block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/elevator.h>
  12 #include <linux/rbtree.h>
  13 #include <linux/ioprio.h>
  14 #include <linux/blktrace_api.h>
  15
  16 /*
  17  * tunables
  18  */
  19 /* max queue in one round of service */
  20 static const int cfq_quantum = 4;
  21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  22 /* maximum backwards seek, in KiB */
  23 static const int cfq_back_max = 16 * 1024;
  24 /* penalty of a backwards seek */
  25 static const int cfq_back_penalty = 2;
  26 static const int cfq_slice_sync = HZ / 10;
  27 static int cfq_slice_async = HZ / 25;
  28 static const int cfq_slice_async_rq = 2;
  29 static int cfq_slice_idle = HZ / 125;
  30
  31 /*
  32  * offset from end of service tree
  33  */
  34 #define CFQ_IDLE_DELAY          (HZ / 5)
  35
  36 /*
  37  * below this threshold, we consider thinktime immediate
  38  */
  39 #define CFQ_MIN_TT              (2)
  40
  41 #define CFQ_SLICE_SCALE         (5)
  42 #define CFQ_HW_QUEUE_MIN        (5)
  43
  44 #define RQ_CIC(rq)              \
  45         ((struct cfq_io_context *) (rq)->elevator_private)
  46 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
  47
  48 static struct kmem_cache *cfq_pool;
  49 static struct kmem_cache *cfq_ioc_pool;
  50
  51 static DEFINE_PER_CPU(unsigned long, ioc_count);
  52 static struct completion *ioc_gone;
  53 static DEFINE_SPINLOCK(ioc_gone_lock);
  54
  55 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  56 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  57 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  58
  59 #define ASYNC                   (0)
  60 #define SYNC                    (1)
  61
  62 #define sample_valid(samples)   ((samples) > 80)
  63
  64 /*
  65  * Most of our rbtree usage is for sorting with min extraction, so
  66  * if we cache the leftmost node we don't have to walk down the tree
  67  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  68  * move this into the elevator for the rq sorting as well.
  69  */
  70 struct cfq_rb_root {
  71         struct rb_root rb;
  72         struct rb_node *left;
  73 };
  74 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
  75
  76 /*
  77  * Per block device queue structure
  78  */
  79 struct cfq_data {
  80         struct request_queue *queue;
  81
  82         /*
  83          * rr list of queues with requests and the count of them
  84          */
  85         struct cfq_rb_root service_tree;
  86         unsigned int busy_queues;
  87         /*
  88          * Used to track any pending rt requests so we can pre-empt current
  89          * non-RT cfqq in service when this value is non-zero.
  90          */
  91         unsigned int busy_rt_queues;
  92
  93         int rq_in_driver;
  94         int sync_flight;
  95
  96         /*
  97          * queue-depth detection
  98          */
  99         int rq_queued;
 100         int hw_tag;
 101         int hw_tag_samples;
 102         int rq_in_driver_peak;
 103
 104         /*
 105          * idle window management
 106          */
 107         struct timer_list idle_slice_timer;
 108         struct work_struct unplug_work;
 109
 110         struct cfq_queue *active_queue;
 111         struct cfq_io_context *active_cic;
 112
 113         /*
 114          * async queue for each priority case
 115          */
 116         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 117         struct cfq_queue *async_idle_cfqq;
 118
 119         sector_t last_position;
 120         unsigned long last_end_request;
 121
 122         /*
 123          * tunables, see top of file
 124          */
 125         unsigned int cfq_quantum;
 126         unsigned int cfq_fifo_expire[2];
 127         unsigned int cfq_back_penalty;
 128         unsigned int cfq_back_max;
 129         unsigned int cfq_slice[2];
 130         unsigned int cfq_slice_async_rq;
 131         unsigned int cfq_slice_idle;
 132
 133         struct list_head cic_list;
 134 };
 135
 136 /*
 137  * Per process-grouping structure
 138  */
 139 struct cfq_queue {
 140         /* reference count */
 141         atomic_t ref;
 142         /* various state flags, see below */
 143         unsigned int flags;
 144         /* parent cfq_data */
 145         struct cfq_data *cfqd;
 146         /* service_tree member */
 147         struct rb_node rb_node;
 148         /* service_tree key */
 149         unsigned long rb_key;
 150         /* sorted list of pending requests */
 151         struct rb_root sort_list;
 152         /* if fifo isn't expired, next request to serve */
 153         struct request *next_rq;
 154         /* requests queued in sort_list */
 155         int queued[2];
 156         /* currently allocated requests */
 157         int allocated[2];
 158         /* fifo list of requests in sort_list */
 159         struct list_head fifo;
 160
 161         unsigned long slice_end;
 162         long slice_resid;
 163         unsigned int slice_dispatch;
 164
 165         /* pending metadata requests */
 166         int meta_pending;
 167         /* number of requests that are on the dispatch list or inside driver */
 168         int dispatched;
 169
 170         /* io prio of this group */
 171         unsigned short ioprio, org_ioprio;
 172         unsigned short ioprio_class, org_ioprio_class;
 173
 174         pid_t pid;
 175 };
 176
 177 enum cfqq_state_flags {
 178         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 179         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 180         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 181         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
 182         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 183         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 184         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 185         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 186         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 187         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 188 };
 189
 190 #define CFQ_CFQQ_FNS(name)                                              \
 191 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 192 {                                                                       \
 193         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 194 }                                                                       \
 195 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 196 {                                                                       \
 197         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 198 }                                                                       \
 199 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 200 {                                                                       \
 201         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 202 }
 203
 204 CFQ_CFQQ_FNS(on_rr);
 205 CFQ_CFQQ_FNS(wait_request);
 206 CFQ_CFQQ_FNS(must_dispatch);
 207 CFQ_CFQQ_FNS(must_alloc);
 208 CFQ_CFQQ_FNS(must_alloc_slice);
 209 CFQ_CFQQ_FNS(fifo_expire);
 210 CFQ_CFQQ_FNS(idle_window);
 211 CFQ_CFQQ_FNS(prio_changed);
 212 CFQ_CFQQ_FNS(slice_new);
 213 CFQ_CFQQ_FNS(sync);
 214 #undef CFQ_CFQQ_FNS
 215
 216 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 217         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
 218 #define cfq_log(cfqd, fmt, args...)     \
 219         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 220
 221 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 222 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
 223                                        struct io_context *, gfp_t);
 224 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
 225                                                 struct io_context *);
 226
 227 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
 228                                             int is_sync)
 229 {
 230         return cic->cfqq[!!is_sync];
 231 }
 232
 233 static inline void cic_set_cfqq(struct cfq_io_context *cic,
 234                                 struct cfq_queue *cfqq, int is_sync)
 235 {
 236         cic->cfqq[!!is_sync] = cfqq;
 237 }
 238
 239 /*
 240  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 241  * set (in which case it could also be direct WRITE).
 242  */
 243 static inline int cfq_bio_sync(struct bio *bio)
 244 {
 245         if (bio_data_dir(bio) == READ || bio_sync(bio))
 246                 return 1;
 247
 248         return 0;
 249 }
 250
 251 /*
 252  * scheduler run of queue, if there are requests pending and no one in the
 253  * driver that will restart queueing
 254  */
 255 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 256 {
 257         if (cfqd->busy_queues) {
 258                 cfq_log(cfqd, "schedule dispatch");
 259                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
 260         }
 261 }
 262
 263 static int cfq_queue_empty(struct request_queue *q)
 264 {
 265         struct cfq_data *cfqd = q->elevator->elevator_data;
 266
 267         return !cfqd->busy_queues;
 268 }
 269
 270 /*
 271  * Scale schedule slice based on io priority. Use the sync time slice only
 272  * if a queue is marked sync and has sync io queued. A sync queue with async
 273  * io only, should not get full sync slice length.
 274  */
 275 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
 276                                  unsigned short prio)
 277 {
 278         const int base_slice = cfqd->cfq_slice[sync];
 279
 280         WARN_ON(prio >= IOPRIO_BE_NR);
 281
 282         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 283 }
 284
 285 static inline int
 286 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 287 {
 288         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 289 }
 290
 291 static inline void
 292 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 293 {
 294         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 295         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 296 }
 297
 298 /*
 299  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 300  * isn't valid until the first request from the dispatch is activated
 301  * and the slice time set.
 302  */
 303 static inline int cfq_slice_used(struct cfq_queue *cfqq)
 304 {
 305         if (cfq_cfqq_slice_new(cfqq))
 306                 return 0;
 307         if (time_before(jiffies, cfqq->slice_end))
 308                 return 0;
 309
 310         return 1;
 311 }
 312
 313 /*
 314  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 315  * We choose the request that is closest to the head right now. Distance
 316  * behind the head is penalized and only allowed to a certain extent.
 317  */
 318 static struct request *
 319 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 320 {
 321         sector_t last, s1, s2, d1 = 0, d2 = 0;
 322         unsigned long back_max;
 323 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 324 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 325         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 326
 327         if (rq1 == NULL || rq1 == rq2)
 328                 return rq2;
 329         if (rq2 == NULL)
 330                 return rq1;
 331
 332         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 333                 return rq1;
 334         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 335                 return rq2;
 336         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 337                 return rq1;
 338         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 339                 return rq2;
 340
 341         s1 = rq1->sector;
 342         s2 = rq2->sector;
 343
 344         last = cfqd->last_position;
 345
 346         /*
 347          * by definition, 1KiB is 2 sectors
 348          */
 349         back_max = cfqd->cfq_back_max * 2;
 350
 351         /*
 352          * Strict one way elevator _except_ in the case where we allow
 353          * short backward seeks which are biased as twice the cost of a
 354          * similar forward seek.
 355          */
 356         if (s1 >= last)
 357                 d1 = s1 - last;
 358         else if (s1 + back_max >= last)
 359                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 360         else
 361                 wrap |= CFQ_RQ1_WRAP;
 362
 363         if (s2 >= last)
 364                 d2 = s2 - last;
 365         else if (s2 + back_max >= last)
 366                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 367         else
 368                 wrap |= CFQ_RQ2_WRAP;
 369
 370         /* Found required data */
 371
 372         /*
 373          * By doing switch() on the bit mask "wrap" we avoid having to
 374          * check two variables for all permutations: --> faster!
 375          */
 376         switch (wrap) {
 377         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 378                 if (d1 < d2)
 379                         return rq1;
 380                 else if (d2 < d1)
 381                         return rq2;
 382                 else {
 383                         if (s1 >= s2)
 384                                 return rq1;
 385                         else
 386                                 return rq2;
 387                 }
 388
 389         case CFQ_RQ2_WRAP:
 390                 return rq1;
 391         case CFQ_RQ1_WRAP:
 392                 return rq2;
 393         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 394         default:
 395                 /*
 396                  * Since both rqs are wrapped,
 397                  * start with the one that's further behind head
 398                  * (--> only *one* back seek required),
 399                  * since back seek takes more time than forward.
 400                  */
 401                 if (s1 <= s2)
 402                         return rq1;
 403                 else
 404                         return rq2;
 405         }
 406 }
 407
 408 /*
 409  * The below is leftmost cache rbtree addon
 410  */
 411 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
 412 {
 413         if (!root->left)
 414                 root->left = rb_first(&root->rb);
 415
 416         if (root->left)
 417                 return rb_entry(root->left, struct cfq_queue, rb_node);
 418
 419         return NULL;
 420 }
 421
 422 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
 423 {
 424         if (root->left == n)
 425                 root->left = NULL;
 426
 427         rb_erase(n, &root->rb);
 428         RB_CLEAR_NODE(n);
 429 }
 430
 431 /*
 432  * would be nice to take fifo expire time into account as well
 433  */
 434 static struct request *
 435 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 436                   struct request *last)
 437 {
 438         struct rb_node *rbnext = rb_next(&last->rb_node);
 439         struct rb_node *rbprev = rb_prev(&last->rb_node);
 440         struct request *next = NULL, *prev = NULL;
 441
 442         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 443
 444         if (rbprev)
 445                 prev = rb_entry_rq(rbprev);
 446
 447         if (rbnext)
 448                 next = rb_entry_rq(rbnext);
 449         else {
 450                 rbnext = rb_first(&cfqq->sort_list);
 451                 if (rbnext && rbnext != &last->rb_node)
 452                         next = rb_entry_rq(rbnext);
 453         }
 454
 455         return cfq_choose_req(cfqd, next, prev);
 456 }
 457
 458 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
 459                                       struct cfq_queue *cfqq)
 460 {
 461         /*
 462          * just an approximation, should be ok.
 463          */
 464         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
 465                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
 466 }
 467
 468 /*
 469  * The cfqd->service_tree holds all pending cfq_queue's that have
 470  * requests waiting to be processed. It is sorted in the order that
 471  * we will service the queues.
 472  */
 473 static void cfq_service_tree_add(struct cfq_data *cfqd,
 474                                     struct cfq_queue *cfqq, int add_front)
 475 {
 476         struct rb_node **p, *parent;
 477         struct cfq_queue *__cfqq;
 478         unsigned long rb_key;
 479         int left;
 480
 481         if (cfq_class_idle(cfqq)) {
 482                 rb_key = CFQ_IDLE_DELAY;
 483                 parent = rb_last(&cfqd->service_tree.rb);
 484                 if (parent && parent != &cfqq->rb_node) {
 485                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 486                         rb_key += __cfqq->rb_key;
 487                 } else
 488                         rb_key += jiffies;
 489         } else if (!add_front) {
 490                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
 491                 rb_key += cfqq->slice_resid;
 492                 cfqq->slice_resid = 0;
 493         } else
 494                 rb_key = 0;
 495
 496         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
 497                 /*
 498                  * same position, nothing more to do
 499                  */
 500                 if (rb_key == cfqq->rb_key)
 501                         return;
 502
 503                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 504         }
 505
 506         left = 1;
 507         parent = NULL;
 508         p = &cfqd->service_tree.rb.rb_node;
 509         while (*p) {
 510                 struct rb_node **n;
 511
 512                 parent = *p;
 513                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 514
 515                 /*
 516                  * sort RT queues first, we always want to give
 517                  * preference to them. IDLE queues goes to the back.
 518                  * after that, sort on the next service time.
 519                  */
 520                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
 521                         n = &(*p)->rb_left;
 522                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
 523                         n = &(*p)->rb_right;
 524                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
 525                         n = &(*p)->rb_left;
 526                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
 527                         n = &(*p)->rb_right;
 528                 else if (rb_key < __cfqq->rb_key)
 529                         n = &(*p)->rb_left;
 530                 else
 531                         n = &(*p)->rb_right;
 532
 533                 if (n == &(*p)->rb_right)
 534                         left = 0;
 535
 536                 p = n;
 537         }
 538
 539         if (left)
 540                 cfqd->service_tree.left = &cfqq->rb_node;
 541
 542         cfqq->rb_key = rb_key;
 543         rb_link_node(&cfqq->rb_node, parent, p);
 544         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
 545 }
 546
 547 /*
 548  * Update cfqq's position in the service tree.
 549  */
 550 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 551 {
 552         /*
 553          * Resorting requires the cfqq to be on the RR list already.
 554          */
 555         if (cfq_cfqq_on_rr(cfqq))
 556                 cfq_service_tree_add(cfqd, cfqq, 0);
 557 }
 558
 559 /*
 560  * add to busy list of queues for service, trying to be fair in ordering
 561  * the pending list according to last request service
 562  */
 563 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 564 {
 565         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
 566         BUG_ON(cfq_cfqq_on_rr(cfqq));
 567         cfq_mark_cfqq_on_rr(cfqq);
 568         cfqd->busy_queues++;
 569         if (cfq_class_rt(cfqq))
 570                 cfqd->busy_rt_queues++;
 571
 572         cfq_resort_rr_list(cfqd, cfqq);
 573 }
 574
 575 /*
 576  * Called when the cfqq no longer has requests pending, remove it from
 577  * the service tree.
 578  */
 579 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 580 {
 581         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
 582         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 583         cfq_clear_cfqq_on_rr(cfqq);
 584
 585         if (!RB_EMPTY_NODE(&cfqq->rb_node))
 586                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 587
 588         BUG_ON(!cfqd->busy_queues);
 589         cfqd->busy_queues--;
 590         if (cfq_class_rt(cfqq))
 591                 cfqd->busy_rt_queues--;
 592 }
 593
 594 /*
 595  * rb tree support functions
 596  */
 597 static void cfq_del_rq_rb(struct request *rq)
 598 {
 599         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 600         struct cfq_data *cfqd = cfqq->cfqd;
 601         const int sync = rq_is_sync(rq);
 602
 603         BUG_ON(!cfqq->queued[sync]);
 604         cfqq->queued[sync]--;
 605
 606         elv_rb_del(&cfqq->sort_list, rq);
 607
 608         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 609                 cfq_del_cfqq_rr(cfqd, cfqq);
 610 }
 611
 612 static void cfq_add_rq_rb(struct request *rq)
 613 {
 614         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 615         struct cfq_data *cfqd = cfqq->cfqd;
 616         struct request *__alias;
 617
 618         cfqq->queued[rq_is_sync(rq)]++;
 619
 620         /*
 621          * looks a little odd, but the first insert might return an alias.
 622          * if that happens, put the alias on the dispatch list
 623          */
 624         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 625                 cfq_dispatch_insert(cfqd->queue, __alias);
 626
 627         if (!cfq_cfqq_on_rr(cfqq))
 628                 cfq_add_cfqq_rr(cfqd, cfqq);
 629
 630         /*
 631          * check if this request is a better next-serve candidate
 632          */
 633         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
 634         BUG_ON(!cfqq->next_rq);
 635 }
 636
 637 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 638 {
 639         elv_rb_del(&cfqq->sort_list, rq);
 640         cfqq->queued[rq_is_sync(rq)]--;
 641         cfq_add_rq_rb(rq);
 642 }
 643
 644 static struct request *
 645 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 646 {
 647         struct task_struct *tsk = current;
 648         struct cfq_io_context *cic;
 649         struct cfq_queue *cfqq;
 650
 651         cic = cfq_cic_lookup(cfqd, tsk->io_context);
 652         if (!cic)
 653                 return NULL;
 654
 655         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 656         if (cfqq) {
 657                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 658
 659                 return elv_rb_find(&cfqq->sort_list, sector);
 660         }
 661
 662         return NULL;
 663 }
 664
 665 static void cfq_activate_request(struct request_queue *q, struct request *rq)
 666 {
 667         struct cfq_data *cfqd = q->elevator->elevator_data;
 668
 669         cfqd->rq_in_driver++;
 670         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
 671                                                 cfqd->rq_in_driver);
 672
 673         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
 674 }
 675
 676 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
 677 {
 678         struct cfq_data *cfqd = q->elevator->elevator_data;
 679
 680         WARN_ON(!cfqd->rq_in_driver);
 681         cfqd->rq_in_driver--;
 682         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
 683                                                 cfqd->rq_in_driver);
 684 }
 685
 686 static void cfq_remove_request(struct request *rq)
 687 {
 688         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 689
 690         if (cfqq->next_rq == rq)
 691                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 692
 693         list_del_init(&rq->queuelist);
 694         cfq_del_rq_rb(rq);
 695
 696         cfqq->cfqd->rq_queued--;
 697         if (rq_is_meta(rq)) {
 698                 WARN_ON(!cfqq->meta_pending);
 699                 cfqq->meta_pending--;
 700         }
 701 }
 702
 703 static int cfq_merge(struct request_queue *q, struct request **req,
 704                      struct bio *bio)
 705 {
 706         struct cfq_data *cfqd = q->elevator->elevator_data;
 707         struct request *__rq;
 708
 709         __rq = cfq_find_rq_fmerge(cfqd, bio);
 710         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 711                 *req = __rq;
 712                 return ELEVATOR_FRONT_MERGE;
 713         }
 714
 715         return ELEVATOR_NO_MERGE;
 716 }
 717
 718 static void cfq_merged_request(struct request_queue *q, struct request *req,
 719                                int type)
 720 {
 721         if (type == ELEVATOR_FRONT_MERGE) {
 722                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 723
 724                 cfq_reposition_rq_rb(cfqq, req);
 725         }
 726 }
 727
 728 static void
 729 cfq_merged_requests(struct request_queue *q, struct request *rq,
 730                     struct request *next)
 731 {
 732         /*
 733          * reposition in fifo if next is older than rq
 734          */
 735         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 736             time_before(next->start_time, rq->start_time))
 737                 list_move(&rq->queuelist, &next->queuelist);
 738
 739         cfq_remove_request(next);
 740 }
 741
 742 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
 743                            struct bio *bio)
 744 {
 745         struct cfq_data *cfqd = q->elevator->elevator_data;
 746         struct cfq_io_context *cic;
 747         struct cfq_queue *cfqq;
 748
 749         /*
 750          * Disallow merge of a sync bio into an async request.
 751          */
 752         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
 753                 return 0;
 754
 755         /*
 756          * Lookup the cfqq that this bio will be queued with. Allow
 757          * merge only if rq is queued there.
 758          */
 759         cic = cfq_cic_lookup(cfqd, current->io_context);
 760         if (!cic)
 761                 return 0;
 762
 763         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 764         if (cfqq == RQ_CFQQ(rq))
 765                 return 1;
 766
 767         return 0;
 768 }
 769
 770 static void __cfq_set_active_queue(struct cfq_data *cfqd,
 771                                    struct cfq_queue *cfqq)
 772 {
 773         if (cfqq) {
 774                 cfq_log_cfqq(cfqd, cfqq, "set_active");
 775                 cfqq->slice_end = 0;
 776                 cfqq->slice_dispatch = 0;
 777
 778                 cfq_clear_cfqq_wait_request(cfqq);
 779                 cfq_clear_cfqq_must_dispatch(cfqq);
 780                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 781                 cfq_clear_cfqq_fifo_expire(cfqq);
 782                 cfq_mark_cfqq_slice_new(cfqq);
 783
 784                 del_timer(&cfqd->idle_slice_timer);
 785         }
 786
 787         cfqd->active_queue = cfqq;
 788 }
 789
 790 /*
 791  * current cfqq expired its slice (or was too idle), select new one
 792  */
 793 static void
 794 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 795                     int timed_out)
 796 {
 797         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
 798
 799         if (cfq_cfqq_wait_request(cfqq))
 800                 del_timer(&cfqd->idle_slice_timer);
 801
 802         cfq_clear_cfqq_wait_request(cfqq);
 803
 804         /*
 805          * store what was left of this slice, if the queue idled/timed out
 806          */
 807         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
 808                 cfqq->slice_resid = cfqq->slice_end - jiffies;
 809                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
 810         }
 811
 812         cfq_resort_rr_list(cfqd, cfqq);
 813
 814         if (cfqq == cfqd->active_queue)
 815                 cfqd->active_queue = NULL;
 816
 817         if (cfqd->active_cic) {
 818                 put_io_context(cfqd->active_cic->ioc);
 819                 cfqd->active_cic = NULL;
 820         }
 821 }
 822
 823 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
 824 {
 825         struct cfq_queue *cfqq = cfqd->active_queue;
 826
 827         if (cfqq)
 828                 __cfq_slice_expired(cfqd, cfqq, timed_out);
 829 }
 830
 831 /*
 832  * Get next queue for service. Unless we have a queue preemption,
 833  * we'll simply select the first cfqq in the service tree.
 834  */
 835 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 836 {
 837         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
 838                 return NULL;
 839
 840         return cfq_rb_first(&cfqd->service_tree);
 841 }
 842
 843 /*
 844  * Get and set a new active queue for service.
 845  */
 846 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 847 {
 848         struct cfq_queue *cfqq;
 849
 850         cfqq = cfq_get_next_queue(cfqd);
 851         __cfq_set_active_queue(cfqd, cfqq);
 852         return cfqq;
 853 }
 854
 855 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
 856                                           struct request *rq)
 857 {
 858         if (rq->sector >= cfqd->last_position)
 859                 return rq->sector - cfqd->last_position;
 860         else
 861                 return cfqd->last_position - rq->sector;
 862 }
 863
 864 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
 865 {
 866         struct cfq_io_context *cic = cfqd->active_cic;
 867
 868         if (!sample_valid(cic->seek_samples))
 869                 return 0;
 870
 871         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
 872 }
 873
 874 static int cfq_close_cooperator(struct cfq_data *cfq_data,
 875                                 struct cfq_queue *cfqq)
 876 {
 877         /*
 878          * We should notice if some of the queues are cooperating, eg
 879          * working closely on the same area of the disk. In that case,
 880          * we can group them together and don't waste time idling.
 881          */
 882         return 0;
 883 }
 884
 885 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
 886
 887 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
 888 {
 889         struct cfq_queue *cfqq = cfqd->active_queue;
 890         struct cfq_io_context *cic;
 891         unsigned long sl;
 892
 893         /*
 894          * SSD device without seek penalty, disable idling. But only do so
 895          * for devices that support queuing, otherwise we still have a problem
 896          * with sync vs async workloads.
 897          */
 898         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
 899                 return;
 900
 901         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 902         WARN_ON(cfq_cfqq_slice_new(cfqq));
 903
 904         /*
 905          * idle is disabled, either manually or by past process history
 906          */
 907         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
 908                 return;
 909
 910         /*
 911          * still requests with the driver, don't idle
 912          */
 913         if (cfqd->rq_in_driver)
 914                 return;
 915
 916         /*
 917          * task has exited, don't wait
 918          */
 919         cic = cfqd->active_cic;
 920         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
 921                 return;
 922
 923         /*
 924          * See if this prio level has a good candidate
 925          */
 926         if (cfq_close_cooperator(cfqd, cfqq) &&
 927             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
 928                 return;
 929
 930         cfq_mark_cfqq_wait_request(cfqq);
 931
 932         /*
 933          * we don't want to idle for seeks, but we do want to allow
 934          * fair distribution of slice time for a process doing back-to-back
 935          * seeks. so allow a little bit of time for him to submit a new rq
 936          */
 937         sl = cfqd->cfq_slice_idle;
 938         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 939                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
 940
 941         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 942         cfq_log(cfqd, "arm_idle: %lu", sl);
 943 }
 944
 945 /*
 946  * Move request from internal lists to the request queue dispatch list.
 947  */
 948 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
 949 {
 950         struct cfq_data *cfqd = q->elevator->elevator_data;
 951         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 952
 953         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
 954
 955         cfq_remove_request(rq);
 956         cfqq->dispatched++;
 957         elv_dispatch_sort(q, rq);
 958
 959         if (cfq_cfqq_sync(cfqq))
 960                 cfqd->sync_flight++;
 961 }
 962
 963 /*
 964  * return expired entry, or NULL to just start from scratch in rbtree
 965  */
 966 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 967 {
 968         struct cfq_data *cfqd = cfqq->cfqd;
 969         struct request *rq;
 970         int fifo;
 971
 972         if (cfq_cfqq_fifo_expire(cfqq))
 973                 return NULL;
 974
 975         cfq_mark_cfqq_fifo_expire(cfqq);
 976
 977         if (list_empty(&cfqq->fifo))
 978                 return NULL;
 979
 980         fifo = cfq_cfqq_sync(cfqq);
 981         rq = rq_entry_fifo(cfqq->fifo.next);
 982
 983         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
 984                 rq = NULL;
 985
 986         cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
 987         return rq;
 988 }
 989
 990 static inline int
 991 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 992 {
 993         const int base_rq = cfqd->cfq_slice_async_rq;
 994
 995         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 996
 997         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 998 }
 999
1000 /*
1001  * Select a queue for service. If we have a current active queue,
1002  * check whether to continue servicing it, or retrieve and set a new one.
1003  */
1004 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1005 {
1006         struct cfq_queue *cfqq;
1007
1008         cfqq = cfqd->active_queue;
1009         if (!cfqq)
1010                 goto new_queue;
1011
1012         /*
1013          * The active queue has run out of time, expire it and select new.
1014          */
1015         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1016                 goto expire;
1017
1018         /*
1019          * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1020          * cfqq.
1021          */
1022         if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1023                 /*
1024                  * We simulate this as cfqq timed out so that it gets to bank
1025                  * the remaining of its time slice.
1026                  */
1027                 cfq_log_cfqq(cfqd, cfqq, "preempt");
1028                 cfq_slice_expired(cfqd, 1);
1029                 goto new_queue;
1030         }
1031
1032         /*
1033          * The active queue has requests and isn't expired, allow it to
1034          * dispatch.
1035          */
1036         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1037                 goto keep_queue;
1038
1039         /*
1040          * No requests pending. If the active queue still has requests in
1041          * flight or is idling for a new request, allow either of these
1042          * conditions to happen (or time out) before selecting a new queue.
1043          */
1044         if (timer_pending(&cfqd->idle_slice_timer) ||
1045             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1046                 cfqq = NULL;
1047                 goto keep_queue;
1048         }
1049
1050 expire:
1051         cfq_slice_expired(cfqd, 0);
1052 new_queue:
1053         cfqq = cfq_set_active_queue(cfqd);
1054 keep_queue:
1055         return cfqq;
1056 }
1057
1058 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1059 {
1060         int dispatched = 0;
1061
1062         while (cfqq->next_rq) {
1063                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1064                 dispatched++;
1065         }
1066
1067         BUG_ON(!list_empty(&cfqq->fifo));
1068         return dispatched;
1069 }
1070
1071 /*
1072  * Drain our current requests. Used for barriers and when switching
1073  * io schedulers on-the-fly.
1074  */
1075 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1076 {
1077         struct cfq_queue *cfqq;
1078         int dispatched = 0;
1079
1080         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1081                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1082
1083         cfq_slice_expired(cfqd, 0);
1084
1085         BUG_ON(cfqd->busy_queues);
1086
1087         cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
1088         return dispatched;
1089 }
1090
1091 /*
1092  * Dispatch a request from cfqq, moving them to the request queue
1093  * dispatch list.
1094  */
1095 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1096 {
1097         struct request *rq;
1098
1099         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1100
1101         /*
1102          * follow expired path, else get first next available
1103          */
1104         rq = cfq_check_fifo(cfqq);
1105         if (!rq)
1106                 rq = cfqq->next_rq;
1107
1108         /*
1109          * insert request into driver dispatch list
1110          */
1111         cfq_dispatch_insert(cfqd->queue, rq);
1112
1113         if (!cfqd->active_cic) {
1114                 struct cfq_io_context *cic = RQ_CIC(rq);
1115
1116                 atomic_inc(&cic->ioc->refcount);
1117                 cfqd->active_cic = cic;
1118         }
1119 }
1120
1121 /*
1122  * Find the cfqq that we need to service and move a request from that to the
1123  * dispatch list
1124  */
1125 static int cfq_dispatch_requests(struct request_queue *q, int force)
1126 {
1127         struct cfq_data *cfqd = q->elevator->elevator_data;
1128         struct cfq_queue *cfqq;
1129         unsigned int max_dispatch;
1130
1131         if (!cfqd->busy_queues)
1132                 return 0;
1133
1134         if (unlikely(force))
1135                 return cfq_forced_dispatch(cfqd);
1136
1137         cfqq = cfq_select_queue(cfqd);
1138         if (!cfqq)
1139                 return 0;
1140
1141         /*
1142          * If this is an async queue and we have sync IO in flight, let it wait
1143          */
1144         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1145                 return 0;
1146
1147         max_dispatch = cfqd->cfq_quantum;
1148         if (cfq_class_idle(cfqq))
1149                 max_dispatch = 1;
1150
1151         /*
1152          * Does this cfqq already have too much IO in flight?
1153          */
1154         if (cfqq->dispatched >= max_dispatch) {
1155                 /*
1156                  * idle queue must always only have a single IO in flight
1157                  */
1158                 if (cfq_class_idle(cfqq))
1159                         return 0;
1160
1161                 /*
1162                  * We have other queues, don't allow more IO from this one
1163                  */
1164                 if (cfqd->busy_queues > 1)
1165                         return 0;
1166
1167                 /*
1168                  * we are the only queue, allow up to 4 times of 'quantum'
1169                  */
1170                 if (cfqq->dispatched >= 4 * max_dispatch)
1171                         return 0;
1172         }
1173
1174         /*
1175          * Dispatch a request from this cfqq
1176          */
1177         cfq_dispatch_request(cfqd, cfqq);
1178         cfqq->slice_dispatch++;
1179         cfq_clear_cfqq_must_dispatch(cfqq);
1180
1181         /*
1182          * expire an async queue immediately if it has used up its slice. idle
1183          * queue always expire after 1 dispatch round.
1184          */
1185         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1186             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1187             cfq_class_idle(cfqq))) {
1188                 cfqq->slice_end = jiffies + 1;
1189                 cfq_slice_expired(cfqd, 0);
1190         }
1191
1192         cfq_log(cfqd, "dispatched a request");
1193         return 1;
1194 }
1195
1196 /*
1197  * task holds one reference to the queue, dropped when task exits. each rq
1198  * in-flight on this queue also holds a reference, dropped when rq is freed.
1199  *
1200  * queue lock must be held here.
1201  */
1202 static void cfq_put_queue(struct cfq_queue *cfqq)
1203 {
1204         struct cfq_data *cfqd = cfqq->cfqd;
1205
1206         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1207
1208         if (!atomic_dec_and_test(&cfqq->ref))
1209                 return;
1210
1211         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1212         BUG_ON(rb_first(&cfqq->sort_list));
1213         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1214         BUG_ON(cfq_cfqq_on_rr(cfqq));
1215
1216         if (unlikely(cfqd->active_queue == cfqq)) {
1217                 __cfq_slice_expired(cfqd, cfqq, 0);
1218                 cfq_schedule_dispatch(cfqd);
1219         }
1220
1221         kmem_cache_free(cfq_pool, cfqq);
1222 }
1223
1224 /*
1225  * Must always be called with the rcu_read_lock() held
1226  */
1227 static void
1228 __call_for_each_cic(struct io_context *ioc,
1229                     void (*func)(struct io_context *, struct cfq_io_context *))
1230 {
1231         struct cfq_io_context *cic;
1232         struct hlist_node *n;
1233
1234         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1235                 func(ioc, cic);
1236 }
1237
1238 /*
1239  * Call func for each cic attached to this ioc.
1240  */
1241 static void
1242 call_for_each_cic(struct io_context *ioc,
1243                   void (*func)(struct io_context *, struct cfq_io_context *))
1244 {
1245         rcu_read_lock();
1246         __call_for_each_cic(ioc, func);
1247         rcu_read_unlock();
1248 }
1249
1250 static void cfq_cic_free_rcu(struct rcu_head *head)
1251 {
1252         struct cfq_io_context *cic;
1253
1254         cic = container_of(head, struct cfq_io_context, rcu_head);
1255
1256         kmem_cache_free(cfq_ioc_pool, cic);
1257         elv_ioc_count_dec(ioc_count);
1258
1259         if (ioc_gone) {
1260                 /*
1261                  * CFQ scheduler is exiting, grab exit lock and check
1262                  * the pending io context count. If it hits zero,
1263                  * complete ioc_gone and set it back to NULL
1264                  */
1265                 spin_lock(&ioc_gone_lock);
1266                 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1267                         complete(ioc_gone);
1268                         ioc_gone = NULL;
1269                 }
1270                 spin_unlock(&ioc_gone_lock);
1271         }
1272 }
1273
1274 static void cfq_cic_free(struct cfq_io_context *cic)
1275 {
1276         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1277 }
1278
1279 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1280 {
1281         unsigned long flags;
1282
1283         BUG_ON(!cic->dead_key);
1284
1285         spin_lock_irqsave(&ioc->lock, flags);
1286         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1287         hlist_del_rcu(&cic->cic_list);
1288         spin_unlock_irqrestore(&ioc->lock, flags);
1289
1290         cfq_cic_free(cic);
1291 }
1292
1293 /*
1294  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1295  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1296  * and ->trim() which is called with the task lock held
1297  */
1298 static void cfq_free_io_context(struct io_context *ioc)
1299 {
1300         /*
1301          * ioc->refcount is zero here, or we are called from elv_unregister(),
1302          * so no more cic's are allowed to be linked into this ioc.  So it
1303          * should be ok to iterate over the known list, we will see all cic's
1304          * since no new ones are added.
1305          */
1306         __call_for_each_cic(ioc, cic_free_func);
1307 }
1308
1309 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1310 {
1311         if (unlikely(cfqq == cfqd->active_queue)) {
1312                 __cfq_slice_expired(cfqd, cfqq, 0);
1313                 cfq_schedule_dispatch(cfqd);
1314         }
1315
1316         cfq_put_queue(cfqq);
1317 }
1318
1319 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1320                                          struct cfq_io_context *cic)
1321 {
1322         struct io_context *ioc = cic->ioc;
1323
1324         list_del_init(&cic->queue_list);
1325
1326         /*
1327          * Make sure key == NULL is seen for dead queues
1328          */
1329         smp_wmb();
1330         cic->dead_key = (unsigned long) cic->key;
1331         cic->key = NULL;
1332
1333         if (ioc->ioc_data == cic)
1334                 rcu_assign_pointer(ioc->ioc_data, NULL);
1335
1336         if (cic->cfqq[ASYNC]) {
1337                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1338                 cic->cfqq[ASYNC] = NULL;
1339         }
1340
1341         if (cic->cfqq[SYNC]) {
1342                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1343                 cic->cfqq[SYNC] = NULL;
1344         }
1345 }
1346
1347 static void cfq_exit_single_io_context(struct io_context *ioc,
1348                                        struct cfq_io_context *cic)
1349 {
1350         struct cfq_data *cfqd = cic->key;
1351
1352         if (cfqd) {
1353                 struct request_queue *q = cfqd->queue;
1354                 unsigned long flags;
1355
1356                 spin_lock_irqsave(q->queue_lock, flags);
1357
1358                 /*
1359                  * Ensure we get a fresh copy of the ->key to prevent
1360                  * race between exiting task and queue
1361                  */
1362                 smp_read_barrier_depends();
1363                 if (cic->key)
1364                         __cfq_exit_single_io_context(cfqd, cic);
1365
1366                 spin_unlock_irqrestore(q->queue_lock, flags);
1367         }
1368 }
1369
1370 /*
1371  * The process that ioc belongs to has exited, we need to clean up
1372  * and put the internal structures we have that belongs to that process.
1373  */
1374 static void cfq_exit_io_context(struct io_context *ioc)
1375 {
1376         call_for_each_cic(ioc, cfq_exit_single_io_context);
1377 }
1378
1379 static struct cfq_io_context *
1380 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1381 {
1382         struct cfq_io_context *cic;
1383
1384         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1385                                                         cfqd->queue->node);
1386         if (cic) {
1387                 cic->last_end_request = jiffies;
1388                 INIT_LIST_HEAD(&cic->queue_list);
1389                 INIT_HLIST_NODE(&cic->cic_list);
1390                 cic->dtor = cfq_free_io_context;
1391                 cic->exit = cfq_exit_io_context;
1392                 elv_ioc_count_inc(ioc_count);
1393         }
1394
1395         return cic;
1396 }
1397
1398 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1399 {
1400         struct task_struct *tsk = current;
1401         int ioprio_class;
1402
1403         if (!cfq_cfqq_prio_changed(cfqq))
1404                 return;
1405
1406         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1407         switch (ioprio_class) {
1408         default:
1409                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1410         case IOPRIO_CLASS_NONE:
1411                 /*
1412                  * no prio set, inherit CPU scheduling settings
1413                  */
1414                 cfqq->ioprio = task_nice_ioprio(tsk);
1415                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1416                 break;
1417         case IOPRIO_CLASS_RT:
1418                 cfqq->ioprio = task_ioprio(ioc);
1419                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1420                 break;
1421         case IOPRIO_CLASS_BE:
1422                 cfqq->ioprio = task_ioprio(ioc);
1423                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1424                 break;
1425         case IOPRIO_CLASS_IDLE:
1426                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1427                 cfqq->ioprio = 7;
1428                 cfq_clear_cfqq_idle_window(cfqq);
1429                 break;
1430         }
1431
1432         /*
1433          * keep track of original prio settings in case we have to temporarily
1434          * elevate the priority of this queue
1435          */
1436         cfqq->org_ioprio = cfqq->ioprio;
1437         cfqq->org_ioprio_class = cfqq->ioprio_class;
1438         cfq_clear_cfqq_prio_changed(cfqq);
1439 }
1440
1441 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1442 {
1443         struct cfq_data *cfqd = cic->key;
1444         struct cfq_queue *cfqq;
1445         unsigned long flags;
1446
1447         if (unlikely(!cfqd))
1448                 return;
1449
1450         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1451
1452         cfqq = cic->cfqq[ASYNC];
1453         if (cfqq) {
1454                 struct cfq_queue *new_cfqq;
1455                 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
1456                 if (new_cfqq) {
1457                         cic->cfqq[ASYNC] = new_cfqq;
1458                         cfq_put_queue(cfqq);
1459                 }
1460         }
1461
1462         cfqq = cic->cfqq[SYNC];
1463         if (cfqq)
1464                 cfq_mark_cfqq_prio_changed(cfqq);
1465
1466         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1467 }
1468
1469 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1470 {
1471         call_for_each_cic(ioc, changed_ioprio);
1472         ioc->ioprio_changed = 0;
1473 }
1474
1475 static struct cfq_queue *
1476 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1477                      struct io_context *ioc, gfp_t gfp_mask)
1478 {
1479         struct cfq_queue *cfqq, *new_cfqq = NULL;
1480         struct cfq_io_context *cic;
1481
1482 retry:
1483         cic = cfq_cic_lookup(cfqd, ioc);
1484         /* cic always exists here */
1485         cfqq = cic_to_cfqq(cic, is_sync);
1486
1487         if (!cfqq) {
1488                 if (new_cfqq) {
1489                         cfqq = new_cfqq;
1490                         new_cfqq = NULL;
1491                 } else if (gfp_mask & __GFP_WAIT) {
1492                         /*
1493                          * Inform the allocator of the fact that we will
1494                          * just repeat this allocation if it fails, to allow
1495                          * the allocator to do whatever it needs to attempt to
1496                          * free memory.
1497                          */
1498                         spin_unlock_irq(cfqd->queue->queue_lock);
1499                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1500                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1501                                         cfqd->queue->node);
1502                         spin_lock_irq(cfqd->queue->queue_lock);
1503                         goto retry;
1504                 } else {
1505                         cfqq = kmem_cache_alloc_node(cfq_pool,
1506                                         gfp_mask | __GFP_ZERO,
1507                                         cfqd->queue->node);
1508                         if (!cfqq)
1509                                 goto out;
1510                 }
1511
1512                 RB_CLEAR_NODE(&cfqq->rb_node);
1513                 INIT_LIST_HEAD(&cfqq->fifo);
1514
1515                 atomic_set(&cfqq->ref, 0);
1516                 cfqq->cfqd = cfqd;
1517
1518                 cfq_mark_cfqq_prio_changed(cfqq);
1519
1520                 cfq_init_prio_data(cfqq, ioc);
1521
1522                 if (is_sync) {
1523                         if (!cfq_class_idle(cfqq))
1524                                 cfq_mark_cfqq_idle_window(cfqq);
1525                         cfq_mark_cfqq_sync(cfqq);
1526                 }
1527                 cfqq->pid = current->pid;
1528                 cfq_log_cfqq(cfqd, cfqq, "alloced");
1529         }
1530
1531         if (new_cfqq)
1532                 kmem_cache_free(cfq_pool, new_cfqq);
1533
1534 out:
1535         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1536         return cfqq;
1537 }
1538
1539 static struct cfq_queue **
1540 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1541 {
1542         switch (ioprio_class) {
1543         case IOPRIO_CLASS_RT:
1544                 return &cfqd->async_cfqq[0][ioprio];
1545         case IOPRIO_CLASS_BE:
1546                 return &cfqd->async_cfqq[1][ioprio];
1547         case IOPRIO_CLASS_IDLE:
1548                 return &cfqd->async_idle_cfqq;
1549         default:
1550                 BUG();
1551         }
1552 }
1553
1554 static struct cfq_queue *
1555 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1556               gfp_t gfp_mask)
1557 {
1558         const int ioprio = task_ioprio(ioc);
1559         const int ioprio_class = task_ioprio_class(ioc);
1560         struct cfq_queue **async_cfqq = NULL;
1561         struct cfq_queue *cfqq = NULL;
1562
1563         if (!is_sync) {
1564                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1565                 cfqq = *async_cfqq;
1566         }
1567
1568         if (!cfqq) {
1569                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1570                 if (!cfqq)
1571                         return NULL;
1572         }
1573
1574         /*
1575          * pin the queue now that it's allocated, scheduler exit will prune it
1576          */
1577         if (!is_sync && !(*async_cfqq)) {
1578                 atomic_inc(&cfqq->ref);
1579                 *async_cfqq = cfqq;
1580         }
1581
1582         atomic_inc(&cfqq->ref);
1583         return cfqq;
1584 }
1585
1586 /*
1587  * We drop cfq io contexts lazily, so we may find a dead one.
1588  */
1589 static void
1590 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1591                   struct cfq_io_context *cic)
1592 {
1593         unsigned long flags;
1594
1595         WARN_ON(!list_empty(&cic->queue_list));
1596
1597         spin_lock_irqsave(&ioc->lock, flags);
1598
1599         BUG_ON(ioc->ioc_data == cic);
1600
1601         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1602         hlist_del_rcu(&cic->cic_list);
1603         spin_unlock_irqrestore(&ioc->lock, flags);
1604
1605         cfq_cic_free(cic);
1606 }
1607
1608 static struct cfq_io_context *
1609 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1610 {
1611         struct cfq_io_context *cic;
1612         unsigned long flags;
1613         void *k;
1614
1615         if (unlikely(!ioc))
1616                 return NULL;
1617
1618         rcu_read_lock();
1619
1620         /*
1621          * we maintain a last-hit cache, to avoid browsing over the tree
1622          */
1623         cic = rcu_dereference(ioc->ioc_data);
1624         if (cic && cic->key == cfqd) {
1625                 rcu_read_unlock();
1626                 return cic;
1627         }
1628
1629         do {
1630                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1631                 rcu_read_unlock();
1632                 if (!cic)
1633                         break;
1634                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1635                 k = cic->key;
1636                 if (unlikely(!k)) {
1637                         cfq_drop_dead_cic(cfqd, ioc, cic);
1638                         rcu_read_lock();
1639                         continue;
1640                 }
1641
1642                 spin_lock_irqsave(&ioc->lock, flags);
1643                 rcu_assign_pointer(ioc->ioc_data, cic);
1644                 spin_unlock_irqrestore(&ioc->lock, flags);
1645                 break;
1646         } while (1);
1647
1648         return cic;
1649 }
1650
1651 /*
1652  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1653  * the process specific cfq io context when entered from the block layer.
1654  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1655  */
1656 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1657                         struct cfq_io_context *cic, gfp_t gfp_mask)
1658 {
1659         unsigned long flags;
1660         int ret;
1661
1662         ret = radix_tree_preload(gfp_mask);
1663         if (!ret) {
1664                 cic->ioc = ioc;
1665                 cic->key = cfqd;
1666
1667                 spin_lock_irqsave(&ioc->lock, flags);
1668                 ret = radix_tree_insert(&ioc->radix_root,
1669                                                 (unsigned long) cfqd, cic);
1670                 if (!ret)
1671                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1672                 spin_unlock_irqrestore(&ioc->lock, flags);
1673
1674                 radix_tree_preload_end();
1675
1676                 if (!ret) {
1677                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1678                         list_add(&cic->queue_list, &cfqd->cic_list);
1679                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1680                 }
1681         }
1682
1683         if (ret)
1684                 printk(KERN_ERR "cfq: cic link failed!\n");
1685
1686         return ret;
1687 }
1688
1689 /*
1690  * Setup general io context and cfq io context. There can be several cfq
1691  * io contexts per general io context, if this process is doing io to more
1692  * than one device managed by cfq.
1693  */
1694 static struct cfq_io_context *
1695 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1696 {
1697         struct io_context *ioc = NULL;
1698         struct cfq_io_context *cic;
1699
1700         might_sleep_if(gfp_mask & __GFP_WAIT);
1701
1702         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1703         if (!ioc)
1704                 return NULL;
1705
1706         cic = cfq_cic_lookup(cfqd, ioc);
1707         if (cic)
1708                 goto out;
1709
1710         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1711         if (cic == NULL)
1712                 goto err;
1713
1714         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1715                 goto err_free;
1716
1717 out:
1718         smp_read_barrier_depends();
1719         if (unlikely(ioc->ioprio_changed))
1720                 cfq_ioc_set_ioprio(ioc);
1721
1722         return cic;
1723 err_free:
1724         cfq_cic_free(cic);
1725 err:
1726         put_io_context(ioc);
1727         return NULL;
1728 }
1729
1730 static void
1731 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1732 {
1733         unsigned long elapsed = jiffies - cic->last_end_request;
1734         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1735
1736         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1737         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1738         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1739 }
1740
1741 static void
1742 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1743                        struct request *rq)
1744 {
1745         sector_t sdist;
1746         u64 total;
1747
1748         if (cic->last_request_pos < rq->sector)
1749                 sdist = rq->sector - cic->last_request_pos;
1750         else
1751                 sdist = cic->last_request_pos - rq->sector;
1752
1753         /*
1754          * Don't allow the seek distance to get too large from the
1755          * odd fragment, pagein, etc
1756          */
1757         if (cic->seek_samples <= 60) /* second&third seek */
1758                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1759         else
1760                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1761
1762         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1763         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1764         total = cic->seek_total + (cic->seek_samples/2);
1765         do_div(total, cic->seek_samples);
1766         cic->seek_mean = (sector_t)total;
1767 }
1768
1769 /*
1770  * Disable idle window if the process thinks too long or seeks so much that
1771  * it doesn't matter
1772  */
1773 static void
1774 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1775                        struct cfq_io_context *cic)
1776 {
1777         int old_idle, enable_idle;
1778
1779         /*
1780          * Don't idle for async or idle io prio class
1781          */
1782         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1783                 return;
1784
1785         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1786
1787         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1788             (cfqd->hw_tag && CIC_SEEKY(cic)))
1789                 enable_idle = 0;
1790         else if (sample_valid(cic->ttime_samples)) {
1791                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1792                         enable_idle = 0;
1793                 else
1794                         enable_idle = 1;
1795         }
1796
1797         if (old_idle != enable_idle) {
1798                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1799                 if (enable_idle)
1800                         cfq_mark_cfqq_idle_window(cfqq);
1801                 else
1802                         cfq_clear_cfqq_idle_window(cfqq);
1803         }
1804 }
1805
1806 /*
1807  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1808  * no or if we aren't sure, a 1 will cause a preempt.
1809  */
1810 static int
1811 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1812                    struct request *rq)
1813 {
1814         struct cfq_queue *cfqq;
1815
1816         cfqq = cfqd->active_queue;
1817         if (!cfqq)
1818                 return 0;
1819
1820         if (cfq_slice_used(cfqq))
1821                 return 1;
1822
1823         if (cfq_class_idle(new_cfqq))
1824                 return 0;
1825
1826         if (cfq_class_idle(cfqq))
1827                 return 1;
1828
1829         /*
1830          * if the new request is sync, but the currently running queue is
1831          * not, let the sync request have priority.
1832          */
1833         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1834                 return 1;
1835
1836         /*
1837          * So both queues are sync. Let the new request get disk time if
1838          * it's a metadata request and the current queue is doing regular IO.
1839          */
1840         if (rq_is_meta(rq) && !cfqq->meta_pending)
1841                 return 1;
1842
1843         /*
1844          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
1845          */
1846         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
1847                 return 1;
1848
1849         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1850                 return 0;
1851
1852         /*
1853          * if this request is as-good as one we would expect from the
1854          * current cfqq, let it preempt
1855          */
1856         if (cfq_rq_close(cfqd, rq))
1857                 return 1;
1858
1859         return 0;
1860 }
1861
1862 /*
1863  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1864  * let it have half of its nominal slice.
1865  */
1866 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1867 {
1868         cfq_log_cfqq(cfqd, cfqq, "preempt");
1869         cfq_slice_expired(cfqd, 1);
1870
1871         /*
1872          * Put the new queue at the front of the of the current list,
1873          * so we know that it will be selected next.
1874          */
1875         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1876
1877         cfq_service_tree_add(cfqd, cfqq, 1);
1878
1879         cfqq->slice_end = 0;
1880         cfq_mark_cfqq_slice_new(cfqq);
1881 }
1882
1883 /*
1884  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1885  * something we should do about it
1886  */
1887 static void
1888 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1889                 struct request *rq)
1890 {
1891         struct cfq_io_context *cic = RQ_CIC(rq);
1892
1893         cfqd->rq_queued++;
1894         if (rq_is_meta(rq))
1895                 cfqq->meta_pending++;
1896
1897         cfq_update_io_thinktime(cfqd, cic);
1898         cfq_update_io_seektime(cfqd, cic, rq);
1899         cfq_update_idle_window(cfqd, cfqq, cic);
1900
1901         cic->last_request_pos = rq->sector + rq->nr_sectors;
1902
1903         if (cfqq == cfqd->active_queue) {
1904                 /*
1905                  * Remember that we saw a request from this process, but
1906                  * don't start queuing just yet. Otherwise we risk seeing lots
1907                  * of tiny requests, because we disrupt the normal plugging
1908                  * and merging.
1909                  */
1910                 if (cfq_cfqq_wait_request(cfqq))
1911                         cfq_mark_cfqq_must_dispatch(cfqq);
1912         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1913                 /*
1914                  * not the active queue - expire current slice if it is
1915                  * idle and has expired it's mean thinktime or this new queue
1916                  * has some old slice time left and is of higher priority or
1917                  * this new queue is RT and the current one is BE
1918                  */
1919                 cfq_preempt_queue(cfqd, cfqq);
1920                 blk_start_queueing(cfqd->queue);
1921         }
1922 }
1923
1924 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1925 {
1926         struct cfq_data *cfqd = q->elevator->elevator_data;
1927         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1928
1929         cfq_log_cfqq(cfqd, cfqq, "insert_request");
1930         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1931
1932         cfq_add_rq_rb(rq);
1933
1934         list_add_tail(&rq->queuelist, &cfqq->fifo);
1935
1936         cfq_rq_enqueued(cfqd, cfqq, rq);
1937 }
1938
1939 /*
1940  * Update hw_tag based on peak queue depth over 50 samples under
1941  * sufficient load.
1942  */
1943 static void cfq_update_hw_tag(struct cfq_data *cfqd)
1944 {
1945         if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1946                 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1947
1948         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1949             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1950                 return;
1951
1952         if (cfqd->hw_tag_samples++ < 50)
1953                 return;
1954
1955         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1956                 cfqd->hw_tag = 1;
1957         else
1958                 cfqd->hw_tag = 0;
1959
1960         cfqd->hw_tag_samples = 0;
1961         cfqd->rq_in_driver_peak = 0;
1962 }
1963
1964 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1965 {
1966         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1967         struct cfq_data *cfqd = cfqq->cfqd;
1968         const int sync = rq_is_sync(rq);
1969         unsigned long now;
1970
1971         now = jiffies;
1972         cfq_log_cfqq(cfqd, cfqq, "complete");
1973
1974         cfq_update_hw_tag(cfqd);
1975
1976         WARN_ON(!cfqd->rq_in_driver);
1977         WARN_ON(!cfqq->dispatched);
1978         cfqd->rq_in_driver--;
1979         cfqq->dispatched--;
1980
1981         if (cfq_cfqq_sync(cfqq))
1982                 cfqd->sync_flight--;
1983
1984         if (!cfq_class_idle(cfqq))
1985                 cfqd->last_end_request = now;
1986
1987         if (sync)
1988                 RQ_CIC(rq)->last_end_request = now;
1989
1990         /*
1991          * If this is the active queue, check if it needs to be expired,
1992          * or if we want to idle in case it has no pending requests.
1993          */
1994         if (cfqd->active_queue == cfqq) {
1995                 if (cfq_cfqq_slice_new(cfqq)) {
1996                         cfq_set_prio_slice(cfqd, cfqq);
1997                         cfq_clear_cfqq_slice_new(cfqq);
1998                 }
1999                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2000                         cfq_slice_expired(cfqd, 1);
2001                 else if (sync && !rq_noidle(rq) &&
2002                          RB_EMPTY_ROOT(&cfqq->sort_list)) {
2003                         cfq_arm_slice_timer(cfqd);
2004                 }
2005         }
2006
2007         if (!cfqd->rq_in_driver)
2008                 cfq_schedule_dispatch(cfqd);
2009 }
2010
2011 /*
2012  * we temporarily boost lower priority queues if they are holding fs exclusive
2013  * resources. they are boosted to normal prio (CLASS_BE/4)
2014  */
2015 static void cfq_prio_boost(struct cfq_queue *cfqq)
2016 {
2017         if (has_fs_excl()) {
2018                 /*
2019                  * boost idle prio on transactions that would lock out other
2020                  * users of the filesystem
2021                  */
2022                 if (cfq_class_idle(cfqq))
2023                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2024                 if (cfqq->ioprio > IOPRIO_NORM)
2025                         cfqq->ioprio = IOPRIO_NORM;
2026         } else {
2027                 /*
2028                  * check if we need to unboost the queue
2029                  */
2030                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2031                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2032                 if (cfqq->ioprio != cfqq->org_ioprio)
2033                         cfqq->ioprio = cfqq->org_ioprio;
2034         }
2035 }
2036
2037 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2038 {
2039         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
2040             !cfq_cfqq_must_alloc_slice(cfqq)) {
2041                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2042                 return ELV_MQUEUE_MUST;
2043         }
2044
2045         return ELV_MQUEUE_MAY;
2046 }
2047
2048 static int cfq_may_queue(struct request_queue *q, int rw)
2049 {
2050         struct cfq_data *cfqd = q->elevator->elevator_data;
2051         struct task_struct *tsk = current;
2052         struct cfq_io_context *cic;
2053         struct cfq_queue *cfqq;
2054
2055         /*
2056          * don't force setup of a queue from here, as a call to may_queue
2057          * does not necessarily imply that a request actually will be queued.
2058          * so just lookup a possibly existing queue, or return 'may queue'
2059          * if that fails
2060          */
2061         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2062         if (!cic)
2063                 return ELV_MQUEUE_MAY;
2064
2065         cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
2066         if (cfqq) {
2067                 cfq_init_prio_data(cfqq, cic->ioc);
2068                 cfq_prio_boost(cfqq);
2069
2070                 return __cfq_may_queue(cfqq);
2071         }
2072
2073         return ELV_MQUEUE_MAY;
2074 }
2075
2076 /*
2077  * queue lock held here
2078  */
2079 static void cfq_put_request(struct request *rq)
2080 {
2081         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2082
2083         if (cfqq) {
2084                 const int rw = rq_data_dir(rq);
2085
2086                 BUG_ON(!cfqq->allocated[rw]);
2087                 cfqq->allocated[rw]--;
2088
2089                 put_io_context(RQ_CIC(rq)->ioc);
2090
2091                 rq->elevator_private = NULL;
2092                 rq->elevator_private2 = NULL;
2093
2094                 cfq_put_queue(cfqq);
2095         }
2096 }
2097
2098 /*
2099  * Allocate cfq data structures associated with this request.
2100  */
2101 static int
2102 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2103 {
2104         struct cfq_data *cfqd = q->elevator->elevator_data;
2105         struct cfq_io_context *cic;
2106         const int rw = rq_data_dir(rq);
2107         const int is_sync = rq_is_sync(rq);
2108         struct cfq_queue *cfqq;
2109         unsigned long flags;
2110
2111         might_sleep_if(gfp_mask & __GFP_WAIT);
2112
2113         cic = cfq_get_io_context(cfqd, gfp_mask);
2114
2115         spin_lock_irqsave(q->queue_lock, flags);
2116
2117         if (!cic)
2118                 goto queue_fail;
2119
2120         cfqq = cic_to_cfqq(cic, is_sync);
2121         if (!cfqq) {
2122                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2123
2124                 if (!cfqq)
2125                         goto queue_fail;
2126
2127                 cic_set_cfqq(cic, cfqq, is_sync);
2128         }
2129
2130         cfqq->allocated[rw]++;
2131         cfq_clear_cfqq_must_alloc(cfqq);
2132         atomic_inc(&cfqq->ref);
2133
2134         spin_unlock_irqrestore(q->queue_lock, flags);
2135
2136         rq->elevator_private = cic;
2137         rq->elevator_private2 = cfqq;
2138         return 0;
2139
2140 queue_fail:
2141         if (cic)
2142                 put_io_context(cic->ioc);
2143
2144         cfq_schedule_dispatch(cfqd);
2145         spin_unlock_irqrestore(q->queue_lock, flags);
2146         cfq_log(cfqd, "set_request fail");
2147         return 1;
2148 }
2149
2150 static void cfq_kick_queue(struct work_struct *work)
2151 {
2152         struct cfq_data *cfqd =
2153                 container_of(work, struct cfq_data, unplug_work);
2154         struct request_queue *q = cfqd->queue;
2155         unsigned long flags;
2156
2157         spin_lock_irqsave(q->queue_lock, flags);
2158         blk_start_queueing(q);
2159         spin_unlock_irqrestore(q->queue_lock, flags);
2160 }
2161
2162 /*
2163  * Timer running if the active_queue is currently idling inside its time slice
2164  */
2165 static void cfq_idle_slice_timer(unsigned long data)
2166 {
2167         struct cfq_data *cfqd = (struct cfq_data *) data;
2168         struct cfq_queue *cfqq;
2169         unsigned long flags;
2170         int timed_out = 1;
2171
2172         cfq_log(cfqd, "idle timer fired");
2173
2174         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2175
2176         cfqq = cfqd->active_queue;
2177         if (cfqq) {
2178                 timed_out = 0;
2179
2180                 /*
2181                  * We saw a request before the queue expired, let it through
2182                  */
2183                 if (cfq_cfqq_must_dispatch(cfqq))
2184                         goto out_kick;
2185
2186                 /*
2187                  * expired
2188                  */
2189                 if (cfq_slice_used(cfqq))
2190                         goto expire;
2191
2192                 /*
2193                  * only expire and reinvoke request handler, if there are
2194                  * other queues with pending requests
2195                  */
2196                 if (!cfqd->busy_queues)
2197                         goto out_cont;
2198
2199                 /*
2200                  * not expired and it has a request pending, let it dispatch
2201                  */
2202                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2203                         goto out_kick;
2204         }
2205 expire:
2206         cfq_slice_expired(cfqd, timed_out);
2207 out_kick:
2208         cfq_schedule_dispatch(cfqd);
2209 out_cont:
2210         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2211 }
2212
2213 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2214 {
2215         del_timer_sync(&cfqd->idle_slice_timer);
2216         cancel_work_sync(&cfqd->unplug_work);
2217 }
2218
2219 static void cfq_put_async_queues(struct cfq_data *cfqd)
2220 {
2221         int i;
2222
2223         for (i = 0; i < IOPRIO_BE_NR; i++) {
2224                 if (cfqd->async_cfqq[0][i])
2225                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2226                 if (cfqd->async_cfqq[1][i])
2227                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2228         }
2229
2230         if (cfqd->async_idle_cfqq)
2231                 cfq_put_queue(cfqd->async_idle_cfqq);
2232 }
2233
2234 static void cfq_exit_queue(struct elevator_queue *e)
2235 {
2236         struct cfq_data *cfqd = e->elevator_data;
2237         struct request_queue *q = cfqd->queue;
2238
2239         cfq_shutdown_timer_wq(cfqd);
2240
2241         spin_lock_irq(q->queue_lock);
2242
2243         if (cfqd->active_queue)
2244                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2245
2246         while (!list_empty(&cfqd->cic_list)) {
2247                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2248                                                         struct cfq_io_context,
2249                                                         queue_list);
2250
2251                 __cfq_exit_single_io_context(cfqd, cic);
2252         }
2253
2254         cfq_put_async_queues(cfqd);
2255
2256         spin_unlock_irq(q->queue_lock);
2257
2258         cfq_shutdown_timer_wq(cfqd);
2259
2260         kfree(cfqd);
2261 }
2262
2263 static void *cfq_init_queue(struct request_queue *q)
2264 {
2265         struct cfq_data *cfqd;
2266
2267         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2268         if (!cfqd)
2269                 return NULL;
2270
2271         cfqd->service_tree = CFQ_RB_ROOT;
2272         INIT_LIST_HEAD(&cfqd->cic_list);
2273
2274         cfqd->queue = q;
2275
2276         init_timer(&cfqd->idle_slice_timer);
2277         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2278         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2279
2280         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2281
2282         cfqd->last_end_request = jiffies;
2283         cfqd->cfq_quantum = cfq_quantum;
2284         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2285         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2286         cfqd->cfq_back_max = cfq_back_max;
2287         cfqd->cfq_back_penalty = cfq_back_penalty;
2288         cfqd->cfq_slice[0] = cfq_slice_async;
2289         cfqd->cfq_slice[1] = cfq_slice_sync;
2290         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2291         cfqd->cfq_slice_idle = cfq_slice_idle;
2292         cfqd->hw_tag = 1;
2293
2294         return cfqd;
2295 }
2296
2297 static void cfq_slab_kill(void)
2298 {
2299         /*
2300          * Caller already ensured that pending RCU callbacks are completed,
2301          * so we should have no busy allocations at this point.
2302          */
2303         if (cfq_pool)
2304                 kmem_cache_destroy(cfq_pool);
2305         if (cfq_ioc_pool)
2306                 kmem_cache_destroy(cfq_ioc_pool);
2307 }
2308
2309 static int __init cfq_slab_setup(void)
2310 {
2311         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2312         if (!cfq_pool)
2313                 goto fail;
2314
2315         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2316         if (!cfq_ioc_pool)
2317                 goto fail;
2318
2319         return 0;
2320 fail:
2321         cfq_slab_kill();
2322         return -ENOMEM;
2323 }
2324
2325 /*
2326  * sysfs parts below -->
2327  */
2328 static ssize_t
2329 cfq_var_show(unsigned int var, char *page)
2330 {
2331         return sprintf(page, "%d\n", var);
2332 }
2333
2334 static ssize_t
2335 cfq_var_store(unsigned int *var, const char *page, size_t count)
2336 {
2337         char *p = (char *) page;
2338
2339         *var = simple_strtoul(p, &p, 10);
2340         return count;
2341 }
2342
2343 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2344 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2345 {                                                                       \
2346         struct cfq_data *cfqd = e->elevator_data;                       \
2347         unsigned int __data = __VAR;                                    \
2348         if (__CONV)                                                     \
2349                 __data = jiffies_to_msecs(__data);                      \
2350         return cfq_var_show(__data, (page));                            \
2351 }
2352 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2353 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2354 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2355 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2356 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2357 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2358 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2359 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2360 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2361 #undef SHOW_FUNCTION
2362
2363 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2364 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2365 {                                                                       \
2366         struct cfq_data *cfqd = e->elevator_data;                       \
2367         unsigned int __data;                                            \
2368         int ret = cfq_var_store(&__data, (page), count);                \
2369         if (__data < (MIN))                                             \
2370                 __data = (MIN);                                         \
2371         else if (__data > (MAX))                                        \
2372                 __data = (MAX);                                         \
2373         if (__CONV)                                                     \
2374                 *(__PTR) = msecs_to_jiffies(__data);                    \
2375         else                                                            \
2376                 *(__PTR) = __data;                                      \
2377         return ret;                                                     \
2378 }
2379 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2380 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2381                 UINT_MAX, 1);
2382 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2383                 UINT_MAX, 1);
2384 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2385 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2386                 UINT_MAX, 0);
2387 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2388 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2389 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2390 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2391                 UINT_MAX, 0);
2392 #undef STORE_FUNCTION
2393
2394 #define CFQ_ATTR(name) \
2395         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2396
2397 static struct elv_fs_entry cfq_attrs[] = {
2398         CFQ_ATTR(quantum),
2399         CFQ_ATTR(fifo_expire_sync),
2400         CFQ_ATTR(fifo_expire_async),
2401         CFQ_ATTR(back_seek_max),
2402         CFQ_ATTR(back_seek_penalty),
2403         CFQ_ATTR(slice_sync),
2404         CFQ_ATTR(slice_async),
2405         CFQ_ATTR(slice_async_rq),
2406         CFQ_ATTR(slice_idle),
2407         __ATTR_NULL
2408 };
2409
2410 static struct elevator_type iosched_cfq = {
2411         .ops = {
2412                 .elevator_merge_fn =            cfq_merge,
2413                 .elevator_merged_fn =           cfq_merged_request,
2414                 .elevator_merge_req_fn =        cfq_merged_requests,
2415                 .elevator_allow_merge_fn =      cfq_allow_merge,
2416                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2417                 .elevator_add_req_fn =          cfq_insert_request,
2418                 .elevator_activate_req_fn =     cfq_activate_request,
2419                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2420                 .elevator_queue_empty_fn =      cfq_queue_empty,
2421                 .elevator_completed_req_fn =    cfq_completed_request,
2422                 .elevator_former_req_fn =       elv_rb_former_request,
2423                 .elevator_latter_req_fn =       elv_rb_latter_request,
2424                 .elevator_set_req_fn =          cfq_set_request,
2425                 .elevator_put_req_fn =          cfq_put_request,
2426                 .elevator_may_queue_fn =        cfq_may_queue,
2427                 .elevator_init_fn =             cfq_init_queue,
2428                 .elevator_exit_fn =             cfq_exit_queue,
2429                 .trim =                         cfq_free_io_context,
2430         },
2431         .elevator_attrs =       cfq_attrs,
2432         .elevator_name =        "cfq",
2433         .elevator_owner =       THIS_MODULE,
2434 };
2435
2436 static int __init cfq_init(void)
2437 {
2438         /*
2439          * could be 0 on HZ < 1000 setups
2440          */
2441         if (!cfq_slice_async)
2442                 cfq_slice_async = 1;
2443         if (!cfq_slice_idle)
2444                 cfq_slice_idle = 1;
2445
2446         if (cfq_slab_setup())
2447                 return -ENOMEM;
2448
2449         elv_register(&iosched_cfq);
2450
2451         return 0;
2452 }
2453
2454 static void __exit cfq_exit(void)
2455 {
2456         DECLARE_COMPLETION_ONSTACK(all_gone);
2457         elv_unregister(&iosched_cfq);
2458         ioc_gone = &all_gone;
2459         /* ioc_gone's update must be visible before reading ioc_count */
2460         smp_wmb();
2461
2462         /*
2463          * this also protects us from entering cfq_slab_kill() with
2464          * pending RCU callbacks
2465          */
2466         if (elv_ioc_count_read(ioc_count))
2467                 wait_for_completion(&all_gone);
2468         cfq_slab_kill();
2469 }
2470
2471 module_init(cfq_init);
2472 module_exit(cfq_exit);
2473
2474 MODULE_AUTHOR("Jens Axboe");
2475 MODULE_LICENSE("GPL");
2476 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");