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         sector_t last_sector;           /* head position */
  43         unsigned int starved;           /* times reads have starved writes */
  44
  45         /*
  46          * settings that change how the i/o scheduler behaves
  47          */
  48         int fifo_expire[2];
  49         int fifo_batch;
  50         int writes_starved;
  51         int front_merges;
  52 };
  53
  54 static void deadline_move_request(struct deadline_data *, struct request *);
  55
  56 #define RQ_RB_ROOT(dd, rq)      (&(dd)->sort_list[rq_data_dir((rq))])
  57
  58 /*
  59  * get the request after `rq' in sector-sorted order
  60  */
  61 static inline struct request *
  62 deadline_latter_request(struct request *rq)
  63 {
  64         struct rb_node *node = rb_next(&rq->rb_node);
  65
  66         if (node)
  67                 return rb_entry_rq(node);
  68
  69         return NULL;
  70 }
  71
  72 static void
  73 deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
  74 {
  75         struct rb_root *root = RQ_RB_ROOT(dd, rq);
  76         struct request *__alias;
  77
  78 retry:
  79         __alias = elv_rb_add(root, rq);
  80         if (unlikely(__alias)) {
  81                 deadline_move_request(dd, __alias);
  82                 goto retry;
  83         }
  84 }
  85
  86 static inline void
  87 deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
  88 {
  89         const int data_dir = rq_data_dir(rq);
  90
  91         if (dd->next_rq[data_dir] == rq)
  92                 dd->next_rq[data_dir] = deadline_latter_request(rq);
  93
  94         elv_rb_del(RQ_RB_ROOT(dd, rq), rq);
  95 }
  96
  97 /*
  98  * add rq to rbtree and fifo
  99  */
 100 static void
 101 deadline_add_request(struct request_queue *q, struct request *rq)
 102 {
 103         struct deadline_data *dd = q->elevator->elevator_data;
 104         const int data_dir = rq_data_dir(rq);
 105
 106         deadline_add_rq_rb(dd, rq);
 107
 108         /*
 109          * set expire time (only used for reads) and add to fifo list
 110          */
 111         rq_set_fifo_time(rq, jiffies + dd->fifo_expire[data_dir]);
 112         list_add_tail(&rq->queuelist, &dd->fifo_list[data_dir]);
 113 }
 114
 115 /*
 116  * remove rq from rbtree and fifo.
 117  */
 118 static void deadline_remove_request(struct request_queue *q, struct request *rq)
 119 {
 120         struct deadline_data *dd = q->elevator->elevator_data;
 121
 122         rq_fifo_clear(rq);
 123         deadline_del_rq_rb(dd, rq);
 124 }
 125
 126 static int
 127 deadline_merge(struct request_queue *q, struct request **req, struct bio *bio)
 128 {
 129         struct deadline_data *dd = q->elevator->elevator_data;
 130         struct request *__rq;
 131         int ret;
 132
 133         /*
 134          * check for front merge
 135          */
 136         if (dd->front_merges) {
 137                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 138
 139                 __rq = elv_rb_find(&dd->sort_list[bio_data_dir(bio)], sector);
 140                 if (__rq) {
 141                         BUG_ON(sector != __rq->sector);
 142
 143                         if (elv_rq_merge_ok(__rq, bio)) {
 144                                 ret = ELEVATOR_FRONT_MERGE;
 145                                 goto out;
 146                         }
 147                 }
 148         }
 149
 150         return ELEVATOR_NO_MERGE;
 151 out:
 152         *req = __rq;
 153         return ret;
 154 }
 155
 156 static void deadline_merged_request(struct request_queue *q,
 157                                     struct request *req, int type)
 158 {
 159         struct deadline_data *dd = q->elevator->elevator_data;
 160
 161         /*
 162          * if the merge was a front merge, we need to reposition request
 163          */
 164         if (type == ELEVATOR_FRONT_MERGE) {
 165                 elv_rb_del(RQ_RB_ROOT(dd, req), req);
 166                 deadline_add_rq_rb(dd, req);
 167         }
 168 }
 169
 170 static void
 171 deadline_merged_requests(struct request_queue *q, struct request *req,
 172                          struct request *next)
 173 {
 174         /*
 175          * if next expires before rq, assign its expire time to rq
 176          * and move into next position (next will be deleted) in fifo
 177          */
 178         if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
 179                 if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
 180                         list_move(&req->queuelist, &next->queuelist);
 181                         rq_set_fifo_time(req, rq_fifo_time(next));
 182                 }
 183         }
 184
 185         /*
 186          * kill knowledge of next, this one is a goner
 187          */
 188         deadline_remove_request(q, next);
 189 }
 190
 191 /*
 192  * move request from sort list to dispatch queue.
 193  */
 194 static inline void
 195 deadline_move_to_dispatch(struct deadline_data *dd, struct request *rq)
 196 {
 197         struct request_queue *q = rq->q;
 198
 199         deadline_remove_request(q, rq);
 200         elv_dispatch_add_tail(q, rq);
 201 }
 202
 203 /*
 204  * move an entry to dispatch queue
 205  */
 206 static void
 207 deadline_move_request(struct deadline_data *dd, struct request *rq)
 208 {
 209         const int data_dir = rq_data_dir(rq);
 210
 211         dd->next_rq[READ] = NULL;
 212         dd->next_rq[WRITE] = NULL;
 213         dd->next_rq[data_dir] = deadline_latter_request(rq);
 214
 215         dd->last_sector = rq->sector + rq->nr_sectors;
 216
 217         /*
 218          * take it off the sort and fifo list, move
 219          * to dispatch queue
 220          */
 221         deadline_move_to_dispatch(dd, rq);
 222 }
 223
 224 /*
 225  * deadline_check_fifo returns 0 if there are no expired reads on the fifo,
 226  * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
 227  */
 228 static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
 229 {
 230         struct request *rq = rq_entry_fifo(dd->fifo_list[ddir].next);
 231
 232         /*
 233          * rq is expired!
 234          */
 235         if (time_after(jiffies, rq_fifo_time(rq)))
 236                 return 1;
 237
 238         return 0;
 239 }
 240
 241 /*
 242  * deadline_dispatch_requests selects the best request according to
 243  * read/write expire, fifo_batch, etc
 244  */
 245 static int deadline_dispatch_requests(struct request_queue *q, int force)
 246 {
 247         struct deadline_data *dd = q->elevator->elevator_data;
 248         const int reads = !list_empty(&dd->fifo_list[READ]);
 249         const int writes = !list_empty(&dd->fifo_list[WRITE]);
 250         struct request *rq;
 251         int data_dir;
 252
 253         /*
 254          * batches are currently reads XOR writes
 255          */
 256         if (dd->next_rq[WRITE])
 257                 rq = dd->next_rq[WRITE];
 258         else
 259                 rq = dd->next_rq[READ];
 260
 261         if (rq) {
 262                 /* we have a "next request" */
 263
 264                 if (dd->last_sector != rq->sector)
 265                         /* end the batch on a non sequential request */
 266                         dd->batching += dd->fifo_batch;
 267
 268                 if (dd->batching < dd->fifo_batch)
 269                         /* we are still entitled to batch */
 270                         goto dispatch_request;
 271         }
 272
 273         /*
 274          * at this point we are not running a batch. select the appropriate
 275          * data direction (read / write)
 276          */
 277
 278         if (reads) {
 279                 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[READ]));
 280
 281                 if (writes && (dd->starved++ >= dd->writes_starved))
 282                         goto dispatch_writes;
 283
 284                 data_dir = READ;
 285
 286                 goto dispatch_find_request;
 287         }
 288
 289         /*
 290          * there are either no reads or writes have been starved
 291          */
 292
 293         if (writes) {
 294 dispatch_writes:
 295                 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[WRITE]));
 296
 297                 dd->starved = 0;
 298
 299                 data_dir = WRITE;
 300
 301                 goto dispatch_find_request;
 302         }
 303
 304         return 0;
 305
 306 dispatch_find_request:
 307         /*
 308          * we are not running a batch, find best request for selected data_dir
 309          * and start a new batch
 310          */
 311         if (deadline_check_fifo(dd, data_dir)) {
 312                 /* An expired request exists - satisfy it */
 313                 rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
 314         } else if (dd->next_rq[data_dir]) {
 315                 /*
 316                  * The last req was the same dir and we have a next request in
 317                  * sort order. No expired requests so continue on from here.
 318                  */
 319                 rq = dd->next_rq[data_dir];
 320         } else {
 321                 struct rb_node *node;
 322                 /*
 323                  * The last req was the other direction or we have run out of
 324                  * higher-sectored requests. Go back to the lowest sectored
 325                  * request (1 way elevator) and start a new batch.
 326                  */
 327                 node = rb_first(&dd->sort_list[data_dir]);
 328                 if (node)
 329                         rq = rb_entry_rq(node);
 330         }
 331
 332         dd->batching = 0;
 333
 334 dispatch_request:
 335         /*
 336          * rq is the selected appropriate request.
 337          */
 338         dd->batching++;
 339         deadline_move_request(dd, rq);
 340
 341         return 1;
 342 }
 343
 344 static int deadline_queue_empty(struct request_queue *q)
 345 {
 346         struct deadline_data *dd = q->elevator->elevator_data;
 347
 348         return list_empty(&dd->fifo_list[WRITE])
 349                 && list_empty(&dd->fifo_list[READ]);
 350 }
 351
 352 static void deadline_exit_queue(elevator_t *e)
 353 {
 354         struct deadline_data *dd = e->elevator_data;
 355
 356         BUG_ON(!list_empty(&dd->fifo_list[READ]));
 357         BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
 358
 359         kfree(dd);
 360 }
 361
 362 /*
 363  * initialize elevator private data (deadline_data).
 364  */
 365 static void *deadline_init_queue(struct request_queue *q)
 366 {
 367         struct deadline_data *dd;
 368
 369         dd = kmalloc_node(sizeof(*dd), GFP_KERNEL | __GFP_ZERO, q->node);
 370         if (!dd)
 371                 return NULL;
 372
 373         INIT_LIST_HEAD(&dd->fifo_list[READ]);
 374         INIT_LIST_HEAD(&dd->fifo_list[WRITE]);
 375         dd->sort_list[READ] = RB_ROOT;
 376         dd->sort_list[WRITE] = RB_ROOT;
 377         dd->fifo_expire[READ] = read_expire;
 378         dd->fifo_expire[WRITE] = write_expire;
 379         dd->writes_starved = writes_starved;
 380         dd->front_merges = 1;
 381         dd->fifo_batch = fifo_batch;
 382         return dd;
 383 }
 384
 385 /*
 386  * sysfs parts below
 387  */
 388
 389 static ssize_t
 390 deadline_var_show(int var, char *page)
 391 {
 392         return sprintf(page, "%d\n", var);
 393 }
 394
 395 static ssize_t
 396 deadline_var_store(int *var, const char *page, size_t count)
 397 {
 398         char *p = (char *) page;
 399
 400         *var = simple_strtol(p, &p, 10);
 401         return count;
 402 }
 403
 404 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
 405 static ssize_t __FUNC(elevator_t *e, char *page)                        \
 406 {                                                                       \
 407         struct deadline_data *dd = e->elevator_data;                    \
 408         int __data = __VAR;                                             \
 409         if (__CONV)                                                     \
 410                 __data = jiffies_to_msecs(__data);                      \
 411         return deadline_var_show(__data, (page));                       \
 412 }
 413 SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
 414 SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
 415 SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
 416 SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
 417 SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
 418 #undef SHOW_FUNCTION
 419
 420 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
 421 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
 422 {                                                                       \
 423         struct deadline_data *dd = e->elevator_data;                    \
 424         int __data;                                                     \
 425         int ret = deadline_var_store(&__data, (page), count);           \
 426         if (__data < (MIN))                                             \
 427                 __data = (MIN);                                         \
 428         else if (__data > (MAX))                                        \
 429                 __data = (MAX);                                         \
 430         if (__CONV)                                                     \
 431                 *(__PTR) = msecs_to_jiffies(__data);                    \
 432         else                                                            \
 433                 *(__PTR) = __data;                                      \
 434         return ret;                                                     \
 435 }
 436 STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
 437 STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
 438 STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
 439 STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
 440 STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
 441 #undef STORE_FUNCTION
 442
 443 #define DD_ATTR(name) \
 444         __ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
 445                                       deadline_##name##_store)
 446
 447 static struct elv_fs_entry deadline_attrs[] = {
 448         DD_ATTR(read_expire),
 449         DD_ATTR(write_expire),
 450         DD_ATTR(writes_starved),
 451         DD_ATTR(front_merges),
 452         DD_ATTR(fifo_batch),
 453         __ATTR_NULL
 454 };
 455
 456 static struct elevator_type iosched_deadline = {
 457         .ops = {
 458                 .elevator_merge_fn =            deadline_merge,
 459                 .elevator_merged_fn =           deadline_merged_request,
 460                 .elevator_merge_req_fn =        deadline_merged_requests,
 461                 .elevator_dispatch_fn =         deadline_dispatch_requests,
 462                 .elevator_add_req_fn =          deadline_add_request,
 463                 .elevator_queue_empty_fn =      deadline_queue_empty,
 464                 .elevator_former_req_fn =       elv_rb_former_request,
 465                 .elevator_latter_req_fn =       elv_rb_latter_request,
 466                 .elevator_init_fn =             deadline_init_queue,
 467                 .elevator_exit_fn =             deadline_exit_queue,
 468         },
 469
 470         .elevator_attrs = deadline_attrs,
 471         .elevator_name = "deadline",
 472         .elevator_owner = THIS_MODULE,
 473 };
 474
 475 static int __init deadline_init(void)
 476 {
 477         return elv_register(&iosched_deadline);
 478 }
 479
 480 static void __exit deadline_exit(void)
 481 {
 482         elv_unregister(&iosched_deadline);
 483 }
 484
 485 module_init(deadline_init);
 486 module_exit(deadline_exit);
 487
 488 MODULE_AUTHOR("Jens Axboe");
 489 MODULE_LICENSE("GPL");
 490 MODULE_DESCRIPTION("deadline IO scheduler");