Merge tag 'metag-for-v4.9-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/jhoga...
[linux-2.6-block.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>
5a0e3ad6 10#include <linux/slab.h>
1cc9be68
AV
11#include <linux/blkdev.h>
12#include <linux/elevator.h>
9a7f38c4 13#include <linux/ktime.h>
1da177e4 14#include <linux/rbtree.h>
22e2c507 15#include <linux/ioprio.h>
7b679138 16#include <linux/blktrace_api.h>
eea8f41c 17#include <linux/blk-cgroup.h>
6e736be7 18#include "blk.h"
1da177e4
LT
19
20/*
21 * tunables
22 */
fe094d98 23/* max queue in one round of service */
abc3c744 24static const int cfq_quantum = 8;
9a7f38c4 25static const u64 cfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
fe094d98
JA
26/* maximum backwards seek, in KiB */
27static const int cfq_back_max = 16 * 1024;
28/* penalty of a backwards seek */
29static const int cfq_back_penalty = 2;
9a7f38c4
JM
30static const u64 cfq_slice_sync = NSEC_PER_SEC / 10;
31static u64 cfq_slice_async = NSEC_PER_SEC / 25;
64100099 32static const int cfq_slice_async_rq = 2;
9a7f38c4
JM
33static u64 cfq_slice_idle = NSEC_PER_SEC / 125;
34static u64 cfq_group_idle = NSEC_PER_SEC / 125;
35static const u64 cfq_target_latency = (u64)NSEC_PER_SEC * 3/10; /* 300 ms */
5db5d642 36static const int cfq_hist_divisor = 4;
22e2c507 37
d9e7620e 38/*
0871714e 39 * offset from end of service tree
d9e7620e 40 */
9a7f38c4 41#define CFQ_IDLE_DELAY (NSEC_PER_SEC / 5)
d9e7620e
JA
42
43/*
44 * below this threshold, we consider thinktime immediate
45 */
9a7f38c4 46#define CFQ_MIN_TT (2 * NSEC_PER_SEC / HZ)
d9e7620e 47
22e2c507 48#define CFQ_SLICE_SCALE (5)
45333d5a 49#define CFQ_HW_QUEUE_MIN (5)
25bc6b07 50#define CFQ_SERVICE_SHIFT 12
22e2c507 51
3dde36dd 52#define CFQQ_SEEK_THR (sector_t)(8 * 100)
e9ce335d 53#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
41647e7a 54#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
3dde36dd 55#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
ae54abed 56
a612fddf
TH
57#define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
1da177e4 60
e18b890b 61static struct kmem_cache *cfq_pool;
1da177e4 62
22e2c507
JA
63#define CFQ_PRIO_LISTS IOPRIO_BE_NR
64#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
22e2c507
JA
65#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
66
206dc69b 67#define sample_valid(samples) ((samples) > 80)
1fa8f6d6 68#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
206dc69b 69
e48453c3 70/* blkio-related constants */
3ecca629
TH
71#define CFQ_WEIGHT_LEGACY_MIN 10
72#define CFQ_WEIGHT_LEGACY_DFL 500
73#define CFQ_WEIGHT_LEGACY_MAX 1000
e48453c3 74
c5869807 75struct cfq_ttime {
9a7f38c4 76 u64 last_end_request;
c5869807 77
9a7f38c4
JM
78 u64 ttime_total;
79 u64 ttime_mean;
c5869807 80 unsigned long ttime_samples;
c5869807
TH
81};
82
cc09e299
JA
83/*
84 * Most of our rbtree usage is for sorting with min extraction, so
85 * if we cache the leftmost node we don't have to walk down the tree
86 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
87 * move this into the elevator for the rq sorting as well.
88 */
89struct cfq_rb_root {
90 struct rb_root rb;
91 struct rb_node *left;
aa6f6a3d 92 unsigned count;
1fa8f6d6 93 u64 min_vdisktime;
f5f2b6ce 94 struct cfq_ttime ttime;
cc09e299 95};
f5f2b6ce 96#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
9a7f38c4 97 .ttime = {.last_end_request = ktime_get_ns(),},}
cc09e299 98
6118b70b
JA
99/*
100 * Per process-grouping structure
101 */
102struct cfq_queue {
103 /* reference count */
30d7b944 104 int ref;
6118b70b
JA
105 /* various state flags, see below */
106 unsigned int flags;
107 /* parent cfq_data */
108 struct cfq_data *cfqd;
109 /* service_tree member */
110 struct rb_node rb_node;
111 /* service_tree key */
9a7f38c4 112 u64 rb_key;
6118b70b
JA
113 /* prio tree member */
114 struct rb_node p_node;
115 /* prio tree root we belong to, if any */
116 struct rb_root *p_root;
117 /* sorted list of pending requests */
118 struct rb_root sort_list;
119 /* if fifo isn't expired, next request to serve */
120 struct request *next_rq;
121 /* requests queued in sort_list */
122 int queued[2];
123 /* currently allocated requests */
124 int allocated[2];
125 /* fifo list of requests in sort_list */
126 struct list_head fifo;
127
dae739eb 128 /* time when queue got scheduled in to dispatch first request. */
9a7f38c4
JM
129 u64 dispatch_start;
130 u64 allocated_slice;
131 u64 slice_dispatch;
dae739eb 132 /* time when first request from queue completed and slice started. */
9a7f38c4
JM
133 u64 slice_start;
134 u64 slice_end;
93fdf147 135 s64 slice_resid;
6118b70b 136
65299a3b
CH
137 /* pending priority requests */
138 int prio_pending;
6118b70b
JA
139 /* number of requests that are on the dispatch list or inside driver */
140 int dispatched;
141
142 /* io prio of this group */
143 unsigned short ioprio, org_ioprio;
b8269db4 144 unsigned short ioprio_class, org_ioprio_class;
6118b70b 145
c4081ba5
RK
146 pid_t pid;
147
3dde36dd 148 u32 seek_history;
b2c18e1e
JM
149 sector_t last_request_pos;
150
aa6f6a3d 151 struct cfq_rb_root *service_tree;
df5fe3e8 152 struct cfq_queue *new_cfqq;
cdb16e8f 153 struct cfq_group *cfqg;
c4e7893e
VG
154 /* Number of sectors dispatched from queue in single dispatch round */
155 unsigned long nr_sectors;
6118b70b
JA
156};
157
c0324a02 158/*
718eee05 159 * First index in the service_trees.
c0324a02
CZ
160 * IDLE is handled separately, so it has negative index
161 */
3bf10fea 162enum wl_class_t {
c0324a02 163 BE_WORKLOAD = 0,
615f0259
VG
164 RT_WORKLOAD = 1,
165 IDLE_WORKLOAD = 2,
b4627321 166 CFQ_PRIO_NR,
c0324a02
CZ
167};
168
718eee05
CZ
169/*
170 * Second index in the service_trees.
171 */
172enum wl_type_t {
173 ASYNC_WORKLOAD = 0,
174 SYNC_NOIDLE_WORKLOAD = 1,
175 SYNC_WORKLOAD = 2
176};
177
155fead9
TH
178struct cfqg_stats {
179#ifdef CONFIG_CFQ_GROUP_IOSCHED
155fead9
TH
180 /* number of ios merged */
181 struct blkg_rwstat merged;
182 /* total time spent on device in ns, may not be accurate w/ queueing */
183 struct blkg_rwstat service_time;
184 /* total time spent waiting in scheduler queue in ns */
185 struct blkg_rwstat wait_time;
186 /* number of IOs queued up */
187 struct blkg_rwstat queued;
155fead9
TH
188 /* total disk time and nr sectors dispatched by this group */
189 struct blkg_stat time;
190#ifdef CONFIG_DEBUG_BLK_CGROUP
191 /* time not charged to this cgroup */
192 struct blkg_stat unaccounted_time;
193 /* sum of number of ios queued across all samples */
194 struct blkg_stat avg_queue_size_sum;
195 /* count of samples taken for average */
196 struct blkg_stat avg_queue_size_samples;
197 /* how many times this group has been removed from service tree */
198 struct blkg_stat dequeue;
199 /* total time spent waiting for it to be assigned a timeslice. */
200 struct blkg_stat group_wait_time;
3c798398 201 /* time spent idling for this blkcg_gq */
155fead9
TH
202 struct blkg_stat idle_time;
203 /* total time with empty current active q with other requests queued */
204 struct blkg_stat empty_time;
205 /* fields after this shouldn't be cleared on stat reset */
206 uint64_t start_group_wait_time;
207 uint64_t start_idle_time;
208 uint64_t start_empty_time;
209 uint16_t flags;
210#endif /* CONFIG_DEBUG_BLK_CGROUP */
211#endif /* CONFIG_CFQ_GROUP_IOSCHED */
212};
213
e48453c3
AA
214/* Per-cgroup data */
215struct cfq_group_data {
216 /* must be the first member */
81437648 217 struct blkcg_policy_data cpd;
e48453c3
AA
218
219 unsigned int weight;
220 unsigned int leaf_weight;
221};
222
cdb16e8f
VG
223/* This is per cgroup per device grouping structure */
224struct cfq_group {
f95a04af
TH
225 /* must be the first member */
226 struct blkg_policy_data pd;
227
1fa8f6d6
VG
228 /* group service_tree member */
229 struct rb_node rb_node;
230
231 /* group service_tree key */
232 u64 vdisktime;
e71357e1 233
7918ffb5
TH
234 /*
235 * The number of active cfqgs and sum of their weights under this
236 * cfqg. This covers this cfqg's leaf_weight and all children's
237 * weights, but does not cover weights of further descendants.
238 *
239 * If a cfqg is on the service tree, it's active. An active cfqg
240 * also activates its parent and contributes to the children_weight
241 * of the parent.
242 */
243 int nr_active;
244 unsigned int children_weight;
245
1d3650f7
TH
246 /*
247 * vfraction is the fraction of vdisktime that the tasks in this
248 * cfqg are entitled to. This is determined by compounding the
249 * ratios walking up from this cfqg to the root.
250 *
251 * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
252 * vfractions on a service tree is approximately 1. The sum may
253 * deviate a bit due to rounding errors and fluctuations caused by
254 * cfqgs entering and leaving the service tree.
255 */
256 unsigned int vfraction;
257
e71357e1
TH
258 /*
259 * There are two weights - (internal) weight is the weight of this
260 * cfqg against the sibling cfqgs. leaf_weight is the wight of
261 * this cfqg against the child cfqgs. For the root cfqg, both
262 * weights are kept in sync for backward compatibility.
263 */
25bc6b07 264 unsigned int weight;
8184f93e 265 unsigned int new_weight;
3381cb8d 266 unsigned int dev_weight;
1fa8f6d6 267
e71357e1
TH
268 unsigned int leaf_weight;
269 unsigned int new_leaf_weight;
270 unsigned int dev_leaf_weight;
271
1fa8f6d6
VG
272 /* number of cfqq currently on this group */
273 int nr_cfqq;
274
cdb16e8f 275 /*
4495a7d4 276 * Per group busy queues average. Useful for workload slice calc. We
b4627321
VG
277 * create the array for each prio class but at run time it is used
278 * only for RT and BE class and slot for IDLE class remains unused.
279 * This is primarily done to avoid confusion and a gcc warning.
280 */
281 unsigned int busy_queues_avg[CFQ_PRIO_NR];
282 /*
283 * rr lists of queues with requests. We maintain service trees for
284 * RT and BE classes. These trees are subdivided in subclasses
285 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
286 * class there is no subclassification and all the cfq queues go on
287 * a single tree service_tree_idle.
cdb16e8f
VG
288 * Counts are embedded in the cfq_rb_root
289 */
290 struct cfq_rb_root service_trees[2][3];
291 struct cfq_rb_root service_tree_idle;
dae739eb 292
9a7f38c4 293 u64 saved_wl_slice;
4d2ceea4
VG
294 enum wl_type_t saved_wl_type;
295 enum wl_class_t saved_wl_class;
4eef3049 296
80bdf0c7
VG
297 /* number of requests that are on the dispatch list or inside driver */
298 int dispatched;
7700fc4f 299 struct cfq_ttime ttime;
0b39920b 300 struct cfqg_stats stats; /* stats for this cfqg */
60a83707
TH
301
302 /* async queue for each priority case */
303 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
304 struct cfq_queue *async_idle_cfqq;
305
cdb16e8f 306};
718eee05 307
c5869807
TH
308struct cfq_io_cq {
309 struct io_cq icq; /* must be the first member */
310 struct cfq_queue *cfqq[2];
311 struct cfq_ttime ttime;
598971bf
TH
312 int ioprio; /* the current ioprio */
313#ifdef CONFIG_CFQ_GROUP_IOSCHED
f4da8072 314 uint64_t blkcg_serial_nr; /* the current blkcg serial */
598971bf 315#endif
c5869807
TH
316};
317
22e2c507
JA
318/*
319 * Per block device queue structure
320 */
1da177e4 321struct cfq_data {
165125e1 322 struct request_queue *queue;
1fa8f6d6
VG
323 /* Root service tree for cfq_groups */
324 struct cfq_rb_root grp_service_tree;
f51b802c 325 struct cfq_group *root_group;
22e2c507 326
c0324a02
CZ
327 /*
328 * The priority currently being served
22e2c507 329 */
4d2ceea4
VG
330 enum wl_class_t serving_wl_class;
331 enum wl_type_t serving_wl_type;
9a7f38c4 332 u64 workload_expires;
cdb16e8f 333 struct cfq_group *serving_group;
a36e71f9
JA
334
335 /*
336 * Each priority tree is sorted by next_request position. These
337 * trees are used when determining if two or more queues are
338 * interleaving requests (see cfq_close_cooperator).
339 */
340 struct rb_root prio_trees[CFQ_PRIO_LISTS];
341
22e2c507 342 unsigned int busy_queues;
ef8a41df 343 unsigned int busy_sync_queues;
22e2c507 344
53c583d2
CZ
345 int rq_in_driver;
346 int rq_in_flight[2];
45333d5a
AC
347
348 /*
349 * queue-depth detection
350 */
351 int rq_queued;
25776e35 352 int hw_tag;
e459dd08
CZ
353 /*
354 * hw_tag can be
355 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
356 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
357 * 0 => no NCQ
358 */
359 int hw_tag_est_depth;
360 unsigned int hw_tag_samples;
1da177e4 361
22e2c507
JA
362 /*
363 * idle window management
364 */
91148325 365 struct hrtimer idle_slice_timer;
23e018a1 366 struct work_struct unplug_work;
1da177e4 367
22e2c507 368 struct cfq_queue *active_queue;
c5869807 369 struct cfq_io_cq *active_cic;
22e2c507 370
6d048f53 371 sector_t last_position;
1da177e4 372
1da177e4
LT
373 /*
374 * tunables, see top of file
375 */
376 unsigned int cfq_quantum;
1da177e4
LT
377 unsigned int cfq_back_penalty;
378 unsigned int cfq_back_max;
22e2c507 379 unsigned int cfq_slice_async_rq;
963b72fc 380 unsigned int cfq_latency;
9a7f38c4
JM
381 u64 cfq_fifo_expire[2];
382 u64 cfq_slice[2];
383 u64 cfq_slice_idle;
384 u64 cfq_group_idle;
385 u64 cfq_target_latency;
d9ff4187 386
6118b70b
JA
387 /*
388 * Fallback dummy cfqq for extreme OOM conditions
389 */
390 struct cfq_queue oom_cfqq;
365722bb 391
9a7f38c4 392 u64 last_delayed_sync;
1da177e4
LT
393};
394
25fb5169 395static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
60a83707 396static void cfq_put_queue(struct cfq_queue *cfqq);
25fb5169 397
34b98d03 398static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
3bf10fea 399 enum wl_class_t class,
65b32a57 400 enum wl_type_t type)
c0324a02 401{
1fa8f6d6
VG
402 if (!cfqg)
403 return NULL;
404
3bf10fea 405 if (class == IDLE_WORKLOAD)
cdb16e8f 406 return &cfqg->service_tree_idle;
c0324a02 407
3bf10fea 408 return &cfqg->service_trees[class][type];
c0324a02
CZ
409}
410
3b18152c 411enum cfqq_state_flags {
b0b8d749
JA
412 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
413 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
b029195d 414 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
b0b8d749 415 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
b0b8d749
JA
416 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
417 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
418 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
44f7c160 419 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
91fac317 420 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
b3b6d040 421 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
ae54abed 422 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
76280aff 423 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
f75edf2d 424 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
3b18152c
JA
425};
426
427#define CFQ_CFQQ_FNS(name) \
428static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
429{ \
fe094d98 430 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
3b18152c
JA
431} \
432static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
433{ \
fe094d98 434 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
3b18152c
JA
435} \
436static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
437{ \
fe094d98 438 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
3b18152c
JA
439}
440
441CFQ_CFQQ_FNS(on_rr);
442CFQ_CFQQ_FNS(wait_request);
b029195d 443CFQ_CFQQ_FNS(must_dispatch);
3b18152c 444CFQ_CFQQ_FNS(must_alloc_slice);
3b18152c
JA
445CFQ_CFQQ_FNS(fifo_expire);
446CFQ_CFQQ_FNS(idle_window);
447CFQ_CFQQ_FNS(prio_changed);
44f7c160 448CFQ_CFQQ_FNS(slice_new);
91fac317 449CFQ_CFQQ_FNS(sync);
a36e71f9 450CFQ_CFQQ_FNS(coop);
ae54abed 451CFQ_CFQQ_FNS(split_coop);
76280aff 452CFQ_CFQQ_FNS(deep);
f75edf2d 453CFQ_CFQQ_FNS(wait_busy);
3b18152c
JA
454#undef CFQ_CFQQ_FNS
455
629ed0b1 456#if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
2ce4d50f 457
155fead9
TH
458/* cfqg stats flags */
459enum cfqg_stats_flags {
460 CFQG_stats_waiting = 0,
461 CFQG_stats_idling,
462 CFQG_stats_empty,
629ed0b1
TH
463};
464
155fead9
TH
465#define CFQG_FLAG_FNS(name) \
466static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
629ed0b1 467{ \
155fead9 468 stats->flags |= (1 << CFQG_stats_##name); \
629ed0b1 469} \
155fead9 470static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
629ed0b1 471{ \
155fead9 472 stats->flags &= ~(1 << CFQG_stats_##name); \
629ed0b1 473} \
155fead9 474static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
629ed0b1 475{ \
155fead9 476 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
629ed0b1
TH
477} \
478
155fead9
TH
479CFQG_FLAG_FNS(waiting)
480CFQG_FLAG_FNS(idling)
481CFQG_FLAG_FNS(empty)
482#undef CFQG_FLAG_FNS
629ed0b1
TH
483
484/* This should be called with the queue_lock held. */
155fead9 485static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
629ed0b1
TH
486{
487 unsigned long long now;
488
155fead9 489 if (!cfqg_stats_waiting(stats))
629ed0b1
TH
490 return;
491
492 now = sched_clock();
493 if (time_after64(now, stats->start_group_wait_time))
494 blkg_stat_add(&stats->group_wait_time,
495 now - stats->start_group_wait_time);
155fead9 496 cfqg_stats_clear_waiting(stats);
629ed0b1
TH
497}
498
499/* This should be called with the queue_lock held. */
155fead9
TH
500static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
501 struct cfq_group *curr_cfqg)
629ed0b1 502{
155fead9 503 struct cfqg_stats *stats = &cfqg->stats;
629ed0b1 504
155fead9 505 if (cfqg_stats_waiting(stats))
629ed0b1 506 return;
155fead9 507 if (cfqg == curr_cfqg)
629ed0b1 508 return;
155fead9
TH
509 stats->start_group_wait_time = sched_clock();
510 cfqg_stats_mark_waiting(stats);
629ed0b1
TH
511}
512
513/* This should be called with the queue_lock held. */
155fead9 514static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
629ed0b1
TH
515{
516 unsigned long long now;
517
155fead9 518 if (!cfqg_stats_empty(stats))
629ed0b1
TH
519 return;
520
521 now = sched_clock();
522 if (time_after64(now, stats->start_empty_time))
523 blkg_stat_add(&stats->empty_time,
524 now - stats->start_empty_time);
155fead9 525 cfqg_stats_clear_empty(stats);
629ed0b1
TH
526}
527
155fead9 528static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
629ed0b1 529{
155fead9 530 blkg_stat_add(&cfqg->stats.dequeue, 1);
629ed0b1
TH
531}
532
155fead9 533static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
629ed0b1 534{
155fead9 535 struct cfqg_stats *stats = &cfqg->stats;
629ed0b1 536
4d5e80a7 537 if (blkg_rwstat_total(&stats->queued))
629ed0b1
TH
538 return;
539
540 /*
541 * group is already marked empty. This can happen if cfqq got new
542 * request in parent group and moved to this group while being added
543 * to service tree. Just ignore the event and move on.
544 */
155fead9 545 if (cfqg_stats_empty(stats))
629ed0b1
TH
546 return;
547
548 stats->start_empty_time = sched_clock();
155fead9 549 cfqg_stats_mark_empty(stats);
629ed0b1
TH
550}
551
155fead9 552static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
629ed0b1 553{
155fead9 554 struct cfqg_stats *stats = &cfqg->stats;
629ed0b1 555
155fead9 556 if (cfqg_stats_idling(stats)) {
629ed0b1
TH
557 unsigned long long now = sched_clock();
558
559 if (time_after64(now, stats->start_idle_time))
560 blkg_stat_add(&stats->idle_time,
561 now - stats->start_idle_time);
155fead9 562 cfqg_stats_clear_idling(stats);
629ed0b1
TH
563 }
564}
565
155fead9 566static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
629ed0b1 567{
155fead9 568 struct cfqg_stats *stats = &cfqg->stats;
629ed0b1 569
155fead9 570 BUG_ON(cfqg_stats_idling(stats));
629ed0b1
TH
571
572 stats->start_idle_time = sched_clock();
155fead9 573 cfqg_stats_mark_idling(stats);
629ed0b1
TH
574}
575
155fead9 576static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
629ed0b1 577{
155fead9 578 struct cfqg_stats *stats = &cfqg->stats;
629ed0b1
TH
579
580 blkg_stat_add(&stats->avg_queue_size_sum,
4d5e80a7 581 blkg_rwstat_total(&stats->queued));
629ed0b1 582 blkg_stat_add(&stats->avg_queue_size_samples, 1);
155fead9 583 cfqg_stats_update_group_wait_time(stats);
629ed0b1
TH
584}
585
586#else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
587
f48ec1d7
TH
588static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
589static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
590static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
591static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
592static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
593static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
594static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
629ed0b1
TH
595
596#endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
597
598#ifdef CONFIG_CFQ_GROUP_IOSCHED
2ce4d50f 599
4ceab71b
JA
600static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
601{
602 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
603}
604
605static struct cfq_group_data
606*cpd_to_cfqgd(struct blkcg_policy_data *cpd)
607{
81437648 608 return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
4ceab71b
JA
609}
610
611static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
612{
613 return pd_to_blkg(&cfqg->pd);
614}
615
ffea73fc
TH
616static struct blkcg_policy blkcg_policy_cfq;
617
618static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
619{
620 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
621}
622
e48453c3
AA
623static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
624{
625 return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
626}
627
d02f7aa8 628static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
7918ffb5 629{
d02f7aa8 630 struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
7918ffb5 631
d02f7aa8 632 return pblkg ? blkg_to_cfqg(pblkg) : NULL;
7918ffb5
TH
633}
634
3984aa55
JK
635static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
636 struct cfq_group *ancestor)
637{
638 return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup,
639 cfqg_to_blkg(ancestor)->blkcg->css.cgroup);
640}
641
eb7d8c07
TH
642static inline void cfqg_get(struct cfq_group *cfqg)
643{
644 return blkg_get(cfqg_to_blkg(cfqg));
645}
646
647static inline void cfqg_put(struct cfq_group *cfqg)
648{
649 return blkg_put(cfqg_to_blkg(cfqg));
650}
651
54e7ed12
TH
652#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
653 char __pbuf[128]; \
654 \
655 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
b226e5c4
VG
656 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
657 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
658 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
54e7ed12
TH
659 __pbuf, ##args); \
660} while (0)
661
662#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
663 char __pbuf[128]; \
664 \
665 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
666 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
667} while (0)
2868ef7b 668
155fead9 669static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
63a4cc24
MC
670 struct cfq_group *curr_cfqg, int op,
671 int op_flags)
2ce4d50f 672{
63a4cc24 673 blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, 1);
155fead9
TH
674 cfqg_stats_end_empty_time(&cfqg->stats);
675 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
2ce4d50f
TH
676}
677
155fead9 678static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
9a7f38c4 679 uint64_t time, unsigned long unaccounted_time)
2ce4d50f 680{
155fead9 681 blkg_stat_add(&cfqg->stats.time, time);
629ed0b1 682#ifdef CONFIG_DEBUG_BLK_CGROUP
155fead9 683 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
629ed0b1 684#endif
2ce4d50f
TH
685}
686
63a4cc24
MC
687static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op,
688 int op_flags)
2ce4d50f 689{
63a4cc24 690 blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, -1);
2ce4d50f
TH
691}
692
63a4cc24
MC
693static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op,
694 int op_flags)
2ce4d50f 695{
63a4cc24 696 blkg_rwstat_add(&cfqg->stats.merged, op, op_flags, 1);
2ce4d50f
TH
697}
698
155fead9 699static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
63a4cc24
MC
700 uint64_t start_time, uint64_t io_start_time, int op,
701 int op_flags)
2ce4d50f 702{
155fead9 703 struct cfqg_stats *stats = &cfqg->stats;
629ed0b1 704 unsigned long long now = sched_clock();
629ed0b1
TH
705
706 if (time_after64(now, io_start_time))
63a4cc24
MC
707 blkg_rwstat_add(&stats->service_time, op, op_flags,
708 now - io_start_time);
629ed0b1 709 if (time_after64(io_start_time, start_time))
63a4cc24 710 blkg_rwstat_add(&stats->wait_time, op, op_flags,
629ed0b1 711 io_start_time - start_time);
2ce4d50f
TH
712}
713
689665af
TH
714/* @stats = 0 */
715static void cfqg_stats_reset(struct cfqg_stats *stats)
155fead9 716{
155fead9 717 /* queued stats shouldn't be cleared */
155fead9
TH
718 blkg_rwstat_reset(&stats->merged);
719 blkg_rwstat_reset(&stats->service_time);
720 blkg_rwstat_reset(&stats->wait_time);
721 blkg_stat_reset(&stats->time);
722#ifdef CONFIG_DEBUG_BLK_CGROUP
723 blkg_stat_reset(&stats->unaccounted_time);
724 blkg_stat_reset(&stats->avg_queue_size_sum);
725 blkg_stat_reset(&stats->avg_queue_size_samples);
726 blkg_stat_reset(&stats->dequeue);
727 blkg_stat_reset(&stats->group_wait_time);
728 blkg_stat_reset(&stats->idle_time);
729 blkg_stat_reset(&stats->empty_time);
730#endif
731}
732
0b39920b 733/* @to += @from */
e6269c44 734static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
0b39920b
TH
735{
736 /* queued stats shouldn't be cleared */
e6269c44
TH
737 blkg_rwstat_add_aux(&to->merged, &from->merged);
738 blkg_rwstat_add_aux(&to->service_time, &from->service_time);
739 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
740 blkg_stat_add_aux(&from->time, &from->time);
0b39920b 741#ifdef CONFIG_DEBUG_BLK_CGROUP
e6269c44
TH
742 blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
743 blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
744 blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
745 blkg_stat_add_aux(&to->dequeue, &from->dequeue);
746 blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
747 blkg_stat_add_aux(&to->idle_time, &from->idle_time);
748 blkg_stat_add_aux(&to->empty_time, &from->empty_time);
0b39920b
TH
749#endif
750}
751
752/*
e6269c44 753 * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
0b39920b
TH
754 * recursive stats can still account for the amount used by this cfqg after
755 * it's gone.
756 */
757static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
758{
759 struct cfq_group *parent = cfqg_parent(cfqg);
760
761 lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
762
763 if (unlikely(!parent))
764 return;
765
e6269c44 766 cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
0b39920b 767 cfqg_stats_reset(&cfqg->stats);
0b39920b
TH
768}
769
eb7d8c07
TH
770#else /* CONFIG_CFQ_GROUP_IOSCHED */
771
d02f7aa8 772static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
3984aa55
JK
773static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
774 struct cfq_group *ancestor)
775{
776 return true;
777}
eb7d8c07
TH
778static inline void cfqg_get(struct cfq_group *cfqg) { }
779static inline void cfqg_put(struct cfq_group *cfqg) { }
780
7b679138 781#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
b226e5c4
VG
782 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
783 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
784 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
785 ##args)
4495a7d4 786#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
eb7d8c07 787
155fead9 788static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
63a4cc24 789 struct cfq_group *curr_cfqg, int op, int op_flags) { }
155fead9 790static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
9a7f38c4 791 uint64_t time, unsigned long unaccounted_time) { }
63a4cc24
MC
792static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op,
793 int op_flags) { }
794static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op,
795 int op_flags) { }
155fead9 796static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
63a4cc24
MC
797 uint64_t start_time, uint64_t io_start_time, int op,
798 int op_flags) { }
2ce4d50f 799
eb7d8c07
TH
800#endif /* CONFIG_CFQ_GROUP_IOSCHED */
801
7b679138
JA
802#define cfq_log(cfqd, fmt, args...) \
803 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
804
615f0259
VG
805/* Traverses through cfq group service trees */
806#define for_each_cfqg_st(cfqg, i, j, st) \
807 for (i = 0; i <= IDLE_WORKLOAD; i++) \
808 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
809 : &cfqg->service_tree_idle; \
810 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
811 (i == IDLE_WORKLOAD && j == 0); \
812 j++, st = i < IDLE_WORKLOAD ? \
813 &cfqg->service_trees[i][j]: NULL) \
814
f5f2b6ce
SL
815static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
816 struct cfq_ttime *ttime, bool group_idle)
817{
9a7f38c4 818 u64 slice;
f5f2b6ce
SL
819 if (!sample_valid(ttime->ttime_samples))
820 return false;
821 if (group_idle)
822 slice = cfqd->cfq_group_idle;
823 else
824 slice = cfqd->cfq_slice_idle;
825 return ttime->ttime_mean > slice;
826}
615f0259 827
02b35081
VG
828static inline bool iops_mode(struct cfq_data *cfqd)
829{
830 /*
831 * If we are not idling on queues and it is a NCQ drive, parallel
832 * execution of requests is on and measuring time is not possible
833 * in most of the cases until and unless we drive shallower queue
834 * depths and that becomes a performance bottleneck. In such cases
835 * switch to start providing fairness in terms of number of IOs.
836 */
837 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
838 return true;
839 else
840 return false;
841}
842
3bf10fea 843static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
c0324a02
CZ
844{
845 if (cfq_class_idle(cfqq))
846 return IDLE_WORKLOAD;
847 if (cfq_class_rt(cfqq))
848 return RT_WORKLOAD;
849 return BE_WORKLOAD;
850}
851
718eee05
CZ
852
853static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
854{
855 if (!cfq_cfqq_sync(cfqq))
856 return ASYNC_WORKLOAD;
857 if (!cfq_cfqq_idle_window(cfqq))
858 return SYNC_NOIDLE_WORKLOAD;
859 return SYNC_WORKLOAD;
860}
861
3bf10fea 862static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
58ff82f3
VG
863 struct cfq_data *cfqd,
864 struct cfq_group *cfqg)
c0324a02 865{
3bf10fea 866 if (wl_class == IDLE_WORKLOAD)
cdb16e8f 867 return cfqg->service_tree_idle.count;
c0324a02 868
34b98d03
VG
869 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
870 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
871 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
c0324a02
CZ
872}
873
f26bd1f0
VG
874static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
875 struct cfq_group *cfqg)
876{
34b98d03
VG
877 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
878 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
f26bd1f0
VG
879}
880
165125e1 881static void cfq_dispatch_insert(struct request_queue *, struct request *);
4f85cb96 882static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
2da8de0b 883 struct cfq_io_cq *cic, struct bio *bio);
91fac317 884
c5869807
TH
885static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
886{
887 /* cic->icq is the first member, %NULL will convert to %NULL */
888 return container_of(icq, struct cfq_io_cq, icq);
889}
890
47fdd4ca
TH
891static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
892 struct io_context *ioc)
893{
894 if (ioc)
895 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
896 return NULL;
897}
898
c5869807 899static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
91fac317 900{
a6151c3a 901 return cic->cfqq[is_sync];
91fac317
VT
902}
903
c5869807
TH
904static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
905 bool is_sync)
91fac317 906{
a6151c3a 907 cic->cfqq[is_sync] = cfqq;
91fac317
VT
908}
909
c5869807 910static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
bca4b914 911{
c5869807 912 return cic->icq.q->elevator->elevator_data;
bca4b914
KK
913}
914
91fac317
VT
915/*
916 * We regard a request as SYNC, if it's either a read or has the SYNC bit
917 * set (in which case it could also be direct WRITE).
918 */
a6151c3a 919static inline bool cfq_bio_sync(struct bio *bio)
91fac317 920{
1eff9d32 921 return bio_data_dir(bio) == READ || (bio->bi_opf & REQ_SYNC);
91fac317 922}
1da177e4 923
99f95e52
AM
924/*
925 * scheduler run of queue, if there are requests pending and no one in the
926 * driver that will restart queueing
927 */
23e018a1 928static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
99f95e52 929{
7b679138
JA
930 if (cfqd->busy_queues) {
931 cfq_log(cfqd, "schedule dispatch");
59c3d45e 932 kblockd_schedule_work(&cfqd->unplug_work);
7b679138 933 }
99f95e52
AM
934}
935
44f7c160
JA
936/*
937 * Scale schedule slice based on io priority. Use the sync time slice only
938 * if a queue is marked sync and has sync io queued. A sync queue with async
939 * io only, should not get full sync slice length.
940 */
9a7f38c4 941static inline u64 cfq_prio_slice(struct cfq_data *cfqd, bool sync,
d9e7620e 942 unsigned short prio)
44f7c160 943{
9a7f38c4
JM
944 u64 base_slice = cfqd->cfq_slice[sync];
945 u64 slice = div_u64(base_slice, CFQ_SLICE_SCALE);
44f7c160 946
d9e7620e
JA
947 WARN_ON(prio >= IOPRIO_BE_NR);
948
9a7f38c4 949 return base_slice + (slice * (4 - prio));
d9e7620e 950}
44f7c160 951
9a7f38c4 952static inline u64
d9e7620e
JA
953cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
954{
955 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
44f7c160
JA
956}
957
1d3650f7
TH
958/**
959 * cfqg_scale_charge - scale disk time charge according to cfqg weight
960 * @charge: disk time being charged
961 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
962 *
963 * Scale @charge according to @vfraction, which is in range (0, 1]. The
964 * scaling is inversely proportional.
965 *
966 * scaled = charge / vfraction
967 *
968 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
969 */
9a7f38c4 970static inline u64 cfqg_scale_charge(u64 charge,
1d3650f7 971 unsigned int vfraction)
25bc6b07 972{
1d3650f7 973 u64 c = charge << CFQ_SERVICE_SHIFT; /* make it fixed point */
25bc6b07 974
1d3650f7
TH
975 /* charge / vfraction */
976 c <<= CFQ_SERVICE_SHIFT;
9a7f38c4 977 return div_u64(c, vfraction);
25bc6b07
VG
978}
979
980static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
981{
982 s64 delta = (s64)(vdisktime - min_vdisktime);
983 if (delta > 0)
984 min_vdisktime = vdisktime;
985
986 return min_vdisktime;
987}
988
989static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
990{
991 s64 delta = (s64)(vdisktime - min_vdisktime);
992 if (delta < 0)
993 min_vdisktime = vdisktime;
994
995 return min_vdisktime;
996}
997
998static void update_min_vdisktime(struct cfq_rb_root *st)
999{
25bc6b07
VG
1000 struct cfq_group *cfqg;
1001
25bc6b07
VG
1002 if (st->left) {
1003 cfqg = rb_entry_cfqg(st->left);
a6032710
GJ
1004 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
1005 cfqg->vdisktime);
25bc6b07 1006 }
25bc6b07
VG
1007}
1008
5db5d642
CZ
1009/*
1010 * get averaged number of queues of RT/BE priority.
1011 * average is updated, with a formula that gives more weight to higher numbers,
1012 * to quickly follows sudden increases and decrease slowly
1013 */
1014
58ff82f3
VG
1015static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1016 struct cfq_group *cfqg, bool rt)
5869619c 1017{
5db5d642
CZ
1018 unsigned min_q, max_q;
1019 unsigned mult = cfq_hist_divisor - 1;
1020 unsigned round = cfq_hist_divisor / 2;
58ff82f3 1021 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
5db5d642 1022
58ff82f3
VG
1023 min_q = min(cfqg->busy_queues_avg[rt], busy);
1024 max_q = max(cfqg->busy_queues_avg[rt], busy);
1025 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
5db5d642 1026 cfq_hist_divisor;
58ff82f3
VG
1027 return cfqg->busy_queues_avg[rt];
1028}
1029
9a7f38c4 1030static inline u64
58ff82f3
VG
1031cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1032{
41cad6ab 1033 return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
5db5d642
CZ
1034}
1035
9a7f38c4 1036static inline u64
ba5bd520 1037cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
44f7c160 1038{
9a7f38c4 1039 u64 slice = cfq_prio_to_slice(cfqd, cfqq);
5db5d642 1040 if (cfqd->cfq_latency) {
58ff82f3
VG
1041 /*
1042 * interested queues (we consider only the ones with the same
1043 * priority class in the cfq group)
1044 */
1045 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1046 cfq_class_rt(cfqq));
9a7f38c4
JM
1047 u64 sync_slice = cfqd->cfq_slice[1];
1048 u64 expect_latency = sync_slice * iq;
1049 u64 group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
58ff82f3
VG
1050
1051 if (expect_latency > group_slice) {
9a7f38c4
JM
1052 u64 base_low_slice = 2 * cfqd->cfq_slice_idle;
1053 u64 low_slice;
1054
5db5d642
CZ
1055 /* scale low_slice according to IO priority
1056 * and sync vs async */
9a7f38c4
JM
1057 low_slice = div64_u64(base_low_slice*slice, sync_slice);
1058 low_slice = min(slice, low_slice);
5db5d642
CZ
1059 /* the adapted slice value is scaled to fit all iqs
1060 * into the target latency */
9a7f38c4
JM
1061 slice = div64_u64(slice*group_slice, expect_latency);
1062 slice = max(slice, low_slice);
5db5d642
CZ
1063 }
1064 }
c553f8e3
SL
1065 return slice;
1066}
1067
1068static inline void
1069cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1070{
9a7f38c4
JM
1071 u64 slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1072 u64 now = ktime_get_ns();
c553f8e3 1073
9a7f38c4
JM
1074 cfqq->slice_start = now;
1075 cfqq->slice_end = now + slice;
f75edf2d 1076 cfqq->allocated_slice = slice;
9a7f38c4 1077 cfq_log_cfqq(cfqd, cfqq, "set_slice=%llu", cfqq->slice_end - now);
44f7c160
JA
1078}
1079
1080/*
1081 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1082 * isn't valid until the first request from the dispatch is activated
1083 * and the slice time set.
1084 */
a6151c3a 1085static inline bool cfq_slice_used(struct cfq_queue *cfqq)
44f7c160
JA
1086{
1087 if (cfq_cfqq_slice_new(cfqq))
c1e44756 1088 return false;
9a7f38c4 1089 if (ktime_get_ns() < cfqq->slice_end)
c1e44756 1090 return false;
44f7c160 1091
c1e44756 1092 return true;
44f7c160
JA
1093}
1094
1da177e4 1095/*
5e705374 1096 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1da177e4 1097 * We choose the request that is closest to the head right now. Distance
e8a99053 1098 * behind the head is penalized and only allowed to a certain extent.
1da177e4 1099 */
5e705374 1100static struct request *
cf7c25cf 1101cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1da177e4 1102{
cf7c25cf 1103 sector_t s1, s2, d1 = 0, d2 = 0;
1da177e4 1104 unsigned long back_max;
e8a99053
AM
1105#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1106#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1107 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1da177e4 1108
5e705374
JA
1109 if (rq1 == NULL || rq1 == rq2)
1110 return rq2;
1111 if (rq2 == NULL)
1112 return rq1;
9c2c38a1 1113
229836bd
NK
1114 if (rq_is_sync(rq1) != rq_is_sync(rq2))
1115 return rq_is_sync(rq1) ? rq1 : rq2;
1116
65299a3b
CH
1117 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1118 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
b53d1ed7 1119
83096ebf
TH
1120 s1 = blk_rq_pos(rq1);
1121 s2 = blk_rq_pos(rq2);
1da177e4 1122
1da177e4
LT
1123 /*
1124 * by definition, 1KiB is 2 sectors
1125 */
1126 back_max = cfqd->cfq_back_max * 2;
1127
1128 /*
1129 * Strict one way elevator _except_ in the case where we allow
1130 * short backward seeks which are biased as twice the cost of a
1131 * similar forward seek.
1132 */
1133 if (s1 >= last)
1134 d1 = s1 - last;
1135 else if (s1 + back_max >= last)
1136 d1 = (last - s1) * cfqd->cfq_back_penalty;
1137 else
e8a99053 1138 wrap |= CFQ_RQ1_WRAP;
1da177e4
LT
1139
1140 if (s2 >= last)
1141 d2 = s2 - last;
1142 else if (s2 + back_max >= last)
1143 d2 = (last - s2) * cfqd->cfq_back_penalty;
1144 else
e8a99053 1145 wrap |= CFQ_RQ2_WRAP;
1da177e4
LT
1146
1147 /* Found required data */
e8a99053
AM
1148
1149 /*
1150 * By doing switch() on the bit mask "wrap" we avoid having to
1151 * check two variables for all permutations: --> faster!
1152 */
1153 switch (wrap) {
5e705374 1154 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
e8a99053 1155 if (d1 < d2)
5e705374 1156 return rq1;
e8a99053 1157 else if (d2 < d1)
5e705374 1158 return rq2;
e8a99053
AM
1159 else {
1160 if (s1 >= s2)
5e705374 1161 return rq1;
e8a99053 1162 else
5e705374 1163 return rq2;
e8a99053 1164 }
1da177e4 1165
e8a99053 1166 case CFQ_RQ2_WRAP:
5e705374 1167 return rq1;
e8a99053 1168 case CFQ_RQ1_WRAP:
5e705374
JA
1169 return rq2;
1170 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
e8a99053
AM
1171 default:
1172 /*
1173 * Since both rqs are wrapped,
1174 * start with the one that's further behind head
1175 * (--> only *one* back seek required),
1176 * since back seek takes more time than forward.
1177 */
1178 if (s1 <= s2)
5e705374 1179 return rq1;
1da177e4 1180 else
5e705374 1181 return rq2;
1da177e4
LT
1182 }
1183}
1184
498d3aa2
JA
1185/*
1186 * The below is leftmost cache rbtree addon
1187 */
0871714e 1188static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
cc09e299 1189{
615f0259
VG
1190 /* Service tree is empty */
1191 if (!root->count)
1192 return NULL;
1193
cc09e299
JA
1194 if (!root->left)
1195 root->left = rb_first(&root->rb);
1196
0871714e
JA
1197 if (root->left)
1198 return rb_entry(root->left, struct cfq_queue, rb_node);
1199
1200 return NULL;
cc09e299
JA
1201}
1202
1fa8f6d6
VG
1203static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1204{
1205 if (!root->left)
1206 root->left = rb_first(&root->rb);
1207
1208 if (root->left)
1209 return rb_entry_cfqg(root->left);
1210
1211 return NULL;
1212}
1213
a36e71f9
JA
1214static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1215{
1216 rb_erase(n, root);
1217 RB_CLEAR_NODE(n);
1218}
1219
cc09e299
JA
1220static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1221{
1222 if (root->left == n)
1223 root->left = NULL;
a36e71f9 1224 rb_erase_init(n, &root->rb);
aa6f6a3d 1225 --root->count;
cc09e299
JA
1226}
1227
1da177e4
LT
1228/*
1229 * would be nice to take fifo expire time into account as well
1230 */
5e705374
JA
1231static struct request *
1232cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1233 struct request *last)
1da177e4 1234{
21183b07
JA
1235 struct rb_node *rbnext = rb_next(&last->rb_node);
1236 struct rb_node *rbprev = rb_prev(&last->rb_node);
5e705374 1237 struct request *next = NULL, *prev = NULL;
1da177e4 1238
21183b07 1239 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1da177e4
LT
1240
1241 if (rbprev)
5e705374 1242 prev = rb_entry_rq(rbprev);
1da177e4 1243
21183b07 1244 if (rbnext)
5e705374 1245 next = rb_entry_rq(rbnext);
21183b07
JA
1246 else {
1247 rbnext = rb_first(&cfqq->sort_list);
1248 if (rbnext && rbnext != &last->rb_node)
5e705374 1249 next = rb_entry_rq(rbnext);
21183b07 1250 }
1da177e4 1251
cf7c25cf 1252 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1da177e4
LT
1253}
1254
9a7f38c4
JM
1255static u64 cfq_slice_offset(struct cfq_data *cfqd,
1256 struct cfq_queue *cfqq)
1da177e4 1257{
d9e7620e
JA
1258 /*
1259 * just an approximation, should be ok.
1260 */
cdb16e8f 1261 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
464191c6 1262 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
d9e7620e
JA
1263}
1264
1fa8f6d6
VG
1265static inline s64
1266cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1267{
1268 return cfqg->vdisktime - st->min_vdisktime;
1269}
1270
1271static void
1272__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1273{
1274 struct rb_node **node = &st->rb.rb_node;
1275 struct rb_node *parent = NULL;
1276 struct cfq_group *__cfqg;
1277 s64 key = cfqg_key(st, cfqg);
1278 int left = 1;
1279
1280 while (*node != NULL) {
1281 parent = *node;
1282 __cfqg = rb_entry_cfqg(parent);
1283
1284 if (key < cfqg_key(st, __cfqg))
1285 node = &parent->rb_left;
1286 else {
1287 node = &parent->rb_right;
1288 left = 0;
1289 }
1290 }
1291
1292 if (left)
1293 st->left = &cfqg->rb_node;
1294
1295 rb_link_node(&cfqg->rb_node, parent, node);
1296 rb_insert_color(&cfqg->rb_node, &st->rb);
1297}
1298
7b5af5cf
TM
1299/*
1300 * This has to be called only on activation of cfqg
1301 */
1fa8f6d6 1302static void
8184f93e
JT
1303cfq_update_group_weight(struct cfq_group *cfqg)
1304{
3381cb8d 1305 if (cfqg->new_weight) {
8184f93e 1306 cfqg->weight = cfqg->new_weight;
3381cb8d 1307 cfqg->new_weight = 0;
8184f93e 1308 }
e15693ef
TM
1309}
1310
1311static void
1312cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1313{
1314 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
e71357e1
TH
1315
1316 if (cfqg->new_leaf_weight) {
1317 cfqg->leaf_weight = cfqg->new_leaf_weight;
1318 cfqg->new_leaf_weight = 0;
1319 }
8184f93e
JT
1320}
1321
1322static void
1323cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1324{
1d3650f7 1325 unsigned int vfr = 1 << CFQ_SERVICE_SHIFT; /* start with 1 */
7918ffb5 1326 struct cfq_group *pos = cfqg;
1d3650f7 1327 struct cfq_group *parent;
7918ffb5
TH
1328 bool propagate;
1329
1330 /* add to the service tree */
8184f93e
JT
1331 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1332
7b5af5cf
TM
1333 /*
1334 * Update leaf_weight. We cannot update weight at this point
1335 * because cfqg might already have been activated and is
1336 * contributing its current weight to the parent's child_weight.
1337 */
e15693ef 1338 cfq_update_group_leaf_weight(cfqg);
8184f93e 1339 __cfq_group_service_tree_add(st, cfqg);
7918ffb5
TH
1340
1341 /*
1d3650f7
TH
1342 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1343 * entitled to. vfraction is calculated by walking the tree
1344 * towards the root calculating the fraction it has at each level.
1345 * The compounded ratio is how much vfraction @cfqg owns.
1346 *
1347 * Start with the proportion tasks in this cfqg has against active
1348 * children cfqgs - its leaf_weight against children_weight.
7918ffb5
TH
1349 */
1350 propagate = !pos->nr_active++;
1351 pos->children_weight += pos->leaf_weight;
1d3650f7 1352 vfr = vfr * pos->leaf_weight / pos->children_weight;
7918ffb5 1353
1d3650f7
TH
1354 /*
1355 * Compound ->weight walking up the tree. Both activation and
1356 * vfraction calculation are done in the same loop. Propagation
1357 * stops once an already activated node is met. vfraction
1358 * calculation should always continue to the root.
1359 */
d02f7aa8 1360 while ((parent = cfqg_parent(pos))) {
1d3650f7 1361 if (propagate) {
e15693ef 1362 cfq_update_group_weight(pos);
1d3650f7
TH
1363 propagate = !parent->nr_active++;
1364 parent->children_weight += pos->weight;
1365 }
1366 vfr = vfr * pos->weight / parent->children_weight;
7918ffb5
TH
1367 pos = parent;
1368 }
1d3650f7
TH
1369
1370 cfqg->vfraction = max_t(unsigned, vfr, 1);
8184f93e
JT
1371}
1372
1373static void
1374cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1fa8f6d6
VG
1375{
1376 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1377 struct cfq_group *__cfqg;
1378 struct rb_node *n;
1379
1380 cfqg->nr_cfqq++;
760701bf 1381 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1fa8f6d6
VG
1382 return;
1383
1384 /*
1385 * Currently put the group at the end. Later implement something
1386 * so that groups get lesser vtime based on their weights, so that
25985edc 1387 * if group does not loose all if it was not continuously backlogged.
1fa8f6d6
VG
1388 */
1389 n = rb_last(&st->rb);
1390 if (n) {
1391 __cfqg = rb_entry_cfqg(n);
1392 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1393 } else
1394 cfqg->vdisktime = st->min_vdisktime;
8184f93e
JT
1395 cfq_group_service_tree_add(st, cfqg);
1396}
1fa8f6d6 1397
8184f93e
JT
1398static void
1399cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1400{
7918ffb5
TH
1401 struct cfq_group *pos = cfqg;
1402 bool propagate;
1403
1404 /*
1405 * Undo activation from cfq_group_service_tree_add(). Deactivate
1406 * @cfqg and propagate deactivation upwards.
1407 */
1408 propagate = !--pos->nr_active;
1409 pos->children_weight -= pos->leaf_weight;
1410
1411 while (propagate) {
d02f7aa8 1412 struct cfq_group *parent = cfqg_parent(pos);
7918ffb5
TH
1413
1414 /* @pos has 0 nr_active at this point */
1415 WARN_ON_ONCE(pos->children_weight);
1d3650f7 1416 pos->vfraction = 0;
7918ffb5
TH
1417
1418 if (!parent)
1419 break;
1420
1421 propagate = !--parent->nr_active;
1422 parent->children_weight -= pos->weight;
1423 pos = parent;
1424 }
1425
1426 /* remove from the service tree */
8184f93e
JT
1427 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1428 cfq_rb_erase(&cfqg->rb_node, st);
1fa8f6d6
VG
1429}
1430
1431static void
8184f93e 1432cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1fa8f6d6
VG
1433{
1434 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1435
1436 BUG_ON(cfqg->nr_cfqq < 1);
1437 cfqg->nr_cfqq--;
25bc6b07 1438
1fa8f6d6
VG
1439 /* If there are other cfq queues under this group, don't delete it */
1440 if (cfqg->nr_cfqq)
1441 return;
1442
2868ef7b 1443 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
8184f93e 1444 cfq_group_service_tree_del(st, cfqg);
4d2ceea4 1445 cfqg->saved_wl_slice = 0;
155fead9 1446 cfqg_stats_update_dequeue(cfqg);
dae739eb
VG
1447}
1448
9a7f38c4
JM
1449static inline u64 cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1450 u64 *unaccounted_time)
dae739eb 1451{
9a7f38c4
JM
1452 u64 slice_used;
1453 u64 now = ktime_get_ns();
dae739eb
VG
1454
1455 /*
1456 * Queue got expired before even a single request completed or
1457 * got expired immediately after first request completion.
1458 */
9a7f38c4 1459 if (!cfqq->slice_start || cfqq->slice_start == now) {
dae739eb
VG
1460 /*
1461 * Also charge the seek time incurred to the group, otherwise
1462 * if there are mutiple queues in the group, each can dispatch
1463 * a single request on seeky media and cause lots of seek time
1464 * and group will never know it.
1465 */
0b31c10c
JK
1466 slice_used = max_t(u64, (now - cfqq->dispatch_start),
1467 jiffies_to_nsecs(1));
dae739eb 1468 } else {
9a7f38c4 1469 slice_used = now - cfqq->slice_start;
167400d3
JT
1470 if (slice_used > cfqq->allocated_slice) {
1471 *unaccounted_time = slice_used - cfqq->allocated_slice;
f75edf2d 1472 slice_used = cfqq->allocated_slice;
167400d3 1473 }
9a7f38c4 1474 if (cfqq->slice_start > cfqq->dispatch_start)
167400d3
JT
1475 *unaccounted_time += cfqq->slice_start -
1476 cfqq->dispatch_start;
dae739eb
VG
1477 }
1478
dae739eb
VG
1479 return slice_used;
1480}
1481
1482static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
e5ff082e 1483 struct cfq_queue *cfqq)
dae739eb
VG
1484{
1485 struct cfq_rb_root *st = &cfqd->grp_service_tree;
9a7f38c4 1486 u64 used_sl, charge, unaccounted_sl = 0;
f26bd1f0
VG
1487 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1488 - cfqg->service_tree_idle.count;
1d3650f7 1489 unsigned int vfr;
9a7f38c4 1490 u64 now = ktime_get_ns();
f26bd1f0
VG
1491
1492 BUG_ON(nr_sync < 0);
167400d3 1493 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
dae739eb 1494
02b35081
VG
1495 if (iops_mode(cfqd))
1496 charge = cfqq->slice_dispatch;
1497 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1498 charge = cfqq->allocated_slice;
dae739eb 1499
1d3650f7
TH
1500 /*
1501 * Can't update vdisktime while on service tree and cfqg->vfraction
1502 * is valid only while on it. Cache vfr, leave the service tree,
1503 * update vdisktime and go back on. The re-addition to the tree
1504 * will also update the weights as necessary.
1505 */
1506 vfr = cfqg->vfraction;
8184f93e 1507 cfq_group_service_tree_del(st, cfqg);
1d3650f7 1508 cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
8184f93e 1509 cfq_group_service_tree_add(st, cfqg);
dae739eb
VG
1510
1511 /* This group is being expired. Save the context */
9a7f38c4
JM
1512 if (cfqd->workload_expires > now) {
1513 cfqg->saved_wl_slice = cfqd->workload_expires - now;
4d2ceea4
VG
1514 cfqg->saved_wl_type = cfqd->serving_wl_type;
1515 cfqg->saved_wl_class = cfqd->serving_wl_class;
dae739eb 1516 } else
4d2ceea4 1517 cfqg->saved_wl_slice = 0;
2868ef7b
VG
1518
1519 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1520 st->min_vdisktime);
fd16d263 1521 cfq_log_cfqq(cfqq->cfqd, cfqq,
9a7f38c4 1522 "sl_used=%llu disp=%llu charge=%llu iops=%u sect=%lu",
fd16d263
JP
1523 used_sl, cfqq->slice_dispatch, charge,
1524 iops_mode(cfqd), cfqq->nr_sectors);
155fead9
TH
1525 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1526 cfqg_stats_set_start_empty_time(cfqg);
1fa8f6d6
VG
1527}
1528
f51b802c
TH
1529/**
1530 * cfq_init_cfqg_base - initialize base part of a cfq_group
1531 * @cfqg: cfq_group to initialize
1532 *
1533 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1534 * is enabled or not.
1535 */
1536static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1537{
1538 struct cfq_rb_root *st;
1539 int i, j;
1540
1541 for_each_cfqg_st(cfqg, i, j, st)
1542 *st = CFQ_RB_ROOT;
1543 RB_CLEAR_NODE(&cfqg->rb_node);
1544
9a7f38c4 1545 cfqg->ttime.last_end_request = ktime_get_ns();
f51b802c
TH
1546}
1547
25fb5169 1548#ifdef CONFIG_CFQ_GROUP_IOSCHED
69d7fde5
TH
1549static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1550 bool on_dfl, bool reset_dev, bool is_leaf_weight);
1551
24bdb8ef 1552static void cfqg_stats_exit(struct cfqg_stats *stats)
90d3839b 1553{
24bdb8ef
TH
1554 blkg_rwstat_exit(&stats->merged);
1555 blkg_rwstat_exit(&stats->service_time);
1556 blkg_rwstat_exit(&stats->wait_time);
1557 blkg_rwstat_exit(&stats->queued);
24bdb8ef
TH
1558 blkg_stat_exit(&stats->time);
1559#ifdef CONFIG_DEBUG_BLK_CGROUP
1560 blkg_stat_exit(&stats->unaccounted_time);
1561 blkg_stat_exit(&stats->avg_queue_size_sum);
1562 blkg_stat_exit(&stats->avg_queue_size_samples);
1563 blkg_stat_exit(&stats->dequeue);
1564 blkg_stat_exit(&stats->group_wait_time);
1565 blkg_stat_exit(&stats->idle_time);
1566 blkg_stat_exit(&stats->empty_time);
1567#endif
1568}
1569
1570static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1571{
77ea7338 1572 if (blkg_rwstat_init(&stats->merged, gfp) ||
24bdb8ef
TH
1573 blkg_rwstat_init(&stats->service_time, gfp) ||
1574 blkg_rwstat_init(&stats->wait_time, gfp) ||
1575 blkg_rwstat_init(&stats->queued, gfp) ||
24bdb8ef
TH
1576 blkg_stat_init(&stats->time, gfp))
1577 goto err;
90d3839b
PZ
1578
1579#ifdef CONFIG_DEBUG_BLK_CGROUP
24bdb8ef
TH
1580 if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1581 blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1582 blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1583 blkg_stat_init(&stats->dequeue, gfp) ||
1584 blkg_stat_init(&stats->group_wait_time, gfp) ||
1585 blkg_stat_init(&stats->idle_time, gfp) ||
1586 blkg_stat_init(&stats->empty_time, gfp))
1587 goto err;
90d3839b 1588#endif
24bdb8ef
TH
1589 return 0;
1590err:
1591 cfqg_stats_exit(stats);
1592 return -ENOMEM;
90d3839b
PZ
1593}
1594
e4a9bde9
TH
1595static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1596{
1597 struct cfq_group_data *cgd;
1598
1599 cgd = kzalloc(sizeof(*cgd), GFP_KERNEL);
1600 if (!cgd)
1601 return NULL;
1602 return &cgd->cpd;
1603}
1604
81437648 1605static void cfq_cpd_init(struct blkcg_policy_data *cpd)
e48453c3 1606{
81437648 1607 struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
9e10a130 1608 unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
69d7fde5 1609 CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
e48453c3 1610
69d7fde5
TH
1611 if (cpd_to_blkcg(cpd) == &blkcg_root)
1612 weight *= 2;
1613
1614 cgd->weight = weight;
1615 cgd->leaf_weight = weight;
e48453c3
AA
1616}
1617
e4a9bde9
TH
1618static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1619{
1620 kfree(cpd_to_cfqgd(cpd));
1621}
1622
69d7fde5
TH
1623static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1624{
1625 struct blkcg *blkcg = cpd_to_blkcg(cpd);
9e10a130 1626 bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
69d7fde5
TH
1627 unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1628
1629 if (blkcg == &blkcg_root)
1630 weight *= 2;
1631
1632 WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1633 WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1634}
1635
001bea73
TH
1636static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1637{
b2ce2643
TH
1638 struct cfq_group *cfqg;
1639
1640 cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1641 if (!cfqg)
1642 return NULL;
1643
1644 cfq_init_cfqg_base(cfqg);
24bdb8ef
TH
1645 if (cfqg_stats_init(&cfqg->stats, gfp)) {
1646 kfree(cfqg);
1647 return NULL;
1648 }
b2ce2643
TH
1649
1650 return &cfqg->pd;
001bea73
TH
1651}
1652
a9520cd6 1653static void cfq_pd_init(struct blkg_policy_data *pd)
f469a7b4 1654{
a9520cd6
TH
1655 struct cfq_group *cfqg = pd_to_cfqg(pd);
1656 struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
25fb5169 1657
e48453c3
AA
1658 cfqg->weight = cgd->weight;
1659 cfqg->leaf_weight = cgd->leaf_weight;
25fb5169
VG
1660}
1661
a9520cd6 1662static void cfq_pd_offline(struct blkg_policy_data *pd)
0b39920b 1663{
a9520cd6 1664 struct cfq_group *cfqg = pd_to_cfqg(pd);
60a83707
TH
1665 int i;
1666
1667 for (i = 0; i < IOPRIO_BE_NR; i++) {
1668 if (cfqg->async_cfqq[0][i])
1669 cfq_put_queue(cfqg->async_cfqq[0][i]);
1670 if (cfqg->async_cfqq[1][i])
1671 cfq_put_queue(cfqg->async_cfqq[1][i]);
1672 }
1673
1674 if (cfqg->async_idle_cfqq)
1675 cfq_put_queue(cfqg->async_idle_cfqq);
1676
0b39920b
TH
1677 /*
1678 * @blkg is going offline and will be ignored by
1679 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
1680 * that they don't get lost. If IOs complete after this point, the
1681 * stats for them will be lost. Oh well...
1682 */
60a83707 1683 cfqg_stats_xfer_dead(cfqg);
0b39920b
TH
1684}
1685
001bea73
TH
1686static void cfq_pd_free(struct blkg_policy_data *pd)
1687{
24bdb8ef
TH
1688 struct cfq_group *cfqg = pd_to_cfqg(pd);
1689
1690 cfqg_stats_exit(&cfqg->stats);
1691 return kfree(cfqg);
001bea73
TH
1692}
1693
a9520cd6 1694static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
689665af 1695{
a9520cd6 1696 struct cfq_group *cfqg = pd_to_cfqg(pd);
689665af
TH
1697
1698 cfqg_stats_reset(&cfqg->stats);
25fb5169
VG
1699}
1700
ae118896
TH
1701static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1702 struct blkcg *blkcg)
25fb5169 1703{
ae118896 1704 struct blkcg_gq *blkg;
f469a7b4 1705
ae118896
TH
1706 blkg = blkg_lookup(blkcg, cfqd->queue);
1707 if (likely(blkg))
1708 return blkg_to_cfqg(blkg);
1709 return NULL;
25fb5169
VG
1710}
1711
1712static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1713{
25fb5169 1714 cfqq->cfqg = cfqg;
b1c35769 1715 /* cfqq reference on cfqg */
eb7d8c07 1716 cfqg_get(cfqg);
b1c35769
VG
1717}
1718
f95a04af
TH
1719static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1720 struct blkg_policy_data *pd, int off)
60c2bc2d 1721{
f95a04af 1722 struct cfq_group *cfqg = pd_to_cfqg(pd);
3381cb8d
TH
1723
1724 if (!cfqg->dev_weight)
60c2bc2d 1725 return 0;
f95a04af 1726 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
60c2bc2d
TH
1727}
1728
2da8ca82 1729static int cfqg_print_weight_device(struct seq_file *sf, void *v)
60c2bc2d 1730{
2da8ca82
TH
1731 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1732 cfqg_prfill_weight_device, &blkcg_policy_cfq,
1733 0, false);
60c2bc2d
TH
1734 return 0;
1735}
1736
e71357e1
TH
1737static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1738 struct blkg_policy_data *pd, int off)
1739{
1740 struct cfq_group *cfqg = pd_to_cfqg(pd);
1741
1742 if (!cfqg->dev_leaf_weight)
1743 return 0;
1744 return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1745}
1746
2da8ca82 1747static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
e71357e1 1748{
2da8ca82
TH
1749 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1750 cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1751 0, false);
e71357e1
TH
1752 return 0;
1753}
1754
2da8ca82 1755static int cfq_print_weight(struct seq_file *sf, void *v)
60c2bc2d 1756{
e48453c3 1757 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
9470e4a6
JA
1758 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1759 unsigned int val = 0;
e48453c3 1760
9470e4a6
JA
1761 if (cgd)
1762 val = cgd->weight;
1763
1764 seq_printf(sf, "%u\n", val);
60c2bc2d
TH
1765 return 0;
1766}
1767
2da8ca82 1768static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
e71357e1 1769{
e48453c3 1770 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
9470e4a6
JA
1771 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1772 unsigned int val = 0;
1773
1774 if (cgd)
1775 val = cgd->leaf_weight;
e48453c3 1776
9470e4a6 1777 seq_printf(sf, "%u\n", val);
e71357e1
TH
1778 return 0;
1779}
1780
451af504
TH
1781static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1782 char *buf, size_t nbytes, loff_t off,
2ee867dc 1783 bool on_dfl, bool is_leaf_weight)
60c2bc2d 1784{
69d7fde5
TH
1785 unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1786 unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
451af504 1787 struct blkcg *blkcg = css_to_blkcg(of_css(of));
60c2bc2d 1788 struct blkg_conf_ctx ctx;
3381cb8d 1789 struct cfq_group *cfqg;
e48453c3 1790 struct cfq_group_data *cfqgd;
60c2bc2d 1791 int ret;
36aa9e5f 1792 u64 v;
60c2bc2d 1793
3c798398 1794 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
60c2bc2d
TH
1795 if (ret)
1796 return ret;
1797
2ee867dc
TH
1798 if (sscanf(ctx.body, "%llu", &v) == 1) {
1799 /* require "default" on dfl */
1800 ret = -ERANGE;
1801 if (!v && on_dfl)
1802 goto out_finish;
1803 } else if (!strcmp(strim(ctx.body), "default")) {
1804 v = 0;
1805 } else {
1806 ret = -EINVAL;
36aa9e5f 1807 goto out_finish;
2ee867dc 1808 }
36aa9e5f 1809
3381cb8d 1810 cfqg = blkg_to_cfqg(ctx.blkg);
e48453c3 1811 cfqgd = blkcg_to_cfqgd(blkcg);
ae994ea9 1812
20386ce0 1813 ret = -ERANGE;
69d7fde5 1814 if (!v || (v >= min && v <= max)) {
e71357e1 1815 if (!is_leaf_weight) {
36aa9e5f
TH
1816 cfqg->dev_weight = v;
1817 cfqg->new_weight = v ?: cfqgd->weight;
e71357e1 1818 } else {
36aa9e5f
TH
1819 cfqg->dev_leaf_weight = v;
1820 cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
e71357e1 1821 }
60c2bc2d
TH
1822 ret = 0;
1823 }
36aa9e5f 1824out_finish:
60c2bc2d 1825 blkg_conf_finish(&ctx);
451af504 1826 return ret ?: nbytes;
60c2bc2d
TH
1827}
1828
451af504
TH
1829static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1830 char *buf, size_t nbytes, loff_t off)
e71357e1 1831{
2ee867dc 1832 return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
e71357e1
TH
1833}
1834
451af504
TH
1835static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1836 char *buf, size_t nbytes, loff_t off)
e71357e1 1837{
2ee867dc 1838 return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
e71357e1
TH
1839}
1840
dd165eb3 1841static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
69d7fde5 1842 bool on_dfl, bool reset_dev, bool is_leaf_weight)
60c2bc2d 1843{
69d7fde5
TH
1844 unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1845 unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
182446d0 1846 struct blkcg *blkcg = css_to_blkcg(css);
3c798398 1847 struct blkcg_gq *blkg;
e48453c3 1848 struct cfq_group_data *cfqgd;
ae994ea9 1849 int ret = 0;
60c2bc2d 1850
69d7fde5
TH
1851 if (val < min || val > max)
1852 return -ERANGE;
60c2bc2d
TH
1853
1854 spin_lock_irq(&blkcg->lock);
e48453c3 1855 cfqgd = blkcg_to_cfqgd(blkcg);
ae994ea9
JA
1856 if (!cfqgd) {
1857 ret = -EINVAL;
1858 goto out;
1859 }
e71357e1
TH
1860
1861 if (!is_leaf_weight)
e48453c3 1862 cfqgd->weight = val;
e71357e1 1863 else
e48453c3 1864 cfqgd->leaf_weight = val;
60c2bc2d 1865
b67bfe0d 1866 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
3381cb8d 1867 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
60c2bc2d 1868
e71357e1
TH
1869 if (!cfqg)
1870 continue;
1871
1872 if (!is_leaf_weight) {
69d7fde5
TH
1873 if (reset_dev)
1874 cfqg->dev_weight = 0;
e71357e1 1875 if (!cfqg->dev_weight)
e48453c3 1876 cfqg->new_weight = cfqgd->weight;
e71357e1 1877 } else {
69d7fde5
TH
1878 if (reset_dev)
1879 cfqg->dev_leaf_weight = 0;
e71357e1 1880 if (!cfqg->dev_leaf_weight)
e48453c3 1881 cfqg->new_leaf_weight = cfqgd->leaf_weight;
e71357e1 1882 }
60c2bc2d
TH
1883 }
1884
ae994ea9 1885out:
60c2bc2d 1886 spin_unlock_irq(&blkcg->lock);
ae994ea9 1887 return ret;
60c2bc2d
TH
1888}
1889
182446d0
TH
1890static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1891 u64 val)
e71357e1 1892{
69d7fde5 1893 return __cfq_set_weight(css, val, false, false, false);
e71357e1
TH
1894}
1895
182446d0
TH
1896static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1897 struct cftype *cft, u64 val)
e71357e1 1898{
69d7fde5 1899 return __cfq_set_weight(css, val, false, false, true);
e71357e1
TH
1900}
1901
2da8ca82 1902static int cfqg_print_stat(struct seq_file *sf, void *v)
5bc4afb1 1903{
2da8ca82
TH
1904 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1905 &blkcg_policy_cfq, seq_cft(sf)->private, false);
5bc4afb1
TH
1906 return 0;
1907}
1908
2da8ca82 1909static int cfqg_print_rwstat(struct seq_file *sf, void *v)
5bc4afb1 1910{
2da8ca82
TH
1911 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1912 &blkcg_policy_cfq, seq_cft(sf)->private, true);
5bc4afb1
TH
1913 return 0;
1914}
1915
43114018
TH
1916static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1917 struct blkg_policy_data *pd, int off)
1918{
f12c74ca
TH
1919 u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1920 &blkcg_policy_cfq, off);
43114018
TH
1921 return __blkg_prfill_u64(sf, pd, sum);
1922}
1923
1924static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1925 struct blkg_policy_data *pd, int off)
1926{
f12c74ca
TH
1927 struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1928 &blkcg_policy_cfq, off);
43114018
TH
1929 return __blkg_prfill_rwstat(sf, pd, &sum);
1930}
1931
2da8ca82 1932static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
43114018 1933{
2da8ca82
TH
1934 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1935 cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1936 seq_cft(sf)->private, false);
43114018
TH
1937 return 0;
1938}
1939
2da8ca82 1940static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
43114018 1941{
2da8ca82
TH
1942 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1943 cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1944 seq_cft(sf)->private, true);
43114018
TH
1945 return 0;
1946}
1947
702747ca
TH
1948static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1949 int off)
1950{
1951 u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1952
1953 return __blkg_prfill_u64(sf, pd, sum >> 9);
1954}
1955
1956static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1957{
1958 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1959 cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1960 return 0;
1961}
1962
1963static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1964 struct blkg_policy_data *pd, int off)
1965{
1966 struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1967 offsetof(struct blkcg_gq, stat_bytes));
1968 u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1969 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1970
1971 return __blkg_prfill_u64(sf, pd, sum >> 9);
1972}
1973
1974static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1975{
1976 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1977 cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1978 false);
1979 return 0;
1980}
1981
60c2bc2d 1982#ifdef CONFIG_DEBUG_BLK_CGROUP
f95a04af
TH
1983static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1984 struct blkg_policy_data *pd, int off)
60c2bc2d 1985{
f95a04af 1986 struct cfq_group *cfqg = pd_to_cfqg(pd);
155fead9 1987 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
60c2bc2d
TH
1988 u64 v = 0;
1989
1990 if (samples) {
155fead9 1991 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
f3cff25f 1992 v = div64_u64(v, samples);
60c2bc2d 1993 }
f95a04af 1994 __blkg_prfill_u64(sf, pd, v);
60c2bc2d
TH
1995 return 0;
1996}
1997
1998/* print avg_queue_size */
2da8ca82 1999static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
60c2bc2d 2000{
2da8ca82
TH
2001 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
2002 cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
2003 0, false);
60c2bc2d
TH
2004 return 0;
2005}
2006#endif /* CONFIG_DEBUG_BLK_CGROUP */
2007
880f50e2 2008static struct cftype cfq_blkcg_legacy_files[] = {
1d3650f7 2009 /* on root, weight is mapped to leaf_weight */
60c2bc2d
TH
2010 {
2011 .name = "weight_device",
1d3650f7 2012 .flags = CFTYPE_ONLY_ON_ROOT,
2da8ca82 2013 .seq_show = cfqg_print_leaf_weight_device,
451af504 2014 .write = cfqg_set_leaf_weight_device,
60c2bc2d
TH
2015 },
2016 {
2017 .name = "weight",
1d3650f7 2018 .flags = CFTYPE_ONLY_ON_ROOT,
2da8ca82 2019 .seq_show = cfq_print_leaf_weight,
1d3650f7 2020 .write_u64 = cfq_set_leaf_weight,
60c2bc2d 2021 },
e71357e1 2022
1d3650f7 2023 /* no such mapping necessary for !roots */
60c2bc2d
TH
2024 {
2025 .name = "weight_device",
1d3650f7 2026 .flags = CFTYPE_NOT_ON_ROOT,
2da8ca82 2027 .seq_show = cfqg_print_weight_device,
451af504 2028 .write = cfqg_set_weight_device,
60c2bc2d
TH
2029 },
2030 {
2031 .name = "weight",
1d3650f7 2032 .flags = CFTYPE_NOT_ON_ROOT,
2da8ca82 2033 .seq_show = cfq_print_weight,
3381cb8d 2034 .write_u64 = cfq_set_weight,
60c2bc2d 2035 },
e71357e1 2036
e71357e1
TH
2037 {
2038 .name = "leaf_weight_device",
2da8ca82 2039 .seq_show = cfqg_print_leaf_weight_device,
451af504 2040 .write = cfqg_set_leaf_weight_device,
e71357e1
TH
2041 },
2042 {
2043 .name = "leaf_weight",
2da8ca82 2044 .seq_show = cfq_print_leaf_weight,
e71357e1
TH
2045 .write_u64 = cfq_set_leaf_weight,
2046 },
2047
43114018 2048 /* statistics, covers only the tasks in the cfqg */
60c2bc2d
TH
2049 {
2050 .name = "time",
5bc4afb1 2051 .private = offsetof(struct cfq_group, stats.time),
2da8ca82 2052 .seq_show = cfqg_print_stat,
60c2bc2d
TH
2053 },
2054 {
2055 .name = "sectors",
702747ca 2056 .seq_show = cfqg_print_stat_sectors,
60c2bc2d
TH
2057 },
2058 {
2059 .name = "io_service_bytes",
77ea7338
TH
2060 .private = (unsigned long)&blkcg_policy_cfq,
2061 .seq_show = blkg_print_stat_bytes,
60c2bc2d
TH
2062 },
2063 {
2064 .name = "io_serviced",
77ea7338
TH
2065 .private = (unsigned long)&blkcg_policy_cfq,
2066 .seq_show = blkg_print_stat_ios,
60c2bc2d
TH
2067 },
2068 {
2069 .name = "io_service_time",
5bc4afb1 2070 .private = offsetof(struct cfq_group, stats.service_time),
2da8ca82 2071 .seq_show = cfqg_print_rwstat,
60c2bc2d
TH
2072 },
2073 {
2074 .name = "io_wait_time",
5bc4afb1 2075 .private = offsetof(struct cfq_group, stats.wait_time),
2da8ca82 2076 .seq_show = cfqg_print_rwstat,
60c2bc2d
TH
2077 },
2078 {
2079 .name = "io_merged",
5bc4afb1 2080 .private = offsetof(struct cfq_group, stats.merged),
2da8ca82 2081 .seq_show = cfqg_print_rwstat,
60c2bc2d
TH
2082 },
2083 {
2084 .name = "io_queued",
5bc4afb1 2085 .private = offsetof(struct cfq_group, stats.queued),
2da8ca82 2086 .seq_show = cfqg_print_rwstat,
60c2bc2d 2087 },
43114018
TH
2088
2089 /* the same statictics which cover the cfqg and its descendants */
2090 {
2091 .name = "time_recursive",
2092 .private = offsetof(struct cfq_group, stats.time),
2da8ca82 2093 .seq_show = cfqg_print_stat_recursive,
43114018
TH
2094 },
2095 {
2096 .name = "sectors_recursive",
702747ca 2097 .seq_show = cfqg_print_stat_sectors_recursive,
43114018
TH
2098 },
2099 {
2100 .name = "io_service_bytes_recursive",
77ea7338
TH
2101 .private = (unsigned long)&blkcg_policy_cfq,
2102 .seq_show = blkg_print_stat_bytes_recursive,
43114018
TH
2103 },
2104 {
2105 .name = "io_serviced_recursive",
77ea7338
TH
2106 .private = (unsigned long)&blkcg_policy_cfq,
2107 .seq_show = blkg_print_stat_ios_recursive,
43114018
TH
2108 },
2109 {
2110 .name = "io_service_time_recursive",
2111 .private = offsetof(struct cfq_group, stats.service_time),
2da8ca82 2112 .seq_show = cfqg_print_rwstat_recursive,
43114018
TH
2113 },
2114 {
2115 .name = "io_wait_time_recursive",
2116 .private = offsetof(struct cfq_group, stats.wait_time),
2da8ca82 2117 .seq_show = cfqg_print_rwstat_recursive,
43114018
TH
2118 },
2119 {
2120 .name = "io_merged_recursive",
2121 .private = offsetof(struct cfq_group, stats.merged),
2da8ca82 2122 .seq_show = cfqg_print_rwstat_recursive,
43114018
TH
2123 },
2124 {
2125 .name = "io_queued_recursive",
2126 .private = offsetof(struct cfq_group, stats.queued),
2da8ca82 2127 .seq_show = cfqg_print_rwstat_recursive,
43114018 2128 },
60c2bc2d
TH
2129#ifdef CONFIG_DEBUG_BLK_CGROUP
2130 {
2131 .name = "avg_queue_size",
2da8ca82 2132 .seq_show = cfqg_print_avg_queue_size,
60c2bc2d
TH
2133 },
2134 {
2135 .name = "group_wait_time",
5bc4afb1 2136 .private = offsetof(struct cfq_group, stats.group_wait_time),
2da8ca82 2137 .seq_show = cfqg_print_stat,
60c2bc2d
TH
2138 },
2139 {
2140 .name = "idle_time",
5bc4afb1 2141 .private = offsetof(struct cfq_group, stats.idle_time),
2da8ca82 2142 .seq_show = cfqg_print_stat,
60c2bc2d
TH
2143 },
2144 {
2145 .name = "empty_time",
5bc4afb1 2146 .private = offsetof(struct cfq_group, stats.empty_time),
2da8ca82 2147 .seq_show = cfqg_print_stat,
60c2bc2d
TH
2148 },
2149 {
2150 .name = "dequeue",
5bc4afb1 2151 .private = offsetof(struct cfq_group, stats.dequeue),
2da8ca82 2152 .seq_show = cfqg_print_stat,
60c2bc2d
TH
2153 },
2154 {
2155 .name = "unaccounted_time",
5bc4afb1 2156 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2da8ca82 2157 .seq_show = cfqg_print_stat,
60c2bc2d
TH
2158 },
2159#endif /* CONFIG_DEBUG_BLK_CGROUP */
2160 { } /* terminate */
2161};
2ee867dc
TH
2162
2163static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2164{
2165 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2166 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2167
2168 seq_printf(sf, "default %u\n", cgd->weight);
2169 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2170 &blkcg_policy_cfq, 0, false);
2171 return 0;
2172}
2173
2174static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2175 char *buf, size_t nbytes, loff_t off)
2176{
2177 char *endp;
2178 int ret;
2179 u64 v;
2180
2181 buf = strim(buf);
2182
2183 /* "WEIGHT" or "default WEIGHT" sets the default weight */
2184 v = simple_strtoull(buf, &endp, 0);
2185 if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
69d7fde5 2186 ret = __cfq_set_weight(of_css(of), v, true, false, false);
2ee867dc
TH
2187 return ret ?: nbytes;
2188 }
2189
2190 /* "MAJ:MIN WEIGHT" */
2191 return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2192}
2193
2194static struct cftype cfq_blkcg_files[] = {
2195 {
2196 .name = "weight",
2197 .flags = CFTYPE_NOT_ON_ROOT,
2198 .seq_show = cfq_print_weight_on_dfl,
2199 .write = cfq_set_weight_on_dfl,
2200 },
2201 { } /* terminate */
2202};
2203
25fb5169 2204#else /* GROUP_IOSCHED */
ae118896
TH
2205static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2206 struct blkcg *blkcg)
25fb5169 2207{
f51b802c 2208 return cfqd->root_group;
25fb5169 2209}
7f1dc8a2 2210
25fb5169
VG
2211static inline void
2212cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2213 cfqq->cfqg = cfqg;
2214}
2215
2216#endif /* GROUP_IOSCHED */
2217
498d3aa2 2218/*
c0324a02 2219 * The cfqd->service_trees holds all pending cfq_queue's that have
498d3aa2
JA
2220 * requests waiting to be processed. It is sorted in the order that
2221 * we will service the queues.
2222 */
a36e71f9 2223static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
a6151c3a 2224 bool add_front)
d9e7620e 2225{
0871714e
JA
2226 struct rb_node **p, *parent;
2227 struct cfq_queue *__cfqq;
9a7f38c4 2228 u64 rb_key;
34b98d03 2229 struct cfq_rb_root *st;
498d3aa2 2230 int left;
dae739eb 2231 int new_cfqq = 1;
9a7f38c4 2232 u64 now = ktime_get_ns();
ae30c286 2233
34b98d03 2234 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
0871714e
JA
2235 if (cfq_class_idle(cfqq)) {
2236 rb_key = CFQ_IDLE_DELAY;
34b98d03 2237 parent = rb_last(&st->rb);
0871714e
JA
2238 if (parent && parent != &cfqq->rb_node) {
2239 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2240 rb_key += __cfqq->rb_key;
2241 } else
9a7f38c4 2242 rb_key += now;
0871714e 2243 } else if (!add_front) {
b9c8946b
JA
2244 /*
2245 * Get our rb key offset. Subtract any residual slice
2246 * value carried from last service. A negative resid
2247 * count indicates slice overrun, and this should position
2248 * the next service time further away in the tree.
2249 */
9a7f38c4 2250 rb_key = cfq_slice_offset(cfqd, cfqq) + now;
b9c8946b 2251 rb_key -= cfqq->slice_resid;
edd75ffd 2252 cfqq->slice_resid = 0;
48e025e6 2253 } else {
9a7f38c4 2254 rb_key = -NSEC_PER_SEC;
34b98d03 2255 __cfqq = cfq_rb_first(st);
9a7f38c4 2256 rb_key += __cfqq ? __cfqq->rb_key : now;
48e025e6 2257 }
1da177e4 2258
d9e7620e 2259 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
dae739eb 2260 new_cfqq = 0;
99f9628a 2261 /*
d9e7620e 2262 * same position, nothing more to do
99f9628a 2263 */
34b98d03 2264 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
d9e7620e 2265 return;
1da177e4 2266
aa6f6a3d
CZ
2267 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2268 cfqq->service_tree = NULL;
1da177e4 2269 }
d9e7620e 2270
498d3aa2 2271 left = 1;
0871714e 2272 parent = NULL;
34b98d03
VG
2273 cfqq->service_tree = st;
2274 p = &st->rb.rb_node;
d9e7620e
JA
2275 while (*p) {
2276 parent = *p;
2277 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2278
0c534e0a 2279 /*
c0324a02 2280 * sort by key, that represents service time.
0c534e0a 2281 */
9a7f38c4 2282 if (rb_key < __cfqq->rb_key)
1f23f121 2283 p = &parent->rb_left;
c0324a02 2284 else {
1f23f121 2285 p = &parent->rb_right;
cc09e299 2286 left = 0;
c0324a02 2287 }
d9e7620e
JA
2288 }
2289
cc09e299 2290 if (left)
34b98d03 2291 st->left = &cfqq->rb_node;
cc09e299 2292
d9e7620e
JA
2293 cfqq->rb_key = rb_key;
2294 rb_link_node(&cfqq->rb_node, parent, p);
34b98d03
VG
2295 rb_insert_color(&cfqq->rb_node, &st->rb);
2296 st->count++;
20359f27 2297 if (add_front || !new_cfqq)
dae739eb 2298 return;
8184f93e 2299 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1da177e4
LT
2300}
2301
a36e71f9 2302static struct cfq_queue *
f2d1f0ae
JA
2303cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2304 sector_t sector, struct rb_node **ret_parent,
2305 struct rb_node ***rb_link)
a36e71f9 2306{
a36e71f9
JA
2307 struct rb_node **p, *parent;
2308 struct cfq_queue *cfqq = NULL;
2309
2310 parent = NULL;
2311 p = &root->rb_node;
2312 while (*p) {
2313 struct rb_node **n;
2314
2315 parent = *p;
2316 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2317
2318 /*
2319 * Sort strictly based on sector. Smallest to the left,
2320 * largest to the right.
2321 */
2e46e8b2 2322 if (sector > blk_rq_pos(cfqq->next_rq))
a36e71f9 2323 n = &(*p)->rb_right;
2e46e8b2 2324 else if (sector < blk_rq_pos(cfqq->next_rq))
a36e71f9
JA
2325 n = &(*p)->rb_left;
2326 else
2327 break;
2328 p = n;
3ac6c9f8 2329 cfqq = NULL;
a36e71f9
JA
2330 }
2331
2332 *ret_parent = parent;
2333 if (rb_link)
2334 *rb_link = p;
3ac6c9f8 2335 return cfqq;
a36e71f9
JA
2336}
2337
2338static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2339{
a36e71f9
JA
2340 struct rb_node **p, *parent;
2341 struct cfq_queue *__cfqq;
2342
f2d1f0ae
JA
2343 if (cfqq->p_root) {
2344 rb_erase(&cfqq->p_node, cfqq->p_root);
2345 cfqq->p_root = NULL;
2346 }
a36e71f9
JA
2347
2348 if (cfq_class_idle(cfqq))
2349 return;
2350 if (!cfqq->next_rq)
2351 return;
2352
f2d1f0ae 2353 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2e46e8b2
TH
2354 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2355 blk_rq_pos(cfqq->next_rq), &parent, &p);
3ac6c9f8
JA
2356 if (!__cfqq) {
2357 rb_link_node(&cfqq->p_node, parent, p);
f2d1f0ae
JA
2358 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2359 } else
2360 cfqq->p_root = NULL;
a36e71f9
JA
2361}
2362
498d3aa2
JA
2363/*
2364 * Update cfqq's position in the service tree.
2365 */
edd75ffd 2366static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
6d048f53 2367{
6d048f53
JA
2368 /*
2369 * Resorting requires the cfqq to be on the RR list already.
2370 */
a36e71f9 2371 if (cfq_cfqq_on_rr(cfqq)) {
edd75ffd 2372 cfq_service_tree_add(cfqd, cfqq, 0);
a36e71f9
JA
2373 cfq_prio_tree_add(cfqd, cfqq);
2374 }
6d048f53
JA
2375}
2376
1da177e4
LT
2377/*
2378 * add to busy list of queues for service, trying to be fair in ordering
22e2c507 2379 * the pending list according to last request service
1da177e4 2380 */
febffd61 2381static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1da177e4 2382{
7b679138 2383 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
3b18152c
JA
2384 BUG_ON(cfq_cfqq_on_rr(cfqq));
2385 cfq_mark_cfqq_on_rr(cfqq);
1da177e4 2386 cfqd->busy_queues++;
ef8a41df
SL
2387 if (cfq_cfqq_sync(cfqq))
2388 cfqd->busy_sync_queues++;
1da177e4 2389
edd75ffd 2390 cfq_resort_rr_list(cfqd, cfqq);
1da177e4
LT
2391}
2392
498d3aa2
JA
2393/*
2394 * Called when the cfqq no longer has requests pending, remove it from
2395 * the service tree.
2396 */
febffd61 2397static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1da177e4 2398{
7b679138 2399 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
3b18152c
JA
2400 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2401 cfq_clear_cfqq_on_rr(cfqq);
1da177e4 2402
aa6f6a3d
CZ
2403 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2404 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2405 cfqq->service_tree = NULL;
2406 }
f2d1f0ae
JA
2407 if (cfqq->p_root) {
2408 rb_erase(&cfqq->p_node, cfqq->p_root);
2409 cfqq->p_root = NULL;
2410 }
d9e7620e 2411
8184f93e 2412 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1da177e4
LT
2413 BUG_ON(!cfqd->busy_queues);
2414 cfqd->busy_queues--;
ef8a41df
SL
2415 if (cfq_cfqq_sync(cfqq))
2416 cfqd->busy_sync_queues--;
1da177e4
LT
2417}
2418
2419/*
2420 * rb tree support functions
2421 */
febffd61 2422static void cfq_del_rq_rb(struct request *rq)
1da177e4 2423{
5e705374 2424 struct cfq_queue *cfqq = RQ_CFQQ(rq);
5e705374 2425 const int sync = rq_is_sync(rq);
1da177e4 2426
b4878f24
JA
2427 BUG_ON(!cfqq->queued[sync]);
2428 cfqq->queued[sync]--;
1da177e4 2429
5e705374 2430 elv_rb_del(&cfqq->sort_list, rq);
1da177e4 2431
f04a6424
VG
2432 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2433 /*
2434 * Queue will be deleted from service tree when we actually
2435 * expire it later. Right now just remove it from prio tree
2436 * as it is empty.
2437 */
2438 if (cfqq->p_root) {
2439 rb_erase(&cfqq->p_node, cfqq->p_root);
2440 cfqq->p_root = NULL;
2441 }
2442 }
1da177e4
LT
2443}
2444
5e705374 2445static void cfq_add_rq_rb(struct request *rq)
1da177e4 2446{
5e705374 2447 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1da177e4 2448 struct cfq_data *cfqd = cfqq->cfqd;
796d5116 2449 struct request *prev;
1da177e4 2450
5380a101 2451 cfqq->queued[rq_is_sync(rq)]++;
1da177e4 2452
796d5116 2453 elv_rb_add(&cfqq->sort_list, rq);
5fccbf61
JA
2454
2455 if (!cfq_cfqq_on_rr(cfqq))
2456 cfq_add_cfqq_rr(cfqd, cfqq);
5044eed4
JA
2457
2458 /*
2459 * check if this request is a better next-serve candidate
2460 */
a36e71f9 2461 prev = cfqq->next_rq;
cf7c25cf 2462 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
a36e71f9
JA
2463
2464 /*
2465 * adjust priority tree position, if ->next_rq changes
2466 */
2467 if (prev != cfqq->next_rq)
2468 cfq_prio_tree_add(cfqd, cfqq);
2469
5044eed4 2470 BUG_ON(!cfqq->next_rq);
1da177e4
LT
2471}
2472
febffd61 2473static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1da177e4 2474{
5380a101
JA
2475 elv_rb_del(&cfqq->sort_list, rq);
2476 cfqq->queued[rq_is_sync(rq)]--;
63a4cc24 2477 cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags);
5e705374 2478 cfq_add_rq_rb(rq);
155fead9 2479 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
63a4cc24 2480 req_op(rq), rq->cmd_flags);
1da177e4
LT
2481}
2482
206dc69b
JA
2483static struct request *
2484cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1da177e4 2485{
206dc69b 2486 struct task_struct *tsk = current;
c5869807 2487 struct cfq_io_cq *cic;
206dc69b 2488 struct cfq_queue *cfqq;
1da177e4 2489
4ac845a2 2490 cic = cfq_cic_lookup(cfqd, tsk->io_context);
91fac317
VT
2491 if (!cic)
2492 return NULL;
2493
2494 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
f73a1c7d
KO
2495 if (cfqq)
2496 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
1da177e4 2497
1da177e4
LT
2498 return NULL;
2499}
2500
165125e1 2501static void cfq_activate_request(struct request_queue *q, struct request *rq)
1da177e4 2502{
22e2c507 2503 struct cfq_data *cfqd = q->elevator->elevator_data;
3b18152c 2504
53c583d2 2505 cfqd->rq_in_driver++;
7b679138 2506 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
53c583d2 2507 cfqd->rq_in_driver);
25776e35 2508
5b93629b 2509 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1da177e4
LT
2510}
2511
165125e1 2512static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1da177e4 2513{
b4878f24
JA
2514 struct cfq_data *cfqd = q->elevator->elevator_data;
2515
53c583d2
CZ
2516 WARN_ON(!cfqd->rq_in_driver);
2517 cfqd->rq_in_driver--;
7b679138 2518 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
53c583d2 2519 cfqd->rq_in_driver);
1da177e4
LT
2520}
2521
b4878f24 2522static void cfq_remove_request(struct request *rq)
1da177e4 2523{
5e705374 2524 struct cfq_queue *cfqq = RQ_CFQQ(rq);
21183b07 2525
5e705374
JA
2526 if (cfqq->next_rq == rq)
2527 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1da177e4 2528
b4878f24 2529 list_del_init(&rq->queuelist);
5e705374 2530 cfq_del_rq_rb(rq);
374f84ac 2531
45333d5a 2532 cfqq->cfqd->rq_queued--;
63a4cc24 2533 cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags);
65299a3b
CH
2534 if (rq->cmd_flags & REQ_PRIO) {
2535 WARN_ON(!cfqq->prio_pending);
2536 cfqq->prio_pending--;
b53d1ed7 2537 }
1da177e4
LT
2538}
2539
165125e1
JA
2540static int cfq_merge(struct request_queue *q, struct request **req,
2541 struct bio *bio)
1da177e4
LT
2542{
2543 struct cfq_data *cfqd = q->elevator->elevator_data;
2544 struct request *__rq;
1da177e4 2545
206dc69b 2546 __rq = cfq_find_rq_fmerge(cfqd, bio);
72ef799b 2547 if (__rq && elv_bio_merge_ok(__rq, bio)) {
9817064b
JA
2548 *req = __rq;
2549 return ELEVATOR_FRONT_MERGE;
1da177e4
LT
2550 }
2551
2552 return ELEVATOR_NO_MERGE;
1da177e4
LT
2553}
2554
165125e1 2555static void cfq_merged_request(struct request_queue *q, struct request *req,
21183b07 2556 int type)
1da177e4 2557{
21183b07 2558 if (type == ELEVATOR_FRONT_MERGE) {
5e705374 2559 struct cfq_queue *cfqq = RQ_CFQQ(req);
1da177e4 2560
5e705374 2561 cfq_reposition_rq_rb(cfqq, req);
1da177e4 2562 }
1da177e4
LT
2563}
2564
812d4026
DS
2565static void cfq_bio_merged(struct request_queue *q, struct request *req,
2566 struct bio *bio)
2567{
1eff9d32 2568 cfqg_stats_update_io_merged(RQ_CFQG(req), bio_op(bio), bio->bi_opf);
812d4026
DS
2569}
2570
1da177e4 2571static void
165125e1 2572cfq_merged_requests(struct request_queue *q, struct request *rq,
1da177e4
LT
2573 struct request *next)
2574{
cf7c25cf 2575 struct cfq_queue *cfqq = RQ_CFQQ(rq);
4a0b75c7
SL
2576 struct cfq_data *cfqd = q->elevator->elevator_data;
2577
22e2c507
JA
2578 /*
2579 * reposition in fifo if next is older than rq
2580 */
2581 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
9a7f38c4 2582 next->fifo_time < rq->fifo_time &&
3d106fba 2583 cfqq == RQ_CFQQ(next)) {
22e2c507 2584 list_move(&rq->queuelist, &next->queuelist);
8b4922d3 2585 rq->fifo_time = next->fifo_time;
30996f40 2586 }
22e2c507 2587
cf7c25cf
CZ
2588 if (cfqq->next_rq == next)
2589 cfqq->next_rq = rq;
b4878f24 2590 cfq_remove_request(next);
63a4cc24 2591 cfqg_stats_update_io_merged(RQ_CFQG(rq), req_op(next), next->cmd_flags);
4a0b75c7
SL
2592
2593 cfqq = RQ_CFQQ(next);
2594 /*
2595 * all requests of this queue are merged to other queues, delete it
2596 * from the service tree. If it's the active_queue,
2597 * cfq_dispatch_requests() will choose to expire it or do idle
2598 */
2599 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2600 cfqq != cfqd->active_queue)
2601 cfq_del_cfqq_rr(cfqd, cfqq);
22e2c507
JA
2602}
2603
72ef799b
TE
2604static int cfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2605 struct bio *bio)
da775265
JA
2606{
2607 struct cfq_data *cfqd = q->elevator->elevator_data;
c5869807 2608 struct cfq_io_cq *cic;
da775265 2609 struct cfq_queue *cfqq;
da775265
JA
2610
2611 /*
ec8acb69 2612 * Disallow merge of a sync bio into an async request.
da775265 2613 */
91fac317 2614 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
a6151c3a 2615 return false;
da775265
JA
2616
2617 /*
f1a4f4d3 2618 * Lookup the cfqq that this bio will be queued with and allow
07c2bd37 2619 * merge only if rq is queued there.
f1a4f4d3 2620 */
07c2bd37
TH
2621 cic = cfq_cic_lookup(cfqd, current->io_context);
2622 if (!cic)
2623 return false;
719d3402 2624
91fac317 2625 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
a6151c3a 2626 return cfqq == RQ_CFQQ(rq);
da775265
JA
2627}
2628
72ef799b
TE
2629static int cfq_allow_rq_merge(struct request_queue *q, struct request *rq,
2630 struct request *next)
2631{
2632 return RQ_CFQQ(rq) == RQ_CFQQ(next);
2633}
2634
812df48d
DS
2635static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2636{
91148325 2637 hrtimer_try_to_cancel(&cfqd->idle_slice_timer);
155fead9 2638 cfqg_stats_update_idle_time(cfqq->cfqg);
812df48d
DS
2639}
2640
febffd61
JA
2641static void __cfq_set_active_queue(struct cfq_data *cfqd,
2642 struct cfq_queue *cfqq)
22e2c507
JA
2643{
2644 if (cfqq) {
3bf10fea 2645 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
4d2ceea4 2646 cfqd->serving_wl_class, cfqd->serving_wl_type);
155fead9 2647 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
62a37f6b 2648 cfqq->slice_start = 0;
9a7f38c4 2649 cfqq->dispatch_start = ktime_get_ns();
62a37f6b
JT
2650 cfqq->allocated_slice = 0;
2651 cfqq->slice_end = 0;
2652 cfqq->slice_dispatch = 0;
2653 cfqq->nr_sectors = 0;
2654
2655 cfq_clear_cfqq_wait_request(cfqq);
2656 cfq_clear_cfqq_must_dispatch(cfqq);
2657 cfq_clear_cfqq_must_alloc_slice(cfqq);
2658 cfq_clear_cfqq_fifo_expire(cfqq);
2659 cfq_mark_cfqq_slice_new(cfqq);
2660
2661 cfq_del_timer(cfqd, cfqq);
22e2c507
JA
2662 }
2663
2664 cfqd->active_queue = cfqq;
2665}
2666
7b14e3b5
JA
2667/*
2668 * current cfqq expired its slice (or was too idle), select new one
2669 */
2670static void
2671__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
e5ff082e 2672 bool timed_out)
7b14e3b5 2673{
7b679138
JA
2674 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2675
7b14e3b5 2676 if (cfq_cfqq_wait_request(cfqq))
812df48d 2677 cfq_del_timer(cfqd, cfqq);
7b14e3b5 2678
7b14e3b5 2679 cfq_clear_cfqq_wait_request(cfqq);
f75edf2d 2680 cfq_clear_cfqq_wait_busy(cfqq);
7b14e3b5 2681
ae54abed
SL
2682 /*
2683 * If this cfqq is shared between multiple processes, check to
2684 * make sure that those processes are still issuing I/Os within
2685 * the mean seek distance. If not, it may be time to break the
2686 * queues apart again.
2687 */
2688 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2689 cfq_mark_cfqq_split_coop(cfqq);
2690
7b14e3b5 2691 /*
6084cdda 2692 * store what was left of this slice, if the queue idled/timed out
7b14e3b5 2693 */
c553f8e3
SL
2694 if (timed_out) {
2695 if (cfq_cfqq_slice_new(cfqq))
ba5bd520 2696 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
c553f8e3 2697 else
9a7f38c4 2698 cfqq->slice_resid = cfqq->slice_end - ktime_get_ns();
93fdf147 2699 cfq_log_cfqq(cfqd, cfqq, "resid=%lld", cfqq->slice_resid);
7b679138 2700 }
7b14e3b5 2701
e5ff082e 2702 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
dae739eb 2703
f04a6424
VG
2704 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2705 cfq_del_cfqq_rr(cfqd, cfqq);
2706
edd75ffd 2707 cfq_resort_rr_list(cfqd, cfqq);
7b14e3b5
JA
2708
2709 if (cfqq == cfqd->active_queue)
2710 cfqd->active_queue = NULL;
2711
2712 if (cfqd->active_cic) {
11a3122f 2713 put_io_context(cfqd->active_cic->icq.ioc);
7b14e3b5
JA
2714 cfqd->active_cic = NULL;
2715 }
7b14e3b5
JA
2716}
2717
e5ff082e 2718static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
7b14e3b5
JA
2719{
2720 struct cfq_queue *cfqq = cfqd->active_queue;
2721
2722 if (cfqq)
e5ff082e 2723 __cfq_slice_expired(cfqd, cfqq, timed_out);
7b14e3b5
JA
2724}
2725
498d3aa2
JA
2726/*
2727 * Get next queue for service. Unless we have a queue preemption,
2728 * we'll simply select the first cfqq in the service tree.
2729 */
6d048f53 2730static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
22e2c507 2731{
34b98d03
VG
2732 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2733 cfqd->serving_wl_class, cfqd->serving_wl_type);
d9e7620e 2734
f04a6424
VG
2735 if (!cfqd->rq_queued)
2736 return NULL;
2737
1fa8f6d6 2738 /* There is nothing to dispatch */
34b98d03 2739 if (!st)
1fa8f6d6 2740 return NULL;
34b98d03 2741 if (RB_EMPTY_ROOT(&st->rb))
c0324a02 2742 return NULL;
34b98d03 2743 return cfq_rb_first(st);
6d048f53
JA
2744}
2745
f04a6424
VG
2746static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2747{
25fb5169 2748 struct cfq_group *cfqg;
f04a6424
VG
2749 struct cfq_queue *cfqq;
2750 int i, j;
2751 struct cfq_rb_root *st;
2752
2753 if (!cfqd->rq_queued)
2754 return NULL;
2755
25fb5169
VG
2756 cfqg = cfq_get_next_cfqg(cfqd);
2757 if (!cfqg)
2758 return NULL;
2759
f04a6424
VG
2760 for_each_cfqg_st(cfqg, i, j, st)
2761 if ((cfqq = cfq_rb_first(st)) != NULL)
2762 return cfqq;
2763 return NULL;
2764}
2765
498d3aa2
JA
2766/*
2767 * Get and set a new active queue for service.
2768 */
a36e71f9
JA
2769static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2770 struct cfq_queue *cfqq)
6d048f53 2771{
e00ef799 2772 if (!cfqq)
a36e71f9 2773 cfqq = cfq_get_next_queue(cfqd);
6d048f53 2774
22e2c507 2775 __cfq_set_active_queue(cfqd, cfqq);
3b18152c 2776 return cfqq;
22e2c507
JA
2777}
2778
d9e7620e
JA
2779static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2780 struct request *rq)
2781{
83096ebf
TH
2782 if (blk_rq_pos(rq) >= cfqd->last_position)
2783 return blk_rq_pos(rq) - cfqd->last_position;
d9e7620e 2784 else
83096ebf 2785 return cfqd->last_position - blk_rq_pos(rq);
d9e7620e
JA
2786}
2787
b2c18e1e 2788static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
e9ce335d 2789 struct request *rq)
6d048f53 2790{
e9ce335d 2791 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
6d048f53
JA
2792}
2793
a36e71f9
JA
2794static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2795 struct cfq_queue *cur_cfqq)
2796{
f2d1f0ae 2797 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
a36e71f9
JA
2798 struct rb_node *parent, *node;
2799 struct cfq_queue *__cfqq;
2800 sector_t sector = cfqd->last_position;
2801
2802 if (RB_EMPTY_ROOT(root))
2803 return NULL;
2804
2805 /*
2806 * First, if we find a request starting at the end of the last
2807 * request, choose it.
2808 */
f2d1f0ae 2809 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
a36e71f9
JA
2810 if (__cfqq)
2811 return __cfqq;
2812
2813 /*
2814 * If the exact sector wasn't found, the parent of the NULL leaf
2815 * will contain the closest sector.
2816 */
2817 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
e9ce335d 2818 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
a36e71f9
JA
2819 return __cfqq;
2820
2e46e8b2 2821 if (blk_rq_pos(__cfqq->next_rq) < sector)
a36e71f9
JA
2822 node = rb_next(&__cfqq->p_node);
2823 else
2824 node = rb_prev(&__cfqq->p_node);
2825 if (!node)
2826 return NULL;
2827
2828 __cfqq = rb_entry(node, struct cfq_queue, p_node);
e9ce335d 2829 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
a36e71f9
JA
2830 return __cfqq;
2831
2832 return NULL;
2833}
2834
2835/*
2836 * cfqd - obvious
2837 * cur_cfqq - passed in so that we don't decide that the current queue is
2838 * closely cooperating with itself.
2839 *
2840 * So, basically we're assuming that that cur_cfqq has dispatched at least
2841 * one request, and that cfqd->last_position reflects a position on the disk
2842 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2843 * assumption.
2844 */
2845static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
b3b6d040 2846 struct cfq_queue *cur_cfqq)
6d048f53 2847{
a36e71f9
JA
2848 struct cfq_queue *cfqq;
2849
39c01b21
DS
2850 if (cfq_class_idle(cur_cfqq))
2851 return NULL;
e6c5bc73
JM
2852 if (!cfq_cfqq_sync(cur_cfqq))
2853 return NULL;
2854 if (CFQQ_SEEKY(cur_cfqq))
2855 return NULL;
2856
b9d8f4c7
GJ
2857 /*
2858 * Don't search priority tree if it's the only queue in the group.
2859 */
2860 if (cur_cfqq->cfqg->nr_cfqq == 1)
2861 return NULL;
2862
6d048f53 2863 /*
d9e7620e
JA
2864 * We should notice if some of the queues are cooperating, eg
2865 * working closely on the same area of the disk. In that case,
2866 * we can group them together and don't waste time idling.
6d048f53 2867 */
a36e71f9
JA
2868 cfqq = cfqq_close(cfqd, cur_cfqq);
2869 if (!cfqq)
2870 return NULL;
2871
8682e1f1
VG
2872 /* If new queue belongs to different cfq_group, don't choose it */
2873 if (cur_cfqq->cfqg != cfqq->cfqg)
2874 return NULL;
2875
df5fe3e8
JM
2876 /*
2877 * It only makes sense to merge sync queues.
2878 */
2879 if (!cfq_cfqq_sync(cfqq))
2880 return NULL;
e6c5bc73
JM
2881 if (CFQQ_SEEKY(cfqq))
2882 return NULL;
df5fe3e8 2883
c0324a02
CZ
2884 /*
2885 * Do not merge queues of different priority classes
2886 */
2887 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2888 return NULL;
2889
a36e71f9 2890 return cfqq;
6d048f53
JA
2891}
2892
a6d44e98
CZ
2893/*
2894 * Determine whether we should enforce idle window for this queue.
2895 */
2896
2897static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2898{
3bf10fea 2899 enum wl_class_t wl_class = cfqq_class(cfqq);
34b98d03 2900 struct cfq_rb_root *st = cfqq->service_tree;
a6d44e98 2901
34b98d03
VG
2902 BUG_ON(!st);
2903 BUG_ON(!st->count);
f04a6424 2904
b6508c16
VG
2905 if (!cfqd->cfq_slice_idle)
2906 return false;
2907
a6d44e98 2908 /* We never do for idle class queues. */
3bf10fea 2909 if (wl_class == IDLE_WORKLOAD)
a6d44e98
CZ
2910 return false;
2911
2912 /* We do for queues that were marked with idle window flag. */
3c764b7a
SL
2913 if (cfq_cfqq_idle_window(cfqq) &&
2914 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
a6d44e98
CZ
2915 return true;
2916
2917 /*
2918 * Otherwise, we do only if they are the last ones
2919 * in their service tree.
2920 */
34b98d03
VG
2921 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2922 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
c1e44756 2923 return true;
34b98d03 2924 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
c1e44756 2925 return false;
a6d44e98
CZ
2926}
2927
6d048f53 2928static void cfq_arm_slice_timer(struct cfq_data *cfqd)
22e2c507 2929{
1792669c 2930 struct cfq_queue *cfqq = cfqd->active_queue;
e795421e 2931 struct cfq_rb_root *st = cfqq->service_tree;
c5869807 2932 struct cfq_io_cq *cic;
9a7f38c4
JM
2933 u64 sl, group_idle = 0;
2934 u64 now = ktime_get_ns();
7b14e3b5 2935
a68bbddb 2936 /*
f7d7b7a7
JA
2937 * SSD device without seek penalty, disable idling. But only do so
2938 * for devices that support queuing, otherwise we still have a problem
2939 * with sync vs async workloads.
a68bbddb 2940 */
f7d7b7a7 2941 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
a68bbddb
JA
2942 return;
2943
dd67d051 2944 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
6d048f53 2945 WARN_ON(cfq_cfqq_slice_new(cfqq));
22e2c507
JA
2946
2947 /*
2948 * idle is disabled, either manually or by past process history
2949 */
80bdf0c7
VG
2950 if (!cfq_should_idle(cfqd, cfqq)) {
2951 /* no queue idling. Check for group idling */
2952 if (cfqd->cfq_group_idle)
2953 group_idle = cfqd->cfq_group_idle;
2954 else
2955 return;
2956 }
6d048f53 2957
7b679138 2958 /*
8e550632 2959 * still active requests from this queue, don't idle
7b679138 2960 */
8e550632 2961 if (cfqq->dispatched)
7b679138
JA
2962 return;
2963
22e2c507
JA
2964 /*
2965 * task has exited, don't wait
2966 */
206dc69b 2967 cic = cfqd->active_cic;
f6e8d01b 2968 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
6d048f53
JA
2969 return;
2970
355b659c
CZ
2971 /*
2972 * If our average think time is larger than the remaining time
2973 * slice, then don't idle. This avoids overrunning the allotted
2974 * time slice.
2975 */
383cd721 2976 if (sample_valid(cic->ttime.ttime_samples) &&
9a7f38c4
JM
2977 (cfqq->slice_end - now < cic->ttime.ttime_mean)) {
2978 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%llu",
383cd721 2979 cic->ttime.ttime_mean);
355b659c 2980 return;
b1ffe737 2981 }
355b659c 2982
e795421e
JK
2983 /*
2984 * There are other queues in the group or this is the only group and
2985 * it has too big thinktime, don't do group idle.
2986 */
2987 if (group_idle &&
2988 (cfqq->cfqg->nr_cfqq > 1 ||
2989 cfq_io_thinktime_big(cfqd, &st->ttime, true)))
80bdf0c7
VG
2990 return;
2991
3b18152c 2992 cfq_mark_cfqq_wait_request(cfqq);
22e2c507 2993
80bdf0c7
VG
2994 if (group_idle)
2995 sl = cfqd->cfq_group_idle;
2996 else
2997 sl = cfqd->cfq_slice_idle;
206dc69b 2998
91148325
JK
2999 hrtimer_start(&cfqd->idle_slice_timer, ns_to_ktime(sl),
3000 HRTIMER_MODE_REL);
155fead9 3001 cfqg_stats_set_start_idle_time(cfqq->cfqg);
9a7f38c4 3002 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %llu group_idle: %d", sl,
80bdf0c7 3003 group_idle ? 1 : 0);
1da177e4
LT
3004}
3005
498d3aa2
JA
3006/*
3007 * Move request from internal lists to the request queue dispatch list.
3008 */
165125e1 3009static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1da177e4 3010{
3ed9a296 3011 struct cfq_data *cfqd = q->elevator->elevator_data;
5e705374 3012 struct cfq_queue *cfqq = RQ_CFQQ(rq);
22e2c507 3013
7b679138
JA
3014 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
3015
06d21886 3016 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
5380a101 3017 cfq_remove_request(rq);
6d048f53 3018 cfqq->dispatched++;
80bdf0c7 3019 (RQ_CFQG(rq))->dispatched++;
5380a101 3020 elv_dispatch_sort(q, rq);
3ed9a296 3021
53c583d2 3022 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
c4e7893e 3023 cfqq->nr_sectors += blk_rq_sectors(rq);
1da177e4
LT
3024}
3025
3026/*
3027 * return expired entry, or NULL to just start from scratch in rbtree
3028 */
febffd61 3029static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1da177e4 3030{
30996f40 3031 struct request *rq = NULL;
1da177e4 3032
3b18152c 3033 if (cfq_cfqq_fifo_expire(cfqq))
1da177e4 3034 return NULL;
cb887411
JA
3035
3036 cfq_mark_cfqq_fifo_expire(cfqq);
3037
89850f7e
JA
3038 if (list_empty(&cfqq->fifo))
3039 return NULL;
1da177e4 3040
89850f7e 3041 rq = rq_entry_fifo(cfqq->fifo.next);
9a7f38c4 3042 if (ktime_get_ns() < rq->fifo_time)
7b679138 3043 rq = NULL;
1da177e4 3044
6d048f53 3045 return rq;
1da177e4
LT
3046}
3047
22e2c507
JA
3048static inline int
3049cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3050{
3051 const int base_rq = cfqd->cfq_slice_async_rq;
1da177e4 3052
22e2c507 3053 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1da177e4 3054
b9f8ce05 3055 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
1da177e4
LT
3056}
3057
df5fe3e8
JM
3058/*
3059 * Must be called with the queue_lock held.
3060 */
3061static int cfqq_process_refs(struct cfq_queue *cfqq)
3062{
3063 int process_refs, io_refs;
3064
3065 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
30d7b944 3066 process_refs = cfqq->ref - io_refs;
df5fe3e8
JM
3067 BUG_ON(process_refs < 0);
3068 return process_refs;
3069}
3070
3071static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3072{
e6c5bc73 3073 int process_refs, new_process_refs;
df5fe3e8
JM
3074 struct cfq_queue *__cfqq;
3075
c10b61f0
JM
3076 /*
3077 * If there are no process references on the new_cfqq, then it is
3078 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3079 * chain may have dropped their last reference (not just their
3080 * last process reference).
3081 */
3082 if (!cfqq_process_refs(new_cfqq))
3083 return;
3084
df5fe3e8
JM
3085 /* Avoid a circular list and skip interim queue merges */
3086 while ((__cfqq = new_cfqq->new_cfqq)) {
3087 if (__cfqq == cfqq)
3088 return;
3089 new_cfqq = __cfqq;
3090 }
3091
3092 process_refs = cfqq_process_refs(cfqq);
c10b61f0 3093 new_process_refs = cfqq_process_refs(new_cfqq);
df5fe3e8
JM
3094 /*
3095 * If the process for the cfqq has gone away, there is no
3096 * sense in merging the queues.
3097 */
c10b61f0 3098 if (process_refs == 0 || new_process_refs == 0)
df5fe3e8
JM
3099 return;
3100
e6c5bc73
JM
3101 /*
3102 * Merge in the direction of the lesser amount of work.
3103 */
e6c5bc73
JM
3104 if (new_process_refs >= process_refs) {
3105 cfqq->new_cfqq = new_cfqq;
30d7b944 3106 new_cfqq->ref += process_refs;
e6c5bc73
JM
3107 } else {
3108 new_cfqq->new_cfqq = cfqq;
30d7b944 3109 cfqq->ref += new_process_refs;
e6c5bc73 3110 }
df5fe3e8
JM
3111}
3112
6d816ec7 3113static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3bf10fea 3114 struct cfq_group *cfqg, enum wl_class_t wl_class)
718eee05
CZ
3115{
3116 struct cfq_queue *queue;
3117 int i;
3118 bool key_valid = false;
9a7f38c4 3119 u64 lowest_key = 0;
718eee05
CZ
3120 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3121
65b32a57
VG
3122 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3123 /* select the one with lowest rb_key */
34b98d03 3124 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
718eee05 3125 if (queue &&
9a7f38c4 3126 (!key_valid || queue->rb_key < lowest_key)) {
718eee05
CZ
3127 lowest_key = queue->rb_key;
3128 cur_best = i;
3129 key_valid = true;
3130 }
3131 }
3132
3133 return cur_best;
3134}
3135
6d816ec7
VG
3136static void
3137choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
718eee05 3138{
9a7f38c4 3139 u64 slice;
718eee05 3140 unsigned count;
cdb16e8f 3141 struct cfq_rb_root *st;
9a7f38c4 3142 u64 group_slice;
4d2ceea4 3143 enum wl_class_t original_class = cfqd->serving_wl_class;
9a7f38c4 3144 u64 now = ktime_get_ns();
1fa8f6d6 3145
718eee05 3146 /* Choose next priority. RT > BE > IDLE */
58ff82f3 3147 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
4d2ceea4 3148 cfqd->serving_wl_class = RT_WORKLOAD;
58ff82f3 3149 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
4d2ceea4 3150 cfqd->serving_wl_class = BE_WORKLOAD;
718eee05 3151 else {
4d2ceea4 3152 cfqd->serving_wl_class = IDLE_WORKLOAD;
9a7f38c4 3153 cfqd->workload_expires = now + jiffies_to_nsecs(1);
718eee05
CZ
3154 return;
3155 }
3156
4d2ceea4 3157 if (original_class != cfqd->serving_wl_class)
e4ea0c16
SL
3158 goto new_workload;
3159
718eee05
CZ
3160 /*
3161 * For RT and BE, we have to choose also the type
3162 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3163 * expiration time
3164 */
34b98d03 3165 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
cdb16e8f 3166 count = st->count;
718eee05
CZ
3167
3168 /*
65b32a57 3169 * check workload expiration, and that we still have other queues ready
718eee05 3170 */
9a7f38c4 3171 if (count && !(now > cfqd->workload_expires))
718eee05
CZ
3172 return;
3173
e4ea0c16 3174new_workload:
718eee05 3175 /* otherwise select new workload type */
6d816ec7 3176 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
4d2ceea4 3177 cfqd->serving_wl_class);
34b98d03 3178 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
cdb16e8f 3179 count = st->count;
718eee05
CZ
3180
3181 /*
3182 * the workload slice is computed as a fraction of target latency
3183 * proportional to the number of queues in that workload, over
3184 * all the queues in the same priority class
3185 */
58ff82f3
VG
3186 group_slice = cfq_group_slice(cfqd, cfqg);
3187
9a7f38c4 3188 slice = div_u64(group_slice * count,
4d2ceea4
VG
3189 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3190 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
9a7f38c4 3191 cfqg)));
718eee05 3192
4d2ceea4 3193 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
9a7f38c4 3194 u64 tmp;
f26bd1f0
VG
3195
3196 /*
3197 * Async queues are currently system wide. Just taking
3198 * proportion of queues with-in same group will lead to higher
3199 * async ratio system wide as generally root group is going
3200 * to have higher weight. A more accurate thing would be to
3201 * calculate system wide asnc/sync ratio.
3202 */
5bf14c07
TM
3203 tmp = cfqd->cfq_target_latency *
3204 cfqg_busy_async_queues(cfqd, cfqg);
9a7f38c4
JM
3205 tmp = div_u64(tmp, cfqd->busy_queues);
3206 slice = min_t(u64, slice, tmp);
f26bd1f0 3207
718eee05
CZ
3208 /* async workload slice is scaled down according to
3209 * the sync/async slice ratio. */
9a7f38c4 3210 slice = div64_u64(slice*cfqd->cfq_slice[0], cfqd->cfq_slice[1]);
f26bd1f0 3211 } else
718eee05
CZ
3212 /* sync workload slice is at least 2 * cfq_slice_idle */
3213 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3214
9a7f38c4
JM
3215 slice = max_t(u64, slice, CFQ_MIN_TT);
3216 cfq_log(cfqd, "workload slice:%llu", slice);
3217 cfqd->workload_expires = now + slice;
718eee05
CZ
3218}
3219
1fa8f6d6
VG
3220static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3221{
3222 struct cfq_rb_root *st = &cfqd->grp_service_tree;
25bc6b07 3223 struct cfq_group *cfqg;
1fa8f6d6
VG
3224
3225 if (RB_EMPTY_ROOT(&st->rb))
3226 return NULL;
25bc6b07 3227 cfqg = cfq_rb_first_group(st);
25bc6b07
VG
3228 update_min_vdisktime(st);
3229 return cfqg;
1fa8f6d6
VG
3230}
3231
cdb16e8f
VG
3232static void cfq_choose_cfqg(struct cfq_data *cfqd)
3233{
1fa8f6d6 3234 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
9a7f38c4 3235 u64 now = ktime_get_ns();
1fa8f6d6
VG
3236
3237 cfqd->serving_group = cfqg;
dae739eb
VG
3238
3239 /* Restore the workload type data */
4d2ceea4 3240 if (cfqg->saved_wl_slice) {
9a7f38c4 3241 cfqd->workload_expires = now + cfqg->saved_wl_slice;
4d2ceea4
VG
3242 cfqd->serving_wl_type = cfqg->saved_wl_type;
3243 cfqd->serving_wl_class = cfqg->saved_wl_class;
66ae2919 3244 } else
9a7f38c4 3245 cfqd->workload_expires = now - 1;
66ae2919 3246
6d816ec7 3247 choose_wl_class_and_type(cfqd, cfqg);
cdb16e8f
VG
3248}
3249
22e2c507 3250/*
498d3aa2
JA
3251 * Select a queue for service. If we have a current active queue,
3252 * check whether to continue servicing it, or retrieve and set a new one.
22e2c507 3253 */
1b5ed5e1 3254static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1da177e4 3255{
a36e71f9 3256 struct cfq_queue *cfqq, *new_cfqq = NULL;
9a7f38c4 3257 u64 now = ktime_get_ns();
1da177e4 3258
22e2c507
JA
3259 cfqq = cfqd->active_queue;
3260 if (!cfqq)
3261 goto new_queue;
1da177e4 3262
f04a6424
VG
3263 if (!cfqd->rq_queued)
3264 return NULL;
c244bb50
VG
3265
3266 /*
3267 * We were waiting for group to get backlogged. Expire the queue
3268 */
3269 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3270 goto expire;
3271
22e2c507 3272 /*
6d048f53 3273 * The active queue has run out of time, expire it and select new.
22e2c507 3274 */
7667aa06
VG
3275 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3276 /*
3277 * If slice had not expired at the completion of last request
3278 * we might not have turned on wait_busy flag. Don't expire
3279 * the queue yet. Allow the group to get backlogged.
3280 *
3281 * The very fact that we have used the slice, that means we
3282 * have been idling all along on this queue and it should be
3283 * ok to wait for this request to complete.
3284 */
82bbbf28
VG
3285 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3286 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3287 cfqq = NULL;
7667aa06 3288 goto keep_queue;
82bbbf28 3289 } else
80bdf0c7 3290 goto check_group_idle;
7667aa06 3291 }
1da177e4 3292
22e2c507 3293 /*
6d048f53
JA
3294 * The active queue has requests and isn't expired, allow it to
3295 * dispatch.
22e2c507 3296 */
dd67d051 3297 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
22e2c507 3298 goto keep_queue;
6d048f53 3299
a36e71f9
JA
3300 /*
3301 * If another queue has a request waiting within our mean seek
3302 * distance, let it run. The expire code will check for close
3303 * cooperators and put the close queue at the front of the service
df5fe3e8 3304 * tree. If possible, merge the expiring queue with the new cfqq.
a36e71f9 3305 */
b3b6d040 3306 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
df5fe3e8
JM
3307 if (new_cfqq) {
3308 if (!cfqq->new_cfqq)
3309 cfq_setup_merge(cfqq, new_cfqq);
a36e71f9 3310 goto expire;
df5fe3e8 3311 }
a36e71f9 3312
6d048f53
JA
3313 /*
3314 * No requests pending. If the active queue still has requests in
3315 * flight or is idling for a new request, allow either of these
3316 * conditions to happen (or time out) before selecting a new queue.
3317 */
91148325 3318 if (hrtimer_active(&cfqd->idle_slice_timer)) {
80bdf0c7
VG
3319 cfqq = NULL;
3320 goto keep_queue;
3321 }
3322
8e1ac665
SL
3323 /*
3324 * This is a deep seek queue, but the device is much faster than
3325 * the queue can deliver, don't idle
3326 **/
3327 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3328 (cfq_cfqq_slice_new(cfqq) ||
9a7f38c4 3329 (cfqq->slice_end - now > now - cfqq->slice_start))) {
8e1ac665
SL
3330 cfq_clear_cfqq_deep(cfqq);
3331 cfq_clear_cfqq_idle_window(cfqq);
3332 }
3333
80bdf0c7
VG
3334 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3335 cfqq = NULL;
3336 goto keep_queue;
3337 }
3338
3339 /*
3340 * If group idle is enabled and there are requests dispatched from
3341 * this group, wait for requests to complete.
3342 */
3343check_group_idle:
7700fc4f
SL
3344 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3345 cfqq->cfqg->dispatched &&
3346 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
caaa5f9f
JA
3347 cfqq = NULL;
3348 goto keep_queue;
22e2c507
JA
3349 }
3350
3b18152c 3351expire:
e5ff082e 3352 cfq_slice_expired(cfqd, 0);
3b18152c 3353new_queue:
718eee05
CZ
3354 /*
3355 * Current queue expired. Check if we have to switch to a new
3356 * service tree
3357 */
3358 if (!new_cfqq)
cdb16e8f 3359 cfq_choose_cfqg(cfqd);
718eee05 3360
a36e71f9 3361 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
22e2c507 3362keep_queue:
3b18152c 3363 return cfqq;
22e2c507
JA
3364}
3365
febffd61 3366static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
d9e7620e
JA
3367{
3368 int dispatched = 0;
3369
3370 while (cfqq->next_rq) {
3371 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3372 dispatched++;
3373 }
3374
3375 BUG_ON(!list_empty(&cfqq->fifo));
f04a6424
VG
3376
3377 /* By default cfqq is not expired if it is empty. Do it explicitly */
e5ff082e 3378 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
d9e7620e
JA
3379 return dispatched;
3380}
3381
498d3aa2
JA
3382/*
3383 * Drain our current requests. Used for barriers and when switching
3384 * io schedulers on-the-fly.
3385 */
d9e7620e 3386static int cfq_forced_dispatch(struct cfq_data *cfqd)
1b5ed5e1 3387{
0871714e 3388 struct cfq_queue *cfqq;
d9e7620e 3389 int dispatched = 0;
cdb16e8f 3390
3440c49f 3391 /* Expire the timeslice of the current active queue first */
e5ff082e 3392 cfq_slice_expired(cfqd, 0);
3440c49f
DS
3393 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3394 __cfq_set_active_queue(cfqd, cfqq);
f04a6424 3395 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3440c49f 3396 }
1b5ed5e1 3397
1b5ed5e1
TH
3398 BUG_ON(cfqd->busy_queues);
3399
6923715a 3400 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1b5ed5e1
TH
3401 return dispatched;
3402}
3403
abc3c744
SL
3404static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3405 struct cfq_queue *cfqq)
3406{
9a7f38c4
JM
3407 u64 now = ktime_get_ns();
3408
abc3c744
SL
3409 /* the queue hasn't finished any request, can't estimate */
3410 if (cfq_cfqq_slice_new(cfqq))
c1e44756 3411 return true;
9a7f38c4 3412 if (now + cfqd->cfq_slice_idle * cfqq->dispatched > cfqq->slice_end)
c1e44756 3413 return true;
abc3c744 3414
c1e44756 3415 return false;
abc3c744
SL
3416}
3417
0b182d61 3418static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2f5cb738 3419{
2f5cb738 3420 unsigned int max_dispatch;
22e2c507 3421
3932a86b
GC
3422 if (cfq_cfqq_must_dispatch(cfqq))
3423 return true;
3424
5ad531db
JA
3425 /*
3426 * Drain async requests before we start sync IO
3427 */
53c583d2 3428 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
0b182d61 3429 return false;
5ad531db 3430
2f5cb738
JA
3431 /*
3432 * If this is an async queue and we have sync IO in flight, let it wait
3433 */
53c583d2 3434 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
0b182d61 3435 return false;
2f5cb738 3436
abc3c744 3437 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2f5cb738
JA
3438 if (cfq_class_idle(cfqq))
3439 max_dispatch = 1;
b4878f24 3440
2f5cb738
JA
3441 /*
3442 * Does this cfqq already have too much IO in flight?
3443 */
3444 if (cfqq->dispatched >= max_dispatch) {
ef8a41df 3445 bool promote_sync = false;
2f5cb738
JA
3446 /*
3447 * idle queue must always only have a single IO in flight
3448 */
3ed9a296 3449 if (cfq_class_idle(cfqq))
0b182d61 3450 return false;
3ed9a296 3451
ef8a41df 3452 /*
c4ade94f
LS
3453 * If there is only one sync queue
3454 * we can ignore async queue here and give the sync
ef8a41df
SL
3455 * queue no dispatch limit. The reason is a sync queue can
3456 * preempt async queue, limiting the sync queue doesn't make
3457 * sense. This is useful for aiostress test.
3458 */
c4ade94f
LS
3459 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3460 promote_sync = true;
ef8a41df 3461
2f5cb738
JA
3462 /*
3463 * We have other queues, don't allow more IO from this one
3464 */
ef8a41df
SL
3465 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3466 !promote_sync)
0b182d61 3467 return false;
9ede209e 3468
365722bb 3469 /*
474b18cc 3470 * Sole queue user, no limit
365722bb 3471 */
ef8a41df 3472 if (cfqd->busy_queues == 1 || promote_sync)
abc3c744
SL
3473 max_dispatch = -1;
3474 else
3475 /*
3476 * Normally we start throttling cfqq when cfq_quantum/2
3477 * requests have been dispatched. But we can drive
3478 * deeper queue depths at the beginning of slice
3479 * subjected to upper limit of cfq_quantum.
3480 * */
3481 max_dispatch = cfqd->cfq_quantum;
8e296755
JA
3482 }
3483
3484 /*
3485 * Async queues must wait a bit before being allowed dispatch.
3486 * We also ramp up the dispatch depth gradually for async IO,
3487 * based on the last sync IO we serviced
3488 */
963b72fc 3489 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
9a7f38c4 3490 u64 last_sync = ktime_get_ns() - cfqd->last_delayed_sync;
8e296755 3491 unsigned int depth;
365722bb 3492
9a7f38c4 3493 depth = div64_u64(last_sync, cfqd->cfq_slice[1]);
e00c54c3
JA
3494 if (!depth && !cfqq->dispatched)
3495 depth = 1;
8e296755
JA
3496 if (depth < max_dispatch)
3497 max_dispatch = depth;
2f5cb738 3498 }
3ed9a296 3499
0b182d61
JA
3500 /*
3501 * If we're below the current max, allow a dispatch
3502 */
3503 return cfqq->dispatched < max_dispatch;
3504}
3505
3506/*
3507 * Dispatch a request from cfqq, moving them to the request queue
3508 * dispatch list.
3509 */
3510static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3511{
3512 struct request *rq;
3513
3514 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3515
3932a86b
GC
3516 rq = cfq_check_fifo(cfqq);
3517 if (rq)
3518 cfq_mark_cfqq_must_dispatch(cfqq);
3519
0b182d61
JA
3520 if (!cfq_may_dispatch(cfqd, cfqq))
3521 return false;
3522
3523 /*
3524 * follow expired path, else get first next available
3525 */
0b182d61
JA
3526 if (!rq)
3527 rq = cfqq->next_rq;
3932a86b
GC
3528 else
3529 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
0b182d61
JA
3530
3531 /*
3532 * insert request into driver dispatch list
3533 */
3534 cfq_dispatch_insert(cfqd->queue, rq);
3535
3536 if (!cfqd->active_cic) {
c5869807 3537 struct cfq_io_cq *cic = RQ_CIC(rq);
0b182d61 3538
c5869807 3539 atomic_long_inc(&cic->icq.ioc->refcount);
0b182d61
JA
3540 cfqd->active_cic = cic;
3541 }
3542
3543 return true;
3544}
3545
3546/*
3547 * Find the cfqq that we need to service and move a request from that to the
3548 * dispatch list
3549 */
3550static int cfq_dispatch_requests(struct request_queue *q, int force)
3551{
3552 struct cfq_data *cfqd = q->elevator->elevator_data;
3553 struct cfq_queue *cfqq;
3554
3555 if (!cfqd->busy_queues)
3556 return 0;
3557
3558 if (unlikely(force))
3559 return cfq_forced_dispatch(cfqd);
3560
3561 cfqq = cfq_select_queue(cfqd);
3562 if (!cfqq)
8e296755
JA
3563 return 0;
3564
2f5cb738 3565 /*
0b182d61 3566 * Dispatch a request from this cfqq, if it is allowed
2f5cb738 3567 */
0b182d61
JA
3568 if (!cfq_dispatch_request(cfqd, cfqq))
3569 return 0;
3570
2f5cb738 3571 cfqq->slice_dispatch++;
b029195d 3572 cfq_clear_cfqq_must_dispatch(cfqq);
22e2c507 3573
2f5cb738
JA
3574 /*
3575 * expire an async queue immediately if it has used up its slice. idle
3576 * queue always expire after 1 dispatch round.
3577 */
3578 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3579 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3580 cfq_class_idle(cfqq))) {
9a7f38c4 3581 cfqq->slice_end = ktime_get_ns() + 1;
e5ff082e 3582 cfq_slice_expired(cfqd, 0);
1da177e4
LT
3583 }
3584
b217a903 3585 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2f5cb738 3586 return 1;
1da177e4
LT
3587}
3588
1da177e4 3589/*
5e705374
JA
3590 * task holds one reference to the queue, dropped when task exits. each rq
3591 * in-flight on this queue also holds a reference, dropped when rq is freed.
1da177e4 3592 *
b1c35769 3593 * Each cfq queue took a reference on the parent group. Drop it now.
1da177e4
LT
3594 * queue lock must be held here.
3595 */
3596static void cfq_put_queue(struct cfq_queue *cfqq)
3597{
22e2c507 3598 struct cfq_data *cfqd = cfqq->cfqd;
0bbfeb83 3599 struct cfq_group *cfqg;
22e2c507 3600
30d7b944 3601 BUG_ON(cfqq->ref <= 0);
1da177e4 3602
30d7b944
SL
3603 cfqq->ref--;
3604 if (cfqq->ref)
1da177e4
LT
3605 return;
3606
7b679138 3607 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1da177e4 3608 BUG_ON(rb_first(&cfqq->sort_list));
22e2c507 3609 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
b1c35769 3610 cfqg = cfqq->cfqg;
1da177e4 3611
28f95cbc 3612 if (unlikely(cfqd->active_queue == cfqq)) {
e5ff082e 3613 __cfq_slice_expired(cfqd, cfqq, 0);
23e018a1 3614 cfq_schedule_dispatch(cfqd);
28f95cbc 3615 }
22e2c507 3616
f04a6424 3617 BUG_ON(cfq_cfqq_on_rr(cfqq));
1da177e4 3618 kmem_cache_free(cfq_pool, cfqq);
eb7d8c07 3619 cfqg_put(cfqg);
1da177e4
LT
3620}
3621
d02a2c07 3622static void cfq_put_cooperator(struct cfq_queue *cfqq)
1da177e4 3623{
df5fe3e8
JM
3624 struct cfq_queue *__cfqq, *next;
3625
df5fe3e8
JM
3626 /*
3627 * If this queue was scheduled to merge with another queue, be
3628 * sure to drop the reference taken on that queue (and others in
3629 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3630 */
3631 __cfqq = cfqq->new_cfqq;
3632 while (__cfqq) {
3633 if (__cfqq == cfqq) {
3634 WARN(1, "cfqq->new_cfqq loop detected\n");
3635 break;
3636 }
3637 next = __cfqq->new_cfqq;
3638 cfq_put_queue(__cfqq);
3639 __cfqq = next;
3640 }
d02a2c07
SL
3641}
3642
3643static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3644{
3645 if (unlikely(cfqq == cfqd->active_queue)) {
3646 __cfq_slice_expired(cfqd, cfqq, 0);
3647 cfq_schedule_dispatch(cfqd);
3648 }
3649
3650 cfq_put_cooperator(cfqq);
df5fe3e8 3651
89850f7e
JA
3652 cfq_put_queue(cfqq);
3653}
22e2c507 3654
9b84cacd
TH
3655static void cfq_init_icq(struct io_cq *icq)
3656{
3657 struct cfq_io_cq *cic = icq_to_cic(icq);
3658
9a7f38c4 3659 cic->ttime.last_end_request = ktime_get_ns();
9b84cacd
TH
3660}
3661
c5869807 3662static void cfq_exit_icq(struct io_cq *icq)
89850f7e 3663{
c5869807 3664 struct cfq_io_cq *cic = icq_to_cic(icq);
283287a5 3665 struct cfq_data *cfqd = cic_to_cfqd(cic);
4faa3c81 3666
563180a4
TH
3667 if (cic_to_cfqq(cic, false)) {
3668 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3669 cic_set_cfqq(cic, NULL, false);
12a05732
AV
3670 }
3671
563180a4
TH
3672 if (cic_to_cfqq(cic, true)) {
3673 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3674 cic_set_cfqq(cic, NULL, true);
12a05732 3675 }
89850f7e
JA
3676}
3677
abede6da 3678static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
22e2c507
JA
3679{
3680 struct task_struct *tsk = current;
3681 int ioprio_class;
3682
3b18152c 3683 if (!cfq_cfqq_prio_changed(cfqq))
22e2c507
JA
3684 return;
3685
598971bf 3686 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
22e2c507 3687 switch (ioprio_class) {
fe094d98
JA
3688 default:
3689 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3690 case IOPRIO_CLASS_NONE:
3691 /*
6d63c275 3692 * no prio set, inherit CPU scheduling settings
fe094d98
JA
3693 */
3694 cfqq->ioprio = task_nice_ioprio(tsk);
6d63c275 3695 cfqq->ioprio_class = task_nice_ioclass(tsk);
fe094d98
JA
3696 break;
3697 case IOPRIO_CLASS_RT:
598971bf 3698 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
fe094d98
JA
3699 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3700 break;
3701 case IOPRIO_CLASS_BE:
598971bf 3702 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
fe094d98
JA
3703 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3704 break;
3705 case IOPRIO_CLASS_IDLE:
3706 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3707 cfqq->ioprio = 7;
3708 cfq_clear_cfqq_idle_window(cfqq);
3709 break;
22e2c507
JA
3710 }
3711
3712 /*
3713 * keep track of original prio settings in case we have to temporarily
3714 * elevate the priority of this queue
3715 */
3716 cfqq->org_ioprio = cfqq->ioprio;
b8269db4 3717 cfqq->org_ioprio_class = cfqq->ioprio_class;
3b18152c 3718 cfq_clear_cfqq_prio_changed(cfqq);
22e2c507
JA
3719}
3720
598971bf 3721static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
22e2c507 3722{
598971bf 3723 int ioprio = cic->icq.ioc->ioprio;
bca4b914 3724 struct cfq_data *cfqd = cic_to_cfqd(cic);
478a82b0 3725 struct cfq_queue *cfqq;
35e6077c 3726
598971bf
TH
3727 /*
3728 * Check whether ioprio has changed. The condition may trigger
3729 * spuriously on a newly created cic but there's no harm.
3730 */
3731 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
caaa5f9f
JA
3732 return;
3733
563180a4 3734 cfqq = cic_to_cfqq(cic, false);
caaa5f9f 3735 if (cfqq) {
563180a4 3736 cfq_put_queue(cfqq);
2da8de0b 3737 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
563180a4 3738 cic_set_cfqq(cic, cfqq, false);
22e2c507 3739 }
caaa5f9f 3740
563180a4 3741 cfqq = cic_to_cfqq(cic, true);
caaa5f9f
JA
3742 if (cfqq)
3743 cfq_mark_cfqq_prio_changed(cfqq);
598971bf
TH
3744
3745 cic->ioprio = ioprio;
22e2c507
JA
3746}
3747
d5036d77 3748static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
a6151c3a 3749 pid_t pid, bool is_sync)
d5036d77
JA
3750{
3751 RB_CLEAR_NODE(&cfqq->rb_node);
3752 RB_CLEAR_NODE(&cfqq->p_node);
3753 INIT_LIST_HEAD(&cfqq->fifo);
3754
30d7b944 3755 cfqq->ref = 0;
d5036d77
JA
3756 cfqq->cfqd = cfqd;
3757
3758 cfq_mark_cfqq_prio_changed(cfqq);
3759
3760 if (is_sync) {
3761 if (!cfq_class_idle(cfqq))
3762 cfq_mark_cfqq_idle_window(cfqq);
3763 cfq_mark_cfqq_sync(cfqq);
3764 }
3765 cfqq->pid = pid;
3766}
3767
24610333 3768#ifdef CONFIG_CFQ_GROUP_IOSCHED
598971bf 3769static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
24610333 3770{
bca4b914 3771 struct cfq_data *cfqd = cic_to_cfqd(cic);
60a83707 3772 struct cfq_queue *cfqq;
f4da8072 3773 uint64_t serial_nr;
24610333 3774
598971bf 3775 rcu_read_lock();
f4da8072 3776 serial_nr = bio_blkcg(bio)->css.serial_nr;
598971bf 3777 rcu_read_unlock();
24610333 3778
598971bf
TH
3779 /*
3780 * Check whether blkcg has changed. The condition may trigger
3781 * spuriously on a newly created cic but there's no harm.
3782 */
f4da8072 3783 if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
598971bf 3784 return;
24610333 3785
60a83707
TH
3786 /*
3787 * Drop reference to queues. New queues will be assigned in new
3788 * group upon arrival of fresh requests.
3789 */
3790 cfqq = cic_to_cfqq(cic, false);
3791 if (cfqq) {
3792 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3793 cic_set_cfqq(cic, NULL, false);
3794 cfq_put_queue(cfqq);
3795 }
3796
3797 cfqq = cic_to_cfqq(cic, true);
3798 if (cfqq) {
3799 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3800 cic_set_cfqq(cic, NULL, true);
3801 cfq_put_queue(cfqq);
24610333 3802 }
598971bf 3803
f4da8072 3804 cic->blkcg_serial_nr = serial_nr;
24610333 3805}
598971bf
TH
3806#else
3807static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
24610333
VG
3808#endif /* CONFIG_CFQ_GROUP_IOSCHED */
3809
c2dea2d1 3810static struct cfq_queue **
60a83707 3811cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
c2dea2d1 3812{
fe094d98 3813 switch (ioprio_class) {
c2dea2d1 3814 case IOPRIO_CLASS_RT:
60a83707 3815 return &cfqg->async_cfqq[0][ioprio];
598971bf
TH
3816 case IOPRIO_CLASS_NONE:
3817 ioprio = IOPRIO_NORM;
3818 /* fall through */
c2dea2d1 3819 case IOPRIO_CLASS_BE:
60a83707 3820 return &cfqg->async_cfqq[1][ioprio];
c2dea2d1 3821 case IOPRIO_CLASS_IDLE:
60a83707 3822 return &cfqg->async_idle_cfqq;
c2dea2d1
VT
3823 default:
3824 BUG();
3825 }
3826}
3827
15c31be4 3828static struct cfq_queue *
abede6da 3829cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
2da8de0b 3830 struct bio *bio)
15c31be4 3831{
c6ce1943
JM
3832 int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3833 int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
d4aad7ff 3834 struct cfq_queue **async_cfqq = NULL;
4ebc1c61 3835 struct cfq_queue *cfqq;
322731ed
TH
3836 struct cfq_group *cfqg;
3837
3838 rcu_read_lock();
ae118896 3839 cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
322731ed
TH
3840 if (!cfqg) {
3841 cfqq = &cfqd->oom_cfqq;
3842 goto out;
3843 }
15c31be4 3844
c2dea2d1 3845 if (!is_sync) {
c6ce1943
JM
3846 if (!ioprio_valid(cic->ioprio)) {
3847 struct task_struct *tsk = current;
3848 ioprio = task_nice_ioprio(tsk);
3849 ioprio_class = task_nice_ioclass(tsk);
3850 }
60a83707 3851 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
c2dea2d1 3852 cfqq = *async_cfqq;
4ebc1c61
TH
3853 if (cfqq)
3854 goto out;
c2dea2d1
VT
3855 }
3856
d4aad7ff
TH
3857 cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3858 cfqd->queue->node);
3859 if (!cfqq) {
3860 cfqq = &cfqd->oom_cfqq;
3861 goto out;
3862 }
3863
3864 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3865 cfq_init_prio_data(cfqq, cic);
3866 cfq_link_cfqq_cfqg(cfqq, cfqg);
3867 cfq_log_cfqq(cfqd, cfqq, "alloced");
15c31be4 3868
d4aad7ff
TH
3869 if (async_cfqq) {
3870 /* a new async queue is created, pin and remember */
30d7b944 3871 cfqq->ref++;
c2dea2d1 3872 *async_cfqq = cfqq;
15c31be4 3873 }
4ebc1c61 3874out:
30d7b944 3875 cfqq->ref++;
322731ed 3876 rcu_read_unlock();
15c31be4
JA
3877 return cfqq;
3878}
3879
22e2c507 3880static void
9a7f38c4 3881__cfq_update_io_thinktime(struct cfq_ttime *ttime, u64 slice_idle)
1da177e4 3882{
9a7f38c4 3883 u64 elapsed = ktime_get_ns() - ttime->last_end_request;
383cd721 3884 elapsed = min(elapsed, 2UL * slice_idle);
db3b5848 3885
383cd721 3886 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
9a7f38c4
JM
3887 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
3888 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
3889 ttime->ttime_samples);
383cd721
SL
3890}
3891
3892static void
3893cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
c5869807 3894 struct cfq_io_cq *cic)
383cd721 3895{
f5f2b6ce 3896 if (cfq_cfqq_sync(cfqq)) {
383cd721 3897 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
f5f2b6ce
SL
3898 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3899 cfqd->cfq_slice_idle);
3900 }
7700fc4f
SL
3901#ifdef CONFIG_CFQ_GROUP_IOSCHED
3902 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3903#endif
22e2c507 3904}
1da177e4 3905
206dc69b 3906static void
b2c18e1e 3907cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
6d048f53 3908 struct request *rq)
206dc69b 3909{
3dde36dd 3910 sector_t sdist = 0;
41647e7a 3911 sector_t n_sec = blk_rq_sectors(rq);
3dde36dd
CZ
3912 if (cfqq->last_request_pos) {
3913 if (cfqq->last_request_pos < blk_rq_pos(rq))
3914 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3915 else
3916 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3917 }
206dc69b 3918
3dde36dd 3919 cfqq->seek_history <<= 1;
41647e7a
CZ
3920 if (blk_queue_nonrot(cfqd->queue))
3921 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3922 else
3923 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
206dc69b 3924}
1da177e4 3925
22e2c507
JA
3926/*
3927 * Disable idle window if the process thinks too long or seeks so much that
3928 * it doesn't matter
3929 */
3930static void
3931cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
c5869807 3932 struct cfq_io_cq *cic)
22e2c507 3933{
7b679138 3934 int old_idle, enable_idle;
1be92f2f 3935
0871714e
JA
3936 /*
3937 * Don't idle for async or idle io prio class
3938 */
3939 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1be92f2f
JA
3940 return;
3941
c265a7f4 3942 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1da177e4 3943
76280aff
CZ
3944 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3945 cfq_mark_cfqq_deep(cfqq);
3946
749ef9f8
CZ
3947 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3948 enable_idle = 0;
f6e8d01b 3949 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
c5869807
TH
3950 !cfqd->cfq_slice_idle ||
3951 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
22e2c507 3952 enable_idle = 0;
383cd721
SL
3953 else if (sample_valid(cic->ttime.ttime_samples)) {
3954 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
22e2c507
JA
3955 enable_idle = 0;
3956 else
3957 enable_idle = 1;
1da177e4
LT
3958 }
3959
7b679138
JA
3960 if (old_idle != enable_idle) {
3961 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3962 if (enable_idle)
3963 cfq_mark_cfqq_idle_window(cfqq);
3964 else
3965 cfq_clear_cfqq_idle_window(cfqq);
3966 }
22e2c507 3967}
1da177e4 3968
22e2c507
JA
3969/*
3970 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3971 * no or if we aren't sure, a 1 will cause a preempt.
3972 */
a6151c3a 3973static bool
22e2c507 3974cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
5e705374 3975 struct request *rq)
22e2c507 3976{
6d048f53 3977 struct cfq_queue *cfqq;
22e2c507 3978
6d048f53
JA
3979 cfqq = cfqd->active_queue;
3980 if (!cfqq)
a6151c3a 3981 return false;
22e2c507 3982
6d048f53 3983 if (cfq_class_idle(new_cfqq))
a6151c3a 3984 return false;
22e2c507
JA
3985
3986 if (cfq_class_idle(cfqq))
a6151c3a 3987 return true;
1e3335de 3988
875feb63
DS
3989 /*
3990 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3991 */
3992 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3993 return false;
3994
374f84ac
JA
3995 /*
3996 * if the new request is sync, but the currently running queue is
3997 * not, let the sync request have priority.
3998 */
3932a86b 3999 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
a6151c3a 4000 return true;
1e3335de 4001
3984aa55
JK
4002 /*
4003 * Treat ancestors of current cgroup the same way as current cgroup.
4004 * For anybody else we disallow preemption to guarantee service
4005 * fairness among cgroups.
4006 */
4007 if (!cfqg_is_descendant(cfqq->cfqg, new_cfqq->cfqg))
8682e1f1
VG
4008 return false;
4009
4010 if (cfq_slice_used(cfqq))
4011 return true;
4012
6c80731c
JK
4013 /*
4014 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
4015 */
4016 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
4017 return true;
4018
4019 WARN_ON_ONCE(cfqq->ioprio_class != new_cfqq->ioprio_class);
8682e1f1 4020 /* Allow preemption only if we are idling on sync-noidle tree */
4d2ceea4 4021 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
8682e1f1 4022 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
8682e1f1
VG
4023 RB_EMPTY_ROOT(&cfqq->sort_list))
4024 return true;
4025
b53d1ed7
JA
4026 /*
4027 * So both queues are sync. Let the new request get disk time if
4028 * it's a metadata request and the current queue is doing regular IO.
4029 */
65299a3b 4030 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
b53d1ed7
JA
4031 return true;
4032
d2d59e18
SL
4033 /* An idle queue should not be idle now for some reason */
4034 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
4035 return true;
4036
1e3335de 4037 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
a6151c3a 4038 return false;
1e3335de
JA
4039
4040 /*
4041 * if this request is as-good as one we would expect from the
4042 * current cfqq, let it preempt
4043 */
e9ce335d 4044 if (cfq_rq_close(cfqd, cfqq, rq))
a6151c3a 4045 return true;
1e3335de 4046
a6151c3a 4047 return false;
22e2c507
JA
4048}
4049
4050/*
4051 * cfqq preempts the active queue. if we allowed preempt with no slice left,
4052 * let it have half of its nominal slice.
4053 */
4054static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4055{
df0793ab
SL
4056 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
4057
7b679138 4058 cfq_log_cfqq(cfqd, cfqq, "preempt");
df0793ab 4059 cfq_slice_expired(cfqd, 1);
22e2c507 4060
f8ae6e3e
SL
4061 /*
4062 * workload type is changed, don't save slice, otherwise preempt
4063 * doesn't happen
4064 */
df0793ab 4065 if (old_type != cfqq_type(cfqq))
4d2ceea4 4066 cfqq->cfqg->saved_wl_slice = 0;
f8ae6e3e 4067
bf572256
JA
4068 /*
4069 * Put the new queue at the front of the of the current list,
4070 * so we know that it will be selected next.
4071 */
4072 BUG_ON(!cfq_cfqq_on_rr(cfqq));
edd75ffd
JA
4073
4074 cfq_service_tree_add(cfqd, cfqq, 1);
eda5e0c9 4075
62a37f6b
JT
4076 cfqq->slice_end = 0;
4077 cfq_mark_cfqq_slice_new(cfqq);
22e2c507
JA
4078}
4079
22e2c507 4080/*
5e705374 4081 * Called when a new fs request (rq) is added (to cfqq). Check if there's
22e2c507
JA
4082 * something we should do about it
4083 */
4084static void
5e705374
JA
4085cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
4086 struct request *rq)
22e2c507 4087{
c5869807 4088 struct cfq_io_cq *cic = RQ_CIC(rq);
12e9fddd 4089
45333d5a 4090 cfqd->rq_queued++;
65299a3b
CH
4091 if (rq->cmd_flags & REQ_PRIO)
4092 cfqq->prio_pending++;
374f84ac 4093
383cd721 4094 cfq_update_io_thinktime(cfqd, cfqq, cic);
b2c18e1e 4095 cfq_update_io_seektime(cfqd, cfqq, rq);
9c2c38a1
JA
4096 cfq_update_idle_window(cfqd, cfqq, cic);
4097
b2c18e1e 4098 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
22e2c507
JA
4099
4100 if (cfqq == cfqd->active_queue) {
4101 /*
b029195d
JA
4102 * Remember that we saw a request from this process, but
4103 * don't start queuing just yet. Otherwise we risk seeing lots
4104 * of tiny requests, because we disrupt the normal plugging
d6ceb25e
JA
4105 * and merging. If the request is already larger than a single
4106 * page, let it rip immediately. For that case we assume that
2d870722
JA
4107 * merging is already done. Ditto for a busy system that
4108 * has other work pending, don't risk delaying until the
4109 * idle timer unplug to continue working.
22e2c507 4110 */
d6ceb25e 4111 if (cfq_cfqq_wait_request(cfqq)) {
09cbfeaf 4112 if (blk_rq_bytes(rq) > PAGE_SIZE ||
2d870722 4113 cfqd->busy_queues > 1) {
812df48d 4114 cfq_del_timer(cfqd, cfqq);
554554f6 4115 cfq_clear_cfqq_wait_request(cfqq);
24ecfbe2 4116 __blk_run_queue(cfqd->queue);
a11cdaa7 4117 } else {
155fead9 4118 cfqg_stats_update_idle_time(cfqq->cfqg);
bf791937 4119 cfq_mark_cfqq_must_dispatch(cfqq);
a11cdaa7 4120 }
d6ceb25e 4121 }
5e705374 4122 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
22e2c507
JA
4123 /*
4124 * not the active queue - expire current slice if it is
4125 * idle and has expired it's mean thinktime or this new queue
3a9a3f6c
DS
4126 * has some old slice time left and is of higher priority or
4127 * this new queue is RT and the current one is BE
22e2c507
JA
4128 */
4129 cfq_preempt_queue(cfqd, cfqq);
24ecfbe2 4130 __blk_run_queue(cfqd->queue);
22e2c507 4131 }
1da177e4
LT
4132}
4133
165125e1 4134static void cfq_insert_request(struct request_queue *q, struct request *rq)
1da177e4 4135{
b4878f24 4136 struct cfq_data *cfqd = q->elevator->elevator_data;
5e705374 4137 struct cfq_queue *cfqq = RQ_CFQQ(rq);
22e2c507 4138
7b679138 4139 cfq_log_cfqq(cfqd, cfqq, "insert_request");
abede6da 4140 cfq_init_prio_data(cfqq, RQ_CIC(rq));
1da177e4 4141
9a7f38c4 4142 rq->fifo_time = ktime_get_ns() + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
22e2c507 4143 list_add_tail(&rq->queuelist, &cfqq->fifo);
aa6f6a3d 4144 cfq_add_rq_rb(rq);
63a4cc24 4145 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group, req_op(rq),
155fead9 4146 rq->cmd_flags);
5e705374 4147 cfq_rq_enqueued(cfqd, cfqq, rq);
1da177e4
LT
4148}
4149
45333d5a
AC
4150/*
4151 * Update hw_tag based on peak queue depth over 50 samples under
4152 * sufficient load.
4153 */
4154static void cfq_update_hw_tag(struct cfq_data *cfqd)
4155{
1a1238a7
SL
4156 struct cfq_queue *cfqq = cfqd->active_queue;
4157
53c583d2
CZ
4158 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4159 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
e459dd08
CZ
4160
4161 if (cfqd->hw_tag == 1)
4162 return;
45333d5a
AC
4163
4164 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
53c583d2 4165 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
45333d5a
AC
4166 return;
4167
1a1238a7
SL
4168 /*
4169 * If active queue hasn't enough requests and can idle, cfq might not
4170 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4171 * case
4172 */
4173 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4174 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
53c583d2 4175 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
1a1238a7
SL
4176 return;
4177
45333d5a
AC
4178 if (cfqd->hw_tag_samples++ < 50)
4179 return;
4180
e459dd08 4181 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
45333d5a
AC
4182 cfqd->hw_tag = 1;
4183 else
4184 cfqd->hw_tag = 0;
45333d5a
AC
4185}
4186
7667aa06
VG
4187static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4188{
c5869807 4189 struct cfq_io_cq *cic = cfqd->active_cic;
9a7f38c4 4190 u64 now = ktime_get_ns();
7667aa06 4191
02a8f01b
JT
4192 /* If the queue already has requests, don't wait */
4193 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4194 return false;
4195
7667aa06
VG
4196 /* If there are other queues in the group, don't wait */
4197 if (cfqq->cfqg->nr_cfqq > 1)
4198 return false;
4199
7700fc4f
SL
4200 /* the only queue in the group, but think time is big */
4201 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4202 return false;
4203
7667aa06
VG
4204 if (cfq_slice_used(cfqq))
4205 return true;
4206
4207 /* if slice left is less than think time, wait busy */
383cd721 4208 if (cic && sample_valid(cic->ttime.ttime_samples)
9a7f38c4 4209 && (cfqq->slice_end - now < cic->ttime.ttime_mean))
7667aa06
VG
4210 return true;
4211
4212 /*
4213 * If think times is less than a jiffy than ttime_mean=0 and above
4214 * will not be true. It might happen that slice has not expired yet
4215 * but will expire soon (4-5 ns) during select_queue(). To cover the
4216 * case where think time is less than a jiffy, mark the queue wait
4217 * busy if only 1 jiffy is left in the slice.
4218 */
9a7f38c4 4219 if (cfqq->slice_end - now <= jiffies_to_nsecs(1))
7667aa06
VG
4220 return true;
4221
4222 return false;
4223}
4224
165125e1 4225static void cfq_completed_request(struct request_queue *q, struct request *rq)
1da177e4 4226{
5e705374 4227 struct cfq_queue *cfqq = RQ_CFQQ(rq);
b4878f24 4228 struct cfq_data *cfqd = cfqq->cfqd;
5380a101 4229 const int sync = rq_is_sync(rq);
9a7f38c4 4230 u64 now = ktime_get_ns();
1da177e4 4231
33659ebb
CH
4232 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4233 !!(rq->cmd_flags & REQ_NOIDLE));
1da177e4 4234
45333d5a
AC
4235 cfq_update_hw_tag(cfqd);
4236
53c583d2 4237 WARN_ON(!cfqd->rq_in_driver);
6d048f53 4238 WARN_ON(!cfqq->dispatched);
53c583d2 4239 cfqd->rq_in_driver--;
6d048f53 4240 cfqq->dispatched--;
80bdf0c7 4241 (RQ_CFQG(rq))->dispatched--;
155fead9 4242 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
63a4cc24
MC
4243 rq_io_start_time_ns(rq), req_op(rq),
4244 rq->cmd_flags);
1da177e4 4245
53c583d2 4246 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3ed9a296 4247
365722bb 4248 if (sync) {
34b98d03 4249 struct cfq_rb_root *st;
f5f2b6ce 4250
383cd721 4251 RQ_CIC(rq)->ttime.last_end_request = now;
f5f2b6ce
SL
4252
4253 if (cfq_cfqq_on_rr(cfqq))
34b98d03 4254 st = cfqq->service_tree;
f5f2b6ce 4255 else
34b98d03
VG
4256 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4257 cfqq_type(cfqq));
4258
4259 st->ttime.last_end_request = now;
149321a6
JK
4260 /*
4261 * We have to do this check in jiffies since start_time is in
4262 * jiffies and it is not trivial to convert to ns. If
4263 * cfq_fifo_expire[1] ever comes close to 1 jiffie, this test
4264 * will become problematic but so far we are fine (the default
4265 * is 128 ms).
4266 */
4267 if (!time_after(rq->start_time +
4268 nsecs_to_jiffies(cfqd->cfq_fifo_expire[1]),
4269 jiffies))
573412b2 4270 cfqd->last_delayed_sync = now;
365722bb 4271 }
caaa5f9f 4272
7700fc4f
SL
4273#ifdef CONFIG_CFQ_GROUP_IOSCHED
4274 cfqq->cfqg->ttime.last_end_request = now;
4275#endif
4276
caaa5f9f
JA
4277 /*
4278 * If this is the active queue, check if it needs to be expired,
4279 * or if we want to idle in case it has no pending requests.
4280 */
4281 if (cfqd->active_queue == cfqq) {
a36e71f9
JA
4282 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4283
44f7c160
JA
4284 if (cfq_cfqq_slice_new(cfqq)) {
4285 cfq_set_prio_slice(cfqd, cfqq);
4286 cfq_clear_cfqq_slice_new(cfqq);
4287 }
f75edf2d
VG
4288
4289 /*
7667aa06
VG
4290 * Should we wait for next request to come in before we expire
4291 * the queue.
f75edf2d 4292 */
7667aa06 4293 if (cfq_should_wait_busy(cfqd, cfqq)) {
9a7f38c4 4294 u64 extend_sl = cfqd->cfq_slice_idle;
80bdf0c7
VG
4295 if (!cfqd->cfq_slice_idle)
4296 extend_sl = cfqd->cfq_group_idle;
9a7f38c4 4297 cfqq->slice_end = now + extend_sl;
f75edf2d 4298 cfq_mark_cfqq_wait_busy(cfqq);
b1ffe737 4299 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
f75edf2d
VG
4300 }
4301
a36e71f9 4302 /*
8e550632
CZ
4303 * Idling is not enabled on:
4304 * - expired queues
4305 * - idle-priority queues
4306 * - async queues
4307 * - queues with still some requests queued
4308 * - when there is a close cooperator
a36e71f9 4309 */
0871714e 4310 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
e5ff082e 4311 cfq_slice_expired(cfqd, 1);
8e550632
CZ
4312 else if (sync && cfqq_empty &&
4313 !cfq_close_cooperator(cfqd, cfqq)) {
749ef9f8 4314 cfq_arm_slice_timer(cfqd);
8e550632 4315 }
caaa5f9f 4316 }
6d048f53 4317
53c583d2 4318 if (!cfqd->rq_in_driver)
23e018a1 4319 cfq_schedule_dispatch(cfqd);
1da177e4
LT
4320}
4321
b8269db4
JA
4322static void cfqq_boost_on_prio(struct cfq_queue *cfqq, int op_flags)
4323{
4324 /*
4325 * If REQ_PRIO is set, boost class and prio level, if it's below
4326 * BE/NORM. If prio is not set, restore the potentially boosted
4327 * class/prio level.
4328 */
4329 if (!(op_flags & REQ_PRIO)) {
4330 cfqq->ioprio_class = cfqq->org_ioprio_class;
4331 cfqq->ioprio = cfqq->org_ioprio;
4332 } else {
4333 if (cfq_class_idle(cfqq))
4334 cfqq->ioprio_class = IOPRIO_CLASS_BE;
4335 if (cfqq->ioprio > IOPRIO_NORM)
4336 cfqq->ioprio = IOPRIO_NORM;
4337 }
4338}
4339
89850f7e 4340static inline int __cfq_may_queue(struct cfq_queue *cfqq)
22e2c507 4341{
1b379d8d 4342 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3b18152c 4343 cfq_mark_cfqq_must_alloc_slice(cfqq);
22e2c507 4344 return ELV_MQUEUE_MUST;
3b18152c 4345 }
1da177e4 4346
22e2c507 4347 return ELV_MQUEUE_MAY;
22e2c507
JA
4348}
4349
ba568ea0 4350static int cfq_may_queue(struct request_queue *q, int op, int op_flags)
22e2c507
JA
4351{
4352 struct cfq_data *cfqd = q->elevator->elevator_data;
4353 struct task_struct *tsk = current;
c5869807 4354 struct cfq_io_cq *cic;
22e2c507
JA
4355 struct cfq_queue *cfqq;
4356
4357 /*
4358 * don't force setup of a queue from here, as a call to may_queue
4359 * does not necessarily imply that a request actually will be queued.
4360 * so just lookup a possibly existing queue, or return 'may queue'
4361 * if that fails
4362 */
4ac845a2 4363 cic = cfq_cic_lookup(cfqd, tsk->io_context);
91fac317
VT
4364 if (!cic)
4365 return ELV_MQUEUE_MAY;
4366
d9d8c5c4 4367 cfqq = cic_to_cfqq(cic, rw_is_sync(op, op_flags));
22e2c507 4368 if (cfqq) {
abede6da 4369 cfq_init_prio_data(cfqq, cic);
b8269db4 4370 cfqq_boost_on_prio(cfqq, op_flags);
22e2c507 4371
89850f7e 4372 return __cfq_may_queue(cfqq);
22e2c507
JA
4373 }
4374
4375 return ELV_MQUEUE_MAY;
1da177e4
LT
4376}
4377
1da177e4
LT
4378/*
4379 * queue lock held here
4380 */
bb37b94c 4381static void cfq_put_request(struct request *rq)
1da177e4 4382{
5e705374 4383 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1da177e4 4384
5e705374 4385 if (cfqq) {
22e2c507 4386 const int rw = rq_data_dir(rq);
1da177e4 4387
22e2c507
JA
4388 BUG_ON(!cfqq->allocated[rw]);
4389 cfqq->allocated[rw]--;
1da177e4 4390
7f1dc8a2 4391 /* Put down rq reference on cfqg */
eb7d8c07 4392 cfqg_put(RQ_CFQG(rq));
a612fddf
TH
4393 rq->elv.priv[0] = NULL;
4394 rq->elv.priv[1] = NULL;
7f1dc8a2 4395
1da177e4
LT
4396 cfq_put_queue(cfqq);
4397 }
4398}
4399
df5fe3e8 4400static struct cfq_queue *
c5869807 4401cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
df5fe3e8
JM
4402 struct cfq_queue *cfqq)
4403{
4404 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4405 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
b3b6d040 4406 cfq_mark_cfqq_coop(cfqq->new_cfqq);
df5fe3e8
JM
4407 cfq_put_queue(cfqq);
4408 return cic_to_cfqq(cic, 1);
4409}
4410
e6c5bc73
JM
4411/*
4412 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4413 * was the last process referring to said cfqq.
4414 */
4415static struct cfq_queue *
c5869807 4416split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
e6c5bc73
JM
4417{
4418 if (cfqq_process_refs(cfqq) == 1) {
e6c5bc73
JM
4419 cfqq->pid = current->pid;
4420 cfq_clear_cfqq_coop(cfqq);
ae54abed 4421 cfq_clear_cfqq_split_coop(cfqq);
e6c5bc73
JM
4422 return cfqq;
4423 }
4424
4425 cic_set_cfqq(cic, NULL, 1);
d02a2c07
SL
4426
4427 cfq_put_cooperator(cfqq);
4428
e6c5bc73
JM
4429 cfq_put_queue(cfqq);
4430 return NULL;
4431}
1da177e4 4432/*
22e2c507 4433 * Allocate cfq data structures associated with this request.
1da177e4 4434 */
22e2c507 4435static int
852c788f
TH
4436cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4437 gfp_t gfp_mask)
1da177e4
LT
4438{
4439 struct cfq_data *cfqd = q->elevator->elevator_data;
f1f8cc94 4440 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
1da177e4 4441 const int rw = rq_data_dir(rq);
a6151c3a 4442 const bool is_sync = rq_is_sync(rq);
22e2c507 4443 struct cfq_queue *cfqq;
1da177e4 4444
216284c3 4445 spin_lock_irq(q->queue_lock);
f1f8cc94 4446
598971bf
TH
4447 check_ioprio_changed(cic, bio);
4448 check_blkcg_changed(cic, bio);
e6c5bc73 4449new_queue:
91fac317 4450 cfqq = cic_to_cfqq(cic, is_sync);
32f2e807 4451 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
bce6133b
TH
4452 if (cfqq)
4453 cfq_put_queue(cfqq);
2da8de0b 4454 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
91fac317 4455 cic_set_cfqq(cic, cfqq, is_sync);
df5fe3e8 4456 } else {
e6c5bc73
JM
4457 /*
4458 * If the queue was seeky for too long, break it apart.
4459 */
ae54abed 4460 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
e6c5bc73
JM
4461 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4462 cfqq = split_cfqq(cic, cfqq);
4463 if (!cfqq)
4464 goto new_queue;
4465 }
4466
df5fe3e8
JM
4467 /*
4468 * Check to see if this queue is scheduled to merge with
4469 * another, closely cooperating queue. The merging of
4470 * queues happens here as it must be done in process context.
4471 * The reference on new_cfqq was taken in merge_cfqqs.
4472 */
4473 if (cfqq->new_cfqq)
4474 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
91fac317 4475 }
1da177e4
LT
4476
4477 cfqq->allocated[rw]++;
1da177e4 4478
6fae9c25 4479 cfqq->ref++;
eb7d8c07 4480 cfqg_get(cfqq->cfqg);
a612fddf 4481 rq->elv.priv[0] = cfqq;
1adaf3dd 4482 rq->elv.priv[1] = cfqq->cfqg;
216284c3 4483 spin_unlock_irq(q->queue_lock);
5e705374 4484 return 0;
1da177e4
LT
4485}
4486
65f27f38 4487static void cfq_kick_queue(struct work_struct *work)
22e2c507 4488{
65f27f38 4489 struct cfq_data *cfqd =
23e018a1 4490 container_of(work, struct cfq_data, unplug_work);
165125e1 4491 struct request_queue *q = cfqd->queue;
22e2c507 4492
40bb54d1 4493 spin_lock_irq(q->queue_lock);
24ecfbe2 4494 __blk_run_queue(cfqd->queue);
40bb54d1 4495 spin_unlock_irq(q->queue_lock);
22e2c507
JA
4496}
4497
4498/*
4499 * Timer running if the active_queue is currently idling inside its time slice
4500 */
91148325 4501static enum hrtimer_restart cfq_idle_slice_timer(struct hrtimer *timer)
22e2c507 4502{
91148325
JK
4503 struct cfq_data *cfqd = container_of(timer, struct cfq_data,
4504 idle_slice_timer);
22e2c507
JA
4505 struct cfq_queue *cfqq;
4506 unsigned long flags;
3c6bd2f8 4507 int timed_out = 1;
22e2c507 4508
7b679138
JA
4509 cfq_log(cfqd, "idle timer fired");
4510
22e2c507
JA
4511 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4512
fe094d98
JA
4513 cfqq = cfqd->active_queue;
4514 if (cfqq) {
3c6bd2f8
JA
4515 timed_out = 0;
4516
b029195d
JA
4517 /*
4518 * We saw a request before the queue expired, let it through
4519 */
4520 if (cfq_cfqq_must_dispatch(cfqq))
4521 goto out_kick;
4522
22e2c507
JA
4523 /*
4524 * expired
4525 */
44f7c160 4526 if (cfq_slice_used(cfqq))
22e2c507
JA
4527 goto expire;
4528
4529 /*
4530 * only expire and reinvoke request handler, if there are
4531 * other queues with pending requests
4532 */
caaa5f9f 4533 if (!cfqd->busy_queues)
22e2c507 4534 goto out_cont;
22e2c507
JA
4535
4536 /*
4537 * not expired and it has a request pending, let it dispatch
4538 */
75e50984 4539 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
22e2c507 4540 goto out_kick;
76280aff
CZ
4541
4542 /*
4543 * Queue depth flag is reset only when the idle didn't succeed
4544 */
4545 cfq_clear_cfqq_deep(cfqq);
22e2c507
JA
4546 }
4547expire:
e5ff082e 4548 cfq_slice_expired(cfqd, timed_out);
22e2c507 4549out_kick:
23e018a1 4550 cfq_schedule_dispatch(cfqd);
22e2c507
JA
4551out_cont:
4552 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
91148325 4553 return HRTIMER_NORESTART;
22e2c507
JA
4554}
4555
3b18152c
JA
4556static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4557{
91148325 4558 hrtimer_cancel(&cfqd->idle_slice_timer);
23e018a1 4559 cancel_work_sync(&cfqd->unplug_work);
3b18152c 4560}
22e2c507 4561
b374d18a 4562static void cfq_exit_queue(struct elevator_queue *e)
1da177e4 4563{
22e2c507 4564 struct cfq_data *cfqd = e->elevator_data;
165125e1 4565 struct request_queue *q = cfqd->queue;
22e2c507 4566
3b18152c 4567 cfq_shutdown_timer_wq(cfqd);
e2d74ac0 4568
d9ff4187 4569 spin_lock_irq(q->queue_lock);
e2d74ac0 4570
d9ff4187 4571 if (cfqd->active_queue)
e5ff082e 4572 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
e2d74ac0 4573
03aa264a
TH
4574 spin_unlock_irq(q->queue_lock);
4575
a90d742e
AV
4576 cfq_shutdown_timer_wq(cfqd);
4577
ffea73fc
TH
4578#ifdef CONFIG_CFQ_GROUP_IOSCHED
4579 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4580#else
f51b802c 4581 kfree(cfqd->root_group);
2abae55f 4582#endif
56edf7d7 4583 kfree(cfqd);
1da177e4
LT
4584}
4585
d50235b7 4586static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
1da177e4
LT
4587{
4588 struct cfq_data *cfqd;
3c798398 4589 struct blkcg_gq *blkg __maybe_unused;
a2b1693b 4590 int i, ret;
d50235b7
JM
4591 struct elevator_queue *eq;
4592
4593 eq = elevator_alloc(q, e);
4594 if (!eq)
4595 return -ENOMEM;
1da177e4 4596
c1b511eb 4597 cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
d50235b7
JM
4598 if (!cfqd) {
4599 kobject_put(&eq->kobj);
b2fab5ac 4600 return -ENOMEM;
d50235b7
JM
4601 }
4602 eq->elevator_data = cfqd;
80b15c73 4603
f51b802c 4604 cfqd->queue = q;
d50235b7
JM
4605 spin_lock_irq(q->queue_lock);
4606 q->elevator = eq;
4607 spin_unlock_irq(q->queue_lock);
f51b802c 4608
1fa8f6d6
VG
4609 /* Init root service tree */
4610 cfqd->grp_service_tree = CFQ_RB_ROOT;
4611
f51b802c 4612 /* Init root group and prefer root group over other groups by default */
25fb5169 4613#ifdef CONFIG_CFQ_GROUP_IOSCHED
3c798398 4614 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
a2b1693b
TH
4615 if (ret)
4616 goto out_free;
f51b802c 4617
a2b1693b 4618 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
f51b802c 4619#else
a2b1693b 4620 ret = -ENOMEM;
f51b802c
TH
4621 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4622 GFP_KERNEL, cfqd->queue->node);
a2b1693b
TH
4623 if (!cfqd->root_group)
4624 goto out_free;
5624a4e4 4625
a2b1693b 4626 cfq_init_cfqg_base(cfqd->root_group);
3ecca629
TH
4627 cfqd->root_group->weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4628 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
69d7fde5 4629#endif
5624a4e4 4630
26a2ac00
JA
4631 /*
4632 * Not strictly needed (since RB_ROOT just clears the node and we
4633 * zeroed cfqd on alloc), but better be safe in case someone decides
4634 * to add magic to the rb code
4635 */
4636 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4637 cfqd->prio_trees[i] = RB_ROOT;
4638
6118b70b 4639 /*
d4aad7ff 4640 * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
6118b70b 4641 * Grab a permanent reference to it, so that the normal code flow
f51b802c
TH
4642 * will not attempt to free it. oom_cfqq is linked to root_group
4643 * but shouldn't hold a reference as it'll never be unlinked. Lose
4644 * the reference from linking right away.
6118b70b
JA
4645 */
4646 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
30d7b944 4647 cfqd->oom_cfqq.ref++;
1adaf3dd
TH
4648
4649 spin_lock_irq(q->queue_lock);
f51b802c 4650 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
eb7d8c07 4651 cfqg_put(cfqd->root_group);
1adaf3dd 4652 spin_unlock_irq(q->queue_lock);
1da177e4 4653
91148325
JK
4654 hrtimer_init(&cfqd->idle_slice_timer, CLOCK_MONOTONIC,
4655 HRTIMER_MODE_REL);
22e2c507 4656 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
22e2c507 4657
23e018a1 4658 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
22e2c507 4659
1da177e4 4660 cfqd->cfq_quantum = cfq_quantum;
22e2c507
JA
4661 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4662 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
1da177e4
LT
4663 cfqd->cfq_back_max = cfq_back_max;
4664 cfqd->cfq_back_penalty = cfq_back_penalty;
22e2c507
JA
4665 cfqd->cfq_slice[0] = cfq_slice_async;
4666 cfqd->cfq_slice[1] = cfq_slice_sync;
5bf14c07 4667 cfqd->cfq_target_latency = cfq_target_latency;
22e2c507 4668 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
0bb97947 4669 cfqd->cfq_slice_idle = cfq_slice_idle;
80bdf0c7 4670 cfqd->cfq_group_idle = cfq_group_idle;
963b72fc 4671 cfqd->cfq_latency = 1;
e459dd08 4672 cfqd->hw_tag = -1;
edc71131
CZ
4673 /*
4674 * we optimistically start assuming sync ops weren't delayed in last
4675 * second, in order to have larger depth for async operations.
4676 */
9a7f38c4 4677 cfqd->last_delayed_sync = ktime_get_ns() - NSEC_PER_SEC;
b2fab5ac 4678 return 0;
a2b1693b
TH
4679
4680out_free:
4681 kfree(cfqd);
d50235b7 4682 kobject_put(&eq->kobj);
a2b1693b 4683 return ret;
1da177e4
LT
4684}
4685
0bb97947
JA
4686static void cfq_registered_queue(struct request_queue *q)
4687{
4688 struct elevator_queue *e = q->elevator;
4689 struct cfq_data *cfqd = e->elevator_data;
4690
4691 /*
4692 * Default to IOPS mode with no idling for SSDs
4693 */
4694 if (blk_queue_nonrot(q))
4695 cfqd->cfq_slice_idle = 0;
4696}
4697
1da177e4
LT
4698/*
4699 * sysfs parts below -->
4700 */
1da177e4
LT
4701static ssize_t
4702cfq_var_show(unsigned int var, char *page)
4703{
176167ad 4704 return sprintf(page, "%u\n", var);
1da177e4
LT
4705}
4706
4707static ssize_t
4708cfq_var_store(unsigned int *var, const char *page, size_t count)
4709{
4710 char *p = (char *) page;
4711
4712 *var = simple_strtoul(p, &p, 10);
4713 return count;
4714}
4715
1da177e4 4716#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
b374d18a 4717static ssize_t __FUNC(struct elevator_queue *e, char *page) \
1da177e4 4718{ \
3d1ab40f 4719 struct cfq_data *cfqd = e->elevator_data; \
9a7f38c4 4720 u64 __data = __VAR; \
1da177e4 4721 if (__CONV) \
9a7f38c4 4722 __data = div_u64(__data, NSEC_PER_MSEC); \
1da177e4
LT
4723 return cfq_var_show(__data, (page)); \
4724}
4725SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
22e2c507
JA
4726SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4727SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
e572ec7e
AV
4728SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4729SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
22e2c507 4730SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
80bdf0c7 4731SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
22e2c507
JA
4732SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4733SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4734SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
963b72fc 4735SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
5bf14c07 4736SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
1da177e4
LT
4737#undef SHOW_FUNCTION
4738
d2d481d0
JM
4739#define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
4740static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4741{ \
4742 struct cfq_data *cfqd = e->elevator_data; \
4743 u64 __data = __VAR; \
4744 __data = div_u64(__data, NSEC_PER_USEC); \
4745 return cfq_var_show(__data, (page)); \
4746}
4747USEC_SHOW_FUNCTION(cfq_slice_idle_us_show, cfqd->cfq_slice_idle);
4748USEC_SHOW_FUNCTION(cfq_group_idle_us_show, cfqd->cfq_group_idle);
4749USEC_SHOW_FUNCTION(cfq_slice_sync_us_show, cfqd->cfq_slice[1]);
4750USEC_SHOW_FUNCTION(cfq_slice_async_us_show, cfqd->cfq_slice[0]);
4751USEC_SHOW_FUNCTION(cfq_target_latency_us_show, cfqd->cfq_target_latency);
4752#undef USEC_SHOW_FUNCTION
4753
1da177e4 4754#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
b374d18a 4755static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
1da177e4 4756{ \
3d1ab40f 4757 struct cfq_data *cfqd = e->elevator_data; \
1da177e4
LT
4758 unsigned int __data; \
4759 int ret = cfq_var_store(&__data, (page), count); \
4760 if (__data < (MIN)) \
4761 __data = (MIN); \
4762 else if (__data > (MAX)) \
4763 __data = (MAX); \
4764 if (__CONV) \
9a7f38c4 4765 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
1da177e4
LT
4766 else \
4767 *(__PTR) = __data; \
4768 return ret; \
4769}
4770STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
fe094d98
JA
4771STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4772 UINT_MAX, 1);
4773STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4774 UINT_MAX, 1);
e572ec7e 4775STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
fe094d98
JA
4776STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4777 UINT_MAX, 0);
22e2c507 4778STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
80bdf0c7 4779STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
22e2c507
JA
4780STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4781STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
fe094d98
JA
4782STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4783 UINT_MAX, 0);
963b72fc 4784STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
5bf14c07 4785STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
1da177e4
LT
4786#undef STORE_FUNCTION
4787
d2d481d0
JM
4788#define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
4789static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4790{ \
4791 struct cfq_data *cfqd = e->elevator_data; \
4792 unsigned int __data; \
4793 int ret = cfq_var_store(&__data, (page), count); \
4794 if (__data < (MIN)) \
4795 __data = (MIN); \
4796 else if (__data > (MAX)) \
4797 __data = (MAX); \
4798 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
4799 return ret; \
4800}
4801USEC_STORE_FUNCTION(cfq_slice_idle_us_store, &cfqd->cfq_slice_idle, 0, UINT_MAX);
4802USEC_STORE_FUNCTION(cfq_group_idle_us_store, &cfqd->cfq_group_idle, 0, UINT_MAX);
4803USEC_STORE_FUNCTION(cfq_slice_sync_us_store, &cfqd->cfq_slice[1], 1, UINT_MAX);
4804USEC_STORE_FUNCTION(cfq_slice_async_us_store, &cfqd->cfq_slice[0], 1, UINT_MAX);
4805USEC_STORE_FUNCTION(cfq_target_latency_us_store, &cfqd->cfq_target_latency, 1, UINT_MAX);
4806#undef USEC_STORE_FUNCTION
4807
e572ec7e
AV
4808#define CFQ_ATTR(name) \
4809 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4810
4811static struct elv_fs_entry cfq_attrs[] = {
4812 CFQ_ATTR(quantum),
e572ec7e
AV
4813 CFQ_ATTR(fifo_expire_sync),
4814 CFQ_ATTR(fifo_expire_async),
4815 CFQ_ATTR(back_seek_max),
4816 CFQ_ATTR(back_seek_penalty),
4817 CFQ_ATTR(slice_sync),
d2d481d0 4818 CFQ_ATTR(slice_sync_us),
e572ec7e 4819 CFQ_ATTR(slice_async),
d2d481d0 4820 CFQ_ATTR(slice_async_us),
e572ec7e
AV
4821 CFQ_ATTR(slice_async_rq),
4822 CFQ_ATTR(slice_idle),
d2d481d0 4823 CFQ_ATTR(slice_idle_us),
80bdf0c7 4824 CFQ_ATTR(group_idle),
d2d481d0 4825 CFQ_ATTR(group_idle_us),
963b72fc 4826 CFQ_ATTR(low_latency),
5bf14c07 4827 CFQ_ATTR(target_latency),
d2d481d0 4828 CFQ_ATTR(target_latency_us),
e572ec7e 4829 __ATTR_NULL
1da177e4
LT
4830};
4831
1da177e4
LT
4832static struct elevator_type iosched_cfq = {
4833 .ops = {
4834 .elevator_merge_fn = cfq_merge,
4835 .elevator_merged_fn = cfq_merged_request,
4836 .elevator_merge_req_fn = cfq_merged_requests,
72ef799b
TE
4837 .elevator_allow_bio_merge_fn = cfq_allow_bio_merge,
4838 .elevator_allow_rq_merge_fn = cfq_allow_rq_merge,
812d4026 4839 .elevator_bio_merged_fn = cfq_bio_merged,
b4878f24 4840 .elevator_dispatch_fn = cfq_dispatch_requests,
1da177e4 4841 .elevator_add_req_fn = cfq_insert_request,
b4878f24 4842 .elevator_activate_req_fn = cfq_activate_request,
1da177e4 4843 .elevator_deactivate_req_fn = cfq_deactivate_request,
1da177e4 4844 .elevator_completed_req_fn = cfq_completed_request,
21183b07
JA
4845 .elevator_former_req_fn = elv_rb_former_request,
4846 .elevator_latter_req_fn = elv_rb_latter_request,
9b84cacd 4847 .elevator_init_icq_fn = cfq_init_icq,
7e5a8794 4848 .elevator_exit_icq_fn = cfq_exit_icq,
1da177e4
LT
4849 .elevator_set_req_fn = cfq_set_request,
4850 .elevator_put_req_fn = cfq_put_request,
4851 .elevator_may_queue_fn = cfq_may_queue,
4852 .elevator_init_fn = cfq_init_queue,
4853 .elevator_exit_fn = cfq_exit_queue,
0bb97947 4854 .elevator_registered_fn = cfq_registered_queue,
1da177e4 4855 },
3d3c2379
TH
4856 .icq_size = sizeof(struct cfq_io_cq),
4857 .icq_align = __alignof__(struct cfq_io_cq),
3d1ab40f 4858 .elevator_attrs = cfq_attrs,
3d3c2379 4859 .elevator_name = "cfq",
1da177e4
LT
4860 .elevator_owner = THIS_MODULE,
4861};
4862
3e252066 4863#ifdef CONFIG_CFQ_GROUP_IOSCHED
3c798398 4864static struct blkcg_policy blkcg_policy_cfq = {
2ee867dc 4865 .dfl_cftypes = cfq_blkcg_files,
880f50e2 4866 .legacy_cftypes = cfq_blkcg_legacy_files,
f9fcc2d3 4867
e4a9bde9 4868 .cpd_alloc_fn = cfq_cpd_alloc,
e48453c3 4869 .cpd_init_fn = cfq_cpd_init,
e4a9bde9 4870 .cpd_free_fn = cfq_cpd_free,
69d7fde5 4871 .cpd_bind_fn = cfq_cpd_bind,
e4a9bde9 4872
001bea73 4873 .pd_alloc_fn = cfq_pd_alloc,
f9fcc2d3 4874 .pd_init_fn = cfq_pd_init,
0b39920b 4875 .pd_offline_fn = cfq_pd_offline,
001bea73 4876 .pd_free_fn = cfq_pd_free,
f9fcc2d3 4877 .pd_reset_stats_fn = cfq_pd_reset_stats,
3e252066 4878};
3e252066
VG
4879#endif
4880
1da177e4
LT
4881static int __init cfq_init(void)
4882{
3d3c2379
TH
4883 int ret;
4884
80bdf0c7 4885#ifdef CONFIG_CFQ_GROUP_IOSCHED
3c798398 4886 ret = blkcg_policy_register(&blkcg_policy_cfq);
8bd435b3
TH
4887 if (ret)
4888 return ret;
ffea73fc
TH
4889#else
4890 cfq_group_idle = 0;
4891#endif
8bd435b3 4892
fd794956 4893 ret = -ENOMEM;
3d3c2379
TH
4894 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4895 if (!cfq_pool)
8bd435b3 4896 goto err_pol_unreg;
1da177e4 4897
3d3c2379 4898 ret = elv_register(&iosched_cfq);
8bd435b3
TH
4899 if (ret)
4900 goto err_free_pool;
3d3c2379 4901
2fdd82bd 4902 return 0;
8bd435b3
TH
4903
4904err_free_pool:
4905 kmem_cache_destroy(cfq_pool);
4906err_pol_unreg:
ffea73fc 4907#ifdef CONFIG_CFQ_GROUP_IOSCHED
3c798398 4908 blkcg_policy_unregister(&blkcg_policy_cfq);
ffea73fc 4909#endif
8bd435b3 4910 return ret;
1da177e4
LT
4911}
4912
4913static void __exit cfq_exit(void)
4914{
ffea73fc 4915#ifdef CONFIG_CFQ_GROUP_IOSCHED
3c798398 4916 blkcg_policy_unregister(&blkcg_policy_cfq);
ffea73fc 4917#endif
1da177e4 4918 elv_unregister(&iosched_cfq);
3d3c2379 4919 kmem_cache_destroy(cfq_pool);
1da177e4
LT
4920}
4921
4922module_init(cfq_init);
4923module_exit(cfq_exit);
4924
4925MODULE_AUTHOR("Jens Axboe");
4926MODULE_LICENSE("GPL");
4927MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");