ_QUIET;
 242                         end_that_request_chunk(rq, 0, nr_bytes);
 243                         end_that_request_last(rq);
 244                 } else {
 245                         printk("%s: bad return=%d\n", __FUNCTION__, ret);
 246                         break;
 247                 }
 248         }
 249
 250         return rq;
 251 }
 252
 253 void elv_remove_request(request_queue_t *q, struct request *rq)
 254 {
 255         elevator_t *e = &q->elevator;
 256
 257         /*
 258          * the time frame between a request being removed from the lists
 259          * and to it is freed is accounted as io that is in progress at
 260          * the driver side. note that we only account requests that the
 261          * driver has seen (REQ_STARTED set), to avoid false accounting
 262          * for request-request merges
 263          */
 264         if (blk_account_rq(rq))
 265                 q->in_flight++;
 266
 267         /*
 268          * the main clearing point for q->last_merge is on retrieval of
 269          * request by driver (it calls elv_next_request()), but it _can_
 270          * also happen here if a request is added to the queue but later
 271          * deleted without ever being given to driver (merged with another
 272          * request).
 273          */
 274         if (rq == q->last_merge)
 275                 q->last_merge = NULL;
 276
 277         if (e->elevator_remove_req_fn)
 278                 e->elevator_remove_req_fn(q, rq);
 279 }
 280
 281 int elv_queue_empty(request_queue_t *q)
 282 {
 283         elevator_t *e = &q->elevator;
 284
 285         if (e->elevator_queue_empty_fn)
 286                 return e->elevator_queue_empty_fn(q);
 287
 288         return list_empty(&q->queue_head);
 289 }
 290
 291 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
 292 {
 293         struct list_head *next;
 294
 295         elevator_t *e = &q->elevator;
 296
 297         if (e->elevator_latter_req_fn)
 298                 return e->elevator_latter_req_fn(q, rq);
 299
 300         next = rq->queuelist.next;
 301         if (next != &q->queue_head && next != &rq->queuelist)
 302                 return list_entry_rq(next);
 303
 304         return NULL;
 305 }
 306
 307 struct request *elv_former_request(request_queue_t *q, struct request *rq)
 308 {
 309         struct list_head *prev;
 310
 311         elevator_t *e = &q->elevator;
 312
 313         if (e->elevator_former_req_fn)
 314                 return e->elevator_former_req_fn(q, rq);
 315
 316         prev = rq->queuelist.prev;
 317         if (prev != &q->queue_head && prev != &rq->queuelist)
 318                 return list_entry_rq(prev);
 319
 320         return NULL;
 321 }
 322
 323 int elv_set_request(request_queue_t *q, struct request *rq, int gfp_mask)
 324 {
 325         elevator_t *e = &q->elevator;
 326
 327         if (e->elevator_set_req_fn)
 328                 return e->elevator_set_req_fn(q, rq, gfp_mask);
 329
 330         rq->elevator_private = NULL;
 331         return 0;
 332 }
 333
 334 void elv_put_request(request_queue_t *q, struct request *rq)
 335 {
 336         elevator_t *e = &q->elevator;
 337
 338         if (e->elevator_put_req_fn)
 339                 e->elevator_put_req_fn(q, rq);
 340 }
 341
 342 int elv_may_queue(request_queue_t *q, int rw)
 343 {
 344         elevator_t *e = &q->elevator;
 345
 346         if (e->elevator_may_queue_fn)
 347                 return e->elevator_may_queue_fn(q, rw);
 348
 349         return 0;
 350 }
 351
 352 void elv_completed_request(request_queue_t *q, struct request *rq)
 353 {
 354         elevator_t *e = &q->elevator;
 355
 356         /*
 357          * request is released from the driver, io must be done
 358          */
 359         if (blk_account_rq(rq))
 360                 q->in_flight--;
 361
 362         if (e->elevator_completed_req_fn)
 363                 e->elevator_completed_req_fn(q, rq);
 364 }
 365
 366 int elv_register_queue(struct request_queue *q)
 367 {
 368         elevator_t *e;
 369
 370         e = &q->elevator;
 371
 372         e->kobj.parent = kobject_get(&q->kobj);
 373         if (!e->kobj.parent)
 374                 return -EBUSY;
 375
 376         snprintf(e->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
 377         e->kobj.ktype = e->elevator_ktype;
 378
 379         return kobject_register(&e->kobj);
 380 }
 381
 382 void elv_unregister_queue(struct request_queue *q)
 383 {
 384         if (q) {
 385                 elevator_t * e = &q->elevator;
 386                 kobject_unregister(&e->kobj);
 387                 kobject_put(&q->kobj);
 388         }
 389 }
 390
 391 module_init(elevator_global_init);
 392
 393 EXPORT_SYMBOL(elv_add_request);
 394 EXPORT_SYMBOL(__elv_add_request);
 395 EXPORT_SYMBOL(elv_requeue_request);
 396 EXPORT_SYMBOL(elv_next_request);
 397 EXPORT_SYMBOL(elv_remove_request);
 398 EXPORT_SYMBOL(elv_queue_empty);
 399 EXPORT_SYMBOL(elv_completed_request);
 400 EXPORT_SYMBOL(elevator_exit);
 401 EXPORT_SYMBOL(elevator_init);