blob: 5617ae030b15864f2d01b3669eabbcd9856922ad [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>
Al Viro1cc9be62006-03-18 12:29:52 -050010#include <linux/blkdev.h>
11#include <linux/elevator.h>
Randy Dunlapad5ebd22009-11-11 13:47:45 +010012#include <linux/jiffies.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020014#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020015#include <linux/blktrace_api.h>
Vivek Goyal25bc6b02009-12-03 12:59:43 -050016#include "blk-cgroup.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070017
18/*
19 * tunables
20 */
Jens Axboefe094d92008-01-31 13:08:54 +010021/* max queue in one round of service */
Shaohua Liabc3c742010-03-01 09:20:54 +010022static const int cfq_quantum = 8;
Arjan van de Ven64100092006-01-06 09:46:02 +010023static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010024/* maximum backwards seek, in KiB */
25static const int cfq_back_max = 16 * 1024;
26/* penalty of a backwards seek */
27static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010028static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020029static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010030static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020031static int cfq_slice_idle = HZ / 125;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +010032static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33static const int cfq_hist_divisor = 4;
Jens Axboe22e2c502005-06-27 10:55:12 +020034
Jens Axboed9e76202007-04-20 14:27:50 +020035/*
Jens Axboe08717142008-01-28 11:38:15 +010036 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020037 */
Jens Axboe08717142008-01-28 11:38:15 +010038#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020039
40/*
41 * below this threshold, we consider thinktime immediate
42 */
43#define CFQ_MIN_TT (2)
44
Jens Axboe22e2c502005-06-27 10:55:12 +020045#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020046#define CFQ_HW_QUEUE_MIN (5)
Vivek Goyal25bc6b02009-12-03 12:59:43 -050047#define CFQ_SERVICE_SHIFT 12
Jens Axboe22e2c502005-06-27 10:55:12 +020048
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010049#define CFQQ_SEEK_THR (sector_t)(8 * 100)
Shaohua Lie9ce3352010-03-19 08:03:04 +010050#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
Corrado Zoccolo41647e72010-02-27 19:45:40 +010051#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010052#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
Shaohua Liae54abe2010-02-05 13:11:45 +010053
Jens Axboefe094d92008-01-31 13:08:54 +010054#define RQ_CIC(rq) \
55 ((struct cfq_io_context *) (rq)->elevator_private)
Jens Axboe7b679132008-05-30 12:23:07 +020056#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
Linus Torvalds1da177e2005-04-16 15:20:36 -070057
Christoph Lametere18b8902006-12-06 20:33:20 -080058static struct kmem_cache *cfq_pool;
59static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Tejun Heo245b2e72009-06-24 15:13:48 +090061static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050062static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020063static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050064
Jens Axboe22e2c502005-06-27 10:55:12 +020065#define CFQ_PRIO_LISTS IOPRIO_BE_NR
66#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020067#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68
Jens Axboe206dc692006-03-28 13:03:44 +020069#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050070#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020071
Jens Axboe22e2c502005-06-27 10:55:12 +020072/*
Jens Axboecc09e292007-04-26 12:53:50 +020073 * Most of our rbtree usage is for sorting with min extraction, so
74 * if we cache the leftmost node we don't have to walk down the tree
75 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76 * move this into the elevator for the rq sorting as well.
77 */
78struct cfq_rb_root {
79 struct rb_root rb;
80 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010081 unsigned count;
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010082 unsigned total_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050083 u64 min_vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -050084 struct rb_node *active;
Jens Axboecc09e292007-04-26 12:53:50 +020085};
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010086#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
87 .count = 0, .min_vdisktime = 0, }
Jens Axboecc09e292007-04-26 12:53:50 +020088
89/*
Jens Axboe6118b702009-06-30 09:34:12 +020090 * Per process-grouping structure
91 */
92struct cfq_queue {
93 /* reference count */
94 atomic_t ref;
95 /* various state flags, see below */
96 unsigned int flags;
97 /* parent cfq_data */
98 struct cfq_data *cfqd;
99 /* service_tree member */
100 struct rb_node rb_node;
101 /* service_tree key */
102 unsigned long rb_key;
103 /* prio tree member */
104 struct rb_node p_node;
105 /* prio tree root we belong to, if any */
106 struct rb_root *p_root;
107 /* sorted list of pending requests */
108 struct rb_root sort_list;
109 /* if fifo isn't expired, next request to serve */
110 struct request *next_rq;
111 /* requests queued in sort_list */
112 int queued[2];
113 /* currently allocated requests */
114 int allocated[2];
115 /* fifo list of requests in sort_list */
116 struct list_head fifo;
117
Vivek Goyaldae739e2009-12-03 12:59:45 -0500118 /* time when queue got scheduled in to dispatch first request. */
119 unsigned long dispatch_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500120 unsigned int allocated_slice;
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100121 unsigned int slice_dispatch;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500122 /* time when first request from queue completed and slice started. */
123 unsigned long slice_start;
Jens Axboe6118b702009-06-30 09:34:12 +0200124 unsigned long slice_end;
125 long slice_resid;
Jens Axboe6118b702009-06-30 09:34:12 +0200126
127 /* pending metadata requests */
128 int meta_pending;
129 /* number of requests that are on the dispatch list or inside driver */
130 int dispatched;
131
132 /* io prio of this group */
133 unsigned short ioprio, org_ioprio;
134 unsigned short ioprio_class, org_ioprio_class;
135
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100136 pid_t pid;
137
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +0100138 u32 seek_history;
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400139 sector_t last_request_pos;
140
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100141 struct cfq_rb_root *service_tree;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400142 struct cfq_queue *new_cfqq;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500143 struct cfq_group *cfqg;
Vivek Goyalae30c282009-12-03 12:59:55 -0500144 struct cfq_group *orig_cfqg;
Jens Axboe6118b702009-06-30 09:34:12 +0200145};
146
147/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100148 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100149 * IDLE is handled separately, so it has negative index
150 */
151enum wl_prio_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100152 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500153 RT_WORKLOAD = 1,
154 IDLE_WORKLOAD = 2,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100155};
156
157/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100158 * Second index in the service_trees.
159 */
160enum wl_type_t {
161 ASYNC_WORKLOAD = 0,
162 SYNC_NOIDLE_WORKLOAD = 1,
163 SYNC_WORKLOAD = 2
164};
165
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500166/* This is per cgroup per device grouping structure */
167struct cfq_group {
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500168 /* group service_tree member */
169 struct rb_node rb_node;
170
171 /* group service_tree key */
172 u64 vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500173 unsigned int weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500174 bool on_st;
175
176 /* number of cfqq currently on this group */
177 int nr_cfqq;
178
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500179 /* Per group busy queus average. Useful for workload slice calc. */
180 unsigned int busy_queues_avg[2];
Jens Axboe22e2c502005-06-27 10:55:12 +0200181 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100182 * rr lists of queues with requests, onle rr for each priority class.
183 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200184 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100185 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100186 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500187
188 unsigned long saved_workload_slice;
189 enum wl_type_t saved_workload;
190 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500191 struct blkio_group blkg;
192#ifdef CONFIG_CFQ_GROUP_IOSCHED
193 struct hlist_node cfqd_node;
Vivek Goyalb1c35762009-12-03 12:59:47 -0500194 atomic_t ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500195#endif
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500196};
197
198/*
199 * Per block device queue structure
200 */
201struct cfq_data {
202 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500203 /* Root service tree for cfq_groups */
204 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500205 struct cfq_group root_group;
206
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100207 /*
208 * The priority currently being served
209 */
210 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100211 enum wl_type_t serving_type;
212 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500213 struct cfq_group *serving_group;
Corrado Zoccolo8e550632009-11-26 10:02:58 +0100214 bool noidle_tree_requires_idle;
Jens Axboea36e71f2009-04-15 12:15:11 +0200215
216 /*
217 * Each priority tree is sorted by next_request position. These
218 * trees are used when determining if two or more queues are
219 * interleaving requests (see cfq_close_cooperator).
220 */
221 struct rb_root prio_trees[CFQ_PRIO_LISTS];
222
Jens Axboe22e2c502005-06-27 10:55:12 +0200223 unsigned int busy_queues;
224
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100225 int rq_in_driver;
226 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200227
228 /*
229 * queue-depth detection
230 */
231 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200232 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100233 /*
234 * hw_tag can be
235 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
236 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
237 * 0 => no NCQ
238 */
239 int hw_tag_est_depth;
240 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200241
242 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200243 * idle window management
244 */
245 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200246 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200247
248 struct cfq_queue *active_queue;
249 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200250
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200251 /*
252 * async queue for each priority case
253 */
254 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
255 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200256
Jens Axboe6d048f52007-04-25 12:44:27 +0200257 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 /*
260 * tunables, see top of file
261 */
262 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200263 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 unsigned int cfq_back_penalty;
265 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200266 unsigned int cfq_slice[2];
267 unsigned int cfq_slice_async_rq;
268 unsigned int cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200269 unsigned int cfq_latency;
Vivek Goyalae30c282009-12-03 12:59:55 -0500270 unsigned int cfq_group_isolation;
Al Virod9ff4182006-03-18 13:51:22 -0500271
272 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
Jens Axboe6118b702009-06-30 09:34:12 +0200274 /*
275 * Fallback dummy cfqq for extreme OOM conditions
276 */
277 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200278
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100279 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500280
281 /* List of cfq groups being managed on this device*/
282 struct hlist_head cfqg_list;
Jens Axboebb729bc2009-12-06 09:54:19 +0100283 struct rcu_head rcu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284};
285
Vivek Goyal25fb5162009-12-03 12:59:46 -0500286static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
287
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500288static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
289 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500290 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100291{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500292 if (!cfqg)
293 return NULL;
294
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100295 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500296 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100297
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500298 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100299}
300
Jens Axboe3b181522005-06-27 10:56:24 +0200301enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100302 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
303 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200304 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100305 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100306 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
307 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
308 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100309 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200310 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400311 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100312 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100313 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500314 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200315};
316
317#define CFQ_CFQQ_FNS(name) \
318static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
319{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100320 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200321} \
322static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
323{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100324 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200325} \
326static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
327{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100328 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200329}
330
331CFQ_CFQQ_FNS(on_rr);
332CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200333CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200334CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200335CFQ_CFQQ_FNS(fifo_expire);
336CFQ_CFQQ_FNS(idle_window);
337CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100338CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200339CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200340CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100341CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100342CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500343CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200344#undef CFQ_CFQQ_FNS
345
Vivek Goyal2868ef72009-12-03 12:59:48 -0500346#ifdef CONFIG_DEBUG_CFQ_IOSCHED
347#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
348 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
349 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
350 blkg_path(&(cfqq)->cfqg->blkg), ##args);
351
352#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
353 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
354 blkg_path(&(cfqg)->blkg), ##args); \
355
356#else
Jens Axboe7b679132008-05-30 12:23:07 +0200357#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
358 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500359#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
360#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200361#define cfq_log(cfqd, fmt, args...) \
362 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
363
Vivek Goyal615f0252009-12-03 12:59:39 -0500364/* Traverses through cfq group service trees */
365#define for_each_cfqg_st(cfqg, i, j, st) \
366 for (i = 0; i <= IDLE_WORKLOAD; i++) \
367 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
368 : &cfqg->service_tree_idle; \
369 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
370 (i == IDLE_WORKLOAD && j == 0); \
371 j++, st = i < IDLE_WORKLOAD ? \
372 &cfqg->service_trees[i][j]: NULL) \
373
374
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100375static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
376{
377 if (cfq_class_idle(cfqq))
378 return IDLE_WORKLOAD;
379 if (cfq_class_rt(cfqq))
380 return RT_WORKLOAD;
381 return BE_WORKLOAD;
382}
383
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100384
385static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
386{
387 if (!cfq_cfqq_sync(cfqq))
388 return ASYNC_WORKLOAD;
389 if (!cfq_cfqq_idle_window(cfqq))
390 return SYNC_NOIDLE_WORKLOAD;
391 return SYNC_WORKLOAD;
392}
393
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500394static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
395 struct cfq_data *cfqd,
396 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100397{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500398 if (wl == IDLE_WORKLOAD)
399 return cfqg->service_tree_idle.count;
400
401 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
402 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
403 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100404}
405
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500406static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
407 struct cfq_group *cfqg)
408{
409 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
410 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
411}
412
Jens Axboe165125e2007-07-24 09:28:11 +0200413static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200414static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100415 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100416static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200417 struct io_context *);
418
419static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200420 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200421{
Jens Axboea6151c32009-10-07 20:02:57 +0200422 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200423}
424
425static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200426 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200427{
Jens Axboea6151c32009-10-07 20:02:57 +0200428 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200429}
430
431/*
432 * We regard a request as SYNC, if it's either a read or has the SYNC bit
433 * set (in which case it could also be direct WRITE).
434 */
Jens Axboea6151c32009-10-07 20:02:57 +0200435static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200436{
Jens Axboea6151c32009-10-07 20:02:57 +0200437 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200438}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700441 * scheduler run of queue, if there are requests pending and no one in the
442 * driver that will restart queueing
443 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200444static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700445{
Jens Axboe7b679132008-05-30 12:23:07 +0200446 if (cfqd->busy_queues) {
447 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200448 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200449 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700450}
451
Jens Axboe165125e2007-07-24 09:28:11 +0200452static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700453{
454 struct cfq_data *cfqd = q->elevator->elevator_data;
455
Vivek Goyalf04a6422009-12-03 12:59:40 -0500456 return !cfqd->rq_queued;
Andrew Morton99f95e52005-06-27 20:14:05 -0700457}
458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100460 * Scale schedule slice based on io priority. Use the sync time slice only
461 * if a queue is marked sync and has sync io queued. A sync queue with async
462 * io only, should not get full sync slice length.
463 */
Jens Axboea6151c32009-10-07 20:02:57 +0200464static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200465 unsigned short prio)
466{
467 const int base_slice = cfqd->cfq_slice[sync];
468
469 WARN_ON(prio >= IOPRIO_BE_NR);
470
471 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
472}
473
Jens Axboe44f7c162007-01-19 11:51:58 +1100474static inline int
475cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
476{
Jens Axboed9e76202007-04-20 14:27:50 +0200477 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100478}
479
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500480static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
481{
482 u64 d = delta << CFQ_SERVICE_SHIFT;
483
484 d = d * BLKIO_WEIGHT_DEFAULT;
485 do_div(d, cfqg->weight);
486 return d;
487}
488
489static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
490{
491 s64 delta = (s64)(vdisktime - min_vdisktime);
492 if (delta > 0)
493 min_vdisktime = vdisktime;
494
495 return min_vdisktime;
496}
497
498static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
499{
500 s64 delta = (s64)(vdisktime - min_vdisktime);
501 if (delta < 0)
502 min_vdisktime = vdisktime;
503
504 return min_vdisktime;
505}
506
507static void update_min_vdisktime(struct cfq_rb_root *st)
508{
509 u64 vdisktime = st->min_vdisktime;
510 struct cfq_group *cfqg;
511
512 if (st->active) {
513 cfqg = rb_entry_cfqg(st->active);
514 vdisktime = cfqg->vdisktime;
515 }
516
517 if (st->left) {
518 cfqg = rb_entry_cfqg(st->left);
519 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
520 }
521
522 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
523}
524
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100525/*
526 * get averaged number of queues of RT/BE priority.
527 * average is updated, with a formula that gives more weight to higher numbers,
528 * to quickly follows sudden increases and decrease slowly
529 */
530
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500531static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
532 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100533{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100534 unsigned min_q, max_q;
535 unsigned mult = cfq_hist_divisor - 1;
536 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500537 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100538
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500539 min_q = min(cfqg->busy_queues_avg[rt], busy);
540 max_q = max(cfqg->busy_queues_avg[rt], busy);
541 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100542 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500543 return cfqg->busy_queues_avg[rt];
544}
545
546static inline unsigned
547cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
548{
549 struct cfq_rb_root *st = &cfqd->grp_service_tree;
550
551 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100552}
553
Jens Axboe44f7c162007-01-19 11:51:58 +1100554static inline void
555cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
556{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100557 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
558 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500559 /*
560 * interested queues (we consider only the ones with the same
561 * priority class in the cfq group)
562 */
563 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
564 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100565 unsigned sync_slice = cfqd->cfq_slice[1];
566 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500567 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
568
569 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100570 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
571 /* scale low_slice according to IO priority
572 * and sync vs async */
573 unsigned low_slice =
574 min(slice, base_low_slice * slice / sync_slice);
575 /* the adapted slice value is scaled to fit all iqs
576 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500577 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100578 low_slice);
579 }
580 }
Vivek Goyaldae739e2009-12-03 12:59:45 -0500581 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100582 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500583 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200584 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100585}
586
587/*
588 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
589 * isn't valid until the first request from the dispatch is activated
590 * and the slice time set.
591 */
Jens Axboea6151c32009-10-07 20:02:57 +0200592static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100593{
594 if (cfq_cfqq_slice_new(cfqq))
595 return 0;
596 if (time_before(jiffies, cfqq->slice_end))
597 return 0;
598
599 return 1;
600}
601
602/*
Jens Axboe5e705372006-07-13 12:39:25 +0200603 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200605 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 */
Jens Axboe5e705372006-07-13 12:39:25 +0200607static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100608cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100610 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200612#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
613#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
614 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615
Jens Axboe5e705372006-07-13 12:39:25 +0200616 if (rq1 == NULL || rq1 == rq2)
617 return rq2;
618 if (rq2 == NULL)
619 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200620
Jens Axboe5e705372006-07-13 12:39:25 +0200621 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
622 return rq1;
623 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
624 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200625 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
626 return rq1;
627 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
628 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
Tejun Heo83096eb2009-05-07 22:24:39 +0900630 s1 = blk_rq_pos(rq1);
631 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633 /*
634 * by definition, 1KiB is 2 sectors
635 */
636 back_max = cfqd->cfq_back_max * 2;
637
638 /*
639 * Strict one way elevator _except_ in the case where we allow
640 * short backward seeks which are biased as twice the cost of a
641 * similar forward seek.
642 */
643 if (s1 >= last)
644 d1 = s1 - last;
645 else if (s1 + back_max >= last)
646 d1 = (last - s1) * cfqd->cfq_back_penalty;
647 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200648 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649
650 if (s2 >= last)
651 d2 = s2 - last;
652 else if (s2 + back_max >= last)
653 d2 = (last - s2) * cfqd->cfq_back_penalty;
654 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200655 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656
657 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658
Andreas Mohre8a99052006-03-28 08:59:49 +0200659 /*
660 * By doing switch() on the bit mask "wrap" we avoid having to
661 * check two variables for all permutations: --> faster!
662 */
663 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200664 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200665 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200666 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200667 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200668 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200669 else {
670 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200671 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200672 else
Jens Axboe5e705372006-07-13 12:39:25 +0200673 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200674 }
675
676 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200677 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200678 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200679 return rq2;
680 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200681 default:
682 /*
683 * Since both rqs are wrapped,
684 * start with the one that's further behind head
685 * (--> only *one* back seek required),
686 * since back seek takes more time than forward.
687 */
688 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200689 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 else
Jens Axboe5e705372006-07-13 12:39:25 +0200691 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 }
693}
694
Jens Axboe498d3aa22007-04-26 12:54:48 +0200695/*
696 * The below is leftmost cache rbtree addon
697 */
Jens Axboe08717142008-01-28 11:38:15 +0100698static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200699{
Vivek Goyal615f0252009-12-03 12:59:39 -0500700 /* Service tree is empty */
701 if (!root->count)
702 return NULL;
703
Jens Axboecc09e292007-04-26 12:53:50 +0200704 if (!root->left)
705 root->left = rb_first(&root->rb);
706
Jens Axboe08717142008-01-28 11:38:15 +0100707 if (root->left)
708 return rb_entry(root->left, struct cfq_queue, rb_node);
709
710 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200711}
712
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500713static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
714{
715 if (!root->left)
716 root->left = rb_first(&root->rb);
717
718 if (root->left)
719 return rb_entry_cfqg(root->left);
720
721 return NULL;
722}
723
Jens Axboea36e71f2009-04-15 12:15:11 +0200724static void rb_erase_init(struct rb_node *n, struct rb_root *root)
725{
726 rb_erase(n, root);
727 RB_CLEAR_NODE(n);
728}
729
Jens Axboecc09e292007-04-26 12:53:50 +0200730static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
731{
732 if (root->left == n)
733 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200734 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100735 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200736}
737
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738/*
739 * would be nice to take fifo expire time into account as well
740 */
Jens Axboe5e705372006-07-13 12:39:25 +0200741static struct request *
742cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
743 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744{
Jens Axboe21183b02006-07-13 12:33:14 +0200745 struct rb_node *rbnext = rb_next(&last->rb_node);
746 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200747 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748
Jens Axboe21183b02006-07-13 12:33:14 +0200749 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750
751 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200752 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200753
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200755 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200756 else {
757 rbnext = rb_first(&cfqq->sort_list);
758 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200759 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200760 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100762 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763}
764
Jens Axboed9e76202007-04-20 14:27:50 +0200765static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
766 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767{
Jens Axboed9e76202007-04-20 14:27:50 +0200768 /*
769 * just an approximation, should be ok.
770 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500771 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100772 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200773}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500775static inline s64
776cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
777{
778 return cfqg->vdisktime - st->min_vdisktime;
779}
780
781static void
782__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
783{
784 struct rb_node **node = &st->rb.rb_node;
785 struct rb_node *parent = NULL;
786 struct cfq_group *__cfqg;
787 s64 key = cfqg_key(st, cfqg);
788 int left = 1;
789
790 while (*node != NULL) {
791 parent = *node;
792 __cfqg = rb_entry_cfqg(parent);
793
794 if (key < cfqg_key(st, __cfqg))
795 node = &parent->rb_left;
796 else {
797 node = &parent->rb_right;
798 left = 0;
799 }
800 }
801
802 if (left)
803 st->left = &cfqg->rb_node;
804
805 rb_link_node(&cfqg->rb_node, parent, node);
806 rb_insert_color(&cfqg->rb_node, &st->rb);
807}
808
809static void
810cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
811{
812 struct cfq_rb_root *st = &cfqd->grp_service_tree;
813 struct cfq_group *__cfqg;
814 struct rb_node *n;
815
816 cfqg->nr_cfqq++;
817 if (cfqg->on_st)
818 return;
819
820 /*
821 * Currently put the group at the end. Later implement something
822 * so that groups get lesser vtime based on their weights, so that
823 * if group does not loose all if it was not continously backlogged.
824 */
825 n = rb_last(&st->rb);
826 if (n) {
827 __cfqg = rb_entry_cfqg(n);
828 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
829 } else
830 cfqg->vdisktime = st->min_vdisktime;
831
832 __cfq_group_service_tree_add(st, cfqg);
833 cfqg->on_st = true;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500834 st->total_weight += cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500835}
836
837static void
838cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
839{
840 struct cfq_rb_root *st = &cfqd->grp_service_tree;
841
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500842 if (st->active == &cfqg->rb_node)
843 st->active = NULL;
844
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500845 BUG_ON(cfqg->nr_cfqq < 1);
846 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500847
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500848 /* If there are other cfq queues under this group, don't delete it */
849 if (cfqg->nr_cfqq)
850 return;
851
Vivek Goyal2868ef72009-12-03 12:59:48 -0500852 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500853 cfqg->on_st = false;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500854 st->total_weight -= cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500855 if (!RB_EMPTY_NODE(&cfqg->rb_node))
856 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500857 cfqg->saved_workload_slice = 0;
Divyesh Shah91952912010-04-01 15:01:41 -0700858 blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500859}
860
861static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
862{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500863 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500864
865 /*
866 * Queue got expired before even a single request completed or
867 * got expired immediately after first request completion.
868 */
869 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
870 /*
871 * Also charge the seek time incurred to the group, otherwise
872 * if there are mutiple queues in the group, each can dispatch
873 * a single request on seeky media and cause lots of seek time
874 * and group will never know it.
875 */
876 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
877 1);
878 } else {
879 slice_used = jiffies - cfqq->slice_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500880 if (slice_used > cfqq->allocated_slice)
881 slice_used = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500882 }
883
Divyesh Shah9a0785b2010-04-01 15:01:04 -0700884 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500885 return slice_used;
886}
887
888static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
889 struct cfq_queue *cfqq)
890{
891 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500892 unsigned int used_sl, charge_sl;
893 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
894 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500895
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500896 BUG_ON(nr_sync < 0);
897 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
898
899 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
900 charge_sl = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500901
902 /* Can't update vdisktime while group is on service tree */
903 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500904 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500905 __cfq_group_service_tree_add(st, cfqg);
906
907 /* This group is being expired. Save the context */
908 if (time_after(cfqd->workload_expires, jiffies)) {
909 cfqg->saved_workload_slice = cfqd->workload_expires
910 - jiffies;
911 cfqg->saved_workload = cfqd->serving_type;
912 cfqg->saved_serving_prio = cfqd->serving_prio;
913 } else
914 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500915
916 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
917 st->min_vdisktime);
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700918 blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500919}
920
Vivek Goyal25fb5162009-12-03 12:59:46 -0500921#ifdef CONFIG_CFQ_GROUP_IOSCHED
922static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
923{
924 if (blkg)
925 return container_of(blkg, struct cfq_group, blkg);
926 return NULL;
927}
928
Vivek Goyalf8d461d2009-12-03 12:59:52 -0500929void
930cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
931{
932 cfqg_of_blkg(blkg)->weight = weight;
933}
934
Vivek Goyal25fb5162009-12-03 12:59:46 -0500935static struct cfq_group *
936cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
937{
938 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
939 struct cfq_group *cfqg = NULL;
940 void *key = cfqd;
941 int i, j;
942 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -0500943 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
944 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500945
Vivek Goyal25fb5162009-12-03 12:59:46 -0500946 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
947 if (cfqg || !create)
948 goto done;
949
950 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
951 if (!cfqg)
952 goto done;
953
954 cfqg->weight = blkcg->weight;
955 for_each_cfqg_st(cfqg, i, j, st)
956 *st = CFQ_RB_ROOT;
957 RB_CLEAR_NODE(&cfqg->rb_node);
Divyesh Shah84c124d2010-04-09 08:31:19 +0200958 blkio_group_init(&cfqg->blkg);
Vivek Goyal25fb5162009-12-03 12:59:46 -0500959
Vivek Goyalb1c35762009-12-03 12:59:47 -0500960 /*
961 * Take the initial reference that will be released on destroy
962 * This can be thought of a joint reference by cgroup and
963 * elevator which will be dropped by either elevator exit
964 * or cgroup deletion path depending on who is exiting first.
965 */
966 atomic_set(&cfqg->ref, 1);
967
Vivek Goyal25fb5162009-12-03 12:59:46 -0500968 /* Add group onto cgroup list */
Vivek Goyal22084192009-12-03 12:59:49 -0500969 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
970 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
971 MKDEV(major, minor));
Vivek Goyal25fb5162009-12-03 12:59:46 -0500972
973 /* Add group on cfqd list */
974 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
975
976done:
Vivek Goyal25fb5162009-12-03 12:59:46 -0500977 return cfqg;
978}
979
980/*
981 * Search for the cfq group current task belongs to. If create = 1, then also
982 * create the cfq group if it does not exist. request_queue lock must be held.
983 */
984static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
985{
986 struct cgroup *cgroup;
987 struct cfq_group *cfqg = NULL;
988
989 rcu_read_lock();
990 cgroup = task_cgroup(current, blkio_subsys_id);
991 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
992 if (!cfqg && create)
993 cfqg = &cfqd->root_group;
994 rcu_read_unlock();
995 return cfqg;
996}
997
998static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
999{
1000 /* Currently, all async queues are mapped to root group */
1001 if (!cfq_cfqq_sync(cfqq))
1002 cfqg = &cfqq->cfqd->root_group;
1003
1004 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001005 /* cfqq reference on cfqg */
1006 atomic_inc(&cfqq->cfqg->ref);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001007}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001008
1009static void cfq_put_cfqg(struct cfq_group *cfqg)
1010{
1011 struct cfq_rb_root *st;
1012 int i, j;
1013
1014 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1015 if (!atomic_dec_and_test(&cfqg->ref))
1016 return;
1017 for_each_cfqg_st(cfqg, i, j, st)
1018 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1019 kfree(cfqg);
1020}
1021
1022static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1023{
1024 /* Something wrong if we are trying to remove same group twice */
1025 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1026
1027 hlist_del_init(&cfqg->cfqd_node);
1028
1029 /*
1030 * Put the reference taken at the time of creation so that when all
1031 * queues are gone, group can be destroyed.
1032 */
1033 cfq_put_cfqg(cfqg);
1034}
1035
1036static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1037{
1038 struct hlist_node *pos, *n;
1039 struct cfq_group *cfqg;
1040
1041 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1042 /*
1043 * If cgroup removal path got to blk_group first and removed
1044 * it from cgroup list, then it will take care of destroying
1045 * cfqg also.
1046 */
1047 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1048 cfq_destroy_cfqg(cfqd, cfqg);
1049 }
1050}
1051
1052/*
1053 * Blk cgroup controller notification saying that blkio_group object is being
1054 * delinked as associated cgroup object is going away. That also means that
1055 * no new IO will come in this group. So get rid of this group as soon as
1056 * any pending IO in the group is finished.
1057 *
1058 * This function is called under rcu_read_lock(). key is the rcu protected
1059 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1060 * read lock.
1061 *
1062 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1063 * it should not be NULL as even if elevator was exiting, cgroup deltion
1064 * path got to it first.
1065 */
1066void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1067{
1068 unsigned long flags;
1069 struct cfq_data *cfqd = key;
1070
1071 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1072 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1073 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1074}
1075
Vivek Goyal25fb5162009-12-03 12:59:46 -05001076#else /* GROUP_IOSCHED */
1077static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1078{
1079 return &cfqd->root_group;
1080}
1081static inline void
1082cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1083 cfqq->cfqg = cfqg;
1084}
1085
Vivek Goyalb1c35762009-12-03 12:59:47 -05001086static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1087static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1088
Vivek Goyal25fb5162009-12-03 12:59:46 -05001089#endif /* GROUP_IOSCHED */
1090
Jens Axboe498d3aa22007-04-26 12:54:48 +02001091/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001092 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001093 * requests waiting to be processed. It is sorted in the order that
1094 * we will service the queues.
1095 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001096static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001097 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001098{
Jens Axboe08717142008-01-28 11:38:15 +01001099 struct rb_node **p, *parent;
1100 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001101 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001102 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001103 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001104 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001105 int group_changed = 0;
1106
1107#ifdef CONFIG_CFQ_GROUP_IOSCHED
1108 if (!cfqd->cfq_group_isolation
1109 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1110 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1111 /* Move this cfq to root group */
1112 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1113 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1114 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1115 cfqq->orig_cfqg = cfqq->cfqg;
1116 cfqq->cfqg = &cfqd->root_group;
1117 atomic_inc(&cfqd->root_group.ref);
1118 group_changed = 1;
1119 } else if (!cfqd->cfq_group_isolation
1120 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1121 /* cfqq is sequential now needs to go to its original group */
1122 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1123 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1124 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1125 cfq_put_cfqg(cfqq->cfqg);
1126 cfqq->cfqg = cfqq->orig_cfqg;
1127 cfqq->orig_cfqg = NULL;
1128 group_changed = 1;
1129 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1130 }
1131#endif
Jens Axboed9e76202007-04-20 14:27:50 +02001132
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001133 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001134 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001135 if (cfq_class_idle(cfqq)) {
1136 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001137 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001138 if (parent && parent != &cfqq->rb_node) {
1139 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1140 rb_key += __cfqq->rb_key;
1141 } else
1142 rb_key += jiffies;
1143 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001144 /*
1145 * Get our rb key offset. Subtract any residual slice
1146 * value carried from last service. A negative resid
1147 * count indicates slice overrun, and this should position
1148 * the next service time further away in the tree.
1149 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001150 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001151 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001152 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001153 } else {
1154 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001155 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001156 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1157 }
Jens Axboed9e76202007-04-20 14:27:50 +02001158
1159 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001160 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001161 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001162 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001163 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001164 if (rb_key == cfqq->rb_key &&
1165 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001166 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001167
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001168 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1169 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001170 }
Jens Axboed9e76202007-04-20 14:27:50 +02001171
Jens Axboe498d3aa22007-04-26 12:54:48 +02001172 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001173 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001174 cfqq->service_tree = service_tree;
1175 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001176 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001177 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001178
Jens Axboed9e76202007-04-20 14:27:50 +02001179 parent = *p;
1180 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1181
Jens Axboe0c534e02007-04-18 20:01:57 +02001182 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001183 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001184 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001185 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001186 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001187 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001188 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001189 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001190 }
Jens Axboe67060e32007-04-18 20:13:32 +02001191
1192 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001193 }
1194
Jens Axboecc09e292007-04-26 12:53:50 +02001195 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001196 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001197
Jens Axboed9e76202007-04-20 14:27:50 +02001198 cfqq->rb_key = rb_key;
1199 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001200 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1201 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001202 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001203 return;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001204 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205}
1206
Jens Axboea36e71f2009-04-15 12:15:11 +02001207static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001208cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1209 sector_t sector, struct rb_node **ret_parent,
1210 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001211{
Jens Axboea36e71f2009-04-15 12:15:11 +02001212 struct rb_node **p, *parent;
1213 struct cfq_queue *cfqq = NULL;
1214
1215 parent = NULL;
1216 p = &root->rb_node;
1217 while (*p) {
1218 struct rb_node **n;
1219
1220 parent = *p;
1221 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1222
1223 /*
1224 * Sort strictly based on sector. Smallest to the left,
1225 * largest to the right.
1226 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001227 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001228 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001229 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001230 n = &(*p)->rb_left;
1231 else
1232 break;
1233 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001234 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001235 }
1236
1237 *ret_parent = parent;
1238 if (rb_link)
1239 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001240 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001241}
1242
1243static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1244{
Jens Axboea36e71f2009-04-15 12:15:11 +02001245 struct rb_node **p, *parent;
1246 struct cfq_queue *__cfqq;
1247
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001248 if (cfqq->p_root) {
1249 rb_erase(&cfqq->p_node, cfqq->p_root);
1250 cfqq->p_root = NULL;
1251 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001252
1253 if (cfq_class_idle(cfqq))
1254 return;
1255 if (!cfqq->next_rq)
1256 return;
1257
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001258 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001259 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1260 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001261 if (!__cfqq) {
1262 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001263 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1264 } else
1265 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001266}
1267
Jens Axboe498d3aa22007-04-26 12:54:48 +02001268/*
1269 * Update cfqq's position in the service tree.
1270 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001271static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001272{
Jens Axboe6d048f52007-04-25 12:44:27 +02001273 /*
1274 * Resorting requires the cfqq to be on the RR list already.
1275 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001276 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001277 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001278 cfq_prio_tree_add(cfqd, cfqq);
1279 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001280}
1281
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282/*
1283 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001284 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 */
Jens Axboefebffd62008-01-28 13:19:43 +01001286static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287{
Jens Axboe7b679132008-05-30 12:23:07 +02001288 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001289 BUG_ON(cfq_cfqq_on_rr(cfqq));
1290 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 cfqd->busy_queues++;
1292
Jens Axboeedd75ff2007-04-19 12:03:34 +02001293 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294}
1295
Jens Axboe498d3aa22007-04-26 12:54:48 +02001296/*
1297 * Called when the cfqq no longer has requests pending, remove it from
1298 * the service tree.
1299 */
Jens Axboefebffd62008-01-28 13:19:43 +01001300static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301{
Jens Axboe7b679132008-05-30 12:23:07 +02001302 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001303 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1304 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001306 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1307 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1308 cfqq->service_tree = NULL;
1309 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001310 if (cfqq->p_root) {
1311 rb_erase(&cfqq->p_node, cfqq->p_root);
1312 cfqq->p_root = NULL;
1313 }
Jens Axboed9e76202007-04-20 14:27:50 +02001314
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001315 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 BUG_ON(!cfqd->busy_queues);
1317 cfqd->busy_queues--;
1318}
1319
1320/*
1321 * rb tree support functions
1322 */
Jens Axboefebffd62008-01-28 13:19:43 +01001323static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324{
Jens Axboe5e705372006-07-13 12:39:25 +02001325 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001326 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327
Jens Axboeb4878f22005-10-20 16:42:29 +02001328 BUG_ON(!cfqq->queued[sync]);
1329 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330
Jens Axboe5e705372006-07-13 12:39:25 +02001331 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332
Vivek Goyalf04a6422009-12-03 12:59:40 -05001333 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1334 /*
1335 * Queue will be deleted from service tree when we actually
1336 * expire it later. Right now just remove it from prio tree
1337 * as it is empty.
1338 */
1339 if (cfqq->p_root) {
1340 rb_erase(&cfqq->p_node, cfqq->p_root);
1341 cfqq->p_root = NULL;
1342 }
1343 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344}
1345
Jens Axboe5e705372006-07-13 12:39:25 +02001346static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347{
Jens Axboe5e705372006-07-13 12:39:25 +02001348 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001350 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351
Jens Axboe5380a102006-07-13 12:37:56 +02001352 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353
1354 /*
1355 * looks a little odd, but the first insert might return an alias.
1356 * if that happens, put the alias on the dispatch list
1357 */
Jens Axboe21183b02006-07-13 12:33:14 +02001358 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001359 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001360
1361 if (!cfq_cfqq_on_rr(cfqq))
1362 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001363
1364 /*
1365 * check if this request is a better next-serve candidate
1366 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001367 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001368 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001369
1370 /*
1371 * adjust priority tree position, if ->next_rq changes
1372 */
1373 if (prev != cfqq->next_rq)
1374 cfq_prio_tree_add(cfqd, cfqq);
1375
Jens Axboe5044eed2007-04-25 11:53:48 +02001376 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377}
1378
Jens Axboefebffd62008-01-28 13:19:43 +01001379static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380{
Jens Axboe5380a102006-07-13 12:37:56 +02001381 elv_rb_del(&cfqq->sort_list, rq);
1382 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +02001383 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384}
1385
Jens Axboe206dc692006-03-28 13:03:44 +02001386static struct request *
1387cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388{
Jens Axboe206dc692006-03-28 13:03:44 +02001389 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001390 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001391 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392
Jens Axboe4ac845a2008-01-24 08:44:49 +01001393 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001394 if (!cic)
1395 return NULL;
1396
1397 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001398 if (cfqq) {
1399 sector_t sector = bio->bi_sector + bio_sectors(bio);
1400
Jens Axboe21183b02006-07-13 12:33:14 +02001401 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001402 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 return NULL;
1405}
1406
Jens Axboe165125e2007-07-24 09:28:11 +02001407static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001408{
1409 struct cfq_data *cfqd = q->elevator->elevator_data;
1410
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001411 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001412 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001413 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001414
Tejun Heo5b936292009-05-07 22:24:38 +09001415 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001416}
1417
Jens Axboe165125e2007-07-24 09:28:11 +02001418static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419{
Jens Axboe22e2c502005-06-27 10:55:12 +02001420 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001422 WARN_ON(!cfqd->rq_in_driver);
1423 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001424 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001425 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426}
1427
Jens Axboeb4878f22005-10-20 16:42:29 +02001428static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429{
Jens Axboe5e705372006-07-13 12:39:25 +02001430 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001431
Jens Axboe5e705372006-07-13 12:39:25 +02001432 if (cfqq->next_rq == rq)
1433 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
Jens Axboeb4878f22005-10-20 16:42:29 +02001435 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001436 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001437
Aaron Carroll45333d52008-08-26 15:52:36 +02001438 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +02001439 if (rq_is_meta(rq)) {
1440 WARN_ON(!cfqq->meta_pending);
1441 cfqq->meta_pending--;
1442 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443}
1444
Jens Axboe165125e2007-07-24 09:28:11 +02001445static int cfq_merge(struct request_queue *q, struct request **req,
1446 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447{
1448 struct cfq_data *cfqd = q->elevator->elevator_data;
1449 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450
Jens Axboe206dc692006-03-28 13:03:44 +02001451 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001452 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001453 *req = __rq;
1454 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455 }
1456
1457 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458}
1459
Jens Axboe165125e2007-07-24 09:28:11 +02001460static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001461 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462{
Jens Axboe21183b02006-07-13 12:33:14 +02001463 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001464 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465
Jens Axboe5e705372006-07-13 12:39:25 +02001466 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468}
1469
1470static void
Jens Axboe165125e2007-07-24 09:28:11 +02001471cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 struct request *next)
1473{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001474 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001475 /*
1476 * reposition in fifo if next is older than rq
1477 */
1478 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001479 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001480 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001481 rq_set_fifo_time(rq, rq_fifo_time(next));
1482 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001483
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001484 if (cfqq->next_rq == next)
1485 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001486 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +02001487}
1488
Jens Axboe165125e2007-07-24 09:28:11 +02001489static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001490 struct bio *bio)
1491{
1492 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001493 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001494 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001495
1496 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001497 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001498 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001499 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001500 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001501
1502 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001503 * Lookup the cfqq that this bio will be queued with. Allow
1504 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001505 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001506 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001507 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001508 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001509
Vasily Tarasov91fac312007-04-25 12:29:51 +02001510 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001511 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001512}
1513
Jens Axboefebffd62008-01-28 13:19:43 +01001514static void __cfq_set_active_queue(struct cfq_data *cfqd,
1515 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001516{
1517 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001518 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1519 cfqd->serving_prio, cfqd->serving_type);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001520 cfqq->slice_start = 0;
1521 cfqq->dispatch_start = jiffies;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001522 cfqq->allocated_slice = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001523 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001524 cfqq->slice_dispatch = 0;
1525
Jens Axboe2f5cb732009-04-07 08:51:19 +02001526 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +02001527 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001528 cfq_clear_cfqq_must_alloc_slice(cfqq);
1529 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +11001530 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001531
1532 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +02001533 }
1534
1535 cfqd->active_queue = cfqq;
1536}
1537
1538/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001539 * current cfqq expired its slice (or was too idle), select new one
1540 */
1541static void
1542__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001543 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001544{
Jens Axboe7b679132008-05-30 12:23:07 +02001545 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1546
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001547 if (cfq_cfqq_wait_request(cfqq))
1548 del_timer(&cfqd->idle_slice_timer);
1549
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001550 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001551 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001552
1553 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001554 * If this cfqq is shared between multiple processes, check to
1555 * make sure that those processes are still issuing I/Os within
1556 * the mean seek distance. If not, it may be time to break the
1557 * queues apart again.
1558 */
1559 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1560 cfq_mark_cfqq_split_coop(cfqq);
1561
1562 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001563 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001564 */
Jens Axboe7b679132008-05-30 12:23:07 +02001565 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +11001566 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001567 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1568 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001569
Vivek Goyaldae739e2009-12-03 12:59:45 -05001570 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1571
Vivek Goyalf04a6422009-12-03 12:59:40 -05001572 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1573 cfq_del_cfqq_rr(cfqd, cfqq);
1574
Jens Axboeedd75ff2007-04-19 12:03:34 +02001575 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001576
1577 if (cfqq == cfqd->active_queue)
1578 cfqd->active_queue = NULL;
1579
Vivek Goyaldae739e2009-12-03 12:59:45 -05001580 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1581 cfqd->grp_service_tree.active = NULL;
1582
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001583 if (cfqd->active_cic) {
1584 put_io_context(cfqd->active_cic->ioc);
1585 cfqd->active_cic = NULL;
1586 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001587}
1588
Jens Axboea6151c32009-10-07 20:02:57 +02001589static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001590{
1591 struct cfq_queue *cfqq = cfqd->active_queue;
1592
1593 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +02001594 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001595}
1596
Jens Axboe498d3aa22007-04-26 12:54:48 +02001597/*
1598 * Get next queue for service. Unless we have a queue preemption,
1599 * we'll simply select the first cfqq in the service tree.
1600 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001601static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001602{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001603 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001604 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001605 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001606
Vivek Goyalf04a6422009-12-03 12:59:40 -05001607 if (!cfqd->rq_queued)
1608 return NULL;
1609
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001610 /* There is nothing to dispatch */
1611 if (!service_tree)
1612 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001613 if (RB_EMPTY_ROOT(&service_tree->rb))
1614 return NULL;
1615 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001616}
1617
Vivek Goyalf04a6422009-12-03 12:59:40 -05001618static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1619{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001620 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001621 struct cfq_queue *cfqq;
1622 int i, j;
1623 struct cfq_rb_root *st;
1624
1625 if (!cfqd->rq_queued)
1626 return NULL;
1627
Vivek Goyal25fb5162009-12-03 12:59:46 -05001628 cfqg = cfq_get_next_cfqg(cfqd);
1629 if (!cfqg)
1630 return NULL;
1631
Vivek Goyalf04a6422009-12-03 12:59:40 -05001632 for_each_cfqg_st(cfqg, i, j, st)
1633 if ((cfqq = cfq_rb_first(st)) != NULL)
1634 return cfqq;
1635 return NULL;
1636}
1637
Jens Axboe498d3aa22007-04-26 12:54:48 +02001638/*
1639 * Get and set a new active queue for service.
1640 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001641static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1642 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001643{
Jens Axboee00ef792009-11-04 08:54:55 +01001644 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001645 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001646
Jens Axboe22e2c502005-06-27 10:55:12 +02001647 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001648 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001649}
1650
Jens Axboed9e76202007-04-20 14:27:50 +02001651static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1652 struct request *rq)
1653{
Tejun Heo83096eb2009-05-07 22:24:39 +09001654 if (blk_rq_pos(rq) >= cfqd->last_position)
1655 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001656 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001657 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001658}
1659
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001660static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001661 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001662{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001663 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001664}
1665
Jens Axboea36e71f2009-04-15 12:15:11 +02001666static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1667 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001668{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001669 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001670 struct rb_node *parent, *node;
1671 struct cfq_queue *__cfqq;
1672 sector_t sector = cfqd->last_position;
1673
1674 if (RB_EMPTY_ROOT(root))
1675 return NULL;
1676
1677 /*
1678 * First, if we find a request starting at the end of the last
1679 * request, choose it.
1680 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001681 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001682 if (__cfqq)
1683 return __cfqq;
1684
1685 /*
1686 * If the exact sector wasn't found, the parent of the NULL leaf
1687 * will contain the closest sector.
1688 */
1689 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001690 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001691 return __cfqq;
1692
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001693 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001694 node = rb_next(&__cfqq->p_node);
1695 else
1696 node = rb_prev(&__cfqq->p_node);
1697 if (!node)
1698 return NULL;
1699
1700 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001701 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001702 return __cfqq;
1703
1704 return NULL;
1705}
1706
1707/*
1708 * cfqd - obvious
1709 * cur_cfqq - passed in so that we don't decide that the current queue is
1710 * closely cooperating with itself.
1711 *
1712 * So, basically we're assuming that that cur_cfqq has dispatched at least
1713 * one request, and that cfqd->last_position reflects a position on the disk
1714 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1715 * assumption.
1716 */
1717static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001718 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001719{
1720 struct cfq_queue *cfqq;
1721
Divyesh Shah39c01b22010-03-25 15:45:57 +01001722 if (cfq_class_idle(cur_cfqq))
1723 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001724 if (!cfq_cfqq_sync(cur_cfqq))
1725 return NULL;
1726 if (CFQQ_SEEKY(cur_cfqq))
1727 return NULL;
1728
Jens Axboea36e71f2009-04-15 12:15:11 +02001729 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001730 * Don't search priority tree if it's the only queue in the group.
1731 */
1732 if (cur_cfqq->cfqg->nr_cfqq == 1)
1733 return NULL;
1734
1735 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001736 * We should notice if some of the queues are cooperating, eg
1737 * working closely on the same area of the disk. In that case,
1738 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001739 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001740 cfqq = cfqq_close(cfqd, cur_cfqq);
1741 if (!cfqq)
1742 return NULL;
1743
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001744 /* If new queue belongs to different cfq_group, don't choose it */
1745 if (cur_cfqq->cfqg != cfqq->cfqg)
1746 return NULL;
1747
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001748 /*
1749 * It only makes sense to merge sync queues.
1750 */
1751 if (!cfq_cfqq_sync(cfqq))
1752 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001753 if (CFQQ_SEEKY(cfqq))
1754 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001755
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001756 /*
1757 * Do not merge queues of different priority classes
1758 */
1759 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1760 return NULL;
1761
Jens Axboea36e71f2009-04-15 12:15:11 +02001762 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001763}
1764
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001765/*
1766 * Determine whether we should enforce idle window for this queue.
1767 */
1768
1769static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1770{
1771 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001772 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001773
Vivek Goyalf04a6422009-12-03 12:59:40 -05001774 BUG_ON(!service_tree);
1775 BUG_ON(!service_tree->count);
1776
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001777 /* We never do for idle class queues. */
1778 if (prio == IDLE_WORKLOAD)
1779 return false;
1780
1781 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001782 if (cfq_cfqq_idle_window(cfqq) &&
1783 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001784 return true;
1785
1786 /*
1787 * Otherwise, we do only if they are the last ones
1788 * in their service tree.
1789 */
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001790 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1791 return 1;
1792 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1793 service_tree->count);
1794 return 0;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001795}
1796
Jens Axboe6d048f52007-04-25 12:44:27 +02001797static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001798{
Jens Axboe17926692007-01-19 11:59:30 +11001799 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001800 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001801 unsigned long sl;
1802
Jens Axboea68bbddba2008-09-24 13:03:33 +02001803 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001804 * SSD device without seek penalty, disable idling. But only do so
1805 * for devices that support queuing, otherwise we still have a problem
1806 * with sync vs async workloads.
Jens Axboea68bbddba2008-09-24 13:03:33 +02001807 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001808 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbddba2008-09-24 13:03:33 +02001809 return;
1810
Jens Axboedd67d052006-06-21 09:36:18 +02001811 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001812 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001813
1814 /*
1815 * idle is disabled, either manually or by past process history
1816 */
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001817 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
Jens Axboe6d048f52007-04-25 12:44:27 +02001818 return;
1819
Jens Axboe22e2c502005-06-27 10:55:12 +02001820 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001821 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001822 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001823 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001824 return;
1825
1826 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001827 * task has exited, don't wait
1828 */
Jens Axboe206dc692006-03-28 13:03:44 +02001829 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001830 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001831 return;
1832
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001833 /*
1834 * If our average think time is larger than the remaining time
1835 * slice, then don't idle. This avoids overrunning the allotted
1836 * time slice.
1837 */
1838 if (sample_valid(cic->ttime_samples) &&
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001839 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1840 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1841 cic->ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001842 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001843 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001844
Jens Axboe3b181522005-06-27 10:56:24 +02001845 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001846
Jens Axboe6d048f52007-04-25 12:44:27 +02001847 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001848
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001849 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001850 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851}
1852
Jens Axboe498d3aa22007-04-26 12:54:48 +02001853/*
1854 * Move request from internal lists to the request queue dispatch list.
1855 */
Jens Axboe165125e2007-07-24 09:28:11 +02001856static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001858 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001859 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001860
Jens Axboe7b679132008-05-30 12:23:07 +02001861 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1862
Jeff Moyer06d21882009-09-11 17:08:59 +02001863 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001864 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001865 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001866 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001867
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001868 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Divyesh Shah84c124d2010-04-09 08:31:19 +02001869 blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1870 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871}
1872
1873/*
1874 * return expired entry, or NULL to just start from scratch in rbtree
1875 */
Jens Axboefebffd62008-01-28 13:19:43 +01001876static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877{
Jens Axboe30996f42009-10-05 11:03:39 +02001878 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879
Jens Axboe3b181522005-06-27 10:56:24 +02001880 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001882
1883 cfq_mark_cfqq_fifo_expire(cfqq);
1884
Jens Axboe89850f72006-07-22 16:48:31 +02001885 if (list_empty(&cfqq->fifo))
1886 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887
Jens Axboe89850f72006-07-22 16:48:31 +02001888 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02001889 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02001890 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891
Jens Axboe30996f42009-10-05 11:03:39 +02001892 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001893 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894}
1895
Jens Axboe22e2c502005-06-27 10:55:12 +02001896static inline int
1897cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1898{
1899 const int base_rq = cfqd->cfq_slice_async_rq;
1900
1901 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1902
1903 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1904}
1905
1906/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001907 * Must be called with the queue_lock held.
1908 */
1909static int cfqq_process_refs(struct cfq_queue *cfqq)
1910{
1911 int process_refs, io_refs;
1912
1913 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1914 process_refs = atomic_read(&cfqq->ref) - io_refs;
1915 BUG_ON(process_refs < 0);
1916 return process_refs;
1917}
1918
1919static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1920{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001921 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001922 struct cfq_queue *__cfqq;
1923
1924 /* Avoid a circular list and skip interim queue merges */
1925 while ((__cfqq = new_cfqq->new_cfqq)) {
1926 if (__cfqq == cfqq)
1927 return;
1928 new_cfqq = __cfqq;
1929 }
1930
1931 process_refs = cfqq_process_refs(cfqq);
1932 /*
1933 * If the process for the cfqq has gone away, there is no
1934 * sense in merging the queues.
1935 */
1936 if (process_refs == 0)
1937 return;
1938
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001939 /*
1940 * Merge in the direction of the lesser amount of work.
1941 */
1942 new_process_refs = cfqq_process_refs(new_cfqq);
1943 if (new_process_refs >= process_refs) {
1944 cfqq->new_cfqq = new_cfqq;
1945 atomic_add(process_refs, &new_cfqq->ref);
1946 } else {
1947 new_cfqq->new_cfqq = cfqq;
1948 atomic_add(new_process_refs, &cfqq->ref);
1949 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001950}
1951
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001952static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001953 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001954{
1955 struct cfq_queue *queue;
1956 int i;
1957 bool key_valid = false;
1958 unsigned long lowest_key = 0;
1959 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1960
Vivek Goyal65b32a52009-12-16 17:52:59 -05001961 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1962 /* select the one with lowest rb_key */
1963 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001964 if (queue &&
1965 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1966 lowest_key = queue->rb_key;
1967 cur_best = i;
1968 key_valid = true;
1969 }
1970 }
1971
1972 return cur_best;
1973}
1974
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001975static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001976{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001977 unsigned slice;
1978 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001979 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001980 unsigned group_slice;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001981
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001982 if (!cfqg) {
1983 cfqd->serving_prio = IDLE_WORKLOAD;
1984 cfqd->workload_expires = jiffies + 1;
1985 return;
1986 }
1987
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001988 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001989 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001990 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001991 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001992 cfqd->serving_prio = BE_WORKLOAD;
1993 else {
1994 cfqd->serving_prio = IDLE_WORKLOAD;
1995 cfqd->workload_expires = jiffies + 1;
1996 return;
1997 }
1998
1999 /*
2000 * For RT and BE, we have to choose also the type
2001 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2002 * expiration time
2003 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002004 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002005 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002006
2007 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002008 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002009 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002010 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002011 return;
2012
2013 /* otherwise select new workload type */
2014 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002015 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2016 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002017 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002018
2019 /*
2020 * the workload slice is computed as a fraction of target latency
2021 * proportional to the number of queues in that workload, over
2022 * all the queues in the same priority class
2023 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002024 group_slice = cfq_group_slice(cfqd, cfqg);
2025
2026 slice = group_slice * count /
2027 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2028 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002029
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002030 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2031 unsigned int tmp;
2032
2033 /*
2034 * Async queues are currently system wide. Just taking
2035 * proportion of queues with-in same group will lead to higher
2036 * async ratio system wide as generally root group is going
2037 * to have higher weight. A more accurate thing would be to
2038 * calculate system wide asnc/sync ratio.
2039 */
2040 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2041 tmp = tmp/cfqd->busy_queues;
2042 slice = min_t(unsigned, slice, tmp);
2043
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002044 /* async workload slice is scaled down according to
2045 * the sync/async slice ratio. */
2046 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002047 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002048 /* sync workload slice is at least 2 * cfq_slice_idle */
2049 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2050
2051 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002052 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002053 cfqd->workload_expires = jiffies + slice;
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002054 cfqd->noidle_tree_requires_idle = false;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002055}
2056
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002057static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2058{
2059 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002060 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002061
2062 if (RB_EMPTY_ROOT(&st->rb))
2063 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002064 cfqg = cfq_rb_first_group(st);
2065 st->active = &cfqg->rb_node;
2066 update_min_vdisktime(st);
2067 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002068}
2069
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002070static void cfq_choose_cfqg(struct cfq_data *cfqd)
2071{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002072 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2073
2074 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002075
2076 /* Restore the workload type data */
2077 if (cfqg->saved_workload_slice) {
2078 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2079 cfqd->serving_type = cfqg->saved_workload;
2080 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002081 } else
2082 cfqd->workload_expires = jiffies - 1;
2083
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002084 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002085}
2086
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002087/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002088 * Select a queue for service. If we have a current active queue,
2089 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002090 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002091static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002092{
Jens Axboea36e71f2009-04-15 12:15:11 +02002093 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002094
2095 cfqq = cfqd->active_queue;
2096 if (!cfqq)
2097 goto new_queue;
2098
Vivek Goyalf04a6422009-12-03 12:59:40 -05002099 if (!cfqd->rq_queued)
2100 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002101
2102 /*
2103 * We were waiting for group to get backlogged. Expire the queue
2104 */
2105 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2106 goto expire;
2107
Jens Axboe22e2c502005-06-27 10:55:12 +02002108 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002109 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002110 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002111 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2112 /*
2113 * If slice had not expired at the completion of last request
2114 * we might not have turned on wait_busy flag. Don't expire
2115 * the queue yet. Allow the group to get backlogged.
2116 *
2117 * The very fact that we have used the slice, that means we
2118 * have been idling all along on this queue and it should be
2119 * ok to wait for this request to complete.
2120 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002121 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2122 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2123 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002124 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002125 } else
Vivek Goyal7667aa02009-12-08 17:52:58 -05002126 goto expire;
2127 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002128
2129 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002130 * The active queue has requests and isn't expired, allow it to
2131 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002132 */
Jens Axboedd67d052006-06-21 09:36:18 +02002133 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002134 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002135
2136 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002137 * If another queue has a request waiting within our mean seek
2138 * distance, let it run. The expire code will check for close
2139 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002140 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002141 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002142 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002143 if (new_cfqq) {
2144 if (!cfqq->new_cfqq)
2145 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002146 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002147 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002148
2149 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002150 * No requests pending. If the active queue still has requests in
2151 * flight or is idling for a new request, allow either of these
2152 * conditions to happen (or time out) before selecting a new queue.
2153 */
Jens Axboecc197472007-04-20 20:45:39 +02002154 if (timer_pending(&cfqd->idle_slice_timer) ||
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002155 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002156 cfqq = NULL;
2157 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002158 }
2159
Jens Axboe3b181522005-06-27 10:56:24 +02002160expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002161 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002162new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002163 /*
2164 * Current queue expired. Check if we have to switch to a new
2165 * service tree
2166 */
2167 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002168 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002169
Jens Axboea36e71f2009-04-15 12:15:11 +02002170 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002171keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002172 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002173}
2174
Jens Axboefebffd62008-01-28 13:19:43 +01002175static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002176{
2177 int dispatched = 0;
2178
2179 while (cfqq->next_rq) {
2180 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2181 dispatched++;
2182 }
2183
2184 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002185
2186 /* By default cfqq is not expired if it is empty. Do it explicitly */
2187 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002188 return dispatched;
2189}
2190
Jens Axboe498d3aa22007-04-26 12:54:48 +02002191/*
2192 * Drain our current requests. Used for barriers and when switching
2193 * io schedulers on-the-fly.
2194 */
Jens Axboed9e76202007-04-20 14:27:50 +02002195static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002196{
Jens Axboe08717142008-01-28 11:38:15 +01002197 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002198 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002199
Vivek Goyalf04a6422009-12-03 12:59:40 -05002200 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2201 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002202
Jens Axboe6084cdd2007-04-23 08:25:00 +02002203 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002204 BUG_ON(cfqd->busy_queues);
2205
Jeff Moyer6923715a2009-06-12 15:29:30 +02002206 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002207 return dispatched;
2208}
2209
Shaohua Liabc3c742010-03-01 09:20:54 +01002210static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2211 struct cfq_queue *cfqq)
2212{
2213 /* the queue hasn't finished any request, can't estimate */
2214 if (cfq_cfqq_slice_new(cfqq))
2215 return 1;
2216 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2217 cfqq->slice_end))
2218 return 1;
2219
2220 return 0;
2221}
2222
Jens Axboe0b182d62009-10-06 20:49:37 +02002223static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002224{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002225 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226
Jens Axboe2f5cb732009-04-07 08:51:19 +02002227 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002228 * Drain async requests before we start sync IO
2229 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002230 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002231 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002232
2233 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002234 * If this is an async queue and we have sync IO in flight, let it wait
2235 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002236 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002237 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002238
Shaohua Liabc3c742010-03-01 09:20:54 +01002239 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002240 if (cfq_class_idle(cfqq))
2241 max_dispatch = 1;
2242
2243 /*
2244 * Does this cfqq already have too much IO in flight?
2245 */
2246 if (cfqq->dispatched >= max_dispatch) {
2247 /*
2248 * idle queue must always only have a single IO in flight
2249 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002250 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002251 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002252
Jens Axboe2f5cb732009-04-07 08:51:19 +02002253 /*
2254 * We have other queues, don't allow more IO from this one
2255 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002256 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002257 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002258
Jens Axboe2f5cb732009-04-07 08:51:19 +02002259 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002260 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002261 */
Shaohua Liabc3c742010-03-01 09:20:54 +01002262 if (cfqd->busy_queues == 1)
2263 max_dispatch = -1;
2264 else
2265 /*
2266 * Normally we start throttling cfqq when cfq_quantum/2
2267 * requests have been dispatched. But we can drive
2268 * deeper queue depths at the beginning of slice
2269 * subjected to upper limit of cfq_quantum.
2270 * */
2271 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002272 }
2273
2274 /*
2275 * Async queues must wait a bit before being allowed dispatch.
2276 * We also ramp up the dispatch depth gradually for async IO,
2277 * based on the last sync IO we serviced
2278 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002279 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002280 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002281 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002282
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002283 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002284 if (!depth && !cfqq->dispatched)
2285 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002286 if (depth < max_dispatch)
2287 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 }
2289
Jens Axboe0b182d62009-10-06 20:49:37 +02002290 /*
2291 * If we're below the current max, allow a dispatch
2292 */
2293 return cfqq->dispatched < max_dispatch;
2294}
2295
2296/*
2297 * Dispatch a request from cfqq, moving them to the request queue
2298 * dispatch list.
2299 */
2300static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2301{
2302 struct request *rq;
2303
2304 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2305
2306 if (!cfq_may_dispatch(cfqd, cfqq))
2307 return false;
2308
2309 /*
2310 * follow expired path, else get first next available
2311 */
2312 rq = cfq_check_fifo(cfqq);
2313 if (!rq)
2314 rq = cfqq->next_rq;
2315
2316 /*
2317 * insert request into driver dispatch list
2318 */
2319 cfq_dispatch_insert(cfqd->queue, rq);
2320
2321 if (!cfqd->active_cic) {
2322 struct cfq_io_context *cic = RQ_CIC(rq);
2323
2324 atomic_long_inc(&cic->ioc->refcount);
2325 cfqd->active_cic = cic;
2326 }
2327
2328 return true;
2329}
2330
2331/*
2332 * Find the cfqq that we need to service and move a request from that to the
2333 * dispatch list
2334 */
2335static int cfq_dispatch_requests(struct request_queue *q, int force)
2336{
2337 struct cfq_data *cfqd = q->elevator->elevator_data;
2338 struct cfq_queue *cfqq;
2339
2340 if (!cfqd->busy_queues)
2341 return 0;
2342
2343 if (unlikely(force))
2344 return cfq_forced_dispatch(cfqd);
2345
2346 cfqq = cfq_select_queue(cfqd);
2347 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002348 return 0;
2349
Jens Axboe2f5cb732009-04-07 08:51:19 +02002350 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002351 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002352 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002353 if (!cfq_dispatch_request(cfqd, cfqq))
2354 return 0;
2355
Jens Axboe2f5cb732009-04-07 08:51:19 +02002356 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002357 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002358
2359 /*
2360 * expire an async queue immediately if it has used up its slice. idle
2361 * queue always expire after 1 dispatch round.
2362 */
2363 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2364 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2365 cfq_class_idle(cfqq))) {
2366 cfqq->slice_end = jiffies + 1;
2367 cfq_slice_expired(cfqd, 0);
2368 }
2369
Shan Weib217a902009-09-01 10:06:42 +02002370 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002371 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372}
2373
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374/*
Jens Axboe5e705372006-07-13 12:39:25 +02002375 * task holds one reference to the queue, dropped when task exits. each rq
2376 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002378 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379 * queue lock must be held here.
2380 */
2381static void cfq_put_queue(struct cfq_queue *cfqq)
2382{
Jens Axboe22e2c502005-06-27 10:55:12 +02002383 struct cfq_data *cfqd = cfqq->cfqd;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002384 struct cfq_group *cfqg, *orig_cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002385
2386 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387
2388 if (!atomic_dec_and_test(&cfqq->ref))
2389 return;
2390
Jens Axboe7b679132008-05-30 12:23:07 +02002391 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002393 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002394 cfqg = cfqq->cfqg;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002395 orig_cfqg = cfqq->orig_cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002397 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02002398 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002399 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002400 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002401
Vivek Goyalf04a6422009-12-03 12:59:40 -05002402 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002404 cfq_put_cfqg(cfqg);
Vivek Goyal878eadd2009-12-07 19:37:15 +01002405 if (orig_cfqg)
2406 cfq_put_cfqg(orig_cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407}
2408
Jens Axboed6de8be2008-05-28 14:46:59 +02002409/*
2410 * Must always be called with the rcu_read_lock() held
2411 */
Jens Axboe07416d22008-05-07 09:17:12 +02002412static void
2413__call_for_each_cic(struct io_context *ioc,
2414 void (*func)(struct io_context *, struct cfq_io_context *))
2415{
2416 struct cfq_io_context *cic;
2417 struct hlist_node *n;
2418
2419 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2420 func(ioc, cic);
2421}
2422
Jens Axboe4ac845a2008-01-24 08:44:49 +01002423/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002424 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002425 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002426static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002427call_for_each_cic(struct io_context *ioc,
2428 void (*func)(struct io_context *, struct cfq_io_context *))
2429{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002430 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002431 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002432 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002433}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002434
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002435static void cfq_cic_free_rcu(struct rcu_head *head)
2436{
2437 struct cfq_io_context *cic;
2438
2439 cic = container_of(head, struct cfq_io_context, rcu_head);
2440
2441 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002442 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002443
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002444 if (ioc_gone) {
2445 /*
2446 * CFQ scheduler is exiting, grab exit lock and check
2447 * the pending io context count. If it hits zero,
2448 * complete ioc_gone and set it back to NULL
2449 */
2450 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002451 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002452 complete(ioc_gone);
2453 ioc_gone = NULL;
2454 }
2455 spin_unlock(&ioc_gone_lock);
2456 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002457}
2458
2459static void cfq_cic_free(struct cfq_io_context *cic)
2460{
2461 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002462}
2463
2464static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2465{
2466 unsigned long flags;
2467
2468 BUG_ON(!cic->dead_key);
2469
2470 spin_lock_irqsave(&ioc->lock, flags);
2471 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01002472 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002473 spin_unlock_irqrestore(&ioc->lock, flags);
2474
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002475 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002476}
2477
Jens Axboed6de8be2008-05-28 14:46:59 +02002478/*
2479 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2480 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2481 * and ->trim() which is called with the task lock held
2482 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002483static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002485 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002486 * ioc->refcount is zero here, or we are called from elv_unregister(),
2487 * so no more cic's are allowed to be linked into this ioc. So it
2488 * should be ok to iterate over the known list, we will see all cic's
2489 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002490 */
Jens Axboe07416d22008-05-07 09:17:12 +02002491 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492}
2493
Jens Axboe89850f72006-07-22 16:48:31 +02002494static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2495{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002496 struct cfq_queue *__cfqq, *next;
2497
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002498 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02002499 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002500 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002501 }
Jens Axboe89850f72006-07-22 16:48:31 +02002502
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002503 /*
2504 * If this queue was scheduled to merge with another queue, be
2505 * sure to drop the reference taken on that queue (and others in
2506 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2507 */
2508 __cfqq = cfqq->new_cfqq;
2509 while (__cfqq) {
2510 if (__cfqq == cfqq) {
2511 WARN(1, "cfqq->new_cfqq loop detected\n");
2512 break;
2513 }
2514 next = __cfqq->new_cfqq;
2515 cfq_put_queue(__cfqq);
2516 __cfqq = next;
2517 }
2518
Jens Axboe89850f72006-07-22 16:48:31 +02002519 cfq_put_queue(cfqq);
2520}
2521
2522static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2523 struct cfq_io_context *cic)
2524{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002525 struct io_context *ioc = cic->ioc;
2526
Jens Axboefc463792006-08-29 09:05:44 +02002527 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002528
2529 /*
2530 * Make sure key == NULL is seen for dead queues
2531 */
Jens Axboefc463792006-08-29 09:05:44 +02002532 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01002533 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02002534 cic->key = NULL;
2535
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002536 if (ioc->ioc_data == cic)
2537 rcu_assign_pointer(ioc->ioc_data, NULL);
2538
Jens Axboeff6657c2009-04-08 10:58:57 +02002539 if (cic->cfqq[BLK_RW_ASYNC]) {
2540 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2541 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002542 }
2543
Jens Axboeff6657c2009-04-08 10:58:57 +02002544 if (cic->cfqq[BLK_RW_SYNC]) {
2545 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2546 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002547 }
Jens Axboe89850f72006-07-22 16:48:31 +02002548}
2549
Jens Axboe4ac845a2008-01-24 08:44:49 +01002550static void cfq_exit_single_io_context(struct io_context *ioc,
2551 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002552{
Al Viro478a82b2006-03-18 13:25:24 -05002553 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02002554
Jens Axboe89850f72006-07-22 16:48:31 +02002555 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002556 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002557 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002558
Jens Axboe4ac845a2008-01-24 08:44:49 +01002559 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002560
2561 /*
2562 * Ensure we get a fresh copy of the ->key to prevent
2563 * race between exiting task and queue
2564 */
2565 smp_read_barrier_depends();
2566 if (cic->key)
2567 __cfq_exit_single_io_context(cfqd, cic);
2568
Jens Axboe4ac845a2008-01-24 08:44:49 +01002569 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002570 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002571}
2572
Jens Axboe498d3aa22007-04-26 12:54:48 +02002573/*
2574 * The process that ioc belongs to has exited, we need to clean up
2575 * and put the internal structures we have that belongs to that process.
2576 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002577static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002579 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580}
2581
Jens Axboe22e2c502005-06-27 10:55:12 +02002582static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002583cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584{
Jens Axboeb5deef92006-07-19 23:39:40 +02002585 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586
Christoph Lameter94f60302007-07-17 04:03:29 -07002587 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2588 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002590 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002591 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002592 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002593 cic->dtor = cfq_free_io_context;
2594 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002595 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596 }
2597
2598 return cic;
2599}
2600
Jens Axboefd0928d2008-01-24 08:52:45 +01002601static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002602{
2603 struct task_struct *tsk = current;
2604 int ioprio_class;
2605
Jens Axboe3b181522005-06-27 10:56:24 +02002606 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002607 return;
2608
Jens Axboefd0928d2008-01-24 08:52:45 +01002609 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002610 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002611 default:
2612 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2613 case IOPRIO_CLASS_NONE:
2614 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002615 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002616 */
2617 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002618 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002619 break;
2620 case IOPRIO_CLASS_RT:
2621 cfqq->ioprio = task_ioprio(ioc);
2622 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2623 break;
2624 case IOPRIO_CLASS_BE:
2625 cfqq->ioprio = task_ioprio(ioc);
2626 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2627 break;
2628 case IOPRIO_CLASS_IDLE:
2629 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2630 cfqq->ioprio = 7;
2631 cfq_clear_cfqq_idle_window(cfqq);
2632 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002633 }
2634
2635 /*
2636 * keep track of original prio settings in case we have to temporarily
2637 * elevate the priority of this queue
2638 */
2639 cfqq->org_ioprio = cfqq->ioprio;
2640 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002641 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002642}
2643
Jens Axboefebffd62008-01-28 13:19:43 +01002644static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002645{
Al Viro478a82b2006-03-18 13:25:24 -05002646 struct cfq_data *cfqd = cic->key;
2647 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002648 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002649
Jens Axboecaaa5f92006-06-16 11:23:00 +02002650 if (unlikely(!cfqd))
2651 return;
2652
Jens Axboec1b707d2006-10-30 19:54:23 +01002653 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002654
Jens Axboeff6657c2009-04-08 10:58:57 +02002655 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002656 if (cfqq) {
2657 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002658 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2659 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002660 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002661 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002662 cfq_put_queue(cfqq);
2663 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002664 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002665
Jens Axboeff6657c2009-04-08 10:58:57 +02002666 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002667 if (cfqq)
2668 cfq_mark_cfqq_prio_changed(cfqq);
2669
Jens Axboec1b707d2006-10-30 19:54:23 +01002670 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002671}
2672
Jens Axboefc463792006-08-29 09:05:44 +02002673static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002675 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002676 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677}
2678
Jens Axboed5036d72009-06-26 10:44:34 +02002679static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002680 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002681{
2682 RB_CLEAR_NODE(&cfqq->rb_node);
2683 RB_CLEAR_NODE(&cfqq->p_node);
2684 INIT_LIST_HEAD(&cfqq->fifo);
2685
2686 atomic_set(&cfqq->ref, 0);
2687 cfqq->cfqd = cfqd;
2688
2689 cfq_mark_cfqq_prio_changed(cfqq);
2690
2691 if (is_sync) {
2692 if (!cfq_class_idle(cfqq))
2693 cfq_mark_cfqq_idle_window(cfqq);
2694 cfq_mark_cfqq_sync(cfqq);
2695 }
2696 cfqq->pid = pid;
2697}
2698
Vivek Goyal246103332009-12-03 12:59:51 -05002699#ifdef CONFIG_CFQ_GROUP_IOSCHED
2700static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2701{
2702 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2703 struct cfq_data *cfqd = cic->key;
2704 unsigned long flags;
2705 struct request_queue *q;
2706
2707 if (unlikely(!cfqd))
2708 return;
2709
2710 q = cfqd->queue;
2711
2712 spin_lock_irqsave(q->queue_lock, flags);
2713
2714 if (sync_cfqq) {
2715 /*
2716 * Drop reference to sync queue. A new sync queue will be
2717 * assigned in new group upon arrival of a fresh request.
2718 */
2719 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2720 cic_set_cfqq(cic, NULL, 1);
2721 cfq_put_queue(sync_cfqq);
2722 }
2723
2724 spin_unlock_irqrestore(q->queue_lock, flags);
2725}
2726
2727static void cfq_ioc_set_cgroup(struct io_context *ioc)
2728{
2729 call_for_each_cic(ioc, changed_cgroup);
2730 ioc->cgroup_changed = 0;
2731}
2732#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2733
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002735cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002736 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002739 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002740 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741
2742retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002743 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002744 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002745 /* cic always exists here */
2746 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747
Jens Axboe6118b702009-06-30 09:34:12 +02002748 /*
2749 * Always try a new alloc if we fell back to the OOM cfqq
2750 * originally, since it should just be a temporary situation.
2751 */
2752 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2753 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754 if (new_cfqq) {
2755 cfqq = new_cfqq;
2756 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002757 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002759 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002760 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002761 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002763 if (new_cfqq)
2764 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002765 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002766 cfqq = kmem_cache_alloc_node(cfq_pool,
2767 gfp_mask | __GFP_ZERO,
2768 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002769 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770
Jens Axboe6118b702009-06-30 09:34:12 +02002771 if (cfqq) {
2772 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2773 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002774 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002775 cfq_log_cfqq(cfqd, cfqq, "alloced");
2776 } else
2777 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 }
2779
2780 if (new_cfqq)
2781 kmem_cache_free(cfq_pool, new_cfqq);
2782
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 return cfqq;
2784}
2785
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002786static struct cfq_queue **
2787cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2788{
Jens Axboefe094d92008-01-31 13:08:54 +01002789 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002790 case IOPRIO_CLASS_RT:
2791 return &cfqd->async_cfqq[0][ioprio];
2792 case IOPRIO_CLASS_BE:
2793 return &cfqd->async_cfqq[1][ioprio];
2794 case IOPRIO_CLASS_IDLE:
2795 return &cfqd->async_idle_cfqq;
2796 default:
2797 BUG();
2798 }
2799}
2800
Jens Axboe15c31be2007-07-10 13:43:25 +02002801static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002802cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002803 gfp_t gfp_mask)
2804{
Jens Axboefd0928d2008-01-24 08:52:45 +01002805 const int ioprio = task_ioprio(ioc);
2806 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002807 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002808 struct cfq_queue *cfqq = NULL;
2809
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002810 if (!is_sync) {
2811 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2812 cfqq = *async_cfqq;
2813 }
2814
Jens Axboe6118b702009-06-30 09:34:12 +02002815 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002816 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002817
2818 /*
2819 * pin the queue now that it's allocated, scheduler exit will prune it
2820 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002821 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02002822 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002823 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002824 }
2825
2826 atomic_inc(&cfqq->ref);
2827 return cfqq;
2828}
2829
Jens Axboe498d3aa22007-04-26 12:54:48 +02002830/*
2831 * We drop cfq io contexts lazily, so we may find a dead one.
2832 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002833static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002834cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2835 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002836{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002837 unsigned long flags;
2838
Jens Axboefc463792006-08-29 09:05:44 +02002839 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02002840
Jens Axboe4ac845a2008-01-24 08:44:49 +01002841 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02002842
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002843 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002844
2845 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01002846 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002847 spin_unlock_irqrestore(&ioc->lock, flags);
2848
2849 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002850}
2851
Jens Axboee2d74ac2006-03-28 08:59:01 +02002852static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01002853cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002854{
Jens Axboee2d74ac2006-03-28 08:59:01 +02002855 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002856 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002857 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002858
Vasily Tarasov91fac312007-04-25 12:29:51 +02002859 if (unlikely(!ioc))
2860 return NULL;
2861
Jens Axboed6de8be2008-05-28 14:46:59 +02002862 rcu_read_lock();
2863
Jens Axboe597bc482007-04-24 21:23:53 +02002864 /*
2865 * we maintain a last-hit cache, to avoid browsing over the tree
2866 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002867 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02002868 if (cic && cic->key == cfqd) {
2869 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02002870 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002871 }
Jens Axboe597bc482007-04-24 21:23:53 +02002872
Jens Axboe4ac845a2008-01-24 08:44:49 +01002873 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01002874 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2875 rcu_read_unlock();
2876 if (!cic)
2877 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02002878 /* ->key must be copied to avoid race with cfq_exit_queue() */
2879 k = cic->key;
2880 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01002881 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002882 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01002883 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002884 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002885
Jens Axboed6de8be2008-05-28 14:46:59 +02002886 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002887 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002888 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002889 break;
2890 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002891
Jens Axboe4ac845a2008-01-24 08:44:49 +01002892 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002893}
2894
Jens Axboe4ac845a2008-01-24 08:44:49 +01002895/*
2896 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2897 * the process specific cfq io context when entered from the block layer.
2898 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2899 */
Jens Axboefebffd62008-01-28 13:19:43 +01002900static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2901 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002902{
Jens Axboe0261d682006-10-30 19:07:48 +01002903 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002904 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002905
Jens Axboe4ac845a2008-01-24 08:44:49 +01002906 ret = radix_tree_preload(gfp_mask);
2907 if (!ret) {
2908 cic->ioc = ioc;
2909 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002910
Jens Axboe4ac845a2008-01-24 08:44:49 +01002911 spin_lock_irqsave(&ioc->lock, flags);
2912 ret = radix_tree_insert(&ioc->radix_root,
2913 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01002914 if (!ret)
2915 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002916 spin_unlock_irqrestore(&ioc->lock, flags);
2917
2918 radix_tree_preload_end();
2919
2920 if (!ret) {
2921 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2922 list_add(&cic->queue_list, &cfqd->cic_list);
2923 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002924 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002925 }
2926
Jens Axboe4ac845a2008-01-24 08:44:49 +01002927 if (ret)
2928 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02002929
Jens Axboe4ac845a2008-01-24 08:44:49 +01002930 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002931}
2932
Jens Axboe22e2c502005-06-27 10:55:12 +02002933/*
2934 * Setup general io context and cfq io context. There can be several cfq
2935 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02002936 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02002937 */
2938static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02002939cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002940{
Jens Axboe22e2c502005-06-27 10:55:12 +02002941 struct io_context *ioc = NULL;
2942 struct cfq_io_context *cic;
2943
2944 might_sleep_if(gfp_mask & __GFP_WAIT);
2945
Jens Axboeb5deef92006-07-19 23:39:40 +02002946 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02002947 if (!ioc)
2948 return NULL;
2949
Jens Axboe4ac845a2008-01-24 08:44:49 +01002950 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002951 if (cic)
2952 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02002953
Jens Axboee2d74ac2006-03-28 08:59:01 +02002954 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2955 if (cic == NULL)
2956 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02002957
Jens Axboe4ac845a2008-01-24 08:44:49 +01002958 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2959 goto err_free;
2960
Jens Axboe22e2c502005-06-27 10:55:12 +02002961out:
Jens Axboefc463792006-08-29 09:05:44 +02002962 smp_read_barrier_depends();
2963 if (unlikely(ioc->ioprio_changed))
2964 cfq_ioc_set_ioprio(ioc);
2965
Vivek Goyal246103332009-12-03 12:59:51 -05002966#ifdef CONFIG_CFQ_GROUP_IOSCHED
2967 if (unlikely(ioc->cgroup_changed))
2968 cfq_ioc_set_cgroup(ioc);
2969#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02002970 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002971err_free:
2972 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002973err:
2974 put_io_context(ioc);
2975 return NULL;
2976}
2977
2978static void
2979cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2980{
Jens Axboeaaf12282007-01-19 11:30:16 +11002981 unsigned long elapsed = jiffies - cic->last_end_request;
2982 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02002983
2984 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2985 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2986 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2987}
2988
Jens Axboe206dc692006-03-28 13:03:44 +02002989static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002990cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02002991 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02002992{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002993 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01002994 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002995 if (cfqq->last_request_pos) {
2996 if (cfqq->last_request_pos < blk_rq_pos(rq))
2997 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2998 else
2999 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3000 }
Jens Axboe206dc692006-03-28 13:03:44 +02003001
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003002 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003003 if (blk_queue_nonrot(cfqd->queue))
3004 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3005 else
3006 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003007}
Jens Axboe22e2c502005-06-27 10:55:12 +02003008
3009/*
3010 * Disable idle window if the process thinks too long or seeks so much that
3011 * it doesn't matter
3012 */
3013static void
3014cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3015 struct cfq_io_context *cic)
3016{
Jens Axboe7b679132008-05-30 12:23:07 +02003017 int old_idle, enable_idle;
Jens Axboe1be92f2f2007-04-19 14:32:26 +02003018
Jens Axboe08717142008-01-28 11:38:15 +01003019 /*
3020 * Don't idle for async or idle io prio class
3021 */
3022 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f2f2007-04-19 14:32:26 +02003023 return;
3024
Jens Axboec265a7f2008-06-26 13:49:33 +02003025 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003026
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003027 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3028 cfq_mark_cfqq_deep(cfqq);
3029
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01003030 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003031 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003032 enable_idle = 0;
3033 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003034 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003035 enable_idle = 0;
3036 else
3037 enable_idle = 1;
3038 }
3039
Jens Axboe7b679132008-05-30 12:23:07 +02003040 if (old_idle != enable_idle) {
3041 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3042 if (enable_idle)
3043 cfq_mark_cfqq_idle_window(cfqq);
3044 else
3045 cfq_clear_cfqq_idle_window(cfqq);
3046 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003047}
3048
Jens Axboe22e2c502005-06-27 10:55:12 +02003049/*
3050 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3051 * no or if we aren't sure, a 1 will cause a preempt.
3052 */
Jens Axboea6151c32009-10-07 20:02:57 +02003053static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003054cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003055 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003056{
Jens Axboe6d048f52007-04-25 12:44:27 +02003057 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003058
Jens Axboe6d048f52007-04-25 12:44:27 +02003059 cfqq = cfqd->active_queue;
3060 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003061 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003062
Jens Axboe6d048f52007-04-25 12:44:27 +02003063 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003064 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003065
3066 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003067 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003068
Jens Axboe22e2c502005-06-27 10:55:12 +02003069 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003070 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3071 */
3072 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3073 return false;
3074
3075 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003076 * if the new request is sync, but the currently running queue is
3077 * not, let the sync request have priority.
3078 */
Jens Axboe5e705372006-07-13 12:39:25 +02003079 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003080 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003081
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003082 if (new_cfqq->cfqg != cfqq->cfqg)
3083 return false;
3084
3085 if (cfq_slice_used(cfqq))
3086 return true;
3087
3088 /* Allow preemption only if we are idling on sync-noidle tree */
3089 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3090 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3091 new_cfqq->service_tree->count == 2 &&
3092 RB_EMPTY_ROOT(&cfqq->sort_list))
3093 return true;
3094
Jens Axboe374f84a2006-07-23 01:42:19 +02003095 /*
3096 * So both queues are sync. Let the new request get disk time if
3097 * it's a metadata request and the current queue is doing regular IO.
3098 */
3099 if (rq_is_meta(rq) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003100 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003101
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003102 /*
3103 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3104 */
3105 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003106 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003107
Jens Axboe1e3335d2007-02-14 19:59:49 +01003108 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003109 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003110
3111 /*
3112 * if this request is as-good as one we would expect from the
3113 * current cfqq, let it preempt
3114 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003115 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003116 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003117
Jens Axboea6151c32009-10-07 20:02:57 +02003118 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003119}
3120
3121/*
3122 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3123 * let it have half of its nominal slice.
3124 */
3125static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3126{
Jens Axboe7b679132008-05-30 12:23:07 +02003127 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02003128 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003129
Jens Axboebf572252006-07-19 20:29:12 +02003130 /*
3131 * Put the new queue at the front of the of the current list,
3132 * so we know that it will be selected next.
3133 */
3134 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003135
3136 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02003137
Jens Axboe44f7c162007-01-19 11:51:58 +11003138 cfqq->slice_end = 0;
3139 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003140}
3141
3142/*
Jens Axboe5e705372006-07-13 12:39:25 +02003143 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003144 * something we should do about it
3145 */
3146static void
Jens Axboe5e705372006-07-13 12:39:25 +02003147cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3148 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003149{
Jens Axboe5e705372006-07-13 12:39:25 +02003150 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003151
Aaron Carroll45333d52008-08-26 15:52:36 +02003152 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003153 if (rq_is_meta(rq))
3154 cfqq->meta_pending++;
3155
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003156 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003157 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003158 cfq_update_idle_window(cfqd, cfqq, cic);
3159
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003160 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003161
3162 if (cfqq == cfqd->active_queue) {
3163 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003164 * Remember that we saw a request from this process, but
3165 * don't start queuing just yet. Otherwise we risk seeing lots
3166 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003167 * and merging. If the request is already larger than a single
3168 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003169 * merging is already done. Ditto for a busy system that
3170 * has other work pending, don't risk delaying until the
3171 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003172 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003173 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003174 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3175 cfqd->busy_queues > 1) {
Jens Axboed6ceb252009-04-14 14:18:16 +02003176 del_timer(&cfqd->idle_slice_timer);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003177 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003178 __blk_run_queue(cfqd->queue);
3179 } else
3180 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboed6ceb252009-04-14 14:18:16 +02003181 }
Jens Axboe5e705372006-07-13 12:39:25 +02003182 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003183 /*
3184 * not the active queue - expire current slice if it is
3185 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003186 * has some old slice time left and is of higher priority or
3187 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003188 */
3189 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09003190 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003191 }
3192}
3193
Jens Axboe165125e2007-07-24 09:28:11 +02003194static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003195{
Jens Axboeb4878f22005-10-20 16:42:29 +02003196 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003197 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003198
Jens Axboe7b679132008-05-30 12:23:07 +02003199 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003200 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201
Jens Axboe30996f42009-10-05 11:03:39 +02003202 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003203 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003204 cfq_add_rq_rb(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003205
Jens Axboe5e705372006-07-13 12:39:25 +02003206 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207}
3208
Aaron Carroll45333d52008-08-26 15:52:36 +02003209/*
3210 * Update hw_tag based on peak queue depth over 50 samples under
3211 * sufficient load.
3212 */
3213static void cfq_update_hw_tag(struct cfq_data *cfqd)
3214{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003215 struct cfq_queue *cfqq = cfqd->active_queue;
3216
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003217 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3218 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003219
3220 if (cfqd->hw_tag == 1)
3221 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003222
3223 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003224 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003225 return;
3226
Shaohua Li1a1238a2009-10-27 08:46:23 +01003227 /*
3228 * If active queue hasn't enough requests and can idle, cfq might not
3229 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3230 * case
3231 */
3232 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3233 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003234 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003235 return;
3236
Aaron Carroll45333d52008-08-26 15:52:36 +02003237 if (cfqd->hw_tag_samples++ < 50)
3238 return;
3239
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003240 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003241 cfqd->hw_tag = 1;
3242 else
3243 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003244}
3245
Vivek Goyal7667aa02009-12-08 17:52:58 -05003246static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3247{
3248 struct cfq_io_context *cic = cfqd->active_cic;
3249
3250 /* If there are other queues in the group, don't wait */
3251 if (cfqq->cfqg->nr_cfqq > 1)
3252 return false;
3253
3254 if (cfq_slice_used(cfqq))
3255 return true;
3256
3257 /* if slice left is less than think time, wait busy */
3258 if (cic && sample_valid(cic->ttime_samples)
3259 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3260 return true;
3261
3262 /*
3263 * If think times is less than a jiffy than ttime_mean=0 and above
3264 * will not be true. It might happen that slice has not expired yet
3265 * but will expire soon (4-5 ns) during select_queue(). To cover the
3266 * case where think time is less than a jiffy, mark the queue wait
3267 * busy if only 1 jiffy is left in the slice.
3268 */
3269 if (cfqq->slice_end - jiffies == 1)
3270 return true;
3271
3272 return false;
3273}
3274
Jens Axboe165125e2007-07-24 09:28:11 +02003275static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276{
Jens Axboe5e705372006-07-13 12:39:25 +02003277 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003278 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003279 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003280 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281
Jens Axboeb4878f22005-10-20 16:42:29 +02003282 now = jiffies;
Vivek Goyal2868ef72009-12-03 12:59:48 -05003283 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284
Aaron Carroll45333d52008-08-26 15:52:36 +02003285 cfq_update_hw_tag(cfqd);
3286
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003287 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003288 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003289 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003290 cfqq->dispatched--;
Divyesh Shah84c124d2010-04-09 08:31:19 +02003291 blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3292 rq_io_start_time_ns(rq), rq_data_dir(rq),
3293 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003295 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003296
Vivek Goyal365722b2009-10-03 15:21:27 +02003297 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003298 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003299 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3300 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003301 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003302
3303 /*
3304 * If this is the active queue, check if it needs to be expired,
3305 * or if we want to idle in case it has no pending requests.
3306 */
3307 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003308 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3309
Jens Axboe44f7c162007-01-19 11:51:58 +11003310 if (cfq_cfqq_slice_new(cfqq)) {
3311 cfq_set_prio_slice(cfqd, cfqq);
3312 cfq_clear_cfqq_slice_new(cfqq);
3313 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003314
3315 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003316 * Should we wait for next request to come in before we expire
3317 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003318 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003319 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyalf75edf22009-12-03 12:59:53 -05003320 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3321 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003322 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003323 }
3324
Jens Axboea36e71f2009-04-15 12:15:11 +02003325 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003326 * Idling is not enabled on:
3327 * - expired queues
3328 * - idle-priority queues
3329 * - async queues
3330 * - queues with still some requests queued
3331 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003332 */
Jens Axboe08717142008-01-28 11:38:15 +01003333 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02003334 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003335 else if (sync && cfqq_empty &&
3336 !cfq_close_cooperator(cfqd, cfqq)) {
3337 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3338 /*
3339 * Idling is enabled for SYNC_WORKLOAD.
3340 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3341 * only if we processed at least one !rq_noidle request
3342 */
3343 if (cfqd->serving_type == SYNC_WORKLOAD
Vivek Goyalc04645e2009-12-03 12:59:56 -05003344 || cfqd->noidle_tree_requires_idle
3345 || cfqq->cfqg->nr_cfqq == 1)
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003346 cfq_arm_slice_timer(cfqd);
3347 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003348 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003349
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003350 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003351 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352}
3353
Jens Axboe22e2c502005-06-27 10:55:12 +02003354/*
3355 * we temporarily boost lower priority queues if they are holding fs exclusive
3356 * resources. they are boosted to normal prio (CLASS_BE/4)
3357 */
3358static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359{
Jens Axboe22e2c502005-06-27 10:55:12 +02003360 if (has_fs_excl()) {
3361 /*
3362 * boost idle prio on transactions that would lock out other
3363 * users of the filesystem
3364 */
3365 if (cfq_class_idle(cfqq))
3366 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3367 if (cfqq->ioprio > IOPRIO_NORM)
3368 cfqq->ioprio = IOPRIO_NORM;
3369 } else {
3370 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003371 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003372 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003373 cfqq->ioprio_class = cfqq->org_ioprio_class;
3374 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003376}
3377
Jens Axboe89850f72006-07-22 16:48:31 +02003378static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003379{
Jens Axboe1b379d82009-08-11 08:26:11 +02003380 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003381 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003382 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003383 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003384
3385 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003386}
3387
Jens Axboe165125e2007-07-24 09:28:11 +02003388static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003389{
3390 struct cfq_data *cfqd = q->elevator->elevator_data;
3391 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003392 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003393 struct cfq_queue *cfqq;
3394
3395 /*
3396 * don't force setup of a queue from here, as a call to may_queue
3397 * does not necessarily imply that a request actually will be queued.
3398 * so just lookup a possibly existing queue, or return 'may queue'
3399 * if that fails
3400 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003401 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003402 if (!cic)
3403 return ELV_MQUEUE_MAY;
3404
Jens Axboeb0b78f82009-04-08 10:56:08 +02003405 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003406 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003407 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003408 cfq_prio_boost(cfqq);
3409
Jens Axboe89850f72006-07-22 16:48:31 +02003410 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003411 }
3412
3413 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414}
3415
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416/*
3417 * queue lock held here
3418 */
Jens Axboebb37b942006-12-01 10:42:33 +01003419static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420{
Jens Axboe5e705372006-07-13 12:39:25 +02003421 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422
Jens Axboe5e705372006-07-13 12:39:25 +02003423 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003424 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425
Jens Axboe22e2c502005-06-27 10:55:12 +02003426 BUG_ON(!cfqq->allocated[rw]);
3427 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428
Jens Axboe5e705372006-07-13 12:39:25 +02003429 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02003432 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434 cfq_put_queue(cfqq);
3435 }
3436}
3437
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003438static struct cfq_queue *
3439cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3440 struct cfq_queue *cfqq)
3441{
3442 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3443 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003444 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003445 cfq_put_queue(cfqq);
3446 return cic_to_cfqq(cic, 1);
3447}
3448
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003449/*
3450 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3451 * was the last process referring to said cfqq.
3452 */
3453static struct cfq_queue *
3454split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3455{
3456 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003457 cfqq->pid = current->pid;
3458 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003459 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003460 return cfqq;
3461 }
3462
3463 cic_set_cfqq(cic, NULL, 1);
3464 cfq_put_queue(cfqq);
3465 return NULL;
3466}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003468 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003470static int
Jens Axboe165125e2007-07-24 09:28:11 +02003471cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472{
3473 struct cfq_data *cfqd = q->elevator->elevator_data;
3474 struct cfq_io_context *cic;
3475 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003476 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003477 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478 unsigned long flags;
3479
3480 might_sleep_if(gfp_mask & __GFP_WAIT);
3481
Jens Axboee2d74ac2006-03-28 08:59:01 +02003482 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003483
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484 spin_lock_irqsave(q->queue_lock, flags);
3485
Jens Axboe22e2c502005-06-27 10:55:12 +02003486 if (!cic)
3487 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003489new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003490 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003491 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003492 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003493 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003494 } else {
3495 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003496 * If the queue was seeky for too long, break it apart.
3497 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003498 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003499 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3500 cfqq = split_cfqq(cic, cfqq);
3501 if (!cfqq)
3502 goto new_queue;
3503 }
3504
3505 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003506 * Check to see if this queue is scheduled to merge with
3507 * another, closely cooperating queue. The merging of
3508 * queues happens here as it must be done in process context.
3509 * The reference on new_cfqq was taken in merge_cfqqs.
3510 */
3511 if (cfqq->new_cfqq)
3512 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003513 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514
3515 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02003516 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02003517
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518 spin_unlock_irqrestore(q->queue_lock, flags);
3519
Jens Axboe5e705372006-07-13 12:39:25 +02003520 rq->elevator_private = cic;
3521 rq->elevator_private2 = cfqq;
3522 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003523
Jens Axboe22e2c502005-06-27 10:55:12 +02003524queue_fail:
3525 if (cic)
3526 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003527
Jens Axboe23e018a2009-10-05 08:52:35 +02003528 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003530 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531 return 1;
3532}
3533
David Howells65f27f32006-11-22 14:55:48 +00003534static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003535{
David Howells65f27f32006-11-22 14:55:48 +00003536 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003537 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003538 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003539
Jens Axboe40bb54d2009-04-15 12:11:10 +02003540 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09003541 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003542 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003543}
3544
3545/*
3546 * Timer running if the active_queue is currently idling inside its time slice
3547 */
3548static void cfq_idle_slice_timer(unsigned long data)
3549{
3550 struct cfq_data *cfqd = (struct cfq_data *) data;
3551 struct cfq_queue *cfqq;
3552 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003553 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003554
Jens Axboe7b679132008-05-30 12:23:07 +02003555 cfq_log(cfqd, "idle timer fired");
3556
Jens Axboe22e2c502005-06-27 10:55:12 +02003557 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3558
Jens Axboefe094d92008-01-31 13:08:54 +01003559 cfqq = cfqd->active_queue;
3560 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003561 timed_out = 0;
3562
Jens Axboe22e2c502005-06-27 10:55:12 +02003563 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003564 * We saw a request before the queue expired, let it through
3565 */
3566 if (cfq_cfqq_must_dispatch(cfqq))
3567 goto out_kick;
3568
3569 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003570 * expired
3571 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003572 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003573 goto expire;
3574
3575 /*
3576 * only expire and reinvoke request handler, if there are
3577 * other queues with pending requests
3578 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003579 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003580 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003581
3582 /*
3583 * not expired and it has a request pending, let it dispatch
3584 */
Jens Axboe75e50982009-04-07 08:56:14 +02003585 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003586 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003587
3588 /*
3589 * Queue depth flag is reset only when the idle didn't succeed
3590 */
3591 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003592 }
3593expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02003594 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003595out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003596 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003597out_cont:
3598 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3599}
3600
Jens Axboe3b181522005-06-27 10:56:24 +02003601static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3602{
3603 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003604 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003605}
Jens Axboe22e2c502005-06-27 10:55:12 +02003606
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003607static void cfq_put_async_queues(struct cfq_data *cfqd)
3608{
3609 int i;
3610
3611 for (i = 0; i < IOPRIO_BE_NR; i++) {
3612 if (cfqd->async_cfqq[0][i])
3613 cfq_put_queue(cfqd->async_cfqq[0][i]);
3614 if (cfqd->async_cfqq[1][i])
3615 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003616 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003617
3618 if (cfqd->async_idle_cfqq)
3619 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003620}
3621
Jens Axboebb729bc2009-12-06 09:54:19 +01003622static void cfq_cfqd_free(struct rcu_head *head)
3623{
3624 kfree(container_of(head, struct cfq_data, rcu));
3625}
3626
Jens Axboeb374d182008-10-31 10:05:07 +01003627static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628{
Jens Axboe22e2c502005-06-27 10:55:12 +02003629 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003630 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003631
Jens Axboe3b181522005-06-27 10:56:24 +02003632 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003633
Al Virod9ff4182006-03-18 13:51:22 -05003634 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003635
Al Virod9ff4182006-03-18 13:51:22 -05003636 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02003637 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003638
3639 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003640 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3641 struct cfq_io_context,
3642 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003643
3644 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003645 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003646
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003647 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003648 cfq_release_cfq_groups(cfqd);
3649 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003650
Al Virod9ff4182006-03-18 13:51:22 -05003651 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003652
3653 cfq_shutdown_timer_wq(cfqd);
3654
Vivek Goyalb1c35762009-12-03 12:59:47 -05003655 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003656 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003657}
3658
Jens Axboe165125e2007-07-24 09:28:11 +02003659static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660{
3661 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003662 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003663 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003664 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003665
Christoph Lameter94f60302007-07-17 04:03:29 -07003666 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003668 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003670 /* Init root service tree */
3671 cfqd->grp_service_tree = CFQ_RB_ROOT;
3672
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003673 /* Init root group */
3674 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003675 for_each_cfqg_st(cfqg, i, j, st)
3676 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003677 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003678
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003679 /* Give preference to root group over other groups */
3680 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3681
Vivek Goyal25fb5162009-12-03 12:59:46 -05003682#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003683 /*
3684 * Take a reference to root group which we never drop. This is just
3685 * to make sure that cfq_put_cfqg() does not try to kfree root group
3686 */
3687 atomic_set(&cfqg->ref, 1);
Vivek Goyal22084192009-12-03 12:59:49 -05003688 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3689 0);
Vivek Goyal25fb5162009-12-03 12:59:46 -05003690#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003691 /*
3692 * Not strictly needed (since RB_ROOT just clears the node and we
3693 * zeroed cfqd on alloc), but better be safe in case someone decides
3694 * to add magic to the rb code
3695 */
3696 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3697 cfqd->prio_trees[i] = RB_ROOT;
3698
Jens Axboe6118b702009-06-30 09:34:12 +02003699 /*
3700 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3701 * Grab a permanent reference to it, so that the normal code flow
3702 * will not attempt to free it.
3703 */
3704 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3705 atomic_inc(&cfqd->oom_cfqq.ref);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003706 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003707
Al Virod9ff4182006-03-18 13:51:22 -05003708 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003709
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711
Jens Axboe22e2c502005-06-27 10:55:12 +02003712 init_timer(&cfqd->idle_slice_timer);
3713 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3714 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3715
Jens Axboe23e018a2009-10-05 08:52:35 +02003716 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003717
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003719 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3720 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721 cfqd->cfq_back_max = cfq_back_max;
3722 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003723 cfqd->cfq_slice[0] = cfq_slice_async;
3724 cfqd->cfq_slice[1] = cfq_slice_sync;
3725 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3726 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003727 cfqd->cfq_latency = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05003728 cfqd->cfq_group_isolation = 0;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003729 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003730 /*
3731 * we optimistically start assuming sync ops weren't delayed in last
3732 * second, in order to have larger depth for async operations.
3733 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003734 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebb729bc2009-12-06 09:54:19 +01003735 INIT_RCU_HEAD(&cfqd->rcu);
Jens Axboebc1c1162006-06-08 08:49:06 +02003736 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003737}
3738
3739static void cfq_slab_kill(void)
3740{
Jens Axboed6de8be2008-05-28 14:46:59 +02003741 /*
3742 * Caller already ensured that pending RCU callbacks are completed,
3743 * so we should have no busy allocations at this point.
3744 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745 if (cfq_pool)
3746 kmem_cache_destroy(cfq_pool);
3747 if (cfq_ioc_pool)
3748 kmem_cache_destroy(cfq_ioc_pool);
3749}
3750
3751static int __init cfq_slab_setup(void)
3752{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003753 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754 if (!cfq_pool)
3755 goto fail;
3756
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003757 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758 if (!cfq_ioc_pool)
3759 goto fail;
3760
3761 return 0;
3762fail:
3763 cfq_slab_kill();
3764 return -ENOMEM;
3765}
3766
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767/*
3768 * sysfs parts below -->
3769 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003770static ssize_t
3771cfq_var_show(unsigned int var, char *page)
3772{
3773 return sprintf(page, "%d\n", var);
3774}
3775
3776static ssize_t
3777cfq_var_store(unsigned int *var, const char *page, size_t count)
3778{
3779 char *p = (char *) page;
3780
3781 *var = simple_strtoul(p, &p, 10);
3782 return count;
3783}
3784
Linus Torvalds1da177e2005-04-16 15:20:36 -07003785#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003786static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003788 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789 unsigned int __data = __VAR; \
3790 if (__CONV) \
3791 __data = jiffies_to_msecs(__data); \
3792 return cfq_var_show(__data, (page)); \
3793}
3794SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003795SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3796SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003797SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3798SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003799SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3800SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3801SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3802SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003803SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003804SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805#undef SHOW_FUNCTION
3806
3807#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003808static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003810 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003811 unsigned int __data; \
3812 int ret = cfq_var_store(&__data, (page), count); \
3813 if (__data < (MIN)) \
3814 __data = (MIN); \
3815 else if (__data > (MAX)) \
3816 __data = (MAX); \
3817 if (__CONV) \
3818 *(__PTR) = msecs_to_jiffies(__data); \
3819 else \
3820 *(__PTR) = __data; \
3821 return ret; \
3822}
3823STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003824STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3825 UINT_MAX, 1);
3826STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3827 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05003828STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003829STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3830 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003831STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3832STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3833STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01003834STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3835 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003836STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003837STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003838#undef STORE_FUNCTION
3839
Al Viroe572ec72006-03-18 22:27:18 -05003840#define CFQ_ATTR(name) \
3841 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02003842
Al Viroe572ec72006-03-18 22:27:18 -05003843static struct elv_fs_entry cfq_attrs[] = {
3844 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05003845 CFQ_ATTR(fifo_expire_sync),
3846 CFQ_ATTR(fifo_expire_async),
3847 CFQ_ATTR(back_seek_max),
3848 CFQ_ATTR(back_seek_penalty),
3849 CFQ_ATTR(slice_sync),
3850 CFQ_ATTR(slice_async),
3851 CFQ_ATTR(slice_async_rq),
3852 CFQ_ATTR(slice_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02003853 CFQ_ATTR(low_latency),
Vivek Goyalae30c282009-12-03 12:59:55 -05003854 CFQ_ATTR(group_isolation),
Al Viroe572ec72006-03-18 22:27:18 -05003855 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856};
3857
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858static struct elevator_type iosched_cfq = {
3859 .ops = {
3860 .elevator_merge_fn = cfq_merge,
3861 .elevator_merged_fn = cfq_merged_request,
3862 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01003863 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02003864 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003865 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02003866 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867 .elevator_deactivate_req_fn = cfq_deactivate_request,
3868 .elevator_queue_empty_fn = cfq_queue_empty,
3869 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02003870 .elevator_former_req_fn = elv_rb_former_request,
3871 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003872 .elevator_set_req_fn = cfq_set_request,
3873 .elevator_put_req_fn = cfq_put_request,
3874 .elevator_may_queue_fn = cfq_may_queue,
3875 .elevator_init_fn = cfq_init_queue,
3876 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02003877 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878 },
Al Viro3d1ab402006-03-18 18:35:43 -05003879 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880 .elevator_name = "cfq",
3881 .elevator_owner = THIS_MODULE,
3882};
3883
Vivek Goyal3e252062009-12-04 10:36:42 -05003884#ifdef CONFIG_CFQ_GROUP_IOSCHED
3885static struct blkio_policy_type blkio_policy_cfq = {
3886 .ops = {
3887 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3888 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3889 },
3890};
3891#else
3892static struct blkio_policy_type blkio_policy_cfq;
3893#endif
3894
Linus Torvalds1da177e2005-04-16 15:20:36 -07003895static int __init cfq_init(void)
3896{
Jens Axboe22e2c502005-06-27 10:55:12 +02003897 /*
3898 * could be 0 on HZ < 1000 setups
3899 */
3900 if (!cfq_slice_async)
3901 cfq_slice_async = 1;
3902 if (!cfq_slice_idle)
3903 cfq_slice_idle = 1;
3904
Linus Torvalds1da177e2005-04-16 15:20:36 -07003905 if (cfq_slab_setup())
3906 return -ENOMEM;
3907
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01003908 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05003909 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01003911 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912}
3913
3914static void __exit cfq_exit(void)
3915{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07003916 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05003917 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05003919 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02003920 /* ioc_gone's update must be visible before reading ioc_count */
3921 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02003922
3923 /*
3924 * this also protects us from entering cfq_slab_kill() with
3925 * pending RCU callbacks
3926 */
Tejun Heo245b2e72009-06-24 15:13:48 +09003927 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02003928 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08003929 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930}
3931
3932module_init(cfq_init);
3933module_exit(cfq_exit);
3934
3935MODULE_AUTHOR("Jens Axboe");
3936MODULE_LICENSE("GPL");
3937MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");