blob: 9eba291eb6fd23854aee14d8a35b5b1108fc0629 [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>
Vivek Goyale98ef892010-06-18 10:39:47 -040017#include "cfq.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070018
19/*
20 * tunables
21 */
Jens Axboefe094d92008-01-31 13:08:54 +010022/* max queue in one round of service */
Shaohua Liabc3c742010-03-01 09:20:54 +010023static const int cfq_quantum = 8;
Arjan van de Ven64100092006-01-06 09:46:02 +010024static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010025/* maximum backwards seek, in KiB */
26static const int cfq_back_max = 16 * 1024;
27/* penalty of a backwards seek */
28static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010029static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020030static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010031static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020032static int cfq_slice_idle = HZ / 125;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +020033static int cfq_group_idle = HZ / 125;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +010034static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35static const int cfq_hist_divisor = 4;
Jens Axboe22e2c502005-06-27 10:55:12 +020036
Jens Axboed9e76202007-04-20 14:27:50 +020037/*
Jens Axboe08717142008-01-28 11:38:15 +010038 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020039 */
Jens Axboe08717142008-01-28 11:38:15 +010040#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020041
42/*
43 * below this threshold, we consider thinktime immediate
44 */
45#define CFQ_MIN_TT (2)
46
Jens Axboe22e2c502005-06-27 10:55:12 +020047#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020048#define CFQ_HW_QUEUE_MIN (5)
Vivek Goyal25bc6b02009-12-03 12:59:43 -050049#define CFQ_SERVICE_SHIFT 12
Jens Axboe22e2c502005-06-27 10:55:12 +020050
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010051#define CFQQ_SEEK_THR (sector_t)(8 * 100)
Shaohua Lie9ce3352010-03-19 08:03:04 +010052#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
Corrado Zoccolo41647e72010-02-27 19:45:40 +010053#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010054#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
Shaohua Liae54abe2010-02-05 13:11:45 +010055
Jens Axboefe094d92008-01-31 13:08:54 +010056#define RQ_CIC(rq) \
57 ((struct cfq_io_context *) (rq)->elevator_private)
Jens Axboe7b679132008-05-30 12:23:07 +020058#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +020059#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private3)
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Christoph Lametere18b8902006-12-06 20:33:20 -080061static struct kmem_cache *cfq_pool;
62static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
Tejun Heo245b2e72009-06-24 15:13:48 +090064static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050065static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020066static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050067
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +040068static DEFINE_SPINLOCK(cic_index_lock);
69static DEFINE_IDA(cic_index_ida);
70
Jens Axboe22e2c502005-06-27 10:55:12 +020071#define CFQ_PRIO_LISTS IOPRIO_BE_NR
72#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020073#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
Jens Axboe206dc692006-03-28 13:03:44 +020075#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050076#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020077
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;
Vivek Goyal25bc6b02009-12-03 12:59:43 -050090 struct rb_node *active;
Jens Axboecc09e292007-04-26 12:53:50 +020091};
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010092#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
93 .count = 0, .min_vdisktime = 0, }
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 */
100 atomic_t ref;
101 /* 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
133 /* pending metadata requests */
134 int meta_pending;
135 /* 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;
140 unsigned short ioprio_class, org_ioprio_class;
141
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 Goyalae30c282009-12-03 12:59:55 -0500150 struct cfq_group *orig_cfqg;
Vivek Goyalc4e78932010-08-23 12:25:03 +0200151 /* Number of sectors dispatched from queue in single dispatch round */
152 unsigned long nr_sectors;
Jens Axboe6118b702009-06-30 09:34:12 +0200153};
154
155/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100156 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100157 * IDLE is handled separately, so it has negative index
158 */
159enum wl_prio_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100160 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500161 RT_WORKLOAD = 1,
162 IDLE_WORKLOAD = 2,
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;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500182 bool on_st;
183
184 /* number of cfqq currently on this group */
185 int nr_cfqq;
186
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500187 /* Per group busy queus average. Useful for workload slice calc. */
188 unsigned int busy_queues_avg[2];
Jens Axboe22e2c502005-06-27 10:55:12 +0200189 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100190 * rr lists of queues with requests, onle rr for each priority class.
191 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200192 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100193 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100194 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500195
196 unsigned long saved_workload_slice;
197 enum wl_type_t saved_workload;
198 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500199 struct blkio_group blkg;
200#ifdef CONFIG_CFQ_GROUP_IOSCHED
201 struct hlist_node cfqd_node;
Vivek Goyalb1c35762009-12-03 12:59:47 -0500202 atomic_t ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500203#endif
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200204 /* number of requests that are on the dispatch list or inside driver */
205 int dispatched;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500206};
207
208/*
209 * Per block device queue structure
210 */
211struct cfq_data {
212 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500213 /* Root service tree for cfq_groups */
214 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500215 struct cfq_group root_group;
216
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100217 /*
218 * The priority currently being served
219 */
220 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100221 enum wl_type_t serving_type;
222 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500223 struct cfq_group *serving_group;
Corrado Zoccolo8e550632009-11-26 10:02:58 +0100224 bool noidle_tree_requires_idle;
Jens Axboea36e71f2009-04-15 12:15:11 +0200225
226 /*
227 * Each priority tree is sorted by next_request position. These
228 * trees are used when determining if two or more queues are
229 * interleaving requests (see cfq_close_cooperator).
230 */
231 struct rb_root prio_trees[CFQ_PRIO_LISTS];
232
Jens Axboe22e2c502005-06-27 10:55:12 +0200233 unsigned int busy_queues;
234
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100235 int rq_in_driver;
236 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200237
238 /*
239 * queue-depth detection
240 */
241 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200242 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100243 /*
244 * hw_tag can be
245 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
246 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
247 * 0 => no NCQ
248 */
249 int hw_tag_est_depth;
250 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200251
252 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200253 * idle window management
254 */
255 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200256 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200257
258 struct cfq_queue *active_queue;
259 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200260
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200261 /*
262 * async queue for each priority case
263 */
264 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
265 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200266
Jens Axboe6d048f52007-04-25 12:44:27 +0200267 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269 /*
270 * tunables, see top of file
271 */
272 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200273 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 unsigned int cfq_back_penalty;
275 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200276 unsigned int cfq_slice[2];
277 unsigned int cfq_slice_async_rq;
278 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200279 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200280 unsigned int cfq_latency;
Vivek Goyalae30c282009-12-03 12:59:55 -0500281 unsigned int cfq_group_isolation;
Al Virod9ff4182006-03-18 13:51:22 -0500282
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400283 unsigned int cic_index;
Al Virod9ff4182006-03-18 13:51:22 -0500284 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285
Jens Axboe6118b702009-06-30 09:34:12 +0200286 /*
287 * Fallback dummy cfqq for extreme OOM conditions
288 */
289 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200290
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100291 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500292
293 /* List of cfq groups being managed on this device*/
294 struct hlist_head cfqg_list;
Jens Axboebb729bc2009-12-06 09:54:19 +0100295 struct rcu_head rcu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296};
297
Vivek Goyal25fb5162009-12-03 12:59:46 -0500298static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
299
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500300static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
301 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500302 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100303{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500304 if (!cfqg)
305 return NULL;
306
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100307 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500308 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100309
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500310 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100311}
312
Jens Axboe3b181522005-06-27 10:56:24 +0200313enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100314 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
315 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200316 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100317 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100318 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
319 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
320 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100321 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200322 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400323 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100324 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100325 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500326 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200327};
328
329#define CFQ_CFQQ_FNS(name) \
330static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
331{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100332 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200333} \
334static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
335{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100336 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200337} \
338static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
339{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100340 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200341}
342
343CFQ_CFQQ_FNS(on_rr);
344CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200345CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200346CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200347CFQ_CFQQ_FNS(fifo_expire);
348CFQ_CFQQ_FNS(idle_window);
349CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100350CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200351CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200352CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100353CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100354CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500355CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200356#undef CFQ_CFQQ_FNS
357
Vivek Goyalafc24d42010-04-26 19:27:56 +0200358#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500359#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
360 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
361 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
362 blkg_path(&(cfqq)->cfqg->blkg), ##args);
363
364#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
365 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
366 blkg_path(&(cfqg)->blkg), ##args); \
367
368#else
Jens Axboe7b679132008-05-30 12:23:07 +0200369#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
370 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500371#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
372#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200373#define cfq_log(cfqd, fmt, args...) \
374 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
375
Vivek Goyal615f0252009-12-03 12:59:39 -0500376/* Traverses through cfq group service trees */
377#define for_each_cfqg_st(cfqg, i, j, st) \
378 for (i = 0; i <= IDLE_WORKLOAD; i++) \
379 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
380 : &cfqg->service_tree_idle; \
381 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
382 (i == IDLE_WORKLOAD && j == 0); \
383 j++, st = i < IDLE_WORKLOAD ? \
384 &cfqg->service_trees[i][j]: NULL) \
385
386
Vivek Goyal02b35082010-08-23 12:23:53 +0200387static inline bool iops_mode(struct cfq_data *cfqd)
388{
389 /*
390 * If we are not idling on queues and it is a NCQ drive, parallel
391 * execution of requests is on and measuring time is not possible
392 * in most of the cases until and unless we drive shallower queue
393 * depths and that becomes a performance bottleneck. In such cases
394 * switch to start providing fairness in terms of number of IOs.
395 */
396 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
397 return true;
398 else
399 return false;
400}
401
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100402static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
403{
404 if (cfq_class_idle(cfqq))
405 return IDLE_WORKLOAD;
406 if (cfq_class_rt(cfqq))
407 return RT_WORKLOAD;
408 return BE_WORKLOAD;
409}
410
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100411
412static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
413{
414 if (!cfq_cfqq_sync(cfqq))
415 return ASYNC_WORKLOAD;
416 if (!cfq_cfqq_idle_window(cfqq))
417 return SYNC_NOIDLE_WORKLOAD;
418 return SYNC_WORKLOAD;
419}
420
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500421static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
422 struct cfq_data *cfqd,
423 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100424{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500425 if (wl == IDLE_WORKLOAD)
426 return cfqg->service_tree_idle.count;
427
428 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
429 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
430 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100431}
432
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500433static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
434 struct cfq_group *cfqg)
435{
436 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
437 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
438}
439
Jens Axboe165125e2007-07-24 09:28:11 +0200440static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200441static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100442 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100443static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200444 struct io_context *);
445
446static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200447 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200448{
Jens Axboea6151c32009-10-07 20:02:57 +0200449 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200450}
451
452static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200453 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200454{
Jens Axboea6151c32009-10-07 20:02:57 +0200455 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200456}
457
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400458#define CIC_DEAD_KEY 1ul
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400459#define CIC_DEAD_INDEX_SHIFT 1
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400460
461static inline void *cfqd_dead_key(struct cfq_data *cfqd)
462{
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400463 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400464}
465
466static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
467{
468 struct cfq_data *cfqd = cic->key;
469
470 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
471 return NULL;
472
473 return cfqd;
474}
475
Vasily Tarasov91fac312007-04-25 12:29:51 +0200476/*
477 * We regard a request as SYNC, if it's either a read or has the SYNC bit
478 * set (in which case it could also be direct WRITE).
479 */
Jens Axboea6151c32009-10-07 20:02:57 +0200480static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200481{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200482 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200483}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700486 * scheduler run of queue, if there are requests pending and no one in the
487 * driver that will restart queueing
488 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200489static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700490{
Jens Axboe7b679132008-05-30 12:23:07 +0200491 if (cfqd->busy_queues) {
492 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200493 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200494 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700495}
496
Jens Axboe165125e2007-07-24 09:28:11 +0200497static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700498{
499 struct cfq_data *cfqd = q->elevator->elevator_data;
500
Vivek Goyalf04a6422009-12-03 12:59:40 -0500501 return !cfqd->rq_queued;
Andrew Morton99f95e52005-06-27 20:14:05 -0700502}
503
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100505 * Scale schedule slice based on io priority. Use the sync time slice only
506 * if a queue is marked sync and has sync io queued. A sync queue with async
507 * io only, should not get full sync slice length.
508 */
Jens Axboea6151c32009-10-07 20:02:57 +0200509static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200510 unsigned short prio)
511{
512 const int base_slice = cfqd->cfq_slice[sync];
513
514 WARN_ON(prio >= IOPRIO_BE_NR);
515
516 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
517}
518
Jens Axboe44f7c162007-01-19 11:51:58 +1100519static inline int
520cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
521{
Jens Axboed9e76202007-04-20 14:27:50 +0200522 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100523}
524
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500525static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
526{
527 u64 d = delta << CFQ_SERVICE_SHIFT;
528
529 d = d * BLKIO_WEIGHT_DEFAULT;
530 do_div(d, cfqg->weight);
531 return d;
532}
533
534static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
535{
536 s64 delta = (s64)(vdisktime - min_vdisktime);
537 if (delta > 0)
538 min_vdisktime = vdisktime;
539
540 return min_vdisktime;
541}
542
543static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
544{
545 s64 delta = (s64)(vdisktime - min_vdisktime);
546 if (delta < 0)
547 min_vdisktime = vdisktime;
548
549 return min_vdisktime;
550}
551
552static void update_min_vdisktime(struct cfq_rb_root *st)
553{
554 u64 vdisktime = st->min_vdisktime;
555 struct cfq_group *cfqg;
556
557 if (st->active) {
558 cfqg = rb_entry_cfqg(st->active);
559 vdisktime = cfqg->vdisktime;
560 }
561
562 if (st->left) {
563 cfqg = rb_entry_cfqg(st->left);
564 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
565 }
566
567 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
568}
569
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100570/*
571 * get averaged number of queues of RT/BE priority.
572 * average is updated, with a formula that gives more weight to higher numbers,
573 * to quickly follows sudden increases and decrease slowly
574 */
575
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500576static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
577 struct cfq_group *cfqg, bool rt)
Jens Axboe58696192009-10-28 09:27:07 +0100578{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100579 unsigned min_q, max_q;
580 unsigned mult = cfq_hist_divisor - 1;
581 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500582 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100583
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500584 min_q = min(cfqg->busy_queues_avg[rt], busy);
585 max_q = max(cfqg->busy_queues_avg[rt], busy);
586 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100587 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500588 return cfqg->busy_queues_avg[rt];
589}
590
591static inline unsigned
592cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
593{
594 struct cfq_rb_root *st = &cfqd->grp_service_tree;
595
596 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100597}
598
Jens Axboe44f7c162007-01-19 11:51:58 +1100599static inline void
600cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
601{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100602 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
603 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500604 /*
605 * interested queues (we consider only the ones with the same
606 * priority class in the cfq group)
607 */
608 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
609 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100610 unsigned sync_slice = cfqd->cfq_slice[1];
611 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500612 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
613
614 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100615 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
616 /* scale low_slice according to IO priority
617 * and sync vs async */
618 unsigned low_slice =
619 min(slice, base_low_slice * slice / sync_slice);
620 /* the adapted slice value is scaled to fit all iqs
621 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500622 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100623 low_slice);
624 }
625 }
Vivek Goyaldae739e2009-12-03 12:59:45 -0500626 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100627 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500628 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200629 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100630}
631
632/*
633 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
634 * isn't valid until the first request from the dispatch is activated
635 * and the slice time set.
636 */
Jens Axboea6151c32009-10-07 20:02:57 +0200637static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100638{
639 if (cfq_cfqq_slice_new(cfqq))
640 return 0;
641 if (time_before(jiffies, cfqq->slice_end))
642 return 0;
643
644 return 1;
645}
646
647/*
Jens Axboe5e705372006-07-13 12:39:25 +0200648 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200650 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 */
Jens Axboe5e705372006-07-13 12:39:25 +0200652static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100653cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100655 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200657#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
658#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
659 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
Jens Axboe5e705372006-07-13 12:39:25 +0200661 if (rq1 == NULL || rq1 == rq2)
662 return rq2;
663 if (rq2 == NULL)
664 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200665
Jens Axboe5e705372006-07-13 12:39:25 +0200666 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
667 return rq1;
668 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
669 return rq2;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200670 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200671 return rq1;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200672 else if ((rq2->cmd_flags & REQ_META) &&
673 !(rq1->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200674 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675
Tejun Heo83096eb2009-05-07 22:24:39 +0900676 s1 = blk_rq_pos(rq1);
677 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 /*
680 * by definition, 1KiB is 2 sectors
681 */
682 back_max = cfqd->cfq_back_max * 2;
683
684 /*
685 * Strict one way elevator _except_ in the case where we allow
686 * short backward seeks which are biased as twice the cost of a
687 * similar forward seek.
688 */
689 if (s1 >= last)
690 d1 = s1 - last;
691 else if (s1 + back_max >= last)
692 d1 = (last - s1) * cfqd->cfq_back_penalty;
693 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200694 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695
696 if (s2 >= last)
697 d2 = s2 - last;
698 else if (s2 + back_max >= last)
699 d2 = (last - s2) * cfqd->cfq_back_penalty;
700 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200701 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702
703 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704
Andreas Mohre8a99052006-03-28 08:59:49 +0200705 /*
706 * By doing switch() on the bit mask "wrap" we avoid having to
707 * check two variables for all permutations: --> faster!
708 */
709 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200710 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200711 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200712 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200713 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200714 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200715 else {
716 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200717 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200718 else
Jens Axboe5e705372006-07-13 12:39:25 +0200719 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200720 }
721
722 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200723 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200724 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200725 return rq2;
726 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200727 default:
728 /*
729 * Since both rqs are wrapped,
730 * start with the one that's further behind head
731 * (--> only *one* back seek required),
732 * since back seek takes more time than forward.
733 */
734 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200735 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 else
Jens Axboe5e705372006-07-13 12:39:25 +0200737 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 }
739}
740
Jens Axboe498d3aa2007-04-26 12:54:48 +0200741/*
742 * The below is leftmost cache rbtree addon
743 */
Jens Axboe08717142008-01-28 11:38:15 +0100744static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200745{
Vivek Goyal615f0252009-12-03 12:59:39 -0500746 /* Service tree is empty */
747 if (!root->count)
748 return NULL;
749
Jens Axboecc09e292007-04-26 12:53:50 +0200750 if (!root->left)
751 root->left = rb_first(&root->rb);
752
Jens Axboe08717142008-01-28 11:38:15 +0100753 if (root->left)
754 return rb_entry(root->left, struct cfq_queue, rb_node);
755
756 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200757}
758
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500759static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
760{
761 if (!root->left)
762 root->left = rb_first(&root->rb);
763
764 if (root->left)
765 return rb_entry_cfqg(root->left);
766
767 return NULL;
768}
769
Jens Axboea36e71f2009-04-15 12:15:11 +0200770static void rb_erase_init(struct rb_node *n, struct rb_root *root)
771{
772 rb_erase(n, root);
773 RB_CLEAR_NODE(n);
774}
775
Jens Axboecc09e292007-04-26 12:53:50 +0200776static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
777{
778 if (root->left == n)
779 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200780 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100781 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200782}
783
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784/*
785 * would be nice to take fifo expire time into account as well
786 */
Jens Axboe5e705372006-07-13 12:39:25 +0200787static struct request *
788cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
789 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790{
Jens Axboe21183b02006-07-13 12:33:14 +0200791 struct rb_node *rbnext = rb_next(&last->rb_node);
792 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200793 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794
Jens Axboe21183b02006-07-13 12:33:14 +0200795 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
797 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200798 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200799
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200801 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200802 else {
803 rbnext = rb_first(&cfqq->sort_list);
804 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200805 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200806 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100808 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809}
810
Jens Axboed9e76202007-04-20 14:27:50 +0200811static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
812 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813{
Jens Axboed9e76202007-04-20 14:27:50 +0200814 /*
815 * just an approximation, should be ok.
816 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500817 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100818 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200819}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500821static inline s64
822cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
823{
824 return cfqg->vdisktime - st->min_vdisktime;
825}
826
827static void
828__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
829{
830 struct rb_node **node = &st->rb.rb_node;
831 struct rb_node *parent = NULL;
832 struct cfq_group *__cfqg;
833 s64 key = cfqg_key(st, cfqg);
834 int left = 1;
835
836 while (*node != NULL) {
837 parent = *node;
838 __cfqg = rb_entry_cfqg(parent);
839
840 if (key < cfqg_key(st, __cfqg))
841 node = &parent->rb_left;
842 else {
843 node = &parent->rb_right;
844 left = 0;
845 }
846 }
847
848 if (left)
849 st->left = &cfqg->rb_node;
850
851 rb_link_node(&cfqg->rb_node, parent, node);
852 rb_insert_color(&cfqg->rb_node, &st->rb);
853}
854
855static void
856cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
857{
858 struct cfq_rb_root *st = &cfqd->grp_service_tree;
859 struct cfq_group *__cfqg;
860 struct rb_node *n;
861
862 cfqg->nr_cfqq++;
863 if (cfqg->on_st)
864 return;
865
866 /*
867 * Currently put the group at the end. Later implement something
868 * so that groups get lesser vtime based on their weights, so that
869 * if group does not loose all if it was not continously backlogged.
870 */
871 n = rb_last(&st->rb);
872 if (n) {
873 __cfqg = rb_entry_cfqg(n);
874 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
875 } else
876 cfqg->vdisktime = st->min_vdisktime;
877
878 __cfq_group_service_tree_add(st, cfqg);
879 cfqg->on_st = true;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500880 st->total_weight += cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500881}
882
883static void
884cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
885{
886 struct cfq_rb_root *st = &cfqd->grp_service_tree;
887
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500888 if (st->active == &cfqg->rb_node)
889 st->active = NULL;
890
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500891 BUG_ON(cfqg->nr_cfqq < 1);
892 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500893
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500894 /* If there are other cfq queues under this group, don't delete it */
895 if (cfqg->nr_cfqq)
896 return;
897
Vivek Goyal2868ef72009-12-03 12:59:48 -0500898 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500899 cfqg->on_st = false;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500900 st->total_weight -= cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500901 if (!RB_EMPTY_NODE(&cfqg->rb_node))
902 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500903 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400904 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500905}
906
907static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
908{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500909 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500910
911 /*
912 * Queue got expired before even a single request completed or
913 * got expired immediately after first request completion.
914 */
915 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
916 /*
917 * Also charge the seek time incurred to the group, otherwise
918 * if there are mutiple queues in the group, each can dispatch
919 * a single request on seeky media and cause lots of seek time
920 * and group will never know it.
921 */
922 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
923 1);
924 } else {
925 slice_used = jiffies - cfqq->slice_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500926 if (slice_used > cfqq->allocated_slice)
927 slice_used = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500928 }
929
Vivek Goyaldae739e2009-12-03 12:59:45 -0500930 return slice_used;
931}
932
933static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200934 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500935{
936 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal02b35082010-08-23 12:23:53 +0200937 unsigned int used_sl, charge;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500938 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
939 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500940
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500941 BUG_ON(nr_sync < 0);
Vivek Goyal02b35082010-08-23 12:23:53 +0200942 used_sl = charge = cfq_cfqq_slice_usage(cfqq);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500943
Vivek Goyal02b35082010-08-23 12:23:53 +0200944 if (iops_mode(cfqd))
945 charge = cfqq->slice_dispatch;
946 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
947 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500948
949 /* Can't update vdisktime while group is on service tree */
950 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyal02b35082010-08-23 12:23:53 +0200951 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500952 __cfq_group_service_tree_add(st, cfqg);
953
954 /* This group is being expired. Save the context */
955 if (time_after(cfqd->workload_expires, jiffies)) {
956 cfqg->saved_workload_slice = cfqd->workload_expires
957 - jiffies;
958 cfqg->saved_workload = cfqd->serving_type;
959 cfqg->saved_serving_prio = cfqd->serving_prio;
960 } else
961 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500962
963 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
964 st->min_vdisktime);
Vivek Goyalc4e78932010-08-23 12:25:03 +0200965 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
966 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
967 iops_mode(cfqd), cfqq->nr_sectors);
Vivek Goyale98ef892010-06-18 10:39:47 -0400968 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
969 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500970}
971
Vivek Goyal25fb5162009-12-03 12:59:46 -0500972#ifdef CONFIG_CFQ_GROUP_IOSCHED
973static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
974{
975 if (blkg)
976 return container_of(blkg, struct cfq_group, blkg);
977 return NULL;
978}
979
Vivek Goyalf8d461d2009-12-03 12:59:52 -0500980void
981cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
982{
983 cfqg_of_blkg(blkg)->weight = weight;
984}
985
Vivek Goyal25fb5162009-12-03 12:59:46 -0500986static struct cfq_group *
987cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
988{
989 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
990 struct cfq_group *cfqg = NULL;
991 void *key = cfqd;
992 int i, j;
993 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -0500994 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
995 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500996
Vivek Goyal25fb5162009-12-03 12:59:46 -0500997 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
Ricky Beniteza74b2ad2010-04-05 18:22:17 +0200998 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
999 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1000 cfqg->blkg.dev = MKDEV(major, minor);
1001 goto done;
1002 }
Vivek Goyal25fb5162009-12-03 12:59:46 -05001003 if (cfqg || !create)
1004 goto done;
1005
1006 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1007 if (!cfqg)
1008 goto done;
1009
Vivek Goyal25fb5162009-12-03 12:59:46 -05001010 for_each_cfqg_st(cfqg, i, j, st)
1011 *st = CFQ_RB_ROOT;
1012 RB_CLEAR_NODE(&cfqg->rb_node);
1013
Vivek Goyalb1c35762009-12-03 12:59:47 -05001014 /*
1015 * Take the initial reference that will be released on destroy
1016 * This can be thought of a joint reference by cgroup and
1017 * elevator which will be dropped by either elevator exit
1018 * or cgroup deletion path depending on who is exiting first.
1019 */
1020 atomic_set(&cfqg->ref, 1);
1021
Vivek Goyal180be2a2010-09-14 08:47:11 +02001022 /*
1023 * Add group onto cgroup list. It might happen that bdi->dev is
1024 * not initiliazed yet. Initialize this new group without major
1025 * and minor info and this info will be filled in once a new thread
1026 * comes for IO. See code above.
1027 */
1028 if (bdi->dev) {
1029 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1030 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
Vivek Goyal22084192009-12-03 12:59:49 -05001031 MKDEV(major, minor));
Vivek Goyal180be2a2010-09-14 08:47:11 +02001032 } else
1033 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1034 0);
1035
Gui Jianfeng34d0f172010-04-13 16:05:49 +08001036 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001037
1038 /* Add group on cfqd list */
1039 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1040
1041done:
Vivek Goyal25fb5162009-12-03 12:59:46 -05001042 return cfqg;
1043}
1044
1045/*
1046 * Search for the cfq group current task belongs to. If create = 1, then also
1047 * create the cfq group if it does not exist. request_queue lock must be held.
1048 */
1049static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1050{
1051 struct cgroup *cgroup;
1052 struct cfq_group *cfqg = NULL;
1053
1054 rcu_read_lock();
1055 cgroup = task_cgroup(current, blkio_subsys_id);
1056 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1057 if (!cfqg && create)
1058 cfqg = &cfqd->root_group;
1059 rcu_read_unlock();
1060 return cfqg;
1061}
1062
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001063static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1064{
1065 atomic_inc(&cfqg->ref);
1066 return cfqg;
1067}
1068
Vivek Goyal25fb5162009-12-03 12:59:46 -05001069static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1070{
1071 /* Currently, all async queues are mapped to root group */
1072 if (!cfq_cfqq_sync(cfqq))
1073 cfqg = &cfqq->cfqd->root_group;
1074
1075 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001076 /* cfqq reference on cfqg */
1077 atomic_inc(&cfqq->cfqg->ref);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001078}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001079
1080static void cfq_put_cfqg(struct cfq_group *cfqg)
1081{
1082 struct cfq_rb_root *st;
1083 int i, j;
1084
1085 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1086 if (!atomic_dec_and_test(&cfqg->ref))
1087 return;
1088 for_each_cfqg_st(cfqg, i, j, st)
1089 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1090 kfree(cfqg);
1091}
1092
1093static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1094{
1095 /* Something wrong if we are trying to remove same group twice */
1096 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1097
1098 hlist_del_init(&cfqg->cfqd_node);
1099
1100 /*
1101 * Put the reference taken at the time of creation so that when all
1102 * queues are gone, group can be destroyed.
1103 */
1104 cfq_put_cfqg(cfqg);
1105}
1106
1107static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1108{
1109 struct hlist_node *pos, *n;
1110 struct cfq_group *cfqg;
1111
1112 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1113 /*
1114 * If cgroup removal path got to blk_group first and removed
1115 * it from cgroup list, then it will take care of destroying
1116 * cfqg also.
1117 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001118 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001119 cfq_destroy_cfqg(cfqd, cfqg);
1120 }
1121}
1122
1123/*
1124 * Blk cgroup controller notification saying that blkio_group object is being
1125 * delinked as associated cgroup object is going away. That also means that
1126 * no new IO will come in this group. So get rid of this group as soon as
1127 * any pending IO in the group is finished.
1128 *
1129 * This function is called under rcu_read_lock(). key is the rcu protected
1130 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1131 * read lock.
1132 *
1133 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1134 * it should not be NULL as even if elevator was exiting, cgroup deltion
1135 * path got to it first.
1136 */
1137void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1138{
1139 unsigned long flags;
1140 struct cfq_data *cfqd = key;
1141
1142 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1143 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1144 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1145}
1146
Vivek Goyal25fb5162009-12-03 12:59:46 -05001147#else /* GROUP_IOSCHED */
1148static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1149{
1150 return &cfqd->root_group;
1151}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001152
1153static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1154{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001155 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001156}
1157
Vivek Goyal25fb5162009-12-03 12:59:46 -05001158static inline void
1159cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1160 cfqq->cfqg = cfqg;
1161}
1162
Vivek Goyalb1c35762009-12-03 12:59:47 -05001163static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1164static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1165
Vivek Goyal25fb5162009-12-03 12:59:46 -05001166#endif /* GROUP_IOSCHED */
1167
Jens Axboe498d3aa2007-04-26 12:54:48 +02001168/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001169 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa2007-04-26 12:54:48 +02001170 * requests waiting to be processed. It is sorted in the order that
1171 * we will service the queues.
1172 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001173static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001174 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001175{
Jens Axboe08717142008-01-28 11:38:15 +01001176 struct rb_node **p, *parent;
1177 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001178 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001179 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa2007-04-26 12:54:48 +02001180 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001181 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001182 int group_changed = 0;
1183
1184#ifdef CONFIG_CFQ_GROUP_IOSCHED
1185 if (!cfqd->cfq_group_isolation
1186 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1187 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1188 /* Move this cfq to root group */
1189 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1190 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1191 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1192 cfqq->orig_cfqg = cfqq->cfqg;
1193 cfqq->cfqg = &cfqd->root_group;
1194 atomic_inc(&cfqd->root_group.ref);
1195 group_changed = 1;
1196 } else if (!cfqd->cfq_group_isolation
1197 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1198 /* cfqq is sequential now needs to go to its original group */
1199 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1200 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1201 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1202 cfq_put_cfqg(cfqq->cfqg);
1203 cfqq->cfqg = cfqq->orig_cfqg;
1204 cfqq->orig_cfqg = NULL;
1205 group_changed = 1;
1206 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1207 }
1208#endif
Jens Axboed9e76202007-04-20 14:27:50 +02001209
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001210 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001211 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001212 if (cfq_class_idle(cfqq)) {
1213 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001214 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001215 if (parent && parent != &cfqq->rb_node) {
1216 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1217 rb_key += __cfqq->rb_key;
1218 } else
1219 rb_key += jiffies;
1220 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001221 /*
1222 * Get our rb key offset. Subtract any residual slice
1223 * value carried from last service. A negative resid
1224 * count indicates slice overrun, and this should position
1225 * the next service time further away in the tree.
1226 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001227 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001228 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001229 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001230 } else {
1231 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001232 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001233 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1234 }
Jens Axboed9e76202007-04-20 14:27:50 +02001235
1236 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001237 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001238 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001239 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001240 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001241 if (rb_key == cfqq->rb_key &&
1242 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001243 return;
Jens Axboe53b03742006-07-28 09:48:51 +02001244
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001245 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1246 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001247 }
Jens Axboed9e76202007-04-20 14:27:50 +02001248
Jens Axboe498d3aa2007-04-26 12:54:48 +02001249 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001250 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001251 cfqq->service_tree = service_tree;
1252 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001253 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001254 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001255
Jens Axboed9e76202007-04-20 14:27:50 +02001256 parent = *p;
1257 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1258
Jens Axboe0c534e02007-04-18 20:01:57 +02001259 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001260 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001261 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001262 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001263 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001264 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001265 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001266 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001267 }
Jens Axboe67060e32007-04-18 20:13:32 +02001268
1269 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001270 }
1271
Jens Axboecc09e292007-04-26 12:53:50 +02001272 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001273 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001274
Jens Axboed9e76202007-04-20 14:27:50 +02001275 cfqq->rb_key = rb_key;
1276 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001277 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1278 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001279 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001280 return;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001281 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282}
1283
Jens Axboea36e71f2009-04-15 12:15:11 +02001284static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001285cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1286 sector_t sector, struct rb_node **ret_parent,
1287 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001288{
Jens Axboea36e71f2009-04-15 12:15:11 +02001289 struct rb_node **p, *parent;
1290 struct cfq_queue *cfqq = NULL;
1291
1292 parent = NULL;
1293 p = &root->rb_node;
1294 while (*p) {
1295 struct rb_node **n;
1296
1297 parent = *p;
1298 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1299
1300 /*
1301 * Sort strictly based on sector. Smallest to the left,
1302 * largest to the right.
1303 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001304 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001305 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001306 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001307 n = &(*p)->rb_left;
1308 else
1309 break;
1310 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001311 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001312 }
1313
1314 *ret_parent = parent;
1315 if (rb_link)
1316 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001317 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001318}
1319
1320static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1321{
Jens Axboea36e71f2009-04-15 12:15:11 +02001322 struct rb_node **p, *parent;
1323 struct cfq_queue *__cfqq;
1324
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001325 if (cfqq->p_root) {
1326 rb_erase(&cfqq->p_node, cfqq->p_root);
1327 cfqq->p_root = NULL;
1328 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001329
1330 if (cfq_class_idle(cfqq))
1331 return;
1332 if (!cfqq->next_rq)
1333 return;
1334
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001335 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001336 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1337 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001338 if (!__cfqq) {
1339 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001340 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1341 } else
1342 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001343}
1344
Jens Axboe498d3aa2007-04-26 12:54:48 +02001345/*
1346 * Update cfqq's position in the service tree.
1347 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001348static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001349{
Jens Axboe6d048f52007-04-25 12:44:27 +02001350 /*
1351 * Resorting requires the cfqq to be on the RR list already.
1352 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001353 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001354 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001355 cfq_prio_tree_add(cfqd, cfqq);
1356 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001357}
1358
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359/*
1360 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001361 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 */
Jens Axboefebffd62008-01-28 13:19:43 +01001363static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364{
Jens Axboe7b679132008-05-30 12:23:07 +02001365 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001366 BUG_ON(cfq_cfqq_on_rr(cfqq));
1367 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 cfqd->busy_queues++;
1369
Jens Axboeedd75ff2007-04-19 12:03:34 +02001370 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371}
1372
Jens Axboe498d3aa2007-04-26 12:54:48 +02001373/*
1374 * Called when the cfqq no longer has requests pending, remove it from
1375 * the service tree.
1376 */
Jens Axboefebffd62008-01-28 13:19:43 +01001377static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378{
Jens Axboe7b679132008-05-30 12:23:07 +02001379 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001380 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1381 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001383 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1384 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1385 cfqq->service_tree = NULL;
1386 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001387 if (cfqq->p_root) {
1388 rb_erase(&cfqq->p_node, cfqq->p_root);
1389 cfqq->p_root = NULL;
1390 }
Jens Axboed9e76202007-04-20 14:27:50 +02001391
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001392 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393 BUG_ON(!cfqd->busy_queues);
1394 cfqd->busy_queues--;
1395}
1396
1397/*
1398 * rb tree support functions
1399 */
Jens Axboefebffd62008-01-28 13:19:43 +01001400static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401{
Jens Axboe5e705372006-07-13 12:39:25 +02001402 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001403 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404
Jens Axboeb4878f22005-10-20 16:42:29 +02001405 BUG_ON(!cfqq->queued[sync]);
1406 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407
Jens Axboe5e705372006-07-13 12:39:25 +02001408 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409
Vivek Goyalf04a6422009-12-03 12:59:40 -05001410 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1411 /*
1412 * Queue will be deleted from service tree when we actually
1413 * expire it later. Right now just remove it from prio tree
1414 * as it is empty.
1415 */
1416 if (cfqq->p_root) {
1417 rb_erase(&cfqq->p_node, cfqq->p_root);
1418 cfqq->p_root = NULL;
1419 }
1420 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421}
1422
Jens Axboe5e705372006-07-13 12:39:25 +02001423static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424{
Jens Axboe5e705372006-07-13 12:39:25 +02001425 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001427 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428
Jens Axboe5380a102006-07-13 12:37:56 +02001429 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430
1431 /*
1432 * looks a little odd, but the first insert might return an alias.
1433 * if that happens, put the alias on the dispatch list
1434 */
Jens Axboe21183b02006-07-13 12:33:14 +02001435 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001436 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001437
1438 if (!cfq_cfqq_on_rr(cfqq))
1439 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001440
1441 /*
1442 * check if this request is a better next-serve candidate
1443 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001444 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001445 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001446
1447 /*
1448 * adjust priority tree position, if ->next_rq changes
1449 */
1450 if (prev != cfqq->next_rq)
1451 cfq_prio_tree_add(cfqd, cfqq);
1452
Jens Axboe5044eed2007-04-25 11:53:48 +02001453 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454}
1455
Jens Axboefebffd62008-01-28 13:19:43 +01001456static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457{
Jens Axboe5380a102006-07-13 12:37:56 +02001458 elv_rb_del(&cfqq->sort_list, rq);
1459 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001460 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1461 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001462 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001463 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001464 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1465 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466}
1467
Jens Axboe206dc692006-03-28 13:03:44 +02001468static struct request *
1469cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470{
Jens Axboe206dc692006-03-28 13:03:44 +02001471 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001472 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001473 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474
Jens Axboe4ac845a2008-01-24 08:44:49 +01001475 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001476 if (!cic)
1477 return NULL;
1478
1479 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001480 if (cfqq) {
1481 sector_t sector = bio->bi_sector + bio_sectors(bio);
1482
Jens Axboe21183b02006-07-13 12:33:14 +02001483 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001484 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 return NULL;
1487}
1488
Jens Axboe165125e2007-07-24 09:28:11 +02001489static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001490{
1491 struct cfq_data *cfqd = q->elevator->elevator_data;
1492
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001493 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001494 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001495 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001496
Tejun Heo5b936292009-05-07 22:24:38 +09001497 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001498}
1499
Jens Axboe165125e2007-07-24 09:28:11 +02001500static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501{
Jens Axboe22e2c502005-06-27 10:55:12 +02001502 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001504 WARN_ON(!cfqd->rq_in_driver);
1505 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001506 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001507 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508}
1509
Jens Axboeb4878f22005-10-20 16:42:29 +02001510static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511{
Jens Axboe5e705372006-07-13 12:39:25 +02001512 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001513
Jens Axboe5e705372006-07-13 12:39:25 +02001514 if (cfqq->next_rq == rq)
1515 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516
Jens Axboeb4878f22005-10-20 16:42:29 +02001517 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001518 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001519
Aaron Carroll45333d52008-08-26 15:52:36 +02001520 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001521 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1522 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02001523 if (rq->cmd_flags & REQ_META) {
Jens Axboe374f84a2006-07-23 01:42:19 +02001524 WARN_ON(!cfqq->meta_pending);
1525 cfqq->meta_pending--;
1526 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527}
1528
Jens Axboe165125e2007-07-24 09:28:11 +02001529static int cfq_merge(struct request_queue *q, struct request **req,
1530 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531{
1532 struct cfq_data *cfqd = q->elevator->elevator_data;
1533 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534
Jens Axboe206dc692006-03-28 13:03:44 +02001535 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001536 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001537 *req = __rq;
1538 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 }
1540
1541 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542}
1543
Jens Axboe165125e2007-07-24 09:28:11 +02001544static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001545 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546{
Jens Axboe21183b02006-07-13 12:33:14 +02001547 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001548 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549
Jens Axboe5e705372006-07-13 12:39:25 +02001550 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552}
1553
Divyesh Shah812d4022010-04-08 21:14:23 -07001554static void cfq_bio_merged(struct request_queue *q, struct request *req,
1555 struct bio *bio)
1556{
Vivek Goyale98ef892010-06-18 10:39:47 -04001557 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1558 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001559}
1560
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561static void
Jens Axboe165125e2007-07-24 09:28:11 +02001562cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 struct request *next)
1564{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001565 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001566 /*
1567 * reposition in fifo if next is older than rq
1568 */
1569 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001570 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001571 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001572 rq_set_fifo_time(rq, rq_fifo_time(next));
1573 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001574
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001575 if (cfqq->next_rq == next)
1576 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001577 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001578 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1579 rq_data_dir(next), rq_is_sync(next));
Jens Axboe22e2c502005-06-27 10:55:12 +02001580}
1581
Jens Axboe165125e2007-07-24 09:28:11 +02001582static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001583 struct bio *bio)
1584{
1585 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001586 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001587 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001588
1589 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001590 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001591 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001592 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001593 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001594
1595 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001596 * Lookup the cfqq that this bio will be queued with. Allow
1597 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001598 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001599 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001600 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001601 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001602
Vasily Tarasov91fac312007-04-25 12:29:51 +02001603 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001604 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001605}
1606
Divyesh Shah812df482010-04-08 21:15:35 -07001607static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1608{
1609 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001610 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001611}
1612
Jens Axboefebffd62008-01-28 13:19:43 +01001613static void __cfq_set_active_queue(struct cfq_data *cfqd,
1614 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001615{
1616 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001617 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1618 cfqd->serving_prio, cfqd->serving_type);
Vivek Goyale98ef892010-06-18 10:39:47 -04001619 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001620 cfqq->slice_start = 0;
1621 cfqq->dispatch_start = jiffies;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001622 cfqq->allocated_slice = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001623 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001624 cfqq->slice_dispatch = 0;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001625 cfqq->nr_sectors = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001626
Jens Axboe2f5cb732009-04-07 08:51:19 +02001627 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +02001628 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001629 cfq_clear_cfqq_must_alloc_slice(cfqq);
1630 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +11001631 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001632
Divyesh Shah812df482010-04-08 21:15:35 -07001633 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001634 }
1635
1636 cfqd->active_queue = cfqq;
1637}
1638
1639/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001640 * current cfqq expired its slice (or was too idle), select new one
1641 */
1642static void
1643__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001644 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001645{
Jens Axboe7b679132008-05-30 12:23:07 +02001646 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1647
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001648 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001649 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001650
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001651 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001652 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001653
1654 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001655 * If this cfqq is shared between multiple processes, check to
1656 * make sure that those processes are still issuing I/Os within
1657 * the mean seek distance. If not, it may be time to break the
1658 * queues apart again.
1659 */
1660 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1661 cfq_mark_cfqq_split_coop(cfqq);
1662
1663 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001664 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001665 */
Jens Axboe7b679132008-05-30 12:23:07 +02001666 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +11001667 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001668 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1669 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001670
Vivek Goyale5ff0822010-04-26 19:25:11 +02001671 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001672
Vivek Goyalf04a6422009-12-03 12:59:40 -05001673 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1674 cfq_del_cfqq_rr(cfqd, cfqq);
1675
Jens Axboeedd75ff2007-04-19 12:03:34 +02001676 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001677
1678 if (cfqq == cfqd->active_queue)
1679 cfqd->active_queue = NULL;
1680
Vivek Goyaldae739e2009-12-03 12:59:45 -05001681 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1682 cfqd->grp_service_tree.active = NULL;
1683
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001684 if (cfqd->active_cic) {
1685 put_io_context(cfqd->active_cic->ioc);
1686 cfqd->active_cic = NULL;
1687 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001688}
1689
Vivek Goyale5ff0822010-04-26 19:25:11 +02001690static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001691{
1692 struct cfq_queue *cfqq = cfqd->active_queue;
1693
1694 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001695 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001696}
1697
Jens Axboe498d3aa2007-04-26 12:54:48 +02001698/*
1699 * Get next queue for service. Unless we have a queue preemption,
1700 * we'll simply select the first cfqq in the service tree.
1701 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001702static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001703{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001704 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001705 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001706 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001707
Vivek Goyalf04a6422009-12-03 12:59:40 -05001708 if (!cfqd->rq_queued)
1709 return NULL;
1710
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001711 /* There is nothing to dispatch */
1712 if (!service_tree)
1713 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001714 if (RB_EMPTY_ROOT(&service_tree->rb))
1715 return NULL;
1716 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001717}
1718
Vivek Goyalf04a6422009-12-03 12:59:40 -05001719static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1720{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001721 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001722 struct cfq_queue *cfqq;
1723 int i, j;
1724 struct cfq_rb_root *st;
1725
1726 if (!cfqd->rq_queued)
1727 return NULL;
1728
Vivek Goyal25fb5162009-12-03 12:59:46 -05001729 cfqg = cfq_get_next_cfqg(cfqd);
1730 if (!cfqg)
1731 return NULL;
1732
Vivek Goyalf04a6422009-12-03 12:59:40 -05001733 for_each_cfqg_st(cfqg, i, j, st)
1734 if ((cfqq = cfq_rb_first(st)) != NULL)
1735 return cfqq;
1736 return NULL;
1737}
1738
Jens Axboe498d3aa2007-04-26 12:54:48 +02001739/*
1740 * Get and set a new active queue for service.
1741 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001742static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1743 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001744{
Jens Axboee00ef792009-11-04 08:54:55 +01001745 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001746 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001747
Jens Axboe22e2c502005-06-27 10:55:12 +02001748 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001749 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001750}
1751
Jens Axboed9e76202007-04-20 14:27:50 +02001752static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1753 struct request *rq)
1754{
Tejun Heo83096eb2009-05-07 22:24:39 +09001755 if (blk_rq_pos(rq) >= cfqd->last_position)
1756 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001757 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001758 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001759}
1760
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001761static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001762 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001763{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001764 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001765}
1766
Jens Axboea36e71f2009-04-15 12:15:11 +02001767static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1768 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001769{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001770 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001771 struct rb_node *parent, *node;
1772 struct cfq_queue *__cfqq;
1773 sector_t sector = cfqd->last_position;
1774
1775 if (RB_EMPTY_ROOT(root))
1776 return NULL;
1777
1778 /*
1779 * First, if we find a request starting at the end of the last
1780 * request, choose it.
1781 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001782 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001783 if (__cfqq)
1784 return __cfqq;
1785
1786 /*
1787 * If the exact sector wasn't found, the parent of the NULL leaf
1788 * will contain the closest sector.
1789 */
1790 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001791 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001792 return __cfqq;
1793
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001794 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001795 node = rb_next(&__cfqq->p_node);
1796 else
1797 node = rb_prev(&__cfqq->p_node);
1798 if (!node)
1799 return NULL;
1800
1801 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001802 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001803 return __cfqq;
1804
1805 return NULL;
1806}
1807
1808/*
1809 * cfqd - obvious
1810 * cur_cfqq - passed in so that we don't decide that the current queue is
1811 * closely cooperating with itself.
1812 *
1813 * So, basically we're assuming that that cur_cfqq has dispatched at least
1814 * one request, and that cfqd->last_position reflects a position on the disk
1815 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1816 * assumption.
1817 */
1818static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001819 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001820{
1821 struct cfq_queue *cfqq;
1822
Divyesh Shah39c01b22010-03-25 15:45:57 +01001823 if (cfq_class_idle(cur_cfqq))
1824 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001825 if (!cfq_cfqq_sync(cur_cfqq))
1826 return NULL;
1827 if (CFQQ_SEEKY(cur_cfqq))
1828 return NULL;
1829
Jens Axboea36e71f2009-04-15 12:15:11 +02001830 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001831 * Don't search priority tree if it's the only queue in the group.
1832 */
1833 if (cur_cfqq->cfqg->nr_cfqq == 1)
1834 return NULL;
1835
1836 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001837 * We should notice if some of the queues are cooperating, eg
1838 * working closely on the same area of the disk. In that case,
1839 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001840 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001841 cfqq = cfqq_close(cfqd, cur_cfqq);
1842 if (!cfqq)
1843 return NULL;
1844
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001845 /* If new queue belongs to different cfq_group, don't choose it */
1846 if (cur_cfqq->cfqg != cfqq->cfqg)
1847 return NULL;
1848
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001849 /*
1850 * It only makes sense to merge sync queues.
1851 */
1852 if (!cfq_cfqq_sync(cfqq))
1853 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001854 if (CFQQ_SEEKY(cfqq))
1855 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001856
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001857 /*
1858 * Do not merge queues of different priority classes
1859 */
1860 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1861 return NULL;
1862
Jens Axboea36e71f2009-04-15 12:15:11 +02001863 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001864}
1865
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001866/*
1867 * Determine whether we should enforce idle window for this queue.
1868 */
1869
1870static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1871{
1872 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001873 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001874
Vivek Goyalf04a6422009-12-03 12:59:40 -05001875 BUG_ON(!service_tree);
1876 BUG_ON(!service_tree->count);
1877
Vivek Goyalb6508c12010-08-23 12:23:33 +02001878 if (!cfqd->cfq_slice_idle)
1879 return false;
1880
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001881 /* We never do for idle class queues. */
1882 if (prio == IDLE_WORKLOAD)
1883 return false;
1884
1885 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001886 if (cfq_cfqq_idle_window(cfqq) &&
1887 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001888 return true;
1889
1890 /*
1891 * Otherwise, we do only if they are the last ones
1892 * in their service tree.
1893 */
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001894 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1895 return 1;
1896 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1897 service_tree->count);
1898 return 0;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001899}
1900
Jens Axboe6d048f52007-04-25 12:44:27 +02001901static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001902{
Jens Axboe17926692007-01-19 11:59:30 +11001903 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001904 struct cfq_io_context *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001905 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001906
Jens Axboea68bbdd2008-09-24 13:03:33 +02001907 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001908 * SSD device without seek penalty, disable idling. But only do so
1909 * for devices that support queuing, otherwise we still have a problem
1910 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001911 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001912 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001913 return;
1914
Jens Axboedd67d052006-06-21 09:36:18 +02001915 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001916 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001917
1918 /*
1919 * idle is disabled, either manually or by past process history
1920 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001921 if (!cfq_should_idle(cfqd, cfqq)) {
1922 /* no queue idling. Check for group idling */
1923 if (cfqd->cfq_group_idle)
1924 group_idle = cfqd->cfq_group_idle;
1925 else
1926 return;
1927 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001928
Jens Axboe22e2c502005-06-27 10:55:12 +02001929 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001930 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001931 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001932 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001933 return;
1934
1935 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001936 * task has exited, don't wait
1937 */
Jens Axboe206dc692006-03-28 13:03:44 +02001938 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001939 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001940 return;
1941
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001942 /*
1943 * If our average think time is larger than the remaining time
1944 * slice, then don't idle. This avoids overrunning the allotted
1945 * time slice.
1946 */
1947 if (sample_valid(cic->ttime_samples) &&
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001948 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1949 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1950 cic->ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001951 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001952 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001953
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001954 /* There are other queues in the group, don't do group idle */
1955 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1956 return;
1957
Jens Axboe3b181522005-06-27 10:56:24 +02001958 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001959
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001960 if (group_idle)
1961 sl = cfqd->cfq_group_idle;
1962 else
1963 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001964
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001965 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001966 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001967 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1968 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969}
1970
Jens Axboe498d3aa2007-04-26 12:54:48 +02001971/*
1972 * Move request from internal lists to the request queue dispatch list.
1973 */
Jens Axboe165125e2007-07-24 09:28:11 +02001974static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001976 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001977 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001978
Jens Axboe7b679132008-05-30 12:23:07 +02001979 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1980
Jeff Moyer06d21882009-09-11 17:08:59 +02001981 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001982 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001983 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001984 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001985 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001986
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001987 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001988 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001989 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02001990 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991}
1992
1993/*
1994 * return expired entry, or NULL to just start from scratch in rbtree
1995 */
Jens Axboefebffd62008-01-28 13:19:43 +01001996static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997{
Jens Axboe30996f42009-10-05 11:03:39 +02001998 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999
Jens Axboe3b181522005-06-27 10:56:24 +02002000 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002002
2003 cfq_mark_cfqq_fifo_expire(cfqq);
2004
Jens Axboe89850f72006-07-22 16:48:31 +02002005 if (list_empty(&cfqq->fifo))
2006 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007
Jens Axboe89850f72006-07-22 16:48:31 +02002008 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002009 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002010 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011
Jens Axboe30996f42009-10-05 11:03:39 +02002012 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002013 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014}
2015
Jens Axboe22e2c502005-06-27 10:55:12 +02002016static inline int
2017cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2018{
2019 const int base_rq = cfqd->cfq_slice_async_rq;
2020
2021 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2022
2023 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2024}
2025
2026/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002027 * Must be called with the queue_lock held.
2028 */
2029static int cfqq_process_refs(struct cfq_queue *cfqq)
2030{
2031 int process_refs, io_refs;
2032
2033 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2034 process_refs = atomic_read(&cfqq->ref) - io_refs;
2035 BUG_ON(process_refs < 0);
2036 return process_refs;
2037}
2038
2039static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2040{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002041 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002042 struct cfq_queue *__cfqq;
2043
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002044 /*
2045 * If there are no process references on the new_cfqq, then it is
2046 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2047 * chain may have dropped their last reference (not just their
2048 * last process reference).
2049 */
2050 if (!cfqq_process_refs(new_cfqq))
2051 return;
2052
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002053 /* Avoid a circular list and skip interim queue merges */
2054 while ((__cfqq = new_cfqq->new_cfqq)) {
2055 if (__cfqq == cfqq)
2056 return;
2057 new_cfqq = __cfqq;
2058 }
2059
2060 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002061 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002062 /*
2063 * If the process for the cfqq has gone away, there is no
2064 * sense in merging the queues.
2065 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002066 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002067 return;
2068
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002069 /*
2070 * Merge in the direction of the lesser amount of work.
2071 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002072 if (new_process_refs >= process_refs) {
2073 cfqq->new_cfqq = new_cfqq;
2074 atomic_add(process_refs, &new_cfqq->ref);
2075 } else {
2076 new_cfqq->new_cfqq = cfqq;
2077 atomic_add(new_process_refs, &cfqq->ref);
2078 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002079}
2080
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002081static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002082 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002083{
2084 struct cfq_queue *queue;
2085 int i;
2086 bool key_valid = false;
2087 unsigned long lowest_key = 0;
2088 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2089
Vivek Goyal65b32a52009-12-16 17:52:59 -05002090 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2091 /* select the one with lowest rb_key */
2092 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002093 if (queue &&
2094 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2095 lowest_key = queue->rb_key;
2096 cur_best = i;
2097 key_valid = true;
2098 }
2099 }
2100
2101 return cur_best;
2102}
2103
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002104static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002105{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002106 unsigned slice;
2107 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002108 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002109 unsigned group_slice;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002110
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002111 if (!cfqg) {
2112 cfqd->serving_prio = IDLE_WORKLOAD;
2113 cfqd->workload_expires = jiffies + 1;
2114 return;
2115 }
2116
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002117 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002118 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002119 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002120 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002121 cfqd->serving_prio = BE_WORKLOAD;
2122 else {
2123 cfqd->serving_prio = IDLE_WORKLOAD;
2124 cfqd->workload_expires = jiffies + 1;
2125 return;
2126 }
2127
2128 /*
2129 * For RT and BE, we have to choose also the type
2130 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2131 * expiration time
2132 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002133 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002134 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002135
2136 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002137 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002138 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002139 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002140 return;
2141
2142 /* otherwise select new workload type */
2143 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002144 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2145 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002146 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002147
2148 /*
2149 * the workload slice is computed as a fraction of target latency
2150 * proportional to the number of queues in that workload, over
2151 * all the queues in the same priority class
2152 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002153 group_slice = cfq_group_slice(cfqd, cfqg);
2154
2155 slice = group_slice * count /
2156 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2157 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002158
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002159 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2160 unsigned int tmp;
2161
2162 /*
2163 * Async queues are currently system wide. Just taking
2164 * proportion of queues with-in same group will lead to higher
2165 * async ratio system wide as generally root group is going
2166 * to have higher weight. A more accurate thing would be to
2167 * calculate system wide asnc/sync ratio.
2168 */
2169 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2170 tmp = tmp/cfqd->busy_queues;
2171 slice = min_t(unsigned, slice, tmp);
2172
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002173 /* async workload slice is scaled down according to
2174 * the sync/async slice ratio. */
2175 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002176 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002177 /* sync workload slice is at least 2 * cfq_slice_idle */
2178 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2179
2180 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002181 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002182 cfqd->workload_expires = jiffies + slice;
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002183 cfqd->noidle_tree_requires_idle = false;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002184}
2185
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002186static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2187{
2188 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002189 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002190
2191 if (RB_EMPTY_ROOT(&st->rb))
2192 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002193 cfqg = cfq_rb_first_group(st);
2194 st->active = &cfqg->rb_node;
2195 update_min_vdisktime(st);
2196 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002197}
2198
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002199static void cfq_choose_cfqg(struct cfq_data *cfqd)
2200{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002201 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2202
2203 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002204
2205 /* Restore the workload type data */
2206 if (cfqg->saved_workload_slice) {
2207 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2208 cfqd->serving_type = cfqg->saved_workload;
2209 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002210 } else
2211 cfqd->workload_expires = jiffies - 1;
2212
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002213 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002214}
2215
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002216/*
Jens Axboe498d3aa2007-04-26 12:54:48 +02002217 * Select a queue for service. If we have a current active queue,
2218 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002219 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002220static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002221{
Jens Axboea36e71f2009-04-15 12:15:11 +02002222 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002223
2224 cfqq = cfqd->active_queue;
2225 if (!cfqq)
2226 goto new_queue;
2227
Vivek Goyalf04a6422009-12-03 12:59:40 -05002228 if (!cfqd->rq_queued)
2229 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002230
2231 /*
2232 * We were waiting for group to get backlogged. Expire the queue
2233 */
2234 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2235 goto expire;
2236
Jens Axboe22e2c502005-06-27 10:55:12 +02002237 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002238 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002239 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002240 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2241 /*
2242 * If slice had not expired at the completion of last request
2243 * we might not have turned on wait_busy flag. Don't expire
2244 * the queue yet. Allow the group to get backlogged.
2245 *
2246 * The very fact that we have used the slice, that means we
2247 * have been idling all along on this queue and it should be
2248 * ok to wait for this request to complete.
2249 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002250 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2251 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2252 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002253 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002254 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002255 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002256 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002257
2258 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002259 * The active queue has requests and isn't expired, allow it to
2260 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002261 */
Jens Axboedd67d052006-06-21 09:36:18 +02002262 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002263 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002264
2265 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002266 * If another queue has a request waiting within our mean seek
2267 * distance, let it run. The expire code will check for close
2268 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002269 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002270 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002271 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002272 if (new_cfqq) {
2273 if (!cfqq->new_cfqq)
2274 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002275 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002276 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002277
2278 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002279 * No requests pending. If the active queue still has requests in
2280 * flight or is idling for a new request, allow either of these
2281 * conditions to happen (or time out) before selecting a new queue.
2282 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002283 if (timer_pending(&cfqd->idle_slice_timer)) {
2284 cfqq = NULL;
2285 goto keep_queue;
2286 }
2287
2288 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2289 cfqq = NULL;
2290 goto keep_queue;
2291 }
2292
2293 /*
2294 * If group idle is enabled and there are requests dispatched from
2295 * this group, wait for requests to complete.
2296 */
2297check_group_idle:
2298 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2299 && cfqq->cfqg->dispatched) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002300 cfqq = NULL;
2301 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002302 }
2303
Jens Axboe3b181522005-06-27 10:56:24 +02002304expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002305 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002306new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002307 /*
2308 * Current queue expired. Check if we have to switch to a new
2309 * service tree
2310 */
2311 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002312 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002313
Jens Axboea36e71f2009-04-15 12:15:11 +02002314 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002315keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002316 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002317}
2318
Jens Axboefebffd62008-01-28 13:19:43 +01002319static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002320{
2321 int dispatched = 0;
2322
2323 while (cfqq->next_rq) {
2324 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2325 dispatched++;
2326 }
2327
2328 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002329
2330 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002331 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002332 return dispatched;
2333}
2334
Jens Axboe498d3aa2007-04-26 12:54:48 +02002335/*
2336 * Drain our current requests. Used for barriers and when switching
2337 * io schedulers on-the-fly.
2338 */
Jens Axboed9e76202007-04-20 14:27:50 +02002339static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002340{
Jens Axboe08717142008-01-28 11:38:15 +01002341 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002342 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002343
Divyesh Shah3440c492010-04-09 09:29:57 +02002344 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002345 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002346 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2347 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002348 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002349 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002350
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002351 BUG_ON(cfqd->busy_queues);
2352
Jeff Moyer69237152009-06-12 15:29:30 +02002353 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002354 return dispatched;
2355}
2356
Shaohua Liabc3c742010-03-01 09:20:54 +01002357static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2358 struct cfq_queue *cfqq)
2359{
2360 /* the queue hasn't finished any request, can't estimate */
2361 if (cfq_cfqq_slice_new(cfqq))
2362 return 1;
2363 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2364 cfqq->slice_end))
2365 return 1;
2366
2367 return 0;
2368}
2369
Jens Axboe0b182d62009-10-06 20:49:37 +02002370static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002371{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002372 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373
Jens Axboe2f5cb732009-04-07 08:51:19 +02002374 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002375 * Drain async requests before we start sync IO
2376 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002377 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002378 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002379
2380 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002381 * If this is an async queue and we have sync IO in flight, let it wait
2382 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002383 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002384 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002385
Shaohua Liabc3c742010-03-01 09:20:54 +01002386 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002387 if (cfq_class_idle(cfqq))
2388 max_dispatch = 1;
2389
2390 /*
2391 * Does this cfqq already have too much IO in flight?
2392 */
2393 if (cfqq->dispatched >= max_dispatch) {
2394 /*
2395 * idle queue must always only have a single IO in flight
2396 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002397 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002398 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002399
Jens Axboe2f5cb732009-04-07 08:51:19 +02002400 /*
2401 * We have other queues, don't allow more IO from this one
2402 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002403 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002404 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002405
Jens Axboe2f5cb732009-04-07 08:51:19 +02002406 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002407 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002408 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002409 if (cfqd->busy_queues == 1)
2410 max_dispatch = -1;
2411 else
2412 /*
2413 * Normally we start throttling cfqq when cfq_quantum/2
2414 * requests have been dispatched. But we can drive
2415 * deeper queue depths at the beginning of slice
2416 * subjected to upper limit of cfq_quantum.
2417 * */
2418 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002419 }
2420
2421 /*
2422 * Async queues must wait a bit before being allowed dispatch.
2423 * We also ramp up the dispatch depth gradually for async IO,
2424 * based on the last sync IO we serviced
2425 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002426 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002427 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002428 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002429
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002430 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002431 if (!depth && !cfqq->dispatched)
2432 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002433 if (depth < max_dispatch)
2434 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 }
2436
Jens Axboe0b182d62009-10-06 20:49:37 +02002437 /*
2438 * If we're below the current max, allow a dispatch
2439 */
2440 return cfqq->dispatched < max_dispatch;
2441}
2442
2443/*
2444 * Dispatch a request from cfqq, moving them to the request queue
2445 * dispatch list.
2446 */
2447static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2448{
2449 struct request *rq;
2450
2451 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2452
2453 if (!cfq_may_dispatch(cfqd, cfqq))
2454 return false;
2455
2456 /*
2457 * follow expired path, else get first next available
2458 */
2459 rq = cfq_check_fifo(cfqq);
2460 if (!rq)
2461 rq = cfqq->next_rq;
2462
2463 /*
2464 * insert request into driver dispatch list
2465 */
2466 cfq_dispatch_insert(cfqd->queue, rq);
2467
2468 if (!cfqd->active_cic) {
2469 struct cfq_io_context *cic = RQ_CIC(rq);
2470
2471 atomic_long_inc(&cic->ioc->refcount);
2472 cfqd->active_cic = cic;
2473 }
2474
2475 return true;
2476}
2477
2478/*
2479 * Find the cfqq that we need to service and move a request from that to the
2480 * dispatch list
2481 */
2482static int cfq_dispatch_requests(struct request_queue *q, int force)
2483{
2484 struct cfq_data *cfqd = q->elevator->elevator_data;
2485 struct cfq_queue *cfqq;
2486
2487 if (!cfqd->busy_queues)
2488 return 0;
2489
2490 if (unlikely(force))
2491 return cfq_forced_dispatch(cfqd);
2492
2493 cfqq = cfq_select_queue(cfqd);
2494 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002495 return 0;
2496
Jens Axboe2f5cb732009-04-07 08:51:19 +02002497 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002498 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002499 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002500 if (!cfq_dispatch_request(cfqd, cfqq))
2501 return 0;
2502
Jens Axboe2f5cb732009-04-07 08:51:19 +02002503 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002504 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002505
2506 /*
2507 * expire an async queue immediately if it has used up its slice. idle
2508 * queue always expire after 1 dispatch round.
2509 */
2510 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2511 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2512 cfq_class_idle(cfqq))) {
2513 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002514 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002515 }
2516
Shan Weib217a902009-09-01 10:06:42 +02002517 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002518 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519}
2520
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521/*
Jens Axboe5e705372006-07-13 12:39:25 +02002522 * task holds one reference to the queue, dropped when task exits. each rq
2523 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002525 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 * queue lock must be held here.
2527 */
2528static void cfq_put_queue(struct cfq_queue *cfqq)
2529{
Jens Axboe22e2c502005-06-27 10:55:12 +02002530 struct cfq_data *cfqd = cfqq->cfqd;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002531 struct cfq_group *cfqg, *orig_cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002532
2533 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534
2535 if (!atomic_dec_and_test(&cfqq->ref))
2536 return;
2537
Jens Axboe7b679132008-05-30 12:23:07 +02002538 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002540 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002541 cfqg = cfqq->cfqg;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002542 orig_cfqg = cfqq->orig_cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002544 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002545 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002546 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002547 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002548
Vivek Goyalf04a6422009-12-03 12:59:40 -05002549 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002551 cfq_put_cfqg(cfqg);
Vivek Goyal878eadd2009-12-07 19:37:15 +01002552 if (orig_cfqg)
2553 cfq_put_cfqg(orig_cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554}
2555
Jens Axboed6de8be2008-05-28 14:46:59 +02002556/*
2557 * Must always be called with the rcu_read_lock() held
2558 */
Jens Axboe07416d22008-05-07 09:17:12 +02002559static void
2560__call_for_each_cic(struct io_context *ioc,
2561 void (*func)(struct io_context *, struct cfq_io_context *))
2562{
2563 struct cfq_io_context *cic;
2564 struct hlist_node *n;
2565
2566 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2567 func(ioc, cic);
2568}
2569
Jens Axboe4ac845a2008-01-24 08:44:49 +01002570/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002571 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002572 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002573static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002574call_for_each_cic(struct io_context *ioc,
2575 void (*func)(struct io_context *, struct cfq_io_context *))
2576{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002577 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002578 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002579 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002580}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002581
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002582static void cfq_cic_free_rcu(struct rcu_head *head)
2583{
2584 struct cfq_io_context *cic;
2585
2586 cic = container_of(head, struct cfq_io_context, rcu_head);
2587
2588 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002589 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002590
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002591 if (ioc_gone) {
2592 /*
2593 * CFQ scheduler is exiting, grab exit lock and check
2594 * the pending io context count. If it hits zero,
2595 * complete ioc_gone and set it back to NULL
2596 */
2597 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002598 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002599 complete(ioc_gone);
2600 ioc_gone = NULL;
2601 }
2602 spin_unlock(&ioc_gone_lock);
2603 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002604}
2605
2606static void cfq_cic_free(struct cfq_io_context *cic)
2607{
2608 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002609}
2610
2611static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2612{
2613 unsigned long flags;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002614 unsigned long dead_key = (unsigned long) cic->key;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002615
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002616 BUG_ON(!(dead_key & CIC_DEAD_KEY));
Jens Axboe4ac845a2008-01-24 08:44:49 +01002617
2618 spin_lock_irqsave(&ioc->lock, flags);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002619 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
Jens Axboeffc4e752008-02-19 10:02:29 +01002620 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002621 spin_unlock_irqrestore(&ioc->lock, flags);
2622
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002623 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002624}
2625
Jens Axboed6de8be2008-05-28 14:46:59 +02002626/*
2627 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2628 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2629 * and ->trim() which is called with the task lock held
2630 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002631static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002633 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002634 * ioc->refcount is zero here, or we are called from elv_unregister(),
2635 * so no more cic's are allowed to be linked into this ioc. So it
2636 * should be ok to iterate over the known list, we will see all cic's
2637 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002638 */
Jens Axboe07416d22008-05-07 09:17:12 +02002639 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640}
2641
Shaohua Lid02a2c02010-05-25 10:16:53 +02002642static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002643{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002644 struct cfq_queue *__cfqq, *next;
2645
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002646 /*
2647 * If this queue was scheduled to merge with another queue, be
2648 * sure to drop the reference taken on that queue (and others in
2649 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2650 */
2651 __cfqq = cfqq->new_cfqq;
2652 while (__cfqq) {
2653 if (__cfqq == cfqq) {
2654 WARN(1, "cfqq->new_cfqq loop detected\n");
2655 break;
2656 }
2657 next = __cfqq->new_cfqq;
2658 cfq_put_queue(__cfqq);
2659 __cfqq = next;
2660 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002661}
2662
2663static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2664{
2665 if (unlikely(cfqq == cfqd->active_queue)) {
2666 __cfq_slice_expired(cfqd, cfqq, 0);
2667 cfq_schedule_dispatch(cfqd);
2668 }
2669
2670 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002671
Jens Axboe89850f72006-07-22 16:48:31 +02002672 cfq_put_queue(cfqq);
2673}
2674
2675static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2676 struct cfq_io_context *cic)
2677{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002678 struct io_context *ioc = cic->ioc;
2679
Jens Axboefc463792006-08-29 09:05:44 +02002680 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002681
2682 /*
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002683 * Make sure dead mark is seen for dead queues
Jens Axboe4ac845a2008-01-24 08:44:49 +01002684 */
Jens Axboefc463792006-08-29 09:05:44 +02002685 smp_wmb();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002686 cic->key = cfqd_dead_key(cfqd);
Jens Axboefc463792006-08-29 09:05:44 +02002687
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002688 if (ioc->ioc_data == cic)
2689 rcu_assign_pointer(ioc->ioc_data, NULL);
2690
Jens Axboeff6657c2009-04-08 10:58:57 +02002691 if (cic->cfqq[BLK_RW_ASYNC]) {
2692 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2693 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002694 }
2695
Jens Axboeff6657c2009-04-08 10:58:57 +02002696 if (cic->cfqq[BLK_RW_SYNC]) {
2697 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2698 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002699 }
Jens Axboe89850f72006-07-22 16:48:31 +02002700}
2701
Jens Axboe4ac845a2008-01-24 08:44:49 +01002702static void cfq_exit_single_io_context(struct io_context *ioc,
2703 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002704{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002705 struct cfq_data *cfqd = cic_to_cfqd(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002706
Jens Axboe89850f72006-07-22 16:48:31 +02002707 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002708 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002709 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002710
Jens Axboe4ac845a2008-01-24 08:44:49 +01002711 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002712
2713 /*
2714 * Ensure we get a fresh copy of the ->key to prevent
2715 * race between exiting task and queue
2716 */
2717 smp_read_barrier_depends();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002718 if (cic->key == cfqd)
Jens Axboe62c1fe92008-12-15 21:19:25 +01002719 __cfq_exit_single_io_context(cfqd, cic);
2720
Jens Axboe4ac845a2008-01-24 08:44:49 +01002721 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002722 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002723}
2724
Jens Axboe498d3aa2007-04-26 12:54:48 +02002725/*
2726 * The process that ioc belongs to has exited, we need to clean up
2727 * and put the internal structures we have that belongs to that process.
2728 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002729static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002731 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732}
2733
Jens Axboe22e2c502005-06-27 10:55:12 +02002734static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002735cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736{
Jens Axboeb5deef92006-07-19 23:39:40 +02002737 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738
Christoph Lameter94f60302007-07-17 04:03:29 -07002739 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2740 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002742 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002743 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002744 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002745 cic->dtor = cfq_free_io_context;
2746 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002747 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 }
2749
2750 return cic;
2751}
2752
Jens Axboefd0928d2008-01-24 08:52:45 +01002753static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002754{
2755 struct task_struct *tsk = current;
2756 int ioprio_class;
2757
Jens Axboe3b181522005-06-27 10:56:24 +02002758 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002759 return;
2760
Jens Axboefd0928d2008-01-24 08:52:45 +01002761 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002762 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002763 default:
2764 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2765 case IOPRIO_CLASS_NONE:
2766 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002767 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002768 */
2769 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002770 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002771 break;
2772 case IOPRIO_CLASS_RT:
2773 cfqq->ioprio = task_ioprio(ioc);
2774 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2775 break;
2776 case IOPRIO_CLASS_BE:
2777 cfqq->ioprio = task_ioprio(ioc);
2778 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2779 break;
2780 case IOPRIO_CLASS_IDLE:
2781 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2782 cfqq->ioprio = 7;
2783 cfq_clear_cfqq_idle_window(cfqq);
2784 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002785 }
2786
2787 /*
2788 * keep track of original prio settings in case we have to temporarily
2789 * elevate the priority of this queue
2790 */
2791 cfqq->org_ioprio = cfqq->ioprio;
2792 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002793 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002794}
2795
Jens Axboefebffd62008-01-28 13:19:43 +01002796static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002797{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002798 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002799 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002800 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002801
Jens Axboecaaa5f92006-06-16 11:23:00 +02002802 if (unlikely(!cfqd))
2803 return;
2804
Jens Axboec1b707d2006-10-30 19:54:23 +01002805 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002806
Jens Axboeff6657c2009-04-08 10:58:57 +02002807 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002808 if (cfqq) {
2809 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002810 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2811 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002812 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002813 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002814 cfq_put_queue(cfqq);
2815 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002816 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002817
Jens Axboeff6657c2009-04-08 10:58:57 +02002818 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002819 if (cfqq)
2820 cfq_mark_cfqq_prio_changed(cfqq);
2821
Jens Axboec1b707d2006-10-30 19:54:23 +01002822 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002823}
2824
Jens Axboefc463792006-08-29 09:05:44 +02002825static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002827 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002828 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829}
2830
Jens Axboed5036d72009-06-26 10:44:34 +02002831static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002832 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002833{
2834 RB_CLEAR_NODE(&cfqq->rb_node);
2835 RB_CLEAR_NODE(&cfqq->p_node);
2836 INIT_LIST_HEAD(&cfqq->fifo);
2837
2838 atomic_set(&cfqq->ref, 0);
2839 cfqq->cfqd = cfqd;
2840
2841 cfq_mark_cfqq_prio_changed(cfqq);
2842
2843 if (is_sync) {
2844 if (!cfq_class_idle(cfqq))
2845 cfq_mark_cfqq_idle_window(cfqq);
2846 cfq_mark_cfqq_sync(cfqq);
2847 }
2848 cfqq->pid = pid;
2849}
2850
Vivek Goyal24610332009-12-03 12:59:51 -05002851#ifdef CONFIG_CFQ_GROUP_IOSCHED
2852static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2853{
2854 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002855 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal24610332009-12-03 12:59:51 -05002856 unsigned long flags;
2857 struct request_queue *q;
2858
2859 if (unlikely(!cfqd))
2860 return;
2861
2862 q = cfqd->queue;
2863
2864 spin_lock_irqsave(q->queue_lock, flags);
2865
2866 if (sync_cfqq) {
2867 /*
2868 * Drop reference to sync queue. A new sync queue will be
2869 * assigned in new group upon arrival of a fresh request.
2870 */
2871 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2872 cic_set_cfqq(cic, NULL, 1);
2873 cfq_put_queue(sync_cfqq);
2874 }
2875
2876 spin_unlock_irqrestore(q->queue_lock, flags);
2877}
2878
2879static void cfq_ioc_set_cgroup(struct io_context *ioc)
2880{
2881 call_for_each_cic(ioc, changed_cgroup);
2882 ioc->cgroup_changed = 0;
2883}
2884#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2885
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002887cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002888 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002891 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002892 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893
2894retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002895 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002896 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002897 /* cic always exists here */
2898 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899
Jens Axboe6118b702009-06-30 09:34:12 +02002900 /*
2901 * Always try a new alloc if we fell back to the OOM cfqq
2902 * originally, since it should just be a temporary situation.
2903 */
2904 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2905 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906 if (new_cfqq) {
2907 cfqq = new_cfqq;
2908 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002909 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002911 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002912 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002913 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002915 if (new_cfqq)
2916 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002917 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002918 cfqq = kmem_cache_alloc_node(cfq_pool,
2919 gfp_mask | __GFP_ZERO,
2920 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002921 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922
Jens Axboe6118b702009-06-30 09:34:12 +02002923 if (cfqq) {
2924 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2925 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002926 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002927 cfq_log_cfqq(cfqd, cfqq, "alloced");
2928 } else
2929 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930 }
2931
2932 if (new_cfqq)
2933 kmem_cache_free(cfq_pool, new_cfqq);
2934
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 return cfqq;
2936}
2937
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002938static struct cfq_queue **
2939cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2940{
Jens Axboefe094d92008-01-31 13:08:54 +01002941 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002942 case IOPRIO_CLASS_RT:
2943 return &cfqd->async_cfqq[0][ioprio];
2944 case IOPRIO_CLASS_BE:
2945 return &cfqd->async_cfqq[1][ioprio];
2946 case IOPRIO_CLASS_IDLE:
2947 return &cfqd->async_idle_cfqq;
2948 default:
2949 BUG();
2950 }
2951}
2952
Jens Axboe15c31be2007-07-10 13:43:25 +02002953static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002954cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002955 gfp_t gfp_mask)
2956{
Jens Axboefd0928d2008-01-24 08:52:45 +01002957 const int ioprio = task_ioprio(ioc);
2958 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002959 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002960 struct cfq_queue *cfqq = NULL;
2961
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002962 if (!is_sync) {
2963 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2964 cfqq = *async_cfqq;
2965 }
2966
Jens Axboe6118b702009-06-30 09:34:12 +02002967 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002968 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002969
2970 /*
2971 * pin the queue now that it's allocated, scheduler exit will prune it
2972 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002973 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02002974 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002975 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002976 }
2977
2978 atomic_inc(&cfqq->ref);
2979 return cfqq;
2980}
2981
Jens Axboe498d3aa2007-04-26 12:54:48 +02002982/*
2983 * We drop cfq io contexts lazily, so we may find a dead one.
2984 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002985static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002986cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2987 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002988{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002989 unsigned long flags;
2990
Jens Axboefc463792006-08-29 09:05:44 +02002991 WARN_ON(!list_empty(&cic->queue_list));
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002992 BUG_ON(cic->key != cfqd_dead_key(cfqd));
Jens Axboe597bc482007-04-24 21:23:53 +02002993
Jens Axboe4ac845a2008-01-24 08:44:49 +01002994 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02002995
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002996 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002997
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002998 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
Jens Axboeffc4e752008-02-19 10:02:29 +01002999 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003000 spin_unlock_irqrestore(&ioc->lock, flags);
3001
3002 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003003}
3004
Jens Axboee2d74ac2006-03-28 08:59:01 +02003005static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01003006cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003007{
Jens Axboee2d74ac2006-03-28 08:59:01 +02003008 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003009 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003010
Vasily Tarasov91fac312007-04-25 12:29:51 +02003011 if (unlikely(!ioc))
3012 return NULL;
3013
Jens Axboed6de8be2008-05-28 14:46:59 +02003014 rcu_read_lock();
3015
Jens Axboe597bc482007-04-24 21:23:53 +02003016 /*
3017 * we maintain a last-hit cache, to avoid browsing over the tree
3018 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003019 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02003020 if (cic && cic->key == cfqd) {
3021 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02003022 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003023 }
Jens Axboe597bc482007-04-24 21:23:53 +02003024
Jens Axboe4ac845a2008-01-24 08:44:49 +01003025 do {
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003026 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003027 rcu_read_unlock();
3028 if (!cic)
3029 break;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003030 if (unlikely(cic->key != cfqd)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01003031 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003032 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01003033 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003034 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003035
Jens Axboed6de8be2008-05-28 14:46:59 +02003036 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003037 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003038 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003039 break;
3040 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003041
Jens Axboe4ac845a2008-01-24 08:44:49 +01003042 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003043}
3044
Jens Axboe4ac845a2008-01-24 08:44:49 +01003045/*
3046 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3047 * the process specific cfq io context when entered from the block layer.
3048 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3049 */
Jens Axboefebffd62008-01-28 13:19:43 +01003050static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3051 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003052{
Jens Axboe0261d682006-10-30 19:07:48 +01003053 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003054 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003055
Jens Axboe4ac845a2008-01-24 08:44:49 +01003056 ret = radix_tree_preload(gfp_mask);
3057 if (!ret) {
3058 cic->ioc = ioc;
3059 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003060
Jens Axboe4ac845a2008-01-24 08:44:49 +01003061 spin_lock_irqsave(&ioc->lock, flags);
3062 ret = radix_tree_insert(&ioc->radix_root,
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003063 cfqd->cic_index, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01003064 if (!ret)
3065 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003066 spin_unlock_irqrestore(&ioc->lock, flags);
3067
3068 radix_tree_preload_end();
3069
3070 if (!ret) {
3071 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3072 list_add(&cic->queue_list, &cfqd->cic_list);
3073 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003074 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003075 }
3076
Jens Axboe4ac845a2008-01-24 08:44:49 +01003077 if (ret)
3078 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02003079
Jens Axboe4ac845a2008-01-24 08:44:49 +01003080 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003081}
3082
Jens Axboe22e2c502005-06-27 10:55:12 +02003083/*
3084 * Setup general io context and cfq io context. There can be several cfq
3085 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02003086 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02003087 */
3088static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02003089cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090{
Jens Axboe22e2c502005-06-27 10:55:12 +02003091 struct io_context *ioc = NULL;
3092 struct cfq_io_context *cic;
3093
3094 might_sleep_if(gfp_mask & __GFP_WAIT);
3095
Jens Axboeb5deef92006-07-19 23:39:40 +02003096 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02003097 if (!ioc)
3098 return NULL;
3099
Jens Axboe4ac845a2008-01-24 08:44:49 +01003100 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003101 if (cic)
3102 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02003103
Jens Axboee2d74ac2006-03-28 08:59:01 +02003104 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3105 if (cic == NULL)
3106 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02003107
Jens Axboe4ac845a2008-01-24 08:44:49 +01003108 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3109 goto err_free;
3110
Jens Axboe22e2c502005-06-27 10:55:12 +02003111out:
Jens Axboefc463792006-08-29 09:05:44 +02003112 smp_read_barrier_depends();
3113 if (unlikely(ioc->ioprio_changed))
3114 cfq_ioc_set_ioprio(ioc);
3115
Vivek Goyal24610332009-12-03 12:59:51 -05003116#ifdef CONFIG_CFQ_GROUP_IOSCHED
3117 if (unlikely(ioc->cgroup_changed))
3118 cfq_ioc_set_cgroup(ioc);
3119#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003120 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003121err_free:
3122 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003123err:
3124 put_io_context(ioc);
3125 return NULL;
3126}
3127
3128static void
3129cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3130{
Jens Axboeaaf12282007-01-19 11:30:16 +11003131 unsigned long elapsed = jiffies - cic->last_end_request;
3132 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003133
3134 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3135 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3136 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3137}
3138
Jens Axboe206dc692006-03-28 13:03:44 +02003139static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003140cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003141 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003142{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003143 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003144 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003145 if (cfqq->last_request_pos) {
3146 if (cfqq->last_request_pos < blk_rq_pos(rq))
3147 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3148 else
3149 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3150 }
Jens Axboe206dc692006-03-28 13:03:44 +02003151
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003152 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003153 if (blk_queue_nonrot(cfqd->queue))
3154 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3155 else
3156 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003157}
Jens Axboe22e2c502005-06-27 10:55:12 +02003158
3159/*
3160 * Disable idle window if the process thinks too long or seeks so much that
3161 * it doesn't matter
3162 */
3163static void
3164cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3165 struct cfq_io_context *cic)
3166{
Jens Axboe7b679132008-05-30 12:23:07 +02003167 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003168
Jens Axboe08717142008-01-28 11:38:15 +01003169 /*
3170 * Don't idle for async or idle io prio class
3171 */
3172 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003173 return;
3174
Jens Axboec265a7f2008-06-26 13:49:33 +02003175 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003176
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003177 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3178 cfq_mark_cfqq_deep(cfqq);
3179
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01003180 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003181 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003182 enable_idle = 0;
3183 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003184 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003185 enable_idle = 0;
3186 else
3187 enable_idle = 1;
3188 }
3189
Jens Axboe7b679132008-05-30 12:23:07 +02003190 if (old_idle != enable_idle) {
3191 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3192 if (enable_idle)
3193 cfq_mark_cfqq_idle_window(cfqq);
3194 else
3195 cfq_clear_cfqq_idle_window(cfqq);
3196 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003197}
3198
Jens Axboe22e2c502005-06-27 10:55:12 +02003199/*
3200 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3201 * no or if we aren't sure, a 1 will cause a preempt.
3202 */
Jens Axboea6151c32009-10-07 20:02:57 +02003203static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003204cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003205 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003206{
Jens Axboe6d048f52007-04-25 12:44:27 +02003207 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003208
Jens Axboe6d048f52007-04-25 12:44:27 +02003209 cfqq = cfqd->active_queue;
3210 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003211 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003212
Jens Axboe6d048f52007-04-25 12:44:27 +02003213 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003214 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003215
3216 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003217 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003218
Jens Axboe22e2c502005-06-27 10:55:12 +02003219 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003220 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3221 */
3222 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3223 return false;
3224
3225 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003226 * if the new request is sync, but the currently running queue is
3227 * not, let the sync request have priority.
3228 */
Jens Axboe5e705372006-07-13 12:39:25 +02003229 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003230 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003231
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003232 if (new_cfqq->cfqg != cfqq->cfqg)
3233 return false;
3234
3235 if (cfq_slice_used(cfqq))
3236 return true;
3237
3238 /* Allow preemption only if we are idling on sync-noidle tree */
3239 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3240 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3241 new_cfqq->service_tree->count == 2 &&
3242 RB_EMPTY_ROOT(&cfqq->sort_list))
3243 return true;
3244
Jens Axboe374f84a2006-07-23 01:42:19 +02003245 /*
3246 * So both queues are sync. Let the new request get disk time if
3247 * it's a metadata request and the current queue is doing regular IO.
3248 */
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003249 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003250 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003251
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003252 /*
3253 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3254 */
3255 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003256 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003257
Jens Axboe1e3335d2007-02-14 19:59:49 +01003258 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003259 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003260
3261 /*
3262 * if this request is as-good as one we would expect from the
3263 * current cfqq, let it preempt
3264 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003265 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003266 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003267
Jens Axboea6151c32009-10-07 20:02:57 +02003268 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003269}
3270
3271/*
3272 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3273 * let it have half of its nominal slice.
3274 */
3275static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3276{
Jens Axboe7b679132008-05-30 12:23:07 +02003277 cfq_log_cfqq(cfqd, cfqq, "preempt");
Vivek Goyale5ff0822010-04-26 19:25:11 +02003278 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003279
Jens Axboebf572252006-07-19 20:29:12 +02003280 /*
3281 * Put the new queue at the front of the of the current list,
3282 * so we know that it will be selected next.
3283 */
3284 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003285
3286 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02003287
Jens Axboe44f7c162007-01-19 11:51:58 +11003288 cfqq->slice_end = 0;
3289 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003290}
3291
3292/*
Jens Axboe5e705372006-07-13 12:39:25 +02003293 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003294 * something we should do about it
3295 */
3296static void
Jens Axboe5e705372006-07-13 12:39:25 +02003297cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3298 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003299{
Jens Axboe5e705372006-07-13 12:39:25 +02003300 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003301
Aaron Carroll45333d52008-08-26 15:52:36 +02003302 cfqd->rq_queued++;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003303 if (rq->cmd_flags & REQ_META)
Jens Axboe374f84a2006-07-23 01:42:19 +02003304 cfqq->meta_pending++;
3305
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003306 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003307 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003308 cfq_update_idle_window(cfqd, cfqq, cic);
3309
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003310 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003311
3312 if (cfqq == cfqd->active_queue) {
3313 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003314 * Remember that we saw a request from this process, but
3315 * don't start queuing just yet. Otherwise we risk seeing lots
3316 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003317 * and merging. If the request is already larger than a single
3318 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003319 * merging is already done. Ditto for a busy system that
3320 * has other work pending, don't risk delaying until the
3321 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003322 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003323 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003324 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3325 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003326 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003327 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalbf791932009-12-03 12:59:37 -05003328 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003329 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003330 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003331 &cfqq->cfqg->blkg);
Vivek Goyalbf791932009-12-03 12:59:37 -05003332 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003333 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003334 }
Jens Axboe5e705372006-07-13 12:39:25 +02003335 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003336 /*
3337 * not the active queue - expire current slice if it is
3338 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003339 * has some old slice time left and is of higher priority or
3340 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003341 */
3342 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09003343 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003344 }
3345}
3346
Jens Axboe165125e2007-07-24 09:28:11 +02003347static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003348{
Jens Axboeb4878f22005-10-20 16:42:29 +02003349 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003350 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003351
Jens Axboe7b679132008-05-30 12:23:07 +02003352 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003353 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354
Jens Axboe30996f42009-10-05 11:03:39 +02003355 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003356 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003357 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003358 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003359 &cfqd->serving_group->blkg, rq_data_dir(rq),
3360 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003361 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362}
3363
Aaron Carroll45333d52008-08-26 15:52:36 +02003364/*
3365 * Update hw_tag based on peak queue depth over 50 samples under
3366 * sufficient load.
3367 */
3368static void cfq_update_hw_tag(struct cfq_data *cfqd)
3369{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003370 struct cfq_queue *cfqq = cfqd->active_queue;
3371
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003372 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3373 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003374
3375 if (cfqd->hw_tag == 1)
3376 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003377
3378 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003379 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003380 return;
3381
Shaohua Li1a1238a2009-10-27 08:46:23 +01003382 /*
3383 * If active queue hasn't enough requests and can idle, cfq might not
3384 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3385 * case
3386 */
3387 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3388 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003389 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003390 return;
3391
Aaron Carroll45333d52008-08-26 15:52:36 +02003392 if (cfqd->hw_tag_samples++ < 50)
3393 return;
3394
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003395 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003396 cfqd->hw_tag = 1;
3397 else
3398 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003399}
3400
Vivek Goyal7667aa02009-12-08 17:52:58 -05003401static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3402{
3403 struct cfq_io_context *cic = cfqd->active_cic;
3404
3405 /* If there are other queues in the group, don't wait */
3406 if (cfqq->cfqg->nr_cfqq > 1)
3407 return false;
3408
3409 if (cfq_slice_used(cfqq))
3410 return true;
3411
3412 /* if slice left is less than think time, wait busy */
3413 if (cic && sample_valid(cic->ttime_samples)
3414 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3415 return true;
3416
3417 /*
3418 * If think times is less than a jiffy than ttime_mean=0 and above
3419 * will not be true. It might happen that slice has not expired yet
3420 * but will expire soon (4-5 ns) during select_queue(). To cover the
3421 * case where think time is less than a jiffy, mark the queue wait
3422 * busy if only 1 jiffy is left in the slice.
3423 */
3424 if (cfqq->slice_end - jiffies == 1)
3425 return true;
3426
3427 return false;
3428}
3429
Jens Axboe165125e2007-07-24 09:28:11 +02003430static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431{
Jens Axboe5e705372006-07-13 12:39:25 +02003432 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003433 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003434 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003435 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436
Jens Axboeb4878f22005-10-20 16:42:29 +02003437 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003438 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3439 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440
Aaron Carroll45333d52008-08-26 15:52:36 +02003441 cfq_update_hw_tag(cfqd);
3442
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003443 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003444 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003445 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003446 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003447 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003448 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3449 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3450 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003452 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003453
Vivek Goyal365722b2009-10-03 15:21:27 +02003454 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003455 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003456 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3457 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003458 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003459
3460 /*
3461 * If this is the active queue, check if it needs to be expired,
3462 * or if we want to idle in case it has no pending requests.
3463 */
3464 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003465 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3466
Jens Axboe44f7c162007-01-19 11:51:58 +11003467 if (cfq_cfqq_slice_new(cfqq)) {
3468 cfq_set_prio_slice(cfqd, cfqq);
3469 cfq_clear_cfqq_slice_new(cfqq);
3470 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003471
3472 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003473 * Should we wait for next request to come in before we expire
3474 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003475 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003476 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003477 unsigned long extend_sl = cfqd->cfq_slice_idle;
3478 if (!cfqd->cfq_slice_idle)
3479 extend_sl = cfqd->cfq_group_idle;
3480 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003481 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003482 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003483 }
3484
Jens Axboea36e71f2009-04-15 12:15:11 +02003485 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003486 * Idling is not enabled on:
3487 * - expired queues
3488 * - idle-priority queues
3489 * - async queues
3490 * - queues with still some requests queued
3491 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003492 */
Jens Axboe08717142008-01-28 11:38:15 +01003493 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003494 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003495 else if (sync && cfqq_empty &&
3496 !cfq_close_cooperator(cfqd, cfqq)) {
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003497 cfqd->noidle_tree_requires_idle |=
3498 !(rq->cmd_flags & REQ_NOIDLE);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003499 /*
3500 * Idling is enabled for SYNC_WORKLOAD.
3501 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003502 * only if we processed at least one !REQ_NOIDLE request
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003503 */
3504 if (cfqd->serving_type == SYNC_WORKLOAD
Vivek Goyalc04645e2009-12-03 12:59:56 -05003505 || cfqd->noidle_tree_requires_idle
3506 || cfqq->cfqg->nr_cfqq == 1)
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003507 cfq_arm_slice_timer(cfqd);
3508 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003509 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003510
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003511 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003512 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513}
3514
Jens Axboe22e2c502005-06-27 10:55:12 +02003515/*
3516 * we temporarily boost lower priority queues if they are holding fs exclusive
3517 * resources. they are boosted to normal prio (CLASS_BE/4)
3518 */
3519static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520{
Jens Axboe22e2c502005-06-27 10:55:12 +02003521 if (has_fs_excl()) {
3522 /*
3523 * boost idle prio on transactions that would lock out other
3524 * users of the filesystem
3525 */
3526 if (cfq_class_idle(cfqq))
3527 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3528 if (cfqq->ioprio > IOPRIO_NORM)
3529 cfqq->ioprio = IOPRIO_NORM;
3530 } else {
3531 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003532 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003533 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003534 cfqq->ioprio_class = cfqq->org_ioprio_class;
3535 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003537}
3538
Jens Axboe89850f72006-07-22 16:48:31 +02003539static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003540{
Jens Axboe1b379d82009-08-11 08:26:11 +02003541 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003542 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003543 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003544 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003545
3546 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003547}
3548
Jens Axboe165125e2007-07-24 09:28:11 +02003549static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003550{
3551 struct cfq_data *cfqd = q->elevator->elevator_data;
3552 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003553 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003554 struct cfq_queue *cfqq;
3555
3556 /*
3557 * don't force setup of a queue from here, as a call to may_queue
3558 * does not necessarily imply that a request actually will be queued.
3559 * so just lookup a possibly existing queue, or return 'may queue'
3560 * if that fails
3561 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003562 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003563 if (!cic)
3564 return ELV_MQUEUE_MAY;
3565
Jens Axboeb0b78f82009-04-08 10:56:08 +02003566 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003567 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003568 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003569 cfq_prio_boost(cfqq);
3570
Jens Axboe89850f72006-07-22 16:48:31 +02003571 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003572 }
3573
3574 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575}
3576
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577/*
3578 * queue lock held here
3579 */
Jens Axboebb37b942006-12-01 10:42:33 +01003580static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581{
Jens Axboe5e705372006-07-13 12:39:25 +02003582 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583
Jens Axboe5e705372006-07-13 12:39:25 +02003584 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003585 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586
Jens Axboe22e2c502005-06-27 10:55:12 +02003587 BUG_ON(!cfqq->allocated[rw]);
3588 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589
Jens Axboe5e705372006-07-13 12:39:25 +02003590 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02003593 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003595 /* Put down rq reference on cfqg */
3596 cfq_put_cfqg(RQ_CFQG(rq));
3597 rq->elevator_private3 = NULL;
3598
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599 cfq_put_queue(cfqq);
3600 }
3601}
3602
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003603static struct cfq_queue *
3604cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3605 struct cfq_queue *cfqq)
3606{
3607 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3608 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003609 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003610 cfq_put_queue(cfqq);
3611 return cic_to_cfqq(cic, 1);
3612}
3613
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003614/*
3615 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3616 * was the last process referring to said cfqq.
3617 */
3618static struct cfq_queue *
3619split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3620{
3621 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003622 cfqq->pid = current->pid;
3623 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003624 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003625 return cfqq;
3626 }
3627
3628 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003629
3630 cfq_put_cooperator(cfqq);
3631
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003632 cfq_put_queue(cfqq);
3633 return NULL;
3634}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003636 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003637 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003638static int
Jens Axboe165125e2007-07-24 09:28:11 +02003639cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003640{
3641 struct cfq_data *cfqd = q->elevator->elevator_data;
3642 struct cfq_io_context *cic;
3643 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003644 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003645 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646 unsigned long flags;
3647
3648 might_sleep_if(gfp_mask & __GFP_WAIT);
3649
Jens Axboee2d74ac2006-03-28 08:59:01 +02003650 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003651
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652 spin_lock_irqsave(q->queue_lock, flags);
3653
Jens Axboe22e2c502005-06-27 10:55:12 +02003654 if (!cic)
3655 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003657new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003658 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003659 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003660 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003661 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003662 } else {
3663 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003664 * If the queue was seeky for too long, break it apart.
3665 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003666 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003667 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3668 cfqq = split_cfqq(cic, cfqq);
3669 if (!cfqq)
3670 goto new_queue;
3671 }
3672
3673 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003674 * Check to see if this queue is scheduled to merge with
3675 * another, closely cooperating queue. The merging of
3676 * queues happens here as it must be done in process context.
3677 * The reference on new_cfqq was taken in merge_cfqqs.
3678 */
3679 if (cfqq->new_cfqq)
3680 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003681 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682
3683 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02003684 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02003685
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686 spin_unlock_irqrestore(q->queue_lock, flags);
3687
Jens Axboe5e705372006-07-13 12:39:25 +02003688 rq->elevator_private = cic;
3689 rq->elevator_private2 = cfqq;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003690 rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
Jens Axboe5e705372006-07-13 12:39:25 +02003691 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003692
Jens Axboe22e2c502005-06-27 10:55:12 +02003693queue_fail:
3694 if (cic)
3695 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003696
Jens Axboe23e018a2009-10-05 08:52:35 +02003697 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003699 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700 return 1;
3701}
3702
David Howells65f27f32006-11-22 14:55:48 +00003703static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003704{
David Howells65f27f32006-11-22 14:55:48 +00003705 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003706 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003707 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003708
Jens Axboe40bb54d2009-04-15 12:11:10 +02003709 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09003710 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003711 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003712}
3713
3714/*
3715 * Timer running if the active_queue is currently idling inside its time slice
3716 */
3717static void cfq_idle_slice_timer(unsigned long data)
3718{
3719 struct cfq_data *cfqd = (struct cfq_data *) data;
3720 struct cfq_queue *cfqq;
3721 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003722 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003723
Jens Axboe7b679132008-05-30 12:23:07 +02003724 cfq_log(cfqd, "idle timer fired");
3725
Jens Axboe22e2c502005-06-27 10:55:12 +02003726 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3727
Jens Axboefe094d92008-01-31 13:08:54 +01003728 cfqq = cfqd->active_queue;
3729 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003730 timed_out = 0;
3731
Jens Axboe22e2c502005-06-27 10:55:12 +02003732 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003733 * We saw a request before the queue expired, let it through
3734 */
3735 if (cfq_cfqq_must_dispatch(cfqq))
3736 goto out_kick;
3737
3738 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003739 * expired
3740 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003741 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003742 goto expire;
3743
3744 /*
3745 * only expire and reinvoke request handler, if there are
3746 * other queues with pending requests
3747 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003748 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003749 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003750
3751 /*
3752 * not expired and it has a request pending, let it dispatch
3753 */
Jens Axboe75e50982009-04-07 08:56:14 +02003754 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003755 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003756
3757 /*
3758 * Queue depth flag is reset only when the idle didn't succeed
3759 */
3760 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003761 }
3762expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003763 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003764out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003765 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003766out_cont:
3767 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3768}
3769
Jens Axboe3b181522005-06-27 10:56:24 +02003770static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3771{
3772 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003773 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003774}
Jens Axboe22e2c502005-06-27 10:55:12 +02003775
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003776static void cfq_put_async_queues(struct cfq_data *cfqd)
3777{
3778 int i;
3779
3780 for (i = 0; i < IOPRIO_BE_NR; i++) {
3781 if (cfqd->async_cfqq[0][i])
3782 cfq_put_queue(cfqd->async_cfqq[0][i]);
3783 if (cfqd->async_cfqq[1][i])
3784 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003785 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003786
3787 if (cfqd->async_idle_cfqq)
3788 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003789}
3790
Jens Axboebb729bc2009-12-06 09:54:19 +01003791static void cfq_cfqd_free(struct rcu_head *head)
3792{
3793 kfree(container_of(head, struct cfq_data, rcu));
3794}
3795
Jens Axboeb374d182008-10-31 10:05:07 +01003796static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797{
Jens Axboe22e2c502005-06-27 10:55:12 +02003798 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003799 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003800
Jens Axboe3b181522005-06-27 10:56:24 +02003801 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003802
Al Virod9ff4182006-03-18 13:51:22 -05003803 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003804
Al Virod9ff4182006-03-18 13:51:22 -05003805 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003806 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003807
3808 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003809 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3810 struct cfq_io_context,
3811 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003812
3813 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003814 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003815
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003816 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003817 cfq_release_cfq_groups(cfqd);
Vivek Goyale98ef892010-06-18 10:39:47 -04003818 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003819
Al Virod9ff4182006-03-18 13:51:22 -05003820 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003821
3822 cfq_shutdown_timer_wq(cfqd);
3823
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003824 spin_lock(&cic_index_lock);
3825 ida_remove(&cic_index_ida, cfqd->cic_index);
3826 spin_unlock(&cic_index_lock);
3827
Vivek Goyalb1c35762009-12-03 12:59:47 -05003828 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003829 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830}
3831
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003832static int cfq_alloc_cic_index(void)
3833{
3834 int index, error;
3835
3836 do {
3837 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3838 return -ENOMEM;
3839
3840 spin_lock(&cic_index_lock);
3841 error = ida_get_new(&cic_index_ida, &index);
3842 spin_unlock(&cic_index_lock);
3843 if (error && error != -EAGAIN)
3844 return error;
3845 } while (error);
3846
3847 return index;
3848}
3849
Jens Axboe165125e2007-07-24 09:28:11 +02003850static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851{
3852 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003853 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003854 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003855 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003857 i = cfq_alloc_cic_index();
3858 if (i < 0)
3859 return NULL;
3860
Christoph Lameter94f60302007-07-17 04:03:29 -07003861 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003863 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003865 cfqd->cic_index = i;
3866
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003867 /* Init root service tree */
3868 cfqd->grp_service_tree = CFQ_RB_ROOT;
3869
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003870 /* Init root group */
3871 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003872 for_each_cfqg_st(cfqg, i, j, st)
3873 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003874 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003875
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003876 /* Give preference to root group over other groups */
3877 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3878
Vivek Goyal25fb5162009-12-03 12:59:46 -05003879#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003880 /*
3881 * Take a reference to root group which we never drop. This is just
3882 * to make sure that cfq_put_cfqg() does not try to kfree root group
3883 */
3884 atomic_set(&cfqg->ref, 1);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003885 rcu_read_lock();
Vivek Goyale98ef892010-06-18 10:39:47 -04003886 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3887 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003888 rcu_read_unlock();
Vivek Goyal25fb5162009-12-03 12:59:46 -05003889#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003890 /*
3891 * Not strictly needed (since RB_ROOT just clears the node and we
3892 * zeroed cfqd on alloc), but better be safe in case someone decides
3893 * to add magic to the rb code
3894 */
3895 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3896 cfqd->prio_trees[i] = RB_ROOT;
3897
Jens Axboe6118b702009-06-30 09:34:12 +02003898 /*
3899 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3900 * Grab a permanent reference to it, so that the normal code flow
3901 * will not attempt to free it.
3902 */
3903 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3904 atomic_inc(&cfqd->oom_cfqq.ref);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003905 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003906
Al Virod9ff4182006-03-18 13:51:22 -05003907 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910
Jens Axboe22e2c502005-06-27 10:55:12 +02003911 init_timer(&cfqd->idle_slice_timer);
3912 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3913 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3914
Jens Axboe23e018a2009-10-05 08:52:35 +02003915 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003916
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003918 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3919 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920 cfqd->cfq_back_max = cfq_back_max;
3921 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003922 cfqd->cfq_slice[0] = cfq_slice_async;
3923 cfqd->cfq_slice[1] = cfq_slice_sync;
3924 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3925 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003926 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003927 cfqd->cfq_latency = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05003928 cfqd->cfq_group_isolation = 0;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003929 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003930 /*
3931 * we optimistically start assuming sync ops weren't delayed in last
3932 * second, in order to have larger depth for async operations.
3933 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003934 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003935 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936}
3937
3938static void cfq_slab_kill(void)
3939{
Jens Axboed6de8be2008-05-28 14:46:59 +02003940 /*
3941 * Caller already ensured that pending RCU callbacks are completed,
3942 * so we should have no busy allocations at this point.
3943 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 if (cfq_pool)
3945 kmem_cache_destroy(cfq_pool);
3946 if (cfq_ioc_pool)
3947 kmem_cache_destroy(cfq_ioc_pool);
3948}
3949
3950static int __init cfq_slab_setup(void)
3951{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003952 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953 if (!cfq_pool)
3954 goto fail;
3955
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003956 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 if (!cfq_ioc_pool)
3958 goto fail;
3959
3960 return 0;
3961fail:
3962 cfq_slab_kill();
3963 return -ENOMEM;
3964}
3965
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966/*
3967 * sysfs parts below -->
3968 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969static ssize_t
3970cfq_var_show(unsigned int var, char *page)
3971{
3972 return sprintf(page, "%d\n", var);
3973}
3974
3975static ssize_t
3976cfq_var_store(unsigned int *var, const char *page, size_t count)
3977{
3978 char *p = (char *) page;
3979
3980 *var = simple_strtoul(p, &p, 10);
3981 return count;
3982}
3983
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003985static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003987 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988 unsigned int __data = __VAR; \
3989 if (__CONV) \
3990 __data = jiffies_to_msecs(__data); \
3991 return cfq_var_show(__data, (page)); \
3992}
3993SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003994SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3995SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003996SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3997SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003998SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003999SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004000SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4001SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4002SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004003SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05004004SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005#undef SHOW_FUNCTION
4006
4007#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004008static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004010 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011 unsigned int __data; \
4012 int ret = cfq_var_store(&__data, (page), count); \
4013 if (__data < (MIN)) \
4014 __data = (MIN); \
4015 else if (__data > (MAX)) \
4016 __data = (MAX); \
4017 if (__CONV) \
4018 *(__PTR) = msecs_to_jiffies(__data); \
4019 else \
4020 *(__PTR) = __data; \
4021 return ret; \
4022}
4023STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004024STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4025 UINT_MAX, 1);
4026STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4027 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004028STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004029STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4030 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004031STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004032STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004033STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4034STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004035STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4036 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004037STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05004038STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039#undef STORE_FUNCTION
4040
Al Viroe572ec72006-03-18 22:27:18 -05004041#define CFQ_ATTR(name) \
4042 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004043
Al Viroe572ec72006-03-18 22:27:18 -05004044static struct elv_fs_entry cfq_attrs[] = {
4045 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004046 CFQ_ATTR(fifo_expire_sync),
4047 CFQ_ATTR(fifo_expire_async),
4048 CFQ_ATTR(back_seek_max),
4049 CFQ_ATTR(back_seek_penalty),
4050 CFQ_ATTR(slice_sync),
4051 CFQ_ATTR(slice_async),
4052 CFQ_ATTR(slice_async_rq),
4053 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004054 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004055 CFQ_ATTR(low_latency),
Vivek Goyalae30c282009-12-03 12:59:55 -05004056 CFQ_ATTR(group_isolation),
Al Viroe572ec72006-03-18 22:27:18 -05004057 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058};
4059
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060static struct elevator_type iosched_cfq = {
4061 .ops = {
4062 .elevator_merge_fn = cfq_merge,
4063 .elevator_merged_fn = cfq_merged_request,
4064 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004065 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004066 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004067 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004069 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004070 .elevator_deactivate_req_fn = cfq_deactivate_request,
4071 .elevator_queue_empty_fn = cfq_queue_empty,
4072 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004073 .elevator_former_req_fn = elv_rb_former_request,
4074 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075 .elevator_set_req_fn = cfq_set_request,
4076 .elevator_put_req_fn = cfq_put_request,
4077 .elevator_may_queue_fn = cfq_may_queue,
4078 .elevator_init_fn = cfq_init_queue,
4079 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02004080 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081 },
Al Viro3d1ab402006-03-18 18:35:43 -05004082 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083 .elevator_name = "cfq",
4084 .elevator_owner = THIS_MODULE,
4085};
4086
Vivek Goyal3e252062009-12-04 10:36:42 -05004087#ifdef CONFIG_CFQ_GROUP_IOSCHED
4088static struct blkio_policy_type blkio_policy_cfq = {
4089 .ops = {
4090 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4091 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4092 },
4093};
4094#else
4095static struct blkio_policy_type blkio_policy_cfq;
4096#endif
4097
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098static int __init cfq_init(void)
4099{
Jens Axboe22e2c502005-06-27 10:55:12 +02004100 /*
4101 * could be 0 on HZ < 1000 setups
4102 */
4103 if (!cfq_slice_async)
4104 cfq_slice_async = 1;
4105 if (!cfq_slice_idle)
4106 cfq_slice_idle = 1;
4107
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004108#ifdef CONFIG_CFQ_GROUP_IOSCHED
4109 if (!cfq_group_idle)
4110 cfq_group_idle = 1;
4111#else
4112 cfq_group_idle = 0;
4113#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114 if (cfq_slab_setup())
4115 return -ENOMEM;
4116
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004117 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05004118 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004120 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121}
4122
4123static void __exit cfq_exit(void)
4124{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07004125 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05004126 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05004128 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02004129 /* ioc_gone's update must be visible before reading ioc_count */
4130 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02004131
4132 /*
4133 * this also protects us from entering cfq_slab_kill() with
4134 * pending RCU callbacks
4135 */
Tejun Heo245b2e72009-06-24 15:13:48 +09004136 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02004137 wait_for_completion(&all_gone);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004138 ida_destroy(&cic_index_ida);
Christoph Hellwig83521d32005-10-30 15:01:39 -08004139 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140}
4141
4142module_init(cfq_init);
4143module_exit(cfq_exit);
4144
4145MODULE_AUTHOR("Jens Axboe");
4146MODULE_LICENSE("GPL");
4147MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");