blob: f27ff3efe6cd2cf92af18aeb0ae063f7dc38ab3a [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) \
Mike Snitzerc1867942011-02-11 11:08:00 +010057 ((struct cfq_io_context *) (rq)->elevator_private[0])
58#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1])
59#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2])
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Christoph Lametere18b8902006-12-06 20:33:20 -080061static struct kmem_cache *cfq_pool;
62static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
Tejun Heo245b2e72009-06-24 15:13:48 +090064static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050065static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020066static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050067
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +040068static DEFINE_SPINLOCK(cic_index_lock);
69static DEFINE_IDA(cic_index_ida);
70
Jens Axboe22e2c502005-06-27 10:55:12 +020071#define CFQ_PRIO_LISTS IOPRIO_BE_NR
72#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020073#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
Jens Axboe206dc692006-03-28 13:03:44 +020075#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050076#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020077
Jens Axboe22e2c502005-06-27 10:55:12 +020078/*
Jens Axboecc09e292007-04-26 12:53:50 +020079 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
83 */
84struct cfq_rb_root {
85 struct rb_root rb;
86 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010087 unsigned count;
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010088 unsigned total_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050089 u64 min_vdisktime;
Jens Axboecc09e292007-04-26 12:53:50 +020090};
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010091#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92 .count = 0, .min_vdisktime = 0, }
Jens Axboecc09e292007-04-26 12:53:50 +020093
94/*
Jens Axboe6118b702009-06-30 09:34:12 +020095 * Per process-grouping structure
96 */
97struct cfq_queue {
98 /* reference count */
Shaohua Li30d7b942011-01-07 08:46:59 +010099 int ref;
Jens Axboe6118b702009-06-30 09:34:12 +0200100 /* various state flags, see below */
101 unsigned int flags;
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
117 int queued[2];
118 /* currently allocated requests */
119 int allocated[2];
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
122
Vivek Goyaldae739e2009-12-03 12:59:45 -0500123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500125 unsigned int allocated_slice;
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100126 unsigned int slice_dispatch;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
Jens Axboe6118b702009-06-30 09:34:12 +0200129 unsigned long slice_end;
130 long slice_resid;
Jens Axboe6118b702009-06-30 09:34:12 +0200131
132 /* pending metadata requests */
133 int meta_pending;
134 /* number of requests that are on the dispatch list or inside driver */
135 int dispatched;
136
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
139 unsigned short ioprio_class, org_ioprio_class;
140
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100141 pid_t pid;
142
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +0100143 u32 seek_history;
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400144 sector_t last_request_pos;
145
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100146 struct cfq_rb_root *service_tree;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400147 struct cfq_queue *new_cfqq;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500148 struct cfq_group *cfqg;
Vivek Goyalae30c282009-12-03 12:59:55 -0500149 struct cfq_group *orig_cfqg;
Vivek Goyalc4e78932010-08-23 12:25:03 +0200150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
Jens Axboe6118b702009-06-30 09:34:12 +0200152};
153
154/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100155 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100156 * IDLE is handled separately, so it has negative index
157 */
158enum wl_prio_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100159 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500160 RT_WORKLOAD = 1,
161 IDLE_WORKLOAD = 2,
Vivek Goyalb4627322010-10-22 09:48:43 +0200162 CFQ_PRIO_NR,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100163};
164
165/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100166 * Second index in the service_trees.
167 */
168enum wl_type_t {
169 ASYNC_WORKLOAD = 0,
170 SYNC_NOIDLE_WORKLOAD = 1,
171 SYNC_WORKLOAD = 2
172};
173
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500174/* This is per cgroup per device grouping structure */
175struct cfq_group {
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500176 /* group service_tree member */
177 struct rb_node rb_node;
178
179 /* group service_tree key */
180 u64 vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500181 unsigned int weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500182
183 /* number of cfqq currently on this group */
184 int nr_cfqq;
185
Jens Axboe22e2c502005-06-27 10:55:12 +0200186 /*
Vivek Goyalb4627322010-10-22 09:48:43 +0200187 * Per group busy queus average. Useful for workload slice calc. We
188 * create the array for each prio class but at run time it is used
189 * only for RT and BE class and slot for IDLE class remains unused.
190 * This is primarily done to avoid confusion and a gcc warning.
191 */
192 unsigned int busy_queues_avg[CFQ_PRIO_NR];
193 /*
194 * rr lists of queues with requests. We maintain service trees for
195 * RT and BE classes. These trees are subdivided in subclasses
196 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
197 * class there is no subclassification and all the cfq queues go on
198 * a single tree service_tree_idle.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100199 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200200 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100201 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100202 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500203
204 unsigned long saved_workload_slice;
205 enum wl_type_t saved_workload;
206 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500207 struct blkio_group blkg;
208#ifdef CONFIG_CFQ_GROUP_IOSCHED
209 struct hlist_node cfqd_node;
Shaohua Li329a6782011-01-07 08:48:28 +0100210 int ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500211#endif
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200212 /* number of requests that are on the dispatch list or inside driver */
213 int dispatched;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500214};
215
216/*
217 * Per block device queue structure
218 */
219struct cfq_data {
220 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500221 /* Root service tree for cfq_groups */
222 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500223 struct cfq_group root_group;
224
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100225 /*
226 * The priority currently being served
227 */
228 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100229 enum wl_type_t serving_type;
230 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500231 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200232
233 /*
234 * Each priority tree is sorted by next_request position. These
235 * trees are used when determining if two or more queues are
236 * interleaving requests (see cfq_close_cooperator).
237 */
238 struct rb_root prio_trees[CFQ_PRIO_LISTS];
239
Jens Axboe22e2c502005-06-27 10:55:12 +0200240 unsigned int busy_queues;
241
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100242 int rq_in_driver;
243 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200244
245 /*
246 * queue-depth detection
247 */
248 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200249 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100250 /*
251 * hw_tag can be
252 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
253 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
254 * 0 => no NCQ
255 */
256 int hw_tag_est_depth;
257 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200258
259 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200260 * idle window management
261 */
262 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200263 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200264
265 struct cfq_queue *active_queue;
266 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200267
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200268 /*
269 * async queue for each priority case
270 */
271 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
272 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200273
Jens Axboe6d048f52007-04-25 12:44:27 +0200274 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 /*
277 * tunables, see top of file
278 */
279 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200280 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 unsigned int cfq_back_penalty;
282 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200283 unsigned int cfq_slice[2];
284 unsigned int cfq_slice_async_rq;
285 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200286 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200287 unsigned int cfq_latency;
Vivek Goyalae30c282009-12-03 12:59:55 -0500288 unsigned int cfq_group_isolation;
Al Virod9ff4182006-03-18 13:51:22 -0500289
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400290 unsigned int cic_index;
Al Virod9ff4182006-03-18 13:51:22 -0500291 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
Jens Axboe6118b702009-06-30 09:34:12 +0200293 /*
294 * Fallback dummy cfqq for extreme OOM conditions
295 */
296 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200297
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100298 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500299
300 /* List of cfq groups being managed on this device*/
301 struct hlist_head cfqg_list;
Jens Axboebb729bc2009-12-06 09:54:19 +0100302 struct rcu_head rcu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303};
304
Vivek Goyal25fb5162009-12-03 12:59:46 -0500305static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
306
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500307static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
308 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500309 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100310{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500311 if (!cfqg)
312 return NULL;
313
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100314 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500315 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100316
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500317 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100318}
319
Jens Axboe3b181522005-06-27 10:56:24 +0200320enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100321 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
322 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200323 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100324 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100325 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
326 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
327 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100328 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200329 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400330 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100331 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100332 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500333 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200334};
335
336#define CFQ_CFQQ_FNS(name) \
337static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
338{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100339 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200340} \
341static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
342{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100343 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200344} \
345static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
346{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100347 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200348}
349
350CFQ_CFQQ_FNS(on_rr);
351CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200352CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200353CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200354CFQ_CFQQ_FNS(fifo_expire);
355CFQ_CFQQ_FNS(idle_window);
356CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100357CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200358CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200359CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100360CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100361CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500362CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200363#undef CFQ_CFQQ_FNS
364
Vivek Goyalafc24d42010-04-26 19:27:56 +0200365#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500366#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
367 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
368 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
369 blkg_path(&(cfqq)->cfqg->blkg), ##args);
370
371#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
372 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
373 blkg_path(&(cfqg)->blkg), ##args); \
374
375#else
Jens Axboe7b679132008-05-30 12:23:07 +0200376#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
377 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500378#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
379#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200380#define cfq_log(cfqd, fmt, args...) \
381 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
382
Vivek Goyal615f0252009-12-03 12:59:39 -0500383/* Traverses through cfq group service trees */
384#define for_each_cfqg_st(cfqg, i, j, st) \
385 for (i = 0; i <= IDLE_WORKLOAD; i++) \
386 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
387 : &cfqg->service_tree_idle; \
388 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
389 (i == IDLE_WORKLOAD && j == 0); \
390 j++, st = i < IDLE_WORKLOAD ? \
391 &cfqg->service_trees[i][j]: NULL) \
392
393
Vivek Goyal02b35082010-08-23 12:23:53 +0200394static inline bool iops_mode(struct cfq_data *cfqd)
395{
396 /*
397 * If we are not idling on queues and it is a NCQ drive, parallel
398 * execution of requests is on and measuring time is not possible
399 * in most of the cases until and unless we drive shallower queue
400 * depths and that becomes a performance bottleneck. In such cases
401 * switch to start providing fairness in terms of number of IOs.
402 */
403 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
404 return true;
405 else
406 return false;
407}
408
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100409static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
410{
411 if (cfq_class_idle(cfqq))
412 return IDLE_WORKLOAD;
413 if (cfq_class_rt(cfqq))
414 return RT_WORKLOAD;
415 return BE_WORKLOAD;
416}
417
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100418
419static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
420{
421 if (!cfq_cfqq_sync(cfqq))
422 return ASYNC_WORKLOAD;
423 if (!cfq_cfqq_idle_window(cfqq))
424 return SYNC_NOIDLE_WORKLOAD;
425 return SYNC_WORKLOAD;
426}
427
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500428static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
429 struct cfq_data *cfqd,
430 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100431{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500432 if (wl == IDLE_WORKLOAD)
433 return cfqg->service_tree_idle.count;
434
435 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
436 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
437 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100438}
439
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500440static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
441 struct cfq_group *cfqg)
442{
443 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
444 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
445}
446
Jens Axboe165125e2007-07-24 09:28:11 +0200447static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200448static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100449 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100450static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200451 struct io_context *);
452
453static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200454 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200455{
Jens Axboea6151c32009-10-07 20:02:57 +0200456 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200457}
458
459static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200460 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200461{
Jens Axboea6151c32009-10-07 20:02:57 +0200462 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200463}
464
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400465#define CIC_DEAD_KEY 1ul
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400466#define CIC_DEAD_INDEX_SHIFT 1
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400467
468static inline void *cfqd_dead_key(struct cfq_data *cfqd)
469{
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400470 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400471}
472
473static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
474{
475 struct cfq_data *cfqd = cic->key;
476
477 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
478 return NULL;
479
480 return cfqd;
481}
482
Vasily Tarasov91fac312007-04-25 12:29:51 +0200483/*
484 * We regard a request as SYNC, if it's either a read or has the SYNC bit
485 * set (in which case it could also be direct WRITE).
486 */
Jens Axboea6151c32009-10-07 20:02:57 +0200487static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200488{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200489 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200490}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700493 * scheduler run of queue, if there are requests pending and no one in the
494 * driver that will restart queueing
495 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200496static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700497{
Jens Axboe7b679132008-05-30 12:23:07 +0200498 if (cfqd->busy_queues) {
499 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200500 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200501 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700502}
503
Jens Axboe165125e2007-07-24 09:28:11 +0200504static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700505{
506 struct cfq_data *cfqd = q->elevator->elevator_data;
507
Vivek Goyalf04a6422009-12-03 12:59:40 -0500508 return !cfqd->rq_queued;
Andrew Morton99f95e52005-06-27 20:14:05 -0700509}
510
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100512 * Scale schedule slice based on io priority. Use the sync time slice only
513 * if a queue is marked sync and has sync io queued. A sync queue with async
514 * io only, should not get full sync slice length.
515 */
Jens Axboea6151c32009-10-07 20:02:57 +0200516static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200517 unsigned short prio)
518{
519 const int base_slice = cfqd->cfq_slice[sync];
520
521 WARN_ON(prio >= IOPRIO_BE_NR);
522
523 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
524}
525
Jens Axboe44f7c162007-01-19 11:51:58 +1100526static inline int
527cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
528{
Jens Axboed9e76202007-04-20 14:27:50 +0200529 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100530}
531
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500532static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
533{
534 u64 d = delta << CFQ_SERVICE_SHIFT;
535
536 d = d * BLKIO_WEIGHT_DEFAULT;
537 do_div(d, cfqg->weight);
538 return d;
539}
540
541static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
542{
543 s64 delta = (s64)(vdisktime - min_vdisktime);
544 if (delta > 0)
545 min_vdisktime = vdisktime;
546
547 return min_vdisktime;
548}
549
550static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
551{
552 s64 delta = (s64)(vdisktime - min_vdisktime);
553 if (delta < 0)
554 min_vdisktime = vdisktime;
555
556 return min_vdisktime;
557}
558
559static void update_min_vdisktime(struct cfq_rb_root *st)
560{
561 u64 vdisktime = st->min_vdisktime;
562 struct cfq_group *cfqg;
563
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500564 if (st->left) {
565 cfqg = rb_entry_cfqg(st->left);
566 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
567 }
568
569 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
570}
571
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100572/*
573 * get averaged number of queues of RT/BE priority.
574 * average is updated, with a formula that gives more weight to higher numbers,
575 * to quickly follows sudden increases and decrease slowly
576 */
577
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500578static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
579 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100580{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100581 unsigned min_q, max_q;
582 unsigned mult = cfq_hist_divisor - 1;
583 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500584 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100585
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500586 min_q = min(cfqg->busy_queues_avg[rt], busy);
587 max_q = max(cfqg->busy_queues_avg[rt], busy);
588 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100589 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500590 return cfqg->busy_queues_avg[rt];
591}
592
593static inline unsigned
594cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
595{
596 struct cfq_rb_root *st = &cfqd->grp_service_tree;
597
598 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100599}
600
Shaohua Lic553f8e2011-01-14 08:41:03 +0100601static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -0700602cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100603{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100604 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
605 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500606 /*
607 * interested queues (we consider only the ones with the same
608 * priority class in the cfq group)
609 */
610 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
611 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100612 unsigned sync_slice = cfqd->cfq_slice[1];
613 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500614 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
615
616 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100617 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
618 /* scale low_slice according to IO priority
619 * and sync vs async */
620 unsigned low_slice =
621 min(slice, base_low_slice * slice / sync_slice);
622 /* the adapted slice value is scaled to fit all iqs
623 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500624 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100625 low_slice);
626 }
627 }
Shaohua Lic553f8e2011-01-14 08:41:03 +0100628 return slice;
629}
630
631static inline void
632cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
633{
Vivek Goyalba5bd522011-01-19 08:25:02 -0700634 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +0100635
Vivek Goyaldae739e2009-12-03 12:59:45 -0500636 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100637 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500638 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200639 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100640}
641
642/*
643 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
644 * isn't valid until the first request from the dispatch is activated
645 * and the slice time set.
646 */
Jens Axboea6151c32009-10-07 20:02:57 +0200647static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100648{
649 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +0100650 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100651 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +0100652 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100653
Shaohua Lic1e44752010-11-08 15:01:02 +0100654 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +1100655}
656
657/*
Jens Axboe5e705372006-07-13 12:39:25 +0200658 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200660 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661 */
Jens Axboe5e705372006-07-13 12:39:25 +0200662static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100663cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100665 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200667#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
668#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
669 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670
Jens Axboe5e705372006-07-13 12:39:25 +0200671 if (rq1 == NULL || rq1 == rq2)
672 return rq2;
673 if (rq2 == NULL)
674 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200675
Jens Axboe5e705372006-07-13 12:39:25 +0200676 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
677 return rq1;
678 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
679 return rq2;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200680 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200681 return rq1;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200682 else if ((rq2->cmd_flags & REQ_META) &&
683 !(rq1->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200684 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685
Tejun Heo83096eb2009-05-07 22:24:39 +0900686 s1 = blk_rq_pos(rq1);
687 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 /*
690 * by definition, 1KiB is 2 sectors
691 */
692 back_max = cfqd->cfq_back_max * 2;
693
694 /*
695 * Strict one way elevator _except_ in the case where we allow
696 * short backward seeks which are biased as twice the cost of a
697 * similar forward seek.
698 */
699 if (s1 >= last)
700 d1 = s1 - last;
701 else if (s1 + back_max >= last)
702 d1 = (last - s1) * cfqd->cfq_back_penalty;
703 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200704 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
706 if (s2 >= last)
707 d2 = s2 - last;
708 else if (s2 + back_max >= last)
709 d2 = (last - s2) * cfqd->cfq_back_penalty;
710 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200711 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712
713 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714
Andreas Mohre8a99052006-03-28 08:59:49 +0200715 /*
716 * By doing switch() on the bit mask "wrap" we avoid having to
717 * check two variables for all permutations: --> faster!
718 */
719 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200720 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200721 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200722 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200723 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200724 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200725 else {
726 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200727 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200728 else
Jens Axboe5e705372006-07-13 12:39:25 +0200729 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200730 }
731
732 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200733 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200734 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200735 return rq2;
736 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200737 default:
738 /*
739 * Since both rqs are wrapped,
740 * start with the one that's further behind head
741 * (--> only *one* back seek required),
742 * since back seek takes more time than forward.
743 */
744 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200745 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 else
Jens Axboe5e705372006-07-13 12:39:25 +0200747 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 }
749}
750
Jens Axboe498d3aa22007-04-26 12:54:48 +0200751/*
752 * The below is leftmost cache rbtree addon
753 */
Jens Axboe08717142008-01-28 11:38:15 +0100754static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200755{
Vivek Goyal615f0252009-12-03 12:59:39 -0500756 /* Service tree is empty */
757 if (!root->count)
758 return NULL;
759
Jens Axboecc09e292007-04-26 12:53:50 +0200760 if (!root->left)
761 root->left = rb_first(&root->rb);
762
Jens Axboe08717142008-01-28 11:38:15 +0100763 if (root->left)
764 return rb_entry(root->left, struct cfq_queue, rb_node);
765
766 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200767}
768
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500769static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
770{
771 if (!root->left)
772 root->left = rb_first(&root->rb);
773
774 if (root->left)
775 return rb_entry_cfqg(root->left);
776
777 return NULL;
778}
779
Jens Axboea36e71f2009-04-15 12:15:11 +0200780static void rb_erase_init(struct rb_node *n, struct rb_root *root)
781{
782 rb_erase(n, root);
783 RB_CLEAR_NODE(n);
784}
785
Jens Axboecc09e292007-04-26 12:53:50 +0200786static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
787{
788 if (root->left == n)
789 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200790 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100791 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200792}
793
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794/*
795 * would be nice to take fifo expire time into account as well
796 */
Jens Axboe5e705372006-07-13 12:39:25 +0200797static struct request *
798cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
799 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800{
Jens Axboe21183b02006-07-13 12:33:14 +0200801 struct rb_node *rbnext = rb_next(&last->rb_node);
802 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200803 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804
Jens Axboe21183b02006-07-13 12:33:14 +0200805 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806
807 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200808 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200809
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200811 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200812 else {
813 rbnext = rb_first(&cfqq->sort_list);
814 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200815 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200816 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100818 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819}
820
Jens Axboed9e76202007-04-20 14:27:50 +0200821static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
822 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823{
Jens Axboed9e76202007-04-20 14:27:50 +0200824 /*
825 * just an approximation, should be ok.
826 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500827 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100828 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200829}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500831static inline s64
832cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
833{
834 return cfqg->vdisktime - st->min_vdisktime;
835}
836
837static void
838__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
839{
840 struct rb_node **node = &st->rb.rb_node;
841 struct rb_node *parent = NULL;
842 struct cfq_group *__cfqg;
843 s64 key = cfqg_key(st, cfqg);
844 int left = 1;
845
846 while (*node != NULL) {
847 parent = *node;
848 __cfqg = rb_entry_cfqg(parent);
849
850 if (key < cfqg_key(st, __cfqg))
851 node = &parent->rb_left;
852 else {
853 node = &parent->rb_right;
854 left = 0;
855 }
856 }
857
858 if (left)
859 st->left = &cfqg->rb_node;
860
861 rb_link_node(&cfqg->rb_node, parent, node);
862 rb_insert_color(&cfqg->rb_node, &st->rb);
863}
864
865static void
866cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
867{
868 struct cfq_rb_root *st = &cfqd->grp_service_tree;
869 struct cfq_group *__cfqg;
870 struct rb_node *n;
871
872 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +0100873 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500874 return;
875
876 /*
877 * Currently put the group at the end. Later implement something
878 * so that groups get lesser vtime based on their weights, so that
879 * if group does not loose all if it was not continously backlogged.
880 */
881 n = rb_last(&st->rb);
882 if (n) {
883 __cfqg = rb_entry_cfqg(n);
884 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
885 } else
886 cfqg->vdisktime = st->min_vdisktime;
887
888 __cfq_group_service_tree_add(st, cfqg);
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
897 BUG_ON(cfqg->nr_cfqq < 1);
898 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500899
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500900 /* If there are other cfq queues under this group, don't delete it */
901 if (cfqg->nr_cfqq)
902 return;
903
Vivek Goyal2868ef72009-12-03 12:59:48 -0500904 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500905 st->total_weight -= cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500906 if (!RB_EMPTY_NODE(&cfqg->rb_node))
907 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500908 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400909 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500910}
911
912static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
913{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500914 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500915
916 /*
917 * Queue got expired before even a single request completed or
918 * got expired immediately after first request completion.
919 */
920 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
921 /*
922 * Also charge the seek time incurred to the group, otherwise
923 * if there are mutiple queues in the group, each can dispatch
924 * a single request on seeky media and cause lots of seek time
925 * and group will never know it.
926 */
927 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
928 1);
929 } else {
930 slice_used = jiffies - cfqq->slice_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500931 if (slice_used > cfqq->allocated_slice)
932 slice_used = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500933 }
934
Vivek Goyaldae739e2009-12-03 12:59:45 -0500935 return slice_used;
936}
937
938static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200939 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500940{
941 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal02b35082010-08-23 12:23:53 +0200942 unsigned int used_sl, charge;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500943 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
944 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500945
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500946 BUG_ON(nr_sync < 0);
Vivek Goyal02b35082010-08-23 12:23:53 +0200947 used_sl = charge = cfq_cfqq_slice_usage(cfqq);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500948
Vivek Goyal02b35082010-08-23 12:23:53 +0200949 if (iops_mode(cfqd))
950 charge = cfqq->slice_dispatch;
951 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
952 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500953
954 /* Can't update vdisktime while group is on service tree */
955 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyal02b35082010-08-23 12:23:53 +0200956 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500957 __cfq_group_service_tree_add(st, cfqg);
958
959 /* This group is being expired. Save the context */
960 if (time_after(cfqd->workload_expires, jiffies)) {
961 cfqg->saved_workload_slice = cfqd->workload_expires
962 - jiffies;
963 cfqg->saved_workload = cfqd->serving_type;
964 cfqg->saved_serving_prio = cfqd->serving_prio;
965 } else
966 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500967
968 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
969 st->min_vdisktime);
Vivek Goyalc4e78932010-08-23 12:25:03 +0200970 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
971 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
972 iops_mode(cfqd), cfqq->nr_sectors);
Vivek Goyale98ef892010-06-18 10:39:47 -0400973 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
974 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500975}
976
Vivek Goyal25fb5162009-12-03 12:59:46 -0500977#ifdef CONFIG_CFQ_GROUP_IOSCHED
978static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
979{
980 if (blkg)
981 return container_of(blkg, struct cfq_group, blkg);
982 return NULL;
983}
984
Vivek Goyalfe071432010-10-01 14:49:49 +0200985void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
986 unsigned int weight)
Vivek Goyalf8d461d2009-12-03 12:59:52 -0500987{
988 cfqg_of_blkg(blkg)->weight = weight;
989}
990
Vivek Goyal25fb5162009-12-03 12:59:46 -0500991static struct cfq_group *
992cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
993{
994 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
995 struct cfq_group *cfqg = NULL;
996 void *key = cfqd;
997 int i, j;
998 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -0500999 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1000 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001001
Vivek Goyal25fb5162009-12-03 12:59:46 -05001002 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
Ricky Beniteza74b2ad2010-04-05 18:22:17 +02001003 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1004 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1005 cfqg->blkg.dev = MKDEV(major, minor);
1006 goto done;
1007 }
Vivek Goyal25fb5162009-12-03 12:59:46 -05001008 if (cfqg || !create)
1009 goto done;
1010
1011 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1012 if (!cfqg)
1013 goto done;
1014
Vivek Goyal25fb5162009-12-03 12:59:46 -05001015 for_each_cfqg_st(cfqg, i, j, st)
1016 *st = CFQ_RB_ROOT;
1017 RB_CLEAR_NODE(&cfqg->rb_node);
1018
Vivek Goyalb1c35762009-12-03 12:59:47 -05001019 /*
1020 * Take the initial reference that will be released on destroy
1021 * This can be thought of a joint reference by cgroup and
1022 * elevator which will be dropped by either elevator exit
1023 * or cgroup deletion path depending on who is exiting first.
1024 */
Shaohua Li329a6782011-01-07 08:48:28 +01001025 cfqg->ref = 1;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001026
Vivek Goyal180be2a2010-09-14 08:47:11 +02001027 /*
1028 * Add group onto cgroup list. It might happen that bdi->dev is
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001029 * not initialized yet. Initialize this new group without major
Vivek Goyal180be2a2010-09-14 08:47:11 +02001030 * and minor info and this info will be filled in once a new thread
1031 * comes for IO. See code above.
1032 */
1033 if (bdi->dev) {
1034 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1035 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
Vivek Goyal22084192009-12-03 12:59:49 -05001036 MKDEV(major, minor));
Vivek Goyal180be2a2010-09-14 08:47:11 +02001037 } else
1038 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1039 0);
1040
Gui Jianfeng34d0f172010-04-13 16:05:49 +08001041 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001042
1043 /* Add group on cfqd list */
1044 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1045
1046done:
Vivek Goyal25fb5162009-12-03 12:59:46 -05001047 return cfqg;
1048}
1049
1050/*
1051 * Search for the cfq group current task belongs to. If create = 1, then also
1052 * create the cfq group if it does not exist. request_queue lock must be held.
1053 */
1054static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1055{
1056 struct cgroup *cgroup;
1057 struct cfq_group *cfqg = NULL;
1058
1059 rcu_read_lock();
1060 cgroup = task_cgroup(current, blkio_subsys_id);
1061 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1062 if (!cfqg && create)
1063 cfqg = &cfqd->root_group;
1064 rcu_read_unlock();
1065 return cfqg;
1066}
1067
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001068static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1069{
Shaohua Li329a6782011-01-07 08:48:28 +01001070 cfqg->ref++;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001071 return cfqg;
1072}
1073
Vivek Goyal25fb5162009-12-03 12:59:46 -05001074static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1075{
1076 /* Currently, all async queues are mapped to root group */
1077 if (!cfq_cfqq_sync(cfqq))
1078 cfqg = &cfqq->cfqd->root_group;
1079
1080 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001081 /* cfqq reference on cfqg */
Shaohua Li329a6782011-01-07 08:48:28 +01001082 cfqq->cfqg->ref++;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001083}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001084
1085static void cfq_put_cfqg(struct cfq_group *cfqg)
1086{
1087 struct cfq_rb_root *st;
1088 int i, j;
1089
Shaohua Li329a6782011-01-07 08:48:28 +01001090 BUG_ON(cfqg->ref <= 0);
1091 cfqg->ref--;
1092 if (cfqg->ref)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001093 return;
1094 for_each_cfqg_st(cfqg, i, j, st)
Gui Jianfengb54ce602010-11-30 20:52:46 +01001095 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
Vivek Goyalb1c35762009-12-03 12:59:47 -05001096 kfree(cfqg);
1097}
1098
1099static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1100{
1101 /* Something wrong if we are trying to remove same group twice */
1102 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1103
1104 hlist_del_init(&cfqg->cfqd_node);
1105
1106 /*
1107 * Put the reference taken at the time of creation so that when all
1108 * queues are gone, group can be destroyed.
1109 */
1110 cfq_put_cfqg(cfqg);
1111}
1112
1113static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1114{
1115 struct hlist_node *pos, *n;
1116 struct cfq_group *cfqg;
1117
1118 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1119 /*
1120 * If cgroup removal path got to blk_group first and removed
1121 * it from cgroup list, then it will take care of destroying
1122 * cfqg also.
1123 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001124 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001125 cfq_destroy_cfqg(cfqd, cfqg);
1126 }
1127}
1128
1129/*
1130 * Blk cgroup controller notification saying that blkio_group object is being
1131 * delinked as associated cgroup object is going away. That also means that
1132 * no new IO will come in this group. So get rid of this group as soon as
1133 * any pending IO in the group is finished.
1134 *
1135 * This function is called under rcu_read_lock(). key is the rcu protected
1136 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1137 * read lock.
1138 *
1139 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1140 * it should not be NULL as even if elevator was exiting, cgroup deltion
1141 * path got to it first.
1142 */
1143void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1144{
1145 unsigned long flags;
1146 struct cfq_data *cfqd = key;
1147
1148 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1149 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1150 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1151}
1152
Vivek Goyal25fb5162009-12-03 12:59:46 -05001153#else /* GROUP_IOSCHED */
1154static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1155{
1156 return &cfqd->root_group;
1157}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001158
1159static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1160{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001161 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001162}
1163
Vivek Goyal25fb5162009-12-03 12:59:46 -05001164static inline void
1165cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1166 cfqq->cfqg = cfqg;
1167}
1168
Vivek Goyalb1c35762009-12-03 12:59:47 -05001169static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1170static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1171
Vivek Goyal25fb5162009-12-03 12:59:46 -05001172#endif /* GROUP_IOSCHED */
1173
Jens Axboe498d3aa22007-04-26 12:54:48 +02001174/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001175 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001176 * requests waiting to be processed. It is sorted in the order that
1177 * we will service the queues.
1178 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001179static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001180 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001181{
Jens Axboe08717142008-01-28 11:38:15 +01001182 struct rb_node **p, *parent;
1183 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001184 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001185 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001186 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001187 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001188 int group_changed = 0;
1189
1190#ifdef CONFIG_CFQ_GROUP_IOSCHED
1191 if (!cfqd->cfq_group_isolation
1192 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1193 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1194 /* Move this cfq to root group */
1195 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1196 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1197 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1198 cfqq->orig_cfqg = cfqq->cfqg;
1199 cfqq->cfqg = &cfqd->root_group;
Shaohua Li329a6782011-01-07 08:48:28 +01001200 cfqd->root_group.ref++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001201 group_changed = 1;
1202 } else if (!cfqd->cfq_group_isolation
1203 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1204 /* cfqq is sequential now needs to go to its original group */
1205 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1206 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1207 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1208 cfq_put_cfqg(cfqq->cfqg);
1209 cfqq->cfqg = cfqq->orig_cfqg;
1210 cfqq->orig_cfqg = NULL;
1211 group_changed = 1;
1212 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1213 }
1214#endif
Jens Axboed9e76202007-04-20 14:27:50 +02001215
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001216 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001217 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001218 if (cfq_class_idle(cfqq)) {
1219 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001220 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001221 if (parent && parent != &cfqq->rb_node) {
1222 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1223 rb_key += __cfqq->rb_key;
1224 } else
1225 rb_key += jiffies;
1226 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001227 /*
1228 * Get our rb key offset. Subtract any residual slice
1229 * value carried from last service. A negative resid
1230 * count indicates slice overrun, and this should position
1231 * the next service time further away in the tree.
1232 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001233 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001234 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001235 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001236 } else {
1237 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001238 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001239 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1240 }
Jens Axboed9e76202007-04-20 14:27:50 +02001241
1242 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001243 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001244 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001245 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001246 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001247 if (rb_key == cfqq->rb_key &&
1248 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001249 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001250
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001251 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1252 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001253 }
Jens Axboed9e76202007-04-20 14:27:50 +02001254
Jens Axboe498d3aa22007-04-26 12:54:48 +02001255 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001256 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001257 cfqq->service_tree = service_tree;
1258 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001259 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001260 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001261
Jens Axboed9e76202007-04-20 14:27:50 +02001262 parent = *p;
1263 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1264
Jens Axboe0c534e02007-04-18 20:01:57 +02001265 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001266 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001267 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001268 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001269 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001270 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001271 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001272 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001273 }
Jens Axboe67060e32007-04-18 20:13:32 +02001274
1275 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001276 }
1277
Jens Axboecc09e292007-04-26 12:53:50 +02001278 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001279 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001280
Jens Axboed9e76202007-04-20 14:27:50 +02001281 cfqq->rb_key = rb_key;
1282 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001283 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1284 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001285 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001286 return;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001287 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288}
1289
Jens Axboea36e71f2009-04-15 12:15:11 +02001290static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001291cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1292 sector_t sector, struct rb_node **ret_parent,
1293 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001294{
Jens Axboea36e71f2009-04-15 12:15:11 +02001295 struct rb_node **p, *parent;
1296 struct cfq_queue *cfqq = NULL;
1297
1298 parent = NULL;
1299 p = &root->rb_node;
1300 while (*p) {
1301 struct rb_node **n;
1302
1303 parent = *p;
1304 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1305
1306 /*
1307 * Sort strictly based on sector. Smallest to the left,
1308 * largest to the right.
1309 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001310 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001311 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001312 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001313 n = &(*p)->rb_left;
1314 else
1315 break;
1316 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001317 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001318 }
1319
1320 *ret_parent = parent;
1321 if (rb_link)
1322 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001323 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001324}
1325
1326static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1327{
Jens Axboea36e71f2009-04-15 12:15:11 +02001328 struct rb_node **p, *parent;
1329 struct cfq_queue *__cfqq;
1330
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001331 if (cfqq->p_root) {
1332 rb_erase(&cfqq->p_node, cfqq->p_root);
1333 cfqq->p_root = NULL;
1334 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001335
1336 if (cfq_class_idle(cfqq))
1337 return;
1338 if (!cfqq->next_rq)
1339 return;
1340
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001341 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001342 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1343 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001344 if (!__cfqq) {
1345 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001346 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1347 } else
1348 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001349}
1350
Jens Axboe498d3aa22007-04-26 12:54:48 +02001351/*
1352 * Update cfqq's position in the service tree.
1353 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001354static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001355{
Jens Axboe6d048f52007-04-25 12:44:27 +02001356 /*
1357 * Resorting requires the cfqq to be on the RR list already.
1358 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001359 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001360 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001361 cfq_prio_tree_add(cfqd, cfqq);
1362 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001363}
1364
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365/*
1366 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001367 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 */
Jens Axboefebffd62008-01-28 13:19:43 +01001369static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370{
Jens Axboe7b679132008-05-30 12:23:07 +02001371 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001372 BUG_ON(cfq_cfqq_on_rr(cfqq));
1373 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 cfqd->busy_queues++;
1375
Jens Axboeedd75ff2007-04-19 12:03:34 +02001376 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377}
1378
Jens Axboe498d3aa22007-04-26 12:54:48 +02001379/*
1380 * Called when the cfqq no longer has requests pending, remove it from
1381 * the service tree.
1382 */
Jens Axboefebffd62008-01-28 13:19:43 +01001383static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384{
Jens Axboe7b679132008-05-30 12:23:07 +02001385 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001386 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1387 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001389 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1390 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1391 cfqq->service_tree = NULL;
1392 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001393 if (cfqq->p_root) {
1394 rb_erase(&cfqq->p_node, cfqq->p_root);
1395 cfqq->p_root = NULL;
1396 }
Jens Axboed9e76202007-04-20 14:27:50 +02001397
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001398 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 BUG_ON(!cfqd->busy_queues);
1400 cfqd->busy_queues--;
1401}
1402
1403/*
1404 * rb tree support functions
1405 */
Jens Axboefebffd62008-01-28 13:19:43 +01001406static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407{
Jens Axboe5e705372006-07-13 12:39:25 +02001408 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001409 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410
Jens Axboeb4878f22005-10-20 16:42:29 +02001411 BUG_ON(!cfqq->queued[sync]);
1412 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413
Jens Axboe5e705372006-07-13 12:39:25 +02001414 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415
Vivek Goyalf04a6422009-12-03 12:59:40 -05001416 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1417 /*
1418 * Queue will be deleted from service tree when we actually
1419 * expire it later. Right now just remove it from prio tree
1420 * as it is empty.
1421 */
1422 if (cfqq->p_root) {
1423 rb_erase(&cfqq->p_node, cfqq->p_root);
1424 cfqq->p_root = NULL;
1425 }
1426 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427}
1428
Jens Axboe5e705372006-07-13 12:39:25 +02001429static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430{
Jens Axboe5e705372006-07-13 12:39:25 +02001431 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001433 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
Jens Axboe5380a102006-07-13 12:37:56 +02001435 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436
1437 /*
1438 * looks a little odd, but the first insert might return an alias.
1439 * if that happens, put the alias on the dispatch list
1440 */
Jens Axboe21183b02006-07-13 12:33:14 +02001441 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001442 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001443
1444 if (!cfq_cfqq_on_rr(cfqq))
1445 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001446
1447 /*
1448 * check if this request is a better next-serve candidate
1449 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001450 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001451 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001452
1453 /*
1454 * adjust priority tree position, if ->next_rq changes
1455 */
1456 if (prev != cfqq->next_rq)
1457 cfq_prio_tree_add(cfqd, cfqq);
1458
Jens Axboe5044eed2007-04-25 11:53:48 +02001459 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460}
1461
Jens Axboefebffd62008-01-28 13:19:43 +01001462static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463{
Jens Axboe5380a102006-07-13 12:37:56 +02001464 elv_rb_del(&cfqq->sort_list, rq);
1465 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001466 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1467 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001468 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001469 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001470 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1471 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472}
1473
Jens Axboe206dc692006-03-28 13:03:44 +02001474static struct request *
1475cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476{
Jens Axboe206dc692006-03-28 13:03:44 +02001477 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001478 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001479 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480
Jens Axboe4ac845a2008-01-24 08:44:49 +01001481 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001482 if (!cic)
1483 return NULL;
1484
1485 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001486 if (cfqq) {
1487 sector_t sector = bio->bi_sector + bio_sectors(bio);
1488
Jens Axboe21183b02006-07-13 12:33:14 +02001489 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001490 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 return NULL;
1493}
1494
Jens Axboe165125e2007-07-24 09:28:11 +02001495static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001496{
1497 struct cfq_data *cfqd = q->elevator->elevator_data;
1498
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001499 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001500 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001501 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001502
Tejun Heo5b936292009-05-07 22:24:38 +09001503 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001504}
1505
Jens Axboe165125e2007-07-24 09:28:11 +02001506static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507{
Jens Axboe22e2c502005-06-27 10:55:12 +02001508 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001510 WARN_ON(!cfqd->rq_in_driver);
1511 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001512 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001513 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514}
1515
Jens Axboeb4878f22005-10-20 16:42:29 +02001516static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517{
Jens Axboe5e705372006-07-13 12:39:25 +02001518 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001519
Jens Axboe5e705372006-07-13 12:39:25 +02001520 if (cfqq->next_rq == rq)
1521 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522
Jens Axboeb4878f22005-10-20 16:42:29 +02001523 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001524 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001525
Aaron Carroll45333d52008-08-26 15:52:36 +02001526 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001527 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1528 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02001529 if (rq->cmd_flags & REQ_META) {
Jens Axboe374f84a2006-07-23 01:42:19 +02001530 WARN_ON(!cfqq->meta_pending);
1531 cfqq->meta_pending--;
1532 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533}
1534
Jens Axboe165125e2007-07-24 09:28:11 +02001535static int cfq_merge(struct request_queue *q, struct request **req,
1536 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537{
1538 struct cfq_data *cfqd = q->elevator->elevator_data;
1539 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540
Jens Axboe206dc692006-03-28 13:03:44 +02001541 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001542 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001543 *req = __rq;
1544 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 }
1546
1547 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548}
1549
Jens Axboe165125e2007-07-24 09:28:11 +02001550static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001551 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552{
Jens Axboe21183b02006-07-13 12:33:14 +02001553 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001554 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555
Jens Axboe5e705372006-07-13 12:39:25 +02001556 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558}
1559
Divyesh Shah812d4022010-04-08 21:14:23 -07001560static void cfq_bio_merged(struct request_queue *q, struct request *req,
1561 struct bio *bio)
1562{
Vivek Goyale98ef892010-06-18 10:39:47 -04001563 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1564 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001565}
1566
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567static void
Jens Axboe165125e2007-07-24 09:28:11 +02001568cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 struct request *next)
1570{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001571 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001572 /*
1573 * reposition in fifo if next is older than rq
1574 */
1575 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001576 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001577 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001578 rq_set_fifo_time(rq, rq_fifo_time(next));
1579 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001580
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001581 if (cfqq->next_rq == next)
1582 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001583 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001584 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1585 rq_data_dir(next), rq_is_sync(next));
Jens Axboe22e2c502005-06-27 10:55:12 +02001586}
1587
Jens Axboe165125e2007-07-24 09:28:11 +02001588static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001589 struct bio *bio)
1590{
1591 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001592 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001593 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001594
1595 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001596 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001597 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001598 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001599 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001600
1601 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001602 * Lookup the cfqq that this bio will be queued with. Allow
1603 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001604 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001605 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001606 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001607 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001608
Vasily Tarasov91fac312007-04-25 12:29:51 +02001609 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001610 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001611}
1612
Divyesh Shah812df482010-04-08 21:15:35 -07001613static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1614{
1615 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001616 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001617}
1618
Jens Axboefebffd62008-01-28 13:19:43 +01001619static void __cfq_set_active_queue(struct cfq_data *cfqd,
1620 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001621{
1622 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001623 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1624 cfqd->serving_prio, cfqd->serving_type);
Vivek Goyale98ef892010-06-18 10:39:47 -04001625 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001626 cfqq->slice_start = 0;
1627 cfqq->dispatch_start = jiffies;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001628 cfqq->allocated_slice = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001629 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001630 cfqq->slice_dispatch = 0;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001631 cfqq->nr_sectors = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001632
Jens Axboe2f5cb732009-04-07 08:51:19 +02001633 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +02001634 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001635 cfq_clear_cfqq_must_alloc_slice(cfqq);
1636 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +11001637 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001638
Divyesh Shah812df482010-04-08 21:15:35 -07001639 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001640 }
1641
1642 cfqd->active_queue = cfqq;
1643}
1644
1645/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001646 * current cfqq expired its slice (or was too idle), select new one
1647 */
1648static void
1649__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001650 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001651{
Jens Axboe7b679132008-05-30 12:23:07 +02001652 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1653
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001654 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001655 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001656
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001657 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001658 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001659
1660 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001661 * If this cfqq is shared between multiple processes, check to
1662 * make sure that those processes are still issuing I/Os within
1663 * the mean seek distance. If not, it may be time to break the
1664 * queues apart again.
1665 */
1666 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1667 cfq_mark_cfqq_split_coop(cfqq);
1668
1669 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001670 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001671 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01001672 if (timed_out) {
1673 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07001674 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001675 else
1676 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
1690 if (cfqd->active_cic) {
1691 put_io_context(cfqd->active_cic->ioc);
1692 cfqd->active_cic = NULL;
1693 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001694}
1695
Vivek Goyale5ff0822010-04-26 19:25:11 +02001696static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001697{
1698 struct cfq_queue *cfqq = cfqd->active_queue;
1699
1700 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001701 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001702}
1703
Jens Axboe498d3aa22007-04-26 12:54:48 +02001704/*
1705 * Get next queue for service. Unless we have a queue preemption,
1706 * we'll simply select the first cfqq in the service tree.
1707 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001708static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001709{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001710 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001711 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001712 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001713
Vivek Goyalf04a6422009-12-03 12:59:40 -05001714 if (!cfqd->rq_queued)
1715 return NULL;
1716
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001717 /* There is nothing to dispatch */
1718 if (!service_tree)
1719 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001720 if (RB_EMPTY_ROOT(&service_tree->rb))
1721 return NULL;
1722 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001723}
1724
Vivek Goyalf04a6422009-12-03 12:59:40 -05001725static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1726{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001727 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001728 struct cfq_queue *cfqq;
1729 int i, j;
1730 struct cfq_rb_root *st;
1731
1732 if (!cfqd->rq_queued)
1733 return NULL;
1734
Vivek Goyal25fb5162009-12-03 12:59:46 -05001735 cfqg = cfq_get_next_cfqg(cfqd);
1736 if (!cfqg)
1737 return NULL;
1738
Vivek Goyalf04a6422009-12-03 12:59:40 -05001739 for_each_cfqg_st(cfqg, i, j, st)
1740 if ((cfqq = cfq_rb_first(st)) != NULL)
1741 return cfqq;
1742 return NULL;
1743}
1744
Jens Axboe498d3aa22007-04-26 12:54:48 +02001745/*
1746 * Get and set a new active queue for service.
1747 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001748static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1749 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001750{
Jens Axboee00ef792009-11-04 08:54:55 +01001751 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001752 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001753
Jens Axboe22e2c502005-06-27 10:55:12 +02001754 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001755 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001756}
1757
Jens Axboed9e76202007-04-20 14:27:50 +02001758static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1759 struct request *rq)
1760{
Tejun Heo83096eb2009-05-07 22:24:39 +09001761 if (blk_rq_pos(rq) >= cfqd->last_position)
1762 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001763 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001764 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001765}
1766
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001767static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001768 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001769{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001770 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001771}
1772
Jens Axboea36e71f2009-04-15 12:15:11 +02001773static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1774 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001775{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001776 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001777 struct rb_node *parent, *node;
1778 struct cfq_queue *__cfqq;
1779 sector_t sector = cfqd->last_position;
1780
1781 if (RB_EMPTY_ROOT(root))
1782 return NULL;
1783
1784 /*
1785 * First, if we find a request starting at the end of the last
1786 * request, choose it.
1787 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001788 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001789 if (__cfqq)
1790 return __cfqq;
1791
1792 /*
1793 * If the exact sector wasn't found, the parent of the NULL leaf
1794 * will contain the closest sector.
1795 */
1796 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001797 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001798 return __cfqq;
1799
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001800 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001801 node = rb_next(&__cfqq->p_node);
1802 else
1803 node = rb_prev(&__cfqq->p_node);
1804 if (!node)
1805 return NULL;
1806
1807 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001808 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001809 return __cfqq;
1810
1811 return NULL;
1812}
1813
1814/*
1815 * cfqd - obvious
1816 * cur_cfqq - passed in so that we don't decide that the current queue is
1817 * closely cooperating with itself.
1818 *
1819 * So, basically we're assuming that that cur_cfqq has dispatched at least
1820 * one request, and that cfqd->last_position reflects a position on the disk
1821 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1822 * assumption.
1823 */
1824static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001825 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001826{
1827 struct cfq_queue *cfqq;
1828
Divyesh Shah39c01b22010-03-25 15:45:57 +01001829 if (cfq_class_idle(cur_cfqq))
1830 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001831 if (!cfq_cfqq_sync(cur_cfqq))
1832 return NULL;
1833 if (CFQQ_SEEKY(cur_cfqq))
1834 return NULL;
1835
Jens Axboea36e71f2009-04-15 12:15:11 +02001836 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001837 * Don't search priority tree if it's the only queue in the group.
1838 */
1839 if (cur_cfqq->cfqg->nr_cfqq == 1)
1840 return NULL;
1841
1842 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001843 * We should notice if some of the queues are cooperating, eg
1844 * working closely on the same area of the disk. In that case,
1845 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001846 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001847 cfqq = cfqq_close(cfqd, cur_cfqq);
1848 if (!cfqq)
1849 return NULL;
1850
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001851 /* If new queue belongs to different cfq_group, don't choose it */
1852 if (cur_cfqq->cfqg != cfqq->cfqg)
1853 return NULL;
1854
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001855 /*
1856 * It only makes sense to merge sync queues.
1857 */
1858 if (!cfq_cfqq_sync(cfqq))
1859 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001860 if (CFQQ_SEEKY(cfqq))
1861 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001862
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001863 /*
1864 * Do not merge queues of different priority classes
1865 */
1866 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1867 return NULL;
1868
Jens Axboea36e71f2009-04-15 12:15:11 +02001869 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001870}
1871
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001872/*
1873 * Determine whether we should enforce idle window for this queue.
1874 */
1875
1876static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1877{
1878 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001879 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001880
Vivek Goyalf04a6422009-12-03 12:59:40 -05001881 BUG_ON(!service_tree);
1882 BUG_ON(!service_tree->count);
1883
Vivek Goyalb6508c12010-08-23 12:23:33 +02001884 if (!cfqd->cfq_slice_idle)
1885 return false;
1886
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001887 /* We never do for idle class queues. */
1888 if (prio == IDLE_WORKLOAD)
1889 return false;
1890
1891 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001892 if (cfq_cfqq_idle_window(cfqq) &&
1893 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001894 return true;
1895
1896 /*
1897 * Otherwise, we do only if they are the last ones
1898 * in their service tree.
1899 */
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001900 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01001901 return true;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001902 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1903 service_tree->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01001904 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001905}
1906
Jens Axboe6d048f52007-04-25 12:44:27 +02001907static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001908{
Jens Axboe17926692007-01-19 11:59:30 +11001909 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001910 struct cfq_io_context *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001911 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001912
Jens Axboea68bbddba2008-09-24 13:03:33 +02001913 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001914 * SSD device without seek penalty, disable idling. But only do so
1915 * for devices that support queuing, otherwise we still have a problem
1916 * with sync vs async workloads.
Jens Axboea68bbddba2008-09-24 13:03:33 +02001917 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001918 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbddba2008-09-24 13:03:33 +02001919 return;
1920
Jens Axboedd67d052006-06-21 09:36:18 +02001921 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001922 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001923
1924 /*
1925 * idle is disabled, either manually or by past process history
1926 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001927 if (!cfq_should_idle(cfqd, cfqq)) {
1928 /* no queue idling. Check for group idling */
1929 if (cfqd->cfq_group_idle)
1930 group_idle = cfqd->cfq_group_idle;
1931 else
1932 return;
1933 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001934
Jens Axboe22e2c502005-06-27 10:55:12 +02001935 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001936 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001937 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001938 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001939 return;
1940
1941 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001942 * task has exited, don't wait
1943 */
Jens Axboe206dc692006-03-28 13:03:44 +02001944 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001945 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001946 return;
1947
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001948 /*
1949 * If our average think time is larger than the remaining time
1950 * slice, then don't idle. This avoids overrunning the allotted
1951 * time slice.
1952 */
1953 if (sample_valid(cic->ttime_samples) &&
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001954 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1955 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1956 cic->ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001957 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001958 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001959
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001960 /* There are other queues in the group, don't do group idle */
1961 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1962 return;
1963
Jens Axboe3b181522005-06-27 10:56:24 +02001964 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001965
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001966 if (group_idle)
1967 sl = cfqd->cfq_group_idle;
1968 else
1969 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001970
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001971 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001972 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001973 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1974 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975}
1976
Jens Axboe498d3aa22007-04-26 12:54:48 +02001977/*
1978 * Move request from internal lists to the request queue dispatch list.
1979 */
Jens Axboe165125e2007-07-24 09:28:11 +02001980static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001982 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001983 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001984
Jens Axboe7b679132008-05-30 12:23:07 +02001985 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1986
Jeff Moyer06d21882009-09-11 17:08:59 +02001987 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001988 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001989 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001990 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001991 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001992
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001993 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001994 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001995 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02001996 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997}
1998
1999/*
2000 * return expired entry, or NULL to just start from scratch in rbtree
2001 */
Jens Axboefebffd62008-01-28 13:19:43 +01002002static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003{
Jens Axboe30996f42009-10-05 11:03:39 +02002004 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005
Jens Axboe3b181522005-06-27 10:56:24 +02002006 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002008
2009 cfq_mark_cfqq_fifo_expire(cfqq);
2010
Jens Axboe89850f72006-07-22 16:48:31 +02002011 if (list_empty(&cfqq->fifo))
2012 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013
Jens Axboe89850f72006-07-22 16:48:31 +02002014 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002015 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002016 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017
Jens Axboe30996f42009-10-05 11:03:39 +02002018 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002019 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020}
2021
Jens Axboe22e2c502005-06-27 10:55:12 +02002022static inline int
2023cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2024{
2025 const int base_rq = cfqd->cfq_slice_async_rq;
2026
2027 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2028
2029 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2030}
2031
2032/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002033 * Must be called with the queue_lock held.
2034 */
2035static int cfqq_process_refs(struct cfq_queue *cfqq)
2036{
2037 int process_refs, io_refs;
2038
2039 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002040 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002041 BUG_ON(process_refs < 0);
2042 return process_refs;
2043}
2044
2045static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2046{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002047 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002048 struct cfq_queue *__cfqq;
2049
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002050 /*
2051 * If there are no process references on the new_cfqq, then it is
2052 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2053 * chain may have dropped their last reference (not just their
2054 * last process reference).
2055 */
2056 if (!cfqq_process_refs(new_cfqq))
2057 return;
2058
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002059 /* Avoid a circular list and skip interim queue merges */
2060 while ((__cfqq = new_cfqq->new_cfqq)) {
2061 if (__cfqq == cfqq)
2062 return;
2063 new_cfqq = __cfqq;
2064 }
2065
2066 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002067 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002068 /*
2069 * If the process for the cfqq has gone away, there is no
2070 * sense in merging the queues.
2071 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002072 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002073 return;
2074
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002075 /*
2076 * Merge in the direction of the lesser amount of work.
2077 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002078 if (new_process_refs >= process_refs) {
2079 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002080 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002081 } else {
2082 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002083 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002084 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002085}
2086
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002087static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002088 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002089{
2090 struct cfq_queue *queue;
2091 int i;
2092 bool key_valid = false;
2093 unsigned long lowest_key = 0;
2094 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2095
Vivek Goyal65b32a52009-12-16 17:52:59 -05002096 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2097 /* select the one with lowest rb_key */
2098 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002099 if (queue &&
2100 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2101 lowest_key = queue->rb_key;
2102 cur_best = i;
2103 key_valid = true;
2104 }
2105 }
2106
2107 return cur_best;
2108}
2109
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002110static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002111{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002112 unsigned slice;
2113 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002114 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002115 unsigned group_slice;
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002116 enum wl_prio_t original_prio = cfqd->serving_prio;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002117
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002118 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002119 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002120 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002121 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002122 cfqd->serving_prio = BE_WORKLOAD;
2123 else {
2124 cfqd->serving_prio = IDLE_WORKLOAD;
2125 cfqd->workload_expires = jiffies + 1;
2126 return;
2127 }
2128
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002129 if (original_prio != cfqd->serving_prio)
2130 goto new_workload;
2131
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002132 /*
2133 * For RT and BE, we have to choose also the type
2134 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2135 * expiration time
2136 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002137 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002138 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002139
2140 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002141 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002142 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002143 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002144 return;
2145
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002146new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002147 /* otherwise select new workload type */
2148 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002149 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2150 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002151 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002152
2153 /*
2154 * the workload slice is computed as a fraction of target latency
2155 * proportional to the number of queues in that workload, over
2156 * all the queues in the same priority class
2157 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002158 group_slice = cfq_group_slice(cfqd, cfqg);
2159
2160 slice = group_slice * count /
2161 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2162 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002163
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002164 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2165 unsigned int tmp;
2166
2167 /*
2168 * Async queues are currently system wide. Just taking
2169 * proportion of queues with-in same group will lead to higher
2170 * async ratio system wide as generally root group is going
2171 * to have higher weight. A more accurate thing would be to
2172 * calculate system wide asnc/sync ratio.
2173 */
2174 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2175 tmp = tmp/cfqd->busy_queues;
2176 slice = min_t(unsigned, slice, tmp);
2177
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002178 /* async workload slice is scaled down according to
2179 * the sync/async slice ratio. */
2180 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002181 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002182 /* sync workload slice is at least 2 * cfq_slice_idle */
2183 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2184
2185 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002186 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002187 cfqd->workload_expires = jiffies + slice;
2188}
2189
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002190static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2191{
2192 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002193 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002194
2195 if (RB_EMPTY_ROOT(&st->rb))
2196 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002197 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002198 update_min_vdisktime(st);
2199 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002200}
2201
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002202static void cfq_choose_cfqg(struct cfq_data *cfqd)
2203{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002204 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2205
2206 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002207
2208 /* Restore the workload type data */
2209 if (cfqg->saved_workload_slice) {
2210 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2211 cfqd->serving_type = cfqg->saved_workload;
2212 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002213 } else
2214 cfqd->workload_expires = jiffies - 1;
2215
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002216 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002217}
2218
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002219/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002220 * Select a queue for service. If we have a current active queue,
2221 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002222 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002223static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002224{
Jens Axboea36e71f2009-04-15 12:15:11 +02002225 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002226
2227 cfqq = cfqd->active_queue;
2228 if (!cfqq)
2229 goto new_queue;
2230
Vivek Goyalf04a6422009-12-03 12:59:40 -05002231 if (!cfqd->rq_queued)
2232 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002233
2234 /*
2235 * We were waiting for group to get backlogged. Expire the queue
2236 */
2237 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2238 goto expire;
2239
Jens Axboe22e2c502005-06-27 10:55:12 +02002240 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002241 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002242 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002243 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2244 /*
2245 * If slice had not expired at the completion of last request
2246 * we might not have turned on wait_busy flag. Don't expire
2247 * the queue yet. Allow the group to get backlogged.
2248 *
2249 * The very fact that we have used the slice, that means we
2250 * have been idling all along on this queue and it should be
2251 * ok to wait for this request to complete.
2252 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002253 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2254 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2255 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002256 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002257 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002258 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002259 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002260
2261 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002262 * The active queue has requests and isn't expired, allow it to
2263 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002264 */
Jens Axboedd67d052006-06-21 09:36:18 +02002265 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002266 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002267
2268 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002269 * If another queue has a request waiting within our mean seek
2270 * distance, let it run. The expire code will check for close
2271 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002272 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002273 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002274 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002275 if (new_cfqq) {
2276 if (!cfqq->new_cfqq)
2277 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002278 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002279 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002280
2281 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002282 * No requests pending. If the active queue still has requests in
2283 * flight or is idling for a new request, allow either of these
2284 * conditions to happen (or time out) before selecting a new queue.
2285 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002286 if (timer_pending(&cfqd->idle_slice_timer)) {
2287 cfqq = NULL;
2288 goto keep_queue;
2289 }
2290
Shaohua Li8e1ac662010-11-08 15:01:04 +01002291 /*
2292 * This is a deep seek queue, but the device is much faster than
2293 * the queue can deliver, don't idle
2294 **/
2295 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2296 (cfq_cfqq_slice_new(cfqq) ||
2297 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2298 cfq_clear_cfqq_deep(cfqq);
2299 cfq_clear_cfqq_idle_window(cfqq);
2300 }
2301
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002302 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2303 cfqq = NULL;
2304 goto keep_queue;
2305 }
2306
2307 /*
2308 * If group idle is enabled and there are requests dispatched from
2309 * this group, wait for requests to complete.
2310 */
2311check_group_idle:
2312 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2313 && cfqq->cfqg->dispatched) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002314 cfqq = NULL;
2315 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002316 }
2317
Jens Axboe3b181522005-06-27 10:56:24 +02002318expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002319 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002320new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002321 /*
2322 * Current queue expired. Check if we have to switch to a new
2323 * service tree
2324 */
2325 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002326 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002327
Jens Axboea36e71f2009-04-15 12:15:11 +02002328 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002329keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002330 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002331}
2332
Jens Axboefebffd62008-01-28 13:19:43 +01002333static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002334{
2335 int dispatched = 0;
2336
2337 while (cfqq->next_rq) {
2338 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2339 dispatched++;
2340 }
2341
2342 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002343
2344 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002345 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002346 return dispatched;
2347}
2348
Jens Axboe498d3aa22007-04-26 12:54:48 +02002349/*
2350 * Drain our current requests. Used for barriers and when switching
2351 * io schedulers on-the-fly.
2352 */
Jens Axboed9e76202007-04-20 14:27:50 +02002353static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002354{
Jens Axboe08717142008-01-28 11:38:15 +01002355 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002356 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002357
Divyesh Shah3440c492010-04-09 09:29:57 +02002358 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002359 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002360 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2361 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002362 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002363 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002364
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002365 BUG_ON(cfqd->busy_queues);
2366
Jeff Moyer6923715a2009-06-12 15:29:30 +02002367 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002368 return dispatched;
2369}
2370
Shaohua Liabc3c742010-03-01 09:20:54 +01002371static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2372 struct cfq_queue *cfqq)
2373{
2374 /* the queue hasn't finished any request, can't estimate */
2375 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01002376 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002377 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2378 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01002379 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002380
Shaohua Lic1e44752010-11-08 15:01:02 +01002381 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01002382}
2383
Jens Axboe0b182d62009-10-06 20:49:37 +02002384static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002385{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002386 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387
Jens Axboe2f5cb732009-04-07 08:51:19 +02002388 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002389 * Drain async requests before we start sync IO
2390 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002391 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002392 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002393
2394 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002395 * If this is an async queue and we have sync IO in flight, let it wait
2396 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002397 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002398 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002399
Shaohua Liabc3c742010-03-01 09:20:54 +01002400 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002401 if (cfq_class_idle(cfqq))
2402 max_dispatch = 1;
2403
2404 /*
2405 * Does this cfqq already have too much IO in flight?
2406 */
2407 if (cfqq->dispatched >= max_dispatch) {
2408 /*
2409 * idle queue must always only have a single IO in flight
2410 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002411 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002412 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002413
Jens Axboe2f5cb732009-04-07 08:51:19 +02002414 /*
2415 * We have other queues, don't allow more IO from this one
2416 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002417 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002418 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002419
Jens Axboe2f5cb732009-04-07 08:51:19 +02002420 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002421 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002422 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002423 if (cfqd->busy_queues == 1)
2424 max_dispatch = -1;
2425 else
2426 /*
2427 * Normally we start throttling cfqq when cfq_quantum/2
2428 * requests have been dispatched. But we can drive
2429 * deeper queue depths at the beginning of slice
2430 * subjected to upper limit of cfq_quantum.
2431 * */
2432 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002433 }
2434
2435 /*
2436 * Async queues must wait a bit before being allowed dispatch.
2437 * We also ramp up the dispatch depth gradually for async IO,
2438 * based on the last sync IO we serviced
2439 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002440 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002441 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002442 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002443
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002444 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002445 if (!depth && !cfqq->dispatched)
2446 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002447 if (depth < max_dispatch)
2448 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 }
2450
Jens Axboe0b182d62009-10-06 20:49:37 +02002451 /*
2452 * If we're below the current max, allow a dispatch
2453 */
2454 return cfqq->dispatched < max_dispatch;
2455}
2456
2457/*
2458 * Dispatch a request from cfqq, moving them to the request queue
2459 * dispatch list.
2460 */
2461static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2462{
2463 struct request *rq;
2464
2465 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2466
2467 if (!cfq_may_dispatch(cfqd, cfqq))
2468 return false;
2469
2470 /*
2471 * follow expired path, else get first next available
2472 */
2473 rq = cfq_check_fifo(cfqq);
2474 if (!rq)
2475 rq = cfqq->next_rq;
2476
2477 /*
2478 * insert request into driver dispatch list
2479 */
2480 cfq_dispatch_insert(cfqd->queue, rq);
2481
2482 if (!cfqd->active_cic) {
2483 struct cfq_io_context *cic = RQ_CIC(rq);
2484
2485 atomic_long_inc(&cic->ioc->refcount);
2486 cfqd->active_cic = cic;
2487 }
2488
2489 return true;
2490}
2491
2492/*
2493 * Find the cfqq that we need to service and move a request from that to the
2494 * dispatch list
2495 */
2496static int cfq_dispatch_requests(struct request_queue *q, int force)
2497{
2498 struct cfq_data *cfqd = q->elevator->elevator_data;
2499 struct cfq_queue *cfqq;
2500
2501 if (!cfqd->busy_queues)
2502 return 0;
2503
2504 if (unlikely(force))
2505 return cfq_forced_dispatch(cfqd);
2506
2507 cfqq = cfq_select_queue(cfqd);
2508 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002509 return 0;
2510
Jens Axboe2f5cb732009-04-07 08:51:19 +02002511 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002512 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002513 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002514 if (!cfq_dispatch_request(cfqd, cfqq))
2515 return 0;
2516
Jens Axboe2f5cb732009-04-07 08:51:19 +02002517 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002518 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002519
2520 /*
2521 * expire an async queue immediately if it has used up its slice. idle
2522 * queue always expire after 1 dispatch round.
2523 */
2524 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2525 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2526 cfq_class_idle(cfqq))) {
2527 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002528 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002529 }
2530
Shan Weib217a902009-09-01 10:06:42 +02002531 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002532 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533}
2534
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535/*
Jens Axboe5e705372006-07-13 12:39:25 +02002536 * task holds one reference to the queue, dropped when task exits. each rq
2537 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002539 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 * queue lock must be held here.
2541 */
2542static void cfq_put_queue(struct cfq_queue *cfqq)
2543{
Jens Axboe22e2c502005-06-27 10:55:12 +02002544 struct cfq_data *cfqd = cfqq->cfqd;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002545 struct cfq_group *cfqg, *orig_cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002546
Shaohua Li30d7b942011-01-07 08:46:59 +01002547 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548
Shaohua Li30d7b942011-01-07 08:46:59 +01002549 cfqq->ref--;
2550 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 return;
2552
Jens Axboe7b679132008-05-30 12:23:07 +02002553 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002555 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002556 cfqg = cfqq->cfqg;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002557 orig_cfqg = cfqq->orig_cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002559 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002560 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002561 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002562 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002563
Vivek Goyalf04a6422009-12-03 12:59:40 -05002564 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002566 cfq_put_cfqg(cfqg);
Vivek Goyal878eadd2009-12-07 19:37:15 +01002567 if (orig_cfqg)
2568 cfq_put_cfqg(orig_cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569}
2570
Jens Axboed6de8be2008-05-28 14:46:59 +02002571/*
2572 * Must always be called with the rcu_read_lock() held
2573 */
Jens Axboe07416d22008-05-07 09:17:12 +02002574static void
2575__call_for_each_cic(struct io_context *ioc,
2576 void (*func)(struct io_context *, struct cfq_io_context *))
2577{
2578 struct cfq_io_context *cic;
2579 struct hlist_node *n;
2580
2581 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2582 func(ioc, cic);
2583}
2584
Jens Axboe4ac845a2008-01-24 08:44:49 +01002585/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002586 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002587 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002588static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002589call_for_each_cic(struct io_context *ioc,
2590 void (*func)(struct io_context *, struct cfq_io_context *))
2591{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002592 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002593 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002594 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002595}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002596
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002597static void cfq_cic_free_rcu(struct rcu_head *head)
2598{
2599 struct cfq_io_context *cic;
2600
2601 cic = container_of(head, struct cfq_io_context, rcu_head);
2602
2603 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002604 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002605
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002606 if (ioc_gone) {
2607 /*
2608 * CFQ scheduler is exiting, grab exit lock and check
2609 * the pending io context count. If it hits zero,
2610 * complete ioc_gone and set it back to NULL
2611 */
2612 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002613 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002614 complete(ioc_gone);
2615 ioc_gone = NULL;
2616 }
2617 spin_unlock(&ioc_gone_lock);
2618 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002619}
2620
2621static void cfq_cic_free(struct cfq_io_context *cic)
2622{
2623 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002624}
2625
2626static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2627{
2628 unsigned long flags;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002629 unsigned long dead_key = (unsigned long) cic->key;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002630
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002631 BUG_ON(!(dead_key & CIC_DEAD_KEY));
Jens Axboe4ac845a2008-01-24 08:44:49 +01002632
2633 spin_lock_irqsave(&ioc->lock, flags);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002634 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
Jens Axboeffc4e752008-02-19 10:02:29 +01002635 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002636 spin_unlock_irqrestore(&ioc->lock, flags);
2637
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002638 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002639}
2640
Jens Axboed6de8be2008-05-28 14:46:59 +02002641/*
2642 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2643 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2644 * and ->trim() which is called with the task lock held
2645 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002646static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002648 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002649 * ioc->refcount is zero here, or we are called from elv_unregister(),
2650 * so no more cic's are allowed to be linked into this ioc. So it
2651 * should be ok to iterate over the known list, we will see all cic's
2652 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002653 */
Jens Axboe07416d22008-05-07 09:17:12 +02002654 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655}
2656
Shaohua Lid02a2c02010-05-25 10:16:53 +02002657static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002658{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002659 struct cfq_queue *__cfqq, *next;
2660
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002661 /*
2662 * If this queue was scheduled to merge with another queue, be
2663 * sure to drop the reference taken on that queue (and others in
2664 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2665 */
2666 __cfqq = cfqq->new_cfqq;
2667 while (__cfqq) {
2668 if (__cfqq == cfqq) {
2669 WARN(1, "cfqq->new_cfqq loop detected\n");
2670 break;
2671 }
2672 next = __cfqq->new_cfqq;
2673 cfq_put_queue(__cfqq);
2674 __cfqq = next;
2675 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002676}
2677
2678static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2679{
2680 if (unlikely(cfqq == cfqd->active_queue)) {
2681 __cfq_slice_expired(cfqd, cfqq, 0);
2682 cfq_schedule_dispatch(cfqd);
2683 }
2684
2685 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002686
Jens Axboe89850f72006-07-22 16:48:31 +02002687 cfq_put_queue(cfqq);
2688}
2689
2690static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2691 struct cfq_io_context *cic)
2692{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002693 struct io_context *ioc = cic->ioc;
2694
Jens Axboefc463792006-08-29 09:05:44 +02002695 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002696
2697 /*
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002698 * Make sure dead mark is seen for dead queues
Jens Axboe4ac845a2008-01-24 08:44:49 +01002699 */
Jens Axboefc463792006-08-29 09:05:44 +02002700 smp_wmb();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002701 cic->key = cfqd_dead_key(cfqd);
Jens Axboefc463792006-08-29 09:05:44 +02002702
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002703 if (ioc->ioc_data == cic)
2704 rcu_assign_pointer(ioc->ioc_data, NULL);
2705
Jens Axboeff6657c2009-04-08 10:58:57 +02002706 if (cic->cfqq[BLK_RW_ASYNC]) {
2707 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2708 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002709 }
2710
Jens Axboeff6657c2009-04-08 10:58:57 +02002711 if (cic->cfqq[BLK_RW_SYNC]) {
2712 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2713 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002714 }
Jens Axboe89850f72006-07-22 16:48:31 +02002715}
2716
Jens Axboe4ac845a2008-01-24 08:44:49 +01002717static void cfq_exit_single_io_context(struct io_context *ioc,
2718 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002719{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002720 struct cfq_data *cfqd = cic_to_cfqd(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002721
Jens Axboe89850f72006-07-22 16:48:31 +02002722 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002723 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002724 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002725
Jens Axboe4ac845a2008-01-24 08:44:49 +01002726 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002727
2728 /*
2729 * Ensure we get a fresh copy of the ->key to prevent
2730 * race between exiting task and queue
2731 */
2732 smp_read_barrier_depends();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002733 if (cic->key == cfqd)
Jens Axboe62c1fe92008-12-15 21:19:25 +01002734 __cfq_exit_single_io_context(cfqd, cic);
2735
Jens Axboe4ac845a2008-01-24 08:44:49 +01002736 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002737 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002738}
2739
Jens Axboe498d3aa22007-04-26 12:54:48 +02002740/*
2741 * The process that ioc belongs to has exited, we need to clean up
2742 * and put the internal structures we have that belongs to that process.
2743 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002744static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002746 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747}
2748
Jens Axboe22e2c502005-06-27 10:55:12 +02002749static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002750cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751{
Jens Axboeb5deef92006-07-19 23:39:40 +02002752 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753
Christoph Lameter94f60302007-07-17 04:03:29 -07002754 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2755 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002757 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002758 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002759 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002760 cic->dtor = cfq_free_io_context;
2761 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002762 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 }
2764
2765 return cic;
2766}
2767
Jens Axboefd0928d2008-01-24 08:52:45 +01002768static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002769{
2770 struct task_struct *tsk = current;
2771 int ioprio_class;
2772
Jens Axboe3b181522005-06-27 10:56:24 +02002773 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002774 return;
2775
Jens Axboefd0928d2008-01-24 08:52:45 +01002776 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002777 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002778 default:
2779 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2780 case IOPRIO_CLASS_NONE:
2781 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002782 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002783 */
2784 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002785 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002786 break;
2787 case IOPRIO_CLASS_RT:
2788 cfqq->ioprio = task_ioprio(ioc);
2789 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2790 break;
2791 case IOPRIO_CLASS_BE:
2792 cfqq->ioprio = task_ioprio(ioc);
2793 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2794 break;
2795 case IOPRIO_CLASS_IDLE:
2796 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2797 cfqq->ioprio = 7;
2798 cfq_clear_cfqq_idle_window(cfqq);
2799 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002800 }
2801
2802 /*
2803 * keep track of original prio settings in case we have to temporarily
2804 * elevate the priority of this queue
2805 */
2806 cfqq->org_ioprio = cfqq->ioprio;
2807 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002808 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002809}
2810
Jens Axboefebffd62008-01-28 13:19:43 +01002811static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002812{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002813 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002814 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002815 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002816
Jens Axboecaaa5f92006-06-16 11:23:00 +02002817 if (unlikely(!cfqd))
2818 return;
2819
Jens Axboec1b707d2006-10-30 19:54:23 +01002820 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002821
Jens Axboeff6657c2009-04-08 10:58:57 +02002822 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002823 if (cfqq) {
2824 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002825 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2826 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002827 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002828 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002829 cfq_put_queue(cfqq);
2830 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002831 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002832
Jens Axboeff6657c2009-04-08 10:58:57 +02002833 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002834 if (cfqq)
2835 cfq_mark_cfqq_prio_changed(cfqq);
2836
Jens Axboec1b707d2006-10-30 19:54:23 +01002837 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002838}
2839
Jens Axboefc463792006-08-29 09:05:44 +02002840static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002842 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002843 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844}
2845
Jens Axboed5036d72009-06-26 10:44:34 +02002846static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002847 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002848{
2849 RB_CLEAR_NODE(&cfqq->rb_node);
2850 RB_CLEAR_NODE(&cfqq->p_node);
2851 INIT_LIST_HEAD(&cfqq->fifo);
2852
Shaohua Li30d7b942011-01-07 08:46:59 +01002853 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02002854 cfqq->cfqd = cfqd;
2855
2856 cfq_mark_cfqq_prio_changed(cfqq);
2857
2858 if (is_sync) {
2859 if (!cfq_class_idle(cfqq))
2860 cfq_mark_cfqq_idle_window(cfqq);
2861 cfq_mark_cfqq_sync(cfqq);
2862 }
2863 cfqq->pid = pid;
2864}
2865
Vivek Goyal246103332009-12-03 12:59:51 -05002866#ifdef CONFIG_CFQ_GROUP_IOSCHED
2867static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2868{
2869 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002870 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal246103332009-12-03 12:59:51 -05002871 unsigned long flags;
2872 struct request_queue *q;
2873
2874 if (unlikely(!cfqd))
2875 return;
2876
2877 q = cfqd->queue;
2878
2879 spin_lock_irqsave(q->queue_lock, flags);
2880
2881 if (sync_cfqq) {
2882 /*
2883 * Drop reference to sync queue. A new sync queue will be
2884 * assigned in new group upon arrival of a fresh request.
2885 */
2886 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2887 cic_set_cfqq(cic, NULL, 1);
2888 cfq_put_queue(sync_cfqq);
2889 }
2890
2891 spin_unlock_irqrestore(q->queue_lock, flags);
2892}
2893
2894static void cfq_ioc_set_cgroup(struct io_context *ioc)
2895{
2896 call_for_each_cic(ioc, changed_cgroup);
2897 ioc->cgroup_changed = 0;
2898}
2899#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2900
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002902cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002903 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002906 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002907 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908
2909retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002910 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002911 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002912 /* cic always exists here */
2913 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914
Jens Axboe6118b702009-06-30 09:34:12 +02002915 /*
2916 * Always try a new alloc if we fell back to the OOM cfqq
2917 * originally, since it should just be a temporary situation.
2918 */
2919 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2920 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 if (new_cfqq) {
2922 cfqq = new_cfqq;
2923 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002924 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002926 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002927 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002928 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002930 if (new_cfqq)
2931 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002932 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002933 cfqq = kmem_cache_alloc_node(cfq_pool,
2934 gfp_mask | __GFP_ZERO,
2935 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002936 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937
Jens Axboe6118b702009-06-30 09:34:12 +02002938 if (cfqq) {
2939 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2940 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002941 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002942 cfq_log_cfqq(cfqd, cfqq, "alloced");
2943 } else
2944 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 }
2946
2947 if (new_cfqq)
2948 kmem_cache_free(cfq_pool, new_cfqq);
2949
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 return cfqq;
2951}
2952
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002953static struct cfq_queue **
2954cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2955{
Jens Axboefe094d92008-01-31 13:08:54 +01002956 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002957 case IOPRIO_CLASS_RT:
2958 return &cfqd->async_cfqq[0][ioprio];
2959 case IOPRIO_CLASS_BE:
2960 return &cfqd->async_cfqq[1][ioprio];
2961 case IOPRIO_CLASS_IDLE:
2962 return &cfqd->async_idle_cfqq;
2963 default:
2964 BUG();
2965 }
2966}
2967
Jens Axboe15c31be2007-07-10 13:43:25 +02002968static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002969cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002970 gfp_t gfp_mask)
2971{
Jens Axboefd0928d2008-01-24 08:52:45 +01002972 const int ioprio = task_ioprio(ioc);
2973 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002974 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002975 struct cfq_queue *cfqq = NULL;
2976
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002977 if (!is_sync) {
2978 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2979 cfqq = *async_cfqq;
2980 }
2981
Jens Axboe6118b702009-06-30 09:34:12 +02002982 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002983 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002984
2985 /*
2986 * pin the queue now that it's allocated, scheduler exit will prune it
2987 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002988 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01002989 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002990 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002991 }
2992
Shaohua Li30d7b942011-01-07 08:46:59 +01002993 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02002994 return cfqq;
2995}
2996
Jens Axboe498d3aa22007-04-26 12:54:48 +02002997/*
2998 * We drop cfq io contexts lazily, so we may find a dead one.
2999 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003000static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01003001cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3002 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003003{
Jens Axboe4ac845a2008-01-24 08:44:49 +01003004 unsigned long flags;
3005
Jens Axboefc463792006-08-29 09:05:44 +02003006 WARN_ON(!list_empty(&cic->queue_list));
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003007 BUG_ON(cic->key != cfqd_dead_key(cfqd));
Jens Axboe597bc482007-04-24 21:23:53 +02003008
Jens Axboe4ac845a2008-01-24 08:44:49 +01003009 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02003010
Fabio Checconi4faa3c82008-04-10 08:28:01 +02003011 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003012
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003013 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
Jens Axboeffc4e752008-02-19 10:02:29 +01003014 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003015 spin_unlock_irqrestore(&ioc->lock, flags);
3016
3017 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003018}
3019
Jens Axboee2d74ac2006-03-28 08:59:01 +02003020static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01003021cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003022{
Jens Axboee2d74ac2006-03-28 08:59:01 +02003023 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003024 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003025
Vasily Tarasov91fac312007-04-25 12:29:51 +02003026 if (unlikely(!ioc))
3027 return NULL;
3028
Jens Axboed6de8be2008-05-28 14:46:59 +02003029 rcu_read_lock();
3030
Jens Axboe597bc482007-04-24 21:23:53 +02003031 /*
3032 * we maintain a last-hit cache, to avoid browsing over the tree
3033 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003034 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02003035 if (cic && cic->key == cfqd) {
3036 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02003037 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003038 }
Jens Axboe597bc482007-04-24 21:23:53 +02003039
Jens Axboe4ac845a2008-01-24 08:44:49 +01003040 do {
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003041 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003042 rcu_read_unlock();
3043 if (!cic)
3044 break;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003045 if (unlikely(cic->key != cfqd)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01003046 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003047 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01003048 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003049 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003050
Jens Axboed6de8be2008-05-28 14:46:59 +02003051 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003052 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003053 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003054 break;
3055 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003056
Jens Axboe4ac845a2008-01-24 08:44:49 +01003057 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003058}
3059
Jens Axboe4ac845a2008-01-24 08:44:49 +01003060/*
3061 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3062 * the process specific cfq io context when entered from the block layer.
3063 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3064 */
Jens Axboefebffd62008-01-28 13:19:43 +01003065static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3066 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003067{
Jens Axboe0261d682006-10-30 19:07:48 +01003068 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003069 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003070
Jens Axboe4ac845a2008-01-24 08:44:49 +01003071 ret = radix_tree_preload(gfp_mask);
3072 if (!ret) {
3073 cic->ioc = ioc;
3074 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003075
Jens Axboe4ac845a2008-01-24 08:44:49 +01003076 spin_lock_irqsave(&ioc->lock, flags);
3077 ret = radix_tree_insert(&ioc->radix_root,
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003078 cfqd->cic_index, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01003079 if (!ret)
3080 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003081 spin_unlock_irqrestore(&ioc->lock, flags);
3082
3083 radix_tree_preload_end();
3084
3085 if (!ret) {
3086 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3087 list_add(&cic->queue_list, &cfqd->cic_list);
3088 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003089 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003090 }
3091
Jens Axboe4ac845a2008-01-24 08:44:49 +01003092 if (ret)
3093 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02003094
Jens Axboe4ac845a2008-01-24 08:44:49 +01003095 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003096}
3097
Jens Axboe22e2c502005-06-27 10:55:12 +02003098/*
3099 * Setup general io context and cfq io context. There can be several cfq
3100 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02003101 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02003102 */
3103static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02003104cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105{
Jens Axboe22e2c502005-06-27 10:55:12 +02003106 struct io_context *ioc = NULL;
3107 struct cfq_io_context *cic;
3108
3109 might_sleep_if(gfp_mask & __GFP_WAIT);
3110
Jens Axboeb5deef92006-07-19 23:39:40 +02003111 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02003112 if (!ioc)
3113 return NULL;
3114
Jens Axboe4ac845a2008-01-24 08:44:49 +01003115 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003116 if (cic)
3117 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02003118
Jens Axboee2d74ac2006-03-28 08:59:01 +02003119 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3120 if (cic == NULL)
3121 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02003122
Jens Axboe4ac845a2008-01-24 08:44:49 +01003123 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3124 goto err_free;
3125
Jens Axboe22e2c502005-06-27 10:55:12 +02003126out:
Jens Axboefc463792006-08-29 09:05:44 +02003127 smp_read_barrier_depends();
3128 if (unlikely(ioc->ioprio_changed))
3129 cfq_ioc_set_ioprio(ioc);
3130
Vivek Goyal246103332009-12-03 12:59:51 -05003131#ifdef CONFIG_CFQ_GROUP_IOSCHED
3132 if (unlikely(ioc->cgroup_changed))
3133 cfq_ioc_set_cgroup(ioc);
3134#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003135 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003136err_free:
3137 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003138err:
3139 put_io_context(ioc);
3140 return NULL;
3141}
3142
3143static void
3144cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3145{
Jens Axboeaaf12282007-01-19 11:30:16 +11003146 unsigned long elapsed = jiffies - cic->last_end_request;
3147 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003148
3149 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3150 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3151 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3152}
3153
Jens Axboe206dc692006-03-28 13:03:44 +02003154static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003155cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003156 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003157{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003158 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003159 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003160 if (cfqq->last_request_pos) {
3161 if (cfqq->last_request_pos < blk_rq_pos(rq))
3162 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3163 else
3164 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3165 }
Jens Axboe206dc692006-03-28 13:03:44 +02003166
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003167 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003168 if (blk_queue_nonrot(cfqd->queue))
3169 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3170 else
3171 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003172}
Jens Axboe22e2c502005-06-27 10:55:12 +02003173
3174/*
3175 * Disable idle window if the process thinks too long or seeks so much that
3176 * it doesn't matter
3177 */
3178static void
3179cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3180 struct cfq_io_context *cic)
3181{
Jens Axboe7b679132008-05-30 12:23:07 +02003182 int old_idle, enable_idle;
Jens Axboe1be92f2f2007-04-19 14:32:26 +02003183
Jens Axboe08717142008-01-28 11:38:15 +01003184 /*
3185 * Don't idle for async or idle io prio class
3186 */
3187 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f2f2007-04-19 14:32:26 +02003188 return;
3189
Jens Axboec265a7f2008-06-26 13:49:33 +02003190 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003191
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003192 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3193 cfq_mark_cfqq_deep(cfqq);
3194
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003195 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3196 enable_idle = 0;
3197 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003198 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003199 enable_idle = 0;
3200 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003201 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003202 enable_idle = 0;
3203 else
3204 enable_idle = 1;
3205 }
3206
Jens Axboe7b679132008-05-30 12:23:07 +02003207 if (old_idle != enable_idle) {
3208 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3209 if (enable_idle)
3210 cfq_mark_cfqq_idle_window(cfqq);
3211 else
3212 cfq_clear_cfqq_idle_window(cfqq);
3213 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003214}
3215
Jens Axboe22e2c502005-06-27 10:55:12 +02003216/*
3217 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3218 * no or if we aren't sure, a 1 will cause a preempt.
3219 */
Jens Axboea6151c32009-10-07 20:02:57 +02003220static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003221cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003222 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003223{
Jens Axboe6d048f52007-04-25 12:44:27 +02003224 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003225
Jens Axboe6d048f52007-04-25 12:44:27 +02003226 cfqq = cfqd->active_queue;
3227 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003228 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003229
Jens Axboe6d048f52007-04-25 12:44:27 +02003230 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003231 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003232
3233 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003234 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003235
Jens Axboe22e2c502005-06-27 10:55:12 +02003236 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003237 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3238 */
3239 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3240 return false;
3241
3242 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003243 * if the new request is sync, but the currently running queue is
3244 * not, let the sync request have priority.
3245 */
Jens Axboe5e705372006-07-13 12:39:25 +02003246 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003247 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003248
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003249 if (new_cfqq->cfqg != cfqq->cfqg)
3250 return false;
3251
3252 if (cfq_slice_used(cfqq))
3253 return true;
3254
3255 /* Allow preemption only if we are idling on sync-noidle tree */
3256 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3257 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3258 new_cfqq->service_tree->count == 2 &&
3259 RB_EMPTY_ROOT(&cfqq->sort_list))
3260 return true;
3261
Jens Axboe374f84a2006-07-23 01:42:19 +02003262 /*
3263 * So both queues are sync. Let the new request get disk time if
3264 * it's a metadata request and the current queue is doing regular IO.
3265 */
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003266 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003267 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003268
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003269 /*
3270 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3271 */
3272 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003273 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003274
Shaohua Lid2d59e12010-11-08 15:01:03 +01003275 /* An idle queue should not be idle now for some reason */
3276 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3277 return true;
3278
Jens Axboe1e3335d2007-02-14 19:59:49 +01003279 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003280 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003281
3282 /*
3283 * if this request is as-good as one we would expect from the
3284 * current cfqq, let it preempt
3285 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003286 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003287 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003288
Jens Axboea6151c32009-10-07 20:02:57 +02003289 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003290}
3291
3292/*
3293 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3294 * let it have half of its nominal slice.
3295 */
3296static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3297{
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003298 struct cfq_queue *old_cfqq = cfqd->active_queue;
3299
Jens Axboe7b679132008-05-30 12:23:07 +02003300 cfq_log_cfqq(cfqd, cfqq, "preempt");
Vivek Goyale5ff0822010-04-26 19:25:11 +02003301 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003302
Jens Axboebf572252006-07-19 20:29:12 +02003303 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003304 * workload type is changed, don't save slice, otherwise preempt
3305 * doesn't happen
3306 */
3307 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3308 cfqq->cfqg->saved_workload_slice = 0;
3309
3310 /*
Jens Axboebf572252006-07-19 20:29:12 +02003311 * Put the new queue at the front of the of the current list,
3312 * so we know that it will be selected next.
3313 */
3314 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003315
3316 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02003317
Jens Axboe44f7c162007-01-19 11:51:58 +11003318 cfqq->slice_end = 0;
3319 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003320}
3321
3322/*
Jens Axboe5e705372006-07-13 12:39:25 +02003323 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003324 * something we should do about it
3325 */
3326static void
Jens Axboe5e705372006-07-13 12:39:25 +02003327cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3328 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003329{
Jens Axboe5e705372006-07-13 12:39:25 +02003330 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003331
Aaron Carroll45333d52008-08-26 15:52:36 +02003332 cfqd->rq_queued++;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003333 if (rq->cmd_flags & REQ_META)
Jens Axboe374f84a2006-07-23 01:42:19 +02003334 cfqq->meta_pending++;
3335
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003336 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003337 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003338 cfq_update_idle_window(cfqd, cfqq, cic);
3339
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003340 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003341
3342 if (cfqq == cfqd->active_queue) {
3343 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003344 * Remember that we saw a request from this process, but
3345 * don't start queuing just yet. Otherwise we risk seeing lots
3346 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003347 * and merging. If the request is already larger than a single
3348 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003349 * merging is already done. Ditto for a busy system that
3350 * has other work pending, don't risk delaying until the
3351 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003352 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003353 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003354 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3355 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003356 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003357 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003358 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003359 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003360 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003361 &cfqq->cfqg->blkg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003362 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003363 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003364 }
Jens Axboe5e705372006-07-13 12:39:25 +02003365 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003366 /*
3367 * not the active queue - expire current slice if it is
3368 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003369 * has some old slice time left and is of higher priority or
3370 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003371 */
3372 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09003373 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003374 }
3375}
3376
Jens Axboe165125e2007-07-24 09:28:11 +02003377static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003378{
Jens Axboeb4878f22005-10-20 16:42:29 +02003379 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003380 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003381
Jens Axboe7b679132008-05-30 12:23:07 +02003382 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003383 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384
Jens Axboe30996f42009-10-05 11:03:39 +02003385 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003386 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003387 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003388 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003389 &cfqd->serving_group->blkg, rq_data_dir(rq),
3390 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003391 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392}
3393
Aaron Carroll45333d52008-08-26 15:52:36 +02003394/*
3395 * Update hw_tag based on peak queue depth over 50 samples under
3396 * sufficient load.
3397 */
3398static void cfq_update_hw_tag(struct cfq_data *cfqd)
3399{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003400 struct cfq_queue *cfqq = cfqd->active_queue;
3401
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003402 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3403 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003404
3405 if (cfqd->hw_tag == 1)
3406 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003407
3408 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003409 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003410 return;
3411
Shaohua Li1a1238a2009-10-27 08:46:23 +01003412 /*
3413 * If active queue hasn't enough requests and can idle, cfq might not
3414 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3415 * case
3416 */
3417 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3418 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003419 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003420 return;
3421
Aaron Carroll45333d52008-08-26 15:52:36 +02003422 if (cfqd->hw_tag_samples++ < 50)
3423 return;
3424
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003425 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003426 cfqd->hw_tag = 1;
3427 else
3428 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003429}
3430
Vivek Goyal7667aa02009-12-08 17:52:58 -05003431static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3432{
3433 struct cfq_io_context *cic = cfqd->active_cic;
3434
Justin TerAvest02a8f012011-02-09 14:20:03 +01003435 /* If the queue already has requests, don't wait */
3436 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3437 return false;
3438
Vivek Goyal7667aa02009-12-08 17:52:58 -05003439 /* If there are other queues in the group, don't wait */
3440 if (cfqq->cfqg->nr_cfqq > 1)
3441 return false;
3442
3443 if (cfq_slice_used(cfqq))
3444 return true;
3445
3446 /* if slice left is less than think time, wait busy */
3447 if (cic && sample_valid(cic->ttime_samples)
3448 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3449 return true;
3450
3451 /*
3452 * If think times is less than a jiffy than ttime_mean=0 and above
3453 * will not be true. It might happen that slice has not expired yet
3454 * but will expire soon (4-5 ns) during select_queue(). To cover the
3455 * case where think time is less than a jiffy, mark the queue wait
3456 * busy if only 1 jiffy is left in the slice.
3457 */
3458 if (cfqq->slice_end - jiffies == 1)
3459 return true;
3460
3461 return false;
3462}
3463
Jens Axboe165125e2007-07-24 09:28:11 +02003464static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465{
Jens Axboe5e705372006-07-13 12:39:25 +02003466 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003467 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003468 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003469 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470
Jens Axboeb4878f22005-10-20 16:42:29 +02003471 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003472 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3473 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474
Aaron Carroll45333d52008-08-26 15:52:36 +02003475 cfq_update_hw_tag(cfqd);
3476
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003477 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003478 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003479 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003480 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003481 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003482 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3483 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3484 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003486 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003487
Vivek Goyal365722b2009-10-03 15:21:27 +02003488 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003489 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003490 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3491 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003492 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003493
3494 /*
3495 * If this is the active queue, check if it needs to be expired,
3496 * or if we want to idle in case it has no pending requests.
3497 */
3498 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003499 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3500
Jens Axboe44f7c162007-01-19 11:51:58 +11003501 if (cfq_cfqq_slice_new(cfqq)) {
3502 cfq_set_prio_slice(cfqd, cfqq);
3503 cfq_clear_cfqq_slice_new(cfqq);
3504 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003505
3506 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003507 * Should we wait for next request to come in before we expire
3508 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003509 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003510 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003511 unsigned long extend_sl = cfqd->cfq_slice_idle;
3512 if (!cfqd->cfq_slice_idle)
3513 extend_sl = cfqd->cfq_group_idle;
3514 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003515 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003516 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003517 }
3518
Jens Axboea36e71f2009-04-15 12:15:11 +02003519 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003520 * Idling is not enabled on:
3521 * - expired queues
3522 * - idle-priority queues
3523 * - async queues
3524 * - queues with still some requests queued
3525 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003526 */
Jens Axboe08717142008-01-28 11:38:15 +01003527 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003528 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003529 else if (sync && cfqq_empty &&
3530 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003531 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003532 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003533 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003534
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003535 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003536 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537}
3538
Jens Axboe22e2c502005-06-27 10:55:12 +02003539/*
3540 * we temporarily boost lower priority queues if they are holding fs exclusive
3541 * resources. they are boosted to normal prio (CLASS_BE/4)
3542 */
3543static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544{
Jens Axboe22e2c502005-06-27 10:55:12 +02003545 if (has_fs_excl()) {
3546 /*
3547 * boost idle prio on transactions that would lock out other
3548 * users of the filesystem
3549 */
3550 if (cfq_class_idle(cfqq))
3551 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3552 if (cfqq->ioprio > IOPRIO_NORM)
3553 cfqq->ioprio = IOPRIO_NORM;
3554 } else {
3555 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003556 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003557 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003558 cfqq->ioprio_class = cfqq->org_ioprio_class;
3559 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003561}
3562
Jens Axboe89850f72006-07-22 16:48:31 +02003563static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003564{
Jens Axboe1b379d82009-08-11 08:26:11 +02003565 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003566 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003567 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003568 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003569
3570 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003571}
3572
Jens Axboe165125e2007-07-24 09:28:11 +02003573static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003574{
3575 struct cfq_data *cfqd = q->elevator->elevator_data;
3576 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003577 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003578 struct cfq_queue *cfqq;
3579
3580 /*
3581 * don't force setup of a queue from here, as a call to may_queue
3582 * does not necessarily imply that a request actually will be queued.
3583 * so just lookup a possibly existing queue, or return 'may queue'
3584 * if that fails
3585 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003586 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003587 if (!cic)
3588 return ELV_MQUEUE_MAY;
3589
Jens Axboeb0b78f82009-04-08 10:56:08 +02003590 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003591 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003592 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003593 cfq_prio_boost(cfqq);
3594
Jens Axboe89850f72006-07-22 16:48:31 +02003595 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003596 }
3597
3598 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599}
3600
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601/*
3602 * queue lock held here
3603 */
Jens Axboebb37b942006-12-01 10:42:33 +01003604static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605{
Jens Axboe5e705372006-07-13 12:39:25 +02003606 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607
Jens Axboe5e705372006-07-13 12:39:25 +02003608 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003609 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610
Jens Axboe22e2c502005-06-27 10:55:12 +02003611 BUG_ON(!cfqq->allocated[rw]);
3612 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003613
Jens Axboe5e705372006-07-13 12:39:25 +02003614 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615
Mike Snitzerc1867942011-02-11 11:08:00 +01003616 rq->elevator_private[0] = NULL;
3617 rq->elevator_private[1] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003619 /* Put down rq reference on cfqg */
3620 cfq_put_cfqg(RQ_CFQG(rq));
Mike Snitzerc1867942011-02-11 11:08:00 +01003621 rq->elevator_private[2] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003622
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623 cfq_put_queue(cfqq);
3624 }
3625}
3626
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003627static struct cfq_queue *
3628cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3629 struct cfq_queue *cfqq)
3630{
3631 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3632 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003633 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003634 cfq_put_queue(cfqq);
3635 return cic_to_cfqq(cic, 1);
3636}
3637
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003638/*
3639 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3640 * was the last process referring to said cfqq.
3641 */
3642static struct cfq_queue *
3643split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3644{
3645 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003646 cfqq->pid = current->pid;
3647 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003648 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003649 return cfqq;
3650 }
3651
3652 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003653
3654 cfq_put_cooperator(cfqq);
3655
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003656 cfq_put_queue(cfqq);
3657 return NULL;
3658}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003660 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003661 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003662static int
Jens Axboe165125e2007-07-24 09:28:11 +02003663cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664{
3665 struct cfq_data *cfqd = q->elevator->elevator_data;
3666 struct cfq_io_context *cic;
3667 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003668 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003669 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 unsigned long flags;
3671
3672 might_sleep_if(gfp_mask & __GFP_WAIT);
3673
Jens Axboee2d74ac2006-03-28 08:59:01 +02003674 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003675
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676 spin_lock_irqsave(q->queue_lock, flags);
3677
Jens Axboe22e2c502005-06-27 10:55:12 +02003678 if (!cic)
3679 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003681new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003682 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003683 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003684 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003685 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003686 } else {
3687 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003688 * If the queue was seeky for too long, break it apart.
3689 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003690 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003691 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3692 cfqq = split_cfqq(cic, cfqq);
3693 if (!cfqq)
3694 goto new_queue;
3695 }
3696
3697 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003698 * Check to see if this queue is scheduled to merge with
3699 * another, closely cooperating queue. The merging of
3700 * queues happens here as it must be done in process context.
3701 * The reference on new_cfqq was taken in merge_cfqqs.
3702 */
3703 if (cfqq->new_cfqq)
3704 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003705 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706
3707 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02003708
Linus Torvalds1da177e2005-04-16 15:20:36 -07003709 spin_unlock_irqrestore(q->queue_lock, flags);
3710
Jens Axboe6fae9c22011-03-01 15:04:39 -05003711 cfqq->ref++;
Mike Snitzerc1867942011-02-11 11:08:00 +01003712 rq->elevator_private[0] = cic;
3713 rq->elevator_private[1] = cfqq;
3714 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
Jens Axboe5e705372006-07-13 12:39:25 +02003715 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003716
Jens Axboe22e2c502005-06-27 10:55:12 +02003717queue_fail:
3718 if (cic)
3719 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003720
Jens Axboe23e018a2009-10-05 08:52:35 +02003721 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003723 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724 return 1;
3725}
3726
David Howells65f27f32006-11-22 14:55:48 +00003727static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003728{
David Howells65f27f32006-11-22 14:55:48 +00003729 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003730 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003731 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003732
Jens Axboe40bb54d2009-04-15 12:11:10 +02003733 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09003734 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003735 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003736}
3737
3738/*
3739 * Timer running if the active_queue is currently idling inside its time slice
3740 */
3741static void cfq_idle_slice_timer(unsigned long data)
3742{
3743 struct cfq_data *cfqd = (struct cfq_data *) data;
3744 struct cfq_queue *cfqq;
3745 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003746 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003747
Jens Axboe7b679132008-05-30 12:23:07 +02003748 cfq_log(cfqd, "idle timer fired");
3749
Jens Axboe22e2c502005-06-27 10:55:12 +02003750 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3751
Jens Axboefe094d92008-01-31 13:08:54 +01003752 cfqq = cfqd->active_queue;
3753 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003754 timed_out = 0;
3755
Jens Axboe22e2c502005-06-27 10:55:12 +02003756 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003757 * We saw a request before the queue expired, let it through
3758 */
3759 if (cfq_cfqq_must_dispatch(cfqq))
3760 goto out_kick;
3761
3762 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003763 * expired
3764 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003765 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003766 goto expire;
3767
3768 /*
3769 * only expire and reinvoke request handler, if there are
3770 * other queues with pending requests
3771 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003772 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003773 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003774
3775 /*
3776 * not expired and it has a request pending, let it dispatch
3777 */
Jens Axboe75e50982009-04-07 08:56:14 +02003778 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003779 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003780
3781 /*
3782 * Queue depth flag is reset only when the idle didn't succeed
3783 */
3784 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003785 }
3786expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003787 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003788out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003789 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003790out_cont:
3791 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3792}
3793
Jens Axboe3b181522005-06-27 10:56:24 +02003794static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3795{
3796 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003797 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003798}
Jens Axboe22e2c502005-06-27 10:55:12 +02003799
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003800static void cfq_put_async_queues(struct cfq_data *cfqd)
3801{
3802 int i;
3803
3804 for (i = 0; i < IOPRIO_BE_NR; i++) {
3805 if (cfqd->async_cfqq[0][i])
3806 cfq_put_queue(cfqd->async_cfqq[0][i]);
3807 if (cfqd->async_cfqq[1][i])
3808 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003809 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003810
3811 if (cfqd->async_idle_cfqq)
3812 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003813}
3814
Jens Axboebb729bc2009-12-06 09:54:19 +01003815static void cfq_cfqd_free(struct rcu_head *head)
3816{
3817 kfree(container_of(head, struct cfq_data, rcu));
3818}
3819
Jens Axboeb374d182008-10-31 10:05:07 +01003820static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821{
Jens Axboe22e2c502005-06-27 10:55:12 +02003822 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003823 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003824
Jens Axboe3b181522005-06-27 10:56:24 +02003825 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003826
Al Virod9ff4182006-03-18 13:51:22 -05003827 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003828
Al Virod9ff4182006-03-18 13:51:22 -05003829 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003830 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003831
3832 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003833 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3834 struct cfq_io_context,
3835 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003836
3837 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003838 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003839
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003840 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003841 cfq_release_cfq_groups(cfqd);
Vivek Goyale98ef892010-06-18 10:39:47 -04003842 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003843
Al Virod9ff4182006-03-18 13:51:22 -05003844 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003845
3846 cfq_shutdown_timer_wq(cfqd);
3847
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003848 spin_lock(&cic_index_lock);
3849 ida_remove(&cic_index_ida, cfqd->cic_index);
3850 spin_unlock(&cic_index_lock);
3851
Vivek Goyalb1c35762009-12-03 12:59:47 -05003852 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003853 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003854}
3855
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003856static int cfq_alloc_cic_index(void)
3857{
3858 int index, error;
3859
3860 do {
3861 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3862 return -ENOMEM;
3863
3864 spin_lock(&cic_index_lock);
3865 error = ida_get_new(&cic_index_ida, &index);
3866 spin_unlock(&cic_index_lock);
3867 if (error && error != -EAGAIN)
3868 return error;
3869 } while (error);
3870
3871 return index;
3872}
3873
Jens Axboe165125e2007-07-24 09:28:11 +02003874static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875{
3876 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003877 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003878 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003879 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003881 i = cfq_alloc_cic_index();
3882 if (i < 0)
3883 return NULL;
3884
Christoph Lameter94f60302007-07-17 04:03:29 -07003885 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003886 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003887 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888
Shaohua Li30d7b942011-01-07 08:46:59 +01003889 /*
3890 * Don't need take queue_lock in the routine, since we are
3891 * initializing the ioscheduler, and nobody is using cfqd
3892 */
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003893 cfqd->cic_index = i;
3894
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003895 /* Init root service tree */
3896 cfqd->grp_service_tree = CFQ_RB_ROOT;
3897
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003898 /* Init root group */
3899 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003900 for_each_cfqg_st(cfqg, i, j, st)
3901 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003902 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003903
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003904 /* Give preference to root group over other groups */
3905 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3906
Vivek Goyal25fb5162009-12-03 12:59:46 -05003907#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003908 /*
3909 * Take a reference to root group which we never drop. This is just
3910 * to make sure that cfq_put_cfqg() does not try to kfree root group
3911 */
Shaohua Li329a6782011-01-07 08:48:28 +01003912 cfqg->ref = 1;
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003913 rcu_read_lock();
Vivek Goyale98ef892010-06-18 10:39:47 -04003914 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3915 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003916 rcu_read_unlock();
Vivek Goyal25fb5162009-12-03 12:59:46 -05003917#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003918 /*
3919 * Not strictly needed (since RB_ROOT just clears the node and we
3920 * zeroed cfqd on alloc), but better be safe in case someone decides
3921 * to add magic to the rb code
3922 */
3923 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3924 cfqd->prio_trees[i] = RB_ROOT;
3925
Jens Axboe6118b702009-06-30 09:34:12 +02003926 /*
3927 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3928 * Grab a permanent reference to it, so that the normal code flow
3929 * will not attempt to free it.
3930 */
3931 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01003932 cfqd->oom_cfqq.ref++;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003933 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003934
Al Virod9ff4182006-03-18 13:51:22 -05003935 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938
Jens Axboe22e2c502005-06-27 10:55:12 +02003939 init_timer(&cfqd->idle_slice_timer);
3940 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3941 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3942
Jens Axboe23e018a2009-10-05 08:52:35 +02003943 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003944
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003946 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3947 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948 cfqd->cfq_back_max = cfq_back_max;
3949 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003950 cfqd->cfq_slice[0] = cfq_slice_async;
3951 cfqd->cfq_slice[1] = cfq_slice_sync;
3952 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3953 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003954 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003955 cfqd->cfq_latency = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05003956 cfqd->cfq_group_isolation = 0;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003957 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003958 /*
3959 * we optimistically start assuming sync ops weren't delayed in last
3960 * second, in order to have larger depth for async operations.
3961 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003962 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003963 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964}
3965
3966static void cfq_slab_kill(void)
3967{
Jens Axboed6de8be2008-05-28 14:46:59 +02003968 /*
3969 * Caller already ensured that pending RCU callbacks are completed,
3970 * so we should have no busy allocations at this point.
3971 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 if (cfq_pool)
3973 kmem_cache_destroy(cfq_pool);
3974 if (cfq_ioc_pool)
3975 kmem_cache_destroy(cfq_ioc_pool);
3976}
3977
3978static int __init cfq_slab_setup(void)
3979{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003980 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981 if (!cfq_pool)
3982 goto fail;
3983
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003984 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985 if (!cfq_ioc_pool)
3986 goto fail;
3987
3988 return 0;
3989fail:
3990 cfq_slab_kill();
3991 return -ENOMEM;
3992}
3993
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994/*
3995 * sysfs parts below -->
3996 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997static ssize_t
3998cfq_var_show(unsigned int var, char *page)
3999{
4000 return sprintf(page, "%d\n", var);
4001}
4002
4003static ssize_t
4004cfq_var_store(unsigned int *var, const char *page, size_t count)
4005{
4006 char *p = (char *) page;
4007
4008 *var = simple_strtoul(p, &p, 10);
4009 return count;
4010}
4011
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004013static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004014{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004015 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016 unsigned int __data = __VAR; \
4017 if (__CONV) \
4018 __data = jiffies_to_msecs(__data); \
4019 return cfq_var_show(__data, (page)); \
4020}
4021SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004022SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4023SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05004024SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4025SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004026SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004027SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004028SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4029SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4030SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004031SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05004032SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033#undef SHOW_FUNCTION
4034
4035#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004036static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004038 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039 unsigned int __data; \
4040 int ret = cfq_var_store(&__data, (page), count); \
4041 if (__data < (MIN)) \
4042 __data = (MIN); \
4043 else if (__data > (MAX)) \
4044 __data = (MAX); \
4045 if (__CONV) \
4046 *(__PTR) = msecs_to_jiffies(__data); \
4047 else \
4048 *(__PTR) = __data; \
4049 return ret; \
4050}
4051STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004052STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4053 UINT_MAX, 1);
4054STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4055 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004056STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004057STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4058 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004059STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004060STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004061STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4062STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004063STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4064 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004065STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05004066STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067#undef STORE_FUNCTION
4068
Al Viroe572ec72006-03-18 22:27:18 -05004069#define CFQ_ATTR(name) \
4070 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004071
Al Viroe572ec72006-03-18 22:27:18 -05004072static struct elv_fs_entry cfq_attrs[] = {
4073 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004074 CFQ_ATTR(fifo_expire_sync),
4075 CFQ_ATTR(fifo_expire_async),
4076 CFQ_ATTR(back_seek_max),
4077 CFQ_ATTR(back_seek_penalty),
4078 CFQ_ATTR(slice_sync),
4079 CFQ_ATTR(slice_async),
4080 CFQ_ATTR(slice_async_rq),
4081 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004082 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004083 CFQ_ATTR(low_latency),
Vivek Goyalae30c282009-12-03 12:59:55 -05004084 CFQ_ATTR(group_isolation),
Al Viroe572ec72006-03-18 22:27:18 -05004085 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086};
4087
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088static struct elevator_type iosched_cfq = {
4089 .ops = {
4090 .elevator_merge_fn = cfq_merge,
4091 .elevator_merged_fn = cfq_merged_request,
4092 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004093 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004094 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004095 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004097 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098 .elevator_deactivate_req_fn = cfq_deactivate_request,
4099 .elevator_queue_empty_fn = cfq_queue_empty,
4100 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004101 .elevator_former_req_fn = elv_rb_former_request,
4102 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004103 .elevator_set_req_fn = cfq_set_request,
4104 .elevator_put_req_fn = cfq_put_request,
4105 .elevator_may_queue_fn = cfq_may_queue,
4106 .elevator_init_fn = cfq_init_queue,
4107 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02004108 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109 },
Al Viro3d1ab402006-03-18 18:35:43 -05004110 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111 .elevator_name = "cfq",
4112 .elevator_owner = THIS_MODULE,
4113};
4114
Vivek Goyal3e252062009-12-04 10:36:42 -05004115#ifdef CONFIG_CFQ_GROUP_IOSCHED
4116static struct blkio_policy_type blkio_policy_cfq = {
4117 .ops = {
4118 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4119 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4120 },
Vivek Goyal062a6442010-09-15 17:06:33 -04004121 .plid = BLKIO_POLICY_PROP,
Vivek Goyal3e252062009-12-04 10:36:42 -05004122};
4123#else
4124static struct blkio_policy_type blkio_policy_cfq;
4125#endif
4126
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127static int __init cfq_init(void)
4128{
Jens Axboe22e2c502005-06-27 10:55:12 +02004129 /*
4130 * could be 0 on HZ < 1000 setups
4131 */
4132 if (!cfq_slice_async)
4133 cfq_slice_async = 1;
4134 if (!cfq_slice_idle)
4135 cfq_slice_idle = 1;
4136
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004137#ifdef CONFIG_CFQ_GROUP_IOSCHED
4138 if (!cfq_group_idle)
4139 cfq_group_idle = 1;
4140#else
4141 cfq_group_idle = 0;
4142#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143 if (cfq_slab_setup())
4144 return -ENOMEM;
4145
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004146 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05004147 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004149 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150}
4151
4152static void __exit cfq_exit(void)
4153{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07004154 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05004155 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05004157 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02004158 /* ioc_gone's update must be visible before reading ioc_count */
4159 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02004160
4161 /*
4162 * this also protects us from entering cfq_slab_kill() with
4163 * pending RCU callbacks
4164 */
Tejun Heo245b2e72009-06-24 15:13:48 +09004165 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02004166 wait_for_completion(&all_gone);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004167 ida_destroy(&cic_index_ida);
Christoph Hellwig83521d32005-10-30 15:01:39 -08004168 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169}
4170
4171module_init(cfq_init);
4172module_exit(cfq_exit);
4173
4174MODULE_AUTHOR("Jens Axboe");
4175MODULE_LICENSE("GPL");
4176MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");