blob: fb2141ec205ca181f8d56c5e4d04f53076096c3c [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;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500181
182 /* number of cfqq currently on this group */
183 int nr_cfqq;
184
Jens Axboe22e2c502005-06-27 10:55:12 +0200185 /*
Vivek Goyalb4627322010-10-22 09:48:43 +0200186 * Per group busy queus average. Useful for workload slice calc. We
187 * create the array for each prio class but at run time it is used
188 * only for RT and BE class and slot for IDLE class remains unused.
189 * This is primarily done to avoid confusion and a gcc warning.
190 */
191 unsigned int busy_queues_avg[CFQ_PRIO_NR];
192 /*
193 * rr lists of queues with requests. We maintain service trees for
194 * RT and BE classes. These trees are subdivided in subclasses
195 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
196 * class there is no subclassification and all the cfq queues go on
197 * a single tree service_tree_idle.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100198 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200199 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100200 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100201 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500202
203 unsigned long saved_workload_slice;
204 enum wl_type_t saved_workload;
205 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500206 struct blkio_group blkg;
207#ifdef CONFIG_CFQ_GROUP_IOSCHED
208 struct hlist_node cfqd_node;
Shaohua Li329a6782011-01-07 08:48:28 +0100209 int ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500210#endif
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200211 /* number of requests that are on the dispatch list or inside driver */
212 int dispatched;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500213};
214
215/*
216 * Per block device queue structure
217 */
218struct cfq_data {
219 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500220 /* Root service tree for cfq_groups */
221 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500222 struct cfq_group root_group;
223
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100224 /*
225 * The priority currently being served
226 */
227 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100228 enum wl_type_t serving_type;
229 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500230 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200231
232 /*
233 * Each priority tree is sorted by next_request position. These
234 * trees are used when determining if two or more queues are
235 * interleaving requests (see cfq_close_cooperator).
236 */
237 struct rb_root prio_trees[CFQ_PRIO_LISTS];
238
Jens Axboe22e2c502005-06-27 10:55:12 +0200239 unsigned int busy_queues;
240
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100241 int rq_in_driver;
242 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200243
244 /*
245 * queue-depth detection
246 */
247 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200248 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100249 /*
250 * hw_tag can be
251 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
252 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
253 * 0 => no NCQ
254 */
255 int hw_tag_est_depth;
256 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200257
258 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200259 * idle window management
260 */
261 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200262 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200263
264 struct cfq_queue *active_queue;
265 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200266
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200267 /*
268 * async queue for each priority case
269 */
270 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
271 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200272
Jens Axboe6d048f52007-04-25 12:44:27 +0200273 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 /*
276 * tunables, see top of file
277 */
278 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200279 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 unsigned int cfq_back_penalty;
281 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200282 unsigned int cfq_slice[2];
283 unsigned int cfq_slice_async_rq;
284 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200285 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200286 unsigned int cfq_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500287
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400288 unsigned int cic_index;
Al Virod9ff4182006-03-18 13:51:22 -0500289 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290
Jens Axboe6118b702009-06-30 09:34:12 +0200291 /*
292 * Fallback dummy cfqq for extreme OOM conditions
293 */
294 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200295
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100296 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500297
298 /* List of cfq groups being managed on this device*/
299 struct hlist_head cfqg_list;
Jens Axboebb729bc2009-12-06 09:54:19 +0100300 struct rcu_head rcu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301};
302
Vivek Goyal25fb5162009-12-03 12:59:46 -0500303static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
304
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500305static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
306 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500307 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100308{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500309 if (!cfqg)
310 return NULL;
311
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100312 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500313 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100314
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500315 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100316}
317
Jens Axboe3b181522005-06-27 10:56:24 +0200318enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100319 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
320 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200321 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100322 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100323 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
324 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
325 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100326 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200327 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400328 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100329 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100330 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500331 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200332};
333
334#define CFQ_CFQQ_FNS(name) \
335static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
336{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100337 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200338} \
339static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
340{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100341 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200342} \
343static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
344{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100345 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200346}
347
348CFQ_CFQQ_FNS(on_rr);
349CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200350CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200351CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200352CFQ_CFQQ_FNS(fifo_expire);
353CFQ_CFQQ_FNS(idle_window);
354CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100355CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200356CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200357CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100358CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100359CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500360CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200361#undef CFQ_CFQQ_FNS
362
Vivek Goyalafc24d42010-04-26 19:27:56 +0200363#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500364#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
365 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
366 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
367 blkg_path(&(cfqq)->cfqg->blkg), ##args);
368
369#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
370 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
371 blkg_path(&(cfqg)->blkg), ##args); \
372
373#else
Jens Axboe7b679132008-05-30 12:23:07 +0200374#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
375 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500376#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
377#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200378#define cfq_log(cfqd, fmt, args...) \
379 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
380
Vivek Goyal615f0252009-12-03 12:59:39 -0500381/* Traverses through cfq group service trees */
382#define for_each_cfqg_st(cfqg, i, j, st) \
383 for (i = 0; i <= IDLE_WORKLOAD; i++) \
384 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
385 : &cfqg->service_tree_idle; \
386 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
387 (i == IDLE_WORKLOAD && j == 0); \
388 j++, st = i < IDLE_WORKLOAD ? \
389 &cfqg->service_trees[i][j]: NULL) \
390
391
Vivek Goyal02b35082010-08-23 12:23:53 +0200392static inline bool iops_mode(struct cfq_data *cfqd)
393{
394 /*
395 * If we are not idling on queues and it is a NCQ drive, parallel
396 * execution of requests is on and measuring time is not possible
397 * in most of the cases until and unless we drive shallower queue
398 * depths and that becomes a performance bottleneck. In such cases
399 * switch to start providing fairness in terms of number of IOs.
400 */
401 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
402 return true;
403 else
404 return false;
405}
406
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100407static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
408{
409 if (cfq_class_idle(cfqq))
410 return IDLE_WORKLOAD;
411 if (cfq_class_rt(cfqq))
412 return RT_WORKLOAD;
413 return BE_WORKLOAD;
414}
415
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100416
417static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
418{
419 if (!cfq_cfqq_sync(cfqq))
420 return ASYNC_WORKLOAD;
421 if (!cfq_cfqq_idle_window(cfqq))
422 return SYNC_NOIDLE_WORKLOAD;
423 return SYNC_WORKLOAD;
424}
425
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500426static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
427 struct cfq_data *cfqd,
428 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100429{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500430 if (wl == IDLE_WORKLOAD)
431 return cfqg->service_tree_idle.count;
432
433 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
434 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
435 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100436}
437
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500438static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
439 struct cfq_group *cfqg)
440{
441 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
442 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
443}
444
Jens Axboe165125e2007-07-24 09:28:11 +0200445static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200446static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100447 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100448static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200449 struct io_context *);
450
451static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200452 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200453{
Jens Axboea6151c32009-10-07 20:02:57 +0200454 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200455}
456
457static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200458 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200459{
Jens Axboea6151c32009-10-07 20:02:57 +0200460 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200461}
462
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400463#define CIC_DEAD_KEY 1ul
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400464#define CIC_DEAD_INDEX_SHIFT 1
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400465
466static inline void *cfqd_dead_key(struct cfq_data *cfqd)
467{
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400468 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400469}
470
471static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
472{
473 struct cfq_data *cfqd = cic->key;
474
475 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
476 return NULL;
477
478 return cfqd;
479}
480
Vasily Tarasov91fac312007-04-25 12:29:51 +0200481/*
482 * We regard a request as SYNC, if it's either a read or has the SYNC bit
483 * set (in which case it could also be direct WRITE).
484 */
Jens Axboea6151c32009-10-07 20:02:57 +0200485static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200486{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200487 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200488}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700491 * scheduler run of queue, if there are requests pending and no one in the
492 * driver that will restart queueing
493 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200494static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700495{
Jens Axboe7b679132008-05-30 12:23:07 +0200496 if (cfqd->busy_queues) {
497 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200498 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200499 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700500}
501
Jens Axboe165125e2007-07-24 09:28:11 +0200502static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700503{
504 struct cfq_data *cfqd = q->elevator->elevator_data;
505
Vivek Goyalf04a6422009-12-03 12:59:40 -0500506 return !cfqd->rq_queued;
Andrew Morton99f95e52005-06-27 20:14:05 -0700507}
508
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100510 * Scale schedule slice based on io priority. Use the sync time slice only
511 * if a queue is marked sync and has sync io queued. A sync queue with async
512 * io only, should not get full sync slice length.
513 */
Jens Axboea6151c32009-10-07 20:02:57 +0200514static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200515 unsigned short prio)
516{
517 const int base_slice = cfqd->cfq_slice[sync];
518
519 WARN_ON(prio >= IOPRIO_BE_NR);
520
521 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
522}
523
Jens Axboe44f7c162007-01-19 11:51:58 +1100524static inline int
525cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
526{
Jens Axboed9e76202007-04-20 14:27:50 +0200527 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100528}
529
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500530static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
531{
532 u64 d = delta << CFQ_SERVICE_SHIFT;
533
534 d = d * BLKIO_WEIGHT_DEFAULT;
535 do_div(d, cfqg->weight);
536 return d;
537}
538
539static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
540{
541 s64 delta = (s64)(vdisktime - min_vdisktime);
542 if (delta > 0)
543 min_vdisktime = vdisktime;
544
545 return min_vdisktime;
546}
547
548static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
549{
550 s64 delta = (s64)(vdisktime - min_vdisktime);
551 if (delta < 0)
552 min_vdisktime = vdisktime;
553
554 return min_vdisktime;
555}
556
557static void update_min_vdisktime(struct cfq_rb_root *st)
558{
559 u64 vdisktime = st->min_vdisktime;
560 struct cfq_group *cfqg;
561
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500562 if (st->left) {
563 cfqg = rb_entry_cfqg(st->left);
564 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
565 }
566
567 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
568}
569
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100570/*
571 * get averaged number of queues of RT/BE priority.
572 * average is updated, with a formula that gives more weight to higher numbers,
573 * to quickly follows sudden increases and decrease slowly
574 */
575
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500576static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
577 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100578{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100579 unsigned min_q, max_q;
580 unsigned mult = cfq_hist_divisor - 1;
581 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500582 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100583
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500584 min_q = min(cfqg->busy_queues_avg[rt], busy);
585 max_q = max(cfqg->busy_queues_avg[rt], busy);
586 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100587 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500588 return cfqg->busy_queues_avg[rt];
589}
590
591static inline unsigned
592cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
593{
594 struct cfq_rb_root *st = &cfqd->grp_service_tree;
595
596 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100597}
598
Shaohua Lic553f8e2011-01-14 08:41:03 +0100599static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -0700600cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100601{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100602 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
603 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500604 /*
605 * interested queues (we consider only the ones with the same
606 * priority class in the cfq group)
607 */
608 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
609 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100610 unsigned sync_slice = cfqd->cfq_slice[1];
611 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500612 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
613
614 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100615 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
616 /* scale low_slice according to IO priority
617 * and sync vs async */
618 unsigned low_slice =
619 min(slice, base_low_slice * slice / sync_slice);
620 /* the adapted slice value is scaled to fit all iqs
621 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500622 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100623 low_slice);
624 }
625 }
Shaohua Lic553f8e2011-01-14 08:41:03 +0100626 return slice;
627}
628
629static inline void
630cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
631{
Vivek Goyalba5bd522011-01-19 08:25:02 -0700632 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +0100633
Vivek Goyaldae739e2009-12-03 12:59:45 -0500634 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100635 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500636 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200637 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100638}
639
640/*
641 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
642 * isn't valid until the first request from the dispatch is activated
643 * and the slice time set.
644 */
Jens Axboea6151c32009-10-07 20:02:57 +0200645static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100646{
647 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +0100648 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100649 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +0100650 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100651
Shaohua Lic1e44752010-11-08 15:01:02 +0100652 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +1100653}
654
655/*
Jens Axboe5e705372006-07-13 12:39:25 +0200656 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200658 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 */
Jens Axboe5e705372006-07-13 12:39:25 +0200660static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100661cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100663 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200665#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
666#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
667 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668
Jens Axboe5e705372006-07-13 12:39:25 +0200669 if (rq1 == NULL || rq1 == rq2)
670 return rq2;
671 if (rq2 == NULL)
672 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200673
Jens Axboe5e705372006-07-13 12:39:25 +0200674 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
675 return rq1;
676 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
677 return rq2;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200678 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200679 return rq1;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200680 else if ((rq2->cmd_flags & REQ_META) &&
681 !(rq1->cmd_flags & REQ_META))
Jens Axboe374f84a2006-07-23 01:42:19 +0200682 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
Tejun Heo83096eb2009-05-07 22:24:39 +0900684 s1 = blk_rq_pos(rq1);
685 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 /*
688 * by definition, 1KiB is 2 sectors
689 */
690 back_max = cfqd->cfq_back_max * 2;
691
692 /*
693 * Strict one way elevator _except_ in the case where we allow
694 * short backward seeks which are biased as twice the cost of a
695 * similar forward seek.
696 */
697 if (s1 >= last)
698 d1 = s1 - last;
699 else if (s1 + back_max >= last)
700 d1 = (last - s1) * cfqd->cfq_back_penalty;
701 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200702 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
704 if (s2 >= last)
705 d2 = s2 - last;
706 else if (s2 + back_max >= last)
707 d2 = (last - s2) * cfqd->cfq_back_penalty;
708 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200709 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
711 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712
Andreas Mohre8a99052006-03-28 08:59:49 +0200713 /*
714 * By doing switch() on the bit mask "wrap" we avoid having to
715 * check two variables for all permutations: --> faster!
716 */
717 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200718 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200719 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200720 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200721 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200722 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200723 else {
724 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200725 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200726 else
Jens Axboe5e705372006-07-13 12:39:25 +0200727 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200728 }
729
730 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200731 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200732 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200733 return rq2;
734 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200735 default:
736 /*
737 * Since both rqs are wrapped,
738 * start with the one that's further behind head
739 * (--> only *one* back seek required),
740 * since back seek takes more time than forward.
741 */
742 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200743 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 else
Jens Axboe5e705372006-07-13 12:39:25 +0200745 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 }
747}
748
Jens Axboe498d3aa22007-04-26 12:54:48 +0200749/*
750 * The below is leftmost cache rbtree addon
751 */
Jens Axboe08717142008-01-28 11:38:15 +0100752static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200753{
Vivek Goyal615f0252009-12-03 12:59:39 -0500754 /* Service tree is empty */
755 if (!root->count)
756 return NULL;
757
Jens Axboecc09e292007-04-26 12:53:50 +0200758 if (!root->left)
759 root->left = rb_first(&root->rb);
760
Jens Axboe08717142008-01-28 11:38:15 +0100761 if (root->left)
762 return rb_entry(root->left, struct cfq_queue, rb_node);
763
764 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200765}
766
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500767static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
768{
769 if (!root->left)
770 root->left = rb_first(&root->rb);
771
772 if (root->left)
773 return rb_entry_cfqg(root->left);
774
775 return NULL;
776}
777
Jens Axboea36e71f2009-04-15 12:15:11 +0200778static void rb_erase_init(struct rb_node *n, struct rb_root *root)
779{
780 rb_erase(n, root);
781 RB_CLEAR_NODE(n);
782}
783
Jens Axboecc09e292007-04-26 12:53:50 +0200784static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
785{
786 if (root->left == n)
787 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200788 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100789 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200790}
791
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792/*
793 * would be nice to take fifo expire time into account as well
794 */
Jens Axboe5e705372006-07-13 12:39:25 +0200795static struct request *
796cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
797 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798{
Jens Axboe21183b02006-07-13 12:33:14 +0200799 struct rb_node *rbnext = rb_next(&last->rb_node);
800 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200801 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802
Jens Axboe21183b02006-07-13 12:33:14 +0200803 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804
805 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200806 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200807
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200809 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200810 else {
811 rbnext = rb_first(&cfqq->sort_list);
812 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200813 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200814 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100816 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817}
818
Jens Axboed9e76202007-04-20 14:27:50 +0200819static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
820 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821{
Jens Axboed9e76202007-04-20 14:27:50 +0200822 /*
823 * just an approximation, should be ok.
824 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500825 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100826 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200827}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500829static inline s64
830cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
831{
832 return cfqg->vdisktime - st->min_vdisktime;
833}
834
835static void
836__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
837{
838 struct rb_node **node = &st->rb.rb_node;
839 struct rb_node *parent = NULL;
840 struct cfq_group *__cfqg;
841 s64 key = cfqg_key(st, cfqg);
842 int left = 1;
843
844 while (*node != NULL) {
845 parent = *node;
846 __cfqg = rb_entry_cfqg(parent);
847
848 if (key < cfqg_key(st, __cfqg))
849 node = &parent->rb_left;
850 else {
851 node = &parent->rb_right;
852 left = 0;
853 }
854 }
855
856 if (left)
857 st->left = &cfqg->rb_node;
858
859 rb_link_node(&cfqg->rb_node, parent, node);
860 rb_insert_color(&cfqg->rb_node, &st->rb);
861}
862
863static void
864cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
865{
866 struct cfq_rb_root *st = &cfqd->grp_service_tree;
867 struct cfq_group *__cfqg;
868 struct rb_node *n;
869
870 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +0100871 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500872 return;
873
874 /*
875 * Currently put the group at the end. Later implement something
876 * so that groups get lesser vtime based on their weights, so that
877 * if group does not loose all if it was not continously backlogged.
878 */
879 n = rb_last(&st->rb);
880 if (n) {
881 __cfqg = rb_entry_cfqg(n);
882 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
883 } else
884 cfqg->vdisktime = st->min_vdisktime;
885
886 __cfq_group_service_tree_add(st, cfqg);
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500887 st->total_weight += cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500888}
889
890static void
891cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
892{
893 struct cfq_rb_root *st = &cfqd->grp_service_tree;
894
895 BUG_ON(cfqg->nr_cfqq < 1);
896 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500897
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500898 /* If there are other cfq queues under this group, don't delete it */
899 if (cfqg->nr_cfqq)
900 return;
901
Vivek Goyal2868ef72009-12-03 12:59:48 -0500902 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500903 st->total_weight -= cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500904 if (!RB_EMPTY_NODE(&cfqg->rb_node))
905 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500906 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400907 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500908}
909
910static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
911{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500912 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500913
914 /*
915 * Queue got expired before even a single request completed or
916 * got expired immediately after first request completion.
917 */
918 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
919 /*
920 * Also charge the seek time incurred to the group, otherwise
921 * if there are mutiple queues in the group, each can dispatch
922 * a single request on seeky media and cause lots of seek time
923 * and group will never know it.
924 */
925 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
926 1);
927 } else {
928 slice_used = jiffies - cfqq->slice_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500929 if (slice_used > cfqq->allocated_slice)
930 slice_used = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500931 }
932
Vivek Goyaldae739e2009-12-03 12:59:45 -0500933 return slice_used;
934}
935
936static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200937 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500938{
939 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal02b35082010-08-23 12:23:53 +0200940 unsigned int used_sl, charge;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500941 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
942 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500943
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500944 BUG_ON(nr_sync < 0);
Vivek Goyal02b35082010-08-23 12:23:53 +0200945 used_sl = charge = cfq_cfqq_slice_usage(cfqq);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500946
Vivek Goyal02b35082010-08-23 12:23:53 +0200947 if (iops_mode(cfqd))
948 charge = cfqq->slice_dispatch;
949 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
950 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500951
952 /* Can't update vdisktime while group is on service tree */
953 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyal02b35082010-08-23 12:23:53 +0200954 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500955 __cfq_group_service_tree_add(st, cfqg);
956
957 /* This group is being expired. Save the context */
958 if (time_after(cfqd->workload_expires, jiffies)) {
959 cfqg->saved_workload_slice = cfqd->workload_expires
960 - jiffies;
961 cfqg->saved_workload = cfqd->serving_type;
962 cfqg->saved_serving_prio = cfqd->serving_prio;
963 } else
964 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500965
966 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
967 st->min_vdisktime);
Vivek Goyalc4e78932010-08-23 12:25:03 +0200968 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
969 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
970 iops_mode(cfqd), cfqq->nr_sectors);
Vivek Goyale98ef892010-06-18 10:39:47 -0400971 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
972 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500973}
974
Vivek Goyal25fb5162009-12-03 12:59:46 -0500975#ifdef CONFIG_CFQ_GROUP_IOSCHED
976static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
977{
978 if (blkg)
979 return container_of(blkg, struct cfq_group, blkg);
980 return NULL;
981}
982
Vivek Goyalfe071432010-10-01 14:49:49 +0200983void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
984 unsigned int weight)
Vivek Goyalf8d461d2009-12-03 12:59:52 -0500985{
986 cfqg_of_blkg(blkg)->weight = weight;
987}
988
Vivek Goyal25fb5162009-12-03 12:59:46 -0500989static struct cfq_group *
990cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
991{
992 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
993 struct cfq_group *cfqg = NULL;
994 void *key = cfqd;
995 int i, j;
996 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -0500997 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
998 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500999
Vivek Goyal25fb5162009-12-03 12:59:46 -05001000 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
Ricky Beniteza74b2ad2010-04-05 18:22:17 +02001001 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1002 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1003 cfqg->blkg.dev = MKDEV(major, minor);
1004 goto done;
1005 }
Vivek Goyal25fb5162009-12-03 12:59:46 -05001006 if (cfqg || !create)
1007 goto done;
1008
1009 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1010 if (!cfqg)
1011 goto done;
1012
Vivek Goyal25fb5162009-12-03 12:59:46 -05001013 for_each_cfqg_st(cfqg, i, j, st)
1014 *st = CFQ_RB_ROOT;
1015 RB_CLEAR_NODE(&cfqg->rb_node);
1016
Vivek Goyalb1c35762009-12-03 12:59:47 -05001017 /*
1018 * Take the initial reference that will be released on destroy
1019 * This can be thought of a joint reference by cgroup and
1020 * elevator which will be dropped by either elevator exit
1021 * or cgroup deletion path depending on who is exiting first.
1022 */
Shaohua Li329a6782011-01-07 08:48:28 +01001023 cfqg->ref = 1;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001024
Vivek Goyal180be2a2010-09-14 08:47:11 +02001025 /*
1026 * Add group onto cgroup list. It might happen that bdi->dev is
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001027 * not initialized yet. Initialize this new group without major
Vivek Goyal180be2a2010-09-14 08:47:11 +02001028 * and minor info and this info will be filled in once a new thread
1029 * comes for IO. See code above.
1030 */
1031 if (bdi->dev) {
1032 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1033 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
Vivek Goyal22084192009-12-03 12:59:49 -05001034 MKDEV(major, minor));
Vivek Goyal180be2a2010-09-14 08:47:11 +02001035 } else
1036 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1037 0);
1038
Gui Jianfeng34d0f172010-04-13 16:05:49 +08001039 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001040
1041 /* Add group on cfqd list */
1042 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1043
1044done:
Vivek Goyal25fb5162009-12-03 12:59:46 -05001045 return cfqg;
1046}
1047
1048/*
1049 * Search for the cfq group current task belongs to. If create = 1, then also
1050 * create the cfq group if it does not exist. request_queue lock must be held.
1051 */
1052static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1053{
1054 struct cgroup *cgroup;
1055 struct cfq_group *cfqg = NULL;
1056
1057 rcu_read_lock();
1058 cgroup = task_cgroup(current, blkio_subsys_id);
1059 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1060 if (!cfqg && create)
1061 cfqg = &cfqd->root_group;
1062 rcu_read_unlock();
1063 return cfqg;
1064}
1065
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001066static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1067{
Shaohua Li329a6782011-01-07 08:48:28 +01001068 cfqg->ref++;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001069 return cfqg;
1070}
1071
Vivek Goyal25fb5162009-12-03 12:59:46 -05001072static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1073{
1074 /* Currently, all async queues are mapped to root group */
1075 if (!cfq_cfqq_sync(cfqq))
1076 cfqg = &cfqq->cfqd->root_group;
1077
1078 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001079 /* cfqq reference on cfqg */
Shaohua Li329a6782011-01-07 08:48:28 +01001080 cfqq->cfqg->ref++;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001081}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001082
1083static void cfq_put_cfqg(struct cfq_group *cfqg)
1084{
1085 struct cfq_rb_root *st;
1086 int i, j;
1087
Shaohua Li329a6782011-01-07 08:48:28 +01001088 BUG_ON(cfqg->ref <= 0);
1089 cfqg->ref--;
1090 if (cfqg->ref)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001091 return;
1092 for_each_cfqg_st(cfqg, i, j, st)
Gui Jianfengb54ce602010-11-30 20:52:46 +01001093 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
Vivek Goyalb1c35762009-12-03 12:59:47 -05001094 kfree(cfqg);
1095}
1096
1097static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1098{
1099 /* Something wrong if we are trying to remove same group twice */
1100 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1101
1102 hlist_del_init(&cfqg->cfqd_node);
1103
1104 /*
1105 * Put the reference taken at the time of creation so that when all
1106 * queues are gone, group can be destroyed.
1107 */
1108 cfq_put_cfqg(cfqg);
1109}
1110
1111static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1112{
1113 struct hlist_node *pos, *n;
1114 struct cfq_group *cfqg;
1115
1116 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1117 /*
1118 * If cgroup removal path got to blk_group first and removed
1119 * it from cgroup list, then it will take care of destroying
1120 * cfqg also.
1121 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001122 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001123 cfq_destroy_cfqg(cfqd, cfqg);
1124 }
1125}
1126
1127/*
1128 * Blk cgroup controller notification saying that blkio_group object is being
1129 * delinked as associated cgroup object is going away. That also means that
1130 * no new IO will come in this group. So get rid of this group as soon as
1131 * any pending IO in the group is finished.
1132 *
1133 * This function is called under rcu_read_lock(). key is the rcu protected
1134 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1135 * read lock.
1136 *
1137 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1138 * it should not be NULL as even if elevator was exiting, cgroup deltion
1139 * path got to it first.
1140 */
1141void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1142{
1143 unsigned long flags;
1144 struct cfq_data *cfqd = key;
1145
1146 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1147 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1148 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1149}
1150
Vivek Goyal25fb5162009-12-03 12:59:46 -05001151#else /* GROUP_IOSCHED */
1152static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1153{
1154 return &cfqd->root_group;
1155}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001156
1157static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1158{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001159 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001160}
1161
Vivek Goyal25fb5162009-12-03 12:59:46 -05001162static inline void
1163cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1164 cfqq->cfqg = cfqg;
1165}
1166
Vivek Goyalb1c35762009-12-03 12:59:47 -05001167static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1168static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1169
Vivek Goyal25fb5162009-12-03 12:59:46 -05001170#endif /* GROUP_IOSCHED */
1171
Jens Axboe498d3aa22007-04-26 12:54:48 +02001172/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001173 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001174 * requests waiting to be processed. It is sorted in the order that
1175 * we will service the queues.
1176 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001177static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001178 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001179{
Jens Axboe08717142008-01-28 11:38:15 +01001180 struct rb_node **p, *parent;
1181 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001182 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001183 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001184 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001185 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001186 int group_changed = 0;
1187
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001188 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001189 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001190 if (cfq_class_idle(cfqq)) {
1191 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001192 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001193 if (parent && parent != &cfqq->rb_node) {
1194 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1195 rb_key += __cfqq->rb_key;
1196 } else
1197 rb_key += jiffies;
1198 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001199 /*
1200 * Get our rb key offset. Subtract any residual slice
1201 * value carried from last service. A negative resid
1202 * count indicates slice overrun, and this should position
1203 * the next service time further away in the tree.
1204 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001205 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001206 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001207 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001208 } else {
1209 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001210 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001211 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1212 }
Jens Axboed9e76202007-04-20 14:27:50 +02001213
1214 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001215 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001216 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001217 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001218 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001219 if (rb_key == cfqq->rb_key &&
1220 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001221 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001222
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001223 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1224 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001225 }
Jens Axboed9e76202007-04-20 14:27:50 +02001226
Jens Axboe498d3aa22007-04-26 12:54:48 +02001227 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001228 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001229 cfqq->service_tree = service_tree;
1230 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001231 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001232 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001233
Jens Axboed9e76202007-04-20 14:27:50 +02001234 parent = *p;
1235 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1236
Jens Axboe0c534e02007-04-18 20:01:57 +02001237 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001238 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001239 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001240 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001241 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001242 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001243 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001244 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001245 }
Jens Axboe67060e32007-04-18 20:13:32 +02001246
1247 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001248 }
1249
Jens Axboecc09e292007-04-26 12:53:50 +02001250 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001251 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001252
Jens Axboed9e76202007-04-20 14:27:50 +02001253 cfqq->rb_key = rb_key;
1254 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001255 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1256 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001257 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001258 return;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001259 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260}
1261
Jens Axboea36e71f2009-04-15 12:15:11 +02001262static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001263cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1264 sector_t sector, struct rb_node **ret_parent,
1265 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001266{
Jens Axboea36e71f2009-04-15 12:15:11 +02001267 struct rb_node **p, *parent;
1268 struct cfq_queue *cfqq = NULL;
1269
1270 parent = NULL;
1271 p = &root->rb_node;
1272 while (*p) {
1273 struct rb_node **n;
1274
1275 parent = *p;
1276 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1277
1278 /*
1279 * Sort strictly based on sector. Smallest to the left,
1280 * largest to the right.
1281 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001282 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001283 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001284 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001285 n = &(*p)->rb_left;
1286 else
1287 break;
1288 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001289 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001290 }
1291
1292 *ret_parent = parent;
1293 if (rb_link)
1294 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001295 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001296}
1297
1298static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1299{
Jens Axboea36e71f2009-04-15 12:15:11 +02001300 struct rb_node **p, *parent;
1301 struct cfq_queue *__cfqq;
1302
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001303 if (cfqq->p_root) {
1304 rb_erase(&cfqq->p_node, cfqq->p_root);
1305 cfqq->p_root = NULL;
1306 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001307
1308 if (cfq_class_idle(cfqq))
1309 return;
1310 if (!cfqq->next_rq)
1311 return;
1312
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001313 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001314 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1315 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001316 if (!__cfqq) {
1317 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001318 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1319 } else
1320 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001321}
1322
Jens Axboe498d3aa22007-04-26 12:54:48 +02001323/*
1324 * Update cfqq's position in the service tree.
1325 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001326static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001327{
Jens Axboe6d048f52007-04-25 12:44:27 +02001328 /*
1329 * Resorting requires the cfqq to be on the RR list already.
1330 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001331 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001332 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001333 cfq_prio_tree_add(cfqd, cfqq);
1334 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001335}
1336
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337/*
1338 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001339 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 */
Jens Axboefebffd62008-01-28 13:19:43 +01001341static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342{
Jens Axboe7b679132008-05-30 12:23:07 +02001343 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001344 BUG_ON(cfq_cfqq_on_rr(cfqq));
1345 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 cfqd->busy_queues++;
1347
Jens Axboeedd75ff2007-04-19 12:03:34 +02001348 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349}
1350
Jens Axboe498d3aa22007-04-26 12:54:48 +02001351/*
1352 * Called when the cfqq no longer has requests pending, remove it from
1353 * the service tree.
1354 */
Jens Axboefebffd62008-01-28 13:19:43 +01001355static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356{
Jens Axboe7b679132008-05-30 12:23:07 +02001357 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001358 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1359 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001361 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1362 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1363 cfqq->service_tree = NULL;
1364 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001365 if (cfqq->p_root) {
1366 rb_erase(&cfqq->p_node, cfqq->p_root);
1367 cfqq->p_root = NULL;
1368 }
Jens Axboed9e76202007-04-20 14:27:50 +02001369
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001370 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 BUG_ON(!cfqd->busy_queues);
1372 cfqd->busy_queues--;
1373}
1374
1375/*
1376 * rb tree support functions
1377 */
Jens Axboefebffd62008-01-28 13:19:43 +01001378static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379{
Jens Axboe5e705372006-07-13 12:39:25 +02001380 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001381 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382
Jens Axboeb4878f22005-10-20 16:42:29 +02001383 BUG_ON(!cfqq->queued[sync]);
1384 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385
Jens Axboe5e705372006-07-13 12:39:25 +02001386 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387
Vivek Goyalf04a6422009-12-03 12:59:40 -05001388 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1389 /*
1390 * Queue will be deleted from service tree when we actually
1391 * expire it later. Right now just remove it from prio tree
1392 * as it is empty.
1393 */
1394 if (cfqq->p_root) {
1395 rb_erase(&cfqq->p_node, cfqq->p_root);
1396 cfqq->p_root = NULL;
1397 }
1398 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399}
1400
Jens Axboe5e705372006-07-13 12:39:25 +02001401static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402{
Jens Axboe5e705372006-07-13 12:39:25 +02001403 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001405 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406
Jens Axboe5380a102006-07-13 12:37:56 +02001407 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408
1409 /*
1410 * looks a little odd, but the first insert might return an alias.
1411 * if that happens, put the alias on the dispatch list
1412 */
Jens Axboe21183b02006-07-13 12:33:14 +02001413 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001414 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001415
1416 if (!cfq_cfqq_on_rr(cfqq))
1417 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001418
1419 /*
1420 * check if this request is a better next-serve candidate
1421 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001422 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001423 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001424
1425 /*
1426 * adjust priority tree position, if ->next_rq changes
1427 */
1428 if (prev != cfqq->next_rq)
1429 cfq_prio_tree_add(cfqd, cfqq);
1430
Jens Axboe5044eed2007-04-25 11:53:48 +02001431 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432}
1433
Jens Axboefebffd62008-01-28 13:19:43 +01001434static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435{
Jens Axboe5380a102006-07-13 12:37:56 +02001436 elv_rb_del(&cfqq->sort_list, rq);
1437 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001438 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1439 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001440 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001441 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001442 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1443 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444}
1445
Jens Axboe206dc692006-03-28 13:03:44 +02001446static struct request *
1447cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448{
Jens Axboe206dc692006-03-28 13:03:44 +02001449 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001450 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001451 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452
Jens Axboe4ac845a2008-01-24 08:44:49 +01001453 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001454 if (!cic)
1455 return NULL;
1456
1457 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001458 if (cfqq) {
1459 sector_t sector = bio->bi_sector + bio_sectors(bio);
1460
Jens Axboe21183b02006-07-13 12:33:14 +02001461 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001462 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 return NULL;
1465}
1466
Jens Axboe165125e2007-07-24 09:28:11 +02001467static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001468{
1469 struct cfq_data *cfqd = q->elevator->elevator_data;
1470
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001471 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001472 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001473 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001474
Tejun Heo5b936292009-05-07 22:24:38 +09001475 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001476}
1477
Jens Axboe165125e2007-07-24 09:28:11 +02001478static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479{
Jens Axboe22e2c502005-06-27 10:55:12 +02001480 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001482 WARN_ON(!cfqd->rq_in_driver);
1483 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001484 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001485 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486}
1487
Jens Axboeb4878f22005-10-20 16:42:29 +02001488static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489{
Jens Axboe5e705372006-07-13 12:39:25 +02001490 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001491
Jens Axboe5e705372006-07-13 12:39:25 +02001492 if (cfqq->next_rq == rq)
1493 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494
Jens Axboeb4878f22005-10-20 16:42:29 +02001495 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001496 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001497
Aaron Carroll45333d52008-08-26 15:52:36 +02001498 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001499 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1500 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02001501 if (rq->cmd_flags & REQ_META) {
Jens Axboe374f84a2006-07-23 01:42:19 +02001502 WARN_ON(!cfqq->meta_pending);
1503 cfqq->meta_pending--;
1504 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505}
1506
Jens Axboe165125e2007-07-24 09:28:11 +02001507static int cfq_merge(struct request_queue *q, struct request **req,
1508 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509{
1510 struct cfq_data *cfqd = q->elevator->elevator_data;
1511 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512
Jens Axboe206dc692006-03-28 13:03:44 +02001513 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001514 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001515 *req = __rq;
1516 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 }
1518
1519 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520}
1521
Jens Axboe165125e2007-07-24 09:28:11 +02001522static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001523 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524{
Jens Axboe21183b02006-07-13 12:33:14 +02001525 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001526 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527
Jens Axboe5e705372006-07-13 12:39:25 +02001528 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530}
1531
Divyesh Shah812d4022010-04-08 21:14:23 -07001532static void cfq_bio_merged(struct request_queue *q, struct request *req,
1533 struct bio *bio)
1534{
Vivek Goyale98ef892010-06-18 10:39:47 -04001535 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1536 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001537}
1538
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539static void
Jens Axboe165125e2007-07-24 09:28:11 +02001540cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 struct request *next)
1542{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001543 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001544 /*
1545 * reposition in fifo if next is older than rq
1546 */
1547 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001548 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001549 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001550 rq_set_fifo_time(rq, rq_fifo_time(next));
1551 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001552
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001553 if (cfqq->next_rq == next)
1554 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001555 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001556 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1557 rq_data_dir(next), rq_is_sync(next));
Jens Axboe22e2c502005-06-27 10:55:12 +02001558}
1559
Jens Axboe165125e2007-07-24 09:28:11 +02001560static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001561 struct bio *bio)
1562{
1563 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001564 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001565 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001566
1567 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001568 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001569 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001570 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001571 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001572
1573 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001574 * Lookup the cfqq that this bio will be queued with. Allow
1575 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001576 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001577 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001578 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001579 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001580
Vasily Tarasov91fac312007-04-25 12:29:51 +02001581 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001582 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001583}
1584
Divyesh Shah812df482010-04-08 21:15:35 -07001585static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1586{
1587 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001588 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001589}
1590
Jens Axboefebffd62008-01-28 13:19:43 +01001591static void __cfq_set_active_queue(struct cfq_data *cfqd,
1592 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001593{
1594 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001595 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1596 cfqd->serving_prio, cfqd->serving_type);
Vivek Goyale98ef892010-06-18 10:39:47 -04001597 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001598 cfqq->slice_start = 0;
1599 cfqq->dispatch_start = jiffies;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001600 cfqq->allocated_slice = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001601 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001602 cfqq->slice_dispatch = 0;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001603 cfqq->nr_sectors = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001604
Jens Axboe2f5cb732009-04-07 08:51:19 +02001605 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +02001606 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001607 cfq_clear_cfqq_must_alloc_slice(cfqq);
1608 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +11001609 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001610
Divyesh Shah812df482010-04-08 21:15:35 -07001611 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001612 }
1613
1614 cfqd->active_queue = cfqq;
1615}
1616
1617/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001618 * current cfqq expired its slice (or was too idle), select new one
1619 */
1620static void
1621__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001622 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001623{
Jens Axboe7b679132008-05-30 12:23:07 +02001624 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1625
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001626 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001627 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001628
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001629 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001630 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001631
1632 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001633 * If this cfqq is shared between multiple processes, check to
1634 * make sure that those processes are still issuing I/Os within
1635 * the mean seek distance. If not, it may be time to break the
1636 * queues apart again.
1637 */
1638 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1639 cfq_mark_cfqq_split_coop(cfqq);
1640
1641 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001642 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001643 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01001644 if (timed_out) {
1645 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07001646 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001647 else
1648 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001649 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1650 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001651
Vivek Goyale5ff0822010-04-26 19:25:11 +02001652 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001653
Vivek Goyalf04a6422009-12-03 12:59:40 -05001654 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1655 cfq_del_cfqq_rr(cfqd, cfqq);
1656
Jens Axboeedd75ff2007-04-19 12:03:34 +02001657 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001658
1659 if (cfqq == cfqd->active_queue)
1660 cfqd->active_queue = NULL;
1661
1662 if (cfqd->active_cic) {
1663 put_io_context(cfqd->active_cic->ioc);
1664 cfqd->active_cic = NULL;
1665 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001666}
1667
Vivek Goyale5ff0822010-04-26 19:25:11 +02001668static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001669{
1670 struct cfq_queue *cfqq = cfqd->active_queue;
1671
1672 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001673 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001674}
1675
Jens Axboe498d3aa22007-04-26 12:54:48 +02001676/*
1677 * Get next queue for service. Unless we have a queue preemption,
1678 * we'll simply select the first cfqq in the service tree.
1679 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001680static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001681{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001682 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001683 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001684 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001685
Vivek Goyalf04a6422009-12-03 12:59:40 -05001686 if (!cfqd->rq_queued)
1687 return NULL;
1688
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001689 /* There is nothing to dispatch */
1690 if (!service_tree)
1691 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001692 if (RB_EMPTY_ROOT(&service_tree->rb))
1693 return NULL;
1694 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001695}
1696
Vivek Goyalf04a6422009-12-03 12:59:40 -05001697static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1698{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001699 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001700 struct cfq_queue *cfqq;
1701 int i, j;
1702 struct cfq_rb_root *st;
1703
1704 if (!cfqd->rq_queued)
1705 return NULL;
1706
Vivek Goyal25fb5162009-12-03 12:59:46 -05001707 cfqg = cfq_get_next_cfqg(cfqd);
1708 if (!cfqg)
1709 return NULL;
1710
Vivek Goyalf04a6422009-12-03 12:59:40 -05001711 for_each_cfqg_st(cfqg, i, j, st)
1712 if ((cfqq = cfq_rb_first(st)) != NULL)
1713 return cfqq;
1714 return NULL;
1715}
1716
Jens Axboe498d3aa22007-04-26 12:54:48 +02001717/*
1718 * Get and set a new active queue for service.
1719 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001720static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1721 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001722{
Jens Axboee00ef792009-11-04 08:54:55 +01001723 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001724 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001725
Jens Axboe22e2c502005-06-27 10:55:12 +02001726 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001727 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001728}
1729
Jens Axboed9e76202007-04-20 14:27:50 +02001730static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1731 struct request *rq)
1732{
Tejun Heo83096eb2009-05-07 22:24:39 +09001733 if (blk_rq_pos(rq) >= cfqd->last_position)
1734 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001735 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001736 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001737}
1738
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001739static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001740 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001741{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001742 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001743}
1744
Jens Axboea36e71f2009-04-15 12:15:11 +02001745static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1746 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001747{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001748 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001749 struct rb_node *parent, *node;
1750 struct cfq_queue *__cfqq;
1751 sector_t sector = cfqd->last_position;
1752
1753 if (RB_EMPTY_ROOT(root))
1754 return NULL;
1755
1756 /*
1757 * First, if we find a request starting at the end of the last
1758 * request, choose it.
1759 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001760 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001761 if (__cfqq)
1762 return __cfqq;
1763
1764 /*
1765 * If the exact sector wasn't found, the parent of the NULL leaf
1766 * will contain the closest sector.
1767 */
1768 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001769 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001770 return __cfqq;
1771
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001772 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001773 node = rb_next(&__cfqq->p_node);
1774 else
1775 node = rb_prev(&__cfqq->p_node);
1776 if (!node)
1777 return NULL;
1778
1779 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001780 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001781 return __cfqq;
1782
1783 return NULL;
1784}
1785
1786/*
1787 * cfqd - obvious
1788 * cur_cfqq - passed in so that we don't decide that the current queue is
1789 * closely cooperating with itself.
1790 *
1791 * So, basically we're assuming that that cur_cfqq has dispatched at least
1792 * one request, and that cfqd->last_position reflects a position on the disk
1793 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1794 * assumption.
1795 */
1796static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001797 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001798{
1799 struct cfq_queue *cfqq;
1800
Divyesh Shah39c01b22010-03-25 15:45:57 +01001801 if (cfq_class_idle(cur_cfqq))
1802 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001803 if (!cfq_cfqq_sync(cur_cfqq))
1804 return NULL;
1805 if (CFQQ_SEEKY(cur_cfqq))
1806 return NULL;
1807
Jens Axboea36e71f2009-04-15 12:15:11 +02001808 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001809 * Don't search priority tree if it's the only queue in the group.
1810 */
1811 if (cur_cfqq->cfqg->nr_cfqq == 1)
1812 return NULL;
1813
1814 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001815 * We should notice if some of the queues are cooperating, eg
1816 * working closely on the same area of the disk. In that case,
1817 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001818 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001819 cfqq = cfqq_close(cfqd, cur_cfqq);
1820 if (!cfqq)
1821 return NULL;
1822
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001823 /* If new queue belongs to different cfq_group, don't choose it */
1824 if (cur_cfqq->cfqg != cfqq->cfqg)
1825 return NULL;
1826
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001827 /*
1828 * It only makes sense to merge sync queues.
1829 */
1830 if (!cfq_cfqq_sync(cfqq))
1831 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001832 if (CFQQ_SEEKY(cfqq))
1833 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001834
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001835 /*
1836 * Do not merge queues of different priority classes
1837 */
1838 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1839 return NULL;
1840
Jens Axboea36e71f2009-04-15 12:15:11 +02001841 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001842}
1843
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001844/*
1845 * Determine whether we should enforce idle window for this queue.
1846 */
1847
1848static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1849{
1850 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001851 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001852
Vivek Goyalf04a6422009-12-03 12:59:40 -05001853 BUG_ON(!service_tree);
1854 BUG_ON(!service_tree->count);
1855
Vivek Goyalb6508c12010-08-23 12:23:33 +02001856 if (!cfqd->cfq_slice_idle)
1857 return false;
1858
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001859 /* We never do for idle class queues. */
1860 if (prio == IDLE_WORKLOAD)
1861 return false;
1862
1863 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001864 if (cfq_cfqq_idle_window(cfqq) &&
1865 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001866 return true;
1867
1868 /*
1869 * Otherwise, we do only if they are the last ones
1870 * in their service tree.
1871 */
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001872 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01001873 return true;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001874 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1875 service_tree->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01001876 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001877}
1878
Jens Axboe6d048f52007-04-25 12:44:27 +02001879static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001880{
Jens Axboe17926692007-01-19 11:59:30 +11001881 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001882 struct cfq_io_context *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001883 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001884
Jens Axboea68bbdd2008-09-24 13:03:33 +02001885 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001886 * SSD device without seek penalty, disable idling. But only do so
1887 * for devices that support queuing, otherwise we still have a problem
1888 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001889 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001890 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001891 return;
1892
Jens Axboedd67d052006-06-21 09:36:18 +02001893 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001894 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001895
1896 /*
1897 * idle is disabled, either manually or by past process history
1898 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001899 if (!cfq_should_idle(cfqd, cfqq)) {
1900 /* no queue idling. Check for group idling */
1901 if (cfqd->cfq_group_idle)
1902 group_idle = cfqd->cfq_group_idle;
1903 else
1904 return;
1905 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001906
Jens Axboe22e2c502005-06-27 10:55:12 +02001907 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001908 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001909 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001910 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001911 return;
1912
1913 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001914 * task has exited, don't wait
1915 */
Jens Axboe206dc692006-03-28 13:03:44 +02001916 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001917 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001918 return;
1919
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001920 /*
1921 * If our average think time is larger than the remaining time
1922 * slice, then don't idle. This avoids overrunning the allotted
1923 * time slice.
1924 */
1925 if (sample_valid(cic->ttime_samples) &&
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001926 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1927 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1928 cic->ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001929 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001930 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001931
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001932 /* There are other queues in the group, don't do group idle */
1933 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1934 return;
1935
Jens Axboe3b181522005-06-27 10:56:24 +02001936 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001937
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001938 if (group_idle)
1939 sl = cfqd->cfq_group_idle;
1940 else
1941 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001942
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001943 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001944 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001945 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1946 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947}
1948
Jens Axboe498d3aa22007-04-26 12:54:48 +02001949/*
1950 * Move request from internal lists to the request queue dispatch list.
1951 */
Jens Axboe165125e2007-07-24 09:28:11 +02001952static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001954 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001955 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001956
Jens Axboe7b679132008-05-30 12:23:07 +02001957 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1958
Jeff Moyer06d21882009-09-11 17:08:59 +02001959 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001960 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001961 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001962 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001963 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001964
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001965 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02001966 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001967 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02001968 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969}
1970
1971/*
1972 * return expired entry, or NULL to just start from scratch in rbtree
1973 */
Jens Axboefebffd62008-01-28 13:19:43 +01001974static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975{
Jens Axboe30996f42009-10-05 11:03:39 +02001976 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977
Jens Axboe3b181522005-06-27 10:56:24 +02001978 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001980
1981 cfq_mark_cfqq_fifo_expire(cfqq);
1982
Jens Axboe89850f72006-07-22 16:48:31 +02001983 if (list_empty(&cfqq->fifo))
1984 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985
Jens Axboe89850f72006-07-22 16:48:31 +02001986 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02001987 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02001988 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989
Jens Axboe30996f42009-10-05 11:03:39 +02001990 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001991 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992}
1993
Jens Axboe22e2c502005-06-27 10:55:12 +02001994static inline int
1995cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1996{
1997 const int base_rq = cfqd->cfq_slice_async_rq;
1998
1999 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2000
2001 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2002}
2003
2004/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002005 * Must be called with the queue_lock held.
2006 */
2007static int cfqq_process_refs(struct cfq_queue *cfqq)
2008{
2009 int process_refs, io_refs;
2010
2011 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002012 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002013 BUG_ON(process_refs < 0);
2014 return process_refs;
2015}
2016
2017static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2018{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002019 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002020 struct cfq_queue *__cfqq;
2021
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002022 /*
2023 * If there are no process references on the new_cfqq, then it is
2024 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2025 * chain may have dropped their last reference (not just their
2026 * last process reference).
2027 */
2028 if (!cfqq_process_refs(new_cfqq))
2029 return;
2030
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002031 /* Avoid a circular list and skip interim queue merges */
2032 while ((__cfqq = new_cfqq->new_cfqq)) {
2033 if (__cfqq == cfqq)
2034 return;
2035 new_cfqq = __cfqq;
2036 }
2037
2038 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002039 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002040 /*
2041 * If the process for the cfqq has gone away, there is no
2042 * sense in merging the queues.
2043 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002044 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002045 return;
2046
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002047 /*
2048 * Merge in the direction of the lesser amount of work.
2049 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002050 if (new_process_refs >= process_refs) {
2051 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002052 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002053 } else {
2054 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002055 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002056 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002057}
2058
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002059static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002060 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002061{
2062 struct cfq_queue *queue;
2063 int i;
2064 bool key_valid = false;
2065 unsigned long lowest_key = 0;
2066 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2067
Vivek Goyal65b32a52009-12-16 17:52:59 -05002068 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2069 /* select the one with lowest rb_key */
2070 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002071 if (queue &&
2072 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2073 lowest_key = queue->rb_key;
2074 cur_best = i;
2075 key_valid = true;
2076 }
2077 }
2078
2079 return cur_best;
2080}
2081
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002082static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002083{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002084 unsigned slice;
2085 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002086 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002087 unsigned group_slice;
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002088 enum wl_prio_t original_prio = cfqd->serving_prio;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002089
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002090 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002091 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002092 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002093 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002094 cfqd->serving_prio = BE_WORKLOAD;
2095 else {
2096 cfqd->serving_prio = IDLE_WORKLOAD;
2097 cfqd->workload_expires = jiffies + 1;
2098 return;
2099 }
2100
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002101 if (original_prio != cfqd->serving_prio)
2102 goto new_workload;
2103
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002104 /*
2105 * For RT and BE, we have to choose also the type
2106 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2107 * expiration time
2108 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002109 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002110 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002111
2112 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002113 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002114 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002115 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002116 return;
2117
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002118new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002119 /* otherwise select new workload type */
2120 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002121 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2122 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002123 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002124
2125 /*
2126 * the workload slice is computed as a fraction of target latency
2127 * proportional to the number of queues in that workload, over
2128 * all the queues in the same priority class
2129 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002130 group_slice = cfq_group_slice(cfqd, cfqg);
2131
2132 slice = group_slice * count /
2133 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2134 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002135
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002136 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2137 unsigned int tmp;
2138
2139 /*
2140 * Async queues are currently system wide. Just taking
2141 * proportion of queues with-in same group will lead to higher
2142 * async ratio system wide as generally root group is going
2143 * to have higher weight. A more accurate thing would be to
2144 * calculate system wide asnc/sync ratio.
2145 */
2146 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2147 tmp = tmp/cfqd->busy_queues;
2148 slice = min_t(unsigned, slice, tmp);
2149
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002150 /* async workload slice is scaled down according to
2151 * the sync/async slice ratio. */
2152 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002153 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002154 /* sync workload slice is at least 2 * cfq_slice_idle */
2155 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2156
2157 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002158 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002159 cfqd->workload_expires = jiffies + slice;
2160}
2161
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002162static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2163{
2164 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002165 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002166
2167 if (RB_EMPTY_ROOT(&st->rb))
2168 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002169 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002170 update_min_vdisktime(st);
2171 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002172}
2173
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002174static void cfq_choose_cfqg(struct cfq_data *cfqd)
2175{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002176 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2177
2178 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002179
2180 /* Restore the workload type data */
2181 if (cfqg->saved_workload_slice) {
2182 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2183 cfqd->serving_type = cfqg->saved_workload;
2184 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002185 } else
2186 cfqd->workload_expires = jiffies - 1;
2187
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002188 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002189}
2190
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002191/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002192 * Select a queue for service. If we have a current active queue,
2193 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002194 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002195static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002196{
Jens Axboea36e71f2009-04-15 12:15:11 +02002197 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002198
2199 cfqq = cfqd->active_queue;
2200 if (!cfqq)
2201 goto new_queue;
2202
Vivek Goyalf04a6422009-12-03 12:59:40 -05002203 if (!cfqd->rq_queued)
2204 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002205
2206 /*
2207 * We were waiting for group to get backlogged. Expire the queue
2208 */
2209 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2210 goto expire;
2211
Jens Axboe22e2c502005-06-27 10:55:12 +02002212 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002213 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002214 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002215 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2216 /*
2217 * If slice had not expired at the completion of last request
2218 * we might not have turned on wait_busy flag. Don't expire
2219 * the queue yet. Allow the group to get backlogged.
2220 *
2221 * The very fact that we have used the slice, that means we
2222 * have been idling all along on this queue and it should be
2223 * ok to wait for this request to complete.
2224 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002225 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2226 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2227 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002228 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002229 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002230 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002231 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002232
2233 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002234 * The active queue has requests and isn't expired, allow it to
2235 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002236 */
Jens Axboedd67d052006-06-21 09:36:18 +02002237 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002238 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002239
2240 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002241 * If another queue has a request waiting within our mean seek
2242 * distance, let it run. The expire code will check for close
2243 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002244 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002245 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002246 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002247 if (new_cfqq) {
2248 if (!cfqq->new_cfqq)
2249 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002250 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002251 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002252
2253 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002254 * No requests pending. If the active queue still has requests in
2255 * flight or is idling for a new request, allow either of these
2256 * conditions to happen (or time out) before selecting a new queue.
2257 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002258 if (timer_pending(&cfqd->idle_slice_timer)) {
2259 cfqq = NULL;
2260 goto keep_queue;
2261 }
2262
Shaohua Li8e1ac662010-11-08 15:01:04 +01002263 /*
2264 * This is a deep seek queue, but the device is much faster than
2265 * the queue can deliver, don't idle
2266 **/
2267 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2268 (cfq_cfqq_slice_new(cfqq) ||
2269 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2270 cfq_clear_cfqq_deep(cfqq);
2271 cfq_clear_cfqq_idle_window(cfqq);
2272 }
2273
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002274 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2275 cfqq = NULL;
2276 goto keep_queue;
2277 }
2278
2279 /*
2280 * If group idle is enabled and there are requests dispatched from
2281 * this group, wait for requests to complete.
2282 */
2283check_group_idle:
2284 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2285 && cfqq->cfqg->dispatched) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002286 cfqq = NULL;
2287 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002288 }
2289
Jens Axboe3b181522005-06-27 10:56:24 +02002290expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002291 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002292new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002293 /*
2294 * Current queue expired. Check if we have to switch to a new
2295 * service tree
2296 */
2297 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002298 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002299
Jens Axboea36e71f2009-04-15 12:15:11 +02002300 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002301keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002302 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002303}
2304
Jens Axboefebffd62008-01-28 13:19:43 +01002305static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002306{
2307 int dispatched = 0;
2308
2309 while (cfqq->next_rq) {
2310 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2311 dispatched++;
2312 }
2313
2314 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002315
2316 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002317 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002318 return dispatched;
2319}
2320
Jens Axboe498d3aa22007-04-26 12:54:48 +02002321/*
2322 * Drain our current requests. Used for barriers and when switching
2323 * io schedulers on-the-fly.
2324 */
Jens Axboed9e76202007-04-20 14:27:50 +02002325static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002326{
Jens Axboe08717142008-01-28 11:38:15 +01002327 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002328 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002329
Divyesh Shah3440c492010-04-09 09:29:57 +02002330 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002331 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002332 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2333 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002334 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002335 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002336
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002337 BUG_ON(cfqd->busy_queues);
2338
Jeff Moyer69237152009-06-12 15:29:30 +02002339 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002340 return dispatched;
2341}
2342
Shaohua Liabc3c742010-03-01 09:20:54 +01002343static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2344 struct cfq_queue *cfqq)
2345{
2346 /* the queue hasn't finished any request, can't estimate */
2347 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01002348 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002349 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2350 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01002351 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002352
Shaohua Lic1e44752010-11-08 15:01:02 +01002353 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01002354}
2355
Jens Axboe0b182d62009-10-06 20:49:37 +02002356static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002357{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002358 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359
Jens Axboe2f5cb732009-04-07 08:51:19 +02002360 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002361 * Drain async requests before we start sync IO
2362 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002363 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002364 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002365
2366 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002367 * If this is an async queue and we have sync IO in flight, let it wait
2368 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002369 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002370 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002371
Shaohua Liabc3c742010-03-01 09:20:54 +01002372 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002373 if (cfq_class_idle(cfqq))
2374 max_dispatch = 1;
2375
2376 /*
2377 * Does this cfqq already have too much IO in flight?
2378 */
2379 if (cfqq->dispatched >= max_dispatch) {
2380 /*
2381 * idle queue must always only have a single IO in flight
2382 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002383 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002384 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002385
Jens Axboe2f5cb732009-04-07 08:51:19 +02002386 /*
2387 * We have other queues, don't allow more IO from this one
2388 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002389 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002390 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002391
Jens Axboe2f5cb732009-04-07 08:51:19 +02002392 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002393 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002394 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002395 if (cfqd->busy_queues == 1)
2396 max_dispatch = -1;
2397 else
2398 /*
2399 * Normally we start throttling cfqq when cfq_quantum/2
2400 * requests have been dispatched. But we can drive
2401 * deeper queue depths at the beginning of slice
2402 * subjected to upper limit of cfq_quantum.
2403 * */
2404 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002405 }
2406
2407 /*
2408 * Async queues must wait a bit before being allowed dispatch.
2409 * We also ramp up the dispatch depth gradually for async IO,
2410 * based on the last sync IO we serviced
2411 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002412 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002413 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002414 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002415
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002416 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002417 if (!depth && !cfqq->dispatched)
2418 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002419 if (depth < max_dispatch)
2420 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421 }
2422
Jens Axboe0b182d62009-10-06 20:49:37 +02002423 /*
2424 * If we're below the current max, allow a dispatch
2425 */
2426 return cfqq->dispatched < max_dispatch;
2427}
2428
2429/*
2430 * Dispatch a request from cfqq, moving them to the request queue
2431 * dispatch list.
2432 */
2433static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2434{
2435 struct request *rq;
2436
2437 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2438
2439 if (!cfq_may_dispatch(cfqd, cfqq))
2440 return false;
2441
2442 /*
2443 * follow expired path, else get first next available
2444 */
2445 rq = cfq_check_fifo(cfqq);
2446 if (!rq)
2447 rq = cfqq->next_rq;
2448
2449 /*
2450 * insert request into driver dispatch list
2451 */
2452 cfq_dispatch_insert(cfqd->queue, rq);
2453
2454 if (!cfqd->active_cic) {
2455 struct cfq_io_context *cic = RQ_CIC(rq);
2456
2457 atomic_long_inc(&cic->ioc->refcount);
2458 cfqd->active_cic = cic;
2459 }
2460
2461 return true;
2462}
2463
2464/*
2465 * Find the cfqq that we need to service and move a request from that to the
2466 * dispatch list
2467 */
2468static int cfq_dispatch_requests(struct request_queue *q, int force)
2469{
2470 struct cfq_data *cfqd = q->elevator->elevator_data;
2471 struct cfq_queue *cfqq;
2472
2473 if (!cfqd->busy_queues)
2474 return 0;
2475
2476 if (unlikely(force))
2477 return cfq_forced_dispatch(cfqd);
2478
2479 cfqq = cfq_select_queue(cfqd);
2480 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002481 return 0;
2482
Jens Axboe2f5cb732009-04-07 08:51:19 +02002483 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002484 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002485 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002486 if (!cfq_dispatch_request(cfqd, cfqq))
2487 return 0;
2488
Jens Axboe2f5cb732009-04-07 08:51:19 +02002489 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002490 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002491
2492 /*
2493 * expire an async queue immediately if it has used up its slice. idle
2494 * queue always expire after 1 dispatch round.
2495 */
2496 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2497 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2498 cfq_class_idle(cfqq))) {
2499 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002500 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002501 }
2502
Shan Weib217a902009-09-01 10:06:42 +02002503 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002504 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505}
2506
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507/*
Jens Axboe5e705372006-07-13 12:39:25 +02002508 * task holds one reference to the queue, dropped when task exits. each rq
2509 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002511 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 * queue lock must be held here.
2513 */
2514static void cfq_put_queue(struct cfq_queue *cfqq)
2515{
Jens Axboe22e2c502005-06-27 10:55:12 +02002516 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05002517 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002518
Shaohua Li30d7b942011-01-07 08:46:59 +01002519 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520
Shaohua Li30d7b942011-01-07 08:46:59 +01002521 cfqq->ref--;
2522 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 return;
2524
Jens Axboe7b679132008-05-30 12:23:07 +02002525 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002527 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002528 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002530 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002531 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002532 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002533 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002534
Vivek Goyalf04a6422009-12-03 12:59:40 -05002535 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002537 cfq_put_cfqg(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538}
2539
Jens Axboed6de8be2008-05-28 14:46:59 +02002540/*
2541 * Must always be called with the rcu_read_lock() held
2542 */
Jens Axboe07416d22008-05-07 09:17:12 +02002543static void
2544__call_for_each_cic(struct io_context *ioc,
2545 void (*func)(struct io_context *, struct cfq_io_context *))
2546{
2547 struct cfq_io_context *cic;
2548 struct hlist_node *n;
2549
2550 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2551 func(ioc, cic);
2552}
2553
Jens Axboe4ac845a2008-01-24 08:44:49 +01002554/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002555 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002556 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002557static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002558call_for_each_cic(struct io_context *ioc,
2559 void (*func)(struct io_context *, struct cfq_io_context *))
2560{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002561 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002562 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002563 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002564}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002565
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002566static void cfq_cic_free_rcu(struct rcu_head *head)
2567{
2568 struct cfq_io_context *cic;
2569
2570 cic = container_of(head, struct cfq_io_context, rcu_head);
2571
2572 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002573 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002574
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002575 if (ioc_gone) {
2576 /*
2577 * CFQ scheduler is exiting, grab exit lock and check
2578 * the pending io context count. If it hits zero,
2579 * complete ioc_gone and set it back to NULL
2580 */
2581 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002582 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002583 complete(ioc_gone);
2584 ioc_gone = NULL;
2585 }
2586 spin_unlock(&ioc_gone_lock);
2587 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002588}
2589
2590static void cfq_cic_free(struct cfq_io_context *cic)
2591{
2592 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002593}
2594
2595static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2596{
2597 unsigned long flags;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002598 unsigned long dead_key = (unsigned long) cic->key;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002599
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002600 BUG_ON(!(dead_key & CIC_DEAD_KEY));
Jens Axboe4ac845a2008-01-24 08:44:49 +01002601
2602 spin_lock_irqsave(&ioc->lock, flags);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002603 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
Jens Axboeffc4e752008-02-19 10:02:29 +01002604 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002605 spin_unlock_irqrestore(&ioc->lock, flags);
2606
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002607 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002608}
2609
Jens Axboed6de8be2008-05-28 14:46:59 +02002610/*
2611 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2612 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2613 * and ->trim() which is called with the task lock held
2614 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002615static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002617 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002618 * ioc->refcount is zero here, or we are called from elv_unregister(),
2619 * so no more cic's are allowed to be linked into this ioc. So it
2620 * should be ok to iterate over the known list, we will see all cic's
2621 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002622 */
Jens Axboe07416d22008-05-07 09:17:12 +02002623 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624}
2625
Shaohua Lid02a2c02010-05-25 10:16:53 +02002626static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002627{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002628 struct cfq_queue *__cfqq, *next;
2629
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002630 /*
2631 * If this queue was scheduled to merge with another queue, be
2632 * sure to drop the reference taken on that queue (and others in
2633 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2634 */
2635 __cfqq = cfqq->new_cfqq;
2636 while (__cfqq) {
2637 if (__cfqq == cfqq) {
2638 WARN(1, "cfqq->new_cfqq loop detected\n");
2639 break;
2640 }
2641 next = __cfqq->new_cfqq;
2642 cfq_put_queue(__cfqq);
2643 __cfqq = next;
2644 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002645}
2646
2647static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2648{
2649 if (unlikely(cfqq == cfqd->active_queue)) {
2650 __cfq_slice_expired(cfqd, cfqq, 0);
2651 cfq_schedule_dispatch(cfqd);
2652 }
2653
2654 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002655
Jens Axboe89850f72006-07-22 16:48:31 +02002656 cfq_put_queue(cfqq);
2657}
2658
2659static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2660 struct cfq_io_context *cic)
2661{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002662 struct io_context *ioc = cic->ioc;
2663
Jens Axboefc463792006-08-29 09:05:44 +02002664 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002665
2666 /*
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002667 * Make sure dead mark is seen for dead queues
Jens Axboe4ac845a2008-01-24 08:44:49 +01002668 */
Jens Axboefc463792006-08-29 09:05:44 +02002669 smp_wmb();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002670 cic->key = cfqd_dead_key(cfqd);
Jens Axboefc463792006-08-29 09:05:44 +02002671
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002672 if (ioc->ioc_data == cic)
2673 rcu_assign_pointer(ioc->ioc_data, NULL);
2674
Jens Axboeff6657c2009-04-08 10:58:57 +02002675 if (cic->cfqq[BLK_RW_ASYNC]) {
2676 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2677 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002678 }
2679
Jens Axboeff6657c2009-04-08 10:58:57 +02002680 if (cic->cfqq[BLK_RW_SYNC]) {
2681 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2682 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002683 }
Jens Axboe89850f72006-07-22 16:48:31 +02002684}
2685
Jens Axboe4ac845a2008-01-24 08:44:49 +01002686static void cfq_exit_single_io_context(struct io_context *ioc,
2687 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002688{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002689 struct cfq_data *cfqd = cic_to_cfqd(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002690
Jens Axboe89850f72006-07-22 16:48:31 +02002691 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002692 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002693 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002694
Jens Axboe4ac845a2008-01-24 08:44:49 +01002695 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002696
2697 /*
2698 * Ensure we get a fresh copy of the ->key to prevent
2699 * race between exiting task and queue
2700 */
2701 smp_read_barrier_depends();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002702 if (cic->key == cfqd)
Jens Axboe62c1fe92008-12-15 21:19:25 +01002703 __cfq_exit_single_io_context(cfqd, cic);
2704
Jens Axboe4ac845a2008-01-24 08:44:49 +01002705 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002706 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002707}
2708
Jens Axboe498d3aa22007-04-26 12:54:48 +02002709/*
2710 * The process that ioc belongs to has exited, we need to clean up
2711 * and put the internal structures we have that belongs to that process.
2712 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002713static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002715 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716}
2717
Jens Axboe22e2c502005-06-27 10:55:12 +02002718static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002719cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720{
Jens Axboeb5deef92006-07-19 23:39:40 +02002721 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722
Christoph Lameter94f60302007-07-17 04:03:29 -07002723 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2724 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002726 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002727 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002728 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002729 cic->dtor = cfq_free_io_context;
2730 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002731 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 }
2733
2734 return cic;
2735}
2736
Jens Axboefd0928d2008-01-24 08:52:45 +01002737static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002738{
2739 struct task_struct *tsk = current;
2740 int ioprio_class;
2741
Jens Axboe3b181522005-06-27 10:56:24 +02002742 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002743 return;
2744
Jens Axboefd0928d2008-01-24 08:52:45 +01002745 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002746 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002747 default:
2748 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2749 case IOPRIO_CLASS_NONE:
2750 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002751 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002752 */
2753 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002754 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002755 break;
2756 case IOPRIO_CLASS_RT:
2757 cfqq->ioprio = task_ioprio(ioc);
2758 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2759 break;
2760 case IOPRIO_CLASS_BE:
2761 cfqq->ioprio = task_ioprio(ioc);
2762 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2763 break;
2764 case IOPRIO_CLASS_IDLE:
2765 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2766 cfqq->ioprio = 7;
2767 cfq_clear_cfqq_idle_window(cfqq);
2768 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002769 }
2770
2771 /*
2772 * keep track of original prio settings in case we have to temporarily
2773 * elevate the priority of this queue
2774 */
2775 cfqq->org_ioprio = cfqq->ioprio;
2776 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002777 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002778}
2779
Jens Axboefebffd62008-01-28 13:19:43 +01002780static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002781{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002782 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002783 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002784 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002785
Jens Axboecaaa5f92006-06-16 11:23:00 +02002786 if (unlikely(!cfqd))
2787 return;
2788
Jens Axboec1b707d2006-10-30 19:54:23 +01002789 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002790
Jens Axboeff6657c2009-04-08 10:58:57 +02002791 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002792 if (cfqq) {
2793 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002794 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2795 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002796 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002797 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002798 cfq_put_queue(cfqq);
2799 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002800 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002801
Jens Axboeff6657c2009-04-08 10:58:57 +02002802 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002803 if (cfqq)
2804 cfq_mark_cfqq_prio_changed(cfqq);
2805
Jens Axboec1b707d2006-10-30 19:54:23 +01002806 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002807}
2808
Jens Axboefc463792006-08-29 09:05:44 +02002809static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002811 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002812 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813}
2814
Jens Axboed5036d72009-06-26 10:44:34 +02002815static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002816 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002817{
2818 RB_CLEAR_NODE(&cfqq->rb_node);
2819 RB_CLEAR_NODE(&cfqq->p_node);
2820 INIT_LIST_HEAD(&cfqq->fifo);
2821
Shaohua Li30d7b942011-01-07 08:46:59 +01002822 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02002823 cfqq->cfqd = cfqd;
2824
2825 cfq_mark_cfqq_prio_changed(cfqq);
2826
2827 if (is_sync) {
2828 if (!cfq_class_idle(cfqq))
2829 cfq_mark_cfqq_idle_window(cfqq);
2830 cfq_mark_cfqq_sync(cfqq);
2831 }
2832 cfqq->pid = pid;
2833}
2834
Vivek Goyal246103332009-12-03 12:59:51 -05002835#ifdef CONFIG_CFQ_GROUP_IOSCHED
2836static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2837{
2838 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002839 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal246103332009-12-03 12:59:51 -05002840 unsigned long flags;
2841 struct request_queue *q;
2842
2843 if (unlikely(!cfqd))
2844 return;
2845
2846 q = cfqd->queue;
2847
2848 spin_lock_irqsave(q->queue_lock, flags);
2849
2850 if (sync_cfqq) {
2851 /*
2852 * Drop reference to sync queue. A new sync queue will be
2853 * assigned in new group upon arrival of a fresh request.
2854 */
2855 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2856 cic_set_cfqq(cic, NULL, 1);
2857 cfq_put_queue(sync_cfqq);
2858 }
2859
2860 spin_unlock_irqrestore(q->queue_lock, flags);
2861}
2862
2863static void cfq_ioc_set_cgroup(struct io_context *ioc)
2864{
2865 call_for_each_cic(ioc, changed_cgroup);
2866 ioc->cgroup_changed = 0;
2867}
2868#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2869
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002871cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002872 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002875 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002876 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877
2878retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002879 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002880 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002881 /* cic always exists here */
2882 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883
Jens Axboe6118b702009-06-30 09:34:12 +02002884 /*
2885 * Always try a new alloc if we fell back to the OOM cfqq
2886 * originally, since it should just be a temporary situation.
2887 */
2888 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2889 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890 if (new_cfqq) {
2891 cfqq = new_cfqq;
2892 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002893 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002895 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002896 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002897 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002899 if (new_cfqq)
2900 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002901 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002902 cfqq = kmem_cache_alloc_node(cfq_pool,
2903 gfp_mask | __GFP_ZERO,
2904 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002905 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906
Jens Axboe6118b702009-06-30 09:34:12 +02002907 if (cfqq) {
2908 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2909 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002910 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002911 cfq_log_cfqq(cfqd, cfqq, "alloced");
2912 } else
2913 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914 }
2915
2916 if (new_cfqq)
2917 kmem_cache_free(cfq_pool, new_cfqq);
2918
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 return cfqq;
2920}
2921
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002922static struct cfq_queue **
2923cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2924{
Jens Axboefe094d92008-01-31 13:08:54 +01002925 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002926 case IOPRIO_CLASS_RT:
2927 return &cfqd->async_cfqq[0][ioprio];
2928 case IOPRIO_CLASS_BE:
2929 return &cfqd->async_cfqq[1][ioprio];
2930 case IOPRIO_CLASS_IDLE:
2931 return &cfqd->async_idle_cfqq;
2932 default:
2933 BUG();
2934 }
2935}
2936
Jens Axboe15c31be2007-07-10 13:43:25 +02002937static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002938cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002939 gfp_t gfp_mask)
2940{
Jens Axboefd0928d2008-01-24 08:52:45 +01002941 const int ioprio = task_ioprio(ioc);
2942 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002943 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002944 struct cfq_queue *cfqq = NULL;
2945
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002946 if (!is_sync) {
2947 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2948 cfqq = *async_cfqq;
2949 }
2950
Jens Axboe6118b702009-06-30 09:34:12 +02002951 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002952 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002953
2954 /*
2955 * pin the queue now that it's allocated, scheduler exit will prune it
2956 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002957 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01002958 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002959 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002960 }
2961
Shaohua Li30d7b942011-01-07 08:46:59 +01002962 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02002963 return cfqq;
2964}
2965
Jens Axboe498d3aa22007-04-26 12:54:48 +02002966/*
2967 * We drop cfq io contexts lazily, so we may find a dead one.
2968 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002969static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002970cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2971 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002972{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002973 unsigned long flags;
2974
Jens Axboefc463792006-08-29 09:05:44 +02002975 WARN_ON(!list_empty(&cic->queue_list));
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002976 BUG_ON(cic->key != cfqd_dead_key(cfqd));
Jens Axboe597bc482007-04-24 21:23:53 +02002977
Jens Axboe4ac845a2008-01-24 08:44:49 +01002978 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02002979
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002980 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002981
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002982 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
Jens Axboeffc4e752008-02-19 10:02:29 +01002983 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002984 spin_unlock_irqrestore(&ioc->lock, flags);
2985
2986 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002987}
2988
Jens Axboee2d74ac2006-03-28 08:59:01 +02002989static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01002990cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002991{
Jens Axboee2d74ac2006-03-28 08:59:01 +02002992 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002993 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002994
Vasily Tarasov91fac312007-04-25 12:29:51 +02002995 if (unlikely(!ioc))
2996 return NULL;
2997
Jens Axboed6de8be2008-05-28 14:46:59 +02002998 rcu_read_lock();
2999
Jens Axboe597bc482007-04-24 21:23:53 +02003000 /*
3001 * we maintain a last-hit cache, to avoid browsing over the tree
3002 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003003 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02003004 if (cic && cic->key == cfqd) {
3005 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02003006 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003007 }
Jens Axboe597bc482007-04-24 21:23:53 +02003008
Jens Axboe4ac845a2008-01-24 08:44:49 +01003009 do {
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003010 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003011 rcu_read_unlock();
3012 if (!cic)
3013 break;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003014 if (unlikely(cic->key != cfqd)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01003015 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003016 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01003017 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003018 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003019
Jens Axboed6de8be2008-05-28 14:46:59 +02003020 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003021 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003022 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003023 break;
3024 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003025
Jens Axboe4ac845a2008-01-24 08:44:49 +01003026 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003027}
3028
Jens Axboe4ac845a2008-01-24 08:44:49 +01003029/*
3030 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3031 * the process specific cfq io context when entered from the block layer.
3032 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3033 */
Jens Axboefebffd62008-01-28 13:19:43 +01003034static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3035 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003036{
Jens Axboe0261d682006-10-30 19:07:48 +01003037 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003038 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003039
Jens Axboe4ac845a2008-01-24 08:44:49 +01003040 ret = radix_tree_preload(gfp_mask);
3041 if (!ret) {
3042 cic->ioc = ioc;
3043 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003044
Jens Axboe4ac845a2008-01-24 08:44:49 +01003045 spin_lock_irqsave(&ioc->lock, flags);
3046 ret = radix_tree_insert(&ioc->radix_root,
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003047 cfqd->cic_index, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01003048 if (!ret)
3049 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003050 spin_unlock_irqrestore(&ioc->lock, flags);
3051
3052 radix_tree_preload_end();
3053
3054 if (!ret) {
3055 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3056 list_add(&cic->queue_list, &cfqd->cic_list);
3057 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003058 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003059 }
3060
Jens Axboe4ac845a2008-01-24 08:44:49 +01003061 if (ret)
3062 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02003063
Jens Axboe4ac845a2008-01-24 08:44:49 +01003064 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003065}
3066
Jens Axboe22e2c502005-06-27 10:55:12 +02003067/*
3068 * Setup general io context and cfq io context. There can be several cfq
3069 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02003070 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02003071 */
3072static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02003073cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074{
Jens Axboe22e2c502005-06-27 10:55:12 +02003075 struct io_context *ioc = NULL;
3076 struct cfq_io_context *cic;
3077
3078 might_sleep_if(gfp_mask & __GFP_WAIT);
3079
Jens Axboeb5deef92006-07-19 23:39:40 +02003080 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02003081 if (!ioc)
3082 return NULL;
3083
Jens Axboe4ac845a2008-01-24 08:44:49 +01003084 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003085 if (cic)
3086 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02003087
Jens Axboee2d74ac2006-03-28 08:59:01 +02003088 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3089 if (cic == NULL)
3090 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02003091
Jens Axboe4ac845a2008-01-24 08:44:49 +01003092 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3093 goto err_free;
3094
Jens Axboe22e2c502005-06-27 10:55:12 +02003095out:
Jens Axboefc463792006-08-29 09:05:44 +02003096 smp_read_barrier_depends();
3097 if (unlikely(ioc->ioprio_changed))
3098 cfq_ioc_set_ioprio(ioc);
3099
Vivek Goyal246103332009-12-03 12:59:51 -05003100#ifdef CONFIG_CFQ_GROUP_IOSCHED
3101 if (unlikely(ioc->cgroup_changed))
3102 cfq_ioc_set_cgroup(ioc);
3103#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003104 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003105err_free:
3106 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003107err:
3108 put_io_context(ioc);
3109 return NULL;
3110}
3111
3112static void
3113cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3114{
Jens Axboeaaf12282007-01-19 11:30:16 +11003115 unsigned long elapsed = jiffies - cic->last_end_request;
3116 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003117
3118 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3119 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3120 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3121}
3122
Jens Axboe206dc692006-03-28 13:03:44 +02003123static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003124cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003125 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003126{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003127 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003128 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003129 if (cfqq->last_request_pos) {
3130 if (cfqq->last_request_pos < blk_rq_pos(rq))
3131 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3132 else
3133 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3134 }
Jens Axboe206dc692006-03-28 13:03:44 +02003135
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003136 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003137 if (blk_queue_nonrot(cfqd->queue))
3138 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3139 else
3140 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003141}
Jens Axboe22e2c502005-06-27 10:55:12 +02003142
3143/*
3144 * Disable idle window if the process thinks too long or seeks so much that
3145 * it doesn't matter
3146 */
3147static void
3148cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3149 struct cfq_io_context *cic)
3150{
Jens Axboe7b679132008-05-30 12:23:07 +02003151 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003152
Jens Axboe08717142008-01-28 11:38:15 +01003153 /*
3154 * Don't idle for async or idle io prio class
3155 */
3156 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003157 return;
3158
Jens Axboec265a7f2008-06-26 13:49:33 +02003159 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003160
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003161 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3162 cfq_mark_cfqq_deep(cfqq);
3163
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003164 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3165 enable_idle = 0;
3166 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003167 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003168 enable_idle = 0;
3169 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003170 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003171 enable_idle = 0;
3172 else
3173 enable_idle = 1;
3174 }
3175
Jens Axboe7b679132008-05-30 12:23:07 +02003176 if (old_idle != enable_idle) {
3177 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3178 if (enable_idle)
3179 cfq_mark_cfqq_idle_window(cfqq);
3180 else
3181 cfq_clear_cfqq_idle_window(cfqq);
3182 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003183}
3184
Jens Axboe22e2c502005-06-27 10:55:12 +02003185/*
3186 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3187 * no or if we aren't sure, a 1 will cause a preempt.
3188 */
Jens Axboea6151c32009-10-07 20:02:57 +02003189static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003190cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003191 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003192{
Jens Axboe6d048f52007-04-25 12:44:27 +02003193 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003194
Jens Axboe6d048f52007-04-25 12:44:27 +02003195 cfqq = cfqd->active_queue;
3196 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003197 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003198
Jens Axboe6d048f52007-04-25 12:44:27 +02003199 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003200 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003201
3202 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003203 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003204
Jens Axboe22e2c502005-06-27 10:55:12 +02003205 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003206 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3207 */
3208 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3209 return false;
3210
3211 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003212 * if the new request is sync, but the currently running queue is
3213 * not, let the sync request have priority.
3214 */
Jens Axboe5e705372006-07-13 12:39:25 +02003215 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003216 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003217
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003218 if (new_cfqq->cfqg != cfqq->cfqg)
3219 return false;
3220
3221 if (cfq_slice_used(cfqq))
3222 return true;
3223
3224 /* Allow preemption only if we are idling on sync-noidle tree */
3225 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3226 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3227 new_cfqq->service_tree->count == 2 &&
3228 RB_EMPTY_ROOT(&cfqq->sort_list))
3229 return true;
3230
Jens Axboe374f84a2006-07-23 01:42:19 +02003231 /*
3232 * So both queues are sync. Let the new request get disk time if
3233 * it's a metadata request and the current queue is doing regular IO.
3234 */
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003235 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003236 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003237
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003238 /*
3239 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3240 */
3241 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003242 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003243
Shaohua Lid2d59e12010-11-08 15:01:03 +01003244 /* An idle queue should not be idle now for some reason */
3245 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3246 return true;
3247
Jens Axboe1e3335d2007-02-14 19:59:49 +01003248 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003249 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003250
3251 /*
3252 * if this request is as-good as one we would expect from the
3253 * current cfqq, let it preempt
3254 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003255 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003256 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003257
Jens Axboea6151c32009-10-07 20:02:57 +02003258 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003259}
3260
3261/*
3262 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3263 * let it have half of its nominal slice.
3264 */
3265static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3266{
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003267 struct cfq_queue *old_cfqq = cfqd->active_queue;
3268
Jens Axboe7b679132008-05-30 12:23:07 +02003269 cfq_log_cfqq(cfqd, cfqq, "preempt");
Vivek Goyale5ff0822010-04-26 19:25:11 +02003270 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003271
Jens Axboebf572252006-07-19 20:29:12 +02003272 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003273 * workload type is changed, don't save slice, otherwise preempt
3274 * doesn't happen
3275 */
3276 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3277 cfqq->cfqg->saved_workload_slice = 0;
3278
3279 /*
Jens Axboebf572252006-07-19 20:29:12 +02003280 * Put the new queue at the front of the of the current list,
3281 * so we know that it will be selected next.
3282 */
3283 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003284
3285 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02003286
Jens Axboe44f7c162007-01-19 11:51:58 +11003287 cfqq->slice_end = 0;
3288 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003289}
3290
3291/*
Jens Axboe5e705372006-07-13 12:39:25 +02003292 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003293 * something we should do about it
3294 */
3295static void
Jens Axboe5e705372006-07-13 12:39:25 +02003296cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3297 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003298{
Jens Axboe5e705372006-07-13 12:39:25 +02003299 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003300
Aaron Carroll45333d52008-08-26 15:52:36 +02003301 cfqd->rq_queued++;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02003302 if (rq->cmd_flags & REQ_META)
Jens Axboe374f84a2006-07-23 01:42:19 +02003303 cfqq->meta_pending++;
3304
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003305 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003306 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003307 cfq_update_idle_window(cfqd, cfqq, cic);
3308
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003309 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003310
3311 if (cfqq == cfqd->active_queue) {
3312 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003313 * Remember that we saw a request from this process, but
3314 * don't start queuing just yet. Otherwise we risk seeing lots
3315 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003316 * and merging. If the request is already larger than a single
3317 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003318 * merging is already done. Ditto for a busy system that
3319 * has other work pending, don't risk delaying until the
3320 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003321 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003322 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003323 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3324 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003325 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003326 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003327 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003328 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003329 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003330 &cfqq->cfqg->blkg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003331 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003332 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003333 }
Jens Axboe5e705372006-07-13 12:39:25 +02003334 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003335 /*
3336 * not the active queue - expire current slice if it is
3337 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003338 * has some old slice time left and is of higher priority or
3339 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003340 */
3341 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09003342 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003343 }
3344}
3345
Jens Axboe165125e2007-07-24 09:28:11 +02003346static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003347{
Jens Axboeb4878f22005-10-20 16:42:29 +02003348 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003349 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003350
Jens Axboe7b679132008-05-30 12:23:07 +02003351 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003352 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353
Jens Axboe30996f42009-10-05 11:03:39 +02003354 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003355 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003356 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003357 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003358 &cfqd->serving_group->blkg, rq_data_dir(rq),
3359 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003360 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361}
3362
Aaron Carroll45333d52008-08-26 15:52:36 +02003363/*
3364 * Update hw_tag based on peak queue depth over 50 samples under
3365 * sufficient load.
3366 */
3367static void cfq_update_hw_tag(struct cfq_data *cfqd)
3368{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003369 struct cfq_queue *cfqq = cfqd->active_queue;
3370
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003371 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3372 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003373
3374 if (cfqd->hw_tag == 1)
3375 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003376
3377 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003378 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003379 return;
3380
Shaohua Li1a1238a2009-10-27 08:46:23 +01003381 /*
3382 * If active queue hasn't enough requests and can idle, cfq might not
3383 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3384 * case
3385 */
3386 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3387 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003388 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003389 return;
3390
Aaron Carroll45333d52008-08-26 15:52:36 +02003391 if (cfqd->hw_tag_samples++ < 50)
3392 return;
3393
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003394 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003395 cfqd->hw_tag = 1;
3396 else
3397 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003398}
3399
Vivek Goyal7667aa02009-12-08 17:52:58 -05003400static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3401{
3402 struct cfq_io_context *cic = cfqd->active_cic;
3403
Justin TerAvest02a8f012011-02-09 14:20:03 +01003404 /* If the queue already has requests, don't wait */
3405 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3406 return false;
3407
Vivek Goyal7667aa02009-12-08 17:52:58 -05003408 /* If there are other queues in the group, don't wait */
3409 if (cfqq->cfqg->nr_cfqq > 1)
3410 return false;
3411
3412 if (cfq_slice_used(cfqq))
3413 return true;
3414
3415 /* if slice left is less than think time, wait busy */
3416 if (cic && sample_valid(cic->ttime_samples)
3417 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3418 return true;
3419
3420 /*
3421 * If think times is less than a jiffy than ttime_mean=0 and above
3422 * will not be true. It might happen that slice has not expired yet
3423 * but will expire soon (4-5 ns) during select_queue(). To cover the
3424 * case where think time is less than a jiffy, mark the queue wait
3425 * busy if only 1 jiffy is left in the slice.
3426 */
3427 if (cfqq->slice_end - jiffies == 1)
3428 return true;
3429
3430 return false;
3431}
3432
Jens Axboe165125e2007-07-24 09:28:11 +02003433static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434{
Jens Axboe5e705372006-07-13 12:39:25 +02003435 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003436 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003437 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003438 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439
Jens Axboeb4878f22005-10-20 16:42:29 +02003440 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003441 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3442 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443
Aaron Carroll45333d52008-08-26 15:52:36 +02003444 cfq_update_hw_tag(cfqd);
3445
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003446 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003447 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003448 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003449 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003450 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003451 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3452 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3453 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003455 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003456
Vivek Goyal365722b2009-10-03 15:21:27 +02003457 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003458 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003459 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3460 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003461 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003462
3463 /*
3464 * If this is the active queue, check if it needs to be expired,
3465 * or if we want to idle in case it has no pending requests.
3466 */
3467 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003468 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3469
Jens Axboe44f7c162007-01-19 11:51:58 +11003470 if (cfq_cfqq_slice_new(cfqq)) {
3471 cfq_set_prio_slice(cfqd, cfqq);
3472 cfq_clear_cfqq_slice_new(cfqq);
3473 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003474
3475 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003476 * Should we wait for next request to come in before we expire
3477 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003478 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003479 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003480 unsigned long extend_sl = cfqd->cfq_slice_idle;
3481 if (!cfqd->cfq_slice_idle)
3482 extend_sl = cfqd->cfq_group_idle;
3483 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003484 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003485 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003486 }
3487
Jens Axboea36e71f2009-04-15 12:15:11 +02003488 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003489 * Idling is not enabled on:
3490 * - expired queues
3491 * - idle-priority queues
3492 * - async queues
3493 * - queues with still some requests queued
3494 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003495 */
Jens Axboe08717142008-01-28 11:38:15 +01003496 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003497 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003498 else if (sync && cfqq_empty &&
3499 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003500 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003501 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003502 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003503
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003504 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003505 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506}
3507
Jens Axboe22e2c502005-06-27 10:55:12 +02003508/*
3509 * we temporarily boost lower priority queues if they are holding fs exclusive
3510 * resources. they are boosted to normal prio (CLASS_BE/4)
3511 */
3512static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513{
Jens Axboe22e2c502005-06-27 10:55:12 +02003514 if (has_fs_excl()) {
3515 /*
3516 * boost idle prio on transactions that would lock out other
3517 * users of the filesystem
3518 */
3519 if (cfq_class_idle(cfqq))
3520 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3521 if (cfqq->ioprio > IOPRIO_NORM)
3522 cfqq->ioprio = IOPRIO_NORM;
3523 } else {
3524 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003525 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003526 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003527 cfqq->ioprio_class = cfqq->org_ioprio_class;
3528 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003530}
3531
Jens Axboe89850f72006-07-22 16:48:31 +02003532static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003533{
Jens Axboe1b379d82009-08-11 08:26:11 +02003534 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003535 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003536 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003537 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003538
3539 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003540}
3541
Jens Axboe165125e2007-07-24 09:28:11 +02003542static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003543{
3544 struct cfq_data *cfqd = q->elevator->elevator_data;
3545 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003546 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003547 struct cfq_queue *cfqq;
3548
3549 /*
3550 * don't force setup of a queue from here, as a call to may_queue
3551 * does not necessarily imply that a request actually will be queued.
3552 * so just lookup a possibly existing queue, or return 'may queue'
3553 * if that fails
3554 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003555 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003556 if (!cic)
3557 return ELV_MQUEUE_MAY;
3558
Jens Axboeb0b78f82009-04-08 10:56:08 +02003559 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003560 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003561 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003562 cfq_prio_boost(cfqq);
3563
Jens Axboe89850f72006-07-22 16:48:31 +02003564 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003565 }
3566
3567 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568}
3569
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570/*
3571 * queue lock held here
3572 */
Jens Axboebb37b942006-12-01 10:42:33 +01003573static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574{
Jens Axboe5e705372006-07-13 12:39:25 +02003575 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576
Jens Axboe5e705372006-07-13 12:39:25 +02003577 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003578 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579
Jens Axboe22e2c502005-06-27 10:55:12 +02003580 BUG_ON(!cfqq->allocated[rw]);
3581 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582
Jens Axboe5e705372006-07-13 12:39:25 +02003583 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584
Mike Snitzerc1867942011-02-11 11:08:00 +01003585 rq->elevator_private[0] = NULL;
3586 rq->elevator_private[1] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003588 /* Put down rq reference on cfqg */
3589 cfq_put_cfqg(RQ_CFQG(rq));
Mike Snitzerc1867942011-02-11 11:08:00 +01003590 rq->elevator_private[2] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003591
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592 cfq_put_queue(cfqq);
3593 }
3594}
3595
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003596static struct cfq_queue *
3597cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3598 struct cfq_queue *cfqq)
3599{
3600 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3601 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003602 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003603 cfq_put_queue(cfqq);
3604 return cic_to_cfqq(cic, 1);
3605}
3606
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003607/*
3608 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3609 * was the last process referring to said cfqq.
3610 */
3611static struct cfq_queue *
3612split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3613{
3614 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003615 cfqq->pid = current->pid;
3616 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003617 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003618 return cfqq;
3619 }
3620
3621 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003622
3623 cfq_put_cooperator(cfqq);
3624
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003625 cfq_put_queue(cfqq);
3626 return NULL;
3627}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003629 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003630 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003631static int
Jens Axboe165125e2007-07-24 09:28:11 +02003632cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003633{
3634 struct cfq_data *cfqd = q->elevator->elevator_data;
3635 struct cfq_io_context *cic;
3636 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003637 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003638 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639 unsigned long flags;
3640
3641 might_sleep_if(gfp_mask & __GFP_WAIT);
3642
Jens Axboee2d74ac2006-03-28 08:59:01 +02003643 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003644
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645 spin_lock_irqsave(q->queue_lock, flags);
3646
Jens Axboe22e2c502005-06-27 10:55:12 +02003647 if (!cic)
3648 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003650new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003651 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003652 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003653 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003654 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003655 } else {
3656 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003657 * If the queue was seeky for too long, break it apart.
3658 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003659 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003660 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3661 cfqq = split_cfqq(cic, cfqq);
3662 if (!cfqq)
3663 goto new_queue;
3664 }
3665
3666 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003667 * Check to see if this queue is scheduled to merge with
3668 * another, closely cooperating queue. The merging of
3669 * queues happens here as it must be done in process context.
3670 * The reference on new_cfqq was taken in merge_cfqqs.
3671 */
3672 if (cfqq->new_cfqq)
3673 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003674 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675
3676 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02003677
Jens Axboe6fae9c22011-03-01 15:04:39 -05003678 cfqq->ref++;
Mike Snitzerc1867942011-02-11 11:08:00 +01003679 rq->elevator_private[0] = cic;
3680 rq->elevator_private[1] = cfqq;
3681 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
Jens Axboe93803e02011-03-07 08:59:06 +01003682 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe5e705372006-07-13 12:39:25 +02003683 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003684
Jens Axboe22e2c502005-06-27 10:55:12 +02003685queue_fail:
3686 if (cic)
3687 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003688
Jens Axboe23e018a2009-10-05 08:52:35 +02003689 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003690 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003691 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003692 return 1;
3693}
3694
David Howells65f27f32006-11-22 14:55:48 +00003695static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003696{
David Howells65f27f32006-11-22 14:55:48 +00003697 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003698 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003699 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003700
Jens Axboe40bb54d2009-04-15 12:11:10 +02003701 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09003702 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003703 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003704}
3705
3706/*
3707 * Timer running if the active_queue is currently idling inside its time slice
3708 */
3709static void cfq_idle_slice_timer(unsigned long data)
3710{
3711 struct cfq_data *cfqd = (struct cfq_data *) data;
3712 struct cfq_queue *cfqq;
3713 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003714 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003715
Jens Axboe7b679132008-05-30 12:23:07 +02003716 cfq_log(cfqd, "idle timer fired");
3717
Jens Axboe22e2c502005-06-27 10:55:12 +02003718 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3719
Jens Axboefe094d92008-01-31 13:08:54 +01003720 cfqq = cfqd->active_queue;
3721 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003722 timed_out = 0;
3723
Jens Axboe22e2c502005-06-27 10:55:12 +02003724 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003725 * We saw a request before the queue expired, let it through
3726 */
3727 if (cfq_cfqq_must_dispatch(cfqq))
3728 goto out_kick;
3729
3730 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003731 * expired
3732 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003733 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003734 goto expire;
3735
3736 /*
3737 * only expire and reinvoke request handler, if there are
3738 * other queues with pending requests
3739 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003740 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003741 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003742
3743 /*
3744 * not expired and it has a request pending, let it dispatch
3745 */
Jens Axboe75e50982009-04-07 08:56:14 +02003746 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003747 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003748
3749 /*
3750 * Queue depth flag is reset only when the idle didn't succeed
3751 */
3752 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003753 }
3754expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003755 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003756out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003757 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003758out_cont:
3759 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3760}
3761
Jens Axboe3b181522005-06-27 10:56:24 +02003762static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3763{
3764 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003765 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003766}
Jens Axboe22e2c502005-06-27 10:55:12 +02003767
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003768static void cfq_put_async_queues(struct cfq_data *cfqd)
3769{
3770 int i;
3771
3772 for (i = 0; i < IOPRIO_BE_NR; i++) {
3773 if (cfqd->async_cfqq[0][i])
3774 cfq_put_queue(cfqd->async_cfqq[0][i]);
3775 if (cfqd->async_cfqq[1][i])
3776 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003777 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003778
3779 if (cfqd->async_idle_cfqq)
3780 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003781}
3782
Jens Axboebb729bc2009-12-06 09:54:19 +01003783static void cfq_cfqd_free(struct rcu_head *head)
3784{
3785 kfree(container_of(head, struct cfq_data, rcu));
3786}
3787
Jens Axboeb374d182008-10-31 10:05:07 +01003788static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789{
Jens Axboe22e2c502005-06-27 10:55:12 +02003790 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003791 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003792
Jens Axboe3b181522005-06-27 10:56:24 +02003793 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003794
Al Virod9ff4182006-03-18 13:51:22 -05003795 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003796
Al Virod9ff4182006-03-18 13:51:22 -05003797 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003798 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003799
3800 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003801 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3802 struct cfq_io_context,
3803 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003804
3805 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003806 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003807
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003808 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003809 cfq_release_cfq_groups(cfqd);
Vivek Goyale98ef892010-06-18 10:39:47 -04003810 cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003811
Al Virod9ff4182006-03-18 13:51:22 -05003812 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003813
3814 cfq_shutdown_timer_wq(cfqd);
3815
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003816 spin_lock(&cic_index_lock);
3817 ida_remove(&cic_index_ida, cfqd->cic_index);
3818 spin_unlock(&cic_index_lock);
3819
Vivek Goyalb1c35762009-12-03 12:59:47 -05003820 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003821 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003822}
3823
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003824static int cfq_alloc_cic_index(void)
3825{
3826 int index, error;
3827
3828 do {
3829 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3830 return -ENOMEM;
3831
3832 spin_lock(&cic_index_lock);
3833 error = ida_get_new(&cic_index_ida, &index);
3834 spin_unlock(&cic_index_lock);
3835 if (error && error != -EAGAIN)
3836 return error;
3837 } while (error);
3838
3839 return index;
3840}
3841
Jens Axboe165125e2007-07-24 09:28:11 +02003842static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003843{
3844 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003845 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003846 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003847 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003849 i = cfq_alloc_cic_index();
3850 if (i < 0)
3851 return NULL;
3852
Christoph Lameter94f60302007-07-17 04:03:29 -07003853 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003854 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003855 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856
Shaohua Li30d7b942011-01-07 08:46:59 +01003857 /*
3858 * Don't need take queue_lock in the routine, since we are
3859 * initializing the ioscheduler, and nobody is using cfqd
3860 */
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003861 cfqd->cic_index = i;
3862
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003863 /* Init root service tree */
3864 cfqd->grp_service_tree = CFQ_RB_ROOT;
3865
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003866 /* Init root group */
3867 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003868 for_each_cfqg_st(cfqg, i, j, st)
3869 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003870 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003871
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003872 /* Give preference to root group over other groups */
3873 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3874
Vivek Goyal25fb5162009-12-03 12:59:46 -05003875#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003876 /*
3877 * Take a reference to root group which we never drop. This is just
3878 * to make sure that cfq_put_cfqg() does not try to kfree root group
3879 */
Shaohua Li329a6782011-01-07 08:48:28 +01003880 cfqg->ref = 1;
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003881 rcu_read_lock();
Vivek Goyale98ef892010-06-18 10:39:47 -04003882 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3883 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003884 rcu_read_unlock();
Vivek Goyal25fb5162009-12-03 12:59:46 -05003885#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003886 /*
3887 * Not strictly needed (since RB_ROOT just clears the node and we
3888 * zeroed cfqd on alloc), but better be safe in case someone decides
3889 * to add magic to the rb code
3890 */
3891 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3892 cfqd->prio_trees[i] = RB_ROOT;
3893
Jens Axboe6118b702009-06-30 09:34:12 +02003894 /*
3895 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3896 * Grab a permanent reference to it, so that the normal code flow
3897 * will not attempt to free it.
3898 */
3899 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01003900 cfqd->oom_cfqq.ref++;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003901 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003902
Al Virod9ff4182006-03-18 13:51:22 -05003903 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904
Linus Torvalds1da177e2005-04-16 15:20:36 -07003905 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906
Jens Axboe22e2c502005-06-27 10:55:12 +02003907 init_timer(&cfqd->idle_slice_timer);
3908 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3909 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3910
Jens Axboe23e018a2009-10-05 08:52:35 +02003911 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003912
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003914 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3915 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916 cfqd->cfq_back_max = cfq_back_max;
3917 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003918 cfqd->cfq_slice[0] = cfq_slice_async;
3919 cfqd->cfq_slice[1] = cfq_slice_sync;
3920 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3921 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003922 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003923 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003924 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003925 /*
3926 * we optimistically start assuming sync ops weren't delayed in last
3927 * second, in order to have larger depth for async operations.
3928 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003929 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003930 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003931}
3932
3933static void cfq_slab_kill(void)
3934{
Jens Axboed6de8be2008-05-28 14:46:59 +02003935 /*
3936 * Caller already ensured that pending RCU callbacks are completed,
3937 * so we should have no busy allocations at this point.
3938 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939 if (cfq_pool)
3940 kmem_cache_destroy(cfq_pool);
3941 if (cfq_ioc_pool)
3942 kmem_cache_destroy(cfq_ioc_pool);
3943}
3944
3945static int __init cfq_slab_setup(void)
3946{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003947 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948 if (!cfq_pool)
3949 goto fail;
3950
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003951 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952 if (!cfq_ioc_pool)
3953 goto fail;
3954
3955 return 0;
3956fail:
3957 cfq_slab_kill();
3958 return -ENOMEM;
3959}
3960
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961/*
3962 * sysfs parts below -->
3963 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964static ssize_t
3965cfq_var_show(unsigned int var, char *page)
3966{
3967 return sprintf(page, "%d\n", var);
3968}
3969
3970static ssize_t
3971cfq_var_store(unsigned int *var, const char *page, size_t count)
3972{
3973 char *p = (char *) page;
3974
3975 *var = simple_strtoul(p, &p, 10);
3976 return count;
3977}
3978
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003980static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003982 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 unsigned int __data = __VAR; \
3984 if (__CONV) \
3985 __data = jiffies_to_msecs(__data); \
3986 return cfq_var_show(__data, (page)); \
3987}
3988SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003989SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3990SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003991SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3992SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003993SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003994SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003995SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3996SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3997SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003998SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999#undef SHOW_FUNCTION
4000
4001#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004002static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004004 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 unsigned int __data; \
4006 int ret = cfq_var_store(&__data, (page), count); \
4007 if (__data < (MIN)) \
4008 __data = (MIN); \
4009 else if (__data > (MAX)) \
4010 __data = (MAX); \
4011 if (__CONV) \
4012 *(__PTR) = msecs_to_jiffies(__data); \
4013 else \
4014 *(__PTR) = __data; \
4015 return ret; \
4016}
4017STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004018STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4019 UINT_MAX, 1);
4020STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4021 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004022STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004023STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4024 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004025STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004026STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004027STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4028STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004029STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4030 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004031STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032#undef STORE_FUNCTION
4033
Al Viroe572ec72006-03-18 22:27:18 -05004034#define CFQ_ATTR(name) \
4035 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004036
Al Viroe572ec72006-03-18 22:27:18 -05004037static struct elv_fs_entry cfq_attrs[] = {
4038 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004039 CFQ_ATTR(fifo_expire_sync),
4040 CFQ_ATTR(fifo_expire_async),
4041 CFQ_ATTR(back_seek_max),
4042 CFQ_ATTR(back_seek_penalty),
4043 CFQ_ATTR(slice_sync),
4044 CFQ_ATTR(slice_async),
4045 CFQ_ATTR(slice_async_rq),
4046 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004047 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004048 CFQ_ATTR(low_latency),
Al Viroe572ec72006-03-18 22:27:18 -05004049 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050};
4051
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052static struct elevator_type iosched_cfq = {
4053 .ops = {
4054 .elevator_merge_fn = cfq_merge,
4055 .elevator_merged_fn = cfq_merged_request,
4056 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004057 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004058 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004059 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004061 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062 .elevator_deactivate_req_fn = cfq_deactivate_request,
4063 .elevator_queue_empty_fn = cfq_queue_empty,
4064 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004065 .elevator_former_req_fn = elv_rb_former_request,
4066 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067 .elevator_set_req_fn = cfq_set_request,
4068 .elevator_put_req_fn = cfq_put_request,
4069 .elevator_may_queue_fn = cfq_may_queue,
4070 .elevator_init_fn = cfq_init_queue,
4071 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02004072 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073 },
Al Viro3d1ab402006-03-18 18:35:43 -05004074 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075 .elevator_name = "cfq",
4076 .elevator_owner = THIS_MODULE,
4077};
4078
Vivek Goyal3e252062009-12-04 10:36:42 -05004079#ifdef CONFIG_CFQ_GROUP_IOSCHED
4080static struct blkio_policy_type blkio_policy_cfq = {
4081 .ops = {
4082 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4083 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4084 },
Vivek Goyal062a6442010-09-15 17:06:33 -04004085 .plid = BLKIO_POLICY_PROP,
Vivek Goyal3e252062009-12-04 10:36:42 -05004086};
4087#else
4088static struct blkio_policy_type blkio_policy_cfq;
4089#endif
4090
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091static int __init cfq_init(void)
4092{
Jens Axboe22e2c502005-06-27 10:55:12 +02004093 /*
4094 * could be 0 on HZ < 1000 setups
4095 */
4096 if (!cfq_slice_async)
4097 cfq_slice_async = 1;
4098 if (!cfq_slice_idle)
4099 cfq_slice_idle = 1;
4100
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004101#ifdef CONFIG_CFQ_GROUP_IOSCHED
4102 if (!cfq_group_idle)
4103 cfq_group_idle = 1;
4104#else
4105 cfq_group_idle = 0;
4106#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 if (cfq_slab_setup())
4108 return -ENOMEM;
4109
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004110 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05004111 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004113 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114}
4115
4116static void __exit cfq_exit(void)
4117{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07004118 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05004119 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05004121 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02004122 /* ioc_gone's update must be visible before reading ioc_count */
4123 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02004124
4125 /*
4126 * this also protects us from entering cfq_slab_kill() with
4127 * pending RCU callbacks
4128 */
Tejun Heo245b2e72009-06-24 15:13:48 +09004129 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02004130 wait_for_completion(&all_gone);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004131 ida_destroy(&cic_index_ida);
Christoph Hellwig83521d32005-10-30 15:01:39 -08004132 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133}
4134
4135module_init(cfq_init);
4136module_exit(cfq_exit);
4137
4138MODULE_AUTHOR("Jens Axboe");
4139MODULE_LICENSE("GPL");
4140MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");