blob: 3be881ec95ad06fa95ce0710dddb81f0cbef1c50 [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 Goyalc4e78932010-08-23 12:25:03 +0200149 /* Number of sectors dispatched from queue in single dispatch round */
150 unsigned long nr_sectors;
Jens Axboe6118b702009-06-30 09:34:12 +0200151};
152
153/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100154 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100155 * IDLE is handled separately, so it has negative index
156 */
157enum wl_prio_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100158 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500159 RT_WORKLOAD = 1,
160 IDLE_WORKLOAD = 2,
Vivek Goyalb4627322010-10-22 09:48:43 +0200161 CFQ_PRIO_NR,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100162};
163
164/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100165 * Second index in the service_trees.
166 */
167enum wl_type_t {
168 ASYNC_WORKLOAD = 0,
169 SYNC_NOIDLE_WORKLOAD = 1,
170 SYNC_WORKLOAD = 2
171};
172
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500173/* This is per cgroup per device grouping structure */
174struct cfq_group {
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500175 /* group service_tree member */
176 struct rb_node rb_node;
177
178 /* group service_tree key */
179 u64 vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500180 unsigned int weight;
Justin TerAvest8184f932011-03-17 16:12:36 +0100181 unsigned int new_weight;
182 bool needs_update;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500183
184 /* number of cfqq currently on this group */
185 int nr_cfqq;
186
Jens Axboe22e2c502005-06-27 10:55:12 +0200187 /*
Vivek Goyalb4627322010-10-22 09:48:43 +0200188 * Per group busy queus average. Useful for workload slice calc. We
189 * create the array for each prio class but at run time it is used
190 * only for RT and BE class and slot for IDLE class remains unused.
191 * This is primarily done to avoid confusion and a gcc warning.
192 */
193 unsigned int busy_queues_avg[CFQ_PRIO_NR];
194 /*
195 * rr lists of queues with requests. We maintain service trees for
196 * RT and BE classes. These trees are subdivided in subclasses
197 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198 * class there is no subclassification and all the cfq queues go on
199 * a single tree service_tree_idle.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100200 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200201 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100202 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100203 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500204
205 unsigned long saved_workload_slice;
206 enum wl_type_t saved_workload;
207 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500208 struct blkio_group blkg;
209#ifdef CONFIG_CFQ_GROUP_IOSCHED
210 struct hlist_node cfqd_node;
Shaohua Li329a6782011-01-07 08:48:28 +0100211 int ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500212#endif
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200213 /* number of requests that are on the dispatch list or inside driver */
214 int dispatched;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500215};
216
217/*
218 * Per block device queue structure
219 */
220struct cfq_data {
221 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500222 /* Root service tree for cfq_groups */
223 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500224 struct cfq_group root_group;
225
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100226 /*
227 * The priority currently being served
228 */
229 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100230 enum wl_type_t serving_type;
231 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500232 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200233
234 /*
235 * Each priority tree is sorted by next_request position. These
236 * trees are used when determining if two or more queues are
237 * interleaving requests (see cfq_close_cooperator).
238 */
239 struct rb_root prio_trees[CFQ_PRIO_LISTS];
240
Jens Axboe22e2c502005-06-27 10:55:12 +0200241 unsigned int busy_queues;
Shaohua Lief8a41d2011-03-07 09:26:29 +0100242 unsigned int busy_sync_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +0200243
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100244 int rq_in_driver;
245 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200246
247 /*
248 * queue-depth detection
249 */
250 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200251 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100252 /*
253 * hw_tag can be
254 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256 * 0 => no NCQ
257 */
258 int hw_tag_est_depth;
259 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200260
261 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200262 * idle window management
263 */
264 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200265 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200266
267 struct cfq_queue *active_queue;
268 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200269
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200270 /*
271 * async queue for each priority case
272 */
273 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200275
Jens Axboe6d048f52007-04-25 12:44:27 +0200276 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 /*
279 * tunables, see top of file
280 */
281 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200282 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 unsigned int cfq_back_penalty;
284 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200285 unsigned int cfq_slice[2];
286 unsigned int cfq_slice_async_rq;
287 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200288 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200289 unsigned int cfq_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500290
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400291 unsigned int cic_index;
Al Virod9ff4182006-03-18 13:51:22 -0500292 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293
Jens Axboe6118b702009-06-30 09:34:12 +0200294 /*
295 * Fallback dummy cfqq for extreme OOM conditions
296 */
297 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200298
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100299 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500300
301 /* List of cfq groups being managed on this device*/
302 struct hlist_head cfqg_list;
Jens Axboebb729bc2009-12-06 09:54:19 +0100303 struct rcu_head rcu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304};
305
Vivek Goyal25fb5162009-12-03 12:59:46 -0500306static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
307
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500308static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
309 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500310 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100311{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500312 if (!cfqg)
313 return NULL;
314
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100315 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500316 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100317
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500318 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100319}
320
Jens Axboe3b181522005-06-27 10:56:24 +0200321enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100322 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
323 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200324 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100325 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100326 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
327 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
328 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100329 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200330 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400331 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100332 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100333 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500334 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200335};
336
337#define CFQ_CFQQ_FNS(name) \
338static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
339{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100340 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200341} \
342static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
343{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100344 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200345} \
346static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
347{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100348 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200349}
350
351CFQ_CFQQ_FNS(on_rr);
352CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200353CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200354CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200355CFQ_CFQQ_FNS(fifo_expire);
356CFQ_CFQQ_FNS(idle_window);
357CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100358CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200359CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200360CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100361CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100362CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500363CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200364#undef CFQ_CFQQ_FNS
365
Vivek Goyalafc24d42010-04-26 19:27:56 +0200366#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500367#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
368 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
369 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
370 blkg_path(&(cfqq)->cfqg->blkg), ##args);
371
372#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
373 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
374 blkg_path(&(cfqg)->blkg), ##args); \
375
376#else
Jens Axboe7b679132008-05-30 12:23:07 +0200377#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
378 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500379#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
380#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200381#define cfq_log(cfqd, fmt, args...) \
382 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
383
Vivek Goyal615f0252009-12-03 12:59:39 -0500384/* Traverses through cfq group service trees */
385#define for_each_cfqg_st(cfqg, i, j, st) \
386 for (i = 0; i <= IDLE_WORKLOAD; i++) \
387 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
388 : &cfqg->service_tree_idle; \
389 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
390 (i == IDLE_WORKLOAD && j == 0); \
391 j++, st = i < IDLE_WORKLOAD ? \
392 &cfqg->service_trees[i][j]: NULL) \
393
394
Vivek Goyal02b35082010-08-23 12:23:53 +0200395static inline bool iops_mode(struct cfq_data *cfqd)
396{
397 /*
398 * If we are not idling on queues and it is a NCQ drive, parallel
399 * execution of requests is on and measuring time is not possible
400 * in most of the cases until and unless we drive shallower queue
401 * depths and that becomes a performance bottleneck. In such cases
402 * switch to start providing fairness in terms of number of IOs.
403 */
404 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
405 return true;
406 else
407 return false;
408}
409
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100410static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
411{
412 if (cfq_class_idle(cfqq))
413 return IDLE_WORKLOAD;
414 if (cfq_class_rt(cfqq))
415 return RT_WORKLOAD;
416 return BE_WORKLOAD;
417}
418
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100419
420static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
421{
422 if (!cfq_cfqq_sync(cfqq))
423 return ASYNC_WORKLOAD;
424 if (!cfq_cfqq_idle_window(cfqq))
425 return SYNC_NOIDLE_WORKLOAD;
426 return SYNC_WORKLOAD;
427}
428
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500429static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
430 struct cfq_data *cfqd,
431 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100432{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500433 if (wl == IDLE_WORKLOAD)
434 return cfqg->service_tree_idle.count;
435
436 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
437 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
438 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100439}
440
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500441static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
442 struct cfq_group *cfqg)
443{
444 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
445 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
446}
447
Jens Axboe165125e2007-07-24 09:28:11 +0200448static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200449static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100450 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100451static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200452 struct io_context *);
453
454static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200455 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200456{
Jens Axboea6151c32009-10-07 20:02:57 +0200457 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200458}
459
460static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200461 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200462{
Jens Axboea6151c32009-10-07 20:02:57 +0200463 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200464}
465
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400466#define CIC_DEAD_KEY 1ul
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400467#define CIC_DEAD_INDEX_SHIFT 1
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400468
469static inline void *cfqd_dead_key(struct cfq_data *cfqd)
470{
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400471 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400472}
473
474static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
475{
476 struct cfq_data *cfqd = cic->key;
477
478 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
479 return NULL;
480
481 return cfqd;
482}
483
Vasily Tarasov91fac312007-04-25 12:29:51 +0200484/*
485 * We regard a request as SYNC, if it's either a read or has the SYNC bit
486 * set (in which case it could also be direct WRITE).
487 */
Jens Axboea6151c32009-10-07 20:02:57 +0200488static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200489{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200490 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200491}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700494 * scheduler run of queue, if there are requests pending and no one in the
495 * driver that will restart queueing
496 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200497static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700498{
Jens Axboe7b679132008-05-30 12:23:07 +0200499 if (cfqd->busy_queues) {
500 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200501 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200502 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700503}
504
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100506 * Scale schedule slice based on io priority. Use the sync time slice only
507 * if a queue is marked sync and has sync io queued. A sync queue with async
508 * io only, should not get full sync slice length.
509 */
Jens Axboea6151c32009-10-07 20:02:57 +0200510static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200511 unsigned short prio)
512{
513 const int base_slice = cfqd->cfq_slice[sync];
514
515 WARN_ON(prio >= IOPRIO_BE_NR);
516
517 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
518}
519
Jens Axboe44f7c162007-01-19 11:51:58 +1100520static inline int
521cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
522{
Jens Axboed9e76202007-04-20 14:27:50 +0200523 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100524}
525
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500526static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
527{
528 u64 d = delta << CFQ_SERVICE_SHIFT;
529
530 d = d * BLKIO_WEIGHT_DEFAULT;
531 do_div(d, cfqg->weight);
532 return d;
533}
534
535static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
536{
537 s64 delta = (s64)(vdisktime - min_vdisktime);
538 if (delta > 0)
539 min_vdisktime = vdisktime;
540
541 return min_vdisktime;
542}
543
544static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
545{
546 s64 delta = (s64)(vdisktime - min_vdisktime);
547 if (delta < 0)
548 min_vdisktime = vdisktime;
549
550 return min_vdisktime;
551}
552
553static void update_min_vdisktime(struct cfq_rb_root *st)
554{
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500555 struct cfq_group *cfqg;
556
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500557 if (st->left) {
558 cfqg = rb_entry_cfqg(st->left);
Gui Jianfenga6032712011-03-07 09:28:09 +0100559 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
560 cfqg->vdisktime);
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500561 }
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500562}
563
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100564/*
565 * get averaged number of queues of RT/BE priority.
566 * average is updated, with a formula that gives more weight to higher numbers,
567 * to quickly follows sudden increases and decrease slowly
568 */
569
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500570static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
571 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100572{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100573 unsigned min_q, max_q;
574 unsigned mult = cfq_hist_divisor - 1;
575 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500576 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100577
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500578 min_q = min(cfqg->busy_queues_avg[rt], busy);
579 max_q = max(cfqg->busy_queues_avg[rt], busy);
580 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100581 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500582 return cfqg->busy_queues_avg[rt];
583}
584
585static inline unsigned
586cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
587{
588 struct cfq_rb_root *st = &cfqd->grp_service_tree;
589
590 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100591}
592
Shaohua Lic553f8e2011-01-14 08:41:03 +0100593static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -0700594cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100595{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100596 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
597 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500598 /*
599 * interested queues (we consider only the ones with the same
600 * priority class in the cfq group)
601 */
602 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
603 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100604 unsigned sync_slice = cfqd->cfq_slice[1];
605 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500606 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
607
608 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100609 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
610 /* scale low_slice according to IO priority
611 * and sync vs async */
612 unsigned low_slice =
613 min(slice, base_low_slice * slice / sync_slice);
614 /* the adapted slice value is scaled to fit all iqs
615 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500616 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100617 low_slice);
618 }
619 }
Shaohua Lic553f8e2011-01-14 08:41:03 +0100620 return slice;
621}
622
623static inline void
624cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
625{
Vivek Goyalba5bd522011-01-19 08:25:02 -0700626 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +0100627
Vivek Goyaldae739e2009-12-03 12:59:45 -0500628 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100629 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500630 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200631 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100632}
633
634/*
635 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
636 * isn't valid until the first request from the dispatch is activated
637 * and the slice time set.
638 */
Jens Axboea6151c32009-10-07 20:02:57 +0200639static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100640{
641 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +0100642 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100643 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +0100644 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100645
Shaohua Lic1e44752010-11-08 15:01:02 +0100646 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +1100647}
648
649/*
Jens Axboe5e705372006-07-13 12:39:25 +0200650 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200652 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 */
Jens Axboe5e705372006-07-13 12:39:25 +0200654static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100655cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100657 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200659#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
660#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
661 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
Jens Axboe5e705372006-07-13 12:39:25 +0200663 if (rq1 == NULL || rq1 == rq2)
664 return rq2;
665 if (rq2 == NULL)
666 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200667
Jens Axboe5e705372006-07-13 12:39:25 +0200668 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
669 return rq1;
670 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
671 return rq2;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200672 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200673 return rq1;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200674 else if ((rq2->cmd_flags & REQ_META) &&
675 !(rq1->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200676 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677
Tejun Heo83096eb2009-05-07 22:24:39 +0900678 s1 = blk_rq_pos(rq1);
679 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 /*
682 * by definition, 1KiB is 2 sectors
683 */
684 back_max = cfqd->cfq_back_max * 2;
685
686 /*
687 * Strict one way elevator _except_ in the case where we allow
688 * short backward seeks which are biased as twice the cost of a
689 * similar forward seek.
690 */
691 if (s1 >= last)
692 d1 = s1 - last;
693 else if (s1 + back_max >= last)
694 d1 = (last - s1) * cfqd->cfq_back_penalty;
695 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200696 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697
698 if (s2 >= last)
699 d2 = s2 - last;
700 else if (s2 + back_max >= last)
701 d2 = (last - s2) * cfqd->cfq_back_penalty;
702 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200703 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704
705 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706
Andreas Mohre8a99052006-03-28 08:59:49 +0200707 /*
708 * By doing switch() on the bit mask "wrap" we avoid having to
709 * check two variables for all permutations: --> faster!
710 */
711 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200712 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200713 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200714 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200715 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200716 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200717 else {
718 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200719 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200720 else
Jens Axboe5e705372006-07-13 12:39:25 +0200721 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200722 }
723
724 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200725 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200726 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200727 return rq2;
728 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200729 default:
730 /*
731 * Since both rqs are wrapped,
732 * start with the one that's further behind head
733 * (--> only *one* back seek required),
734 * since back seek takes more time than forward.
735 */
736 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200737 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 else
Jens Axboe5e705372006-07-13 12:39:25 +0200739 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 }
741}
742
Jens Axboe498d3aa22007-04-26 12:54:48 +0200743/*
744 * The below is leftmost cache rbtree addon
745 */
Jens Axboe08717142008-01-28 11:38:15 +0100746static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200747{
Vivek Goyal615f0252009-12-03 12:59:39 -0500748 /* Service tree is empty */
749 if (!root->count)
750 return NULL;
751
Jens Axboecc09e292007-04-26 12:53:50 +0200752 if (!root->left)
753 root->left = rb_first(&root->rb);
754
Jens Axboe08717142008-01-28 11:38:15 +0100755 if (root->left)
756 return rb_entry(root->left, struct cfq_queue, rb_node);
757
758 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200759}
760
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500761static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
762{
763 if (!root->left)
764 root->left = rb_first(&root->rb);
765
766 if (root->left)
767 return rb_entry_cfqg(root->left);
768
769 return NULL;
770}
771
Jens Axboea36e71f2009-04-15 12:15:11 +0200772static void rb_erase_init(struct rb_node *n, struct rb_root *root)
773{
774 rb_erase(n, root);
775 RB_CLEAR_NODE(n);
776}
777
Jens Axboecc09e292007-04-26 12:53:50 +0200778static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
779{
780 if (root->left == n)
781 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200782 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100783 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200784}
785
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786/*
787 * would be nice to take fifo expire time into account as well
788 */
Jens Axboe5e705372006-07-13 12:39:25 +0200789static struct request *
790cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
791 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792{
Jens Axboe21183b02006-07-13 12:33:14 +0200793 struct rb_node *rbnext = rb_next(&last->rb_node);
794 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200795 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
Jens Axboe21183b02006-07-13 12:33:14 +0200797 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
799 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200800 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200801
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200803 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200804 else {
805 rbnext = rb_first(&cfqq->sort_list);
806 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200807 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200808 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100810 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811}
812
Jens Axboed9e76202007-04-20 14:27:50 +0200813static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
814 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815{
Jens Axboed9e76202007-04-20 14:27:50 +0200816 /*
817 * just an approximation, should be ok.
818 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500819 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100820 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200821}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500823static inline s64
824cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
825{
826 return cfqg->vdisktime - st->min_vdisktime;
827}
828
829static void
830__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
831{
832 struct rb_node **node = &st->rb.rb_node;
833 struct rb_node *parent = NULL;
834 struct cfq_group *__cfqg;
835 s64 key = cfqg_key(st, cfqg);
836 int left = 1;
837
838 while (*node != NULL) {
839 parent = *node;
840 __cfqg = rb_entry_cfqg(parent);
841
842 if (key < cfqg_key(st, __cfqg))
843 node = &parent->rb_left;
844 else {
845 node = &parent->rb_right;
846 left = 0;
847 }
848 }
849
850 if (left)
851 st->left = &cfqg->rb_node;
852
853 rb_link_node(&cfqg->rb_node, parent, node);
854 rb_insert_color(&cfqg->rb_node, &st->rb);
855}
856
857static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100858cfq_update_group_weight(struct cfq_group *cfqg)
859{
860 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
861 if (cfqg->needs_update) {
862 cfqg->weight = cfqg->new_weight;
863 cfqg->needs_update = false;
864 }
865}
866
867static void
868cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
869{
870 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
871
872 cfq_update_group_weight(cfqg);
873 __cfq_group_service_tree_add(st, cfqg);
874 st->total_weight += cfqg->weight;
875}
876
877static void
878cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500879{
880 struct cfq_rb_root *st = &cfqd->grp_service_tree;
881 struct cfq_group *__cfqg;
882 struct rb_node *n;
883
884 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +0100885 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500886 return;
887
888 /*
889 * Currently put the group at the end. Later implement something
890 * so that groups get lesser vtime based on their weights, so that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300891 * if group does not loose all if it was not continuously backlogged.
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500892 */
893 n = rb_last(&st->rb);
894 if (n) {
895 __cfqg = rb_entry_cfqg(n);
896 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
897 } else
898 cfqg->vdisktime = st->min_vdisktime;
Justin TerAvest8184f932011-03-17 16:12:36 +0100899 cfq_group_service_tree_add(st, cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500900}
901
902static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100903cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
904{
905 st->total_weight -= cfqg->weight;
906 if (!RB_EMPTY_NODE(&cfqg->rb_node))
907 cfq_rb_erase(&cfqg->rb_node, st);
908}
909
910static void
911cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500912{
913 struct cfq_rb_root *st = &cfqd->grp_service_tree;
914
915 BUG_ON(cfqg->nr_cfqq < 1);
916 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500917
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500918 /* If there are other cfq queues under this group, don't delete it */
919 if (cfqg->nr_cfqq)
920 return;
921
Vivek Goyal2868ef72009-12-03 12:59:48 -0500922 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Justin TerAvest8184f932011-03-17 16:12:36 +0100923 cfq_group_service_tree_del(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500924 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400925 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500926}
927
Justin TerAvest167400d2011-03-12 16:54:00 +0100928static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
929 unsigned int *unaccounted_time)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500930{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500931 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500932
933 /*
934 * Queue got expired before even a single request completed or
935 * got expired immediately after first request completion.
936 */
937 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
938 /*
939 * Also charge the seek time incurred to the group, otherwise
940 * if there are mutiple queues in the group, each can dispatch
941 * a single request on seeky media and cause lots of seek time
942 * and group will never know it.
943 */
944 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
945 1);
946 } else {
947 slice_used = jiffies - cfqq->slice_start;
Justin TerAvest167400d2011-03-12 16:54:00 +0100948 if (slice_used > cfqq->allocated_slice) {
949 *unaccounted_time = slice_used - cfqq->allocated_slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500950 slice_used = cfqq->allocated_slice;
Justin TerAvest167400d2011-03-12 16:54:00 +0100951 }
952 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
953 *unaccounted_time += cfqq->slice_start -
954 cfqq->dispatch_start;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500955 }
956
Vivek Goyaldae739e2009-12-03 12:59:45 -0500957 return slice_used;
958}
959
960static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200961 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500962{
963 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Justin TerAvest167400d2011-03-12 16:54:00 +0100964 unsigned int used_sl, charge, unaccounted_sl = 0;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500965 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
966 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500967
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500968 BUG_ON(nr_sync < 0);
Justin TerAvest167400d2011-03-12 16:54:00 +0100969 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500970
Vivek Goyal02b35082010-08-23 12:23:53 +0200971 if (iops_mode(cfqd))
972 charge = cfqq->slice_dispatch;
973 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
974 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500975
976 /* Can't update vdisktime while group is on service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +0100977 cfq_group_service_tree_del(st, cfqg);
Vivek Goyal02b35082010-08-23 12:23:53 +0200978 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Justin TerAvest8184f932011-03-17 16:12:36 +0100979 /* If a new weight was requested, update now, off tree */
980 cfq_group_service_tree_add(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500981
982 /* This group is being expired. Save the context */
983 if (time_after(cfqd->workload_expires, jiffies)) {
984 cfqg->saved_workload_slice = cfqd->workload_expires
985 - jiffies;
986 cfqg->saved_workload = cfqd->serving_type;
987 cfqg->saved_serving_prio = cfqd->serving_prio;
988 } else
989 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500990
991 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
992 st->min_vdisktime);
Vivek Goyalc4e78932010-08-23 12:25:03 +0200993 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
994 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
995 iops_mode(cfqd), cfqq->nr_sectors);
Justin TerAvest167400d2011-03-12 16:54:00 +0100996 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
997 unaccounted_sl);
Vivek Goyale98ef892010-06-18 10:39:47 -0400998 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500999}
1000
Vivek Goyal25fb5162009-12-03 12:59:46 -05001001#ifdef CONFIG_CFQ_GROUP_IOSCHED
1002static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1003{
1004 if (blkg)
1005 return container_of(blkg, struct cfq_group, blkg);
1006 return NULL;
1007}
1008
Vivek Goyalfe071432010-10-01 14:49:49 +02001009void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1010 unsigned int weight)
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001011{
Justin TerAvest8184f932011-03-17 16:12:36 +01001012 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1013 cfqg->new_weight = weight;
1014 cfqg->needs_update = true;
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001015}
1016
Vivek Goyal25fb5162009-12-03 12:59:46 -05001017static struct cfq_group *
1018cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
1019{
1020 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
1021 struct cfq_group *cfqg = NULL;
1022 void *key = cfqd;
1023 int i, j;
1024 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -05001025 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1026 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001027
Vivek Goyal25fb5162009-12-03 12:59:46 -05001028 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
Ricky Beniteza74b2ad2010-04-05 18:22:17 +02001029 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1030 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1031 cfqg->blkg.dev = MKDEV(major, minor);
1032 goto done;
1033 }
Vivek Goyal25fb5162009-12-03 12:59:46 -05001034 if (cfqg || !create)
1035 goto done;
1036
1037 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1038 if (!cfqg)
1039 goto done;
1040
Vivek Goyal25fb5162009-12-03 12:59:46 -05001041 for_each_cfqg_st(cfqg, i, j, st)
1042 *st = CFQ_RB_ROOT;
1043 RB_CLEAR_NODE(&cfqg->rb_node);
1044
Vivek Goyalb1c35762009-12-03 12:59:47 -05001045 /*
1046 * Take the initial reference that will be released on destroy
1047 * This can be thought of a joint reference by cgroup and
1048 * elevator which will be dropped by either elevator exit
1049 * or cgroup deletion path depending on who is exiting first.
1050 */
Shaohua Li329a6782011-01-07 08:48:28 +01001051 cfqg->ref = 1;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001052
Vivek Goyal180be2a2010-09-14 08:47:11 +02001053 /*
1054 * Add group onto cgroup list. It might happen that bdi->dev is
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001055 * not initialized yet. Initialize this new group without major
Vivek Goyal180be2a2010-09-14 08:47:11 +02001056 * and minor info and this info will be filled in once a new thread
1057 * comes for IO. See code above.
1058 */
1059 if (bdi->dev) {
1060 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1061 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
Vivek Goyal22084192009-12-03 12:59:49 -05001062 MKDEV(major, minor));
Vivek Goyal180be2a2010-09-14 08:47:11 +02001063 } else
1064 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1065 0);
1066
Gui Jianfeng34d0f172010-04-13 16:05:49 +08001067 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001068
1069 /* Add group on cfqd list */
1070 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1071
1072done:
Vivek Goyal25fb5162009-12-03 12:59:46 -05001073 return cfqg;
1074}
1075
1076/*
1077 * Search for the cfq group current task belongs to. If create = 1, then also
1078 * create the cfq group if it does not exist. request_queue lock must be held.
1079 */
1080static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1081{
1082 struct cgroup *cgroup;
1083 struct cfq_group *cfqg = NULL;
1084
1085 rcu_read_lock();
1086 cgroup = task_cgroup(current, blkio_subsys_id);
1087 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1088 if (!cfqg && create)
1089 cfqg = &cfqd->root_group;
1090 rcu_read_unlock();
1091 return cfqg;
1092}
1093
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001094static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1095{
Shaohua Li329a6782011-01-07 08:48:28 +01001096 cfqg->ref++;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001097 return cfqg;
1098}
1099
Vivek Goyal25fb5162009-12-03 12:59:46 -05001100static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1101{
1102 /* Currently, all async queues are mapped to root group */
1103 if (!cfq_cfqq_sync(cfqq))
1104 cfqg = &cfqq->cfqd->root_group;
1105
1106 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001107 /* cfqq reference on cfqg */
Shaohua Li329a6782011-01-07 08:48:28 +01001108 cfqq->cfqg->ref++;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001109}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001110
1111static void cfq_put_cfqg(struct cfq_group *cfqg)
1112{
1113 struct cfq_rb_root *st;
1114 int i, j;
1115
Shaohua Li329a6782011-01-07 08:48:28 +01001116 BUG_ON(cfqg->ref <= 0);
1117 cfqg->ref--;
1118 if (cfqg->ref)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001119 return;
1120 for_each_cfqg_st(cfqg, i, j, st)
Gui Jianfengb54ce602010-11-30 20:52:46 +01001121 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
Vivek Goyalb1c35762009-12-03 12:59:47 -05001122 kfree(cfqg);
1123}
1124
1125static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1126{
1127 /* Something wrong if we are trying to remove same group twice */
1128 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1129
1130 hlist_del_init(&cfqg->cfqd_node);
1131
1132 /*
1133 * Put the reference taken at the time of creation so that when all
1134 * queues are gone, group can be destroyed.
1135 */
1136 cfq_put_cfqg(cfqg);
1137}
1138
1139static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1140{
1141 struct hlist_node *pos, *n;
1142 struct cfq_group *cfqg;
1143
1144 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1145 /*
1146 * If cgroup removal path got to blk_group first and removed
1147 * it from cgroup list, then it will take care of destroying
1148 * cfqg also.
1149 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001150 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001151 cfq_destroy_cfqg(cfqd, cfqg);
1152 }
1153}
1154
1155/*
1156 * Blk cgroup controller notification saying that blkio_group object is being
1157 * delinked as associated cgroup object is going away. That also means that
1158 * no new IO will come in this group. So get rid of this group as soon as
1159 * any pending IO in the group is finished.
1160 *
1161 * This function is called under rcu_read_lock(). key is the rcu protected
1162 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1163 * read lock.
1164 *
1165 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1166 * it should not be NULL as even if elevator was exiting, cgroup deltion
1167 * path got to it first.
1168 */
1169void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1170{
1171 unsigned long flags;
1172 struct cfq_data *cfqd = key;
1173
1174 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1175 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1176 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1177}
1178
Vivek Goyal25fb5162009-12-03 12:59:46 -05001179#else /* GROUP_IOSCHED */
1180static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1181{
1182 return &cfqd->root_group;
1183}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001184
1185static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1186{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001187 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001188}
1189
Vivek Goyal25fb5162009-12-03 12:59:46 -05001190static inline void
1191cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1192 cfqq->cfqg = cfqg;
1193}
1194
Vivek Goyalb1c35762009-12-03 12:59:47 -05001195static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1196static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1197
Vivek Goyal25fb5162009-12-03 12:59:46 -05001198#endif /* GROUP_IOSCHED */
1199
Jens Axboe498d3aa22007-04-26 12:54:48 +02001200/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001201 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001202 * requests waiting to be processed. It is sorted in the order that
1203 * we will service the queues.
1204 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001205static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001206 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001207{
Jens Axboe08717142008-01-28 11:38:15 +01001208 struct rb_node **p, *parent;
1209 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001210 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001211 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001212 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001213 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001214 int group_changed = 0;
1215
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;
Justin TerAvest8184f932011-03-17 16:12:36 +01001287 cfq_group_notify_queue_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++;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001375 if (cfq_cfqq_sync(cfqq))
1376 cfqd->busy_sync_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377
Jens Axboeedd75ff2007-04-19 12:03:34 +02001378 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379}
1380
Jens Axboe498d3aa22007-04-26 12:54:48 +02001381/*
1382 * Called when the cfqq no longer has requests pending, remove it from
1383 * the service tree.
1384 */
Jens Axboefebffd62008-01-28 13:19:43 +01001385static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386{
Jens Axboe7b679132008-05-30 12:23:07 +02001387 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001388 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1389 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001391 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1392 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1393 cfqq->service_tree = NULL;
1394 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001395 if (cfqq->p_root) {
1396 rb_erase(&cfqq->p_node, cfqq->p_root);
1397 cfqq->p_root = NULL;
1398 }
Jens Axboed9e76202007-04-20 14:27:50 +02001399
Justin TerAvest8184f932011-03-17 16:12:36 +01001400 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401 BUG_ON(!cfqd->busy_queues);
1402 cfqd->busy_queues--;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001403 if (cfq_cfqq_sync(cfqq))
1404 cfqd->busy_sync_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405}
1406
1407/*
1408 * rb tree support functions
1409 */
Jens Axboefebffd62008-01-28 13:19:43 +01001410static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411{
Jens Axboe5e705372006-07-13 12:39:25 +02001412 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001413 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414
Jens Axboeb4878f22005-10-20 16:42:29 +02001415 BUG_ON(!cfqq->queued[sync]);
1416 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417
Jens Axboe5e705372006-07-13 12:39:25 +02001418 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419
Vivek Goyalf04a6422009-12-03 12:59:40 -05001420 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1421 /*
1422 * Queue will be deleted from service tree when we actually
1423 * expire it later. Right now just remove it from prio tree
1424 * as it is empty.
1425 */
1426 if (cfqq->p_root) {
1427 rb_erase(&cfqq->p_node, cfqq->p_root);
1428 cfqq->p_root = NULL;
1429 }
1430 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431}
1432
Jens Axboe5e705372006-07-13 12:39:25 +02001433static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434{
Jens Axboe5e705372006-07-13 12:39:25 +02001435 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001437 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438
Jens Axboe5380a102006-07-13 12:37:56 +02001439 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440
1441 /*
1442 * looks a little odd, but the first insert might return an alias.
1443 * if that happens, put the alias on the dispatch list
1444 */
Jens Axboe21183b02006-07-13 12:33:14 +02001445 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001446 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001447
1448 if (!cfq_cfqq_on_rr(cfqq))
1449 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001450
1451 /*
1452 * check if this request is a better next-serve candidate
1453 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001454 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001455 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001456
1457 /*
1458 * adjust priority tree position, if ->next_rq changes
1459 */
1460 if (prev != cfqq->next_rq)
1461 cfq_prio_tree_add(cfqd, cfqq);
1462
Jens Axboe5044eed2007-04-25 11:53:48 +02001463 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464}
1465
Jens Axboefebffd62008-01-28 13:19:43 +01001466static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467{
Jens Axboe5380a102006-07-13 12:37:56 +02001468 elv_rb_del(&cfqq->sort_list, rq);
1469 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001470 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1471 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001472 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001473 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001474 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1475 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476}
1477
Jens Axboe206dc692006-03-28 13:03:44 +02001478static struct request *
1479cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480{
Jens Axboe206dc692006-03-28 13:03:44 +02001481 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001482 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001483 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484
Jens Axboe4ac845a2008-01-24 08:44:49 +01001485 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001486 if (!cic)
1487 return NULL;
1488
1489 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001490 if (cfqq) {
1491 sector_t sector = bio->bi_sector + bio_sectors(bio);
1492
Jens Axboe21183b02006-07-13 12:33:14 +02001493 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001494 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 return NULL;
1497}
1498
Jens Axboe165125e2007-07-24 09:28:11 +02001499static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001500{
1501 struct cfq_data *cfqd = q->elevator->elevator_data;
1502
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001503 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001504 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001505 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001506
Tejun Heo5b936292009-05-07 22:24:38 +09001507 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001508}
1509
Jens Axboe165125e2007-07-24 09:28:11 +02001510static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511{
Jens Axboe22e2c502005-06-27 10:55:12 +02001512 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001514 WARN_ON(!cfqd->rq_in_driver);
1515 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001516 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001517 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518}
1519
Jens Axboeb4878f22005-10-20 16:42:29 +02001520static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521{
Jens Axboe5e705372006-07-13 12:39:25 +02001522 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001523
Jens Axboe5e705372006-07-13 12:39:25 +02001524 if (cfqq->next_rq == rq)
1525 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526
Jens Axboeb4878f22005-10-20 16:42:29 +02001527 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001528 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001529
Aaron Carroll45333d52008-08-26 15:52:36 +02001530 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001531 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1532 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02001533 if (rq->cmd_flags & REQ_META) {
Jens Axboe374f84a2006-07-23 01:42:19 +02001534 WARN_ON(!cfqq->meta_pending);
1535 cfqq->meta_pending--;
1536 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537}
1538
Jens Axboe165125e2007-07-24 09:28:11 +02001539static int cfq_merge(struct request_queue *q, struct request **req,
1540 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541{
1542 struct cfq_data *cfqd = q->elevator->elevator_data;
1543 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544
Jens Axboe206dc692006-03-28 13:03:44 +02001545 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001546 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001547 *req = __rq;
1548 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549 }
1550
1551 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552}
1553
Jens Axboe165125e2007-07-24 09:28:11 +02001554static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001555 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556{
Jens Axboe21183b02006-07-13 12:33:14 +02001557 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001558 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559
Jens Axboe5e705372006-07-13 12:39:25 +02001560 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562}
1563
Divyesh Shah812d4022010-04-08 21:14:23 -07001564static void cfq_bio_merged(struct request_queue *q, struct request *req,
1565 struct bio *bio)
1566{
Vivek Goyale98ef892010-06-18 10:39:47 -04001567 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1568 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001569}
1570
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571static void
Jens Axboe165125e2007-07-24 09:28:11 +02001572cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573 struct request *next)
1574{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001575 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001576 /*
1577 * reposition in fifo if next is older than rq
1578 */
1579 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001580 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001581 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001582 rq_set_fifo_time(rq, rq_fifo_time(next));
1583 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001584
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001585 if (cfqq->next_rq == next)
1586 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001587 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001588 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1589 rq_data_dir(next), rq_is_sync(next));
Jens Axboe22e2c502005-06-27 10:55:12 +02001590}
1591
Jens Axboe165125e2007-07-24 09:28:11 +02001592static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001593 struct bio *bio)
1594{
1595 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001596 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001597 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001598
1599 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001600 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001601 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001602 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001603 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001604
1605 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001606 * Lookup the cfqq that this bio will be queued with. Allow
1607 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001608 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001609 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001610 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001611 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001612
Vasily Tarasov91fac312007-04-25 12:29:51 +02001613 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001614 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001615}
1616
Divyesh Shah812df482010-04-08 21:15:35 -07001617static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1618{
1619 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001620 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001621}
1622
Jens Axboefebffd62008-01-28 13:19:43 +01001623static void __cfq_set_active_queue(struct cfq_data *cfqd,
1624 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001625{
1626 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001627 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1628 cfqd->serving_prio, cfqd->serving_type);
Justin TerAvest62a37f62011-03-23 08:25:44 +01001629 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1630 cfqq->slice_start = 0;
1631 cfqq->dispatch_start = jiffies;
1632 cfqq->allocated_slice = 0;
1633 cfqq->slice_end = 0;
1634 cfqq->slice_dispatch = 0;
1635 cfqq->nr_sectors = 0;
1636
1637 cfq_clear_cfqq_wait_request(cfqq);
1638 cfq_clear_cfqq_must_dispatch(cfqq);
1639 cfq_clear_cfqq_must_alloc_slice(cfqq);
1640 cfq_clear_cfqq_fifo_expire(cfqq);
1641 cfq_mark_cfqq_slice_new(cfqq);
1642
1643 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001644 }
1645
1646 cfqd->active_queue = cfqq;
1647}
1648
1649/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001650 * current cfqq expired its slice (or was too idle), select new one
1651 */
1652static void
1653__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001654 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001655{
Jens Axboe7b679132008-05-30 12:23:07 +02001656 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1657
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001658 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001659 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001660
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001661 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001662 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001663
1664 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001665 * If this cfqq is shared between multiple processes, check to
1666 * make sure that those processes are still issuing I/Os within
1667 * the mean seek distance. If not, it may be time to break the
1668 * queues apart again.
1669 */
1670 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1671 cfq_mark_cfqq_split_coop(cfqq);
1672
1673 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001674 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001675 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01001676 if (timed_out) {
1677 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07001678 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001679 else
1680 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001681 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1682 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001683
Vivek Goyale5ff0822010-04-26 19:25:11 +02001684 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001685
Vivek Goyalf04a6422009-12-03 12:59:40 -05001686 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1687 cfq_del_cfqq_rr(cfqd, cfqq);
1688
Jens Axboeedd75ff2007-04-19 12:03:34 +02001689 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001690
1691 if (cfqq == cfqd->active_queue)
1692 cfqd->active_queue = NULL;
1693
1694 if (cfqd->active_cic) {
1695 put_io_context(cfqd->active_cic->ioc);
1696 cfqd->active_cic = NULL;
1697 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001698}
1699
Vivek Goyale5ff0822010-04-26 19:25:11 +02001700static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001701{
1702 struct cfq_queue *cfqq = cfqd->active_queue;
1703
1704 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001705 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001706}
1707
Jens Axboe498d3aa22007-04-26 12:54:48 +02001708/*
1709 * Get next queue for service. Unless we have a queue preemption,
1710 * we'll simply select the first cfqq in the service tree.
1711 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001712static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001713{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001714 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001715 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001716 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001717
Vivek Goyalf04a6422009-12-03 12:59:40 -05001718 if (!cfqd->rq_queued)
1719 return NULL;
1720
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001721 /* There is nothing to dispatch */
1722 if (!service_tree)
1723 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001724 if (RB_EMPTY_ROOT(&service_tree->rb))
1725 return NULL;
1726 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001727}
1728
Vivek Goyalf04a6422009-12-03 12:59:40 -05001729static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1730{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001731 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001732 struct cfq_queue *cfqq;
1733 int i, j;
1734 struct cfq_rb_root *st;
1735
1736 if (!cfqd->rq_queued)
1737 return NULL;
1738
Vivek Goyal25fb5162009-12-03 12:59:46 -05001739 cfqg = cfq_get_next_cfqg(cfqd);
1740 if (!cfqg)
1741 return NULL;
1742
Vivek Goyalf04a6422009-12-03 12:59:40 -05001743 for_each_cfqg_st(cfqg, i, j, st)
1744 if ((cfqq = cfq_rb_first(st)) != NULL)
1745 return cfqq;
1746 return NULL;
1747}
1748
Jens Axboe498d3aa22007-04-26 12:54:48 +02001749/*
1750 * Get and set a new active queue for service.
1751 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001752static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1753 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001754{
Jens Axboee00ef792009-11-04 08:54:55 +01001755 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001756 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001757
Jens Axboe22e2c502005-06-27 10:55:12 +02001758 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001759 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001760}
1761
Jens Axboed9e76202007-04-20 14:27:50 +02001762static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1763 struct request *rq)
1764{
Tejun Heo83096eb2009-05-07 22:24:39 +09001765 if (blk_rq_pos(rq) >= cfqd->last_position)
1766 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001767 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001768 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001769}
1770
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001771static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001772 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001773{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001774 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001775}
1776
Jens Axboea36e71f2009-04-15 12:15:11 +02001777static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1778 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001779{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001780 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001781 struct rb_node *parent, *node;
1782 struct cfq_queue *__cfqq;
1783 sector_t sector = cfqd->last_position;
1784
1785 if (RB_EMPTY_ROOT(root))
1786 return NULL;
1787
1788 /*
1789 * First, if we find a request starting at the end of the last
1790 * request, choose it.
1791 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001792 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001793 if (__cfqq)
1794 return __cfqq;
1795
1796 /*
1797 * If the exact sector wasn't found, the parent of the NULL leaf
1798 * will contain the closest sector.
1799 */
1800 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001801 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001802 return __cfqq;
1803
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001804 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001805 node = rb_next(&__cfqq->p_node);
1806 else
1807 node = rb_prev(&__cfqq->p_node);
1808 if (!node)
1809 return NULL;
1810
1811 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001812 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001813 return __cfqq;
1814
1815 return NULL;
1816}
1817
1818/*
1819 * cfqd - obvious
1820 * cur_cfqq - passed in so that we don't decide that the current queue is
1821 * closely cooperating with itself.
1822 *
1823 * So, basically we're assuming that that cur_cfqq has dispatched at least
1824 * one request, and that cfqd->last_position reflects a position on the disk
1825 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1826 * assumption.
1827 */
1828static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001829 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001830{
1831 struct cfq_queue *cfqq;
1832
Divyesh Shah39c01b22010-03-25 15:45:57 +01001833 if (cfq_class_idle(cur_cfqq))
1834 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001835 if (!cfq_cfqq_sync(cur_cfqq))
1836 return NULL;
1837 if (CFQQ_SEEKY(cur_cfqq))
1838 return NULL;
1839
Jens Axboea36e71f2009-04-15 12:15:11 +02001840 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001841 * Don't search priority tree if it's the only queue in the group.
1842 */
1843 if (cur_cfqq->cfqg->nr_cfqq == 1)
1844 return NULL;
1845
1846 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001847 * We should notice if some of the queues are cooperating, eg
1848 * working closely on the same area of the disk. In that case,
1849 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001850 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001851 cfqq = cfqq_close(cfqd, cur_cfqq);
1852 if (!cfqq)
1853 return NULL;
1854
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001855 /* If new queue belongs to different cfq_group, don't choose it */
1856 if (cur_cfqq->cfqg != cfqq->cfqg)
1857 return NULL;
1858
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001859 /*
1860 * It only makes sense to merge sync queues.
1861 */
1862 if (!cfq_cfqq_sync(cfqq))
1863 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001864 if (CFQQ_SEEKY(cfqq))
1865 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001866
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001867 /*
1868 * Do not merge queues of different priority classes
1869 */
1870 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1871 return NULL;
1872
Jens Axboea36e71f2009-04-15 12:15:11 +02001873 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001874}
1875
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001876/*
1877 * Determine whether we should enforce idle window for this queue.
1878 */
1879
1880static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1881{
1882 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001883 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001884
Vivek Goyalf04a6422009-12-03 12:59:40 -05001885 BUG_ON(!service_tree);
1886 BUG_ON(!service_tree->count);
1887
Vivek Goyalb6508c12010-08-23 12:23:33 +02001888 if (!cfqd->cfq_slice_idle)
1889 return false;
1890
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001891 /* We never do for idle class queues. */
1892 if (prio == IDLE_WORKLOAD)
1893 return false;
1894
1895 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001896 if (cfq_cfqq_idle_window(cfqq) &&
1897 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001898 return true;
1899
1900 /*
1901 * Otherwise, we do only if they are the last ones
1902 * in their service tree.
1903 */
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001904 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01001905 return true;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001906 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1907 service_tree->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01001908 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001909}
1910
Jens Axboe6d048f52007-04-25 12:44:27 +02001911static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001912{
Jens Axboe17926692007-01-19 11:59:30 +11001913 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001914 struct cfq_io_context *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001915 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001916
Jens Axboea68bbdd2008-09-24 13:03:33 +02001917 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001918 * SSD device without seek penalty, disable idling. But only do so
1919 * for devices that support queuing, otherwise we still have a problem
1920 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001921 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001922 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001923 return;
1924
Jens Axboedd67d052006-06-21 09:36:18 +02001925 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001926 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001927
1928 /*
1929 * idle is disabled, either manually or by past process history
1930 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001931 if (!cfq_should_idle(cfqd, cfqq)) {
1932 /* no queue idling. Check for group idling */
1933 if (cfqd->cfq_group_idle)
1934 group_idle = cfqd->cfq_group_idle;
1935 else
1936 return;
1937 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001938
Jens Axboe22e2c502005-06-27 10:55:12 +02001939 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001940 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001941 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001942 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001943 return;
1944
1945 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001946 * task has exited, don't wait
1947 */
Jens Axboe206dc692006-03-28 13:03:44 +02001948 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001949 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001950 return;
1951
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001952 /*
1953 * If our average think time is larger than the remaining time
1954 * slice, then don't idle. This avoids overrunning the allotted
1955 * time slice.
1956 */
1957 if (sample_valid(cic->ttime_samples) &&
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001958 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1959 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1960 cic->ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001961 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001962 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001963
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001964 /* There are other queues in the group, don't do group idle */
1965 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1966 return;
1967
Jens Axboe3b181522005-06-27 10:56:24 +02001968 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001969
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001970 if (group_idle)
1971 sl = cfqd->cfq_group_idle;
1972 else
1973 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001974
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001975 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001976 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001977 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1978 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979}
1980
Jens Axboe498d3aa22007-04-26 12:54:48 +02001981/*
1982 * Move request from internal lists to the request queue dispatch list.
1983 */
Jens Axboe165125e2007-07-24 09:28:11 +02001984static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001986 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001987 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001988
Jens Axboe7b679132008-05-30 12:23:07 +02001989 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1990
Jeff Moyer06d21882009-09-11 17:08:59 +02001991 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001992 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001993 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001994 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001995 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001996
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001997 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001998 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001999 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02002000 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001}
2002
2003/*
2004 * return expired entry, or NULL to just start from scratch in rbtree
2005 */
Jens Axboefebffd62008-01-28 13:19:43 +01002006static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007{
Jens Axboe30996f42009-10-05 11:03:39 +02002008 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009
Jens Axboe3b181522005-06-27 10:56:24 +02002010 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002012
2013 cfq_mark_cfqq_fifo_expire(cfqq);
2014
Jens Axboe89850f72006-07-22 16:48:31 +02002015 if (list_empty(&cfqq->fifo))
2016 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017
Jens Axboe89850f72006-07-22 16:48:31 +02002018 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002019 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002020 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021
Jens Axboe30996f42009-10-05 11:03:39 +02002022 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002023 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024}
2025
Jens Axboe22e2c502005-06-27 10:55:12 +02002026static inline int
2027cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2028{
2029 const int base_rq = cfqd->cfq_slice_async_rq;
2030
2031 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2032
2033 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2034}
2035
2036/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002037 * Must be called with the queue_lock held.
2038 */
2039static int cfqq_process_refs(struct cfq_queue *cfqq)
2040{
2041 int process_refs, io_refs;
2042
2043 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002044 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002045 BUG_ON(process_refs < 0);
2046 return process_refs;
2047}
2048
2049static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2050{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002051 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002052 struct cfq_queue *__cfqq;
2053
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002054 /*
2055 * If there are no process references on the new_cfqq, then it is
2056 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2057 * chain may have dropped their last reference (not just their
2058 * last process reference).
2059 */
2060 if (!cfqq_process_refs(new_cfqq))
2061 return;
2062
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002063 /* Avoid a circular list and skip interim queue merges */
2064 while ((__cfqq = new_cfqq->new_cfqq)) {
2065 if (__cfqq == cfqq)
2066 return;
2067 new_cfqq = __cfqq;
2068 }
2069
2070 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002071 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002072 /*
2073 * If the process for the cfqq has gone away, there is no
2074 * sense in merging the queues.
2075 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002076 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002077 return;
2078
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002079 /*
2080 * Merge in the direction of the lesser amount of work.
2081 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002082 if (new_process_refs >= process_refs) {
2083 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002084 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002085 } else {
2086 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002087 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002088 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002089}
2090
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002091static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002092 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002093{
2094 struct cfq_queue *queue;
2095 int i;
2096 bool key_valid = false;
2097 unsigned long lowest_key = 0;
2098 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2099
Vivek Goyal65b32a52009-12-16 17:52:59 -05002100 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2101 /* select the one with lowest rb_key */
2102 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002103 if (queue &&
2104 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2105 lowest_key = queue->rb_key;
2106 cur_best = i;
2107 key_valid = true;
2108 }
2109 }
2110
2111 return cur_best;
2112}
2113
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002114static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002115{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002116 unsigned slice;
2117 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002118 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002119 unsigned group_slice;
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002120 enum wl_prio_t original_prio = cfqd->serving_prio;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002121
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002122 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002123 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002124 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002125 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002126 cfqd->serving_prio = BE_WORKLOAD;
2127 else {
2128 cfqd->serving_prio = IDLE_WORKLOAD;
2129 cfqd->workload_expires = jiffies + 1;
2130 return;
2131 }
2132
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002133 if (original_prio != cfqd->serving_prio)
2134 goto new_workload;
2135
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002136 /*
2137 * For RT and BE, we have to choose also the type
2138 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2139 * expiration time
2140 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002141 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002142 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002143
2144 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002145 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002146 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002147 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002148 return;
2149
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002150new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002151 /* otherwise select new workload type */
2152 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002153 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2154 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002155 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002156
2157 /*
2158 * the workload slice is computed as a fraction of target latency
2159 * proportional to the number of queues in that workload, over
2160 * all the queues in the same priority class
2161 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002162 group_slice = cfq_group_slice(cfqd, cfqg);
2163
2164 slice = group_slice * count /
2165 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2166 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002167
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002168 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2169 unsigned int tmp;
2170
2171 /*
2172 * Async queues are currently system wide. Just taking
2173 * proportion of queues with-in same group will lead to higher
2174 * async ratio system wide as generally root group is going
2175 * to have higher weight. A more accurate thing would be to
2176 * calculate system wide asnc/sync ratio.
2177 */
2178 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2179 tmp = tmp/cfqd->busy_queues;
2180 slice = min_t(unsigned, slice, tmp);
2181
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002182 /* async workload slice is scaled down according to
2183 * the sync/async slice ratio. */
2184 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002185 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002186 /* sync workload slice is at least 2 * cfq_slice_idle */
2187 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2188
2189 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002190 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002191 cfqd->workload_expires = jiffies + slice;
2192}
2193
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002194static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2195{
2196 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002197 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002198
2199 if (RB_EMPTY_ROOT(&st->rb))
2200 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002201 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002202 update_min_vdisktime(st);
2203 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002204}
2205
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002206static void cfq_choose_cfqg(struct cfq_data *cfqd)
2207{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002208 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2209
2210 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002211
2212 /* Restore the workload type data */
2213 if (cfqg->saved_workload_slice) {
2214 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2215 cfqd->serving_type = cfqg->saved_workload;
2216 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002217 } else
2218 cfqd->workload_expires = jiffies - 1;
2219
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002220 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002221}
2222
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002223/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002224 * Select a queue for service. If we have a current active queue,
2225 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002226 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002227static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002228{
Jens Axboea36e71f2009-04-15 12:15:11 +02002229 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002230
2231 cfqq = cfqd->active_queue;
2232 if (!cfqq)
2233 goto new_queue;
2234
Vivek Goyalf04a6422009-12-03 12:59:40 -05002235 if (!cfqd->rq_queued)
2236 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002237
2238 /*
2239 * We were waiting for group to get backlogged. Expire the queue
2240 */
2241 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2242 goto expire;
2243
Jens Axboe22e2c502005-06-27 10:55:12 +02002244 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002245 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002246 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002247 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2248 /*
2249 * If slice had not expired at the completion of last request
2250 * we might not have turned on wait_busy flag. Don't expire
2251 * the queue yet. Allow the group to get backlogged.
2252 *
2253 * The very fact that we have used the slice, that means we
2254 * have been idling all along on this queue and it should be
2255 * ok to wait for this request to complete.
2256 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002257 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2258 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2259 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002260 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002261 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002262 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002263 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002264
2265 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002266 * The active queue has requests and isn't expired, allow it to
2267 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002268 */
Jens Axboedd67d052006-06-21 09:36:18 +02002269 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002270 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002271
2272 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002273 * If another queue has a request waiting within our mean seek
2274 * distance, let it run. The expire code will check for close
2275 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002276 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002277 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002278 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002279 if (new_cfqq) {
2280 if (!cfqq->new_cfqq)
2281 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002282 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002283 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002284
2285 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002286 * No requests pending. If the active queue still has requests in
2287 * flight or is idling for a new request, allow either of these
2288 * conditions to happen (or time out) before selecting a new queue.
2289 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002290 if (timer_pending(&cfqd->idle_slice_timer)) {
2291 cfqq = NULL;
2292 goto keep_queue;
2293 }
2294
Shaohua Li8e1ac662010-11-08 15:01:04 +01002295 /*
2296 * This is a deep seek queue, but the device is much faster than
2297 * the queue can deliver, don't idle
2298 **/
2299 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2300 (cfq_cfqq_slice_new(cfqq) ||
2301 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2302 cfq_clear_cfqq_deep(cfqq);
2303 cfq_clear_cfqq_idle_window(cfqq);
2304 }
2305
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002306 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2307 cfqq = NULL;
2308 goto keep_queue;
2309 }
2310
2311 /*
2312 * If group idle is enabled and there are requests dispatched from
2313 * this group, wait for requests to complete.
2314 */
2315check_group_idle:
2316 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2317 && cfqq->cfqg->dispatched) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002318 cfqq = NULL;
2319 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002320 }
2321
Jens Axboe3b181522005-06-27 10:56:24 +02002322expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002323 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002324new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002325 /*
2326 * Current queue expired. Check if we have to switch to a new
2327 * service tree
2328 */
2329 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002330 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002331
Jens Axboea36e71f2009-04-15 12:15:11 +02002332 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002333keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002334 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002335}
2336
Jens Axboefebffd62008-01-28 13:19:43 +01002337static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002338{
2339 int dispatched = 0;
2340
2341 while (cfqq->next_rq) {
2342 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2343 dispatched++;
2344 }
2345
2346 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002347
2348 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002349 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002350 return dispatched;
2351}
2352
Jens Axboe498d3aa22007-04-26 12:54:48 +02002353/*
2354 * Drain our current requests. Used for barriers and when switching
2355 * io schedulers on-the-fly.
2356 */
Jens Axboed9e76202007-04-20 14:27:50 +02002357static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002358{
Jens Axboe08717142008-01-28 11:38:15 +01002359 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002360 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002361
Divyesh Shah3440c492010-04-09 09:29:57 +02002362 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002363 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002364 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2365 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002366 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002367 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002368
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002369 BUG_ON(cfqd->busy_queues);
2370
Jeff Moyer69237152009-06-12 15:29:30 +02002371 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002372 return dispatched;
2373}
2374
Shaohua Liabc3c742010-03-01 09:20:54 +01002375static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2376 struct cfq_queue *cfqq)
2377{
2378 /* the queue hasn't finished any request, can't estimate */
2379 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01002380 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002381 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2382 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01002383 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002384
Shaohua Lic1e44752010-11-08 15:01:02 +01002385 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01002386}
2387
Jens Axboe0b182d62009-10-06 20:49:37 +02002388static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002389{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002390 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391
Jens Axboe2f5cb732009-04-07 08:51:19 +02002392 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002393 * Drain async requests before we start sync IO
2394 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002395 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002396 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002397
2398 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002399 * If this is an async queue and we have sync IO in flight, let it wait
2400 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002401 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002402 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002403
Shaohua Liabc3c742010-03-01 09:20:54 +01002404 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002405 if (cfq_class_idle(cfqq))
2406 max_dispatch = 1;
2407
2408 /*
2409 * Does this cfqq already have too much IO in flight?
2410 */
2411 if (cfqq->dispatched >= max_dispatch) {
Shaohua Lief8a41d2011-03-07 09:26:29 +01002412 bool promote_sync = false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002413 /*
2414 * idle queue must always only have a single IO in flight
2415 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002416 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002417 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002418
Jens Axboe2f5cb732009-04-07 08:51:19 +02002419 /*
Li, Shaohuac4ade942011-03-23 08:30:34 +01002420 * If there is only one sync queue
2421 * we can ignore async queue here and give the sync
Shaohua Lief8a41d2011-03-07 09:26:29 +01002422 * queue no dispatch limit. The reason is a sync queue can
2423 * preempt async queue, limiting the sync queue doesn't make
2424 * sense. This is useful for aiostress test.
2425 */
Li, Shaohuac4ade942011-03-23 08:30:34 +01002426 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2427 promote_sync = true;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002428
2429 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002430 * We have other queues, don't allow more IO from this one
2431 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002432 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2433 !promote_sync)
Jens Axboe0b182d62009-10-06 20:49:37 +02002434 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002435
Jens Axboe2f5cb732009-04-07 08:51:19 +02002436 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002437 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002438 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002439 if (cfqd->busy_queues == 1 || promote_sync)
Shaohua Liabc3c742010-03-01 09:20:54 +01002440 max_dispatch = -1;
2441 else
2442 /*
2443 * Normally we start throttling cfqq when cfq_quantum/2
2444 * requests have been dispatched. But we can drive
2445 * deeper queue depths at the beginning of slice
2446 * subjected to upper limit of cfq_quantum.
2447 * */
2448 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002449 }
2450
2451 /*
2452 * Async queues must wait a bit before being allowed dispatch.
2453 * We also ramp up the dispatch depth gradually for async IO,
2454 * based on the last sync IO we serviced
2455 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002456 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002457 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002458 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002459
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002460 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002461 if (!depth && !cfqq->dispatched)
2462 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002463 if (depth < max_dispatch)
2464 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 }
2466
Jens Axboe0b182d62009-10-06 20:49:37 +02002467 /*
2468 * If we're below the current max, allow a dispatch
2469 */
2470 return cfqq->dispatched < max_dispatch;
2471}
2472
2473/*
2474 * Dispatch a request from cfqq, moving them to the request queue
2475 * dispatch list.
2476 */
2477static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2478{
2479 struct request *rq;
2480
2481 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2482
2483 if (!cfq_may_dispatch(cfqd, cfqq))
2484 return false;
2485
2486 /*
2487 * follow expired path, else get first next available
2488 */
2489 rq = cfq_check_fifo(cfqq);
2490 if (!rq)
2491 rq = cfqq->next_rq;
2492
2493 /*
2494 * insert request into driver dispatch list
2495 */
2496 cfq_dispatch_insert(cfqd->queue, rq);
2497
2498 if (!cfqd->active_cic) {
2499 struct cfq_io_context *cic = RQ_CIC(rq);
2500
2501 atomic_long_inc(&cic->ioc->refcount);
2502 cfqd->active_cic = cic;
2503 }
2504
2505 return true;
2506}
2507
2508/*
2509 * Find the cfqq that we need to service and move a request from that to the
2510 * dispatch list
2511 */
2512static int cfq_dispatch_requests(struct request_queue *q, int force)
2513{
2514 struct cfq_data *cfqd = q->elevator->elevator_data;
2515 struct cfq_queue *cfqq;
2516
2517 if (!cfqd->busy_queues)
2518 return 0;
2519
2520 if (unlikely(force))
2521 return cfq_forced_dispatch(cfqd);
2522
2523 cfqq = cfq_select_queue(cfqd);
2524 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002525 return 0;
2526
Jens Axboe2f5cb732009-04-07 08:51:19 +02002527 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002528 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002529 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002530 if (!cfq_dispatch_request(cfqd, cfqq))
2531 return 0;
2532
Jens Axboe2f5cb732009-04-07 08:51:19 +02002533 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002534 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002535
2536 /*
2537 * expire an async queue immediately if it has used up its slice. idle
2538 * queue always expire after 1 dispatch round.
2539 */
2540 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2541 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2542 cfq_class_idle(cfqq))) {
2543 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002544 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002545 }
2546
Shan Weib217a902009-09-01 10:06:42 +02002547 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002548 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549}
2550
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551/*
Jens Axboe5e705372006-07-13 12:39:25 +02002552 * task holds one reference to the queue, dropped when task exits. each rq
2553 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002555 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 * queue lock must be held here.
2557 */
2558static void cfq_put_queue(struct cfq_queue *cfqq)
2559{
Jens Axboe22e2c502005-06-27 10:55:12 +02002560 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05002561 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002562
Shaohua Li30d7b942011-01-07 08:46:59 +01002563 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564
Shaohua Li30d7b942011-01-07 08:46:59 +01002565 cfqq->ref--;
2566 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 return;
2568
Jens Axboe7b679132008-05-30 12:23:07 +02002569 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002571 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002572 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002574 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002575 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002576 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002577 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002578
Vivek Goyalf04a6422009-12-03 12:59:40 -05002579 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002581 cfq_put_cfqg(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582}
2583
Jens Axboed6de8be2008-05-28 14:46:59 +02002584/*
2585 * Must always be called with the rcu_read_lock() held
2586 */
Jens Axboe07416d22008-05-07 09:17:12 +02002587static void
2588__call_for_each_cic(struct io_context *ioc,
2589 void (*func)(struct io_context *, struct cfq_io_context *))
2590{
2591 struct cfq_io_context *cic;
2592 struct hlist_node *n;
2593
2594 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2595 func(ioc, cic);
2596}
2597
Jens Axboe4ac845a2008-01-24 08:44:49 +01002598/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002599 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002600 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002601static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002602call_for_each_cic(struct io_context *ioc,
2603 void (*func)(struct io_context *, struct cfq_io_context *))
2604{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002605 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002606 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002607 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002608}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002609
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002610static void cfq_cic_free_rcu(struct rcu_head *head)
2611{
2612 struct cfq_io_context *cic;
2613
2614 cic = container_of(head, struct cfq_io_context, rcu_head);
2615
2616 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002617 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002618
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002619 if (ioc_gone) {
2620 /*
2621 * CFQ scheduler is exiting, grab exit lock and check
2622 * the pending io context count. If it hits zero,
2623 * complete ioc_gone and set it back to NULL
2624 */
2625 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002626 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002627 complete(ioc_gone);
2628 ioc_gone = NULL;
2629 }
2630 spin_unlock(&ioc_gone_lock);
2631 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002632}
2633
2634static void cfq_cic_free(struct cfq_io_context *cic)
2635{
2636 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002637}
2638
2639static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2640{
2641 unsigned long flags;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002642 unsigned long dead_key = (unsigned long) cic->key;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002643
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002644 BUG_ON(!(dead_key & CIC_DEAD_KEY));
Jens Axboe4ac845a2008-01-24 08:44:49 +01002645
2646 spin_lock_irqsave(&ioc->lock, flags);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002647 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
Jens Axboeffc4e752008-02-19 10:02:29 +01002648 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002649 spin_unlock_irqrestore(&ioc->lock, flags);
2650
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002651 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002652}
2653
Jens Axboed6de8be2008-05-28 14:46:59 +02002654/*
2655 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2656 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2657 * and ->trim() which is called with the task lock held
2658 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002659static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002661 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002662 * ioc->refcount is zero here, or we are called from elv_unregister(),
2663 * so no more cic's are allowed to be linked into this ioc. So it
2664 * should be ok to iterate over the known list, we will see all cic's
2665 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002666 */
Jens Axboe07416d22008-05-07 09:17:12 +02002667 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668}
2669
Shaohua Lid02a2c02010-05-25 10:16:53 +02002670static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002671{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002672 struct cfq_queue *__cfqq, *next;
2673
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002674 /*
2675 * If this queue was scheduled to merge with another queue, be
2676 * sure to drop the reference taken on that queue (and others in
2677 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2678 */
2679 __cfqq = cfqq->new_cfqq;
2680 while (__cfqq) {
2681 if (__cfqq == cfqq) {
2682 WARN(1, "cfqq->new_cfqq loop detected\n");
2683 break;
2684 }
2685 next = __cfqq->new_cfqq;
2686 cfq_put_queue(__cfqq);
2687 __cfqq = next;
2688 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002689}
2690
2691static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2692{
2693 if (unlikely(cfqq == cfqd->active_queue)) {
2694 __cfq_slice_expired(cfqd, cfqq, 0);
2695 cfq_schedule_dispatch(cfqd);
2696 }
2697
2698 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002699
Jens Axboe89850f72006-07-22 16:48:31 +02002700 cfq_put_queue(cfqq);
2701}
2702
2703static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2704 struct cfq_io_context *cic)
2705{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002706 struct io_context *ioc = cic->ioc;
2707
Jens Axboefc463792006-08-29 09:05:44 +02002708 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002709
2710 /*
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002711 * Make sure dead mark is seen for dead queues
Jens Axboe4ac845a2008-01-24 08:44:49 +01002712 */
Jens Axboefc463792006-08-29 09:05:44 +02002713 smp_wmb();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002714 cic->key = cfqd_dead_key(cfqd);
Jens Axboefc463792006-08-29 09:05:44 +02002715
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002716 if (ioc->ioc_data == cic)
2717 rcu_assign_pointer(ioc->ioc_data, NULL);
2718
Jens Axboeff6657c2009-04-08 10:58:57 +02002719 if (cic->cfqq[BLK_RW_ASYNC]) {
2720 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2721 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002722 }
2723
Jens Axboeff6657c2009-04-08 10:58:57 +02002724 if (cic->cfqq[BLK_RW_SYNC]) {
2725 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2726 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002727 }
Jens Axboe89850f72006-07-22 16:48:31 +02002728}
2729
Jens Axboe4ac845a2008-01-24 08:44:49 +01002730static void cfq_exit_single_io_context(struct io_context *ioc,
2731 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002732{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002733 struct cfq_data *cfqd = cic_to_cfqd(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002734
Jens Axboe89850f72006-07-22 16:48:31 +02002735 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002736 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002737 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002738
Jens Axboe4ac845a2008-01-24 08:44:49 +01002739 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002740
2741 /*
2742 * Ensure we get a fresh copy of the ->key to prevent
2743 * race between exiting task and queue
2744 */
2745 smp_read_barrier_depends();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002746 if (cic->key == cfqd)
Jens Axboe62c1fe92008-12-15 21:19:25 +01002747 __cfq_exit_single_io_context(cfqd, cic);
2748
Jens Axboe4ac845a2008-01-24 08:44:49 +01002749 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002750 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002751}
2752
Jens Axboe498d3aa22007-04-26 12:54:48 +02002753/*
2754 * The process that ioc belongs to has exited, we need to clean up
2755 * and put the internal structures we have that belongs to that process.
2756 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002757static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002759 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760}
2761
Jens Axboe22e2c502005-06-27 10:55:12 +02002762static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002763cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764{
Jens Axboeb5deef92006-07-19 23:39:40 +02002765 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766
Christoph Lameter94f60302007-07-17 04:03:29 -07002767 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2768 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002770 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002771 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002772 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002773 cic->dtor = cfq_free_io_context;
2774 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002775 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 }
2777
2778 return cic;
2779}
2780
Jens Axboefd0928d2008-01-24 08:52:45 +01002781static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002782{
2783 struct task_struct *tsk = current;
2784 int ioprio_class;
2785
Jens Axboe3b181522005-06-27 10:56:24 +02002786 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002787 return;
2788
Jens Axboefd0928d2008-01-24 08:52:45 +01002789 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002790 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002791 default:
2792 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2793 case IOPRIO_CLASS_NONE:
2794 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002795 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002796 */
2797 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002798 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002799 break;
2800 case IOPRIO_CLASS_RT:
2801 cfqq->ioprio = task_ioprio(ioc);
2802 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2803 break;
2804 case IOPRIO_CLASS_BE:
2805 cfqq->ioprio = task_ioprio(ioc);
2806 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2807 break;
2808 case IOPRIO_CLASS_IDLE:
2809 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2810 cfqq->ioprio = 7;
2811 cfq_clear_cfqq_idle_window(cfqq);
2812 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002813 }
2814
2815 /*
2816 * keep track of original prio settings in case we have to temporarily
2817 * elevate the priority of this queue
2818 */
2819 cfqq->org_ioprio = cfqq->ioprio;
2820 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002821 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002822}
2823
Jens Axboefebffd62008-01-28 13:19:43 +01002824static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002825{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002826 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002827 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002828 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002829
Jens Axboecaaa5f92006-06-16 11:23:00 +02002830 if (unlikely(!cfqd))
2831 return;
2832
Jens Axboec1b707d2006-10-30 19:54:23 +01002833 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002834
Jens Axboeff6657c2009-04-08 10:58:57 +02002835 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002836 if (cfqq) {
2837 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002838 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2839 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002840 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002841 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002842 cfq_put_queue(cfqq);
2843 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002844 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002845
Jens Axboeff6657c2009-04-08 10:58:57 +02002846 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002847 if (cfqq)
2848 cfq_mark_cfqq_prio_changed(cfqq);
2849
Jens Axboec1b707d2006-10-30 19:54:23 +01002850 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002851}
2852
Jens Axboefc463792006-08-29 09:05:44 +02002853static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002855 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002856 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857}
2858
Jens Axboed5036d72009-06-26 10:44:34 +02002859static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002860 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002861{
2862 RB_CLEAR_NODE(&cfqq->rb_node);
2863 RB_CLEAR_NODE(&cfqq->p_node);
2864 INIT_LIST_HEAD(&cfqq->fifo);
2865
Shaohua Li30d7b942011-01-07 08:46:59 +01002866 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02002867 cfqq->cfqd = cfqd;
2868
2869 cfq_mark_cfqq_prio_changed(cfqq);
2870
2871 if (is_sync) {
2872 if (!cfq_class_idle(cfqq))
2873 cfq_mark_cfqq_idle_window(cfqq);
2874 cfq_mark_cfqq_sync(cfqq);
2875 }
2876 cfqq->pid = pid;
2877}
2878
Vivek Goyal246103332009-12-03 12:59:51 -05002879#ifdef CONFIG_CFQ_GROUP_IOSCHED
2880static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2881{
2882 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002883 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal246103332009-12-03 12:59:51 -05002884 unsigned long flags;
2885 struct request_queue *q;
2886
2887 if (unlikely(!cfqd))
2888 return;
2889
2890 q = cfqd->queue;
2891
2892 spin_lock_irqsave(q->queue_lock, flags);
2893
2894 if (sync_cfqq) {
2895 /*
2896 * Drop reference to sync queue. A new sync queue will be
2897 * assigned in new group upon arrival of a fresh request.
2898 */
2899 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2900 cic_set_cfqq(cic, NULL, 1);
2901 cfq_put_queue(sync_cfqq);
2902 }
2903
2904 spin_unlock_irqrestore(q->queue_lock, flags);
2905}
2906
2907static void cfq_ioc_set_cgroup(struct io_context *ioc)
2908{
2909 call_for_each_cic(ioc, changed_cgroup);
2910 ioc->cgroup_changed = 0;
2911}
2912#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2913
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002915cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002916 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002919 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002920 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921
2922retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002923 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002924 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002925 /* cic always exists here */
2926 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927
Jens Axboe6118b702009-06-30 09:34:12 +02002928 /*
2929 * Always try a new alloc if we fell back to the OOM cfqq
2930 * originally, since it should just be a temporary situation.
2931 */
2932 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2933 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 if (new_cfqq) {
2935 cfqq = new_cfqq;
2936 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002937 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002939 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002940 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002941 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002943 if (new_cfqq)
2944 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002945 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002946 cfqq = kmem_cache_alloc_node(cfq_pool,
2947 gfp_mask | __GFP_ZERO,
2948 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002949 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950
Jens Axboe6118b702009-06-30 09:34:12 +02002951 if (cfqq) {
2952 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2953 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002954 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002955 cfq_log_cfqq(cfqd, cfqq, "alloced");
2956 } else
2957 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958 }
2959
2960 if (new_cfqq)
2961 kmem_cache_free(cfq_pool, new_cfqq);
2962
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 return cfqq;
2964}
2965
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002966static struct cfq_queue **
2967cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2968{
Jens Axboefe094d92008-01-31 13:08:54 +01002969 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002970 case IOPRIO_CLASS_RT:
2971 return &cfqd->async_cfqq[0][ioprio];
2972 case IOPRIO_CLASS_BE:
2973 return &cfqd->async_cfqq[1][ioprio];
2974 case IOPRIO_CLASS_IDLE:
2975 return &cfqd->async_idle_cfqq;
2976 default:
2977 BUG();
2978 }
2979}
2980
Jens Axboe15c31be2007-07-10 13:43:25 +02002981static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002982cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002983 gfp_t gfp_mask)
2984{
Jens Axboefd0928d2008-01-24 08:52:45 +01002985 const int ioprio = task_ioprio(ioc);
2986 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002987 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002988 struct cfq_queue *cfqq = NULL;
2989
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002990 if (!is_sync) {
2991 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2992 cfqq = *async_cfqq;
2993 }
2994
Jens Axboe6118b702009-06-30 09:34:12 +02002995 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002996 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002997
2998 /*
2999 * pin the queue now that it's allocated, scheduler exit will prune it
3000 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003001 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01003002 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003003 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02003004 }
3005
Shaohua Li30d7b942011-01-07 08:46:59 +01003006 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02003007 return cfqq;
3008}
3009
Jens Axboe498d3aa22007-04-26 12:54:48 +02003010/*
3011 * We drop cfq io contexts lazily, so we may find a dead one.
3012 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003013static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01003014cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3015 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003016{
Jens Axboe4ac845a2008-01-24 08:44:49 +01003017 unsigned long flags;
3018
Jens Axboefc463792006-08-29 09:05:44 +02003019 WARN_ON(!list_empty(&cic->queue_list));
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003020 BUG_ON(cic->key != cfqd_dead_key(cfqd));
Jens Axboe597bc482007-04-24 21:23:53 +02003021
Jens Axboe4ac845a2008-01-24 08:44:49 +01003022 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02003023
Fabio Checconi4faa3c82008-04-10 08:28:01 +02003024 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003025
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003026 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
Jens Axboeffc4e752008-02-19 10:02:29 +01003027 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003028 spin_unlock_irqrestore(&ioc->lock, flags);
3029
3030 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003031}
3032
Jens Axboee2d74ac2006-03-28 08:59:01 +02003033static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01003034cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003035{
Jens Axboee2d74ac2006-03-28 08:59:01 +02003036 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003037 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003038
Vasily Tarasov91fac312007-04-25 12:29:51 +02003039 if (unlikely(!ioc))
3040 return NULL;
3041
Jens Axboed6de8be2008-05-28 14:46:59 +02003042 rcu_read_lock();
3043
Jens Axboe597bc482007-04-24 21:23:53 +02003044 /*
3045 * we maintain a last-hit cache, to avoid browsing over the tree
3046 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003047 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02003048 if (cic && cic->key == cfqd) {
3049 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02003050 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003051 }
Jens Axboe597bc482007-04-24 21:23:53 +02003052
Jens Axboe4ac845a2008-01-24 08:44:49 +01003053 do {
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003054 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003055 rcu_read_unlock();
3056 if (!cic)
3057 break;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003058 if (unlikely(cic->key != cfqd)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01003059 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003060 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01003061 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003062 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003063
Jens Axboed6de8be2008-05-28 14:46:59 +02003064 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003065 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003066 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003067 break;
3068 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003069
Jens Axboe4ac845a2008-01-24 08:44:49 +01003070 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003071}
3072
Jens Axboe4ac845a2008-01-24 08:44:49 +01003073/*
3074 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3075 * the process specific cfq io context when entered from the block layer.
3076 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3077 */
Jens Axboefebffd62008-01-28 13:19:43 +01003078static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3079 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003080{
Jens Axboe0261d682006-10-30 19:07:48 +01003081 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003082 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003083
Jens Axboe4ac845a2008-01-24 08:44:49 +01003084 ret = radix_tree_preload(gfp_mask);
3085 if (!ret) {
3086 cic->ioc = ioc;
3087 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003088
Jens Axboe4ac845a2008-01-24 08:44:49 +01003089 spin_lock_irqsave(&ioc->lock, flags);
3090 ret = radix_tree_insert(&ioc->radix_root,
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003091 cfqd->cic_index, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01003092 if (!ret)
3093 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003094 spin_unlock_irqrestore(&ioc->lock, flags);
3095
3096 radix_tree_preload_end();
3097
3098 if (!ret) {
3099 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3100 list_add(&cic->queue_list, &cfqd->cic_list);
3101 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003102 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003103 }
3104
Jens Axboe4ac845a2008-01-24 08:44:49 +01003105 if (ret)
3106 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02003107
Jens Axboe4ac845a2008-01-24 08:44:49 +01003108 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003109}
3110
Jens Axboe22e2c502005-06-27 10:55:12 +02003111/*
3112 * Setup general io context and cfq io context. There can be several cfq
3113 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02003114 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02003115 */
3116static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02003117cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118{
Jens Axboe22e2c502005-06-27 10:55:12 +02003119 struct io_context *ioc = NULL;
3120 struct cfq_io_context *cic;
3121
3122 might_sleep_if(gfp_mask & __GFP_WAIT);
3123
Jens Axboeb5deef92006-07-19 23:39:40 +02003124 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02003125 if (!ioc)
3126 return NULL;
3127
Jens Axboe4ac845a2008-01-24 08:44:49 +01003128 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003129 if (cic)
3130 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02003131
Jens Axboee2d74ac2006-03-28 08:59:01 +02003132 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3133 if (cic == NULL)
3134 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02003135
Jens Axboe4ac845a2008-01-24 08:44:49 +01003136 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3137 goto err_free;
3138
Jens Axboe22e2c502005-06-27 10:55:12 +02003139out:
Jens Axboefc463792006-08-29 09:05:44 +02003140 smp_read_barrier_depends();
3141 if (unlikely(ioc->ioprio_changed))
3142 cfq_ioc_set_ioprio(ioc);
3143
Vivek Goyal246103332009-12-03 12:59:51 -05003144#ifdef CONFIG_CFQ_GROUP_IOSCHED
3145 if (unlikely(ioc->cgroup_changed))
3146 cfq_ioc_set_cgroup(ioc);
3147#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003148 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003149err_free:
3150 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003151err:
3152 put_io_context(ioc);
3153 return NULL;
3154}
3155
3156static void
3157cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3158{
Jens Axboeaaf12282007-01-19 11:30:16 +11003159 unsigned long elapsed = jiffies - cic->last_end_request;
3160 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003161
3162 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3163 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3164 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3165}
3166
Jens Axboe206dc692006-03-28 13:03:44 +02003167static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003168cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003169 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003170{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003171 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003172 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003173 if (cfqq->last_request_pos) {
3174 if (cfqq->last_request_pos < blk_rq_pos(rq))
3175 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3176 else
3177 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3178 }
Jens Axboe206dc692006-03-28 13:03:44 +02003179
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003180 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003181 if (blk_queue_nonrot(cfqd->queue))
3182 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3183 else
3184 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003185}
Jens Axboe22e2c502005-06-27 10:55:12 +02003186
3187/*
3188 * Disable idle window if the process thinks too long or seeks so much that
3189 * it doesn't matter
3190 */
3191static void
3192cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3193 struct cfq_io_context *cic)
3194{
Jens Axboe7b679132008-05-30 12:23:07 +02003195 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003196
Jens Axboe08717142008-01-28 11:38:15 +01003197 /*
3198 * Don't idle for async or idle io prio class
3199 */
3200 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003201 return;
3202
Jens Axboec265a7f2008-06-26 13:49:33 +02003203 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003204
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003205 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3206 cfq_mark_cfqq_deep(cfqq);
3207
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003208 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3209 enable_idle = 0;
3210 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003211 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003212 enable_idle = 0;
3213 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003214 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003215 enable_idle = 0;
3216 else
3217 enable_idle = 1;
3218 }
3219
Jens Axboe7b679132008-05-30 12:23:07 +02003220 if (old_idle != enable_idle) {
3221 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3222 if (enable_idle)
3223 cfq_mark_cfqq_idle_window(cfqq);
3224 else
3225 cfq_clear_cfqq_idle_window(cfqq);
3226 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003227}
3228
Jens Axboe22e2c502005-06-27 10:55:12 +02003229/*
3230 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3231 * no or if we aren't sure, a 1 will cause a preempt.
3232 */
Jens Axboea6151c32009-10-07 20:02:57 +02003233static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003234cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003235 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003236{
Jens Axboe6d048f52007-04-25 12:44:27 +02003237 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003238
Jens Axboe6d048f52007-04-25 12:44:27 +02003239 cfqq = cfqd->active_queue;
3240 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003241 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003242
Jens Axboe6d048f52007-04-25 12:44:27 +02003243 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003244 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003245
3246 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003247 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003248
Jens Axboe22e2c502005-06-27 10:55:12 +02003249 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003250 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3251 */
3252 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3253 return false;
3254
3255 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003256 * if the new request is sync, but the currently running queue is
3257 * not, let the sync request have priority.
3258 */
Jens Axboe5e705372006-07-13 12:39:25 +02003259 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003260 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003261
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003262 if (new_cfqq->cfqg != cfqq->cfqg)
3263 return false;
3264
3265 if (cfq_slice_used(cfqq))
3266 return true;
3267
3268 /* Allow preemption only if we are idling on sync-noidle tree */
3269 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3270 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3271 new_cfqq->service_tree->count == 2 &&
3272 RB_EMPTY_ROOT(&cfqq->sort_list))
3273 return true;
3274
Jens Axboe374f84a2006-07-23 01:42:19 +02003275 /*
3276 * So both queues are sync. Let the new request get disk time if
3277 * it's a metadata request and the current queue is doing regular IO.
3278 */
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003279 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003280 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003281
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003282 /*
3283 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3284 */
3285 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003286 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003287
Shaohua Lid2d59e12010-11-08 15:01:03 +01003288 /* An idle queue should not be idle now for some reason */
3289 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3290 return true;
3291
Jens Axboe1e3335d2007-02-14 19:59:49 +01003292 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003293 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003294
3295 /*
3296 * if this request is as-good as one we would expect from the
3297 * current cfqq, let it preempt
3298 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003299 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003300 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003301
Jens Axboea6151c32009-10-07 20:02:57 +02003302 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003303}
3304
3305/*
3306 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3307 * let it have half of its nominal slice.
3308 */
3309static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3310{
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003311 struct cfq_queue *old_cfqq = cfqd->active_queue;
3312
Jens Axboe7b679132008-05-30 12:23:07 +02003313 cfq_log_cfqq(cfqd, cfqq, "preempt");
Vivek Goyale5ff0822010-04-26 19:25:11 +02003314 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003315
Jens Axboebf572252006-07-19 20:29:12 +02003316 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003317 * workload type is changed, don't save slice, otherwise preempt
3318 * doesn't happen
3319 */
3320 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3321 cfqq->cfqg->saved_workload_slice = 0;
3322
3323 /*
Jens Axboebf572252006-07-19 20:29:12 +02003324 * Put the new queue at the front of the of the current list,
3325 * so we know that it will be selected next.
3326 */
3327 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003328
3329 cfq_service_tree_add(cfqd, cfqq, 1);
Justin TerAvesteda5e0c2011-03-22 21:26:49 +01003330
Justin TerAvest62a37f62011-03-23 08:25:44 +01003331 cfqq->slice_end = 0;
3332 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003333}
3334
3335/*
Jens Axboe5e705372006-07-13 12:39:25 +02003336 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003337 * something we should do about it
3338 */
3339static void
Jens Axboe5e705372006-07-13 12:39:25 +02003340cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3341 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003342{
Jens Axboe5e705372006-07-13 12:39:25 +02003343 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003344
Aaron Carroll45333d52008-08-26 15:52:36 +02003345 cfqd->rq_queued++;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003346 if (rq->cmd_flags & REQ_META)
Jens Axboe374f84a2006-07-23 01:42:19 +02003347 cfqq->meta_pending++;
3348
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003349 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003350 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003351 cfq_update_idle_window(cfqd, cfqq, cic);
3352
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003353 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003354
3355 if (cfqq == cfqd->active_queue) {
3356 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003357 * Remember that we saw a request from this process, but
3358 * don't start queuing just yet. Otherwise we risk seeing lots
3359 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003360 * and merging. If the request is already larger than a single
3361 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003362 * merging is already done. Ditto for a busy system that
3363 * has other work pending, don't risk delaying until the
3364 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003365 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003366 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003367 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3368 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003369 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003370 cfq_clear_cfqq_wait_request(cfqq);
Tejun Heo1654e742011-03-02 08:48:05 -05003371 __blk_run_queue(cfqd->queue, false);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003372 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003373 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003374 &cfqq->cfqg->blkg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003375 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003376 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003377 }
Jens Axboe5e705372006-07-13 12:39:25 +02003378 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003379 /*
3380 * not the active queue - expire current slice if it is
3381 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003382 * has some old slice time left and is of higher priority or
3383 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003384 */
3385 cfq_preempt_queue(cfqd, cfqq);
Tejun Heo1654e742011-03-02 08:48:05 -05003386 __blk_run_queue(cfqd->queue, false);
Jens Axboe22e2c502005-06-27 10:55:12 +02003387 }
3388}
3389
Jens Axboe165125e2007-07-24 09:28:11 +02003390static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003391{
Jens Axboeb4878f22005-10-20 16:42:29 +02003392 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003393 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003394
Jens Axboe7b679132008-05-30 12:23:07 +02003395 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003396 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397
Jens Axboe30996f42009-10-05 11:03:39 +02003398 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003399 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003400 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003401 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003402 &cfqd->serving_group->blkg, rq_data_dir(rq),
3403 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003404 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405}
3406
Aaron Carroll45333d52008-08-26 15:52:36 +02003407/*
3408 * Update hw_tag based on peak queue depth over 50 samples under
3409 * sufficient load.
3410 */
3411static void cfq_update_hw_tag(struct cfq_data *cfqd)
3412{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003413 struct cfq_queue *cfqq = cfqd->active_queue;
3414
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003415 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3416 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003417
3418 if (cfqd->hw_tag == 1)
3419 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003420
3421 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003422 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003423 return;
3424
Shaohua Li1a1238a2009-10-27 08:46:23 +01003425 /*
3426 * If active queue hasn't enough requests and can idle, cfq might not
3427 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3428 * case
3429 */
3430 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3431 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003432 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003433 return;
3434
Aaron Carroll45333d52008-08-26 15:52:36 +02003435 if (cfqd->hw_tag_samples++ < 50)
3436 return;
3437
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003438 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003439 cfqd->hw_tag = 1;
3440 else
3441 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003442}
3443
Vivek Goyal7667aa02009-12-08 17:52:58 -05003444static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3445{
3446 struct cfq_io_context *cic = cfqd->active_cic;
3447
Justin TerAvest02a8f012011-02-09 14:20:03 +01003448 /* If the queue already has requests, don't wait */
3449 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3450 return false;
3451
Vivek Goyal7667aa02009-12-08 17:52:58 -05003452 /* If there are other queues in the group, don't wait */
3453 if (cfqq->cfqg->nr_cfqq > 1)
3454 return false;
3455
3456 if (cfq_slice_used(cfqq))
3457 return true;
3458
3459 /* if slice left is less than think time, wait busy */
3460 if (cic && sample_valid(cic->ttime_samples)
3461 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3462 return true;
3463
3464 /*
3465 * If think times is less than a jiffy than ttime_mean=0 and above
3466 * will not be true. It might happen that slice has not expired yet
3467 * but will expire soon (4-5 ns) during select_queue(). To cover the
3468 * case where think time is less than a jiffy, mark the queue wait
3469 * busy if only 1 jiffy is left in the slice.
3470 */
3471 if (cfqq->slice_end - jiffies == 1)
3472 return true;
3473
3474 return false;
3475}
3476
Jens Axboe165125e2007-07-24 09:28:11 +02003477static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478{
Jens Axboe5e705372006-07-13 12:39:25 +02003479 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003480 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003481 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003482 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483
Jens Axboeb4878f22005-10-20 16:42:29 +02003484 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003485 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3486 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487
Aaron Carroll45333d52008-08-26 15:52:36 +02003488 cfq_update_hw_tag(cfqd);
3489
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003490 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003491 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003492 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003493 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003494 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003495 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3496 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3497 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003499 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003500
Vivek Goyal365722b2009-10-03 15:21:27 +02003501 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003502 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003503 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3504 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003505 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003506
3507 /*
3508 * If this is the active queue, check if it needs to be expired,
3509 * or if we want to idle in case it has no pending requests.
3510 */
3511 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003512 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3513
Jens Axboe44f7c162007-01-19 11:51:58 +11003514 if (cfq_cfqq_slice_new(cfqq)) {
3515 cfq_set_prio_slice(cfqd, cfqq);
3516 cfq_clear_cfqq_slice_new(cfqq);
3517 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003518
3519 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003520 * Should we wait for next request to come in before we expire
3521 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003522 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003523 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003524 unsigned long extend_sl = cfqd->cfq_slice_idle;
3525 if (!cfqd->cfq_slice_idle)
3526 extend_sl = cfqd->cfq_group_idle;
3527 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003528 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003529 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003530 }
3531
Jens Axboea36e71f2009-04-15 12:15:11 +02003532 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003533 * Idling is not enabled on:
3534 * - expired queues
3535 * - idle-priority queues
3536 * - async queues
3537 * - queues with still some requests queued
3538 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003539 */
Jens Axboe08717142008-01-28 11:38:15 +01003540 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003541 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003542 else if (sync && cfqq_empty &&
3543 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003544 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003545 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003546 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003547
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003548 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003549 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550}
3551
Jens Axboe22e2c502005-06-27 10:55:12 +02003552/*
3553 * we temporarily boost lower priority queues if they are holding fs exclusive
3554 * resources. they are boosted to normal prio (CLASS_BE/4)
3555 */
3556static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557{
Jens Axboe22e2c502005-06-27 10:55:12 +02003558 if (has_fs_excl()) {
3559 /*
3560 * boost idle prio on transactions that would lock out other
3561 * users of the filesystem
3562 */
3563 if (cfq_class_idle(cfqq))
3564 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3565 if (cfqq->ioprio > IOPRIO_NORM)
3566 cfqq->ioprio = IOPRIO_NORM;
3567 } else {
3568 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003569 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003570 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003571 cfqq->ioprio_class = cfqq->org_ioprio_class;
3572 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003574}
3575
Jens Axboe89850f72006-07-22 16:48:31 +02003576static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003577{
Jens Axboe1b379d82009-08-11 08:26:11 +02003578 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003579 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003580 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003581 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003582
3583 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003584}
3585
Jens Axboe165125e2007-07-24 09:28:11 +02003586static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003587{
3588 struct cfq_data *cfqd = q->elevator->elevator_data;
3589 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003590 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003591 struct cfq_queue *cfqq;
3592
3593 /*
3594 * don't force setup of a queue from here, as a call to may_queue
3595 * does not necessarily imply that a request actually will be queued.
3596 * so just lookup a possibly existing queue, or return 'may queue'
3597 * if that fails
3598 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003599 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003600 if (!cic)
3601 return ELV_MQUEUE_MAY;
3602
Jens Axboeb0b78f82009-04-08 10:56:08 +02003603 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003604 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003605 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003606 cfq_prio_boost(cfqq);
3607
Jens Axboe89850f72006-07-22 16:48:31 +02003608 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003609 }
3610
3611 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612}
3613
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614/*
3615 * queue lock held here
3616 */
Jens Axboebb37b942006-12-01 10:42:33 +01003617static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618{
Jens Axboe5e705372006-07-13 12:39:25 +02003619 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003620
Jens Axboe5e705372006-07-13 12:39:25 +02003621 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003622 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623
Jens Axboe22e2c502005-06-27 10:55:12 +02003624 BUG_ON(!cfqq->allocated[rw]);
3625 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626
Jens Axboe5e705372006-07-13 12:39:25 +02003627 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628
Mike Snitzerc1867942011-02-11 11:08:00 +01003629 rq->elevator_private[0] = NULL;
3630 rq->elevator_private[1] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003632 /* Put down rq reference on cfqg */
3633 cfq_put_cfqg(RQ_CFQG(rq));
Mike Snitzerc1867942011-02-11 11:08:00 +01003634 rq->elevator_private[2] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003635
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 cfq_put_queue(cfqq);
3637 }
3638}
3639
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003640static struct cfq_queue *
3641cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3642 struct cfq_queue *cfqq)
3643{
3644 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3645 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003646 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003647 cfq_put_queue(cfqq);
3648 return cic_to_cfqq(cic, 1);
3649}
3650
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003651/*
3652 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3653 * was the last process referring to said cfqq.
3654 */
3655static struct cfq_queue *
3656split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3657{
3658 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003659 cfqq->pid = current->pid;
3660 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003661 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003662 return cfqq;
3663 }
3664
3665 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003666
3667 cfq_put_cooperator(cfqq);
3668
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003669 cfq_put_queue(cfqq);
3670 return NULL;
3671}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003672/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003673 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003675static int
Jens Axboe165125e2007-07-24 09:28:11 +02003676cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677{
3678 struct cfq_data *cfqd = q->elevator->elevator_data;
3679 struct cfq_io_context *cic;
3680 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003681 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003682 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683 unsigned long flags;
3684
3685 might_sleep_if(gfp_mask & __GFP_WAIT);
3686
Jens Axboee2d74ac2006-03-28 08:59:01 +02003687 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003688
Linus Torvalds1da177e2005-04-16 15:20:36 -07003689 spin_lock_irqsave(q->queue_lock, flags);
3690
Jens Axboe22e2c502005-06-27 10:55:12 +02003691 if (!cic)
3692 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003694new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003695 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003696 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003697 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003698 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003699 } else {
3700 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003701 * If the queue was seeky for too long, break it apart.
3702 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003703 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003704 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3705 cfqq = split_cfqq(cic, cfqq);
3706 if (!cfqq)
3707 goto new_queue;
3708 }
3709
3710 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003711 * Check to see if this queue is scheduled to merge with
3712 * another, closely cooperating queue. The merging of
3713 * queues happens here as it must be done in process context.
3714 * The reference on new_cfqq was taken in merge_cfqqs.
3715 */
3716 if (cfqq->new_cfqq)
3717 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003718 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719
3720 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02003721
Jens Axboe6fae9c22011-03-01 15:04:39 -05003722 cfqq->ref++;
Mike Snitzerc1867942011-02-11 11:08:00 +01003723 rq->elevator_private[0] = cic;
3724 rq->elevator_private[1] = cfqq;
3725 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
Jens Axboe93803e02011-03-07 08:59:06 +01003726 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe5e705372006-07-13 12:39:25 +02003727 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003728
Jens Axboe22e2c502005-06-27 10:55:12 +02003729queue_fail:
3730 if (cic)
3731 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003732
Jens Axboe23e018a2009-10-05 08:52:35 +02003733 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003735 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736 return 1;
3737}
3738
David Howells65f27f32006-11-22 14:55:48 +00003739static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003740{
David Howells65f27f32006-11-22 14:55:48 +00003741 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003742 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003743 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003744
Jens Axboe40bb54d2009-04-15 12:11:10 +02003745 spin_lock_irq(q->queue_lock);
Tejun Heo1654e742011-03-02 08:48:05 -05003746 __blk_run_queue(cfqd->queue, false);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003747 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003748}
3749
3750/*
3751 * Timer running if the active_queue is currently idling inside its time slice
3752 */
3753static void cfq_idle_slice_timer(unsigned long data)
3754{
3755 struct cfq_data *cfqd = (struct cfq_data *) data;
3756 struct cfq_queue *cfqq;
3757 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003758 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003759
Jens Axboe7b679132008-05-30 12:23:07 +02003760 cfq_log(cfqd, "idle timer fired");
3761
Jens Axboe22e2c502005-06-27 10:55:12 +02003762 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3763
Jens Axboefe094d92008-01-31 13:08:54 +01003764 cfqq = cfqd->active_queue;
3765 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003766 timed_out = 0;
3767
Jens Axboe22e2c502005-06-27 10:55:12 +02003768 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003769 * We saw a request before the queue expired, let it through
3770 */
3771 if (cfq_cfqq_must_dispatch(cfqq))
3772 goto out_kick;
3773
3774 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003775 * expired
3776 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003777 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003778 goto expire;
3779
3780 /*
3781 * only expire and reinvoke request handler, if there are
3782 * other queues with pending requests
3783 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003784 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003785 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003786
3787 /*
3788 * not expired and it has a request pending, let it dispatch
3789 */
Jens Axboe75e50982009-04-07 08:56:14 +02003790 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003791 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003792
3793 /*
3794 * Queue depth flag is reset only when the idle didn't succeed
3795 */
3796 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003797 }
3798expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003799 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003800out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003801 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003802out_cont:
3803 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3804}
3805
Jens Axboe3b181522005-06-27 10:56:24 +02003806static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3807{
3808 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003809 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003810}
Jens Axboe22e2c502005-06-27 10:55:12 +02003811
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003812static void cfq_put_async_queues(struct cfq_data *cfqd)
3813{
3814 int i;
3815
3816 for (i = 0; i < IOPRIO_BE_NR; i++) {
3817 if (cfqd->async_cfqq[0][i])
3818 cfq_put_queue(cfqd->async_cfqq[0][i]);
3819 if (cfqd->async_cfqq[1][i])
3820 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003821 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003822
3823 if (cfqd->async_idle_cfqq)
3824 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003825}
3826
Jens Axboebb729bc2009-12-06 09:54:19 +01003827static void cfq_cfqd_free(struct rcu_head *head)
3828{
3829 kfree(container_of(head, struct cfq_data, rcu));
3830}
3831
Jens Axboeb374d182008-10-31 10:05:07 +01003832static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833{
Jens Axboe22e2c502005-06-27 10:55:12 +02003834 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003835 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003836
Jens Axboe3b181522005-06-27 10:56:24 +02003837 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003838
Al Virod9ff4182006-03-18 13:51:22 -05003839 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003840
Al Virod9ff4182006-03-18 13:51:22 -05003841 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003842 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003843
3844 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003845 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3846 struct cfq_io_context,
3847 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003848
3849 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003850 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003851
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003852 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003853 cfq_release_cfq_groups(cfqd);
Vivek Goyale98ef892010-06-18 10:39:47 -04003854 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003855
Al Virod9ff4182006-03-18 13:51:22 -05003856 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003857
3858 cfq_shutdown_timer_wq(cfqd);
3859
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003860 spin_lock(&cic_index_lock);
3861 ida_remove(&cic_index_ida, cfqd->cic_index);
3862 spin_unlock(&cic_index_lock);
3863
Vivek Goyalb1c35762009-12-03 12:59:47 -05003864 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003865 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866}
3867
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003868static int cfq_alloc_cic_index(void)
3869{
3870 int index, error;
3871
3872 do {
3873 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3874 return -ENOMEM;
3875
3876 spin_lock(&cic_index_lock);
3877 error = ida_get_new(&cic_index_ida, &index);
3878 spin_unlock(&cic_index_lock);
3879 if (error && error != -EAGAIN)
3880 return error;
3881 } while (error);
3882
3883 return index;
3884}
3885
Jens Axboe165125e2007-07-24 09:28:11 +02003886static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887{
3888 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003889 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003890 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003891 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003893 i = cfq_alloc_cic_index();
3894 if (i < 0)
3895 return NULL;
3896
Christoph Lameter94f60302007-07-17 04:03:29 -07003897 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003899 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900
Shaohua Li30d7b942011-01-07 08:46:59 +01003901 /*
3902 * Don't need take queue_lock in the routine, since we are
3903 * initializing the ioscheduler, and nobody is using cfqd
3904 */
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003905 cfqd->cic_index = i;
3906
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003907 /* Init root service tree */
3908 cfqd->grp_service_tree = CFQ_RB_ROOT;
3909
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003910 /* Init root group */
3911 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003912 for_each_cfqg_st(cfqg, i, j, st)
3913 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003914 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003915
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003916 /* Give preference to root group over other groups */
3917 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3918
Vivek Goyal25fb5162009-12-03 12:59:46 -05003919#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003920 /*
3921 * Take a reference to root group which we never drop. This is just
3922 * to make sure that cfq_put_cfqg() does not try to kfree root group
3923 */
Shaohua Li329a6782011-01-07 08:48:28 +01003924 cfqg->ref = 1;
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003925 rcu_read_lock();
Vivek Goyale98ef892010-06-18 10:39:47 -04003926 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3927 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003928 rcu_read_unlock();
Vivek Goyal25fb5162009-12-03 12:59:46 -05003929#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003930 /*
3931 * Not strictly needed (since RB_ROOT just clears the node and we
3932 * zeroed cfqd on alloc), but better be safe in case someone decides
3933 * to add magic to the rb code
3934 */
3935 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3936 cfqd->prio_trees[i] = RB_ROOT;
3937
Jens Axboe6118b702009-06-30 09:34:12 +02003938 /*
3939 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3940 * Grab a permanent reference to it, so that the normal code flow
3941 * will not attempt to free it.
3942 */
3943 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01003944 cfqd->oom_cfqq.ref++;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003945 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003946
Al Virod9ff4182006-03-18 13:51:22 -05003947 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948
Linus Torvalds1da177e2005-04-16 15:20:36 -07003949 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950
Jens Axboe22e2c502005-06-27 10:55:12 +02003951 init_timer(&cfqd->idle_slice_timer);
3952 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3953 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3954
Jens Axboe23e018a2009-10-05 08:52:35 +02003955 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003956
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003958 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3959 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960 cfqd->cfq_back_max = cfq_back_max;
3961 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003962 cfqd->cfq_slice[0] = cfq_slice_async;
3963 cfqd->cfq_slice[1] = cfq_slice_sync;
3964 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3965 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003966 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003967 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003968 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003969 /*
3970 * we optimistically start assuming sync ops weren't delayed in last
3971 * second, in order to have larger depth for async operations.
3972 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003973 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003974 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975}
3976
3977static void cfq_slab_kill(void)
3978{
Jens Axboed6de8be2008-05-28 14:46:59 +02003979 /*
3980 * Caller already ensured that pending RCU callbacks are completed,
3981 * so we should have no busy allocations at this point.
3982 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 if (cfq_pool)
3984 kmem_cache_destroy(cfq_pool);
3985 if (cfq_ioc_pool)
3986 kmem_cache_destroy(cfq_ioc_pool);
3987}
3988
3989static int __init cfq_slab_setup(void)
3990{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003991 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003992 if (!cfq_pool)
3993 goto fail;
3994
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003995 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996 if (!cfq_ioc_pool)
3997 goto fail;
3998
3999 return 0;
4000fail:
4001 cfq_slab_kill();
4002 return -ENOMEM;
4003}
4004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005/*
4006 * sysfs parts below -->
4007 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008static ssize_t
4009cfq_var_show(unsigned int var, char *page)
4010{
4011 return sprintf(page, "%d\n", var);
4012}
4013
4014static ssize_t
4015cfq_var_store(unsigned int *var, const char *page, size_t count)
4016{
4017 char *p = (char *) page;
4018
4019 *var = simple_strtoul(p, &p, 10);
4020 return count;
4021}
4022
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004024static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004026 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027 unsigned int __data = __VAR; \
4028 if (__CONV) \
4029 __data = jiffies_to_msecs(__data); \
4030 return cfq_var_show(__data, (page)); \
4031}
4032SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004033SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4034SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05004035SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4036SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004037SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004038SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004039SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4040SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4041SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004042SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043#undef SHOW_FUNCTION
4044
4045#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004046static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004048 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049 unsigned int __data; \
4050 int ret = cfq_var_store(&__data, (page), count); \
4051 if (__data < (MIN)) \
4052 __data = (MIN); \
4053 else if (__data > (MAX)) \
4054 __data = (MAX); \
4055 if (__CONV) \
4056 *(__PTR) = msecs_to_jiffies(__data); \
4057 else \
4058 *(__PTR) = __data; \
4059 return ret; \
4060}
4061STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004062STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4063 UINT_MAX, 1);
4064STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4065 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004066STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004067STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4068 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004069STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004070STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004071STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4072STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004073STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4074 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004075STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076#undef STORE_FUNCTION
4077
Al Viroe572ec72006-03-18 22:27:18 -05004078#define CFQ_ATTR(name) \
4079 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004080
Al Viroe572ec72006-03-18 22:27:18 -05004081static struct elv_fs_entry cfq_attrs[] = {
4082 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004083 CFQ_ATTR(fifo_expire_sync),
4084 CFQ_ATTR(fifo_expire_async),
4085 CFQ_ATTR(back_seek_max),
4086 CFQ_ATTR(back_seek_penalty),
4087 CFQ_ATTR(slice_sync),
4088 CFQ_ATTR(slice_async),
4089 CFQ_ATTR(slice_async_rq),
4090 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004091 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004092 CFQ_ATTR(low_latency),
Al Viroe572ec72006-03-18 22:27:18 -05004093 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094};
4095
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096static struct elevator_type iosched_cfq = {
4097 .ops = {
4098 .elevator_merge_fn = cfq_merge,
4099 .elevator_merged_fn = cfq_merged_request,
4100 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004101 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004102 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004103 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004105 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106 .elevator_deactivate_req_fn = cfq_deactivate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004108 .elevator_former_req_fn = elv_rb_former_request,
4109 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110 .elevator_set_req_fn = cfq_set_request,
4111 .elevator_put_req_fn = cfq_put_request,
4112 .elevator_may_queue_fn = cfq_may_queue,
4113 .elevator_init_fn = cfq_init_queue,
4114 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02004115 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004116 },
Al Viro3d1ab402006-03-18 18:35:43 -05004117 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004118 .elevator_name = "cfq",
4119 .elevator_owner = THIS_MODULE,
4120};
4121
Vivek Goyal3e252062009-12-04 10:36:42 -05004122#ifdef CONFIG_CFQ_GROUP_IOSCHED
4123static struct blkio_policy_type blkio_policy_cfq = {
4124 .ops = {
4125 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4126 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4127 },
Vivek Goyal062a6442010-09-15 17:06:33 -04004128 .plid = BLKIO_POLICY_PROP,
Vivek Goyal3e252062009-12-04 10:36:42 -05004129};
4130#else
4131static struct blkio_policy_type blkio_policy_cfq;
4132#endif
4133
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134static int __init cfq_init(void)
4135{
Jens Axboe22e2c502005-06-27 10:55:12 +02004136 /*
4137 * could be 0 on HZ < 1000 setups
4138 */
4139 if (!cfq_slice_async)
4140 cfq_slice_async = 1;
4141 if (!cfq_slice_idle)
4142 cfq_slice_idle = 1;
4143
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004144#ifdef CONFIG_CFQ_GROUP_IOSCHED
4145 if (!cfq_group_idle)
4146 cfq_group_idle = 1;
4147#else
4148 cfq_group_idle = 0;
4149#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150 if (cfq_slab_setup())
4151 return -ENOMEM;
4152
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004153 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05004154 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004156 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004157}
4158
4159static void __exit cfq_exit(void)
4160{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07004161 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05004162 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05004164 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02004165 /* ioc_gone's update must be visible before reading ioc_count */
4166 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02004167
4168 /*
4169 * this also protects us from entering cfq_slab_kill() with
4170 * pending RCU callbacks
4171 */
Tejun Heo245b2e72009-06-24 15:13:48 +09004172 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02004173 wait_for_completion(&all_gone);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004174 ida_destroy(&cic_index_ida);
Christoph Hellwig83521d32005-10-30 15:01:39 -08004175 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004176}
4177
4178module_init(cfq_init);
4179module_exit(cfq_exit);
4180
4181MODULE_AUTHOR("Jens Axboe");
4182MODULE_LICENSE("GPL");
4183MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");