blob: 3c38536bd52c3e3a25f1b8152e40a6ebe5192d36 [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"
Vivek Goyale98ef892010-06-18 10:39:47 -040018#include "cfq.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
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500174/* This is per cgroup per device grouping structure */
175struct cfq_group {
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500176 /* group service_tree member */
177 struct rb_node rb_node;
178
179 /* group service_tree key */
180 u64 vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500181 unsigned int weight;
Justin TerAvest8184f932011-03-17 16:12:36 +0100182 unsigned int new_weight;
183 bool needs_update;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500184
185 /* number of cfqq currently on this group */
186 int nr_cfqq;
187
Jens Axboe22e2c502005-06-27 10:55:12 +0200188 /*
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200189 * Per group busy queues average. Useful for workload slice calc. We
Vivek Goyalb4627322010-10-22 09:48:43 +0200190 * create the array for each prio class but at run time it is used
191 * only for RT and BE class and slot for IDLE class remains unused.
192 * This is primarily done to avoid confusion and a gcc warning.
193 */
194 unsigned int busy_queues_avg[CFQ_PRIO_NR];
195 /*
196 * rr lists of queues with requests. We maintain service trees for
197 * RT and BE classes. These trees are subdivided in subclasses
198 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199 * class there is no subclassification and all the cfq queues go on
200 * a single tree service_tree_idle.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100201 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200202 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100203 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100204 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500205
206 unsigned long saved_workload_slice;
207 enum wl_type_t saved_workload;
208 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500209 struct blkio_group blkg;
210#ifdef CONFIG_CFQ_GROUP_IOSCHED
211 struct hlist_node cfqd_node;
Shaohua Li329a6782011-01-07 08:48:28 +0100212 int ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500213#endif
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200214 /* number of requests that are on the dispatch list or inside driver */
215 int dispatched;
Shaohua Li7700fc42011-07-12 14:24:56 +0200216 struct cfq_ttime ttime;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500217};
218
Tejun Heoc5869802011-12-14 00:33:41 +0100219struct cfq_io_cq {
220 struct io_cq icq; /* must be the first member */
221 struct cfq_queue *cfqq[2];
222 struct cfq_ttime ttime;
223};
224
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500225/*
226 * Per block device queue structure
227 */
228struct cfq_data {
229 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500230 /* Root service tree for cfq_groups */
231 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500232 struct cfq_group root_group;
233
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100234 /*
235 * The priority currently being served
236 */
237 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100238 enum wl_type_t serving_type;
239 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500240 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200241
242 /*
243 * Each priority tree is sorted by next_request position. These
244 * trees are used when determining if two or more queues are
245 * interleaving requests (see cfq_close_cooperator).
246 */
247 struct rb_root prio_trees[CFQ_PRIO_LISTS];
248
Jens Axboe22e2c502005-06-27 10:55:12 +0200249 unsigned int busy_queues;
Shaohua Lief8a41d2011-03-07 09:26:29 +0100250 unsigned int busy_sync_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +0200251
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100252 int rq_in_driver;
253 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200254
255 /*
256 * queue-depth detection
257 */
258 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200259 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100260 /*
261 * hw_tag can be
262 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
263 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
264 * 0 => no NCQ
265 */
266 int hw_tag_est_depth;
267 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200268
269 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200270 * idle window management
271 */
272 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200273 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200274
275 struct cfq_queue *active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +0100276 struct cfq_io_cq *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200277
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200278 /*
279 * async queue for each priority case
280 */
281 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
282 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200283
Jens Axboe6d048f52007-04-25 12:44:27 +0200284 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286 /*
287 * tunables, see top of file
288 */
289 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200290 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 unsigned int cfq_back_penalty;
292 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200293 unsigned int cfq_slice[2];
294 unsigned int cfq_slice_async_rq;
295 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200296 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200297 unsigned int cfq_latency;
Tao Ma5bf14c02012-04-01 14:33:39 -0700298 unsigned int cfq_target_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500299
Jens Axboe6118b702009-06-30 09:34:12 +0200300 /*
301 * Fallback dummy cfqq for extreme OOM conditions
302 */
303 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200304
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100305 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500306
307 /* List of cfq groups being managed on this device*/
308 struct hlist_head cfqg_list;
Vivek Goyal56edf7d2011-05-19 15:38:22 -0400309
310 /* Number of groups which are on blkcg->blkg_list */
311 unsigned int nr_blkcg_linked_grps;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312};
313
Vivek Goyal25fb5162009-12-03 12:59:46 -0500314static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
315
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500316static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
317 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500318 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100319{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500320 if (!cfqg)
321 return NULL;
322
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100323 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500324 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100325
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500326 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100327}
328
Jens Axboe3b181522005-06-27 10:56:24 +0200329enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100330 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
331 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200332 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100333 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100334 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
335 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
336 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100337 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200338 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400339 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100340 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100341 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500342 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200343};
344
345#define CFQ_CFQQ_FNS(name) \
346static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
347{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100348 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200349} \
350static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
351{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100352 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200353} \
354static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
355{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100356 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200357}
358
359CFQ_CFQQ_FNS(on_rr);
360CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200361CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200362CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200363CFQ_CFQQ_FNS(fifo_expire);
364CFQ_CFQQ_FNS(idle_window);
365CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100366CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200367CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200368CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100369CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100370CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500371CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200372#undef CFQ_CFQQ_FNS
373
Vivek Goyalafc24d42010-04-26 19:27:56 +0200374#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500375#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
376 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
377 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200378 blkg_path(&(cfqq)->cfqg->blkg), ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500379
380#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
381 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200382 blkg_path(&(cfqg)->blkg), ##args) \
Vivek Goyal2868ef72009-12-03 12:59:48 -0500383
384#else
Jens Axboe7b679132008-05-30 12:23:07 +0200385#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
386 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200387#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500388#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200389#define cfq_log(cfqd, fmt, args...) \
390 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
391
Vivek Goyal615f0252009-12-03 12:59:39 -0500392/* Traverses through cfq group service trees */
393#define for_each_cfqg_st(cfqg, i, j, st) \
394 for (i = 0; i <= IDLE_WORKLOAD; i++) \
395 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
396 : &cfqg->service_tree_idle; \
397 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
398 (i == IDLE_WORKLOAD && j == 0); \
399 j++, st = i < IDLE_WORKLOAD ? \
400 &cfqg->service_trees[i][j]: NULL) \
401
Shaohua Lif5f2b6c2011-07-12 14:24:55 +0200402static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
403 struct cfq_ttime *ttime, bool group_idle)
404{
405 unsigned long slice;
406 if (!sample_valid(ttime->ttime_samples))
407 return false;
408 if (group_idle)
409 slice = cfqd->cfq_group_idle;
410 else
411 slice = cfqd->cfq_slice_idle;
412 return ttime->ttime_mean > slice;
413}
Vivek Goyal615f0252009-12-03 12:59:39 -0500414
Vivek Goyal02b35082010-08-23 12:23:53 +0200415static inline bool iops_mode(struct cfq_data *cfqd)
416{
417 /*
418 * If we are not idling on queues and it is a NCQ drive, parallel
419 * execution of requests is on and measuring time is not possible
420 * in most of the cases until and unless we drive shallower queue
421 * depths and that becomes a performance bottleneck. In such cases
422 * switch to start providing fairness in terms of number of IOs.
423 */
424 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
425 return true;
426 else
427 return false;
428}
429
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100430static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
431{
432 if (cfq_class_idle(cfqq))
433 return IDLE_WORKLOAD;
434 if (cfq_class_rt(cfqq))
435 return RT_WORKLOAD;
436 return BE_WORKLOAD;
437}
438
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100439
440static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
441{
442 if (!cfq_cfqq_sync(cfqq))
443 return ASYNC_WORKLOAD;
444 if (!cfq_cfqq_idle_window(cfqq))
445 return SYNC_NOIDLE_WORKLOAD;
446 return SYNC_WORKLOAD;
447}
448
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500449static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
450 struct cfq_data *cfqd,
451 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100452{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500453 if (wl == IDLE_WORKLOAD)
454 return cfqg->service_tree_idle.count;
455
456 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
457 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
458 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100459}
460
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500461static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
462 struct cfq_group *cfqg)
463{
464 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
465 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
466}
467
Jens Axboe165125e2007-07-24 09:28:11 +0200468static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200469static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100470 struct io_context *, gfp_t);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200471
Tejun Heoc5869802011-12-14 00:33:41 +0100472static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
473{
474 /* cic->icq is the first member, %NULL will convert to %NULL */
475 return container_of(icq, struct cfq_io_cq, icq);
476}
477
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100478static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
479 struct io_context *ioc)
480{
481 if (ioc)
482 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
483 return NULL;
484}
485
Tejun Heoc5869802011-12-14 00:33:41 +0100486static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200487{
Jens Axboea6151c32009-10-07 20:02:57 +0200488 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200489}
490
Tejun Heoc5869802011-12-14 00:33:41 +0100491static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
492 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200493{
Jens Axboea6151c32009-10-07 20:02:57 +0200494 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200495}
496
Tejun Heoc5869802011-12-14 00:33:41 +0100497static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400498{
Tejun Heoc5869802011-12-14 00:33:41 +0100499 return cic->icq.q->elevator->elevator_data;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400500}
501
Vasily Tarasov91fac312007-04-25 12:29:51 +0200502/*
503 * We regard a request as SYNC, if it's either a read or has the SYNC bit
504 * set (in which case it could also be direct WRITE).
505 */
Jens Axboea6151c32009-10-07 20:02:57 +0200506static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200507{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200508 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200509}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700512 * scheduler run of queue, if there are requests pending and no one in the
513 * driver that will restart queueing
514 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200515static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700516{
Jens Axboe7b679132008-05-30 12:23:07 +0200517 if (cfqd->busy_queues) {
518 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200519 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200520 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700521}
522
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100524 * Scale schedule slice based on io priority. Use the sync time slice only
525 * if a queue is marked sync and has sync io queued. A sync queue with async
526 * io only, should not get full sync slice length.
527 */
Jens Axboea6151c32009-10-07 20:02:57 +0200528static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200529 unsigned short prio)
530{
531 const int base_slice = cfqd->cfq_slice[sync];
532
533 WARN_ON(prio >= IOPRIO_BE_NR);
534
535 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
536}
537
Jens Axboe44f7c162007-01-19 11:51:58 +1100538static inline int
539cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
540{
Jens Axboed9e76202007-04-20 14:27:50 +0200541 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100542}
543
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500544static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
545{
546 u64 d = delta << CFQ_SERVICE_SHIFT;
547
548 d = d * BLKIO_WEIGHT_DEFAULT;
549 do_div(d, cfqg->weight);
550 return d;
551}
552
553static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
554{
555 s64 delta = (s64)(vdisktime - min_vdisktime);
556 if (delta > 0)
557 min_vdisktime = vdisktime;
558
559 return min_vdisktime;
560}
561
562static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
563{
564 s64 delta = (s64)(vdisktime - min_vdisktime);
565 if (delta < 0)
566 min_vdisktime = vdisktime;
567
568 return min_vdisktime;
569}
570
571static void update_min_vdisktime(struct cfq_rb_root *st)
572{
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500573 struct cfq_group *cfqg;
574
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500575 if (st->left) {
576 cfqg = rb_entry_cfqg(st->left);
Gui Jianfenga6032712011-03-07 09:28:09 +0100577 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
578 cfqg->vdisktime);
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500579 }
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500580}
581
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100582/*
583 * get averaged number of queues of RT/BE priority.
584 * average is updated, with a formula that gives more weight to higher numbers,
585 * to quickly follows sudden increases and decrease slowly
586 */
587
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500588static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
589 struct cfq_group *cfqg, bool rt)
Jens Axboe58696192009-10-28 09:27:07 +0100590{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100591 unsigned min_q, max_q;
592 unsigned mult = cfq_hist_divisor - 1;
593 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500594 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100595
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500596 min_q = min(cfqg->busy_queues_avg[rt], busy);
597 max_q = max(cfqg->busy_queues_avg[rt], busy);
598 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100599 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500600 return cfqg->busy_queues_avg[rt];
601}
602
603static inline unsigned
604cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
605{
606 struct cfq_rb_root *st = &cfqd->grp_service_tree;
607
Tao Ma5bf14c02012-04-01 14:33:39 -0700608 return cfqd->cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100609}
610
Shaohua Lic553f8e2011-01-14 08:41:03 +0100611static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -0700612cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100613{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100614 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
615 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500616 /*
617 * interested queues (we consider only the ones with the same
618 * priority class in the cfq group)
619 */
620 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
621 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100622 unsigned sync_slice = cfqd->cfq_slice[1];
623 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500624 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
625
626 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100627 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
628 /* scale low_slice according to IO priority
629 * and sync vs async */
630 unsigned low_slice =
631 min(slice, base_low_slice * slice / sync_slice);
632 /* the adapted slice value is scaled to fit all iqs
633 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500634 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100635 low_slice);
636 }
637 }
Shaohua Lic553f8e2011-01-14 08:41:03 +0100638 return slice;
639}
640
641static inline void
642cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
643{
Vivek Goyalba5bd522011-01-19 08:25:02 -0700644 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +0100645
Vivek Goyaldae739e2009-12-03 12:59:45 -0500646 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100647 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500648 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200649 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100650}
651
652/*
653 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
654 * isn't valid until the first request from the dispatch is activated
655 * and the slice time set.
656 */
Jens Axboea6151c32009-10-07 20:02:57 +0200657static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100658{
659 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +0100660 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100661 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +0100662 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100663
Shaohua Lic1e44752010-11-08 15:01:02 +0100664 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +1100665}
666
667/*
Jens Axboe5e705372006-07-13 12:39:25 +0200668 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200670 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 */
Jens Axboe5e705372006-07-13 12:39:25 +0200672static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100673cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100675 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200677#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
678#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
679 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
Jens Axboe5e705372006-07-13 12:39:25 +0200681 if (rq1 == NULL || rq1 == rq2)
682 return rq2;
683 if (rq2 == NULL)
684 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200685
Namhyung Kim229836b2011-05-24 10:23:21 +0200686 if (rq_is_sync(rq1) != rq_is_sync(rq2))
687 return rq_is_sync(rq1) ? rq1 : rq2;
688
Christoph Hellwig65299a32011-08-23 14:50:29 +0200689 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
690 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
Jens Axboeb53d1ed2011-08-19 08:34:48 +0200691
Tejun Heo83096eb2009-05-07 22:24:39 +0900692 s1 = blk_rq_pos(rq1);
693 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 /*
696 * by definition, 1KiB is 2 sectors
697 */
698 back_max = cfqd->cfq_back_max * 2;
699
700 /*
701 * Strict one way elevator _except_ in the case where we allow
702 * short backward seeks which are biased as twice the cost of a
703 * similar forward seek.
704 */
705 if (s1 >= last)
706 d1 = s1 - last;
707 else if (s1 + back_max >= last)
708 d1 = (last - s1) * cfqd->cfq_back_penalty;
709 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200710 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711
712 if (s2 >= last)
713 d2 = s2 - last;
714 else if (s2 + back_max >= last)
715 d2 = (last - s2) * cfqd->cfq_back_penalty;
716 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200717 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718
719 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720
Andreas Mohre8a99052006-03-28 08:59:49 +0200721 /*
722 * By doing switch() on the bit mask "wrap" we avoid having to
723 * check two variables for all permutations: --> faster!
724 */
725 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200726 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200727 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200728 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200729 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200730 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200731 else {
732 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200733 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200734 else
Jens Axboe5e705372006-07-13 12:39:25 +0200735 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200736 }
737
738 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200739 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200740 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200741 return rq2;
742 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200743 default:
744 /*
745 * Since both rqs are wrapped,
746 * start with the one that's further behind head
747 * (--> only *one* back seek required),
748 * since back seek takes more time than forward.
749 */
750 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200751 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 else
Jens Axboe5e705372006-07-13 12:39:25 +0200753 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 }
755}
756
Jens Axboe498d3aa2007-04-26 12:54:48 +0200757/*
758 * The below is leftmost cache rbtree addon
759 */
Jens Axboe08717142008-01-28 11:38:15 +0100760static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200761{
Vivek Goyal615f0252009-12-03 12:59:39 -0500762 /* Service tree is empty */
763 if (!root->count)
764 return NULL;
765
Jens Axboecc09e292007-04-26 12:53:50 +0200766 if (!root->left)
767 root->left = rb_first(&root->rb);
768
Jens Axboe08717142008-01-28 11:38:15 +0100769 if (root->left)
770 return rb_entry(root->left, struct cfq_queue, rb_node);
771
772 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200773}
774
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500775static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
776{
777 if (!root->left)
778 root->left = rb_first(&root->rb);
779
780 if (root->left)
781 return rb_entry_cfqg(root->left);
782
783 return NULL;
784}
785
Jens Axboea36e71f2009-04-15 12:15:11 +0200786static void rb_erase_init(struct rb_node *n, struct rb_root *root)
787{
788 rb_erase(n, root);
789 RB_CLEAR_NODE(n);
790}
791
Jens Axboecc09e292007-04-26 12:53:50 +0200792static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
793{
794 if (root->left == n)
795 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200796 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100797 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200798}
799
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800/*
801 * would be nice to take fifo expire time into account as well
802 */
Jens Axboe5e705372006-07-13 12:39:25 +0200803static struct request *
804cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
805 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806{
Jens Axboe21183b02006-07-13 12:33:14 +0200807 struct rb_node *rbnext = rb_next(&last->rb_node);
808 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200809 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810
Jens Axboe21183b02006-07-13 12:33:14 +0200811 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812
813 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200814 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200815
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200817 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200818 else {
819 rbnext = rb_first(&cfqq->sort_list);
820 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200821 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200822 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100824 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825}
826
Jens Axboed9e76202007-04-20 14:27:50 +0200827static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
828 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829{
Jens Axboed9e76202007-04-20 14:27:50 +0200830 /*
831 * just an approximation, should be ok.
832 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500833 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100834 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200835}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500837static inline s64
838cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
839{
840 return cfqg->vdisktime - st->min_vdisktime;
841}
842
843static void
844__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
845{
846 struct rb_node **node = &st->rb.rb_node;
847 struct rb_node *parent = NULL;
848 struct cfq_group *__cfqg;
849 s64 key = cfqg_key(st, cfqg);
850 int left = 1;
851
852 while (*node != NULL) {
853 parent = *node;
854 __cfqg = rb_entry_cfqg(parent);
855
856 if (key < cfqg_key(st, __cfqg))
857 node = &parent->rb_left;
858 else {
859 node = &parent->rb_right;
860 left = 0;
861 }
862 }
863
864 if (left)
865 st->left = &cfqg->rb_node;
866
867 rb_link_node(&cfqg->rb_node, parent, node);
868 rb_insert_color(&cfqg->rb_node, &st->rb);
869}
870
871static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100872cfq_update_group_weight(struct cfq_group *cfqg)
873{
874 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
875 if (cfqg->needs_update) {
876 cfqg->weight = cfqg->new_weight;
877 cfqg->needs_update = false;
878 }
879}
880
881static void
882cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
883{
884 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
885
886 cfq_update_group_weight(cfqg);
887 __cfq_group_service_tree_add(st, cfqg);
888 st->total_weight += cfqg->weight;
889}
890
891static void
892cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500893{
894 struct cfq_rb_root *st = &cfqd->grp_service_tree;
895 struct cfq_group *__cfqg;
896 struct rb_node *n;
897
898 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +0100899 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500900 return;
901
902 /*
903 * Currently put the group at the end. Later implement something
904 * so that groups get lesser vtime based on their weights, so that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300905 * if group does not loose all if it was not continuously backlogged.
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500906 */
907 n = rb_last(&st->rb);
908 if (n) {
909 __cfqg = rb_entry_cfqg(n);
910 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
911 } else
912 cfqg->vdisktime = st->min_vdisktime;
Justin TerAvest8184f932011-03-17 16:12:36 +0100913 cfq_group_service_tree_add(st, cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500914}
915
916static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100917cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
918{
919 st->total_weight -= cfqg->weight;
920 if (!RB_EMPTY_NODE(&cfqg->rb_node))
921 cfq_rb_erase(&cfqg->rb_node, st);
922}
923
924static void
925cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500926{
927 struct cfq_rb_root *st = &cfqd->grp_service_tree;
928
929 BUG_ON(cfqg->nr_cfqq < 1);
930 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500931
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500932 /* If there are other cfq queues under this group, don't delete it */
933 if (cfqg->nr_cfqq)
934 return;
935
Vivek Goyal2868ef72009-12-03 12:59:48 -0500936 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Justin TerAvest8184f932011-03-17 16:12:36 +0100937 cfq_group_service_tree_del(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500938 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400939 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500940}
941
Justin TerAvest167400d2011-03-12 16:54:00 +0100942static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
943 unsigned int *unaccounted_time)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500944{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500945 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500946
947 /*
948 * Queue got expired before even a single request completed or
949 * got expired immediately after first request completion.
950 */
951 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
952 /*
953 * Also charge the seek time incurred to the group, otherwise
954 * if there are mutiple queues in the group, each can dispatch
955 * a single request on seeky media and cause lots of seek time
956 * and group will never know it.
957 */
958 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
959 1);
960 } else {
961 slice_used = jiffies - cfqq->slice_start;
Justin TerAvest167400d2011-03-12 16:54:00 +0100962 if (slice_used > cfqq->allocated_slice) {
963 *unaccounted_time = slice_used - cfqq->allocated_slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500964 slice_used = cfqq->allocated_slice;
Justin TerAvest167400d2011-03-12 16:54:00 +0100965 }
966 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
967 *unaccounted_time += cfqq->slice_start -
968 cfqq->dispatch_start;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500969 }
970
Vivek Goyaldae739e2009-12-03 12:59:45 -0500971 return slice_used;
972}
973
974static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200975 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500976{
977 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Justin TerAvest167400d2011-03-12 16:54:00 +0100978 unsigned int used_sl, charge, unaccounted_sl = 0;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500979 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
980 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500981
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500982 BUG_ON(nr_sync < 0);
Justin TerAvest167400d2011-03-12 16:54:00 +0100983 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500984
Vivek Goyal02b35082010-08-23 12:23:53 +0200985 if (iops_mode(cfqd))
986 charge = cfqq->slice_dispatch;
987 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
988 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500989
990 /* Can't update vdisktime while group is on service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +0100991 cfq_group_service_tree_del(st, cfqg);
Vivek Goyal02b35082010-08-23 12:23:53 +0200992 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Justin TerAvest8184f932011-03-17 16:12:36 +0100993 /* If a new weight was requested, update now, off tree */
994 cfq_group_service_tree_add(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500995
996 /* This group is being expired. Save the context */
997 if (time_after(cfqd->workload_expires, jiffies)) {
998 cfqg->saved_workload_slice = cfqd->workload_expires
999 - jiffies;
1000 cfqg->saved_workload = cfqd->serving_type;
1001 cfqg->saved_serving_prio = cfqd->serving_prio;
1002 } else
1003 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -05001004
1005 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1006 st->min_vdisktime);
Joe Perchesfd16d262011-06-13 10:42:49 +02001007 cfq_log_cfqq(cfqq->cfqd, cfqq,
1008 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1009 used_sl, cfqq->slice_dispatch, charge,
1010 iops_mode(cfqd), cfqq->nr_sectors);
Justin TerAvest167400d2011-03-12 16:54:00 +01001011 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1012 unaccounted_sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001013 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001014}
1015
Vivek Goyal25fb5162009-12-03 12:59:46 -05001016#ifdef CONFIG_CFQ_GROUP_IOSCHED
1017static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1018{
1019 if (blkg)
1020 return container_of(blkg, struct cfq_group, blkg);
1021 return NULL;
1022}
1023
Paul Bolle8aea4542011-06-06 05:07:54 +02001024static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1025 unsigned int weight)
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001026{
Justin TerAvest8184f932011-03-17 16:12:36 +01001027 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1028 cfqg->new_weight = weight;
1029 cfqg->needs_update = true;
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001030}
1031
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001032static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1033 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001034{
Vivek Goyal22084192009-12-03 12:59:49 -05001035 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1036 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001037
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001038 /*
1039 * Add group onto cgroup list. It might happen that bdi->dev is
1040 * not initialized yet. Initialize this new group without major
1041 * and minor info and this info will be filled in once a new thread
1042 * comes for IO.
1043 */
1044 if (bdi->dev) {
Ricky Beniteza74b2ad2010-04-05 18:22:17 +02001045 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001046 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1047 (void *)cfqd, MKDEV(major, minor));
1048 } else
1049 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1050 (void *)cfqd, 0);
1051
1052 cfqd->nr_blkcg_linked_grps++;
1053 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1054
1055 /* Add group on cfqd list */
1056 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1057}
1058
1059/*
1060 * Should be called from sleepable context. No request queue lock as per
1061 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1062 * from sleepable context.
1063 */
1064static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1065{
1066 struct cfq_group *cfqg = NULL;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001067 int i, j, ret;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001068 struct cfq_rb_root *st;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001069
1070 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1071 if (!cfqg)
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001072 return NULL;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001073
Vivek Goyal25fb5162009-12-03 12:59:46 -05001074 for_each_cfqg_st(cfqg, i, j, st)
1075 *st = CFQ_RB_ROOT;
1076 RB_CLEAR_NODE(&cfqg->rb_node);
1077
Shaohua Li7700fc42011-07-12 14:24:56 +02001078 cfqg->ttime.last_end_request = jiffies;
1079
Vivek Goyalb1c35762009-12-03 12:59:47 -05001080 /*
1081 * Take the initial reference that will be released on destroy
1082 * This can be thought of a joint reference by cgroup and
1083 * elevator which will be dropped by either elevator exit
1084 * or cgroup deletion path depending on who is exiting first.
1085 */
Shaohua Li329a6782011-01-07 08:48:28 +01001086 cfqg->ref = 1;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001087
1088 ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1089 if (ret) {
1090 kfree(cfqg);
1091 return NULL;
1092 }
1093
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001094 return cfqg;
1095}
1096
1097static struct cfq_group *
1098cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1099{
1100 struct cfq_group *cfqg = NULL;
1101 void *key = cfqd;
1102 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1103 unsigned int major, minor;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001104
Vivek Goyal180be2a2010-09-14 08:47:11 +02001105 /*
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001106 * This is the common case when there are no blkio cgroups.
1107 * Avoid lookup in this case
Vivek Goyal180be2a2010-09-14 08:47:11 +02001108 */
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001109 if (blkcg == &blkio_root_cgroup)
1110 cfqg = &cfqd->root_group;
1111 else
1112 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1113
1114 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
Vivek Goyal180be2a2010-09-14 08:47:11 +02001115 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001116 cfqg->blkg.dev = MKDEV(major, minor);
1117 }
Vivek Goyal180be2a2010-09-14 08:47:11 +02001118
Vivek Goyal25fb5162009-12-03 12:59:46 -05001119 return cfqg;
1120}
1121
1122/*
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001123 * Search for the cfq group current task belongs to. request_queue lock must
1124 * be held.
Vivek Goyal25fb5162009-12-03 12:59:46 -05001125 */
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001126static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001127{
Vivek Goyal70087dc2011-05-16 15:24:08 +02001128 struct blkio_cgroup *blkcg;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001129 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1130 struct request_queue *q = cfqd->queue;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001131
1132 rcu_read_lock();
Vivek Goyal70087dc2011-05-16 15:24:08 +02001133 blkcg = task_blkio_cgroup(current);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001134 cfqg = cfq_find_cfqg(cfqd, blkcg);
1135 if (cfqg) {
1136 rcu_read_unlock();
1137 return cfqg;
1138 }
1139
1140 /*
1141 * Need to allocate a group. Allocation of group also needs allocation
1142 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1143 * we need to drop rcu lock and queue_lock before we call alloc.
1144 *
1145 * Not taking any queue reference here and assuming that queue is
1146 * around by the time we return. CFQ queue allocation code does
1147 * the same. It might be racy though.
1148 */
1149
1150 rcu_read_unlock();
1151 spin_unlock_irq(q->queue_lock);
1152
1153 cfqg = cfq_alloc_cfqg(cfqd);
1154
1155 spin_lock_irq(q->queue_lock);
1156
1157 rcu_read_lock();
1158 blkcg = task_blkio_cgroup(current);
1159
1160 /*
1161 * If some other thread already allocated the group while we were
1162 * not holding queue lock, free up the group
1163 */
1164 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1165
1166 if (__cfqg) {
1167 kfree(cfqg);
1168 rcu_read_unlock();
1169 return __cfqg;
1170 }
1171
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001172 if (!cfqg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001173 cfqg = &cfqd->root_group;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001174
1175 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001176 rcu_read_unlock();
1177 return cfqg;
1178}
1179
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001180static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1181{
Shaohua Li329a6782011-01-07 08:48:28 +01001182 cfqg->ref++;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001183 return cfqg;
1184}
1185
Vivek Goyal25fb5162009-12-03 12:59:46 -05001186static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1187{
1188 /* Currently, all async queues are mapped to root group */
1189 if (!cfq_cfqq_sync(cfqq))
1190 cfqg = &cfqq->cfqd->root_group;
1191
1192 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001193 /* cfqq reference on cfqg */
Shaohua Li329a6782011-01-07 08:48:28 +01001194 cfqq->cfqg->ref++;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001195}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001196
1197static void cfq_put_cfqg(struct cfq_group *cfqg)
1198{
1199 struct cfq_rb_root *st;
1200 int i, j;
1201
Shaohua Li329a6782011-01-07 08:48:28 +01001202 BUG_ON(cfqg->ref <= 0);
1203 cfqg->ref--;
1204 if (cfqg->ref)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001205 return;
1206 for_each_cfqg_st(cfqg, i, j, st)
Gui Jianfengb54ce602010-11-30 20:52:46 +01001207 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001208 free_percpu(cfqg->blkg.stats_cpu);
Vivek Goyalb1c35762009-12-03 12:59:47 -05001209 kfree(cfqg);
1210}
1211
1212static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1213{
1214 /* Something wrong if we are trying to remove same group twice */
1215 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1216
1217 hlist_del_init(&cfqg->cfqd_node);
1218
Vivek Goyala5395b82011-08-02 09:24:09 +02001219 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1220 cfqd->nr_blkcg_linked_grps--;
1221
Vivek Goyalb1c35762009-12-03 12:59:47 -05001222 /*
1223 * Put the reference taken at the time of creation so that when all
1224 * queues are gone, group can be destroyed.
1225 */
1226 cfq_put_cfqg(cfqg);
1227}
1228
1229static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1230{
1231 struct hlist_node *pos, *n;
1232 struct cfq_group *cfqg;
1233
1234 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1235 /*
1236 * If cgroup removal path got to blk_group first and removed
1237 * it from cgroup list, then it will take care of destroying
1238 * cfqg also.
1239 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001240 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001241 cfq_destroy_cfqg(cfqd, cfqg);
1242 }
1243}
1244
1245/*
1246 * Blk cgroup controller notification saying that blkio_group object is being
1247 * delinked as associated cgroup object is going away. That also means that
1248 * no new IO will come in this group. So get rid of this group as soon as
1249 * any pending IO in the group is finished.
1250 *
1251 * This function is called under rcu_read_lock(). key is the rcu protected
1252 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1253 * read lock.
1254 *
1255 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1256 * it should not be NULL as even if elevator was exiting, cgroup deltion
1257 * path got to it first.
1258 */
Paul Bolle8aea4542011-06-06 05:07:54 +02001259static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001260{
1261 unsigned long flags;
1262 struct cfq_data *cfqd = key;
1263
1264 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1265 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1266 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1267}
1268
Vivek Goyal25fb5162009-12-03 12:59:46 -05001269#else /* GROUP_IOSCHED */
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001270static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001271{
1272 return &cfqd->root_group;
1273}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001274
1275static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1276{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001277 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001278}
1279
Vivek Goyal25fb5162009-12-03 12:59:46 -05001280static inline void
1281cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1282 cfqq->cfqg = cfqg;
1283}
1284
Vivek Goyalb1c35762009-12-03 12:59:47 -05001285static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1286static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1287
Vivek Goyal25fb5162009-12-03 12:59:46 -05001288#endif /* GROUP_IOSCHED */
1289
Jens Axboe498d3aa2007-04-26 12:54:48 +02001290/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001291 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa2007-04-26 12:54:48 +02001292 * requests waiting to be processed. It is sorted in the order that
1293 * we will service the queues.
1294 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001295static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001296 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001297{
Jens Axboe08717142008-01-28 11:38:15 +01001298 struct rb_node **p, *parent;
1299 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001300 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001301 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa2007-04-26 12:54:48 +02001302 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001303 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001304
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001305 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001306 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001307 if (cfq_class_idle(cfqq)) {
1308 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001309 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001310 if (parent && parent != &cfqq->rb_node) {
1311 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1312 rb_key += __cfqq->rb_key;
1313 } else
1314 rb_key += jiffies;
1315 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001316 /*
1317 * Get our rb key offset. Subtract any residual slice
1318 * value carried from last service. A negative resid
1319 * count indicates slice overrun, and this should position
1320 * the next service time further away in the tree.
1321 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001322 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001323 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001324 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001325 } else {
1326 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001327 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001328 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1329 }
Jens Axboed9e76202007-04-20 14:27:50 +02001330
1331 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001332 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001333 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001334 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001335 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001336 if (rb_key == cfqq->rb_key &&
1337 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001338 return;
Jens Axboe53b03742006-07-28 09:48:51 +02001339
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001340 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1341 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001342 }
Jens Axboed9e76202007-04-20 14:27:50 +02001343
Jens Axboe498d3aa2007-04-26 12:54:48 +02001344 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001345 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001346 cfqq->service_tree = service_tree;
1347 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001348 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001349 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001350
Jens Axboed9e76202007-04-20 14:27:50 +02001351 parent = *p;
1352 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1353
Jens Axboe0c534e02007-04-18 20:01:57 +02001354 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001355 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001356 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001357 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001358 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001359 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001360 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001361 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001362 }
Jens Axboe67060e32007-04-18 20:13:32 +02001363
1364 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001365 }
1366
Jens Axboecc09e292007-04-26 12:53:50 +02001367 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001368 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001369
Jens Axboed9e76202007-04-20 14:27:50 +02001370 cfqq->rb_key = rb_key;
1371 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001372 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1373 service_tree->count++;
Namhyung Kim20359f22011-05-24 10:23:22 +02001374 if (add_front || !new_cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001375 return;
Justin TerAvest8184f932011-03-17 16:12:36 +01001376 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377}
1378
Jens Axboea36e71f2009-04-15 12:15:11 +02001379static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001380cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1381 sector_t sector, struct rb_node **ret_parent,
1382 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001383{
Jens Axboea36e71f2009-04-15 12:15:11 +02001384 struct rb_node **p, *parent;
1385 struct cfq_queue *cfqq = NULL;
1386
1387 parent = NULL;
1388 p = &root->rb_node;
1389 while (*p) {
1390 struct rb_node **n;
1391
1392 parent = *p;
1393 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1394
1395 /*
1396 * Sort strictly based on sector. Smallest to the left,
1397 * largest to the right.
1398 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001399 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001400 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001401 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001402 n = &(*p)->rb_left;
1403 else
1404 break;
1405 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001406 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001407 }
1408
1409 *ret_parent = parent;
1410 if (rb_link)
1411 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001412 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001413}
1414
1415static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1416{
Jens Axboea36e71f2009-04-15 12:15:11 +02001417 struct rb_node **p, *parent;
1418 struct cfq_queue *__cfqq;
1419
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001420 if (cfqq->p_root) {
1421 rb_erase(&cfqq->p_node, cfqq->p_root);
1422 cfqq->p_root = NULL;
1423 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001424
1425 if (cfq_class_idle(cfqq))
1426 return;
1427 if (!cfqq->next_rq)
1428 return;
1429
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001430 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001431 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1432 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001433 if (!__cfqq) {
1434 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001435 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1436 } else
1437 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001438}
1439
Jens Axboe498d3aa2007-04-26 12:54:48 +02001440/*
1441 * Update cfqq's position in the service tree.
1442 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001443static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001444{
Jens Axboe6d048f52007-04-25 12:44:27 +02001445 /*
1446 * Resorting requires the cfqq to be on the RR list already.
1447 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001448 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001449 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001450 cfq_prio_tree_add(cfqd, cfqq);
1451 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001452}
1453
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454/*
1455 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001456 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 */
Jens Axboefebffd62008-01-28 13:19:43 +01001458static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459{
Jens Axboe7b679132008-05-30 12:23:07 +02001460 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001461 BUG_ON(cfq_cfqq_on_rr(cfqq));
1462 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 cfqd->busy_queues++;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001464 if (cfq_cfqq_sync(cfqq))
1465 cfqd->busy_sync_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466
Jens Axboeedd75ff2007-04-19 12:03:34 +02001467 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468}
1469
Jens Axboe498d3aa2007-04-26 12:54:48 +02001470/*
1471 * Called when the cfqq no longer has requests pending, remove it from
1472 * the service tree.
1473 */
Jens Axboefebffd62008-01-28 13:19:43 +01001474static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475{
Jens Axboe7b679132008-05-30 12:23:07 +02001476 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001477 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1478 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001480 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1481 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1482 cfqq->service_tree = NULL;
1483 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001484 if (cfqq->p_root) {
1485 rb_erase(&cfqq->p_node, cfqq->p_root);
1486 cfqq->p_root = NULL;
1487 }
Jens Axboed9e76202007-04-20 14:27:50 +02001488
Justin TerAvest8184f932011-03-17 16:12:36 +01001489 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 BUG_ON(!cfqd->busy_queues);
1491 cfqd->busy_queues--;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001492 if (cfq_cfqq_sync(cfqq))
1493 cfqd->busy_sync_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494}
1495
1496/*
1497 * rb tree support functions
1498 */
Jens Axboefebffd62008-01-28 13:19:43 +01001499static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500{
Jens Axboe5e705372006-07-13 12:39:25 +02001501 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001502 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503
Jens Axboeb4878f22005-10-20 16:42:29 +02001504 BUG_ON(!cfqq->queued[sync]);
1505 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506
Jens Axboe5e705372006-07-13 12:39:25 +02001507 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508
Vivek Goyalf04a6422009-12-03 12:59:40 -05001509 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1510 /*
1511 * Queue will be deleted from service tree when we actually
1512 * expire it later. Right now just remove it from prio tree
1513 * as it is empty.
1514 */
1515 if (cfqq->p_root) {
1516 rb_erase(&cfqq->p_node, cfqq->p_root);
1517 cfqq->p_root = NULL;
1518 }
1519 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520}
1521
Jens Axboe5e705372006-07-13 12:39:25 +02001522static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523{
Jens Axboe5e705372006-07-13 12:39:25 +02001524 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 struct cfq_data *cfqd = cfqq->cfqd;
Jeff Moyer796d5112011-06-02 21:19:05 +02001526 struct request *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527
Jens Axboe5380a102006-07-13 12:37:56 +02001528 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529
Jeff Moyer796d5112011-06-02 21:19:05 +02001530 elv_rb_add(&cfqq->sort_list, rq);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001531
1532 if (!cfq_cfqq_on_rr(cfqq))
1533 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001534
1535 /*
1536 * check if this request is a better next-serve candidate
1537 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001538 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001539 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001540
1541 /*
1542 * adjust priority tree position, if ->next_rq changes
1543 */
1544 if (prev != cfqq->next_rq)
1545 cfq_prio_tree_add(cfqd, cfqq);
1546
Jens Axboe5044eed2007-04-25 11:53:48 +02001547 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548}
1549
Jens Axboefebffd62008-01-28 13:19:43 +01001550static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551{
Jens Axboe5380a102006-07-13 12:37:56 +02001552 elv_rb_del(&cfqq->sort_list, rq);
1553 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001554 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1555 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001556 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001557 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001558 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1559 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560}
1561
Jens Axboe206dc692006-03-28 13:03:44 +02001562static struct request *
1563cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564{
Jens Axboe206dc692006-03-28 13:03:44 +02001565 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01001566 struct cfq_io_cq *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001567 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568
Jens Axboe4ac845a2008-01-24 08:44:49 +01001569 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001570 if (!cic)
1571 return NULL;
1572
1573 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001574 if (cfqq) {
1575 sector_t sector = bio->bi_sector + bio_sectors(bio);
1576
Jens Axboe21183b02006-07-13 12:33:14 +02001577 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001578 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 return NULL;
1581}
1582
Jens Axboe165125e2007-07-24 09:28:11 +02001583static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001584{
1585 struct cfq_data *cfqd = q->elevator->elevator_data;
1586
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001587 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001588 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001589 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001590
Tejun Heo5b936292009-05-07 22:24:38 +09001591 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001592}
1593
Jens Axboe165125e2007-07-24 09:28:11 +02001594static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595{
Jens Axboe22e2c502005-06-27 10:55:12 +02001596 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001598 WARN_ON(!cfqd->rq_in_driver);
1599 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001600 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001601 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602}
1603
Jens Axboeb4878f22005-10-20 16:42:29 +02001604static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605{
Jens Axboe5e705372006-07-13 12:39:25 +02001606 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001607
Jens Axboe5e705372006-07-13 12:39:25 +02001608 if (cfqq->next_rq == rq)
1609 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610
Jens Axboeb4878f22005-10-20 16:42:29 +02001611 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001612 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001613
Aaron Carroll45333d52008-08-26 15:52:36 +02001614 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001615 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1616 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig65299a32011-08-23 14:50:29 +02001617 if (rq->cmd_flags & REQ_PRIO) {
1618 WARN_ON(!cfqq->prio_pending);
1619 cfqq->prio_pending--;
Jens Axboeb53d1ed2011-08-19 08:34:48 +02001620 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621}
1622
Jens Axboe165125e2007-07-24 09:28:11 +02001623static int cfq_merge(struct request_queue *q, struct request **req,
1624 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625{
1626 struct cfq_data *cfqd = q->elevator->elevator_data;
1627 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628
Jens Axboe206dc692006-03-28 13:03:44 +02001629 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001630 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001631 *req = __rq;
1632 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 }
1634
1635 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636}
1637
Jens Axboe165125e2007-07-24 09:28:11 +02001638static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001639 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640{
Jens Axboe21183b02006-07-13 12:33:14 +02001641 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001642 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643
Jens Axboe5e705372006-07-13 12:39:25 +02001644 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646}
1647
Divyesh Shah812d4022010-04-08 21:14:23 -07001648static void cfq_bio_merged(struct request_queue *q, struct request *req,
1649 struct bio *bio)
1650{
Vivek Goyale98ef892010-06-18 10:39:47 -04001651 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1652 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001653}
1654
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655static void
Jens Axboe165125e2007-07-24 09:28:11 +02001656cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 struct request *next)
1658{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001659 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Shaohua Li4a0b75c2011-12-16 14:00:22 +01001660 struct cfq_data *cfqd = q->elevator->elevator_data;
1661
Jens Axboe22e2c502005-06-27 10:55:12 +02001662 /*
1663 * reposition in fifo if next is older than rq
1664 */
1665 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001666 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001667 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001668 rq_set_fifo_time(rq, rq_fifo_time(next));
1669 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001670
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001671 if (cfqq->next_rq == next)
1672 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001673 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001674 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1675 rq_data_dir(next), rq_is_sync(next));
Shaohua Li4a0b75c2011-12-16 14:00:22 +01001676
1677 cfqq = RQ_CFQQ(next);
1678 /*
1679 * all requests of this queue are merged to other queues, delete it
1680 * from the service tree. If it's the active_queue,
1681 * cfq_dispatch_requests() will choose to expire it or do idle
1682 */
1683 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1684 cfqq != cfqd->active_queue)
1685 cfq_del_cfqq_rr(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001686}
1687
Jens Axboe165125e2007-07-24 09:28:11 +02001688static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001689 struct bio *bio)
1690{
1691 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heoc5869802011-12-14 00:33:41 +01001692 struct cfq_io_cq *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001693 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001694
1695 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001696 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001697 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001698 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001699 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001700
1701 /*
Tejun Heof1a4f4d2011-12-14 00:33:39 +01001702 * Lookup the cfqq that this bio will be queued with and allow
Tejun Heo07c2bd32012-02-08 09:19:42 +01001703 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001704 */
Tejun Heo07c2bd32012-02-08 09:19:42 +01001705 cic = cfq_cic_lookup(cfqd, current->io_context);
1706 if (!cic)
1707 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001708
Vasily Tarasov91fac312007-04-25 12:29:51 +02001709 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001710 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001711}
1712
Divyesh Shah812df482010-04-08 21:15:35 -07001713static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1714{
1715 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001716 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001717}
1718
Jens Axboefebffd62008-01-28 13:19:43 +01001719static void __cfq_set_active_queue(struct cfq_data *cfqd,
1720 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001721{
1722 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001723 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1724 cfqd->serving_prio, cfqd->serving_type);
Justin TerAvest62a37f62011-03-23 08:25:44 +01001725 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1726 cfqq->slice_start = 0;
1727 cfqq->dispatch_start = jiffies;
1728 cfqq->allocated_slice = 0;
1729 cfqq->slice_end = 0;
1730 cfqq->slice_dispatch = 0;
1731 cfqq->nr_sectors = 0;
1732
1733 cfq_clear_cfqq_wait_request(cfqq);
1734 cfq_clear_cfqq_must_dispatch(cfqq);
1735 cfq_clear_cfqq_must_alloc_slice(cfqq);
1736 cfq_clear_cfqq_fifo_expire(cfqq);
1737 cfq_mark_cfqq_slice_new(cfqq);
1738
1739 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001740 }
1741
1742 cfqd->active_queue = cfqq;
1743}
1744
1745/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001746 * current cfqq expired its slice (or was too idle), select new one
1747 */
1748static void
1749__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001750 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001751{
Jens Axboe7b679132008-05-30 12:23:07 +02001752 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1753
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001754 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001755 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001756
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001757 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001758 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001759
1760 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001761 * If this cfqq is shared between multiple processes, check to
1762 * make sure that those processes are still issuing I/Os within
1763 * the mean seek distance. If not, it may be time to break the
1764 * queues apart again.
1765 */
1766 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1767 cfq_mark_cfqq_split_coop(cfqq);
1768
1769 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001770 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001771 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01001772 if (timed_out) {
1773 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07001774 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001775 else
1776 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001777 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1778 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001779
Vivek Goyale5ff0822010-04-26 19:25:11 +02001780 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001781
Vivek Goyalf04a6422009-12-03 12:59:40 -05001782 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1783 cfq_del_cfqq_rr(cfqd, cfqq);
1784
Jens Axboeedd75ff2007-04-19 12:03:34 +02001785 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001786
1787 if (cfqq == cfqd->active_queue)
1788 cfqd->active_queue = NULL;
1789
1790 if (cfqd->active_cic) {
Tejun Heo11a31222012-02-07 07:51:30 +01001791 put_io_context(cfqd->active_cic->icq.ioc);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001792 cfqd->active_cic = NULL;
1793 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001794}
1795
Vivek Goyale5ff0822010-04-26 19:25:11 +02001796static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001797{
1798 struct cfq_queue *cfqq = cfqd->active_queue;
1799
1800 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001801 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001802}
1803
Jens Axboe498d3aa2007-04-26 12:54:48 +02001804/*
1805 * Get next queue for service. Unless we have a queue preemption,
1806 * we'll simply select the first cfqq in the service tree.
1807 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001808static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001809{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001810 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001811 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001812 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001813
Vivek Goyalf04a6422009-12-03 12:59:40 -05001814 if (!cfqd->rq_queued)
1815 return NULL;
1816
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001817 /* There is nothing to dispatch */
1818 if (!service_tree)
1819 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001820 if (RB_EMPTY_ROOT(&service_tree->rb))
1821 return NULL;
1822 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001823}
1824
Vivek Goyalf04a6422009-12-03 12:59:40 -05001825static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1826{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001827 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001828 struct cfq_queue *cfqq;
1829 int i, j;
1830 struct cfq_rb_root *st;
1831
1832 if (!cfqd->rq_queued)
1833 return NULL;
1834
Vivek Goyal25fb5162009-12-03 12:59:46 -05001835 cfqg = cfq_get_next_cfqg(cfqd);
1836 if (!cfqg)
1837 return NULL;
1838
Vivek Goyalf04a6422009-12-03 12:59:40 -05001839 for_each_cfqg_st(cfqg, i, j, st)
1840 if ((cfqq = cfq_rb_first(st)) != NULL)
1841 return cfqq;
1842 return NULL;
1843}
1844
Jens Axboe498d3aa2007-04-26 12:54:48 +02001845/*
1846 * Get and set a new active queue for service.
1847 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001848static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1849 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001850{
Jens Axboee00ef792009-11-04 08:54:55 +01001851 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001852 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001853
Jens Axboe22e2c502005-06-27 10:55:12 +02001854 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001855 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001856}
1857
Jens Axboed9e76202007-04-20 14:27:50 +02001858static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1859 struct request *rq)
1860{
Tejun Heo83096eb2009-05-07 22:24:39 +09001861 if (blk_rq_pos(rq) >= cfqd->last_position)
1862 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001863 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001864 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001865}
1866
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001867static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001868 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001869{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001870 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001871}
1872
Jens Axboea36e71f2009-04-15 12:15:11 +02001873static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1874 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001875{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001876 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001877 struct rb_node *parent, *node;
1878 struct cfq_queue *__cfqq;
1879 sector_t sector = cfqd->last_position;
1880
1881 if (RB_EMPTY_ROOT(root))
1882 return NULL;
1883
1884 /*
1885 * First, if we find a request starting at the end of the last
1886 * request, choose it.
1887 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001888 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001889 if (__cfqq)
1890 return __cfqq;
1891
1892 /*
1893 * If the exact sector wasn't found, the parent of the NULL leaf
1894 * will contain the closest sector.
1895 */
1896 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001897 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001898 return __cfqq;
1899
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001900 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001901 node = rb_next(&__cfqq->p_node);
1902 else
1903 node = rb_prev(&__cfqq->p_node);
1904 if (!node)
1905 return NULL;
1906
1907 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001908 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001909 return __cfqq;
1910
1911 return NULL;
1912}
1913
1914/*
1915 * cfqd - obvious
1916 * cur_cfqq - passed in so that we don't decide that the current queue is
1917 * closely cooperating with itself.
1918 *
1919 * So, basically we're assuming that that cur_cfqq has dispatched at least
1920 * one request, and that cfqd->last_position reflects a position on the disk
1921 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1922 * assumption.
1923 */
1924static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001925 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001926{
1927 struct cfq_queue *cfqq;
1928
Divyesh Shah39c01b22010-03-25 15:45:57 +01001929 if (cfq_class_idle(cur_cfqq))
1930 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001931 if (!cfq_cfqq_sync(cur_cfqq))
1932 return NULL;
1933 if (CFQQ_SEEKY(cur_cfqq))
1934 return NULL;
1935
Jens Axboea36e71f2009-04-15 12:15:11 +02001936 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001937 * Don't search priority tree if it's the only queue in the group.
1938 */
1939 if (cur_cfqq->cfqg->nr_cfqq == 1)
1940 return NULL;
1941
1942 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001943 * We should notice if some of the queues are cooperating, eg
1944 * working closely on the same area of the disk. In that case,
1945 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001946 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001947 cfqq = cfqq_close(cfqd, cur_cfqq);
1948 if (!cfqq)
1949 return NULL;
1950
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001951 /* If new queue belongs to different cfq_group, don't choose it */
1952 if (cur_cfqq->cfqg != cfqq->cfqg)
1953 return NULL;
1954
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001955 /*
1956 * It only makes sense to merge sync queues.
1957 */
1958 if (!cfq_cfqq_sync(cfqq))
1959 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001960 if (CFQQ_SEEKY(cfqq))
1961 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001962
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001963 /*
1964 * Do not merge queues of different priority classes
1965 */
1966 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1967 return NULL;
1968
Jens Axboea36e71f2009-04-15 12:15:11 +02001969 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001970}
1971
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001972/*
1973 * Determine whether we should enforce idle window for this queue.
1974 */
1975
1976static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1977{
1978 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001979 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001980
Vivek Goyalf04a6422009-12-03 12:59:40 -05001981 BUG_ON(!service_tree);
1982 BUG_ON(!service_tree->count);
1983
Vivek Goyalb6508c12010-08-23 12:23:33 +02001984 if (!cfqd->cfq_slice_idle)
1985 return false;
1986
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001987 /* We never do for idle class queues. */
1988 if (prio == IDLE_WORKLOAD)
1989 return false;
1990
1991 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001992 if (cfq_cfqq_idle_window(cfqq) &&
1993 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001994 return true;
1995
1996 /*
1997 * Otherwise, we do only if they are the last ones
1998 * in their service tree.
1999 */
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002000 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2001 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
Shaohua Lic1e44752010-11-08 15:01:02 +01002002 return true;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002003 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2004 service_tree->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01002005 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002006}
2007
Jens Axboe6d048f52007-04-25 12:44:27 +02002008static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002009{
Jens Axboe17926692007-01-19 11:59:30 +11002010 struct cfq_queue *cfqq = cfqd->active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +01002011 struct cfq_io_cq *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002012 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002013
Jens Axboea68bbdd2008-09-24 13:03:33 +02002014 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002015 * SSD device without seek penalty, disable idling. But only do so
2016 * for devices that support queuing, otherwise we still have a problem
2017 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02002018 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002019 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02002020 return;
2021
Jens Axboedd67d052006-06-21 09:36:18 +02002022 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02002023 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02002024
2025 /*
2026 * idle is disabled, either manually or by past process history
2027 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002028 if (!cfq_should_idle(cfqd, cfqq)) {
2029 /* no queue idling. Check for group idling */
2030 if (cfqd->cfq_group_idle)
2031 group_idle = cfqd->cfq_group_idle;
2032 else
2033 return;
2034 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002035
Jens Axboe22e2c502005-06-27 10:55:12 +02002036 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002037 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02002038 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002039 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02002040 return;
2041
2042 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002043 * task has exited, don't wait
2044 */
Jens Axboe206dc692006-03-28 13:03:44 +02002045 cic = cfqd->active_cic;
Tejun Heoc5869802011-12-14 00:33:41 +01002046 if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02002047 return;
2048
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002049 /*
2050 * If our average think time is larger than the remaining time
2051 * slice, then don't idle. This avoids overrunning the allotted
2052 * time slice.
2053 */
Shaohua Li383cd722011-07-12 14:24:35 +02002054 if (sample_valid(cic->ttime.ttime_samples) &&
2055 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
Joe Perchesfd16d262011-06-13 10:42:49 +02002056 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
Shaohua Li383cd722011-07-12 14:24:35 +02002057 cic->ttime.ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002058 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002059 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002060
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002061 /* There are other queues in the group, don't do group idle */
2062 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2063 return;
2064
Jens Axboe3b181522005-06-27 10:56:24 +02002065 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002066
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002067 if (group_idle)
2068 sl = cfqd->cfq_group_idle;
2069 else
2070 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02002071
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002072 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04002073 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002074 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2075 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076}
2077
Jens Axboe498d3aa2007-04-26 12:54:48 +02002078/*
2079 * Move request from internal lists to the request queue dispatch list.
2080 */
Jens Axboe165125e2007-07-24 09:28:11 +02002081static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082{
Jens Axboe3ed9a292007-04-23 08:33:33 +02002083 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002084 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002085
Jens Axboe7b679132008-05-30 12:23:07 +02002086 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2087
Jeff Moyer06d21882009-09-11 17:08:59 +02002088 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02002089 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002090 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002091 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02002092 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02002093
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002094 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02002095 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04002096 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02002097 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098}
2099
2100/*
2101 * return expired entry, or NULL to just start from scratch in rbtree
2102 */
Jens Axboefebffd62008-01-28 13:19:43 +01002103static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104{
Jens Axboe30996f42009-10-05 11:03:39 +02002105 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106
Jens Axboe3b181522005-06-27 10:56:24 +02002107 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002109
2110 cfq_mark_cfqq_fifo_expire(cfqq);
2111
Jens Axboe89850f72006-07-22 16:48:31 +02002112 if (list_empty(&cfqq->fifo))
2113 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114
Jens Axboe89850f72006-07-22 16:48:31 +02002115 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002116 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002117 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118
Jens Axboe30996f42009-10-05 11:03:39 +02002119 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002120 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121}
2122
Jens Axboe22e2c502005-06-27 10:55:12 +02002123static inline int
2124cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2125{
2126 const int base_rq = cfqd->cfq_slice_async_rq;
2127
2128 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2129
Namhyung Kimb9f8ce02011-05-24 10:23:21 +02002130 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002131}
2132
2133/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002134 * Must be called with the queue_lock held.
2135 */
2136static int cfqq_process_refs(struct cfq_queue *cfqq)
2137{
2138 int process_refs, io_refs;
2139
2140 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002141 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002142 BUG_ON(process_refs < 0);
2143 return process_refs;
2144}
2145
2146static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2147{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002148 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002149 struct cfq_queue *__cfqq;
2150
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002151 /*
2152 * If there are no process references on the new_cfqq, then it is
2153 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2154 * chain may have dropped their last reference (not just their
2155 * last process reference).
2156 */
2157 if (!cfqq_process_refs(new_cfqq))
2158 return;
2159
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002160 /* Avoid a circular list and skip interim queue merges */
2161 while ((__cfqq = new_cfqq->new_cfqq)) {
2162 if (__cfqq == cfqq)
2163 return;
2164 new_cfqq = __cfqq;
2165 }
2166
2167 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002168 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002169 /*
2170 * If the process for the cfqq has gone away, there is no
2171 * sense in merging the queues.
2172 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002173 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002174 return;
2175
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002176 /*
2177 * Merge in the direction of the lesser amount of work.
2178 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002179 if (new_process_refs >= process_refs) {
2180 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002181 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002182 } else {
2183 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002184 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002185 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002186}
2187
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002188static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002189 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002190{
2191 struct cfq_queue *queue;
2192 int i;
2193 bool key_valid = false;
2194 unsigned long lowest_key = 0;
2195 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2196
Vivek Goyal65b32a52009-12-16 17:52:59 -05002197 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2198 /* select the one with lowest rb_key */
2199 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002200 if (queue &&
2201 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2202 lowest_key = queue->rb_key;
2203 cur_best = i;
2204 key_valid = true;
2205 }
2206 }
2207
2208 return cur_best;
2209}
2210
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002211static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002212{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002213 unsigned slice;
2214 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002215 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002216 unsigned group_slice;
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002217 enum wl_prio_t original_prio = cfqd->serving_prio;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002218
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002219 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002220 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002221 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002222 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002223 cfqd->serving_prio = BE_WORKLOAD;
2224 else {
2225 cfqd->serving_prio = IDLE_WORKLOAD;
2226 cfqd->workload_expires = jiffies + 1;
2227 return;
2228 }
2229
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002230 if (original_prio != cfqd->serving_prio)
2231 goto new_workload;
2232
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002233 /*
2234 * For RT and BE, we have to choose also the type
2235 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2236 * expiration time
2237 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002238 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002239 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002240
2241 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002242 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002243 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002244 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002245 return;
2246
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002247new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002248 /* otherwise select new workload type */
2249 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002250 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2251 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002252 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002253
2254 /*
2255 * the workload slice is computed as a fraction of target latency
2256 * proportional to the number of queues in that workload, over
2257 * all the queues in the same priority class
2258 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002259 group_slice = cfq_group_slice(cfqd, cfqg);
2260
2261 slice = group_slice * count /
2262 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2263 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002264
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002265 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2266 unsigned int tmp;
2267
2268 /*
2269 * Async queues are currently system wide. Just taking
2270 * proportion of queues with-in same group will lead to higher
2271 * async ratio system wide as generally root group is going
2272 * to have higher weight. A more accurate thing would be to
2273 * calculate system wide asnc/sync ratio.
2274 */
Tao Ma5bf14c02012-04-01 14:33:39 -07002275 tmp = cfqd->cfq_target_latency *
2276 cfqg_busy_async_queues(cfqd, cfqg);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002277 tmp = tmp/cfqd->busy_queues;
2278 slice = min_t(unsigned, slice, tmp);
2279
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002280 /* async workload slice is scaled down according to
2281 * the sync/async slice ratio. */
2282 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002283 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002284 /* sync workload slice is at least 2 * cfq_slice_idle */
2285 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2286
2287 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002288 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002289 cfqd->workload_expires = jiffies + slice;
2290}
2291
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002292static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2293{
2294 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002295 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002296
2297 if (RB_EMPTY_ROOT(&st->rb))
2298 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002299 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002300 update_min_vdisktime(st);
2301 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002302}
2303
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002304static void cfq_choose_cfqg(struct cfq_data *cfqd)
2305{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002306 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2307
2308 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002309
2310 /* Restore the workload type data */
2311 if (cfqg->saved_workload_slice) {
2312 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2313 cfqd->serving_type = cfqg->saved_workload;
2314 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002315 } else
2316 cfqd->workload_expires = jiffies - 1;
2317
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002318 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002319}
2320
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002321/*
Jens Axboe498d3aa2007-04-26 12:54:48 +02002322 * Select a queue for service. If we have a current active queue,
2323 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002324 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002325static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002326{
Jens Axboea36e71f2009-04-15 12:15:11 +02002327 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002328
2329 cfqq = cfqd->active_queue;
2330 if (!cfqq)
2331 goto new_queue;
2332
Vivek Goyalf04a6422009-12-03 12:59:40 -05002333 if (!cfqd->rq_queued)
2334 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002335
2336 /*
2337 * We were waiting for group to get backlogged. Expire the queue
2338 */
2339 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2340 goto expire;
2341
Jens Axboe22e2c502005-06-27 10:55:12 +02002342 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002343 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002344 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002345 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2346 /*
2347 * If slice had not expired at the completion of last request
2348 * we might not have turned on wait_busy flag. Don't expire
2349 * the queue yet. Allow the group to get backlogged.
2350 *
2351 * The very fact that we have used the slice, that means we
2352 * have been idling all along on this queue and it should be
2353 * ok to wait for this request to complete.
2354 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002355 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2356 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2357 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002358 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002359 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002360 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002361 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002362
2363 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002364 * The active queue has requests and isn't expired, allow it to
2365 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002366 */
Jens Axboedd67d052006-06-21 09:36:18 +02002367 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002368 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002369
2370 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002371 * If another queue has a request waiting within our mean seek
2372 * distance, let it run. The expire code will check for close
2373 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002374 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002375 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002376 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002377 if (new_cfqq) {
2378 if (!cfqq->new_cfqq)
2379 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002380 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002381 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002382
2383 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002384 * No requests pending. If the active queue still has requests in
2385 * flight or is idling for a new request, allow either of these
2386 * conditions to happen (or time out) before selecting a new queue.
2387 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002388 if (timer_pending(&cfqd->idle_slice_timer)) {
2389 cfqq = NULL;
2390 goto keep_queue;
2391 }
2392
Shaohua Li8e1ac662010-11-08 15:01:04 +01002393 /*
2394 * This is a deep seek queue, but the device is much faster than
2395 * the queue can deliver, don't idle
2396 **/
2397 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2398 (cfq_cfqq_slice_new(cfqq) ||
2399 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2400 cfq_clear_cfqq_deep(cfqq);
2401 cfq_clear_cfqq_idle_window(cfqq);
2402 }
2403
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002404 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2405 cfqq = NULL;
2406 goto keep_queue;
2407 }
2408
2409 /*
2410 * If group idle is enabled and there are requests dispatched from
2411 * this group, wait for requests to complete.
2412 */
2413check_group_idle:
Shaohua Li7700fc42011-07-12 14:24:56 +02002414 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2415 cfqq->cfqg->dispatched &&
2416 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002417 cfqq = NULL;
2418 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002419 }
2420
Jens Axboe3b181522005-06-27 10:56:24 +02002421expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002422 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002423new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002424 /*
2425 * Current queue expired. Check if we have to switch to a new
2426 * service tree
2427 */
2428 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002429 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002430
Jens Axboea36e71f2009-04-15 12:15:11 +02002431 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002432keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002433 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002434}
2435
Jens Axboefebffd62008-01-28 13:19:43 +01002436static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002437{
2438 int dispatched = 0;
2439
2440 while (cfqq->next_rq) {
2441 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2442 dispatched++;
2443 }
2444
2445 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002446
2447 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002448 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002449 return dispatched;
2450}
2451
Jens Axboe498d3aa2007-04-26 12:54:48 +02002452/*
2453 * Drain our current requests. Used for barriers and when switching
2454 * io schedulers on-the-fly.
2455 */
Jens Axboed9e76202007-04-20 14:27:50 +02002456static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002457{
Jens Axboe08717142008-01-28 11:38:15 +01002458 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002459 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002460
Divyesh Shah3440c492010-04-09 09:29:57 +02002461 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002462 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002463 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2464 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002465 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002466 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002467
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002468 BUG_ON(cfqd->busy_queues);
2469
Jeff Moyer69237152009-06-12 15:29:30 +02002470 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002471 return dispatched;
2472}
2473
Shaohua Liabc3c742010-03-01 09:20:54 +01002474static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2475 struct cfq_queue *cfqq)
2476{
2477 /* the queue hasn't finished any request, can't estimate */
2478 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01002479 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002480 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2481 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01002482 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002483
Shaohua Lic1e44752010-11-08 15:01:02 +01002484 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01002485}
2486
Jens Axboe0b182d62009-10-06 20:49:37 +02002487static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002488{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002489 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490
Jens Axboe2f5cb732009-04-07 08:51:19 +02002491 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002492 * Drain async requests before we start sync IO
2493 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002494 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002495 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002496
2497 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002498 * If this is an async queue and we have sync IO in flight, let it wait
2499 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002500 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002501 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002502
Shaohua Liabc3c742010-03-01 09:20:54 +01002503 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002504 if (cfq_class_idle(cfqq))
2505 max_dispatch = 1;
2506
2507 /*
2508 * Does this cfqq already have too much IO in flight?
2509 */
2510 if (cfqq->dispatched >= max_dispatch) {
Shaohua Lief8a41d2011-03-07 09:26:29 +01002511 bool promote_sync = false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002512 /*
2513 * idle queue must always only have a single IO in flight
2514 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002515 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002516 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002517
Jens Axboe2f5cb732009-04-07 08:51:19 +02002518 /*
Li, Shaohuac4ade942011-03-23 08:30:34 +01002519 * If there is only one sync queue
2520 * we can ignore async queue here and give the sync
Shaohua Lief8a41d2011-03-07 09:26:29 +01002521 * queue no dispatch limit. The reason is a sync queue can
2522 * preempt async queue, limiting the sync queue doesn't make
2523 * sense. This is useful for aiostress test.
2524 */
Li, Shaohuac4ade942011-03-23 08:30:34 +01002525 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2526 promote_sync = true;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002527
2528 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002529 * We have other queues, don't allow more IO from this one
2530 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002531 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2532 !promote_sync)
Jens Axboe0b182d62009-10-06 20:49:37 +02002533 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002534
Jens Axboe2f5cb732009-04-07 08:51:19 +02002535 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002536 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002537 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002538 if (cfqd->busy_queues == 1 || promote_sync)
Shaohua Liabc3c742010-03-01 09:20:54 +01002539 max_dispatch = -1;
2540 else
2541 /*
2542 * Normally we start throttling cfqq when cfq_quantum/2
2543 * requests have been dispatched. But we can drive
2544 * deeper queue depths at the beginning of slice
2545 * subjected to upper limit of cfq_quantum.
2546 * */
2547 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002548 }
2549
2550 /*
2551 * Async queues must wait a bit before being allowed dispatch.
2552 * We also ramp up the dispatch depth gradually for async IO,
2553 * based on the last sync IO we serviced
2554 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002555 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002556 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002557 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002558
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002559 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002560 if (!depth && !cfqq->dispatched)
2561 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002562 if (depth < max_dispatch)
2563 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 }
2565
Jens Axboe0b182d62009-10-06 20:49:37 +02002566 /*
2567 * If we're below the current max, allow a dispatch
2568 */
2569 return cfqq->dispatched < max_dispatch;
2570}
2571
2572/*
2573 * Dispatch a request from cfqq, moving them to the request queue
2574 * dispatch list.
2575 */
2576static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2577{
2578 struct request *rq;
2579
2580 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2581
2582 if (!cfq_may_dispatch(cfqd, cfqq))
2583 return false;
2584
2585 /*
2586 * follow expired path, else get first next available
2587 */
2588 rq = cfq_check_fifo(cfqq);
2589 if (!rq)
2590 rq = cfqq->next_rq;
2591
2592 /*
2593 * insert request into driver dispatch list
2594 */
2595 cfq_dispatch_insert(cfqd->queue, rq);
2596
2597 if (!cfqd->active_cic) {
Tejun Heoc5869802011-12-14 00:33:41 +01002598 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe0b182d62009-10-06 20:49:37 +02002599
Tejun Heoc5869802011-12-14 00:33:41 +01002600 atomic_long_inc(&cic->icq.ioc->refcount);
Jens Axboe0b182d62009-10-06 20:49:37 +02002601 cfqd->active_cic = cic;
2602 }
2603
2604 return true;
2605}
2606
2607/*
2608 * Find the cfqq that we need to service and move a request from that to the
2609 * dispatch list
2610 */
2611static int cfq_dispatch_requests(struct request_queue *q, int force)
2612{
2613 struct cfq_data *cfqd = q->elevator->elevator_data;
2614 struct cfq_queue *cfqq;
2615
2616 if (!cfqd->busy_queues)
2617 return 0;
2618
2619 if (unlikely(force))
2620 return cfq_forced_dispatch(cfqd);
2621
2622 cfqq = cfq_select_queue(cfqd);
2623 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002624 return 0;
2625
Jens Axboe2f5cb732009-04-07 08:51:19 +02002626 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002627 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002628 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002629 if (!cfq_dispatch_request(cfqd, cfqq))
2630 return 0;
2631
Jens Axboe2f5cb732009-04-07 08:51:19 +02002632 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002633 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002634
2635 /*
2636 * expire an async queue immediately if it has used up its slice. idle
2637 * queue always expire after 1 dispatch round.
2638 */
2639 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2640 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2641 cfq_class_idle(cfqq))) {
2642 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002643 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002644 }
2645
Shan Weib217a902009-09-01 10:06:42 +02002646 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002647 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648}
2649
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650/*
Jens Axboe5e705372006-07-13 12:39:25 +02002651 * task holds one reference to the queue, dropped when task exits. each rq
2652 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002654 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 * queue lock must be held here.
2656 */
2657static void cfq_put_queue(struct cfq_queue *cfqq)
2658{
Jens Axboe22e2c502005-06-27 10:55:12 +02002659 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05002660 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002661
Shaohua Li30d7b942011-01-07 08:46:59 +01002662 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663
Shaohua Li30d7b942011-01-07 08:46:59 +01002664 cfqq->ref--;
2665 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 return;
2667
Jens Axboe7b679132008-05-30 12:23:07 +02002668 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002670 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002671 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002673 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002674 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002675 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002676 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002677
Vivek Goyalf04a6422009-12-03 12:59:40 -05002678 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002680 cfq_put_cfqg(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681}
2682
Shaohua Lid02a2c02010-05-25 10:16:53 +02002683static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002684{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002685 struct cfq_queue *__cfqq, *next;
2686
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002687 /*
2688 * If this queue was scheduled to merge with another queue, be
2689 * sure to drop the reference taken on that queue (and others in
2690 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2691 */
2692 __cfqq = cfqq->new_cfqq;
2693 while (__cfqq) {
2694 if (__cfqq == cfqq) {
2695 WARN(1, "cfqq->new_cfqq loop detected\n");
2696 break;
2697 }
2698 next = __cfqq->new_cfqq;
2699 cfq_put_queue(__cfqq);
2700 __cfqq = next;
2701 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002702}
2703
2704static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2705{
2706 if (unlikely(cfqq == cfqd->active_queue)) {
2707 __cfq_slice_expired(cfqd, cfqq, 0);
2708 cfq_schedule_dispatch(cfqd);
2709 }
2710
2711 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002712
Jens Axboe89850f72006-07-22 16:48:31 +02002713 cfq_put_queue(cfqq);
2714}
2715
Tejun Heo9b84cac2011-12-14 00:33:42 +01002716static void cfq_init_icq(struct io_cq *icq)
2717{
2718 struct cfq_io_cq *cic = icq_to_cic(icq);
2719
2720 cic->ttime.last_end_request = jiffies;
2721}
2722
Tejun Heoc5869802011-12-14 00:33:41 +01002723static void cfq_exit_icq(struct io_cq *icq)
Jens Axboe89850f72006-07-22 16:48:31 +02002724{
Tejun Heoc5869802011-12-14 00:33:41 +01002725 struct cfq_io_cq *cic = icq_to_cic(icq);
Tejun Heo283287a2011-12-14 00:33:38 +01002726 struct cfq_data *cfqd = cic_to_cfqd(cic);
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002727
Jens Axboeff6657c2009-04-08 10:58:57 +02002728 if (cic->cfqq[BLK_RW_ASYNC]) {
2729 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2730 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002731 }
2732
Jens Axboeff6657c2009-04-08 10:58:57 +02002733 if (cic->cfqq[BLK_RW_SYNC]) {
2734 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2735 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002736 }
Jens Axboe89850f72006-07-22 16:48:31 +02002737}
2738
Jens Axboefd0928d2008-01-24 08:52:45 +01002739static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002740{
2741 struct task_struct *tsk = current;
2742 int ioprio_class;
2743
Jens Axboe3b181522005-06-27 10:56:24 +02002744 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002745 return;
2746
Jens Axboefd0928d2008-01-24 08:52:45 +01002747 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002748 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002749 default:
2750 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2751 case IOPRIO_CLASS_NONE:
2752 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002753 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002754 */
2755 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002756 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002757 break;
2758 case IOPRIO_CLASS_RT:
2759 cfqq->ioprio = task_ioprio(ioc);
2760 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2761 break;
2762 case IOPRIO_CLASS_BE:
2763 cfqq->ioprio = task_ioprio(ioc);
2764 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2765 break;
2766 case IOPRIO_CLASS_IDLE:
2767 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2768 cfqq->ioprio = 7;
2769 cfq_clear_cfqq_idle_window(cfqq);
2770 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002771 }
2772
2773 /*
2774 * keep track of original prio settings in case we have to temporarily
2775 * elevate the priority of this queue
2776 */
2777 cfqq->org_ioprio = cfqq->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02002778 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002779}
2780
Tejun Heoc5869802011-12-14 00:33:41 +01002781static void changed_ioprio(struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002782{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002783 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002784 struct cfq_queue *cfqq;
Jens Axboe35e60772006-06-14 09:10:45 +02002785
Jens Axboecaaa5f92006-06-16 11:23:00 +02002786 if (unlikely(!cfqd))
2787 return;
2788
Jens Axboeff6657c2009-04-08 10:58:57 +02002789 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002790 if (cfqq) {
2791 struct cfq_queue *new_cfqq;
Tejun Heoc5869802011-12-14 00:33:41 +01002792 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
Jens Axboeff6657c2009-04-08 10:58:57 +02002793 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002794 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002795 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002796 cfq_put_queue(cfqq);
2797 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002798 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002799
Jens Axboeff6657c2009-04-08 10:58:57 +02002800 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002801 if (cfqq)
2802 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002803}
2804
Jens Axboed5036d72009-06-26 10:44:34 +02002805static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002806 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002807{
2808 RB_CLEAR_NODE(&cfqq->rb_node);
2809 RB_CLEAR_NODE(&cfqq->p_node);
2810 INIT_LIST_HEAD(&cfqq->fifo);
2811
Shaohua Li30d7b942011-01-07 08:46:59 +01002812 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02002813 cfqq->cfqd = cfqd;
2814
2815 cfq_mark_cfqq_prio_changed(cfqq);
2816
2817 if (is_sync) {
2818 if (!cfq_class_idle(cfqq))
2819 cfq_mark_cfqq_idle_window(cfqq);
2820 cfq_mark_cfqq_sync(cfqq);
2821 }
2822 cfqq->pid = pid;
2823}
2824
Vivek Goyal24610332009-12-03 12:59:51 -05002825#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heoc5869802011-12-14 00:33:41 +01002826static void changed_cgroup(struct cfq_io_cq *cic)
Vivek Goyal24610332009-12-03 12:59:51 -05002827{
2828 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002829 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal24610332009-12-03 12:59:51 -05002830 struct request_queue *q;
2831
2832 if (unlikely(!cfqd))
2833 return;
2834
2835 q = cfqd->queue;
2836
Vivek Goyal24610332009-12-03 12:59:51 -05002837 if (sync_cfqq) {
2838 /*
2839 * Drop reference to sync queue. A new sync queue will be
2840 * assigned in new group upon arrival of a fresh request.
2841 */
2842 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2843 cic_set_cfqq(cic, NULL, 1);
2844 cfq_put_queue(sync_cfqq);
2845 }
Vivek Goyal24610332009-12-03 12:59:51 -05002846}
Vivek Goyal24610332009-12-03 12:59:51 -05002847#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2848
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002850cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002851 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 struct cfq_queue *cfqq, *new_cfqq = NULL;
Tejun Heoc5869802011-12-14 00:33:41 +01002854 struct cfq_io_cq *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002855 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856
2857retry:
Vivek Goyal3e59cf92011-05-19 15:38:21 -04002858 cfqg = cfq_get_cfqg(cfqd);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002859 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002860 /* cic always exists here */
2861 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862
Jens Axboe6118b702009-06-30 09:34:12 +02002863 /*
2864 * Always try a new alloc if we fell back to the OOM cfqq
2865 * originally, since it should just be a temporary situation.
2866 */
2867 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2868 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869 if (new_cfqq) {
2870 cfqq = new_cfqq;
2871 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002872 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002874 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002875 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002876 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002878 if (new_cfqq)
2879 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002880 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002881 cfqq = kmem_cache_alloc_node(cfq_pool,
2882 gfp_mask | __GFP_ZERO,
2883 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002884 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885
Jens Axboe6118b702009-06-30 09:34:12 +02002886 if (cfqq) {
2887 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2888 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002889 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002890 cfq_log_cfqq(cfqd, cfqq, "alloced");
2891 } else
2892 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893 }
2894
2895 if (new_cfqq)
2896 kmem_cache_free(cfq_pool, new_cfqq);
2897
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 return cfqq;
2899}
2900
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002901static struct cfq_queue **
2902cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2903{
Jens Axboefe094d92008-01-31 13:08:54 +01002904 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002905 case IOPRIO_CLASS_RT:
2906 return &cfqd->async_cfqq[0][ioprio];
2907 case IOPRIO_CLASS_BE:
2908 return &cfqd->async_cfqq[1][ioprio];
2909 case IOPRIO_CLASS_IDLE:
2910 return &cfqd->async_idle_cfqq;
2911 default:
2912 BUG();
2913 }
2914}
2915
Jens Axboe15c31be2007-07-10 13:43:25 +02002916static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002917cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002918 gfp_t gfp_mask)
2919{
Jens Axboefd0928d2008-01-24 08:52:45 +01002920 const int ioprio = task_ioprio(ioc);
2921 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002922 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002923 struct cfq_queue *cfqq = NULL;
2924
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002925 if (!is_sync) {
2926 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2927 cfqq = *async_cfqq;
2928 }
2929
Jens Axboe6118b702009-06-30 09:34:12 +02002930 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002931 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002932
2933 /*
2934 * pin the queue now that it's allocated, scheduler exit will prune it
2935 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002936 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01002937 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002938 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002939 }
2940
Shaohua Li30d7b942011-01-07 08:46:59 +01002941 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02002942 return cfqq;
2943}
2944
Jens Axboe22e2c502005-06-27 10:55:12 +02002945static void
Shaohua Li383cd722011-07-12 14:24:35 +02002946__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02002947{
Shaohua Li383cd722011-07-12 14:24:35 +02002948 unsigned long elapsed = jiffies - ttime->last_end_request;
2949 elapsed = min(elapsed, 2UL * slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02002950
Shaohua Li383cd722011-07-12 14:24:35 +02002951 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
2952 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
2953 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
2954}
2955
2956static void
2957cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01002958 struct cfq_io_cq *cic)
Shaohua Li383cd722011-07-12 14:24:35 +02002959{
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002960 if (cfq_cfqq_sync(cfqq)) {
Shaohua Li383cd722011-07-12 14:24:35 +02002961 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002962 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
2963 cfqd->cfq_slice_idle);
2964 }
Shaohua Li7700fc42011-07-12 14:24:56 +02002965#ifdef CONFIG_CFQ_GROUP_IOSCHED
2966 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
2967#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02002968}
2969
Jens Axboe206dc692006-03-28 13:03:44 +02002970static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002971cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02002972 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02002973{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002974 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01002975 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002976 if (cfqq->last_request_pos) {
2977 if (cfqq->last_request_pos < blk_rq_pos(rq))
2978 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2979 else
2980 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2981 }
Jens Axboe206dc692006-03-28 13:03:44 +02002982
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002983 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01002984 if (blk_queue_nonrot(cfqd->queue))
2985 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
2986 else
2987 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02002988}
Jens Axboe22e2c502005-06-27 10:55:12 +02002989
2990/*
2991 * Disable idle window if the process thinks too long or seeks so much that
2992 * it doesn't matter
2993 */
2994static void
2995cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01002996 struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002997{
Jens Axboe7b679132008-05-30 12:23:07 +02002998 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02002999
Jens Axboe08717142008-01-28 11:38:15 +01003000 /*
3001 * Don't idle for async or idle io prio class
3002 */
3003 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003004 return;
3005
Jens Axboec265a7f2008-06-26 13:49:33 +02003006 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003007
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003008 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3009 cfq_mark_cfqq_deep(cfqq);
3010
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003011 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3012 enable_idle = 0;
Tejun Heoc5869802011-12-14 00:33:41 +01003013 else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
3014 !cfqd->cfq_slice_idle ||
3015 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003016 enable_idle = 0;
Shaohua Li383cd722011-07-12 14:24:35 +02003017 else if (sample_valid(cic->ttime.ttime_samples)) {
3018 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003019 enable_idle = 0;
3020 else
3021 enable_idle = 1;
3022 }
3023
Jens Axboe7b679132008-05-30 12:23:07 +02003024 if (old_idle != enable_idle) {
3025 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3026 if (enable_idle)
3027 cfq_mark_cfqq_idle_window(cfqq);
3028 else
3029 cfq_clear_cfqq_idle_window(cfqq);
3030 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003031}
3032
Jens Axboe22e2c502005-06-27 10:55:12 +02003033/*
3034 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3035 * no or if we aren't sure, a 1 will cause a preempt.
3036 */
Jens Axboea6151c32009-10-07 20:02:57 +02003037static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003038cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003039 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003040{
Jens Axboe6d048f52007-04-25 12:44:27 +02003041 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003042
Jens Axboe6d048f52007-04-25 12:44:27 +02003043 cfqq = cfqd->active_queue;
3044 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003045 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003046
Jens Axboe6d048f52007-04-25 12:44:27 +02003047 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003048 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003049
3050 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003051 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003052
Jens Axboe22e2c502005-06-27 10:55:12 +02003053 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003054 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3055 */
3056 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3057 return false;
3058
3059 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003060 * if the new request is sync, but the currently running queue is
3061 * not, let the sync request have priority.
3062 */
Jens Axboe5e705372006-07-13 12:39:25 +02003063 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003064 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003065
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003066 if (new_cfqq->cfqg != cfqq->cfqg)
3067 return false;
3068
3069 if (cfq_slice_used(cfqq))
3070 return true;
3071
3072 /* Allow preemption only if we are idling on sync-noidle tree */
3073 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3074 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3075 new_cfqq->service_tree->count == 2 &&
3076 RB_EMPTY_ROOT(&cfqq->sort_list))
3077 return true;
3078
Jens Axboe374f84a2006-07-23 01:42:19 +02003079 /*
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003080 * So both queues are sync. Let the new request get disk time if
3081 * it's a metadata request and the current queue is doing regular IO.
3082 */
Christoph Hellwig65299a32011-08-23 14:50:29 +02003083 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003084 return true;
3085
3086 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003087 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3088 */
3089 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003090 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003091
Shaohua Lid2d59e12010-11-08 15:01:03 +01003092 /* An idle queue should not be idle now for some reason */
3093 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3094 return true;
3095
Jens Axboe1e3335d2007-02-14 19:59:49 +01003096 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003097 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003098
3099 /*
3100 * if this request is as-good as one we would expect from the
3101 * current cfqq, let it preempt
3102 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003103 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003104 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003105
Jens Axboea6151c32009-10-07 20:02:57 +02003106 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003107}
3108
3109/*
3110 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3111 * let it have half of its nominal slice.
3112 */
3113static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3114{
Shaohua Lidf0793a2012-01-19 09:20:09 +01003115 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3116
Jens Axboe7b679132008-05-30 12:23:07 +02003117 cfq_log_cfqq(cfqd, cfqq, "preempt");
Shaohua Lidf0793a2012-01-19 09:20:09 +01003118 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003119
Jens Axboebf572252006-07-19 20:29:12 +02003120 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003121 * workload type is changed, don't save slice, otherwise preempt
3122 * doesn't happen
3123 */
Shaohua Lidf0793a2012-01-19 09:20:09 +01003124 if (old_type != cfqq_type(cfqq))
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003125 cfqq->cfqg->saved_workload_slice = 0;
3126
3127 /*
Jens Axboebf572252006-07-19 20:29:12 +02003128 * Put the new queue at the front of the of the current list,
3129 * so we know that it will be selected next.
3130 */
3131 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003132
3133 cfq_service_tree_add(cfqd, cfqq, 1);
Justin TerAvesteda5e0c2011-03-22 21:26:49 +01003134
Justin TerAvest62a37f62011-03-23 08:25:44 +01003135 cfqq->slice_end = 0;
3136 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003137}
3138
3139/*
Jens Axboe5e705372006-07-13 12:39:25 +02003140 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003141 * something we should do about it
3142 */
3143static void
Jens Axboe5e705372006-07-13 12:39:25 +02003144cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3145 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003146{
Tejun Heoc5869802011-12-14 00:33:41 +01003147 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003148
Aaron Carroll45333d52008-08-26 15:52:36 +02003149 cfqd->rq_queued++;
Christoph Hellwig65299a32011-08-23 14:50:29 +02003150 if (rq->cmd_flags & REQ_PRIO)
3151 cfqq->prio_pending++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003152
Shaohua Li383cd722011-07-12 14:24:35 +02003153 cfq_update_io_thinktime(cfqd, cfqq, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003154 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003155 cfq_update_idle_window(cfqd, cfqq, cic);
3156
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003157 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003158
3159 if (cfqq == cfqd->active_queue) {
3160 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003161 * Remember that we saw a request from this process, but
3162 * don't start queuing just yet. Otherwise we risk seeing lots
3163 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003164 * and merging. If the request is already larger than a single
3165 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003166 * merging is already done. Ditto for a busy system that
3167 * has other work pending, don't risk delaying until the
3168 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003169 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003170 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003171 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3172 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003173 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003174 cfq_clear_cfqq_wait_request(cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003175 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003176 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003177 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003178 &cfqq->cfqg->blkg);
Vivek Goyalbf791932009-12-03 12:59:37 -05003179 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003180 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003181 }
Jens Axboe5e705372006-07-13 12:39:25 +02003182 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003183 /*
3184 * not the active queue - expire current slice if it is
3185 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003186 * has some old slice time left and is of higher priority or
3187 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003188 */
3189 cfq_preempt_queue(cfqd, cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003190 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003191 }
3192}
3193
Jens Axboe165125e2007-07-24 09:28:11 +02003194static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003195{
Jens Axboeb4878f22005-10-20 16:42:29 +02003196 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003197 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003198
Jens Axboe7b679132008-05-30 12:23:07 +02003199 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Tejun Heoc5869802011-12-14 00:33:41 +01003200 cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201
Jens Axboe30996f42009-10-05 11:03:39 +02003202 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003203 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003204 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003205 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003206 &cfqd->serving_group->blkg, rq_data_dir(rq),
3207 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003208 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209}
3210
Aaron Carroll45333d52008-08-26 15:52:36 +02003211/*
3212 * Update hw_tag based on peak queue depth over 50 samples under
3213 * sufficient load.
3214 */
3215static void cfq_update_hw_tag(struct cfq_data *cfqd)
3216{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003217 struct cfq_queue *cfqq = cfqd->active_queue;
3218
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003219 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3220 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003221
3222 if (cfqd->hw_tag == 1)
3223 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003224
3225 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003226 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003227 return;
3228
Shaohua Li1a1238a2009-10-27 08:46:23 +01003229 /*
3230 * If active queue hasn't enough requests and can idle, cfq might not
3231 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3232 * case
3233 */
3234 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3235 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003236 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003237 return;
3238
Aaron Carroll45333d52008-08-26 15:52:36 +02003239 if (cfqd->hw_tag_samples++ < 50)
3240 return;
3241
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003242 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003243 cfqd->hw_tag = 1;
3244 else
3245 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003246}
3247
Vivek Goyal7667aa02009-12-08 17:52:58 -05003248static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3249{
Tejun Heoc5869802011-12-14 00:33:41 +01003250 struct cfq_io_cq *cic = cfqd->active_cic;
Vivek Goyal7667aa02009-12-08 17:52:58 -05003251
Justin TerAvest02a8f012011-02-09 14:20:03 +01003252 /* If the queue already has requests, don't wait */
3253 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3254 return false;
3255
Vivek Goyal7667aa02009-12-08 17:52:58 -05003256 /* If there are other queues in the group, don't wait */
3257 if (cfqq->cfqg->nr_cfqq > 1)
3258 return false;
3259
Shaohua Li7700fc42011-07-12 14:24:56 +02003260 /* the only queue in the group, but think time is big */
3261 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3262 return false;
3263
Vivek Goyal7667aa02009-12-08 17:52:58 -05003264 if (cfq_slice_used(cfqq))
3265 return true;
3266
3267 /* if slice left is less than think time, wait busy */
Shaohua Li383cd722011-07-12 14:24:35 +02003268 if (cic && sample_valid(cic->ttime.ttime_samples)
3269 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
Vivek Goyal7667aa02009-12-08 17:52:58 -05003270 return true;
3271
3272 /*
3273 * If think times is less than a jiffy than ttime_mean=0 and above
3274 * will not be true. It might happen that slice has not expired yet
3275 * but will expire soon (4-5 ns) during select_queue(). To cover the
3276 * case where think time is less than a jiffy, mark the queue wait
3277 * busy if only 1 jiffy is left in the slice.
3278 */
3279 if (cfqq->slice_end - jiffies == 1)
3280 return true;
3281
3282 return false;
3283}
3284
Jens Axboe165125e2007-07-24 09:28:11 +02003285static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286{
Jens Axboe5e705372006-07-13 12:39:25 +02003287 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003288 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003289 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003290 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291
Jens Axboeb4878f22005-10-20 16:42:29 +02003292 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003293 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3294 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295
Aaron Carroll45333d52008-08-26 15:52:36 +02003296 cfq_update_hw_tag(cfqd);
3297
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003298 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003299 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003300 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003301 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003302 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003303 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3304 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3305 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003307 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003308
Vivek Goyal365722b2009-10-03 15:21:27 +02003309 if (sync) {
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003310 struct cfq_rb_root *service_tree;
3311
Shaohua Li383cd722011-07-12 14:24:35 +02003312 RQ_CIC(rq)->ttime.last_end_request = now;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003313
3314 if (cfq_cfqq_on_rr(cfqq))
3315 service_tree = cfqq->service_tree;
3316 else
3317 service_tree = service_tree_for(cfqq->cfqg,
3318 cfqq_prio(cfqq), cfqq_type(cfqq));
3319 service_tree->ttime.last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003320 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3321 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003322 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003323
Shaohua Li7700fc42011-07-12 14:24:56 +02003324#ifdef CONFIG_CFQ_GROUP_IOSCHED
3325 cfqq->cfqg->ttime.last_end_request = now;
3326#endif
3327
Jens Axboecaaa5f92006-06-16 11:23:00 +02003328 /*
3329 * If this is the active queue, check if it needs to be expired,
3330 * or if we want to idle in case it has no pending requests.
3331 */
3332 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003333 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3334
Jens Axboe44f7c162007-01-19 11:51:58 +11003335 if (cfq_cfqq_slice_new(cfqq)) {
3336 cfq_set_prio_slice(cfqd, cfqq);
3337 cfq_clear_cfqq_slice_new(cfqq);
3338 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003339
3340 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003341 * Should we wait for next request to come in before we expire
3342 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003343 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003344 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003345 unsigned long extend_sl = cfqd->cfq_slice_idle;
3346 if (!cfqd->cfq_slice_idle)
3347 extend_sl = cfqd->cfq_group_idle;
3348 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003349 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003350 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003351 }
3352
Jens Axboea36e71f2009-04-15 12:15:11 +02003353 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003354 * Idling is not enabled on:
3355 * - expired queues
3356 * - idle-priority queues
3357 * - async queues
3358 * - queues with still some requests queued
3359 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003360 */
Jens Axboe08717142008-01-28 11:38:15 +01003361 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003362 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003363 else if (sync && cfqq_empty &&
3364 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003365 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003366 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003367 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003368
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003369 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003370 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371}
3372
Jens Axboe89850f72006-07-22 16:48:31 +02003373static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003374{
Jens Axboe1b379d82009-08-11 08:26:11 +02003375 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003376 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003377 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003378 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003379
3380 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003381}
3382
Jens Axboe165125e2007-07-24 09:28:11 +02003383static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003384{
3385 struct cfq_data *cfqd = q->elevator->elevator_data;
3386 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01003387 struct cfq_io_cq *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003388 struct cfq_queue *cfqq;
3389
3390 /*
3391 * don't force setup of a queue from here, as a call to may_queue
3392 * does not necessarily imply that a request actually will be queued.
3393 * so just lookup a possibly existing queue, or return 'may queue'
3394 * if that fails
3395 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003396 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003397 if (!cic)
3398 return ELV_MQUEUE_MAY;
3399
Jens Axboeb0b78f82009-04-08 10:56:08 +02003400 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003401 if (cfqq) {
Tejun Heoc5869802011-12-14 00:33:41 +01003402 cfq_init_prio_data(cfqq, cic->icq.ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003403
Jens Axboe89850f72006-07-22 16:48:31 +02003404 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003405 }
3406
3407 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408}
3409
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410/*
3411 * queue lock held here
3412 */
Jens Axboebb37b942006-12-01 10:42:33 +01003413static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414{
Jens Axboe5e705372006-07-13 12:39:25 +02003415 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416
Jens Axboe5e705372006-07-13 12:39:25 +02003417 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003418 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419
Jens Axboe22e2c502005-06-27 10:55:12 +02003420 BUG_ON(!cfqq->allocated[rw]);
3421 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003423 /* Put down rq reference on cfqg */
3424 cfq_put_cfqg(RQ_CFQG(rq));
Tejun Heoa612fdd2011-12-14 00:33:41 +01003425 rq->elv.priv[0] = NULL;
3426 rq->elv.priv[1] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003427
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428 cfq_put_queue(cfqq);
3429 }
3430}
3431
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003432static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01003433cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003434 struct cfq_queue *cfqq)
3435{
3436 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3437 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003438 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003439 cfq_put_queue(cfqq);
3440 return cic_to_cfqq(cic, 1);
3441}
3442
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003443/*
3444 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3445 * was the last process referring to said cfqq.
3446 */
3447static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01003448split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003449{
3450 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003451 cfqq->pid = current->pid;
3452 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003453 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003454 return cfqq;
3455 }
3456
3457 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003458
3459 cfq_put_cooperator(cfqq);
3460
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003461 cfq_put_queue(cfqq);
3462 return NULL;
3463}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003465 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003467static int
Jens Axboe165125e2007-07-24 09:28:11 +02003468cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469{
3470 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heof1f8cc92011-12-14 00:33:42 +01003471 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003473 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003474 struct cfq_queue *cfqq;
Tejun Heod705ae62012-02-15 09:45:49 +01003475 unsigned int changed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476
3477 might_sleep_if(gfp_mask & __GFP_WAIT);
3478
Tejun Heo216284c2011-12-14 00:33:38 +01003479 spin_lock_irq(q->queue_lock);
Tejun Heof1f8cc92011-12-14 00:33:42 +01003480
3481 /* handle changed notifications */
Tejun Heod705ae62012-02-15 09:45:49 +01003482 changed = icq_get_changed(&cic->icq);
3483 if (unlikely(changed & ICQ_IOPRIO_CHANGED))
3484 changed_ioprio(cic);
Tejun Heof1f8cc92011-12-14 00:33:42 +01003485#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heod705ae62012-02-15 09:45:49 +01003486 if (unlikely(changed & ICQ_CGROUP_CHANGED))
3487 changed_cgroup(cic);
Tejun Heof1f8cc92011-12-14 00:33:42 +01003488#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003490new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003491 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003492 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Tejun Heoc5869802011-12-14 00:33:41 +01003493 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003494 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003495 } else {
3496 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003497 * If the queue was seeky for too long, break it apart.
3498 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003499 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003500 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3501 cfqq = split_cfqq(cic, cfqq);
3502 if (!cfqq)
3503 goto new_queue;
3504 }
3505
3506 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003507 * Check to see if this queue is scheduled to merge with
3508 * another, closely cooperating queue. The merging of
3509 * queues happens here as it must be done in process context.
3510 * The reference on new_cfqq was taken in merge_cfqqs.
3511 */
3512 if (cfqq->new_cfqq)
3513 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003514 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515
3516 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02003517
Jens Axboe6fae9c22011-03-01 15:04:39 -05003518 cfqq->ref++;
Tejun Heoa612fdd2011-12-14 00:33:41 +01003519 rq->elv.priv[0] = cfqq;
3520 rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
Tejun Heo216284c2011-12-14 00:33:38 +01003521 spin_unlock_irq(q->queue_lock);
Jens Axboe5e705372006-07-13 12:39:25 +02003522 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523}
3524
David Howells65f27f32006-11-22 14:55:48 +00003525static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003526{
David Howells65f27f32006-11-22 14:55:48 +00003527 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003528 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003529 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003530
Jens Axboe40bb54d2009-04-15 12:11:10 +02003531 spin_lock_irq(q->queue_lock);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003532 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003533 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003534}
3535
3536/*
3537 * Timer running if the active_queue is currently idling inside its time slice
3538 */
3539static void cfq_idle_slice_timer(unsigned long data)
3540{
3541 struct cfq_data *cfqd = (struct cfq_data *) data;
3542 struct cfq_queue *cfqq;
3543 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003544 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003545
Jens Axboe7b679132008-05-30 12:23:07 +02003546 cfq_log(cfqd, "idle timer fired");
3547
Jens Axboe22e2c502005-06-27 10:55:12 +02003548 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3549
Jens Axboefe094d92008-01-31 13:08:54 +01003550 cfqq = cfqd->active_queue;
3551 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003552 timed_out = 0;
3553
Jens Axboe22e2c502005-06-27 10:55:12 +02003554 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003555 * We saw a request before the queue expired, let it through
3556 */
3557 if (cfq_cfqq_must_dispatch(cfqq))
3558 goto out_kick;
3559
3560 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003561 * expired
3562 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003563 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003564 goto expire;
3565
3566 /*
3567 * only expire and reinvoke request handler, if there are
3568 * other queues with pending requests
3569 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003570 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003571 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003572
3573 /*
3574 * not expired and it has a request pending, let it dispatch
3575 */
Jens Axboe75e50982009-04-07 08:56:14 +02003576 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003577 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003578
3579 /*
3580 * Queue depth flag is reset only when the idle didn't succeed
3581 */
3582 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003583 }
3584expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003585 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003586out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003587 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003588out_cont:
3589 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3590}
3591
Jens Axboe3b181522005-06-27 10:56:24 +02003592static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3593{
3594 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003595 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003596}
Jens Axboe22e2c502005-06-27 10:55:12 +02003597
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003598static void cfq_put_async_queues(struct cfq_data *cfqd)
3599{
3600 int i;
3601
3602 for (i = 0; i < IOPRIO_BE_NR; i++) {
3603 if (cfqd->async_cfqq[0][i])
3604 cfq_put_queue(cfqd->async_cfqq[0][i]);
3605 if (cfqd->async_cfqq[1][i])
3606 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003607 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003608
3609 if (cfqd->async_idle_cfqq)
3610 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003611}
3612
Jens Axboeb374d182008-10-31 10:05:07 +01003613static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614{
Jens Axboe22e2c502005-06-27 10:55:12 +02003615 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003616 struct request_queue *q = cfqd->queue;
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003617 bool wait = false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003618
Jens Axboe3b181522005-06-27 10:56:24 +02003619 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003620
Al Virod9ff4182006-03-18 13:51:22 -05003621 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003622
Al Virod9ff4182006-03-18 13:51:22 -05003623 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003624 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003625
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003626 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003627 cfq_release_cfq_groups(cfqd);
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003628
3629 /*
3630 * If there are groups which we could not unlink from blkcg list,
3631 * wait for a rcu period for them to be freed.
3632 */
3633 if (cfqd->nr_blkcg_linked_grps)
3634 wait = true;
Jens Axboe15c31be2007-07-10 13:43:25 +02003635
Al Virod9ff4182006-03-18 13:51:22 -05003636 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003637
3638 cfq_shutdown_timer_wq(cfqd);
3639
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003640 /*
3641 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3642 * Do this wait only if there are other unlinked groups out
3643 * there. This can happen if cgroup deletion path claimed the
3644 * responsibility of cleaning up a group before queue cleanup code
3645 * get to the group.
3646 *
3647 * Do not call synchronize_rcu() unconditionally as there are drivers
3648 * which create/delete request queue hundreds of times during scan/boot
3649 * and synchronize_rcu() can take significant time and slow down boot.
3650 */
3651 if (wait)
3652 synchronize_rcu();
Vivek Goyal2abae552011-05-23 10:02:19 +02003653
3654#ifdef CONFIG_CFQ_GROUP_IOSCHED
3655 /* Free up per cpu stats for root group */
3656 free_percpu(cfqd->root_group.blkg.stats_cpu);
3657#endif
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003658 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659}
3660
Jens Axboe165125e2007-07-24 09:28:11 +02003661static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003662{
3663 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003664 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003665 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003666 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667
Christoph Lameter94f60302007-07-17 04:03:29 -07003668 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Tejun Heoa73f7302011-12-14 00:33:37 +01003669 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003670 return NULL;
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003671
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003672 /* Init root service tree */
3673 cfqd->grp_service_tree = CFQ_RB_ROOT;
3674
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003675 /* Init root group */
3676 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003677 for_each_cfqg_st(cfqg, i, j, st)
3678 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003679 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003680
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003681 /* Give preference to root group over other groups */
3682 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3683
Vivek Goyal25fb5162009-12-03 12:59:46 -05003684#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003685 /*
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003686 * Set root group reference to 2. One reference will be dropped when
3687 * all groups on cfqd->cfqg_list are being deleted during queue exit.
3688 * Other reference will remain there as we don't want to delete this
3689 * group as it is statically allocated and gets destroyed when
3690 * throtl_data goes away.
Vivek Goyalb1c35762009-12-03 12:59:47 -05003691 */
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003692 cfqg->ref = 2;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04003693
3694 if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3695 kfree(cfqg);
3696 kfree(cfqd);
3697 return NULL;
3698 }
3699
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003700 rcu_read_lock();
Vivek Goyal5624a4e2011-05-19 15:38:28 -04003701
Vivek Goyale98ef892010-06-18 10:39:47 -04003702 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3703 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003704 rcu_read_unlock();
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003705 cfqd->nr_blkcg_linked_grps++;
3706
3707 /* Add group on cfqd->cfqg_list */
3708 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
Vivek Goyal25fb5162009-12-03 12:59:46 -05003709#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003710 /*
3711 * Not strictly needed (since RB_ROOT just clears the node and we
3712 * zeroed cfqd on alloc), but better be safe in case someone decides
3713 * to add magic to the rb code
3714 */
3715 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3716 cfqd->prio_trees[i] = RB_ROOT;
3717
Jens Axboe6118b702009-06-30 09:34:12 +02003718 /*
3719 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3720 * Grab a permanent reference to it, so that the normal code flow
3721 * will not attempt to free it.
3722 */
3723 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01003724 cfqd->oom_cfqq.ref++;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003725 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003726
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003728
Jens Axboe22e2c502005-06-27 10:55:12 +02003729 init_timer(&cfqd->idle_slice_timer);
3730 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3731 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3732
Jens Axboe23e018a2009-10-05 08:52:35 +02003733 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003734
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003736 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3737 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738 cfqd->cfq_back_max = cfq_back_max;
3739 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003740 cfqd->cfq_slice[0] = cfq_slice_async;
3741 cfqd->cfq_slice[1] = cfq_slice_sync;
Tao Ma5bf14c02012-04-01 14:33:39 -07003742 cfqd->cfq_target_latency = cfq_target_latency;
Jens Axboe22e2c502005-06-27 10:55:12 +02003743 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3744 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003745 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003746 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003747 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003748 /*
3749 * we optimistically start assuming sync ops weren't delayed in last
3750 * second, in order to have larger depth for async operations.
3751 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003752 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003753 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754}
3755
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756/*
3757 * sysfs parts below -->
3758 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759static ssize_t
3760cfq_var_show(unsigned int var, char *page)
3761{
3762 return sprintf(page, "%d\n", var);
3763}
3764
3765static ssize_t
3766cfq_var_store(unsigned int *var, const char *page, size_t count)
3767{
3768 char *p = (char *) page;
3769
3770 *var = simple_strtoul(p, &p, 10);
3771 return count;
3772}
3773
Linus Torvalds1da177e2005-04-16 15:20:36 -07003774#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003775static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003777 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778 unsigned int __data = __VAR; \
3779 if (__CONV) \
3780 __data = jiffies_to_msecs(__data); \
3781 return cfq_var_show(__data, (page)); \
3782}
3783SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003784SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3785SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003786SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3787SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003788SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003789SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003790SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3791SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3792SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003793SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07003794SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795#undef SHOW_FUNCTION
3796
3797#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003798static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003800 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801 unsigned int __data; \
3802 int ret = cfq_var_store(&__data, (page), count); \
3803 if (__data < (MIN)) \
3804 __data = (MIN); \
3805 else if (__data > (MAX)) \
3806 __data = (MAX); \
3807 if (__CONV) \
3808 *(__PTR) = msecs_to_jiffies(__data); \
3809 else \
3810 *(__PTR) = __data; \
3811 return ret; \
3812}
3813STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003814STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3815 UINT_MAX, 1);
3816STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3817 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05003818STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003819STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3820 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003821STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003822STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003823STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3824STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01003825STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3826 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003827STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07003828STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829#undef STORE_FUNCTION
3830
Al Viroe572ec72006-03-18 22:27:18 -05003831#define CFQ_ATTR(name) \
3832 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02003833
Al Viroe572ec72006-03-18 22:27:18 -05003834static struct elv_fs_entry cfq_attrs[] = {
3835 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05003836 CFQ_ATTR(fifo_expire_sync),
3837 CFQ_ATTR(fifo_expire_async),
3838 CFQ_ATTR(back_seek_max),
3839 CFQ_ATTR(back_seek_penalty),
3840 CFQ_ATTR(slice_sync),
3841 CFQ_ATTR(slice_async),
3842 CFQ_ATTR(slice_async_rq),
3843 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003844 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02003845 CFQ_ATTR(low_latency),
Tao Ma5bf14c02012-04-01 14:33:39 -07003846 CFQ_ATTR(target_latency),
Al Viroe572ec72006-03-18 22:27:18 -05003847 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848};
3849
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850static struct elevator_type iosched_cfq = {
3851 .ops = {
3852 .elevator_merge_fn = cfq_merge,
3853 .elevator_merged_fn = cfq_merged_request,
3854 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01003855 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07003856 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02003857 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02003859 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860 .elevator_deactivate_req_fn = cfq_deactivate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02003862 .elevator_former_req_fn = elv_rb_former_request,
3863 .elevator_latter_req_fn = elv_rb_latter_request,
Tejun Heo9b84cac2011-12-14 00:33:42 +01003864 .elevator_init_icq_fn = cfq_init_icq,
Tejun Heo7e5a8792011-12-14 00:33:42 +01003865 .elevator_exit_icq_fn = cfq_exit_icq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866 .elevator_set_req_fn = cfq_set_request,
3867 .elevator_put_req_fn = cfq_put_request,
3868 .elevator_may_queue_fn = cfq_may_queue,
3869 .elevator_init_fn = cfq_init_queue,
3870 .elevator_exit_fn = cfq_exit_queue,
3871 },
Tejun Heo3d3c2372011-12-14 00:33:42 +01003872 .icq_size = sizeof(struct cfq_io_cq),
3873 .icq_align = __alignof__(struct cfq_io_cq),
Al Viro3d1ab402006-03-18 18:35:43 -05003874 .elevator_attrs = cfq_attrs,
Tejun Heo3d3c2372011-12-14 00:33:42 +01003875 .elevator_name = "cfq",
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876 .elevator_owner = THIS_MODULE,
3877};
3878
Vivek Goyal3e252062009-12-04 10:36:42 -05003879#ifdef CONFIG_CFQ_GROUP_IOSCHED
3880static struct blkio_policy_type blkio_policy_cfq = {
3881 .ops = {
3882 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3883 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3884 },
Vivek Goyal062a6442010-09-15 17:06:33 -04003885 .plid = BLKIO_POLICY_PROP,
Vivek Goyal3e252062009-12-04 10:36:42 -05003886};
3887#else
3888static struct blkio_policy_type blkio_policy_cfq;
3889#endif
3890
Linus Torvalds1da177e2005-04-16 15:20:36 -07003891static int __init cfq_init(void)
3892{
Tejun Heo3d3c2372011-12-14 00:33:42 +01003893 int ret;
3894
Jens Axboe22e2c502005-06-27 10:55:12 +02003895 /*
3896 * could be 0 on HZ < 1000 setups
3897 */
3898 if (!cfq_slice_async)
3899 cfq_slice_async = 1;
3900 if (!cfq_slice_idle)
3901 cfq_slice_idle = 1;
3902
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003903#ifdef CONFIG_CFQ_GROUP_IOSCHED
3904 if (!cfq_group_idle)
3905 cfq_group_idle = 1;
3906#else
3907 cfq_group_idle = 0;
3908#endif
Tejun Heo3d3c2372011-12-14 00:33:42 +01003909 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3910 if (!cfq_pool)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 return -ENOMEM;
3912
Tejun Heo3d3c2372011-12-14 00:33:42 +01003913 ret = elv_register(&iosched_cfq);
3914 if (ret) {
3915 kmem_cache_destroy(cfq_pool);
3916 return ret;
3917 }
Tejun Heo3d3c2372011-12-14 00:33:42 +01003918
Vivek Goyal3e252062009-12-04 10:36:42 -05003919 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01003921 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922}
3923
3924static void __exit cfq_exit(void)
3925{
Vivek Goyal3e252062009-12-04 10:36:42 -05003926 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003927 elv_unregister(&iosched_cfq);
Tejun Heo3d3c2372011-12-14 00:33:42 +01003928 kmem_cache_destroy(cfq_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003929}
3930
3931module_init(cfq_init);
3932module_exit(cfq_exit);
3933
3934MODULE_AUTHOR("Jens Axboe");
3935MODULE_LICENSE("GPL");
3936MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");