blob: ab7a9e6a9b1cb12c1952f5fd725ca2ea0dd20fd1 [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 Goyal70087dc2011-05-16 15:24:08 +02001017static struct cfq_group * cfq_find_alloc_cfqg(struct cfq_data *cfqd,
1018 struct blkio_cgroup *blkcg, int create)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001019{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001020 struct cfq_group *cfqg = NULL;
1021 void *key = cfqd;
1022 int i, j;
1023 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -05001024 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1025 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001026
Vivek Goyal25fb5162009-12-03 12:59:46 -05001027 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
Ricky Beniteza74b2ad2010-04-05 18:22:17 +02001028 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1029 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1030 cfqg->blkg.dev = MKDEV(major, minor);
1031 goto done;
1032 }
Vivek Goyal25fb5162009-12-03 12:59:46 -05001033 if (cfqg || !create)
1034 goto done;
1035
1036 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1037 if (!cfqg)
1038 goto done;
1039
Vivek Goyal25fb5162009-12-03 12:59:46 -05001040 for_each_cfqg_st(cfqg, i, j, st)
1041 *st = CFQ_RB_ROOT;
1042 RB_CLEAR_NODE(&cfqg->rb_node);
1043
Vivek Goyalb1c35762009-12-03 12:59:47 -05001044 /*
1045 * Take the initial reference that will be released on destroy
1046 * This can be thought of a joint reference by cgroup and
1047 * elevator which will be dropped by either elevator exit
1048 * or cgroup deletion path depending on who is exiting first.
1049 */
Shaohua Li329a6782011-01-07 08:48:28 +01001050 cfqg->ref = 1;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001051
Vivek Goyal180be2a2010-09-14 08:47:11 +02001052 /*
1053 * Add group onto cgroup list. It might happen that bdi->dev is
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001054 * not initialized yet. Initialize this new group without major
Vivek Goyal180be2a2010-09-14 08:47:11 +02001055 * and minor info and this info will be filled in once a new thread
1056 * comes for IO. See code above.
1057 */
1058 if (bdi->dev) {
1059 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1060 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
Vivek Goyal22084192009-12-03 12:59:49 -05001061 MKDEV(major, minor));
Vivek Goyal180be2a2010-09-14 08:47:11 +02001062 } else
1063 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1064 0);
1065
Gui Jianfeng34d0f172010-04-13 16:05:49 +08001066 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001067
1068 /* Add group on cfqd list */
1069 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1070
1071done:
Vivek Goyal25fb5162009-12-03 12:59:46 -05001072 return cfqg;
1073}
1074
1075/*
1076 * Search for the cfq group current task belongs to. If create = 1, then also
1077 * create the cfq group if it does not exist. request_queue lock must be held.
1078 */
1079static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1080{
Vivek Goyal70087dc2011-05-16 15:24:08 +02001081 struct blkio_cgroup *blkcg;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001082 struct cfq_group *cfqg = NULL;
1083
1084 rcu_read_lock();
Vivek Goyal70087dc2011-05-16 15:24:08 +02001085 blkcg = task_blkio_cgroup(current);
1086 cfqg = cfq_find_alloc_cfqg(cfqd, blkcg, create);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001087 if (!cfqg && create)
1088 cfqg = &cfqd->root_group;
1089 rcu_read_unlock();
1090 return cfqg;
1091}
1092
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001093static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1094{
Shaohua Li329a6782011-01-07 08:48:28 +01001095 cfqg->ref++;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001096 return cfqg;
1097}
1098
Vivek Goyal25fb5162009-12-03 12:59:46 -05001099static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1100{
1101 /* Currently, all async queues are mapped to root group */
1102 if (!cfq_cfqq_sync(cfqq))
1103 cfqg = &cfqq->cfqd->root_group;
1104
1105 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001106 /* cfqq reference on cfqg */
Shaohua Li329a6782011-01-07 08:48:28 +01001107 cfqq->cfqg->ref++;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001108}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001109
1110static void cfq_put_cfqg(struct cfq_group *cfqg)
1111{
1112 struct cfq_rb_root *st;
1113 int i, j;
1114
Shaohua Li329a6782011-01-07 08:48:28 +01001115 BUG_ON(cfqg->ref <= 0);
1116 cfqg->ref--;
1117 if (cfqg->ref)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001118 return;
1119 for_each_cfqg_st(cfqg, i, j, st)
Gui Jianfengb54ce602010-11-30 20:52:46 +01001120 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
Vivek Goyalb1c35762009-12-03 12:59:47 -05001121 kfree(cfqg);
1122}
1123
1124static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1125{
1126 /* Something wrong if we are trying to remove same group twice */
1127 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1128
1129 hlist_del_init(&cfqg->cfqd_node);
1130
1131 /*
1132 * Put the reference taken at the time of creation so that when all
1133 * queues are gone, group can be destroyed.
1134 */
1135 cfq_put_cfqg(cfqg);
1136}
1137
1138static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1139{
1140 struct hlist_node *pos, *n;
1141 struct cfq_group *cfqg;
1142
1143 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1144 /*
1145 * If cgroup removal path got to blk_group first and removed
1146 * it from cgroup list, then it will take care of destroying
1147 * cfqg also.
1148 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001149 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001150 cfq_destroy_cfqg(cfqd, cfqg);
1151 }
1152}
1153
1154/*
1155 * Blk cgroup controller notification saying that blkio_group object is being
1156 * delinked as associated cgroup object is going away. That also means that
1157 * no new IO will come in this group. So get rid of this group as soon as
1158 * any pending IO in the group is finished.
1159 *
1160 * This function is called under rcu_read_lock(). key is the rcu protected
1161 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1162 * read lock.
1163 *
1164 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1165 * it should not be NULL as even if elevator was exiting, cgroup deltion
1166 * path got to it first.
1167 */
1168void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1169{
1170 unsigned long flags;
1171 struct cfq_data *cfqd = key;
1172
1173 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1174 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1175 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1176}
1177
Vivek Goyal25fb5162009-12-03 12:59:46 -05001178#else /* GROUP_IOSCHED */
1179static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1180{
1181 return &cfqd->root_group;
1182}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001183
1184static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1185{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001186 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001187}
1188
Vivek Goyal25fb5162009-12-03 12:59:46 -05001189static inline void
1190cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1191 cfqq->cfqg = cfqg;
1192}
1193
Vivek Goyalb1c35762009-12-03 12:59:47 -05001194static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1195static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1196
Vivek Goyal25fb5162009-12-03 12:59:46 -05001197#endif /* GROUP_IOSCHED */
1198
Jens Axboe498d3aa22007-04-26 12:54:48 +02001199/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001200 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001201 * requests waiting to be processed. It is sorted in the order that
1202 * we will service the queues.
1203 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001204static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001205 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001206{
Jens Axboe08717142008-01-28 11:38:15 +01001207 struct rb_node **p, *parent;
1208 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001209 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001210 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001211 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001212 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001213 int group_changed = 0;
1214
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001215 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001216 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001217 if (cfq_class_idle(cfqq)) {
1218 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001219 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001220 if (parent && parent != &cfqq->rb_node) {
1221 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1222 rb_key += __cfqq->rb_key;
1223 } else
1224 rb_key += jiffies;
1225 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001226 /*
1227 * Get our rb key offset. Subtract any residual slice
1228 * value carried from last service. A negative resid
1229 * count indicates slice overrun, and this should position
1230 * the next service time further away in the tree.
1231 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001232 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001233 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001234 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001235 } else {
1236 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001237 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001238 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1239 }
Jens Axboed9e76202007-04-20 14:27:50 +02001240
1241 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001242 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001243 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001244 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001245 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001246 if (rb_key == cfqq->rb_key &&
1247 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001248 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001249
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001250 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1251 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001252 }
Jens Axboed9e76202007-04-20 14:27:50 +02001253
Jens Axboe498d3aa22007-04-26 12:54:48 +02001254 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001255 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001256 cfqq->service_tree = service_tree;
1257 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001258 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001259 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001260
Jens Axboed9e76202007-04-20 14:27:50 +02001261 parent = *p;
1262 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1263
Jens Axboe0c534e02007-04-18 20:01:57 +02001264 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001265 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001266 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001267 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001268 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001269 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001270 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001271 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001272 }
Jens Axboe67060e32007-04-18 20:13:32 +02001273
1274 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001275 }
1276
Jens Axboecc09e292007-04-26 12:53:50 +02001277 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001278 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001279
Jens Axboed9e76202007-04-20 14:27:50 +02001280 cfqq->rb_key = rb_key;
1281 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001282 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1283 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001284 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001285 return;
Justin TerAvest8184f932011-03-17 16:12:36 +01001286 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287}
1288
Jens Axboea36e71f2009-04-15 12:15:11 +02001289static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001290cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1291 sector_t sector, struct rb_node **ret_parent,
1292 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001293{
Jens Axboea36e71f2009-04-15 12:15:11 +02001294 struct rb_node **p, *parent;
1295 struct cfq_queue *cfqq = NULL;
1296
1297 parent = NULL;
1298 p = &root->rb_node;
1299 while (*p) {
1300 struct rb_node **n;
1301
1302 parent = *p;
1303 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1304
1305 /*
1306 * Sort strictly based on sector. Smallest to the left,
1307 * largest to the right.
1308 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001309 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001310 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001311 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001312 n = &(*p)->rb_left;
1313 else
1314 break;
1315 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001316 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001317 }
1318
1319 *ret_parent = parent;
1320 if (rb_link)
1321 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001322 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001323}
1324
1325static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1326{
Jens Axboea36e71f2009-04-15 12:15:11 +02001327 struct rb_node **p, *parent;
1328 struct cfq_queue *__cfqq;
1329
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001330 if (cfqq->p_root) {
1331 rb_erase(&cfqq->p_node, cfqq->p_root);
1332 cfqq->p_root = NULL;
1333 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001334
1335 if (cfq_class_idle(cfqq))
1336 return;
1337 if (!cfqq->next_rq)
1338 return;
1339
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001340 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001341 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1342 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001343 if (!__cfqq) {
1344 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001345 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1346 } else
1347 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001348}
1349
Jens Axboe498d3aa22007-04-26 12:54:48 +02001350/*
1351 * Update cfqq's position in the service tree.
1352 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001353static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001354{
Jens Axboe6d048f52007-04-25 12:44:27 +02001355 /*
1356 * Resorting requires the cfqq to be on the RR list already.
1357 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001358 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001359 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001360 cfq_prio_tree_add(cfqd, cfqq);
1361 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001362}
1363
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364/*
1365 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001366 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 */
Jens Axboefebffd62008-01-28 13:19:43 +01001368static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369{
Jens Axboe7b679132008-05-30 12:23:07 +02001370 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001371 BUG_ON(cfq_cfqq_on_rr(cfqq));
1372 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 cfqd->busy_queues++;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001374 if (cfq_cfqq_sync(cfqq))
1375 cfqd->busy_sync_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376
Jens Axboeedd75ff2007-04-19 12:03:34 +02001377 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378}
1379
Jens Axboe498d3aa22007-04-26 12:54:48 +02001380/*
1381 * Called when the cfqq no longer has requests pending, remove it from
1382 * the service tree.
1383 */
Jens Axboefebffd62008-01-28 13:19:43 +01001384static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385{
Jens Axboe7b679132008-05-30 12:23:07 +02001386 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001387 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1388 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001390 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1391 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1392 cfqq->service_tree = NULL;
1393 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001394 if (cfqq->p_root) {
1395 rb_erase(&cfqq->p_node, cfqq->p_root);
1396 cfqq->p_root = NULL;
1397 }
Jens Axboed9e76202007-04-20 14:27:50 +02001398
Justin TerAvest8184f932011-03-17 16:12:36 +01001399 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 BUG_ON(!cfqd->busy_queues);
1401 cfqd->busy_queues--;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001402 if (cfq_cfqq_sync(cfqq))
1403 cfqd->busy_sync_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404}
1405
1406/*
1407 * rb tree support functions
1408 */
Jens Axboefebffd62008-01-28 13:19:43 +01001409static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410{
Jens Axboe5e705372006-07-13 12:39:25 +02001411 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001412 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413
Jens Axboeb4878f22005-10-20 16:42:29 +02001414 BUG_ON(!cfqq->queued[sync]);
1415 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416
Jens Axboe5e705372006-07-13 12:39:25 +02001417 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
Vivek Goyalf04a6422009-12-03 12:59:40 -05001419 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1420 /*
1421 * Queue will be deleted from service tree when we actually
1422 * expire it later. Right now just remove it from prio tree
1423 * as it is empty.
1424 */
1425 if (cfqq->p_root) {
1426 rb_erase(&cfqq->p_node, cfqq->p_root);
1427 cfqq->p_root = NULL;
1428 }
1429 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430}
1431
Jens Axboe5e705372006-07-13 12:39:25 +02001432static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433{
Jens Axboe5e705372006-07-13 12:39:25 +02001434 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001436 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437
Jens Axboe5380a102006-07-13 12:37:56 +02001438 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439
1440 /*
1441 * looks a little odd, but the first insert might return an alias.
1442 * if that happens, put the alias on the dispatch list
1443 */
Jens Axboe21183b02006-07-13 12:33:14 +02001444 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001445 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001446
1447 if (!cfq_cfqq_on_rr(cfqq))
1448 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001449
1450 /*
1451 * check if this request is a better next-serve candidate
1452 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001453 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001454 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001455
1456 /*
1457 * adjust priority tree position, if ->next_rq changes
1458 */
1459 if (prev != cfqq->next_rq)
1460 cfq_prio_tree_add(cfqd, cfqq);
1461
Jens Axboe5044eed2007-04-25 11:53:48 +02001462 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463}
1464
Jens Axboefebffd62008-01-28 13:19:43 +01001465static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466{
Jens Axboe5380a102006-07-13 12:37:56 +02001467 elv_rb_del(&cfqq->sort_list, rq);
1468 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001469 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1470 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001471 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001472 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001473 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1474 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475}
1476
Jens Axboe206dc692006-03-28 13:03:44 +02001477static struct request *
1478cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479{
Jens Axboe206dc692006-03-28 13:03:44 +02001480 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001481 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001482 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483
Jens Axboe4ac845a2008-01-24 08:44:49 +01001484 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001485 if (!cic)
1486 return NULL;
1487
1488 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001489 if (cfqq) {
1490 sector_t sector = bio->bi_sector + bio_sectors(bio);
1491
Jens Axboe21183b02006-07-13 12:33:14 +02001492 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001493 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 return NULL;
1496}
1497
Jens Axboe165125e2007-07-24 09:28:11 +02001498static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001499{
1500 struct cfq_data *cfqd = q->elevator->elevator_data;
1501
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001502 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001503 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001504 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001505
Tejun Heo5b936292009-05-07 22:24:38 +09001506 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001507}
1508
Jens Axboe165125e2007-07-24 09:28:11 +02001509static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510{
Jens Axboe22e2c502005-06-27 10:55:12 +02001511 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001513 WARN_ON(!cfqd->rq_in_driver);
1514 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001515 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001516 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517}
1518
Jens Axboeb4878f22005-10-20 16:42:29 +02001519static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520{
Jens Axboe5e705372006-07-13 12:39:25 +02001521 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001522
Jens Axboe5e705372006-07-13 12:39:25 +02001523 if (cfqq->next_rq == rq)
1524 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525
Jens Axboeb4878f22005-10-20 16:42:29 +02001526 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001527 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001528
Aaron Carroll45333d52008-08-26 15:52:36 +02001529 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001530 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1531 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02001532 if (rq->cmd_flags & REQ_META) {
Jens Axboe374f84a2006-07-23 01:42:19 +02001533 WARN_ON(!cfqq->meta_pending);
1534 cfqq->meta_pending--;
1535 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536}
1537
Jens Axboe165125e2007-07-24 09:28:11 +02001538static int cfq_merge(struct request_queue *q, struct request **req,
1539 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540{
1541 struct cfq_data *cfqd = q->elevator->elevator_data;
1542 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543
Jens Axboe206dc692006-03-28 13:03:44 +02001544 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001545 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001546 *req = __rq;
1547 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 }
1549
1550 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551}
1552
Jens Axboe165125e2007-07-24 09:28:11 +02001553static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001554 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555{
Jens Axboe21183b02006-07-13 12:33:14 +02001556 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001557 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558
Jens Axboe5e705372006-07-13 12:39:25 +02001559 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561}
1562
Divyesh Shah812d4022010-04-08 21:14:23 -07001563static void cfq_bio_merged(struct request_queue *q, struct request *req,
1564 struct bio *bio)
1565{
Vivek Goyale98ef892010-06-18 10:39:47 -04001566 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1567 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001568}
1569
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570static void
Jens Axboe165125e2007-07-24 09:28:11 +02001571cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 struct request *next)
1573{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001574 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001575 /*
1576 * reposition in fifo if next is older than rq
1577 */
1578 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001579 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001580 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001581 rq_set_fifo_time(rq, rq_fifo_time(next));
1582 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001583
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001584 if (cfqq->next_rq == next)
1585 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001586 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001587 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1588 rq_data_dir(next), rq_is_sync(next));
Jens Axboe22e2c502005-06-27 10:55:12 +02001589}
1590
Jens Axboe165125e2007-07-24 09:28:11 +02001591static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001592 struct bio *bio)
1593{
1594 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001595 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001596 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001597
1598 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001599 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001600 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001601 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001602 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001603
1604 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001605 * Lookup the cfqq that this bio will be queued with. Allow
1606 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001607 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001608 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001609 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001610 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001611
Vasily Tarasov91fac312007-04-25 12:29:51 +02001612 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001613 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001614}
1615
Divyesh Shah812df482010-04-08 21:15:35 -07001616static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1617{
1618 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001619 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001620}
1621
Jens Axboefebffd62008-01-28 13:19:43 +01001622static void __cfq_set_active_queue(struct cfq_data *cfqd,
1623 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001624{
1625 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001626 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1627 cfqd->serving_prio, cfqd->serving_type);
Justin TerAvest62a37f62011-03-23 08:25:44 +01001628 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1629 cfqq->slice_start = 0;
1630 cfqq->dispatch_start = jiffies;
1631 cfqq->allocated_slice = 0;
1632 cfqq->slice_end = 0;
1633 cfqq->slice_dispatch = 0;
1634 cfqq->nr_sectors = 0;
1635
1636 cfq_clear_cfqq_wait_request(cfqq);
1637 cfq_clear_cfqq_must_dispatch(cfqq);
1638 cfq_clear_cfqq_must_alloc_slice(cfqq);
1639 cfq_clear_cfqq_fifo_expire(cfqq);
1640 cfq_mark_cfqq_slice_new(cfqq);
1641
1642 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001643 }
1644
1645 cfqd->active_queue = cfqq;
1646}
1647
1648/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001649 * current cfqq expired its slice (or was too idle), select new one
1650 */
1651static void
1652__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001653 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001654{
Jens Axboe7b679132008-05-30 12:23:07 +02001655 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1656
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001657 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001658 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001659
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001660 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001661 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001662
1663 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001664 * If this cfqq is shared between multiple processes, check to
1665 * make sure that those processes are still issuing I/Os within
1666 * the mean seek distance. If not, it may be time to break the
1667 * queues apart again.
1668 */
1669 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1670 cfq_mark_cfqq_split_coop(cfqq);
1671
1672 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001673 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001674 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01001675 if (timed_out) {
1676 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07001677 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001678 else
1679 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001680 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1681 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001682
Vivek Goyale5ff0822010-04-26 19:25:11 +02001683 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001684
Vivek Goyalf04a6422009-12-03 12:59:40 -05001685 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1686 cfq_del_cfqq_rr(cfqd, cfqq);
1687
Jens Axboeedd75ff2007-04-19 12:03:34 +02001688 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001689
1690 if (cfqq == cfqd->active_queue)
1691 cfqd->active_queue = NULL;
1692
1693 if (cfqd->active_cic) {
1694 put_io_context(cfqd->active_cic->ioc);
1695 cfqd->active_cic = NULL;
1696 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001697}
1698
Vivek Goyale5ff0822010-04-26 19:25:11 +02001699static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001700{
1701 struct cfq_queue *cfqq = cfqd->active_queue;
1702
1703 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001704 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001705}
1706
Jens Axboe498d3aa22007-04-26 12:54:48 +02001707/*
1708 * Get next queue for service. Unless we have a queue preemption,
1709 * we'll simply select the first cfqq in the service tree.
1710 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001711static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001712{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001713 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001714 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001715 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001716
Vivek Goyalf04a6422009-12-03 12:59:40 -05001717 if (!cfqd->rq_queued)
1718 return NULL;
1719
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001720 /* There is nothing to dispatch */
1721 if (!service_tree)
1722 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001723 if (RB_EMPTY_ROOT(&service_tree->rb))
1724 return NULL;
1725 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001726}
1727
Vivek Goyalf04a6422009-12-03 12:59:40 -05001728static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1729{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001730 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001731 struct cfq_queue *cfqq;
1732 int i, j;
1733 struct cfq_rb_root *st;
1734
1735 if (!cfqd->rq_queued)
1736 return NULL;
1737
Vivek Goyal25fb5162009-12-03 12:59:46 -05001738 cfqg = cfq_get_next_cfqg(cfqd);
1739 if (!cfqg)
1740 return NULL;
1741
Vivek Goyalf04a6422009-12-03 12:59:40 -05001742 for_each_cfqg_st(cfqg, i, j, st)
1743 if ((cfqq = cfq_rb_first(st)) != NULL)
1744 return cfqq;
1745 return NULL;
1746}
1747
Jens Axboe498d3aa22007-04-26 12:54:48 +02001748/*
1749 * Get and set a new active queue for service.
1750 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001751static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1752 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001753{
Jens Axboee00ef792009-11-04 08:54:55 +01001754 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001755 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001756
Jens Axboe22e2c502005-06-27 10:55:12 +02001757 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001758 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001759}
1760
Jens Axboed9e76202007-04-20 14:27:50 +02001761static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1762 struct request *rq)
1763{
Tejun Heo83096eb2009-05-07 22:24:39 +09001764 if (blk_rq_pos(rq) >= cfqd->last_position)
1765 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001766 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001767 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001768}
1769
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001770static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001771 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001772{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001773 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001774}
1775
Jens Axboea36e71f2009-04-15 12:15:11 +02001776static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1777 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001778{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001779 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001780 struct rb_node *parent, *node;
1781 struct cfq_queue *__cfqq;
1782 sector_t sector = cfqd->last_position;
1783
1784 if (RB_EMPTY_ROOT(root))
1785 return NULL;
1786
1787 /*
1788 * First, if we find a request starting at the end of the last
1789 * request, choose it.
1790 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001791 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001792 if (__cfqq)
1793 return __cfqq;
1794
1795 /*
1796 * If the exact sector wasn't found, the parent of the NULL leaf
1797 * will contain the closest sector.
1798 */
1799 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001800 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001801 return __cfqq;
1802
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001803 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001804 node = rb_next(&__cfqq->p_node);
1805 else
1806 node = rb_prev(&__cfqq->p_node);
1807 if (!node)
1808 return NULL;
1809
1810 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001811 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001812 return __cfqq;
1813
1814 return NULL;
1815}
1816
1817/*
1818 * cfqd - obvious
1819 * cur_cfqq - passed in so that we don't decide that the current queue is
1820 * closely cooperating with itself.
1821 *
1822 * So, basically we're assuming that that cur_cfqq has dispatched at least
1823 * one request, and that cfqd->last_position reflects a position on the disk
1824 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1825 * assumption.
1826 */
1827static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001828 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001829{
1830 struct cfq_queue *cfqq;
1831
Divyesh Shah39c01b22010-03-25 15:45:57 +01001832 if (cfq_class_idle(cur_cfqq))
1833 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001834 if (!cfq_cfqq_sync(cur_cfqq))
1835 return NULL;
1836 if (CFQQ_SEEKY(cur_cfqq))
1837 return NULL;
1838
Jens Axboea36e71f2009-04-15 12:15:11 +02001839 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001840 * Don't search priority tree if it's the only queue in the group.
1841 */
1842 if (cur_cfqq->cfqg->nr_cfqq == 1)
1843 return NULL;
1844
1845 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001846 * We should notice if some of the queues are cooperating, eg
1847 * working closely on the same area of the disk. In that case,
1848 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001849 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001850 cfqq = cfqq_close(cfqd, cur_cfqq);
1851 if (!cfqq)
1852 return NULL;
1853
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001854 /* If new queue belongs to different cfq_group, don't choose it */
1855 if (cur_cfqq->cfqg != cfqq->cfqg)
1856 return NULL;
1857
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001858 /*
1859 * It only makes sense to merge sync queues.
1860 */
1861 if (!cfq_cfqq_sync(cfqq))
1862 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001863 if (CFQQ_SEEKY(cfqq))
1864 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001865
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001866 /*
1867 * Do not merge queues of different priority classes
1868 */
1869 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1870 return NULL;
1871
Jens Axboea36e71f2009-04-15 12:15:11 +02001872 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001873}
1874
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001875/*
1876 * Determine whether we should enforce idle window for this queue.
1877 */
1878
1879static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1880{
1881 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001882 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001883
Vivek Goyalf04a6422009-12-03 12:59:40 -05001884 BUG_ON(!service_tree);
1885 BUG_ON(!service_tree->count);
1886
Vivek Goyalb6508c12010-08-23 12:23:33 +02001887 if (!cfqd->cfq_slice_idle)
1888 return false;
1889
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001890 /* We never do for idle class queues. */
1891 if (prio == IDLE_WORKLOAD)
1892 return false;
1893
1894 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001895 if (cfq_cfqq_idle_window(cfqq) &&
1896 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001897 return true;
1898
1899 /*
1900 * Otherwise, we do only if they are the last ones
1901 * in their service tree.
1902 */
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001903 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01001904 return true;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001905 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1906 service_tree->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01001907 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001908}
1909
Jens Axboe6d048f52007-04-25 12:44:27 +02001910static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001911{
Jens Axboe17926692007-01-19 11:59:30 +11001912 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001913 struct cfq_io_context *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001914 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001915
Jens Axboea68bbdd2008-09-24 13:03:33 +02001916 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001917 * SSD device without seek penalty, disable idling. But only do so
1918 * for devices that support queuing, otherwise we still have a problem
1919 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001920 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001921 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001922 return;
1923
Jens Axboedd67d052006-06-21 09:36:18 +02001924 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001925 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001926
1927 /*
1928 * idle is disabled, either manually or by past process history
1929 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001930 if (!cfq_should_idle(cfqd, cfqq)) {
1931 /* no queue idling. Check for group idling */
1932 if (cfqd->cfq_group_idle)
1933 group_idle = cfqd->cfq_group_idle;
1934 else
1935 return;
1936 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001937
Jens Axboe22e2c502005-06-27 10:55:12 +02001938 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001939 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001940 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001941 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001942 return;
1943
1944 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001945 * task has exited, don't wait
1946 */
Jens Axboe206dc692006-03-28 13:03:44 +02001947 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001948 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001949 return;
1950
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001951 /*
1952 * If our average think time is larger than the remaining time
1953 * slice, then don't idle. This avoids overrunning the allotted
1954 * time slice.
1955 */
1956 if (sample_valid(cic->ttime_samples) &&
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001957 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1958 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1959 cic->ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001960 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001961 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001962
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001963 /* There are other queues in the group, don't do group idle */
1964 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1965 return;
1966
Jens Axboe3b181522005-06-27 10:56:24 +02001967 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001968
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001969 if (group_idle)
1970 sl = cfqd->cfq_group_idle;
1971 else
1972 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001973
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001974 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001975 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001976 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1977 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978}
1979
Jens Axboe498d3aa22007-04-26 12:54:48 +02001980/*
1981 * Move request from internal lists to the request queue dispatch list.
1982 */
Jens Axboe165125e2007-07-24 09:28:11 +02001983static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001985 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001986 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001987
Jens Axboe7b679132008-05-30 12:23:07 +02001988 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1989
Jeff Moyer06d21882009-09-11 17:08:59 +02001990 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001991 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001992 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001993 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001994 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001995
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001996 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001997 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001998 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02001999 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000}
2001
2002/*
2003 * return expired entry, or NULL to just start from scratch in rbtree
2004 */
Jens Axboefebffd62008-01-28 13:19:43 +01002005static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006{
Jens Axboe30996f42009-10-05 11:03:39 +02002007 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008
Jens Axboe3b181522005-06-27 10:56:24 +02002009 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002011
2012 cfq_mark_cfqq_fifo_expire(cfqq);
2013
Jens Axboe89850f72006-07-22 16:48:31 +02002014 if (list_empty(&cfqq->fifo))
2015 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016
Jens Axboe89850f72006-07-22 16:48:31 +02002017 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002018 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002019 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020
Jens Axboe30996f42009-10-05 11:03:39 +02002021 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002022 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023}
2024
Jens Axboe22e2c502005-06-27 10:55:12 +02002025static inline int
2026cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2027{
2028 const int base_rq = cfqd->cfq_slice_async_rq;
2029
2030 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2031
2032 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2033}
2034
2035/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002036 * Must be called with the queue_lock held.
2037 */
2038static int cfqq_process_refs(struct cfq_queue *cfqq)
2039{
2040 int process_refs, io_refs;
2041
2042 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002043 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002044 BUG_ON(process_refs < 0);
2045 return process_refs;
2046}
2047
2048static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2049{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002050 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002051 struct cfq_queue *__cfqq;
2052
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002053 /*
2054 * If there are no process references on the new_cfqq, then it is
2055 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2056 * chain may have dropped their last reference (not just their
2057 * last process reference).
2058 */
2059 if (!cfqq_process_refs(new_cfqq))
2060 return;
2061
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002062 /* Avoid a circular list and skip interim queue merges */
2063 while ((__cfqq = new_cfqq->new_cfqq)) {
2064 if (__cfqq == cfqq)
2065 return;
2066 new_cfqq = __cfqq;
2067 }
2068
2069 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002070 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002071 /*
2072 * If the process for the cfqq has gone away, there is no
2073 * sense in merging the queues.
2074 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002075 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002076 return;
2077
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002078 /*
2079 * Merge in the direction of the lesser amount of work.
2080 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002081 if (new_process_refs >= process_refs) {
2082 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002083 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002084 } else {
2085 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002086 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002087 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002088}
2089
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002090static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002091 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002092{
2093 struct cfq_queue *queue;
2094 int i;
2095 bool key_valid = false;
2096 unsigned long lowest_key = 0;
2097 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2098
Vivek Goyal65b32a52009-12-16 17:52:59 -05002099 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2100 /* select the one with lowest rb_key */
2101 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002102 if (queue &&
2103 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2104 lowest_key = queue->rb_key;
2105 cur_best = i;
2106 key_valid = true;
2107 }
2108 }
2109
2110 return cur_best;
2111}
2112
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002113static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002114{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002115 unsigned slice;
2116 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002117 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002118 unsigned group_slice;
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002119 enum wl_prio_t original_prio = cfqd->serving_prio;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002120
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002121 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002122 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002123 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002124 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002125 cfqd->serving_prio = BE_WORKLOAD;
2126 else {
2127 cfqd->serving_prio = IDLE_WORKLOAD;
2128 cfqd->workload_expires = jiffies + 1;
2129 return;
2130 }
2131
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002132 if (original_prio != cfqd->serving_prio)
2133 goto new_workload;
2134
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002135 /*
2136 * For RT and BE, we have to choose also the type
2137 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2138 * expiration time
2139 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002140 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002141 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002142
2143 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002144 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002145 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002146 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002147 return;
2148
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002149new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002150 /* otherwise select new workload type */
2151 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002152 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2153 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002154 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002155
2156 /*
2157 * the workload slice is computed as a fraction of target latency
2158 * proportional to the number of queues in that workload, over
2159 * all the queues in the same priority class
2160 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002161 group_slice = cfq_group_slice(cfqd, cfqg);
2162
2163 slice = group_slice * count /
2164 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2165 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002166
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002167 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2168 unsigned int tmp;
2169
2170 /*
2171 * Async queues are currently system wide. Just taking
2172 * proportion of queues with-in same group will lead to higher
2173 * async ratio system wide as generally root group is going
2174 * to have higher weight. A more accurate thing would be to
2175 * calculate system wide asnc/sync ratio.
2176 */
2177 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2178 tmp = tmp/cfqd->busy_queues;
2179 slice = min_t(unsigned, slice, tmp);
2180
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002181 /* async workload slice is scaled down according to
2182 * the sync/async slice ratio. */
2183 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002184 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002185 /* sync workload slice is at least 2 * cfq_slice_idle */
2186 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2187
2188 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002189 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002190 cfqd->workload_expires = jiffies + slice;
2191}
2192
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002193static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2194{
2195 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002196 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002197
2198 if (RB_EMPTY_ROOT(&st->rb))
2199 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002200 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002201 update_min_vdisktime(st);
2202 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002203}
2204
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002205static void cfq_choose_cfqg(struct cfq_data *cfqd)
2206{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002207 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2208
2209 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002210
2211 /* Restore the workload type data */
2212 if (cfqg->saved_workload_slice) {
2213 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2214 cfqd->serving_type = cfqg->saved_workload;
2215 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002216 } else
2217 cfqd->workload_expires = jiffies - 1;
2218
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002219 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002220}
2221
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002222/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002223 * Select a queue for service. If we have a current active queue,
2224 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002225 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002226static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002227{
Jens Axboea36e71f2009-04-15 12:15:11 +02002228 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002229
2230 cfqq = cfqd->active_queue;
2231 if (!cfqq)
2232 goto new_queue;
2233
Vivek Goyalf04a6422009-12-03 12:59:40 -05002234 if (!cfqd->rq_queued)
2235 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002236
2237 /*
2238 * We were waiting for group to get backlogged. Expire the queue
2239 */
2240 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2241 goto expire;
2242
Jens Axboe22e2c502005-06-27 10:55:12 +02002243 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002244 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002245 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002246 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2247 /*
2248 * If slice had not expired at the completion of last request
2249 * we might not have turned on wait_busy flag. Don't expire
2250 * the queue yet. Allow the group to get backlogged.
2251 *
2252 * The very fact that we have used the slice, that means we
2253 * have been idling all along on this queue and it should be
2254 * ok to wait for this request to complete.
2255 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002256 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2257 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2258 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002259 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002260 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002261 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002262 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002263
2264 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002265 * The active queue has requests and isn't expired, allow it to
2266 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002267 */
Jens Axboedd67d052006-06-21 09:36:18 +02002268 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002269 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002270
2271 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002272 * If another queue has a request waiting within our mean seek
2273 * distance, let it run. The expire code will check for close
2274 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002275 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002276 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002277 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002278 if (new_cfqq) {
2279 if (!cfqq->new_cfqq)
2280 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002281 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002282 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002283
2284 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002285 * No requests pending. If the active queue still has requests in
2286 * flight or is idling for a new request, allow either of these
2287 * conditions to happen (or time out) before selecting a new queue.
2288 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002289 if (timer_pending(&cfqd->idle_slice_timer)) {
2290 cfqq = NULL;
2291 goto keep_queue;
2292 }
2293
Shaohua Li8e1ac662010-11-08 15:01:04 +01002294 /*
2295 * This is a deep seek queue, but the device is much faster than
2296 * the queue can deliver, don't idle
2297 **/
2298 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2299 (cfq_cfqq_slice_new(cfqq) ||
2300 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2301 cfq_clear_cfqq_deep(cfqq);
2302 cfq_clear_cfqq_idle_window(cfqq);
2303 }
2304
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002305 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2306 cfqq = NULL;
2307 goto keep_queue;
2308 }
2309
2310 /*
2311 * If group idle is enabled and there are requests dispatched from
2312 * this group, wait for requests to complete.
2313 */
2314check_group_idle:
2315 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2316 && cfqq->cfqg->dispatched) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002317 cfqq = NULL;
2318 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002319 }
2320
Jens Axboe3b181522005-06-27 10:56:24 +02002321expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002322 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002323new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002324 /*
2325 * Current queue expired. Check if we have to switch to a new
2326 * service tree
2327 */
2328 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002329 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002330
Jens Axboea36e71f2009-04-15 12:15:11 +02002331 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002332keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002333 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002334}
2335
Jens Axboefebffd62008-01-28 13:19:43 +01002336static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002337{
2338 int dispatched = 0;
2339
2340 while (cfqq->next_rq) {
2341 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2342 dispatched++;
2343 }
2344
2345 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002346
2347 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002348 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002349 return dispatched;
2350}
2351
Jens Axboe498d3aa22007-04-26 12:54:48 +02002352/*
2353 * Drain our current requests. Used for barriers and when switching
2354 * io schedulers on-the-fly.
2355 */
Jens Axboed9e76202007-04-20 14:27:50 +02002356static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002357{
Jens Axboe08717142008-01-28 11:38:15 +01002358 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002359 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002360
Divyesh Shah3440c492010-04-09 09:29:57 +02002361 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002362 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002363 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2364 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002365 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002366 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002367
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002368 BUG_ON(cfqd->busy_queues);
2369
Jeff Moyer69237152009-06-12 15:29:30 +02002370 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002371 return dispatched;
2372}
2373
Shaohua Liabc3c742010-03-01 09:20:54 +01002374static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2375 struct cfq_queue *cfqq)
2376{
2377 /* the queue hasn't finished any request, can't estimate */
2378 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01002379 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002380 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2381 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01002382 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002383
Shaohua Lic1e44752010-11-08 15:01:02 +01002384 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01002385}
2386
Jens Axboe0b182d62009-10-06 20:49:37 +02002387static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002388{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002389 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390
Jens Axboe2f5cb732009-04-07 08:51:19 +02002391 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002392 * Drain async requests before we start sync IO
2393 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002394 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002395 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002396
2397 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002398 * If this is an async queue and we have sync IO in flight, let it wait
2399 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002400 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002401 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002402
Shaohua Liabc3c742010-03-01 09:20:54 +01002403 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002404 if (cfq_class_idle(cfqq))
2405 max_dispatch = 1;
2406
2407 /*
2408 * Does this cfqq already have too much IO in flight?
2409 */
2410 if (cfqq->dispatched >= max_dispatch) {
Shaohua Lief8a41d2011-03-07 09:26:29 +01002411 bool promote_sync = false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002412 /*
2413 * idle queue must always only have a single IO in flight
2414 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002415 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002416 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002417
Jens Axboe2f5cb732009-04-07 08:51:19 +02002418 /*
Li, Shaohuac4ade942011-03-23 08:30:34 +01002419 * If there is only one sync queue
2420 * we can ignore async queue here and give the sync
Shaohua Lief8a41d2011-03-07 09:26:29 +01002421 * queue no dispatch limit. The reason is a sync queue can
2422 * preempt async queue, limiting the sync queue doesn't make
2423 * sense. This is useful for aiostress test.
2424 */
Li, Shaohuac4ade942011-03-23 08:30:34 +01002425 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2426 promote_sync = true;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002427
2428 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002429 * We have other queues, don't allow more IO from this one
2430 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002431 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2432 !promote_sync)
Jens Axboe0b182d62009-10-06 20:49:37 +02002433 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002434
Jens Axboe2f5cb732009-04-07 08:51:19 +02002435 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002436 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002437 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002438 if (cfqd->busy_queues == 1 || promote_sync)
Shaohua Liabc3c742010-03-01 09:20:54 +01002439 max_dispatch = -1;
2440 else
2441 /*
2442 * Normally we start throttling cfqq when cfq_quantum/2
2443 * requests have been dispatched. But we can drive
2444 * deeper queue depths at the beginning of slice
2445 * subjected to upper limit of cfq_quantum.
2446 * */
2447 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002448 }
2449
2450 /*
2451 * Async queues must wait a bit before being allowed dispatch.
2452 * We also ramp up the dispatch depth gradually for async IO,
2453 * based on the last sync IO we serviced
2454 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002455 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002456 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002457 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002458
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002459 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002460 if (!depth && !cfqq->dispatched)
2461 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002462 if (depth < max_dispatch)
2463 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 }
2465
Jens Axboe0b182d62009-10-06 20:49:37 +02002466 /*
2467 * If we're below the current max, allow a dispatch
2468 */
2469 return cfqq->dispatched < max_dispatch;
2470}
2471
2472/*
2473 * Dispatch a request from cfqq, moving them to the request queue
2474 * dispatch list.
2475 */
2476static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2477{
2478 struct request *rq;
2479
2480 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2481
2482 if (!cfq_may_dispatch(cfqd, cfqq))
2483 return false;
2484
2485 /*
2486 * follow expired path, else get first next available
2487 */
2488 rq = cfq_check_fifo(cfqq);
2489 if (!rq)
2490 rq = cfqq->next_rq;
2491
2492 /*
2493 * insert request into driver dispatch list
2494 */
2495 cfq_dispatch_insert(cfqd->queue, rq);
2496
2497 if (!cfqd->active_cic) {
2498 struct cfq_io_context *cic = RQ_CIC(rq);
2499
2500 atomic_long_inc(&cic->ioc->refcount);
2501 cfqd->active_cic = cic;
2502 }
2503
2504 return true;
2505}
2506
2507/*
2508 * Find the cfqq that we need to service and move a request from that to the
2509 * dispatch list
2510 */
2511static int cfq_dispatch_requests(struct request_queue *q, int force)
2512{
2513 struct cfq_data *cfqd = q->elevator->elevator_data;
2514 struct cfq_queue *cfqq;
2515
2516 if (!cfqd->busy_queues)
2517 return 0;
2518
2519 if (unlikely(force))
2520 return cfq_forced_dispatch(cfqd);
2521
2522 cfqq = cfq_select_queue(cfqd);
2523 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002524 return 0;
2525
Jens Axboe2f5cb732009-04-07 08:51:19 +02002526 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002527 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002528 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002529 if (!cfq_dispatch_request(cfqd, cfqq))
2530 return 0;
2531
Jens Axboe2f5cb732009-04-07 08:51:19 +02002532 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002533 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002534
2535 /*
2536 * expire an async queue immediately if it has used up its slice. idle
2537 * queue always expire after 1 dispatch round.
2538 */
2539 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2540 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2541 cfq_class_idle(cfqq))) {
2542 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002543 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002544 }
2545
Shan Weib217a902009-09-01 10:06:42 +02002546 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002547 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548}
2549
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550/*
Jens Axboe5e705372006-07-13 12:39:25 +02002551 * task holds one reference to the queue, dropped when task exits. each rq
2552 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002554 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 * queue lock must be held here.
2556 */
2557static void cfq_put_queue(struct cfq_queue *cfqq)
2558{
Jens Axboe22e2c502005-06-27 10:55:12 +02002559 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05002560 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002561
Shaohua Li30d7b942011-01-07 08:46:59 +01002562 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563
Shaohua Li30d7b942011-01-07 08:46:59 +01002564 cfqq->ref--;
2565 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566 return;
2567
Jens Axboe7b679132008-05-30 12:23:07 +02002568 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002570 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002571 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002573 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002574 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002575 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002576 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002577
Vivek Goyalf04a6422009-12-03 12:59:40 -05002578 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002580 cfq_put_cfqg(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581}
2582
Jens Axboed6de8be2008-05-28 14:46:59 +02002583/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002584 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002585 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002586static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002587call_for_each_cic(struct io_context *ioc,
2588 void (*func)(struct io_context *, struct cfq_io_context *))
2589{
Jens Axboe5f45c692011-04-19 09:10:35 +02002590 struct cfq_io_context *cic;
2591 struct hlist_node *n;
2592
Jens Axboe4ac845a2008-01-24 08:44:49 +01002593 rcu_read_lock();
Jens Axboe5f45c692011-04-19 09:10:35 +02002594
2595 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2596 func(ioc, cic);
2597
Jens Axboe4ac845a2008-01-24 08:44:49 +01002598 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002599}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002600
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002601static void cfq_cic_free_rcu(struct rcu_head *head)
2602{
2603 struct cfq_io_context *cic;
2604
2605 cic = container_of(head, struct cfq_io_context, rcu_head);
2606
2607 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002608 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002609
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002610 if (ioc_gone) {
2611 /*
2612 * CFQ scheduler is exiting, grab exit lock and check
2613 * the pending io context count. If it hits zero,
2614 * complete ioc_gone and set it back to NULL
2615 */
2616 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002617 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002618 complete(ioc_gone);
2619 ioc_gone = NULL;
2620 }
2621 spin_unlock(&ioc_gone_lock);
2622 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002623}
2624
2625static void cfq_cic_free(struct cfq_io_context *cic)
2626{
2627 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002628}
2629
2630static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2631{
2632 unsigned long flags;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002633 unsigned long dead_key = (unsigned long) cic->key;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002634
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002635 BUG_ON(!(dead_key & CIC_DEAD_KEY));
Jens Axboe4ac845a2008-01-24 08:44:49 +01002636
2637 spin_lock_irqsave(&ioc->lock, flags);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002638 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
Jens Axboeffc4e752008-02-19 10:02:29 +01002639 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002640 spin_unlock_irqrestore(&ioc->lock, flags);
2641
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002642 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002643}
2644
Jens Axboed6de8be2008-05-28 14:46:59 +02002645/*
2646 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2647 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2648 * and ->trim() which is called with the task lock held
2649 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002650static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002652 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002653 * ioc->refcount is zero here, or we are called from elv_unregister(),
2654 * so no more cic's are allowed to be linked into this ioc. So it
2655 * should be ok to iterate over the known list, we will see all cic's
2656 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002657 */
Jens Axboe5f45c692011-04-19 09:10:35 +02002658 call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659}
2660
Shaohua Lid02a2c02010-05-25 10:16:53 +02002661static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002662{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002663 struct cfq_queue *__cfqq, *next;
2664
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002665 /*
2666 * If this queue was scheduled to merge with another queue, be
2667 * sure to drop the reference taken on that queue (and others in
2668 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2669 */
2670 __cfqq = cfqq->new_cfqq;
2671 while (__cfqq) {
2672 if (__cfqq == cfqq) {
2673 WARN(1, "cfqq->new_cfqq loop detected\n");
2674 break;
2675 }
2676 next = __cfqq->new_cfqq;
2677 cfq_put_queue(__cfqq);
2678 __cfqq = next;
2679 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002680}
2681
2682static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2683{
2684 if (unlikely(cfqq == cfqd->active_queue)) {
2685 __cfq_slice_expired(cfqd, cfqq, 0);
2686 cfq_schedule_dispatch(cfqd);
2687 }
2688
2689 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002690
Jens Axboe89850f72006-07-22 16:48:31 +02002691 cfq_put_queue(cfqq);
2692}
2693
2694static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2695 struct cfq_io_context *cic)
2696{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002697 struct io_context *ioc = cic->ioc;
2698
Jens Axboefc463792006-08-29 09:05:44 +02002699 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002700
2701 /*
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002702 * Make sure dead mark is seen for dead queues
Jens Axboe4ac845a2008-01-24 08:44:49 +01002703 */
Jens Axboefc463792006-08-29 09:05:44 +02002704 smp_wmb();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002705 cic->key = cfqd_dead_key(cfqd);
Jens Axboefc463792006-08-29 09:05:44 +02002706
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002707 if (ioc->ioc_data == cic)
2708 rcu_assign_pointer(ioc->ioc_data, NULL);
2709
Jens Axboeff6657c2009-04-08 10:58:57 +02002710 if (cic->cfqq[BLK_RW_ASYNC]) {
2711 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2712 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002713 }
2714
Jens Axboeff6657c2009-04-08 10:58:57 +02002715 if (cic->cfqq[BLK_RW_SYNC]) {
2716 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2717 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002718 }
Jens Axboe89850f72006-07-22 16:48:31 +02002719}
2720
Jens Axboe4ac845a2008-01-24 08:44:49 +01002721static void cfq_exit_single_io_context(struct io_context *ioc,
2722 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002723{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002724 struct cfq_data *cfqd = cic_to_cfqd(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002725
Jens Axboe89850f72006-07-22 16:48:31 +02002726 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002727 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002728 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002729
Jens Axboe4ac845a2008-01-24 08:44:49 +01002730 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002731
2732 /*
2733 * Ensure we get a fresh copy of the ->key to prevent
2734 * race between exiting task and queue
2735 */
2736 smp_read_barrier_depends();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002737 if (cic->key == cfqd)
Jens Axboe62c1fe92008-12-15 21:19:25 +01002738 __cfq_exit_single_io_context(cfqd, cic);
2739
Jens Axboe4ac845a2008-01-24 08:44:49 +01002740 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002741 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002742}
2743
Jens Axboe498d3aa22007-04-26 12:54:48 +02002744/*
2745 * The process that ioc belongs to has exited, we need to clean up
2746 * and put the internal structures we have that belongs to that process.
2747 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002748static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002750 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751}
2752
Jens Axboe22e2c502005-06-27 10:55:12 +02002753static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002754cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755{
Jens Axboeb5deef92006-07-19 23:39:40 +02002756 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757
Christoph Lameter94f60302007-07-17 04:03:29 -07002758 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2759 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002761 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002762 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002763 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002764 cic->dtor = cfq_free_io_context;
2765 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002766 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 }
2768
2769 return cic;
2770}
2771
Jens Axboefd0928d2008-01-24 08:52:45 +01002772static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002773{
2774 struct task_struct *tsk = current;
2775 int ioprio_class;
2776
Jens Axboe3b181522005-06-27 10:56:24 +02002777 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002778 return;
2779
Jens Axboefd0928d2008-01-24 08:52:45 +01002780 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002781 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002782 default:
2783 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2784 case IOPRIO_CLASS_NONE:
2785 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002786 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002787 */
2788 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002789 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002790 break;
2791 case IOPRIO_CLASS_RT:
2792 cfqq->ioprio = task_ioprio(ioc);
2793 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2794 break;
2795 case IOPRIO_CLASS_BE:
2796 cfqq->ioprio = task_ioprio(ioc);
2797 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2798 break;
2799 case IOPRIO_CLASS_IDLE:
2800 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2801 cfqq->ioprio = 7;
2802 cfq_clear_cfqq_idle_window(cfqq);
2803 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002804 }
2805
2806 /*
2807 * keep track of original prio settings in case we have to temporarily
2808 * elevate the priority of this queue
2809 */
2810 cfqq->org_ioprio = cfqq->ioprio;
2811 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002812 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002813}
2814
Jens Axboefebffd62008-01-28 13:19:43 +01002815static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002816{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002817 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002818 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002819 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002820
Jens Axboecaaa5f92006-06-16 11:23:00 +02002821 if (unlikely(!cfqd))
2822 return;
2823
Jens Axboec1b707d2006-10-30 19:54:23 +01002824 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002825
Jens Axboeff6657c2009-04-08 10:58:57 +02002826 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002827 if (cfqq) {
2828 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002829 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2830 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002831 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002832 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002833 cfq_put_queue(cfqq);
2834 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002835 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002836
Jens Axboeff6657c2009-04-08 10:58:57 +02002837 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002838 if (cfqq)
2839 cfq_mark_cfqq_prio_changed(cfqq);
2840
Jens Axboec1b707d2006-10-30 19:54:23 +01002841 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002842}
2843
Jens Axboefc463792006-08-29 09:05:44 +02002844static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002846 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002847 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848}
2849
Jens Axboed5036d72009-06-26 10:44:34 +02002850static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002851 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002852{
2853 RB_CLEAR_NODE(&cfqq->rb_node);
2854 RB_CLEAR_NODE(&cfqq->p_node);
2855 INIT_LIST_HEAD(&cfqq->fifo);
2856
Shaohua Li30d7b942011-01-07 08:46:59 +01002857 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02002858 cfqq->cfqd = cfqd;
2859
2860 cfq_mark_cfqq_prio_changed(cfqq);
2861
2862 if (is_sync) {
2863 if (!cfq_class_idle(cfqq))
2864 cfq_mark_cfqq_idle_window(cfqq);
2865 cfq_mark_cfqq_sync(cfqq);
2866 }
2867 cfqq->pid = pid;
2868}
2869
Vivek Goyal246103332009-12-03 12:59:51 -05002870#ifdef CONFIG_CFQ_GROUP_IOSCHED
2871static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2872{
2873 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002874 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal246103332009-12-03 12:59:51 -05002875 unsigned long flags;
2876 struct request_queue *q;
2877
2878 if (unlikely(!cfqd))
2879 return;
2880
2881 q = cfqd->queue;
2882
2883 spin_lock_irqsave(q->queue_lock, flags);
2884
2885 if (sync_cfqq) {
2886 /*
2887 * Drop reference to sync queue. A new sync queue will be
2888 * assigned in new group upon arrival of a fresh request.
2889 */
2890 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2891 cic_set_cfqq(cic, NULL, 1);
2892 cfq_put_queue(sync_cfqq);
2893 }
2894
2895 spin_unlock_irqrestore(q->queue_lock, flags);
2896}
2897
2898static void cfq_ioc_set_cgroup(struct io_context *ioc)
2899{
2900 call_for_each_cic(ioc, changed_cgroup);
2901 ioc->cgroup_changed = 0;
2902}
2903#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2904
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002906cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002907 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002910 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002911 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912
2913retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002914 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002915 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002916 /* cic always exists here */
2917 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918
Jens Axboe6118b702009-06-30 09:34:12 +02002919 /*
2920 * Always try a new alloc if we fell back to the OOM cfqq
2921 * originally, since it should just be a temporary situation.
2922 */
2923 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2924 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925 if (new_cfqq) {
2926 cfqq = new_cfqq;
2927 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002928 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002930 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002931 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002932 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002934 if (new_cfqq)
2935 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002936 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002937 cfqq = kmem_cache_alloc_node(cfq_pool,
2938 gfp_mask | __GFP_ZERO,
2939 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002940 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941
Jens Axboe6118b702009-06-30 09:34:12 +02002942 if (cfqq) {
2943 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2944 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002945 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002946 cfq_log_cfqq(cfqd, cfqq, "alloced");
2947 } else
2948 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 }
2950
2951 if (new_cfqq)
2952 kmem_cache_free(cfq_pool, new_cfqq);
2953
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 return cfqq;
2955}
2956
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002957static struct cfq_queue **
2958cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2959{
Jens Axboefe094d92008-01-31 13:08:54 +01002960 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002961 case IOPRIO_CLASS_RT:
2962 return &cfqd->async_cfqq[0][ioprio];
2963 case IOPRIO_CLASS_BE:
2964 return &cfqd->async_cfqq[1][ioprio];
2965 case IOPRIO_CLASS_IDLE:
2966 return &cfqd->async_idle_cfqq;
2967 default:
2968 BUG();
2969 }
2970}
2971
Jens Axboe15c31be2007-07-10 13:43:25 +02002972static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002973cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002974 gfp_t gfp_mask)
2975{
Jens Axboefd0928d2008-01-24 08:52:45 +01002976 const int ioprio = task_ioprio(ioc);
2977 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002978 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002979 struct cfq_queue *cfqq = NULL;
2980
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002981 if (!is_sync) {
2982 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2983 cfqq = *async_cfqq;
2984 }
2985
Jens Axboe6118b702009-06-30 09:34:12 +02002986 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002987 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002988
2989 /*
2990 * pin the queue now that it's allocated, scheduler exit will prune it
2991 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002992 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01002993 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002994 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002995 }
2996
Shaohua Li30d7b942011-01-07 08:46:59 +01002997 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02002998 return cfqq;
2999}
3000
Jens Axboe498d3aa22007-04-26 12:54:48 +02003001/*
3002 * We drop cfq io contexts lazily, so we may find a dead one.
3003 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003004static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01003005cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3006 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003007{
Jens Axboe4ac845a2008-01-24 08:44:49 +01003008 unsigned long flags;
3009
Jens Axboefc463792006-08-29 09:05:44 +02003010 WARN_ON(!list_empty(&cic->queue_list));
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003011 BUG_ON(cic->key != cfqd_dead_key(cfqd));
Jens Axboe597bc482007-04-24 21:23:53 +02003012
Jens Axboe4ac845a2008-01-24 08:44:49 +01003013 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02003014
Fabio Checconi4faa3c82008-04-10 08:28:01 +02003015 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003016
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003017 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
Jens Axboeffc4e752008-02-19 10:02:29 +01003018 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003019 spin_unlock_irqrestore(&ioc->lock, flags);
3020
3021 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003022}
3023
Jens Axboee2d74ac2006-03-28 08:59:01 +02003024static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01003025cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003026{
Jens Axboee2d74ac2006-03-28 08:59:01 +02003027 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003028 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003029
Vasily Tarasov91fac312007-04-25 12:29:51 +02003030 if (unlikely(!ioc))
3031 return NULL;
3032
Jens Axboed6de8be2008-05-28 14:46:59 +02003033 rcu_read_lock();
3034
Jens Axboe597bc482007-04-24 21:23:53 +02003035 /*
3036 * we maintain a last-hit cache, to avoid browsing over the tree
3037 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003038 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02003039 if (cic && cic->key == cfqd) {
3040 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02003041 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003042 }
Jens Axboe597bc482007-04-24 21:23:53 +02003043
Jens Axboe4ac845a2008-01-24 08:44:49 +01003044 do {
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003045 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003046 rcu_read_unlock();
3047 if (!cic)
3048 break;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003049 if (unlikely(cic->key != cfqd)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01003050 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003051 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01003052 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003053 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003054
Jens Axboed6de8be2008-05-28 14:46:59 +02003055 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003056 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003057 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003058 break;
3059 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003060
Jens Axboe4ac845a2008-01-24 08:44:49 +01003061 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003062}
3063
Jens Axboe4ac845a2008-01-24 08:44:49 +01003064/*
3065 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3066 * the process specific cfq io context when entered from the block layer.
3067 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3068 */
Jens Axboefebffd62008-01-28 13:19:43 +01003069static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3070 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003071{
Jens Axboe0261d682006-10-30 19:07:48 +01003072 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003073 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003074
Jens Axboe4ac845a2008-01-24 08:44:49 +01003075 ret = radix_tree_preload(gfp_mask);
3076 if (!ret) {
3077 cic->ioc = ioc;
3078 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003079
Jens Axboe4ac845a2008-01-24 08:44:49 +01003080 spin_lock_irqsave(&ioc->lock, flags);
3081 ret = radix_tree_insert(&ioc->radix_root,
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003082 cfqd->cic_index, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01003083 if (!ret)
3084 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003085 spin_unlock_irqrestore(&ioc->lock, flags);
3086
3087 radix_tree_preload_end();
3088
3089 if (!ret) {
3090 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3091 list_add(&cic->queue_list, &cfqd->cic_list);
3092 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003093 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003094 }
3095
Jens Axboe4ac845a2008-01-24 08:44:49 +01003096 if (ret)
3097 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02003098
Jens Axboe4ac845a2008-01-24 08:44:49 +01003099 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003100}
3101
Jens Axboe22e2c502005-06-27 10:55:12 +02003102/*
3103 * Setup general io context and cfq io context. There can be several cfq
3104 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02003105 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02003106 */
3107static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02003108cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109{
Jens Axboe22e2c502005-06-27 10:55:12 +02003110 struct io_context *ioc = NULL;
3111 struct cfq_io_context *cic;
3112
3113 might_sleep_if(gfp_mask & __GFP_WAIT);
3114
Jens Axboeb5deef92006-07-19 23:39:40 +02003115 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02003116 if (!ioc)
3117 return NULL;
3118
Jens Axboe4ac845a2008-01-24 08:44:49 +01003119 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003120 if (cic)
3121 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02003122
Jens Axboee2d74ac2006-03-28 08:59:01 +02003123 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3124 if (cic == NULL)
3125 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02003126
Jens Axboe4ac845a2008-01-24 08:44:49 +01003127 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3128 goto err_free;
3129
Jens Axboe22e2c502005-06-27 10:55:12 +02003130out:
Jens Axboefc463792006-08-29 09:05:44 +02003131 smp_read_barrier_depends();
3132 if (unlikely(ioc->ioprio_changed))
3133 cfq_ioc_set_ioprio(ioc);
3134
Vivek Goyal246103332009-12-03 12:59:51 -05003135#ifdef CONFIG_CFQ_GROUP_IOSCHED
3136 if (unlikely(ioc->cgroup_changed))
3137 cfq_ioc_set_cgroup(ioc);
3138#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003139 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003140err_free:
3141 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003142err:
3143 put_io_context(ioc);
3144 return NULL;
3145}
3146
3147static void
3148cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3149{
Jens Axboeaaf12282007-01-19 11:30:16 +11003150 unsigned long elapsed = jiffies - cic->last_end_request;
3151 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003152
3153 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3154 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3155 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3156}
3157
Jens Axboe206dc692006-03-28 13:03:44 +02003158static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003159cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003160 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003161{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003162 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003163 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003164 if (cfqq->last_request_pos) {
3165 if (cfqq->last_request_pos < blk_rq_pos(rq))
3166 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3167 else
3168 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3169 }
Jens Axboe206dc692006-03-28 13:03:44 +02003170
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003171 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003172 if (blk_queue_nonrot(cfqd->queue))
3173 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3174 else
3175 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003176}
Jens Axboe22e2c502005-06-27 10:55:12 +02003177
3178/*
3179 * Disable idle window if the process thinks too long or seeks so much that
3180 * it doesn't matter
3181 */
3182static void
3183cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3184 struct cfq_io_context *cic)
3185{
Jens Axboe7b679132008-05-30 12:23:07 +02003186 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003187
Jens Axboe08717142008-01-28 11:38:15 +01003188 /*
3189 * Don't idle for async or idle io prio class
3190 */
3191 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003192 return;
3193
Jens Axboec265a7f2008-06-26 13:49:33 +02003194 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003195
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003196 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3197 cfq_mark_cfqq_deep(cfqq);
3198
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003199 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3200 enable_idle = 0;
3201 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003202 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003203 enable_idle = 0;
3204 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003205 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003206 enable_idle = 0;
3207 else
3208 enable_idle = 1;
3209 }
3210
Jens Axboe7b679132008-05-30 12:23:07 +02003211 if (old_idle != enable_idle) {
3212 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3213 if (enable_idle)
3214 cfq_mark_cfqq_idle_window(cfqq);
3215 else
3216 cfq_clear_cfqq_idle_window(cfqq);
3217 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003218}
3219
Jens Axboe22e2c502005-06-27 10:55:12 +02003220/*
3221 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3222 * no or if we aren't sure, a 1 will cause a preempt.
3223 */
Jens Axboea6151c32009-10-07 20:02:57 +02003224static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003225cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003226 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003227{
Jens Axboe6d048f52007-04-25 12:44:27 +02003228 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003229
Jens Axboe6d048f52007-04-25 12:44:27 +02003230 cfqq = cfqd->active_queue;
3231 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003232 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003233
Jens Axboe6d048f52007-04-25 12:44:27 +02003234 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003235 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003236
3237 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003238 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003239
Jens Axboe22e2c502005-06-27 10:55:12 +02003240 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003241 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3242 */
3243 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3244 return false;
3245
3246 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003247 * if the new request is sync, but the currently running queue is
3248 * not, let the sync request have priority.
3249 */
Jens Axboe5e705372006-07-13 12:39:25 +02003250 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003251 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003252
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003253 if (new_cfqq->cfqg != cfqq->cfqg)
3254 return false;
3255
3256 if (cfq_slice_used(cfqq))
3257 return true;
3258
3259 /* Allow preemption only if we are idling on sync-noidle tree */
3260 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3261 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3262 new_cfqq->service_tree->count == 2 &&
3263 RB_EMPTY_ROOT(&cfqq->sort_list))
3264 return true;
3265
Jens Axboe374f84a2006-07-23 01:42:19 +02003266 /*
3267 * So both queues are sync. Let the new request get disk time if
3268 * it's a metadata request and the current queue is doing regular IO.
3269 */
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003270 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003271 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003272
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003273 /*
3274 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3275 */
3276 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003277 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003278
Shaohua Lid2d59e12010-11-08 15:01:03 +01003279 /* An idle queue should not be idle now for some reason */
3280 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3281 return true;
3282
Jens Axboe1e3335d2007-02-14 19:59:49 +01003283 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003284 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003285
3286 /*
3287 * if this request is as-good as one we would expect from the
3288 * current cfqq, let it preempt
3289 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003290 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003291 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003292
Jens Axboea6151c32009-10-07 20:02:57 +02003293 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003294}
3295
3296/*
3297 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3298 * let it have half of its nominal slice.
3299 */
3300static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3301{
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003302 struct cfq_queue *old_cfqq = cfqd->active_queue;
3303
Jens Axboe7b679132008-05-30 12:23:07 +02003304 cfq_log_cfqq(cfqd, cfqq, "preempt");
Vivek Goyale5ff0822010-04-26 19:25:11 +02003305 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003306
Jens Axboebf572252006-07-19 20:29:12 +02003307 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003308 * workload type is changed, don't save slice, otherwise preempt
3309 * doesn't happen
3310 */
3311 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3312 cfqq->cfqg->saved_workload_slice = 0;
3313
3314 /*
Jens Axboebf572252006-07-19 20:29:12 +02003315 * Put the new queue at the front of the of the current list,
3316 * so we know that it will be selected next.
3317 */
3318 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003319
3320 cfq_service_tree_add(cfqd, cfqq, 1);
Justin TerAvesteda5e0c2011-03-22 21:26:49 +01003321
Justin TerAvest62a37f62011-03-23 08:25:44 +01003322 cfqq->slice_end = 0;
3323 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003324}
3325
3326/*
Jens Axboe5e705372006-07-13 12:39:25 +02003327 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003328 * something we should do about it
3329 */
3330static void
Jens Axboe5e705372006-07-13 12:39:25 +02003331cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3332 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003333{
Jens Axboe5e705372006-07-13 12:39:25 +02003334 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003335
Aaron Carroll45333d52008-08-26 15:52:36 +02003336 cfqd->rq_queued++;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003337 if (rq->cmd_flags & REQ_META)
Jens Axboe374f84a2006-07-23 01:42:19 +02003338 cfqq->meta_pending++;
3339
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003340 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003341 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003342 cfq_update_idle_window(cfqd, cfqq, cic);
3343
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003344 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003345
3346 if (cfqq == cfqd->active_queue) {
3347 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003348 * Remember that we saw a request from this process, but
3349 * don't start queuing just yet. Otherwise we risk seeing lots
3350 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003351 * and merging. If the request is already larger than a single
3352 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003353 * merging is already done. Ditto for a busy system that
3354 * has other work pending, don't risk delaying until the
3355 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003356 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003357 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003358 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3359 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003360 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003361 cfq_clear_cfqq_wait_request(cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003362 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003363 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003364 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003365 &cfqq->cfqg->blkg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003366 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003367 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003368 }
Jens Axboe5e705372006-07-13 12:39:25 +02003369 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003370 /*
3371 * not the active queue - expire current slice if it is
3372 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003373 * has some old slice time left and is of higher priority or
3374 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003375 */
3376 cfq_preempt_queue(cfqd, cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003377 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003378 }
3379}
3380
Jens Axboe165125e2007-07-24 09:28:11 +02003381static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003382{
Jens Axboeb4878f22005-10-20 16:42:29 +02003383 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003384 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003385
Jens Axboe7b679132008-05-30 12:23:07 +02003386 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003387 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388
Jens Axboe30996f42009-10-05 11:03:39 +02003389 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003390 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003391 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003392 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003393 &cfqd->serving_group->blkg, rq_data_dir(rq),
3394 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003395 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396}
3397
Aaron Carroll45333d52008-08-26 15:52:36 +02003398/*
3399 * Update hw_tag based on peak queue depth over 50 samples under
3400 * sufficient load.
3401 */
3402static void cfq_update_hw_tag(struct cfq_data *cfqd)
3403{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003404 struct cfq_queue *cfqq = cfqd->active_queue;
3405
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003406 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3407 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003408
3409 if (cfqd->hw_tag == 1)
3410 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003411
3412 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003413 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003414 return;
3415
Shaohua Li1a1238a2009-10-27 08:46:23 +01003416 /*
3417 * If active queue hasn't enough requests and can idle, cfq might not
3418 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3419 * case
3420 */
3421 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3422 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003423 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003424 return;
3425
Aaron Carroll45333d52008-08-26 15:52:36 +02003426 if (cfqd->hw_tag_samples++ < 50)
3427 return;
3428
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003429 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003430 cfqd->hw_tag = 1;
3431 else
3432 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003433}
3434
Vivek Goyal7667aa02009-12-08 17:52:58 -05003435static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3436{
3437 struct cfq_io_context *cic = cfqd->active_cic;
3438
Justin TerAvest02a8f012011-02-09 14:20:03 +01003439 /* If the queue already has requests, don't wait */
3440 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3441 return false;
3442
Vivek Goyal7667aa02009-12-08 17:52:58 -05003443 /* If there are other queues in the group, don't wait */
3444 if (cfqq->cfqg->nr_cfqq > 1)
3445 return false;
3446
3447 if (cfq_slice_used(cfqq))
3448 return true;
3449
3450 /* if slice left is less than think time, wait busy */
3451 if (cic && sample_valid(cic->ttime_samples)
3452 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3453 return true;
3454
3455 /*
3456 * If think times is less than a jiffy than ttime_mean=0 and above
3457 * will not be true. It might happen that slice has not expired yet
3458 * but will expire soon (4-5 ns) during select_queue(). To cover the
3459 * case where think time is less than a jiffy, mark the queue wait
3460 * busy if only 1 jiffy is left in the slice.
3461 */
3462 if (cfqq->slice_end - jiffies == 1)
3463 return true;
3464
3465 return false;
3466}
3467
Jens Axboe165125e2007-07-24 09:28:11 +02003468static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469{
Jens Axboe5e705372006-07-13 12:39:25 +02003470 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003471 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003472 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003473 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474
Jens Axboeb4878f22005-10-20 16:42:29 +02003475 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003476 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3477 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478
Aaron Carroll45333d52008-08-26 15:52:36 +02003479 cfq_update_hw_tag(cfqd);
3480
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003481 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003482 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003483 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003484 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003485 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003486 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3487 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3488 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003490 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003491
Vivek Goyal365722b2009-10-03 15:21:27 +02003492 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003493 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003494 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3495 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003496 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003497
3498 /*
3499 * If this is the active queue, check if it needs to be expired,
3500 * or if we want to idle in case it has no pending requests.
3501 */
3502 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003503 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3504
Jens Axboe44f7c162007-01-19 11:51:58 +11003505 if (cfq_cfqq_slice_new(cfqq)) {
3506 cfq_set_prio_slice(cfqd, cfqq);
3507 cfq_clear_cfqq_slice_new(cfqq);
3508 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003509
3510 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003511 * Should we wait for next request to come in before we expire
3512 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003513 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003514 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003515 unsigned long extend_sl = cfqd->cfq_slice_idle;
3516 if (!cfqd->cfq_slice_idle)
3517 extend_sl = cfqd->cfq_group_idle;
3518 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003519 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003520 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003521 }
3522
Jens Axboea36e71f2009-04-15 12:15:11 +02003523 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003524 * Idling is not enabled on:
3525 * - expired queues
3526 * - idle-priority queues
3527 * - async queues
3528 * - queues with still some requests queued
3529 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003530 */
Jens Axboe08717142008-01-28 11:38:15 +01003531 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003532 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003533 else if (sync && cfqq_empty &&
3534 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003535 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003536 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003537 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003538
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003539 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003540 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541}
3542
Jens Axboe22e2c502005-06-27 10:55:12 +02003543/*
3544 * we temporarily boost lower priority queues if they are holding fs exclusive
3545 * resources. they are boosted to normal prio (CLASS_BE/4)
3546 */
3547static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548{
Jens Axboe22e2c502005-06-27 10:55:12 +02003549 if (has_fs_excl()) {
3550 /*
3551 * boost idle prio on transactions that would lock out other
3552 * users of the filesystem
3553 */
3554 if (cfq_class_idle(cfqq))
3555 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3556 if (cfqq->ioprio > IOPRIO_NORM)
3557 cfqq->ioprio = IOPRIO_NORM;
3558 } else {
3559 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003560 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003561 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003562 cfqq->ioprio_class = cfqq->org_ioprio_class;
3563 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003565}
3566
Jens Axboe89850f72006-07-22 16:48:31 +02003567static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003568{
Jens Axboe1b379d82009-08-11 08:26:11 +02003569 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003570 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003571 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003572 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003573
3574 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003575}
3576
Jens Axboe165125e2007-07-24 09:28:11 +02003577static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003578{
3579 struct cfq_data *cfqd = q->elevator->elevator_data;
3580 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003581 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003582 struct cfq_queue *cfqq;
3583
3584 /*
3585 * don't force setup of a queue from here, as a call to may_queue
3586 * does not necessarily imply that a request actually will be queued.
3587 * so just lookup a possibly existing queue, or return 'may queue'
3588 * if that fails
3589 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003590 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003591 if (!cic)
3592 return ELV_MQUEUE_MAY;
3593
Jens Axboeb0b78f82009-04-08 10:56:08 +02003594 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003595 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003596 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003597 cfq_prio_boost(cfqq);
3598
Jens Axboe89850f72006-07-22 16:48:31 +02003599 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003600 }
3601
3602 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603}
3604
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605/*
3606 * queue lock held here
3607 */
Jens Axboebb37b942006-12-01 10:42:33 +01003608static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609{
Jens Axboe5e705372006-07-13 12:39:25 +02003610 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611
Jens Axboe5e705372006-07-13 12:39:25 +02003612 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003613 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614
Jens Axboe22e2c502005-06-27 10:55:12 +02003615 BUG_ON(!cfqq->allocated[rw]);
3616 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617
Jens Axboe5e705372006-07-13 12:39:25 +02003618 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619
Mike Snitzerc1867942011-02-11 11:08:00 +01003620 rq->elevator_private[0] = NULL;
3621 rq->elevator_private[1] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003623 /* Put down rq reference on cfqg */
3624 cfq_put_cfqg(RQ_CFQG(rq));
Mike Snitzerc1867942011-02-11 11:08:00 +01003625 rq->elevator_private[2] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003626
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627 cfq_put_queue(cfqq);
3628 }
3629}
3630
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003631static struct cfq_queue *
3632cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3633 struct cfq_queue *cfqq)
3634{
3635 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3636 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003637 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003638 cfq_put_queue(cfqq);
3639 return cic_to_cfqq(cic, 1);
3640}
3641
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003642/*
3643 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3644 * was the last process referring to said cfqq.
3645 */
3646static struct cfq_queue *
3647split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3648{
3649 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003650 cfqq->pid = current->pid;
3651 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003652 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003653 return cfqq;
3654 }
3655
3656 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003657
3658 cfq_put_cooperator(cfqq);
3659
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003660 cfq_put_queue(cfqq);
3661 return NULL;
3662}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003663/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003664 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003665 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003666static int
Jens Axboe165125e2007-07-24 09:28:11 +02003667cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668{
3669 struct cfq_data *cfqd = q->elevator->elevator_data;
3670 struct cfq_io_context *cic;
3671 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003672 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003673 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674 unsigned long flags;
3675
3676 might_sleep_if(gfp_mask & __GFP_WAIT);
3677
Jens Axboee2d74ac2006-03-28 08:59:01 +02003678 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003679
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680 spin_lock_irqsave(q->queue_lock, flags);
3681
Jens Axboe22e2c502005-06-27 10:55:12 +02003682 if (!cic)
3683 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003685new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003686 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003687 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003688 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003689 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003690 } else {
3691 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003692 * If the queue was seeky for too long, break it apart.
3693 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003694 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003695 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3696 cfqq = split_cfqq(cic, cfqq);
3697 if (!cfqq)
3698 goto new_queue;
3699 }
3700
3701 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003702 * Check to see if this queue is scheduled to merge with
3703 * another, closely cooperating queue. The merging of
3704 * queues happens here as it must be done in process context.
3705 * The reference on new_cfqq was taken in merge_cfqqs.
3706 */
3707 if (cfqq->new_cfqq)
3708 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003709 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710
3711 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02003712
Jens Axboe6fae9c22011-03-01 15:04:39 -05003713 cfqq->ref++;
Mike Snitzerc1867942011-02-11 11:08:00 +01003714 rq->elevator_private[0] = cic;
3715 rq->elevator_private[1] = cfqq;
3716 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
Jens Axboe93803e02011-03-07 08:59:06 +01003717 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe5e705372006-07-13 12:39:25 +02003718 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003719
Jens Axboe22e2c502005-06-27 10:55:12 +02003720queue_fail:
3721 if (cic)
3722 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003723
Jens Axboe23e018a2009-10-05 08:52:35 +02003724 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003726 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727 return 1;
3728}
3729
David Howells65f27f32006-11-22 14:55:48 +00003730static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003731{
David Howells65f27f32006-11-22 14:55:48 +00003732 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003733 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003734 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003735
Jens Axboe40bb54d2009-04-15 12:11:10 +02003736 spin_lock_irq(q->queue_lock);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003737 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003738 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003739}
3740
3741/*
3742 * Timer running if the active_queue is currently idling inside its time slice
3743 */
3744static void cfq_idle_slice_timer(unsigned long data)
3745{
3746 struct cfq_data *cfqd = (struct cfq_data *) data;
3747 struct cfq_queue *cfqq;
3748 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003749 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003750
Jens Axboe7b679132008-05-30 12:23:07 +02003751 cfq_log(cfqd, "idle timer fired");
3752
Jens Axboe22e2c502005-06-27 10:55:12 +02003753 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3754
Jens Axboefe094d92008-01-31 13:08:54 +01003755 cfqq = cfqd->active_queue;
3756 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003757 timed_out = 0;
3758
Jens Axboe22e2c502005-06-27 10:55:12 +02003759 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003760 * We saw a request before the queue expired, let it through
3761 */
3762 if (cfq_cfqq_must_dispatch(cfqq))
3763 goto out_kick;
3764
3765 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003766 * expired
3767 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003768 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003769 goto expire;
3770
3771 /*
3772 * only expire and reinvoke request handler, if there are
3773 * other queues with pending requests
3774 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003775 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003776 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003777
3778 /*
3779 * not expired and it has a request pending, let it dispatch
3780 */
Jens Axboe75e50982009-04-07 08:56:14 +02003781 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003782 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003783
3784 /*
3785 * Queue depth flag is reset only when the idle didn't succeed
3786 */
3787 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003788 }
3789expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003790 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003791out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003792 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003793out_cont:
3794 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3795}
3796
Jens Axboe3b181522005-06-27 10:56:24 +02003797static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3798{
3799 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003800 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003801}
Jens Axboe22e2c502005-06-27 10:55:12 +02003802
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003803static void cfq_put_async_queues(struct cfq_data *cfqd)
3804{
3805 int i;
3806
3807 for (i = 0; i < IOPRIO_BE_NR; i++) {
3808 if (cfqd->async_cfqq[0][i])
3809 cfq_put_queue(cfqd->async_cfqq[0][i]);
3810 if (cfqd->async_cfqq[1][i])
3811 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003812 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003813
3814 if (cfqd->async_idle_cfqq)
3815 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003816}
3817
Jens Axboebb729bc2009-12-06 09:54:19 +01003818static void cfq_cfqd_free(struct rcu_head *head)
3819{
3820 kfree(container_of(head, struct cfq_data, rcu));
3821}
3822
Jens Axboeb374d182008-10-31 10:05:07 +01003823static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824{
Jens Axboe22e2c502005-06-27 10:55:12 +02003825 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003826 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003827
Jens Axboe3b181522005-06-27 10:56:24 +02003828 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003829
Al Virod9ff4182006-03-18 13:51:22 -05003830 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003831
Al Virod9ff4182006-03-18 13:51:22 -05003832 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003833 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003834
3835 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003836 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3837 struct cfq_io_context,
3838 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003839
3840 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003841 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003842
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003843 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003844 cfq_release_cfq_groups(cfqd);
Vivek Goyale98ef892010-06-18 10:39:47 -04003845 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003846
Al Virod9ff4182006-03-18 13:51:22 -05003847 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003848
3849 cfq_shutdown_timer_wq(cfqd);
3850
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003851 spin_lock(&cic_index_lock);
3852 ida_remove(&cic_index_ida, cfqd->cic_index);
3853 spin_unlock(&cic_index_lock);
3854
Vivek Goyalb1c35762009-12-03 12:59:47 -05003855 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003856 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003857}
3858
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003859static int cfq_alloc_cic_index(void)
3860{
3861 int index, error;
3862
3863 do {
3864 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3865 return -ENOMEM;
3866
3867 spin_lock(&cic_index_lock);
3868 error = ida_get_new(&cic_index_ida, &index);
3869 spin_unlock(&cic_index_lock);
3870 if (error && error != -EAGAIN)
3871 return error;
3872 } while (error);
3873
3874 return index;
3875}
3876
Jens Axboe165125e2007-07-24 09:28:11 +02003877static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878{
3879 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003880 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003881 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003882 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003884 i = cfq_alloc_cic_index();
3885 if (i < 0)
3886 return NULL;
3887
Christoph Lameter94f60302007-07-17 04:03:29 -07003888 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003890 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003891
Shaohua Li30d7b942011-01-07 08:46:59 +01003892 /*
3893 * Don't need take queue_lock in the routine, since we are
3894 * initializing the ioscheduler, and nobody is using cfqd
3895 */
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003896 cfqd->cic_index = i;
3897
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003898 /* Init root service tree */
3899 cfqd->grp_service_tree = CFQ_RB_ROOT;
3900
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003901 /* Init root group */
3902 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003903 for_each_cfqg_st(cfqg, i, j, st)
3904 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003905 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003906
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003907 /* Give preference to root group over other groups */
3908 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3909
Vivek Goyal25fb5162009-12-03 12:59:46 -05003910#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003911 /*
3912 * Take a reference to root group which we never drop. This is just
3913 * to make sure that cfq_put_cfqg() does not try to kfree root group
3914 */
Shaohua Li329a6782011-01-07 08:48:28 +01003915 cfqg->ref = 1;
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003916 rcu_read_lock();
Vivek Goyale98ef892010-06-18 10:39:47 -04003917 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3918 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003919 rcu_read_unlock();
Vivek Goyal25fb5162009-12-03 12:59:46 -05003920#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003921 /*
3922 * Not strictly needed (since RB_ROOT just clears the node and we
3923 * zeroed cfqd on alloc), but better be safe in case someone decides
3924 * to add magic to the rb code
3925 */
3926 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3927 cfqd->prio_trees[i] = RB_ROOT;
3928
Jens Axboe6118b702009-06-30 09:34:12 +02003929 /*
3930 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3931 * Grab a permanent reference to it, so that the normal code flow
3932 * will not attempt to free it.
3933 */
3934 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01003935 cfqd->oom_cfqq.ref++;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003936 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003937
Al Virod9ff4182006-03-18 13:51:22 -05003938 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941
Jens Axboe22e2c502005-06-27 10:55:12 +02003942 init_timer(&cfqd->idle_slice_timer);
3943 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3944 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3945
Jens Axboe23e018a2009-10-05 08:52:35 +02003946 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003947
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003949 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3950 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 cfqd->cfq_back_max = cfq_back_max;
3952 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003953 cfqd->cfq_slice[0] = cfq_slice_async;
3954 cfqd->cfq_slice[1] = cfq_slice_sync;
3955 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3956 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003957 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003958 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003959 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003960 /*
3961 * we optimistically start assuming sync ops weren't delayed in last
3962 * second, in order to have larger depth for async operations.
3963 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003964 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003965 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966}
3967
3968static void cfq_slab_kill(void)
3969{
Jens Axboed6de8be2008-05-28 14:46:59 +02003970 /*
3971 * Caller already ensured that pending RCU callbacks are completed,
3972 * so we should have no busy allocations at this point.
3973 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974 if (cfq_pool)
3975 kmem_cache_destroy(cfq_pool);
3976 if (cfq_ioc_pool)
3977 kmem_cache_destroy(cfq_ioc_pool);
3978}
3979
3980static int __init cfq_slab_setup(void)
3981{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003982 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 if (!cfq_pool)
3984 goto fail;
3985
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003986 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987 if (!cfq_ioc_pool)
3988 goto fail;
3989
3990 return 0;
3991fail:
3992 cfq_slab_kill();
3993 return -ENOMEM;
3994}
3995
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996/*
3997 * sysfs parts below -->
3998 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999static ssize_t
4000cfq_var_show(unsigned int var, char *page)
4001{
4002 return sprintf(page, "%d\n", var);
4003}
4004
4005static ssize_t
4006cfq_var_store(unsigned int *var, const char *page, size_t count)
4007{
4008 char *p = (char *) page;
4009
4010 *var = simple_strtoul(p, &p, 10);
4011 return count;
4012}
4013
Linus Torvalds1da177e2005-04-16 15:20:36 -07004014#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004015static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004017 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018 unsigned int __data = __VAR; \
4019 if (__CONV) \
4020 __data = jiffies_to_msecs(__data); \
4021 return cfq_var_show(__data, (page)); \
4022}
4023SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004024SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4025SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05004026SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4027SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004028SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004029SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004030SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4031SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4032SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004033SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004034#undef SHOW_FUNCTION
4035
4036#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004037static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004039 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040 unsigned int __data; \
4041 int ret = cfq_var_store(&__data, (page), count); \
4042 if (__data < (MIN)) \
4043 __data = (MIN); \
4044 else if (__data > (MAX)) \
4045 __data = (MAX); \
4046 if (__CONV) \
4047 *(__PTR) = msecs_to_jiffies(__data); \
4048 else \
4049 *(__PTR) = __data; \
4050 return ret; \
4051}
4052STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004053STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4054 UINT_MAX, 1);
4055STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4056 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004057STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004058STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4059 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004060STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004061STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004062STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4063STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004064STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4065 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004066STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067#undef STORE_FUNCTION
4068
Al Viroe572ec72006-03-18 22:27:18 -05004069#define CFQ_ATTR(name) \
4070 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004071
Al Viroe572ec72006-03-18 22:27:18 -05004072static struct elv_fs_entry cfq_attrs[] = {
4073 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004074 CFQ_ATTR(fifo_expire_sync),
4075 CFQ_ATTR(fifo_expire_async),
4076 CFQ_ATTR(back_seek_max),
4077 CFQ_ATTR(back_seek_penalty),
4078 CFQ_ATTR(slice_sync),
4079 CFQ_ATTR(slice_async),
4080 CFQ_ATTR(slice_async_rq),
4081 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004082 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004083 CFQ_ATTR(low_latency),
Al Viroe572ec72006-03-18 22:27:18 -05004084 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085};
4086
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087static struct elevator_type iosched_cfq = {
4088 .ops = {
4089 .elevator_merge_fn = cfq_merge,
4090 .elevator_merged_fn = cfq_merged_request,
4091 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004092 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004093 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004094 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004095 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004096 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097 .elevator_deactivate_req_fn = cfq_deactivate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004099 .elevator_former_req_fn = elv_rb_former_request,
4100 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101 .elevator_set_req_fn = cfq_set_request,
4102 .elevator_put_req_fn = cfq_put_request,
4103 .elevator_may_queue_fn = cfq_may_queue,
4104 .elevator_init_fn = cfq_init_queue,
4105 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02004106 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 },
Al Viro3d1ab402006-03-18 18:35:43 -05004108 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109 .elevator_name = "cfq",
4110 .elevator_owner = THIS_MODULE,
4111};
4112
Vivek Goyal3e252062009-12-04 10:36:42 -05004113#ifdef CONFIG_CFQ_GROUP_IOSCHED
4114static struct blkio_policy_type blkio_policy_cfq = {
4115 .ops = {
4116 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4117 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4118 },
Vivek Goyal062a6442010-09-15 17:06:33 -04004119 .plid = BLKIO_POLICY_PROP,
Vivek Goyal3e252062009-12-04 10:36:42 -05004120};
4121#else
4122static struct blkio_policy_type blkio_policy_cfq;
4123#endif
4124
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125static int __init cfq_init(void)
4126{
Jens Axboe22e2c502005-06-27 10:55:12 +02004127 /*
4128 * could be 0 on HZ < 1000 setups
4129 */
4130 if (!cfq_slice_async)
4131 cfq_slice_async = 1;
4132 if (!cfq_slice_idle)
4133 cfq_slice_idle = 1;
4134
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004135#ifdef CONFIG_CFQ_GROUP_IOSCHED
4136 if (!cfq_group_idle)
4137 cfq_group_idle = 1;
4138#else
4139 cfq_group_idle = 0;
4140#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004141 if (cfq_slab_setup())
4142 return -ENOMEM;
4143
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004144 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05004145 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004147 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148}
4149
4150static void __exit cfq_exit(void)
4151{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07004152 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05004153 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004154 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05004155 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02004156 /* ioc_gone's update must be visible before reading ioc_count */
4157 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02004158
4159 /*
4160 * this also protects us from entering cfq_slab_kill() with
4161 * pending RCU callbacks
4162 */
Tejun Heo245b2e72009-06-24 15:13:48 +09004163 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02004164 wait_for_completion(&all_gone);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004165 ida_destroy(&cic_index_ida);
Christoph Hellwig83521d32005-10-30 15:01:39 -08004166 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167}
4168
4169module_init(cfq_init);
4170module_exit(cfq_exit);
4171
4172MODULE_AUTHOR("Jens Axboe");
4173MODULE_LICENSE("GPL");
4174MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");