blob: e62e9205b80a8047193cd32a690a0e5d5b689b85 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * 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 *
Jens Axboe0fe23472006-09-04 15:41:16 +02007 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090010#include <linux/slab.h>
Al Viro1cc9be62006-03-18 12:29:52 -050011#include <linux/blkdev.h>
12#include <linux/elevator.h>
Randy Dunlapad5ebd22009-11-11 13:47:45 +010013#include <linux/jiffies.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020015#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020016#include <linux/blktrace_api.h>
Tejun Heo6e736be2011-12-14 00:33:38 +010017#include "blk.h"
Tejun Heo629ed0b2012-04-01 14:38:44 -070018#include "blk-cgroup.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070019
20/*
21 * tunables
22 */
Jens Axboefe094d92008-01-31 13:08:54 +010023/* max queue in one round of service */
Shaohua Liabc3c742010-03-01 09:20:54 +010024static const int cfq_quantum = 8;
Arjan van de Ven64100092006-01-06 09:46:02 +010025static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010026/* 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;
Arjan van de Ven64100092006-01-06 09:46:02 +010030static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020031static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010032static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020033static int cfq_slice_idle = HZ / 125;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +020034static int cfq_group_idle = HZ / 125;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +010035static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36static const int cfq_hist_divisor = 4;
Jens Axboe22e2c502005-06-27 10:55:12 +020037
Jens Axboed9e76202007-04-20 14:27:50 +020038/*
Jens Axboe08717142008-01-28 11:38:15 +010039 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020040 */
Jens Axboe08717142008-01-28 11:38:15 +010041#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020042
43/*
44 * below this threshold, we consider thinktime immediate
45 */
46#define CFQ_MIN_TT (2)
47
Jens Axboe22e2c502005-06-27 10:55:12 +020048#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020049#define CFQ_HW_QUEUE_MIN (5)
Vivek Goyal25bc6b02009-12-03 12:59:43 -050050#define CFQ_SERVICE_SHIFT 12
Jens Axboe22e2c502005-06-27 10:55:12 +020051
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010052#define CFQQ_SEEK_THR (sector_t)(8 * 100)
Shaohua Lie9ce3352010-03-19 08:03:04 +010053#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
Corrado Zoccolo41647e72010-02-27 19:45:40 +010054#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010055#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
Shaohua Liae54abe2010-02-05 13:11:45 +010056
Tejun Heoa612fdd2011-12-14 00:33:41 +010057#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])
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Christoph Lametere18b8902006-12-06 20:33:20 -080061static struct kmem_cache *cfq_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
Jens Axboe22e2c502005-06-27 10:55:12 +020063#define CFQ_PRIO_LISTS IOPRIO_BE_NR
64#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020065#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
66
Jens Axboe206dc692006-03-28 13:03:44 +020067#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050068#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020069
Tejun Heoc5869802011-12-14 00:33:41 +010070struct cfq_ttime {
71 unsigned long last_end_request;
72
73 unsigned long ttime_total;
74 unsigned long ttime_samples;
75 unsigned long ttime_mean;
76};
77
Jens Axboe22e2c502005-06-27 10:55:12 +020078/*
Jens Axboecc09e292007-04-26 12:53:50 +020079 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
83 */
84struct cfq_rb_root {
85 struct rb_root rb;
86 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010087 unsigned count;
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010088 unsigned total_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050089 u64 min_vdisktime;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +020090 struct cfq_ttime ttime;
Jens Axboecc09e292007-04-26 12:53:50 +020091};
Shaohua Lif5f2b6c2011-07-12 14:24:55 +020092#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
93 .ttime = {.last_end_request = jiffies,},}
Jens Axboecc09e292007-04-26 12:53:50 +020094
95/*
Jens Axboe6118b702009-06-30 09:34:12 +020096 * Per process-grouping structure
97 */
98struct cfq_queue {
99 /* reference count */
Shaohua Li30d7b942011-01-07 08:46:59 +0100100 int ref;
Jens Axboe6118b702009-06-30 09:34:12 +0200101 /* various state flags, see below */
102 unsigned int flags;
103 /* parent cfq_data */
104 struct cfq_data *cfqd;
105 /* service_tree member */
106 struct rb_node rb_node;
107 /* service_tree key */
108 unsigned long rb_key;
109 /* prio tree member */
110 struct rb_node p_node;
111 /* prio tree root we belong to, if any */
112 struct rb_root *p_root;
113 /* sorted list of pending requests */
114 struct rb_root sort_list;
115 /* if fifo isn't expired, next request to serve */
116 struct request *next_rq;
117 /* requests queued in sort_list */
118 int queued[2];
119 /* currently allocated requests */
120 int allocated[2];
121 /* fifo list of requests in sort_list */
122 struct list_head fifo;
123
Vivek Goyaldae739e2009-12-03 12:59:45 -0500124 /* time when queue got scheduled in to dispatch first request. */
125 unsigned long dispatch_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500126 unsigned int allocated_slice;
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100127 unsigned int slice_dispatch;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500128 /* time when first request from queue completed and slice started. */
129 unsigned long slice_start;
Jens Axboe6118b702009-06-30 09:34:12 +0200130 unsigned long slice_end;
131 long slice_resid;
Jens Axboe6118b702009-06-30 09:34:12 +0200132
Christoph Hellwig65299a32011-08-23 14:50:29 +0200133 /* pending priority requests */
134 int prio_pending;
Jens Axboe6118b702009-06-30 09:34:12 +0200135 /* number of requests that are on the dispatch list or inside driver */
136 int dispatched;
137
138 /* io prio of this group */
139 unsigned short ioprio, org_ioprio;
Justin TerAvest4aede842011-07-12 08:31:45 +0200140 unsigned short ioprio_class;
Jens Axboe6118b702009-06-30 09:34:12 +0200141
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100142 pid_t pid;
143
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +0100144 u32 seek_history;
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400145 sector_t last_request_pos;
146
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100147 struct cfq_rb_root *service_tree;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400148 struct cfq_queue *new_cfqq;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500149 struct cfq_group *cfqg;
Vivek Goyalc4e78932010-08-23 12:25:03 +0200150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
Jens Axboe6118b702009-06-30 09:34:12 +0200152};
153
154/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100155 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100156 * IDLE is handled separately, so it has negative index
157 */
158enum wl_prio_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100159 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500160 RT_WORKLOAD = 1,
161 IDLE_WORKLOAD = 2,
Vivek Goyalb4627322010-10-22 09:48:43 +0200162 CFQ_PRIO_NR,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100163};
164
165/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100166 * Second index in the service_trees.
167 */
168enum wl_type_t {
169 ASYNC_WORKLOAD = 0,
170 SYNC_NOIDLE_WORKLOAD = 1,
171 SYNC_WORKLOAD = 2
172};
173
Tejun Heo155fead2012-04-01 14:38:44 -0700174struct cfqg_stats {
175#ifdef CONFIG_CFQ_GROUP_IOSCHED
176 /* total bytes transferred */
177 struct blkg_rwstat service_bytes;
178 /* total IOs serviced, post merge */
179 struct blkg_rwstat serviced;
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;
188 /* total sectors transferred */
189 struct blkg_stat sectors;
190 /* total disk time and nr sectors dispatched by this group */
191 struct blkg_stat time;
192#ifdef CONFIG_DEBUG_BLK_CGROUP
193 /* time not charged to this cgroup */
194 struct blkg_stat unaccounted_time;
195 /* sum of number of ios queued across all samples */
196 struct blkg_stat avg_queue_size_sum;
197 /* count of samples taken for average */
198 struct blkg_stat avg_queue_size_samples;
199 /* how many times this group has been removed from service tree */
200 struct blkg_stat dequeue;
201 /* total time spent waiting for it to be assigned a timeslice. */
202 struct blkg_stat group_wait_time;
Tejun Heo3c798392012-04-16 13:57:25 -0700203 /* time spent idling for this blkcg_gq */
Tejun Heo155fead2012-04-01 14:38:44 -0700204 struct blkg_stat idle_time;
205 /* total time with empty current active q with other requests queued */
206 struct blkg_stat empty_time;
207 /* fields after this shouldn't be cleared on stat reset */
208 uint64_t start_group_wait_time;
209 uint64_t start_idle_time;
210 uint64_t start_empty_time;
211 uint16_t flags;
212#endif /* CONFIG_DEBUG_BLK_CGROUP */
213#endif /* CONFIG_CFQ_GROUP_IOSCHED */
214};
215
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500216/* This is per cgroup per device grouping structure */
217struct cfq_group {
Tejun Heof95a04a2012-04-16 13:57:26 -0700218 /* must be the first member */
219 struct blkg_policy_data pd;
220
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500221 /* group service_tree member */
222 struct rb_node rb_node;
223
224 /* group service_tree key */
225 u64 vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500226 unsigned int weight;
Justin TerAvest8184f932011-03-17 16:12:36 +0100227 unsigned int new_weight;
Tejun Heo3381cb82012-04-01 14:38:44 -0700228 unsigned int dev_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500229
230 /* number of cfqq currently on this group */
231 int nr_cfqq;
232
Jens Axboe22e2c502005-06-27 10:55:12 +0200233 /*
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200234 * Per group busy queues average. Useful for workload slice calc. We
Vivek Goyalb4627322010-10-22 09:48:43 +0200235 * create the array for each prio class but at run time it is used
236 * only for RT and BE class and slot for IDLE class remains unused.
237 * This is primarily done to avoid confusion and a gcc warning.
238 */
239 unsigned int busy_queues_avg[CFQ_PRIO_NR];
240 /*
241 * rr lists of queues with requests. We maintain service trees for
242 * RT and BE classes. These trees are subdivided in subclasses
243 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
244 * class there is no subclassification and all the cfq queues go on
245 * a single tree service_tree_idle.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100246 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200247 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100248 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100249 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500250
251 unsigned long saved_workload_slice;
252 enum wl_type_t saved_workload;
253 enum wl_prio_t saved_serving_prio;
Tejun Heo4eef3042012-03-05 13:15:18 -0800254
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200255 /* number of requests that are on the dispatch list or inside driver */
256 int dispatched;
Shaohua Li7700fc42011-07-12 14:24:56 +0200257 struct cfq_ttime ttime;
Tejun Heo155fead2012-04-01 14:38:44 -0700258 struct cfqg_stats stats;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500259};
260
Tejun Heoc5869802011-12-14 00:33:41 +0100261struct cfq_io_cq {
262 struct io_cq icq; /* must be the first member */
263 struct cfq_queue *cfqq[2];
264 struct cfq_ttime ttime;
Tejun Heo598971b2012-03-19 15:10:58 -0700265 int ioprio; /* the current ioprio */
266#ifdef CONFIG_CFQ_GROUP_IOSCHED
267 uint64_t blkcg_id; /* the current blkcg ID */
268#endif
Tejun Heoc5869802011-12-14 00:33:41 +0100269};
270
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500271/*
272 * Per block device queue structure
273 */
274struct cfq_data {
275 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500276 /* Root service tree for cfq_groups */
277 struct cfq_rb_root grp_service_tree;
Tejun Heof51b8022012-03-05 13:15:05 -0800278 struct cfq_group *root_group;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500279
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100280 /*
281 * The priority currently being served
282 */
283 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100284 enum wl_type_t serving_type;
285 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500286 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200287
288 /*
289 * Each priority tree is sorted by next_request position. These
290 * trees are used when determining if two or more queues are
291 * interleaving requests (see cfq_close_cooperator).
292 */
293 struct rb_root prio_trees[CFQ_PRIO_LISTS];
294
Jens Axboe22e2c502005-06-27 10:55:12 +0200295 unsigned int busy_queues;
Shaohua Lief8a41d2011-03-07 09:26:29 +0100296 unsigned int busy_sync_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +0200297
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100298 int rq_in_driver;
299 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200300
301 /*
302 * queue-depth detection
303 */
304 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200305 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100306 /*
307 * hw_tag can be
308 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
309 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
310 * 0 => no NCQ
311 */
312 int hw_tag_est_depth;
313 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200314
315 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200316 * idle window management
317 */
318 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200319 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200320
321 struct cfq_queue *active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +0100322 struct cfq_io_cq *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200323
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200324 /*
325 * async queue for each priority case
326 */
327 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
328 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200329
Jens Axboe6d048f52007-04-25 12:44:27 +0200330 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 /*
333 * tunables, see top of file
334 */
335 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200336 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 unsigned int cfq_back_penalty;
338 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200339 unsigned int cfq_slice[2];
340 unsigned int cfq_slice_async_rq;
341 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200342 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200343 unsigned int cfq_latency;
Tao Ma5bf14c02012-04-01 14:33:39 -0700344 unsigned int cfq_target_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500345
Jens Axboe6118b702009-06-30 09:34:12 +0200346 /*
347 * Fallback dummy cfqq for extreme OOM conditions
348 */
349 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200350
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100351 unsigned long last_delayed_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352};
353
Vivek Goyal25fb5162009-12-03 12:59:46 -0500354static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
355
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500356static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
357 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500358 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100359{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500360 if (!cfqg)
361 return NULL;
362
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100363 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500364 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100365
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500366 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100367}
368
Jens Axboe3b181522005-06-27 10:56:24 +0200369enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100370 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
371 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200372 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100373 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100374 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
375 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
376 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100377 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200378 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400379 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100380 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100381 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500382 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200383};
384
385#define CFQ_CFQQ_FNS(name) \
386static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
387{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100388 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200389} \
390static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
391{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100392 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200393} \
394static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
395{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100396 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200397}
398
399CFQ_CFQQ_FNS(on_rr);
400CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200401CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200402CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200403CFQ_CFQQ_FNS(fifo_expire);
404CFQ_CFQQ_FNS(idle_window);
405CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100406CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200407CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200408CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100409CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100410CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500411CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200412#undef CFQ_CFQQ_FNS
413
Tejun Heof95a04a2012-04-16 13:57:26 -0700414static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
415{
416 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
417}
418
Tejun Heof95a04a2012-04-16 13:57:26 -0700419static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
420{
421 return pd_to_blkg(&cfqg->pd);
422}
423
Tejun Heo629ed0b2012-04-01 14:38:44 -0700424#if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700425
Tejun Heo155fead2012-04-01 14:38:44 -0700426/* cfqg stats flags */
427enum cfqg_stats_flags {
428 CFQG_stats_waiting = 0,
429 CFQG_stats_idling,
430 CFQG_stats_empty,
Tejun Heo629ed0b2012-04-01 14:38:44 -0700431};
432
Tejun Heo155fead2012-04-01 14:38:44 -0700433#define CFQG_FLAG_FNS(name) \
434static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700435{ \
Tejun Heo155fead2012-04-01 14:38:44 -0700436 stats->flags |= (1 << CFQG_stats_##name); \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700437} \
Tejun Heo155fead2012-04-01 14:38:44 -0700438static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700439{ \
Tejun Heo155fead2012-04-01 14:38:44 -0700440 stats->flags &= ~(1 << CFQG_stats_##name); \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700441} \
Tejun Heo155fead2012-04-01 14:38:44 -0700442static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700443{ \
Tejun Heo155fead2012-04-01 14:38:44 -0700444 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700445} \
446
Tejun Heo155fead2012-04-01 14:38:44 -0700447CFQG_FLAG_FNS(waiting)
448CFQG_FLAG_FNS(idling)
449CFQG_FLAG_FNS(empty)
450#undef CFQG_FLAG_FNS
Tejun Heo629ed0b2012-04-01 14:38:44 -0700451
452/* This should be called with the queue_lock held. */
Tejun Heo155fead2012-04-01 14:38:44 -0700453static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700454{
455 unsigned long long now;
456
Tejun Heo155fead2012-04-01 14:38:44 -0700457 if (!cfqg_stats_waiting(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700458 return;
459
460 now = sched_clock();
461 if (time_after64(now, stats->start_group_wait_time))
462 blkg_stat_add(&stats->group_wait_time,
463 now - stats->start_group_wait_time);
Tejun Heo155fead2012-04-01 14:38:44 -0700464 cfqg_stats_clear_waiting(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700465}
466
467/* This should be called with the queue_lock held. */
Tejun Heo155fead2012-04-01 14:38:44 -0700468static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
469 struct cfq_group *curr_cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700470{
Tejun Heo155fead2012-04-01 14:38:44 -0700471 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700472
Tejun Heo155fead2012-04-01 14:38:44 -0700473 if (cfqg_stats_waiting(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700474 return;
Tejun Heo155fead2012-04-01 14:38:44 -0700475 if (cfqg == curr_cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700476 return;
Tejun Heo155fead2012-04-01 14:38:44 -0700477 stats->start_group_wait_time = sched_clock();
478 cfqg_stats_mark_waiting(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700479}
480
481/* This should be called with the queue_lock held. */
Tejun Heo155fead2012-04-01 14:38:44 -0700482static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700483{
484 unsigned long long now;
485
Tejun Heo155fead2012-04-01 14:38:44 -0700486 if (!cfqg_stats_empty(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700487 return;
488
489 now = sched_clock();
490 if (time_after64(now, stats->start_empty_time))
491 blkg_stat_add(&stats->empty_time,
492 now - stats->start_empty_time);
Tejun Heo155fead2012-04-01 14:38:44 -0700493 cfqg_stats_clear_empty(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700494}
495
Tejun Heo155fead2012-04-01 14:38:44 -0700496static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700497{
Tejun Heo155fead2012-04-01 14:38:44 -0700498 blkg_stat_add(&cfqg->stats.dequeue, 1);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700499}
500
Tejun Heo155fead2012-04-01 14:38:44 -0700501static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700502{
Tejun Heo155fead2012-04-01 14:38:44 -0700503 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700504
505 if (blkg_rwstat_sum(&stats->queued))
506 return;
507
508 /*
509 * group is already marked empty. This can happen if cfqq got new
510 * request in parent group and moved to this group while being added
511 * to service tree. Just ignore the event and move on.
512 */
Tejun Heo155fead2012-04-01 14:38:44 -0700513 if (cfqg_stats_empty(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700514 return;
515
516 stats->start_empty_time = sched_clock();
Tejun Heo155fead2012-04-01 14:38:44 -0700517 cfqg_stats_mark_empty(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700518}
519
Tejun Heo155fead2012-04-01 14:38:44 -0700520static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700521{
Tejun Heo155fead2012-04-01 14:38:44 -0700522 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700523
Tejun Heo155fead2012-04-01 14:38:44 -0700524 if (cfqg_stats_idling(stats)) {
Tejun Heo629ed0b2012-04-01 14:38:44 -0700525 unsigned long long now = sched_clock();
526
527 if (time_after64(now, stats->start_idle_time))
528 blkg_stat_add(&stats->idle_time,
529 now - stats->start_idle_time);
Tejun Heo155fead2012-04-01 14:38:44 -0700530 cfqg_stats_clear_idling(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700531 }
532}
533
Tejun Heo155fead2012-04-01 14:38:44 -0700534static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700535{
Tejun Heo155fead2012-04-01 14:38:44 -0700536 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700537
Tejun Heo155fead2012-04-01 14:38:44 -0700538 BUG_ON(cfqg_stats_idling(stats));
Tejun Heo629ed0b2012-04-01 14:38:44 -0700539
540 stats->start_idle_time = sched_clock();
Tejun Heo155fead2012-04-01 14:38:44 -0700541 cfqg_stats_mark_idling(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700542}
543
Tejun Heo155fead2012-04-01 14:38:44 -0700544static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700545{
Tejun Heo155fead2012-04-01 14:38:44 -0700546 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700547
548 blkg_stat_add(&stats->avg_queue_size_sum,
549 blkg_rwstat_sum(&stats->queued));
550 blkg_stat_add(&stats->avg_queue_size_samples, 1);
Tejun Heo155fead2012-04-01 14:38:44 -0700551 cfqg_stats_update_group_wait_time(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700552}
553
554#else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
555
Tejun Heof48ec1d2012-04-13 13:11:25 -0700556static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
557static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
558static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
559static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
560static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
561static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
562static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
Tejun Heo629ed0b2012-04-01 14:38:44 -0700563
564#endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
565
566#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo2ce4d502012-04-01 14:38:43 -0700567
Tejun Heoffea73f2012-06-04 10:02:29 +0200568static struct blkcg_policy blkcg_policy_cfq;
569
570static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
571{
572 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
573}
574
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100575static inline void cfqg_get(struct cfq_group *cfqg)
576{
577 return blkg_get(cfqg_to_blkg(cfqg));
578}
579
580static inline void cfqg_put(struct cfq_group *cfqg)
581{
582 return blkg_put(cfqg_to_blkg(cfqg));
583}
584
Tejun Heo54e7ed12012-04-16 13:57:23 -0700585#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
586 char __pbuf[128]; \
587 \
588 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
Vivek Goyal2868ef72009-12-03 12:59:48 -0500589 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
Tejun Heo54e7ed12012-04-16 13:57:23 -0700590 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
591 __pbuf, ##args); \
592} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500593
Tejun Heo54e7ed12012-04-16 13:57:23 -0700594#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
595 char __pbuf[128]; \
596 \
597 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
598 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
599} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500600
Tejun Heo155fead2012-04-01 14:38:44 -0700601static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
602 struct cfq_group *curr_cfqg, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700603{
Tejun Heo155fead2012-04-01 14:38:44 -0700604 blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
605 cfqg_stats_end_empty_time(&cfqg->stats);
606 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700607}
608
Tejun Heo155fead2012-04-01 14:38:44 -0700609static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
610 unsigned long time, unsigned long unaccounted_time)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700611{
Tejun Heo155fead2012-04-01 14:38:44 -0700612 blkg_stat_add(&cfqg->stats.time, time);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700613#ifdef CONFIG_DEBUG_BLK_CGROUP
Tejun Heo155fead2012-04-01 14:38:44 -0700614 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700615#endif
Tejun Heo2ce4d502012-04-01 14:38:43 -0700616}
617
Tejun Heo155fead2012-04-01 14:38:44 -0700618static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700619{
Tejun Heo155fead2012-04-01 14:38:44 -0700620 blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700621}
622
Tejun Heo155fead2012-04-01 14:38:44 -0700623static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700624{
Tejun Heo155fead2012-04-01 14:38:44 -0700625 blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700626}
627
Tejun Heo155fead2012-04-01 14:38:44 -0700628static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
629 uint64_t bytes, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700630{
Tejun Heo155fead2012-04-01 14:38:44 -0700631 blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
632 blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
633 blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700634}
635
Tejun Heo155fead2012-04-01 14:38:44 -0700636static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
637 uint64_t start_time, uint64_t io_start_time, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700638{
Tejun Heo155fead2012-04-01 14:38:44 -0700639 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700640 unsigned long long now = sched_clock();
Tejun Heo629ed0b2012-04-01 14:38:44 -0700641
642 if (time_after64(now, io_start_time))
643 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
644 if (time_after64(io_start_time, start_time))
645 blkg_rwstat_add(&stats->wait_time, rw,
646 io_start_time - start_time);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700647}
648
Tejun Heo3c798392012-04-16 13:57:25 -0700649static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
Tejun Heo155fead2012-04-01 14:38:44 -0700650{
651 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
652 struct cfqg_stats *stats = &cfqg->stats;
653
654 /* queued stats shouldn't be cleared */
655 blkg_rwstat_reset(&stats->service_bytes);
656 blkg_rwstat_reset(&stats->serviced);
657 blkg_rwstat_reset(&stats->merged);
658 blkg_rwstat_reset(&stats->service_time);
659 blkg_rwstat_reset(&stats->wait_time);
660 blkg_stat_reset(&stats->time);
661#ifdef CONFIG_DEBUG_BLK_CGROUP
662 blkg_stat_reset(&stats->unaccounted_time);
663 blkg_stat_reset(&stats->avg_queue_size_sum);
664 blkg_stat_reset(&stats->avg_queue_size_samples);
665 blkg_stat_reset(&stats->dequeue);
666 blkg_stat_reset(&stats->group_wait_time);
667 blkg_stat_reset(&stats->idle_time);
668 blkg_stat_reset(&stats->empty_time);
669#endif
670}
671
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100672#else /* CONFIG_CFQ_GROUP_IOSCHED */
673
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100674static inline void cfqg_get(struct cfq_group *cfqg) { }
675static inline void cfqg_put(struct cfq_group *cfqg) { }
676
Jens Axboe7b679132008-05-30 12:23:07 +0200677#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
678 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200679#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100680
Tejun Heo155fead2012-04-01 14:38:44 -0700681static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
682 struct cfq_group *curr_cfqg, int rw) { }
683static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
684 unsigned long time, unsigned long unaccounted_time) { }
685static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
686static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
687static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
688 uint64_t bytes, int rw) { }
689static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
690 uint64_t start_time, uint64_t io_start_time, int rw) { }
Tejun Heo2ce4d502012-04-01 14:38:43 -0700691
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100692#endif /* CONFIG_CFQ_GROUP_IOSCHED */
693
Jens Axboe7b679132008-05-30 12:23:07 +0200694#define cfq_log(cfqd, fmt, args...) \
695 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
696
Vivek Goyal615f0252009-12-03 12:59:39 -0500697/* Traverses through cfq group service trees */
698#define for_each_cfqg_st(cfqg, i, j, st) \
699 for (i = 0; i <= IDLE_WORKLOAD; i++) \
700 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
701 : &cfqg->service_tree_idle; \
702 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
703 (i == IDLE_WORKLOAD && j == 0); \
704 j++, st = i < IDLE_WORKLOAD ? \
705 &cfqg->service_trees[i][j]: NULL) \
706
Shaohua Lif5f2b6c2011-07-12 14:24:55 +0200707static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
708 struct cfq_ttime *ttime, bool group_idle)
709{
710 unsigned long slice;
711 if (!sample_valid(ttime->ttime_samples))
712 return false;
713 if (group_idle)
714 slice = cfqd->cfq_group_idle;
715 else
716 slice = cfqd->cfq_slice_idle;
717 return ttime->ttime_mean > slice;
718}
Vivek Goyal615f0252009-12-03 12:59:39 -0500719
Vivek Goyal02b35082010-08-23 12:23:53 +0200720static inline bool iops_mode(struct cfq_data *cfqd)
721{
722 /*
723 * If we are not idling on queues and it is a NCQ drive, parallel
724 * execution of requests is on and measuring time is not possible
725 * in most of the cases until and unless we drive shallower queue
726 * depths and that becomes a performance bottleneck. In such cases
727 * switch to start providing fairness in terms of number of IOs.
728 */
729 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
730 return true;
731 else
732 return false;
733}
734
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100735static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
736{
737 if (cfq_class_idle(cfqq))
738 return IDLE_WORKLOAD;
739 if (cfq_class_rt(cfqq))
740 return RT_WORKLOAD;
741 return BE_WORKLOAD;
742}
743
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100744
745static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
746{
747 if (!cfq_cfqq_sync(cfqq))
748 return ASYNC_WORKLOAD;
749 if (!cfq_cfqq_idle_window(cfqq))
750 return SYNC_NOIDLE_WORKLOAD;
751 return SYNC_WORKLOAD;
752}
753
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500754static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
755 struct cfq_data *cfqd,
756 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100757{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500758 if (wl == IDLE_WORKLOAD)
759 return cfqg->service_tree_idle.count;
760
761 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
762 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
763 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100764}
765
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500766static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
767 struct cfq_group *cfqg)
768{
769 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
770 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
771}
772
Jens Axboe165125e2007-07-24 09:28:11 +0200773static void cfq_dispatch_insert(struct request_queue *, struct request *);
Tejun Heo4f85cb92012-03-05 13:15:28 -0800774static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
Tejun Heoabede6d2012-03-19 15:10:57 -0700775 struct cfq_io_cq *cic, struct bio *bio,
Tejun Heo4f85cb92012-03-05 13:15:28 -0800776 gfp_t gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200777
Tejun Heoc5869802011-12-14 00:33:41 +0100778static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
779{
780 /* cic->icq is the first member, %NULL will convert to %NULL */
781 return container_of(icq, struct cfq_io_cq, icq);
782}
783
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100784static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
785 struct io_context *ioc)
786{
787 if (ioc)
788 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
789 return NULL;
790}
791
Tejun Heoc5869802011-12-14 00:33:41 +0100792static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200793{
Jens Axboea6151c32009-10-07 20:02:57 +0200794 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200795}
796
Tejun Heoc5869802011-12-14 00:33:41 +0100797static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
798 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200799{
Jens Axboea6151c32009-10-07 20:02:57 +0200800 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200801}
802
Tejun Heoc5869802011-12-14 00:33:41 +0100803static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400804{
Tejun Heoc5869802011-12-14 00:33:41 +0100805 return cic->icq.q->elevator->elevator_data;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400806}
807
Vasily Tarasov91fac312007-04-25 12:29:51 +0200808/*
809 * We regard a request as SYNC, if it's either a read or has the SYNC bit
810 * set (in which case it could also be direct WRITE).
811 */
Jens Axboea6151c32009-10-07 20:02:57 +0200812static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200813{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200814 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200815}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700818 * scheduler run of queue, if there are requests pending and no one in the
819 * driver that will restart queueing
820 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200821static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700822{
Jens Axboe7b679132008-05-30 12:23:07 +0200823 if (cfqd->busy_queues) {
824 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200825 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200826 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700827}
828
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100830 * Scale schedule slice based on io priority. Use the sync time slice only
831 * if a queue is marked sync and has sync io queued. A sync queue with async
832 * io only, should not get full sync slice length.
833 */
Jens Axboea6151c32009-10-07 20:02:57 +0200834static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200835 unsigned short prio)
836{
837 const int base_slice = cfqd->cfq_slice[sync];
838
839 WARN_ON(prio >= IOPRIO_BE_NR);
840
841 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
842}
843
Jens Axboe44f7c162007-01-19 11:51:58 +1100844static inline int
845cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
846{
Jens Axboed9e76202007-04-20 14:27:50 +0200847 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100848}
849
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500850static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
851{
852 u64 d = delta << CFQ_SERVICE_SHIFT;
853
Tejun Heo3381cb82012-04-01 14:38:44 -0700854 d = d * CFQ_WEIGHT_DEFAULT;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500855 do_div(d, cfqg->weight);
856 return d;
857}
858
859static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
860{
861 s64 delta = (s64)(vdisktime - min_vdisktime);
862 if (delta > 0)
863 min_vdisktime = vdisktime;
864
865 return min_vdisktime;
866}
867
868static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
869{
870 s64 delta = (s64)(vdisktime - min_vdisktime);
871 if (delta < 0)
872 min_vdisktime = vdisktime;
873
874 return min_vdisktime;
875}
876
877static void update_min_vdisktime(struct cfq_rb_root *st)
878{
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500879 struct cfq_group *cfqg;
880
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500881 if (st->left) {
882 cfqg = rb_entry_cfqg(st->left);
Gui Jianfenga6032712011-03-07 09:28:09 +0100883 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
884 cfqg->vdisktime);
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500885 }
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500886}
887
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100888/*
889 * get averaged number of queues of RT/BE priority.
890 * average is updated, with a formula that gives more weight to higher numbers,
891 * to quickly follows sudden increases and decrease slowly
892 */
893
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500894static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
895 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100896{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100897 unsigned min_q, max_q;
898 unsigned mult = cfq_hist_divisor - 1;
899 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500900 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100901
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500902 min_q = min(cfqg->busy_queues_avg[rt], busy);
903 max_q = max(cfqg->busy_queues_avg[rt], busy);
904 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100905 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500906 return cfqg->busy_queues_avg[rt];
907}
908
909static inline unsigned
910cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
911{
912 struct cfq_rb_root *st = &cfqd->grp_service_tree;
913
Tao Ma5bf14c02012-04-01 14:33:39 -0700914 return cfqd->cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100915}
916
Shaohua Lic553f8e2011-01-14 08:41:03 +0100917static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -0700918cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100919{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100920 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
921 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500922 /*
923 * interested queues (we consider only the ones with the same
924 * priority class in the cfq group)
925 */
926 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
927 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100928 unsigned sync_slice = cfqd->cfq_slice[1];
929 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500930 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
931
932 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100933 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
934 /* scale low_slice according to IO priority
935 * and sync vs async */
936 unsigned low_slice =
937 min(slice, base_low_slice * slice / sync_slice);
938 /* the adapted slice value is scaled to fit all iqs
939 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500940 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100941 low_slice);
942 }
943 }
Shaohua Lic553f8e2011-01-14 08:41:03 +0100944 return slice;
945}
946
947static inline void
948cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
949{
Vivek Goyalba5bd522011-01-19 08:25:02 -0700950 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +0100951
Vivek Goyaldae739e2009-12-03 12:59:45 -0500952 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100953 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500954 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200955 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100956}
957
958/*
959 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
960 * isn't valid until the first request from the dispatch is activated
961 * and the slice time set.
962 */
Jens Axboea6151c32009-10-07 20:02:57 +0200963static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100964{
965 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +0100966 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100967 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +0100968 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100969
Shaohua Lic1e44752010-11-08 15:01:02 +0100970 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +1100971}
972
973/*
Jens Axboe5e705372006-07-13 12:39:25 +0200974 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200976 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 */
Jens Axboe5e705372006-07-13 12:39:25 +0200978static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100979cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100981 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200983#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
984#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
985 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986
Jens Axboe5e705372006-07-13 12:39:25 +0200987 if (rq1 == NULL || rq1 == rq2)
988 return rq2;
989 if (rq2 == NULL)
990 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200991
Namhyung Kim229836b2011-05-24 10:23:21 +0200992 if (rq_is_sync(rq1) != rq_is_sync(rq2))
993 return rq_is_sync(rq1) ? rq1 : rq2;
994
Christoph Hellwig65299a32011-08-23 14:50:29 +0200995 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
996 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
Jens Axboeb53d1ed2011-08-19 08:34:48 +0200997
Tejun Heo83096eb2009-05-07 22:24:39 +0900998 s1 = blk_rq_pos(rq1);
999 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001 /*
1002 * by definition, 1KiB is 2 sectors
1003 */
1004 back_max = cfqd->cfq_back_max * 2;
1005
1006 /*
1007 * Strict one way elevator _except_ in the case where we allow
1008 * short backward seeks which are biased as twice the cost of a
1009 * similar forward seek.
1010 */
1011 if (s1 >= last)
1012 d1 = s1 - last;
1013 else if (s1 + back_max >= last)
1014 d1 = (last - s1) * cfqd->cfq_back_penalty;
1015 else
Andreas Mohre8a99052006-03-28 08:59:49 +02001016 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017
1018 if (s2 >= last)
1019 d2 = s2 - last;
1020 else if (s2 + back_max >= last)
1021 d2 = (last - s2) * cfqd->cfq_back_penalty;
1022 else
Andreas Mohre8a99052006-03-28 08:59:49 +02001023 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024
1025 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026
Andreas Mohre8a99052006-03-28 08:59:49 +02001027 /*
1028 * By doing switch() on the bit mask "wrap" we avoid having to
1029 * check two variables for all permutations: --> faster!
1030 */
1031 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +02001032 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +02001033 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +02001034 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +02001035 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +02001036 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +02001037 else {
1038 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +02001039 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +02001040 else
Jens Axboe5e705372006-07-13 12:39:25 +02001041 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +02001042 }
1043
1044 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +02001045 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +02001046 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +02001047 return rq2;
1048 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +02001049 default:
1050 /*
1051 * Since both rqs are wrapped,
1052 * start with the one that's further behind head
1053 * (--> only *one* back seek required),
1054 * since back seek takes more time than forward.
1055 */
1056 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +02001057 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058 else
Jens Axboe5e705372006-07-13 12:39:25 +02001059 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060 }
1061}
1062
Jens Axboe498d3aa22007-04-26 12:54:48 +02001063/*
1064 * The below is leftmost cache rbtree addon
1065 */
Jens Axboe08717142008-01-28 11:38:15 +01001066static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +02001067{
Vivek Goyal615f0252009-12-03 12:59:39 -05001068 /* Service tree is empty */
1069 if (!root->count)
1070 return NULL;
1071
Jens Axboecc09e292007-04-26 12:53:50 +02001072 if (!root->left)
1073 root->left = rb_first(&root->rb);
1074
Jens Axboe08717142008-01-28 11:38:15 +01001075 if (root->left)
1076 return rb_entry(root->left, struct cfq_queue, rb_node);
1077
1078 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +02001079}
1080
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001081static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1082{
1083 if (!root->left)
1084 root->left = rb_first(&root->rb);
1085
1086 if (root->left)
1087 return rb_entry_cfqg(root->left);
1088
1089 return NULL;
1090}
1091
Jens Axboea36e71f2009-04-15 12:15:11 +02001092static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1093{
1094 rb_erase(n, root);
1095 RB_CLEAR_NODE(n);
1096}
1097
Jens Axboecc09e292007-04-26 12:53:50 +02001098static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1099{
1100 if (root->left == n)
1101 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001102 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001103 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +02001104}
1105
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106/*
1107 * would be nice to take fifo expire time into account as well
1108 */
Jens Axboe5e705372006-07-13 12:39:25 +02001109static struct request *
1110cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1111 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112{
Jens Axboe21183b02006-07-13 12:33:14 +02001113 struct rb_node *rbnext = rb_next(&last->rb_node);
1114 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +02001115 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116
Jens Axboe21183b02006-07-13 12:33:14 +02001117 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118
1119 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +02001120 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +02001121
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +02001123 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +02001124 else {
1125 rbnext = rb_first(&cfqq->sort_list);
1126 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +02001127 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +02001128 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001130 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131}
1132
Jens Axboed9e76202007-04-20 14:27:50 +02001133static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1134 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135{
Jens Axboed9e76202007-04-20 14:27:50 +02001136 /*
1137 * just an approximation, should be ok.
1138 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001139 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +01001140 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +02001141}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001143static inline s64
1144cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1145{
1146 return cfqg->vdisktime - st->min_vdisktime;
1147}
1148
1149static void
1150__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1151{
1152 struct rb_node **node = &st->rb.rb_node;
1153 struct rb_node *parent = NULL;
1154 struct cfq_group *__cfqg;
1155 s64 key = cfqg_key(st, cfqg);
1156 int left = 1;
1157
1158 while (*node != NULL) {
1159 parent = *node;
1160 __cfqg = rb_entry_cfqg(parent);
1161
1162 if (key < cfqg_key(st, __cfqg))
1163 node = &parent->rb_left;
1164 else {
1165 node = &parent->rb_right;
1166 left = 0;
1167 }
1168 }
1169
1170 if (left)
1171 st->left = &cfqg->rb_node;
1172
1173 rb_link_node(&cfqg->rb_node, parent, node);
1174 rb_insert_color(&cfqg->rb_node, &st->rb);
1175}
1176
1177static void
Justin TerAvest8184f932011-03-17 16:12:36 +01001178cfq_update_group_weight(struct cfq_group *cfqg)
1179{
1180 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
Tejun Heo3381cb82012-04-01 14:38:44 -07001181 if (cfqg->new_weight) {
Justin TerAvest8184f932011-03-17 16:12:36 +01001182 cfqg->weight = cfqg->new_weight;
Tejun Heo3381cb82012-04-01 14:38:44 -07001183 cfqg->new_weight = 0;
Justin TerAvest8184f932011-03-17 16:12:36 +01001184 }
1185}
1186
1187static void
1188cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1189{
1190 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1191
1192 cfq_update_group_weight(cfqg);
1193 __cfq_group_service_tree_add(st, cfqg);
1194 st->total_weight += cfqg->weight;
1195}
1196
1197static void
1198cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001199{
1200 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1201 struct cfq_group *__cfqg;
1202 struct rb_node *n;
1203
1204 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +01001205 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001206 return;
1207
1208 /*
1209 * Currently put the group at the end. Later implement something
1210 * so that groups get lesser vtime based on their weights, so that
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001211 * if group does not loose all if it was not continuously backlogged.
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001212 */
1213 n = rb_last(&st->rb);
1214 if (n) {
1215 __cfqg = rb_entry_cfqg(n);
1216 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1217 } else
1218 cfqg->vdisktime = st->min_vdisktime;
Justin TerAvest8184f932011-03-17 16:12:36 +01001219 cfq_group_service_tree_add(st, cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001220}
1221
1222static void
Justin TerAvest8184f932011-03-17 16:12:36 +01001223cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1224{
1225 st->total_weight -= cfqg->weight;
1226 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1227 cfq_rb_erase(&cfqg->rb_node, st);
1228}
1229
1230static void
1231cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001232{
1233 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1234
1235 BUG_ON(cfqg->nr_cfqq < 1);
1236 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05001237
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001238 /* If there are other cfq queues under this group, don't delete it */
1239 if (cfqg->nr_cfqq)
1240 return;
1241
Vivek Goyal2868ef72009-12-03 12:59:48 -05001242 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Justin TerAvest8184f932011-03-17 16:12:36 +01001243 cfq_group_service_tree_del(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001244 cfqg->saved_workload_slice = 0;
Tejun Heo155fead2012-04-01 14:38:44 -07001245 cfqg_stats_update_dequeue(cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001246}
1247
Justin TerAvest167400d2011-03-12 16:54:00 +01001248static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1249 unsigned int *unaccounted_time)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001250{
Vivek Goyalf75edf22009-12-03 12:59:53 -05001251 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001252
1253 /*
1254 * Queue got expired before even a single request completed or
1255 * got expired immediately after first request completion.
1256 */
1257 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1258 /*
1259 * Also charge the seek time incurred to the group, otherwise
1260 * if there are mutiple queues in the group, each can dispatch
1261 * a single request on seeky media and cause lots of seek time
1262 * and group will never know it.
1263 */
1264 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1265 1);
1266 } else {
1267 slice_used = jiffies - cfqq->slice_start;
Justin TerAvest167400d2011-03-12 16:54:00 +01001268 if (slice_used > cfqq->allocated_slice) {
1269 *unaccounted_time = slice_used - cfqq->allocated_slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001270 slice_used = cfqq->allocated_slice;
Justin TerAvest167400d2011-03-12 16:54:00 +01001271 }
1272 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1273 *unaccounted_time += cfqq->slice_start -
1274 cfqq->dispatch_start;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001275 }
1276
Vivek Goyaldae739e2009-12-03 12:59:45 -05001277 return slice_used;
1278}
1279
1280static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001281 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001282{
1283 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Justin TerAvest167400d2011-03-12 16:54:00 +01001284 unsigned int used_sl, charge, unaccounted_sl = 0;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05001285 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1286 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001287
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05001288 BUG_ON(nr_sync < 0);
Justin TerAvest167400d2011-03-12 16:54:00 +01001289 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05001290
Vivek Goyal02b35082010-08-23 12:23:53 +02001291 if (iops_mode(cfqd))
1292 charge = cfqq->slice_dispatch;
1293 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1294 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001295
1296 /* Can't update vdisktime while group is on service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +01001297 cfq_group_service_tree_del(st, cfqg);
Vivek Goyal02b35082010-08-23 12:23:53 +02001298 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Justin TerAvest8184f932011-03-17 16:12:36 +01001299 /* If a new weight was requested, update now, off tree */
1300 cfq_group_service_tree_add(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001301
1302 /* This group is being expired. Save the context */
1303 if (time_after(cfqd->workload_expires, jiffies)) {
1304 cfqg->saved_workload_slice = cfqd->workload_expires
1305 - jiffies;
1306 cfqg->saved_workload = cfqd->serving_type;
1307 cfqg->saved_serving_prio = cfqd->serving_prio;
1308 } else
1309 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -05001310
1311 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1312 st->min_vdisktime);
Joe Perchesfd16d262011-06-13 10:42:49 +02001313 cfq_log_cfqq(cfqq->cfqd, cfqq,
1314 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1315 used_sl, cfqq->slice_dispatch, charge,
1316 iops_mode(cfqd), cfqq->nr_sectors);
Tejun Heo155fead2012-04-01 14:38:44 -07001317 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1318 cfqg_stats_set_start_empty_time(cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001319}
1320
Tejun Heof51b8022012-03-05 13:15:05 -08001321/**
1322 * cfq_init_cfqg_base - initialize base part of a cfq_group
1323 * @cfqg: cfq_group to initialize
1324 *
1325 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1326 * is enabled or not.
1327 */
1328static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1329{
1330 struct cfq_rb_root *st;
1331 int i, j;
1332
1333 for_each_cfqg_st(cfqg, i, j, st)
1334 *st = CFQ_RB_ROOT;
1335 RB_CLEAR_NODE(&cfqg->rb_node);
1336
1337 cfqg->ttime.last_end_request = jiffies;
1338}
1339
Vivek Goyal25fb5162009-12-03 12:59:46 -05001340#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07001341static void cfq_pd_init(struct blkcg_gq *blkg)
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001342{
Tejun Heo03814112012-03-05 13:15:14 -08001343 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001344
Tejun Heof51b8022012-03-05 13:15:05 -08001345 cfq_init_cfqg_base(cfqg);
Tejun Heo3381cb82012-04-01 14:38:44 -07001346 cfqg->weight = blkg->blkcg->cfq_weight;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001347}
1348
1349/*
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001350 * Search for the cfq group current task belongs to. request_queue lock must
1351 * be held.
Vivek Goyal25fb5162009-12-03 12:59:46 -05001352 */
Tejun Heocd1604f2012-03-05 13:15:06 -08001353static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
Tejun Heo3c798392012-04-16 13:57:25 -07001354 struct blkcg *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001355{
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001356 struct request_queue *q = cfqd->queue;
Tejun Heocd1604f2012-03-05 13:15:06 -08001357 struct cfq_group *cfqg = NULL;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001358
Tejun Heo3c798392012-04-16 13:57:25 -07001359 /* avoid lookup for the common case where there's no blkcg */
1360 if (blkcg == &blkcg_root) {
Tejun Heocd1604f2012-03-05 13:15:06 -08001361 cfqg = cfqd->root_group;
1362 } else {
Tejun Heo3c798392012-04-16 13:57:25 -07001363 struct blkcg_gq *blkg;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001364
Tejun Heo3c96cb32012-04-13 13:11:34 -07001365 blkg = blkg_lookup_create(blkcg, q);
Tejun Heocd1604f2012-03-05 13:15:06 -08001366 if (!IS_ERR(blkg))
Tejun Heo03814112012-03-05 13:15:14 -08001367 cfqg = blkg_to_cfqg(blkg);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001368 }
1369
Vivek Goyal25fb5162009-12-03 12:59:46 -05001370 return cfqg;
1371}
1372
1373static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1374{
1375 /* Currently, all async queues are mapped to root group */
1376 if (!cfq_cfqq_sync(cfqq))
Tejun Heof51b8022012-03-05 13:15:05 -08001377 cfqg = cfqq->cfqd->root_group;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001378
1379 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001380 /* cfqq reference on cfqg */
Tejun Heoeb7d8c072012-03-23 14:02:53 +01001381 cfqg_get(cfqg);
Vivek Goyalb1c35762009-12-03 12:59:47 -05001382}
1383
Tejun Heof95a04a2012-04-16 13:57:26 -07001384static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1385 struct blkg_policy_data *pd, int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001386{
Tejun Heof95a04a2012-04-16 13:57:26 -07001387 struct cfq_group *cfqg = pd_to_cfqg(pd);
Tejun Heo3381cb82012-04-01 14:38:44 -07001388
1389 if (!cfqg->dev_weight)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001390 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -07001391 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001392}
1393
Tejun Heo3381cb82012-04-01 14:38:44 -07001394static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1395 struct seq_file *sf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001396{
Tejun Heo3c798392012-04-16 13:57:25 -07001397 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1398 cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001399 false);
1400 return 0;
1401}
1402
Tejun Heo3381cb82012-04-01 14:38:44 -07001403static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1404 struct seq_file *sf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001405{
Tejun Heo3c798392012-04-16 13:57:25 -07001406 seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001407 return 0;
1408}
1409
Tejun Heo3381cb82012-04-01 14:38:44 -07001410static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1411 const char *buf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001412{
Tejun Heo3c798392012-04-16 13:57:25 -07001413 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001414 struct blkg_conf_ctx ctx;
Tejun Heo3381cb82012-04-01 14:38:44 -07001415 struct cfq_group *cfqg;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001416 int ret;
1417
Tejun Heo3c798392012-04-16 13:57:25 -07001418 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001419 if (ret)
1420 return ret;
1421
1422 ret = -EINVAL;
Tejun Heo3381cb82012-04-01 14:38:44 -07001423 cfqg = blkg_to_cfqg(ctx.blkg);
Tejun Heoa2b16932012-04-13 13:11:33 -07001424 if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
Tejun Heo3381cb82012-04-01 14:38:44 -07001425 cfqg->dev_weight = ctx.v;
1426 cfqg->new_weight = cfqg->dev_weight ?: blkcg->cfq_weight;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001427 ret = 0;
1428 }
1429
1430 blkg_conf_finish(&ctx);
1431 return ret;
1432}
1433
Tejun Heo3381cb82012-04-01 14:38:44 -07001434static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001435{
Tejun Heo3c798392012-04-16 13:57:25 -07001436 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1437 struct blkcg_gq *blkg;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001438 struct hlist_node *n;
1439
Tejun Heo3381cb82012-04-01 14:38:44 -07001440 if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001441 return -EINVAL;
1442
1443 spin_lock_irq(&blkcg->lock);
Tejun Heo3381cb82012-04-01 14:38:44 -07001444 blkcg->cfq_weight = (unsigned int)val;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001445
1446 hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
Tejun Heo3381cb82012-04-01 14:38:44 -07001447 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001448
Tejun Heo3381cb82012-04-01 14:38:44 -07001449 if (cfqg && !cfqg->dev_weight)
1450 cfqg->new_weight = blkcg->cfq_weight;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001451 }
1452
1453 spin_unlock_irq(&blkcg->lock);
1454 return 0;
1455}
1456
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001457static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1458 struct seq_file *sf)
1459{
Tejun Heo3c798392012-04-16 13:57:25 -07001460 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001461
Tejun Heo3c798392012-04-16 13:57:25 -07001462 blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001463 cft->private, false);
1464 return 0;
1465}
1466
1467static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1468 struct seq_file *sf)
1469{
Tejun Heo3c798392012-04-16 13:57:25 -07001470 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001471
Tejun Heo3c798392012-04-16 13:57:25 -07001472 blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001473 cft->private, true);
1474 return 0;
1475}
1476
Tejun Heo60c2bc22012-04-01 14:38:43 -07001477#ifdef CONFIG_DEBUG_BLK_CGROUP
Tejun Heof95a04a2012-04-16 13:57:26 -07001478static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1479 struct blkg_policy_data *pd, int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001480{
Tejun Heof95a04a2012-04-16 13:57:26 -07001481 struct cfq_group *cfqg = pd_to_cfqg(pd);
Tejun Heo155fead2012-04-01 14:38:44 -07001482 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001483 u64 v = 0;
1484
1485 if (samples) {
Tejun Heo155fead2012-04-01 14:38:44 -07001486 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001487 do_div(v, samples);
1488 }
Tejun Heof95a04a2012-04-16 13:57:26 -07001489 __blkg_prfill_u64(sf, pd, v);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001490 return 0;
1491}
1492
1493/* print avg_queue_size */
Tejun Heo155fead2012-04-01 14:38:44 -07001494static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1495 struct seq_file *sf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001496{
Tejun Heo3c798392012-04-16 13:57:25 -07001497 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001498
Tejun Heo155fead2012-04-01 14:38:44 -07001499 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
Tejun Heo3c798392012-04-16 13:57:25 -07001500 &blkcg_policy_cfq, 0, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001501 return 0;
1502}
1503#endif /* CONFIG_DEBUG_BLK_CGROUP */
1504
1505static struct cftype cfq_blkcg_files[] = {
1506 {
1507 .name = "weight_device",
Tejun Heo3381cb82012-04-01 14:38:44 -07001508 .read_seq_string = cfqg_print_weight_device,
1509 .write_string = cfqg_set_weight_device,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001510 .max_write_len = 256,
1511 },
1512 {
1513 .name = "weight",
Tejun Heo3381cb82012-04-01 14:38:44 -07001514 .read_seq_string = cfq_print_weight,
1515 .write_u64 = cfq_set_weight,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001516 },
1517 {
1518 .name = "time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001519 .private = offsetof(struct cfq_group, stats.time),
1520 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001521 },
1522 {
1523 .name = "sectors",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001524 .private = offsetof(struct cfq_group, stats.sectors),
1525 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001526 },
1527 {
1528 .name = "io_service_bytes",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001529 .private = offsetof(struct cfq_group, stats.service_bytes),
1530 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001531 },
1532 {
1533 .name = "io_serviced",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001534 .private = offsetof(struct cfq_group, stats.serviced),
1535 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001536 },
1537 {
1538 .name = "io_service_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001539 .private = offsetof(struct cfq_group, stats.service_time),
1540 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001541 },
1542 {
1543 .name = "io_wait_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001544 .private = offsetof(struct cfq_group, stats.wait_time),
1545 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001546 },
1547 {
1548 .name = "io_merged",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001549 .private = offsetof(struct cfq_group, stats.merged),
1550 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001551 },
1552 {
1553 .name = "io_queued",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001554 .private = offsetof(struct cfq_group, stats.queued),
1555 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001556 },
1557#ifdef CONFIG_DEBUG_BLK_CGROUP
1558 {
1559 .name = "avg_queue_size",
Tejun Heo155fead2012-04-01 14:38:44 -07001560 .read_seq_string = cfqg_print_avg_queue_size,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001561 },
1562 {
1563 .name = "group_wait_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001564 .private = offsetof(struct cfq_group, stats.group_wait_time),
1565 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001566 },
1567 {
1568 .name = "idle_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001569 .private = offsetof(struct cfq_group, stats.idle_time),
1570 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001571 },
1572 {
1573 .name = "empty_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001574 .private = offsetof(struct cfq_group, stats.empty_time),
1575 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001576 },
1577 {
1578 .name = "dequeue",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001579 .private = offsetof(struct cfq_group, stats.dequeue),
1580 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001581 },
1582 {
1583 .name = "unaccounted_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001584 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1585 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001586 },
1587#endif /* CONFIG_DEBUG_BLK_CGROUP */
1588 { } /* terminate */
1589};
Vivek Goyal25fb5162009-12-03 12:59:46 -05001590#else /* GROUP_IOSCHED */
Tejun Heocd1604f2012-03-05 13:15:06 -08001591static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
Tejun Heo3c798392012-04-16 13:57:25 -07001592 struct blkcg *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001593{
Tejun Heof51b8022012-03-05 13:15:05 -08001594 return cfqd->root_group;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001595}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001596
Vivek Goyal25fb5162009-12-03 12:59:46 -05001597static inline void
1598cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1599 cfqq->cfqg = cfqg;
1600}
1601
1602#endif /* GROUP_IOSCHED */
1603
Jens Axboe498d3aa22007-04-26 12:54:48 +02001604/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001605 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001606 * requests waiting to be processed. It is sorted in the order that
1607 * we will service the queues.
1608 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001609static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001610 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001611{
Jens Axboe08717142008-01-28 11:38:15 +01001612 struct rb_node **p, *parent;
1613 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001614 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001615 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001616 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001617 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001618
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001619 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001620 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001621 if (cfq_class_idle(cfqq)) {
1622 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001623 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001624 if (parent && parent != &cfqq->rb_node) {
1625 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1626 rb_key += __cfqq->rb_key;
1627 } else
1628 rb_key += jiffies;
1629 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001630 /*
1631 * Get our rb key offset. Subtract any residual slice
1632 * value carried from last service. A negative resid
1633 * count indicates slice overrun, and this should position
1634 * the next service time further away in the tree.
1635 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001636 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001637 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001638 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001639 } else {
1640 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001641 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001642 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1643 }
Jens Axboed9e76202007-04-20 14:27:50 +02001644
1645 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001646 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001647 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001648 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001649 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001650 if (rb_key == cfqq->rb_key &&
1651 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001652 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001653
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001654 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1655 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001656 }
Jens Axboed9e76202007-04-20 14:27:50 +02001657
Jens Axboe498d3aa22007-04-26 12:54:48 +02001658 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001659 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001660 cfqq->service_tree = service_tree;
1661 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001662 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001663 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001664
Jens Axboed9e76202007-04-20 14:27:50 +02001665 parent = *p;
1666 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1667
Jens Axboe0c534e02007-04-18 20:01:57 +02001668 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001669 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001670 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001671 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001672 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001673 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001674 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001675 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001676 }
Jens Axboe67060e32007-04-18 20:13:32 +02001677
1678 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001679 }
1680
Jens Axboecc09e292007-04-26 12:53:50 +02001681 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001682 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001683
Jens Axboed9e76202007-04-20 14:27:50 +02001684 cfqq->rb_key = rb_key;
1685 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001686 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1687 service_tree->count++;
Namhyung Kim20359f22011-05-24 10:23:22 +02001688 if (add_front || !new_cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001689 return;
Justin TerAvest8184f932011-03-17 16:12:36 +01001690 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691}
1692
Jens Axboea36e71f2009-04-15 12:15:11 +02001693static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001694cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1695 sector_t sector, struct rb_node **ret_parent,
1696 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001697{
Jens Axboea36e71f2009-04-15 12:15:11 +02001698 struct rb_node **p, *parent;
1699 struct cfq_queue *cfqq = NULL;
1700
1701 parent = NULL;
1702 p = &root->rb_node;
1703 while (*p) {
1704 struct rb_node **n;
1705
1706 parent = *p;
1707 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1708
1709 /*
1710 * Sort strictly based on sector. Smallest to the left,
1711 * largest to the right.
1712 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001713 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001714 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001715 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001716 n = &(*p)->rb_left;
1717 else
1718 break;
1719 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001720 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001721 }
1722
1723 *ret_parent = parent;
1724 if (rb_link)
1725 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001726 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001727}
1728
1729static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1730{
Jens Axboea36e71f2009-04-15 12:15:11 +02001731 struct rb_node **p, *parent;
1732 struct cfq_queue *__cfqq;
1733
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001734 if (cfqq->p_root) {
1735 rb_erase(&cfqq->p_node, cfqq->p_root);
1736 cfqq->p_root = NULL;
1737 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001738
1739 if (cfq_class_idle(cfqq))
1740 return;
1741 if (!cfqq->next_rq)
1742 return;
1743
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001744 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001745 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1746 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001747 if (!__cfqq) {
1748 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001749 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1750 } else
1751 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001752}
1753
Jens Axboe498d3aa22007-04-26 12:54:48 +02001754/*
1755 * Update cfqq's position in the service tree.
1756 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001757static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001758{
Jens Axboe6d048f52007-04-25 12:44:27 +02001759 /*
1760 * Resorting requires the cfqq to be on the RR list already.
1761 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001762 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001763 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001764 cfq_prio_tree_add(cfqd, cfqq);
1765 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001766}
1767
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768/*
1769 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001770 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771 */
Jens Axboefebffd62008-01-28 13:19:43 +01001772static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773{
Jens Axboe7b679132008-05-30 12:23:07 +02001774 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001775 BUG_ON(cfq_cfqq_on_rr(cfqq));
1776 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 cfqd->busy_queues++;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001778 if (cfq_cfqq_sync(cfqq))
1779 cfqd->busy_sync_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780
Jens Axboeedd75ff2007-04-19 12:03:34 +02001781 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782}
1783
Jens Axboe498d3aa22007-04-26 12:54:48 +02001784/*
1785 * Called when the cfqq no longer has requests pending, remove it from
1786 * the service tree.
1787 */
Jens Axboefebffd62008-01-28 13:19:43 +01001788static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789{
Jens Axboe7b679132008-05-30 12:23:07 +02001790 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001791 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1792 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001794 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1795 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1796 cfqq->service_tree = NULL;
1797 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001798 if (cfqq->p_root) {
1799 rb_erase(&cfqq->p_node, cfqq->p_root);
1800 cfqq->p_root = NULL;
1801 }
Jens Axboed9e76202007-04-20 14:27:50 +02001802
Justin TerAvest8184f932011-03-17 16:12:36 +01001803 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804 BUG_ON(!cfqd->busy_queues);
1805 cfqd->busy_queues--;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001806 if (cfq_cfqq_sync(cfqq))
1807 cfqd->busy_sync_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808}
1809
1810/*
1811 * rb tree support functions
1812 */
Jens Axboefebffd62008-01-28 13:19:43 +01001813static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814{
Jens Axboe5e705372006-07-13 12:39:25 +02001815 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001816 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817
Jens Axboeb4878f22005-10-20 16:42:29 +02001818 BUG_ON(!cfqq->queued[sync]);
1819 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820
Jens Axboe5e705372006-07-13 12:39:25 +02001821 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822
Vivek Goyalf04a6422009-12-03 12:59:40 -05001823 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1824 /*
1825 * Queue will be deleted from service tree when we actually
1826 * expire it later. Right now just remove it from prio tree
1827 * as it is empty.
1828 */
1829 if (cfqq->p_root) {
1830 rb_erase(&cfqq->p_node, cfqq->p_root);
1831 cfqq->p_root = NULL;
1832 }
1833 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834}
1835
Jens Axboe5e705372006-07-13 12:39:25 +02001836static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837{
Jens Axboe5e705372006-07-13 12:39:25 +02001838 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839 struct cfq_data *cfqd = cfqq->cfqd;
Jeff Moyer796d5112011-06-02 21:19:05 +02001840 struct request *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841
Jens Axboe5380a102006-07-13 12:37:56 +02001842 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843
Jeff Moyer796d5112011-06-02 21:19:05 +02001844 elv_rb_add(&cfqq->sort_list, rq);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001845
1846 if (!cfq_cfqq_on_rr(cfqq))
1847 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001848
1849 /*
1850 * check if this request is a better next-serve candidate
1851 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001852 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001853 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001854
1855 /*
1856 * adjust priority tree position, if ->next_rq changes
1857 */
1858 if (prev != cfqq->next_rq)
1859 cfq_prio_tree_add(cfqd, cfqq);
1860
Jens Axboe5044eed2007-04-25 11:53:48 +02001861 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862}
1863
Jens Axboefebffd62008-01-28 13:19:43 +01001864static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865{
Jens Axboe5380a102006-07-13 12:37:56 +02001866 elv_rb_del(&cfqq->sort_list, rq);
1867 cfqq->queued[rq_is_sync(rq)]--;
Tejun Heo155fead2012-04-01 14:38:44 -07001868 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
Jens Axboe5e705372006-07-13 12:39:25 +02001869 cfq_add_rq_rb(rq);
Tejun Heo155fead2012-04-01 14:38:44 -07001870 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
1871 rq->cmd_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872}
1873
Jens Axboe206dc692006-03-28 13:03:44 +02001874static struct request *
1875cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876{
Jens Axboe206dc692006-03-28 13:03:44 +02001877 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01001878 struct cfq_io_cq *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001879 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880
Jens Axboe4ac845a2008-01-24 08:44:49 +01001881 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001882 if (!cic)
1883 return NULL;
1884
1885 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001886 if (cfqq) {
1887 sector_t sector = bio->bi_sector + bio_sectors(bio);
1888
Jens Axboe21183b02006-07-13 12:33:14 +02001889 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001890 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 return NULL;
1893}
1894
Jens Axboe165125e2007-07-24 09:28:11 +02001895static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001896{
1897 struct cfq_data *cfqd = q->elevator->elevator_data;
1898
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001899 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001900 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001901 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001902
Tejun Heo5b936292009-05-07 22:24:38 +09001903 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001904}
1905
Jens Axboe165125e2007-07-24 09:28:11 +02001906static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907{
Jens Axboe22e2c502005-06-27 10:55:12 +02001908 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001910 WARN_ON(!cfqd->rq_in_driver);
1911 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001912 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001913 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914}
1915
Jens Axboeb4878f22005-10-20 16:42:29 +02001916static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917{
Jens Axboe5e705372006-07-13 12:39:25 +02001918 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001919
Jens Axboe5e705372006-07-13 12:39:25 +02001920 if (cfqq->next_rq == rq)
1921 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922
Jens Axboeb4878f22005-10-20 16:42:29 +02001923 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001924 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001925
Aaron Carroll45333d52008-08-26 15:52:36 +02001926 cfqq->cfqd->rq_queued--;
Tejun Heo155fead2012-04-01 14:38:44 -07001927 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
Christoph Hellwig65299a32011-08-23 14:50:29 +02001928 if (rq->cmd_flags & REQ_PRIO) {
1929 WARN_ON(!cfqq->prio_pending);
1930 cfqq->prio_pending--;
Jens Axboeb53d1ed2011-08-19 08:34:48 +02001931 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932}
1933
Jens Axboe165125e2007-07-24 09:28:11 +02001934static int cfq_merge(struct request_queue *q, struct request **req,
1935 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936{
1937 struct cfq_data *cfqd = q->elevator->elevator_data;
1938 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939
Jens Axboe206dc692006-03-28 13:03:44 +02001940 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001941 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001942 *req = __rq;
1943 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944 }
1945
1946 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947}
1948
Jens Axboe165125e2007-07-24 09:28:11 +02001949static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001950 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951{
Jens Axboe21183b02006-07-13 12:33:14 +02001952 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001953 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954
Jens Axboe5e705372006-07-13 12:39:25 +02001955 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957}
1958
Divyesh Shah812d4022010-04-08 21:14:23 -07001959static void cfq_bio_merged(struct request_queue *q, struct request *req,
1960 struct bio *bio)
1961{
Tejun Heo155fead2012-04-01 14:38:44 -07001962 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
Divyesh Shah812d4022010-04-08 21:14:23 -07001963}
1964
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965static void
Jens Axboe165125e2007-07-24 09:28:11 +02001966cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967 struct request *next)
1968{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001969 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Shaohua Li4a0b75c2011-12-16 14:00:22 +01001970 struct cfq_data *cfqd = q->elevator->elevator_data;
1971
Jens Axboe22e2c502005-06-27 10:55:12 +02001972 /*
1973 * reposition in fifo if next is older than rq
1974 */
1975 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Shaohua Li3d106fba2012-11-06 12:39:51 +01001976 time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
1977 cfqq == RQ_CFQQ(next)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001978 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001979 rq_set_fifo_time(rq, rq_fifo_time(next));
1980 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001981
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001982 if (cfqq->next_rq == next)
1983 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001984 cfq_remove_request(next);
Tejun Heo155fead2012-04-01 14:38:44 -07001985 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
Shaohua Li4a0b75c2011-12-16 14:00:22 +01001986
1987 cfqq = RQ_CFQQ(next);
1988 /*
1989 * all requests of this queue are merged to other queues, delete it
1990 * from the service tree. If it's the active_queue,
1991 * cfq_dispatch_requests() will choose to expire it or do idle
1992 */
1993 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1994 cfqq != cfqd->active_queue)
1995 cfq_del_cfqq_rr(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001996}
1997
Jens Axboe165125e2007-07-24 09:28:11 +02001998static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001999 struct bio *bio)
2000{
2001 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heoc5869802011-12-14 00:33:41 +01002002 struct cfq_io_cq *cic;
Jens Axboeda775262006-12-20 11:04:12 +01002003 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01002004
2005 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01002006 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01002007 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02002008 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02002009 return false;
Jens Axboeda775262006-12-20 11:04:12 +01002010
2011 /*
Tejun Heof1a4f4d2011-12-14 00:33:39 +01002012 * Lookup the cfqq that this bio will be queued with and allow
Tejun Heo07c2bd32012-02-08 09:19:42 +01002013 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01002014 */
Tejun Heo07c2bd32012-02-08 09:19:42 +01002015 cic = cfq_cic_lookup(cfqd, current->io_context);
2016 if (!cic)
2017 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01002018
Vasily Tarasov91fac312007-04-25 12:29:51 +02002019 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02002020 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01002021}
2022
Divyesh Shah812df482010-04-08 21:15:35 -07002023static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2024{
2025 del_timer(&cfqd->idle_slice_timer);
Tejun Heo155fead2012-04-01 14:38:44 -07002026 cfqg_stats_update_idle_time(cfqq->cfqg);
Divyesh Shah812df482010-04-08 21:15:35 -07002027}
2028
Jens Axboefebffd62008-01-28 13:19:43 +01002029static void __cfq_set_active_queue(struct cfq_data *cfqd,
2030 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002031{
2032 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002033 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
2034 cfqd->serving_prio, cfqd->serving_type);
Tejun Heo155fead2012-04-01 14:38:44 -07002035 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
Justin TerAvest62a37f62011-03-23 08:25:44 +01002036 cfqq->slice_start = 0;
2037 cfqq->dispatch_start = jiffies;
2038 cfqq->allocated_slice = 0;
2039 cfqq->slice_end = 0;
2040 cfqq->slice_dispatch = 0;
2041 cfqq->nr_sectors = 0;
2042
2043 cfq_clear_cfqq_wait_request(cfqq);
2044 cfq_clear_cfqq_must_dispatch(cfqq);
2045 cfq_clear_cfqq_must_alloc_slice(cfqq);
2046 cfq_clear_cfqq_fifo_expire(cfqq);
2047 cfq_mark_cfqq_slice_new(cfqq);
2048
2049 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002050 }
2051
2052 cfqd->active_queue = cfqq;
2053}
2054
2055/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002056 * current cfqq expired its slice (or was too idle), select new one
2057 */
2058static void
2059__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02002060 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002061{
Jens Axboe7b679132008-05-30 12:23:07 +02002062 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2063
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002064 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07002065 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002066
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002067 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05002068 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002069
2070 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01002071 * If this cfqq is shared between multiple processes, check to
2072 * make sure that those processes are still issuing I/Os within
2073 * the mean seek distance. If not, it may be time to break the
2074 * queues apart again.
2075 */
2076 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2077 cfq_mark_cfqq_split_coop(cfqq);
2078
2079 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02002080 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002081 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01002082 if (timed_out) {
2083 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07002084 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01002085 else
2086 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02002087 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2088 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002089
Vivek Goyale5ff0822010-04-26 19:25:11 +02002090 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05002091
Vivek Goyalf04a6422009-12-03 12:59:40 -05002092 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2093 cfq_del_cfqq_rr(cfqd, cfqq);
2094
Jens Axboeedd75ff2007-04-19 12:03:34 +02002095 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002096
2097 if (cfqq == cfqd->active_queue)
2098 cfqd->active_queue = NULL;
2099
2100 if (cfqd->active_cic) {
Tejun Heo11a31222012-02-07 07:51:30 +01002101 put_io_context(cfqd->active_cic->icq.ioc);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002102 cfqd->active_cic = NULL;
2103 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002104}
2105
Vivek Goyale5ff0822010-04-26 19:25:11 +02002106static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002107{
2108 struct cfq_queue *cfqq = cfqd->active_queue;
2109
2110 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02002111 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002112}
2113
Jens Axboe498d3aa22007-04-26 12:54:48 +02002114/*
2115 * Get next queue for service. Unless we have a queue preemption,
2116 * we'll simply select the first cfqq in the service tree.
2117 */
Jens Axboe6d048f52007-04-25 12:44:27 +02002118static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002119{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002120 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002121 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002122 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02002123
Vivek Goyalf04a6422009-12-03 12:59:40 -05002124 if (!cfqd->rq_queued)
2125 return NULL;
2126
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002127 /* There is nothing to dispatch */
2128 if (!service_tree)
2129 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002130 if (RB_EMPTY_ROOT(&service_tree->rb))
2131 return NULL;
2132 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02002133}
2134
Vivek Goyalf04a6422009-12-03 12:59:40 -05002135static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2136{
Vivek Goyal25fb5162009-12-03 12:59:46 -05002137 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05002138 struct cfq_queue *cfqq;
2139 int i, j;
2140 struct cfq_rb_root *st;
2141
2142 if (!cfqd->rq_queued)
2143 return NULL;
2144
Vivek Goyal25fb5162009-12-03 12:59:46 -05002145 cfqg = cfq_get_next_cfqg(cfqd);
2146 if (!cfqg)
2147 return NULL;
2148
Vivek Goyalf04a6422009-12-03 12:59:40 -05002149 for_each_cfqg_st(cfqg, i, j, st)
2150 if ((cfqq = cfq_rb_first(st)) != NULL)
2151 return cfqq;
2152 return NULL;
2153}
2154
Jens Axboe498d3aa22007-04-26 12:54:48 +02002155/*
2156 * Get and set a new active queue for service.
2157 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002158static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2159 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002160{
Jens Axboee00ef792009-11-04 08:54:55 +01002161 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02002162 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02002163
Jens Axboe22e2c502005-06-27 10:55:12 +02002164 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02002165 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002166}
2167
Jens Axboed9e76202007-04-20 14:27:50 +02002168static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2169 struct request *rq)
2170{
Tejun Heo83096eb2009-05-07 22:24:39 +09002171 if (blk_rq_pos(rq) >= cfqd->last_position)
2172 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02002173 else
Tejun Heo83096eb2009-05-07 22:24:39 +09002174 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02002175}
2176
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002177static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01002178 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002179{
Shaohua Lie9ce3352010-03-19 08:03:04 +01002180 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02002181}
2182
Jens Axboea36e71f2009-04-15 12:15:11 +02002183static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2184 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002185{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002186 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02002187 struct rb_node *parent, *node;
2188 struct cfq_queue *__cfqq;
2189 sector_t sector = cfqd->last_position;
2190
2191 if (RB_EMPTY_ROOT(root))
2192 return NULL;
2193
2194 /*
2195 * First, if we find a request starting at the end of the last
2196 * request, choose it.
2197 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002198 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02002199 if (__cfqq)
2200 return __cfqq;
2201
2202 /*
2203 * If the exact sector wasn't found, the parent of the NULL leaf
2204 * will contain the closest sector.
2205 */
2206 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01002207 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002208 return __cfqq;
2209
Tejun Heo2e46e8b2009-05-07 22:24:41 +09002210 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02002211 node = rb_next(&__cfqq->p_node);
2212 else
2213 node = rb_prev(&__cfqq->p_node);
2214 if (!node)
2215 return NULL;
2216
2217 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01002218 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002219 return __cfqq;
2220
2221 return NULL;
2222}
2223
2224/*
2225 * cfqd - obvious
2226 * cur_cfqq - passed in so that we don't decide that the current queue is
2227 * closely cooperating with itself.
2228 *
2229 * So, basically we're assuming that that cur_cfqq has dispatched at least
2230 * one request, and that cfqd->last_position reflects a position on the disk
2231 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2232 * assumption.
2233 */
2234static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002235 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02002236{
2237 struct cfq_queue *cfqq;
2238
Divyesh Shah39c01b22010-03-25 15:45:57 +01002239 if (cfq_class_idle(cur_cfqq))
2240 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002241 if (!cfq_cfqq_sync(cur_cfqq))
2242 return NULL;
2243 if (CFQQ_SEEKY(cur_cfqq))
2244 return NULL;
2245
Jens Axboea36e71f2009-04-15 12:15:11 +02002246 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01002247 * Don't search priority tree if it's the only queue in the group.
2248 */
2249 if (cur_cfqq->cfqg->nr_cfqq == 1)
2250 return NULL;
2251
2252 /*
Jens Axboed9e76202007-04-20 14:27:50 +02002253 * We should notice if some of the queues are cooperating, eg
2254 * working closely on the same area of the disk. In that case,
2255 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02002256 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002257 cfqq = cfqq_close(cfqd, cur_cfqq);
2258 if (!cfqq)
2259 return NULL;
2260
Vivek Goyal8682e1f2009-12-03 12:59:50 -05002261 /* If new queue belongs to different cfq_group, don't choose it */
2262 if (cur_cfqq->cfqg != cfqq->cfqg)
2263 return NULL;
2264
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002265 /*
2266 * It only makes sense to merge sync queues.
2267 */
2268 if (!cfq_cfqq_sync(cfqq))
2269 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002270 if (CFQQ_SEEKY(cfqq))
2271 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002272
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002273 /*
2274 * Do not merge queues of different priority classes
2275 */
2276 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2277 return NULL;
2278
Jens Axboea36e71f2009-04-15 12:15:11 +02002279 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02002280}
2281
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002282/*
2283 * Determine whether we should enforce idle window for this queue.
2284 */
2285
2286static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2287{
2288 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002289 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002290
Vivek Goyalf04a6422009-12-03 12:59:40 -05002291 BUG_ON(!service_tree);
2292 BUG_ON(!service_tree->count);
2293
Vivek Goyalb6508c12010-08-23 12:23:33 +02002294 if (!cfqd->cfq_slice_idle)
2295 return false;
2296
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002297 /* We never do for idle class queues. */
2298 if (prio == IDLE_WORKLOAD)
2299 return false;
2300
2301 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01002302 if (cfq_cfqq_idle_window(cfqq) &&
2303 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002304 return true;
2305
2306 /*
2307 * Otherwise, we do only if they are the last ones
2308 * in their service tree.
2309 */
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002310 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2311 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
Shaohua Lic1e44752010-11-08 15:01:02 +01002312 return true;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002313 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2314 service_tree->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01002315 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002316}
2317
Jens Axboe6d048f52007-04-25 12:44:27 +02002318static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002319{
Jens Axboe17926692007-01-19 11:59:30 +11002320 struct cfq_queue *cfqq = cfqd->active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +01002321 struct cfq_io_cq *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002322 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002323
Jens Axboea68bbddba2008-09-24 13:03:33 +02002324 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002325 * SSD device without seek penalty, disable idling. But only do so
2326 * for devices that support queuing, otherwise we still have a problem
2327 * with sync vs async workloads.
Jens Axboea68bbddba2008-09-24 13:03:33 +02002328 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002329 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbddba2008-09-24 13:03:33 +02002330 return;
2331
Jens Axboedd67d052006-06-21 09:36:18 +02002332 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02002333 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02002334
2335 /*
2336 * idle is disabled, either manually or by past process history
2337 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002338 if (!cfq_should_idle(cfqd, cfqq)) {
2339 /* no queue idling. Check for group idling */
2340 if (cfqd->cfq_group_idle)
2341 group_idle = cfqd->cfq_group_idle;
2342 else
2343 return;
2344 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002345
Jens Axboe22e2c502005-06-27 10:55:12 +02002346 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002347 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02002348 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002349 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02002350 return;
2351
2352 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002353 * task has exited, don't wait
2354 */
Jens Axboe206dc692006-03-28 13:03:44 +02002355 cic = cfqd->active_cic;
Tejun Heof6e8d012012-03-05 13:15:26 -08002356 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
Jens Axboe6d048f52007-04-25 12:44:27 +02002357 return;
2358
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002359 /*
2360 * If our average think time is larger than the remaining time
2361 * slice, then don't idle. This avoids overrunning the allotted
2362 * time slice.
2363 */
Shaohua Li383cd722011-07-12 14:24:35 +02002364 if (sample_valid(cic->ttime.ttime_samples) &&
2365 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
Joe Perchesfd16d262011-06-13 10:42:49 +02002366 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
Shaohua Li383cd722011-07-12 14:24:35 +02002367 cic->ttime.ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002368 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002369 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002370
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002371 /* There are other queues in the group, don't do group idle */
2372 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2373 return;
2374
Jens Axboe3b181522005-06-27 10:56:24 +02002375 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002376
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002377 if (group_idle)
2378 sl = cfqd->cfq_group_idle;
2379 else
2380 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02002381
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002382 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Tejun Heo155fead2012-04-01 14:38:44 -07002383 cfqg_stats_set_start_idle_time(cfqq->cfqg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002384 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2385 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386}
2387
Jens Axboe498d3aa22007-04-26 12:54:48 +02002388/*
2389 * Move request from internal lists to the request queue dispatch list.
2390 */
Jens Axboe165125e2007-07-24 09:28:11 +02002391static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392{
Jens Axboe3ed9a292007-04-23 08:33:33 +02002393 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002394 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002395
Jens Axboe7b679132008-05-30 12:23:07 +02002396 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2397
Jeff Moyer06d21882009-09-11 17:08:59 +02002398 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02002399 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002400 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002401 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02002402 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02002403
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002404 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02002405 cfqq->nr_sectors += blk_rq_sectors(rq);
Tejun Heo155fead2012-04-01 14:38:44 -07002406 cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407}
2408
2409/*
2410 * return expired entry, or NULL to just start from scratch in rbtree
2411 */
Jens Axboefebffd62008-01-28 13:19:43 +01002412static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413{
Jens Axboe30996f42009-10-05 11:03:39 +02002414 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415
Jens Axboe3b181522005-06-27 10:56:24 +02002416 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002418
2419 cfq_mark_cfqq_fifo_expire(cfqq);
2420
Jens Axboe89850f72006-07-22 16:48:31 +02002421 if (list_empty(&cfqq->fifo))
2422 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423
Jens Axboe89850f72006-07-22 16:48:31 +02002424 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002425 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002426 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427
Jens Axboe30996f42009-10-05 11:03:39 +02002428 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002429 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430}
2431
Jens Axboe22e2c502005-06-27 10:55:12 +02002432static inline int
2433cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2434{
2435 const int base_rq = cfqd->cfq_slice_async_rq;
2436
2437 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2438
Namhyung Kimb9f8ce02011-05-24 10:23:21 +02002439 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002440}
2441
2442/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002443 * Must be called with the queue_lock held.
2444 */
2445static int cfqq_process_refs(struct cfq_queue *cfqq)
2446{
2447 int process_refs, io_refs;
2448
2449 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002450 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002451 BUG_ON(process_refs < 0);
2452 return process_refs;
2453}
2454
2455static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2456{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002457 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002458 struct cfq_queue *__cfqq;
2459
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002460 /*
2461 * If there are no process references on the new_cfqq, then it is
2462 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2463 * chain may have dropped their last reference (not just their
2464 * last process reference).
2465 */
2466 if (!cfqq_process_refs(new_cfqq))
2467 return;
2468
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002469 /* Avoid a circular list and skip interim queue merges */
2470 while ((__cfqq = new_cfqq->new_cfqq)) {
2471 if (__cfqq == cfqq)
2472 return;
2473 new_cfqq = __cfqq;
2474 }
2475
2476 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002477 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002478 /*
2479 * If the process for the cfqq has gone away, there is no
2480 * sense in merging the queues.
2481 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002482 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002483 return;
2484
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002485 /*
2486 * Merge in the direction of the lesser amount of work.
2487 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002488 if (new_process_refs >= process_refs) {
2489 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002490 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002491 } else {
2492 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002493 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002494 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002495}
2496
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002497static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002498 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002499{
2500 struct cfq_queue *queue;
2501 int i;
2502 bool key_valid = false;
2503 unsigned long lowest_key = 0;
2504 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2505
Vivek Goyal65b32a52009-12-16 17:52:59 -05002506 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2507 /* select the one with lowest rb_key */
2508 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002509 if (queue &&
2510 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2511 lowest_key = queue->rb_key;
2512 cur_best = i;
2513 key_valid = true;
2514 }
2515 }
2516
2517 return cur_best;
2518}
2519
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002520static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002521{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002522 unsigned slice;
2523 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002524 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002525 unsigned group_slice;
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002526 enum wl_prio_t original_prio = cfqd->serving_prio;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002527
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002528 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002529 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002530 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002531 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002532 cfqd->serving_prio = BE_WORKLOAD;
2533 else {
2534 cfqd->serving_prio = IDLE_WORKLOAD;
2535 cfqd->workload_expires = jiffies + 1;
2536 return;
2537 }
2538
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002539 if (original_prio != cfqd->serving_prio)
2540 goto new_workload;
2541
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002542 /*
2543 * For RT and BE, we have to choose also the type
2544 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2545 * expiration time
2546 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002547 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002548 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002549
2550 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002551 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002552 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002553 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002554 return;
2555
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002556new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002557 /* otherwise select new workload type */
2558 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002559 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2560 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002561 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002562
2563 /*
2564 * the workload slice is computed as a fraction of target latency
2565 * proportional to the number of queues in that workload, over
2566 * all the queues in the same priority class
2567 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002568 group_slice = cfq_group_slice(cfqd, cfqg);
2569
2570 slice = group_slice * count /
2571 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2572 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002573
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002574 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2575 unsigned int tmp;
2576
2577 /*
2578 * Async queues are currently system wide. Just taking
2579 * proportion of queues with-in same group will lead to higher
2580 * async ratio system wide as generally root group is going
2581 * to have higher weight. A more accurate thing would be to
2582 * calculate system wide asnc/sync ratio.
2583 */
Tao Ma5bf14c02012-04-01 14:33:39 -07002584 tmp = cfqd->cfq_target_latency *
2585 cfqg_busy_async_queues(cfqd, cfqg);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002586 tmp = tmp/cfqd->busy_queues;
2587 slice = min_t(unsigned, slice, tmp);
2588
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002589 /* async workload slice is scaled down according to
2590 * the sync/async slice ratio. */
2591 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002592 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002593 /* sync workload slice is at least 2 * cfq_slice_idle */
2594 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2595
2596 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002597 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002598 cfqd->workload_expires = jiffies + slice;
2599}
2600
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002601static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2602{
2603 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002604 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002605
2606 if (RB_EMPTY_ROOT(&st->rb))
2607 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002608 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002609 update_min_vdisktime(st);
2610 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002611}
2612
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002613static void cfq_choose_cfqg(struct cfq_data *cfqd)
2614{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002615 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2616
2617 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002618
2619 /* Restore the workload type data */
2620 if (cfqg->saved_workload_slice) {
2621 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2622 cfqd->serving_type = cfqg->saved_workload;
2623 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002624 } else
2625 cfqd->workload_expires = jiffies - 1;
2626
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002627 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002628}
2629
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002630/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002631 * Select a queue for service. If we have a current active queue,
2632 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002633 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002634static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002635{
Jens Axboea36e71f2009-04-15 12:15:11 +02002636 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002637
2638 cfqq = cfqd->active_queue;
2639 if (!cfqq)
2640 goto new_queue;
2641
Vivek Goyalf04a6422009-12-03 12:59:40 -05002642 if (!cfqd->rq_queued)
2643 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002644
2645 /*
2646 * We were waiting for group to get backlogged. Expire the queue
2647 */
2648 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2649 goto expire;
2650
Jens Axboe22e2c502005-06-27 10:55:12 +02002651 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002652 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002653 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002654 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2655 /*
2656 * If slice had not expired at the completion of last request
2657 * we might not have turned on wait_busy flag. Don't expire
2658 * the queue yet. Allow the group to get backlogged.
2659 *
2660 * The very fact that we have used the slice, that means we
2661 * have been idling all along on this queue and it should be
2662 * ok to wait for this request to complete.
2663 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002664 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2665 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2666 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002667 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002668 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002669 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002670 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002671
2672 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002673 * The active queue has requests and isn't expired, allow it to
2674 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002675 */
Jens Axboedd67d052006-06-21 09:36:18 +02002676 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002677 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002678
2679 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002680 * If another queue has a request waiting within our mean seek
2681 * distance, let it run. The expire code will check for close
2682 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002683 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002684 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002685 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002686 if (new_cfqq) {
2687 if (!cfqq->new_cfqq)
2688 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002689 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002690 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002691
2692 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002693 * No requests pending. If the active queue still has requests in
2694 * flight or is idling for a new request, allow either of these
2695 * conditions to happen (or time out) before selecting a new queue.
2696 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002697 if (timer_pending(&cfqd->idle_slice_timer)) {
2698 cfqq = NULL;
2699 goto keep_queue;
2700 }
2701
Shaohua Li8e1ac662010-11-08 15:01:04 +01002702 /*
2703 * This is a deep seek queue, but the device is much faster than
2704 * the queue can deliver, don't idle
2705 **/
2706 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2707 (cfq_cfqq_slice_new(cfqq) ||
2708 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2709 cfq_clear_cfqq_deep(cfqq);
2710 cfq_clear_cfqq_idle_window(cfqq);
2711 }
2712
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002713 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2714 cfqq = NULL;
2715 goto keep_queue;
2716 }
2717
2718 /*
2719 * If group idle is enabled and there are requests dispatched from
2720 * this group, wait for requests to complete.
2721 */
2722check_group_idle:
Shaohua Li7700fc42011-07-12 14:24:56 +02002723 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2724 cfqq->cfqg->dispatched &&
2725 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002726 cfqq = NULL;
2727 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002728 }
2729
Jens Axboe3b181522005-06-27 10:56:24 +02002730expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002731 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002732new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002733 /*
2734 * Current queue expired. Check if we have to switch to a new
2735 * service tree
2736 */
2737 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002738 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002739
Jens Axboea36e71f2009-04-15 12:15:11 +02002740 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002741keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002742 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002743}
2744
Jens Axboefebffd62008-01-28 13:19:43 +01002745static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002746{
2747 int dispatched = 0;
2748
2749 while (cfqq->next_rq) {
2750 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2751 dispatched++;
2752 }
2753
2754 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002755
2756 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002757 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002758 return dispatched;
2759}
2760
Jens Axboe498d3aa22007-04-26 12:54:48 +02002761/*
2762 * Drain our current requests. Used for barriers and when switching
2763 * io schedulers on-the-fly.
2764 */
Jens Axboed9e76202007-04-20 14:27:50 +02002765static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002766{
Jens Axboe08717142008-01-28 11:38:15 +01002767 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002768 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002769
Divyesh Shah3440c492010-04-09 09:29:57 +02002770 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002771 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002772 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2773 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002774 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002775 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002776
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002777 BUG_ON(cfqd->busy_queues);
2778
Jeff Moyer69237152009-06-12 15:29:30 +02002779 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002780 return dispatched;
2781}
2782
Shaohua Liabc3c742010-03-01 09:20:54 +01002783static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2784 struct cfq_queue *cfqq)
2785{
2786 /* the queue hasn't finished any request, can't estimate */
2787 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01002788 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002789 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2790 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01002791 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002792
Shaohua Lic1e44752010-11-08 15:01:02 +01002793 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01002794}
2795
Jens Axboe0b182d62009-10-06 20:49:37 +02002796static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002797{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002798 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799
Jens Axboe2f5cb732009-04-07 08:51:19 +02002800 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002801 * Drain async requests before we start sync IO
2802 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002803 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002804 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002805
2806 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002807 * If this is an async queue and we have sync IO in flight, let it wait
2808 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002809 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002810 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002811
Shaohua Liabc3c742010-03-01 09:20:54 +01002812 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002813 if (cfq_class_idle(cfqq))
2814 max_dispatch = 1;
2815
2816 /*
2817 * Does this cfqq already have too much IO in flight?
2818 */
2819 if (cfqq->dispatched >= max_dispatch) {
Shaohua Lief8a41d2011-03-07 09:26:29 +01002820 bool promote_sync = false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002821 /*
2822 * idle queue must always only have a single IO in flight
2823 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002824 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002825 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002826
Jens Axboe2f5cb732009-04-07 08:51:19 +02002827 /*
Li, Shaohuac4ade942011-03-23 08:30:34 +01002828 * If there is only one sync queue
2829 * we can ignore async queue here and give the sync
Shaohua Lief8a41d2011-03-07 09:26:29 +01002830 * queue no dispatch limit. The reason is a sync queue can
2831 * preempt async queue, limiting the sync queue doesn't make
2832 * sense. This is useful for aiostress test.
2833 */
Li, Shaohuac4ade942011-03-23 08:30:34 +01002834 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2835 promote_sync = true;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002836
2837 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002838 * We have other queues, don't allow more IO from this one
2839 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002840 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2841 !promote_sync)
Jens Axboe0b182d62009-10-06 20:49:37 +02002842 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002843
Jens Axboe2f5cb732009-04-07 08:51:19 +02002844 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002845 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002846 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002847 if (cfqd->busy_queues == 1 || promote_sync)
Shaohua Liabc3c742010-03-01 09:20:54 +01002848 max_dispatch = -1;
2849 else
2850 /*
2851 * Normally we start throttling cfqq when cfq_quantum/2
2852 * requests have been dispatched. But we can drive
2853 * deeper queue depths at the beginning of slice
2854 * subjected to upper limit of cfq_quantum.
2855 * */
2856 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002857 }
2858
2859 /*
2860 * Async queues must wait a bit before being allowed dispatch.
2861 * We also ramp up the dispatch depth gradually for async IO,
2862 * based on the last sync IO we serviced
2863 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002864 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002865 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002866 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002867
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002868 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002869 if (!depth && !cfqq->dispatched)
2870 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002871 if (depth < max_dispatch)
2872 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873 }
2874
Jens Axboe0b182d62009-10-06 20:49:37 +02002875 /*
2876 * If we're below the current max, allow a dispatch
2877 */
2878 return cfqq->dispatched < max_dispatch;
2879}
2880
2881/*
2882 * Dispatch a request from cfqq, moving them to the request queue
2883 * dispatch list.
2884 */
2885static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2886{
2887 struct request *rq;
2888
2889 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2890
2891 if (!cfq_may_dispatch(cfqd, cfqq))
2892 return false;
2893
2894 /*
2895 * follow expired path, else get first next available
2896 */
2897 rq = cfq_check_fifo(cfqq);
2898 if (!rq)
2899 rq = cfqq->next_rq;
2900
2901 /*
2902 * insert request into driver dispatch list
2903 */
2904 cfq_dispatch_insert(cfqd->queue, rq);
2905
2906 if (!cfqd->active_cic) {
Tejun Heoc5869802011-12-14 00:33:41 +01002907 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe0b182d62009-10-06 20:49:37 +02002908
Tejun Heoc5869802011-12-14 00:33:41 +01002909 atomic_long_inc(&cic->icq.ioc->refcount);
Jens Axboe0b182d62009-10-06 20:49:37 +02002910 cfqd->active_cic = cic;
2911 }
2912
2913 return true;
2914}
2915
2916/*
2917 * Find the cfqq that we need to service and move a request from that to the
2918 * dispatch list
2919 */
2920static int cfq_dispatch_requests(struct request_queue *q, int force)
2921{
2922 struct cfq_data *cfqd = q->elevator->elevator_data;
2923 struct cfq_queue *cfqq;
2924
2925 if (!cfqd->busy_queues)
2926 return 0;
2927
2928 if (unlikely(force))
2929 return cfq_forced_dispatch(cfqd);
2930
2931 cfqq = cfq_select_queue(cfqd);
2932 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002933 return 0;
2934
Jens Axboe2f5cb732009-04-07 08:51:19 +02002935 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002936 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002937 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002938 if (!cfq_dispatch_request(cfqd, cfqq))
2939 return 0;
2940
Jens Axboe2f5cb732009-04-07 08:51:19 +02002941 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002942 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002943
2944 /*
2945 * expire an async queue immediately if it has used up its slice. idle
2946 * queue always expire after 1 dispatch round.
2947 */
2948 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2949 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2950 cfq_class_idle(cfqq))) {
2951 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002952 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002953 }
2954
Shan Weib217a902009-09-01 10:06:42 +02002955 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002956 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957}
2958
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959/*
Jens Axboe5e705372006-07-13 12:39:25 +02002960 * task holds one reference to the queue, dropped when task exits. each rq
2961 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002963 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964 * queue lock must be held here.
2965 */
2966static void cfq_put_queue(struct cfq_queue *cfqq)
2967{
Jens Axboe22e2c502005-06-27 10:55:12 +02002968 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05002969 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002970
Shaohua Li30d7b942011-01-07 08:46:59 +01002971 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972
Shaohua Li30d7b942011-01-07 08:46:59 +01002973 cfqq->ref--;
2974 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975 return;
2976
Jens Axboe7b679132008-05-30 12:23:07 +02002977 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002979 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002980 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002982 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002983 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002984 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002985 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002986
Vivek Goyalf04a6422009-12-03 12:59:40 -05002987 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 kmem_cache_free(cfq_pool, cfqq);
Tejun Heoeb7d8c072012-03-23 14:02:53 +01002989 cfqg_put(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990}
2991
Shaohua Lid02a2c02010-05-25 10:16:53 +02002992static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002993{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002994 struct cfq_queue *__cfqq, *next;
2995
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002996 /*
2997 * If this queue was scheduled to merge with another queue, be
2998 * sure to drop the reference taken on that queue (and others in
2999 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3000 */
3001 __cfqq = cfqq->new_cfqq;
3002 while (__cfqq) {
3003 if (__cfqq == cfqq) {
3004 WARN(1, "cfqq->new_cfqq loop detected\n");
3005 break;
3006 }
3007 next = __cfqq->new_cfqq;
3008 cfq_put_queue(__cfqq);
3009 __cfqq = next;
3010 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02003011}
3012
3013static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3014{
3015 if (unlikely(cfqq == cfqd->active_queue)) {
3016 __cfq_slice_expired(cfqd, cfqq, 0);
3017 cfq_schedule_dispatch(cfqd);
3018 }
3019
3020 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003021
Jens Axboe89850f72006-07-22 16:48:31 +02003022 cfq_put_queue(cfqq);
3023}
3024
Tejun Heo9b84cac2011-12-14 00:33:42 +01003025static void cfq_init_icq(struct io_cq *icq)
3026{
3027 struct cfq_io_cq *cic = icq_to_cic(icq);
3028
3029 cic->ttime.last_end_request = jiffies;
3030}
3031
Tejun Heoc5869802011-12-14 00:33:41 +01003032static void cfq_exit_icq(struct io_cq *icq)
Jens Axboe89850f72006-07-22 16:48:31 +02003033{
Tejun Heoc5869802011-12-14 00:33:41 +01003034 struct cfq_io_cq *cic = icq_to_cic(icq);
Tejun Heo283287a2011-12-14 00:33:38 +01003035 struct cfq_data *cfqd = cic_to_cfqd(cic);
Fabio Checconi4faa3c82008-04-10 08:28:01 +02003036
Jens Axboeff6657c2009-04-08 10:58:57 +02003037 if (cic->cfqq[BLK_RW_ASYNC]) {
3038 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3039 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02003040 }
3041
Jens Axboeff6657c2009-04-08 10:58:57 +02003042 if (cic->cfqq[BLK_RW_SYNC]) {
3043 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3044 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02003045 }
Jens Axboe89850f72006-07-22 16:48:31 +02003046}
3047
Tejun Heoabede6d2012-03-19 15:10:57 -07003048static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02003049{
3050 struct task_struct *tsk = current;
3051 int ioprio_class;
3052
Jens Axboe3b181522005-06-27 10:56:24 +02003053 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003054 return;
3055
Tejun Heo598971b2012-03-19 15:10:58 -07003056 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02003057 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01003058 default:
3059 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3060 case IOPRIO_CLASS_NONE:
3061 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02003062 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01003063 */
3064 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02003065 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01003066 break;
3067 case IOPRIO_CLASS_RT:
Tejun Heo598971b2012-03-19 15:10:58 -07003068 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
Jens Axboefe094d92008-01-31 13:08:54 +01003069 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3070 break;
3071 case IOPRIO_CLASS_BE:
Tejun Heo598971b2012-03-19 15:10:58 -07003072 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
Jens Axboefe094d92008-01-31 13:08:54 +01003073 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3074 break;
3075 case IOPRIO_CLASS_IDLE:
3076 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3077 cfqq->ioprio = 7;
3078 cfq_clear_cfqq_idle_window(cfqq);
3079 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02003080 }
3081
3082 /*
3083 * keep track of original prio settings in case we have to temporarily
3084 * elevate the priority of this queue
3085 */
3086 cfqq->org_ioprio = cfqq->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02003087 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003088}
3089
Tejun Heo598971b2012-03-19 15:10:58 -07003090static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
Jens Axboe22e2c502005-06-27 10:55:12 +02003091{
Tejun Heo598971b2012-03-19 15:10:58 -07003092 int ioprio = cic->icq.ioc->ioprio;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003093 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05003094 struct cfq_queue *cfqq;
Jens Axboe35e60772006-06-14 09:10:45 +02003095
Tejun Heo598971b2012-03-19 15:10:58 -07003096 /*
3097 * Check whether ioprio has changed. The condition may trigger
3098 * spuriously on a newly created cic but there's no harm.
3099 */
3100 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
Jens Axboecaaa5f92006-06-16 11:23:00 +02003101 return;
3102
Jens Axboeff6657c2009-04-08 10:58:57 +02003103 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02003104 if (cfqq) {
3105 struct cfq_queue *new_cfqq;
Tejun Heoabede6d2012-03-19 15:10:57 -07003106 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3107 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02003108 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02003109 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02003110 cfq_put_queue(cfqq);
3111 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003112 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003113
Jens Axboeff6657c2009-04-08 10:58:57 +02003114 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02003115 if (cfqq)
3116 cfq_mark_cfqq_prio_changed(cfqq);
Tejun Heo598971b2012-03-19 15:10:58 -07003117
3118 cic->ioprio = ioprio;
Jens Axboe22e2c502005-06-27 10:55:12 +02003119}
3120
Jens Axboed5036d72009-06-26 10:44:34 +02003121static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02003122 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02003123{
3124 RB_CLEAR_NODE(&cfqq->rb_node);
3125 RB_CLEAR_NODE(&cfqq->p_node);
3126 INIT_LIST_HEAD(&cfqq->fifo);
3127
Shaohua Li30d7b942011-01-07 08:46:59 +01003128 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02003129 cfqq->cfqd = cfqd;
3130
3131 cfq_mark_cfqq_prio_changed(cfqq);
3132
3133 if (is_sync) {
3134 if (!cfq_class_idle(cfqq))
3135 cfq_mark_cfqq_idle_window(cfqq);
3136 cfq_mark_cfqq_sync(cfqq);
3137 }
3138 cfqq->pid = pid;
3139}
3140
Vivek Goyal246103332009-12-03 12:59:51 -05003141#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo598971b2012-03-19 15:10:58 -07003142static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
Vivek Goyal246103332009-12-03 12:59:51 -05003143{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003144 struct cfq_data *cfqd = cic_to_cfqd(cic);
Tejun Heo598971b2012-03-19 15:10:58 -07003145 struct cfq_queue *sync_cfqq;
3146 uint64_t id;
Vivek Goyal246103332009-12-03 12:59:51 -05003147
Tejun Heo598971b2012-03-19 15:10:58 -07003148 rcu_read_lock();
Tejun Heo3c798392012-04-16 13:57:25 -07003149 id = bio_blkcg(bio)->id;
Tejun Heo598971b2012-03-19 15:10:58 -07003150 rcu_read_unlock();
3151
3152 /*
3153 * Check whether blkcg has changed. The condition may trigger
3154 * spuriously on a newly created cic but there's no harm.
3155 */
3156 if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
Vivek Goyal246103332009-12-03 12:59:51 -05003157 return;
3158
Tejun Heo598971b2012-03-19 15:10:58 -07003159 sync_cfqq = cic_to_cfqq(cic, 1);
Vivek Goyal246103332009-12-03 12:59:51 -05003160 if (sync_cfqq) {
3161 /*
3162 * Drop reference to sync queue. A new sync queue will be
3163 * assigned in new group upon arrival of a fresh request.
3164 */
3165 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3166 cic_set_cfqq(cic, NULL, 1);
3167 cfq_put_queue(sync_cfqq);
3168 }
Tejun Heo598971b2012-03-19 15:10:58 -07003169
3170 cic->blkcg_id = id;
Vivek Goyal246103332009-12-03 12:59:51 -05003171}
Tejun Heo598971b2012-03-19 15:10:58 -07003172#else
3173static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
Vivek Goyal246103332009-12-03 12:59:51 -05003174#endif /* CONFIG_CFQ_GROUP_IOSCHED */
3175
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176static struct cfq_queue *
Tejun Heoabede6d2012-03-19 15:10:57 -07003177cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3178 struct bio *bio, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003179{
Tejun Heo3c798392012-04-16 13:57:25 -07003180 struct blkcg *blkcg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003181 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003182 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183
3184retry:
Tejun Heo2a7f1242012-03-05 13:15:01 -08003185 rcu_read_lock();
3186
Tejun Heo3c798392012-04-16 13:57:25 -07003187 blkcg = bio_blkcg(bio);
Tejun Heocd1604f2012-03-05 13:15:06 -08003188 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003189 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190
Jens Axboe6118b702009-06-30 09:34:12 +02003191 /*
3192 * Always try a new alloc if we fell back to the OOM cfqq
3193 * originally, since it should just be a temporary situation.
3194 */
3195 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3196 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 if (new_cfqq) {
3198 cfqq = new_cfqq;
3199 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02003200 } else if (gfp_mask & __GFP_WAIT) {
Tejun Heo2a7f1242012-03-05 13:15:01 -08003201 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07003203 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02003204 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07003205 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02003207 if (new_cfqq)
3208 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02003209 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07003210 cfqq = kmem_cache_alloc_node(cfq_pool,
3211 gfp_mask | __GFP_ZERO,
3212 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02003213 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214
Jens Axboe6118b702009-06-30 09:34:12 +02003215 if (cfqq) {
3216 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
Tejun Heoabede6d2012-03-19 15:10:57 -07003217 cfq_init_prio_data(cfqq, cic);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003218 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02003219 cfq_log_cfqq(cfqd, cfqq, "alloced");
3220 } else
3221 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003222 }
3223
3224 if (new_cfqq)
3225 kmem_cache_free(cfq_pool, new_cfqq);
3226
Tejun Heo2a7f1242012-03-05 13:15:01 -08003227 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228 return cfqq;
3229}
3230
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003231static struct cfq_queue **
3232cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3233{
Jens Axboefe094d92008-01-31 13:08:54 +01003234 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003235 case IOPRIO_CLASS_RT:
3236 return &cfqd->async_cfqq[0][ioprio];
Tejun Heo598971b2012-03-19 15:10:58 -07003237 case IOPRIO_CLASS_NONE:
3238 ioprio = IOPRIO_NORM;
3239 /* fall through */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003240 case IOPRIO_CLASS_BE:
3241 return &cfqd->async_cfqq[1][ioprio];
3242 case IOPRIO_CLASS_IDLE:
3243 return &cfqd->async_idle_cfqq;
3244 default:
3245 BUG();
3246 }
3247}
3248
Jens Axboe15c31be2007-07-10 13:43:25 +02003249static struct cfq_queue *
Tejun Heoabede6d2012-03-19 15:10:57 -07003250cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
Tejun Heo4f85cb92012-03-05 13:15:28 -08003251 struct bio *bio, gfp_t gfp_mask)
Jens Axboe15c31be2007-07-10 13:43:25 +02003252{
Tejun Heo598971b2012-03-19 15:10:58 -07003253 const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3254 const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003255 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02003256 struct cfq_queue *cfqq = NULL;
3257
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003258 if (!is_sync) {
3259 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3260 cfqq = *async_cfqq;
3261 }
3262
Jens Axboe6118b702009-06-30 09:34:12 +02003263 if (!cfqq)
Tejun Heoabede6d2012-03-19 15:10:57 -07003264 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02003265
3266 /*
3267 * pin the queue now that it's allocated, scheduler exit will prune it
3268 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003269 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01003270 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003271 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02003272 }
3273
Shaohua Li30d7b942011-01-07 08:46:59 +01003274 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02003275 return cfqq;
3276}
3277
Jens Axboe22e2c502005-06-27 10:55:12 +02003278static void
Shaohua Li383cd722011-07-12 14:24:35 +02003279__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003280{
Shaohua Li383cd722011-07-12 14:24:35 +02003281 unsigned long elapsed = jiffies - ttime->last_end_request;
3282 elapsed = min(elapsed, 2UL * slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003283
Shaohua Li383cd722011-07-12 14:24:35 +02003284 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3285 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3286 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3287}
3288
3289static void
3290cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01003291 struct cfq_io_cq *cic)
Shaohua Li383cd722011-07-12 14:24:35 +02003292{
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003293 if (cfq_cfqq_sync(cfqq)) {
Shaohua Li383cd722011-07-12 14:24:35 +02003294 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003295 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3296 cfqd->cfq_slice_idle);
3297 }
Shaohua Li7700fc42011-07-12 14:24:56 +02003298#ifdef CONFIG_CFQ_GROUP_IOSCHED
3299 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3300#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003301}
3302
Jens Axboe206dc692006-03-28 13:03:44 +02003303static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003304cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003305 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003306{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003307 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003308 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003309 if (cfqq->last_request_pos) {
3310 if (cfqq->last_request_pos < blk_rq_pos(rq))
3311 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3312 else
3313 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3314 }
Jens Axboe206dc692006-03-28 13:03:44 +02003315
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003316 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003317 if (blk_queue_nonrot(cfqd->queue))
3318 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3319 else
3320 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003321}
Jens Axboe22e2c502005-06-27 10:55:12 +02003322
3323/*
3324 * Disable idle window if the process thinks too long or seeks so much that
3325 * it doesn't matter
3326 */
3327static void
3328cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01003329 struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02003330{
Jens Axboe7b679132008-05-30 12:23:07 +02003331 int old_idle, enable_idle;
Jens Axboe1be92f2f2007-04-19 14:32:26 +02003332
Jens Axboe08717142008-01-28 11:38:15 +01003333 /*
3334 * Don't idle for async or idle io prio class
3335 */
3336 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f2f2007-04-19 14:32:26 +02003337 return;
3338
Jens Axboec265a7f2008-06-26 13:49:33 +02003339 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003340
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003341 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3342 cfq_mark_cfqq_deep(cfqq);
3343
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003344 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3345 enable_idle = 0;
Tejun Heof6e8d012012-03-05 13:15:26 -08003346 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
Tejun Heoc5869802011-12-14 00:33:41 +01003347 !cfqd->cfq_slice_idle ||
3348 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003349 enable_idle = 0;
Shaohua Li383cd722011-07-12 14:24:35 +02003350 else if (sample_valid(cic->ttime.ttime_samples)) {
3351 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003352 enable_idle = 0;
3353 else
3354 enable_idle = 1;
3355 }
3356
Jens Axboe7b679132008-05-30 12:23:07 +02003357 if (old_idle != enable_idle) {
3358 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3359 if (enable_idle)
3360 cfq_mark_cfqq_idle_window(cfqq);
3361 else
3362 cfq_clear_cfqq_idle_window(cfqq);
3363 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003364}
3365
Jens Axboe22e2c502005-06-27 10:55:12 +02003366/*
3367 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3368 * no or if we aren't sure, a 1 will cause a preempt.
3369 */
Jens Axboea6151c32009-10-07 20:02:57 +02003370static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003371cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003372 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003373{
Jens Axboe6d048f52007-04-25 12:44:27 +02003374 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003375
Jens Axboe6d048f52007-04-25 12:44:27 +02003376 cfqq = cfqd->active_queue;
3377 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003378 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003379
Jens Axboe6d048f52007-04-25 12:44:27 +02003380 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003381 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003382
3383 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003384 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003385
Jens Axboe22e2c502005-06-27 10:55:12 +02003386 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003387 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3388 */
3389 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3390 return false;
3391
3392 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003393 * if the new request is sync, but the currently running queue is
3394 * not, let the sync request have priority.
3395 */
Jens Axboe5e705372006-07-13 12:39:25 +02003396 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003397 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003398
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003399 if (new_cfqq->cfqg != cfqq->cfqg)
3400 return false;
3401
3402 if (cfq_slice_used(cfqq))
3403 return true;
3404
3405 /* Allow preemption only if we are idling on sync-noidle tree */
3406 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3407 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3408 new_cfqq->service_tree->count == 2 &&
3409 RB_EMPTY_ROOT(&cfqq->sort_list))
3410 return true;
3411
Jens Axboe374f84a2006-07-23 01:42:19 +02003412 /*
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003413 * So both queues are sync. Let the new request get disk time if
3414 * it's a metadata request and the current queue is doing regular IO.
3415 */
Christoph Hellwig65299a32011-08-23 14:50:29 +02003416 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003417 return true;
3418
3419 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003420 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3421 */
3422 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003423 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003424
Shaohua Lid2d59e12010-11-08 15:01:03 +01003425 /* An idle queue should not be idle now for some reason */
3426 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3427 return true;
3428
Jens Axboe1e3335d2007-02-14 19:59:49 +01003429 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003430 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003431
3432 /*
3433 * if this request is as-good as one we would expect from the
3434 * current cfqq, let it preempt
3435 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003436 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003437 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003438
Jens Axboea6151c32009-10-07 20:02:57 +02003439 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003440}
3441
3442/*
3443 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3444 * let it have half of its nominal slice.
3445 */
3446static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3447{
Shaohua Lidf0793a2012-01-19 09:20:09 +01003448 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3449
Jens Axboe7b679132008-05-30 12:23:07 +02003450 cfq_log_cfqq(cfqd, cfqq, "preempt");
Shaohua Lidf0793a2012-01-19 09:20:09 +01003451 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003452
Jens Axboebf572252006-07-19 20:29:12 +02003453 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003454 * workload type is changed, don't save slice, otherwise preempt
3455 * doesn't happen
3456 */
Shaohua Lidf0793a2012-01-19 09:20:09 +01003457 if (old_type != cfqq_type(cfqq))
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003458 cfqq->cfqg->saved_workload_slice = 0;
3459
3460 /*
Jens Axboebf572252006-07-19 20:29:12 +02003461 * Put the new queue at the front of the of the current list,
3462 * so we know that it will be selected next.
3463 */
3464 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003465
3466 cfq_service_tree_add(cfqd, cfqq, 1);
Justin TerAvesteda5e0c2011-03-22 21:26:49 +01003467
Justin TerAvest62a37f62011-03-23 08:25:44 +01003468 cfqq->slice_end = 0;
3469 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003470}
3471
3472/*
Jens Axboe5e705372006-07-13 12:39:25 +02003473 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003474 * something we should do about it
3475 */
3476static void
Jens Axboe5e705372006-07-13 12:39:25 +02003477cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3478 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003479{
Tejun Heoc5869802011-12-14 00:33:41 +01003480 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003481
Aaron Carroll45333d52008-08-26 15:52:36 +02003482 cfqd->rq_queued++;
Christoph Hellwig65299a32011-08-23 14:50:29 +02003483 if (rq->cmd_flags & REQ_PRIO)
3484 cfqq->prio_pending++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003485
Shaohua Li383cd722011-07-12 14:24:35 +02003486 cfq_update_io_thinktime(cfqd, cfqq, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003487 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003488 cfq_update_idle_window(cfqd, cfqq, cic);
3489
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003490 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003491
3492 if (cfqq == cfqd->active_queue) {
3493 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003494 * Remember that we saw a request from this process, but
3495 * don't start queuing just yet. Otherwise we risk seeing lots
3496 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003497 * and merging. If the request is already larger than a single
3498 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003499 * merging is already done. Ditto for a busy system that
3500 * has other work pending, don't risk delaying until the
3501 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003502 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003503 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003504 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3505 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003506 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003507 cfq_clear_cfqq_wait_request(cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003508 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003509 } else {
Tejun Heo155fead2012-04-01 14:38:44 -07003510 cfqg_stats_update_idle_time(cfqq->cfqg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003511 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003512 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003513 }
Jens Axboe5e705372006-07-13 12:39:25 +02003514 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003515 /*
3516 * not the active queue - expire current slice if it is
3517 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003518 * has some old slice time left and is of higher priority or
3519 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003520 */
3521 cfq_preempt_queue(cfqd, cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003522 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003523 }
3524}
3525
Jens Axboe165125e2007-07-24 09:28:11 +02003526static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003527{
Jens Axboeb4878f22005-10-20 16:42:29 +02003528 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003529 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003530
Jens Axboe7b679132008-05-30 12:23:07 +02003531 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Tejun Heoabede6d2012-03-19 15:10:57 -07003532 cfq_init_prio_data(cfqq, RQ_CIC(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533
Jens Axboe30996f42009-10-05 11:03:39 +02003534 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003535 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003536 cfq_add_rq_rb(rq);
Tejun Heo155fead2012-04-01 14:38:44 -07003537 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3538 rq->cmd_flags);
Jens Axboe5e705372006-07-13 12:39:25 +02003539 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540}
3541
Aaron Carroll45333d52008-08-26 15:52:36 +02003542/*
3543 * Update hw_tag based on peak queue depth over 50 samples under
3544 * sufficient load.
3545 */
3546static void cfq_update_hw_tag(struct cfq_data *cfqd)
3547{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003548 struct cfq_queue *cfqq = cfqd->active_queue;
3549
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003550 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3551 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003552
3553 if (cfqd->hw_tag == 1)
3554 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003555
3556 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003557 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003558 return;
3559
Shaohua Li1a1238a2009-10-27 08:46:23 +01003560 /*
3561 * If active queue hasn't enough requests and can idle, cfq might not
3562 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3563 * case
3564 */
3565 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3566 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003567 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003568 return;
3569
Aaron Carroll45333d52008-08-26 15:52:36 +02003570 if (cfqd->hw_tag_samples++ < 50)
3571 return;
3572
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003573 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003574 cfqd->hw_tag = 1;
3575 else
3576 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003577}
3578
Vivek Goyal7667aa02009-12-08 17:52:58 -05003579static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3580{
Tejun Heoc5869802011-12-14 00:33:41 +01003581 struct cfq_io_cq *cic = cfqd->active_cic;
Vivek Goyal7667aa02009-12-08 17:52:58 -05003582
Justin TerAvest02a8f012011-02-09 14:20:03 +01003583 /* If the queue already has requests, don't wait */
3584 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3585 return false;
3586
Vivek Goyal7667aa02009-12-08 17:52:58 -05003587 /* If there are other queues in the group, don't wait */
3588 if (cfqq->cfqg->nr_cfqq > 1)
3589 return false;
3590
Shaohua Li7700fc42011-07-12 14:24:56 +02003591 /* the only queue in the group, but think time is big */
3592 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3593 return false;
3594
Vivek Goyal7667aa02009-12-08 17:52:58 -05003595 if (cfq_slice_used(cfqq))
3596 return true;
3597
3598 /* if slice left is less than think time, wait busy */
Shaohua Li383cd722011-07-12 14:24:35 +02003599 if (cic && sample_valid(cic->ttime.ttime_samples)
3600 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
Vivek Goyal7667aa02009-12-08 17:52:58 -05003601 return true;
3602
3603 /*
3604 * If think times is less than a jiffy than ttime_mean=0 and above
3605 * will not be true. It might happen that slice has not expired yet
3606 * but will expire soon (4-5 ns) during select_queue(). To cover the
3607 * case where think time is less than a jiffy, mark the queue wait
3608 * busy if only 1 jiffy is left in the slice.
3609 */
3610 if (cfqq->slice_end - jiffies == 1)
3611 return true;
3612
3613 return false;
3614}
3615
Jens Axboe165125e2007-07-24 09:28:11 +02003616static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617{
Jens Axboe5e705372006-07-13 12:39:25 +02003618 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003619 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003620 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003621 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622
Jens Axboeb4878f22005-10-20 16:42:29 +02003623 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003624 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3625 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626
Aaron Carroll45333d52008-08-26 15:52:36 +02003627 cfq_update_hw_tag(cfqd);
3628
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003629 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003630 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003631 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003632 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003633 (RQ_CFQG(rq))->dispatched--;
Tejun Heo155fead2012-04-01 14:38:44 -07003634 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
3635 rq_io_start_time_ns(rq), rq->cmd_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003637 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003638
Vivek Goyal365722b2009-10-03 15:21:27 +02003639 if (sync) {
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003640 struct cfq_rb_root *service_tree;
3641
Shaohua Li383cd722011-07-12 14:24:35 +02003642 RQ_CIC(rq)->ttime.last_end_request = now;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003643
3644 if (cfq_cfqq_on_rr(cfqq))
3645 service_tree = cfqq->service_tree;
3646 else
3647 service_tree = service_tree_for(cfqq->cfqg,
3648 cfqq_prio(cfqq), cfqq_type(cfqq));
3649 service_tree->ttime.last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003650 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3651 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003652 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003653
Shaohua Li7700fc42011-07-12 14:24:56 +02003654#ifdef CONFIG_CFQ_GROUP_IOSCHED
3655 cfqq->cfqg->ttime.last_end_request = now;
3656#endif
3657
Jens Axboecaaa5f92006-06-16 11:23:00 +02003658 /*
3659 * If this is the active queue, check if it needs to be expired,
3660 * or if we want to idle in case it has no pending requests.
3661 */
3662 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003663 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3664
Jens Axboe44f7c162007-01-19 11:51:58 +11003665 if (cfq_cfqq_slice_new(cfqq)) {
3666 cfq_set_prio_slice(cfqd, cfqq);
3667 cfq_clear_cfqq_slice_new(cfqq);
3668 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003669
3670 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003671 * Should we wait for next request to come in before we expire
3672 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003673 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003674 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003675 unsigned long extend_sl = cfqd->cfq_slice_idle;
3676 if (!cfqd->cfq_slice_idle)
3677 extend_sl = cfqd->cfq_group_idle;
3678 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003679 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003680 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003681 }
3682
Jens Axboea36e71f2009-04-15 12:15:11 +02003683 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003684 * Idling is not enabled on:
3685 * - expired queues
3686 * - idle-priority queues
3687 * - async queues
3688 * - queues with still some requests queued
3689 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003690 */
Jens Axboe08717142008-01-28 11:38:15 +01003691 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003692 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003693 else if (sync && cfqq_empty &&
3694 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003695 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003696 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003697 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003698
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003699 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003700 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701}
3702
Jens Axboe89850f72006-07-22 16:48:31 +02003703static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003704{
Jens Axboe1b379d82009-08-11 08:26:11 +02003705 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003706 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003707 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003708 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003709
3710 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003711}
3712
Jens Axboe165125e2007-07-24 09:28:11 +02003713static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003714{
3715 struct cfq_data *cfqd = q->elevator->elevator_data;
3716 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01003717 struct cfq_io_cq *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003718 struct cfq_queue *cfqq;
3719
3720 /*
3721 * don't force setup of a queue from here, as a call to may_queue
3722 * does not necessarily imply that a request actually will be queued.
3723 * so just lookup a possibly existing queue, or return 'may queue'
3724 * if that fails
3725 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003726 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003727 if (!cic)
3728 return ELV_MQUEUE_MAY;
3729
Jens Axboeb0b78f82009-04-08 10:56:08 +02003730 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003731 if (cfqq) {
Tejun Heoabede6d2012-03-19 15:10:57 -07003732 cfq_init_prio_data(cfqq, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003733
Jens Axboe89850f72006-07-22 16:48:31 +02003734 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003735 }
3736
3737 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738}
3739
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740/*
3741 * queue lock held here
3742 */
Jens Axboebb37b942006-12-01 10:42:33 +01003743static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003744{
Jens Axboe5e705372006-07-13 12:39:25 +02003745 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746
Jens Axboe5e705372006-07-13 12:39:25 +02003747 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003748 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003749
Jens Axboe22e2c502005-06-27 10:55:12 +02003750 BUG_ON(!cfqq->allocated[rw]);
3751 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003753 /* Put down rq reference on cfqg */
Tejun Heoeb7d8c072012-03-23 14:02:53 +01003754 cfqg_put(RQ_CFQG(rq));
Tejun Heoa612fdd2011-12-14 00:33:41 +01003755 rq->elv.priv[0] = NULL;
3756 rq->elv.priv[1] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003757
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758 cfq_put_queue(cfqq);
3759 }
3760}
3761
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003762static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01003763cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003764 struct cfq_queue *cfqq)
3765{
3766 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3767 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003768 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003769 cfq_put_queue(cfqq);
3770 return cic_to_cfqq(cic, 1);
3771}
3772
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003773/*
3774 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3775 * was the last process referring to said cfqq.
3776 */
3777static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01003778split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003779{
3780 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003781 cfqq->pid = current->pid;
3782 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003783 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003784 return cfqq;
3785 }
3786
3787 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003788
3789 cfq_put_cooperator(cfqq);
3790
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003791 cfq_put_queue(cfqq);
3792 return NULL;
3793}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003795 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003797static int
Tejun Heo852c7882012-03-05 13:15:27 -08003798cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
3799 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800{
3801 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heof1f8cc92011-12-14 00:33:42 +01003802 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003804 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003805 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003806
3807 might_sleep_if(gfp_mask & __GFP_WAIT);
3808
Tejun Heo216284c32011-12-14 00:33:38 +01003809 spin_lock_irq(q->queue_lock);
Tejun Heof1f8cc92011-12-14 00:33:42 +01003810
Tejun Heo598971b2012-03-19 15:10:58 -07003811 check_ioprio_changed(cic, bio);
3812 check_blkcg_changed(cic, bio);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003813new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003814 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003815 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Tejun Heoabede6d2012-03-19 15:10:57 -07003816 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003817 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003818 } else {
3819 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003820 * If the queue was seeky for too long, break it apart.
3821 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003822 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003823 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3824 cfqq = split_cfqq(cic, cfqq);
3825 if (!cfqq)
3826 goto new_queue;
3827 }
3828
3829 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003830 * Check to see if this queue is scheduled to merge with
3831 * another, closely cooperating queue. The merging of
3832 * queues happens here as it must be done in process context.
3833 * The reference on new_cfqq was taken in merge_cfqqs.
3834 */
3835 if (cfqq->new_cfqq)
3836 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003837 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003838
3839 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02003840
Jens Axboe6fae9c22011-03-01 15:04:39 -05003841 cfqq->ref++;
Tejun Heoeb7d8c072012-03-23 14:02:53 +01003842 cfqg_get(cfqq->cfqg);
Tejun Heoa612fdd2011-12-14 00:33:41 +01003843 rq->elv.priv[0] = cfqq;
Tejun Heo1adaf3d2012-03-05 13:15:15 -08003844 rq->elv.priv[1] = cfqq->cfqg;
Tejun Heo216284c32011-12-14 00:33:38 +01003845 spin_unlock_irq(q->queue_lock);
Jens Axboe5e705372006-07-13 12:39:25 +02003846 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847}
3848
David Howells65f27f32006-11-22 14:55:48 +00003849static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003850{
David Howells65f27f32006-11-22 14:55:48 +00003851 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003852 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003853 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003854
Jens Axboe40bb54d2009-04-15 12:11:10 +02003855 spin_lock_irq(q->queue_lock);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003856 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003857 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003858}
3859
3860/*
3861 * Timer running if the active_queue is currently idling inside its time slice
3862 */
3863static void cfq_idle_slice_timer(unsigned long data)
3864{
3865 struct cfq_data *cfqd = (struct cfq_data *) data;
3866 struct cfq_queue *cfqq;
3867 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003868 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003869
Jens Axboe7b679132008-05-30 12:23:07 +02003870 cfq_log(cfqd, "idle timer fired");
3871
Jens Axboe22e2c502005-06-27 10:55:12 +02003872 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3873
Jens Axboefe094d92008-01-31 13:08:54 +01003874 cfqq = cfqd->active_queue;
3875 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003876 timed_out = 0;
3877
Jens Axboe22e2c502005-06-27 10:55:12 +02003878 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003879 * We saw a request before the queue expired, let it through
3880 */
3881 if (cfq_cfqq_must_dispatch(cfqq))
3882 goto out_kick;
3883
3884 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003885 * expired
3886 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003887 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003888 goto expire;
3889
3890 /*
3891 * only expire and reinvoke request handler, if there are
3892 * other queues with pending requests
3893 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003894 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003895 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003896
3897 /*
3898 * not expired and it has a request pending, let it dispatch
3899 */
Jens Axboe75e50982009-04-07 08:56:14 +02003900 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003901 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003902
3903 /*
3904 * Queue depth flag is reset only when the idle didn't succeed
3905 */
3906 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003907 }
3908expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003909 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003910out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003911 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003912out_cont:
3913 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3914}
3915
Jens Axboe3b181522005-06-27 10:56:24 +02003916static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3917{
3918 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003919 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003920}
Jens Axboe22e2c502005-06-27 10:55:12 +02003921
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003922static void cfq_put_async_queues(struct cfq_data *cfqd)
3923{
3924 int i;
3925
3926 for (i = 0; i < IOPRIO_BE_NR; i++) {
3927 if (cfqd->async_cfqq[0][i])
3928 cfq_put_queue(cfqd->async_cfqq[0][i]);
3929 if (cfqd->async_cfqq[1][i])
3930 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003931 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003932
3933 if (cfqd->async_idle_cfqq)
3934 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003935}
3936
Jens Axboeb374d182008-10-31 10:05:07 +01003937static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938{
Jens Axboe22e2c502005-06-27 10:55:12 +02003939 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003940 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003941
Jens Axboe3b181522005-06-27 10:56:24 +02003942 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003943
Al Virod9ff4182006-03-18 13:51:22 -05003944 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003945
Al Virod9ff4182006-03-18 13:51:22 -05003946 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003947 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003948
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003949 cfq_put_async_queues(cfqd);
Tejun Heo03aa2642012-03-05 13:15:19 -08003950
3951 spin_unlock_irq(q->queue_lock);
3952
Al Viroa90d7422006-03-18 12:05:37 -05003953 cfq_shutdown_timer_wq(cfqd);
3954
Tejun Heoffea73f2012-06-04 10:02:29 +02003955#ifdef CONFIG_CFQ_GROUP_IOSCHED
3956 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
3957#else
Tejun Heof51b8022012-03-05 13:15:05 -08003958 kfree(cfqd->root_group);
Vivek Goyal2abae552011-05-23 10:02:19 +02003959#endif
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003960 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961}
3962
Tejun Heob2fab5a2012-03-05 13:14:57 -08003963static int cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964{
3965 struct cfq_data *cfqd;
Tejun Heo3c798392012-04-16 13:57:25 -07003966 struct blkcg_gq *blkg __maybe_unused;
Tejun Heoa2b16932012-04-13 13:11:33 -07003967 int i, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968
Christoph Lameter94f60302007-07-17 04:03:29 -07003969 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Tejun Heoa73f7302011-12-14 00:33:37 +01003970 if (!cfqd)
Tejun Heob2fab5a2012-03-05 13:14:57 -08003971 return -ENOMEM;
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003972
Tejun Heof51b8022012-03-05 13:15:05 -08003973 cfqd->queue = q;
3974 q->elevator->elevator_data = cfqd;
3975
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003976 /* Init root service tree */
3977 cfqd->grp_service_tree = CFQ_RB_ROOT;
3978
Tejun Heof51b8022012-03-05 13:15:05 -08003979 /* Init root group and prefer root group over other groups by default */
Vivek Goyal25fb5162009-12-03 12:59:46 -05003980#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07003981 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
Tejun Heoa2b16932012-04-13 13:11:33 -07003982 if (ret)
3983 goto out_free;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04003984
Tejun Heoa2b16932012-04-13 13:11:33 -07003985 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
Tejun Heof51b8022012-03-05 13:15:05 -08003986#else
Tejun Heoa2b16932012-04-13 13:11:33 -07003987 ret = -ENOMEM;
Tejun Heof51b8022012-03-05 13:15:05 -08003988 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
3989 GFP_KERNEL, cfqd->queue->node);
Tejun Heoa2b16932012-04-13 13:11:33 -07003990 if (!cfqd->root_group)
3991 goto out_free;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04003992
Tejun Heoa2b16932012-04-13 13:11:33 -07003993 cfq_init_cfqg_base(cfqd->root_group);
3994#endif
Tejun Heo3381cb82012-04-01 14:38:44 -07003995 cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04003996
Jens Axboe26a2ac02009-04-23 12:13:27 +02003997 /*
3998 * Not strictly needed (since RB_ROOT just clears the node and we
3999 * zeroed cfqd on alloc), but better be safe in case someone decides
4000 * to add magic to the rb code
4001 */
4002 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4003 cfqd->prio_trees[i] = RB_ROOT;
4004
Jens Axboe6118b702009-06-30 09:34:12 +02004005 /*
4006 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4007 * Grab a permanent reference to it, so that the normal code flow
Tejun Heof51b8022012-03-05 13:15:05 -08004008 * will not attempt to free it. oom_cfqq is linked to root_group
4009 * but shouldn't hold a reference as it'll never be unlinked. Lose
4010 * the reference from linking right away.
Jens Axboe6118b702009-06-30 09:34:12 +02004011 */
4012 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01004013 cfqd->oom_cfqq.ref++;
Tejun Heo1adaf3d2012-03-05 13:15:15 -08004014
4015 spin_lock_irq(q->queue_lock);
Tejun Heof51b8022012-03-05 13:15:05 -08004016 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
Tejun Heoeb7d8c072012-03-23 14:02:53 +01004017 cfqg_put(cfqd->root_group);
Tejun Heo1adaf3d2012-03-05 13:15:15 -08004018 spin_unlock_irq(q->queue_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019
Jens Axboe22e2c502005-06-27 10:55:12 +02004020 init_timer(&cfqd->idle_slice_timer);
4021 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4022 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4023
Jens Axboe23e018a2009-10-05 08:52:35 +02004024 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02004025
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02004027 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4028 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 cfqd->cfq_back_max = cfq_back_max;
4030 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02004031 cfqd->cfq_slice[0] = cfq_slice_async;
4032 cfqd->cfq_slice[1] = cfq_slice_sync;
Tao Ma5bf14c02012-04-01 14:33:39 -07004033 cfqd->cfq_target_latency = cfq_target_latency;
Jens Axboe22e2c502005-06-27 10:55:12 +02004034 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4035 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004036 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02004037 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01004038 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01004039 /*
4040 * we optimistically start assuming sync ops weren't delayed in last
4041 * second, in order to have larger depth for async operations.
4042 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01004043 cfqd->last_delayed_sync = jiffies - HZ;
Tejun Heob2fab5a2012-03-05 13:14:57 -08004044 return 0;
Tejun Heoa2b16932012-04-13 13:11:33 -07004045
4046out_free:
4047 kfree(cfqd);
4048 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049}
4050
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051/*
4052 * sysfs parts below -->
4053 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054static ssize_t
4055cfq_var_show(unsigned int var, char *page)
4056{
4057 return sprintf(page, "%d\n", var);
4058}
4059
4060static ssize_t
4061cfq_var_store(unsigned int *var, const char *page, size_t count)
4062{
4063 char *p = (char *) page;
4064
4065 *var = simple_strtoul(p, &p, 10);
4066 return count;
4067}
4068
Linus Torvalds1da177e2005-04-16 15:20:36 -07004069#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004070static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004072 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073 unsigned int __data = __VAR; \
4074 if (__CONV) \
4075 __data = jiffies_to_msecs(__data); \
4076 return cfq_var_show(__data, (page)); \
4077}
4078SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004079SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4080SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05004081SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4082SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004083SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004084SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004085SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4086SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4087SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004088SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07004089SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090#undef SHOW_FUNCTION
4091
4092#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004093static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004095 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096 unsigned int __data; \
4097 int ret = cfq_var_store(&__data, (page), count); \
4098 if (__data < (MIN)) \
4099 __data = (MIN); \
4100 else if (__data > (MAX)) \
4101 __data = (MAX); \
4102 if (__CONV) \
4103 *(__PTR) = msecs_to_jiffies(__data); \
4104 else \
4105 *(__PTR) = __data; \
4106 return ret; \
4107}
4108STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004109STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4110 UINT_MAX, 1);
4111STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4112 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004113STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004114STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4115 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004116STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004117STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004118STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4119STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004120STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4121 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004122STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07004123STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124#undef STORE_FUNCTION
4125
Al Viroe572ec72006-03-18 22:27:18 -05004126#define CFQ_ATTR(name) \
4127 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004128
Al Viroe572ec72006-03-18 22:27:18 -05004129static struct elv_fs_entry cfq_attrs[] = {
4130 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004131 CFQ_ATTR(fifo_expire_sync),
4132 CFQ_ATTR(fifo_expire_async),
4133 CFQ_ATTR(back_seek_max),
4134 CFQ_ATTR(back_seek_penalty),
4135 CFQ_ATTR(slice_sync),
4136 CFQ_ATTR(slice_async),
4137 CFQ_ATTR(slice_async_rq),
4138 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004139 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004140 CFQ_ATTR(low_latency),
Tao Ma5bf14c02012-04-01 14:33:39 -07004141 CFQ_ATTR(target_latency),
Al Viroe572ec72006-03-18 22:27:18 -05004142 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143};
4144
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145static struct elevator_type iosched_cfq = {
4146 .ops = {
4147 .elevator_merge_fn = cfq_merge,
4148 .elevator_merged_fn = cfq_merged_request,
4149 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004150 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004151 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004152 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004154 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155 .elevator_deactivate_req_fn = cfq_deactivate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004157 .elevator_former_req_fn = elv_rb_former_request,
4158 .elevator_latter_req_fn = elv_rb_latter_request,
Tejun Heo9b84cac2011-12-14 00:33:42 +01004159 .elevator_init_icq_fn = cfq_init_icq,
Tejun Heo7e5a8792011-12-14 00:33:42 +01004160 .elevator_exit_icq_fn = cfq_exit_icq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161 .elevator_set_req_fn = cfq_set_request,
4162 .elevator_put_req_fn = cfq_put_request,
4163 .elevator_may_queue_fn = cfq_may_queue,
4164 .elevator_init_fn = cfq_init_queue,
4165 .elevator_exit_fn = cfq_exit_queue,
4166 },
Tejun Heo3d3c2372011-12-14 00:33:42 +01004167 .icq_size = sizeof(struct cfq_io_cq),
4168 .icq_align = __alignof__(struct cfq_io_cq),
Al Viro3d1ab402006-03-18 18:35:43 -05004169 .elevator_attrs = cfq_attrs,
Tejun Heo3d3c2372011-12-14 00:33:42 +01004170 .elevator_name = "cfq",
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171 .elevator_owner = THIS_MODULE,
4172};
4173
Vivek Goyal3e252062009-12-04 10:36:42 -05004174#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004175static struct blkcg_policy blkcg_policy_cfq = {
Tejun Heof9fcc2d2012-04-16 13:57:27 -07004176 .pd_size = sizeof(struct cfq_group),
4177 .cftypes = cfq_blkcg_files,
4178
4179 .pd_init_fn = cfq_pd_init,
4180 .pd_reset_stats_fn = cfq_pd_reset_stats,
Vivek Goyal3e252062009-12-04 10:36:42 -05004181};
Vivek Goyal3e252062009-12-04 10:36:42 -05004182#endif
4183
Linus Torvalds1da177e2005-04-16 15:20:36 -07004184static int __init cfq_init(void)
4185{
Tejun Heo3d3c2372011-12-14 00:33:42 +01004186 int ret;
4187
Jens Axboe22e2c502005-06-27 10:55:12 +02004188 /*
4189 * could be 0 on HZ < 1000 setups
4190 */
4191 if (!cfq_slice_async)
4192 cfq_slice_async = 1;
4193 if (!cfq_slice_idle)
4194 cfq_slice_idle = 1;
4195
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004196#ifdef CONFIG_CFQ_GROUP_IOSCHED
4197 if (!cfq_group_idle)
4198 cfq_group_idle = 1;
Tejun Heo8bd435b2012-04-13 13:11:28 -07004199
Tejun Heo3c798392012-04-16 13:57:25 -07004200 ret = blkcg_policy_register(&blkcg_policy_cfq);
Tejun Heo8bd435b2012-04-13 13:11:28 -07004201 if (ret)
4202 return ret;
Tejun Heoffea73f2012-06-04 10:02:29 +02004203#else
4204 cfq_group_idle = 0;
4205#endif
Tejun Heo8bd435b2012-04-13 13:11:28 -07004206
Tejun Heofd794952012-06-04 10:01:38 +02004207 ret = -ENOMEM;
Tejun Heo3d3c2372011-12-14 00:33:42 +01004208 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4209 if (!cfq_pool)
Tejun Heo8bd435b2012-04-13 13:11:28 -07004210 goto err_pol_unreg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211
Tejun Heo3d3c2372011-12-14 00:33:42 +01004212 ret = elv_register(&iosched_cfq);
Tejun Heo8bd435b2012-04-13 13:11:28 -07004213 if (ret)
4214 goto err_free_pool;
Tejun Heo3d3c2372011-12-14 00:33:42 +01004215
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004216 return 0;
Tejun Heo8bd435b2012-04-13 13:11:28 -07004217
4218err_free_pool:
4219 kmem_cache_destroy(cfq_pool);
4220err_pol_unreg:
Tejun Heoffea73f2012-06-04 10:02:29 +02004221#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004222 blkcg_policy_unregister(&blkcg_policy_cfq);
Tejun Heoffea73f2012-06-04 10:02:29 +02004223#endif
Tejun Heo8bd435b2012-04-13 13:11:28 -07004224 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225}
4226
4227static void __exit cfq_exit(void)
4228{
Tejun Heoffea73f2012-06-04 10:02:29 +02004229#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004230 blkcg_policy_unregister(&blkcg_policy_cfq);
Tejun Heoffea73f2012-06-04 10:02:29 +02004231#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004232 elv_unregister(&iosched_cfq);
Tejun Heo3d3c2372011-12-14 00:33:42 +01004233 kmem_cache_destroy(cfq_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004234}
4235
4236module_init(cfq_init);
4237module_exit(cfq_exit);
4238
4239MODULE_AUTHOR("Jens Axboe");
4240MODULE_LICENSE("GPL");
4241MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");