blob: 5ff4f4850e717ddb319423e9678e0e44cd7f265c [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 Goyal25bc6b02009-12-03 12:59:43 -050017#include "blk-cgroup.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;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +010033static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
34static const int cfq_hist_divisor = 4;
Jens Axboe22e2c502005-06-27 10:55:12 +020035
Jens Axboed9e76202007-04-20 14:27:50 +020036/*
Jens Axboe08717142008-01-28 11:38:15 +010037 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020038 */
Jens Axboe08717142008-01-28 11:38:15 +010039#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020040
41/*
42 * below this threshold, we consider thinktime immediate
43 */
44#define CFQ_MIN_TT (2)
45
Jens Axboe22e2c502005-06-27 10:55:12 +020046#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020047#define CFQ_HW_QUEUE_MIN (5)
Vivek Goyal25bc6b02009-12-03 12:59:43 -050048#define CFQ_SERVICE_SHIFT 12
Jens Axboe22e2c502005-06-27 10:55:12 +020049
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010050#define CFQQ_SEEK_THR (sector_t)(8 * 100)
Shaohua Lie9ce3352010-03-19 08:03:04 +010051#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
Corrado Zoccolo41647e72010-02-27 19:45:40 +010052#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010053#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
Shaohua Liae54abe2010-02-05 13:11:45 +010054
Jens Axboefe094d92008-01-31 13:08:54 +010055#define RQ_CIC(rq) \
56 ((struct cfq_io_context *) (rq)->elevator_private)
Jens Axboe7b679132008-05-30 12:23:07 +020057#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +020058#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private3)
Linus Torvalds1da177e2005-04-16 15:20:36 -070059
Christoph Lametere18b8902006-12-06 20:33:20 -080060static struct kmem_cache *cfq_pool;
61static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
Tejun Heo245b2e72009-06-24 15:13:48 +090063static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050064static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020065static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050066
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +040067static DEFINE_SPINLOCK(cic_index_lock);
68static DEFINE_IDA(cic_index_ida);
69
Jens Axboe22e2c502005-06-27 10:55:12 +020070#define CFQ_PRIO_LISTS IOPRIO_BE_NR
71#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020072#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
73
Jens Axboe206dc692006-03-28 13:03:44 +020074#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050075#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020076
Jens Axboe22e2c502005-06-27 10:55:12 +020077/*
Jens Axboecc09e292007-04-26 12:53:50 +020078 * Most of our rbtree usage is for sorting with min extraction, so
79 * if we cache the leftmost node we don't have to walk down the tree
80 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
81 * move this into the elevator for the rq sorting as well.
82 */
83struct cfq_rb_root {
84 struct rb_root rb;
85 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010086 unsigned count;
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010087 unsigned total_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050088 u64 min_vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -050089 struct rb_node *active;
Jens Axboecc09e292007-04-26 12:53:50 +020090};
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010091#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92 .count = 0, .min_vdisktime = 0, }
Jens Axboecc09e292007-04-26 12:53:50 +020093
94/*
Jens Axboe6118b702009-06-30 09:34:12 +020095 * Per process-grouping structure
96 */
97struct cfq_queue {
98 /* reference count */
99 atomic_t ref;
100 /* various state flags, see below */
101 unsigned int flags;
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
117 int queued[2];
118 /* currently allocated requests */
119 int allocated[2];
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
122
Vivek Goyaldae739e2009-12-03 12:59:45 -0500123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500125 unsigned int allocated_slice;
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100126 unsigned int slice_dispatch;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
Jens Axboe6118b702009-06-30 09:34:12 +0200129 unsigned long slice_end;
130 long slice_resid;
Jens Axboe6118b702009-06-30 09:34:12 +0200131
132 /* pending metadata requests */
133 int meta_pending;
134 /* number of requests that are on the dispatch list or inside driver */
135 int dispatched;
136
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
139 unsigned short ioprio_class, org_ioprio_class;
140
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100141 pid_t pid;
142
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +0100143 u32 seek_history;
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400144 sector_t last_request_pos;
145
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100146 struct cfq_rb_root *service_tree;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400147 struct cfq_queue *new_cfqq;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500148 struct cfq_group *cfqg;
Vivek Goyalae30c282009-12-03 12:59:55 -0500149 struct cfq_group *orig_cfqg;
Jens Axboe6118b702009-06-30 09:34:12 +0200150};
151
152/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100153 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100154 * IDLE is handled separately, so it has negative index
155 */
156enum wl_prio_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100157 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500158 RT_WORKLOAD = 1,
159 IDLE_WORKLOAD = 2,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100160};
161
162/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100163 * Second index in the service_trees.
164 */
165enum wl_type_t {
166 ASYNC_WORKLOAD = 0,
167 SYNC_NOIDLE_WORKLOAD = 1,
168 SYNC_WORKLOAD = 2
169};
170
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500171/* This is per cgroup per device grouping structure */
172struct cfq_group {
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500173 /* group service_tree member */
174 struct rb_node rb_node;
175
176 /* group service_tree key */
177 u64 vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500178 unsigned int weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500179 bool on_st;
180
181 /* number of cfqq currently on this group */
182 int nr_cfqq;
183
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500184 /* Per group busy queus average. Useful for workload slice calc. */
185 unsigned int busy_queues_avg[2];
Jens Axboe22e2c502005-06-27 10:55:12 +0200186 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100187 * rr lists of queues with requests, onle rr for each priority class.
188 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200189 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100190 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100191 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500192
193 unsigned long saved_workload_slice;
194 enum wl_type_t saved_workload;
195 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500196 struct blkio_group blkg;
197#ifdef CONFIG_CFQ_GROUP_IOSCHED
198 struct hlist_node cfqd_node;
Vivek Goyalb1c35762009-12-03 12:59:47 -0500199 atomic_t ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500200#endif
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500201};
202
203/*
204 * Per block device queue structure
205 */
206struct cfq_data {
207 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500208 /* Root service tree for cfq_groups */
209 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500210 struct cfq_group root_group;
211
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100212 /*
213 * The priority currently being served
214 */
215 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100216 enum wl_type_t serving_type;
217 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500218 struct cfq_group *serving_group;
Corrado Zoccolo8e550632009-11-26 10:02:58 +0100219 bool noidle_tree_requires_idle;
Jens Axboea36e71f2009-04-15 12:15:11 +0200220
221 /*
222 * Each priority tree is sorted by next_request position. These
223 * trees are used when determining if two or more queues are
224 * interleaving requests (see cfq_close_cooperator).
225 */
226 struct rb_root prio_trees[CFQ_PRIO_LISTS];
227
Jens Axboe22e2c502005-06-27 10:55:12 +0200228 unsigned int busy_queues;
229
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100230 int rq_in_driver;
231 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200232
233 /*
234 * queue-depth detection
235 */
236 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200237 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100238 /*
239 * hw_tag can be
240 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
241 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
242 * 0 => no NCQ
243 */
244 int hw_tag_est_depth;
245 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200246
247 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200248 * idle window management
249 */
250 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200251 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200252
253 struct cfq_queue *active_queue;
254 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200255
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200256 /*
257 * async queue for each priority case
258 */
259 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
260 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200261
Jens Axboe6d048f52007-04-25 12:44:27 +0200262 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 /*
265 * tunables, see top of file
266 */
267 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200268 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269 unsigned int cfq_back_penalty;
270 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200271 unsigned int cfq_slice[2];
272 unsigned int cfq_slice_async_rq;
273 unsigned int cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200274 unsigned int cfq_latency;
Vivek Goyalae30c282009-12-03 12:59:55 -0500275 unsigned int cfq_group_isolation;
Al Virod9ff4182006-03-18 13:51:22 -0500276
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400277 unsigned int cic_index;
Al Virod9ff4182006-03-18 13:51:22 -0500278 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
Jens Axboe6118b702009-06-30 09:34:12 +0200280 /*
281 * Fallback dummy cfqq for extreme OOM conditions
282 */
283 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200284
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100285 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500286
287 /* List of cfq groups being managed on this device*/
288 struct hlist_head cfqg_list;
Jens Axboebb729bc2009-12-06 09:54:19 +0100289 struct rcu_head rcu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290};
291
Vivek Goyal25fb5162009-12-03 12:59:46 -0500292static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
293
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500294static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
295 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500296 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100297{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500298 if (!cfqg)
299 return NULL;
300
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100301 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500302 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100303
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500304 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100305}
306
Jens Axboe3b181522005-06-27 10:56:24 +0200307enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100308 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
309 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200310 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100311 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100312 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
313 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
314 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100315 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200316 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400317 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100318 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100319 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500320 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200321};
322
323#define CFQ_CFQQ_FNS(name) \
324static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
325{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100326 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200327} \
328static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
329{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100330 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200331} \
332static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
333{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100334 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200335}
336
337CFQ_CFQQ_FNS(on_rr);
338CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200339CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200340CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200341CFQ_CFQQ_FNS(fifo_expire);
342CFQ_CFQQ_FNS(idle_window);
343CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100344CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200345CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200346CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100347CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100348CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500349CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200350#undef CFQ_CFQQ_FNS
351
Vivek Goyalafc24d42010-04-26 19:27:56 +0200352#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500353#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
354 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
355 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
356 blkg_path(&(cfqq)->cfqg->blkg), ##args);
357
358#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
359 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
360 blkg_path(&(cfqg)->blkg), ##args); \
361
362#else
Jens Axboe7b679132008-05-30 12:23:07 +0200363#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
364 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500365#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
366#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200367#define cfq_log(cfqd, fmt, args...) \
368 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
369
Vivek Goyal615f0252009-12-03 12:59:39 -0500370/* Traverses through cfq group service trees */
371#define for_each_cfqg_st(cfqg, i, j, st) \
372 for (i = 0; i <= IDLE_WORKLOAD; i++) \
373 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
374 : &cfqg->service_tree_idle; \
375 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
376 (i == IDLE_WORKLOAD && j == 0); \
377 j++, st = i < IDLE_WORKLOAD ? \
378 &cfqg->service_trees[i][j]: NULL) \
379
380
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100381static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
382{
383 if (cfq_class_idle(cfqq))
384 return IDLE_WORKLOAD;
385 if (cfq_class_rt(cfqq))
386 return RT_WORKLOAD;
387 return BE_WORKLOAD;
388}
389
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100390
391static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
392{
393 if (!cfq_cfqq_sync(cfqq))
394 return ASYNC_WORKLOAD;
395 if (!cfq_cfqq_idle_window(cfqq))
396 return SYNC_NOIDLE_WORKLOAD;
397 return SYNC_WORKLOAD;
398}
399
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500400static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
401 struct cfq_data *cfqd,
402 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100403{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500404 if (wl == IDLE_WORKLOAD)
405 return cfqg->service_tree_idle.count;
406
407 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
408 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
409 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100410}
411
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500412static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
413 struct cfq_group *cfqg)
414{
415 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
416 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
417}
418
Jens Axboe165125e2007-07-24 09:28:11 +0200419static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200420static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100421 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100422static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200423 struct io_context *);
424
425static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200426 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200427{
Jens Axboea6151c32009-10-07 20:02:57 +0200428 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200429}
430
431static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200432 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200433{
Jens Axboea6151c32009-10-07 20:02:57 +0200434 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200435}
436
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400437#define CIC_DEAD_KEY 1ul
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400438#define CIC_DEAD_INDEX_SHIFT 1
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400439
440static inline void *cfqd_dead_key(struct cfq_data *cfqd)
441{
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400442 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400443}
444
445static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
446{
447 struct cfq_data *cfqd = cic->key;
448
449 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
450 return NULL;
451
452 return cfqd;
453}
454
Vasily Tarasov91fac312007-04-25 12:29:51 +0200455/*
456 * We regard a request as SYNC, if it's either a read or has the SYNC bit
457 * set (in which case it could also be direct WRITE).
458 */
Jens Axboea6151c32009-10-07 20:02:57 +0200459static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200460{
Jens Axboea6151c32009-10-07 20:02:57 +0200461 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200462}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700465 * scheduler run of queue, if there are requests pending and no one in the
466 * driver that will restart queueing
467 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200468static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700469{
Jens Axboe7b679132008-05-30 12:23:07 +0200470 if (cfqd->busy_queues) {
471 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200472 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200473 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700474}
475
Jens Axboe165125e2007-07-24 09:28:11 +0200476static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700477{
478 struct cfq_data *cfqd = q->elevator->elevator_data;
479
Vivek Goyalf04a6422009-12-03 12:59:40 -0500480 return !cfqd->rq_queued;
Andrew Morton99f95e52005-06-27 20:14:05 -0700481}
482
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100484 * Scale schedule slice based on io priority. Use the sync time slice only
485 * if a queue is marked sync and has sync io queued. A sync queue with async
486 * io only, should not get full sync slice length.
487 */
Jens Axboea6151c32009-10-07 20:02:57 +0200488static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200489 unsigned short prio)
490{
491 const int base_slice = cfqd->cfq_slice[sync];
492
493 WARN_ON(prio >= IOPRIO_BE_NR);
494
495 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
496}
497
Jens Axboe44f7c162007-01-19 11:51:58 +1100498static inline int
499cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
500{
Jens Axboed9e76202007-04-20 14:27:50 +0200501 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100502}
503
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500504static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
505{
506 u64 d = delta << CFQ_SERVICE_SHIFT;
507
508 d = d * BLKIO_WEIGHT_DEFAULT;
509 do_div(d, cfqg->weight);
510 return d;
511}
512
513static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
514{
515 s64 delta = (s64)(vdisktime - min_vdisktime);
516 if (delta > 0)
517 min_vdisktime = vdisktime;
518
519 return min_vdisktime;
520}
521
522static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
523{
524 s64 delta = (s64)(vdisktime - min_vdisktime);
525 if (delta < 0)
526 min_vdisktime = vdisktime;
527
528 return min_vdisktime;
529}
530
531static void update_min_vdisktime(struct cfq_rb_root *st)
532{
533 u64 vdisktime = st->min_vdisktime;
534 struct cfq_group *cfqg;
535
536 if (st->active) {
537 cfqg = rb_entry_cfqg(st->active);
538 vdisktime = cfqg->vdisktime;
539 }
540
541 if (st->left) {
542 cfqg = rb_entry_cfqg(st->left);
543 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
544 }
545
546 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
547}
548
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100549/*
550 * get averaged number of queues of RT/BE priority.
551 * average is updated, with a formula that gives more weight to higher numbers,
552 * to quickly follows sudden increases and decrease slowly
553 */
554
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500555static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
556 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100557{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100558 unsigned min_q, max_q;
559 unsigned mult = cfq_hist_divisor - 1;
560 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500561 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100562
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500563 min_q = min(cfqg->busy_queues_avg[rt], busy);
564 max_q = max(cfqg->busy_queues_avg[rt], busy);
565 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100566 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500567 return cfqg->busy_queues_avg[rt];
568}
569
570static inline unsigned
571cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
572{
573 struct cfq_rb_root *st = &cfqd->grp_service_tree;
574
575 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100576}
577
Jens Axboe44f7c162007-01-19 11:51:58 +1100578static inline void
579cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
580{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100581 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
582 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500583 /*
584 * interested queues (we consider only the ones with the same
585 * priority class in the cfq group)
586 */
587 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
588 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100589 unsigned sync_slice = cfqd->cfq_slice[1];
590 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500591 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
592
593 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100594 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
595 /* scale low_slice according to IO priority
596 * and sync vs async */
597 unsigned low_slice =
598 min(slice, base_low_slice * slice / sync_slice);
599 /* the adapted slice value is scaled to fit all iqs
600 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500601 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100602 low_slice);
603 }
604 }
Vivek Goyaldae739e2009-12-03 12:59:45 -0500605 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100606 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500607 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200608 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100609}
610
611/*
612 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
613 * isn't valid until the first request from the dispatch is activated
614 * and the slice time set.
615 */
Jens Axboea6151c32009-10-07 20:02:57 +0200616static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100617{
618 if (cfq_cfqq_slice_new(cfqq))
619 return 0;
620 if (time_before(jiffies, cfqq->slice_end))
621 return 0;
622
623 return 1;
624}
625
626/*
Jens Axboe5e705372006-07-13 12:39:25 +0200627 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200629 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 */
Jens Axboe5e705372006-07-13 12:39:25 +0200631static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100632cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100634 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200636#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
637#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
638 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Jens Axboe5e705372006-07-13 12:39:25 +0200640 if (rq1 == NULL || rq1 == rq2)
641 return rq2;
642 if (rq2 == NULL)
643 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200644
Jens Axboe5e705372006-07-13 12:39:25 +0200645 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
646 return rq1;
647 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
648 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200649 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
650 return rq1;
651 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
652 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653
Tejun Heo83096eb2009-05-07 22:24:39 +0900654 s1 = blk_rq_pos(rq1);
655 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 /*
658 * by definition, 1KiB is 2 sectors
659 */
660 back_max = cfqd->cfq_back_max * 2;
661
662 /*
663 * Strict one way elevator _except_ in the case where we allow
664 * short backward seeks which are biased as twice the cost of a
665 * similar forward seek.
666 */
667 if (s1 >= last)
668 d1 = s1 - last;
669 else if (s1 + back_max >= last)
670 d1 = (last - s1) * cfqd->cfq_back_penalty;
671 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200672 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
674 if (s2 >= last)
675 d2 = s2 - last;
676 else if (s2 + back_max >= last)
677 d2 = (last - s2) * cfqd->cfq_back_penalty;
678 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200679 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
681 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682
Andreas Mohre8a99052006-03-28 08:59:49 +0200683 /*
684 * By doing switch() on the bit mask "wrap" we avoid having to
685 * check two variables for all permutations: --> faster!
686 */
687 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200688 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200689 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200690 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200691 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200692 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200693 else {
694 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200695 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200696 else
Jens Axboe5e705372006-07-13 12:39:25 +0200697 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200698 }
699
700 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200701 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200702 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200703 return rq2;
704 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200705 default:
706 /*
707 * Since both rqs are wrapped,
708 * start with the one that's further behind head
709 * (--> only *one* back seek required),
710 * since back seek takes more time than forward.
711 */
712 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200713 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 else
Jens Axboe5e705372006-07-13 12:39:25 +0200715 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 }
717}
718
Jens Axboe498d3aa22007-04-26 12:54:48 +0200719/*
720 * The below is leftmost cache rbtree addon
721 */
Jens Axboe08717142008-01-28 11:38:15 +0100722static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200723{
Vivek Goyal615f0252009-12-03 12:59:39 -0500724 /* Service tree is empty */
725 if (!root->count)
726 return NULL;
727
Jens Axboecc09e292007-04-26 12:53:50 +0200728 if (!root->left)
729 root->left = rb_first(&root->rb);
730
Jens Axboe08717142008-01-28 11:38:15 +0100731 if (root->left)
732 return rb_entry(root->left, struct cfq_queue, rb_node);
733
734 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200735}
736
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500737static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
738{
739 if (!root->left)
740 root->left = rb_first(&root->rb);
741
742 if (root->left)
743 return rb_entry_cfqg(root->left);
744
745 return NULL;
746}
747
Jens Axboea36e71f2009-04-15 12:15:11 +0200748static void rb_erase_init(struct rb_node *n, struct rb_root *root)
749{
750 rb_erase(n, root);
751 RB_CLEAR_NODE(n);
752}
753
Jens Axboecc09e292007-04-26 12:53:50 +0200754static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
755{
756 if (root->left == n)
757 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200758 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100759 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200760}
761
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762/*
763 * would be nice to take fifo expire time into account as well
764 */
Jens Axboe5e705372006-07-13 12:39:25 +0200765static struct request *
766cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
767 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
Jens Axboe21183b02006-07-13 12:33:14 +0200769 struct rb_node *rbnext = rb_next(&last->rb_node);
770 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200771 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772
Jens Axboe21183b02006-07-13 12:33:14 +0200773 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
775 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200776 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200777
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200779 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200780 else {
781 rbnext = rb_first(&cfqq->sort_list);
782 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200783 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200784 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100786 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787}
788
Jens Axboed9e76202007-04-20 14:27:50 +0200789static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
790 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791{
Jens Axboed9e76202007-04-20 14:27:50 +0200792 /*
793 * just an approximation, should be ok.
794 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500795 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100796 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200797}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500799static inline s64
800cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
801{
802 return cfqg->vdisktime - st->min_vdisktime;
803}
804
805static void
806__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
807{
808 struct rb_node **node = &st->rb.rb_node;
809 struct rb_node *parent = NULL;
810 struct cfq_group *__cfqg;
811 s64 key = cfqg_key(st, cfqg);
812 int left = 1;
813
814 while (*node != NULL) {
815 parent = *node;
816 __cfqg = rb_entry_cfqg(parent);
817
818 if (key < cfqg_key(st, __cfqg))
819 node = &parent->rb_left;
820 else {
821 node = &parent->rb_right;
822 left = 0;
823 }
824 }
825
826 if (left)
827 st->left = &cfqg->rb_node;
828
829 rb_link_node(&cfqg->rb_node, parent, node);
830 rb_insert_color(&cfqg->rb_node, &st->rb);
831}
832
833static void
834cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
835{
836 struct cfq_rb_root *st = &cfqd->grp_service_tree;
837 struct cfq_group *__cfqg;
838 struct rb_node *n;
839
840 cfqg->nr_cfqq++;
841 if (cfqg->on_st)
842 return;
843
844 /*
845 * Currently put the group at the end. Later implement something
846 * so that groups get lesser vtime based on their weights, so that
847 * if group does not loose all if it was not continously backlogged.
848 */
849 n = rb_last(&st->rb);
850 if (n) {
851 __cfqg = rb_entry_cfqg(n);
852 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
853 } else
854 cfqg->vdisktime = st->min_vdisktime;
855
856 __cfq_group_service_tree_add(st, cfqg);
857 cfqg->on_st = true;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500858 st->total_weight += cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500859}
860
861static void
862cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
863{
864 struct cfq_rb_root *st = &cfqd->grp_service_tree;
865
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500866 if (st->active == &cfqg->rb_node)
867 st->active = NULL;
868
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500869 BUG_ON(cfqg->nr_cfqq < 1);
870 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500871
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500872 /* If there are other cfq queues under this group, don't delete it */
873 if (cfqg->nr_cfqq)
874 return;
875
Vivek Goyal2868ef72009-12-03 12:59:48 -0500876 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500877 cfqg->on_st = false;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500878 st->total_weight -= cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500879 if (!RB_EMPTY_NODE(&cfqg->rb_node))
880 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500881 cfqg->saved_workload_slice = 0;
Divyesh Shah91952912010-04-01 15:01:41 -0700882 blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500883}
884
885static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
886{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500887 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500888
889 /*
890 * Queue got expired before even a single request completed or
891 * got expired immediately after first request completion.
892 */
893 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
894 /*
895 * Also charge the seek time incurred to the group, otherwise
896 * if there are mutiple queues in the group, each can dispatch
897 * a single request on seeky media and cause lots of seek time
898 * and group will never know it.
899 */
900 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
901 1);
902 } else {
903 slice_used = jiffies - cfqq->slice_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500904 if (slice_used > cfqq->allocated_slice)
905 slice_used = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500906 }
907
Divyesh Shah9a0785b2010-04-01 15:01:04 -0700908 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500909 return slice_used;
910}
911
912static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200913 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500914{
915 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500916 unsigned int used_sl, charge_sl;
917 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
918 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500919
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500920 BUG_ON(nr_sync < 0);
921 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
922
923 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
924 charge_sl = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500925
926 /* Can't update vdisktime while group is on service tree */
927 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500928 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500929 __cfq_group_service_tree_add(st, cfqg);
930
931 /* This group is being expired. Save the context */
932 if (time_after(cfqd->workload_expires, jiffies)) {
933 cfqg->saved_workload_slice = cfqd->workload_expires
934 - jiffies;
935 cfqg->saved_workload = cfqd->serving_type;
936 cfqg->saved_serving_prio = cfqd->serving_prio;
937 } else
938 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500939
940 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
941 st->min_vdisktime);
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700942 blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
Vivek Goyale5ff0822010-04-26 19:25:11 +0200943 blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500944}
945
Vivek Goyal25fb5162009-12-03 12:59:46 -0500946#ifdef CONFIG_CFQ_GROUP_IOSCHED
947static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
948{
949 if (blkg)
950 return container_of(blkg, struct cfq_group, blkg);
951 return NULL;
952}
953
Vivek Goyalf8d461d2009-12-03 12:59:52 -0500954void
955cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
956{
957 cfqg_of_blkg(blkg)->weight = weight;
958}
959
Vivek Goyal25fb5162009-12-03 12:59:46 -0500960static struct cfq_group *
961cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
962{
963 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
964 struct cfq_group *cfqg = NULL;
965 void *key = cfqd;
966 int i, j;
967 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -0500968 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
969 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500970
Vivek Goyal25fb5162009-12-03 12:59:46 -0500971 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
Ricky Beniteza74b2ad2010-04-05 18:22:17 +0200972 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
973 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
974 cfqg->blkg.dev = MKDEV(major, minor);
975 goto done;
976 }
Vivek Goyal25fb5162009-12-03 12:59:46 -0500977 if (cfqg || !create)
978 goto done;
979
980 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
981 if (!cfqg)
982 goto done;
983
Vivek Goyal25fb5162009-12-03 12:59:46 -0500984 for_each_cfqg_st(cfqg, i, j, st)
985 *st = CFQ_RB_ROOT;
986 RB_CLEAR_NODE(&cfqg->rb_node);
987
Vivek Goyalb1c35762009-12-03 12:59:47 -0500988 /*
989 * Take the initial reference that will be released on destroy
990 * This can be thought of a joint reference by cgroup and
991 * elevator which will be dropped by either elevator exit
992 * or cgroup deletion path depending on who is exiting first.
993 */
994 atomic_set(&cfqg->ref, 1);
995
Vivek Goyal25fb5162009-12-03 12:59:46 -0500996 /* Add group onto cgroup list */
Vivek Goyal22084192009-12-03 12:59:49 -0500997 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
998 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
999 MKDEV(major, minor));
Gui Jianfeng34d0f172010-04-13 16:05:49 +08001000 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001001
1002 /* Add group on cfqd list */
1003 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1004
1005done:
Vivek Goyal25fb5162009-12-03 12:59:46 -05001006 return cfqg;
1007}
1008
1009/*
1010 * Search for the cfq group current task belongs to. If create = 1, then also
1011 * create the cfq group if it does not exist. request_queue lock must be held.
1012 */
1013static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1014{
1015 struct cgroup *cgroup;
1016 struct cfq_group *cfqg = NULL;
1017
1018 rcu_read_lock();
1019 cgroup = task_cgroup(current, blkio_subsys_id);
1020 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1021 if (!cfqg && create)
1022 cfqg = &cfqd->root_group;
1023 rcu_read_unlock();
1024 return cfqg;
1025}
1026
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001027static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1028{
1029 atomic_inc(&cfqg->ref);
1030 return cfqg;
1031}
1032
Vivek Goyal25fb5162009-12-03 12:59:46 -05001033static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1034{
1035 /* Currently, all async queues are mapped to root group */
1036 if (!cfq_cfqq_sync(cfqq))
1037 cfqg = &cfqq->cfqd->root_group;
1038
1039 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001040 /* cfqq reference on cfqg */
1041 atomic_inc(&cfqq->cfqg->ref);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001042}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001043
1044static void cfq_put_cfqg(struct cfq_group *cfqg)
1045{
1046 struct cfq_rb_root *st;
1047 int i, j;
1048
1049 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1050 if (!atomic_dec_and_test(&cfqg->ref))
1051 return;
1052 for_each_cfqg_st(cfqg, i, j, st)
1053 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1054 kfree(cfqg);
1055}
1056
1057static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1058{
1059 /* Something wrong if we are trying to remove same group twice */
1060 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1061
1062 hlist_del_init(&cfqg->cfqd_node);
1063
1064 /*
1065 * Put the reference taken at the time of creation so that when all
1066 * queues are gone, group can be destroyed.
1067 */
1068 cfq_put_cfqg(cfqg);
1069}
1070
1071static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1072{
1073 struct hlist_node *pos, *n;
1074 struct cfq_group *cfqg;
1075
1076 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1077 /*
1078 * If cgroup removal path got to blk_group first and removed
1079 * it from cgroup list, then it will take care of destroying
1080 * cfqg also.
1081 */
1082 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1083 cfq_destroy_cfqg(cfqd, cfqg);
1084 }
1085}
1086
1087/*
1088 * Blk cgroup controller notification saying that blkio_group object is being
1089 * delinked as associated cgroup object is going away. That also means that
1090 * no new IO will come in this group. So get rid of this group as soon as
1091 * any pending IO in the group is finished.
1092 *
1093 * This function is called under rcu_read_lock(). key is the rcu protected
1094 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1095 * read lock.
1096 *
1097 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1098 * it should not be NULL as even if elevator was exiting, cgroup deltion
1099 * path got to it first.
1100 */
1101void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1102{
1103 unsigned long flags;
1104 struct cfq_data *cfqd = key;
1105
1106 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1107 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1108 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1109}
1110
Vivek Goyal25fb5162009-12-03 12:59:46 -05001111#else /* GROUP_IOSCHED */
1112static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1113{
1114 return &cfqd->root_group;
1115}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001116
1117static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1118{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001119 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001120}
1121
Vivek Goyal25fb5162009-12-03 12:59:46 -05001122static inline void
1123cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1124 cfqq->cfqg = cfqg;
1125}
1126
Vivek Goyalb1c35762009-12-03 12:59:47 -05001127static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1128static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1129
Vivek Goyal25fb5162009-12-03 12:59:46 -05001130#endif /* GROUP_IOSCHED */
1131
Jens Axboe498d3aa22007-04-26 12:54:48 +02001132/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001133 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001134 * requests waiting to be processed. It is sorted in the order that
1135 * we will service the queues.
1136 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001137static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001138 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001139{
Jens Axboe08717142008-01-28 11:38:15 +01001140 struct rb_node **p, *parent;
1141 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001142 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001143 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001144 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001145 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001146 int group_changed = 0;
1147
1148#ifdef CONFIG_CFQ_GROUP_IOSCHED
1149 if (!cfqd->cfq_group_isolation
1150 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1151 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1152 /* Move this cfq to root group */
1153 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1154 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1155 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1156 cfqq->orig_cfqg = cfqq->cfqg;
1157 cfqq->cfqg = &cfqd->root_group;
1158 atomic_inc(&cfqd->root_group.ref);
1159 group_changed = 1;
1160 } else if (!cfqd->cfq_group_isolation
1161 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1162 /* cfqq is sequential now needs to go to its original group */
1163 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1164 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1165 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1166 cfq_put_cfqg(cfqq->cfqg);
1167 cfqq->cfqg = cfqq->orig_cfqg;
1168 cfqq->orig_cfqg = NULL;
1169 group_changed = 1;
1170 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1171 }
1172#endif
Jens Axboed9e76202007-04-20 14:27:50 +02001173
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001174 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001175 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001176 if (cfq_class_idle(cfqq)) {
1177 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001178 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001179 if (parent && parent != &cfqq->rb_node) {
1180 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1181 rb_key += __cfqq->rb_key;
1182 } else
1183 rb_key += jiffies;
1184 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001185 /*
1186 * Get our rb key offset. Subtract any residual slice
1187 * value carried from last service. A negative resid
1188 * count indicates slice overrun, and this should position
1189 * the next service time further away in the tree.
1190 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001191 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001192 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001193 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001194 } else {
1195 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001196 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001197 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1198 }
Jens Axboed9e76202007-04-20 14:27:50 +02001199
1200 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001201 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001202 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001203 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001204 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001205 if (rb_key == cfqq->rb_key &&
1206 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001207 return;
Jens Axboe53b03742006-07-28 09:48:51 +02001208
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001209 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1210 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001211 }
Jens Axboed9e76202007-04-20 14:27:50 +02001212
Jens Axboe498d3aa22007-04-26 12:54:48 +02001213 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001214 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001215 cfqq->service_tree = service_tree;
1216 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001217 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001218 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001219
Jens Axboed9e76202007-04-20 14:27:50 +02001220 parent = *p;
1221 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1222
Jens Axboe0c534e02007-04-18 20:01:57 +02001223 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001224 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001225 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001226 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001227 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001228 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001229 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001230 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001231 }
Jens Axboe67060e32007-04-18 20:13:32 +02001232
1233 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001234 }
1235
Jens Axboecc09e292007-04-26 12:53:50 +02001236 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001237 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001238
Jens Axboed9e76202007-04-20 14:27:50 +02001239 cfqq->rb_key = rb_key;
1240 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001241 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1242 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001243 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001244 return;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001245 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246}
1247
Jens Axboea36e71f2009-04-15 12:15:11 +02001248static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001249cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1250 sector_t sector, struct rb_node **ret_parent,
1251 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001252{
Jens Axboea36e71f2009-04-15 12:15:11 +02001253 struct rb_node **p, *parent;
1254 struct cfq_queue *cfqq = NULL;
1255
1256 parent = NULL;
1257 p = &root->rb_node;
1258 while (*p) {
1259 struct rb_node **n;
1260
1261 parent = *p;
1262 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1263
1264 /*
1265 * Sort strictly based on sector. Smallest to the left,
1266 * largest to the right.
1267 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001268 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001269 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001270 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001271 n = &(*p)->rb_left;
1272 else
1273 break;
1274 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001275 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001276 }
1277
1278 *ret_parent = parent;
1279 if (rb_link)
1280 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001281 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001282}
1283
1284static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1285{
Jens Axboea36e71f2009-04-15 12:15:11 +02001286 struct rb_node **p, *parent;
1287 struct cfq_queue *__cfqq;
1288
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001289 if (cfqq->p_root) {
1290 rb_erase(&cfqq->p_node, cfqq->p_root);
1291 cfqq->p_root = NULL;
1292 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001293
1294 if (cfq_class_idle(cfqq))
1295 return;
1296 if (!cfqq->next_rq)
1297 return;
1298
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001299 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001300 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1301 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001302 if (!__cfqq) {
1303 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001304 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1305 } else
1306 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001307}
1308
Jens Axboe498d3aa22007-04-26 12:54:48 +02001309/*
1310 * Update cfqq's position in the service tree.
1311 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001312static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001313{
Jens Axboe6d048f52007-04-25 12:44:27 +02001314 /*
1315 * Resorting requires the cfqq to be on the RR list already.
1316 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001317 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001318 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001319 cfq_prio_tree_add(cfqd, cfqq);
1320 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001321}
1322
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323/*
1324 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001325 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 */
Jens Axboefebffd62008-01-28 13:19:43 +01001327static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328{
Jens Axboe7b679132008-05-30 12:23:07 +02001329 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001330 BUG_ON(cfq_cfqq_on_rr(cfqq));
1331 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 cfqd->busy_queues++;
1333
Jens Axboeedd75ff2007-04-19 12:03:34 +02001334 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335}
1336
Jens Axboe498d3aa22007-04-26 12:54:48 +02001337/*
1338 * Called when the cfqq no longer has requests pending, remove it from
1339 * the service tree.
1340 */
Jens Axboefebffd62008-01-28 13:19:43 +01001341static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342{
Jens Axboe7b679132008-05-30 12:23:07 +02001343 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001344 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1345 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001347 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1348 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1349 cfqq->service_tree = NULL;
1350 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001351 if (cfqq->p_root) {
1352 rb_erase(&cfqq->p_node, cfqq->p_root);
1353 cfqq->p_root = NULL;
1354 }
Jens Axboed9e76202007-04-20 14:27:50 +02001355
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001356 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 BUG_ON(!cfqd->busy_queues);
1358 cfqd->busy_queues--;
1359}
1360
1361/*
1362 * rb tree support functions
1363 */
Jens Axboefebffd62008-01-28 13:19:43 +01001364static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365{
Jens Axboe5e705372006-07-13 12:39:25 +02001366 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001367 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368
Jens Axboeb4878f22005-10-20 16:42:29 +02001369 BUG_ON(!cfqq->queued[sync]);
1370 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371
Jens Axboe5e705372006-07-13 12:39:25 +02001372 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373
Vivek Goyalf04a6422009-12-03 12:59:40 -05001374 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1375 /*
1376 * Queue will be deleted from service tree when we actually
1377 * expire it later. Right now just remove it from prio tree
1378 * as it is empty.
1379 */
1380 if (cfqq->p_root) {
1381 rb_erase(&cfqq->p_node, cfqq->p_root);
1382 cfqq->p_root = NULL;
1383 }
1384 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385}
1386
Jens Axboe5e705372006-07-13 12:39:25 +02001387static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388{
Jens Axboe5e705372006-07-13 12:39:25 +02001389 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001391 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392
Jens Axboe5380a102006-07-13 12:37:56 +02001393 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394
1395 /*
1396 * looks a little odd, but the first insert might return an alias.
1397 * if that happens, put the alias on the dispatch list
1398 */
Jens Axboe21183b02006-07-13 12:33:14 +02001399 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001400 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001401
1402 if (!cfq_cfqq_on_rr(cfqq))
1403 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001404
1405 /*
1406 * check if this request is a better next-serve candidate
1407 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001408 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001409 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001410
1411 /*
1412 * adjust priority tree position, if ->next_rq changes
1413 */
1414 if (prev != cfqq->next_rq)
1415 cfq_prio_tree_add(cfqd, cfqq);
1416
Jens Axboe5044eed2007-04-25 11:53:48 +02001417 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418}
1419
Jens Axboefebffd62008-01-28 13:19:43 +01001420static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421{
Jens Axboe5380a102006-07-13 12:37:56 +02001422 elv_rb_del(&cfqq->sort_list, rq);
1423 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001424 blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(rq),
Divyesh Shahcdc11842010-04-08 21:15:10 -07001425 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001426 cfq_add_rq_rb(rq);
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001427 blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1428 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1429 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430}
1431
Jens Axboe206dc692006-03-28 13:03:44 +02001432static struct request *
1433cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434{
Jens Axboe206dc692006-03-28 13:03:44 +02001435 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001436 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001437 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438
Jens Axboe4ac845a2008-01-24 08:44:49 +01001439 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001440 if (!cic)
1441 return NULL;
1442
1443 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001444 if (cfqq) {
1445 sector_t sector = bio->bi_sector + bio_sectors(bio);
1446
Jens Axboe21183b02006-07-13 12:33:14 +02001447 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001448 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 return NULL;
1451}
1452
Jens Axboe165125e2007-07-24 09:28:11 +02001453static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001454{
1455 struct cfq_data *cfqd = q->elevator->elevator_data;
1456
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001457 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001458 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001459 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001460
Tejun Heo5b936292009-05-07 22:24:38 +09001461 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001462}
1463
Jens Axboe165125e2007-07-24 09:28:11 +02001464static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465{
Jens Axboe22e2c502005-06-27 10:55:12 +02001466 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001468 WARN_ON(!cfqd->rq_in_driver);
1469 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001470 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001471 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472}
1473
Jens Axboeb4878f22005-10-20 16:42:29 +02001474static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475{
Jens Axboe5e705372006-07-13 12:39:25 +02001476 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001477
Jens Axboe5e705372006-07-13 12:39:25 +02001478 if (cfqq->next_rq == rq)
1479 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480
Jens Axboeb4878f22005-10-20 16:42:29 +02001481 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001482 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001483
Aaron Carroll45333d52008-08-26 15:52:36 +02001484 cfqq->cfqd->rq_queued--;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001485 blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(rq),
Divyesh Shahcdc11842010-04-08 21:15:10 -07001486 rq_is_sync(rq));
Jens Axboe374f84a2006-07-23 01:42:19 +02001487 if (rq_is_meta(rq)) {
1488 WARN_ON(!cfqq->meta_pending);
1489 cfqq->meta_pending--;
1490 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491}
1492
Jens Axboe165125e2007-07-24 09:28:11 +02001493static int cfq_merge(struct request_queue *q, struct request **req,
1494 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495{
1496 struct cfq_data *cfqd = q->elevator->elevator_data;
1497 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498
Jens Axboe206dc692006-03-28 13:03:44 +02001499 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001500 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001501 *req = __rq;
1502 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 }
1504
1505 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506}
1507
Jens Axboe165125e2007-07-24 09:28:11 +02001508static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001509 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510{
Jens Axboe21183b02006-07-13 12:33:14 +02001511 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001512 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513
Jens Axboe5e705372006-07-13 12:39:25 +02001514 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516}
1517
Divyesh Shah812d4022010-04-08 21:14:23 -07001518static void cfq_bio_merged(struct request_queue *q, struct request *req,
1519 struct bio *bio)
1520{
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001521 blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg, bio_data_dir(bio),
Divyesh Shah812d4022010-04-08 21:14:23 -07001522 cfq_bio_sync(bio));
1523}
1524
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525static void
Jens Axboe165125e2007-07-24 09:28:11 +02001526cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527 struct request *next)
1528{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001529 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001530 /*
1531 * reposition in fifo if next is older than rq
1532 */
1533 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001534 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001535 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001536 rq_set_fifo_time(rq, rq_fifo_time(next));
1537 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001538
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001539 if (cfqq->next_rq == next)
1540 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001541 cfq_remove_request(next);
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001542 blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(next),
Divyesh Shah812d4022010-04-08 21:14:23 -07001543 rq_is_sync(next));
Jens Axboe22e2c502005-06-27 10:55:12 +02001544}
1545
Jens Axboe165125e2007-07-24 09:28:11 +02001546static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001547 struct bio *bio)
1548{
1549 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001550 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001551 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001552
1553 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001554 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001555 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001556 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001557 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001558
1559 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001560 * Lookup the cfqq that this bio will be queued with. Allow
1561 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001562 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001563 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001564 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001565 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001566
Vasily Tarasov91fac312007-04-25 12:29:51 +02001567 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001568 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001569}
1570
Divyesh Shah812df482010-04-08 21:15:35 -07001571static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1572{
1573 del_timer(&cfqd->idle_slice_timer);
1574 blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1575}
1576
Jens Axboefebffd62008-01-28 13:19:43 +01001577static void __cfq_set_active_queue(struct cfq_data *cfqd,
1578 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001579{
1580 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001581 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1582 cfqd->serving_prio, cfqd->serving_type);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02001583 blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001584 cfqq->slice_start = 0;
1585 cfqq->dispatch_start = jiffies;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001586 cfqq->allocated_slice = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001587 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001588 cfqq->slice_dispatch = 0;
1589
Jens Axboe2f5cb732009-04-07 08:51:19 +02001590 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +02001591 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001592 cfq_clear_cfqq_must_alloc_slice(cfqq);
1593 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +11001594 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001595
Divyesh Shah812df482010-04-08 21:15:35 -07001596 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001597 }
1598
1599 cfqd->active_queue = cfqq;
1600}
1601
1602/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001603 * current cfqq expired its slice (or was too idle), select new one
1604 */
1605static void
1606__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001607 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001608{
Jens Axboe7b679132008-05-30 12:23:07 +02001609 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1610
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001611 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001612 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001613
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001614 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001615 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001616
1617 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001618 * If this cfqq is shared between multiple processes, check to
1619 * make sure that those processes are still issuing I/Os within
1620 * the mean seek distance. If not, it may be time to break the
1621 * queues apart again.
1622 */
1623 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1624 cfq_mark_cfqq_split_coop(cfqq);
1625
1626 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001627 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001628 */
Jens Axboe7b679132008-05-30 12:23:07 +02001629 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +11001630 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001631 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1632 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001633
Vivek Goyale5ff0822010-04-26 19:25:11 +02001634 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001635
Vivek Goyalf04a6422009-12-03 12:59:40 -05001636 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1637 cfq_del_cfqq_rr(cfqd, cfqq);
1638
Jens Axboeedd75ff2007-04-19 12:03:34 +02001639 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001640
1641 if (cfqq == cfqd->active_queue)
1642 cfqd->active_queue = NULL;
1643
Vivek Goyaldae739e2009-12-03 12:59:45 -05001644 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1645 cfqd->grp_service_tree.active = NULL;
1646
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001647 if (cfqd->active_cic) {
1648 put_io_context(cfqd->active_cic->ioc);
1649 cfqd->active_cic = NULL;
1650 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001651}
1652
Vivek Goyale5ff0822010-04-26 19:25:11 +02001653static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001654{
1655 struct cfq_queue *cfqq = cfqd->active_queue;
1656
1657 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001658 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001659}
1660
Jens Axboe498d3aa22007-04-26 12:54:48 +02001661/*
1662 * Get next queue for service. Unless we have a queue preemption,
1663 * we'll simply select the first cfqq in the service tree.
1664 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001665static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001666{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001667 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001668 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001669 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001670
Vivek Goyalf04a6422009-12-03 12:59:40 -05001671 if (!cfqd->rq_queued)
1672 return NULL;
1673
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001674 /* There is nothing to dispatch */
1675 if (!service_tree)
1676 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001677 if (RB_EMPTY_ROOT(&service_tree->rb))
1678 return NULL;
1679 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001680}
1681
Vivek Goyalf04a6422009-12-03 12:59:40 -05001682static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1683{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001684 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001685 struct cfq_queue *cfqq;
1686 int i, j;
1687 struct cfq_rb_root *st;
1688
1689 if (!cfqd->rq_queued)
1690 return NULL;
1691
Vivek Goyal25fb5162009-12-03 12:59:46 -05001692 cfqg = cfq_get_next_cfqg(cfqd);
1693 if (!cfqg)
1694 return NULL;
1695
Vivek Goyalf04a6422009-12-03 12:59:40 -05001696 for_each_cfqg_st(cfqg, i, j, st)
1697 if ((cfqq = cfq_rb_first(st)) != NULL)
1698 return cfqq;
1699 return NULL;
1700}
1701
Jens Axboe498d3aa22007-04-26 12:54:48 +02001702/*
1703 * Get and set a new active queue for service.
1704 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001705static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1706 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001707{
Jens Axboee00ef792009-11-04 08:54:55 +01001708 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001709 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001710
Jens Axboe22e2c502005-06-27 10:55:12 +02001711 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001712 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001713}
1714
Jens Axboed9e76202007-04-20 14:27:50 +02001715static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1716 struct request *rq)
1717{
Tejun Heo83096eb2009-05-07 22:24:39 +09001718 if (blk_rq_pos(rq) >= cfqd->last_position)
1719 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001720 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001721 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001722}
1723
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001724static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001725 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001726{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001727 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001728}
1729
Jens Axboea36e71f2009-04-15 12:15:11 +02001730static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1731 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001732{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001733 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001734 struct rb_node *parent, *node;
1735 struct cfq_queue *__cfqq;
1736 sector_t sector = cfqd->last_position;
1737
1738 if (RB_EMPTY_ROOT(root))
1739 return NULL;
1740
1741 /*
1742 * First, if we find a request starting at the end of the last
1743 * request, choose it.
1744 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001745 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001746 if (__cfqq)
1747 return __cfqq;
1748
1749 /*
1750 * If the exact sector wasn't found, the parent of the NULL leaf
1751 * will contain the closest sector.
1752 */
1753 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001754 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001755 return __cfqq;
1756
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001757 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001758 node = rb_next(&__cfqq->p_node);
1759 else
1760 node = rb_prev(&__cfqq->p_node);
1761 if (!node)
1762 return NULL;
1763
1764 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001765 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001766 return __cfqq;
1767
1768 return NULL;
1769}
1770
1771/*
1772 * cfqd - obvious
1773 * cur_cfqq - passed in so that we don't decide that the current queue is
1774 * closely cooperating with itself.
1775 *
1776 * So, basically we're assuming that that cur_cfqq has dispatched at least
1777 * one request, and that cfqd->last_position reflects a position on the disk
1778 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1779 * assumption.
1780 */
1781static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001782 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001783{
1784 struct cfq_queue *cfqq;
1785
Divyesh Shah39c01b22010-03-25 15:45:57 +01001786 if (cfq_class_idle(cur_cfqq))
1787 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001788 if (!cfq_cfqq_sync(cur_cfqq))
1789 return NULL;
1790 if (CFQQ_SEEKY(cur_cfqq))
1791 return NULL;
1792
Jens Axboea36e71f2009-04-15 12:15:11 +02001793 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001794 * Don't search priority tree if it's the only queue in the group.
1795 */
1796 if (cur_cfqq->cfqg->nr_cfqq == 1)
1797 return NULL;
1798
1799 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001800 * We should notice if some of the queues are cooperating, eg
1801 * working closely on the same area of the disk. In that case,
1802 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001803 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001804 cfqq = cfqq_close(cfqd, cur_cfqq);
1805 if (!cfqq)
1806 return NULL;
1807
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001808 /* If new queue belongs to different cfq_group, don't choose it */
1809 if (cur_cfqq->cfqg != cfqq->cfqg)
1810 return NULL;
1811
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001812 /*
1813 * It only makes sense to merge sync queues.
1814 */
1815 if (!cfq_cfqq_sync(cfqq))
1816 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001817 if (CFQQ_SEEKY(cfqq))
1818 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001819
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001820 /*
1821 * Do not merge queues of different priority classes
1822 */
1823 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1824 return NULL;
1825
Jens Axboea36e71f2009-04-15 12:15:11 +02001826 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001827}
1828
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001829/*
1830 * Determine whether we should enforce idle window for this queue.
1831 */
1832
1833static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1834{
1835 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001836 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001837
Vivek Goyalf04a6422009-12-03 12:59:40 -05001838 BUG_ON(!service_tree);
1839 BUG_ON(!service_tree->count);
1840
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001841 /* We never do for idle class queues. */
1842 if (prio == IDLE_WORKLOAD)
1843 return false;
1844
1845 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001846 if (cfq_cfqq_idle_window(cfqq) &&
1847 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001848 return true;
1849
1850 /*
1851 * Otherwise, we do only if they are the last ones
1852 * in their service tree.
1853 */
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001854 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1855 return 1;
1856 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1857 service_tree->count);
1858 return 0;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001859}
1860
Jens Axboe6d048f52007-04-25 12:44:27 +02001861static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001862{
Jens Axboe17926692007-01-19 11:59:30 +11001863 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001864 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001865 unsigned long sl;
1866
Jens Axboea68bbdd2008-09-24 13:03:33 +02001867 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001868 * SSD device without seek penalty, disable idling. But only do so
1869 * for devices that support queuing, otherwise we still have a problem
1870 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001871 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001872 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001873 return;
1874
Jens Axboedd67d052006-06-21 09:36:18 +02001875 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001876 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001877
1878 /*
1879 * idle is disabled, either manually or by past process history
1880 */
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001881 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
Jens Axboe6d048f52007-04-25 12:44:27 +02001882 return;
1883
Jens Axboe22e2c502005-06-27 10:55:12 +02001884 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001885 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001886 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001887 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001888 return;
1889
1890 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001891 * task has exited, don't wait
1892 */
Jens Axboe206dc692006-03-28 13:03:44 +02001893 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001894 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001895 return;
1896
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001897 /*
1898 * If our average think time is larger than the remaining time
1899 * slice, then don't idle. This avoids overrunning the allotted
1900 * time slice.
1901 */
1902 if (sample_valid(cic->ttime_samples) &&
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001903 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1904 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1905 cic->ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001906 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001907 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001908
Jens Axboe3b181522005-06-27 10:56:24 +02001909 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001910
Jens Axboe6d048f52007-04-25 12:44:27 +02001911 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001912
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001913 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Divyesh Shah812df482010-04-08 21:15:35 -07001914 blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001915 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916}
1917
Jens Axboe498d3aa22007-04-26 12:54:48 +02001918/*
1919 * Move request from internal lists to the request queue dispatch list.
1920 */
Jens Axboe165125e2007-07-24 09:28:11 +02001921static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001923 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001924 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001925
Jens Axboe7b679132008-05-30 12:23:07 +02001926 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1927
Jeff Moyer06d21882009-09-11 17:08:59 +02001928 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001929 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001930 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001931 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001932
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001933 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Divyesh Shah84c124d2010-04-09 08:31:19 +02001934 blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1935 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936}
1937
1938/*
1939 * return expired entry, or NULL to just start from scratch in rbtree
1940 */
Jens Axboefebffd62008-01-28 13:19:43 +01001941static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942{
Jens Axboe30996f42009-10-05 11:03:39 +02001943 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944
Jens Axboe3b181522005-06-27 10:56:24 +02001945 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001947
1948 cfq_mark_cfqq_fifo_expire(cfqq);
1949
Jens Axboe89850f72006-07-22 16:48:31 +02001950 if (list_empty(&cfqq->fifo))
1951 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952
Jens Axboe89850f72006-07-22 16:48:31 +02001953 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02001954 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02001955 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956
Jens Axboe30996f42009-10-05 11:03:39 +02001957 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001958 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959}
1960
Jens Axboe22e2c502005-06-27 10:55:12 +02001961static inline int
1962cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1963{
1964 const int base_rq = cfqd->cfq_slice_async_rq;
1965
1966 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1967
1968 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1969}
1970
1971/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001972 * Must be called with the queue_lock held.
1973 */
1974static int cfqq_process_refs(struct cfq_queue *cfqq)
1975{
1976 int process_refs, io_refs;
1977
1978 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1979 process_refs = atomic_read(&cfqq->ref) - io_refs;
1980 BUG_ON(process_refs < 0);
1981 return process_refs;
1982}
1983
1984static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1985{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001986 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001987 struct cfq_queue *__cfqq;
1988
1989 /* Avoid a circular list and skip interim queue merges */
1990 while ((__cfqq = new_cfqq->new_cfqq)) {
1991 if (__cfqq == cfqq)
1992 return;
1993 new_cfqq = __cfqq;
1994 }
1995
1996 process_refs = cfqq_process_refs(cfqq);
1997 /*
1998 * If the process for the cfqq has gone away, there is no
1999 * sense in merging the queues.
2000 */
2001 if (process_refs == 0)
2002 return;
2003
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002004 /*
2005 * Merge in the direction of the lesser amount of work.
2006 */
2007 new_process_refs = cfqq_process_refs(new_cfqq);
2008 if (new_process_refs >= process_refs) {
2009 cfqq->new_cfqq = new_cfqq;
2010 atomic_add(process_refs, &new_cfqq->ref);
2011 } else {
2012 new_cfqq->new_cfqq = cfqq;
2013 atomic_add(new_process_refs, &cfqq->ref);
2014 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002015}
2016
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002017static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002018 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002019{
2020 struct cfq_queue *queue;
2021 int i;
2022 bool key_valid = false;
2023 unsigned long lowest_key = 0;
2024 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2025
Vivek Goyal65b32a52009-12-16 17:52:59 -05002026 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2027 /* select the one with lowest rb_key */
2028 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002029 if (queue &&
2030 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2031 lowest_key = queue->rb_key;
2032 cur_best = i;
2033 key_valid = true;
2034 }
2035 }
2036
2037 return cur_best;
2038}
2039
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002040static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002041{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002042 unsigned slice;
2043 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002044 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002045 unsigned group_slice;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002046
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002047 if (!cfqg) {
2048 cfqd->serving_prio = IDLE_WORKLOAD;
2049 cfqd->workload_expires = jiffies + 1;
2050 return;
2051 }
2052
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002053 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002054 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002055 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002056 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002057 cfqd->serving_prio = BE_WORKLOAD;
2058 else {
2059 cfqd->serving_prio = IDLE_WORKLOAD;
2060 cfqd->workload_expires = jiffies + 1;
2061 return;
2062 }
2063
2064 /*
2065 * For RT and BE, we have to choose also the type
2066 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2067 * expiration time
2068 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002069 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002070 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002071
2072 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002073 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002074 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002075 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002076 return;
2077
2078 /* otherwise select new workload type */
2079 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002080 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2081 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002082 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002083
2084 /*
2085 * the workload slice is computed as a fraction of target latency
2086 * proportional to the number of queues in that workload, over
2087 * all the queues in the same priority class
2088 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002089 group_slice = cfq_group_slice(cfqd, cfqg);
2090
2091 slice = group_slice * count /
2092 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2093 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002094
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002095 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2096 unsigned int tmp;
2097
2098 /*
2099 * Async queues are currently system wide. Just taking
2100 * proportion of queues with-in same group will lead to higher
2101 * async ratio system wide as generally root group is going
2102 * to have higher weight. A more accurate thing would be to
2103 * calculate system wide asnc/sync ratio.
2104 */
2105 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2106 tmp = tmp/cfqd->busy_queues;
2107 slice = min_t(unsigned, slice, tmp);
2108
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002109 /* async workload slice is scaled down according to
2110 * the sync/async slice ratio. */
2111 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002112 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002113 /* sync workload slice is at least 2 * cfq_slice_idle */
2114 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2115
2116 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002117 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002118 cfqd->workload_expires = jiffies + slice;
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002119 cfqd->noidle_tree_requires_idle = false;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002120}
2121
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002122static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2123{
2124 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002125 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002126
2127 if (RB_EMPTY_ROOT(&st->rb))
2128 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002129 cfqg = cfq_rb_first_group(st);
2130 st->active = &cfqg->rb_node;
2131 update_min_vdisktime(st);
2132 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002133}
2134
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002135static void cfq_choose_cfqg(struct cfq_data *cfqd)
2136{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002137 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2138
2139 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002140
2141 /* Restore the workload type data */
2142 if (cfqg->saved_workload_slice) {
2143 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2144 cfqd->serving_type = cfqg->saved_workload;
2145 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002146 } else
2147 cfqd->workload_expires = jiffies - 1;
2148
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002149 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002150}
2151
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002152/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002153 * Select a queue for service. If we have a current active queue,
2154 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002155 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002156static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002157{
Jens Axboea36e71f2009-04-15 12:15:11 +02002158 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002159
2160 cfqq = cfqd->active_queue;
2161 if (!cfqq)
2162 goto new_queue;
2163
Vivek Goyalf04a6422009-12-03 12:59:40 -05002164 if (!cfqd->rq_queued)
2165 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002166
2167 /*
2168 * We were waiting for group to get backlogged. Expire the queue
2169 */
2170 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2171 goto expire;
2172
Jens Axboe22e2c502005-06-27 10:55:12 +02002173 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002174 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002175 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002176 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2177 /*
2178 * If slice had not expired at the completion of last request
2179 * we might not have turned on wait_busy flag. Don't expire
2180 * the queue yet. Allow the group to get backlogged.
2181 *
2182 * The very fact that we have used the slice, that means we
2183 * have been idling all along on this queue and it should be
2184 * ok to wait for this request to complete.
2185 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002186 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2187 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2188 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002189 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002190 } else
Vivek Goyal7667aa02009-12-08 17:52:58 -05002191 goto expire;
2192 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002193
2194 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002195 * The active queue has requests and isn't expired, allow it to
2196 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002197 */
Jens Axboedd67d052006-06-21 09:36:18 +02002198 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002199 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002200
2201 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002202 * If another queue has a request waiting within our mean seek
2203 * distance, let it run. The expire code will check for close
2204 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002205 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002206 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002207 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002208 if (new_cfqq) {
2209 if (!cfqq->new_cfqq)
2210 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002211 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002212 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002213
2214 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002215 * No requests pending. If the active queue still has requests in
2216 * flight or is idling for a new request, allow either of these
2217 * conditions to happen (or time out) before selecting a new queue.
2218 */
Jens Axboecc197472007-04-20 20:45:39 +02002219 if (timer_pending(&cfqd->idle_slice_timer) ||
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002220 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002221 cfqq = NULL;
2222 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002223 }
2224
Jens Axboe3b181522005-06-27 10:56:24 +02002225expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002226 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002227new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002228 /*
2229 * Current queue expired. Check if we have to switch to a new
2230 * service tree
2231 */
2232 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002233 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002234
Jens Axboea36e71f2009-04-15 12:15:11 +02002235 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002236keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002237 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002238}
2239
Jens Axboefebffd62008-01-28 13:19:43 +01002240static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002241{
2242 int dispatched = 0;
2243
2244 while (cfqq->next_rq) {
2245 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2246 dispatched++;
2247 }
2248
2249 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002250
2251 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002252 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002253 return dispatched;
2254}
2255
Jens Axboe498d3aa22007-04-26 12:54:48 +02002256/*
2257 * Drain our current requests. Used for barriers and when switching
2258 * io schedulers on-the-fly.
2259 */
Jens Axboed9e76202007-04-20 14:27:50 +02002260static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002261{
Jens Axboe08717142008-01-28 11:38:15 +01002262 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002263 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002264
Divyesh Shah3440c492010-04-09 09:29:57 +02002265 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002266 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002267 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2268 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002269 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002270 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002271
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002272 BUG_ON(cfqd->busy_queues);
2273
Jeff Moyer69237152009-06-12 15:29:30 +02002274 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002275 return dispatched;
2276}
2277
Shaohua Liabc3c742010-03-01 09:20:54 +01002278static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2279 struct cfq_queue *cfqq)
2280{
2281 /* the queue hasn't finished any request, can't estimate */
2282 if (cfq_cfqq_slice_new(cfqq))
2283 return 1;
2284 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2285 cfqq->slice_end))
2286 return 1;
2287
2288 return 0;
2289}
2290
Jens Axboe0b182d62009-10-06 20:49:37 +02002291static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002292{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002293 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294
Jens Axboe2f5cb732009-04-07 08:51:19 +02002295 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002296 * Drain async requests before we start sync IO
2297 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002298 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002299 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002300
2301 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002302 * If this is an async queue and we have sync IO in flight, let it wait
2303 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002304 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002305 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002306
Shaohua Liabc3c742010-03-01 09:20:54 +01002307 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002308 if (cfq_class_idle(cfqq))
2309 max_dispatch = 1;
2310
2311 /*
2312 * Does this cfqq already have too much IO in flight?
2313 */
2314 if (cfqq->dispatched >= max_dispatch) {
2315 /*
2316 * idle queue must always only have a single IO in flight
2317 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002318 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002319 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002320
Jens Axboe2f5cb732009-04-07 08:51:19 +02002321 /*
2322 * We have other queues, don't allow more IO from this one
2323 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002324 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002325 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002326
Jens Axboe2f5cb732009-04-07 08:51:19 +02002327 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002328 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002329 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002330 if (cfqd->busy_queues == 1)
2331 max_dispatch = -1;
2332 else
2333 /*
2334 * Normally we start throttling cfqq when cfq_quantum/2
2335 * requests have been dispatched. But we can drive
2336 * deeper queue depths at the beginning of slice
2337 * subjected to upper limit of cfq_quantum.
2338 * */
2339 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002340 }
2341
2342 /*
2343 * Async queues must wait a bit before being allowed dispatch.
2344 * We also ramp up the dispatch depth gradually for async IO,
2345 * based on the last sync IO we serviced
2346 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002347 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002348 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002349 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002350
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002351 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002352 if (!depth && !cfqq->dispatched)
2353 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002354 if (depth < max_dispatch)
2355 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 }
2357
Jens Axboe0b182d62009-10-06 20:49:37 +02002358 /*
2359 * If we're below the current max, allow a dispatch
2360 */
2361 return cfqq->dispatched < max_dispatch;
2362}
2363
2364/*
2365 * Dispatch a request from cfqq, moving them to the request queue
2366 * dispatch list.
2367 */
2368static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2369{
2370 struct request *rq;
2371
2372 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2373
2374 if (!cfq_may_dispatch(cfqd, cfqq))
2375 return false;
2376
2377 /*
2378 * follow expired path, else get first next available
2379 */
2380 rq = cfq_check_fifo(cfqq);
2381 if (!rq)
2382 rq = cfqq->next_rq;
2383
2384 /*
2385 * insert request into driver dispatch list
2386 */
2387 cfq_dispatch_insert(cfqd->queue, rq);
2388
2389 if (!cfqd->active_cic) {
2390 struct cfq_io_context *cic = RQ_CIC(rq);
2391
2392 atomic_long_inc(&cic->ioc->refcount);
2393 cfqd->active_cic = cic;
2394 }
2395
2396 return true;
2397}
2398
2399/*
2400 * Find the cfqq that we need to service and move a request from that to the
2401 * dispatch list
2402 */
2403static int cfq_dispatch_requests(struct request_queue *q, int force)
2404{
2405 struct cfq_data *cfqd = q->elevator->elevator_data;
2406 struct cfq_queue *cfqq;
2407
2408 if (!cfqd->busy_queues)
2409 return 0;
2410
2411 if (unlikely(force))
2412 return cfq_forced_dispatch(cfqd);
2413
2414 cfqq = cfq_select_queue(cfqd);
2415 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002416 return 0;
2417
Jens Axboe2f5cb732009-04-07 08:51:19 +02002418 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002419 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002420 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002421 if (!cfq_dispatch_request(cfqd, cfqq))
2422 return 0;
2423
Jens Axboe2f5cb732009-04-07 08:51:19 +02002424 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002425 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002426
2427 /*
2428 * expire an async queue immediately if it has used up its slice. idle
2429 * queue always expire after 1 dispatch round.
2430 */
2431 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2432 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2433 cfq_class_idle(cfqq))) {
2434 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002435 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002436 }
2437
Shan Weib217a902009-09-01 10:06:42 +02002438 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002439 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440}
2441
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442/*
Jens Axboe5e705372006-07-13 12:39:25 +02002443 * task holds one reference to the queue, dropped when task exits. each rq
2444 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002446 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 * queue lock must be held here.
2448 */
2449static void cfq_put_queue(struct cfq_queue *cfqq)
2450{
Jens Axboe22e2c502005-06-27 10:55:12 +02002451 struct cfq_data *cfqd = cfqq->cfqd;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002452 struct cfq_group *cfqg, *orig_cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002453
2454 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455
2456 if (!atomic_dec_and_test(&cfqq->ref))
2457 return;
2458
Jens Axboe7b679132008-05-30 12:23:07 +02002459 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002461 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002462 cfqg = cfqq->cfqg;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002463 orig_cfqg = cfqq->orig_cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002465 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002466 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002467 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002468 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002469
Vivek Goyalf04a6422009-12-03 12:59:40 -05002470 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002472 cfq_put_cfqg(cfqg);
Vivek Goyal878eadd2009-12-07 19:37:15 +01002473 if (orig_cfqg)
2474 cfq_put_cfqg(orig_cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475}
2476
Jens Axboed6de8be2008-05-28 14:46:59 +02002477/*
2478 * Must always be called with the rcu_read_lock() held
2479 */
Jens Axboe07416d22008-05-07 09:17:12 +02002480static void
2481__call_for_each_cic(struct io_context *ioc,
2482 void (*func)(struct io_context *, struct cfq_io_context *))
2483{
2484 struct cfq_io_context *cic;
2485 struct hlist_node *n;
2486
2487 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2488 func(ioc, cic);
2489}
2490
Jens Axboe4ac845a2008-01-24 08:44:49 +01002491/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002492 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002493 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002494static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002495call_for_each_cic(struct io_context *ioc,
2496 void (*func)(struct io_context *, struct cfq_io_context *))
2497{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002498 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002499 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002500 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002501}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002502
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002503static void cfq_cic_free_rcu(struct rcu_head *head)
2504{
2505 struct cfq_io_context *cic;
2506
2507 cic = container_of(head, struct cfq_io_context, rcu_head);
2508
2509 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002510 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002511
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002512 if (ioc_gone) {
2513 /*
2514 * CFQ scheduler is exiting, grab exit lock and check
2515 * the pending io context count. If it hits zero,
2516 * complete ioc_gone and set it back to NULL
2517 */
2518 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002519 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002520 complete(ioc_gone);
2521 ioc_gone = NULL;
2522 }
2523 spin_unlock(&ioc_gone_lock);
2524 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002525}
2526
2527static void cfq_cic_free(struct cfq_io_context *cic)
2528{
2529 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002530}
2531
2532static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2533{
2534 unsigned long flags;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002535 unsigned long dead_key = (unsigned long) cic->key;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002536
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002537 BUG_ON(!(dead_key & CIC_DEAD_KEY));
Jens Axboe4ac845a2008-01-24 08:44:49 +01002538
2539 spin_lock_irqsave(&ioc->lock, flags);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002540 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
Jens Axboeffc4e752008-02-19 10:02:29 +01002541 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002542 spin_unlock_irqrestore(&ioc->lock, flags);
2543
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002544 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002545}
2546
Jens Axboed6de8be2008-05-28 14:46:59 +02002547/*
2548 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2549 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2550 * and ->trim() which is called with the task lock held
2551 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002552static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002554 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002555 * ioc->refcount is zero here, or we are called from elv_unregister(),
2556 * so no more cic's are allowed to be linked into this ioc. So it
2557 * should be ok to iterate over the known list, we will see all cic's
2558 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002559 */
Jens Axboe07416d22008-05-07 09:17:12 +02002560 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561}
2562
Shaohua Lid02a2c02010-05-25 10:16:53 +02002563static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002564{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002565 struct cfq_queue *__cfqq, *next;
2566
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002567 /*
2568 * If this queue was scheduled to merge with another queue, be
2569 * sure to drop the reference taken on that queue (and others in
2570 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2571 */
2572 __cfqq = cfqq->new_cfqq;
2573 while (__cfqq) {
2574 if (__cfqq == cfqq) {
2575 WARN(1, "cfqq->new_cfqq loop detected\n");
2576 break;
2577 }
2578 next = __cfqq->new_cfqq;
2579 cfq_put_queue(__cfqq);
2580 __cfqq = next;
2581 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002582}
2583
2584static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2585{
2586 if (unlikely(cfqq == cfqd->active_queue)) {
2587 __cfq_slice_expired(cfqd, cfqq, 0);
2588 cfq_schedule_dispatch(cfqd);
2589 }
2590
2591 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002592
Jens Axboe89850f72006-07-22 16:48:31 +02002593 cfq_put_queue(cfqq);
2594}
2595
2596static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2597 struct cfq_io_context *cic)
2598{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002599 struct io_context *ioc = cic->ioc;
2600
Jens Axboefc463792006-08-29 09:05:44 +02002601 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002602
2603 /*
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002604 * Make sure dead mark is seen for dead queues
Jens Axboe4ac845a2008-01-24 08:44:49 +01002605 */
Jens Axboefc463792006-08-29 09:05:44 +02002606 smp_wmb();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002607 cic->key = cfqd_dead_key(cfqd);
Jens Axboefc463792006-08-29 09:05:44 +02002608
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002609 if (ioc->ioc_data == cic)
2610 rcu_assign_pointer(ioc->ioc_data, NULL);
2611
Jens Axboeff6657c2009-04-08 10:58:57 +02002612 if (cic->cfqq[BLK_RW_ASYNC]) {
2613 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2614 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002615 }
2616
Jens Axboeff6657c2009-04-08 10:58:57 +02002617 if (cic->cfqq[BLK_RW_SYNC]) {
2618 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2619 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002620 }
Jens Axboe89850f72006-07-22 16:48:31 +02002621}
2622
Jens Axboe4ac845a2008-01-24 08:44:49 +01002623static void cfq_exit_single_io_context(struct io_context *ioc,
2624 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002625{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002626 struct cfq_data *cfqd = cic_to_cfqd(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002627
Jens Axboe89850f72006-07-22 16:48:31 +02002628 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002629 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002630 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002631
Jens Axboe4ac845a2008-01-24 08:44:49 +01002632 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002633
2634 /*
2635 * Ensure we get a fresh copy of the ->key to prevent
2636 * race between exiting task and queue
2637 */
2638 smp_read_barrier_depends();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002639 if (cic->key == cfqd)
Jens Axboe62c1fe92008-12-15 21:19:25 +01002640 __cfq_exit_single_io_context(cfqd, cic);
2641
Jens Axboe4ac845a2008-01-24 08:44:49 +01002642 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002643 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002644}
2645
Jens Axboe498d3aa22007-04-26 12:54:48 +02002646/*
2647 * The process that ioc belongs to has exited, we need to clean up
2648 * and put the internal structures we have that belongs to that process.
2649 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002650static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002652 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653}
2654
Jens Axboe22e2c502005-06-27 10:55:12 +02002655static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002656cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657{
Jens Axboeb5deef92006-07-19 23:39:40 +02002658 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659
Christoph Lameter94f60302007-07-17 04:03:29 -07002660 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2661 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002663 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002664 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002665 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002666 cic->dtor = cfq_free_io_context;
2667 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002668 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 }
2670
2671 return cic;
2672}
2673
Jens Axboefd0928d2008-01-24 08:52:45 +01002674static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002675{
2676 struct task_struct *tsk = current;
2677 int ioprio_class;
2678
Jens Axboe3b181522005-06-27 10:56:24 +02002679 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002680 return;
2681
Jens Axboefd0928d2008-01-24 08:52:45 +01002682 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002683 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002684 default:
2685 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2686 case IOPRIO_CLASS_NONE:
2687 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002688 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002689 */
2690 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002691 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002692 break;
2693 case IOPRIO_CLASS_RT:
2694 cfqq->ioprio = task_ioprio(ioc);
2695 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2696 break;
2697 case IOPRIO_CLASS_BE:
2698 cfqq->ioprio = task_ioprio(ioc);
2699 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2700 break;
2701 case IOPRIO_CLASS_IDLE:
2702 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2703 cfqq->ioprio = 7;
2704 cfq_clear_cfqq_idle_window(cfqq);
2705 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002706 }
2707
2708 /*
2709 * keep track of original prio settings in case we have to temporarily
2710 * elevate the priority of this queue
2711 */
2712 cfqq->org_ioprio = cfqq->ioprio;
2713 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002714 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002715}
2716
Jens Axboefebffd62008-01-28 13:19:43 +01002717static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002718{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002719 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002720 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002721 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002722
Jens Axboecaaa5f92006-06-16 11:23:00 +02002723 if (unlikely(!cfqd))
2724 return;
2725
Jens Axboec1b707d2006-10-30 19:54:23 +01002726 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002727
Jens Axboeff6657c2009-04-08 10:58:57 +02002728 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002729 if (cfqq) {
2730 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002731 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2732 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002733 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002734 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002735 cfq_put_queue(cfqq);
2736 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002737 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002738
Jens Axboeff6657c2009-04-08 10:58:57 +02002739 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002740 if (cfqq)
2741 cfq_mark_cfqq_prio_changed(cfqq);
2742
Jens Axboec1b707d2006-10-30 19:54:23 +01002743 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002744}
2745
Jens Axboefc463792006-08-29 09:05:44 +02002746static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002748 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002749 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750}
2751
Jens Axboed5036d72009-06-26 10:44:34 +02002752static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002753 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002754{
2755 RB_CLEAR_NODE(&cfqq->rb_node);
2756 RB_CLEAR_NODE(&cfqq->p_node);
2757 INIT_LIST_HEAD(&cfqq->fifo);
2758
2759 atomic_set(&cfqq->ref, 0);
2760 cfqq->cfqd = cfqd;
2761
2762 cfq_mark_cfqq_prio_changed(cfqq);
2763
2764 if (is_sync) {
2765 if (!cfq_class_idle(cfqq))
2766 cfq_mark_cfqq_idle_window(cfqq);
2767 cfq_mark_cfqq_sync(cfqq);
2768 }
2769 cfqq->pid = pid;
2770}
2771
Vivek Goyal246103332009-12-03 12:59:51 -05002772#ifdef CONFIG_CFQ_GROUP_IOSCHED
2773static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2774{
2775 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002776 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal246103332009-12-03 12:59:51 -05002777 unsigned long flags;
2778 struct request_queue *q;
2779
2780 if (unlikely(!cfqd))
2781 return;
2782
2783 q = cfqd->queue;
2784
2785 spin_lock_irqsave(q->queue_lock, flags);
2786
2787 if (sync_cfqq) {
2788 /*
2789 * Drop reference to sync queue. A new sync queue will be
2790 * assigned in new group upon arrival of a fresh request.
2791 */
2792 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2793 cic_set_cfqq(cic, NULL, 1);
2794 cfq_put_queue(sync_cfqq);
2795 }
2796
2797 spin_unlock_irqrestore(q->queue_lock, flags);
2798}
2799
2800static void cfq_ioc_set_cgroup(struct io_context *ioc)
2801{
2802 call_for_each_cic(ioc, changed_cgroup);
2803 ioc->cgroup_changed = 0;
2804}
2805#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2806
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002808cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002809 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002812 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002813 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814
2815retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002816 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002817 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002818 /* cic always exists here */
2819 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820
Jens Axboe6118b702009-06-30 09:34:12 +02002821 /*
2822 * Always try a new alloc if we fell back to the OOM cfqq
2823 * originally, since it should just be a temporary situation.
2824 */
2825 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2826 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 if (new_cfqq) {
2828 cfqq = new_cfqq;
2829 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002830 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002832 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002833 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002834 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002836 if (new_cfqq)
2837 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002838 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002839 cfqq = kmem_cache_alloc_node(cfq_pool,
2840 gfp_mask | __GFP_ZERO,
2841 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002842 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843
Jens Axboe6118b702009-06-30 09:34:12 +02002844 if (cfqq) {
2845 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2846 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002847 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002848 cfq_log_cfqq(cfqd, cfqq, "alloced");
2849 } else
2850 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 }
2852
2853 if (new_cfqq)
2854 kmem_cache_free(cfq_pool, new_cfqq);
2855
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 return cfqq;
2857}
2858
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002859static struct cfq_queue **
2860cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2861{
Jens Axboefe094d92008-01-31 13:08:54 +01002862 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002863 case IOPRIO_CLASS_RT:
2864 return &cfqd->async_cfqq[0][ioprio];
2865 case IOPRIO_CLASS_BE:
2866 return &cfqd->async_cfqq[1][ioprio];
2867 case IOPRIO_CLASS_IDLE:
2868 return &cfqd->async_idle_cfqq;
2869 default:
2870 BUG();
2871 }
2872}
2873
Jens Axboe15c31be2007-07-10 13:43:25 +02002874static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002875cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002876 gfp_t gfp_mask)
2877{
Jens Axboefd0928d2008-01-24 08:52:45 +01002878 const int ioprio = task_ioprio(ioc);
2879 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002880 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002881 struct cfq_queue *cfqq = NULL;
2882
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002883 if (!is_sync) {
2884 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2885 cfqq = *async_cfqq;
2886 }
2887
Jens Axboe6118b702009-06-30 09:34:12 +02002888 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002889 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002890
2891 /*
2892 * pin the queue now that it's allocated, scheduler exit will prune it
2893 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002894 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02002895 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002896 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002897 }
2898
2899 atomic_inc(&cfqq->ref);
2900 return cfqq;
2901}
2902
Jens Axboe498d3aa22007-04-26 12:54:48 +02002903/*
2904 * We drop cfq io contexts lazily, so we may find a dead one.
2905 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002906static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002907cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2908 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002909{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002910 unsigned long flags;
2911
Jens Axboefc463792006-08-29 09:05:44 +02002912 WARN_ON(!list_empty(&cic->queue_list));
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002913 BUG_ON(cic->key != cfqd_dead_key(cfqd));
Jens Axboe597bc482007-04-24 21:23:53 +02002914
Jens Axboe4ac845a2008-01-24 08:44:49 +01002915 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02002916
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002917 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002918
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002919 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
Jens Axboeffc4e752008-02-19 10:02:29 +01002920 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002921 spin_unlock_irqrestore(&ioc->lock, flags);
2922
2923 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002924}
2925
Jens Axboee2d74ac2006-03-28 08:59:01 +02002926static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01002927cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002928{
Jens Axboee2d74ac2006-03-28 08:59:01 +02002929 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002930 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002931
Vasily Tarasov91fac312007-04-25 12:29:51 +02002932 if (unlikely(!ioc))
2933 return NULL;
2934
Jens Axboed6de8be2008-05-28 14:46:59 +02002935 rcu_read_lock();
2936
Jens Axboe597bc482007-04-24 21:23:53 +02002937 /*
2938 * we maintain a last-hit cache, to avoid browsing over the tree
2939 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002940 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02002941 if (cic && cic->key == cfqd) {
2942 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02002943 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002944 }
Jens Axboe597bc482007-04-24 21:23:53 +02002945
Jens Axboe4ac845a2008-01-24 08:44:49 +01002946 do {
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002947 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002948 rcu_read_unlock();
2949 if (!cic)
2950 break;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002951 if (unlikely(cic->key != cfqd)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01002952 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002953 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01002954 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002955 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002956
Jens Axboed6de8be2008-05-28 14:46:59 +02002957 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002958 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002959 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002960 break;
2961 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002962
Jens Axboe4ac845a2008-01-24 08:44:49 +01002963 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002964}
2965
Jens Axboe4ac845a2008-01-24 08:44:49 +01002966/*
2967 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2968 * the process specific cfq io context when entered from the block layer.
2969 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2970 */
Jens Axboefebffd62008-01-28 13:19:43 +01002971static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2972 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002973{
Jens Axboe0261d682006-10-30 19:07:48 +01002974 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002975 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002976
Jens Axboe4ac845a2008-01-24 08:44:49 +01002977 ret = radix_tree_preload(gfp_mask);
2978 if (!ret) {
2979 cic->ioc = ioc;
2980 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002981
Jens Axboe4ac845a2008-01-24 08:44:49 +01002982 spin_lock_irqsave(&ioc->lock, flags);
2983 ret = radix_tree_insert(&ioc->radix_root,
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002984 cfqd->cic_index, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01002985 if (!ret)
2986 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002987 spin_unlock_irqrestore(&ioc->lock, flags);
2988
2989 radix_tree_preload_end();
2990
2991 if (!ret) {
2992 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2993 list_add(&cic->queue_list, &cfqd->cic_list);
2994 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002995 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002996 }
2997
Jens Axboe4ac845a2008-01-24 08:44:49 +01002998 if (ret)
2999 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02003000
Jens Axboe4ac845a2008-01-24 08:44:49 +01003001 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003002}
3003
Jens Axboe22e2c502005-06-27 10:55:12 +02003004/*
3005 * Setup general io context and cfq io context. There can be several cfq
3006 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02003007 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02003008 */
3009static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02003010cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011{
Jens Axboe22e2c502005-06-27 10:55:12 +02003012 struct io_context *ioc = NULL;
3013 struct cfq_io_context *cic;
3014
3015 might_sleep_if(gfp_mask & __GFP_WAIT);
3016
Jens Axboeb5deef92006-07-19 23:39:40 +02003017 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02003018 if (!ioc)
3019 return NULL;
3020
Jens Axboe4ac845a2008-01-24 08:44:49 +01003021 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003022 if (cic)
3023 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02003024
Jens Axboee2d74ac2006-03-28 08:59:01 +02003025 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3026 if (cic == NULL)
3027 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02003028
Jens Axboe4ac845a2008-01-24 08:44:49 +01003029 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3030 goto err_free;
3031
Jens Axboe22e2c502005-06-27 10:55:12 +02003032out:
Jens Axboefc463792006-08-29 09:05:44 +02003033 smp_read_barrier_depends();
3034 if (unlikely(ioc->ioprio_changed))
3035 cfq_ioc_set_ioprio(ioc);
3036
Vivek Goyal246103332009-12-03 12:59:51 -05003037#ifdef CONFIG_CFQ_GROUP_IOSCHED
3038 if (unlikely(ioc->cgroup_changed))
3039 cfq_ioc_set_cgroup(ioc);
3040#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003041 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003042err_free:
3043 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003044err:
3045 put_io_context(ioc);
3046 return NULL;
3047}
3048
3049static void
3050cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3051{
Jens Axboeaaf12282007-01-19 11:30:16 +11003052 unsigned long elapsed = jiffies - cic->last_end_request;
3053 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003054
3055 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3056 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3057 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3058}
3059
Jens Axboe206dc692006-03-28 13:03:44 +02003060static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003061cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003062 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003063{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003064 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003065 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003066 if (cfqq->last_request_pos) {
3067 if (cfqq->last_request_pos < blk_rq_pos(rq))
3068 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3069 else
3070 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3071 }
Jens Axboe206dc692006-03-28 13:03:44 +02003072
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003073 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003074 if (blk_queue_nonrot(cfqd->queue))
3075 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3076 else
3077 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003078}
Jens Axboe22e2c502005-06-27 10:55:12 +02003079
3080/*
3081 * Disable idle window if the process thinks too long or seeks so much that
3082 * it doesn't matter
3083 */
3084static void
3085cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3086 struct cfq_io_context *cic)
3087{
Jens Axboe7b679132008-05-30 12:23:07 +02003088 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003089
Jens Axboe08717142008-01-28 11:38:15 +01003090 /*
3091 * Don't idle for async or idle io prio class
3092 */
3093 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003094 return;
3095
Jens Axboec265a7f2008-06-26 13:49:33 +02003096 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003097
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003098 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3099 cfq_mark_cfqq_deep(cfqq);
3100
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01003101 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003102 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003103 enable_idle = 0;
3104 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003105 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003106 enable_idle = 0;
3107 else
3108 enable_idle = 1;
3109 }
3110
Jens Axboe7b679132008-05-30 12:23:07 +02003111 if (old_idle != enable_idle) {
3112 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3113 if (enable_idle)
3114 cfq_mark_cfqq_idle_window(cfqq);
3115 else
3116 cfq_clear_cfqq_idle_window(cfqq);
3117 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003118}
3119
Jens Axboe22e2c502005-06-27 10:55:12 +02003120/*
3121 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3122 * no or if we aren't sure, a 1 will cause a preempt.
3123 */
Jens Axboea6151c32009-10-07 20:02:57 +02003124static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003125cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003126 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003127{
Jens Axboe6d048f52007-04-25 12:44:27 +02003128 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003129
Jens Axboe6d048f52007-04-25 12:44:27 +02003130 cfqq = cfqd->active_queue;
3131 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003132 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003133
Jens Axboe6d048f52007-04-25 12:44:27 +02003134 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003135 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003136
3137 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003138 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003139
Jens Axboe22e2c502005-06-27 10:55:12 +02003140 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003141 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3142 */
3143 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3144 return false;
3145
3146 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003147 * if the new request is sync, but the currently running queue is
3148 * not, let the sync request have priority.
3149 */
Jens Axboe5e705372006-07-13 12:39:25 +02003150 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003151 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003152
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003153 if (new_cfqq->cfqg != cfqq->cfqg)
3154 return false;
3155
3156 if (cfq_slice_used(cfqq))
3157 return true;
3158
3159 /* Allow preemption only if we are idling on sync-noidle tree */
3160 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3161 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3162 new_cfqq->service_tree->count == 2 &&
3163 RB_EMPTY_ROOT(&cfqq->sort_list))
3164 return true;
3165
Jens Axboe374f84a2006-07-23 01:42:19 +02003166 /*
3167 * So both queues are sync. Let the new request get disk time if
3168 * it's a metadata request and the current queue is doing regular IO.
3169 */
3170 if (rq_is_meta(rq) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003171 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003172
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003173 /*
3174 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3175 */
3176 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003177 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003178
Jens Axboe1e3335d2007-02-14 19:59:49 +01003179 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003180 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003181
3182 /*
3183 * if this request is as-good as one we would expect from the
3184 * current cfqq, let it preempt
3185 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003186 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003187 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003188
Jens Axboea6151c32009-10-07 20:02:57 +02003189 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003190}
3191
3192/*
3193 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3194 * let it have half of its nominal slice.
3195 */
3196static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3197{
Jens Axboe7b679132008-05-30 12:23:07 +02003198 cfq_log_cfqq(cfqd, cfqq, "preempt");
Vivek Goyale5ff0822010-04-26 19:25:11 +02003199 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003200
Jens Axboebf572252006-07-19 20:29:12 +02003201 /*
3202 * Put the new queue at the front of the of the current list,
3203 * so we know that it will be selected next.
3204 */
3205 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003206
3207 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02003208
Jens Axboe44f7c162007-01-19 11:51:58 +11003209 cfqq->slice_end = 0;
3210 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003211}
3212
3213/*
Jens Axboe5e705372006-07-13 12:39:25 +02003214 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003215 * something we should do about it
3216 */
3217static void
Jens Axboe5e705372006-07-13 12:39:25 +02003218cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3219 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003220{
Jens Axboe5e705372006-07-13 12:39:25 +02003221 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003222
Aaron Carroll45333d52008-08-26 15:52:36 +02003223 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003224 if (rq_is_meta(rq))
3225 cfqq->meta_pending++;
3226
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003227 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003228 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003229 cfq_update_idle_window(cfqd, cfqq, cic);
3230
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003231 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003232
3233 if (cfqq == cfqd->active_queue) {
3234 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003235 * Remember that we saw a request from this process, but
3236 * don't start queuing just yet. Otherwise we risk seeing lots
3237 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003238 * and merging. If the request is already larger than a single
3239 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003240 * merging is already done. Ditto for a busy system that
3241 * has other work pending, don't risk delaying until the
3242 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003243 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003244 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003245 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3246 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003247 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003248 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003249 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003250 } else {
3251 blkiocg_update_idle_time_stats(
3252 &cfqq->cfqg->blkg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003253 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003254 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003255 }
Jens Axboe5e705372006-07-13 12:39:25 +02003256 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003257 /*
3258 * not the active queue - expire current slice if it is
3259 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003260 * has some old slice time left and is of higher priority or
3261 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003262 */
3263 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09003264 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003265 }
3266}
3267
Jens Axboe165125e2007-07-24 09:28:11 +02003268static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003269{
Jens Axboeb4878f22005-10-20 16:42:29 +02003270 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003271 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003272
Jens Axboe7b679132008-05-30 12:23:07 +02003273 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003274 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275
Jens Axboe30996f42009-10-05 11:03:39 +02003276 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003277 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003278 cfq_add_rq_rb(rq);
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003279 blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003280 &cfqd->serving_group->blkg, rq_data_dir(rq),
3281 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003282 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283}
3284
Aaron Carroll45333d52008-08-26 15:52:36 +02003285/*
3286 * Update hw_tag based on peak queue depth over 50 samples under
3287 * sufficient load.
3288 */
3289static void cfq_update_hw_tag(struct cfq_data *cfqd)
3290{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003291 struct cfq_queue *cfqq = cfqd->active_queue;
3292
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003293 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3294 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003295
3296 if (cfqd->hw_tag == 1)
3297 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003298
3299 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003300 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003301 return;
3302
Shaohua Li1a1238a2009-10-27 08:46:23 +01003303 /*
3304 * If active queue hasn't enough requests and can idle, cfq might not
3305 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3306 * case
3307 */
3308 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3309 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003310 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003311 return;
3312
Aaron Carroll45333d52008-08-26 15:52:36 +02003313 if (cfqd->hw_tag_samples++ < 50)
3314 return;
3315
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003316 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003317 cfqd->hw_tag = 1;
3318 else
3319 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003320}
3321
Vivek Goyal7667aa02009-12-08 17:52:58 -05003322static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3323{
3324 struct cfq_io_context *cic = cfqd->active_cic;
3325
3326 /* If there are other queues in the group, don't wait */
3327 if (cfqq->cfqg->nr_cfqq > 1)
3328 return false;
3329
3330 if (cfq_slice_used(cfqq))
3331 return true;
3332
3333 /* if slice left is less than think time, wait busy */
3334 if (cic && sample_valid(cic->ttime_samples)
3335 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3336 return true;
3337
3338 /*
3339 * If think times is less than a jiffy than ttime_mean=0 and above
3340 * will not be true. It might happen that slice has not expired yet
3341 * but will expire soon (4-5 ns) during select_queue(). To cover the
3342 * case where think time is less than a jiffy, mark the queue wait
3343 * busy if only 1 jiffy is left in the slice.
3344 */
3345 if (cfqq->slice_end - jiffies == 1)
3346 return true;
3347
3348 return false;
3349}
3350
Jens Axboe165125e2007-07-24 09:28:11 +02003351static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352{
Jens Axboe5e705372006-07-13 12:39:25 +02003353 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003354 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003355 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003356 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357
Jens Axboeb4878f22005-10-20 16:42:29 +02003358 now = jiffies;
Vivek Goyal2868ef72009-12-03 12:59:48 -05003359 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360
Aaron Carroll45333d52008-08-26 15:52:36 +02003361 cfq_update_hw_tag(cfqd);
3362
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003363 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003364 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003365 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003366 cfqq->dispatched--;
Divyesh Shah84c124d2010-04-09 08:31:19 +02003367 blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3368 rq_io_start_time_ns(rq), rq_data_dir(rq),
3369 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003371 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003372
Vivek Goyal365722b2009-10-03 15:21:27 +02003373 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003374 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003375 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3376 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003377 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003378
3379 /*
3380 * If this is the active queue, check if it needs to be expired,
3381 * or if we want to idle in case it has no pending requests.
3382 */
3383 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003384 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3385
Jens Axboe44f7c162007-01-19 11:51:58 +11003386 if (cfq_cfqq_slice_new(cfqq)) {
3387 cfq_set_prio_slice(cfqd, cfqq);
3388 cfq_clear_cfqq_slice_new(cfqq);
3389 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003390
3391 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003392 * Should we wait for next request to come in before we expire
3393 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003394 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003395 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyalf75edf22009-12-03 12:59:53 -05003396 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3397 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003398 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003399 }
3400
Jens Axboea36e71f2009-04-15 12:15:11 +02003401 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003402 * Idling is not enabled on:
3403 * - expired queues
3404 * - idle-priority queues
3405 * - async queues
3406 * - queues with still some requests queued
3407 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003408 */
Jens Axboe08717142008-01-28 11:38:15 +01003409 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003410 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003411 else if (sync && cfqq_empty &&
3412 !cfq_close_cooperator(cfqd, cfqq)) {
3413 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3414 /*
3415 * Idling is enabled for SYNC_WORKLOAD.
3416 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3417 * only if we processed at least one !rq_noidle request
3418 */
3419 if (cfqd->serving_type == SYNC_WORKLOAD
Vivek Goyalc04645e2009-12-03 12:59:56 -05003420 || cfqd->noidle_tree_requires_idle
3421 || cfqq->cfqg->nr_cfqq == 1)
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003422 cfq_arm_slice_timer(cfqd);
3423 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003424 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003425
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003426 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003427 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428}
3429
Jens Axboe22e2c502005-06-27 10:55:12 +02003430/*
3431 * we temporarily boost lower priority queues if they are holding fs exclusive
3432 * resources. they are boosted to normal prio (CLASS_BE/4)
3433 */
3434static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435{
Jens Axboe22e2c502005-06-27 10:55:12 +02003436 if (has_fs_excl()) {
3437 /*
3438 * boost idle prio on transactions that would lock out other
3439 * users of the filesystem
3440 */
3441 if (cfq_class_idle(cfqq))
3442 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3443 if (cfqq->ioprio > IOPRIO_NORM)
3444 cfqq->ioprio = IOPRIO_NORM;
3445 } else {
3446 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003447 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003448 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003449 cfqq->ioprio_class = cfqq->org_ioprio_class;
3450 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003452}
3453
Jens Axboe89850f72006-07-22 16:48:31 +02003454static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003455{
Jens Axboe1b379d82009-08-11 08:26:11 +02003456 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003457 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003458 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003459 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003460
3461 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003462}
3463
Jens Axboe165125e2007-07-24 09:28:11 +02003464static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003465{
3466 struct cfq_data *cfqd = q->elevator->elevator_data;
3467 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003468 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003469 struct cfq_queue *cfqq;
3470
3471 /*
3472 * don't force setup of a queue from here, as a call to may_queue
3473 * does not necessarily imply that a request actually will be queued.
3474 * so just lookup a possibly existing queue, or return 'may queue'
3475 * if that fails
3476 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003477 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003478 if (!cic)
3479 return ELV_MQUEUE_MAY;
3480
Jens Axboeb0b78f82009-04-08 10:56:08 +02003481 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003482 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003483 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003484 cfq_prio_boost(cfqq);
3485
Jens Axboe89850f72006-07-22 16:48:31 +02003486 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003487 }
3488
3489 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490}
3491
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492/*
3493 * queue lock held here
3494 */
Jens Axboebb37b942006-12-01 10:42:33 +01003495static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496{
Jens Axboe5e705372006-07-13 12:39:25 +02003497 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498
Jens Axboe5e705372006-07-13 12:39:25 +02003499 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003500 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501
Jens Axboe22e2c502005-06-27 10:55:12 +02003502 BUG_ON(!cfqq->allocated[rw]);
3503 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504
Jens Axboe5e705372006-07-13 12:39:25 +02003505 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02003508 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003510 /* Put down rq reference on cfqg */
3511 cfq_put_cfqg(RQ_CFQG(rq));
3512 rq->elevator_private3 = NULL;
3513
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 cfq_put_queue(cfqq);
3515 }
3516}
3517
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003518static struct cfq_queue *
3519cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3520 struct cfq_queue *cfqq)
3521{
3522 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3523 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003524 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003525 cfq_put_queue(cfqq);
3526 return cic_to_cfqq(cic, 1);
3527}
3528
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003529/*
3530 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3531 * was the last process referring to said cfqq.
3532 */
3533static struct cfq_queue *
3534split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3535{
3536 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003537 cfqq->pid = current->pid;
3538 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003539 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003540 return cfqq;
3541 }
3542
3543 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003544
3545 cfq_put_cooperator(cfqq);
3546
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003547 cfq_put_queue(cfqq);
3548 return NULL;
3549}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003551 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003553static int
Jens Axboe165125e2007-07-24 09:28:11 +02003554cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555{
3556 struct cfq_data *cfqd = q->elevator->elevator_data;
3557 struct cfq_io_context *cic;
3558 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003559 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003560 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561 unsigned long flags;
3562
3563 might_sleep_if(gfp_mask & __GFP_WAIT);
3564
Jens Axboee2d74ac2006-03-28 08:59:01 +02003565 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003566
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 spin_lock_irqsave(q->queue_lock, flags);
3568
Jens Axboe22e2c502005-06-27 10:55:12 +02003569 if (!cic)
3570 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003572new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003573 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003574 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003575 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003576 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003577 } else {
3578 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003579 * If the queue was seeky for too long, break it apart.
3580 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003581 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003582 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3583 cfqq = split_cfqq(cic, cfqq);
3584 if (!cfqq)
3585 goto new_queue;
3586 }
3587
3588 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003589 * Check to see if this queue is scheduled to merge with
3590 * another, closely cooperating queue. The merging of
3591 * queues happens here as it must be done in process context.
3592 * The reference on new_cfqq was taken in merge_cfqqs.
3593 */
3594 if (cfqq->new_cfqq)
3595 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003596 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597
3598 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02003599 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02003600
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601 spin_unlock_irqrestore(q->queue_lock, flags);
3602
Jens Axboe5e705372006-07-13 12:39:25 +02003603 rq->elevator_private = cic;
3604 rq->elevator_private2 = cfqq;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003605 rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
Jens Axboe5e705372006-07-13 12:39:25 +02003606 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003607
Jens Axboe22e2c502005-06-27 10:55:12 +02003608queue_fail:
3609 if (cic)
3610 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003611
Jens Axboe23e018a2009-10-05 08:52:35 +02003612 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003613 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003614 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615 return 1;
3616}
3617
David Howells65f27f32006-11-22 14:55:48 +00003618static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003619{
David Howells65f27f32006-11-22 14:55:48 +00003620 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003621 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003622 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003623
Jens Axboe40bb54d2009-04-15 12:11:10 +02003624 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09003625 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003626 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003627}
3628
3629/*
3630 * Timer running if the active_queue is currently idling inside its time slice
3631 */
3632static void cfq_idle_slice_timer(unsigned long data)
3633{
3634 struct cfq_data *cfqd = (struct cfq_data *) data;
3635 struct cfq_queue *cfqq;
3636 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003637 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003638
Jens Axboe7b679132008-05-30 12:23:07 +02003639 cfq_log(cfqd, "idle timer fired");
3640
Jens Axboe22e2c502005-06-27 10:55:12 +02003641 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3642
Jens Axboefe094d92008-01-31 13:08:54 +01003643 cfqq = cfqd->active_queue;
3644 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003645 timed_out = 0;
3646
Jens Axboe22e2c502005-06-27 10:55:12 +02003647 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003648 * We saw a request before the queue expired, let it through
3649 */
3650 if (cfq_cfqq_must_dispatch(cfqq))
3651 goto out_kick;
3652
3653 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003654 * expired
3655 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003656 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003657 goto expire;
3658
3659 /*
3660 * only expire and reinvoke request handler, if there are
3661 * other queues with pending requests
3662 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003663 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003664 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003665
3666 /*
3667 * not expired and it has a request pending, let it dispatch
3668 */
Jens Axboe75e50982009-04-07 08:56:14 +02003669 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003670 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003671
3672 /*
3673 * Queue depth flag is reset only when the idle didn't succeed
3674 */
3675 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003676 }
3677expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003678 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003679out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003680 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003681out_cont:
3682 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3683}
3684
Jens Axboe3b181522005-06-27 10:56:24 +02003685static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3686{
3687 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003688 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003689}
Jens Axboe22e2c502005-06-27 10:55:12 +02003690
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003691static void cfq_put_async_queues(struct cfq_data *cfqd)
3692{
3693 int i;
3694
3695 for (i = 0; i < IOPRIO_BE_NR; i++) {
3696 if (cfqd->async_cfqq[0][i])
3697 cfq_put_queue(cfqd->async_cfqq[0][i]);
3698 if (cfqd->async_cfqq[1][i])
3699 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003700 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003701
3702 if (cfqd->async_idle_cfqq)
3703 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003704}
3705
Jens Axboebb729bc2009-12-06 09:54:19 +01003706static void cfq_cfqd_free(struct rcu_head *head)
3707{
3708 kfree(container_of(head, struct cfq_data, rcu));
3709}
3710
Jens Axboeb374d182008-10-31 10:05:07 +01003711static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712{
Jens Axboe22e2c502005-06-27 10:55:12 +02003713 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003714 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003715
Jens Axboe3b181522005-06-27 10:56:24 +02003716 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003717
Al Virod9ff4182006-03-18 13:51:22 -05003718 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003719
Al Virod9ff4182006-03-18 13:51:22 -05003720 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003721 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003722
3723 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003724 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3725 struct cfq_io_context,
3726 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003727
3728 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003729 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003730
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003731 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003732 cfq_release_cfq_groups(cfqd);
3733 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003734
Al Virod9ff4182006-03-18 13:51:22 -05003735 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003736
3737 cfq_shutdown_timer_wq(cfqd);
3738
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003739 spin_lock(&cic_index_lock);
3740 ida_remove(&cic_index_ida, cfqd->cic_index);
3741 spin_unlock(&cic_index_lock);
3742
Vivek Goyalb1c35762009-12-03 12:59:47 -05003743 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003744 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745}
3746
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003747static int cfq_alloc_cic_index(void)
3748{
3749 int index, error;
3750
3751 do {
3752 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3753 return -ENOMEM;
3754
3755 spin_lock(&cic_index_lock);
3756 error = ida_get_new(&cic_index_ida, &index);
3757 spin_unlock(&cic_index_lock);
3758 if (error && error != -EAGAIN)
3759 return error;
3760 } while (error);
3761
3762 return index;
3763}
3764
Jens Axboe165125e2007-07-24 09:28:11 +02003765static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766{
3767 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003768 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003769 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003770 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003772 i = cfq_alloc_cic_index();
3773 if (i < 0)
3774 return NULL;
3775
Christoph Lameter94f60302007-07-17 04:03:29 -07003776 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003778 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003780 cfqd->cic_index = i;
3781
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003782 /* Init root service tree */
3783 cfqd->grp_service_tree = CFQ_RB_ROOT;
3784
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003785 /* Init root group */
3786 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003787 for_each_cfqg_st(cfqg, i, j, st)
3788 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003789 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003790
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003791 /* Give preference to root group over other groups */
3792 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3793
Vivek Goyal25fb5162009-12-03 12:59:46 -05003794#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003795 /*
3796 * Take a reference to root group which we never drop. This is just
3797 * to make sure that cfq_put_cfqg() does not try to kfree root group
3798 */
3799 atomic_set(&cfqg->ref, 1);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003800 rcu_read_lock();
Vivek Goyal22084192009-12-03 12:59:49 -05003801 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3802 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003803 rcu_read_unlock();
Vivek Goyal25fb5162009-12-03 12:59:46 -05003804#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003805 /*
3806 * Not strictly needed (since RB_ROOT just clears the node and we
3807 * zeroed cfqd on alloc), but better be safe in case someone decides
3808 * to add magic to the rb code
3809 */
3810 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3811 cfqd->prio_trees[i] = RB_ROOT;
3812
Jens Axboe6118b702009-06-30 09:34:12 +02003813 /*
3814 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3815 * Grab a permanent reference to it, so that the normal code flow
3816 * will not attempt to free it.
3817 */
3818 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3819 atomic_inc(&cfqd->oom_cfqq.ref);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003820 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003821
Al Virod9ff4182006-03-18 13:51:22 -05003822 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825
Jens Axboe22e2c502005-06-27 10:55:12 +02003826 init_timer(&cfqd->idle_slice_timer);
3827 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3828 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3829
Jens Axboe23e018a2009-10-05 08:52:35 +02003830 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003831
Linus Torvalds1da177e2005-04-16 15:20:36 -07003832 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003833 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3834 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003835 cfqd->cfq_back_max = cfq_back_max;
3836 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003837 cfqd->cfq_slice[0] = cfq_slice_async;
3838 cfqd->cfq_slice[1] = cfq_slice_sync;
3839 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3840 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003841 cfqd->cfq_latency = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05003842 cfqd->cfq_group_isolation = 0;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003843 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003844 /*
3845 * we optimistically start assuming sync ops weren't delayed in last
3846 * second, in order to have larger depth for async operations.
3847 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003848 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003849 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850}
3851
3852static void cfq_slab_kill(void)
3853{
Jens Axboed6de8be2008-05-28 14:46:59 +02003854 /*
3855 * Caller already ensured that pending RCU callbacks are completed,
3856 * so we should have no busy allocations at this point.
3857 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858 if (cfq_pool)
3859 kmem_cache_destroy(cfq_pool);
3860 if (cfq_ioc_pool)
3861 kmem_cache_destroy(cfq_ioc_pool);
3862}
3863
3864static int __init cfq_slab_setup(void)
3865{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003866 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867 if (!cfq_pool)
3868 goto fail;
3869
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003870 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871 if (!cfq_ioc_pool)
3872 goto fail;
3873
3874 return 0;
3875fail:
3876 cfq_slab_kill();
3877 return -ENOMEM;
3878}
3879
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880/*
3881 * sysfs parts below -->
3882 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883static ssize_t
3884cfq_var_show(unsigned int var, char *page)
3885{
3886 return sprintf(page, "%d\n", var);
3887}
3888
3889static ssize_t
3890cfq_var_store(unsigned int *var, const char *page, size_t count)
3891{
3892 char *p = (char *) page;
3893
3894 *var = simple_strtoul(p, &p, 10);
3895 return count;
3896}
3897
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003899static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003901 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902 unsigned int __data = __VAR; \
3903 if (__CONV) \
3904 __data = jiffies_to_msecs(__data); \
3905 return cfq_var_show(__data, (page)); \
3906}
3907SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003908SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3909SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003910SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3911SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003912SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3913SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3914SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3915SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003916SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003917SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918#undef SHOW_FUNCTION
3919
3920#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003921static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003923 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924 unsigned int __data; \
3925 int ret = cfq_var_store(&__data, (page), count); \
3926 if (__data < (MIN)) \
3927 __data = (MIN); \
3928 else if (__data > (MAX)) \
3929 __data = (MAX); \
3930 if (__CONV) \
3931 *(__PTR) = msecs_to_jiffies(__data); \
3932 else \
3933 *(__PTR) = __data; \
3934 return ret; \
3935}
3936STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003937STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3938 UINT_MAX, 1);
3939STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3940 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05003941STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003942STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3943 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003944STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3945STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3946STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01003947STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3948 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003949STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003950STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951#undef STORE_FUNCTION
3952
Al Viroe572ec72006-03-18 22:27:18 -05003953#define CFQ_ATTR(name) \
3954 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02003955
Al Viroe572ec72006-03-18 22:27:18 -05003956static struct elv_fs_entry cfq_attrs[] = {
3957 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05003958 CFQ_ATTR(fifo_expire_sync),
3959 CFQ_ATTR(fifo_expire_async),
3960 CFQ_ATTR(back_seek_max),
3961 CFQ_ATTR(back_seek_penalty),
3962 CFQ_ATTR(slice_sync),
3963 CFQ_ATTR(slice_async),
3964 CFQ_ATTR(slice_async_rq),
3965 CFQ_ATTR(slice_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02003966 CFQ_ATTR(low_latency),
Vivek Goyalae30c282009-12-03 12:59:55 -05003967 CFQ_ATTR(group_isolation),
Al Viroe572ec72006-03-18 22:27:18 -05003968 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969};
3970
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971static struct elevator_type iosched_cfq = {
3972 .ops = {
3973 .elevator_merge_fn = cfq_merge,
3974 .elevator_merged_fn = cfq_merged_request,
3975 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01003976 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07003977 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02003978 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02003980 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981 .elevator_deactivate_req_fn = cfq_deactivate_request,
3982 .elevator_queue_empty_fn = cfq_queue_empty,
3983 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02003984 .elevator_former_req_fn = elv_rb_former_request,
3985 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986 .elevator_set_req_fn = cfq_set_request,
3987 .elevator_put_req_fn = cfq_put_request,
3988 .elevator_may_queue_fn = cfq_may_queue,
3989 .elevator_init_fn = cfq_init_queue,
3990 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02003991 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003992 },
Al Viro3d1ab402006-03-18 18:35:43 -05003993 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994 .elevator_name = "cfq",
3995 .elevator_owner = THIS_MODULE,
3996};
3997
Vivek Goyal3e252062009-12-04 10:36:42 -05003998#ifdef CONFIG_CFQ_GROUP_IOSCHED
3999static struct blkio_policy_type blkio_policy_cfq = {
4000 .ops = {
4001 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4002 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4003 },
4004};
4005#else
4006static struct blkio_policy_type blkio_policy_cfq;
4007#endif
4008
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009static int __init cfq_init(void)
4010{
Jens Axboe22e2c502005-06-27 10:55:12 +02004011 /*
4012 * could be 0 on HZ < 1000 setups
4013 */
4014 if (!cfq_slice_async)
4015 cfq_slice_async = 1;
4016 if (!cfq_slice_idle)
4017 cfq_slice_idle = 1;
4018
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019 if (cfq_slab_setup())
4020 return -ENOMEM;
4021
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004022 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05004023 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004024
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004025 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026}
4027
4028static void __exit cfq_exit(void)
4029{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07004030 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05004031 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05004033 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02004034 /* ioc_gone's update must be visible before reading ioc_count */
4035 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02004036
4037 /*
4038 * this also protects us from entering cfq_slab_kill() with
4039 * pending RCU callbacks
4040 */
Tejun Heo245b2e72009-06-24 15:13:48 +09004041 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02004042 wait_for_completion(&all_gone);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004043 ida_destroy(&cic_index_ida);
Christoph Hellwig83521d32005-10-30 15:01:39 -08004044 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004045}
4046
4047module_init(cfq_init);
4048module_exit(cfq_exit);
4049
4050MODULE_AUTHOR("Jens Axboe");
4051MODULE_LICENSE("GPL");
4052MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");