]> bbs.cooldavid.org Git - net-next-2.6.git/blame - block/cfq-iosched.c
backing-dev: Handle class_create() failure
[net-next-2.6.git] / block / cfq-iosched.c
CommitLineData
1da177e4 1/*
1da177e4
LT
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 *
0fe23479 7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
1da177e4 8 */
1da177e4 9#include <linux/module.h>
1cc9be68
AV
10#include <linux/blkdev.h>
11#include <linux/elevator.h>
ad5ebd2f 12#include <linux/jiffies.h>
1da177e4 13#include <linux/rbtree.h>
22e2c507 14#include <linux/ioprio.h>
7b679138 15#include <linux/blktrace_api.h>
25bc6b07 16#include "blk-cgroup.h"
1da177e4
LT
17
18/*
19 * tunables
20 */
fe094d98 21/* max queue in one round of service */
abc3c744 22static const int cfq_quantum = 8;
64100099 23static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
fe094d98
JA
24/* maximum backwards seek, in KiB */
25static const int cfq_back_max = 16 * 1024;
26/* penalty of a backwards seek */
27static const int cfq_back_penalty = 2;
64100099 28static const int cfq_slice_sync = HZ / 10;
3b18152c 29static int cfq_slice_async = HZ / 25;
64100099 30static const int cfq_slice_async_rq = 2;
caaa5f9f 31static int cfq_slice_idle = HZ / 125;
5db5d642
CZ
32static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33static const int cfq_hist_divisor = 4;
22e2c507 34
d9e7620e 35/*
0871714e 36 * offset from end of service tree
d9e7620e 37 */
0871714e 38#define CFQ_IDLE_DELAY (HZ / 5)
d9e7620e
JA
39
40/*
41 * below this threshold, we consider thinktime immediate
42 */
43#define CFQ_MIN_TT (2)
44
22e2c507 45#define CFQ_SLICE_SCALE (5)
45333d5a 46#define CFQ_HW_QUEUE_MIN (5)
25bc6b07 47#define CFQ_SERVICE_SHIFT 12
22e2c507 48
3dde36dd 49#define CFQQ_SEEK_THR (sector_t)(8 * 100)
e9ce335d 50#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
41647e7a 51#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
3dde36dd 52#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
ae54abed 53
fe094d98
JA
54#define RQ_CIC(rq) \
55 ((struct cfq_io_context *) (rq)->elevator_private)
7b679138 56#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
1da177e4 57
e18b890b
CL
58static struct kmem_cache *cfq_pool;
59static struct kmem_cache *cfq_ioc_pool;
1da177e4 60
245b2e70 61static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
334e94de 62static struct completion *ioc_gone;
9a11b4ed 63static DEFINE_SPINLOCK(ioc_gone_lock);
334e94de 64
22e2c507
JA
65#define CFQ_PRIO_LISTS IOPRIO_BE_NR
66#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
22e2c507
JA
67#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68
206dc69b 69#define sample_valid(samples) ((samples) > 80)
1fa8f6d6 70#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
206dc69b 71
cc09e299
JA
72/*
73 * Most of our rbtree usage is for sorting with min extraction, so
74 * if we cache the leftmost node we don't have to walk down the tree
75 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76 * move this into the elevator for the rq sorting as well.
77 */
78struct cfq_rb_root {
79 struct rb_root rb;
80 struct rb_node *left;
aa6f6a3d 81 unsigned count;
73e9ffdd 82 unsigned total_weight;
1fa8f6d6 83 u64 min_vdisktime;
25bc6b07 84 struct rb_node *active;
cc09e299 85};
73e9ffdd
RK
86#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
87 .count = 0, .min_vdisktime = 0, }
cc09e299 88
6118b70b
JA
89/*
90 * Per process-grouping structure
91 */
92struct cfq_queue {
93 /* reference count */
94 atomic_t ref;
95 /* various state flags, see below */
96 unsigned int flags;
97 /* parent cfq_data */
98 struct cfq_data *cfqd;
99 /* service_tree member */
100 struct rb_node rb_node;
101 /* service_tree key */
102 unsigned long rb_key;
103 /* prio tree member */
104 struct rb_node p_node;
105 /* prio tree root we belong to, if any */
106 struct rb_root *p_root;
107 /* sorted list of pending requests */
108 struct rb_root sort_list;
109 /* if fifo isn't expired, next request to serve */
110 struct request *next_rq;
111 /* requests queued in sort_list */
112 int queued[2];
113 /* currently allocated requests */
114 int allocated[2];
115 /* fifo list of requests in sort_list */
116 struct list_head fifo;
117
dae739eb
VG
118 /* time when queue got scheduled in to dispatch first request. */
119 unsigned long dispatch_start;
f75edf2d 120 unsigned int allocated_slice;
c4081ba5 121 unsigned int slice_dispatch;
dae739eb
VG
122 /* time when first request from queue completed and slice started. */
123 unsigned long slice_start;
6118b70b
JA
124 unsigned long slice_end;
125 long slice_resid;
6118b70b
JA
126
127 /* pending metadata requests */
128 int meta_pending;
129 /* number of requests that are on the dispatch list or inside driver */
130 int dispatched;
131
132 /* io prio of this group */
133 unsigned short ioprio, org_ioprio;
134 unsigned short ioprio_class, org_ioprio_class;
135
c4081ba5
RK
136 pid_t pid;
137
3dde36dd 138 u32 seek_history;
b2c18e1e
JM
139 sector_t last_request_pos;
140
aa6f6a3d 141 struct cfq_rb_root *service_tree;
df5fe3e8 142 struct cfq_queue *new_cfqq;
cdb16e8f 143 struct cfq_group *cfqg;
ae30c286 144 struct cfq_group *orig_cfqg;
22084190
VG
145 /* Sectors dispatched in current dispatch round */
146 unsigned long nr_sectors;
6118b70b
JA
147};
148
c0324a02 149/*
718eee05 150 * First index in the service_trees.
c0324a02
CZ
151 * IDLE is handled separately, so it has negative index
152 */
153enum wl_prio_t {
c0324a02 154 BE_WORKLOAD = 0,
615f0259
VG
155 RT_WORKLOAD = 1,
156 IDLE_WORKLOAD = 2,
c0324a02
CZ
157};
158
718eee05
CZ
159/*
160 * Second index in the service_trees.
161 */
162enum wl_type_t {
163 ASYNC_WORKLOAD = 0,
164 SYNC_NOIDLE_WORKLOAD = 1,
165 SYNC_WORKLOAD = 2
166};
167
cdb16e8f
VG
168/* This is per cgroup per device grouping structure */
169struct cfq_group {
1fa8f6d6
VG
170 /* group service_tree member */
171 struct rb_node rb_node;
172
173 /* group service_tree key */
174 u64 vdisktime;
25bc6b07 175 unsigned int weight;
1fa8f6d6
VG
176 bool on_st;
177
178 /* number of cfqq currently on this group */
179 int nr_cfqq;
180
58ff82f3
VG
181 /* Per group busy queus average. Useful for workload slice calc. */
182 unsigned int busy_queues_avg[2];
cdb16e8f
VG
183 /*
184 * rr lists of queues with requests, onle rr for each priority class.
185 * Counts are embedded in the cfq_rb_root
186 */
187 struct cfq_rb_root service_trees[2][3];
188 struct cfq_rb_root service_tree_idle;
dae739eb
VG
189
190 unsigned long saved_workload_slice;
191 enum wl_type_t saved_workload;
192 enum wl_prio_t saved_serving_prio;
25fb5169
VG
193 struct blkio_group blkg;
194#ifdef CONFIG_CFQ_GROUP_IOSCHED
195 struct hlist_node cfqd_node;
b1c35769 196 atomic_t ref;
25fb5169 197#endif
cdb16e8f 198};
718eee05 199
22e2c507
JA
200/*
201 * Per block device queue structure
202 */
1da177e4 203struct cfq_data {
165125e1 204 struct request_queue *queue;
1fa8f6d6
VG
205 /* Root service tree for cfq_groups */
206 struct cfq_rb_root grp_service_tree;
cdb16e8f 207 struct cfq_group root_group;
22e2c507 208
c0324a02
CZ
209 /*
210 * The priority currently being served
22e2c507 211 */
c0324a02 212 enum wl_prio_t serving_prio;
718eee05
CZ
213 enum wl_type_t serving_type;
214 unsigned long workload_expires;
cdb16e8f 215 struct cfq_group *serving_group;
8e550632 216 bool noidle_tree_requires_idle;
a36e71f9
JA
217
218 /*
219 * Each priority tree is sorted by next_request position. These
220 * trees are used when determining if two or more queues are
221 * interleaving requests (see cfq_close_cooperator).
222 */
223 struct rb_root prio_trees[CFQ_PRIO_LISTS];
224
22e2c507
JA
225 unsigned int busy_queues;
226
53c583d2
CZ
227 int rq_in_driver;
228 int rq_in_flight[2];
45333d5a
AC
229
230 /*
231 * queue-depth detection
232 */
233 int rq_queued;
25776e35 234 int hw_tag;
e459dd08
CZ
235 /*
236 * hw_tag can be
237 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
238 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
239 * 0 => no NCQ
240 */
241 int hw_tag_est_depth;
242 unsigned int hw_tag_samples;
1da177e4 243
22e2c507
JA
244 /*
245 * idle window management
246 */
247 struct timer_list idle_slice_timer;
23e018a1 248 struct work_struct unplug_work;
1da177e4 249
22e2c507
JA
250 struct cfq_queue *active_queue;
251 struct cfq_io_context *active_cic;
22e2c507 252
c2dea2d1
VT
253 /*
254 * async queue for each priority case
255 */
256 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
257 struct cfq_queue *async_idle_cfqq;
15c31be4 258
6d048f53 259 sector_t last_position;
1da177e4 260
1da177e4
LT
261 /*
262 * tunables, see top of file
263 */
264 unsigned int cfq_quantum;
22e2c507 265 unsigned int cfq_fifo_expire[2];
1da177e4
LT
266 unsigned int cfq_back_penalty;
267 unsigned int cfq_back_max;
22e2c507
JA
268 unsigned int cfq_slice[2];
269 unsigned int cfq_slice_async_rq;
270 unsigned int cfq_slice_idle;
963b72fc 271 unsigned int cfq_latency;
ae30c286 272 unsigned int cfq_group_isolation;
d9ff4187
AV
273
274 struct list_head cic_list;
1da177e4 275
6118b70b
JA
276 /*
277 * Fallback dummy cfqq for extreme OOM conditions
278 */
279 struct cfq_queue oom_cfqq;
365722bb 280
573412b2 281 unsigned long last_delayed_sync;
25fb5169
VG
282
283 /* List of cfq groups being managed on this device*/
284 struct hlist_head cfqg_list;
bb729bc9 285 struct rcu_head rcu;
1da177e4
LT
286};
287
25fb5169
VG
288static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
289
cdb16e8f
VG
290static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
291 enum wl_prio_t prio,
65b32a57 292 enum wl_type_t type)
c0324a02 293{
1fa8f6d6
VG
294 if (!cfqg)
295 return NULL;
296
c0324a02 297 if (prio == IDLE_WORKLOAD)
cdb16e8f 298 return &cfqg->service_tree_idle;
c0324a02 299
cdb16e8f 300 return &cfqg->service_trees[prio][type];
c0324a02
CZ
301}
302
3b18152c 303enum cfqq_state_flags {
b0b8d749
JA
304 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
305 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
b029195d 306 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
b0b8d749 307 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
b0b8d749
JA
308 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
309 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
310 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
44f7c160 311 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
91fac317 312 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
b3b6d040 313 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
ae54abed 314 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
76280aff 315 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
f75edf2d 316 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
3b18152c
JA
317};
318
319#define CFQ_CFQQ_FNS(name) \
320static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
321{ \
fe094d98 322 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
3b18152c
JA
323} \
324static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
325{ \
fe094d98 326 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
3b18152c
JA
327} \
328static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
329{ \
fe094d98 330 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
3b18152c
JA
331}
332
333CFQ_CFQQ_FNS(on_rr);
334CFQ_CFQQ_FNS(wait_request);
b029195d 335CFQ_CFQQ_FNS(must_dispatch);
3b18152c 336CFQ_CFQQ_FNS(must_alloc_slice);
3b18152c
JA
337CFQ_CFQQ_FNS(fifo_expire);
338CFQ_CFQQ_FNS(idle_window);
339CFQ_CFQQ_FNS(prio_changed);
44f7c160 340CFQ_CFQQ_FNS(slice_new);
91fac317 341CFQ_CFQQ_FNS(sync);
a36e71f9 342CFQ_CFQQ_FNS(coop);
ae54abed 343CFQ_CFQQ_FNS(split_coop);
76280aff 344CFQ_CFQQ_FNS(deep);
f75edf2d 345CFQ_CFQQ_FNS(wait_busy);
3b18152c
JA
346#undef CFQ_CFQQ_FNS
347
2868ef7b
VG
348#ifdef CONFIG_DEBUG_CFQ_IOSCHED
349#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
350 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
351 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
352 blkg_path(&(cfqq)->cfqg->blkg), ##args);
353
354#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
355 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
356 blkg_path(&(cfqg)->blkg), ##args); \
357
358#else
7b679138
JA
359#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
360 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
2868ef7b
VG
361#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
362#endif
7b679138
JA
363#define cfq_log(cfqd, fmt, args...) \
364 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
365
615f0259
VG
366/* Traverses through cfq group service trees */
367#define for_each_cfqg_st(cfqg, i, j, st) \
368 for (i = 0; i <= IDLE_WORKLOAD; i++) \
369 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
370 : &cfqg->service_tree_idle; \
371 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
372 (i == IDLE_WORKLOAD && j == 0); \
373 j++, st = i < IDLE_WORKLOAD ? \
374 &cfqg->service_trees[i][j]: NULL) \
375
376
c0324a02
CZ
377static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
378{
379 if (cfq_class_idle(cfqq))
380 return IDLE_WORKLOAD;
381 if (cfq_class_rt(cfqq))
382 return RT_WORKLOAD;
383 return BE_WORKLOAD;
384}
385
718eee05
CZ
386
387static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
388{
389 if (!cfq_cfqq_sync(cfqq))
390 return ASYNC_WORKLOAD;
391 if (!cfq_cfqq_idle_window(cfqq))
392 return SYNC_NOIDLE_WORKLOAD;
393 return SYNC_WORKLOAD;
394}
395
58ff82f3
VG
396static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
397 struct cfq_data *cfqd,
398 struct cfq_group *cfqg)
c0324a02
CZ
399{
400 if (wl == IDLE_WORKLOAD)
cdb16e8f 401 return cfqg->service_tree_idle.count;
c0324a02 402
cdb16e8f
VG
403 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
404 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
405 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
c0324a02
CZ
406}
407
f26bd1f0
VG
408static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
409 struct cfq_group *cfqg)
410{
411 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
412 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
413}
414
165125e1 415static void cfq_dispatch_insert(struct request_queue *, struct request *);
a6151c3a 416static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
fd0928df 417 struct io_context *, gfp_t);
4ac845a2 418static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
91fac317
VT
419 struct io_context *);
420
421static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
a6151c3a 422 bool is_sync)
91fac317 423{
a6151c3a 424 return cic->cfqq[is_sync];
91fac317
VT
425}
426
427static inline void cic_set_cfqq(struct cfq_io_context *cic,
a6151c3a 428 struct cfq_queue *cfqq, bool is_sync)
91fac317 429{
a6151c3a 430 cic->cfqq[is_sync] = cfqq;
91fac317
VT
431}
432
433/*
434 * We regard a request as SYNC, if it's either a read or has the SYNC bit
435 * set (in which case it could also be direct WRITE).
436 */
a6151c3a 437static inline bool cfq_bio_sync(struct bio *bio)
91fac317 438{
a6151c3a 439 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
91fac317 440}
1da177e4 441
99f95e52
AM
442/*
443 * scheduler run of queue, if there are requests pending and no one in the
444 * driver that will restart queueing
445 */
23e018a1 446static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
99f95e52 447{
7b679138
JA
448 if (cfqd->busy_queues) {
449 cfq_log(cfqd, "schedule dispatch");
23e018a1 450 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
7b679138 451 }
99f95e52
AM
452}
453
165125e1 454static int cfq_queue_empty(struct request_queue *q)
99f95e52
AM
455{
456 struct cfq_data *cfqd = q->elevator->elevator_data;
457
f04a6424 458 return !cfqd->rq_queued;
99f95e52
AM
459}
460
44f7c160
JA
461/*
462 * Scale schedule slice based on io priority. Use the sync time slice only
463 * if a queue is marked sync and has sync io queued. A sync queue with async
464 * io only, should not get full sync slice length.
465 */
a6151c3a 466static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
d9e7620e 467 unsigned short prio)
44f7c160 468{
d9e7620e 469 const int base_slice = cfqd->cfq_slice[sync];
44f7c160 470
d9e7620e
JA
471 WARN_ON(prio >= IOPRIO_BE_NR);
472
473 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
474}
44f7c160 475
d9e7620e
JA
476static inline int
477cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
478{
479 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
44f7c160
JA
480}
481
25bc6b07
VG
482static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
483{
484 u64 d = delta << CFQ_SERVICE_SHIFT;
485
486 d = d * BLKIO_WEIGHT_DEFAULT;
487 do_div(d, cfqg->weight);
488 return d;
489}
490
491static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
492{
493 s64 delta = (s64)(vdisktime - min_vdisktime);
494 if (delta > 0)
495 min_vdisktime = vdisktime;
496
497 return min_vdisktime;
498}
499
500static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
501{
502 s64 delta = (s64)(vdisktime - min_vdisktime);
503 if (delta < 0)
504 min_vdisktime = vdisktime;
505
506 return min_vdisktime;
507}
508
509static void update_min_vdisktime(struct cfq_rb_root *st)
510{
511 u64 vdisktime = st->min_vdisktime;
512 struct cfq_group *cfqg;
513
514 if (st->active) {
515 cfqg = rb_entry_cfqg(st->active);
516 vdisktime = cfqg->vdisktime;
517 }
518
519 if (st->left) {
520 cfqg = rb_entry_cfqg(st->left);
521 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
522 }
523
524 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
525}
526
5db5d642
CZ
527/*
528 * get averaged number of queues of RT/BE priority.
529 * average is updated, with a formula that gives more weight to higher numbers,
530 * to quickly follows sudden increases and decrease slowly
531 */
532
58ff82f3
VG
533static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
534 struct cfq_group *cfqg, bool rt)
5869619c 535{
5db5d642
CZ
536 unsigned min_q, max_q;
537 unsigned mult = cfq_hist_divisor - 1;
538 unsigned round = cfq_hist_divisor / 2;
58ff82f3 539 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
5db5d642 540
58ff82f3
VG
541 min_q = min(cfqg->busy_queues_avg[rt], busy);
542 max_q = max(cfqg->busy_queues_avg[rt], busy);
543 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
5db5d642 544 cfq_hist_divisor;
58ff82f3
VG
545 return cfqg->busy_queues_avg[rt];
546}
547
548static inline unsigned
549cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
550{
551 struct cfq_rb_root *st = &cfqd->grp_service_tree;
552
553 return cfq_target_latency * cfqg->weight / st->total_weight;
5db5d642
CZ
554}
555
44f7c160
JA
556static inline void
557cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
558{
5db5d642
CZ
559 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
560 if (cfqd->cfq_latency) {
58ff82f3
VG
561 /*
562 * interested queues (we consider only the ones with the same
563 * priority class in the cfq group)
564 */
565 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
566 cfq_class_rt(cfqq));
5db5d642
CZ
567 unsigned sync_slice = cfqd->cfq_slice[1];
568 unsigned expect_latency = sync_slice * iq;
58ff82f3
VG
569 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
570
571 if (expect_latency > group_slice) {
5db5d642
CZ
572 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
573 /* scale low_slice according to IO priority
574 * and sync vs async */
575 unsigned low_slice =
576 min(slice, base_low_slice * slice / sync_slice);
577 /* the adapted slice value is scaled to fit all iqs
578 * into the target latency */
58ff82f3 579 slice = max(slice * group_slice / expect_latency,
5db5d642
CZ
580 low_slice);
581 }
582 }
dae739eb 583 cfqq->slice_start = jiffies;
5db5d642 584 cfqq->slice_end = jiffies + slice;
f75edf2d 585 cfqq->allocated_slice = slice;
7b679138 586 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
44f7c160
JA
587}
588
589/*
590 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
591 * isn't valid until the first request from the dispatch is activated
592 * and the slice time set.
593 */
a6151c3a 594static inline bool cfq_slice_used(struct cfq_queue *cfqq)
44f7c160
JA
595{
596 if (cfq_cfqq_slice_new(cfqq))
597 return 0;
598 if (time_before(jiffies, cfqq->slice_end))
599 return 0;
600
601 return 1;
602}
603
1da177e4 604/*
5e705374 605 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1da177e4 606 * We choose the request that is closest to the head right now. Distance
e8a99053 607 * behind the head is penalized and only allowed to a certain extent.
1da177e4 608 */
5e705374 609static struct request *
cf7c25cf 610cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1da177e4 611{
cf7c25cf 612 sector_t s1, s2, d1 = 0, d2 = 0;
1da177e4 613 unsigned long back_max;
e8a99053
AM
614#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
615#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
616 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1da177e4 617
5e705374
JA
618 if (rq1 == NULL || rq1 == rq2)
619 return rq2;
620 if (rq2 == NULL)
621 return rq1;
9c2c38a1 622
5e705374
JA
623 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
624 return rq1;
625 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
626 return rq2;
374f84ac
JA
627 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
628 return rq1;
629 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
630 return rq2;
1da177e4 631
83096ebf
TH
632 s1 = blk_rq_pos(rq1);
633 s2 = blk_rq_pos(rq2);
1da177e4 634
1da177e4
LT
635 /*
636 * by definition, 1KiB is 2 sectors
637 */
638 back_max = cfqd->cfq_back_max * 2;
639
640 /*
641 * Strict one way elevator _except_ in the case where we allow
642 * short backward seeks which are biased as twice the cost of a
643 * similar forward seek.
644 */
645 if (s1 >= last)
646 d1 = s1 - last;
647 else if (s1 + back_max >= last)
648 d1 = (last - s1) * cfqd->cfq_back_penalty;
649 else
e8a99053 650 wrap |= CFQ_RQ1_WRAP;
1da177e4
LT
651
652 if (s2 >= last)
653 d2 = s2 - last;
654 else if (s2 + back_max >= last)
655 d2 = (last - s2) * cfqd->cfq_back_penalty;
656 else
e8a99053 657 wrap |= CFQ_RQ2_WRAP;
1da177e4
LT
658
659 /* Found required data */
e8a99053
AM
660
661 /*
662 * By doing switch() on the bit mask "wrap" we avoid having to
663 * check two variables for all permutations: --> faster!
664 */
665 switch (wrap) {
5e705374 666 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
e8a99053 667 if (d1 < d2)
5e705374 668 return rq1;
e8a99053 669 else if (d2 < d1)
5e705374 670 return rq2;
e8a99053
AM
671 else {
672 if (s1 >= s2)
5e705374 673 return rq1;
e8a99053 674 else
5e705374 675 return rq2;
e8a99053 676 }
1da177e4 677
e8a99053 678 case CFQ_RQ2_WRAP:
5e705374 679 return rq1;
e8a99053 680 case CFQ_RQ1_WRAP:
5e705374
JA
681 return rq2;
682 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
e8a99053
AM
683 default:
684 /*
685 * Since both rqs are wrapped,
686 * start with the one that's further behind head
687 * (--> only *one* back seek required),
688 * since back seek takes more time than forward.
689 */
690 if (s1 <= s2)
5e705374 691 return rq1;
1da177e4 692 else
5e705374 693 return rq2;
1da177e4
LT
694 }
695}
696
498d3aa2
JA
697/*
698 * The below is leftmost cache rbtree addon
699 */
0871714e 700static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
cc09e299 701{
615f0259
VG
702 /* Service tree is empty */
703 if (!root->count)
704 return NULL;
705
cc09e299
JA
706 if (!root->left)
707 root->left = rb_first(&root->rb);
708
0871714e
JA
709 if (root->left)
710 return rb_entry(root->left, struct cfq_queue, rb_node);
711
712 return NULL;
cc09e299
JA
713}
714
1fa8f6d6
VG
715static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
716{
717 if (!root->left)
718 root->left = rb_first(&root->rb);
719
720 if (root->left)
721 return rb_entry_cfqg(root->left);
722
723 return NULL;
724}
725
a36e71f9
JA
726static void rb_erase_init(struct rb_node *n, struct rb_root *root)
727{
728 rb_erase(n, root);
729 RB_CLEAR_NODE(n);
730}
731
cc09e299
JA
732static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
733{
734 if (root->left == n)
735 root->left = NULL;
a36e71f9 736 rb_erase_init(n, &root->rb);
aa6f6a3d 737 --root->count;
cc09e299
JA
738}
739
1da177e4
LT
740/*
741 * would be nice to take fifo expire time into account as well
742 */
5e705374
JA
743static struct request *
744cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
745 struct request *last)
1da177e4 746{
21183b07
JA
747 struct rb_node *rbnext = rb_next(&last->rb_node);
748 struct rb_node *rbprev = rb_prev(&last->rb_node);
5e705374 749 struct request *next = NULL, *prev = NULL;
1da177e4 750
21183b07 751 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1da177e4
LT
752
753 if (rbprev)
5e705374 754 prev = rb_entry_rq(rbprev);
1da177e4 755
21183b07 756 if (rbnext)
5e705374 757 next = rb_entry_rq(rbnext);
21183b07
JA
758 else {
759 rbnext = rb_first(&cfqq->sort_list);
760 if (rbnext && rbnext != &last->rb_node)
5e705374 761 next = rb_entry_rq(rbnext);
21183b07 762 }
1da177e4 763
cf7c25cf 764 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1da177e4
LT
765}
766
d9e7620e
JA
767static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
768 struct cfq_queue *cfqq)
1da177e4 769{
d9e7620e
JA
770 /*
771 * just an approximation, should be ok.
772 */
cdb16e8f 773 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
464191c6 774 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
d9e7620e
JA
775}
776
1fa8f6d6
VG
777static inline s64
778cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
779{
780 return cfqg->vdisktime - st->min_vdisktime;
781}
782
783static void
784__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
785{
786 struct rb_node **node = &st->rb.rb_node;
787 struct rb_node *parent = NULL;
788 struct cfq_group *__cfqg;
789 s64 key = cfqg_key(st, cfqg);
790 int left = 1;
791
792 while (*node != NULL) {
793 parent = *node;
794 __cfqg = rb_entry_cfqg(parent);
795
796 if (key < cfqg_key(st, __cfqg))
797 node = &parent->rb_left;
798 else {
799 node = &parent->rb_right;
800 left = 0;
801 }
802 }
803
804 if (left)
805 st->left = &cfqg->rb_node;
806
807 rb_link_node(&cfqg->rb_node, parent, node);
808 rb_insert_color(&cfqg->rb_node, &st->rb);
809}
810
811static void
812cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
813{
814 struct cfq_rb_root *st = &cfqd->grp_service_tree;
815 struct cfq_group *__cfqg;
816 struct rb_node *n;
817
818 cfqg->nr_cfqq++;
819 if (cfqg->on_st)
820 return;
821
822 /*
823 * Currently put the group at the end. Later implement something
824 * so that groups get lesser vtime based on their weights, so that
825 * if group does not loose all if it was not continously backlogged.
826 */
827 n = rb_last(&st->rb);
828 if (n) {
829 __cfqg = rb_entry_cfqg(n);
830 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
831 } else
832 cfqg->vdisktime = st->min_vdisktime;
833
834 __cfq_group_service_tree_add(st, cfqg);
835 cfqg->on_st = true;
58ff82f3 836 st->total_weight += cfqg->weight;
1fa8f6d6
VG
837}
838
839static void
840cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
841{
842 struct cfq_rb_root *st = &cfqd->grp_service_tree;
843
25bc6b07
VG
844 if (st->active == &cfqg->rb_node)
845 st->active = NULL;
846
1fa8f6d6
VG
847 BUG_ON(cfqg->nr_cfqq < 1);
848 cfqg->nr_cfqq--;
25bc6b07 849
1fa8f6d6
VG
850 /* If there are other cfq queues under this group, don't delete it */
851 if (cfqg->nr_cfqq)
852 return;
853
2868ef7b 854 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1fa8f6d6 855 cfqg->on_st = false;
58ff82f3 856 st->total_weight -= cfqg->weight;
1fa8f6d6
VG
857 if (!RB_EMPTY_NODE(&cfqg->rb_node))
858 cfq_rb_erase(&cfqg->rb_node, st);
dae739eb 859 cfqg->saved_workload_slice = 0;
22084190 860 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
dae739eb
VG
861}
862
863static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
864{
f75edf2d 865 unsigned int slice_used;
dae739eb
VG
866
867 /*
868 * Queue got expired before even a single request completed or
869 * got expired immediately after first request completion.
870 */
871 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
872 /*
873 * Also charge the seek time incurred to the group, otherwise
874 * if there are mutiple queues in the group, each can dispatch
875 * a single request on seeky media and cause lots of seek time
876 * and group will never know it.
877 */
878 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
879 1);
880 } else {
881 slice_used = jiffies - cfqq->slice_start;
f75edf2d
VG
882 if (slice_used > cfqq->allocated_slice)
883 slice_used = cfqq->allocated_slice;
dae739eb
VG
884 }
885
22084190
VG
886 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
887 cfqq->nr_sectors);
dae739eb
VG
888 return slice_used;
889}
890
891static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
892 struct cfq_queue *cfqq)
893{
894 struct cfq_rb_root *st = &cfqd->grp_service_tree;
f26bd1f0
VG
895 unsigned int used_sl, charge_sl;
896 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
897 - cfqg->service_tree_idle.count;
898
899 BUG_ON(nr_sync < 0);
900 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
dae739eb 901
f26bd1f0
VG
902 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
903 charge_sl = cfqq->allocated_slice;
dae739eb
VG
904
905 /* Can't update vdisktime while group is on service tree */
906 cfq_rb_erase(&cfqg->rb_node, st);
f26bd1f0 907 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
dae739eb
VG
908 __cfq_group_service_tree_add(st, cfqg);
909
910 /* This group is being expired. Save the context */
911 if (time_after(cfqd->workload_expires, jiffies)) {
912 cfqg->saved_workload_slice = cfqd->workload_expires
913 - jiffies;
914 cfqg->saved_workload = cfqd->serving_type;
915 cfqg->saved_serving_prio = cfqd->serving_prio;
916 } else
917 cfqg->saved_workload_slice = 0;
2868ef7b
VG
918
919 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
920 st->min_vdisktime);
22084190
VG
921 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
922 cfqq->nr_sectors);
1fa8f6d6
VG
923}
924
25fb5169
VG
925#ifdef CONFIG_CFQ_GROUP_IOSCHED
926static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
927{
928 if (blkg)
929 return container_of(blkg, struct cfq_group, blkg);
930 return NULL;
931}
932
f8d461d6
VG
933void
934cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
935{
936 cfqg_of_blkg(blkg)->weight = weight;
937}
938
25fb5169
VG
939static struct cfq_group *
940cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
941{
942 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
943 struct cfq_group *cfqg = NULL;
944 void *key = cfqd;
945 int i, j;
946 struct cfq_rb_root *st;
22084190
VG
947 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
948 unsigned int major, minor;
25fb5169 949
25fb5169
VG
950 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
951 if (cfqg || !create)
952 goto done;
953
954 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
955 if (!cfqg)
956 goto done;
957
958 cfqg->weight = blkcg->weight;
959 for_each_cfqg_st(cfqg, i, j, st)
960 *st = CFQ_RB_ROOT;
961 RB_CLEAR_NODE(&cfqg->rb_node);
962
b1c35769
VG
963 /*
964 * Take the initial reference that will be released on destroy
965 * This can be thought of a joint reference by cgroup and
966 * elevator which will be dropped by either elevator exit
967 * or cgroup deletion path depending on who is exiting first.
968 */
969 atomic_set(&cfqg->ref, 1);
970
25fb5169 971 /* Add group onto cgroup list */
22084190
VG
972 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
973 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
974 MKDEV(major, minor));
25fb5169
VG
975
976 /* Add group on cfqd list */
977 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
978
979done:
25fb5169
VG
980 return cfqg;
981}
982
983/*
984 * Search for the cfq group current task belongs to. If create = 1, then also
985 * create the cfq group if it does not exist. request_queue lock must be held.
986 */
987static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
988{
989 struct cgroup *cgroup;
990 struct cfq_group *cfqg = NULL;
991
992 rcu_read_lock();
993 cgroup = task_cgroup(current, blkio_subsys_id);
994 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
995 if (!cfqg && create)
996 cfqg = &cfqd->root_group;
997 rcu_read_unlock();
998 return cfqg;
999}
1000
1001static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1002{
1003 /* Currently, all async queues are mapped to root group */
1004 if (!cfq_cfqq_sync(cfqq))
1005 cfqg = &cfqq->cfqd->root_group;
1006
1007 cfqq->cfqg = cfqg;
b1c35769
VG
1008 /* cfqq reference on cfqg */
1009 atomic_inc(&cfqq->cfqg->ref);
1010}
1011
1012static void cfq_put_cfqg(struct cfq_group *cfqg)
1013{
1014 struct cfq_rb_root *st;
1015 int i, j;
1016
1017 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1018 if (!atomic_dec_and_test(&cfqg->ref))
1019 return;
1020 for_each_cfqg_st(cfqg, i, j, st)
1021 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1022 kfree(cfqg);
1023}
1024
1025static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1026{
1027 /* Something wrong if we are trying to remove same group twice */
1028 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1029
1030 hlist_del_init(&cfqg->cfqd_node);
1031
1032 /*
1033 * Put the reference taken at the time of creation so that when all
1034 * queues are gone, group can be destroyed.
1035 */
1036 cfq_put_cfqg(cfqg);
1037}
1038
1039static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1040{
1041 struct hlist_node *pos, *n;
1042 struct cfq_group *cfqg;
1043
1044 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1045 /*
1046 * If cgroup removal path got to blk_group first and removed
1047 * it from cgroup list, then it will take care of destroying
1048 * cfqg also.
1049 */
1050 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1051 cfq_destroy_cfqg(cfqd, cfqg);
1052 }
25fb5169 1053}
b1c35769
VG
1054
1055/*
1056 * Blk cgroup controller notification saying that blkio_group object is being
1057 * delinked as associated cgroup object is going away. That also means that
1058 * no new IO will come in this group. So get rid of this group as soon as
1059 * any pending IO in the group is finished.
1060 *
1061 * This function is called under rcu_read_lock(). key is the rcu protected
1062 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1063 * read lock.
1064 *
1065 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1066 * it should not be NULL as even if elevator was exiting, cgroup deltion
1067 * path got to it first.
1068 */
1069void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1070{
1071 unsigned long flags;
1072 struct cfq_data *cfqd = key;
1073
1074 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1075 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1076 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1077}
1078
25fb5169
VG
1079#else /* GROUP_IOSCHED */
1080static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1081{
1082 return &cfqd->root_group;
1083}
1084static inline void
1085cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1086 cfqq->cfqg = cfqg;
1087}
1088
b1c35769
VG
1089static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1090static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1091
25fb5169
VG
1092#endif /* GROUP_IOSCHED */
1093
498d3aa2 1094/*
c0324a02 1095 * The cfqd->service_trees holds all pending cfq_queue's that have
498d3aa2
JA
1096 * requests waiting to be processed. It is sorted in the order that
1097 * we will service the queues.
1098 */
a36e71f9 1099static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
a6151c3a 1100 bool add_front)
d9e7620e 1101{
0871714e
JA
1102 struct rb_node **p, *parent;
1103 struct cfq_queue *__cfqq;
d9e7620e 1104 unsigned long rb_key;
c0324a02 1105 struct cfq_rb_root *service_tree;
498d3aa2 1106 int left;
dae739eb 1107 int new_cfqq = 1;
ae30c286
VG
1108 int group_changed = 0;
1109
1110#ifdef CONFIG_CFQ_GROUP_IOSCHED
1111 if (!cfqd->cfq_group_isolation
1112 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1113 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1114 /* Move this cfq to root group */
1115 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1116 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1117 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1118 cfqq->orig_cfqg = cfqq->cfqg;
1119 cfqq->cfqg = &cfqd->root_group;
1120 atomic_inc(&cfqd->root_group.ref);
1121 group_changed = 1;
1122 } else if (!cfqd->cfq_group_isolation
1123 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1124 /* cfqq is sequential now needs to go to its original group */
1125 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1126 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1127 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1128 cfq_put_cfqg(cfqq->cfqg);
1129 cfqq->cfqg = cfqq->orig_cfqg;
1130 cfqq->orig_cfqg = NULL;
1131 group_changed = 1;
1132 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1133 }
1134#endif
d9e7620e 1135
cdb16e8f 1136 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
65b32a57 1137 cfqq_type(cfqq));
0871714e
JA
1138 if (cfq_class_idle(cfqq)) {
1139 rb_key = CFQ_IDLE_DELAY;
aa6f6a3d 1140 parent = rb_last(&service_tree->rb);
0871714e
JA
1141 if (parent && parent != &cfqq->rb_node) {
1142 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1143 rb_key += __cfqq->rb_key;
1144 } else
1145 rb_key += jiffies;
1146 } else if (!add_front) {
b9c8946b
JA
1147 /*
1148 * Get our rb key offset. Subtract any residual slice
1149 * value carried from last service. A negative resid
1150 * count indicates slice overrun, and this should position
1151 * the next service time further away in the tree.
1152 */
edd75ffd 1153 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
b9c8946b 1154 rb_key -= cfqq->slice_resid;
edd75ffd 1155 cfqq->slice_resid = 0;
48e025e6
CZ
1156 } else {
1157 rb_key = -HZ;
aa6f6a3d 1158 __cfqq = cfq_rb_first(service_tree);
48e025e6
CZ
1159 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1160 }
1da177e4 1161
d9e7620e 1162 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
dae739eb 1163 new_cfqq = 0;
99f9628a 1164 /*
d9e7620e 1165 * same position, nothing more to do
99f9628a 1166 */
c0324a02
CZ
1167 if (rb_key == cfqq->rb_key &&
1168 cfqq->service_tree == service_tree)
d9e7620e 1169 return;
1da177e4 1170
aa6f6a3d
CZ
1171 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1172 cfqq->service_tree = NULL;
1da177e4 1173 }
d9e7620e 1174
498d3aa2 1175 left = 1;
0871714e 1176 parent = NULL;
aa6f6a3d
CZ
1177 cfqq->service_tree = service_tree;
1178 p = &service_tree->rb.rb_node;
d9e7620e 1179 while (*p) {
67060e37 1180 struct rb_node **n;
cc09e299 1181
d9e7620e
JA
1182 parent = *p;
1183 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1184
0c534e0a 1185 /*
c0324a02 1186 * sort by key, that represents service time.
0c534e0a 1187 */
c0324a02 1188 if (time_before(rb_key, __cfqq->rb_key))
67060e37 1189 n = &(*p)->rb_left;
c0324a02 1190 else {
67060e37 1191 n = &(*p)->rb_right;
cc09e299 1192 left = 0;
c0324a02 1193 }
67060e37
JA
1194
1195 p = n;
d9e7620e
JA
1196 }
1197
cc09e299 1198 if (left)
aa6f6a3d 1199 service_tree->left = &cfqq->rb_node;
cc09e299 1200
d9e7620e
JA
1201 cfqq->rb_key = rb_key;
1202 rb_link_node(&cfqq->rb_node, parent, p);
aa6f6a3d
CZ
1203 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1204 service_tree->count++;
ae30c286 1205 if ((add_front || !new_cfqq) && !group_changed)
dae739eb 1206 return;
1fa8f6d6 1207 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1da177e4
LT
1208}
1209
a36e71f9 1210static struct cfq_queue *
f2d1f0ae
JA
1211cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1212 sector_t sector, struct rb_node **ret_parent,
1213 struct rb_node ***rb_link)
a36e71f9 1214{
a36e71f9
JA
1215 struct rb_node **p, *parent;
1216 struct cfq_queue *cfqq = NULL;
1217
1218 parent = NULL;
1219 p = &root->rb_node;
1220 while (*p) {
1221 struct rb_node **n;
1222
1223 parent = *p;
1224 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1225
1226 /*
1227 * Sort strictly based on sector. Smallest to the left,
1228 * largest to the right.
1229 */
2e46e8b2 1230 if (sector > blk_rq_pos(cfqq->next_rq))
a36e71f9 1231 n = &(*p)->rb_right;
2e46e8b2 1232 else if (sector < blk_rq_pos(cfqq->next_rq))
a36e71f9
JA
1233 n = &(*p)->rb_left;
1234 else
1235 break;
1236 p = n;
3ac6c9f8 1237 cfqq = NULL;
a36e71f9
JA
1238 }
1239
1240 *ret_parent = parent;
1241 if (rb_link)
1242 *rb_link = p;
3ac6c9f8 1243 return cfqq;
a36e71f9
JA
1244}
1245
1246static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1247{
a36e71f9
JA
1248 struct rb_node **p, *parent;
1249 struct cfq_queue *__cfqq;
1250
f2d1f0ae
JA
1251 if (cfqq->p_root) {
1252 rb_erase(&cfqq->p_node, cfqq->p_root);
1253 cfqq->p_root = NULL;
1254 }
a36e71f9
JA
1255
1256 if (cfq_class_idle(cfqq))
1257 return;
1258 if (!cfqq->next_rq)
1259 return;
1260
f2d1f0ae 1261 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2e46e8b2
TH
1262 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1263 blk_rq_pos(cfqq->next_rq), &parent, &p);
3ac6c9f8
JA
1264 if (!__cfqq) {
1265 rb_link_node(&cfqq->p_node, parent, p);
f2d1f0ae
JA
1266 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1267 } else
1268 cfqq->p_root = NULL;
a36e71f9
JA
1269}
1270
498d3aa2
JA
1271/*
1272 * Update cfqq's position in the service tree.
1273 */
edd75ffd 1274static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
6d048f53 1275{
6d048f53
JA
1276 /*
1277 * Resorting requires the cfqq to be on the RR list already.
1278 */
a36e71f9 1279 if (cfq_cfqq_on_rr(cfqq)) {
edd75ffd 1280 cfq_service_tree_add(cfqd, cfqq, 0);
a36e71f9
JA
1281 cfq_prio_tree_add(cfqd, cfqq);
1282 }
6d048f53
JA
1283}
1284
1da177e4
LT
1285/*
1286 * add to busy list of queues for service, trying to be fair in ordering
22e2c507 1287 * the pending list according to last request service
1da177e4 1288 */
febffd61 1289static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1da177e4 1290{
7b679138 1291 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
3b18152c
JA
1292 BUG_ON(cfq_cfqq_on_rr(cfqq));
1293 cfq_mark_cfqq_on_rr(cfqq);
1da177e4
LT
1294 cfqd->busy_queues++;
1295
edd75ffd 1296 cfq_resort_rr_list(cfqd, cfqq);
1da177e4
LT
1297}
1298
498d3aa2
JA
1299/*
1300 * Called when the cfqq no longer has requests pending, remove it from
1301 * the service tree.
1302 */
febffd61 1303static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1da177e4 1304{
7b679138 1305 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
3b18152c
JA
1306 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1307 cfq_clear_cfqq_on_rr(cfqq);
1da177e4 1308
aa6f6a3d
CZ
1309 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1310 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1311 cfqq->service_tree = NULL;
1312 }
f2d1f0ae
JA
1313 if (cfqq->p_root) {
1314 rb_erase(&cfqq->p_node, cfqq->p_root);
1315 cfqq->p_root = NULL;
1316 }
d9e7620e 1317
1fa8f6d6 1318 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1da177e4
LT
1319 BUG_ON(!cfqd->busy_queues);
1320 cfqd->busy_queues--;
1321}
1322
1323/*
1324 * rb tree support functions
1325 */
febffd61 1326static void cfq_del_rq_rb(struct request *rq)
1da177e4 1327{
5e705374 1328 struct cfq_queue *cfqq = RQ_CFQQ(rq);
5e705374 1329 const int sync = rq_is_sync(rq);
1da177e4 1330
b4878f24
JA
1331 BUG_ON(!cfqq->queued[sync]);
1332 cfqq->queued[sync]--;
1da177e4 1333
5e705374 1334 elv_rb_del(&cfqq->sort_list, rq);
1da177e4 1335
f04a6424
VG
1336 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1337 /*
1338 * Queue will be deleted from service tree when we actually
1339 * expire it later. Right now just remove it from prio tree
1340 * as it is empty.
1341 */
1342 if (cfqq->p_root) {
1343 rb_erase(&cfqq->p_node, cfqq->p_root);
1344 cfqq->p_root = NULL;
1345 }
1346 }
1da177e4
LT
1347}
1348
5e705374 1349static void cfq_add_rq_rb(struct request *rq)
1da177e4 1350{
5e705374 1351 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1da177e4 1352 struct cfq_data *cfqd = cfqq->cfqd;
a36e71f9 1353 struct request *__alias, *prev;
1da177e4 1354
5380a101 1355 cfqq->queued[rq_is_sync(rq)]++;
1da177e4
LT
1356
1357 /*
1358 * looks a little odd, but the first insert might return an alias.
1359 * if that happens, put the alias on the dispatch list
1360 */
21183b07 1361 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
5e705374 1362 cfq_dispatch_insert(cfqd->queue, __alias);
5fccbf61
JA
1363
1364 if (!cfq_cfqq_on_rr(cfqq))
1365 cfq_add_cfqq_rr(cfqd, cfqq);
5044eed4
JA
1366
1367 /*
1368 * check if this request is a better next-serve candidate
1369 */
a36e71f9 1370 prev = cfqq->next_rq;
cf7c25cf 1371 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
a36e71f9
JA
1372
1373 /*
1374 * adjust priority tree position, if ->next_rq changes
1375 */
1376 if (prev != cfqq->next_rq)
1377 cfq_prio_tree_add(cfqd, cfqq);
1378
5044eed4 1379 BUG_ON(!cfqq->next_rq);
1da177e4
LT
1380}
1381
febffd61 1382static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1da177e4 1383{
5380a101
JA
1384 elv_rb_del(&cfqq->sort_list, rq);
1385 cfqq->queued[rq_is_sync(rq)]--;
5e705374 1386 cfq_add_rq_rb(rq);
1da177e4
LT
1387}
1388
206dc69b
JA
1389static struct request *
1390cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1da177e4 1391{
206dc69b 1392 struct task_struct *tsk = current;
91fac317 1393 struct cfq_io_context *cic;
206dc69b 1394 struct cfq_queue *cfqq;
1da177e4 1395
4ac845a2 1396 cic = cfq_cic_lookup(cfqd, tsk->io_context);
91fac317
VT
1397 if (!cic)
1398 return NULL;
1399
1400 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
89850f7e
JA
1401 if (cfqq) {
1402 sector_t sector = bio->bi_sector + bio_sectors(bio);
1403
21183b07 1404 return elv_rb_find(&cfqq->sort_list, sector);
89850f7e 1405 }
1da177e4 1406
1da177e4
LT
1407 return NULL;
1408}
1409
165125e1 1410static void cfq_activate_request(struct request_queue *q, struct request *rq)
1da177e4 1411{
22e2c507 1412 struct cfq_data *cfqd = q->elevator->elevator_data;
3b18152c 1413
53c583d2 1414 cfqd->rq_in_driver++;
7b679138 1415 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
53c583d2 1416 cfqd->rq_in_driver);
25776e35 1417
5b93629b 1418 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1da177e4
LT
1419}
1420
165125e1 1421static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1da177e4 1422{
b4878f24
JA
1423 struct cfq_data *cfqd = q->elevator->elevator_data;
1424
53c583d2
CZ
1425 WARN_ON(!cfqd->rq_in_driver);
1426 cfqd->rq_in_driver--;
7b679138 1427 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
53c583d2 1428 cfqd->rq_in_driver);
1da177e4
LT
1429}
1430
b4878f24 1431static void cfq_remove_request(struct request *rq)
1da177e4 1432{
5e705374 1433 struct cfq_queue *cfqq = RQ_CFQQ(rq);
21183b07 1434
5e705374
JA
1435 if (cfqq->next_rq == rq)
1436 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1da177e4 1437
b4878f24 1438 list_del_init(&rq->queuelist);
5e705374 1439 cfq_del_rq_rb(rq);
374f84ac 1440
45333d5a 1441 cfqq->cfqd->rq_queued--;
374f84ac
JA
1442 if (rq_is_meta(rq)) {
1443 WARN_ON(!cfqq->meta_pending);
1444 cfqq->meta_pending--;
1445 }
1da177e4
LT
1446}
1447
165125e1
JA
1448static int cfq_merge(struct request_queue *q, struct request **req,
1449 struct bio *bio)
1da177e4
LT
1450{
1451 struct cfq_data *cfqd = q->elevator->elevator_data;
1452 struct request *__rq;
1da177e4 1453
206dc69b 1454 __rq = cfq_find_rq_fmerge(cfqd, bio);
22e2c507 1455 if (__rq && elv_rq_merge_ok(__rq, bio)) {
9817064b
JA
1456 *req = __rq;
1457 return ELEVATOR_FRONT_MERGE;
1da177e4
LT
1458 }
1459
1460 return ELEVATOR_NO_MERGE;
1da177e4
LT
1461}
1462
165125e1 1463static void cfq_merged_request(struct request_queue *q, struct request *req,
21183b07 1464 int type)
1da177e4 1465{
21183b07 1466 if (type == ELEVATOR_FRONT_MERGE) {
5e705374 1467 struct cfq_queue *cfqq = RQ_CFQQ(req);
1da177e4 1468
5e705374 1469 cfq_reposition_rq_rb(cfqq, req);
1da177e4 1470 }
1da177e4
LT
1471}
1472
1473static void
165125e1 1474cfq_merged_requests(struct request_queue *q, struct request *rq,
1da177e4
LT
1475 struct request *next)
1476{
cf7c25cf 1477 struct cfq_queue *cfqq = RQ_CFQQ(rq);
22e2c507
JA
1478 /*
1479 * reposition in fifo if next is older than rq
1480 */
1481 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
30996f40 1482 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
22e2c507 1483 list_move(&rq->queuelist, &next->queuelist);
30996f40
JA
1484 rq_set_fifo_time(rq, rq_fifo_time(next));
1485 }
22e2c507 1486
cf7c25cf
CZ
1487 if (cfqq->next_rq == next)
1488 cfqq->next_rq = rq;
b4878f24 1489 cfq_remove_request(next);
22e2c507
JA
1490}
1491
165125e1 1492static int cfq_allow_merge(struct request_queue *q, struct request *rq,
da775265
JA
1493 struct bio *bio)
1494{
1495 struct cfq_data *cfqd = q->elevator->elevator_data;
91fac317 1496 struct cfq_io_context *cic;
da775265 1497 struct cfq_queue *cfqq;
da775265
JA
1498
1499 /*
ec8acb69 1500 * Disallow merge of a sync bio into an async request.
da775265 1501 */
91fac317 1502 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
a6151c3a 1503 return false;
da775265
JA
1504
1505 /*
719d3402
JA
1506 * Lookup the cfqq that this bio will be queued with. Allow
1507 * merge only if rq is queued there.
da775265 1508 */
4ac845a2 1509 cic = cfq_cic_lookup(cfqd, current->io_context);
91fac317 1510 if (!cic)
a6151c3a 1511 return false;
719d3402 1512
91fac317 1513 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
a6151c3a 1514 return cfqq == RQ_CFQQ(rq);
da775265
JA
1515}
1516
febffd61
JA
1517static void __cfq_set_active_queue(struct cfq_data *cfqd,
1518 struct cfq_queue *cfqq)
22e2c507
JA
1519{
1520 if (cfqq) {
b1ffe737
DS
1521 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1522 cfqd->serving_prio, cfqd->serving_type);
dae739eb
VG
1523 cfqq->slice_start = 0;
1524 cfqq->dispatch_start = jiffies;
f75edf2d 1525 cfqq->allocated_slice = 0;
22e2c507 1526 cfqq->slice_end = 0;
2f5cb738 1527 cfqq->slice_dispatch = 0;
22084190 1528 cfqq->nr_sectors = 0;
2f5cb738 1529
2f5cb738 1530 cfq_clear_cfqq_wait_request(cfqq);
b029195d 1531 cfq_clear_cfqq_must_dispatch(cfqq);
3b18152c
JA
1532 cfq_clear_cfqq_must_alloc_slice(cfqq);
1533 cfq_clear_cfqq_fifo_expire(cfqq);
44f7c160 1534 cfq_mark_cfqq_slice_new(cfqq);
2f5cb738
JA
1535
1536 del_timer(&cfqd->idle_slice_timer);
22e2c507
JA
1537 }
1538
1539 cfqd->active_queue = cfqq;
1540}
1541
7b14e3b5
JA
1542/*
1543 * current cfqq expired its slice (or was too idle), select new one
1544 */
1545static void
1546__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
a6151c3a 1547 bool timed_out)
7b14e3b5 1548{
7b679138
JA
1549 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1550
7b14e3b5
JA
1551 if (cfq_cfqq_wait_request(cfqq))
1552 del_timer(&cfqd->idle_slice_timer);
1553
7b14e3b5 1554 cfq_clear_cfqq_wait_request(cfqq);
f75edf2d 1555 cfq_clear_cfqq_wait_busy(cfqq);
7b14e3b5 1556
ae54abed
SL
1557 /*
1558 * If this cfqq is shared between multiple processes, check to
1559 * make sure that those processes are still issuing I/Os within
1560 * the mean seek distance. If not, it may be time to break the
1561 * queues apart again.
1562 */
1563 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1564 cfq_mark_cfqq_split_coop(cfqq);
1565
7b14e3b5 1566 /*
6084cdda 1567 * store what was left of this slice, if the queue idled/timed out
7b14e3b5 1568 */
7b679138 1569 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
c5b680f3 1570 cfqq->slice_resid = cfqq->slice_end - jiffies;
7b679138
JA
1571 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1572 }
7b14e3b5 1573
dae739eb
VG
1574 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1575
f04a6424
VG
1576 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1577 cfq_del_cfqq_rr(cfqd, cfqq);
1578
edd75ffd 1579 cfq_resort_rr_list(cfqd, cfqq);
7b14e3b5
JA
1580
1581 if (cfqq == cfqd->active_queue)
1582 cfqd->active_queue = NULL;
1583
dae739eb
VG
1584 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1585 cfqd->grp_service_tree.active = NULL;
1586
7b14e3b5
JA
1587 if (cfqd->active_cic) {
1588 put_io_context(cfqd->active_cic->ioc);
1589 cfqd->active_cic = NULL;
1590 }
7b14e3b5
JA
1591}
1592
a6151c3a 1593static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
7b14e3b5
JA
1594{
1595 struct cfq_queue *cfqq = cfqd->active_queue;
1596
1597 if (cfqq)
6084cdda 1598 __cfq_slice_expired(cfqd, cfqq, timed_out);
7b14e3b5
JA
1599}
1600
498d3aa2
JA
1601/*
1602 * Get next queue for service. Unless we have a queue preemption,
1603 * we'll simply select the first cfqq in the service tree.
1604 */
6d048f53 1605static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
22e2c507 1606{
c0324a02 1607 struct cfq_rb_root *service_tree =
cdb16e8f 1608 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
65b32a57 1609 cfqd->serving_type);
d9e7620e 1610
f04a6424
VG
1611 if (!cfqd->rq_queued)
1612 return NULL;
1613
1fa8f6d6
VG
1614 /* There is nothing to dispatch */
1615 if (!service_tree)
1616 return NULL;
c0324a02
CZ
1617 if (RB_EMPTY_ROOT(&service_tree->rb))
1618 return NULL;
1619 return cfq_rb_first(service_tree);
6d048f53
JA
1620}
1621
f04a6424
VG
1622static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1623{
25fb5169 1624 struct cfq_group *cfqg;
f04a6424
VG
1625 struct cfq_queue *cfqq;
1626 int i, j;
1627 struct cfq_rb_root *st;
1628
1629 if (!cfqd->rq_queued)
1630 return NULL;
1631
25fb5169
VG
1632 cfqg = cfq_get_next_cfqg(cfqd);
1633 if (!cfqg)
1634 return NULL;
1635
f04a6424
VG
1636 for_each_cfqg_st(cfqg, i, j, st)
1637 if ((cfqq = cfq_rb_first(st)) != NULL)
1638 return cfqq;
1639 return NULL;
1640}
1641
498d3aa2
JA
1642/*
1643 * Get and set a new active queue for service.
1644 */
a36e71f9
JA
1645static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1646 struct cfq_queue *cfqq)
6d048f53 1647{
e00ef799 1648 if (!cfqq)
a36e71f9 1649 cfqq = cfq_get_next_queue(cfqd);
6d048f53 1650
22e2c507 1651 __cfq_set_active_queue(cfqd, cfqq);
3b18152c 1652 return cfqq;
22e2c507
JA
1653}
1654
d9e7620e
JA
1655static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1656 struct request *rq)
1657{
83096ebf
TH
1658 if (blk_rq_pos(rq) >= cfqd->last_position)
1659 return blk_rq_pos(rq) - cfqd->last_position;
d9e7620e 1660 else
83096ebf 1661 return cfqd->last_position - blk_rq_pos(rq);
d9e7620e
JA
1662}
1663
b2c18e1e 1664static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
e9ce335d 1665 struct request *rq)
6d048f53 1666{
e9ce335d 1667 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
6d048f53
JA
1668}
1669
a36e71f9
JA
1670static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1671 struct cfq_queue *cur_cfqq)
1672{
f2d1f0ae 1673 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
a36e71f9
JA
1674 struct rb_node *parent, *node;
1675 struct cfq_queue *__cfqq;
1676 sector_t sector = cfqd->last_position;
1677
1678 if (RB_EMPTY_ROOT(root))
1679 return NULL;
1680
1681 /*
1682 * First, if we find a request starting at the end of the last
1683 * request, choose it.
1684 */
f2d1f0ae 1685 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
a36e71f9
JA
1686 if (__cfqq)
1687 return __cfqq;
1688
1689 /*
1690 * If the exact sector wasn't found, the parent of the NULL leaf
1691 * will contain the closest sector.
1692 */
1693 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
e9ce335d 1694 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
a36e71f9
JA
1695 return __cfqq;
1696
2e46e8b2 1697 if (blk_rq_pos(__cfqq->next_rq) < sector)
a36e71f9
JA
1698 node = rb_next(&__cfqq->p_node);
1699 else
1700 node = rb_prev(&__cfqq->p_node);
1701 if (!node)
1702 return NULL;
1703
1704 __cfqq = rb_entry(node, struct cfq_queue, p_node);
e9ce335d 1705 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
a36e71f9
JA
1706 return __cfqq;
1707
1708 return NULL;
1709}
1710
1711/*
1712 * cfqd - obvious
1713 * cur_cfqq - passed in so that we don't decide that the current queue is
1714 * closely cooperating with itself.
1715 *
1716 * So, basically we're assuming that that cur_cfqq has dispatched at least
1717 * one request, and that cfqd->last_position reflects a position on the disk
1718 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1719 * assumption.
1720 */
1721static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
b3b6d040 1722 struct cfq_queue *cur_cfqq)
6d048f53 1723{
a36e71f9
JA
1724 struct cfq_queue *cfqq;
1725
39c01b21
DS
1726 if (cfq_class_idle(cur_cfqq))
1727 return NULL;
e6c5bc73
JM
1728 if (!cfq_cfqq_sync(cur_cfqq))
1729 return NULL;
1730 if (CFQQ_SEEKY(cur_cfqq))
1731 return NULL;
1732
b9d8f4c7
GJ
1733 /*
1734 * Don't search priority tree if it's the only queue in the group.
1735 */
1736 if (cur_cfqq->cfqg->nr_cfqq == 1)
1737 return NULL;
1738
6d048f53 1739 /*
d9e7620e
JA
1740 * We should notice if some of the queues are cooperating, eg
1741 * working closely on the same area of the disk. In that case,
1742 * we can group them together and don't waste time idling.
6d048f53 1743 */
a36e71f9
JA
1744 cfqq = cfqq_close(cfqd, cur_cfqq);
1745 if (!cfqq)
1746 return NULL;
1747
8682e1f1
VG
1748 /* If new queue belongs to different cfq_group, don't choose it */
1749 if (cur_cfqq->cfqg != cfqq->cfqg)
1750 return NULL;
1751
df5fe3e8
JM
1752 /*
1753 * It only makes sense to merge sync queues.
1754 */
1755 if (!cfq_cfqq_sync(cfqq))
1756 return NULL;
e6c5bc73
JM
1757 if (CFQQ_SEEKY(cfqq))
1758 return NULL;
df5fe3e8 1759
c0324a02
CZ
1760 /*
1761 * Do not merge queues of different priority classes
1762 */
1763 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1764 return NULL;
1765
a36e71f9 1766 return cfqq;
6d048f53
JA
1767}
1768
a6d44e98
CZ
1769/*
1770 * Determine whether we should enforce idle window for this queue.
1771 */
1772
1773static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1774{
1775 enum wl_prio_t prio = cfqq_prio(cfqq);
718eee05 1776 struct cfq_rb_root *service_tree = cfqq->service_tree;
a6d44e98 1777
f04a6424
VG
1778 BUG_ON(!service_tree);
1779 BUG_ON(!service_tree->count);
1780
a6d44e98
CZ
1781 /* We never do for idle class queues. */
1782 if (prio == IDLE_WORKLOAD)
1783 return false;
1784
1785 /* We do for queues that were marked with idle window flag. */
3c764b7a
SL
1786 if (cfq_cfqq_idle_window(cfqq) &&
1787 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
a6d44e98
CZ
1788 return true;
1789
1790 /*
1791 * Otherwise, we do only if they are the last ones
1792 * in their service tree.
1793 */
b1ffe737
DS
1794 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1795 return 1;
1796 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1797 service_tree->count);
1798 return 0;
a6d44e98
CZ
1799}
1800
6d048f53 1801static void cfq_arm_slice_timer(struct cfq_data *cfqd)
22e2c507 1802{
1792669c 1803 struct cfq_queue *cfqq = cfqd->active_queue;
206dc69b 1804 struct cfq_io_context *cic;
7b14e3b5
JA
1805 unsigned long sl;
1806
a68bbddb 1807 /*
f7d7b7a7
JA
1808 * SSD device without seek penalty, disable idling. But only do so
1809 * for devices that support queuing, otherwise we still have a problem
1810 * with sync vs async workloads.
a68bbddb 1811 */
f7d7b7a7 1812 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
a68bbddb
JA
1813 return;
1814
dd67d051 1815 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
6d048f53 1816 WARN_ON(cfq_cfqq_slice_new(cfqq));
22e2c507
JA
1817
1818 /*
1819 * idle is disabled, either manually or by past process history
1820 */
a6d44e98 1821 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
6d048f53
JA
1822 return;
1823
7b679138 1824 /*
8e550632 1825 * still active requests from this queue, don't idle
7b679138 1826 */
8e550632 1827 if (cfqq->dispatched)
7b679138
JA
1828 return;
1829
22e2c507
JA
1830 /*
1831 * task has exited, don't wait
1832 */
206dc69b 1833 cic = cfqd->active_cic;
66dac98e 1834 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
6d048f53
JA
1835 return;
1836
355b659c
CZ
1837 /*
1838 * If our average think time is larger than the remaining time
1839 * slice, then don't idle. This avoids overrunning the allotted
1840 * time slice.
1841 */
1842 if (sample_valid(cic->ttime_samples) &&
b1ffe737
DS
1843 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1844 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1845 cic->ttime_mean);
355b659c 1846 return;
b1ffe737 1847 }
355b659c 1848
3b18152c 1849 cfq_mark_cfqq_wait_request(cfqq);
22e2c507 1850
6d048f53 1851 sl = cfqd->cfq_slice_idle;
206dc69b 1852
7b14e3b5 1853 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
9481ffdc 1854 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1da177e4
LT
1855}
1856
498d3aa2
JA
1857/*
1858 * Move request from internal lists to the request queue dispatch list.
1859 */
165125e1 1860static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1da177e4 1861{
3ed9a296 1862 struct cfq_data *cfqd = q->elevator->elevator_data;
5e705374 1863 struct cfq_queue *cfqq = RQ_CFQQ(rq);
22e2c507 1864
7b679138
JA
1865 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1866
06d21886 1867 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
5380a101 1868 cfq_remove_request(rq);
6d048f53 1869 cfqq->dispatched++;
5380a101 1870 elv_dispatch_sort(q, rq);
3ed9a296 1871
53c583d2 1872 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
22084190 1873 cfqq->nr_sectors += blk_rq_sectors(rq);
1da177e4
LT
1874}
1875
1876/*
1877 * return expired entry, or NULL to just start from scratch in rbtree
1878 */
febffd61 1879static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1da177e4 1880{
30996f40 1881 struct request *rq = NULL;
1da177e4 1882
3b18152c 1883 if (cfq_cfqq_fifo_expire(cfqq))
1da177e4 1884 return NULL;
cb887411
JA
1885
1886 cfq_mark_cfqq_fifo_expire(cfqq);
1887
89850f7e
JA
1888 if (list_empty(&cfqq->fifo))
1889 return NULL;
1da177e4 1890
89850f7e 1891 rq = rq_entry_fifo(cfqq->fifo.next);
30996f40 1892 if (time_before(jiffies, rq_fifo_time(rq)))
7b679138 1893 rq = NULL;
1da177e4 1894
30996f40 1895 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
6d048f53 1896 return rq;
1da177e4
LT
1897}
1898
22e2c507
JA
1899static inline int
1900cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1901{
1902 const int base_rq = cfqd->cfq_slice_async_rq;
1da177e4 1903
22e2c507 1904 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1da177e4 1905
22e2c507 1906 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1da177e4
LT
1907}
1908
df5fe3e8
JM
1909/*
1910 * Must be called with the queue_lock held.
1911 */
1912static int cfqq_process_refs(struct cfq_queue *cfqq)
1913{
1914 int process_refs, io_refs;
1915
1916 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1917 process_refs = atomic_read(&cfqq->ref) - io_refs;
1918 BUG_ON(process_refs < 0);
1919 return process_refs;
1920}
1921
1922static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1923{
e6c5bc73 1924 int process_refs, new_process_refs;
df5fe3e8
JM
1925 struct cfq_queue *__cfqq;
1926
1927 /* Avoid a circular list and skip interim queue merges */
1928 while ((__cfqq = new_cfqq->new_cfqq)) {
1929 if (__cfqq == cfqq)
1930 return;
1931 new_cfqq = __cfqq;
1932 }
1933
1934 process_refs = cfqq_process_refs(cfqq);
1935 /*
1936 * If the process for the cfqq has gone away, there is no
1937 * sense in merging the queues.
1938 */
1939 if (process_refs == 0)
1940 return;
1941
e6c5bc73
JM
1942 /*
1943 * Merge in the direction of the lesser amount of work.
1944 */
1945 new_process_refs = cfqq_process_refs(new_cfqq);
1946 if (new_process_refs >= process_refs) {
1947 cfqq->new_cfqq = new_cfqq;
1948 atomic_add(process_refs, &new_cfqq->ref);
1949 } else {
1950 new_cfqq->new_cfqq = cfqq;
1951 atomic_add(new_process_refs, &cfqq->ref);
1952 }
df5fe3e8
JM
1953}
1954
cdb16e8f 1955static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
65b32a57 1956 struct cfq_group *cfqg, enum wl_prio_t prio)
718eee05
CZ
1957{
1958 struct cfq_queue *queue;
1959 int i;
1960 bool key_valid = false;
1961 unsigned long lowest_key = 0;
1962 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1963
65b32a57
VG
1964 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1965 /* select the one with lowest rb_key */
1966 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
718eee05
CZ
1967 if (queue &&
1968 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1969 lowest_key = queue->rb_key;
1970 cur_best = i;
1971 key_valid = true;
1972 }
1973 }
1974
1975 return cur_best;
1976}
1977
cdb16e8f 1978static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
718eee05 1979{
718eee05
CZ
1980 unsigned slice;
1981 unsigned count;
cdb16e8f 1982 struct cfq_rb_root *st;
58ff82f3 1983 unsigned group_slice;
718eee05 1984
1fa8f6d6
VG
1985 if (!cfqg) {
1986 cfqd->serving_prio = IDLE_WORKLOAD;
1987 cfqd->workload_expires = jiffies + 1;
1988 return;
1989 }
1990
718eee05 1991 /* Choose next priority. RT > BE > IDLE */
58ff82f3 1992 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
718eee05 1993 cfqd->serving_prio = RT_WORKLOAD;
58ff82f3 1994 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
718eee05
CZ
1995 cfqd->serving_prio = BE_WORKLOAD;
1996 else {
1997 cfqd->serving_prio = IDLE_WORKLOAD;
1998 cfqd->workload_expires = jiffies + 1;
1999 return;
2000 }
2001
2002 /*
2003 * For RT and BE, we have to choose also the type
2004 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2005 * expiration time
2006 */
65b32a57 2007 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
cdb16e8f 2008 count = st->count;
718eee05
CZ
2009
2010 /*
65b32a57 2011 * check workload expiration, and that we still have other queues ready
718eee05 2012 */
65b32a57 2013 if (count && !time_after(jiffies, cfqd->workload_expires))
718eee05
CZ
2014 return;
2015
2016 /* otherwise select new workload type */
2017 cfqd->serving_type =
65b32a57
VG
2018 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2019 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
cdb16e8f 2020 count = st->count;
718eee05
CZ
2021
2022 /*
2023 * the workload slice is computed as a fraction of target latency
2024 * proportional to the number of queues in that workload, over
2025 * all the queues in the same priority class
2026 */
58ff82f3
VG
2027 group_slice = cfq_group_slice(cfqd, cfqg);
2028
2029 slice = group_slice * count /
2030 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2031 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
718eee05 2032
f26bd1f0
VG
2033 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2034 unsigned int tmp;
2035
2036 /*
2037 * Async queues are currently system wide. Just taking
2038 * proportion of queues with-in same group will lead to higher
2039 * async ratio system wide as generally root group is going
2040 * to have higher weight. A more accurate thing would be to
2041 * calculate system wide asnc/sync ratio.
2042 */
2043 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2044 tmp = tmp/cfqd->busy_queues;
2045 slice = min_t(unsigned, slice, tmp);
2046
718eee05
CZ
2047 /* async workload slice is scaled down according to
2048 * the sync/async slice ratio. */
2049 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
f26bd1f0 2050 } else
718eee05
CZ
2051 /* sync workload slice is at least 2 * cfq_slice_idle */
2052 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2053
2054 slice = max_t(unsigned, slice, CFQ_MIN_TT);
b1ffe737 2055 cfq_log(cfqd, "workload slice:%d", slice);
718eee05 2056 cfqd->workload_expires = jiffies + slice;
8e550632 2057 cfqd->noidle_tree_requires_idle = false;
718eee05
CZ
2058}
2059
1fa8f6d6
VG
2060static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2061{
2062 struct cfq_rb_root *st = &cfqd->grp_service_tree;
25bc6b07 2063 struct cfq_group *cfqg;
1fa8f6d6
VG
2064
2065 if (RB_EMPTY_ROOT(&st->rb))
2066 return NULL;
25bc6b07
VG
2067 cfqg = cfq_rb_first_group(st);
2068 st->active = &cfqg->rb_node;
2069 update_min_vdisktime(st);
2070 return cfqg;
1fa8f6d6
VG
2071}
2072
cdb16e8f
VG
2073static void cfq_choose_cfqg(struct cfq_data *cfqd)
2074{
1fa8f6d6
VG
2075 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2076
2077 cfqd->serving_group = cfqg;
dae739eb
VG
2078
2079 /* Restore the workload type data */
2080 if (cfqg->saved_workload_slice) {
2081 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2082 cfqd->serving_type = cfqg->saved_workload;
2083 cfqd->serving_prio = cfqg->saved_serving_prio;
66ae2919
GJ
2084 } else
2085 cfqd->workload_expires = jiffies - 1;
2086
1fa8f6d6 2087 choose_service_tree(cfqd, cfqg);
cdb16e8f
VG
2088}
2089
22e2c507 2090/*
498d3aa2
JA
2091 * Select a queue for service. If we have a current active queue,
2092 * check whether to continue servicing it, or retrieve and set a new one.
22e2c507 2093 */
1b5ed5e1 2094static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1da177e4 2095{
a36e71f9 2096 struct cfq_queue *cfqq, *new_cfqq = NULL;
1da177e4 2097
22e2c507
JA
2098 cfqq = cfqd->active_queue;
2099 if (!cfqq)
2100 goto new_queue;
1da177e4 2101
f04a6424
VG
2102 if (!cfqd->rq_queued)
2103 return NULL;
c244bb50
VG
2104
2105 /*
2106 * We were waiting for group to get backlogged. Expire the queue
2107 */
2108 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2109 goto expire;
2110
22e2c507 2111 /*
6d048f53 2112 * The active queue has run out of time, expire it and select new.
22e2c507 2113 */
7667aa06
VG
2114 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2115 /*
2116 * If slice had not expired at the completion of last request
2117 * we might not have turned on wait_busy flag. Don't expire
2118 * the queue yet. Allow the group to get backlogged.
2119 *
2120 * The very fact that we have used the slice, that means we
2121 * have been idling all along on this queue and it should be
2122 * ok to wait for this request to complete.
2123 */
82bbbf28
VG
2124 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2125 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2126 cfqq = NULL;
7667aa06 2127 goto keep_queue;
82bbbf28 2128 } else
7667aa06
VG
2129 goto expire;
2130 }
1da177e4 2131
22e2c507 2132 /*
6d048f53
JA
2133 * The active queue has requests and isn't expired, allow it to
2134 * dispatch.
22e2c507 2135 */
dd67d051 2136 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
22e2c507 2137 goto keep_queue;
6d048f53 2138
a36e71f9
JA
2139 /*
2140 * If another queue has a request waiting within our mean seek
2141 * distance, let it run. The expire code will check for close
2142 * cooperators and put the close queue at the front of the service
df5fe3e8 2143 * tree. If possible, merge the expiring queue with the new cfqq.
a36e71f9 2144 */
b3b6d040 2145 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
df5fe3e8
JM
2146 if (new_cfqq) {
2147 if (!cfqq->new_cfqq)
2148 cfq_setup_merge(cfqq, new_cfqq);
a36e71f9 2149 goto expire;
df5fe3e8 2150 }
a36e71f9 2151
6d048f53
JA
2152 /*
2153 * No requests pending. If the active queue still has requests in
2154 * flight or is idling for a new request, allow either of these
2155 * conditions to happen (or time out) before selecting a new queue.
2156 */
cc197479 2157 if (timer_pending(&cfqd->idle_slice_timer) ||
a6d44e98 2158 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
caaa5f9f
JA
2159 cfqq = NULL;
2160 goto keep_queue;
22e2c507
JA
2161 }
2162
3b18152c 2163expire:
6084cdda 2164 cfq_slice_expired(cfqd, 0);
3b18152c 2165new_queue:
718eee05
CZ
2166 /*
2167 * Current queue expired. Check if we have to switch to a new
2168 * service tree
2169 */
2170 if (!new_cfqq)
cdb16e8f 2171 cfq_choose_cfqg(cfqd);
718eee05 2172
a36e71f9 2173 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
22e2c507 2174keep_queue:
3b18152c 2175 return cfqq;
22e2c507
JA
2176}
2177
febffd61 2178static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
d9e7620e
JA
2179{
2180 int dispatched = 0;
2181
2182 while (cfqq->next_rq) {
2183 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2184 dispatched++;
2185 }
2186
2187 BUG_ON(!list_empty(&cfqq->fifo));
f04a6424
VG
2188
2189 /* By default cfqq is not expired if it is empty. Do it explicitly */
2190 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
d9e7620e
JA
2191 return dispatched;
2192}
2193
498d3aa2
JA
2194/*
2195 * Drain our current requests. Used for barriers and when switching
2196 * io schedulers on-the-fly.
2197 */
d9e7620e 2198static int cfq_forced_dispatch(struct cfq_data *cfqd)
1b5ed5e1 2199{
0871714e 2200 struct cfq_queue *cfqq;
d9e7620e 2201 int dispatched = 0;
cdb16e8f 2202
f04a6424
VG
2203 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2204 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1b5ed5e1 2205
6084cdda 2206 cfq_slice_expired(cfqd, 0);
1b5ed5e1
TH
2207 BUG_ON(cfqd->busy_queues);
2208
6923715a 2209 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1b5ed5e1
TH
2210 return dispatched;
2211}
2212
abc3c744
SL
2213static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2214 struct cfq_queue *cfqq)
2215{
2216 /* the queue hasn't finished any request, can't estimate */
2217 if (cfq_cfqq_slice_new(cfqq))
2218 return 1;
2219 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2220 cfqq->slice_end))
2221 return 1;
2222
2223 return 0;
2224}
2225
0b182d61 2226static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2f5cb738 2227{
2f5cb738 2228 unsigned int max_dispatch;
22e2c507 2229
5ad531db
JA
2230 /*
2231 * Drain async requests before we start sync IO
2232 */
53c583d2 2233 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
0b182d61 2234 return false;
5ad531db 2235
2f5cb738
JA
2236 /*
2237 * If this is an async queue and we have sync IO in flight, let it wait
2238 */
53c583d2 2239 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
0b182d61 2240 return false;
2f5cb738 2241
abc3c744 2242 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2f5cb738
JA
2243 if (cfq_class_idle(cfqq))
2244 max_dispatch = 1;
b4878f24 2245
2f5cb738
JA
2246 /*
2247 * Does this cfqq already have too much IO in flight?
2248 */
2249 if (cfqq->dispatched >= max_dispatch) {
2250 /*
2251 * idle queue must always only have a single IO in flight
2252 */
3ed9a296 2253 if (cfq_class_idle(cfqq))
0b182d61 2254 return false;
3ed9a296 2255
2f5cb738
JA
2256 /*
2257 * We have other queues, don't allow more IO from this one
2258 */
abc3c744 2259 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
0b182d61 2260 return false;
9ede209e 2261
365722bb 2262 /*
474b18cc 2263 * Sole queue user, no limit
365722bb 2264 */
abc3c744
SL
2265 if (cfqd->busy_queues == 1)
2266 max_dispatch = -1;
2267 else
2268 /*
2269 * Normally we start throttling cfqq when cfq_quantum/2
2270 * requests have been dispatched. But we can drive
2271 * deeper queue depths at the beginning of slice
2272 * subjected to upper limit of cfq_quantum.
2273 * */
2274 max_dispatch = cfqd->cfq_quantum;
8e296755
JA
2275 }
2276
2277 /*
2278 * Async queues must wait a bit before being allowed dispatch.
2279 * We also ramp up the dispatch depth gradually for async IO,
2280 * based on the last sync IO we serviced
2281 */
963b72fc 2282 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
573412b2 2283 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
8e296755 2284 unsigned int depth;
365722bb 2285
61f0c1dc 2286 depth = last_sync / cfqd->cfq_slice[1];
e00c54c3
JA
2287 if (!depth && !cfqq->dispatched)
2288 depth = 1;
8e296755
JA
2289 if (depth < max_dispatch)
2290 max_dispatch = depth;
2f5cb738 2291 }
3ed9a296 2292
0b182d61
JA
2293 /*
2294 * If we're below the current max, allow a dispatch
2295 */
2296 return cfqq->dispatched < max_dispatch;
2297}
2298
2299/*
2300 * Dispatch a request from cfqq, moving them to the request queue
2301 * dispatch list.
2302 */
2303static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2304{
2305 struct request *rq;
2306
2307 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2308
2309 if (!cfq_may_dispatch(cfqd, cfqq))
2310 return false;
2311
2312 /*
2313 * follow expired path, else get first next available
2314 */
2315 rq = cfq_check_fifo(cfqq);
2316 if (!rq)
2317 rq = cfqq->next_rq;
2318
2319 /*
2320 * insert request into driver dispatch list
2321 */
2322 cfq_dispatch_insert(cfqd->queue, rq);
2323
2324 if (!cfqd->active_cic) {
2325 struct cfq_io_context *cic = RQ_CIC(rq);
2326
2327 atomic_long_inc(&cic->ioc->refcount);
2328 cfqd->active_cic = cic;
2329 }
2330
2331 return true;
2332}
2333
2334/*
2335 * Find the cfqq that we need to service and move a request from that to the
2336 * dispatch list
2337 */
2338static int cfq_dispatch_requests(struct request_queue *q, int force)
2339{
2340 struct cfq_data *cfqd = q->elevator->elevator_data;
2341 struct cfq_queue *cfqq;
2342
2343 if (!cfqd->busy_queues)
2344 return 0;
2345
2346 if (unlikely(force))
2347 return cfq_forced_dispatch(cfqd);
2348
2349 cfqq = cfq_select_queue(cfqd);
2350 if (!cfqq)
8e296755
JA
2351 return 0;
2352
2f5cb738 2353 /*
0b182d61 2354 * Dispatch a request from this cfqq, if it is allowed
2f5cb738 2355 */
0b182d61
JA
2356 if (!cfq_dispatch_request(cfqd, cfqq))
2357 return 0;
2358
2f5cb738 2359 cfqq->slice_dispatch++;
b029195d 2360 cfq_clear_cfqq_must_dispatch(cfqq);
22e2c507 2361
2f5cb738
JA
2362 /*
2363 * expire an async queue immediately if it has used up its slice. idle
2364 * queue always expire after 1 dispatch round.
2365 */
2366 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2367 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2368 cfq_class_idle(cfqq))) {
2369 cfqq->slice_end = jiffies + 1;
2370 cfq_slice_expired(cfqd, 0);
1da177e4
LT
2371 }
2372
b217a903 2373 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2f5cb738 2374 return 1;
1da177e4
LT
2375}
2376
1da177e4 2377/*
5e705374
JA
2378 * task holds one reference to the queue, dropped when task exits. each rq
2379 * in-flight on this queue also holds a reference, dropped when rq is freed.
1da177e4 2380 *
b1c35769 2381 * Each cfq queue took a reference on the parent group. Drop it now.
1da177e4
LT
2382 * queue lock must be held here.
2383 */
2384static void cfq_put_queue(struct cfq_queue *cfqq)
2385{
22e2c507 2386 struct cfq_data *cfqd = cfqq->cfqd;
878eaddd 2387 struct cfq_group *cfqg, *orig_cfqg;
22e2c507
JA
2388
2389 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1da177e4
LT
2390
2391 if (!atomic_dec_and_test(&cfqq->ref))
2392 return;
2393
7b679138 2394 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1da177e4 2395 BUG_ON(rb_first(&cfqq->sort_list));
22e2c507 2396 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
b1c35769 2397 cfqg = cfqq->cfqg;
878eaddd 2398 orig_cfqg = cfqq->orig_cfqg;
1da177e4 2399
28f95cbc 2400 if (unlikely(cfqd->active_queue == cfqq)) {
6084cdda 2401 __cfq_slice_expired(cfqd, cfqq, 0);
23e018a1 2402 cfq_schedule_dispatch(cfqd);
28f95cbc 2403 }
22e2c507 2404
f04a6424 2405 BUG_ON(cfq_cfqq_on_rr(cfqq));
1da177e4 2406 kmem_cache_free(cfq_pool, cfqq);
b1c35769 2407 cfq_put_cfqg(cfqg);
878eaddd
VG
2408 if (orig_cfqg)
2409 cfq_put_cfqg(orig_cfqg);
1da177e4
LT
2410}
2411
d6de8be7
JA
2412/*
2413 * Must always be called with the rcu_read_lock() held
2414 */
07416d29
JA
2415static void
2416__call_for_each_cic(struct io_context *ioc,
2417 void (*func)(struct io_context *, struct cfq_io_context *))
2418{
2419 struct cfq_io_context *cic;
2420 struct hlist_node *n;
2421
2422 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2423 func(ioc, cic);
2424}
2425
4ac845a2 2426/*
34e6bbf2 2427 * Call func for each cic attached to this ioc.
4ac845a2 2428 */
34e6bbf2 2429static void
4ac845a2
JA
2430call_for_each_cic(struct io_context *ioc,
2431 void (*func)(struct io_context *, struct cfq_io_context *))
1da177e4 2432{
4ac845a2 2433 rcu_read_lock();
07416d29 2434 __call_for_each_cic(ioc, func);
4ac845a2 2435 rcu_read_unlock();
34e6bbf2
FC
2436}
2437
2438static void cfq_cic_free_rcu(struct rcu_head *head)
2439{
2440 struct cfq_io_context *cic;
2441
2442 cic = container_of(head, struct cfq_io_context, rcu_head);
2443
2444 kmem_cache_free(cfq_ioc_pool, cic);
245b2e70 2445 elv_ioc_count_dec(cfq_ioc_count);
34e6bbf2 2446
9a11b4ed
JA
2447 if (ioc_gone) {
2448 /*
2449 * CFQ scheduler is exiting, grab exit lock and check
2450 * the pending io context count. If it hits zero,
2451 * complete ioc_gone and set it back to NULL
2452 */
2453 spin_lock(&ioc_gone_lock);
245b2e70 2454 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
9a11b4ed
JA
2455 complete(ioc_gone);
2456 ioc_gone = NULL;
2457 }
2458 spin_unlock(&ioc_gone_lock);
2459 }
34e6bbf2 2460}
4ac845a2 2461
34e6bbf2
FC
2462static void cfq_cic_free(struct cfq_io_context *cic)
2463{
2464 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
4ac845a2
JA
2465}
2466
2467static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2468{
2469 unsigned long flags;
2470
2471 BUG_ON(!cic->dead_key);
2472
2473 spin_lock_irqsave(&ioc->lock, flags);
2474 radix_tree_delete(&ioc->radix_root, cic->dead_key);
ffc4e759 2475 hlist_del_rcu(&cic->cic_list);
4ac845a2
JA
2476 spin_unlock_irqrestore(&ioc->lock, flags);
2477
34e6bbf2 2478 cfq_cic_free(cic);
4ac845a2
JA
2479}
2480
d6de8be7
JA
2481/*
2482 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2483 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2484 * and ->trim() which is called with the task lock held
2485 */
4ac845a2
JA
2486static void cfq_free_io_context(struct io_context *ioc)
2487{
4ac845a2 2488 /*
34e6bbf2
FC
2489 * ioc->refcount is zero here, or we are called from elv_unregister(),
2490 * so no more cic's are allowed to be linked into this ioc. So it
2491 * should be ok to iterate over the known list, we will see all cic's
2492 * since no new ones are added.
4ac845a2 2493 */
07416d29 2494 __call_for_each_cic(ioc, cic_free_func);
1da177e4
LT
2495}
2496
89850f7e 2497static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1da177e4 2498{
df5fe3e8
JM
2499 struct cfq_queue *__cfqq, *next;
2500
28f95cbc 2501 if (unlikely(cfqq == cfqd->active_queue)) {
6084cdda 2502 __cfq_slice_expired(cfqd, cfqq, 0);
23e018a1 2503 cfq_schedule_dispatch(cfqd);
28f95cbc 2504 }
22e2c507 2505
df5fe3e8
JM
2506 /*
2507 * If this queue was scheduled to merge with another queue, be
2508 * sure to drop the reference taken on that queue (and others in
2509 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2510 */
2511 __cfqq = cfqq->new_cfqq;
2512 while (__cfqq) {
2513 if (__cfqq == cfqq) {
2514 WARN(1, "cfqq->new_cfqq loop detected\n");
2515 break;
2516 }
2517 next = __cfqq->new_cfqq;
2518 cfq_put_queue(__cfqq);
2519 __cfqq = next;
2520 }
2521
89850f7e
JA
2522 cfq_put_queue(cfqq);
2523}
22e2c507 2524
89850f7e
JA
2525static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2526 struct cfq_io_context *cic)
2527{
4faa3c81
FC
2528 struct io_context *ioc = cic->ioc;
2529
fc46379d 2530 list_del_init(&cic->queue_list);
4ac845a2
JA
2531
2532 /*
2533 * Make sure key == NULL is seen for dead queues
2534 */
fc46379d 2535 smp_wmb();
4ac845a2 2536 cic->dead_key = (unsigned long) cic->key;
fc46379d
JA
2537 cic->key = NULL;
2538
4faa3c81
FC
2539 if (ioc->ioc_data == cic)
2540 rcu_assign_pointer(ioc->ioc_data, NULL);
2541
ff6657c6
JA
2542 if (cic->cfqq[BLK_RW_ASYNC]) {
2543 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2544 cic->cfqq[BLK_RW_ASYNC] = NULL;
12a05732
AV
2545 }
2546
ff6657c6
JA
2547 if (cic->cfqq[BLK_RW_SYNC]) {
2548 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2549 cic->cfqq[BLK_RW_SYNC] = NULL;
12a05732 2550 }
89850f7e
JA
2551}
2552
4ac845a2
JA
2553static void cfq_exit_single_io_context(struct io_context *ioc,
2554 struct cfq_io_context *cic)
89850f7e
JA
2555{
2556 struct cfq_data *cfqd = cic->key;
2557
89850f7e 2558 if (cfqd) {
165125e1 2559 struct request_queue *q = cfqd->queue;
4ac845a2 2560 unsigned long flags;
89850f7e 2561
4ac845a2 2562 spin_lock_irqsave(q->queue_lock, flags);
62c1fe9d
JA
2563
2564 /*
2565 * Ensure we get a fresh copy of the ->key to prevent
2566 * race between exiting task and queue
2567 */
2568 smp_read_barrier_depends();
2569 if (cic->key)
2570 __cfq_exit_single_io_context(cfqd, cic);
2571
4ac845a2 2572 spin_unlock_irqrestore(q->queue_lock, flags);
89850f7e 2573 }
1da177e4
LT
2574}
2575
498d3aa2
JA
2576/*
2577 * The process that ioc belongs to has exited, we need to clean up
2578 * and put the internal structures we have that belongs to that process.
2579 */
e2d74ac0 2580static void cfq_exit_io_context(struct io_context *ioc)
1da177e4 2581{
4ac845a2 2582 call_for_each_cic(ioc, cfq_exit_single_io_context);
1da177e4
LT
2583}
2584
22e2c507 2585static struct cfq_io_context *
8267e268 2586cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1da177e4 2587{
b5deef90 2588 struct cfq_io_context *cic;
1da177e4 2589
94f6030c
CL
2590 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2591 cfqd->queue->node);
1da177e4 2592 if (cic) {
22e2c507 2593 cic->last_end_request = jiffies;
553698f9 2594 INIT_LIST_HEAD(&cic->queue_list);
ffc4e759 2595 INIT_HLIST_NODE(&cic->cic_list);
22e2c507
JA
2596 cic->dtor = cfq_free_io_context;
2597 cic->exit = cfq_exit_io_context;
245b2e70 2598 elv_ioc_count_inc(cfq_ioc_count);
1da177e4
LT
2599 }
2600
2601 return cic;
2602}
2603
fd0928df 2604static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
22e2c507
JA
2605{
2606 struct task_struct *tsk = current;
2607 int ioprio_class;
2608
3b18152c 2609 if (!cfq_cfqq_prio_changed(cfqq))
22e2c507
JA
2610 return;
2611
fd0928df 2612 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
22e2c507 2613 switch (ioprio_class) {
fe094d98
JA
2614 default:
2615 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2616 case IOPRIO_CLASS_NONE:
2617 /*
6d63c275 2618 * no prio set, inherit CPU scheduling settings
fe094d98
JA
2619 */
2620 cfqq->ioprio = task_nice_ioprio(tsk);
6d63c275 2621 cfqq->ioprio_class = task_nice_ioclass(tsk);
fe094d98
JA
2622 break;
2623 case IOPRIO_CLASS_RT:
2624 cfqq->ioprio = task_ioprio(ioc);
2625 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2626 break;
2627 case IOPRIO_CLASS_BE:
2628 cfqq->ioprio = task_ioprio(ioc);
2629 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2630 break;
2631 case IOPRIO_CLASS_IDLE:
2632 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2633 cfqq->ioprio = 7;
2634 cfq_clear_cfqq_idle_window(cfqq);
2635 break;
22e2c507
JA
2636 }
2637
2638 /*
2639 * keep track of original prio settings in case we have to temporarily
2640 * elevate the priority of this queue
2641 */
2642 cfqq->org_ioprio = cfqq->ioprio;
2643 cfqq->org_ioprio_class = cfqq->ioprio_class;
3b18152c 2644 cfq_clear_cfqq_prio_changed(cfqq);
22e2c507
JA
2645}
2646
febffd61 2647static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
22e2c507 2648{
478a82b0
AV
2649 struct cfq_data *cfqd = cic->key;
2650 struct cfq_queue *cfqq;
c1b707d2 2651 unsigned long flags;
35e6077c 2652
caaa5f9f
JA
2653 if (unlikely(!cfqd))
2654 return;
2655
c1b707d2 2656 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
caaa5f9f 2657
ff6657c6 2658 cfqq = cic->cfqq[BLK_RW_ASYNC];
caaa5f9f
JA
2659 if (cfqq) {
2660 struct cfq_queue *new_cfqq;
ff6657c6
JA
2661 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2662 GFP_ATOMIC);
caaa5f9f 2663 if (new_cfqq) {
ff6657c6 2664 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
caaa5f9f
JA
2665 cfq_put_queue(cfqq);
2666 }
22e2c507 2667 }
caaa5f9f 2668
ff6657c6 2669 cfqq = cic->cfqq[BLK_RW_SYNC];
caaa5f9f
JA
2670 if (cfqq)
2671 cfq_mark_cfqq_prio_changed(cfqq);
2672
c1b707d2 2673 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
22e2c507
JA
2674}
2675
fc46379d 2676static void cfq_ioc_set_ioprio(struct io_context *ioc)
22e2c507 2677{
4ac845a2 2678 call_for_each_cic(ioc, changed_ioprio);
fc46379d 2679 ioc->ioprio_changed = 0;
22e2c507
JA
2680}
2681
d5036d77 2682static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
a6151c3a 2683 pid_t pid, bool is_sync)
d5036d77
JA
2684{
2685 RB_CLEAR_NODE(&cfqq->rb_node);
2686 RB_CLEAR_NODE(&cfqq->p_node);
2687 INIT_LIST_HEAD(&cfqq->fifo);
2688
2689 atomic_set(&cfqq->ref, 0);
2690 cfqq->cfqd = cfqd;
2691
2692 cfq_mark_cfqq_prio_changed(cfqq);
2693
2694 if (is_sync) {
2695 if (!cfq_class_idle(cfqq))
2696 cfq_mark_cfqq_idle_window(cfqq);
2697 cfq_mark_cfqq_sync(cfqq);
2698 }
2699 cfqq->pid = pid;
2700}
2701
24610333
VG
2702#ifdef CONFIG_CFQ_GROUP_IOSCHED
2703static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2704{
2705 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2706 struct cfq_data *cfqd = cic->key;
2707 unsigned long flags;
2708 struct request_queue *q;
2709
2710 if (unlikely(!cfqd))
2711 return;
2712
2713 q = cfqd->queue;
2714
2715 spin_lock_irqsave(q->queue_lock, flags);
2716
2717 if (sync_cfqq) {
2718 /*
2719 * Drop reference to sync queue. A new sync queue will be
2720 * assigned in new group upon arrival of a fresh request.
2721 */
2722 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2723 cic_set_cfqq(cic, NULL, 1);
2724 cfq_put_queue(sync_cfqq);
2725 }
2726
2727 spin_unlock_irqrestore(q->queue_lock, flags);
2728}
2729
2730static void cfq_ioc_set_cgroup(struct io_context *ioc)
2731{
2732 call_for_each_cic(ioc, changed_cgroup);
2733 ioc->cgroup_changed = 0;
2734}
2735#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2736
22e2c507 2737static struct cfq_queue *
a6151c3a 2738cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
fd0928df 2739 struct io_context *ioc, gfp_t gfp_mask)
22e2c507 2740{
22e2c507 2741 struct cfq_queue *cfqq, *new_cfqq = NULL;
91fac317 2742 struct cfq_io_context *cic;
cdb16e8f 2743 struct cfq_group *cfqg;
22e2c507
JA
2744
2745retry:
cdb16e8f 2746 cfqg = cfq_get_cfqg(cfqd, 1);
4ac845a2 2747 cic = cfq_cic_lookup(cfqd, ioc);
91fac317
VT
2748 /* cic always exists here */
2749 cfqq = cic_to_cfqq(cic, is_sync);
22e2c507 2750
6118b70b
JA
2751 /*
2752 * Always try a new alloc if we fell back to the OOM cfqq
2753 * originally, since it should just be a temporary situation.
2754 */
2755 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2756 cfqq = NULL;
22e2c507
JA
2757 if (new_cfqq) {
2758 cfqq = new_cfqq;
2759 new_cfqq = NULL;
2760 } else if (gfp_mask & __GFP_WAIT) {
2761 spin_unlock_irq(cfqd->queue->queue_lock);
94f6030c 2762 new_cfqq = kmem_cache_alloc_node(cfq_pool,
6118b70b 2763 gfp_mask | __GFP_ZERO,
94f6030c 2764 cfqd->queue->node);
22e2c507 2765 spin_lock_irq(cfqd->queue->queue_lock);
6118b70b
JA
2766 if (new_cfqq)
2767 goto retry;
22e2c507 2768 } else {
94f6030c
CL
2769 cfqq = kmem_cache_alloc_node(cfq_pool,
2770 gfp_mask | __GFP_ZERO,
2771 cfqd->queue->node);
22e2c507
JA
2772 }
2773
6118b70b
JA
2774 if (cfqq) {
2775 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2776 cfq_init_prio_data(cfqq, ioc);
cdb16e8f 2777 cfq_link_cfqq_cfqg(cfqq, cfqg);
6118b70b
JA
2778 cfq_log_cfqq(cfqd, cfqq, "alloced");
2779 } else
2780 cfqq = &cfqd->oom_cfqq;
22e2c507
JA
2781 }
2782
2783 if (new_cfqq)
2784 kmem_cache_free(cfq_pool, new_cfqq);
2785
22e2c507
JA
2786 return cfqq;
2787}
2788
c2dea2d1
VT
2789static struct cfq_queue **
2790cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2791{
fe094d98 2792 switch (ioprio_class) {
c2dea2d1
VT
2793 case IOPRIO_CLASS_RT:
2794 return &cfqd->async_cfqq[0][ioprio];
2795 case IOPRIO_CLASS_BE:
2796 return &cfqd->async_cfqq[1][ioprio];
2797 case IOPRIO_CLASS_IDLE:
2798 return &cfqd->async_idle_cfqq;
2799 default:
2800 BUG();
2801 }
2802}
2803
15c31be4 2804static struct cfq_queue *
a6151c3a 2805cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
15c31be4
JA
2806 gfp_t gfp_mask)
2807{
fd0928df
JA
2808 const int ioprio = task_ioprio(ioc);
2809 const int ioprio_class = task_ioprio_class(ioc);
c2dea2d1 2810 struct cfq_queue **async_cfqq = NULL;
15c31be4
JA
2811 struct cfq_queue *cfqq = NULL;
2812
c2dea2d1
VT
2813 if (!is_sync) {
2814 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2815 cfqq = *async_cfqq;
2816 }
2817
6118b70b 2818 if (!cfqq)
fd0928df 2819 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
15c31be4
JA
2820
2821 /*
2822 * pin the queue now that it's allocated, scheduler exit will prune it
2823 */
c2dea2d1 2824 if (!is_sync && !(*async_cfqq)) {
15c31be4 2825 atomic_inc(&cfqq->ref);
c2dea2d1 2826 *async_cfqq = cfqq;
15c31be4
JA
2827 }
2828
2829 atomic_inc(&cfqq->ref);
2830 return cfqq;
2831}
2832
498d3aa2
JA
2833/*
2834 * We drop cfq io contexts lazily, so we may find a dead one.
2835 */
dbecf3ab 2836static void
4ac845a2
JA
2837cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2838 struct cfq_io_context *cic)
dbecf3ab 2839{
4ac845a2
JA
2840 unsigned long flags;
2841
fc46379d 2842 WARN_ON(!list_empty(&cic->queue_list));
597bc485 2843
4ac845a2
JA
2844 spin_lock_irqsave(&ioc->lock, flags);
2845
4faa3c81 2846 BUG_ON(ioc->ioc_data == cic);
597bc485 2847
4ac845a2 2848 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
ffc4e759 2849 hlist_del_rcu(&cic->cic_list);
4ac845a2
JA
2850 spin_unlock_irqrestore(&ioc->lock, flags);
2851
2852 cfq_cic_free(cic);
dbecf3ab
OH
2853}
2854
e2d74ac0 2855static struct cfq_io_context *
4ac845a2 2856cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
e2d74ac0 2857{
e2d74ac0 2858 struct cfq_io_context *cic;
d6de8be7 2859 unsigned long flags;
4ac845a2 2860 void *k;
e2d74ac0 2861
91fac317
VT
2862 if (unlikely(!ioc))
2863 return NULL;
2864
d6de8be7
JA
2865 rcu_read_lock();
2866
597bc485
JA
2867 /*
2868 * we maintain a last-hit cache, to avoid browsing over the tree
2869 */
4ac845a2 2870 cic = rcu_dereference(ioc->ioc_data);
d6de8be7
JA
2871 if (cic && cic->key == cfqd) {
2872 rcu_read_unlock();
597bc485 2873 return cic;
d6de8be7 2874 }
597bc485 2875
4ac845a2 2876 do {
4ac845a2
JA
2877 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2878 rcu_read_unlock();
2879 if (!cic)
2880 break;
be3b0753
OH
2881 /* ->key must be copied to avoid race with cfq_exit_queue() */
2882 k = cic->key;
2883 if (unlikely(!k)) {
4ac845a2 2884 cfq_drop_dead_cic(cfqd, ioc, cic);
d6de8be7 2885 rcu_read_lock();
4ac845a2 2886 continue;
dbecf3ab 2887 }
e2d74ac0 2888
d6de8be7 2889 spin_lock_irqsave(&ioc->lock, flags);
4ac845a2 2890 rcu_assign_pointer(ioc->ioc_data, cic);
d6de8be7 2891 spin_unlock_irqrestore(&ioc->lock, flags);
4ac845a2
JA
2892 break;
2893 } while (1);
e2d74ac0 2894
4ac845a2 2895 return cic;
e2d74ac0
JA
2896}
2897
4ac845a2
JA
2898/*
2899 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2900 * the process specific cfq io context when entered from the block layer.
2901 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2902 */
febffd61
JA
2903static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2904 struct cfq_io_context *cic, gfp_t gfp_mask)
e2d74ac0 2905{
0261d688 2906 unsigned long flags;
4ac845a2 2907 int ret;
e2d74ac0 2908
4ac845a2
JA
2909 ret = radix_tree_preload(gfp_mask);
2910 if (!ret) {
2911 cic->ioc = ioc;
2912 cic->key = cfqd;
e2d74ac0 2913
4ac845a2
JA
2914 spin_lock_irqsave(&ioc->lock, flags);
2915 ret = radix_tree_insert(&ioc->radix_root,
2916 (unsigned long) cfqd, cic);
ffc4e759
JA
2917 if (!ret)
2918 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
4ac845a2 2919 spin_unlock_irqrestore(&ioc->lock, flags);
e2d74ac0 2920
4ac845a2
JA
2921 radix_tree_preload_end();
2922
2923 if (!ret) {
2924 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2925 list_add(&cic->queue_list, &cfqd->cic_list);
2926 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2927 }
e2d74ac0
JA
2928 }
2929
4ac845a2
JA
2930 if (ret)
2931 printk(KERN_ERR "cfq: cic link failed!\n");
fc46379d 2932
4ac845a2 2933 return ret;
e2d74ac0
JA
2934}
2935
1da177e4
LT
2936/*
2937 * Setup general io context and cfq io context. There can be several cfq
2938 * io contexts per general io context, if this process is doing io to more
e2d74ac0 2939 * than one device managed by cfq.
1da177e4
LT
2940 */
2941static struct cfq_io_context *
e2d74ac0 2942cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1da177e4 2943{
22e2c507 2944 struct io_context *ioc = NULL;
1da177e4 2945 struct cfq_io_context *cic;
1da177e4 2946
22e2c507 2947 might_sleep_if(gfp_mask & __GFP_WAIT);
1da177e4 2948
b5deef90 2949 ioc = get_io_context(gfp_mask, cfqd->queue->node);
1da177e4
LT
2950 if (!ioc)
2951 return NULL;
2952
4ac845a2 2953 cic = cfq_cic_lookup(cfqd, ioc);
e2d74ac0
JA
2954 if (cic)
2955 goto out;
1da177e4 2956
e2d74ac0
JA
2957 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2958 if (cic == NULL)
2959 goto err;
1da177e4 2960
4ac845a2
JA
2961 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2962 goto err_free;
2963
1da177e4 2964out:
fc46379d
JA
2965 smp_read_barrier_depends();
2966 if (unlikely(ioc->ioprio_changed))
2967 cfq_ioc_set_ioprio(ioc);
2968
24610333
VG
2969#ifdef CONFIG_CFQ_GROUP_IOSCHED
2970 if (unlikely(ioc->cgroup_changed))
2971 cfq_ioc_set_cgroup(ioc);
2972#endif
1da177e4 2973 return cic;
4ac845a2
JA
2974err_free:
2975 cfq_cic_free(cic);
1da177e4
LT
2976err:
2977 put_io_context(ioc);
2978 return NULL;
2979}
2980
22e2c507
JA
2981static void
2982cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1da177e4 2983{
aaf1228d
JA
2984 unsigned long elapsed = jiffies - cic->last_end_request;
2985 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
db3b5848 2986
22e2c507
JA
2987 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2988 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2989 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2990}
1da177e4 2991
206dc69b 2992static void
b2c18e1e 2993cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
6d048f53 2994 struct request *rq)
206dc69b 2995{
3dde36dd 2996 sector_t sdist = 0;
41647e7a 2997 sector_t n_sec = blk_rq_sectors(rq);
3dde36dd
CZ
2998 if (cfqq->last_request_pos) {
2999 if (cfqq->last_request_pos < blk_rq_pos(rq))
3000 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3001 else
3002 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3003 }
206dc69b 3004
3dde36dd 3005 cfqq->seek_history <<= 1;
41647e7a
CZ
3006 if (blk_queue_nonrot(cfqd->queue))
3007 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3008 else
3009 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
206dc69b 3010}
1da177e4 3011
22e2c507
JA
3012/*
3013 * Disable idle window if the process thinks too long or seeks so much that
3014 * it doesn't matter
3015 */
3016static void
3017cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3018 struct cfq_io_context *cic)
3019{
7b679138 3020 int old_idle, enable_idle;
1be92f2f 3021
0871714e
JA
3022 /*
3023 * Don't idle for async or idle io prio class
3024 */
3025 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1be92f2f
JA
3026 return;
3027
c265a7f4 3028 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1da177e4 3029
76280aff
CZ
3030 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3031 cfq_mark_cfqq_deep(cfqq);
3032
66dac98e 3033 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3dde36dd 3034 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
22e2c507
JA
3035 enable_idle = 0;
3036 else if (sample_valid(cic->ttime_samples)) {
718eee05 3037 if (cic->ttime_mean > cfqd->cfq_slice_idle)
22e2c507
JA
3038 enable_idle = 0;
3039 else
3040 enable_idle = 1;
1da177e4
LT
3041 }
3042
7b679138
JA
3043 if (old_idle != enable_idle) {
3044 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3045 if (enable_idle)
3046 cfq_mark_cfqq_idle_window(cfqq);
3047 else
3048 cfq_clear_cfqq_idle_window(cfqq);
3049 }
22e2c507 3050}
1da177e4 3051
22e2c507
JA
3052/*
3053 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3054 * no or if we aren't sure, a 1 will cause a preempt.
3055 */
a6151c3a 3056static bool
22e2c507 3057cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
5e705374 3058 struct request *rq)
22e2c507 3059{
6d048f53 3060 struct cfq_queue *cfqq;
22e2c507 3061
6d048f53
JA
3062 cfqq = cfqd->active_queue;
3063 if (!cfqq)
a6151c3a 3064 return false;
22e2c507 3065
6d048f53 3066 if (cfq_class_idle(new_cfqq))
a6151c3a 3067 return false;
22e2c507
JA
3068
3069 if (cfq_class_idle(cfqq))
a6151c3a 3070 return true;
1e3335de 3071
875feb63
DS
3072 /*
3073 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3074 */
3075 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3076 return false;
3077
374f84ac
JA
3078 /*
3079 * if the new request is sync, but the currently running queue is
3080 * not, let the sync request have priority.
3081 */
5e705374 3082 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
a6151c3a 3083 return true;
1e3335de 3084
8682e1f1
VG
3085 if (new_cfqq->cfqg != cfqq->cfqg)
3086 return false;
3087
3088 if (cfq_slice_used(cfqq))
3089 return true;
3090
3091 /* Allow preemption only if we are idling on sync-noidle tree */
3092 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3093 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3094 new_cfqq->service_tree->count == 2 &&
3095 RB_EMPTY_ROOT(&cfqq->sort_list))
3096 return true;
3097
374f84ac
JA
3098 /*
3099 * So both queues are sync. Let the new request get disk time if
3100 * it's a metadata request and the current queue is doing regular IO.
3101 */
3102 if (rq_is_meta(rq) && !cfqq->meta_pending)
e6ec4fe2 3103 return true;
22e2c507 3104
3a9a3f6c
DS
3105 /*
3106 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3107 */
3108 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
a6151c3a 3109 return true;
3a9a3f6c 3110
1e3335de 3111 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
a6151c3a 3112 return false;
1e3335de
JA
3113
3114 /*
3115 * if this request is as-good as one we would expect from the
3116 * current cfqq, let it preempt
3117 */
e9ce335d 3118 if (cfq_rq_close(cfqd, cfqq, rq))
a6151c3a 3119 return true;
1e3335de 3120
a6151c3a 3121 return false;
22e2c507
JA
3122}
3123
3124/*
3125 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3126 * let it have half of its nominal slice.
3127 */
3128static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3129{
7b679138 3130 cfq_log_cfqq(cfqd, cfqq, "preempt");
6084cdda 3131 cfq_slice_expired(cfqd, 1);
22e2c507 3132
bf572256
JA
3133 /*
3134 * Put the new queue at the front of the of the current list,
3135 * so we know that it will be selected next.
3136 */
3137 BUG_ON(!cfq_cfqq_on_rr(cfqq));
edd75ffd
JA
3138
3139 cfq_service_tree_add(cfqd, cfqq, 1);
bf572256 3140
44f7c160
JA
3141 cfqq->slice_end = 0;
3142 cfq_mark_cfqq_slice_new(cfqq);
22e2c507
JA
3143}
3144
22e2c507 3145/*
5e705374 3146 * Called when a new fs request (rq) is added (to cfqq). Check if there's
22e2c507
JA
3147 * something we should do about it
3148 */
3149static void
5e705374
JA
3150cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3151 struct request *rq)
22e2c507 3152{
5e705374 3153 struct cfq_io_context *cic = RQ_CIC(rq);
12e9fddd 3154
45333d5a 3155 cfqd->rq_queued++;
374f84ac
JA
3156 if (rq_is_meta(rq))
3157 cfqq->meta_pending++;
3158
9c2c38a1 3159 cfq_update_io_thinktime(cfqd, cic);
b2c18e1e 3160 cfq_update_io_seektime(cfqd, cfqq, rq);
9c2c38a1
JA
3161 cfq_update_idle_window(cfqd, cfqq, cic);
3162
b2c18e1e 3163 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
22e2c507
JA
3164
3165 if (cfqq == cfqd->active_queue) {
3166 /*
b029195d
JA
3167 * Remember that we saw a request from this process, but
3168 * don't start queuing just yet. Otherwise we risk seeing lots
3169 * of tiny requests, because we disrupt the normal plugging
d6ceb25e
JA
3170 * and merging. If the request is already larger than a single
3171 * page, let it rip immediately. For that case we assume that
2d870722
JA
3172 * merging is already done. Ditto for a busy system that
3173 * has other work pending, don't risk delaying until the
3174 * idle timer unplug to continue working.
22e2c507 3175 */
d6ceb25e 3176 if (cfq_cfqq_wait_request(cfqq)) {
2d870722
JA
3177 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3178 cfqd->busy_queues > 1) {
d6ceb25e 3179 del_timer(&cfqd->idle_slice_timer);
554554f6 3180 cfq_clear_cfqq_wait_request(cfqq);
bf791937
VG
3181 __blk_run_queue(cfqd->queue);
3182 } else
3183 cfq_mark_cfqq_must_dispatch(cfqq);
d6ceb25e 3184 }
5e705374 3185 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
22e2c507
JA
3186 /*
3187 * not the active queue - expire current slice if it is
3188 * idle and has expired it's mean thinktime or this new queue
3a9a3f6c
DS
3189 * has some old slice time left and is of higher priority or
3190 * this new queue is RT and the current one is BE
22e2c507
JA
3191 */
3192 cfq_preempt_queue(cfqd, cfqq);
a7f55792 3193 __blk_run_queue(cfqd->queue);
22e2c507 3194 }
1da177e4
LT
3195}
3196
165125e1 3197static void cfq_insert_request(struct request_queue *q, struct request *rq)
1da177e4 3198{
b4878f24 3199 struct cfq_data *cfqd = q->elevator->elevator_data;
5e705374 3200 struct cfq_queue *cfqq = RQ_CFQQ(rq);
22e2c507 3201
7b679138 3202 cfq_log_cfqq(cfqd, cfqq, "insert_request");
fd0928df 3203 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1da177e4 3204
30996f40 3205 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
22e2c507 3206 list_add_tail(&rq->queuelist, &cfqq->fifo);
aa6f6a3d 3207 cfq_add_rq_rb(rq);
22e2c507 3208
5e705374 3209 cfq_rq_enqueued(cfqd, cfqq, rq);
1da177e4
LT
3210}
3211
45333d5a
AC
3212/*
3213 * Update hw_tag based on peak queue depth over 50 samples under
3214 * sufficient load.
3215 */
3216static void cfq_update_hw_tag(struct cfq_data *cfqd)
3217{
1a1238a7
SL
3218 struct cfq_queue *cfqq = cfqd->active_queue;
3219
53c583d2
CZ
3220 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3221 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
e459dd08
CZ
3222
3223 if (cfqd->hw_tag == 1)
3224 return;
45333d5a
AC
3225
3226 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
53c583d2 3227 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
45333d5a
AC
3228 return;
3229
1a1238a7
SL
3230 /*
3231 * If active queue hasn't enough requests and can idle, cfq might not
3232 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3233 * case
3234 */
3235 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3236 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
53c583d2 3237 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
1a1238a7
SL
3238 return;
3239
45333d5a
AC
3240 if (cfqd->hw_tag_samples++ < 50)
3241 return;
3242
e459dd08 3243 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
45333d5a
AC
3244 cfqd->hw_tag = 1;
3245 else
3246 cfqd->hw_tag = 0;
45333d5a
AC
3247}
3248
7667aa06
VG
3249static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3250{
3251 struct cfq_io_context *cic = cfqd->active_cic;
3252
3253 /* If there are other queues in the group, don't wait */
3254 if (cfqq->cfqg->nr_cfqq > 1)
3255 return false;
3256
3257 if (cfq_slice_used(cfqq))
3258 return true;
3259
3260 /* if slice left is less than think time, wait busy */
3261 if (cic && sample_valid(cic->ttime_samples)
3262 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3263 return true;
3264
3265 /*
3266 * If think times is less than a jiffy than ttime_mean=0 and above
3267 * will not be true. It might happen that slice has not expired yet
3268 * but will expire soon (4-5 ns) during select_queue(). To cover the
3269 * case where think time is less than a jiffy, mark the queue wait
3270 * busy if only 1 jiffy is left in the slice.
3271 */
3272 if (cfqq->slice_end - jiffies == 1)
3273 return true;
3274
3275 return false;
3276}
3277
165125e1 3278static void cfq_completed_request(struct request_queue *q, struct request *rq)
1da177e4 3279{
5e705374 3280 struct cfq_queue *cfqq = RQ_CFQQ(rq);
b4878f24 3281 struct cfq_data *cfqd = cfqq->cfqd;
5380a101 3282 const int sync = rq_is_sync(rq);
b4878f24 3283 unsigned long now;
1da177e4 3284
b4878f24 3285 now = jiffies;
2868ef7b 3286 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
1da177e4 3287
45333d5a
AC
3288 cfq_update_hw_tag(cfqd);
3289
53c583d2 3290 WARN_ON(!cfqd->rq_in_driver);
6d048f53 3291 WARN_ON(!cfqq->dispatched);
53c583d2 3292 cfqd->rq_in_driver--;
6d048f53 3293 cfqq->dispatched--;
1da177e4 3294
53c583d2 3295 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3ed9a296 3296
365722bb 3297 if (sync) {
5e705374 3298 RQ_CIC(rq)->last_end_request = now;
573412b2
CZ
3299 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3300 cfqd->last_delayed_sync = now;
365722bb 3301 }
caaa5f9f
JA
3302
3303 /*
3304 * If this is the active queue, check if it needs to be expired,
3305 * or if we want to idle in case it has no pending requests.
3306 */
3307 if (cfqd->active_queue == cfqq) {
a36e71f9
JA
3308 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3309
44f7c160
JA
3310 if (cfq_cfqq_slice_new(cfqq)) {
3311 cfq_set_prio_slice(cfqd, cfqq);
3312 cfq_clear_cfqq_slice_new(cfqq);
3313 }
f75edf2d
VG
3314
3315 /*
7667aa06
VG
3316 * Should we wait for next request to come in before we expire
3317 * the queue.
f75edf2d 3318 */
7667aa06 3319 if (cfq_should_wait_busy(cfqd, cfqq)) {
f75edf2d
VG
3320 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3321 cfq_mark_cfqq_wait_busy(cfqq);
b1ffe737 3322 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
f75edf2d
VG
3323 }
3324
a36e71f9 3325 /*
8e550632
CZ
3326 * Idling is not enabled on:
3327 * - expired queues
3328 * - idle-priority queues
3329 * - async queues
3330 * - queues with still some requests queued
3331 * - when there is a close cooperator
a36e71f9 3332 */
0871714e 3333 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
6084cdda 3334 cfq_slice_expired(cfqd, 1);
8e550632
CZ
3335 else if (sync && cfqq_empty &&
3336 !cfq_close_cooperator(cfqd, cfqq)) {
3337 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3338 /*
3339 * Idling is enabled for SYNC_WORKLOAD.
3340 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3341 * only if we processed at least one !rq_noidle request
3342 */
3343 if (cfqd->serving_type == SYNC_WORKLOAD
c04645e5
VG
3344 || cfqd->noidle_tree_requires_idle
3345 || cfqq->cfqg->nr_cfqq == 1)
8e550632
CZ
3346 cfq_arm_slice_timer(cfqd);
3347 }
caaa5f9f 3348 }
6d048f53 3349
53c583d2 3350 if (!cfqd->rq_in_driver)
23e018a1 3351 cfq_schedule_dispatch(cfqd);
1da177e4
LT
3352}
3353
22e2c507
JA
3354/*
3355 * we temporarily boost lower priority queues if they are holding fs exclusive
3356 * resources. they are boosted to normal prio (CLASS_BE/4)
3357 */
3358static void cfq_prio_boost(struct cfq_queue *cfqq)
1da177e4 3359{
22e2c507
JA
3360 if (has_fs_excl()) {
3361 /*
3362 * boost idle prio on transactions that would lock out other
3363 * users of the filesystem
3364 */
3365 if (cfq_class_idle(cfqq))
3366 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3367 if (cfqq->ioprio > IOPRIO_NORM)
3368 cfqq->ioprio = IOPRIO_NORM;
3369 } else {
3370 /*
dddb7451 3371 * unboost the queue (if needed)
22e2c507 3372 */
dddb7451
CZ
3373 cfqq->ioprio_class = cfqq->org_ioprio_class;
3374 cfqq->ioprio = cfqq->org_ioprio;
22e2c507 3375 }
22e2c507 3376}
1da177e4 3377
89850f7e 3378static inline int __cfq_may_queue(struct cfq_queue *cfqq)
22e2c507 3379{
1b379d8d 3380 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3b18152c 3381 cfq_mark_cfqq_must_alloc_slice(cfqq);
22e2c507 3382 return ELV_MQUEUE_MUST;
3b18152c 3383 }
1da177e4 3384
22e2c507 3385 return ELV_MQUEUE_MAY;
22e2c507
JA
3386}
3387
165125e1 3388static int cfq_may_queue(struct request_queue *q, int rw)
22e2c507
JA
3389{
3390 struct cfq_data *cfqd = q->elevator->elevator_data;
3391 struct task_struct *tsk = current;
91fac317 3392 struct cfq_io_context *cic;
22e2c507
JA
3393 struct cfq_queue *cfqq;
3394
3395 /*
3396 * don't force setup of a queue from here, as a call to may_queue
3397 * does not necessarily imply that a request actually will be queued.
3398 * so just lookup a possibly existing queue, or return 'may queue'
3399 * if that fails
3400 */
4ac845a2 3401 cic = cfq_cic_lookup(cfqd, tsk->io_context);
91fac317
VT
3402 if (!cic)
3403 return ELV_MQUEUE_MAY;
3404
b0b78f81 3405 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
22e2c507 3406 if (cfqq) {
fd0928df 3407 cfq_init_prio_data(cfqq, cic->ioc);
22e2c507
JA
3408 cfq_prio_boost(cfqq);
3409
89850f7e 3410 return __cfq_may_queue(cfqq);
22e2c507
JA
3411 }
3412
3413 return ELV_MQUEUE_MAY;
1da177e4
LT
3414}
3415
1da177e4
LT
3416/*
3417 * queue lock held here
3418 */
bb37b94c 3419static void cfq_put_request(struct request *rq)
1da177e4 3420{
5e705374 3421 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1da177e4 3422
5e705374 3423 if (cfqq) {
22e2c507 3424 const int rw = rq_data_dir(rq);
1da177e4 3425
22e2c507
JA
3426 BUG_ON(!cfqq->allocated[rw]);
3427 cfqq->allocated[rw]--;
1da177e4 3428
5e705374 3429 put_io_context(RQ_CIC(rq)->ioc);
1da177e4 3430
1da177e4 3431 rq->elevator_private = NULL;
5e705374 3432 rq->elevator_private2 = NULL;
1da177e4 3433
1da177e4
LT
3434 cfq_put_queue(cfqq);
3435 }
3436}
3437
df5fe3e8
JM
3438static struct cfq_queue *
3439cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3440 struct cfq_queue *cfqq)
3441{
3442 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3443 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
b3b6d040 3444 cfq_mark_cfqq_coop(cfqq->new_cfqq);
df5fe3e8
JM
3445 cfq_put_queue(cfqq);
3446 return cic_to_cfqq(cic, 1);
3447}
3448
e6c5bc73
JM
3449/*
3450 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3451 * was the last process referring to said cfqq.
3452 */
3453static struct cfq_queue *
3454split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3455{
3456 if (cfqq_process_refs(cfqq) == 1) {
e6c5bc73
JM
3457 cfqq->pid = current->pid;
3458 cfq_clear_cfqq_coop(cfqq);
ae54abed 3459 cfq_clear_cfqq_split_coop(cfqq);
e6c5bc73
JM
3460 return cfqq;
3461 }
3462
3463 cic_set_cfqq(cic, NULL, 1);
3464 cfq_put_queue(cfqq);
3465 return NULL;
3466}
1da177e4 3467/*
22e2c507 3468 * Allocate cfq data structures associated with this request.
1da177e4 3469 */
22e2c507 3470static int
165125e1 3471cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
1da177e4
LT
3472{
3473 struct cfq_data *cfqd = q->elevator->elevator_data;
3474 struct cfq_io_context *cic;
3475 const int rw = rq_data_dir(rq);
a6151c3a 3476 const bool is_sync = rq_is_sync(rq);
22e2c507 3477 struct cfq_queue *cfqq;
1da177e4
LT
3478 unsigned long flags;
3479
3480 might_sleep_if(gfp_mask & __GFP_WAIT);
3481
e2d74ac0 3482 cic = cfq_get_io_context(cfqd, gfp_mask);
22e2c507 3483
1da177e4
LT
3484 spin_lock_irqsave(q->queue_lock, flags);
3485
22e2c507
JA
3486 if (!cic)
3487 goto queue_fail;
3488
e6c5bc73 3489new_queue:
91fac317 3490 cfqq = cic_to_cfqq(cic, is_sync);
32f2e807 3491 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
fd0928df 3492 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
91fac317 3493 cic_set_cfqq(cic, cfqq, is_sync);
df5fe3e8 3494 } else {
e6c5bc73
JM
3495 /*
3496 * If the queue was seeky for too long, break it apart.
3497 */
ae54abed 3498 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
e6c5bc73
JM
3499 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3500 cfqq = split_cfqq(cic, cfqq);
3501 if (!cfqq)
3502 goto new_queue;
3503 }
3504
df5fe3e8
JM
3505 /*
3506 * Check to see if this queue is scheduled to merge with
3507 * another, closely cooperating queue. The merging of
3508 * queues happens here as it must be done in process context.
3509 * The reference on new_cfqq was taken in merge_cfqqs.
3510 */
3511 if (cfqq->new_cfqq)
3512 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
91fac317 3513 }
1da177e4
LT
3514
3515 cfqq->allocated[rw]++;
22e2c507 3516 atomic_inc(&cfqq->ref);
1da177e4 3517
5e705374 3518 spin_unlock_irqrestore(q->queue_lock, flags);
3b18152c 3519
5e705374
JA
3520 rq->elevator_private = cic;
3521 rq->elevator_private2 = cfqq;
3522 return 0;
1da177e4 3523
22e2c507
JA
3524queue_fail:
3525 if (cic)
3526 put_io_context(cic->ioc);
89850f7e 3527
23e018a1 3528 cfq_schedule_dispatch(cfqd);
1da177e4 3529 spin_unlock_irqrestore(q->queue_lock, flags);
7b679138 3530 cfq_log(cfqd, "set_request fail");
1da177e4
LT
3531 return 1;
3532}
3533
65f27f38 3534static void cfq_kick_queue(struct work_struct *work)
22e2c507 3535{
65f27f38 3536 struct cfq_data *cfqd =
23e018a1 3537 container_of(work, struct cfq_data, unplug_work);
165125e1 3538 struct request_queue *q = cfqd->queue;
22e2c507 3539
40bb54d1 3540 spin_lock_irq(q->queue_lock);
a7f55792 3541 __blk_run_queue(cfqd->queue);
40bb54d1 3542 spin_unlock_irq(q->queue_lock);
22e2c507
JA
3543}
3544
3545/*
3546 * Timer running if the active_queue is currently idling inside its time slice
3547 */
3548static void cfq_idle_slice_timer(unsigned long data)
3549{
3550 struct cfq_data *cfqd = (struct cfq_data *) data;
3551 struct cfq_queue *cfqq;
3552 unsigned long flags;
3c6bd2f8 3553 int timed_out = 1;
22e2c507 3554
7b679138
JA
3555 cfq_log(cfqd, "idle timer fired");
3556
22e2c507
JA
3557 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3558
fe094d98
JA
3559 cfqq = cfqd->active_queue;
3560 if (cfqq) {
3c6bd2f8
JA
3561 timed_out = 0;
3562
b029195d
JA
3563 /*
3564 * We saw a request before the queue expired, let it through
3565 */
3566 if (cfq_cfqq_must_dispatch(cfqq))
3567 goto out_kick;
3568
22e2c507
JA
3569 /*
3570 * expired
3571 */
44f7c160 3572 if (cfq_slice_used(cfqq))
22e2c507
JA
3573 goto expire;
3574
3575 /*
3576 * only expire and reinvoke request handler, if there are
3577 * other queues with pending requests
3578 */
caaa5f9f 3579 if (!cfqd->busy_queues)
22e2c507 3580 goto out_cont;
22e2c507
JA
3581
3582 /*
3583 * not expired and it has a request pending, let it dispatch
3584 */
75e50984 3585 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
22e2c507 3586 goto out_kick;
76280aff
CZ
3587
3588 /*
3589 * Queue depth flag is reset only when the idle didn't succeed
3590 */
3591 cfq_clear_cfqq_deep(cfqq);
22e2c507
JA
3592 }
3593expire:
6084cdda 3594 cfq_slice_expired(cfqd, timed_out);
22e2c507 3595out_kick:
23e018a1 3596 cfq_schedule_dispatch(cfqd);
22e2c507
JA
3597out_cont:
3598 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3599}
3600
3b18152c
JA
3601static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3602{
3603 del_timer_sync(&cfqd->idle_slice_timer);
23e018a1 3604 cancel_work_sync(&cfqd->unplug_work);
3b18152c 3605}
22e2c507 3606
c2dea2d1
VT
3607static void cfq_put_async_queues(struct cfq_data *cfqd)
3608{
3609 int i;
3610
3611 for (i = 0; i < IOPRIO_BE_NR; i++) {
3612 if (cfqd->async_cfqq[0][i])
3613 cfq_put_queue(cfqd->async_cfqq[0][i]);
3614 if (cfqd->async_cfqq[1][i])
3615 cfq_put_queue(cfqd->async_cfqq[1][i]);
c2dea2d1 3616 }
2389d1ef
ON
3617
3618 if (cfqd->async_idle_cfqq)
3619 cfq_put_queue(cfqd->async_idle_cfqq);
c2dea2d1
VT
3620}
3621
bb729bc9
JA
3622static void cfq_cfqd_free(struct rcu_head *head)
3623{
3624 kfree(container_of(head, struct cfq_data, rcu));
3625}
3626
b374d18a 3627static void cfq_exit_queue(struct elevator_queue *e)
1da177e4 3628{
22e2c507 3629 struct cfq_data *cfqd = e->elevator_data;
165125e1 3630 struct request_queue *q = cfqd->queue;
22e2c507 3631
3b18152c 3632 cfq_shutdown_timer_wq(cfqd);
e2d74ac0 3633
d9ff4187 3634 spin_lock_irq(q->queue_lock);
e2d74ac0 3635
d9ff4187 3636 if (cfqd->active_queue)
6084cdda 3637 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
e2d74ac0
JA
3638
3639 while (!list_empty(&cfqd->cic_list)) {
d9ff4187
AV
3640 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3641 struct cfq_io_context,
3642 queue_list);
89850f7e
JA
3643
3644 __cfq_exit_single_io_context(cfqd, cic);
d9ff4187 3645 }
e2d74ac0 3646
c2dea2d1 3647 cfq_put_async_queues(cfqd);
b1c35769
VG
3648 cfq_release_cfq_groups(cfqd);
3649 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
15c31be4 3650
d9ff4187 3651 spin_unlock_irq(q->queue_lock);
a90d742e
AV
3652
3653 cfq_shutdown_timer_wq(cfqd);
3654
b1c35769 3655 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
bb729bc9 3656 call_rcu(&cfqd->rcu, cfq_cfqd_free);
1da177e4
LT
3657}
3658
165125e1 3659static void *cfq_init_queue(struct request_queue *q)
1da177e4
LT
3660{
3661 struct cfq_data *cfqd;
718eee05 3662 int i, j;
cdb16e8f 3663 struct cfq_group *cfqg;
615f0259 3664 struct cfq_rb_root *st;
1da177e4 3665
94f6030c 3666 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
1da177e4 3667 if (!cfqd)
bc1c1169 3668 return NULL;
1da177e4 3669
1fa8f6d6
VG
3670 /* Init root service tree */
3671 cfqd->grp_service_tree = CFQ_RB_ROOT;
3672
cdb16e8f
VG
3673 /* Init root group */
3674 cfqg = &cfqd->root_group;
615f0259
VG
3675 for_each_cfqg_st(cfqg, i, j, st)
3676 *st = CFQ_RB_ROOT;
1fa8f6d6 3677 RB_CLEAR_NODE(&cfqg->rb_node);
26a2ac00 3678
25bc6b07
VG
3679 /* Give preference to root group over other groups */
3680 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3681
25fb5169 3682#ifdef CONFIG_CFQ_GROUP_IOSCHED
b1c35769
VG
3683 /*
3684 * Take a reference to root group which we never drop. This is just
3685 * to make sure that cfq_put_cfqg() does not try to kfree root group
3686 */
3687 atomic_set(&cfqg->ref, 1);
22084190
VG
3688 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3689 0);
25fb5169 3690#endif
26a2ac00
JA
3691 /*
3692 * Not strictly needed (since RB_ROOT just clears the node and we
3693 * zeroed cfqd on alloc), but better be safe in case someone decides
3694 * to add magic to the rb code
3695 */
3696 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3697 cfqd->prio_trees[i] = RB_ROOT;
3698
6118b70b
JA
3699 /*
3700 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3701 * Grab a permanent reference to it, so that the normal code flow
3702 * will not attempt to free it.
3703 */
3704 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3705 atomic_inc(&cfqd->oom_cfqq.ref);
cdb16e8f 3706 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
6118b70b 3707
d9ff4187 3708 INIT_LIST_HEAD(&cfqd->cic_list);
1da177e4 3709
1da177e4 3710 cfqd->queue = q;
1da177e4 3711
22e2c507
JA
3712 init_timer(&cfqd->idle_slice_timer);
3713 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3714 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3715
23e018a1 3716 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
22e2c507 3717
1da177e4 3718 cfqd->cfq_quantum = cfq_quantum;
22e2c507
JA
3719 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3720 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
1da177e4
LT
3721 cfqd->cfq_back_max = cfq_back_max;
3722 cfqd->cfq_back_penalty = cfq_back_penalty;
22e2c507
JA
3723 cfqd->cfq_slice[0] = cfq_slice_async;
3724 cfqd->cfq_slice[1] = cfq_slice_sync;
3725 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3726 cfqd->cfq_slice_idle = cfq_slice_idle;
963b72fc 3727 cfqd->cfq_latency = 1;
ae30c286 3728 cfqd->cfq_group_isolation = 0;
e459dd08 3729 cfqd->hw_tag = -1;
edc71131
CZ
3730 /*
3731 * we optimistically start assuming sync ops weren't delayed in last
3732 * second, in order to have larger depth for async operations.
3733 */
573412b2 3734 cfqd->last_delayed_sync = jiffies - HZ;
bb729bc9 3735 INIT_RCU_HEAD(&cfqd->rcu);
bc1c1169 3736 return cfqd;
1da177e4
LT
3737}
3738
3739static void cfq_slab_kill(void)
3740{
d6de8be7
JA
3741 /*
3742 * Caller already ensured that pending RCU callbacks are completed,
3743 * so we should have no busy allocations at this point.
3744 */
1da177e4
LT
3745 if (cfq_pool)
3746 kmem_cache_destroy(cfq_pool);
3747 if (cfq_ioc_pool)
3748 kmem_cache_destroy(cfq_ioc_pool);
3749}
3750
3751static int __init cfq_slab_setup(void)
3752{
0a31bd5f 3753 cfq_pool = KMEM_CACHE(cfq_queue, 0);
1da177e4
LT
3754 if (!cfq_pool)
3755 goto fail;
3756
34e6bbf2 3757 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
1da177e4
LT
3758 if (!cfq_ioc_pool)
3759 goto fail;
3760
3761 return 0;
3762fail:
3763 cfq_slab_kill();
3764 return -ENOMEM;
3765}
3766
1da177e4
LT
3767/*
3768 * sysfs parts below -->
3769 */
1da177e4
LT
3770static ssize_t
3771cfq_var_show(unsigned int var, char *page)
3772{
3773 return sprintf(page, "%d\n", var);
3774}
3775
3776static ssize_t
3777cfq_var_store(unsigned int *var, const char *page, size_t count)
3778{
3779 char *p = (char *) page;
3780
3781 *var = simple_strtoul(p, &p, 10);
3782 return count;
3783}
3784
1da177e4 3785#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
b374d18a 3786static ssize_t __FUNC(struct elevator_queue *e, char *page) \
1da177e4 3787{ \
3d1ab40f 3788 struct cfq_data *cfqd = e->elevator_data; \
1da177e4
LT
3789 unsigned int __data = __VAR; \
3790 if (__CONV) \
3791 __data = jiffies_to_msecs(__data); \
3792 return cfq_var_show(__data, (page)); \
3793}
3794SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
22e2c507
JA
3795SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3796SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
e572ec7e
AV
3797SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3798SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
22e2c507
JA
3799SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3800SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3801SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3802SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
963b72fc 3803SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
ae30c286 3804SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
1da177e4
LT
3805#undef SHOW_FUNCTION
3806
3807#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
b374d18a 3808static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
1da177e4 3809{ \
3d1ab40f 3810 struct cfq_data *cfqd = e->elevator_data; \
1da177e4
LT
3811 unsigned int __data; \
3812 int ret = cfq_var_store(&__data, (page), count); \
3813 if (__data < (MIN)) \
3814 __data = (MIN); \
3815 else if (__data > (MAX)) \
3816 __data = (MAX); \
3817 if (__CONV) \
3818 *(__PTR) = msecs_to_jiffies(__data); \
3819 else \
3820 *(__PTR) = __data; \
3821 return ret; \
3822}
3823STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
fe094d98
JA
3824STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3825 UINT_MAX, 1);
3826STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3827 UINT_MAX, 1);
e572ec7e 3828STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
fe094d98
JA
3829STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3830 UINT_MAX, 0);
22e2c507
JA
3831STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3832STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3833STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
fe094d98
JA
3834STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3835 UINT_MAX, 0);
963b72fc 3836STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
ae30c286 3837STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
1da177e4
LT
3838#undef STORE_FUNCTION
3839
e572ec7e
AV
3840#define CFQ_ATTR(name) \
3841 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3842
3843static struct elv_fs_entry cfq_attrs[] = {
3844 CFQ_ATTR(quantum),
e572ec7e
AV
3845 CFQ_ATTR(fifo_expire_sync),
3846 CFQ_ATTR(fifo_expire_async),
3847 CFQ_ATTR(back_seek_max),
3848 CFQ_ATTR(back_seek_penalty),
3849 CFQ_ATTR(slice_sync),
3850 CFQ_ATTR(slice_async),
3851 CFQ_ATTR(slice_async_rq),
3852 CFQ_ATTR(slice_idle),
963b72fc 3853 CFQ_ATTR(low_latency),
ae30c286 3854 CFQ_ATTR(group_isolation),
e572ec7e 3855 __ATTR_NULL
1da177e4
LT
3856};
3857
1da177e4
LT
3858static struct elevator_type iosched_cfq = {
3859 .ops = {
3860 .elevator_merge_fn = cfq_merge,
3861 .elevator_merged_fn = cfq_merged_request,
3862 .elevator_merge_req_fn = cfq_merged_requests,
da775265 3863 .elevator_allow_merge_fn = cfq_allow_merge,
b4878f24 3864 .elevator_dispatch_fn = cfq_dispatch_requests,
1da177e4 3865 .elevator_add_req_fn = cfq_insert_request,
b4878f24 3866 .elevator_activate_req_fn = cfq_activate_request,
1da177e4
LT
3867 .elevator_deactivate_req_fn = cfq_deactivate_request,
3868 .elevator_queue_empty_fn = cfq_queue_empty,
3869 .elevator_completed_req_fn = cfq_completed_request,
21183b07
JA
3870 .elevator_former_req_fn = elv_rb_former_request,
3871 .elevator_latter_req_fn = elv_rb_latter_request,
1da177e4
LT
3872 .elevator_set_req_fn = cfq_set_request,
3873 .elevator_put_req_fn = cfq_put_request,
3874 .elevator_may_queue_fn = cfq_may_queue,
3875 .elevator_init_fn = cfq_init_queue,
3876 .elevator_exit_fn = cfq_exit_queue,
fc46379d 3877 .trim = cfq_free_io_context,
1da177e4 3878 },
3d1ab40f 3879 .elevator_attrs = cfq_attrs,
1da177e4
LT
3880 .elevator_name = "cfq",
3881 .elevator_owner = THIS_MODULE,
3882};
3883
3e252066
VG
3884#ifdef CONFIG_CFQ_GROUP_IOSCHED
3885static struct blkio_policy_type blkio_policy_cfq = {
3886 .ops = {
3887 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3888 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3889 },
3890};
3891#else
3892static struct blkio_policy_type blkio_policy_cfq;
3893#endif
3894
1da177e4
LT
3895static int __init cfq_init(void)
3896{
22e2c507
JA
3897 /*
3898 * could be 0 on HZ < 1000 setups
3899 */
3900 if (!cfq_slice_async)
3901 cfq_slice_async = 1;
3902 if (!cfq_slice_idle)
3903 cfq_slice_idle = 1;
3904
1da177e4
LT
3905 if (cfq_slab_setup())
3906 return -ENOMEM;
3907
2fdd82bd 3908 elv_register(&iosched_cfq);
3e252066 3909 blkio_policy_register(&blkio_policy_cfq);
1da177e4 3910
2fdd82bd 3911 return 0;
1da177e4
LT
3912}
3913
3914static void __exit cfq_exit(void)
3915{
6e9a4738 3916 DECLARE_COMPLETION_ONSTACK(all_gone);
3e252066 3917 blkio_policy_unregister(&blkio_policy_cfq);
1da177e4 3918 elv_unregister(&iosched_cfq);
334e94de 3919 ioc_gone = &all_gone;
fba82272
OH
3920 /* ioc_gone's update must be visible before reading ioc_count */
3921 smp_wmb();
d6de8be7
JA
3922
3923 /*
3924 * this also protects us from entering cfq_slab_kill() with
3925 * pending RCU callbacks
3926 */
245b2e70 3927 if (elv_ioc_count_read(cfq_ioc_count))
9a11b4ed 3928 wait_for_completion(&all_gone);
83521d3e 3929 cfq_slab_kill();
1da177e4
LT
3930}
3931
3932module_init(cfq_init);
3933module_exit(cfq_exit);
3934
3935MODULE_AUTHOR("Jens Axboe");
3936MODULE_LICENSE("GPL");
3937MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");