block/deadline-iosched.c

   1 /*
   2  *  Deadline i/o scheduler.
   3  *
   4  *  Copyright (C) 2002 Jens Axboe <axboe@kernel.dk>
   5  */
   6 #include <linux/kernel.h>
   7 #include <linux/fs.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/elevator.h>
  10 #include <linux/bio.h>
  11 #include <linux/module.h>
  12 #include <linux/slab.h>
  13 #include <linux/init.h>
  14 #include <linux/compiler.h>
  15 #include <linux/rbtree.h>
  16
  17 /*
  18  * See Documentation/block/deadline-iosched.txt
  19  */
  20 static const int read_expire = HZ / 2;  /* max time before a read is submitted. */
  21 static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
  22 static const int writes_starved = 2;    /* max times reads can starve a write */
  23 static const int fifo_batch = 16;       /* # of sequential requests treated as one
  24                                      by the above parameters. For throughput. */
  25
  26 struct deadline_data {
  27         /*
  28          * run time data
  29          */
  30
  31         /*
  32          * requests (deadline_rq s) are present on both sort_list and fifo_list
  33          */
  34         struct rb_root sort_list[2];
  35         struct list_head fifo_list[2];
  36
  37         /*
  38          * next in sort order. read, write or both are NULL
  39          */
  40         struct request *next_rq[2];
  41         unsigned int batching;          /* number of sequential requests made */
  42         unsigned int starved;           /* times reads have starved writes */
  43
  44         /*
  45          * settings that change how the i/o scheduler behaves
  46          */
  47         int fifo_expire[2];
  48         int fifo_batch;
  49         int writes_starved;
  50         int front_merges;
  51 };
  52
  53 static void deadline_move_request(struct deadline_data *, struct request *);
  54
  55 static inline struct rb_root *
  56 deadline_rb_root(struct deadline_data *dd, struct request *rq)
  57 {
  58         return &dd->sort_list[rq_data_dir(rq)];
  59 }
  60
  61 /*
  62  * get the request after `rq' in sector-sorted order
  63  */
  64 static inline struct request *
  65 deadline_latter_request(struct request *rq)
  66 {
  67         struct rb_node *node = rb_next(&rq->rb_node);
  68
  69         if (node)
  70                 return rb_entry_rq(node);
  71
  72         return NULL;
  73 }
  74
  75 static void
  76 deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
  77 {
  78         struct rb_root *root = deadline_rb_root(dd, rq);
  79
  80         elv_rb_add(root, rq);
  81 }
  82
  83 static inline void
  84 deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
  85 {
  86         const int data_dir = rq_data_dir(rq);
  87
  88         if (dd->next_rq[data_dir] == rq)
  89                 dd->next_rq[data_dir] = deadline_latter_request(rq);
  90
  91         elv_rb_del(deadline_rb_root(dd, rq), rq);
  92 }
  93
  94 /*
  95  * add rq to rbtree and fifo
  96  */
  97 static void
  98 deadline_add_request(struct request_queue *q, struct request *rq)
  99 {
 100         struct deadline_data *dd = q->elevator->elevator_data;
 101         const int data_dir = rq_data_dir(rq);
 102
 103         deadline_add_rq_rb(dd, rq);
 104
 105         /*
 106          * set expire time and add to fifo list
 107          */
 108         rq->fifo_time = jiffies + dd->fifo_expire[data_dir];
 109         list_add_tail(&rq->queuelist, &dd->fifo_list[data_dir]);
 110 }
 111
 112 /*
 113  * remove rq from rbtree and fifo.
 114  */
 115 static void deadline_remove_request(struct request_queue *q, struct request *rq)
 116 {
 117         struct deadline_data *dd = q->elevator->elevator_data;
 118
 119         rq_fifo_clear(rq);
 120         deadline_del_rq_rb(dd, rq);
 121 }
 122
 123 static int
 124 deadline_merge(struct request_queue *q, struct request **req, struct bio *bio)
 125 {
 126         struct deadline_data *dd = q->elevator->elevator_data;
 127         struct request *__rq;
 128         int ret;
 129
 130         /*
 131          * check for front merge
 132          */
 133         if (dd->front_merges) {
 134                 sector_t sector = bio_end_sector(bio);
 135
 136                 __rq = elv_rb_find(&dd->sort_list[bio_data_dir(bio)], sector);
 137                 if (__rq) {
 138                         BUG_ON(sector != blk_rq_pos(__rq));
 139
 140                         if (elv_rq_merge_ok(__rq, bio)) {
 141                                 ret = ELEVATOR_FRONT_MERGE;
 142                                 goto out;
 143                         }
 144                 }
 145         }
 146
 147         return ELEVATOR_NO_MERGE;
 148 out:
 149         *req = __rq;
 150         return ret;
 151 }
 152
 153 static void deadline_merged_request(struct request_queue *q,
 154                                     struct request *req, int type)
 155 {
 156         struct deadline_data *dd = q->elevator->elevator_data;
 157
 158         /*
 159          * if the merge was a front merge, we need to reposition request
 160          */
 161         if (type == ELEVATOR_FRONT_MERGE) {
 162                 elv_rb_del(deadline_rb_root(dd, req), req);
 163                 deadline_add_rq_rb(dd, req);
 164         }
 165 }
 166
 167 static void
 168 deadline_merged_requests(struct request_queue *q, struct request *req,
 169                          struct request *next)
 170 {
 171         /*
 172          * if next expires before rq, assign its expire time to rq
 173          * and move into next position (next will be deleted) in fifo
 174          */
 175         if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
 176                 if (time_before(next->fifo_time, req->fifo_time)) {
 177                         list_move(&req->queuelist, &next->queuelist);
 178                         req->fifo_time = next->fifo_time;
 179                 }
 180         }
 181
 182         /*
 183          * kill knowledge of next, this one is a goner
 184          */
 185         deadline_remove_request(q, next);
 186 }
 187
 188 /*
 189  * move request from sort list to dispatch queue.
 190  */
 191 static inline void
 192 deadline_move_to_dispatch(struct deadline_data *dd, struct request *rq)
 193 {
 194         struct request_queue *q = rq->q;
 195
 196         deadline_remove_request(q, rq);
 197         elv_dispatch_add_tail(q, rq);
 198 }
 199
 200 /*
 201  * move an entry to dispatch queue
 202  */
 203 static void
 204 deadline_move_request(struct deadline_data *dd, struct request *rq)
 205 {
 206         const int data_dir = rq_data_dir(rq);
 207
 208         dd->next_rq[READ] = NULL;
 209         dd->next_rq[WRITE] = NULL;
 210         dd->next_rq[data_dir] = deadline_latter_request(rq);
 211
 212         /*
 213          * take it off the sort and fifo list, move
 214          * to dispatch queue
 215          */
 216         deadline_move_to_dispatch(dd, rq);
 217 }
 218
 219 /*
 220  * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
 221  * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
 222  */
 223 static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
 224 {
 225         struct request *rq = rq_entry_fifo(dd->fifo_list[ddir].next);
 226
 227         /*
 228          * rq is expired!
 229          */
 230         if (time_after_eq(jiffies, rq->fifo_time))
 231                 return 1;
 232
 233         return 0;
 234 }
 235
 236 /*
 237  * deadline_dispatch_requests selects the best request according to
 238  * read/write expire, fifo_batch, etc
 239  */
 240 static int deadline_dispatch_requests(struct request_queue *q, int force)
 241 {
 242         struct deadline_data *dd = q->elevator->elevator_data;
 243         const int reads = !list_empty(&dd->fifo_list[READ]);
 244         const int writes = !list_empty(&dd->fifo_list[WRITE]);
 245         struct request *rq;
 246         int data_dir;
 247
 248         /*
 249          * batches are currently reads XOR writes
 250          */
 251         if (dd->next_rq[WRITE])
 252                 rq = dd->next_rq[WRITE];
 253         else
 254                 rq = dd->next_rq[READ];
 255
 256         if (rq && dd->batching < dd->fifo_batch)
 257                 /* we have a next request are still entitled to batch */
 258                 goto dispatch_request;
 259
 260         /*
 261          * at this point we are not running a batch. select the appropriate
 262          * data direction (read / write)
 263          */
 264
 265         if (reads) {
 266                 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[READ]));
 267
 268                 if (writes && (dd->starved++ >= dd->writes_starved))
 269                         goto dispatch_writes;
 270
 271                 data_dir = READ;
 272
 273                 goto dispatch_find_request;
 274         }
 275
 276         /*
 277          * there are either no reads or writes have been starved
 278          */
 279
 280         if (writes) {
 281 dispatch_writes:
 282                 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[WRITE]));
 283
 284                 dd->starved = 0;
 285
 286                 data_dir = WRITE;
 287
 288                 goto dispatch_find_request;
 289         }
 290
 291         return 0;
 292
 293 dispatch_find_request:
 294         /*
 295          * we are not running a batch, find best request for selected data_dir
 296          */
 297         if (deadline_check_fifo(dd, data_dir) || !dd->next_rq[data_dir]) {
 298                 /*
 299                  * A deadline has expired, the last request was in the other
 300                  * direction, or we have run out of higher-sectored requests.
 301                  * Start again from the request with the earliest expiry time.
 302                  */
 303                 rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
 304         } else {
 305                 /*
 306                  * The last req was the same dir and we have a next request in
 307                  * sort order. No expired requests so continue on from here.
 308                  */
 309                 rq = dd->next_rq[data_dir];
 310         }
 311
 312         dd->batching = 0;
 313
 314 dispatch_request:
 315         /*
 316          * rq is the selected appropriate request.
 317          */
 318         dd->batching++;
 319         deadline_move_request(dd, rq);
 320
 321         return 1;
 322 }
 323
 324 static void deadline_exit_queue(struct elevator_queue *e)
 325 {
 326         struct deadline_data *dd = e->elevator_data;
 327
 328         BUG_ON(!list_empty(&dd->fifo_list[READ]));
 329         BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
 330
 331         kfree(dd);
 332 }
 333
 334 /*
 335  * initialize elevator private data (deadline_data).
 336  */
 337 static int deadline_init_queue(struct request_queue *q, struct elevator_type *e)
 338 {
 339         struct deadline_data *dd;
 340         struct elevator_queue *eq;
 341
 342         eq = elevator_alloc(q, e);
 343         if (!eq)
 344                 return -ENOMEM;
 345
 346         dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
 347         if (!dd) {
 348                 kobject_put(&eq->kobj);
 349                 return -ENOMEM;
 350         }
 351         eq->elevator_data = dd;
 352
 353         INIT_LIST_HEAD(&dd->fifo_list[READ]);
 354         INIT_LIST_HEAD(&dd->fifo_list[WRITE]);
 355         dd->sort_list[READ] = RB_ROOT;
 356         dd->sort_list[WRITE] = RB_ROOT;
 357         dd->fifo_expire[READ] = read_expire;
 358         dd->fifo_expire[WRITE] = write_expire;
 359         dd->writes_starved = writes_starved;
 360         dd->front_merges = 1;
 361         dd->fifo_batch = fifo_batch;
 362
 363         spin_lock_irq(q->queue_lock);
 364         q->elevator = eq;
 365         spin_unlock_irq(q->queue_lock);
 366         return 0;
 367 }
 368
 369 /*
 370  * sysfs parts below
 371  */
 372
 373 static ssize_t
 374 deadline_var_show(int var, char *page)
 375 {
 376         return sprintf(page, "%d\n", var);
 377 }
 378
 379 static ssize_t
 380 deadline_var_store(int *var, const char *page, size_t count)
 381 {
 382         char *p = (char *) page;
 383
 384         *var = simple_strtol(p, &p, 10);
 385         return count;
 386 }
 387
 388 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
 389 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
 390 {                                                                       \
 391         struct deadline_data *dd = e->elevator_data;                    \
 392         int __data = __VAR;                                             \
 393         if (__CONV)                                                     \
 394                 __data = jiffies_to_msecs(__data);                      \
 395         return deadline_var_show(__data, (page));                       \
 396 }
 397 SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
 398 SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
 399 SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
 400 SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
 401 SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
 402 #undef SHOW_FUNCTION
 403
 404 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
 405 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
 406 {                                                                       \
 407         struct deadline_data *dd = e->elevator_data;                    \
 408         int __data;                                                     \
 409         int ret = deadline_var_store(&__data, (page), count);           \
 410         if (__data < (MIN))                                             \
 411                 __data = (MIN);                                         \
 412         else if (__data > (MAX))                                        \
 413                 __data = (MAX);                                         \
 414         if (__CONV)                                                     \
 415                 *(__PTR) = msecs_to_jiffies(__data);                    \
 416         else                                                            \
 417                 *(__PTR) = __data;                                      \
 418         return ret;                                                     \
 419 }
 420 STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
 421 STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
 422 STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
 423 STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
 424 STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
 425 #undef STORE_FUNCTION
 426
 427 #define DD_ATTR(name) \
 428         __ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
 429                                       deadline_##name##_store)
 430
 431 static struct elv_fs_entry deadline_attrs[] = {
 432         DD_ATTR(read_expire),
 433         DD_ATTR(write_expire),
 434         DD_ATTR(writes_starved),
 435         DD_ATTR(front_merges),
 436         DD_ATTR(fifo_batch),
 437         __ATTR_NULL
 438 };
 439
 440 static struct elevator_type iosched_deadline = {
 441         .ops = {
 442                 .elevator_merge_fn =            deadline_merge,
 443                 .elevator_merged_fn =           deadline_merged_request,
 444                 .elevator_merge_req_fn =        deadline_merged_requests,
 445                 .elevator_dispatch_fn =         deadline_dispatch_requests,
 446                 .elevator_add_req_fn =          deadline_add_request,
 447                 .elevator_former_req_fn =       elv_rb_former_request,
 448                 .elevator_latter_req_fn =       elv_rb_latter_request,
 449                 .elevator_init_fn =             deadline_init_queue,
 450                 .elevator_exit_fn =             deadline_exit_queue,
 451         },
 452
 453         .elevator_attrs = deadline_attrs,
 454         .elevator_name = "deadline",
 455         .elevator_owner = THIS_MODULE,
 456 };
 457
 458 static int __init deadline_init(void)
 459 {
 460         return elv_register(&iosched_deadline);
 461 }
 462
 463 static void __exit deadline_exit(void)
 464 {
 465         elv_unregister(&iosched_deadline);
 466 }
 467
 468 module_init(deadline_init);
 469 module_exit(deadline_exit);
 470
 471 MODULE_AUTHOR("Jens Axboe");
 472 MODULE_LICENSE("GPL");
 473 MODULE_DESCRIPTION("deadline IO scheduler");