| 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/blktrace_api.h> |
| 35 | #include <linux/hash.h> |
| 36 | #include <linux/uaccess.h> |
| 37 | #include <linux/pm_runtime.h> |
| 38 | |
| 39 | #include <trace/events/block.h> |
| 40 | |
| 41 | #include "blk.h" |
| 42 | #include "blk-cgroup.h" |
| 43 | |
| 44 | static DEFINE_SPINLOCK(elv_list_lock); |
| 45 | static LIST_HEAD(elv_list); |
| 46 | |
| 47 | /* |
| 48 | * Merge hash stuff. |
| 49 | */ |
| 50 | #define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq)) |
| 51 | |
| 52 | /* |
| 53 | * Query io scheduler to see if the current process issuing bio may be |
| 54 | * merged with rq. |
| 55 | */ |
| 56 | static int elv_iosched_allow_merge(struct request *rq, struct bio *bio) |
| 57 | { |
| 58 | struct request_queue *q = rq->q; |
| 59 | struct elevator_queue *e = q->elevator; |
| 60 | |
| 61 | if (e->type->ops.elevator_allow_merge_fn) |
| 62 | return e->type->ops.elevator_allow_merge_fn(q, rq, bio); |
| 63 | |
| 64 | return 1; |
| 65 | } |
| 66 | |
| 67 | /* |
| 68 | * can we safely merge with this request? |
| 69 | */ |
| 70 | bool elv_rq_merge_ok(struct request *rq, struct bio *bio) |
| 71 | { |
| 72 | if (!blk_rq_merge_ok(rq, bio)) |
| 73 | return 0; |
| 74 | |
| 75 | if (!elv_iosched_allow_merge(rq, bio)) |
| 76 | return 0; |
| 77 | |
| 78 | return 1; |
| 79 | } |
| 80 | EXPORT_SYMBOL(elv_rq_merge_ok); |
| 81 | |
| 82 | static struct elevator_type *elevator_find(const char *name) |
| 83 | { |
| 84 | struct elevator_type *e; |
| 85 | |
| 86 | list_for_each_entry(e, &elv_list, list) { |
| 87 | if (!strcmp(e->elevator_name, name)) |
| 88 | return e; |
| 89 | } |
| 90 | |
| 91 | return NULL; |
| 92 | } |
| 93 | |
| 94 | static void elevator_put(struct elevator_type *e) |
| 95 | { |
| 96 | module_put(e->elevator_owner); |
| 97 | } |
| 98 | |
| 99 | static struct elevator_type *elevator_get(const char *name, bool try_loading) |
| 100 | { |
| 101 | struct elevator_type *e; |
| 102 | |
| 103 | spin_lock(&elv_list_lock); |
| 104 | |
| 105 | e = elevator_find(name); |
| 106 | if (!e && try_loading) { |
| 107 | spin_unlock(&elv_list_lock); |
| 108 | request_module("%s-iosched", name); |
| 109 | spin_lock(&elv_list_lock); |
| 110 | e = elevator_find(name); |
| 111 | } |
| 112 | |
| 113 | if (e && !try_module_get(e->elevator_owner)) |
| 114 | e = NULL; |
| 115 | |
| 116 | spin_unlock(&elv_list_lock); |
| 117 | |
| 118 | return e; |
| 119 | } |
| 120 | |
| 121 | static char chosen_elevator[ELV_NAME_MAX]; |
| 122 | |
| 123 | static int __init elevator_setup(char *str) |
| 124 | { |
| 125 | /* |
| 126 | * Be backwards-compatible with previous kernels, so users |
| 127 | * won't get the wrong elevator. |
| 128 | */ |
| 129 | strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1); |
| 130 | return 1; |
| 131 | } |
| 132 | |
| 133 | __setup("elevator=", elevator_setup); |
| 134 | |
| 135 | /* called during boot to load the elevator chosen by the elevator param */ |
| 136 | void __init load_default_elevator_module(void) |
| 137 | { |
| 138 | struct elevator_type *e; |
| 139 | |
| 140 | if (!chosen_elevator[0]) |
| 141 | return; |
| 142 | |
| 143 | spin_lock(&elv_list_lock); |
| 144 | e = elevator_find(chosen_elevator); |
| 145 | spin_unlock(&elv_list_lock); |
| 146 | |
| 147 | if (!e) |
| 148 | request_module("%s-iosched", chosen_elevator); |
| 149 | } |
| 150 | |
| 151 | static struct kobj_type elv_ktype; |
| 152 | |
| 153 | struct elevator_queue *elevator_alloc(struct request_queue *q, |
| 154 | struct elevator_type *e) |
| 155 | { |
| 156 | struct elevator_queue *eq; |
| 157 | |
| 158 | eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node); |
| 159 | if (unlikely(!eq)) |
| 160 | goto err; |
| 161 | |
| 162 | eq->type = e; |
| 163 | kobject_init(&eq->kobj, &elv_ktype); |
| 164 | mutex_init(&eq->sysfs_lock); |
| 165 | hash_init(eq->hash); |
| 166 | |
| 167 | return eq; |
| 168 | err: |
| 169 | kfree(eq); |
| 170 | elevator_put(e); |
| 171 | return NULL; |
| 172 | } |
| 173 | EXPORT_SYMBOL(elevator_alloc); |
| 174 | |
| 175 | static void elevator_release(struct kobject *kobj) |
| 176 | { |
| 177 | struct elevator_queue *e; |
| 178 | |
| 179 | e = container_of(kobj, struct elevator_queue, kobj); |
| 180 | elevator_put(e->type); |
| 181 | kfree(e); |
| 182 | } |
| 183 | |
| 184 | int elevator_init(struct request_queue *q, char *name) |
| 185 | { |
| 186 | struct elevator_type *e = NULL; |
| 187 | int err; |
| 188 | |
| 189 | /* |
| 190 | * q->sysfs_lock must be held to provide mutual exclusion between |
| 191 | * elevator_switch() and here. |
| 192 | */ |
| 193 | lockdep_assert_held(&q->sysfs_lock); |
| 194 | |
| 195 | if (unlikely(q->elevator)) |
| 196 | return 0; |
| 197 | |
| 198 | INIT_LIST_HEAD(&q->queue_head); |
| 199 | q->last_merge = NULL; |
| 200 | q->end_sector = 0; |
| 201 | q->boundary_rq = NULL; |
| 202 | |
| 203 | if (name) { |
| 204 | e = elevator_get(name, true); |
| 205 | if (!e) |
| 206 | return -EINVAL; |
| 207 | } |
| 208 | |
| 209 | /* |
| 210 | * Use the default elevator specified by config boot param or |
| 211 | * config option. Don't try to load modules as we could be running |
| 212 | * off async and request_module() isn't allowed from async. |
| 213 | */ |
| 214 | if (!e && *chosen_elevator) { |
| 215 | e = elevator_get(chosen_elevator, false); |
| 216 | if (!e) |
| 217 | printk(KERN_ERR "I/O scheduler %s not found\n", |
| 218 | chosen_elevator); |
| 219 | } |
| 220 | |
| 221 | if (!e) { |
| 222 | e = elevator_get(CONFIG_DEFAULT_IOSCHED, false); |
| 223 | if (!e) { |
| 224 | printk(KERN_ERR |
| 225 | "Default I/O scheduler not found. " \ |
| 226 | "Using noop.\n"); |
| 227 | e = elevator_get("noop", false); |
| 228 | } |
| 229 | } |
| 230 | |
| 231 | err = e->ops.elevator_init_fn(q, e); |
| 232 | return 0; |
| 233 | } |
| 234 | EXPORT_SYMBOL(elevator_init); |
| 235 | |
| 236 | void elevator_exit(struct elevator_queue *e) |
| 237 | { |
| 238 | mutex_lock(&e->sysfs_lock); |
| 239 | if (e->type->ops.elevator_exit_fn) |
| 240 | e->type->ops.elevator_exit_fn(e); |
| 241 | mutex_unlock(&e->sysfs_lock); |
| 242 | |
| 243 | kobject_put(&e->kobj); |
| 244 | } |
| 245 | EXPORT_SYMBOL(elevator_exit); |
| 246 | |
| 247 | static inline void __elv_rqhash_del(struct request *rq) |
| 248 | { |
| 249 | hash_del(&rq->hash); |
| 250 | } |
| 251 | |
| 252 | static void elv_rqhash_del(struct request_queue *q, struct request *rq) |
| 253 | { |
| 254 | if (ELV_ON_HASH(rq)) |
| 255 | __elv_rqhash_del(rq); |
| 256 | } |
| 257 | |
| 258 | static void elv_rqhash_add(struct request_queue *q, struct request *rq) |
| 259 | { |
| 260 | struct elevator_queue *e = q->elevator; |
| 261 | |
| 262 | BUG_ON(ELV_ON_HASH(rq)); |
| 263 | hash_add(e->hash, &rq->hash, rq_hash_key(rq)); |
| 264 | } |
| 265 | |
| 266 | static void elv_rqhash_reposition(struct request_queue *q, struct request *rq) |
| 267 | { |
| 268 | __elv_rqhash_del(rq); |
| 269 | elv_rqhash_add(q, rq); |
| 270 | } |
| 271 | |
| 272 | static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset) |
| 273 | { |
| 274 | struct elevator_queue *e = q->elevator; |
| 275 | struct hlist_node *next; |
| 276 | struct request *rq; |
| 277 | |
| 278 | hash_for_each_possible_safe(e->hash, rq, next, hash, offset) { |
| 279 | BUG_ON(!ELV_ON_HASH(rq)); |
| 280 | |
| 281 | if (unlikely(!rq_mergeable(rq))) { |
| 282 | __elv_rqhash_del(rq); |
| 283 | continue; |
| 284 | } |
| 285 | |
| 286 | if (rq_hash_key(rq) == offset) |
| 287 | return rq; |
| 288 | } |
| 289 | |
| 290 | return NULL; |
| 291 | } |
| 292 | |
| 293 | /* |
| 294 | * RB-tree support functions for inserting/lookup/removal of requests |
| 295 | * in a sorted RB tree. |
| 296 | */ |
| 297 | void elv_rb_add(struct rb_root *root, struct request *rq) |
| 298 | { |
| 299 | struct rb_node **p = &root->rb_node; |
| 300 | struct rb_node *parent = NULL; |
| 301 | struct request *__rq; |
| 302 | |
| 303 | while (*p) { |
| 304 | parent = *p; |
| 305 | __rq = rb_entry(parent, struct request, rb_node); |
| 306 | |
| 307 | if (blk_rq_pos(rq) < blk_rq_pos(__rq)) |
| 308 | p = &(*p)->rb_left; |
| 309 | else if (blk_rq_pos(rq) >= blk_rq_pos(__rq)) |
| 310 | p = &(*p)->rb_right; |
| 311 | } |
| 312 | |
| 313 | rb_link_node(&rq->rb_node, parent, p); |
| 314 | rb_insert_color(&rq->rb_node, root); |
| 315 | } |
| 316 | EXPORT_SYMBOL(elv_rb_add); |
| 317 | |
| 318 | void elv_rb_del(struct rb_root *root, struct request *rq) |
| 319 | { |
| 320 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); |
| 321 | rb_erase(&rq->rb_node, root); |
| 322 | RB_CLEAR_NODE(&rq->rb_node); |
| 323 | } |
| 324 | EXPORT_SYMBOL(elv_rb_del); |
| 325 | |
| 326 | struct request *elv_rb_find(struct rb_root *root, sector_t sector) |
| 327 | { |
| 328 | struct rb_node *n = root->rb_node; |
| 329 | struct request *rq; |
| 330 | |
| 331 | while (n) { |
| 332 | rq = rb_entry(n, struct request, rb_node); |
| 333 | |
| 334 | if (sector < blk_rq_pos(rq)) |
| 335 | n = n->rb_left; |
| 336 | else if (sector > blk_rq_pos(rq)) |
| 337 | n = n->rb_right; |
| 338 | else |
| 339 | return rq; |
| 340 | } |
| 341 | |
| 342 | return NULL; |
| 343 | } |
| 344 | EXPORT_SYMBOL(elv_rb_find); |
| 345 | |
| 346 | /* |
| 347 | * Insert rq into dispatch queue of q. Queue lock must be held on |
| 348 | * entry. rq is sort instead into the dispatch queue. To be used by |
| 349 | * specific elevators. |
| 350 | */ |
| 351 | void elv_dispatch_sort(struct request_queue *q, struct request *rq) |
| 352 | { |
| 353 | sector_t boundary; |
| 354 | struct list_head *entry; |
| 355 | int stop_flags; |
| 356 | |
| 357 | if (q->last_merge == rq) |
| 358 | q->last_merge = NULL; |
| 359 | |
| 360 | elv_rqhash_del(q, rq); |
| 361 | |
| 362 | q->nr_sorted--; |
| 363 | |
| 364 | boundary = q->end_sector; |
| 365 | stop_flags = REQ_SOFTBARRIER | REQ_STARTED; |
| 366 | list_for_each_prev(entry, &q->queue_head) { |
| 367 | struct request *pos = list_entry_rq(entry); |
| 368 | |
| 369 | if ((rq->cmd_flags & REQ_DISCARD) != |
| 370 | (pos->cmd_flags & REQ_DISCARD)) |
| 371 | break; |
| 372 | if (rq_data_dir(rq) != rq_data_dir(pos)) |
| 373 | break; |
| 374 | if (pos->cmd_flags & stop_flags) |
| 375 | break; |
| 376 | if (blk_rq_pos(rq) >= boundary) { |
| 377 | if (blk_rq_pos(pos) < boundary) |
| 378 | continue; |
| 379 | } else { |
| 380 | if (blk_rq_pos(pos) >= boundary) |
| 381 | break; |
| 382 | } |
| 383 | if (blk_rq_pos(rq) >= blk_rq_pos(pos)) |
| 384 | break; |
| 385 | } |
| 386 | |
| 387 | list_add(&rq->queuelist, entry); |
| 388 | } |
| 389 | EXPORT_SYMBOL(elv_dispatch_sort); |
| 390 | |
| 391 | /* |
| 392 | * Insert rq into dispatch queue of q. Queue lock must be held on |
| 393 | * entry. rq is added to the back of the dispatch queue. To be used by |
| 394 | * specific elevators. |
| 395 | */ |
| 396 | void elv_dispatch_add_tail(struct request_queue *q, struct request *rq) |
| 397 | { |
| 398 | if (q->last_merge == rq) |
| 399 | q->last_merge = NULL; |
| 400 | |
| 401 | elv_rqhash_del(q, rq); |
| 402 | |
| 403 | q->nr_sorted--; |
| 404 | |
| 405 | q->end_sector = rq_end_sector(rq); |
| 406 | q->boundary_rq = rq; |
| 407 | list_add_tail(&rq->queuelist, &q->queue_head); |
| 408 | } |
| 409 | EXPORT_SYMBOL(elv_dispatch_add_tail); |
| 410 | |
| 411 | int elv_merge(struct request_queue *q, struct request **req, struct bio *bio) |
| 412 | { |
| 413 | struct elevator_queue *e = q->elevator; |
| 414 | struct request *__rq; |
| 415 | int ret; |
| 416 | |
| 417 | /* |
| 418 | * Levels of merges: |
| 419 | * nomerges: No merges at all attempted |
| 420 | * noxmerges: Only simple one-hit cache try |
| 421 | * merges: All merge tries attempted |
| 422 | */ |
| 423 | if (blk_queue_nomerges(q)) |
| 424 | return ELEVATOR_NO_MERGE; |
| 425 | |
| 426 | /* |
| 427 | * First try one-hit cache. |
| 428 | */ |
| 429 | if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) { |
| 430 | ret = blk_try_merge(q->last_merge, bio); |
| 431 | if (ret != ELEVATOR_NO_MERGE) { |
| 432 | *req = q->last_merge; |
| 433 | return ret; |
| 434 | } |
| 435 | } |
| 436 | |
| 437 | if (blk_queue_noxmerges(q)) |
| 438 | return ELEVATOR_NO_MERGE; |
| 439 | |
| 440 | /* |
| 441 | * See if our hash lookup can find a potential backmerge. |
| 442 | */ |
| 443 | __rq = elv_rqhash_find(q, bio->bi_sector); |
| 444 | if (__rq && elv_rq_merge_ok(__rq, bio)) { |
| 445 | *req = __rq; |
| 446 | return ELEVATOR_BACK_MERGE; |
| 447 | } |
| 448 | |
| 449 | if (e->type->ops.elevator_merge_fn) |
| 450 | return e->type->ops.elevator_merge_fn(q, req, bio); |
| 451 | |
| 452 | return ELEVATOR_NO_MERGE; |
| 453 | } |
| 454 | |
| 455 | /* |
| 456 | * Attempt to do an insertion back merge. Only check for the case where |
| 457 | * we can append 'rq' to an existing request, so we can throw 'rq' away |
| 458 | * afterwards. |
| 459 | * |
| 460 | * Returns true if we merged, false otherwise |
| 461 | */ |
| 462 | static bool elv_attempt_insert_merge(struct request_queue *q, |
| 463 | struct request *rq) |
| 464 | { |
| 465 | struct request *__rq; |
| 466 | bool ret; |
| 467 | |
| 468 | if (blk_queue_nomerges(q)) |
| 469 | return false; |
| 470 | |
| 471 | /* |
| 472 | * First try one-hit cache. |
| 473 | */ |
| 474 | if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) |
| 475 | return true; |
| 476 | |
| 477 | if (blk_queue_noxmerges(q)) |
| 478 | return false; |
| 479 | |
| 480 | ret = false; |
| 481 | /* |
| 482 | * See if our hash lookup can find a potential backmerge. |
| 483 | */ |
| 484 | while (1) { |
| 485 | __rq = elv_rqhash_find(q, blk_rq_pos(rq)); |
| 486 | if (!__rq || !blk_attempt_req_merge(q, __rq, rq)) |
| 487 | break; |
| 488 | |
| 489 | /* The merged request could be merged with others, try again */ |
| 490 | ret = true; |
| 491 | rq = __rq; |
| 492 | } |
| 493 | |
| 494 | return ret; |
| 495 | } |
| 496 | |
| 497 | void elv_merged_request(struct request_queue *q, struct request *rq, int type) |
| 498 | { |
| 499 | struct elevator_queue *e = q->elevator; |
| 500 | |
| 501 | if (e->type->ops.elevator_merged_fn) |
| 502 | e->type->ops.elevator_merged_fn(q, rq, type); |
| 503 | |
| 504 | if (type == ELEVATOR_BACK_MERGE) |
| 505 | elv_rqhash_reposition(q, rq); |
| 506 | |
| 507 | q->last_merge = rq; |
| 508 | } |
| 509 | |
| 510 | void elv_merge_requests(struct request_queue *q, struct request *rq, |
| 511 | struct request *next) |
| 512 | { |
| 513 | struct elevator_queue *e = q->elevator; |
| 514 | const int next_sorted = next->cmd_flags & REQ_SORTED; |
| 515 | |
| 516 | if (next_sorted && e->type->ops.elevator_merge_req_fn) |
| 517 | e->type->ops.elevator_merge_req_fn(q, rq, next); |
| 518 | |
| 519 | elv_rqhash_reposition(q, rq); |
| 520 | |
| 521 | if (next_sorted) { |
| 522 | elv_rqhash_del(q, next); |
| 523 | q->nr_sorted--; |
| 524 | } |
| 525 | |
| 526 | q->last_merge = rq; |
| 527 | } |
| 528 | |
| 529 | void elv_bio_merged(struct request_queue *q, struct request *rq, |
| 530 | struct bio *bio) |
| 531 | { |
| 532 | struct elevator_queue *e = q->elevator; |
| 533 | |
| 534 | if (e->type->ops.elevator_bio_merged_fn) |
| 535 | e->type->ops.elevator_bio_merged_fn(q, rq, bio); |
| 536 | } |
| 537 | |
| 538 | #ifdef CONFIG_PM_RUNTIME |
| 539 | static void blk_pm_requeue_request(struct request *rq) |
| 540 | { |
| 541 | if (rq->q->dev && !(rq->cmd_flags & REQ_PM)) |
| 542 | rq->q->nr_pending--; |
| 543 | } |
| 544 | |
| 545 | static void blk_pm_add_request(struct request_queue *q, struct request *rq) |
| 546 | { |
| 547 | if (q->dev && !(rq->cmd_flags & REQ_PM) && q->nr_pending++ == 0 && |
| 548 | (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING)) |
| 549 | pm_request_resume(q->dev); |
| 550 | } |
| 551 | #else |
| 552 | static inline void blk_pm_requeue_request(struct request *rq) {} |
| 553 | static inline void blk_pm_add_request(struct request_queue *q, |
| 554 | struct request *rq) |
| 555 | { |
| 556 | } |
| 557 | #endif |
| 558 | |
| 559 | void elv_requeue_request(struct request_queue *q, struct request *rq) |
| 560 | { |
| 561 | /* |
| 562 | * it already went through dequeue, we need to decrement the |
| 563 | * in_flight count again |
| 564 | */ |
| 565 | if (blk_account_rq(rq)) { |
| 566 | q->in_flight[rq_is_sync(rq)]--; |
| 567 | if (rq->cmd_flags & REQ_SORTED) |
| 568 | elv_deactivate_rq(q, rq); |
| 569 | } |
| 570 | |
| 571 | rq->cmd_flags &= ~REQ_STARTED; |
| 572 | |
| 573 | blk_pm_requeue_request(rq); |
| 574 | |
| 575 | __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE); |
| 576 | } |
| 577 | |
| 578 | void elv_drain_elevator(struct request_queue *q) |
| 579 | { |
| 580 | static int printed; |
| 581 | |
| 582 | lockdep_assert_held(q->queue_lock); |
| 583 | |
| 584 | while (q->elevator->type->ops.elevator_dispatch_fn(q, 1)) |
| 585 | ; |
| 586 | if (q->nr_sorted && printed++ < 10) { |
| 587 | printk(KERN_ERR "%s: forced dispatching is broken " |
| 588 | "(nr_sorted=%u), please report this\n", |
| 589 | q->elevator->type->elevator_name, q->nr_sorted); |
| 590 | } |
| 591 | } |
| 592 | |
| 593 | void __elv_add_request(struct request_queue *q, struct request *rq, int where) |
| 594 | { |
| 595 | trace_block_rq_insert(q, rq); |
| 596 | |
| 597 | blk_pm_add_request(q, rq); |
| 598 | |
| 599 | rq->q = q; |
| 600 | |
| 601 | if (rq->cmd_flags & REQ_SOFTBARRIER) { |
| 602 | /* barriers are scheduling boundary, update end_sector */ |
| 603 | if (rq->cmd_type == REQ_TYPE_FS) { |
| 604 | q->end_sector = rq_end_sector(rq); |
| 605 | q->boundary_rq = rq; |
| 606 | } |
| 607 | } else if (!(rq->cmd_flags & REQ_ELVPRIV) && |
| 608 | (where == ELEVATOR_INSERT_SORT || |
| 609 | where == ELEVATOR_INSERT_SORT_MERGE)) |
| 610 | where = ELEVATOR_INSERT_BACK; |
| 611 | |
| 612 | switch (where) { |
| 613 | case ELEVATOR_INSERT_REQUEUE: |
| 614 | case ELEVATOR_INSERT_FRONT: |
| 615 | rq->cmd_flags |= REQ_SOFTBARRIER; |
| 616 | list_add(&rq->queuelist, &q->queue_head); |
| 617 | break; |
| 618 | |
| 619 | case ELEVATOR_INSERT_BACK: |
| 620 | rq->cmd_flags |= REQ_SOFTBARRIER; |
| 621 | elv_drain_elevator(q); |
| 622 | list_add_tail(&rq->queuelist, &q->queue_head); |
| 623 | /* |
| 624 | * We kick the queue here for the following reasons. |
| 625 | * - The elevator might have returned NULL previously |
| 626 | * to delay requests and returned them now. As the |
| 627 | * queue wasn't empty before this request, ll_rw_blk |
| 628 | * won't run the queue on return, resulting in hang. |
| 629 | * - Usually, back inserted requests won't be merged |
| 630 | * with anything. There's no point in delaying queue |
| 631 | * processing. |
| 632 | */ |
| 633 | __blk_run_queue(q); |
| 634 | break; |
| 635 | |
| 636 | case ELEVATOR_INSERT_SORT_MERGE: |
| 637 | /* |
| 638 | * If we succeed in merging this request with one in the |
| 639 | * queue already, we are done - rq has now been freed, |
| 640 | * so no need to do anything further. |
| 641 | */ |
| 642 | if (elv_attempt_insert_merge(q, rq)) |
| 643 | break; |
| 644 | case ELEVATOR_INSERT_SORT: |
| 645 | BUG_ON(rq->cmd_type != REQ_TYPE_FS); |
| 646 | rq->cmd_flags |= REQ_SORTED; |
| 647 | q->nr_sorted++; |
| 648 | if (rq_mergeable(rq)) { |
| 649 | elv_rqhash_add(q, rq); |
| 650 | if (!q->last_merge) |
| 651 | q->last_merge = rq; |
| 652 | } |
| 653 | |
| 654 | /* |
| 655 | * Some ioscheds (cfq) run q->request_fn directly, so |
| 656 | * rq cannot be accessed after calling |
| 657 | * elevator_add_req_fn. |
| 658 | */ |
| 659 | q->elevator->type->ops.elevator_add_req_fn(q, rq); |
| 660 | break; |
| 661 | |
| 662 | case ELEVATOR_INSERT_FLUSH: |
| 663 | rq->cmd_flags |= REQ_SOFTBARRIER; |
| 664 | blk_insert_flush(rq); |
| 665 | break; |
| 666 | default: |
| 667 | printk(KERN_ERR "%s: bad insertion point %d\n", |
| 668 | __func__, where); |
| 669 | BUG(); |
| 670 | } |
| 671 | } |
| 672 | EXPORT_SYMBOL(__elv_add_request); |
| 673 | |
| 674 | void elv_add_request(struct request_queue *q, struct request *rq, int where) |
| 675 | { |
| 676 | unsigned long flags; |
| 677 | |
| 678 | spin_lock_irqsave(q->queue_lock, flags); |
| 679 | __elv_add_request(q, rq, where); |
| 680 | spin_unlock_irqrestore(q->queue_lock, flags); |
| 681 | } |
| 682 | EXPORT_SYMBOL(elv_add_request); |
| 683 | |
| 684 | struct request *elv_latter_request(struct request_queue *q, struct request *rq) |
| 685 | { |
| 686 | struct elevator_queue *e = q->elevator; |
| 687 | |
| 688 | if (e->type->ops.elevator_latter_req_fn) |
| 689 | return e->type->ops.elevator_latter_req_fn(q, rq); |
| 690 | return NULL; |
| 691 | } |
| 692 | |
| 693 | struct request *elv_former_request(struct request_queue *q, struct request *rq) |
| 694 | { |
| 695 | struct elevator_queue *e = q->elevator; |
| 696 | |
| 697 | if (e->type->ops.elevator_former_req_fn) |
| 698 | return e->type->ops.elevator_former_req_fn(q, rq); |
| 699 | return NULL; |
| 700 | } |
| 701 | |
| 702 | int elv_set_request(struct request_queue *q, struct request *rq, |
| 703 | struct bio *bio, gfp_t gfp_mask) |
| 704 | { |
| 705 | struct elevator_queue *e = q->elevator; |
| 706 | |
| 707 | if (e->type->ops.elevator_set_req_fn) |
| 708 | return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask); |
| 709 | return 0; |
| 710 | } |
| 711 | |
| 712 | void elv_put_request(struct request_queue *q, struct request *rq) |
| 713 | { |
| 714 | struct elevator_queue *e = q->elevator; |
| 715 | |
| 716 | if (e->type->ops.elevator_put_req_fn) |
| 717 | e->type->ops.elevator_put_req_fn(rq); |
| 718 | } |
| 719 | |
| 720 | int elv_may_queue(struct request_queue *q, int rw) |
| 721 | { |
| 722 | struct elevator_queue *e = q->elevator; |
| 723 | |
| 724 | if (e->type->ops.elevator_may_queue_fn) |
| 725 | return e->type->ops.elevator_may_queue_fn(q, rw); |
| 726 | |
| 727 | return ELV_MQUEUE_MAY; |
| 728 | } |
| 729 | |
| 730 | void elv_abort_queue(struct request_queue *q) |
| 731 | { |
| 732 | struct request *rq; |
| 733 | |
| 734 | blk_abort_flushes(q); |
| 735 | |
| 736 | while (!list_empty(&q->queue_head)) { |
| 737 | rq = list_entry_rq(q->queue_head.next); |
| 738 | rq->cmd_flags |= REQ_QUIET; |
| 739 | trace_block_rq_abort(q, rq); |
| 740 | /* |
| 741 | * Mark this request as started so we don't trigger |
| 742 | * any debug logic in the end I/O path. |
| 743 | */ |
| 744 | blk_start_request(rq); |
| 745 | __blk_end_request_all(rq, -EIO); |
| 746 | } |
| 747 | } |
| 748 | EXPORT_SYMBOL(elv_abort_queue); |
| 749 | |
| 750 | void elv_completed_request(struct request_queue *q, struct request *rq) |
| 751 | { |
| 752 | struct elevator_queue *e = q->elevator; |
| 753 | |
| 754 | /* |
| 755 | * request is released from the driver, io must be done |
| 756 | */ |
| 757 | if (blk_account_rq(rq)) { |
| 758 | q->in_flight[rq_is_sync(rq)]--; |
| 759 | if ((rq->cmd_flags & REQ_SORTED) && |
| 760 | e->type->ops.elevator_completed_req_fn) |
| 761 | e->type->ops.elevator_completed_req_fn(q, rq); |
| 762 | } |
| 763 | } |
| 764 | |
| 765 | #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr) |
| 766 | |
| 767 | static ssize_t |
| 768 | elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page) |
| 769 | { |
| 770 | struct elv_fs_entry *entry = to_elv(attr); |
| 771 | struct elevator_queue *e; |
| 772 | ssize_t error; |
| 773 | |
| 774 | if (!entry->show) |
| 775 | return -EIO; |
| 776 | |
| 777 | e = container_of(kobj, struct elevator_queue, kobj); |
| 778 | mutex_lock(&e->sysfs_lock); |
| 779 | error = e->type ? entry->show(e, page) : -ENOENT; |
| 780 | mutex_unlock(&e->sysfs_lock); |
| 781 | return error; |
| 782 | } |
| 783 | |
| 784 | static ssize_t |
| 785 | elv_attr_store(struct kobject *kobj, struct attribute *attr, |
| 786 | const char *page, size_t length) |
| 787 | { |
| 788 | struct elv_fs_entry *entry = to_elv(attr); |
| 789 | struct elevator_queue *e; |
| 790 | ssize_t error; |
| 791 | |
| 792 | if (!entry->store) |
| 793 | return -EIO; |
| 794 | |
| 795 | e = container_of(kobj, struct elevator_queue, kobj); |
| 796 | mutex_lock(&e->sysfs_lock); |
| 797 | error = e->type ? entry->store(e, page, length) : -ENOENT; |
| 798 | mutex_unlock(&e->sysfs_lock); |
| 799 | return error; |
| 800 | } |
| 801 | |
| 802 | static const struct sysfs_ops elv_sysfs_ops = { |
| 803 | .show = elv_attr_show, |
| 804 | .store = elv_attr_store, |
| 805 | }; |
| 806 | |
| 807 | static struct kobj_type elv_ktype = { |
| 808 | .sysfs_ops = &elv_sysfs_ops, |
| 809 | .release = elevator_release, |
| 810 | }; |
| 811 | |
| 812 | int elv_register_queue(struct request_queue *q) |
| 813 | { |
| 814 | struct elevator_queue *e = q->elevator; |
| 815 | int error; |
| 816 | |
| 817 | error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched"); |
| 818 | if (!error) { |
| 819 | struct elv_fs_entry *attr = e->type->elevator_attrs; |
| 820 | if (attr) { |
| 821 | while (attr->attr.name) { |
| 822 | if (sysfs_create_file(&e->kobj, &attr->attr)) |
| 823 | break; |
| 824 | attr++; |
| 825 | } |
| 826 | } |
| 827 | kobject_uevent(&e->kobj, KOBJ_ADD); |
| 828 | e->registered = 1; |
| 829 | } |
| 830 | return error; |
| 831 | } |
| 832 | EXPORT_SYMBOL(elv_register_queue); |
| 833 | |
| 834 | void elv_unregister_queue(struct request_queue *q) |
| 835 | { |
| 836 | if (q) { |
| 837 | struct elevator_queue *e = q->elevator; |
| 838 | |
| 839 | kobject_uevent(&e->kobj, KOBJ_REMOVE); |
| 840 | kobject_del(&e->kobj); |
| 841 | e->registered = 0; |
| 842 | } |
| 843 | } |
| 844 | EXPORT_SYMBOL(elv_unregister_queue); |
| 845 | |
| 846 | int elv_register(struct elevator_type *e) |
| 847 | { |
| 848 | char *def = ""; |
| 849 | |
| 850 | /* create icq_cache if requested */ |
| 851 | if (e->icq_size) { |
| 852 | if (WARN_ON(e->icq_size < sizeof(struct io_cq)) || |
| 853 | WARN_ON(e->icq_align < __alignof__(struct io_cq))) |
| 854 | return -EINVAL; |
| 855 | |
| 856 | snprintf(e->icq_cache_name, sizeof(e->icq_cache_name), |
| 857 | "%s_io_cq", e->elevator_name); |
| 858 | e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size, |
| 859 | e->icq_align, 0, NULL); |
| 860 | if (!e->icq_cache) |
| 861 | return -ENOMEM; |
| 862 | } |
| 863 | |
| 864 | /* register, don't allow duplicate names */ |
| 865 | spin_lock(&elv_list_lock); |
| 866 | if (elevator_find(e->elevator_name)) { |
| 867 | spin_unlock(&elv_list_lock); |
| 868 | if (e->icq_cache) |
| 869 | kmem_cache_destroy(e->icq_cache); |
| 870 | return -EBUSY; |
| 871 | } |
| 872 | list_add_tail(&e->list, &elv_list); |
| 873 | spin_unlock(&elv_list_lock); |
| 874 | |
| 875 | /* print pretty message */ |
| 876 | if (!strcmp(e->elevator_name, chosen_elevator) || |
| 877 | (!*chosen_elevator && |
| 878 | !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED))) |
| 879 | def = " (default)"; |
| 880 | |
| 881 | printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name, |
| 882 | def); |
| 883 | return 0; |
| 884 | } |
| 885 | EXPORT_SYMBOL_GPL(elv_register); |
| 886 | |
| 887 | void elv_unregister(struct elevator_type *e) |
| 888 | { |
| 889 | /* unregister */ |
| 890 | spin_lock(&elv_list_lock); |
| 891 | list_del_init(&e->list); |
| 892 | spin_unlock(&elv_list_lock); |
| 893 | |
| 894 | /* |
| 895 | * Destroy icq_cache if it exists. icq's are RCU managed. Make |
| 896 | * sure all RCU operations are complete before proceeding. |
| 897 | */ |
| 898 | if (e->icq_cache) { |
| 899 | rcu_barrier(); |
| 900 | kmem_cache_destroy(e->icq_cache); |
| 901 | e->icq_cache = NULL; |
| 902 | } |
| 903 | } |
| 904 | EXPORT_SYMBOL_GPL(elv_unregister); |
| 905 | |
| 906 | /* |
| 907 | * switch to new_e io scheduler. be careful not to introduce deadlocks - |
| 908 | * we don't free the old io scheduler, before we have allocated what we |
| 909 | * need for the new one. this way we have a chance of going back to the old |
| 910 | * one, if the new one fails init for some reason. |
| 911 | */ |
| 912 | static int elevator_switch(struct request_queue *q, struct elevator_type *new_e) |
| 913 | { |
| 914 | struct elevator_queue *old = q->elevator; |
| 915 | bool registered = old->registered; |
| 916 | int err; |
| 917 | |
| 918 | /* |
| 919 | * Turn on BYPASS and drain all requests w/ elevator private data. |
| 920 | * Block layer doesn't call into a quiesced elevator - all requests |
| 921 | * are directly put on the dispatch list without elevator data |
| 922 | * using INSERT_BACK. All requests have SOFTBARRIER set and no |
| 923 | * merge happens either. |
| 924 | */ |
| 925 | blk_queue_bypass_start(q); |
| 926 | |
| 927 | /* unregister and clear all auxiliary data of the old elevator */ |
| 928 | if (registered) |
| 929 | elv_unregister_queue(q); |
| 930 | |
| 931 | spin_lock_irq(q->queue_lock); |
| 932 | ioc_clear_queue(q); |
| 933 | spin_unlock_irq(q->queue_lock); |
| 934 | |
| 935 | /* allocate, init and register new elevator */ |
| 936 | err = new_e->ops.elevator_init_fn(q, new_e); |
| 937 | if (err) |
| 938 | goto fail_init; |
| 939 | |
| 940 | if (registered) { |
| 941 | err = elv_register_queue(q); |
| 942 | if (err) |
| 943 | goto fail_register; |
| 944 | } |
| 945 | |
| 946 | /* done, kill the old one and finish */ |
| 947 | elevator_exit(old); |
| 948 | blk_queue_bypass_end(q); |
| 949 | |
| 950 | blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name); |
| 951 | |
| 952 | return 0; |
| 953 | |
| 954 | fail_register: |
| 955 | elevator_exit(q->elevator); |
| 956 | fail_init: |
| 957 | /* switch failed, restore and re-register old elevator */ |
| 958 | q->elevator = old; |
| 959 | elv_register_queue(q); |
| 960 | blk_queue_bypass_end(q); |
| 961 | |
| 962 | return err; |
| 963 | } |
| 964 | |
| 965 | /* |
| 966 | * Switch this queue to the given IO scheduler. |
| 967 | */ |
| 968 | static int __elevator_change(struct request_queue *q, const char *name) |
| 969 | { |
| 970 | char elevator_name[ELV_NAME_MAX]; |
| 971 | struct elevator_type *e; |
| 972 | |
| 973 | if (!q->elevator) |
| 974 | return -ENXIO; |
| 975 | |
| 976 | strlcpy(elevator_name, name, sizeof(elevator_name)); |
| 977 | e = elevator_get(strstrip(elevator_name), true); |
| 978 | if (!e) { |
| 979 | printk(KERN_ERR "elevator: type %s not found\n", elevator_name); |
| 980 | return -EINVAL; |
| 981 | } |
| 982 | |
| 983 | if (!strcmp(elevator_name, q->elevator->type->elevator_name)) { |
| 984 | elevator_put(e); |
| 985 | return 0; |
| 986 | } |
| 987 | |
| 988 | return elevator_switch(q, e); |
| 989 | } |
| 990 | |
| 991 | int elevator_change(struct request_queue *q, const char *name) |
| 992 | { |
| 993 | int ret; |
| 994 | |
| 995 | /* Protect q->elevator from elevator_init() */ |
| 996 | mutex_lock(&q->sysfs_lock); |
| 997 | ret = __elevator_change(q, name); |
| 998 | mutex_unlock(&q->sysfs_lock); |
| 999 | |
| 1000 | return ret; |
| 1001 | } |
| 1002 | EXPORT_SYMBOL(elevator_change); |
| 1003 | |
| 1004 | ssize_t elv_iosched_store(struct request_queue *q, const char *name, |
| 1005 | size_t count) |
| 1006 | { |
| 1007 | int ret; |
| 1008 | |
| 1009 | if (!q->elevator) |
| 1010 | return count; |
| 1011 | |
| 1012 | ret = __elevator_change(q, name); |
| 1013 | if (!ret) |
| 1014 | return count; |
| 1015 | |
| 1016 | printk(KERN_ERR "elevator: switch to %s failed\n", name); |
| 1017 | return ret; |
| 1018 | } |
| 1019 | |
| 1020 | ssize_t elv_iosched_show(struct request_queue *q, char *name) |
| 1021 | { |
| 1022 | struct elevator_queue *e = q->elevator; |
| 1023 | struct elevator_type *elv; |
| 1024 | struct elevator_type *__e; |
| 1025 | int len = 0; |
| 1026 | |
| 1027 | if (!q->elevator || !blk_queue_stackable(q)) |
| 1028 | return sprintf(name, "none\n"); |
| 1029 | |
| 1030 | elv = e->type; |
| 1031 | |
| 1032 | spin_lock(&elv_list_lock); |
| 1033 | list_for_each_entry(__e, &elv_list, list) { |
| 1034 | if (!strcmp(elv->elevator_name, __e->elevator_name)) |
| 1035 | len += sprintf(name+len, "[%s] ", elv->elevator_name); |
| 1036 | else |
| 1037 | len += sprintf(name+len, "%s ", __e->elevator_name); |
| 1038 | } |
| 1039 | spin_unlock(&elv_list_lock); |
| 1040 | |
| 1041 | len += sprintf(len+name, "\n"); |
| 1042 | return len; |
| 1043 | } |
| 1044 | |
| 1045 | struct request *elv_rb_former_request(struct request_queue *q, |
| 1046 | struct request *rq) |
| 1047 | { |
| 1048 | struct rb_node *rbprev = rb_prev(&rq->rb_node); |
| 1049 | |
| 1050 | if (rbprev) |
| 1051 | return rb_entry_rq(rbprev); |
| 1052 | |
| 1053 | return NULL; |
| 1054 | } |
| 1055 | EXPORT_SYMBOL(elv_rb_former_request); |
| 1056 | |
| 1057 | struct request *elv_rb_latter_request(struct request_queue *q, |
| 1058 | struct request *rq) |
| 1059 | { |
| 1060 | struct rb_node *rbnext = rb_next(&rq->rb_node); |
| 1061 | |
| 1062 | if (rbnext) |
| 1063 | return rb_entry_rq(rbnext); |
| 1064 | |
| 1065 | return NULL; |
| 1066 | } |
| 1067 | EXPORT_SYMBOL(elv_rb_latter_request); |