blob: eb4086f7dfef9eb7efc6d202fb7a979919a551b8 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * CFQ, or complete fairness queueing, disk scheduler.
3 *
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6 *
Jens Axboe0fe23472006-09-04 15:41:16 +02007 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090010#include <linux/slab.h>
Al Viro1cc9be62006-03-18 12:29:52 -050011#include <linux/blkdev.h>
12#include <linux/elevator.h>
Randy Dunlapad5ebd22009-11-11 13:47:45 +010013#include <linux/jiffies.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020015#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020016#include <linux/blktrace_api.h>
Vivek Goyale98ef892010-06-18 10:39:47 -040017#include "cfq.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070018
19/*
20 * tunables
21 */
Jens Axboefe094d92008-01-31 13:08:54 +010022/* max queue in one round of service */
Shaohua Liabc3c742010-03-01 09:20:54 +010023static const int cfq_quantum = 8;
Arjan van de Ven64100092006-01-06 09:46:02 +010024static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010025/* maximum backwards seek, in KiB */
26static const int cfq_back_max = 16 * 1024;
27/* penalty of a backwards seek */
28static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010029static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020030static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010031static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020032static int cfq_slice_idle = HZ / 125;
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{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200461 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
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;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200649 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200650 return rq1;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200651 else if ((rq2->cmd_flags & REQ_META) &&
652 !(rq1->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200653 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
Tejun Heo83096eb2009-05-07 22:24:39 +0900655 s1 = blk_rq_pos(rq1);
656 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 /*
659 * by definition, 1KiB is 2 sectors
660 */
661 back_max = cfqd->cfq_back_max * 2;
662
663 /*
664 * Strict one way elevator _except_ in the case where we allow
665 * short backward seeks which are biased as twice the cost of a
666 * similar forward seek.
667 */
668 if (s1 >= last)
669 d1 = s1 - last;
670 else if (s1 + back_max >= last)
671 d1 = (last - s1) * cfqd->cfq_back_penalty;
672 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200673 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674
675 if (s2 >= last)
676 d2 = s2 - last;
677 else if (s2 + back_max >= last)
678 d2 = (last - s2) * cfqd->cfq_back_penalty;
679 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200680 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681
682 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
Andreas Mohre8a99052006-03-28 08:59:49 +0200684 /*
685 * By doing switch() on the bit mask "wrap" we avoid having to
686 * check two variables for all permutations: --> faster!
687 */
688 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200689 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200690 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200691 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200692 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200693 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200694 else {
695 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200696 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200697 else
Jens Axboe5e705372006-07-13 12:39:25 +0200698 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200699 }
700
701 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200702 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200703 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200704 return rq2;
705 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200706 default:
707 /*
708 * Since both rqs are wrapped,
709 * start with the one that's further behind head
710 * (--> only *one* back seek required),
711 * since back seek takes more time than forward.
712 */
713 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200714 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 else
Jens Axboe5e705372006-07-13 12:39:25 +0200716 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 }
718}
719
Jens Axboe498d3aa22007-04-26 12:54:48 +0200720/*
721 * The below is leftmost cache rbtree addon
722 */
Jens Axboe08717142008-01-28 11:38:15 +0100723static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200724{
Vivek Goyal615f0252009-12-03 12:59:39 -0500725 /* Service tree is empty */
726 if (!root->count)
727 return NULL;
728
Jens Axboecc09e292007-04-26 12:53:50 +0200729 if (!root->left)
730 root->left = rb_first(&root->rb);
731
Jens Axboe08717142008-01-28 11:38:15 +0100732 if (root->left)
733 return rb_entry(root->left, struct cfq_queue, rb_node);
734
735 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200736}
737
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500738static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
739{
740 if (!root->left)
741 root->left = rb_first(&root->rb);
742
743 if (root->left)
744 return rb_entry_cfqg(root->left);
745
746 return NULL;
747}
748
Jens Axboea36e71f2009-04-15 12:15:11 +0200749static void rb_erase_init(struct rb_node *n, struct rb_root *root)
750{
751 rb_erase(n, root);
752 RB_CLEAR_NODE(n);
753}
754
Jens Axboecc09e292007-04-26 12:53:50 +0200755static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
756{
757 if (root->left == n)
758 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200759 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100760 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200761}
762
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763/*
764 * would be nice to take fifo expire time into account as well
765 */
Jens Axboe5e705372006-07-13 12:39:25 +0200766static struct request *
767cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
768 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769{
Jens Axboe21183b02006-07-13 12:33:14 +0200770 struct rb_node *rbnext = rb_next(&last->rb_node);
771 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200772 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773
Jens Axboe21183b02006-07-13 12:33:14 +0200774 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775
776 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200777 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200778
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200780 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200781 else {
782 rbnext = rb_first(&cfqq->sort_list);
783 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200784 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200785 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100787 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788}
789
Jens Axboed9e76202007-04-20 14:27:50 +0200790static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
791 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792{
Jens Axboed9e76202007-04-20 14:27:50 +0200793 /*
794 * just an approximation, should be ok.
795 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500796 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100797 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200798}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500800static inline s64
801cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
802{
803 return cfqg->vdisktime - st->min_vdisktime;
804}
805
806static void
807__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
808{
809 struct rb_node **node = &st->rb.rb_node;
810 struct rb_node *parent = NULL;
811 struct cfq_group *__cfqg;
812 s64 key = cfqg_key(st, cfqg);
813 int left = 1;
814
815 while (*node != NULL) {
816 parent = *node;
817 __cfqg = rb_entry_cfqg(parent);
818
819 if (key < cfqg_key(st, __cfqg))
820 node = &parent->rb_left;
821 else {
822 node = &parent->rb_right;
823 left = 0;
824 }
825 }
826
827 if (left)
828 st->left = &cfqg->rb_node;
829
830 rb_link_node(&cfqg->rb_node, parent, node);
831 rb_insert_color(&cfqg->rb_node, &st->rb);
832}
833
834static void
835cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
836{
837 struct cfq_rb_root *st = &cfqd->grp_service_tree;
838 struct cfq_group *__cfqg;
839 struct rb_node *n;
840
841 cfqg->nr_cfqq++;
842 if (cfqg->on_st)
843 return;
844
845 /*
846 * Currently put the group at the end. Later implement something
847 * so that groups get lesser vtime based on their weights, so that
848 * if group does not loose all if it was not continously backlogged.
849 */
850 n = rb_last(&st->rb);
851 if (n) {
852 __cfqg = rb_entry_cfqg(n);
853 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
854 } else
855 cfqg->vdisktime = st->min_vdisktime;
856
857 __cfq_group_service_tree_add(st, cfqg);
858 cfqg->on_st = true;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500859 st->total_weight += cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500860}
861
862static void
863cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
864{
865 struct cfq_rb_root *st = &cfqd->grp_service_tree;
866
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500867 if (st->active == &cfqg->rb_node)
868 st->active = NULL;
869
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500870 BUG_ON(cfqg->nr_cfqq < 1);
871 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500872
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500873 /* If there are other cfq queues under this group, don't delete it */
874 if (cfqg->nr_cfqq)
875 return;
876
Vivek Goyal2868ef72009-12-03 12:59:48 -0500877 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500878 cfqg->on_st = false;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500879 st->total_weight -= cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500880 if (!RB_EMPTY_NODE(&cfqg->rb_node))
881 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500882 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400883 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500884}
885
886static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
887{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500888 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500889
890 /*
891 * Queue got expired before even a single request completed or
892 * got expired immediately after first request completion.
893 */
894 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
895 /*
896 * Also charge the seek time incurred to the group, otherwise
897 * if there are mutiple queues in the group, each can dispatch
898 * a single request on seeky media and cause lots of seek time
899 * and group will never know it.
900 */
901 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
902 1);
903 } else {
904 slice_used = jiffies - cfqq->slice_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500905 if (slice_used > cfqq->allocated_slice)
906 slice_used = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500907 }
908
Divyesh Shah9a0785b2010-04-01 15:01:04 -0700909 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500910 return slice_used;
911}
912
913static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200914 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500915{
916 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500917 unsigned int used_sl, charge_sl;
918 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
919 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500920
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500921 BUG_ON(nr_sync < 0);
922 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
923
924 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
925 charge_sl = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500926
927 /* Can't update vdisktime while group is on service tree */
928 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500929 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500930 __cfq_group_service_tree_add(st, cfqg);
931
932 /* This group is being expired. Save the context */
933 if (time_after(cfqd->workload_expires, jiffies)) {
934 cfqg->saved_workload_slice = cfqd->workload_expires
935 - jiffies;
936 cfqg->saved_workload = cfqd->serving_type;
937 cfqg->saved_serving_prio = cfqd->serving_prio;
938 } else
939 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500940
941 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
942 st->min_vdisktime);
Vivek Goyale98ef892010-06-18 10:39:47 -0400943 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
944 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500945}
946
Vivek Goyal25fb5162009-12-03 12:59:46 -0500947#ifdef CONFIG_CFQ_GROUP_IOSCHED
948static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
949{
950 if (blkg)
951 return container_of(blkg, struct cfq_group, blkg);
952 return NULL;
953}
954
Vivek Goyalf8d461d2009-12-03 12:59:52 -0500955void
956cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
957{
958 cfqg_of_blkg(blkg)->weight = weight;
959}
960
Vivek Goyal25fb5162009-12-03 12:59:46 -0500961static struct cfq_group *
962cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
963{
964 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
965 struct cfq_group *cfqg = NULL;
966 void *key = cfqd;
967 int i, j;
968 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -0500969 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
970 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500971
Vivek Goyal25fb5162009-12-03 12:59:46 -0500972 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
Ricky Beniteza74b2ad2010-04-05 18:22:17 +0200973 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
974 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
975 cfqg->blkg.dev = MKDEV(major, minor);
976 goto done;
977 }
Vivek Goyal25fb5162009-12-03 12:59:46 -0500978 if (cfqg || !create)
979 goto done;
980
981 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
982 if (!cfqg)
983 goto done;
984
Vivek Goyal25fb5162009-12-03 12:59:46 -0500985 for_each_cfqg_st(cfqg, i, j, st)
986 *st = CFQ_RB_ROOT;
987 RB_CLEAR_NODE(&cfqg->rb_node);
988
Vivek Goyalb1c35762009-12-03 12:59:47 -0500989 /*
990 * Take the initial reference that will be released on destroy
991 * This can be thought of a joint reference by cgroup and
992 * elevator which will be dropped by either elevator exit
993 * or cgroup deletion path depending on who is exiting first.
994 */
995 atomic_set(&cfqg->ref, 1);
996
Vivek Goyal25fb5162009-12-03 12:59:46 -0500997 /* Add group onto cgroup list */
Vivek Goyal22084192009-12-03 12:59:49 -0500998 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
Vivek Goyale98ef892010-06-18 10:39:47 -0400999 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
Vivek Goyal22084192009-12-03 12:59:49 -05001000 MKDEV(major, minor));
Gui Jianfeng34d0f172010-04-13 16:05:49 +08001001 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001002
1003 /* Add group on cfqd list */
1004 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1005
1006done:
Vivek Goyal25fb5162009-12-03 12:59:46 -05001007 return cfqg;
1008}
1009
1010/*
1011 * Search for the cfq group current task belongs to. If create = 1, then also
1012 * create the cfq group if it does not exist. request_queue lock must be held.
1013 */
1014static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1015{
1016 struct cgroup *cgroup;
1017 struct cfq_group *cfqg = NULL;
1018
1019 rcu_read_lock();
1020 cgroup = task_cgroup(current, blkio_subsys_id);
1021 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1022 if (!cfqg && create)
1023 cfqg = &cfqd->root_group;
1024 rcu_read_unlock();
1025 return cfqg;
1026}
1027
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001028static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1029{
1030 atomic_inc(&cfqg->ref);
1031 return cfqg;
1032}
1033
Vivek Goyal25fb5162009-12-03 12:59:46 -05001034static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1035{
1036 /* Currently, all async queues are mapped to root group */
1037 if (!cfq_cfqq_sync(cfqq))
1038 cfqg = &cfqq->cfqd->root_group;
1039
1040 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001041 /* cfqq reference on cfqg */
1042 atomic_inc(&cfqq->cfqg->ref);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001043}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001044
1045static void cfq_put_cfqg(struct cfq_group *cfqg)
1046{
1047 struct cfq_rb_root *st;
1048 int i, j;
1049
1050 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1051 if (!atomic_dec_and_test(&cfqg->ref))
1052 return;
1053 for_each_cfqg_st(cfqg, i, j, st)
1054 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1055 kfree(cfqg);
1056}
1057
1058static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1059{
1060 /* Something wrong if we are trying to remove same group twice */
1061 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1062
1063 hlist_del_init(&cfqg->cfqd_node);
1064
1065 /*
1066 * Put the reference taken at the time of creation so that when all
1067 * queues are gone, group can be destroyed.
1068 */
1069 cfq_put_cfqg(cfqg);
1070}
1071
1072static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1073{
1074 struct hlist_node *pos, *n;
1075 struct cfq_group *cfqg;
1076
1077 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1078 /*
1079 * If cgroup removal path got to blk_group first and removed
1080 * it from cgroup list, then it will take care of destroying
1081 * cfqg also.
1082 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001083 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001084 cfq_destroy_cfqg(cfqd, cfqg);
1085 }
1086}
1087
1088/*
1089 * Blk cgroup controller notification saying that blkio_group object is being
1090 * delinked as associated cgroup object is going away. That also means that
1091 * no new IO will come in this group. So get rid of this group as soon as
1092 * any pending IO in the group is finished.
1093 *
1094 * This function is called under rcu_read_lock(). key is the rcu protected
1095 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1096 * read lock.
1097 *
1098 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1099 * it should not be NULL as even if elevator was exiting, cgroup deltion
1100 * path got to it first.
1101 */
1102void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1103{
1104 unsigned long flags;
1105 struct cfq_data *cfqd = key;
1106
1107 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1108 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1109 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1110}
1111
Vivek Goyal25fb5162009-12-03 12:59:46 -05001112#else /* GROUP_IOSCHED */
1113static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1114{
1115 return &cfqd->root_group;
1116}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001117
1118static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1119{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001120 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001121}
1122
Vivek Goyal25fb5162009-12-03 12:59:46 -05001123static inline void
1124cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1125 cfqq->cfqg = cfqg;
1126}
1127
Vivek Goyalb1c35762009-12-03 12:59:47 -05001128static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1129static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1130
Vivek Goyal25fb5162009-12-03 12:59:46 -05001131#endif /* GROUP_IOSCHED */
1132
Jens Axboe498d3aa22007-04-26 12:54:48 +02001133/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001134 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001135 * requests waiting to be processed. It is sorted in the order that
1136 * we will service the queues.
1137 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001138static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001139 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001140{
Jens Axboe08717142008-01-28 11:38:15 +01001141 struct rb_node **p, *parent;
1142 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001143 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001144 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001145 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001146 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001147 int group_changed = 0;
1148
1149#ifdef CONFIG_CFQ_GROUP_IOSCHED
1150 if (!cfqd->cfq_group_isolation
1151 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1152 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1153 /* Move this cfq to root group */
1154 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1155 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1156 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1157 cfqq->orig_cfqg = cfqq->cfqg;
1158 cfqq->cfqg = &cfqd->root_group;
1159 atomic_inc(&cfqd->root_group.ref);
1160 group_changed = 1;
1161 } else if (!cfqd->cfq_group_isolation
1162 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1163 /* cfqq is sequential now needs to go to its original group */
1164 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1165 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1166 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1167 cfq_put_cfqg(cfqq->cfqg);
1168 cfqq->cfqg = cfqq->orig_cfqg;
1169 cfqq->orig_cfqg = NULL;
1170 group_changed = 1;
1171 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1172 }
1173#endif
Jens Axboed9e76202007-04-20 14:27:50 +02001174
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001175 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001176 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001177 if (cfq_class_idle(cfqq)) {
1178 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001179 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001180 if (parent && parent != &cfqq->rb_node) {
1181 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1182 rb_key += __cfqq->rb_key;
1183 } else
1184 rb_key += jiffies;
1185 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001186 /*
1187 * Get our rb key offset. Subtract any residual slice
1188 * value carried from last service. A negative resid
1189 * count indicates slice overrun, and this should position
1190 * the next service time further away in the tree.
1191 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001192 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001193 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001194 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001195 } else {
1196 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001197 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001198 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1199 }
Jens Axboed9e76202007-04-20 14:27:50 +02001200
1201 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001202 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001203 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001204 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001205 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001206 if (rb_key == cfqq->rb_key &&
1207 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001208 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001209
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001210 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1211 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001212 }
Jens Axboed9e76202007-04-20 14:27:50 +02001213
Jens Axboe498d3aa22007-04-26 12:54:48 +02001214 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001215 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001216 cfqq->service_tree = service_tree;
1217 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001218 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001219 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001220
Jens Axboed9e76202007-04-20 14:27:50 +02001221 parent = *p;
1222 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1223
Jens Axboe0c534e02007-04-18 20:01:57 +02001224 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001225 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001226 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001227 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001228 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001229 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001230 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001231 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001232 }
Jens Axboe67060e32007-04-18 20:13:32 +02001233
1234 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001235 }
1236
Jens Axboecc09e292007-04-26 12:53:50 +02001237 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001238 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001239
Jens Axboed9e76202007-04-20 14:27:50 +02001240 cfqq->rb_key = rb_key;
1241 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001242 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1243 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001244 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001245 return;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001246 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247}
1248
Jens Axboea36e71f2009-04-15 12:15:11 +02001249static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001250cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1251 sector_t sector, struct rb_node **ret_parent,
1252 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001253{
Jens Axboea36e71f2009-04-15 12:15:11 +02001254 struct rb_node **p, *parent;
1255 struct cfq_queue *cfqq = NULL;
1256
1257 parent = NULL;
1258 p = &root->rb_node;
1259 while (*p) {
1260 struct rb_node **n;
1261
1262 parent = *p;
1263 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1264
1265 /*
1266 * Sort strictly based on sector. Smallest to the left,
1267 * largest to the right.
1268 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001269 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001270 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001271 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001272 n = &(*p)->rb_left;
1273 else
1274 break;
1275 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001276 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001277 }
1278
1279 *ret_parent = parent;
1280 if (rb_link)
1281 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001282 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001283}
1284
1285static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1286{
Jens Axboea36e71f2009-04-15 12:15:11 +02001287 struct rb_node **p, *parent;
1288 struct cfq_queue *__cfqq;
1289
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001290 if (cfqq->p_root) {
1291 rb_erase(&cfqq->p_node, cfqq->p_root);
1292 cfqq->p_root = NULL;
1293 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001294
1295 if (cfq_class_idle(cfqq))
1296 return;
1297 if (!cfqq->next_rq)
1298 return;
1299
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001300 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001301 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1302 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001303 if (!__cfqq) {
1304 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001305 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1306 } else
1307 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001308}
1309
Jens Axboe498d3aa22007-04-26 12:54:48 +02001310/*
1311 * Update cfqq's position in the service tree.
1312 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001313static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001314{
Jens Axboe6d048f52007-04-25 12:44:27 +02001315 /*
1316 * Resorting requires the cfqq to be on the RR list already.
1317 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001318 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001319 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001320 cfq_prio_tree_add(cfqd, cfqq);
1321 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001322}
1323
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324/*
1325 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001326 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327 */
Jens Axboefebffd62008-01-28 13:19:43 +01001328static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329{
Jens Axboe7b679132008-05-30 12:23:07 +02001330 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001331 BUG_ON(cfq_cfqq_on_rr(cfqq));
1332 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 cfqd->busy_queues++;
1334
Jens Axboeedd75ff2007-04-19 12:03:34 +02001335 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336}
1337
Jens Axboe498d3aa22007-04-26 12:54:48 +02001338/*
1339 * Called when the cfqq no longer has requests pending, remove it from
1340 * the service tree.
1341 */
Jens Axboefebffd62008-01-28 13:19:43 +01001342static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343{
Jens Axboe7b679132008-05-30 12:23:07 +02001344 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001345 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1346 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001348 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1349 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1350 cfqq->service_tree = NULL;
1351 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001352 if (cfqq->p_root) {
1353 rb_erase(&cfqq->p_node, cfqq->p_root);
1354 cfqq->p_root = NULL;
1355 }
Jens Axboed9e76202007-04-20 14:27:50 +02001356
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001357 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 BUG_ON(!cfqd->busy_queues);
1359 cfqd->busy_queues--;
1360}
1361
1362/*
1363 * rb tree support functions
1364 */
Jens Axboefebffd62008-01-28 13:19:43 +01001365static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366{
Jens Axboe5e705372006-07-13 12:39:25 +02001367 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001368 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369
Jens Axboeb4878f22005-10-20 16:42:29 +02001370 BUG_ON(!cfqq->queued[sync]);
1371 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
Jens Axboe5e705372006-07-13 12:39:25 +02001373 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374
Vivek Goyalf04a6422009-12-03 12:59:40 -05001375 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1376 /*
1377 * Queue will be deleted from service tree when we actually
1378 * expire it later. Right now just remove it from prio tree
1379 * as it is empty.
1380 */
1381 if (cfqq->p_root) {
1382 rb_erase(&cfqq->p_node, cfqq->p_root);
1383 cfqq->p_root = NULL;
1384 }
1385 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386}
1387
Jens Axboe5e705372006-07-13 12:39:25 +02001388static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389{
Jens Axboe5e705372006-07-13 12:39:25 +02001390 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001392 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393
Jens Axboe5380a102006-07-13 12:37:56 +02001394 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395
1396 /*
1397 * looks a little odd, but the first insert might return an alias.
1398 * if that happens, put the alias on the dispatch list
1399 */
Jens Axboe21183b02006-07-13 12:33:14 +02001400 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001401 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001402
1403 if (!cfq_cfqq_on_rr(cfqq))
1404 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001405
1406 /*
1407 * check if this request is a better next-serve candidate
1408 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001409 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001410 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001411
1412 /*
1413 * adjust priority tree position, if ->next_rq changes
1414 */
1415 if (prev != cfqq->next_rq)
1416 cfq_prio_tree_add(cfqd, cfqq);
1417
Jens Axboe5044eed2007-04-25 11:53:48 +02001418 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419}
1420
Jens Axboefebffd62008-01-28 13:19:43 +01001421static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422{
Jens Axboe5380a102006-07-13 12:37:56 +02001423 elv_rb_del(&cfqq->sort_list, rq);
1424 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001425 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1426 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001427 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001428 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001429 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1430 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431}
1432
Jens Axboe206dc692006-03-28 13:03:44 +02001433static struct request *
1434cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435{
Jens Axboe206dc692006-03-28 13:03:44 +02001436 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001437 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001438 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439
Jens Axboe4ac845a2008-01-24 08:44:49 +01001440 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001441 if (!cic)
1442 return NULL;
1443
1444 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001445 if (cfqq) {
1446 sector_t sector = bio->bi_sector + bio_sectors(bio);
1447
Jens Axboe21183b02006-07-13 12:33:14 +02001448 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001449 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451 return NULL;
1452}
1453
Jens Axboe165125e2007-07-24 09:28:11 +02001454static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001455{
1456 struct cfq_data *cfqd = q->elevator->elevator_data;
1457
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001458 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001459 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001460 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001461
Tejun Heo5b936292009-05-07 22:24:38 +09001462 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001463}
1464
Jens Axboe165125e2007-07-24 09:28:11 +02001465static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466{
Jens Axboe22e2c502005-06-27 10:55:12 +02001467 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001469 WARN_ON(!cfqd->rq_in_driver);
1470 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001471 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001472 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473}
1474
Jens Axboeb4878f22005-10-20 16:42:29 +02001475static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476{
Jens Axboe5e705372006-07-13 12:39:25 +02001477 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001478
Jens Axboe5e705372006-07-13 12:39:25 +02001479 if (cfqq->next_rq == rq)
1480 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481
Jens Axboeb4878f22005-10-20 16:42:29 +02001482 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001483 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001484
Aaron Carroll45333d52008-08-26 15:52:36 +02001485 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001486 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1487 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02001488 if (rq->cmd_flags & REQ_META) {
Jens Axboe374f84a2006-07-23 01:42:19 +02001489 WARN_ON(!cfqq->meta_pending);
1490 cfqq->meta_pending--;
1491 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492}
1493
Jens Axboe165125e2007-07-24 09:28:11 +02001494static int cfq_merge(struct request_queue *q, struct request **req,
1495 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496{
1497 struct cfq_data *cfqd = q->elevator->elevator_data;
1498 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499
Jens Axboe206dc692006-03-28 13:03:44 +02001500 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001501 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001502 *req = __rq;
1503 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504 }
1505
1506 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507}
1508
Jens Axboe165125e2007-07-24 09:28:11 +02001509static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001510 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511{
Jens Axboe21183b02006-07-13 12:33:14 +02001512 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001513 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514
Jens Axboe5e705372006-07-13 12:39:25 +02001515 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517}
1518
Divyesh Shah812d4022010-04-08 21:14:23 -07001519static void cfq_bio_merged(struct request_queue *q, struct request *req,
1520 struct bio *bio)
1521{
Vivek Goyale98ef892010-06-18 10:39:47 -04001522 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1523 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001524}
1525
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526static void
Jens Axboe165125e2007-07-24 09:28:11 +02001527cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 struct request *next)
1529{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001530 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001531 /*
1532 * reposition in fifo if next is older than rq
1533 */
1534 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001535 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001536 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001537 rq_set_fifo_time(rq, rq_fifo_time(next));
1538 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001539
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001540 if (cfqq->next_rq == next)
1541 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001542 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001543 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1544 rq_data_dir(next), rq_is_sync(next));
Jens Axboe22e2c502005-06-27 10:55:12 +02001545}
1546
Jens Axboe165125e2007-07-24 09:28:11 +02001547static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001548 struct bio *bio)
1549{
1550 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001551 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001552 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001553
1554 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001555 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001556 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001557 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001558 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001559
1560 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001561 * Lookup the cfqq that this bio will be queued with. Allow
1562 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001563 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001564 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001565 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001566 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001567
Vasily Tarasov91fac312007-04-25 12:29:51 +02001568 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001569 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001570}
1571
Divyesh Shah812df482010-04-08 21:15:35 -07001572static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1573{
1574 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001575 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001576}
1577
Jens Axboefebffd62008-01-28 13:19:43 +01001578static void __cfq_set_active_queue(struct cfq_data *cfqd,
1579 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001580{
1581 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001582 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1583 cfqd->serving_prio, cfqd->serving_type);
Vivek Goyale98ef892010-06-18 10:39:47 -04001584 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001585 cfqq->slice_start = 0;
1586 cfqq->dispatch_start = jiffies;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001587 cfqq->allocated_slice = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001588 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001589 cfqq->slice_dispatch = 0;
1590
Jens Axboe2f5cb732009-04-07 08:51:19 +02001591 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +02001592 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001593 cfq_clear_cfqq_must_alloc_slice(cfqq);
1594 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +11001595 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001596
Divyesh Shah812df482010-04-08 21:15:35 -07001597 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001598 }
1599
1600 cfqd->active_queue = cfqq;
1601}
1602
1603/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001604 * current cfqq expired its slice (or was too idle), select new one
1605 */
1606static void
1607__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001608 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001609{
Jens Axboe7b679132008-05-30 12:23:07 +02001610 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1611
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001612 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001613 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001614
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001615 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001616 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001617
1618 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001619 * If this cfqq is shared between multiple processes, check to
1620 * make sure that those processes are still issuing I/Os within
1621 * the mean seek distance. If not, it may be time to break the
1622 * queues apart again.
1623 */
1624 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1625 cfq_mark_cfqq_split_coop(cfqq);
1626
1627 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001628 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001629 */
Jens Axboe7b679132008-05-30 12:23:07 +02001630 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +11001631 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001632 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1633 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001634
Vivek Goyale5ff0822010-04-26 19:25:11 +02001635 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001636
Vivek Goyalf04a6422009-12-03 12:59:40 -05001637 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1638 cfq_del_cfqq_rr(cfqd, cfqq);
1639
Jens Axboeedd75ff2007-04-19 12:03:34 +02001640 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001641
1642 if (cfqq == cfqd->active_queue)
1643 cfqd->active_queue = NULL;
1644
Vivek Goyaldae739e2009-12-03 12:59:45 -05001645 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1646 cfqd->grp_service_tree.active = NULL;
1647
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001648 if (cfqd->active_cic) {
1649 put_io_context(cfqd->active_cic->ioc);
1650 cfqd->active_cic = NULL;
1651 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001652}
1653
Vivek Goyale5ff0822010-04-26 19:25:11 +02001654static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001655{
1656 struct cfq_queue *cfqq = cfqd->active_queue;
1657
1658 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001659 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001660}
1661
Jens Axboe498d3aa22007-04-26 12:54:48 +02001662/*
1663 * Get next queue for service. Unless we have a queue preemption,
1664 * we'll simply select the first cfqq in the service tree.
1665 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001666static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001667{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001668 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001669 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001670 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001671
Vivek Goyalf04a6422009-12-03 12:59:40 -05001672 if (!cfqd->rq_queued)
1673 return NULL;
1674
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001675 /* There is nothing to dispatch */
1676 if (!service_tree)
1677 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001678 if (RB_EMPTY_ROOT(&service_tree->rb))
1679 return NULL;
1680 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001681}
1682
Vivek Goyalf04a6422009-12-03 12:59:40 -05001683static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1684{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001685 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001686 struct cfq_queue *cfqq;
1687 int i, j;
1688 struct cfq_rb_root *st;
1689
1690 if (!cfqd->rq_queued)
1691 return NULL;
1692
Vivek Goyal25fb5162009-12-03 12:59:46 -05001693 cfqg = cfq_get_next_cfqg(cfqd);
1694 if (!cfqg)
1695 return NULL;
1696
Vivek Goyalf04a6422009-12-03 12:59:40 -05001697 for_each_cfqg_st(cfqg, i, j, st)
1698 if ((cfqq = cfq_rb_first(st)) != NULL)
1699 return cfqq;
1700 return NULL;
1701}
1702
Jens Axboe498d3aa22007-04-26 12:54:48 +02001703/*
1704 * Get and set a new active queue for service.
1705 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001706static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1707 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001708{
Jens Axboee00ef792009-11-04 08:54:55 +01001709 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001710 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001711
Jens Axboe22e2c502005-06-27 10:55:12 +02001712 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001713 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001714}
1715
Jens Axboed9e76202007-04-20 14:27:50 +02001716static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1717 struct request *rq)
1718{
Tejun Heo83096eb2009-05-07 22:24:39 +09001719 if (blk_rq_pos(rq) >= cfqd->last_position)
1720 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001721 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001722 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001723}
1724
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001725static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001726 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001727{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001728 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001729}
1730
Jens Axboea36e71f2009-04-15 12:15:11 +02001731static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1732 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001733{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001734 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001735 struct rb_node *parent, *node;
1736 struct cfq_queue *__cfqq;
1737 sector_t sector = cfqd->last_position;
1738
1739 if (RB_EMPTY_ROOT(root))
1740 return NULL;
1741
1742 /*
1743 * First, if we find a request starting at the end of the last
1744 * request, choose it.
1745 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001746 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001747 if (__cfqq)
1748 return __cfqq;
1749
1750 /*
1751 * If the exact sector wasn't found, the parent of the NULL leaf
1752 * will contain the closest sector.
1753 */
1754 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001755 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001756 return __cfqq;
1757
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001758 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001759 node = rb_next(&__cfqq->p_node);
1760 else
1761 node = rb_prev(&__cfqq->p_node);
1762 if (!node)
1763 return NULL;
1764
1765 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001766 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001767 return __cfqq;
1768
1769 return NULL;
1770}
1771
1772/*
1773 * cfqd - obvious
1774 * cur_cfqq - passed in so that we don't decide that the current queue is
1775 * closely cooperating with itself.
1776 *
1777 * So, basically we're assuming that that cur_cfqq has dispatched at least
1778 * one request, and that cfqd->last_position reflects a position on the disk
1779 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1780 * assumption.
1781 */
1782static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001783 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001784{
1785 struct cfq_queue *cfqq;
1786
Divyesh Shah39c01b22010-03-25 15:45:57 +01001787 if (cfq_class_idle(cur_cfqq))
1788 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001789 if (!cfq_cfqq_sync(cur_cfqq))
1790 return NULL;
1791 if (CFQQ_SEEKY(cur_cfqq))
1792 return NULL;
1793
Jens Axboea36e71f2009-04-15 12:15:11 +02001794 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001795 * Don't search priority tree if it's the only queue in the group.
1796 */
1797 if (cur_cfqq->cfqg->nr_cfqq == 1)
1798 return NULL;
1799
1800 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001801 * We should notice if some of the queues are cooperating, eg
1802 * working closely on the same area of the disk. In that case,
1803 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001804 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001805 cfqq = cfqq_close(cfqd, cur_cfqq);
1806 if (!cfqq)
1807 return NULL;
1808
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001809 /* If new queue belongs to different cfq_group, don't choose it */
1810 if (cur_cfqq->cfqg != cfqq->cfqg)
1811 return NULL;
1812
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001813 /*
1814 * It only makes sense to merge sync queues.
1815 */
1816 if (!cfq_cfqq_sync(cfqq))
1817 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001818 if (CFQQ_SEEKY(cfqq))
1819 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001820
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001821 /*
1822 * Do not merge queues of different priority classes
1823 */
1824 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1825 return NULL;
1826
Jens Axboea36e71f2009-04-15 12:15:11 +02001827 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001828}
1829
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001830/*
1831 * Determine whether we should enforce idle window for this queue.
1832 */
1833
1834static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1835{
1836 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001837 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001838
Vivek Goyalf04a6422009-12-03 12:59:40 -05001839 BUG_ON(!service_tree);
1840 BUG_ON(!service_tree->count);
1841
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001842 /* We never do for idle class queues. */
1843 if (prio == IDLE_WORKLOAD)
1844 return false;
1845
1846 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001847 if (cfq_cfqq_idle_window(cfqq) &&
1848 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001849 return true;
1850
1851 /*
1852 * Otherwise, we do only if they are the last ones
1853 * in their service tree.
1854 */
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001855 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1856 return 1;
1857 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1858 service_tree->count);
1859 return 0;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001860}
1861
Jens Axboe6d048f52007-04-25 12:44:27 +02001862static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001863{
Jens Axboe17926692007-01-19 11:59:30 +11001864 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001865 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001866 unsigned long sl;
1867
Jens Axboea68bbdd2008-09-24 13:03:33 +02001868 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001869 * SSD device without seek penalty, disable idling. But only do so
1870 * for devices that support queuing, otherwise we still have a problem
1871 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001872 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001873 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001874 return;
1875
Jens Axboedd67d052006-06-21 09:36:18 +02001876 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001877 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001878
1879 /*
1880 * idle is disabled, either manually or by past process history
1881 */
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001882 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
Jens Axboe6d048f52007-04-25 12:44:27 +02001883 return;
1884
Jens Axboe22e2c502005-06-27 10:55:12 +02001885 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001886 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001887 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001888 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001889 return;
1890
1891 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001892 * task has exited, don't wait
1893 */
Jens Axboe206dc692006-03-28 13:03:44 +02001894 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001895 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001896 return;
1897
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001898 /*
1899 * If our average think time is larger than the remaining time
1900 * slice, then don't idle. This avoids overrunning the allotted
1901 * time slice.
1902 */
1903 if (sample_valid(cic->ttime_samples) &&
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001904 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1905 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1906 cic->ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001907 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001908 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001909
Jens Axboe3b181522005-06-27 10:56:24 +02001910 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001911
Jens Axboe6d048f52007-04-25 12:44:27 +02001912 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001913
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001914 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001915 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001916 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917}
1918
Jens Axboe498d3aa22007-04-26 12:54:48 +02001919/*
1920 * Move request from internal lists to the request queue dispatch list.
1921 */
Jens Axboe165125e2007-07-24 09:28:11 +02001922static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001924 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001925 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001926
Jens Axboe7b679132008-05-30 12:23:07 +02001927 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1928
Jeff Moyer06d21882009-09-11 17:08:59 +02001929 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001930 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001931 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001932 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001933
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001934 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyale98ef892010-06-18 10:39:47 -04001935 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02001936 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937}
1938
1939/*
1940 * return expired entry, or NULL to just start from scratch in rbtree
1941 */
Jens Axboefebffd62008-01-28 13:19:43 +01001942static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943{
Jens Axboe30996f42009-10-05 11:03:39 +02001944 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945
Jens Axboe3b181522005-06-27 10:56:24 +02001946 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001948
1949 cfq_mark_cfqq_fifo_expire(cfqq);
1950
Jens Axboe89850f72006-07-22 16:48:31 +02001951 if (list_empty(&cfqq->fifo))
1952 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953
Jens Axboe89850f72006-07-22 16:48:31 +02001954 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02001955 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02001956 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957
Jens Axboe30996f42009-10-05 11:03:39 +02001958 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001959 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960}
1961
Jens Axboe22e2c502005-06-27 10:55:12 +02001962static inline int
1963cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1964{
1965 const int base_rq = cfqd->cfq_slice_async_rq;
1966
1967 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1968
1969 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1970}
1971
1972/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001973 * Must be called with the queue_lock held.
1974 */
1975static int cfqq_process_refs(struct cfq_queue *cfqq)
1976{
1977 int process_refs, io_refs;
1978
1979 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1980 process_refs = atomic_read(&cfqq->ref) - io_refs;
1981 BUG_ON(process_refs < 0);
1982 return process_refs;
1983}
1984
1985static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1986{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001987 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001988 struct cfq_queue *__cfqq;
1989
Jeff Moyerc10b61f2010-06-17 10:19:11 -04001990 /*
1991 * If there are no process references on the new_cfqq, then it is
1992 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
1993 * chain may have dropped their last reference (not just their
1994 * last process reference).
1995 */
1996 if (!cfqq_process_refs(new_cfqq))
1997 return;
1998
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001999 /* Avoid a circular list and skip interim queue merges */
2000 while ((__cfqq = new_cfqq->new_cfqq)) {
2001 if (__cfqq == cfqq)
2002 return;
2003 new_cfqq = __cfqq;
2004 }
2005
2006 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002007 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002008 /*
2009 * If the process for the cfqq has gone away, there is no
2010 * sense in merging the queues.
2011 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002012 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002013 return;
2014
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002015 /*
2016 * Merge in the direction of the lesser amount of work.
2017 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002018 if (new_process_refs >= process_refs) {
2019 cfqq->new_cfqq = new_cfqq;
2020 atomic_add(process_refs, &new_cfqq->ref);
2021 } else {
2022 new_cfqq->new_cfqq = cfqq;
2023 atomic_add(new_process_refs, &cfqq->ref);
2024 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002025}
2026
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002027static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002028 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002029{
2030 struct cfq_queue *queue;
2031 int i;
2032 bool key_valid = false;
2033 unsigned long lowest_key = 0;
2034 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2035
Vivek Goyal65b32a52009-12-16 17:52:59 -05002036 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2037 /* select the one with lowest rb_key */
2038 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002039 if (queue &&
2040 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2041 lowest_key = queue->rb_key;
2042 cur_best = i;
2043 key_valid = true;
2044 }
2045 }
2046
2047 return cur_best;
2048}
2049
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002050static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002051{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002052 unsigned slice;
2053 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002054 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002055 unsigned group_slice;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002056
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002057 if (!cfqg) {
2058 cfqd->serving_prio = IDLE_WORKLOAD;
2059 cfqd->workload_expires = jiffies + 1;
2060 return;
2061 }
2062
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002063 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002064 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002065 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002066 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002067 cfqd->serving_prio = BE_WORKLOAD;
2068 else {
2069 cfqd->serving_prio = IDLE_WORKLOAD;
2070 cfqd->workload_expires = jiffies + 1;
2071 return;
2072 }
2073
2074 /*
2075 * For RT and BE, we have to choose also the type
2076 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2077 * expiration time
2078 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002079 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002080 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002081
2082 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002083 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002084 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002085 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002086 return;
2087
2088 /* otherwise select new workload type */
2089 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002090 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2091 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002092 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002093
2094 /*
2095 * the workload slice is computed as a fraction of target latency
2096 * proportional to the number of queues in that workload, over
2097 * all the queues in the same priority class
2098 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002099 group_slice = cfq_group_slice(cfqd, cfqg);
2100
2101 slice = group_slice * count /
2102 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2103 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002104
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002105 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2106 unsigned int tmp;
2107
2108 /*
2109 * Async queues are currently system wide. Just taking
2110 * proportion of queues with-in same group will lead to higher
2111 * async ratio system wide as generally root group is going
2112 * to have higher weight. A more accurate thing would be to
2113 * calculate system wide asnc/sync ratio.
2114 */
2115 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2116 tmp = tmp/cfqd->busy_queues;
2117 slice = min_t(unsigned, slice, tmp);
2118
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002119 /* async workload slice is scaled down according to
2120 * the sync/async slice ratio. */
2121 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002122 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002123 /* sync workload slice is at least 2 * cfq_slice_idle */
2124 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2125
2126 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002127 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002128 cfqd->workload_expires = jiffies + slice;
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002129 cfqd->noidle_tree_requires_idle = false;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002130}
2131
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002132static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2133{
2134 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002135 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002136
2137 if (RB_EMPTY_ROOT(&st->rb))
2138 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002139 cfqg = cfq_rb_first_group(st);
2140 st->active = &cfqg->rb_node;
2141 update_min_vdisktime(st);
2142 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002143}
2144
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002145static void cfq_choose_cfqg(struct cfq_data *cfqd)
2146{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002147 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2148
2149 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002150
2151 /* Restore the workload type data */
2152 if (cfqg->saved_workload_slice) {
2153 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2154 cfqd->serving_type = cfqg->saved_workload;
2155 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002156 } else
2157 cfqd->workload_expires = jiffies - 1;
2158
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002159 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002160}
2161
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002162/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002163 * Select a queue for service. If we have a current active queue,
2164 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002165 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002166static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002167{
Jens Axboea36e71f2009-04-15 12:15:11 +02002168 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002169
2170 cfqq = cfqd->active_queue;
2171 if (!cfqq)
2172 goto new_queue;
2173
Vivek Goyalf04a6422009-12-03 12:59:40 -05002174 if (!cfqd->rq_queued)
2175 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002176
2177 /*
2178 * We were waiting for group to get backlogged. Expire the queue
2179 */
2180 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2181 goto expire;
2182
Jens Axboe22e2c502005-06-27 10:55:12 +02002183 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002184 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002185 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002186 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2187 /*
2188 * If slice had not expired at the completion of last request
2189 * we might not have turned on wait_busy flag. Don't expire
2190 * the queue yet. Allow the group to get backlogged.
2191 *
2192 * The very fact that we have used the slice, that means we
2193 * have been idling all along on this queue and it should be
2194 * ok to wait for this request to complete.
2195 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002196 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2197 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2198 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002199 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002200 } else
Vivek Goyal7667aa02009-12-08 17:52:58 -05002201 goto expire;
2202 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002203
2204 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002205 * The active queue has requests and isn't expired, allow it to
2206 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002207 */
Jens Axboedd67d052006-06-21 09:36:18 +02002208 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002209 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002210
2211 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002212 * If another queue has a request waiting within our mean seek
2213 * distance, let it run. The expire code will check for close
2214 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002215 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002216 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002217 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002218 if (new_cfqq) {
2219 if (!cfqq->new_cfqq)
2220 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002221 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002222 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002223
2224 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002225 * No requests pending. If the active queue still has requests in
2226 * flight or is idling for a new request, allow either of these
2227 * conditions to happen (or time out) before selecting a new queue.
2228 */
Jens Axboecc197472007-04-20 20:45:39 +02002229 if (timer_pending(&cfqd->idle_slice_timer) ||
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002230 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002231 cfqq = NULL;
2232 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002233 }
2234
Jens Axboe3b181522005-06-27 10:56:24 +02002235expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002236 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002237new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002238 /*
2239 * Current queue expired. Check if we have to switch to a new
2240 * service tree
2241 */
2242 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002243 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002244
Jens Axboea36e71f2009-04-15 12:15:11 +02002245 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002246keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002247 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002248}
2249
Jens Axboefebffd62008-01-28 13:19:43 +01002250static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002251{
2252 int dispatched = 0;
2253
2254 while (cfqq->next_rq) {
2255 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2256 dispatched++;
2257 }
2258
2259 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002260
2261 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002262 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002263 return dispatched;
2264}
2265
Jens Axboe498d3aa22007-04-26 12:54:48 +02002266/*
2267 * Drain our current requests. Used for barriers and when switching
2268 * io schedulers on-the-fly.
2269 */
Jens Axboed9e76202007-04-20 14:27:50 +02002270static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002271{
Jens Axboe08717142008-01-28 11:38:15 +01002272 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002273 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002274
Divyesh Shah3440c492010-04-09 09:29:57 +02002275 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002276 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002277 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2278 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002279 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002280 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002281
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002282 BUG_ON(cfqd->busy_queues);
2283
Jeff Moyer69237152009-06-12 15:29:30 +02002284 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002285 return dispatched;
2286}
2287
Shaohua Liabc3c742010-03-01 09:20:54 +01002288static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2289 struct cfq_queue *cfqq)
2290{
2291 /* the queue hasn't finished any request, can't estimate */
2292 if (cfq_cfqq_slice_new(cfqq))
2293 return 1;
2294 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2295 cfqq->slice_end))
2296 return 1;
2297
2298 return 0;
2299}
2300
Jens Axboe0b182d62009-10-06 20:49:37 +02002301static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002302{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002303 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304
Jens Axboe2f5cb732009-04-07 08:51:19 +02002305 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002306 * Drain async requests before we start sync IO
2307 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002308 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002309 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002310
2311 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002312 * If this is an async queue and we have sync IO in flight, let it wait
2313 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002314 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002315 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002316
Shaohua Liabc3c742010-03-01 09:20:54 +01002317 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002318 if (cfq_class_idle(cfqq))
2319 max_dispatch = 1;
2320
2321 /*
2322 * Does this cfqq already have too much IO in flight?
2323 */
2324 if (cfqq->dispatched >= max_dispatch) {
2325 /*
2326 * idle queue must always only have a single IO in flight
2327 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002328 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002329 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002330
Jens Axboe2f5cb732009-04-07 08:51:19 +02002331 /*
2332 * We have other queues, don't allow more IO from this one
2333 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002334 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002335 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002336
Jens Axboe2f5cb732009-04-07 08:51:19 +02002337 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002338 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002339 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002340 if (cfqd->busy_queues == 1)
2341 max_dispatch = -1;
2342 else
2343 /*
2344 * Normally we start throttling cfqq when cfq_quantum/2
2345 * requests have been dispatched. But we can drive
2346 * deeper queue depths at the beginning of slice
2347 * subjected to upper limit of cfq_quantum.
2348 * */
2349 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002350 }
2351
2352 /*
2353 * Async queues must wait a bit before being allowed dispatch.
2354 * We also ramp up the dispatch depth gradually for async IO,
2355 * based on the last sync IO we serviced
2356 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002357 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002358 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002359 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002360
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002361 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002362 if (!depth && !cfqq->dispatched)
2363 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002364 if (depth < max_dispatch)
2365 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 }
2367
Jens Axboe0b182d62009-10-06 20:49:37 +02002368 /*
2369 * If we're below the current max, allow a dispatch
2370 */
2371 return cfqq->dispatched < max_dispatch;
2372}
2373
2374/*
2375 * Dispatch a request from cfqq, moving them to the request queue
2376 * dispatch list.
2377 */
2378static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2379{
2380 struct request *rq;
2381
2382 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2383
2384 if (!cfq_may_dispatch(cfqd, cfqq))
2385 return false;
2386
2387 /*
2388 * follow expired path, else get first next available
2389 */
2390 rq = cfq_check_fifo(cfqq);
2391 if (!rq)
2392 rq = cfqq->next_rq;
2393
2394 /*
2395 * insert request into driver dispatch list
2396 */
2397 cfq_dispatch_insert(cfqd->queue, rq);
2398
2399 if (!cfqd->active_cic) {
2400 struct cfq_io_context *cic = RQ_CIC(rq);
2401
2402 atomic_long_inc(&cic->ioc->refcount);
2403 cfqd->active_cic = cic;
2404 }
2405
2406 return true;
2407}
2408
2409/*
2410 * Find the cfqq that we need to service and move a request from that to the
2411 * dispatch list
2412 */
2413static int cfq_dispatch_requests(struct request_queue *q, int force)
2414{
2415 struct cfq_data *cfqd = q->elevator->elevator_data;
2416 struct cfq_queue *cfqq;
2417
2418 if (!cfqd->busy_queues)
2419 return 0;
2420
2421 if (unlikely(force))
2422 return cfq_forced_dispatch(cfqd);
2423
2424 cfqq = cfq_select_queue(cfqd);
2425 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002426 return 0;
2427
Jens Axboe2f5cb732009-04-07 08:51:19 +02002428 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002429 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002430 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002431 if (!cfq_dispatch_request(cfqd, cfqq))
2432 return 0;
2433
Jens Axboe2f5cb732009-04-07 08:51:19 +02002434 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002435 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002436
2437 /*
2438 * expire an async queue immediately if it has used up its slice. idle
2439 * queue always expire after 1 dispatch round.
2440 */
2441 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2442 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2443 cfq_class_idle(cfqq))) {
2444 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002445 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002446 }
2447
Shan Weib217a902009-09-01 10:06:42 +02002448 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002449 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450}
2451
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452/*
Jens Axboe5e705372006-07-13 12:39:25 +02002453 * task holds one reference to the queue, dropped when task exits. each rq
2454 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002456 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 * queue lock must be held here.
2458 */
2459static void cfq_put_queue(struct cfq_queue *cfqq)
2460{
Jens Axboe22e2c502005-06-27 10:55:12 +02002461 struct cfq_data *cfqd = cfqq->cfqd;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002462 struct cfq_group *cfqg, *orig_cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002463
2464 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465
2466 if (!atomic_dec_and_test(&cfqq->ref))
2467 return;
2468
Jens Axboe7b679132008-05-30 12:23:07 +02002469 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002471 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002472 cfqg = cfqq->cfqg;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002473 orig_cfqg = cfqq->orig_cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002475 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002476 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002477 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002478 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002479
Vivek Goyalf04a6422009-12-03 12:59:40 -05002480 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002482 cfq_put_cfqg(cfqg);
Vivek Goyal878eadd2009-12-07 19:37:15 +01002483 if (orig_cfqg)
2484 cfq_put_cfqg(orig_cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485}
2486
Jens Axboed6de8be2008-05-28 14:46:59 +02002487/*
2488 * Must always be called with the rcu_read_lock() held
2489 */
Jens Axboe07416d22008-05-07 09:17:12 +02002490static void
2491__call_for_each_cic(struct io_context *ioc,
2492 void (*func)(struct io_context *, struct cfq_io_context *))
2493{
2494 struct cfq_io_context *cic;
2495 struct hlist_node *n;
2496
2497 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2498 func(ioc, cic);
2499}
2500
Jens Axboe4ac845a2008-01-24 08:44:49 +01002501/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002502 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002503 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002504static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002505call_for_each_cic(struct io_context *ioc,
2506 void (*func)(struct io_context *, struct cfq_io_context *))
2507{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002508 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002509 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002510 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002511}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002512
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002513static void cfq_cic_free_rcu(struct rcu_head *head)
2514{
2515 struct cfq_io_context *cic;
2516
2517 cic = container_of(head, struct cfq_io_context, rcu_head);
2518
2519 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002520 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002521
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002522 if (ioc_gone) {
2523 /*
2524 * CFQ scheduler is exiting, grab exit lock and check
2525 * the pending io context count. If it hits zero,
2526 * complete ioc_gone and set it back to NULL
2527 */
2528 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002529 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002530 complete(ioc_gone);
2531 ioc_gone = NULL;
2532 }
2533 spin_unlock(&ioc_gone_lock);
2534 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002535}
2536
2537static void cfq_cic_free(struct cfq_io_context *cic)
2538{
2539 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002540}
2541
2542static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2543{
2544 unsigned long flags;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002545 unsigned long dead_key = (unsigned long) cic->key;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002546
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002547 BUG_ON(!(dead_key & CIC_DEAD_KEY));
Jens Axboe4ac845a2008-01-24 08:44:49 +01002548
2549 spin_lock_irqsave(&ioc->lock, flags);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002550 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
Jens Axboeffc4e752008-02-19 10:02:29 +01002551 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002552 spin_unlock_irqrestore(&ioc->lock, flags);
2553
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002554 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002555}
2556
Jens Axboed6de8be2008-05-28 14:46:59 +02002557/*
2558 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2559 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2560 * and ->trim() which is called with the task lock held
2561 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002562static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002564 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002565 * ioc->refcount is zero here, or we are called from elv_unregister(),
2566 * so no more cic's are allowed to be linked into this ioc. So it
2567 * should be ok to iterate over the known list, we will see all cic's
2568 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002569 */
Jens Axboe07416d22008-05-07 09:17:12 +02002570 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571}
2572
Shaohua Lid02a2c02010-05-25 10:16:53 +02002573static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002574{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002575 struct cfq_queue *__cfqq, *next;
2576
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002577 /*
2578 * If this queue was scheduled to merge with another queue, be
2579 * sure to drop the reference taken on that queue (and others in
2580 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2581 */
2582 __cfqq = cfqq->new_cfqq;
2583 while (__cfqq) {
2584 if (__cfqq == cfqq) {
2585 WARN(1, "cfqq->new_cfqq loop detected\n");
2586 break;
2587 }
2588 next = __cfqq->new_cfqq;
2589 cfq_put_queue(__cfqq);
2590 __cfqq = next;
2591 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002592}
2593
2594static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2595{
2596 if (unlikely(cfqq == cfqd->active_queue)) {
2597 __cfq_slice_expired(cfqd, cfqq, 0);
2598 cfq_schedule_dispatch(cfqd);
2599 }
2600
2601 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002602
Jens Axboe89850f72006-07-22 16:48:31 +02002603 cfq_put_queue(cfqq);
2604}
2605
2606static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2607 struct cfq_io_context *cic)
2608{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002609 struct io_context *ioc = cic->ioc;
2610
Jens Axboefc463792006-08-29 09:05:44 +02002611 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002612
2613 /*
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002614 * Make sure dead mark is seen for dead queues
Jens Axboe4ac845a2008-01-24 08:44:49 +01002615 */
Jens Axboefc463792006-08-29 09:05:44 +02002616 smp_wmb();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002617 cic->key = cfqd_dead_key(cfqd);
Jens Axboefc463792006-08-29 09:05:44 +02002618
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002619 if (ioc->ioc_data == cic)
2620 rcu_assign_pointer(ioc->ioc_data, NULL);
2621
Jens Axboeff6657c2009-04-08 10:58:57 +02002622 if (cic->cfqq[BLK_RW_ASYNC]) {
2623 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2624 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002625 }
2626
Jens Axboeff6657c2009-04-08 10:58:57 +02002627 if (cic->cfqq[BLK_RW_SYNC]) {
2628 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2629 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002630 }
Jens Axboe89850f72006-07-22 16:48:31 +02002631}
2632
Jens Axboe4ac845a2008-01-24 08:44:49 +01002633static void cfq_exit_single_io_context(struct io_context *ioc,
2634 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002635{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002636 struct cfq_data *cfqd = cic_to_cfqd(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002637
Jens Axboe89850f72006-07-22 16:48:31 +02002638 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002639 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002640 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002641
Jens Axboe4ac845a2008-01-24 08:44:49 +01002642 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002643
2644 /*
2645 * Ensure we get a fresh copy of the ->key to prevent
2646 * race between exiting task and queue
2647 */
2648 smp_read_barrier_depends();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002649 if (cic->key == cfqd)
Jens Axboe62c1fe92008-12-15 21:19:25 +01002650 __cfq_exit_single_io_context(cfqd, cic);
2651
Jens Axboe4ac845a2008-01-24 08:44:49 +01002652 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002653 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002654}
2655
Jens Axboe498d3aa22007-04-26 12:54:48 +02002656/*
2657 * The process that ioc belongs to has exited, we need to clean up
2658 * and put the internal structures we have that belongs to that process.
2659 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002660static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002662 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663}
2664
Jens Axboe22e2c502005-06-27 10:55:12 +02002665static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002666cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667{
Jens Axboeb5deef92006-07-19 23:39:40 +02002668 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669
Christoph Lameter94f60302007-07-17 04:03:29 -07002670 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2671 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002673 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002674 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002675 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002676 cic->dtor = cfq_free_io_context;
2677 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002678 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 }
2680
2681 return cic;
2682}
2683
Jens Axboefd0928d2008-01-24 08:52:45 +01002684static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002685{
2686 struct task_struct *tsk = current;
2687 int ioprio_class;
2688
Jens Axboe3b181522005-06-27 10:56:24 +02002689 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002690 return;
2691
Jens Axboefd0928d2008-01-24 08:52:45 +01002692 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002693 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002694 default:
2695 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2696 case IOPRIO_CLASS_NONE:
2697 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002698 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002699 */
2700 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002701 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002702 break;
2703 case IOPRIO_CLASS_RT:
2704 cfqq->ioprio = task_ioprio(ioc);
2705 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2706 break;
2707 case IOPRIO_CLASS_BE:
2708 cfqq->ioprio = task_ioprio(ioc);
2709 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2710 break;
2711 case IOPRIO_CLASS_IDLE:
2712 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2713 cfqq->ioprio = 7;
2714 cfq_clear_cfqq_idle_window(cfqq);
2715 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002716 }
2717
2718 /*
2719 * keep track of original prio settings in case we have to temporarily
2720 * elevate the priority of this queue
2721 */
2722 cfqq->org_ioprio = cfqq->ioprio;
2723 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002724 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002725}
2726
Jens Axboefebffd62008-01-28 13:19:43 +01002727static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002728{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002729 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002730 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002731 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002732
Jens Axboecaaa5f92006-06-16 11:23:00 +02002733 if (unlikely(!cfqd))
2734 return;
2735
Jens Axboec1b707d2006-10-30 19:54:23 +01002736 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002737
Jens Axboeff6657c2009-04-08 10:58:57 +02002738 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002739 if (cfqq) {
2740 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002741 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2742 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002743 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002744 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002745 cfq_put_queue(cfqq);
2746 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002747 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002748
Jens Axboeff6657c2009-04-08 10:58:57 +02002749 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002750 if (cfqq)
2751 cfq_mark_cfqq_prio_changed(cfqq);
2752
Jens Axboec1b707d2006-10-30 19:54:23 +01002753 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002754}
2755
Jens Axboefc463792006-08-29 09:05:44 +02002756static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002758 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002759 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760}
2761
Jens Axboed5036d72009-06-26 10:44:34 +02002762static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002763 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002764{
2765 RB_CLEAR_NODE(&cfqq->rb_node);
2766 RB_CLEAR_NODE(&cfqq->p_node);
2767 INIT_LIST_HEAD(&cfqq->fifo);
2768
2769 atomic_set(&cfqq->ref, 0);
2770 cfqq->cfqd = cfqd;
2771
2772 cfq_mark_cfqq_prio_changed(cfqq);
2773
2774 if (is_sync) {
2775 if (!cfq_class_idle(cfqq))
2776 cfq_mark_cfqq_idle_window(cfqq);
2777 cfq_mark_cfqq_sync(cfqq);
2778 }
2779 cfqq->pid = pid;
2780}
2781
Vivek Goyal246103332009-12-03 12:59:51 -05002782#ifdef CONFIG_CFQ_GROUP_IOSCHED
2783static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2784{
2785 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002786 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal246103332009-12-03 12:59:51 -05002787 unsigned long flags;
2788 struct request_queue *q;
2789
2790 if (unlikely(!cfqd))
2791 return;
2792
2793 q = cfqd->queue;
2794
2795 spin_lock_irqsave(q->queue_lock, flags);
2796
2797 if (sync_cfqq) {
2798 /*
2799 * Drop reference to sync queue. A new sync queue will be
2800 * assigned in new group upon arrival of a fresh request.
2801 */
2802 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2803 cic_set_cfqq(cic, NULL, 1);
2804 cfq_put_queue(sync_cfqq);
2805 }
2806
2807 spin_unlock_irqrestore(q->queue_lock, flags);
2808}
2809
2810static void cfq_ioc_set_cgroup(struct io_context *ioc)
2811{
2812 call_for_each_cic(ioc, changed_cgroup);
2813 ioc->cgroup_changed = 0;
2814}
2815#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2816
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002818cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002819 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002822 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002823 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824
2825retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002826 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002827 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002828 /* cic always exists here */
2829 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830
Jens Axboe6118b702009-06-30 09:34:12 +02002831 /*
2832 * Always try a new alloc if we fell back to the OOM cfqq
2833 * originally, since it should just be a temporary situation.
2834 */
2835 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2836 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837 if (new_cfqq) {
2838 cfqq = new_cfqq;
2839 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002840 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002842 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002843 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002844 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002846 if (new_cfqq)
2847 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002848 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002849 cfqq = kmem_cache_alloc_node(cfq_pool,
2850 gfp_mask | __GFP_ZERO,
2851 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002852 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853
Jens Axboe6118b702009-06-30 09:34:12 +02002854 if (cfqq) {
2855 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2856 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002857 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002858 cfq_log_cfqq(cfqd, cfqq, "alloced");
2859 } else
2860 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 }
2862
2863 if (new_cfqq)
2864 kmem_cache_free(cfq_pool, new_cfqq);
2865
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 return cfqq;
2867}
2868
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002869static struct cfq_queue **
2870cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2871{
Jens Axboefe094d92008-01-31 13:08:54 +01002872 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002873 case IOPRIO_CLASS_RT:
2874 return &cfqd->async_cfqq[0][ioprio];
2875 case IOPRIO_CLASS_BE:
2876 return &cfqd->async_cfqq[1][ioprio];
2877 case IOPRIO_CLASS_IDLE:
2878 return &cfqd->async_idle_cfqq;
2879 default:
2880 BUG();
2881 }
2882}
2883
Jens Axboe15c31be2007-07-10 13:43:25 +02002884static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002885cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002886 gfp_t gfp_mask)
2887{
Jens Axboefd0928d2008-01-24 08:52:45 +01002888 const int ioprio = task_ioprio(ioc);
2889 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002890 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002891 struct cfq_queue *cfqq = NULL;
2892
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002893 if (!is_sync) {
2894 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2895 cfqq = *async_cfqq;
2896 }
2897
Jens Axboe6118b702009-06-30 09:34:12 +02002898 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002899 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002900
2901 /*
2902 * pin the queue now that it's allocated, scheduler exit will prune it
2903 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002904 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02002905 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002906 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002907 }
2908
2909 atomic_inc(&cfqq->ref);
2910 return cfqq;
2911}
2912
Jens Axboe498d3aa22007-04-26 12:54:48 +02002913/*
2914 * We drop cfq io contexts lazily, so we may find a dead one.
2915 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002916static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002917cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2918 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002919{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002920 unsigned long flags;
2921
Jens Axboefc463792006-08-29 09:05:44 +02002922 WARN_ON(!list_empty(&cic->queue_list));
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002923 BUG_ON(cic->key != cfqd_dead_key(cfqd));
Jens Axboe597bc482007-04-24 21:23:53 +02002924
Jens Axboe4ac845a2008-01-24 08:44:49 +01002925 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02002926
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002927 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002928
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002929 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
Jens Axboeffc4e752008-02-19 10:02:29 +01002930 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002931 spin_unlock_irqrestore(&ioc->lock, flags);
2932
2933 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002934}
2935
Jens Axboee2d74ac2006-03-28 08:59:01 +02002936static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01002937cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002938{
Jens Axboee2d74ac2006-03-28 08:59:01 +02002939 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002940 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002941
Vasily Tarasov91fac312007-04-25 12:29:51 +02002942 if (unlikely(!ioc))
2943 return NULL;
2944
Jens Axboed6de8be2008-05-28 14:46:59 +02002945 rcu_read_lock();
2946
Jens Axboe597bc482007-04-24 21:23:53 +02002947 /*
2948 * we maintain a last-hit cache, to avoid browsing over the tree
2949 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002950 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02002951 if (cic && cic->key == cfqd) {
2952 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02002953 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002954 }
Jens Axboe597bc482007-04-24 21:23:53 +02002955
Jens Axboe4ac845a2008-01-24 08:44:49 +01002956 do {
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002957 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002958 rcu_read_unlock();
2959 if (!cic)
2960 break;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002961 if (unlikely(cic->key != cfqd)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01002962 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002963 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01002964 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002965 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002966
Jens Axboed6de8be2008-05-28 14:46:59 +02002967 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002968 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002969 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002970 break;
2971 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002972
Jens Axboe4ac845a2008-01-24 08:44:49 +01002973 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002974}
2975
Jens Axboe4ac845a2008-01-24 08:44:49 +01002976/*
2977 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2978 * the process specific cfq io context when entered from the block layer.
2979 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2980 */
Jens Axboefebffd62008-01-28 13:19:43 +01002981static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2982 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002983{
Jens Axboe0261d682006-10-30 19:07:48 +01002984 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002985 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002986
Jens Axboe4ac845a2008-01-24 08:44:49 +01002987 ret = radix_tree_preload(gfp_mask);
2988 if (!ret) {
2989 cic->ioc = ioc;
2990 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002991
Jens Axboe4ac845a2008-01-24 08:44:49 +01002992 spin_lock_irqsave(&ioc->lock, flags);
2993 ret = radix_tree_insert(&ioc->radix_root,
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002994 cfqd->cic_index, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01002995 if (!ret)
2996 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002997 spin_unlock_irqrestore(&ioc->lock, flags);
2998
2999 radix_tree_preload_end();
3000
3001 if (!ret) {
3002 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3003 list_add(&cic->queue_list, &cfqd->cic_list);
3004 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003005 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003006 }
3007
Jens Axboe4ac845a2008-01-24 08:44:49 +01003008 if (ret)
3009 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02003010
Jens Axboe4ac845a2008-01-24 08:44:49 +01003011 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003012}
3013
Jens Axboe22e2c502005-06-27 10:55:12 +02003014/*
3015 * Setup general io context and cfq io context. There can be several cfq
3016 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02003017 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02003018 */
3019static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02003020cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021{
Jens Axboe22e2c502005-06-27 10:55:12 +02003022 struct io_context *ioc = NULL;
3023 struct cfq_io_context *cic;
3024
3025 might_sleep_if(gfp_mask & __GFP_WAIT);
3026
Jens Axboeb5deef92006-07-19 23:39:40 +02003027 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02003028 if (!ioc)
3029 return NULL;
3030
Jens Axboe4ac845a2008-01-24 08:44:49 +01003031 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003032 if (cic)
3033 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02003034
Jens Axboee2d74ac2006-03-28 08:59:01 +02003035 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3036 if (cic == NULL)
3037 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02003038
Jens Axboe4ac845a2008-01-24 08:44:49 +01003039 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3040 goto err_free;
3041
Jens Axboe22e2c502005-06-27 10:55:12 +02003042out:
Jens Axboefc463792006-08-29 09:05:44 +02003043 smp_read_barrier_depends();
3044 if (unlikely(ioc->ioprio_changed))
3045 cfq_ioc_set_ioprio(ioc);
3046
Vivek Goyal246103332009-12-03 12:59:51 -05003047#ifdef CONFIG_CFQ_GROUP_IOSCHED
3048 if (unlikely(ioc->cgroup_changed))
3049 cfq_ioc_set_cgroup(ioc);
3050#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003051 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003052err_free:
3053 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003054err:
3055 put_io_context(ioc);
3056 return NULL;
3057}
3058
3059static void
3060cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3061{
Jens Axboeaaf12282007-01-19 11:30:16 +11003062 unsigned long elapsed = jiffies - cic->last_end_request;
3063 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003064
3065 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3066 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3067 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3068}
3069
Jens Axboe206dc692006-03-28 13:03:44 +02003070static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003071cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003072 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003073{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003074 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003075 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003076 if (cfqq->last_request_pos) {
3077 if (cfqq->last_request_pos < blk_rq_pos(rq))
3078 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3079 else
3080 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3081 }
Jens Axboe206dc692006-03-28 13:03:44 +02003082
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003083 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003084 if (blk_queue_nonrot(cfqd->queue))
3085 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3086 else
3087 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003088}
Jens Axboe22e2c502005-06-27 10:55:12 +02003089
3090/*
3091 * Disable idle window if the process thinks too long or seeks so much that
3092 * it doesn't matter
3093 */
3094static void
3095cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3096 struct cfq_io_context *cic)
3097{
Jens Axboe7b679132008-05-30 12:23:07 +02003098 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003099
Jens Axboe08717142008-01-28 11:38:15 +01003100 /*
3101 * Don't idle for async or idle io prio class
3102 */
3103 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003104 return;
3105
Jens Axboec265a7f2008-06-26 13:49:33 +02003106 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003107
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003108 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3109 cfq_mark_cfqq_deep(cfqq);
3110
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01003111 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003112 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003113 enable_idle = 0;
3114 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003115 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003116 enable_idle = 0;
3117 else
3118 enable_idle = 1;
3119 }
3120
Jens Axboe7b679132008-05-30 12:23:07 +02003121 if (old_idle != enable_idle) {
3122 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3123 if (enable_idle)
3124 cfq_mark_cfqq_idle_window(cfqq);
3125 else
3126 cfq_clear_cfqq_idle_window(cfqq);
3127 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003128}
3129
Jens Axboe22e2c502005-06-27 10:55:12 +02003130/*
3131 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3132 * no or if we aren't sure, a 1 will cause a preempt.
3133 */
Jens Axboea6151c32009-10-07 20:02:57 +02003134static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003135cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003136 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003137{
Jens Axboe6d048f52007-04-25 12:44:27 +02003138 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003139
Jens Axboe6d048f52007-04-25 12:44:27 +02003140 cfqq = cfqd->active_queue;
3141 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003142 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003143
Jens Axboe6d048f52007-04-25 12:44:27 +02003144 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003145 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003146
3147 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003148 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003149
Jens Axboe22e2c502005-06-27 10:55:12 +02003150 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003151 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3152 */
3153 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3154 return false;
3155
3156 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003157 * if the new request is sync, but the currently running queue is
3158 * not, let the sync request have priority.
3159 */
Jens Axboe5e705372006-07-13 12:39:25 +02003160 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003161 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003162
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003163 if (new_cfqq->cfqg != cfqq->cfqg)
3164 return false;
3165
3166 if (cfq_slice_used(cfqq))
3167 return true;
3168
3169 /* Allow preemption only if we are idling on sync-noidle tree */
3170 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3171 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3172 new_cfqq->service_tree->count == 2 &&
3173 RB_EMPTY_ROOT(&cfqq->sort_list))
3174 return true;
3175
Jens Axboe374f84a2006-07-23 01:42:19 +02003176 /*
3177 * So both queues are sync. Let the new request get disk time if
3178 * it's a metadata request and the current queue is doing regular IO.
3179 */
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003180 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003181 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003182
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003183 /*
3184 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3185 */
3186 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003187 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003188
Jens Axboe1e3335d2007-02-14 19:59:49 +01003189 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003190 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003191
3192 /*
3193 * if this request is as-good as one we would expect from the
3194 * current cfqq, let it preempt
3195 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003196 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003197 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003198
Jens Axboea6151c32009-10-07 20:02:57 +02003199 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003200}
3201
3202/*
3203 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3204 * let it have half of its nominal slice.
3205 */
3206static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3207{
Jens Axboe7b679132008-05-30 12:23:07 +02003208 cfq_log_cfqq(cfqd, cfqq, "preempt");
Vivek Goyale5ff0822010-04-26 19:25:11 +02003209 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003210
Jens Axboebf572252006-07-19 20:29:12 +02003211 /*
3212 * Put the new queue at the front of the of the current list,
3213 * so we know that it will be selected next.
3214 */
3215 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003216
3217 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02003218
Jens Axboe44f7c162007-01-19 11:51:58 +11003219 cfqq->slice_end = 0;
3220 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003221}
3222
3223/*
Jens Axboe5e705372006-07-13 12:39:25 +02003224 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003225 * something we should do about it
3226 */
3227static void
Jens Axboe5e705372006-07-13 12:39:25 +02003228cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3229 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003230{
Jens Axboe5e705372006-07-13 12:39:25 +02003231 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003232
Aaron Carroll45333d52008-08-26 15:52:36 +02003233 cfqd->rq_queued++;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003234 if (rq->cmd_flags & REQ_META)
Jens Axboe374f84a2006-07-23 01:42:19 +02003235 cfqq->meta_pending++;
3236
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003237 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003238 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003239 cfq_update_idle_window(cfqd, cfqq, cic);
3240
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003241 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003242
3243 if (cfqq == cfqd->active_queue) {
3244 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003245 * Remember that we saw a request from this process, but
3246 * don't start queuing just yet. Otherwise we risk seeing lots
3247 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003248 * and merging. If the request is already larger than a single
3249 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003250 * merging is already done. Ditto for a busy system that
3251 * has other work pending, don't risk delaying until the
3252 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003253 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003254 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003255 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3256 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003257 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003258 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003259 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003260 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003261 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003262 &cfqq->cfqg->blkg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003263 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003264 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003265 }
Jens Axboe5e705372006-07-13 12:39:25 +02003266 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003267 /*
3268 * not the active queue - expire current slice if it is
3269 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003270 * has some old slice time left and is of higher priority or
3271 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003272 */
3273 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09003274 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003275 }
3276}
3277
Jens Axboe165125e2007-07-24 09:28:11 +02003278static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003279{
Jens Axboeb4878f22005-10-20 16:42:29 +02003280 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003281 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003282
Jens Axboe7b679132008-05-30 12:23:07 +02003283 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003284 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285
Jens Axboe30996f42009-10-05 11:03:39 +02003286 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003287 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003288 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003289 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003290 &cfqd->serving_group->blkg, rq_data_dir(rq),
3291 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003292 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293}
3294
Aaron Carroll45333d52008-08-26 15:52:36 +02003295/*
3296 * Update hw_tag based on peak queue depth over 50 samples under
3297 * sufficient load.
3298 */
3299static void cfq_update_hw_tag(struct cfq_data *cfqd)
3300{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003301 struct cfq_queue *cfqq = cfqd->active_queue;
3302
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003303 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3304 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003305
3306 if (cfqd->hw_tag == 1)
3307 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003308
3309 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003310 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003311 return;
3312
Shaohua Li1a1238a2009-10-27 08:46:23 +01003313 /*
3314 * If active queue hasn't enough requests and can idle, cfq might not
3315 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3316 * case
3317 */
3318 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3319 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003320 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003321 return;
3322
Aaron Carroll45333d52008-08-26 15:52:36 +02003323 if (cfqd->hw_tag_samples++ < 50)
3324 return;
3325
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003326 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003327 cfqd->hw_tag = 1;
3328 else
3329 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003330}
3331
Vivek Goyal7667aa02009-12-08 17:52:58 -05003332static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3333{
3334 struct cfq_io_context *cic = cfqd->active_cic;
3335
3336 /* If there are other queues in the group, don't wait */
3337 if (cfqq->cfqg->nr_cfqq > 1)
3338 return false;
3339
3340 if (cfq_slice_used(cfqq))
3341 return true;
3342
3343 /* if slice left is less than think time, wait busy */
3344 if (cic && sample_valid(cic->ttime_samples)
3345 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3346 return true;
3347
3348 /*
3349 * If think times is less than a jiffy than ttime_mean=0 and above
3350 * will not be true. It might happen that slice has not expired yet
3351 * but will expire soon (4-5 ns) during select_queue(). To cover the
3352 * case where think time is less than a jiffy, mark the queue wait
3353 * busy if only 1 jiffy is left in the slice.
3354 */
3355 if (cfqq->slice_end - jiffies == 1)
3356 return true;
3357
3358 return false;
3359}
3360
Jens Axboe165125e2007-07-24 09:28:11 +02003361static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362{
Jens Axboe5e705372006-07-13 12:39:25 +02003363 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003364 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003365 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003366 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367
Jens Axboeb4878f22005-10-20 16:42:29 +02003368 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003369 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3370 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371
Aaron Carroll45333d52008-08-26 15:52:36 +02003372 cfq_update_hw_tag(cfqd);
3373
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003374 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003375 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003376 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003377 cfqq->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003378 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3379 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3380 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003382 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003383
Vivek Goyal365722b2009-10-03 15:21:27 +02003384 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003385 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003386 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3387 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003388 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003389
3390 /*
3391 * If this is the active queue, check if it needs to be expired,
3392 * or if we want to idle in case it has no pending requests.
3393 */
3394 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003395 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3396
Jens Axboe44f7c162007-01-19 11:51:58 +11003397 if (cfq_cfqq_slice_new(cfqq)) {
3398 cfq_set_prio_slice(cfqd, cfqq);
3399 cfq_clear_cfqq_slice_new(cfqq);
3400 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003401
3402 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003403 * Should we wait for next request to come in before we expire
3404 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003405 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003406 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyalf75edf22009-12-03 12:59:53 -05003407 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3408 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003409 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003410 }
3411
Jens Axboea36e71f2009-04-15 12:15:11 +02003412 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003413 * Idling is not enabled on:
3414 * - expired queues
3415 * - idle-priority queues
3416 * - async queues
3417 * - queues with still some requests queued
3418 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003419 */
Jens Axboe08717142008-01-28 11:38:15 +01003420 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003421 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003422 else if (sync && cfqq_empty &&
3423 !cfq_close_cooperator(cfqd, cfqq)) {
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003424 cfqd->noidle_tree_requires_idle |=
3425 !(rq->cmd_flags & REQ_NOIDLE);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003426 /*
3427 * Idling is enabled for SYNC_WORKLOAD.
3428 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003429 * only if we processed at least one !REQ_NOIDLE request
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003430 */
3431 if (cfqd->serving_type == SYNC_WORKLOAD
Vivek Goyalc04645e2009-12-03 12:59:56 -05003432 || cfqd->noidle_tree_requires_idle
3433 || cfqq->cfqg->nr_cfqq == 1)
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003434 cfq_arm_slice_timer(cfqd);
3435 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003436 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003437
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003438 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003439 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440}
3441
Jens Axboe22e2c502005-06-27 10:55:12 +02003442/*
3443 * we temporarily boost lower priority queues if they are holding fs exclusive
3444 * resources. they are boosted to normal prio (CLASS_BE/4)
3445 */
3446static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447{
Jens Axboe22e2c502005-06-27 10:55:12 +02003448 if (has_fs_excl()) {
3449 /*
3450 * boost idle prio on transactions that would lock out other
3451 * users of the filesystem
3452 */
3453 if (cfq_class_idle(cfqq))
3454 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3455 if (cfqq->ioprio > IOPRIO_NORM)
3456 cfqq->ioprio = IOPRIO_NORM;
3457 } else {
3458 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003459 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003460 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003461 cfqq->ioprio_class = cfqq->org_ioprio_class;
3462 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003464}
3465
Jens Axboe89850f72006-07-22 16:48:31 +02003466static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003467{
Jens Axboe1b379d82009-08-11 08:26:11 +02003468 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003469 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003470 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003471 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003472
3473 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003474}
3475
Jens Axboe165125e2007-07-24 09:28:11 +02003476static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003477{
3478 struct cfq_data *cfqd = q->elevator->elevator_data;
3479 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003480 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003481 struct cfq_queue *cfqq;
3482
3483 /*
3484 * don't force setup of a queue from here, as a call to may_queue
3485 * does not necessarily imply that a request actually will be queued.
3486 * so just lookup a possibly existing queue, or return 'may queue'
3487 * if that fails
3488 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003489 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003490 if (!cic)
3491 return ELV_MQUEUE_MAY;
3492
Jens Axboeb0b78f82009-04-08 10:56:08 +02003493 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003494 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003495 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003496 cfq_prio_boost(cfqq);
3497
Jens Axboe89850f72006-07-22 16:48:31 +02003498 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003499 }
3500
3501 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502}
3503
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504/*
3505 * queue lock held here
3506 */
Jens Axboebb37b942006-12-01 10:42:33 +01003507static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508{
Jens Axboe5e705372006-07-13 12:39:25 +02003509 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510
Jens Axboe5e705372006-07-13 12:39:25 +02003511 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003512 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513
Jens Axboe22e2c502005-06-27 10:55:12 +02003514 BUG_ON(!cfqq->allocated[rw]);
3515 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516
Jens Axboe5e705372006-07-13 12:39:25 +02003517 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02003520 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003522 /* Put down rq reference on cfqg */
3523 cfq_put_cfqg(RQ_CFQG(rq));
3524 rq->elevator_private3 = NULL;
3525
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526 cfq_put_queue(cfqq);
3527 }
3528}
3529
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003530static struct cfq_queue *
3531cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3532 struct cfq_queue *cfqq)
3533{
3534 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3535 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003536 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003537 cfq_put_queue(cfqq);
3538 return cic_to_cfqq(cic, 1);
3539}
3540
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003541/*
3542 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3543 * was the last process referring to said cfqq.
3544 */
3545static struct cfq_queue *
3546split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3547{
3548 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003549 cfqq->pid = current->pid;
3550 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003551 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003552 return cfqq;
3553 }
3554
3555 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003556
3557 cfq_put_cooperator(cfqq);
3558
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003559 cfq_put_queue(cfqq);
3560 return NULL;
3561}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003563 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003565static int
Jens Axboe165125e2007-07-24 09:28:11 +02003566cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567{
3568 struct cfq_data *cfqd = q->elevator->elevator_data;
3569 struct cfq_io_context *cic;
3570 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003571 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003572 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 unsigned long flags;
3574
3575 might_sleep_if(gfp_mask & __GFP_WAIT);
3576
Jens Axboee2d74ac2006-03-28 08:59:01 +02003577 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003578
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579 spin_lock_irqsave(q->queue_lock, flags);
3580
Jens Axboe22e2c502005-06-27 10:55:12 +02003581 if (!cic)
3582 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003584new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003585 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003586 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003587 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003588 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003589 } else {
3590 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003591 * If the queue was seeky for too long, break it apart.
3592 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003593 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003594 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3595 cfqq = split_cfqq(cic, cfqq);
3596 if (!cfqq)
3597 goto new_queue;
3598 }
3599
3600 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003601 * Check to see if this queue is scheduled to merge with
3602 * another, closely cooperating queue. The merging of
3603 * queues happens here as it must be done in process context.
3604 * The reference on new_cfqq was taken in merge_cfqqs.
3605 */
3606 if (cfqq->new_cfqq)
3607 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003608 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609
3610 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02003611 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02003612
Linus Torvalds1da177e2005-04-16 15:20:36 -07003613 spin_unlock_irqrestore(q->queue_lock, flags);
3614
Jens Axboe5e705372006-07-13 12:39:25 +02003615 rq->elevator_private = cic;
3616 rq->elevator_private2 = cfqq;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003617 rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
Jens Axboe5e705372006-07-13 12:39:25 +02003618 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003619
Jens Axboe22e2c502005-06-27 10:55:12 +02003620queue_fail:
3621 if (cic)
3622 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003623
Jens Axboe23e018a2009-10-05 08:52:35 +02003624 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003625 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003626 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627 return 1;
3628}
3629
David Howells65f27f32006-11-22 14:55:48 +00003630static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003631{
David Howells65f27f32006-11-22 14:55:48 +00003632 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003633 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003634 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003635
Jens Axboe40bb54d2009-04-15 12:11:10 +02003636 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09003637 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003638 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003639}
3640
3641/*
3642 * Timer running if the active_queue is currently idling inside its time slice
3643 */
3644static void cfq_idle_slice_timer(unsigned long data)
3645{
3646 struct cfq_data *cfqd = (struct cfq_data *) data;
3647 struct cfq_queue *cfqq;
3648 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003649 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003650
Jens Axboe7b679132008-05-30 12:23:07 +02003651 cfq_log(cfqd, "idle timer fired");
3652
Jens Axboe22e2c502005-06-27 10:55:12 +02003653 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3654
Jens Axboefe094d92008-01-31 13:08:54 +01003655 cfqq = cfqd->active_queue;
3656 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003657 timed_out = 0;
3658
Jens Axboe22e2c502005-06-27 10:55:12 +02003659 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003660 * We saw a request before the queue expired, let it through
3661 */
3662 if (cfq_cfqq_must_dispatch(cfqq))
3663 goto out_kick;
3664
3665 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003666 * expired
3667 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003668 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003669 goto expire;
3670
3671 /*
3672 * only expire and reinvoke request handler, if there are
3673 * other queues with pending requests
3674 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003675 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003676 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003677
3678 /*
3679 * not expired and it has a request pending, let it dispatch
3680 */
Jens Axboe75e50982009-04-07 08:56:14 +02003681 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003682 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003683
3684 /*
3685 * Queue depth flag is reset only when the idle didn't succeed
3686 */
3687 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003688 }
3689expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003690 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003691out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003692 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003693out_cont:
3694 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3695}
3696
Jens Axboe3b181522005-06-27 10:56:24 +02003697static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3698{
3699 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003700 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003701}
Jens Axboe22e2c502005-06-27 10:55:12 +02003702
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003703static void cfq_put_async_queues(struct cfq_data *cfqd)
3704{
3705 int i;
3706
3707 for (i = 0; i < IOPRIO_BE_NR; i++) {
3708 if (cfqd->async_cfqq[0][i])
3709 cfq_put_queue(cfqd->async_cfqq[0][i]);
3710 if (cfqd->async_cfqq[1][i])
3711 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003712 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003713
3714 if (cfqd->async_idle_cfqq)
3715 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003716}
3717
Jens Axboebb729bc2009-12-06 09:54:19 +01003718static void cfq_cfqd_free(struct rcu_head *head)
3719{
3720 kfree(container_of(head, struct cfq_data, rcu));
3721}
3722
Jens Axboeb374d182008-10-31 10:05:07 +01003723static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724{
Jens Axboe22e2c502005-06-27 10:55:12 +02003725 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003726 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003727
Jens Axboe3b181522005-06-27 10:56:24 +02003728 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003729
Al Virod9ff4182006-03-18 13:51:22 -05003730 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003731
Al Virod9ff4182006-03-18 13:51:22 -05003732 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003733 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003734
3735 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003736 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3737 struct cfq_io_context,
3738 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003739
3740 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003741 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003742
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003743 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003744 cfq_release_cfq_groups(cfqd);
Vivek Goyale98ef892010-06-18 10:39:47 -04003745 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003746
Al Virod9ff4182006-03-18 13:51:22 -05003747 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003748
3749 cfq_shutdown_timer_wq(cfqd);
3750
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003751 spin_lock(&cic_index_lock);
3752 ida_remove(&cic_index_ida, cfqd->cic_index);
3753 spin_unlock(&cic_index_lock);
3754
Vivek Goyalb1c35762009-12-03 12:59:47 -05003755 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003756 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757}
3758
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003759static int cfq_alloc_cic_index(void)
3760{
3761 int index, error;
3762
3763 do {
3764 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3765 return -ENOMEM;
3766
3767 spin_lock(&cic_index_lock);
3768 error = ida_get_new(&cic_index_ida, &index);
3769 spin_unlock(&cic_index_lock);
3770 if (error && error != -EAGAIN)
3771 return error;
3772 } while (error);
3773
3774 return index;
3775}
3776
Jens Axboe165125e2007-07-24 09:28:11 +02003777static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778{
3779 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003780 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003781 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003782 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003784 i = cfq_alloc_cic_index();
3785 if (i < 0)
3786 return NULL;
3787
Christoph Lameter94f60302007-07-17 04:03:29 -07003788 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003790 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003792 cfqd->cic_index = i;
3793
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003794 /* Init root service tree */
3795 cfqd->grp_service_tree = CFQ_RB_ROOT;
3796
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003797 /* Init root group */
3798 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003799 for_each_cfqg_st(cfqg, i, j, st)
3800 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003801 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003802
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003803 /* Give preference to root group over other groups */
3804 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3805
Vivek Goyal25fb5162009-12-03 12:59:46 -05003806#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003807 /*
3808 * Take a reference to root group which we never drop. This is just
3809 * to make sure that cfq_put_cfqg() does not try to kfree root group
3810 */
3811 atomic_set(&cfqg->ref, 1);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003812 rcu_read_lock();
Vivek Goyale98ef892010-06-18 10:39:47 -04003813 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3814 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003815 rcu_read_unlock();
Vivek Goyal25fb5162009-12-03 12:59:46 -05003816#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003817 /*
3818 * Not strictly needed (since RB_ROOT just clears the node and we
3819 * zeroed cfqd on alloc), but better be safe in case someone decides
3820 * to add magic to the rb code
3821 */
3822 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3823 cfqd->prio_trees[i] = RB_ROOT;
3824
Jens Axboe6118b702009-06-30 09:34:12 +02003825 /*
3826 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3827 * Grab a permanent reference to it, so that the normal code flow
3828 * will not attempt to free it.
3829 */
3830 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3831 atomic_inc(&cfqd->oom_cfqq.ref);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003832 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003833
Al Virod9ff4182006-03-18 13:51:22 -05003834 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003835
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003837
Jens Axboe22e2c502005-06-27 10:55:12 +02003838 init_timer(&cfqd->idle_slice_timer);
3839 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3840 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3841
Jens Axboe23e018a2009-10-05 08:52:35 +02003842 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003843
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003845 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3846 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847 cfqd->cfq_back_max = cfq_back_max;
3848 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003849 cfqd->cfq_slice[0] = cfq_slice_async;
3850 cfqd->cfq_slice[1] = cfq_slice_sync;
3851 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3852 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003853 cfqd->cfq_latency = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05003854 cfqd->cfq_group_isolation = 0;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003855 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003856 /*
3857 * we optimistically start assuming sync ops weren't delayed in last
3858 * second, in order to have larger depth for async operations.
3859 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003860 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003861 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862}
3863
3864static void cfq_slab_kill(void)
3865{
Jens Axboed6de8be2008-05-28 14:46:59 +02003866 /*
3867 * Caller already ensured that pending RCU callbacks are completed,
3868 * so we should have no busy allocations at this point.
3869 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870 if (cfq_pool)
3871 kmem_cache_destroy(cfq_pool);
3872 if (cfq_ioc_pool)
3873 kmem_cache_destroy(cfq_ioc_pool);
3874}
3875
3876static int __init cfq_slab_setup(void)
3877{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003878 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879 if (!cfq_pool)
3880 goto fail;
3881
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003882 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883 if (!cfq_ioc_pool)
3884 goto fail;
3885
3886 return 0;
3887fail:
3888 cfq_slab_kill();
3889 return -ENOMEM;
3890}
3891
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892/*
3893 * sysfs parts below -->
3894 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003895static ssize_t
3896cfq_var_show(unsigned int var, char *page)
3897{
3898 return sprintf(page, "%d\n", var);
3899}
3900
3901static ssize_t
3902cfq_var_store(unsigned int *var, const char *page, size_t count)
3903{
3904 char *p = (char *) page;
3905
3906 *var = simple_strtoul(p, &p, 10);
3907 return count;
3908}
3909
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003911static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003913 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914 unsigned int __data = __VAR; \
3915 if (__CONV) \
3916 __data = jiffies_to_msecs(__data); \
3917 return cfq_var_show(__data, (page)); \
3918}
3919SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003920SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3921SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003922SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3923SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003924SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3925SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3926SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3927SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003928SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003929SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930#undef SHOW_FUNCTION
3931
3932#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003933static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003935 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936 unsigned int __data; \
3937 int ret = cfq_var_store(&__data, (page), count); \
3938 if (__data < (MIN)) \
3939 __data = (MIN); \
3940 else if (__data > (MAX)) \
3941 __data = (MAX); \
3942 if (__CONV) \
3943 *(__PTR) = msecs_to_jiffies(__data); \
3944 else \
3945 *(__PTR) = __data; \
3946 return ret; \
3947}
3948STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003949STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3950 UINT_MAX, 1);
3951STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3952 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05003953STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003954STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3955 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003956STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3957STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3958STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01003959STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3960 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003961STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003962STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003963#undef STORE_FUNCTION
3964
Al Viroe572ec72006-03-18 22:27:18 -05003965#define CFQ_ATTR(name) \
3966 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02003967
Al Viroe572ec72006-03-18 22:27:18 -05003968static struct elv_fs_entry cfq_attrs[] = {
3969 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05003970 CFQ_ATTR(fifo_expire_sync),
3971 CFQ_ATTR(fifo_expire_async),
3972 CFQ_ATTR(back_seek_max),
3973 CFQ_ATTR(back_seek_penalty),
3974 CFQ_ATTR(slice_sync),
3975 CFQ_ATTR(slice_async),
3976 CFQ_ATTR(slice_async_rq),
3977 CFQ_ATTR(slice_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02003978 CFQ_ATTR(low_latency),
Vivek Goyalae30c282009-12-03 12:59:55 -05003979 CFQ_ATTR(group_isolation),
Al Viroe572ec72006-03-18 22:27:18 -05003980 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981};
3982
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983static struct elevator_type iosched_cfq = {
3984 .ops = {
3985 .elevator_merge_fn = cfq_merge,
3986 .elevator_merged_fn = cfq_merged_request,
3987 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01003988 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07003989 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02003990 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02003992 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993 .elevator_deactivate_req_fn = cfq_deactivate_request,
3994 .elevator_queue_empty_fn = cfq_queue_empty,
3995 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02003996 .elevator_former_req_fn = elv_rb_former_request,
3997 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998 .elevator_set_req_fn = cfq_set_request,
3999 .elevator_put_req_fn = cfq_put_request,
4000 .elevator_may_queue_fn = cfq_may_queue,
4001 .elevator_init_fn = cfq_init_queue,
4002 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02004003 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004004 },
Al Viro3d1ab402006-03-18 18:35:43 -05004005 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006 .elevator_name = "cfq",
4007 .elevator_owner = THIS_MODULE,
4008};
4009
Vivek Goyal3e252062009-12-04 10:36:42 -05004010#ifdef CONFIG_CFQ_GROUP_IOSCHED
4011static struct blkio_policy_type blkio_policy_cfq = {
4012 .ops = {
4013 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4014 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4015 },
4016};
4017#else
4018static struct blkio_policy_type blkio_policy_cfq;
4019#endif
4020
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021static int __init cfq_init(void)
4022{
Jens Axboe22e2c502005-06-27 10:55:12 +02004023 /*
4024 * could be 0 on HZ < 1000 setups
4025 */
4026 if (!cfq_slice_async)
4027 cfq_slice_async = 1;
4028 if (!cfq_slice_idle)
4029 cfq_slice_idle = 1;
4030
Linus Torvalds1da177e2005-04-16 15:20:36 -07004031 if (cfq_slab_setup())
4032 return -ENOMEM;
4033
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004034 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05004035 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004037 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038}
4039
4040static void __exit cfq_exit(void)
4041{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07004042 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05004043 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05004045 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02004046 /* ioc_gone's update must be visible before reading ioc_count */
4047 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02004048
4049 /*
4050 * this also protects us from entering cfq_slab_kill() with
4051 * pending RCU callbacks
4052 */
Tejun Heo245b2e72009-06-24 15:13:48 +09004053 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02004054 wait_for_completion(&all_gone);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004055 ida_destroy(&cic_index_ida);
Christoph Hellwig83521d32005-10-30 15:01:39 -08004056 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057}
4058
4059module_init(cfq_init);
4060module_exit(cfq_exit);
4061
4062MODULE_AUTHOR("Jens Axboe");
4063MODULE_LICENSE("GPL");
4064MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");