blob: 4cd59b0d7c15b2b0df8c6e7323afe5cd51359969 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * CFQ, or complete fairness queueing, disk scheduler.
3 *
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6 *
Jens Axboe0fe23472006-09-04 15:41:16 +02007 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090010#include <linux/slab.h>
Al Viro1cc9be62006-03-18 12:29:52 -050011#include <linux/blkdev.h>
12#include <linux/elevator.h>
Randy Dunlapad5ebd22009-11-11 13:47:45 +010013#include <linux/jiffies.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020015#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020016#include <linux/blktrace_api.h>
Vivek Goyale98ef892010-06-18 10:39:47 -040017#include "cfq.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070018
19/*
20 * tunables
21 */
Jens Axboefe094d92008-01-31 13:08:54 +010022/* max queue in one round of service */
Shaohua Liabc3c742010-03-01 09:20:54 +010023static const int cfq_quantum = 8;
Arjan van de Ven64100092006-01-06 09:46:02 +010024static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010025/* maximum backwards seek, in KiB */
26static const int cfq_back_max = 16 * 1024;
27/* penalty of a backwards seek */
28static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010029static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020030static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010031static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020032static int cfq_slice_idle = HZ / 125;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +020033static int cfq_group_idle = HZ / 125;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +010034static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35static const int cfq_hist_divisor = 4;
Jens Axboe22e2c502005-06-27 10:55:12 +020036
Jens Axboed9e76202007-04-20 14:27:50 +020037/*
Jens Axboe08717142008-01-28 11:38:15 +010038 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020039 */
Jens Axboe08717142008-01-28 11:38:15 +010040#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020041
42/*
43 * below this threshold, we consider thinktime immediate
44 */
45#define CFQ_MIN_TT (2)
46
Jens Axboe22e2c502005-06-27 10:55:12 +020047#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020048#define CFQ_HW_QUEUE_MIN (5)
Vivek Goyal25bc6b02009-12-03 12:59:43 -050049#define CFQ_SERVICE_SHIFT 12
Jens Axboe22e2c502005-06-27 10:55:12 +020050
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010051#define CFQQ_SEEK_THR (sector_t)(8 * 100)
Shaohua Lie9ce3352010-03-19 08:03:04 +010052#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
Corrado Zoccolo41647e72010-02-27 19:45:40 +010053#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010054#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
Shaohua Liae54abe2010-02-05 13:11:45 +010055
Jens Axboefe094d92008-01-31 13:08:54 +010056#define RQ_CIC(rq) \
57 ((struct cfq_io_context *) (rq)->elevator_private)
Jens Axboe7b679132008-05-30 12:23:07 +020058#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +020059#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private3)
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Christoph Lametere18b8902006-12-06 20:33:20 -080061static struct kmem_cache *cfq_pool;
62static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
Tejun Heo245b2e72009-06-24 15:13:48 +090064static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
Al Viro334e94de2006-03-18 15:05:53 -050065static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020066static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94de2006-03-18 15:05:53 -050067
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +040068static DEFINE_SPINLOCK(cic_index_lock);
69static DEFINE_IDA(cic_index_ida);
70
Jens Axboe22e2c502005-06-27 10:55:12 +020071#define CFQ_PRIO_LISTS IOPRIO_BE_NR
72#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020073#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
Jens Axboe206dc692006-03-28 13:03:44 +020075#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050076#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020077
Jens Axboe22e2c502005-06-27 10:55:12 +020078/*
Jens Axboecc09e292007-04-26 12:53:50 +020079 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
83 */
84struct cfq_rb_root {
85 struct rb_root rb;
86 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010087 unsigned count;
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010088 unsigned total_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050089 u64 min_vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -050090 struct rb_node *active;
Jens Axboecc09e292007-04-26 12:53:50 +020091};
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010092#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
93 .count = 0, .min_vdisktime = 0, }
Jens Axboecc09e292007-04-26 12:53:50 +020094
95/*
Jens Axboe6118b702009-06-30 09:34:12 +020096 * Per process-grouping structure
97 */
98struct cfq_queue {
99 /* reference count */
100 atomic_t ref;
101 /* various state flags, see below */
102 unsigned int flags;
103 /* parent cfq_data */
104 struct cfq_data *cfqd;
105 /* service_tree member */
106 struct rb_node rb_node;
107 /* service_tree key */
108 unsigned long rb_key;
109 /* prio tree member */
110 struct rb_node p_node;
111 /* prio tree root we belong to, if any */
112 struct rb_root *p_root;
113 /* sorted list of pending requests */
114 struct rb_root sort_list;
115 /* if fifo isn't expired, next request to serve */
116 struct request *next_rq;
117 /* requests queued in sort_list */
118 int queued[2];
119 /* currently allocated requests */
120 int allocated[2];
121 /* fifo list of requests in sort_list */
122 struct list_head fifo;
123
Vivek Goyaldae739e2009-12-03 12:59:45 -0500124 /* time when queue got scheduled in to dispatch first request. */
125 unsigned long dispatch_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500126 unsigned int allocated_slice;
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100127 unsigned int slice_dispatch;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500128 /* time when first request from queue completed and slice started. */
129 unsigned long slice_start;
Jens Axboe6118b702009-06-30 09:34:12 +0200130 unsigned long slice_end;
131 long slice_resid;
Jens Axboe6118b702009-06-30 09:34:12 +0200132
133 /* pending metadata requests */
134 int meta_pending;
135 /* number of requests that are on the dispatch list or inside driver */
136 int dispatched;
137
138 /* io prio of this group */
139 unsigned short ioprio, org_ioprio;
140 unsigned short ioprio_class, org_ioprio_class;
141
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100142 pid_t pid;
143
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +0100144 u32 seek_history;
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400145 sector_t last_request_pos;
146
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100147 struct cfq_rb_root *service_tree;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400148 struct cfq_queue *new_cfqq;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500149 struct cfq_group *cfqg;
Vivek Goyalae30c282009-12-03 12:59:55 -0500150 struct cfq_group *orig_cfqg;
Vivek Goyalc4e78932010-08-23 12:25:03 +0200151 /* Number of sectors dispatched from queue in single dispatch round */
152 unsigned long nr_sectors;
Jens Axboe6118b702009-06-30 09:34:12 +0200153};
154
155/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100156 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100157 * IDLE is handled separately, so it has negative index
158 */
159enum wl_prio_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100160 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500161 RT_WORKLOAD = 1,
162 IDLE_WORKLOAD = 2,
Vivek Goyalb4627322010-10-22 09:48:43 +0200163 CFQ_PRIO_NR,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100164};
165
166/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100167 * Second index in the service_trees.
168 */
169enum wl_type_t {
170 ASYNC_WORKLOAD = 0,
171 SYNC_NOIDLE_WORKLOAD = 1,
172 SYNC_WORKLOAD = 2
173};
174
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500175/* This is per cgroup per device grouping structure */
176struct cfq_group {
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500177 /* group service_tree member */
178 struct rb_node rb_node;
179
180 /* group service_tree key */
181 u64 vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500182 unsigned int weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500183 bool on_st;
184
185 /* number of cfqq currently on this group */
186 int nr_cfqq;
187
Jens Axboe22e2c502005-06-27 10:55:12 +0200188 /*
Vivek Goyalb4627322010-10-22 09:48:43 +0200189 * Per group busy queus average. Useful for workload slice calc. We
190 * create the array for each prio class but at run time it is used
191 * only for RT and BE class and slot for IDLE class remains unused.
192 * This is primarily done to avoid confusion and a gcc warning.
193 */
194 unsigned int busy_queues_avg[CFQ_PRIO_NR];
195 /*
196 * rr lists of queues with requests. We maintain service trees for
197 * RT and BE classes. These trees are subdivided in subclasses
198 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199 * class there is no subclassification and all the cfq queues go on
200 * a single tree service_tree_idle.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100201 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200202 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100203 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100204 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500205
206 unsigned long saved_workload_slice;
207 enum wl_type_t saved_workload;
208 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500209 struct blkio_group blkg;
210#ifdef CONFIG_CFQ_GROUP_IOSCHED
211 struct hlist_node cfqd_node;
Vivek Goyalb1c35762009-12-03 12:59:47 -0500212 atomic_t ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500213#endif
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200214 /* number of requests that are on the dispatch list or inside driver */
215 int dispatched;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500216};
217
218/*
219 * Per block device queue structure
220 */
221struct cfq_data {
222 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500223 /* Root service tree for cfq_groups */
224 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500225 struct cfq_group root_group;
226
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100227 /*
228 * The priority currently being served
229 */
230 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100231 enum wl_type_t serving_type;
232 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500233 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200234
235 /*
236 * Each priority tree is sorted by next_request position. These
237 * trees are used when determining if two or more queues are
238 * interleaving requests (see cfq_close_cooperator).
239 */
240 struct rb_root prio_trees[CFQ_PRIO_LISTS];
241
Jens Axboe22e2c502005-06-27 10:55:12 +0200242 unsigned int busy_queues;
243
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100244 int rq_in_driver;
245 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200246
247 /*
248 * queue-depth detection
249 */
250 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200251 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100252 /*
253 * hw_tag can be
254 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256 * 0 => no NCQ
257 */
258 int hw_tag_est_depth;
259 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200260
261 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200262 * idle window management
263 */
264 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200265 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200266
267 struct cfq_queue *active_queue;
268 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200269
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200270 /*
271 * async queue for each priority case
272 */
273 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200275
Jens Axboe6d048f52007-04-25 12:44:27 +0200276 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 /*
279 * tunables, see top of file
280 */
281 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200282 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 unsigned int cfq_back_penalty;
284 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200285 unsigned int cfq_slice[2];
286 unsigned int cfq_slice_async_rq;
287 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200288 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200289 unsigned int cfq_latency;
Vivek Goyalae30c282009-12-03 12:59:55 -0500290 unsigned int cfq_group_isolation;
Al Virod9ff4182006-03-18 13:51:22 -0500291
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400292 unsigned int cic_index;
Al Virod9ff4182006-03-18 13:51:22 -0500293 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294
Jens Axboe6118b702009-06-30 09:34:12 +0200295 /*
296 * Fallback dummy cfqq for extreme OOM conditions
297 */
298 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200299
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100300 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500301
302 /* List of cfq groups being managed on this device*/
303 struct hlist_head cfqg_list;
Jens Axboebb729bc2009-12-06 09:54:19 +0100304 struct rcu_head rcu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305};
306
Vivek Goyal25fb5162009-12-03 12:59:46 -0500307static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
308
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500309static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
310 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500311 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100312{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500313 if (!cfqg)
314 return NULL;
315
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100316 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500317 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100318
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500319 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100320}
321
Jens Axboe3b181522005-06-27 10:56:24 +0200322enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100323 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
324 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200325 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100326 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100327 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
328 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
329 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100330 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200331 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400332 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100333 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100334 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500335 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200336};
337
338#define CFQ_CFQQ_FNS(name) \
339static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
340{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100341 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200342} \
343static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
344{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100345 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200346} \
347static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
348{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100349 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200350}
351
352CFQ_CFQQ_FNS(on_rr);
353CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200354CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200355CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200356CFQ_CFQQ_FNS(fifo_expire);
357CFQ_CFQQ_FNS(idle_window);
358CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100359CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200360CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200361CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100362CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100363CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500364CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200365#undef CFQ_CFQQ_FNS
366
Vivek Goyalafc24d42010-04-26 19:27:56 +0200367#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500368#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
369 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
370 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
371 blkg_path(&(cfqq)->cfqg->blkg), ##args);
372
373#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
374 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
375 blkg_path(&(cfqg)->blkg), ##args); \
376
377#else
Jens Axboe7b679132008-05-30 12:23:07 +0200378#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
379 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500380#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
381#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200382#define cfq_log(cfqd, fmt, args...) \
383 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
384
Vivek Goyal615f0252009-12-03 12:59:39 -0500385/* Traverses through cfq group service trees */
386#define for_each_cfqg_st(cfqg, i, j, st) \
387 for (i = 0; i <= IDLE_WORKLOAD; i++) \
388 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
389 : &cfqg->service_tree_idle; \
390 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
391 (i == IDLE_WORKLOAD && j == 0); \
392 j++, st = i < IDLE_WORKLOAD ? \
393 &cfqg->service_trees[i][j]: NULL) \
394
395
Vivek Goyal02b35082010-08-23 12:23:53 +0200396static inline bool iops_mode(struct cfq_data *cfqd)
397{
398 /*
399 * If we are not idling on queues and it is a NCQ drive, parallel
400 * execution of requests is on and measuring time is not possible
401 * in most of the cases until and unless we drive shallower queue
402 * depths and that becomes a performance bottleneck. In such cases
403 * switch to start providing fairness in terms of number of IOs.
404 */
405 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
406 return true;
407 else
408 return false;
409}
410
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100411static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
412{
413 if (cfq_class_idle(cfqq))
414 return IDLE_WORKLOAD;
415 if (cfq_class_rt(cfqq))
416 return RT_WORKLOAD;
417 return BE_WORKLOAD;
418}
419
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100420
421static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
422{
423 if (!cfq_cfqq_sync(cfqq))
424 return ASYNC_WORKLOAD;
425 if (!cfq_cfqq_idle_window(cfqq))
426 return SYNC_NOIDLE_WORKLOAD;
427 return SYNC_WORKLOAD;
428}
429
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500430static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
431 struct cfq_data *cfqd,
432 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100433{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500434 if (wl == IDLE_WORKLOAD)
435 return cfqg->service_tree_idle.count;
436
437 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
438 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
439 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100440}
441
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500442static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
443 struct cfq_group *cfqg)
444{
445 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
446 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
447}
448
Jens Axboe165125e2007-07-24 09:28:11 +0200449static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200450static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100451 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100452static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200453 struct io_context *);
454
455static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200456 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200457{
Jens Axboea6151c32009-10-07 20:02:57 +0200458 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200459}
460
461static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200462 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200463{
Jens Axboea6151c32009-10-07 20:02:57 +0200464 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200465}
466
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400467#define CIC_DEAD_KEY 1ul
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400468#define CIC_DEAD_INDEX_SHIFT 1
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400469
470static inline void *cfqd_dead_key(struct cfq_data *cfqd)
471{
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400472 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400473}
474
475static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
476{
477 struct cfq_data *cfqd = cic->key;
478
479 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
480 return NULL;
481
482 return cfqd;
483}
484
Vasily Tarasov91fac312007-04-25 12:29:51 +0200485/*
486 * We regard a request as SYNC, if it's either a read or has the SYNC bit
487 * set (in which case it could also be direct WRITE).
488 */
Jens Axboea6151c32009-10-07 20:02:57 +0200489static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200490{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200491 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200492}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700495 * scheduler run of queue, if there are requests pending and no one in the
496 * driver that will restart queueing
497 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200498static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700499{
Jens Axboe7b679132008-05-30 12:23:07 +0200500 if (cfqd->busy_queues) {
501 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200502 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200503 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700504}
505
Jens Axboe165125e2007-07-24 09:28:11 +0200506static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700507{
508 struct cfq_data *cfqd = q->elevator->elevator_data;
509
Vivek Goyalf04a6422009-12-03 12:59:40 -0500510 return !cfqd->rq_queued;
Andrew Morton99f95e52005-06-27 20:14:05 -0700511}
512
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100514 * Scale schedule slice based on io priority. Use the sync time slice only
515 * if a queue is marked sync and has sync io queued. A sync queue with async
516 * io only, should not get full sync slice length.
517 */
Jens Axboea6151c32009-10-07 20:02:57 +0200518static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200519 unsigned short prio)
520{
521 const int base_slice = cfqd->cfq_slice[sync];
522
523 WARN_ON(prio >= IOPRIO_BE_NR);
524
525 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
526}
527
Jens Axboe44f7c162007-01-19 11:51:58 +1100528static inline int
529cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
530{
Jens Axboed9e76202007-04-20 14:27:50 +0200531 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100532}
533
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500534static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
535{
536 u64 d = delta << CFQ_SERVICE_SHIFT;
537
538 d = d * BLKIO_WEIGHT_DEFAULT;
539 do_div(d, cfqg->weight);
540 return d;
541}
542
543static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
544{
545 s64 delta = (s64)(vdisktime - min_vdisktime);
546 if (delta > 0)
547 min_vdisktime = vdisktime;
548
549 return min_vdisktime;
550}
551
552static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
553{
554 s64 delta = (s64)(vdisktime - min_vdisktime);
555 if (delta < 0)
556 min_vdisktime = vdisktime;
557
558 return min_vdisktime;
559}
560
561static void update_min_vdisktime(struct cfq_rb_root *st)
562{
563 u64 vdisktime = st->min_vdisktime;
564 struct cfq_group *cfqg;
565
566 if (st->active) {
567 cfqg = rb_entry_cfqg(st->active);
568 vdisktime = cfqg->vdisktime;
569 }
570
571 if (st->left) {
572 cfqg = rb_entry_cfqg(st->left);
573 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
574 }
575
576 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
577}
578
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100579/*
580 * get averaged number of queues of RT/BE priority.
581 * average is updated, with a formula that gives more weight to higher numbers,
582 * to quickly follows sudden increases and decrease slowly
583 */
584
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500585static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
586 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100587{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100588 unsigned min_q, max_q;
589 unsigned mult = cfq_hist_divisor - 1;
590 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500591 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100592
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500593 min_q = min(cfqg->busy_queues_avg[rt], busy);
594 max_q = max(cfqg->busy_queues_avg[rt], busy);
595 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100596 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500597 return cfqg->busy_queues_avg[rt];
598}
599
600static inline unsigned
601cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
602{
603 struct cfq_rb_root *st = &cfqd->grp_service_tree;
604
605 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100606}
607
Jens Axboe44f7c162007-01-19 11:51:58 +1100608static inline void
609cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
610{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100611 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
612 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500613 /*
614 * interested queues (we consider only the ones with the same
615 * priority class in the cfq group)
616 */
617 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
618 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100619 unsigned sync_slice = cfqd->cfq_slice[1];
620 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500621 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
622
623 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100624 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
625 /* scale low_slice according to IO priority
626 * and sync vs async */
627 unsigned low_slice =
628 min(slice, base_low_slice * slice / sync_slice);
629 /* the adapted slice value is scaled to fit all iqs
630 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500631 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100632 low_slice);
633 }
634 }
Vivek Goyaldae739e2009-12-03 12:59:45 -0500635 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100636 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500637 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200638 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100639}
640
641/*
642 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
643 * isn't valid until the first request from the dispatch is activated
644 * and the slice time set.
645 */
Jens Axboea6151c32009-10-07 20:02:57 +0200646static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100647{
648 if (cfq_cfqq_slice_new(cfqq))
649 return 0;
650 if (time_before(jiffies, cfqq->slice_end))
651 return 0;
652
653 return 1;
654}
655
656/*
Jens Axboe5e7053742006-07-13 12:39:25 +0200657 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200659 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 */
Jens Axboe5e7053742006-07-13 12:39:25 +0200661static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100662cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100664 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200666#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
667#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
668 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
Jens Axboe5e7053742006-07-13 12:39:25 +0200670 if (rq1 == NULL || rq1 == rq2)
671 return rq2;
672 if (rq2 == NULL)
673 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200674
Jens Axboe5e7053742006-07-13 12:39:25 +0200675 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
676 return rq1;
677 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
678 return rq2;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200679 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200680 return rq1;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200681 else if ((rq2->cmd_flags & REQ_META) &&
682 !(rq1->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200683 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684
Tejun Heo83096eb2009-05-07 22:24:39 +0900685 s1 = blk_rq_pos(rq1);
686 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 /*
689 * by definition, 1KiB is 2 sectors
690 */
691 back_max = cfqd->cfq_back_max * 2;
692
693 /*
694 * Strict one way elevator _except_ in the case where we allow
695 * short backward seeks which are biased as twice the cost of a
696 * similar forward seek.
697 */
698 if (s1 >= last)
699 d1 = s1 - last;
700 else if (s1 + back_max >= last)
701 d1 = (last - s1) * cfqd->cfq_back_penalty;
702 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200703 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704
705 if (s2 >= last)
706 d2 = s2 - last;
707 else if (s2 + back_max >= last)
708 d2 = (last - s2) * cfqd->cfq_back_penalty;
709 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200710 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711
712 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713
Andreas Mohre8a99052006-03-28 08:59:49 +0200714 /*
715 * By doing switch() on the bit mask "wrap" we avoid having to
716 * check two variables for all permutations: --> faster!
717 */
718 switch (wrap) {
Jens Axboe5e7053742006-07-13 12:39:25 +0200719 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200720 if (d1 < d2)
Jens Axboe5e7053742006-07-13 12:39:25 +0200721 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200722 else if (d2 < d1)
Jens Axboe5e7053742006-07-13 12:39:25 +0200723 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200724 else {
725 if (s1 >= s2)
Jens Axboe5e7053742006-07-13 12:39:25 +0200726 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200727 else
Jens Axboe5e7053742006-07-13 12:39:25 +0200728 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200729 }
730
731 case CFQ_RQ2_WRAP:
Jens Axboe5e7053742006-07-13 12:39:25 +0200732 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200733 case CFQ_RQ1_WRAP:
Jens Axboe5e7053742006-07-13 12:39:25 +0200734 return rq2;
735 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200736 default:
737 /*
738 * Since both rqs are wrapped,
739 * start with the one that's further behind head
740 * (--> only *one* back seek required),
741 * since back seek takes more time than forward.
742 */
743 if (s1 <= s2)
Jens Axboe5e7053742006-07-13 12:39:25 +0200744 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 else
Jens Axboe5e7053742006-07-13 12:39:25 +0200746 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 }
748}
749
Jens Axboe498d3aa22007-04-26 12:54:48 +0200750/*
751 * The below is leftmost cache rbtree addon
752 */
Jens Axboe08717142008-01-28 11:38:15 +0100753static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200754{
Vivek Goyal615f0252009-12-03 12:59:39 -0500755 /* Service tree is empty */
756 if (!root->count)
757 return NULL;
758
Jens Axboecc09e292007-04-26 12:53:50 +0200759 if (!root->left)
760 root->left = rb_first(&root->rb);
761
Jens Axboe08717142008-01-28 11:38:15 +0100762 if (root->left)
763 return rb_entry(root->left, struct cfq_queue, rb_node);
764
765 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200766}
767
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500768static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
769{
770 if (!root->left)
771 root->left = rb_first(&root->rb);
772
773 if (root->left)
774 return rb_entry_cfqg(root->left);
775
776 return NULL;
777}
778
Jens Axboea36e71f2009-04-15 12:15:11 +0200779static void rb_erase_init(struct rb_node *n, struct rb_root *root)
780{
781 rb_erase(n, root);
782 RB_CLEAR_NODE(n);
783}
784
Jens Axboecc09e292007-04-26 12:53:50 +0200785static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
786{
787 if (root->left == n)
788 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200789 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100790 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200791}
792
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793/*
794 * would be nice to take fifo expire time into account as well
795 */
Jens Axboe5e7053742006-07-13 12:39:25 +0200796static struct request *
797cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
798 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799{
Jens Axboe21183b02006-07-13 12:33:14 +0200800 struct rb_node *rbnext = rb_next(&last->rb_node);
801 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e7053742006-07-13 12:39:25 +0200802 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803
Jens Axboe21183b02006-07-13 12:33:14 +0200804 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805
806 if (rbprev)
Jens Axboe5e7053742006-07-13 12:39:25 +0200807 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200808
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 if (rbnext)
Jens Axboe5e7053742006-07-13 12:39:25 +0200810 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200811 else {
812 rbnext = rb_first(&cfqq->sort_list);
813 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e7053742006-07-13 12:39:25 +0200814 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200815 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100817 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818}
819
Jens Axboed9e76202007-04-20 14:27:50 +0200820static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
821 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822{
Jens Axboed9e76202007-04-20 14:27:50 +0200823 /*
824 * just an approximation, should be ok.
825 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500826 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100827 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200828}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500830static inline s64
831cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
832{
833 return cfqg->vdisktime - st->min_vdisktime;
834}
835
836static void
837__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
838{
839 struct rb_node **node = &st->rb.rb_node;
840 struct rb_node *parent = NULL;
841 struct cfq_group *__cfqg;
842 s64 key = cfqg_key(st, cfqg);
843 int left = 1;
844
845 while (*node != NULL) {
846 parent = *node;
847 __cfqg = rb_entry_cfqg(parent);
848
849 if (key < cfqg_key(st, __cfqg))
850 node = &parent->rb_left;
851 else {
852 node = &parent->rb_right;
853 left = 0;
854 }
855 }
856
857 if (left)
858 st->left = &cfqg->rb_node;
859
860 rb_link_node(&cfqg->rb_node, parent, node);
861 rb_insert_color(&cfqg->rb_node, &st->rb);
862}
863
864static void
865cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
866{
867 struct cfq_rb_root *st = &cfqd->grp_service_tree;
868 struct cfq_group *__cfqg;
869 struct rb_node *n;
870
871 cfqg->nr_cfqq++;
872 if (cfqg->on_st)
873 return;
874
875 /*
876 * Currently put the group at the end. Later implement something
877 * so that groups get lesser vtime based on their weights, so that
878 * if group does not loose all if it was not continously backlogged.
879 */
880 n = rb_last(&st->rb);
881 if (n) {
882 __cfqg = rb_entry_cfqg(n);
883 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
884 } else
885 cfqg->vdisktime = st->min_vdisktime;
886
887 __cfq_group_service_tree_add(st, cfqg);
888 cfqg->on_st = true;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500889 st->total_weight += cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500890}
891
892static void
893cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
894{
895 struct cfq_rb_root *st = &cfqd->grp_service_tree;
896
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500897 if (st->active == &cfqg->rb_node)
898 st->active = NULL;
899
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500900 BUG_ON(cfqg->nr_cfqq < 1);
901 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500902
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500903 /* If there are other cfq queues under this group, don't delete it */
904 if (cfqg->nr_cfqq)
905 return;
906
Vivek Goyal2868ef72009-12-03 12:59:48 -0500907 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500908 cfqg->on_st = false;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500909 st->total_weight -= cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500910 if (!RB_EMPTY_NODE(&cfqg->rb_node))
911 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500912 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400913 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500914}
915
916static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
917{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500918 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500919
920 /*
921 * Queue got expired before even a single request completed or
922 * got expired immediately after first request completion.
923 */
924 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
925 /*
926 * Also charge the seek time incurred to the group, otherwise
927 * if there are mutiple queues in the group, each can dispatch
928 * a single request on seeky media and cause lots of seek time
929 * and group will never know it.
930 */
931 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
932 1);
933 } else {
934 slice_used = jiffies - cfqq->slice_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500935 if (slice_used > cfqq->allocated_slice)
936 slice_used = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500937 }
938
Vivek Goyaldae739e2009-12-03 12:59:45 -0500939 return slice_used;
940}
941
942static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200943 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500944{
945 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal02b35082010-08-23 12:23:53 +0200946 unsigned int used_sl, charge;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500947 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
948 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500949
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500950 BUG_ON(nr_sync < 0);
Vivek Goyal02b35082010-08-23 12:23:53 +0200951 used_sl = charge = cfq_cfqq_slice_usage(cfqq);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500952
Vivek Goyal02b35082010-08-23 12:23:53 +0200953 if (iops_mode(cfqd))
954 charge = cfqq->slice_dispatch;
955 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
956 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500957
958 /* Can't update vdisktime while group is on service tree */
959 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyal02b35082010-08-23 12:23:53 +0200960 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500961 __cfq_group_service_tree_add(st, cfqg);
962
963 /* This group is being expired. Save the context */
964 if (time_after(cfqd->workload_expires, jiffies)) {
965 cfqg->saved_workload_slice = cfqd->workload_expires
966 - jiffies;
967 cfqg->saved_workload = cfqd->serving_type;
968 cfqg->saved_serving_prio = cfqd->serving_prio;
969 } else
970 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500971
972 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
973 st->min_vdisktime);
Vivek Goyalc4e78932010-08-23 12:25:03 +0200974 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
975 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
976 iops_mode(cfqd), cfqq->nr_sectors);
Vivek Goyale98ef892010-06-18 10:39:47 -0400977 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
978 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500979}
980
Vivek Goyal25fb5162009-12-03 12:59:46 -0500981#ifdef CONFIG_CFQ_GROUP_IOSCHED
982static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
983{
984 if (blkg)
985 return container_of(blkg, struct cfq_group, blkg);
986 return NULL;
987}
988
Vivek Goyalfe0714372010-10-01 14:49:49 +0200989void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
990 unsigned int weight)
Vivek Goyalf8d461d2009-12-03 12:59:52 -0500991{
992 cfqg_of_blkg(blkg)->weight = weight;
993}
994
Vivek Goyal25fb5162009-12-03 12:59:46 -0500995static struct cfq_group *
996cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
997{
998 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
999 struct cfq_group *cfqg = NULL;
1000 void *key = cfqd;
1001 int i, j;
1002 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -05001003 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1004 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001005
Vivek Goyal25fb5162009-12-03 12:59:46 -05001006 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
Ricky Beniteza74b2ad2010-04-05 18:22:17 +02001007 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1008 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1009 cfqg->blkg.dev = MKDEV(major, minor);
1010 goto done;
1011 }
Vivek Goyal25fb5162009-12-03 12:59:46 -05001012 if (cfqg || !create)
1013 goto done;
1014
1015 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1016 if (!cfqg)
1017 goto done;
1018
Vivek Goyal25fb5162009-12-03 12:59:46 -05001019 for_each_cfqg_st(cfqg, i, j, st)
1020 *st = CFQ_RB_ROOT;
1021 RB_CLEAR_NODE(&cfqg->rb_node);
1022
Vivek Goyalb1c35762009-12-03 12:59:47 -05001023 /*
1024 * Take the initial reference that will be released on destroy
1025 * This can be thought of a joint reference by cgroup and
1026 * elevator which will be dropped by either elevator exit
1027 * or cgroup deletion path depending on who is exiting first.
1028 */
1029 atomic_set(&cfqg->ref, 1);
1030
Vivek Goyal180be2a2010-09-14 08:47:11 +02001031 /*
1032 * Add group onto cgroup list. It might happen that bdi->dev is
1033 * not initiliazed yet. Initialize this new group without major
1034 * and minor info and this info will be filled in once a new thread
1035 * comes for IO. See code above.
1036 */
1037 if (bdi->dev) {
1038 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1039 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
Vivek Goyal22084192009-12-03 12:59:49 -05001040 MKDEV(major, minor));
Vivek Goyal180be2a2010-09-14 08:47:11 +02001041 } else
1042 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1043 0);
1044
Gui Jianfeng34d0f172010-04-13 16:05:49 +08001045 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001046
1047 /* Add group on cfqd list */
1048 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1049
1050done:
Vivek Goyal25fb5162009-12-03 12:59:46 -05001051 return cfqg;
1052}
1053
1054/*
1055 * Search for the cfq group current task belongs to. If create = 1, then also
1056 * create the cfq group if it does not exist. request_queue lock must be held.
1057 */
1058static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1059{
1060 struct cgroup *cgroup;
1061 struct cfq_group *cfqg = NULL;
1062
1063 rcu_read_lock();
1064 cgroup = task_cgroup(current, blkio_subsys_id);
1065 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1066 if (!cfqg && create)
1067 cfqg = &cfqd->root_group;
1068 rcu_read_unlock();
1069 return cfqg;
1070}
1071
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001072static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1073{
1074 atomic_inc(&cfqg->ref);
1075 return cfqg;
1076}
1077
Vivek Goyal25fb5162009-12-03 12:59:46 -05001078static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1079{
1080 /* Currently, all async queues are mapped to root group */
1081 if (!cfq_cfqq_sync(cfqq))
1082 cfqg = &cfqq->cfqd->root_group;
1083
1084 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001085 /* cfqq reference on cfqg */
1086 atomic_inc(&cfqq->cfqg->ref);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001087}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001088
1089static void cfq_put_cfqg(struct cfq_group *cfqg)
1090{
1091 struct cfq_rb_root *st;
1092 int i, j;
1093
1094 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1095 if (!atomic_dec_and_test(&cfqg->ref))
1096 return;
1097 for_each_cfqg_st(cfqg, i, j, st)
1098 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1099 kfree(cfqg);
1100}
1101
1102static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1103{
1104 /* Something wrong if we are trying to remove same group twice */
1105 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1106
1107 hlist_del_init(&cfqg->cfqd_node);
1108
1109 /*
1110 * Put the reference taken at the time of creation so that when all
1111 * queues are gone, group can be destroyed.
1112 */
1113 cfq_put_cfqg(cfqg);
1114}
1115
1116static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1117{
1118 struct hlist_node *pos, *n;
1119 struct cfq_group *cfqg;
1120
1121 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1122 /*
1123 * If cgroup removal path got to blk_group first and removed
1124 * it from cgroup list, then it will take care of destroying
1125 * cfqg also.
1126 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001127 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001128 cfq_destroy_cfqg(cfqd, cfqg);
1129 }
1130}
1131
1132/*
1133 * Blk cgroup controller notification saying that blkio_group object is being
1134 * delinked as associated cgroup object is going away. That also means that
1135 * no new IO will come in this group. So get rid of this group as soon as
1136 * any pending IO in the group is finished.
1137 *
1138 * This function is called under rcu_read_lock(). key is the rcu protected
1139 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1140 * read lock.
1141 *
1142 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1143 * it should not be NULL as even if elevator was exiting, cgroup deltion
1144 * path got to it first.
1145 */
1146void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1147{
1148 unsigned long flags;
1149 struct cfq_data *cfqd = key;
1150
1151 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1152 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1153 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1154}
1155
Vivek Goyal25fb5162009-12-03 12:59:46 -05001156#else /* GROUP_IOSCHED */
1157static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1158{
1159 return &cfqd->root_group;
1160}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001161
1162static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1163{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001164 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001165}
1166
Vivek Goyal25fb5162009-12-03 12:59:46 -05001167static inline void
1168cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1169 cfqq->cfqg = cfqg;
1170}
1171
Vivek Goyalb1c35762009-12-03 12:59:47 -05001172static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1173static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1174
Vivek Goyal25fb5162009-12-03 12:59:46 -05001175#endif /* GROUP_IOSCHED */
1176
Jens Axboe498d3aa22007-04-26 12:54:48 +02001177/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001178 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001179 * requests waiting to be processed. It is sorted in the order that
1180 * we will service the queues.
1181 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001182static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001183 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001184{
Jens Axboe08717142008-01-28 11:38:15 +01001185 struct rb_node **p, *parent;
1186 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001187 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001188 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001189 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001190 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001191 int group_changed = 0;
1192
1193#ifdef CONFIG_CFQ_GROUP_IOSCHED
1194 if (!cfqd->cfq_group_isolation
1195 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1196 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1197 /* Move this cfq to root group */
1198 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1199 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1200 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1201 cfqq->orig_cfqg = cfqq->cfqg;
1202 cfqq->cfqg = &cfqd->root_group;
1203 atomic_inc(&cfqd->root_group.ref);
1204 group_changed = 1;
1205 } else if (!cfqd->cfq_group_isolation
1206 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1207 /* cfqq is sequential now needs to go to its original group */
1208 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1209 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1210 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1211 cfq_put_cfqg(cfqq->cfqg);
1212 cfqq->cfqg = cfqq->orig_cfqg;
1213 cfqq->orig_cfqg = NULL;
1214 group_changed = 1;
1215 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1216 }
1217#endif
Jens Axboed9e76202007-04-20 14:27:50 +02001218
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001219 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001220 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001221 if (cfq_class_idle(cfqq)) {
1222 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001223 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001224 if (parent && parent != &cfqq->rb_node) {
1225 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1226 rb_key += __cfqq->rb_key;
1227 } else
1228 rb_key += jiffies;
1229 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001230 /*
1231 * Get our rb key offset. Subtract any residual slice
1232 * value carried from last service. A negative resid
1233 * count indicates slice overrun, and this should position
1234 * the next service time further away in the tree.
1235 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001236 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001237 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001238 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001239 } else {
1240 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001241 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001242 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1243 }
Jens Axboed9e76202007-04-20 14:27:50 +02001244
1245 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001246 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001247 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001248 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001249 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001250 if (rb_key == cfqq->rb_key &&
1251 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001252 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001253
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001254 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1255 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001256 }
Jens Axboed9e76202007-04-20 14:27:50 +02001257
Jens Axboe498d3aa22007-04-26 12:54:48 +02001258 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001259 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001260 cfqq->service_tree = service_tree;
1261 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001262 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001263 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001264
Jens Axboed9e76202007-04-20 14:27:50 +02001265 parent = *p;
1266 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1267
Jens Axboe0c534e02007-04-18 20:01:57 +02001268 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001269 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001270 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001271 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001272 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001273 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001274 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001275 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001276 }
Jens Axboe67060e32007-04-18 20:13:32 +02001277
1278 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001279 }
1280
Jens Axboecc09e292007-04-26 12:53:50 +02001281 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001282 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001283
Jens Axboed9e76202007-04-20 14:27:50 +02001284 cfqq->rb_key = rb_key;
1285 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001286 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1287 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001288 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001289 return;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001290 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291}
1292
Jens Axboea36e71f2009-04-15 12:15:11 +02001293static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001294cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1295 sector_t sector, struct rb_node **ret_parent,
1296 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001297{
Jens Axboea36e71f2009-04-15 12:15:11 +02001298 struct rb_node **p, *parent;
1299 struct cfq_queue *cfqq = NULL;
1300
1301 parent = NULL;
1302 p = &root->rb_node;
1303 while (*p) {
1304 struct rb_node **n;
1305
1306 parent = *p;
1307 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1308
1309 /*
1310 * Sort strictly based on sector. Smallest to the left,
1311 * largest to the right.
1312 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001313 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001314 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001315 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001316 n = &(*p)->rb_left;
1317 else
1318 break;
1319 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001320 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001321 }
1322
1323 *ret_parent = parent;
1324 if (rb_link)
1325 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001326 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001327}
1328
1329static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1330{
Jens Axboea36e71f2009-04-15 12:15:11 +02001331 struct rb_node **p, *parent;
1332 struct cfq_queue *__cfqq;
1333
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001334 if (cfqq->p_root) {
1335 rb_erase(&cfqq->p_node, cfqq->p_root);
1336 cfqq->p_root = NULL;
1337 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001338
1339 if (cfq_class_idle(cfqq))
1340 return;
1341 if (!cfqq->next_rq)
1342 return;
1343
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001344 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001345 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1346 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001347 if (!__cfqq) {
1348 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001349 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1350 } else
1351 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001352}
1353
Jens Axboe498d3aa22007-04-26 12:54:48 +02001354/*
1355 * Update cfqq's position in the service tree.
1356 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001357static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001358{
Jens Axboe6d048f52007-04-25 12:44:27 +02001359 /*
1360 * Resorting requires the cfqq to be on the RR list already.
1361 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001362 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001363 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001364 cfq_prio_tree_add(cfqd, cfqq);
1365 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001366}
1367
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368/*
1369 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001370 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 */
Jens Axboefebffd62008-01-28 13:19:43 +01001372static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373{
Jens Axboe7b679132008-05-30 12:23:07 +02001374 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001375 BUG_ON(cfq_cfqq_on_rr(cfqq));
1376 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 cfqd->busy_queues++;
1378
Jens Axboeedd75ff2007-04-19 12:03:34 +02001379 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380}
1381
Jens Axboe498d3aa22007-04-26 12:54:48 +02001382/*
1383 * Called when the cfqq no longer has requests pending, remove it from
1384 * the service tree.
1385 */
Jens Axboefebffd62008-01-28 13:19:43 +01001386static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387{
Jens Axboe7b679132008-05-30 12:23:07 +02001388 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001389 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1390 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001392 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1393 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1394 cfqq->service_tree = NULL;
1395 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001396 if (cfqq->p_root) {
1397 rb_erase(&cfqq->p_node, cfqq->p_root);
1398 cfqq->p_root = NULL;
1399 }
Jens Axboed9e76202007-04-20 14:27:50 +02001400
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001401 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 BUG_ON(!cfqd->busy_queues);
1403 cfqd->busy_queues--;
1404}
1405
1406/*
1407 * rb tree support functions
1408 */
Jens Axboefebffd62008-01-28 13:19:43 +01001409static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410{
Jens Axboe5e7053742006-07-13 12:39:25 +02001411 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e7053742006-07-13 12:39:25 +02001412 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413
Jens Axboeb4878f22005-10-20 16:42:29 +02001414 BUG_ON(!cfqq->queued[sync]);
1415 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416
Jens Axboe5e7053742006-07-13 12:39:25 +02001417 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
Vivek Goyalf04a6422009-12-03 12:59:40 -05001419 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1420 /*
1421 * Queue will be deleted from service tree when we actually
1422 * expire it later. Right now just remove it from prio tree
1423 * as it is empty.
1424 */
1425 if (cfqq->p_root) {
1426 rb_erase(&cfqq->p_node, cfqq->p_root);
1427 cfqq->p_root = NULL;
1428 }
1429 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430}
1431
Jens Axboe5e7053742006-07-13 12:39:25 +02001432static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433{
Jens Axboe5e7053742006-07-13 12:39:25 +02001434 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001436 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437
Jens Axboe5380a102006-07-13 12:37:56 +02001438 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439
1440 /*
1441 * looks a little odd, but the first insert might return an alias.
1442 * if that happens, put the alias on the dispatch list
1443 */
Jens Axboe21183b02006-07-13 12:33:14 +02001444 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e7053742006-07-13 12:39:25 +02001445 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001446
1447 if (!cfq_cfqq_on_rr(cfqq))
1448 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001449
1450 /*
1451 * check if this request is a better next-serve candidate
1452 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001453 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001454 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001455
1456 /*
1457 * adjust priority tree position, if ->next_rq changes
1458 */
1459 if (prev != cfqq->next_rq)
1460 cfq_prio_tree_add(cfqd, cfqq);
1461
Jens Axboe5044eed2007-04-25 11:53:48 +02001462 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463}
1464
Jens Axboefebffd62008-01-28 13:19:43 +01001465static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466{
Jens Axboe5380a102006-07-13 12:37:56 +02001467 elv_rb_del(&cfqq->sort_list, rq);
1468 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001469 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1470 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e7053742006-07-13 12:39:25 +02001471 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001472 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001473 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1474 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475}
1476
Jens Axboe206dc692006-03-28 13:03:44 +02001477static struct request *
1478cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479{
Jens Axboe206dc692006-03-28 13:03:44 +02001480 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001481 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001482 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483
Jens Axboe4ac845a2008-01-24 08:44:49 +01001484 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001485 if (!cic)
1486 return NULL;
1487
1488 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001489 if (cfqq) {
1490 sector_t sector = bio->bi_sector + bio_sectors(bio);
1491
Jens Axboe21183b02006-07-13 12:33:14 +02001492 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001493 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 return NULL;
1496}
1497
Jens Axboe165125e2007-07-24 09:28:11 +02001498static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001499{
1500 struct cfq_data *cfqd = q->elevator->elevator_data;
1501
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001502 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001503 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001504 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001505
Tejun Heo5b936292009-05-07 22:24:38 +09001506 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001507}
1508
Jens Axboe165125e2007-07-24 09:28:11 +02001509static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510{
Jens Axboe22e2c502005-06-27 10:55:12 +02001511 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001513 WARN_ON(!cfqd->rq_in_driver);
1514 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001515 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001516 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517}
1518
Jens Axboeb4878f22005-10-20 16:42:29 +02001519static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520{
Jens Axboe5e7053742006-07-13 12:39:25 +02001521 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001522
Jens Axboe5e7053742006-07-13 12:39:25 +02001523 if (cfqq->next_rq == rq)
1524 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525
Jens Axboeb4878f22005-10-20 16:42:29 +02001526 list_del_init(&rq->queuelist);
Jens Axboe5e7053742006-07-13 12:39:25 +02001527 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001528
Aaron Carroll45333d52008-08-26 15:52:36 +02001529 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001530 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1531 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02001532 if (rq->cmd_flags & REQ_META) {
Jens Axboe374f84a2006-07-23 01:42:19 +02001533 WARN_ON(!cfqq->meta_pending);
1534 cfqq->meta_pending--;
1535 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536}
1537
Jens Axboe165125e2007-07-24 09:28:11 +02001538static int cfq_merge(struct request_queue *q, struct request **req,
1539 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540{
1541 struct cfq_data *cfqd = q->elevator->elevator_data;
1542 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543
Jens Axboe206dc692006-03-28 13:03:44 +02001544 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001545 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001546 *req = __rq;
1547 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 }
1549
1550 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551}
1552
Jens Axboe165125e2007-07-24 09:28:11 +02001553static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001554 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555{
Jens Axboe21183b02006-07-13 12:33:14 +02001556 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e7053742006-07-13 12:39:25 +02001557 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558
Jens Axboe5e7053742006-07-13 12:39:25 +02001559 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561}
1562
Divyesh Shah812d4022010-04-08 21:14:23 -07001563static void cfq_bio_merged(struct request_queue *q, struct request *req,
1564 struct bio *bio)
1565{
Vivek Goyale98ef892010-06-18 10:39:47 -04001566 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1567 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001568}
1569
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570static void
Jens Axboe165125e2007-07-24 09:28:11 +02001571cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 struct request *next)
1573{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001574 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001575 /*
1576 * reposition in fifo if next is older than rq
1577 */
1578 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001579 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001580 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001581 rq_set_fifo_time(rq, rq_fifo_time(next));
1582 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001583
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001584 if (cfqq->next_rq == next)
1585 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001586 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001587 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1588 rq_data_dir(next), rq_is_sync(next));
Jens Axboe22e2c502005-06-27 10:55:12 +02001589}
1590
Jens Axboe165125e2007-07-24 09:28:11 +02001591static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001592 struct bio *bio)
1593{
1594 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001595 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001596 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001597
1598 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001599 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001600 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001601 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001602 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001603
1604 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001605 * Lookup the cfqq that this bio will be queued with. Allow
1606 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001607 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001608 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001609 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001610 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001611
Vasily Tarasov91fac312007-04-25 12:29:51 +02001612 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001613 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001614}
1615
Divyesh Shah812df482010-04-08 21:15:35 -07001616static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1617{
1618 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001619 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001620}
1621
Jens Axboefebffd62008-01-28 13:19:43 +01001622static void __cfq_set_active_queue(struct cfq_data *cfqd,
1623 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001624{
1625 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001626 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1627 cfqd->serving_prio, cfqd->serving_type);
Vivek Goyale98ef892010-06-18 10:39:47 -04001628 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001629 cfqq->slice_start = 0;
1630 cfqq->dispatch_start = jiffies;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001631 cfqq->allocated_slice = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001632 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001633 cfqq->slice_dispatch = 0;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001634 cfqq->nr_sectors = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001635
Jens Axboe2f5cb732009-04-07 08:51:19 +02001636 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +02001637 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001638 cfq_clear_cfqq_must_alloc_slice(cfqq);
1639 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +11001640 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001641
Divyesh Shah812df482010-04-08 21:15:35 -07001642 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001643 }
1644
1645 cfqd->active_queue = cfqq;
1646}
1647
1648/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001649 * current cfqq expired its slice (or was too idle), select new one
1650 */
1651static void
1652__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001653 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001654{
Jens Axboe7b679132008-05-30 12:23:07 +02001655 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1656
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001657 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001658 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001659
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001660 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001661 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001662
1663 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001664 * If this cfqq is shared between multiple processes, check to
1665 * make sure that those processes are still issuing I/Os within
1666 * the mean seek distance. If not, it may be time to break the
1667 * queues apart again.
1668 */
1669 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1670 cfq_mark_cfqq_split_coop(cfqq);
1671
1672 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001673 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001674 */
Jens Axboe7b679132008-05-30 12:23:07 +02001675 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +11001676 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001677 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1678 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001679
Vivek Goyale5ff0822010-04-26 19:25:11 +02001680 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001681
Vivek Goyalf04a6422009-12-03 12:59:40 -05001682 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1683 cfq_del_cfqq_rr(cfqd, cfqq);
1684
Jens Axboeedd75ff2007-04-19 12:03:34 +02001685 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001686
1687 if (cfqq == cfqd->active_queue)
1688 cfqd->active_queue = NULL;
1689
Vivek Goyaldae739e2009-12-03 12:59:45 -05001690 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1691 cfqd->grp_service_tree.active = NULL;
1692
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001693 if (cfqd->active_cic) {
1694 put_io_context(cfqd->active_cic->ioc);
1695 cfqd->active_cic = NULL;
1696 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001697}
1698
Vivek Goyale5ff0822010-04-26 19:25:11 +02001699static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001700{
1701 struct cfq_queue *cfqq = cfqd->active_queue;
1702
1703 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001704 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001705}
1706
Jens Axboe498d3aa22007-04-26 12:54:48 +02001707/*
1708 * Get next queue for service. Unless we have a queue preemption,
1709 * we'll simply select the first cfqq in the service tree.
1710 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001711static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001712{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001713 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001714 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001715 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001716
Vivek Goyalf04a6422009-12-03 12:59:40 -05001717 if (!cfqd->rq_queued)
1718 return NULL;
1719
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001720 /* There is nothing to dispatch */
1721 if (!service_tree)
1722 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001723 if (RB_EMPTY_ROOT(&service_tree->rb))
1724 return NULL;
1725 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001726}
1727
Vivek Goyalf04a6422009-12-03 12:59:40 -05001728static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1729{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001730 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001731 struct cfq_queue *cfqq;
1732 int i, j;
1733 struct cfq_rb_root *st;
1734
1735 if (!cfqd->rq_queued)
1736 return NULL;
1737
Vivek Goyal25fb5162009-12-03 12:59:46 -05001738 cfqg = cfq_get_next_cfqg(cfqd);
1739 if (!cfqg)
1740 return NULL;
1741
Vivek Goyalf04a6422009-12-03 12:59:40 -05001742 for_each_cfqg_st(cfqg, i, j, st)
1743 if ((cfqq = cfq_rb_first(st)) != NULL)
1744 return cfqq;
1745 return NULL;
1746}
1747
Jens Axboe498d3aa22007-04-26 12:54:48 +02001748/*
1749 * Get and set a new active queue for service.
1750 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001751static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1752 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001753{
Jens Axboee00ef792009-11-04 08:54:55 +01001754 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001755 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001756
Jens Axboe22e2c502005-06-27 10:55:12 +02001757 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001758 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001759}
1760
Jens Axboed9e76202007-04-20 14:27:50 +02001761static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1762 struct request *rq)
1763{
Tejun Heo83096eb2009-05-07 22:24:39 +09001764 if (blk_rq_pos(rq) >= cfqd->last_position)
1765 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001766 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001767 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001768}
1769
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001770static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001771 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001772{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001773 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001774}
1775
Jens Axboea36e71f2009-04-15 12:15:11 +02001776static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1777 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001778{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001779 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001780 struct rb_node *parent, *node;
1781 struct cfq_queue *__cfqq;
1782 sector_t sector = cfqd->last_position;
1783
1784 if (RB_EMPTY_ROOT(root))
1785 return NULL;
1786
1787 /*
1788 * First, if we find a request starting at the end of the last
1789 * request, choose it.
1790 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001791 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001792 if (__cfqq)
1793 return __cfqq;
1794
1795 /*
1796 * If the exact sector wasn't found, the parent of the NULL leaf
1797 * will contain the closest sector.
1798 */
1799 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001800 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001801 return __cfqq;
1802
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001803 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001804 node = rb_next(&__cfqq->p_node);
1805 else
1806 node = rb_prev(&__cfqq->p_node);
1807 if (!node)
1808 return NULL;
1809
1810 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001811 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001812 return __cfqq;
1813
1814 return NULL;
1815}
1816
1817/*
1818 * cfqd - obvious
1819 * cur_cfqq - passed in so that we don't decide that the current queue is
1820 * closely cooperating with itself.
1821 *
1822 * So, basically we're assuming that that cur_cfqq has dispatched at least
1823 * one request, and that cfqd->last_position reflects a position on the disk
1824 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1825 * assumption.
1826 */
1827static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001828 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001829{
1830 struct cfq_queue *cfqq;
1831
Divyesh Shah39c01b22010-03-25 15:45:57 +01001832 if (cfq_class_idle(cur_cfqq))
1833 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001834 if (!cfq_cfqq_sync(cur_cfqq))
1835 return NULL;
1836 if (CFQQ_SEEKY(cur_cfqq))
1837 return NULL;
1838
Jens Axboea36e71f2009-04-15 12:15:11 +02001839 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001840 * Don't search priority tree if it's the only queue in the group.
1841 */
1842 if (cur_cfqq->cfqg->nr_cfqq == 1)
1843 return NULL;
1844
1845 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001846 * We should notice if some of the queues are cooperating, eg
1847 * working closely on the same area of the disk. In that case,
1848 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001849 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001850 cfqq = cfqq_close(cfqd, cur_cfqq);
1851 if (!cfqq)
1852 return NULL;
1853
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001854 /* If new queue belongs to different cfq_group, don't choose it */
1855 if (cur_cfqq->cfqg != cfqq->cfqg)
1856 return NULL;
1857
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001858 /*
1859 * It only makes sense to merge sync queues.
1860 */
1861 if (!cfq_cfqq_sync(cfqq))
1862 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001863 if (CFQQ_SEEKY(cfqq))
1864 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001865
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001866 /*
1867 * Do not merge queues of different priority classes
1868 */
1869 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1870 return NULL;
1871
Jens Axboea36e71f2009-04-15 12:15:11 +02001872 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001873}
1874
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001875/*
1876 * Determine whether we should enforce idle window for this queue.
1877 */
1878
1879static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1880{
1881 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001882 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001883
Vivek Goyalf04a6422009-12-03 12:59:40 -05001884 BUG_ON(!service_tree);
1885 BUG_ON(!service_tree->count);
1886
Vivek Goyalb6508c12010-08-23 12:23:33 +02001887 if (!cfqd->cfq_slice_idle)
1888 return false;
1889
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001890 /* We never do for idle class queues. */
1891 if (prio == IDLE_WORKLOAD)
1892 return false;
1893
1894 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001895 if (cfq_cfqq_idle_window(cfqq) &&
1896 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001897 return true;
1898
1899 /*
1900 * Otherwise, we do only if they are the last ones
1901 * in their service tree.
1902 */
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001903 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1904 return 1;
1905 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1906 service_tree->count);
1907 return 0;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001908}
1909
Jens Axboe6d048f52007-04-25 12:44:27 +02001910static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001911{
Jens Axboe17926692007-01-19 11:59:30 +11001912 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001913 struct cfq_io_context *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001914 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001915
Jens Axboea68bbddba2008-09-24 13:03:33 +02001916 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001917 * SSD device without seek penalty, disable idling. But only do so
1918 * for devices that support queuing, otherwise we still have a problem
1919 * with sync vs async workloads.
Jens Axboea68bbddba2008-09-24 13:03:33 +02001920 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001921 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbddba2008-09-24 13:03:33 +02001922 return;
1923
Jens Axboedd67d052006-06-21 09:36:18 +02001924 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001925 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001926
1927 /*
1928 * idle is disabled, either manually or by past process history
1929 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001930 if (!cfq_should_idle(cfqd, cfqq)) {
1931 /* no queue idling. Check for group idling */
1932 if (cfqd->cfq_group_idle)
1933 group_idle = cfqd->cfq_group_idle;
1934 else
1935 return;
1936 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001937
Jens Axboe22e2c502005-06-27 10:55:12 +02001938 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001939 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001940 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001941 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001942 return;
1943
1944 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001945 * task has exited, don't wait
1946 */
Jens Axboe206dc692006-03-28 13:03:44 +02001947 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001948 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001949 return;
1950
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001951 /*
1952 * If our average think time is larger than the remaining time
1953 * slice, then don't idle. This avoids overrunning the allotted
1954 * time slice.
1955 */
1956 if (sample_valid(cic->ttime_samples) &&
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001957 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1958 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1959 cic->ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001960 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001961 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001962
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001963 /* There are other queues in the group, don't do group idle */
1964 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1965 return;
1966
Jens Axboe3b181522005-06-27 10:56:24 +02001967 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001968
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001969 if (group_idle)
1970 sl = cfqd->cfq_group_idle;
1971 else
1972 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001973
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001974 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001975 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001976 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1977 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978}
1979
Jens Axboe498d3aa22007-04-26 12:54:48 +02001980/*
1981 * Move request from internal lists to the request queue dispatch list.
1982 */
Jens Axboe165125e2007-07-24 09:28:11 +02001983static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001985 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e7053742006-07-13 12:39:25 +02001986 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001987
Jens Axboe7b679132008-05-30 12:23:07 +02001988 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1989
Jeff Moyer06d21882009-09-11 17:08:59 +02001990 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001991 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001992 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001993 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001994 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001995
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001996 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001997 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001998 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02001999 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000}
2001
2002/*
2003 * return expired entry, or NULL to just start from scratch in rbtree
2004 */
Jens Axboefebffd62008-01-28 13:19:43 +01002005static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006{
Jens Axboe30996f42009-10-05 11:03:39 +02002007 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008
Jens Axboe3b181522005-06-27 10:56:24 +02002009 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002011
2012 cfq_mark_cfqq_fifo_expire(cfqq);
2013
Jens Axboe89850f72006-07-22 16:48:31 +02002014 if (list_empty(&cfqq->fifo))
2015 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016
Jens Axboe89850f72006-07-22 16:48:31 +02002017 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002018 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002019 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020
Jens Axboe30996f42009-10-05 11:03:39 +02002021 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002022 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023}
2024
Jens Axboe22e2c502005-06-27 10:55:12 +02002025static inline int
2026cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2027{
2028 const int base_rq = cfqd->cfq_slice_async_rq;
2029
2030 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2031
2032 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2033}
2034
2035/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002036 * Must be called with the queue_lock held.
2037 */
2038static int cfqq_process_refs(struct cfq_queue *cfqq)
2039{
2040 int process_refs, io_refs;
2041
2042 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2043 process_refs = atomic_read(&cfqq->ref) - io_refs;
2044 BUG_ON(process_refs < 0);
2045 return process_refs;
2046}
2047
2048static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2049{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002050 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002051 struct cfq_queue *__cfqq;
2052
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002053 /*
2054 * If there are no process references on the new_cfqq, then it is
2055 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2056 * chain may have dropped their last reference (not just their
2057 * last process reference).
2058 */
2059 if (!cfqq_process_refs(new_cfqq))
2060 return;
2061
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002062 /* Avoid a circular list and skip interim queue merges */
2063 while ((__cfqq = new_cfqq->new_cfqq)) {
2064 if (__cfqq == cfqq)
2065 return;
2066 new_cfqq = __cfqq;
2067 }
2068
2069 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002070 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002071 /*
2072 * If the process for the cfqq has gone away, there is no
2073 * sense in merging the queues.
2074 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002075 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002076 return;
2077
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002078 /*
2079 * Merge in the direction of the lesser amount of work.
2080 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002081 if (new_process_refs >= process_refs) {
2082 cfqq->new_cfqq = new_cfqq;
2083 atomic_add(process_refs, &new_cfqq->ref);
2084 } else {
2085 new_cfqq->new_cfqq = cfqq;
2086 atomic_add(new_process_refs, &cfqq->ref);
2087 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002088}
2089
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002090static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002091 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002092{
2093 struct cfq_queue *queue;
2094 int i;
2095 bool key_valid = false;
2096 unsigned long lowest_key = 0;
2097 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2098
Vivek Goyal65b32a52009-12-16 17:52:59 -05002099 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2100 /* select the one with lowest rb_key */
2101 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002102 if (queue &&
2103 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2104 lowest_key = queue->rb_key;
2105 cur_best = i;
2106 key_valid = true;
2107 }
2108 }
2109
2110 return cur_best;
2111}
2112
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002113static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002114{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002115 unsigned slice;
2116 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002117 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002118 unsigned group_slice;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002119
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002120 if (!cfqg) {
2121 cfqd->serving_prio = IDLE_WORKLOAD;
2122 cfqd->workload_expires = jiffies + 1;
2123 return;
2124 }
2125
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002126 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002127 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002128 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002129 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002130 cfqd->serving_prio = BE_WORKLOAD;
2131 else {
2132 cfqd->serving_prio = IDLE_WORKLOAD;
2133 cfqd->workload_expires = jiffies + 1;
2134 return;
2135 }
2136
2137 /*
2138 * For RT and BE, we have to choose also the type
2139 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2140 * expiration time
2141 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002142 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002143 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002144
2145 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002146 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002147 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002148 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002149 return;
2150
2151 /* otherwise select new workload type */
2152 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002153 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2154 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002155 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002156
2157 /*
2158 * the workload slice is computed as a fraction of target latency
2159 * proportional to the number of queues in that workload, over
2160 * all the queues in the same priority class
2161 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002162 group_slice = cfq_group_slice(cfqd, cfqg);
2163
2164 slice = group_slice * count /
2165 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2166 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002167
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002168 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2169 unsigned int tmp;
2170
2171 /*
2172 * Async queues are currently system wide. Just taking
2173 * proportion of queues with-in same group will lead to higher
2174 * async ratio system wide as generally root group is going
2175 * to have higher weight. A more accurate thing would be to
2176 * calculate system wide asnc/sync ratio.
2177 */
2178 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2179 tmp = tmp/cfqd->busy_queues;
2180 slice = min_t(unsigned, slice, tmp);
2181
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002182 /* async workload slice is scaled down according to
2183 * the sync/async slice ratio. */
2184 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002185 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002186 /* sync workload slice is at least 2 * cfq_slice_idle */
2187 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2188
2189 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002190 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002191 cfqd->workload_expires = jiffies + slice;
2192}
2193
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002194static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2195{
2196 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002197 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002198
2199 if (RB_EMPTY_ROOT(&st->rb))
2200 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002201 cfqg = cfq_rb_first_group(st);
2202 st->active = &cfqg->rb_node;
2203 update_min_vdisktime(st);
2204 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002205}
2206
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002207static void cfq_choose_cfqg(struct cfq_data *cfqd)
2208{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002209 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2210
2211 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002212
2213 /* Restore the workload type data */
2214 if (cfqg->saved_workload_slice) {
2215 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2216 cfqd->serving_type = cfqg->saved_workload;
2217 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002218 } else
2219 cfqd->workload_expires = jiffies - 1;
2220
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002221 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002222}
2223
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002224/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002225 * Select a queue for service. If we have a current active queue,
2226 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002227 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002228static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002229{
Jens Axboea36e71f2009-04-15 12:15:11 +02002230 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002231
2232 cfqq = cfqd->active_queue;
2233 if (!cfqq)
2234 goto new_queue;
2235
Vivek Goyalf04a6422009-12-03 12:59:40 -05002236 if (!cfqd->rq_queued)
2237 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002238
2239 /*
2240 * We were waiting for group to get backlogged. Expire the queue
2241 */
2242 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2243 goto expire;
2244
Jens Axboe22e2c502005-06-27 10:55:12 +02002245 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002246 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002247 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002248 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2249 /*
2250 * If slice had not expired at the completion of last request
2251 * we might not have turned on wait_busy flag. Don't expire
2252 * the queue yet. Allow the group to get backlogged.
2253 *
2254 * The very fact that we have used the slice, that means we
2255 * have been idling all along on this queue and it should be
2256 * ok to wait for this request to complete.
2257 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002258 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2259 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2260 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002261 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002262 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002263 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002264 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002265
2266 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002267 * The active queue has requests and isn't expired, allow it to
2268 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002269 */
Jens Axboedd67d052006-06-21 09:36:18 +02002270 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002271 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002272
2273 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002274 * If another queue has a request waiting within our mean seek
2275 * distance, let it run. The expire code will check for close
2276 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002277 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002278 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002279 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002280 if (new_cfqq) {
2281 if (!cfqq->new_cfqq)
2282 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002283 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002284 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002285
2286 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002287 * No requests pending. If the active queue still has requests in
2288 * flight or is idling for a new request, allow either of these
2289 * conditions to happen (or time out) before selecting a new queue.
2290 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002291 if (timer_pending(&cfqd->idle_slice_timer)) {
2292 cfqq = NULL;
2293 goto keep_queue;
2294 }
2295
2296 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2297 cfqq = NULL;
2298 goto keep_queue;
2299 }
2300
2301 /*
2302 * If group idle is enabled and there are requests dispatched from
2303 * this group, wait for requests to complete.
2304 */
2305check_group_idle:
2306 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2307 && cfqq->cfqg->dispatched) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002308 cfqq = NULL;
2309 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002310 }
2311
Jens Axboe3b181522005-06-27 10:56:24 +02002312expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002313 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002314new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002315 /*
2316 * Current queue expired. Check if we have to switch to a new
2317 * service tree
2318 */
2319 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002320 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002321
Jens Axboea36e71f2009-04-15 12:15:11 +02002322 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002323keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002324 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002325}
2326
Jens Axboefebffd62008-01-28 13:19:43 +01002327static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002328{
2329 int dispatched = 0;
2330
2331 while (cfqq->next_rq) {
2332 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2333 dispatched++;
2334 }
2335
2336 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002337
2338 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002339 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002340 return dispatched;
2341}
2342
Jens Axboe498d3aa22007-04-26 12:54:48 +02002343/*
2344 * Drain our current requests. Used for barriers and when switching
2345 * io schedulers on-the-fly.
2346 */
Jens Axboed9e76202007-04-20 14:27:50 +02002347static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002348{
Jens Axboe08717142008-01-28 11:38:15 +01002349 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002350 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002351
Divyesh Shah3440c492010-04-09 09:29:57 +02002352 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002353 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002354 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2355 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002356 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002357 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002358
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002359 BUG_ON(cfqd->busy_queues);
2360
Jeff Moyer6923715a2009-06-12 15:29:30 +02002361 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002362 return dispatched;
2363}
2364
Shaohua Liabc3c742010-03-01 09:20:54 +01002365static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2366 struct cfq_queue *cfqq)
2367{
2368 /* the queue hasn't finished any request, can't estimate */
2369 if (cfq_cfqq_slice_new(cfqq))
2370 return 1;
2371 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2372 cfqq->slice_end))
2373 return 1;
2374
2375 return 0;
2376}
2377
Jens Axboe0b182d62009-10-06 20:49:37 +02002378static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002379{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002380 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381
Jens Axboe2f5cb732009-04-07 08:51:19 +02002382 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002383 * Drain async requests before we start sync IO
2384 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002385 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002386 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002387
2388 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002389 * If this is an async queue and we have sync IO in flight, let it wait
2390 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002391 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002392 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002393
Shaohua Liabc3c742010-03-01 09:20:54 +01002394 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002395 if (cfq_class_idle(cfqq))
2396 max_dispatch = 1;
2397
2398 /*
2399 * Does this cfqq already have too much IO in flight?
2400 */
2401 if (cfqq->dispatched >= max_dispatch) {
2402 /*
2403 * idle queue must always only have a single IO in flight
2404 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002405 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002406 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002407
Jens Axboe2f5cb732009-04-07 08:51:19 +02002408 /*
2409 * We have other queues, don't allow more IO from this one
2410 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002411 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002412 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002413
Jens Axboe2f5cb732009-04-07 08:51:19 +02002414 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002415 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002416 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002417 if (cfqd->busy_queues == 1)
2418 max_dispatch = -1;
2419 else
2420 /*
2421 * Normally we start throttling cfqq when cfq_quantum/2
2422 * requests have been dispatched. But we can drive
2423 * deeper queue depths at the beginning of slice
2424 * subjected to upper limit of cfq_quantum.
2425 * */
2426 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002427 }
2428
2429 /*
2430 * Async queues must wait a bit before being allowed dispatch.
2431 * We also ramp up the dispatch depth gradually for async IO,
2432 * based on the last sync IO we serviced
2433 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002434 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002435 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002436 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002437
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002438 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002439 if (!depth && !cfqq->dispatched)
2440 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002441 if (depth < max_dispatch)
2442 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 }
2444
Jens Axboe0b182d62009-10-06 20:49:37 +02002445 /*
2446 * If we're below the current max, allow a dispatch
2447 */
2448 return cfqq->dispatched < max_dispatch;
2449}
2450
2451/*
2452 * Dispatch a request from cfqq, moving them to the request queue
2453 * dispatch list.
2454 */
2455static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2456{
2457 struct request *rq;
2458
2459 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2460
2461 if (!cfq_may_dispatch(cfqd, cfqq))
2462 return false;
2463
2464 /*
2465 * follow expired path, else get first next available
2466 */
2467 rq = cfq_check_fifo(cfqq);
2468 if (!rq)
2469 rq = cfqq->next_rq;
2470
2471 /*
2472 * insert request into driver dispatch list
2473 */
2474 cfq_dispatch_insert(cfqd->queue, rq);
2475
2476 if (!cfqd->active_cic) {
2477 struct cfq_io_context *cic = RQ_CIC(rq);
2478
2479 atomic_long_inc(&cic->ioc->refcount);
2480 cfqd->active_cic = cic;
2481 }
2482
2483 return true;
2484}
2485
2486/*
2487 * Find the cfqq that we need to service and move a request from that to the
2488 * dispatch list
2489 */
2490static int cfq_dispatch_requests(struct request_queue *q, int force)
2491{
2492 struct cfq_data *cfqd = q->elevator->elevator_data;
2493 struct cfq_queue *cfqq;
2494
2495 if (!cfqd->busy_queues)
2496 return 0;
2497
2498 if (unlikely(force))
2499 return cfq_forced_dispatch(cfqd);
2500
2501 cfqq = cfq_select_queue(cfqd);
2502 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002503 return 0;
2504
Jens Axboe2f5cb732009-04-07 08:51:19 +02002505 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002506 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002507 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002508 if (!cfq_dispatch_request(cfqd, cfqq))
2509 return 0;
2510
Jens Axboe2f5cb732009-04-07 08:51:19 +02002511 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002512 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002513
2514 /*
2515 * expire an async queue immediately if it has used up its slice. idle
2516 * queue always expire after 1 dispatch round.
2517 */
2518 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2519 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2520 cfq_class_idle(cfqq))) {
2521 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002522 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002523 }
2524
Shan Weib217a902009-09-01 10:06:42 +02002525 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002526 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527}
2528
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529/*
Jens Axboe5e7053742006-07-13 12:39:25 +02002530 * task holds one reference to the queue, dropped when task exits. each rq
2531 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002533 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 * queue lock must be held here.
2535 */
2536static void cfq_put_queue(struct cfq_queue *cfqq)
2537{
Jens Axboe22e2c502005-06-27 10:55:12 +02002538 struct cfq_data *cfqd = cfqq->cfqd;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002539 struct cfq_group *cfqg, *orig_cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002540
2541 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542
2543 if (!atomic_dec_and_test(&cfqq->ref))
2544 return;
2545
Jens Axboe7b679132008-05-30 12:23:07 +02002546 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002548 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002549 cfqg = cfqq->cfqg;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002550 orig_cfqg = cfqq->orig_cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002552 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002553 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002554 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002555 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002556
Vivek Goyalf04a6422009-12-03 12:59:40 -05002557 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002559 cfq_put_cfqg(cfqg);
Vivek Goyal878eadd2009-12-07 19:37:15 +01002560 if (orig_cfqg)
2561 cfq_put_cfqg(orig_cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562}
2563
Jens Axboed6de8be2008-05-28 14:46:59 +02002564/*
2565 * Must always be called with the rcu_read_lock() held
2566 */
Jens Axboe07416d22008-05-07 09:17:12 +02002567static void
2568__call_for_each_cic(struct io_context *ioc,
2569 void (*func)(struct io_context *, struct cfq_io_context *))
2570{
2571 struct cfq_io_context *cic;
2572 struct hlist_node *n;
2573
2574 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2575 func(ioc, cic);
2576}
2577
Jens Axboe4ac845a2008-01-24 08:44:49 +01002578/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002579 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002580 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002581static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002582call_for_each_cic(struct io_context *ioc,
2583 void (*func)(struct io_context *, struct cfq_io_context *))
2584{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002585 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002586 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002587 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002588}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002589
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002590static void cfq_cic_free_rcu(struct rcu_head *head)
2591{
2592 struct cfq_io_context *cic;
2593
2594 cic = container_of(head, struct cfq_io_context, rcu_head);
2595
2596 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002597 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002598
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002599 if (ioc_gone) {
2600 /*
2601 * CFQ scheduler is exiting, grab exit lock and check
2602 * the pending io context count. If it hits zero,
2603 * complete ioc_gone and set it back to NULL
2604 */
2605 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002606 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002607 complete(ioc_gone);
2608 ioc_gone = NULL;
2609 }
2610 spin_unlock(&ioc_gone_lock);
2611 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002612}
2613
2614static void cfq_cic_free(struct cfq_io_context *cic)
2615{
2616 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002617}
2618
2619static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2620{
2621 unsigned long flags;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002622 unsigned long dead_key = (unsigned long) cic->key;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002623
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002624 BUG_ON(!(dead_key & CIC_DEAD_KEY));
Jens Axboe4ac845a2008-01-24 08:44:49 +01002625
2626 spin_lock_irqsave(&ioc->lock, flags);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002627 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
Jens Axboeffc4e752008-02-19 10:02:29 +01002628 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002629 spin_unlock_irqrestore(&ioc->lock, flags);
2630
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002631 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002632}
2633
Jens Axboed6de8be2008-05-28 14:46:59 +02002634/*
2635 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2636 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2637 * and ->trim() which is called with the task lock held
2638 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002639static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002641 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002642 * ioc->refcount is zero here, or we are called from elv_unregister(),
2643 * so no more cic's are allowed to be linked into this ioc. So it
2644 * should be ok to iterate over the known list, we will see all cic's
2645 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002646 */
Jens Axboe07416d22008-05-07 09:17:12 +02002647 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648}
2649
Shaohua Lid02a2c02010-05-25 10:16:53 +02002650static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002651{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002652 struct cfq_queue *__cfqq, *next;
2653
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002654 /*
2655 * If this queue was scheduled to merge with another queue, be
2656 * sure to drop the reference taken on that queue (and others in
2657 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2658 */
2659 __cfqq = cfqq->new_cfqq;
2660 while (__cfqq) {
2661 if (__cfqq == cfqq) {
2662 WARN(1, "cfqq->new_cfqq loop detected\n");
2663 break;
2664 }
2665 next = __cfqq->new_cfqq;
2666 cfq_put_queue(__cfqq);
2667 __cfqq = next;
2668 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002669}
2670
2671static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2672{
2673 if (unlikely(cfqq == cfqd->active_queue)) {
2674 __cfq_slice_expired(cfqd, cfqq, 0);
2675 cfq_schedule_dispatch(cfqd);
2676 }
2677
2678 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002679
Jens Axboe89850f72006-07-22 16:48:31 +02002680 cfq_put_queue(cfqq);
2681}
2682
2683static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2684 struct cfq_io_context *cic)
2685{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002686 struct io_context *ioc = cic->ioc;
2687
Jens Axboefc463792006-08-29 09:05:44 +02002688 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002689
2690 /*
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002691 * Make sure dead mark is seen for dead queues
Jens Axboe4ac845a2008-01-24 08:44:49 +01002692 */
Jens Axboefc463792006-08-29 09:05:44 +02002693 smp_wmb();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002694 cic->key = cfqd_dead_key(cfqd);
Jens Axboefc463792006-08-29 09:05:44 +02002695
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002696 if (ioc->ioc_data == cic)
2697 rcu_assign_pointer(ioc->ioc_data, NULL);
2698
Jens Axboeff6657c2009-04-08 10:58:57 +02002699 if (cic->cfqq[BLK_RW_ASYNC]) {
2700 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2701 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002702 }
2703
Jens Axboeff6657c2009-04-08 10:58:57 +02002704 if (cic->cfqq[BLK_RW_SYNC]) {
2705 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2706 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002707 }
Jens Axboe89850f72006-07-22 16:48:31 +02002708}
2709
Jens Axboe4ac845a2008-01-24 08:44:49 +01002710static void cfq_exit_single_io_context(struct io_context *ioc,
2711 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002712{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002713 struct cfq_data *cfqd = cic_to_cfqd(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002714
Jens Axboe89850f72006-07-22 16:48:31 +02002715 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002716 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002717 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002718
Jens Axboe4ac845a2008-01-24 08:44:49 +01002719 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002720
2721 /*
2722 * Ensure we get a fresh copy of the ->key to prevent
2723 * race between exiting task and queue
2724 */
2725 smp_read_barrier_depends();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002726 if (cic->key == cfqd)
Jens Axboe62c1fe92008-12-15 21:19:25 +01002727 __cfq_exit_single_io_context(cfqd, cic);
2728
Jens Axboe4ac845a2008-01-24 08:44:49 +01002729 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002730 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002731}
2732
Jens Axboe498d3aa22007-04-26 12:54:48 +02002733/*
2734 * The process that ioc belongs to has exited, we need to clean up
2735 * and put the internal structures we have that belongs to that process.
2736 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002737static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002739 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740}
2741
Jens Axboe22e2c502005-06-27 10:55:12 +02002742static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002743cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744{
Jens Axboeb5deef92006-07-19 23:39:40 +02002745 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746
Christoph Lameter94f60302007-07-17 04:03:29 -07002747 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2748 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002750 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002751 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002752 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002753 cic->dtor = cfq_free_io_context;
2754 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002755 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 }
2757
2758 return cic;
2759}
2760
Jens Axboefd0928d2008-01-24 08:52:45 +01002761static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002762{
2763 struct task_struct *tsk = current;
2764 int ioprio_class;
2765
Jens Axboe3b181522005-06-27 10:56:24 +02002766 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002767 return;
2768
Jens Axboefd0928d2008-01-24 08:52:45 +01002769 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002770 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002771 default:
2772 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2773 case IOPRIO_CLASS_NONE:
2774 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002775 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002776 */
2777 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002778 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002779 break;
2780 case IOPRIO_CLASS_RT:
2781 cfqq->ioprio = task_ioprio(ioc);
2782 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2783 break;
2784 case IOPRIO_CLASS_BE:
2785 cfqq->ioprio = task_ioprio(ioc);
2786 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2787 break;
2788 case IOPRIO_CLASS_IDLE:
2789 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2790 cfqq->ioprio = 7;
2791 cfq_clear_cfqq_idle_window(cfqq);
2792 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002793 }
2794
2795 /*
2796 * keep track of original prio settings in case we have to temporarily
2797 * elevate the priority of this queue
2798 */
2799 cfqq->org_ioprio = cfqq->ioprio;
2800 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002801 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002802}
2803
Jens Axboefebffd62008-01-28 13:19:43 +01002804static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002805{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002806 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002807 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002808 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002809
Jens Axboecaaa5f92006-06-16 11:23:00 +02002810 if (unlikely(!cfqd))
2811 return;
2812
Jens Axboec1b707d2006-10-30 19:54:23 +01002813 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002814
Jens Axboeff6657c2009-04-08 10:58:57 +02002815 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002816 if (cfqq) {
2817 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002818 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2819 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002820 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002821 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002822 cfq_put_queue(cfqq);
2823 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002824 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002825
Jens Axboeff6657c2009-04-08 10:58:57 +02002826 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002827 if (cfqq)
2828 cfq_mark_cfqq_prio_changed(cfqq);
2829
Jens Axboec1b707d2006-10-30 19:54:23 +01002830 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002831}
2832
Jens Axboefc463792006-08-29 09:05:44 +02002833static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002835 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002836 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837}
2838
Jens Axboed5036d72009-06-26 10:44:34 +02002839static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002840 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002841{
2842 RB_CLEAR_NODE(&cfqq->rb_node);
2843 RB_CLEAR_NODE(&cfqq->p_node);
2844 INIT_LIST_HEAD(&cfqq->fifo);
2845
2846 atomic_set(&cfqq->ref, 0);
2847 cfqq->cfqd = cfqd;
2848
2849 cfq_mark_cfqq_prio_changed(cfqq);
2850
2851 if (is_sync) {
2852 if (!cfq_class_idle(cfqq))
2853 cfq_mark_cfqq_idle_window(cfqq);
2854 cfq_mark_cfqq_sync(cfqq);
2855 }
2856 cfqq->pid = pid;
2857}
2858
Vivek Goyal246103332009-12-03 12:59:51 -05002859#ifdef CONFIG_CFQ_GROUP_IOSCHED
2860static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2861{
2862 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002863 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal246103332009-12-03 12:59:51 -05002864 unsigned long flags;
2865 struct request_queue *q;
2866
2867 if (unlikely(!cfqd))
2868 return;
2869
2870 q = cfqd->queue;
2871
2872 spin_lock_irqsave(q->queue_lock, flags);
2873
2874 if (sync_cfqq) {
2875 /*
2876 * Drop reference to sync queue. A new sync queue will be
2877 * assigned in new group upon arrival of a fresh request.
2878 */
2879 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2880 cic_set_cfqq(cic, NULL, 1);
2881 cfq_put_queue(sync_cfqq);
2882 }
2883
2884 spin_unlock_irqrestore(q->queue_lock, flags);
2885}
2886
2887static void cfq_ioc_set_cgroup(struct io_context *ioc)
2888{
2889 call_for_each_cic(ioc, changed_cgroup);
2890 ioc->cgroup_changed = 0;
2891}
2892#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2893
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002895cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002896 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002899 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002900 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901
2902retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002903 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002904 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002905 /* cic always exists here */
2906 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907
Jens Axboe6118b702009-06-30 09:34:12 +02002908 /*
2909 * Always try a new alloc if we fell back to the OOM cfqq
2910 * originally, since it should just be a temporary situation.
2911 */
2912 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2913 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914 if (new_cfqq) {
2915 cfqq = new_cfqq;
2916 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002917 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002919 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002920 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002921 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002923 if (new_cfqq)
2924 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002925 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002926 cfqq = kmem_cache_alloc_node(cfq_pool,
2927 gfp_mask | __GFP_ZERO,
2928 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002929 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930
Jens Axboe6118b702009-06-30 09:34:12 +02002931 if (cfqq) {
2932 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2933 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002934 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002935 cfq_log_cfqq(cfqd, cfqq, "alloced");
2936 } else
2937 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 }
2939
2940 if (new_cfqq)
2941 kmem_cache_free(cfq_pool, new_cfqq);
2942
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 return cfqq;
2944}
2945
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002946static struct cfq_queue **
2947cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2948{
Jens Axboefe094d92008-01-31 13:08:54 +01002949 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002950 case IOPRIO_CLASS_RT:
2951 return &cfqd->async_cfqq[0][ioprio];
2952 case IOPRIO_CLASS_BE:
2953 return &cfqd->async_cfqq[1][ioprio];
2954 case IOPRIO_CLASS_IDLE:
2955 return &cfqd->async_idle_cfqq;
2956 default:
2957 BUG();
2958 }
2959}
2960
Jens Axboe15c31be2007-07-10 13:43:25 +02002961static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002962cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002963 gfp_t gfp_mask)
2964{
Jens Axboefd0928d2008-01-24 08:52:45 +01002965 const int ioprio = task_ioprio(ioc);
2966 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002967 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002968 struct cfq_queue *cfqq = NULL;
2969
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002970 if (!is_sync) {
2971 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2972 cfqq = *async_cfqq;
2973 }
2974
Jens Axboe6118b702009-06-30 09:34:12 +02002975 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002976 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002977
2978 /*
2979 * pin the queue now that it's allocated, scheduler exit will prune it
2980 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002981 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02002982 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002983 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002984 }
2985
2986 atomic_inc(&cfqq->ref);
2987 return cfqq;
2988}
2989
Jens Axboe498d3aa22007-04-26 12:54:48 +02002990/*
2991 * We drop cfq io contexts lazily, so we may find a dead one.
2992 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002993static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002994cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2995 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002996{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002997 unsigned long flags;
2998
Jens Axboefc463792006-08-29 09:05:44 +02002999 WARN_ON(!list_empty(&cic->queue_list));
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003000 BUG_ON(cic->key != cfqd_dead_key(cfqd));
Jens Axboe597bc482007-04-24 21:23:53 +02003001
Jens Axboe4ac845a2008-01-24 08:44:49 +01003002 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02003003
Fabio Checconi4faa3c82008-04-10 08:28:01 +02003004 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003005
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003006 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
Jens Axboeffc4e752008-02-19 10:02:29 +01003007 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003008 spin_unlock_irqrestore(&ioc->lock, flags);
3009
3010 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003011}
3012
Jens Axboee2d74ac2006-03-28 08:59:01 +02003013static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01003014cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003015{
Jens Axboee2d74ac2006-03-28 08:59:01 +02003016 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003017 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003018
Vasily Tarasov91fac312007-04-25 12:29:51 +02003019 if (unlikely(!ioc))
3020 return NULL;
3021
Jens Axboed6de8be2008-05-28 14:46:59 +02003022 rcu_read_lock();
3023
Jens Axboe597bc482007-04-24 21:23:53 +02003024 /*
3025 * we maintain a last-hit cache, to avoid browsing over the tree
3026 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003027 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02003028 if (cic && cic->key == cfqd) {
3029 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02003030 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003031 }
Jens Axboe597bc482007-04-24 21:23:53 +02003032
Jens Axboe4ac845a2008-01-24 08:44:49 +01003033 do {
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003034 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003035 rcu_read_unlock();
3036 if (!cic)
3037 break;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003038 if (unlikely(cic->key != cfqd)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01003039 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003040 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01003041 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003042 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003043
Jens Axboed6de8be2008-05-28 14:46:59 +02003044 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003045 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003046 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003047 break;
3048 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003049
Jens Axboe4ac845a2008-01-24 08:44:49 +01003050 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003051}
3052
Jens Axboe4ac845a2008-01-24 08:44:49 +01003053/*
3054 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3055 * the process specific cfq io context when entered from the block layer.
3056 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3057 */
Jens Axboefebffd62008-01-28 13:19:43 +01003058static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3059 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003060{
Jens Axboe0261d682006-10-30 19:07:48 +01003061 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003062 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003063
Jens Axboe4ac845a2008-01-24 08:44:49 +01003064 ret = radix_tree_preload(gfp_mask);
3065 if (!ret) {
3066 cic->ioc = ioc;
3067 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003068
Jens Axboe4ac845a2008-01-24 08:44:49 +01003069 spin_lock_irqsave(&ioc->lock, flags);
3070 ret = radix_tree_insert(&ioc->radix_root,
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003071 cfqd->cic_index, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01003072 if (!ret)
3073 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003074 spin_unlock_irqrestore(&ioc->lock, flags);
3075
3076 radix_tree_preload_end();
3077
3078 if (!ret) {
3079 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3080 list_add(&cic->queue_list, &cfqd->cic_list);
3081 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003082 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003083 }
3084
Jens Axboe4ac845a2008-01-24 08:44:49 +01003085 if (ret)
3086 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02003087
Jens Axboe4ac845a2008-01-24 08:44:49 +01003088 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003089}
3090
Jens Axboe22e2c502005-06-27 10:55:12 +02003091/*
3092 * Setup general io context and cfq io context. There can be several cfq
3093 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02003094 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02003095 */
3096static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02003097cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098{
Jens Axboe22e2c502005-06-27 10:55:12 +02003099 struct io_context *ioc = NULL;
3100 struct cfq_io_context *cic;
3101
3102 might_sleep_if(gfp_mask & __GFP_WAIT);
3103
Jens Axboeb5deef92006-07-19 23:39:40 +02003104 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02003105 if (!ioc)
3106 return NULL;
3107
Jens Axboe4ac845a2008-01-24 08:44:49 +01003108 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003109 if (cic)
3110 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02003111
Jens Axboee2d74ac2006-03-28 08:59:01 +02003112 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3113 if (cic == NULL)
3114 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02003115
Jens Axboe4ac845a2008-01-24 08:44:49 +01003116 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3117 goto err_free;
3118
Jens Axboe22e2c502005-06-27 10:55:12 +02003119out:
Jens Axboefc463792006-08-29 09:05:44 +02003120 smp_read_barrier_depends();
3121 if (unlikely(ioc->ioprio_changed))
3122 cfq_ioc_set_ioprio(ioc);
3123
Vivek Goyal246103332009-12-03 12:59:51 -05003124#ifdef CONFIG_CFQ_GROUP_IOSCHED
3125 if (unlikely(ioc->cgroup_changed))
3126 cfq_ioc_set_cgroup(ioc);
3127#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003128 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003129err_free:
3130 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003131err:
3132 put_io_context(ioc);
3133 return NULL;
3134}
3135
3136static void
3137cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3138{
Jens Axboeaaf12282007-01-19 11:30:16 +11003139 unsigned long elapsed = jiffies - cic->last_end_request;
3140 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003141
3142 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3143 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3144 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3145}
3146
Jens Axboe206dc692006-03-28 13:03:44 +02003147static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003148cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003149 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003150{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003151 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003152 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003153 if (cfqq->last_request_pos) {
3154 if (cfqq->last_request_pos < blk_rq_pos(rq))
3155 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3156 else
3157 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3158 }
Jens Axboe206dc692006-03-28 13:03:44 +02003159
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003160 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003161 if (blk_queue_nonrot(cfqd->queue))
3162 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3163 else
3164 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003165}
Jens Axboe22e2c502005-06-27 10:55:12 +02003166
3167/*
3168 * Disable idle window if the process thinks too long or seeks so much that
3169 * it doesn't matter
3170 */
3171static void
3172cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3173 struct cfq_io_context *cic)
3174{
Jens Axboe7b679132008-05-30 12:23:07 +02003175 int old_idle, enable_idle;
Jens Axboe1be92f2f2007-04-19 14:32:26 +02003176
Jens Axboe08717142008-01-28 11:38:15 +01003177 /*
3178 * Don't idle for async or idle io prio class
3179 */
3180 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f2f2007-04-19 14:32:26 +02003181 return;
3182
Jens Axboec265a7f2008-06-26 13:49:33 +02003183 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003184
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003185 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3186 cfq_mark_cfqq_deep(cfqq);
3187
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003188 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3189 enable_idle = 0;
3190 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003191 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003192 enable_idle = 0;
3193 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003194 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003195 enable_idle = 0;
3196 else
3197 enable_idle = 1;
3198 }
3199
Jens Axboe7b679132008-05-30 12:23:07 +02003200 if (old_idle != enable_idle) {
3201 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3202 if (enable_idle)
3203 cfq_mark_cfqq_idle_window(cfqq);
3204 else
3205 cfq_clear_cfqq_idle_window(cfqq);
3206 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003207}
3208
Jens Axboe22e2c502005-06-27 10:55:12 +02003209/*
3210 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3211 * no or if we aren't sure, a 1 will cause a preempt.
3212 */
Jens Axboea6151c32009-10-07 20:02:57 +02003213static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003214cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e7053742006-07-13 12:39:25 +02003215 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003216{
Jens Axboe6d048f52007-04-25 12:44:27 +02003217 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003218
Jens Axboe6d048f52007-04-25 12:44:27 +02003219 cfqq = cfqd->active_queue;
3220 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003221 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003222
Jens Axboe6d048f52007-04-25 12:44:27 +02003223 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003224 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003225
3226 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003227 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003228
Jens Axboe22e2c502005-06-27 10:55:12 +02003229 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003230 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3231 */
3232 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3233 return false;
3234
3235 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003236 * if the new request is sync, but the currently running queue is
3237 * not, let the sync request have priority.
3238 */
Jens Axboe5e7053742006-07-13 12:39:25 +02003239 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003240 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003241
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003242 if (new_cfqq->cfqg != cfqq->cfqg)
3243 return false;
3244
3245 if (cfq_slice_used(cfqq))
3246 return true;
3247
3248 /* Allow preemption only if we are idling on sync-noidle tree */
3249 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3250 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3251 new_cfqq->service_tree->count == 2 &&
3252 RB_EMPTY_ROOT(&cfqq->sort_list))
3253 return true;
3254
Jens Axboe374f84a2006-07-23 01:42:19 +02003255 /*
3256 * So both queues are sync. Let the new request get disk time if
3257 * it's a metadata request and the current queue is doing regular IO.
3258 */
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003259 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003260 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003261
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003262 /*
3263 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3264 */
3265 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003266 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003267
Jens Axboe1e3335d2007-02-14 19:59:49 +01003268 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003269 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003270
3271 /*
3272 * if this request is as-good as one we would expect from the
3273 * current cfqq, let it preempt
3274 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003275 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003276 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003277
Jens Axboea6151c32009-10-07 20:02:57 +02003278 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003279}
3280
3281/*
3282 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3283 * let it have half of its nominal slice.
3284 */
3285static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3286{
Jens Axboe7b679132008-05-30 12:23:07 +02003287 cfq_log_cfqq(cfqd, cfqq, "preempt");
Vivek Goyale5ff0822010-04-26 19:25:11 +02003288 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003289
Jens Axboebf572252006-07-19 20:29:12 +02003290 /*
3291 * Put the new queue at the front of the of the current list,
3292 * so we know that it will be selected next.
3293 */
3294 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003295
3296 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02003297
Jens Axboe44f7c162007-01-19 11:51:58 +11003298 cfqq->slice_end = 0;
3299 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003300}
3301
3302/*
Jens Axboe5e7053742006-07-13 12:39:25 +02003303 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003304 * something we should do about it
3305 */
3306static void
Jens Axboe5e7053742006-07-13 12:39:25 +02003307cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3308 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003309{
Jens Axboe5e7053742006-07-13 12:39:25 +02003310 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003311
Aaron Carroll45333d52008-08-26 15:52:36 +02003312 cfqd->rq_queued++;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003313 if (rq->cmd_flags & REQ_META)
Jens Axboe374f84a2006-07-23 01:42:19 +02003314 cfqq->meta_pending++;
3315
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003316 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003317 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003318 cfq_update_idle_window(cfqd, cfqq, cic);
3319
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003320 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003321
3322 if (cfqq == cfqd->active_queue) {
3323 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003324 * Remember that we saw a request from this process, but
3325 * don't start queuing just yet. Otherwise we risk seeing lots
3326 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003327 * and merging. If the request is already larger than a single
3328 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003329 * merging is already done. Ditto for a busy system that
3330 * has other work pending, don't risk delaying until the
3331 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003332 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003333 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003334 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3335 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003336 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003337 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003338 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003339 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003340 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003341 &cfqq->cfqg->blkg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003342 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003343 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003344 }
Jens Axboe5e7053742006-07-13 12:39:25 +02003345 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003346 /*
3347 * not the active queue - expire current slice if it is
3348 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003349 * has some old slice time left and is of higher priority or
3350 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003351 */
3352 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09003353 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003354 }
3355}
3356
Jens Axboe165125e2007-07-24 09:28:11 +02003357static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003358{
Jens Axboeb4878f22005-10-20 16:42:29 +02003359 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e7053742006-07-13 12:39:25 +02003360 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003361
Jens Axboe7b679132008-05-30 12:23:07 +02003362 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003363 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364
Jens Axboe30996f42009-10-05 11:03:39 +02003365 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003366 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003367 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003368 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003369 &cfqd->serving_group->blkg, rq_data_dir(rq),
3370 rq_is_sync(rq));
Jens Axboe5e7053742006-07-13 12:39:25 +02003371 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003372}
3373
Aaron Carroll45333d52008-08-26 15:52:36 +02003374/*
3375 * Update hw_tag based on peak queue depth over 50 samples under
3376 * sufficient load.
3377 */
3378static void cfq_update_hw_tag(struct cfq_data *cfqd)
3379{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003380 struct cfq_queue *cfqq = cfqd->active_queue;
3381
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003382 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3383 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003384
3385 if (cfqd->hw_tag == 1)
3386 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003387
3388 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003389 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003390 return;
3391
Shaohua Li1a1238a2009-10-27 08:46:23 +01003392 /*
3393 * If active queue hasn't enough requests and can idle, cfq might not
3394 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3395 * case
3396 */
3397 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3398 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003399 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003400 return;
3401
Aaron Carroll45333d52008-08-26 15:52:36 +02003402 if (cfqd->hw_tag_samples++ < 50)
3403 return;
3404
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003405 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003406 cfqd->hw_tag = 1;
3407 else
3408 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003409}
3410
Vivek Goyal7667aa02009-12-08 17:52:58 -05003411static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3412{
3413 struct cfq_io_context *cic = cfqd->active_cic;
3414
3415 /* If there are other queues in the group, don't wait */
3416 if (cfqq->cfqg->nr_cfqq > 1)
3417 return false;
3418
3419 if (cfq_slice_used(cfqq))
3420 return true;
3421
3422 /* if slice left is less than think time, wait busy */
3423 if (cic && sample_valid(cic->ttime_samples)
3424 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3425 return true;
3426
3427 /*
3428 * If think times is less than a jiffy than ttime_mean=0 and above
3429 * will not be true. It might happen that slice has not expired yet
3430 * but will expire soon (4-5 ns) during select_queue(). To cover the
3431 * case where think time is less than a jiffy, mark the queue wait
3432 * busy if only 1 jiffy is left in the slice.
3433 */
3434 if (cfqq->slice_end - jiffies == 1)
3435 return true;
3436
3437 return false;
3438}
3439
Jens Axboe165125e2007-07-24 09:28:11 +02003440static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441{
Jens Axboe5e7053742006-07-13 12:39:25 +02003442 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003443 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003444 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003445 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446
Jens Axboeb4878f22005-10-20 16:42:29 +02003447 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003448 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3449 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450
Aaron Carroll45333d52008-08-26 15:52:36 +02003451 cfq_update_hw_tag(cfqd);
3452
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003453 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003454 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003455 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003456 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003457 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003458 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3459 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3460 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003462 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003463
Vivek Goyal365722b2009-10-03 15:21:27 +02003464 if (sync) {
Jens Axboe5e7053742006-07-13 12:39:25 +02003465 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003466 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3467 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003468 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003469
3470 /*
3471 * If this is the active queue, check if it needs to be expired,
3472 * or if we want to idle in case it has no pending requests.
3473 */
3474 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003475 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3476
Jens Axboe44f7c162007-01-19 11:51:58 +11003477 if (cfq_cfqq_slice_new(cfqq)) {
3478 cfq_set_prio_slice(cfqd, cfqq);
3479 cfq_clear_cfqq_slice_new(cfqq);
3480 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003481
3482 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003483 * Should we wait for next request to come in before we expire
3484 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003485 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003486 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003487 unsigned long extend_sl = cfqd->cfq_slice_idle;
3488 if (!cfqd->cfq_slice_idle)
3489 extend_sl = cfqd->cfq_group_idle;
3490 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003491 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003492 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003493 }
3494
Jens Axboea36e71f2009-04-15 12:15:11 +02003495 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003496 * Idling is not enabled on:
3497 * - expired queues
3498 * - idle-priority queues
3499 * - async queues
3500 * - queues with still some requests queued
3501 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003502 */
Jens Axboe08717142008-01-28 11:38:15 +01003503 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003504 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003505 else if (sync && cfqq_empty &&
3506 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003507 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003508 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003509 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003510
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003511 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003512 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513}
3514
Jens Axboe22e2c502005-06-27 10:55:12 +02003515/*
3516 * we temporarily boost lower priority queues if they are holding fs exclusive
3517 * resources. they are boosted to normal prio (CLASS_BE/4)
3518 */
3519static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520{
Jens Axboe22e2c502005-06-27 10:55:12 +02003521 if (has_fs_excl()) {
3522 /*
3523 * boost idle prio on transactions that would lock out other
3524 * users of the filesystem
3525 */
3526 if (cfq_class_idle(cfqq))
3527 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3528 if (cfqq->ioprio > IOPRIO_NORM)
3529 cfqq->ioprio = IOPRIO_NORM;
3530 } else {
3531 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003532 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003533 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003534 cfqq->ioprio_class = cfqq->org_ioprio_class;
3535 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003537}
3538
Jens Axboe89850f72006-07-22 16:48:31 +02003539static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003540{
Jens Axboe1b379d82009-08-11 08:26:11 +02003541 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003542 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003543 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003544 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003545
3546 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003547}
3548
Jens Axboe165125e2007-07-24 09:28:11 +02003549static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003550{
3551 struct cfq_data *cfqd = q->elevator->elevator_data;
3552 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003553 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003554 struct cfq_queue *cfqq;
3555
3556 /*
3557 * don't force setup of a queue from here, as a call to may_queue
3558 * does not necessarily imply that a request actually will be queued.
3559 * so just lookup a possibly existing queue, or return 'may queue'
3560 * if that fails
3561 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003562 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003563 if (!cic)
3564 return ELV_MQUEUE_MAY;
3565
Jens Axboeb0b78f82009-04-08 10:56:08 +02003566 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003567 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003568 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003569 cfq_prio_boost(cfqq);
3570
Jens Axboe89850f72006-07-22 16:48:31 +02003571 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003572 }
3573
3574 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575}
3576
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577/*
3578 * queue lock held here
3579 */
Jens Axboebb37b942006-12-01 10:42:33 +01003580static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581{
Jens Axboe5e7053742006-07-13 12:39:25 +02003582 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583
Jens Axboe5e7053742006-07-13 12:39:25 +02003584 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003585 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586
Jens Axboe22e2c502005-06-27 10:55:12 +02003587 BUG_ON(!cfqq->allocated[rw]);
3588 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589
Jens Axboe5e7053742006-07-13 12:39:25 +02003590 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592 rq->elevator_private = NULL;
Jens Axboe5e7053742006-07-13 12:39:25 +02003593 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003595 /* Put down rq reference on cfqg */
3596 cfq_put_cfqg(RQ_CFQG(rq));
3597 rq->elevator_private3 = NULL;
3598
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599 cfq_put_queue(cfqq);
3600 }
3601}
3602
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003603static struct cfq_queue *
3604cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3605 struct cfq_queue *cfqq)
3606{
3607 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3608 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003609 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003610 cfq_put_queue(cfqq);
3611 return cic_to_cfqq(cic, 1);
3612}
3613
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003614/*
3615 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3616 * was the last process referring to said cfqq.
3617 */
3618static struct cfq_queue *
3619split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3620{
3621 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003622 cfqq->pid = current->pid;
3623 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003624 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003625 return cfqq;
3626 }
3627
3628 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003629
3630 cfq_put_cooperator(cfqq);
3631
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003632 cfq_put_queue(cfqq);
3633 return NULL;
3634}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003636 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003637 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003638static int
Jens Axboe165125e2007-07-24 09:28:11 +02003639cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003640{
3641 struct cfq_data *cfqd = q->elevator->elevator_data;
3642 struct cfq_io_context *cic;
3643 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003644 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003645 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646 unsigned long flags;
3647
3648 might_sleep_if(gfp_mask & __GFP_WAIT);
3649
Jens Axboee2d74ac2006-03-28 08:59:01 +02003650 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003651
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652 spin_lock_irqsave(q->queue_lock, flags);
3653
Jens Axboe22e2c502005-06-27 10:55:12 +02003654 if (!cic)
3655 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003657new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003658 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003659 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003660 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003661 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003662 } else {
3663 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003664 * If the queue was seeky for too long, break it apart.
3665 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003666 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003667 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3668 cfqq = split_cfqq(cic, cfqq);
3669 if (!cfqq)
3670 goto new_queue;
3671 }
3672
3673 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003674 * Check to see if this queue is scheduled to merge with
3675 * another, closely cooperating queue. The merging of
3676 * queues happens here as it must be done in process context.
3677 * The reference on new_cfqq was taken in merge_cfqqs.
3678 */
3679 if (cfqq->new_cfqq)
3680 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003681 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682
3683 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02003684 atomic_inc(&cfqq->ref);
Jens Axboe5e7053742006-07-13 12:39:25 +02003685
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686 spin_unlock_irqrestore(q->queue_lock, flags);
3687
Jens Axboe5e7053742006-07-13 12:39:25 +02003688 rq->elevator_private = cic;
3689 rq->elevator_private2 = cfqq;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003690 rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
Jens Axboe5e7053742006-07-13 12:39:25 +02003691 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003692
Jens Axboe22e2c502005-06-27 10:55:12 +02003693queue_fail:
3694 if (cic)
3695 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003696
Jens Axboe23e018a2009-10-05 08:52:35 +02003697 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003699 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700 return 1;
3701}
3702
David Howells65f27f32006-11-22 14:55:48 +00003703static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003704{
David Howells65f27f32006-11-22 14:55:48 +00003705 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003706 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003707 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003708
Jens Axboe40bb54d2009-04-15 12:11:10 +02003709 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09003710 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003711 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003712}
3713
3714/*
3715 * Timer running if the active_queue is currently idling inside its time slice
3716 */
3717static void cfq_idle_slice_timer(unsigned long data)
3718{
3719 struct cfq_data *cfqd = (struct cfq_data *) data;
3720 struct cfq_queue *cfqq;
3721 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003722 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003723
Jens Axboe7b679132008-05-30 12:23:07 +02003724 cfq_log(cfqd, "idle timer fired");
3725
Jens Axboe22e2c502005-06-27 10:55:12 +02003726 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3727
Jens Axboefe094d92008-01-31 13:08:54 +01003728 cfqq = cfqd->active_queue;
3729 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003730 timed_out = 0;
3731
Jens Axboe22e2c502005-06-27 10:55:12 +02003732 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003733 * We saw a request before the queue expired, let it through
3734 */
3735 if (cfq_cfqq_must_dispatch(cfqq))
3736 goto out_kick;
3737
3738 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003739 * expired
3740 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003741 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003742 goto expire;
3743
3744 /*
3745 * only expire and reinvoke request handler, if there are
3746 * other queues with pending requests
3747 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003748 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003749 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003750
3751 /*
3752 * not expired and it has a request pending, let it dispatch
3753 */
Jens Axboe75e50982009-04-07 08:56:14 +02003754 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003755 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003756
3757 /*
3758 * Queue depth flag is reset only when the idle didn't succeed
3759 */
3760 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003761 }
3762expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003763 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003764out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003765 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003766out_cont:
3767 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3768}
3769
Jens Axboe3b181522005-06-27 10:56:24 +02003770static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3771{
3772 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003773 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003774}
Jens Axboe22e2c502005-06-27 10:55:12 +02003775
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003776static void cfq_put_async_queues(struct cfq_data *cfqd)
3777{
3778 int i;
3779
3780 for (i = 0; i < IOPRIO_BE_NR; i++) {
3781 if (cfqd->async_cfqq[0][i])
3782 cfq_put_queue(cfqd->async_cfqq[0][i]);
3783 if (cfqd->async_cfqq[1][i])
3784 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003785 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003786
3787 if (cfqd->async_idle_cfqq)
3788 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003789}
3790
Jens Axboebb729bc2009-12-06 09:54:19 +01003791static void cfq_cfqd_free(struct rcu_head *head)
3792{
3793 kfree(container_of(head, struct cfq_data, rcu));
3794}
3795
Jens Axboeb374d182008-10-31 10:05:07 +01003796static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797{
Jens Axboe22e2c502005-06-27 10:55:12 +02003798 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003799 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003800
Jens Axboe3b181522005-06-27 10:56:24 +02003801 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003802
Al Virod9ff4182006-03-18 13:51:22 -05003803 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003804
Al Virod9ff4182006-03-18 13:51:22 -05003805 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003806 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003807
3808 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003809 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3810 struct cfq_io_context,
3811 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003812
3813 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003814 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003815
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003816 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003817 cfq_release_cfq_groups(cfqd);
Vivek Goyale98ef892010-06-18 10:39:47 -04003818 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003819
Al Virod9ff4182006-03-18 13:51:22 -05003820 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003821
3822 cfq_shutdown_timer_wq(cfqd);
3823
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003824 spin_lock(&cic_index_lock);
3825 ida_remove(&cic_index_ida, cfqd->cic_index);
3826 spin_unlock(&cic_index_lock);
3827
Vivek Goyalb1c35762009-12-03 12:59:47 -05003828 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003829 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830}
3831
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003832static int cfq_alloc_cic_index(void)
3833{
3834 int index, error;
3835
3836 do {
3837 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3838 return -ENOMEM;
3839
3840 spin_lock(&cic_index_lock);
3841 error = ida_get_new(&cic_index_ida, &index);
3842 spin_unlock(&cic_index_lock);
3843 if (error && error != -EAGAIN)
3844 return error;
3845 } while (error);
3846
3847 return index;
3848}
3849
Jens Axboe165125e2007-07-24 09:28:11 +02003850static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851{
3852 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003853 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003854 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003855 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003857 i = cfq_alloc_cic_index();
3858 if (i < 0)
3859 return NULL;
3860
Christoph Lameter94f60302007-07-17 04:03:29 -07003861 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003863 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003865 cfqd->cic_index = i;
3866
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003867 /* Init root service tree */
3868 cfqd->grp_service_tree = CFQ_RB_ROOT;
3869
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003870 /* Init root group */
3871 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003872 for_each_cfqg_st(cfqg, i, j, st)
3873 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003874 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003875
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003876 /* Give preference to root group over other groups */
3877 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3878
Vivek Goyal25fb5162009-12-03 12:59:46 -05003879#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003880 /*
3881 * Take a reference to root group which we never drop. This is just
3882 * to make sure that cfq_put_cfqg() does not try to kfree root group
3883 */
3884 atomic_set(&cfqg->ref, 1);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003885 rcu_read_lock();
Vivek Goyale98ef892010-06-18 10:39:47 -04003886 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3887 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003888 rcu_read_unlock();
Vivek Goyal25fb5162009-12-03 12:59:46 -05003889#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003890 /*
3891 * Not strictly needed (since RB_ROOT just clears the node and we
3892 * zeroed cfqd on alloc), but better be safe in case someone decides
3893 * to add magic to the rb code
3894 */
3895 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3896 cfqd->prio_trees[i] = RB_ROOT;
3897
Jens Axboe6118b702009-06-30 09:34:12 +02003898 /*
3899 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3900 * Grab a permanent reference to it, so that the normal code flow
3901 * will not attempt to free it.
3902 */
3903 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3904 atomic_inc(&cfqd->oom_cfqq.ref);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003905 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003906
Al Virod9ff4182006-03-18 13:51:22 -05003907 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910
Jens Axboe22e2c502005-06-27 10:55:12 +02003911 init_timer(&cfqd->idle_slice_timer);
3912 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3913 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3914
Jens Axboe23e018a2009-10-05 08:52:35 +02003915 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003916
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003918 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3919 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920 cfqd->cfq_back_max = cfq_back_max;
3921 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003922 cfqd->cfq_slice[0] = cfq_slice_async;
3923 cfqd->cfq_slice[1] = cfq_slice_sync;
3924 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3925 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003926 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003927 cfqd->cfq_latency = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05003928 cfqd->cfq_group_isolation = 0;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003929 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003930 /*
3931 * we optimistically start assuming sync ops weren't delayed in last
3932 * second, in order to have larger depth for async operations.
3933 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003934 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003935 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936}
3937
3938static void cfq_slab_kill(void)
3939{
Jens Axboed6de8be2008-05-28 14:46:59 +02003940 /*
3941 * Caller already ensured that pending RCU callbacks are completed,
3942 * so we should have no busy allocations at this point.
3943 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 if (cfq_pool)
3945 kmem_cache_destroy(cfq_pool);
3946 if (cfq_ioc_pool)
3947 kmem_cache_destroy(cfq_ioc_pool);
3948}
3949
3950static int __init cfq_slab_setup(void)
3951{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003952 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953 if (!cfq_pool)
3954 goto fail;
3955
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003956 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 if (!cfq_ioc_pool)
3958 goto fail;
3959
3960 return 0;
3961fail:
3962 cfq_slab_kill();
3963 return -ENOMEM;
3964}
3965
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966/*
3967 * sysfs parts below -->
3968 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969static ssize_t
3970cfq_var_show(unsigned int var, char *page)
3971{
3972 return sprintf(page, "%d\n", var);
3973}
3974
3975static ssize_t
3976cfq_var_store(unsigned int *var, const char *page, size_t count)
3977{
3978 char *p = (char *) page;
3979
3980 *var = simple_strtoul(p, &p, 10);
3981 return count;
3982}
3983
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003985static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003987 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988 unsigned int __data = __VAR; \
3989 if (__CONV) \
3990 __data = jiffies_to_msecs(__data); \
3991 return cfq_var_show(__data, (page)); \
3992}
3993SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003994SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3995SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003996SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3997SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003998SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003999SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004000SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4001SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4002SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004003SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05004004SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005#undef SHOW_FUNCTION
4006
4007#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004008static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004010 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011 unsigned int __data; \
4012 int ret = cfq_var_store(&__data, (page), count); \
4013 if (__data < (MIN)) \
4014 __data = (MIN); \
4015 else if (__data > (MAX)) \
4016 __data = (MAX); \
4017 if (__CONV) \
4018 *(__PTR) = msecs_to_jiffies(__data); \
4019 else \
4020 *(__PTR) = __data; \
4021 return ret; \
4022}
4023STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004024STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4025 UINT_MAX, 1);
4026STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4027 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004028STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004029STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4030 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004031STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004032STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004033STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4034STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004035STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4036 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004037STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05004038STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039#undef STORE_FUNCTION
4040
Al Viroe572ec72006-03-18 22:27:18 -05004041#define CFQ_ATTR(name) \
4042 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004043
Al Viroe572ec72006-03-18 22:27:18 -05004044static struct elv_fs_entry cfq_attrs[] = {
4045 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004046 CFQ_ATTR(fifo_expire_sync),
4047 CFQ_ATTR(fifo_expire_async),
4048 CFQ_ATTR(back_seek_max),
4049 CFQ_ATTR(back_seek_penalty),
4050 CFQ_ATTR(slice_sync),
4051 CFQ_ATTR(slice_async),
4052 CFQ_ATTR(slice_async_rq),
4053 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004054 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004055 CFQ_ATTR(low_latency),
Vivek Goyalae30c282009-12-03 12:59:55 -05004056 CFQ_ATTR(group_isolation),
Al Viroe572ec72006-03-18 22:27:18 -05004057 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058};
4059
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060static struct elevator_type iosched_cfq = {
4061 .ops = {
4062 .elevator_merge_fn = cfq_merge,
4063 .elevator_merged_fn = cfq_merged_request,
4064 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004065 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004066 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004067 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004069 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004070 .elevator_deactivate_req_fn = cfq_deactivate_request,
4071 .elevator_queue_empty_fn = cfq_queue_empty,
4072 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004073 .elevator_former_req_fn = elv_rb_former_request,
4074 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075 .elevator_set_req_fn = cfq_set_request,
4076 .elevator_put_req_fn = cfq_put_request,
4077 .elevator_may_queue_fn = cfq_may_queue,
4078 .elevator_init_fn = cfq_init_queue,
4079 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02004080 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081 },
Al Viro3d1ab402006-03-18 18:35:43 -05004082 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083 .elevator_name = "cfq",
4084 .elevator_owner = THIS_MODULE,
4085};
4086
Vivek Goyal3e252062009-12-04 10:36:42 -05004087#ifdef CONFIG_CFQ_GROUP_IOSCHED
4088static struct blkio_policy_type blkio_policy_cfq = {
4089 .ops = {
4090 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4091 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4092 },
Vivek Goyal062a6442010-09-15 17:06:33 -04004093 .plid = BLKIO_POLICY_PROP,
Vivek Goyal3e252062009-12-04 10:36:42 -05004094};
4095#else
4096static struct blkio_policy_type blkio_policy_cfq;
4097#endif
4098
Linus Torvalds1da177e2005-04-16 15:20:36 -07004099static int __init cfq_init(void)
4100{
Jens Axboe22e2c502005-06-27 10:55:12 +02004101 /*
4102 * could be 0 on HZ < 1000 setups
4103 */
4104 if (!cfq_slice_async)
4105 cfq_slice_async = 1;
4106 if (!cfq_slice_idle)
4107 cfq_slice_idle = 1;
4108
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004109#ifdef CONFIG_CFQ_GROUP_IOSCHED
4110 if (!cfq_group_idle)
4111 cfq_group_idle = 1;
4112#else
4113 cfq_group_idle = 0;
4114#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115 if (cfq_slab_setup())
4116 return -ENOMEM;
4117
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004118 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05004119 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004121 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004122}
4123
4124static void __exit cfq_exit(void)
4125{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07004126 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05004127 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128 elv_unregister(&iosched_cfq);
Al Viro334e94de2006-03-18 15:05:53 -05004129 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02004130 /* ioc_gone's update must be visible before reading ioc_count */
4131 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02004132
4133 /*
4134 * this also protects us from entering cfq_slab_kill() with
4135 * pending RCU callbacks
4136 */
Tejun Heo245b2e72009-06-24 15:13:48 +09004137 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02004138 wait_for_completion(&all_gone);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004139 ida_destroy(&cic_index_ida);
Christoph Hellwig83521d32005-10-30 15:01:39 -08004140 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004141}
4142
4143module_init(cfq_init);
4144module_exit(cfq_exit);
4145
4146MODULE_AUTHOR("Jens Axboe");
4147MODULE_LICENSE("GPL");
4148MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");