blob: ee55019066a19500c6df54addf90d91d2ace60fa [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * CFQ, or complete fairness queueing, disk scheduler.
3 *
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6 *
Jens Axboe0fe23472006-09-04 15:41:16 +02007 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090010#include <linux/slab.h>
Al Viro1cc9be62006-03-18 12:29:52 -050011#include <linux/blkdev.h>
12#include <linux/elevator.h>
Randy Dunlapad5ebd22009-11-11 13:47:45 +010013#include <linux/jiffies.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020015#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020016#include <linux/blktrace_api.h>
Tejun Heo6e736be2011-12-14 00:33:38 +010017#include "blk.h"
Vivek Goyale98ef892010-06-18 10:39:47 -040018#include "cfq.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070019
20/*
21 * tunables
22 */
Jens Axboefe094d92008-01-31 13:08:54 +010023/* max queue in one round of service */
Shaohua Liabc3c742010-03-01 09:20:54 +010024static const int cfq_quantum = 8;
Arjan van de Ven64100092006-01-06 09:46:02 +010025static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010026/* maximum backwards seek, in KiB */
27static const int cfq_back_max = 16 * 1024;
28/* penalty of a backwards seek */
29static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010030static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020031static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010032static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020033static int cfq_slice_idle = HZ / 125;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +020034static int cfq_group_idle = HZ / 125;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +010035static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36static const int cfq_hist_divisor = 4;
Jens Axboe22e2c502005-06-27 10:55:12 +020037
Jens Axboed9e76202007-04-20 14:27:50 +020038/*
Jens Axboe08717142008-01-28 11:38:15 +010039 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020040 */
Jens Axboe08717142008-01-28 11:38:15 +010041#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020042
43/*
44 * below this threshold, we consider thinktime immediate
45 */
46#define CFQ_MIN_TT (2)
47
Jens Axboe22e2c502005-06-27 10:55:12 +020048#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020049#define CFQ_HW_QUEUE_MIN (5)
Vivek Goyal25bc6b02009-12-03 12:59:43 -050050#define CFQ_SERVICE_SHIFT 12
Jens Axboe22e2c502005-06-27 10:55:12 +020051
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010052#define CFQQ_SEEK_THR (sector_t)(8 * 100)
Shaohua Lie9ce3352010-03-19 08:03:04 +010053#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
Corrado Zoccolo41647e72010-02-27 19:45:40 +010054#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010055#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
Shaohua Liae54abe2010-02-05 13:11:45 +010056
Tejun Heoa612fdd2011-12-14 00:33:41 +010057#define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Christoph Lametere18b8902006-12-06 20:33:20 -080061static struct kmem_cache *cfq_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
Jens Axboe22e2c502005-06-27 10:55:12 +020063#define CFQ_PRIO_LISTS IOPRIO_BE_NR
64#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020065#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
66
Jens Axboe206dc692006-03-28 13:03:44 +020067#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050068#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020069
Tejun Heoc5869802011-12-14 00:33:41 +010070struct cfq_ttime {
71 unsigned long last_end_request;
72
73 unsigned long ttime_total;
74 unsigned long ttime_samples;
75 unsigned long ttime_mean;
76};
77
Jens Axboe22e2c502005-06-27 10:55:12 +020078/*
Jens Axboecc09e292007-04-26 12:53:50 +020079 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
83 */
84struct cfq_rb_root {
85 struct rb_root rb;
86 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010087 unsigned count;
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010088 unsigned total_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050089 u64 min_vdisktime;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +020090 struct cfq_ttime ttime;
Jens Axboecc09e292007-04-26 12:53:50 +020091};
Shaohua Lif5f2b6c2011-07-12 14:24:55 +020092#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
93 .ttime = {.last_end_request = jiffies,},}
Jens Axboecc09e292007-04-26 12:53:50 +020094
95/*
Jens Axboe6118b702009-06-30 09:34:12 +020096 * Per process-grouping structure
97 */
98struct cfq_queue {
99 /* reference count */
Shaohua Li30d7b942011-01-07 08:46:59 +0100100 int ref;
Jens Axboe6118b702009-06-30 09:34:12 +0200101 /* various state flags, see below */
102 unsigned int flags;
103 /* parent cfq_data */
104 struct cfq_data *cfqd;
105 /* service_tree member */
106 struct rb_node rb_node;
107 /* service_tree key */
108 unsigned long rb_key;
109 /* prio tree member */
110 struct rb_node p_node;
111 /* prio tree root we belong to, if any */
112 struct rb_root *p_root;
113 /* sorted list of pending requests */
114 struct rb_root sort_list;
115 /* if fifo isn't expired, next request to serve */
116 struct request *next_rq;
117 /* requests queued in sort_list */
118 int queued[2];
119 /* currently allocated requests */
120 int allocated[2];
121 /* fifo list of requests in sort_list */
122 struct list_head fifo;
123
Vivek Goyaldae739e2009-12-03 12:59:45 -0500124 /* time when queue got scheduled in to dispatch first request. */
125 unsigned long dispatch_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500126 unsigned int allocated_slice;
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100127 unsigned int slice_dispatch;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500128 /* time when first request from queue completed and slice started. */
129 unsigned long slice_start;
Jens Axboe6118b702009-06-30 09:34:12 +0200130 unsigned long slice_end;
131 long slice_resid;
Jens Axboe6118b702009-06-30 09:34:12 +0200132
Christoph Hellwig65299a32011-08-23 14:50:29 +0200133 /* pending priority requests */
134 int prio_pending;
Jens Axboe6118b702009-06-30 09:34:12 +0200135 /* number of requests that are on the dispatch list or inside driver */
136 int dispatched;
137
138 /* io prio of this group */
139 unsigned short ioprio, org_ioprio;
Justin TerAvest4aede842011-07-12 08:31:45 +0200140 unsigned short ioprio_class;
Jens Axboe6118b702009-06-30 09:34:12 +0200141
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100142 pid_t pid;
143
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +0100144 u32 seek_history;
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400145 sector_t last_request_pos;
146
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100147 struct cfq_rb_root *service_tree;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400148 struct cfq_queue *new_cfqq;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500149 struct cfq_group *cfqg;
Vivek Goyalc4e78932010-08-23 12:25:03 +0200150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
Jens Axboe6118b702009-06-30 09:34:12 +0200152};
153
154/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100155 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100156 * IDLE is handled separately, so it has negative index
157 */
158enum wl_prio_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100159 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500160 RT_WORKLOAD = 1,
161 IDLE_WORKLOAD = 2,
Vivek Goyalb4627322010-10-22 09:48:43 +0200162 CFQ_PRIO_NR,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100163};
164
165/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100166 * Second index in the service_trees.
167 */
168enum wl_type_t {
169 ASYNC_WORKLOAD = 0,
170 SYNC_NOIDLE_WORKLOAD = 1,
171 SYNC_WORKLOAD = 2
172};
173
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500174/* This is per cgroup per device grouping structure */
175struct cfq_group {
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500176 /* group service_tree member */
177 struct rb_node rb_node;
178
179 /* group service_tree key */
180 u64 vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500181 unsigned int weight;
Justin TerAvest8184f932011-03-17 16:12:36 +0100182 unsigned int new_weight;
183 bool needs_update;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500184
185 /* number of cfqq currently on this group */
186 int nr_cfqq;
187
Jens Axboe22e2c502005-06-27 10:55:12 +0200188 /*
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200189 * Per group busy queues average. Useful for workload slice calc. We
Vivek Goyalb4627322010-10-22 09:48:43 +0200190 * create the array for each prio class but at run time it is used
191 * only for RT and BE class and slot for IDLE class remains unused.
192 * This is primarily done to avoid confusion and a gcc warning.
193 */
194 unsigned int busy_queues_avg[CFQ_PRIO_NR];
195 /*
196 * rr lists of queues with requests. We maintain service trees for
197 * RT and BE classes. These trees are subdivided in subclasses
198 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199 * class there is no subclassification and all the cfq queues go on
200 * a single tree service_tree_idle.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100201 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200202 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100203 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100204 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500205
206 unsigned long saved_workload_slice;
207 enum wl_type_t saved_workload;
208 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500209 struct blkio_group blkg;
210#ifdef CONFIG_CFQ_GROUP_IOSCHED
211 struct hlist_node cfqd_node;
Shaohua Li329a6782011-01-07 08:48:28 +0100212 int ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500213#endif
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200214 /* number of requests that are on the dispatch list or inside driver */
215 int dispatched;
Shaohua Li7700fc42011-07-12 14:24:56 +0200216 struct cfq_ttime ttime;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500217};
218
Tejun Heoc5869802011-12-14 00:33:41 +0100219struct cfq_io_cq {
220 struct io_cq icq; /* must be the first member */
221 struct cfq_queue *cfqq[2];
222 struct cfq_ttime ttime;
223};
224
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500225/*
226 * Per block device queue structure
227 */
228struct cfq_data {
229 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500230 /* Root service tree for cfq_groups */
231 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500232 struct cfq_group root_group;
233
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100234 /*
235 * The priority currently being served
236 */
237 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100238 enum wl_type_t serving_type;
239 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500240 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200241
242 /*
243 * Each priority tree is sorted by next_request position. These
244 * trees are used when determining if two or more queues are
245 * interleaving requests (see cfq_close_cooperator).
246 */
247 struct rb_root prio_trees[CFQ_PRIO_LISTS];
248
Jens Axboe22e2c502005-06-27 10:55:12 +0200249 unsigned int busy_queues;
Shaohua Lief8a41d2011-03-07 09:26:29 +0100250 unsigned int busy_sync_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +0200251
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100252 int rq_in_driver;
253 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200254
255 /*
256 * queue-depth detection
257 */
258 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200259 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100260 /*
261 * hw_tag can be
262 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
263 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
264 * 0 => no NCQ
265 */
266 int hw_tag_est_depth;
267 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200268
269 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200270 * idle window management
271 */
272 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200273 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200274
275 struct cfq_queue *active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +0100276 struct cfq_io_cq *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200277
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200278 /*
279 * async queue for each priority case
280 */
281 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
282 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200283
Jens Axboe6d048f52007-04-25 12:44:27 +0200284 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286 /*
287 * tunables, see top of file
288 */
289 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200290 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 unsigned int cfq_back_penalty;
292 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200293 unsigned int cfq_slice[2];
294 unsigned int cfq_slice_async_rq;
295 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200296 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200297 unsigned int cfq_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500298
Jens Axboe6118b702009-06-30 09:34:12 +0200299 /*
300 * Fallback dummy cfqq for extreme OOM conditions
301 */
302 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200303
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100304 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500305
306 /* List of cfq groups being managed on this device*/
307 struct hlist_head cfqg_list;
Vivek Goyal56edf7d2011-05-19 15:38:22 -0400308
309 /* Number of groups which are on blkcg->blkg_list */
310 unsigned int nr_blkcg_linked_grps;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311};
312
Vivek Goyal25fb5162009-12-03 12:59:46 -0500313static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
314
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500315static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
316 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500317 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100318{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500319 if (!cfqg)
320 return NULL;
321
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100322 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500323 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100324
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500325 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100326}
327
Jens Axboe3b181522005-06-27 10:56:24 +0200328enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100329 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
330 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200331 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100332 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100333 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
334 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
335 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100336 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200337 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400338 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100339 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100340 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500341 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200342};
343
344#define CFQ_CFQQ_FNS(name) \
345static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
346{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100347 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200348} \
349static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
350{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100351 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200352} \
353static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
354{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100355 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200356}
357
358CFQ_CFQQ_FNS(on_rr);
359CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200360CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200361CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200362CFQ_CFQQ_FNS(fifo_expire);
363CFQ_CFQQ_FNS(idle_window);
364CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100365CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200366CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200367CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100368CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100369CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500370CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200371#undef CFQ_CFQQ_FNS
372
Vivek Goyalafc24d42010-04-26 19:27:56 +0200373#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500374#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
375 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
376 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200377 blkg_path(&(cfqq)->cfqg->blkg), ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500378
379#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
380 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200381 blkg_path(&(cfqg)->blkg), ##args) \
Vivek Goyal2868ef72009-12-03 12:59:48 -0500382
383#else
Jens Axboe7b679132008-05-30 12:23:07 +0200384#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
385 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200386#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500387#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200388#define cfq_log(cfqd, fmt, args...) \
389 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
390
Vivek Goyal615f0252009-12-03 12:59:39 -0500391/* Traverses through cfq group service trees */
392#define for_each_cfqg_st(cfqg, i, j, st) \
393 for (i = 0; i <= IDLE_WORKLOAD; i++) \
394 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
395 : &cfqg->service_tree_idle; \
396 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
397 (i == IDLE_WORKLOAD && j == 0); \
398 j++, st = i < IDLE_WORKLOAD ? \
399 &cfqg->service_trees[i][j]: NULL) \
400
Shaohua Lif5f2b6c2011-07-12 14:24:55 +0200401static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
402 struct cfq_ttime *ttime, bool group_idle)
403{
404 unsigned long slice;
405 if (!sample_valid(ttime->ttime_samples))
406 return false;
407 if (group_idle)
408 slice = cfqd->cfq_group_idle;
409 else
410 slice = cfqd->cfq_slice_idle;
411 return ttime->ttime_mean > slice;
412}
Vivek Goyal615f0252009-12-03 12:59:39 -0500413
Vivek Goyal02b35082010-08-23 12:23:53 +0200414static inline bool iops_mode(struct cfq_data *cfqd)
415{
416 /*
417 * If we are not idling on queues and it is a NCQ drive, parallel
418 * execution of requests is on and measuring time is not possible
419 * in most of the cases until and unless we drive shallower queue
420 * depths and that becomes a performance bottleneck. In such cases
421 * switch to start providing fairness in terms of number of IOs.
422 */
423 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
424 return true;
425 else
426 return false;
427}
428
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100429static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
430{
431 if (cfq_class_idle(cfqq))
432 return IDLE_WORKLOAD;
433 if (cfq_class_rt(cfqq))
434 return RT_WORKLOAD;
435 return BE_WORKLOAD;
436}
437
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100438
439static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
440{
441 if (!cfq_cfqq_sync(cfqq))
442 return ASYNC_WORKLOAD;
443 if (!cfq_cfqq_idle_window(cfqq))
444 return SYNC_NOIDLE_WORKLOAD;
445 return SYNC_WORKLOAD;
446}
447
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500448static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
449 struct cfq_data *cfqd,
450 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100451{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500452 if (wl == IDLE_WORKLOAD)
453 return cfqg->service_tree_idle.count;
454
455 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
456 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
457 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100458}
459
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500460static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
461 struct cfq_group *cfqg)
462{
463 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
464 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
465}
466
Jens Axboe165125e2007-07-24 09:28:11 +0200467static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200468static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100469 struct io_context *, gfp_t);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200470
Tejun Heoc5869802011-12-14 00:33:41 +0100471static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
472{
473 /* cic->icq is the first member, %NULL will convert to %NULL */
474 return container_of(icq, struct cfq_io_cq, icq);
475}
476
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100477static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
478 struct io_context *ioc)
479{
480 if (ioc)
481 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
482 return NULL;
483}
484
Tejun Heoc5869802011-12-14 00:33:41 +0100485static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200486{
Jens Axboea6151c32009-10-07 20:02:57 +0200487 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200488}
489
Tejun Heoc5869802011-12-14 00:33:41 +0100490static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
491 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200492{
Jens Axboea6151c32009-10-07 20:02:57 +0200493 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200494}
495
Tejun Heoc5869802011-12-14 00:33:41 +0100496static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400497{
Tejun Heoc5869802011-12-14 00:33:41 +0100498 return cic->icq.q->elevator->elevator_data;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400499}
500
Vasily Tarasov91fac312007-04-25 12:29:51 +0200501/*
502 * We regard a request as SYNC, if it's either a read or has the SYNC bit
503 * set (in which case it could also be direct WRITE).
504 */
Jens Axboea6151c32009-10-07 20:02:57 +0200505static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200506{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200507 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200508}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700511 * scheduler run of queue, if there are requests pending and no one in the
512 * driver that will restart queueing
513 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200514static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700515{
Jens Axboe7b679132008-05-30 12:23:07 +0200516 if (cfqd->busy_queues) {
517 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200518 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200519 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700520}
521
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100523 * Scale schedule slice based on io priority. Use the sync time slice only
524 * if a queue is marked sync and has sync io queued. A sync queue with async
525 * io only, should not get full sync slice length.
526 */
Jens Axboea6151c32009-10-07 20:02:57 +0200527static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200528 unsigned short prio)
529{
530 const int base_slice = cfqd->cfq_slice[sync];
531
532 WARN_ON(prio >= IOPRIO_BE_NR);
533
534 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
535}
536
Jens Axboe44f7c162007-01-19 11:51:58 +1100537static inline int
538cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
539{
Jens Axboed9e76202007-04-20 14:27:50 +0200540 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100541}
542
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500543static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
544{
545 u64 d = delta << CFQ_SERVICE_SHIFT;
546
547 d = d * BLKIO_WEIGHT_DEFAULT;
548 do_div(d, cfqg->weight);
549 return d;
550}
551
552static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
553{
554 s64 delta = (s64)(vdisktime - min_vdisktime);
555 if (delta > 0)
556 min_vdisktime = vdisktime;
557
558 return min_vdisktime;
559}
560
561static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
562{
563 s64 delta = (s64)(vdisktime - min_vdisktime);
564 if (delta < 0)
565 min_vdisktime = vdisktime;
566
567 return min_vdisktime;
568}
569
570static void update_min_vdisktime(struct cfq_rb_root *st)
571{
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500572 struct cfq_group *cfqg;
573
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500574 if (st->left) {
575 cfqg = rb_entry_cfqg(st->left);
Gui Jianfenga6032712011-03-07 09:28:09 +0100576 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
577 cfqg->vdisktime);
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500578 }
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500579}
580
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100581/*
582 * get averaged number of queues of RT/BE priority.
583 * average is updated, with a formula that gives more weight to higher numbers,
584 * to quickly follows sudden increases and decrease slowly
585 */
586
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500587static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
588 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100589{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100590 unsigned min_q, max_q;
591 unsigned mult = cfq_hist_divisor - 1;
592 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500593 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100594
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500595 min_q = min(cfqg->busy_queues_avg[rt], busy);
596 max_q = max(cfqg->busy_queues_avg[rt], busy);
597 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100598 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500599 return cfqg->busy_queues_avg[rt];
600}
601
602static inline unsigned
603cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
604{
605 struct cfq_rb_root *st = &cfqd->grp_service_tree;
606
607 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100608}
609
Shaohua Lic553f8e2011-01-14 08:41:03 +0100610static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -0700611cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100612{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100613 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
614 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500615 /*
616 * interested queues (we consider only the ones with the same
617 * priority class in the cfq group)
618 */
619 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
620 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100621 unsigned sync_slice = cfqd->cfq_slice[1];
622 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500623 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
624
625 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100626 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
627 /* scale low_slice according to IO priority
628 * and sync vs async */
629 unsigned low_slice =
630 min(slice, base_low_slice * slice / sync_slice);
631 /* the adapted slice value is scaled to fit all iqs
632 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500633 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100634 low_slice);
635 }
636 }
Shaohua Lic553f8e2011-01-14 08:41:03 +0100637 return slice;
638}
639
640static inline void
641cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
642{
Vivek Goyalba5bd522011-01-19 08:25:02 -0700643 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +0100644
Vivek Goyaldae739e2009-12-03 12:59:45 -0500645 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100646 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500647 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200648 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100649}
650
651/*
652 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
653 * isn't valid until the first request from the dispatch is activated
654 * and the slice time set.
655 */
Jens Axboea6151c32009-10-07 20:02:57 +0200656static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100657{
658 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +0100659 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100660 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +0100661 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100662
Shaohua Lic1e44752010-11-08 15:01:02 +0100663 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +1100664}
665
666/*
Jens Axboe5e705372006-07-13 12:39:25 +0200667 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200669 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670 */
Jens Axboe5e705372006-07-13 12:39:25 +0200671static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100672cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100674 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200676#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
677#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
678 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679
Jens Axboe5e705372006-07-13 12:39:25 +0200680 if (rq1 == NULL || rq1 == rq2)
681 return rq2;
682 if (rq2 == NULL)
683 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200684
Namhyung Kim229836b2011-05-24 10:23:21 +0200685 if (rq_is_sync(rq1) != rq_is_sync(rq2))
686 return rq_is_sync(rq1) ? rq1 : rq2;
687
Christoph Hellwig65299a32011-08-23 14:50:29 +0200688 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
689 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
Jens Axboeb53d1ed2011-08-19 08:34:48 +0200690
Tejun Heo83096eb2009-05-07 22:24:39 +0900691 s1 = blk_rq_pos(rq1);
692 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 /*
695 * by definition, 1KiB is 2 sectors
696 */
697 back_max = cfqd->cfq_back_max * 2;
698
699 /*
700 * Strict one way elevator _except_ in the case where we allow
701 * short backward seeks which are biased as twice the cost of a
702 * similar forward seek.
703 */
704 if (s1 >= last)
705 d1 = s1 - last;
706 else if (s1 + back_max >= last)
707 d1 = (last - s1) * cfqd->cfq_back_penalty;
708 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200709 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
711 if (s2 >= last)
712 d2 = s2 - last;
713 else if (s2 + back_max >= last)
714 d2 = (last - s2) * cfqd->cfq_back_penalty;
715 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200716 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717
718 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719
Andreas Mohre8a99052006-03-28 08:59:49 +0200720 /*
721 * By doing switch() on the bit mask "wrap" we avoid having to
722 * check two variables for all permutations: --> faster!
723 */
724 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200725 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200726 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200727 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200728 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200729 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200730 else {
731 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200732 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200733 else
Jens Axboe5e705372006-07-13 12:39:25 +0200734 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200735 }
736
737 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200738 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200739 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200740 return rq2;
741 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200742 default:
743 /*
744 * Since both rqs are wrapped,
745 * start with the one that's further behind head
746 * (--> only *one* back seek required),
747 * since back seek takes more time than forward.
748 */
749 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200750 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 else
Jens Axboe5e705372006-07-13 12:39:25 +0200752 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 }
754}
755
Jens Axboe498d3aa22007-04-26 12:54:48 +0200756/*
757 * The below is leftmost cache rbtree addon
758 */
Jens Axboe08717142008-01-28 11:38:15 +0100759static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200760{
Vivek Goyal615f0252009-12-03 12:59:39 -0500761 /* Service tree is empty */
762 if (!root->count)
763 return NULL;
764
Jens Axboecc09e292007-04-26 12:53:50 +0200765 if (!root->left)
766 root->left = rb_first(&root->rb);
767
Jens Axboe08717142008-01-28 11:38:15 +0100768 if (root->left)
769 return rb_entry(root->left, struct cfq_queue, rb_node);
770
771 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200772}
773
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500774static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
775{
776 if (!root->left)
777 root->left = rb_first(&root->rb);
778
779 if (root->left)
780 return rb_entry_cfqg(root->left);
781
782 return NULL;
783}
784
Jens Axboea36e71f2009-04-15 12:15:11 +0200785static void rb_erase_init(struct rb_node *n, struct rb_root *root)
786{
787 rb_erase(n, root);
788 RB_CLEAR_NODE(n);
789}
790
Jens Axboecc09e292007-04-26 12:53:50 +0200791static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
792{
793 if (root->left == n)
794 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200795 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100796 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200797}
798
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799/*
800 * would be nice to take fifo expire time into account as well
801 */
Jens Axboe5e705372006-07-13 12:39:25 +0200802static struct request *
803cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
804 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805{
Jens Axboe21183b02006-07-13 12:33:14 +0200806 struct rb_node *rbnext = rb_next(&last->rb_node);
807 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200808 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809
Jens Axboe21183b02006-07-13 12:33:14 +0200810 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811
812 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200813 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200814
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200816 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200817 else {
818 rbnext = rb_first(&cfqq->sort_list);
819 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200820 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200821 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100823 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824}
825
Jens Axboed9e76202007-04-20 14:27:50 +0200826static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
827 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828{
Jens Axboed9e76202007-04-20 14:27:50 +0200829 /*
830 * just an approximation, should be ok.
831 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500832 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100833 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200834}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500836static inline s64
837cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
838{
839 return cfqg->vdisktime - st->min_vdisktime;
840}
841
842static void
843__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
844{
845 struct rb_node **node = &st->rb.rb_node;
846 struct rb_node *parent = NULL;
847 struct cfq_group *__cfqg;
848 s64 key = cfqg_key(st, cfqg);
849 int left = 1;
850
851 while (*node != NULL) {
852 parent = *node;
853 __cfqg = rb_entry_cfqg(parent);
854
855 if (key < cfqg_key(st, __cfqg))
856 node = &parent->rb_left;
857 else {
858 node = &parent->rb_right;
859 left = 0;
860 }
861 }
862
863 if (left)
864 st->left = &cfqg->rb_node;
865
866 rb_link_node(&cfqg->rb_node, parent, node);
867 rb_insert_color(&cfqg->rb_node, &st->rb);
868}
869
870static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100871cfq_update_group_weight(struct cfq_group *cfqg)
872{
873 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
874 if (cfqg->needs_update) {
875 cfqg->weight = cfqg->new_weight;
876 cfqg->needs_update = false;
877 }
878}
879
880static void
881cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
882{
883 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
884
885 cfq_update_group_weight(cfqg);
886 __cfq_group_service_tree_add(st, cfqg);
887 st->total_weight += cfqg->weight;
888}
889
890static void
891cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500892{
893 struct cfq_rb_root *st = &cfqd->grp_service_tree;
894 struct cfq_group *__cfqg;
895 struct rb_node *n;
896
897 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +0100898 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500899 return;
900
901 /*
902 * Currently put the group at the end. Later implement something
903 * so that groups get lesser vtime based on their weights, so that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300904 * if group does not loose all if it was not continuously backlogged.
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500905 */
906 n = rb_last(&st->rb);
907 if (n) {
908 __cfqg = rb_entry_cfqg(n);
909 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
910 } else
911 cfqg->vdisktime = st->min_vdisktime;
Justin TerAvest8184f932011-03-17 16:12:36 +0100912 cfq_group_service_tree_add(st, cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500913}
914
915static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100916cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
917{
918 st->total_weight -= cfqg->weight;
919 if (!RB_EMPTY_NODE(&cfqg->rb_node))
920 cfq_rb_erase(&cfqg->rb_node, st);
921}
922
923static void
924cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500925{
926 struct cfq_rb_root *st = &cfqd->grp_service_tree;
927
928 BUG_ON(cfqg->nr_cfqq < 1);
929 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500930
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500931 /* If there are other cfq queues under this group, don't delete it */
932 if (cfqg->nr_cfqq)
933 return;
934
Vivek Goyal2868ef72009-12-03 12:59:48 -0500935 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Justin TerAvest8184f932011-03-17 16:12:36 +0100936 cfq_group_service_tree_del(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500937 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400938 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500939}
940
Justin TerAvest167400d2011-03-12 16:54:00 +0100941static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
942 unsigned int *unaccounted_time)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500943{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500944 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500945
946 /*
947 * Queue got expired before even a single request completed or
948 * got expired immediately after first request completion.
949 */
950 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
951 /*
952 * Also charge the seek time incurred to the group, otherwise
953 * if there are mutiple queues in the group, each can dispatch
954 * a single request on seeky media and cause lots of seek time
955 * and group will never know it.
956 */
957 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
958 1);
959 } else {
960 slice_used = jiffies - cfqq->slice_start;
Justin TerAvest167400d2011-03-12 16:54:00 +0100961 if (slice_used > cfqq->allocated_slice) {
962 *unaccounted_time = slice_used - cfqq->allocated_slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500963 slice_used = cfqq->allocated_slice;
Justin TerAvest167400d2011-03-12 16:54:00 +0100964 }
965 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
966 *unaccounted_time += cfqq->slice_start -
967 cfqq->dispatch_start;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500968 }
969
Vivek Goyaldae739e2009-12-03 12:59:45 -0500970 return slice_used;
971}
972
973static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200974 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500975{
976 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Justin TerAvest167400d2011-03-12 16:54:00 +0100977 unsigned int used_sl, charge, unaccounted_sl = 0;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500978 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
979 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500980
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500981 BUG_ON(nr_sync < 0);
Justin TerAvest167400d2011-03-12 16:54:00 +0100982 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500983
Vivek Goyal02b35082010-08-23 12:23:53 +0200984 if (iops_mode(cfqd))
985 charge = cfqq->slice_dispatch;
986 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
987 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500988
989 /* Can't update vdisktime while group is on service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +0100990 cfq_group_service_tree_del(st, cfqg);
Vivek Goyal02b35082010-08-23 12:23:53 +0200991 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Justin TerAvest8184f932011-03-17 16:12:36 +0100992 /* If a new weight was requested, update now, off tree */
993 cfq_group_service_tree_add(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500994
995 /* This group is being expired. Save the context */
996 if (time_after(cfqd->workload_expires, jiffies)) {
997 cfqg->saved_workload_slice = cfqd->workload_expires
998 - jiffies;
999 cfqg->saved_workload = cfqd->serving_type;
1000 cfqg->saved_serving_prio = cfqd->serving_prio;
1001 } else
1002 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -05001003
1004 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1005 st->min_vdisktime);
Joe Perchesfd16d262011-06-13 10:42:49 +02001006 cfq_log_cfqq(cfqq->cfqd, cfqq,
1007 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1008 used_sl, cfqq->slice_dispatch, charge,
1009 iops_mode(cfqd), cfqq->nr_sectors);
Justin TerAvest167400d2011-03-12 16:54:00 +01001010 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1011 unaccounted_sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001012 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001013}
1014
Vivek Goyal25fb5162009-12-03 12:59:46 -05001015#ifdef CONFIG_CFQ_GROUP_IOSCHED
1016static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1017{
1018 if (blkg)
1019 return container_of(blkg, struct cfq_group, blkg);
1020 return NULL;
1021}
1022
Paul Bolle8aea4542011-06-06 05:07:54 +02001023static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1024 unsigned int weight)
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001025{
Justin TerAvest8184f932011-03-17 16:12:36 +01001026 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1027 cfqg->new_weight = weight;
1028 cfqg->needs_update = true;
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001029}
1030
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001031static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1032 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001033{
Vivek Goyal22084192009-12-03 12:59:49 -05001034 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1035 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001036
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001037 /*
1038 * Add group onto cgroup list. It might happen that bdi->dev is
1039 * not initialized yet. Initialize this new group without major
1040 * and minor info and this info will be filled in once a new thread
1041 * comes for IO.
1042 */
1043 if (bdi->dev) {
Ricky Beniteza74b2ad2010-04-05 18:22:17 +02001044 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001045 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1046 (void *)cfqd, MKDEV(major, minor));
1047 } else
1048 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1049 (void *)cfqd, 0);
1050
1051 cfqd->nr_blkcg_linked_grps++;
1052 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1053
1054 /* Add group on cfqd list */
1055 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1056}
1057
1058/*
1059 * Should be called from sleepable context. No request queue lock as per
1060 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1061 * from sleepable context.
1062 */
1063static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1064{
1065 struct cfq_group *cfqg = NULL;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001066 int i, j, ret;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001067 struct cfq_rb_root *st;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001068
1069 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1070 if (!cfqg)
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001071 return NULL;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001072
Vivek Goyal25fb5162009-12-03 12:59:46 -05001073 for_each_cfqg_st(cfqg, i, j, st)
1074 *st = CFQ_RB_ROOT;
1075 RB_CLEAR_NODE(&cfqg->rb_node);
1076
Shaohua Li7700fc42011-07-12 14:24:56 +02001077 cfqg->ttime.last_end_request = jiffies;
1078
Vivek Goyalb1c35762009-12-03 12:59:47 -05001079 /*
1080 * Take the initial reference that will be released on destroy
1081 * This can be thought of a joint reference by cgroup and
1082 * elevator which will be dropped by either elevator exit
1083 * or cgroup deletion path depending on who is exiting first.
1084 */
Shaohua Li329a6782011-01-07 08:48:28 +01001085 cfqg->ref = 1;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001086
1087 ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1088 if (ret) {
1089 kfree(cfqg);
1090 return NULL;
1091 }
1092
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001093 return cfqg;
1094}
1095
1096static struct cfq_group *
1097cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1098{
1099 struct cfq_group *cfqg = NULL;
1100 void *key = cfqd;
1101 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1102 unsigned int major, minor;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001103
Vivek Goyal180be2a2010-09-14 08:47:11 +02001104 /*
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001105 * This is the common case when there are no blkio cgroups.
1106 * Avoid lookup in this case
Vivek Goyal180be2a2010-09-14 08:47:11 +02001107 */
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001108 if (blkcg == &blkio_root_cgroup)
1109 cfqg = &cfqd->root_group;
1110 else
1111 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1112
1113 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
Vivek Goyal180be2a2010-09-14 08:47:11 +02001114 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001115 cfqg->blkg.dev = MKDEV(major, minor);
1116 }
Vivek Goyal180be2a2010-09-14 08:47:11 +02001117
Vivek Goyal25fb5162009-12-03 12:59:46 -05001118 return cfqg;
1119}
1120
1121/*
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001122 * Search for the cfq group current task belongs to. request_queue lock must
1123 * be held.
Vivek Goyal25fb5162009-12-03 12:59:46 -05001124 */
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001125static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001126{
Vivek Goyal70087dc2011-05-16 15:24:08 +02001127 struct blkio_cgroup *blkcg;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001128 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1129 struct request_queue *q = cfqd->queue;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001130
1131 rcu_read_lock();
Vivek Goyal70087dc2011-05-16 15:24:08 +02001132 blkcg = task_blkio_cgroup(current);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001133 cfqg = cfq_find_cfqg(cfqd, blkcg);
1134 if (cfqg) {
1135 rcu_read_unlock();
1136 return cfqg;
1137 }
1138
1139 /*
1140 * Need to allocate a group. Allocation of group also needs allocation
1141 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1142 * we need to drop rcu lock and queue_lock before we call alloc.
1143 *
1144 * Not taking any queue reference here and assuming that queue is
1145 * around by the time we return. CFQ queue allocation code does
1146 * the same. It might be racy though.
1147 */
1148
1149 rcu_read_unlock();
1150 spin_unlock_irq(q->queue_lock);
1151
1152 cfqg = cfq_alloc_cfqg(cfqd);
1153
1154 spin_lock_irq(q->queue_lock);
1155
1156 rcu_read_lock();
1157 blkcg = task_blkio_cgroup(current);
1158
1159 /*
1160 * If some other thread already allocated the group while we were
1161 * not holding queue lock, free up the group
1162 */
1163 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1164
1165 if (__cfqg) {
1166 kfree(cfqg);
1167 rcu_read_unlock();
1168 return __cfqg;
1169 }
1170
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001171 if (!cfqg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001172 cfqg = &cfqd->root_group;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001173
1174 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001175 rcu_read_unlock();
1176 return cfqg;
1177}
1178
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001179static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1180{
Shaohua Li329a6782011-01-07 08:48:28 +01001181 cfqg->ref++;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001182 return cfqg;
1183}
1184
Vivek Goyal25fb5162009-12-03 12:59:46 -05001185static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1186{
1187 /* Currently, all async queues are mapped to root group */
1188 if (!cfq_cfqq_sync(cfqq))
1189 cfqg = &cfqq->cfqd->root_group;
1190
1191 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001192 /* cfqq reference on cfqg */
Shaohua Li329a6782011-01-07 08:48:28 +01001193 cfqq->cfqg->ref++;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001194}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001195
1196static void cfq_put_cfqg(struct cfq_group *cfqg)
1197{
1198 struct cfq_rb_root *st;
1199 int i, j;
1200
Shaohua Li329a6782011-01-07 08:48:28 +01001201 BUG_ON(cfqg->ref <= 0);
1202 cfqg->ref--;
1203 if (cfqg->ref)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001204 return;
1205 for_each_cfqg_st(cfqg, i, j, st)
Gui Jianfengb54ce602010-11-30 20:52:46 +01001206 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001207 free_percpu(cfqg->blkg.stats_cpu);
Vivek Goyalb1c35762009-12-03 12:59:47 -05001208 kfree(cfqg);
1209}
1210
1211static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1212{
1213 /* Something wrong if we are trying to remove same group twice */
1214 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1215
1216 hlist_del_init(&cfqg->cfqd_node);
1217
Vivek Goyala5395b82011-08-02 09:24:09 +02001218 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1219 cfqd->nr_blkcg_linked_grps--;
1220
Vivek Goyalb1c35762009-12-03 12:59:47 -05001221 /*
1222 * Put the reference taken at the time of creation so that when all
1223 * queues are gone, group can be destroyed.
1224 */
1225 cfq_put_cfqg(cfqg);
1226}
1227
1228static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1229{
1230 struct hlist_node *pos, *n;
1231 struct cfq_group *cfqg;
1232
1233 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1234 /*
1235 * If cgroup removal path got to blk_group first and removed
1236 * it from cgroup list, then it will take care of destroying
1237 * cfqg also.
1238 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001239 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001240 cfq_destroy_cfqg(cfqd, cfqg);
1241 }
1242}
1243
1244/*
1245 * Blk cgroup controller notification saying that blkio_group object is being
1246 * delinked as associated cgroup object is going away. That also means that
1247 * no new IO will come in this group. So get rid of this group as soon as
1248 * any pending IO in the group is finished.
1249 *
1250 * This function is called under rcu_read_lock(). key is the rcu protected
1251 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1252 * read lock.
1253 *
1254 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1255 * it should not be NULL as even if elevator was exiting, cgroup deltion
1256 * path got to it first.
1257 */
Paul Bolle8aea4542011-06-06 05:07:54 +02001258static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001259{
1260 unsigned long flags;
1261 struct cfq_data *cfqd = key;
1262
1263 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1264 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1265 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1266}
1267
Vivek Goyal25fb5162009-12-03 12:59:46 -05001268#else /* GROUP_IOSCHED */
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001269static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001270{
1271 return &cfqd->root_group;
1272}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001273
1274static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1275{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001276 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001277}
1278
Vivek Goyal25fb5162009-12-03 12:59:46 -05001279static inline void
1280cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1281 cfqq->cfqg = cfqg;
1282}
1283
Vivek Goyalb1c35762009-12-03 12:59:47 -05001284static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1285static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1286
Vivek Goyal25fb5162009-12-03 12:59:46 -05001287#endif /* GROUP_IOSCHED */
1288
Jens Axboe498d3aa22007-04-26 12:54:48 +02001289/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001290 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001291 * requests waiting to be processed. It is sorted in the order that
1292 * we will service the queues.
1293 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001294static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001295 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001296{
Jens Axboe08717142008-01-28 11:38:15 +01001297 struct rb_node **p, *parent;
1298 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001299 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001300 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001301 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001302 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001303
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001304 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001305 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001306 if (cfq_class_idle(cfqq)) {
1307 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001308 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001309 if (parent && parent != &cfqq->rb_node) {
1310 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1311 rb_key += __cfqq->rb_key;
1312 } else
1313 rb_key += jiffies;
1314 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001315 /*
1316 * Get our rb key offset. Subtract any residual slice
1317 * value carried from last service. A negative resid
1318 * count indicates slice overrun, and this should position
1319 * the next service time further away in the tree.
1320 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001321 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001322 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001323 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001324 } else {
1325 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001326 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001327 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1328 }
Jens Axboed9e76202007-04-20 14:27:50 +02001329
1330 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001331 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001332 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001333 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001334 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001335 if (rb_key == cfqq->rb_key &&
1336 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001337 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001338
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001339 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1340 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001341 }
Jens Axboed9e76202007-04-20 14:27:50 +02001342
Jens Axboe498d3aa22007-04-26 12:54:48 +02001343 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001344 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001345 cfqq->service_tree = service_tree;
1346 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001347 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001348 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001349
Jens Axboed9e76202007-04-20 14:27:50 +02001350 parent = *p;
1351 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1352
Jens Axboe0c534e02007-04-18 20:01:57 +02001353 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001354 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001355 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001356 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001357 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001358 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001359 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001360 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001361 }
Jens Axboe67060e32007-04-18 20:13:32 +02001362
1363 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001364 }
1365
Jens Axboecc09e292007-04-26 12:53:50 +02001366 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001367 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001368
Jens Axboed9e76202007-04-20 14:27:50 +02001369 cfqq->rb_key = rb_key;
1370 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001371 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1372 service_tree->count++;
Namhyung Kim20359f22011-05-24 10:23:22 +02001373 if (add_front || !new_cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001374 return;
Justin TerAvest8184f932011-03-17 16:12:36 +01001375 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376}
1377
Jens Axboea36e71f2009-04-15 12:15:11 +02001378static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001379cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1380 sector_t sector, struct rb_node **ret_parent,
1381 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001382{
Jens Axboea36e71f2009-04-15 12:15:11 +02001383 struct rb_node **p, *parent;
1384 struct cfq_queue *cfqq = NULL;
1385
1386 parent = NULL;
1387 p = &root->rb_node;
1388 while (*p) {
1389 struct rb_node **n;
1390
1391 parent = *p;
1392 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1393
1394 /*
1395 * Sort strictly based on sector. Smallest to the left,
1396 * largest to the right.
1397 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001398 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001399 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001400 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001401 n = &(*p)->rb_left;
1402 else
1403 break;
1404 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001405 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001406 }
1407
1408 *ret_parent = parent;
1409 if (rb_link)
1410 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001411 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001412}
1413
1414static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1415{
Jens Axboea36e71f2009-04-15 12:15:11 +02001416 struct rb_node **p, *parent;
1417 struct cfq_queue *__cfqq;
1418
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001419 if (cfqq->p_root) {
1420 rb_erase(&cfqq->p_node, cfqq->p_root);
1421 cfqq->p_root = NULL;
1422 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001423
1424 if (cfq_class_idle(cfqq))
1425 return;
1426 if (!cfqq->next_rq)
1427 return;
1428
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001429 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001430 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1431 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001432 if (!__cfqq) {
1433 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001434 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1435 } else
1436 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001437}
1438
Jens Axboe498d3aa22007-04-26 12:54:48 +02001439/*
1440 * Update cfqq's position in the service tree.
1441 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001442static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001443{
Jens Axboe6d048f52007-04-25 12:44:27 +02001444 /*
1445 * Resorting requires the cfqq to be on the RR list already.
1446 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001447 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001448 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001449 cfq_prio_tree_add(cfqd, cfqq);
1450 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001451}
1452
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453/*
1454 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001455 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 */
Jens Axboefebffd62008-01-28 13:19:43 +01001457static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458{
Jens Axboe7b679132008-05-30 12:23:07 +02001459 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001460 BUG_ON(cfq_cfqq_on_rr(cfqq));
1461 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462 cfqd->busy_queues++;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001463 if (cfq_cfqq_sync(cfqq))
1464 cfqd->busy_sync_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465
Jens Axboeedd75ff2007-04-19 12:03:34 +02001466 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467}
1468
Jens Axboe498d3aa22007-04-26 12:54:48 +02001469/*
1470 * Called when the cfqq no longer has requests pending, remove it from
1471 * the service tree.
1472 */
Jens Axboefebffd62008-01-28 13:19:43 +01001473static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474{
Jens Axboe7b679132008-05-30 12:23:07 +02001475 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001476 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1477 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001479 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1480 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1481 cfqq->service_tree = NULL;
1482 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001483 if (cfqq->p_root) {
1484 rb_erase(&cfqq->p_node, cfqq->p_root);
1485 cfqq->p_root = NULL;
1486 }
Jens Axboed9e76202007-04-20 14:27:50 +02001487
Justin TerAvest8184f932011-03-17 16:12:36 +01001488 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 BUG_ON(!cfqd->busy_queues);
1490 cfqd->busy_queues--;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001491 if (cfq_cfqq_sync(cfqq))
1492 cfqd->busy_sync_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493}
1494
1495/*
1496 * rb tree support functions
1497 */
Jens Axboefebffd62008-01-28 13:19:43 +01001498static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499{
Jens Axboe5e705372006-07-13 12:39:25 +02001500 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001501 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502
Jens Axboeb4878f22005-10-20 16:42:29 +02001503 BUG_ON(!cfqq->queued[sync]);
1504 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505
Jens Axboe5e705372006-07-13 12:39:25 +02001506 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507
Vivek Goyalf04a6422009-12-03 12:59:40 -05001508 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1509 /*
1510 * Queue will be deleted from service tree when we actually
1511 * expire it later. Right now just remove it from prio tree
1512 * as it is empty.
1513 */
1514 if (cfqq->p_root) {
1515 rb_erase(&cfqq->p_node, cfqq->p_root);
1516 cfqq->p_root = NULL;
1517 }
1518 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519}
1520
Jens Axboe5e705372006-07-13 12:39:25 +02001521static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522{
Jens Axboe5e705372006-07-13 12:39:25 +02001523 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 struct cfq_data *cfqd = cfqq->cfqd;
Jeff Moyer796d5112011-06-02 21:19:05 +02001525 struct request *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526
Jens Axboe5380a102006-07-13 12:37:56 +02001527 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528
Jeff Moyer796d5112011-06-02 21:19:05 +02001529 elv_rb_add(&cfqq->sort_list, rq);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001530
1531 if (!cfq_cfqq_on_rr(cfqq))
1532 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001533
1534 /*
1535 * check if this request is a better next-serve candidate
1536 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001537 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001538 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001539
1540 /*
1541 * adjust priority tree position, if ->next_rq changes
1542 */
1543 if (prev != cfqq->next_rq)
1544 cfq_prio_tree_add(cfqd, cfqq);
1545
Jens Axboe5044eed2007-04-25 11:53:48 +02001546 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547}
1548
Jens Axboefebffd62008-01-28 13:19:43 +01001549static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550{
Jens Axboe5380a102006-07-13 12:37:56 +02001551 elv_rb_del(&cfqq->sort_list, rq);
1552 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001553 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1554 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001555 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001556 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001557 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1558 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559}
1560
Jens Axboe206dc692006-03-28 13:03:44 +02001561static struct request *
1562cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563{
Jens Axboe206dc692006-03-28 13:03:44 +02001564 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01001565 struct cfq_io_cq *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001566 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567
Jens Axboe4ac845a2008-01-24 08:44:49 +01001568 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001569 if (!cic)
1570 return NULL;
1571
1572 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001573 if (cfqq) {
1574 sector_t sector = bio->bi_sector + bio_sectors(bio);
1575
Jens Axboe21183b02006-07-13 12:33:14 +02001576 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001577 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579 return NULL;
1580}
1581
Jens Axboe165125e2007-07-24 09:28:11 +02001582static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001583{
1584 struct cfq_data *cfqd = q->elevator->elevator_data;
1585
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001586 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001587 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001588 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001589
Tejun Heo5b936292009-05-07 22:24:38 +09001590 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001591}
1592
Jens Axboe165125e2007-07-24 09:28:11 +02001593static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594{
Jens Axboe22e2c502005-06-27 10:55:12 +02001595 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001597 WARN_ON(!cfqd->rq_in_driver);
1598 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001599 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001600 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601}
1602
Jens Axboeb4878f22005-10-20 16:42:29 +02001603static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604{
Jens Axboe5e705372006-07-13 12:39:25 +02001605 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001606
Jens Axboe5e705372006-07-13 12:39:25 +02001607 if (cfqq->next_rq == rq)
1608 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609
Jens Axboeb4878f22005-10-20 16:42:29 +02001610 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001611 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001612
Aaron Carroll45333d52008-08-26 15:52:36 +02001613 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001614 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1615 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig65299a32011-08-23 14:50:29 +02001616 if (rq->cmd_flags & REQ_PRIO) {
1617 WARN_ON(!cfqq->prio_pending);
1618 cfqq->prio_pending--;
Jens Axboeb53d1ed2011-08-19 08:34:48 +02001619 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620}
1621
Jens Axboe165125e2007-07-24 09:28:11 +02001622static int cfq_merge(struct request_queue *q, struct request **req,
1623 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624{
1625 struct cfq_data *cfqd = q->elevator->elevator_data;
1626 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627
Jens Axboe206dc692006-03-28 13:03:44 +02001628 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001629 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001630 *req = __rq;
1631 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 }
1633
1634 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635}
1636
Jens Axboe165125e2007-07-24 09:28:11 +02001637static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001638 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639{
Jens Axboe21183b02006-07-13 12:33:14 +02001640 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001641 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642
Jens Axboe5e705372006-07-13 12:39:25 +02001643 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645}
1646
Divyesh Shah812d4022010-04-08 21:14:23 -07001647static void cfq_bio_merged(struct request_queue *q, struct request *req,
1648 struct bio *bio)
1649{
Vivek Goyale98ef892010-06-18 10:39:47 -04001650 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1651 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001652}
1653
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654static void
Jens Axboe165125e2007-07-24 09:28:11 +02001655cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 struct request *next)
1657{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001658 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Shaohua Li4a0b75c2011-12-16 14:00:22 +01001659 struct cfq_data *cfqd = q->elevator->elevator_data;
1660
Jens Axboe22e2c502005-06-27 10:55:12 +02001661 /*
1662 * reposition in fifo if next is older than rq
1663 */
1664 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001665 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001666 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001667 rq_set_fifo_time(rq, rq_fifo_time(next));
1668 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001669
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001670 if (cfqq->next_rq == next)
1671 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001672 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001673 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1674 rq_data_dir(next), rq_is_sync(next));
Shaohua Li4a0b75c2011-12-16 14:00:22 +01001675
1676 cfqq = RQ_CFQQ(next);
1677 /*
1678 * all requests of this queue are merged to other queues, delete it
1679 * from the service tree. If it's the active_queue,
1680 * cfq_dispatch_requests() will choose to expire it or do idle
1681 */
1682 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1683 cfqq != cfqd->active_queue)
1684 cfq_del_cfqq_rr(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001685}
1686
Jens Axboe165125e2007-07-24 09:28:11 +02001687static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001688 struct bio *bio)
1689{
1690 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heoc5869802011-12-14 00:33:41 +01001691 struct cfq_io_cq *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001692 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001693
1694 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001695 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001696 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001697 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001698 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001699
1700 /*
Tejun Heof1a4f4d2011-12-14 00:33:39 +01001701 * Lookup the cfqq that this bio will be queued with and allow
1702 * merge only if rq is queued there. This function can be called
1703 * from plug merge without queue_lock. In such cases, ioc of @rq
1704 * and %current are guaranteed to be equal. Avoid lookup which
1705 * requires queue_lock by using @rq's cic.
Jens Axboeda775262006-12-20 11:04:12 +01001706 */
Tejun Heoc5869802011-12-14 00:33:41 +01001707 if (current->io_context == RQ_CIC(rq)->icq.ioc) {
Tejun Heof1a4f4d2011-12-14 00:33:39 +01001708 cic = RQ_CIC(rq);
1709 } else {
1710 cic = cfq_cic_lookup(cfqd, current->io_context);
1711 if (!cic)
1712 return false;
1713 }
Jens Axboe719d3402006-12-22 09:38:53 +01001714
Vasily Tarasov91fac312007-04-25 12:29:51 +02001715 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001716 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001717}
1718
Divyesh Shah812df482010-04-08 21:15:35 -07001719static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1720{
1721 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001722 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001723}
1724
Jens Axboefebffd62008-01-28 13:19:43 +01001725static void __cfq_set_active_queue(struct cfq_data *cfqd,
1726 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001727{
1728 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001729 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1730 cfqd->serving_prio, cfqd->serving_type);
Justin TerAvest62a37f62011-03-23 08:25:44 +01001731 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1732 cfqq->slice_start = 0;
1733 cfqq->dispatch_start = jiffies;
1734 cfqq->allocated_slice = 0;
1735 cfqq->slice_end = 0;
1736 cfqq->slice_dispatch = 0;
1737 cfqq->nr_sectors = 0;
1738
1739 cfq_clear_cfqq_wait_request(cfqq);
1740 cfq_clear_cfqq_must_dispatch(cfqq);
1741 cfq_clear_cfqq_must_alloc_slice(cfqq);
1742 cfq_clear_cfqq_fifo_expire(cfqq);
1743 cfq_mark_cfqq_slice_new(cfqq);
1744
1745 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001746 }
1747
1748 cfqd->active_queue = cfqq;
1749}
1750
1751/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001752 * current cfqq expired its slice (or was too idle), select new one
1753 */
1754static void
1755__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001756 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001757{
Jens Axboe7b679132008-05-30 12:23:07 +02001758 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1759
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001760 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001761 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001762
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001763 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001764 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001765
1766 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001767 * If this cfqq is shared between multiple processes, check to
1768 * make sure that those processes are still issuing I/Os within
1769 * the mean seek distance. If not, it may be time to break the
1770 * queues apart again.
1771 */
1772 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1773 cfq_mark_cfqq_split_coop(cfqq);
1774
1775 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001776 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001777 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01001778 if (timed_out) {
1779 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07001780 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001781 else
1782 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001783 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1784 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001785
Vivek Goyale5ff0822010-04-26 19:25:11 +02001786 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001787
Vivek Goyalf04a6422009-12-03 12:59:40 -05001788 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1789 cfq_del_cfqq_rr(cfqd, cfqq);
1790
Jens Axboeedd75ff2007-04-19 12:03:34 +02001791 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001792
1793 if (cfqq == cfqd->active_queue)
1794 cfqd->active_queue = NULL;
1795
1796 if (cfqd->active_cic) {
Tejun Heoc5869802011-12-14 00:33:41 +01001797 put_io_context(cfqd->active_cic->icq.ioc, cfqd->queue);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001798 cfqd->active_cic = NULL;
1799 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001800}
1801
Vivek Goyale5ff0822010-04-26 19:25:11 +02001802static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001803{
1804 struct cfq_queue *cfqq = cfqd->active_queue;
1805
1806 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001807 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001808}
1809
Jens Axboe498d3aa22007-04-26 12:54:48 +02001810/*
1811 * Get next queue for service. Unless we have a queue preemption,
1812 * we'll simply select the first cfqq in the service tree.
1813 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001814static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001815{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001816 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001817 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001818 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001819
Vivek Goyalf04a6422009-12-03 12:59:40 -05001820 if (!cfqd->rq_queued)
1821 return NULL;
1822
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001823 /* There is nothing to dispatch */
1824 if (!service_tree)
1825 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001826 if (RB_EMPTY_ROOT(&service_tree->rb))
1827 return NULL;
1828 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001829}
1830
Vivek Goyalf04a6422009-12-03 12:59:40 -05001831static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1832{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001833 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001834 struct cfq_queue *cfqq;
1835 int i, j;
1836 struct cfq_rb_root *st;
1837
1838 if (!cfqd->rq_queued)
1839 return NULL;
1840
Vivek Goyal25fb5162009-12-03 12:59:46 -05001841 cfqg = cfq_get_next_cfqg(cfqd);
1842 if (!cfqg)
1843 return NULL;
1844
Vivek Goyalf04a6422009-12-03 12:59:40 -05001845 for_each_cfqg_st(cfqg, i, j, st)
1846 if ((cfqq = cfq_rb_first(st)) != NULL)
1847 return cfqq;
1848 return NULL;
1849}
1850
Jens Axboe498d3aa22007-04-26 12:54:48 +02001851/*
1852 * Get and set a new active queue for service.
1853 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001854static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1855 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001856{
Jens Axboee00ef792009-11-04 08:54:55 +01001857 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001858 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001859
Jens Axboe22e2c502005-06-27 10:55:12 +02001860 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001861 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001862}
1863
Jens Axboed9e76202007-04-20 14:27:50 +02001864static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1865 struct request *rq)
1866{
Tejun Heo83096eb2009-05-07 22:24:39 +09001867 if (blk_rq_pos(rq) >= cfqd->last_position)
1868 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001869 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001870 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001871}
1872
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001873static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001874 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001875{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001876 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001877}
1878
Jens Axboea36e71f2009-04-15 12:15:11 +02001879static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1880 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001881{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001882 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001883 struct rb_node *parent, *node;
1884 struct cfq_queue *__cfqq;
1885 sector_t sector = cfqd->last_position;
1886
1887 if (RB_EMPTY_ROOT(root))
1888 return NULL;
1889
1890 /*
1891 * First, if we find a request starting at the end of the last
1892 * request, choose it.
1893 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001894 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001895 if (__cfqq)
1896 return __cfqq;
1897
1898 /*
1899 * If the exact sector wasn't found, the parent of the NULL leaf
1900 * will contain the closest sector.
1901 */
1902 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001903 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001904 return __cfqq;
1905
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001906 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001907 node = rb_next(&__cfqq->p_node);
1908 else
1909 node = rb_prev(&__cfqq->p_node);
1910 if (!node)
1911 return NULL;
1912
1913 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001914 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001915 return __cfqq;
1916
1917 return NULL;
1918}
1919
1920/*
1921 * cfqd - obvious
1922 * cur_cfqq - passed in so that we don't decide that the current queue is
1923 * closely cooperating with itself.
1924 *
1925 * So, basically we're assuming that that cur_cfqq has dispatched at least
1926 * one request, and that cfqd->last_position reflects a position on the disk
1927 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1928 * assumption.
1929 */
1930static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001931 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001932{
1933 struct cfq_queue *cfqq;
1934
Divyesh Shah39c01b22010-03-25 15:45:57 +01001935 if (cfq_class_idle(cur_cfqq))
1936 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001937 if (!cfq_cfqq_sync(cur_cfqq))
1938 return NULL;
1939 if (CFQQ_SEEKY(cur_cfqq))
1940 return NULL;
1941
Jens Axboea36e71f2009-04-15 12:15:11 +02001942 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001943 * Don't search priority tree if it's the only queue in the group.
1944 */
1945 if (cur_cfqq->cfqg->nr_cfqq == 1)
1946 return NULL;
1947
1948 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001949 * We should notice if some of the queues are cooperating, eg
1950 * working closely on the same area of the disk. In that case,
1951 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001952 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001953 cfqq = cfqq_close(cfqd, cur_cfqq);
1954 if (!cfqq)
1955 return NULL;
1956
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001957 /* If new queue belongs to different cfq_group, don't choose it */
1958 if (cur_cfqq->cfqg != cfqq->cfqg)
1959 return NULL;
1960
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001961 /*
1962 * It only makes sense to merge sync queues.
1963 */
1964 if (!cfq_cfqq_sync(cfqq))
1965 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001966 if (CFQQ_SEEKY(cfqq))
1967 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001968
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001969 /*
1970 * Do not merge queues of different priority classes
1971 */
1972 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1973 return NULL;
1974
Jens Axboea36e71f2009-04-15 12:15:11 +02001975 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001976}
1977
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001978/*
1979 * Determine whether we should enforce idle window for this queue.
1980 */
1981
1982static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1983{
1984 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001985 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001986
Vivek Goyalf04a6422009-12-03 12:59:40 -05001987 BUG_ON(!service_tree);
1988 BUG_ON(!service_tree->count);
1989
Vivek Goyalb6508c12010-08-23 12:23:33 +02001990 if (!cfqd->cfq_slice_idle)
1991 return false;
1992
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001993 /* We never do for idle class queues. */
1994 if (prio == IDLE_WORKLOAD)
1995 return false;
1996
1997 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001998 if (cfq_cfqq_idle_window(cfqq) &&
1999 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002000 return true;
2001
2002 /*
2003 * Otherwise, we do only if they are the last ones
2004 * in their service tree.
2005 */
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002006 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2007 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
Shaohua Lic1e44752010-11-08 15:01:02 +01002008 return true;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002009 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2010 service_tree->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01002011 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002012}
2013
Jens Axboe6d048f52007-04-25 12:44:27 +02002014static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002015{
Jens Axboe17926692007-01-19 11:59:30 +11002016 struct cfq_queue *cfqq = cfqd->active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +01002017 struct cfq_io_cq *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002018 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002019
Jens Axboea68bbdd2008-09-24 13:03:33 +02002020 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002021 * SSD device without seek penalty, disable idling. But only do so
2022 * for devices that support queuing, otherwise we still have a problem
2023 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02002024 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002025 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02002026 return;
2027
Jens Axboedd67d052006-06-21 09:36:18 +02002028 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02002029 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02002030
2031 /*
2032 * idle is disabled, either manually or by past process history
2033 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002034 if (!cfq_should_idle(cfqd, cfqq)) {
2035 /* no queue idling. Check for group idling */
2036 if (cfqd->cfq_group_idle)
2037 group_idle = cfqd->cfq_group_idle;
2038 else
2039 return;
2040 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002041
Jens Axboe22e2c502005-06-27 10:55:12 +02002042 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002043 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02002044 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002045 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02002046 return;
2047
2048 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002049 * task has exited, don't wait
2050 */
Jens Axboe206dc692006-03-28 13:03:44 +02002051 cic = cfqd->active_cic;
Tejun Heoc5869802011-12-14 00:33:41 +01002052 if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02002053 return;
2054
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002055 /*
2056 * If our average think time is larger than the remaining time
2057 * slice, then don't idle. This avoids overrunning the allotted
2058 * time slice.
2059 */
Shaohua Li383cd722011-07-12 14:24:35 +02002060 if (sample_valid(cic->ttime.ttime_samples) &&
2061 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
Joe Perchesfd16d262011-06-13 10:42:49 +02002062 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
Shaohua Li383cd722011-07-12 14:24:35 +02002063 cic->ttime.ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002064 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002065 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002066
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002067 /* There are other queues in the group, don't do group idle */
2068 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2069 return;
2070
Jens Axboe3b181522005-06-27 10:56:24 +02002071 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002072
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002073 if (group_idle)
2074 sl = cfqd->cfq_group_idle;
2075 else
2076 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02002077
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002078 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04002079 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002080 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2081 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082}
2083
Jens Axboe498d3aa22007-04-26 12:54:48 +02002084/*
2085 * Move request from internal lists to the request queue dispatch list.
2086 */
Jens Axboe165125e2007-07-24 09:28:11 +02002087static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088{
Jens Axboe3ed9a292007-04-23 08:33:33 +02002089 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002090 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002091
Jens Axboe7b679132008-05-30 12:23:07 +02002092 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2093
Jeff Moyer06d21882009-09-11 17:08:59 +02002094 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02002095 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002096 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002097 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02002098 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02002099
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002100 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02002101 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04002102 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02002103 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104}
2105
2106/*
2107 * return expired entry, or NULL to just start from scratch in rbtree
2108 */
Jens Axboefebffd62008-01-28 13:19:43 +01002109static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110{
Jens Axboe30996f42009-10-05 11:03:39 +02002111 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112
Jens Axboe3b181522005-06-27 10:56:24 +02002113 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002115
2116 cfq_mark_cfqq_fifo_expire(cfqq);
2117
Jens Axboe89850f72006-07-22 16:48:31 +02002118 if (list_empty(&cfqq->fifo))
2119 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120
Jens Axboe89850f72006-07-22 16:48:31 +02002121 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002122 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002123 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124
Jens Axboe30996f42009-10-05 11:03:39 +02002125 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002126 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127}
2128
Jens Axboe22e2c502005-06-27 10:55:12 +02002129static inline int
2130cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2131{
2132 const int base_rq = cfqd->cfq_slice_async_rq;
2133
2134 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2135
Namhyung Kimb9f8ce02011-05-24 10:23:21 +02002136 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002137}
2138
2139/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002140 * Must be called with the queue_lock held.
2141 */
2142static int cfqq_process_refs(struct cfq_queue *cfqq)
2143{
2144 int process_refs, io_refs;
2145
2146 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002147 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002148 BUG_ON(process_refs < 0);
2149 return process_refs;
2150}
2151
2152static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2153{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002154 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002155 struct cfq_queue *__cfqq;
2156
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002157 /*
2158 * If there are no process references on the new_cfqq, then it is
2159 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2160 * chain may have dropped their last reference (not just their
2161 * last process reference).
2162 */
2163 if (!cfqq_process_refs(new_cfqq))
2164 return;
2165
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002166 /* Avoid a circular list and skip interim queue merges */
2167 while ((__cfqq = new_cfqq->new_cfqq)) {
2168 if (__cfqq == cfqq)
2169 return;
2170 new_cfqq = __cfqq;
2171 }
2172
2173 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002174 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002175 /*
2176 * If the process for the cfqq has gone away, there is no
2177 * sense in merging the queues.
2178 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002179 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002180 return;
2181
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002182 /*
2183 * Merge in the direction of the lesser amount of work.
2184 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002185 if (new_process_refs >= process_refs) {
2186 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002187 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002188 } else {
2189 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002190 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002191 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002192}
2193
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002194static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002195 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002196{
2197 struct cfq_queue *queue;
2198 int i;
2199 bool key_valid = false;
2200 unsigned long lowest_key = 0;
2201 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2202
Vivek Goyal65b32a52009-12-16 17:52:59 -05002203 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2204 /* select the one with lowest rb_key */
2205 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002206 if (queue &&
2207 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2208 lowest_key = queue->rb_key;
2209 cur_best = i;
2210 key_valid = true;
2211 }
2212 }
2213
2214 return cur_best;
2215}
2216
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002217static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002218{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002219 unsigned slice;
2220 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002221 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002222 unsigned group_slice;
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002223 enum wl_prio_t original_prio = cfqd->serving_prio;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002224
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002225 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002226 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002227 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002228 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002229 cfqd->serving_prio = BE_WORKLOAD;
2230 else {
2231 cfqd->serving_prio = IDLE_WORKLOAD;
2232 cfqd->workload_expires = jiffies + 1;
2233 return;
2234 }
2235
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002236 if (original_prio != cfqd->serving_prio)
2237 goto new_workload;
2238
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002239 /*
2240 * For RT and BE, we have to choose also the type
2241 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2242 * expiration time
2243 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002244 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002245 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002246
2247 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002248 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002249 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002250 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002251 return;
2252
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002253new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002254 /* otherwise select new workload type */
2255 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002256 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2257 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002258 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002259
2260 /*
2261 * the workload slice is computed as a fraction of target latency
2262 * proportional to the number of queues in that workload, over
2263 * all the queues in the same priority class
2264 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002265 group_slice = cfq_group_slice(cfqd, cfqg);
2266
2267 slice = group_slice * count /
2268 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2269 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002270
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002271 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2272 unsigned int tmp;
2273
2274 /*
2275 * Async queues are currently system wide. Just taking
2276 * proportion of queues with-in same group will lead to higher
2277 * async ratio system wide as generally root group is going
2278 * to have higher weight. A more accurate thing would be to
2279 * calculate system wide asnc/sync ratio.
2280 */
2281 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2282 tmp = tmp/cfqd->busy_queues;
2283 slice = min_t(unsigned, slice, tmp);
2284
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002285 /* async workload slice is scaled down according to
2286 * the sync/async slice ratio. */
2287 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002288 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002289 /* sync workload slice is at least 2 * cfq_slice_idle */
2290 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2291
2292 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002293 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002294 cfqd->workload_expires = jiffies + slice;
2295}
2296
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002297static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2298{
2299 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002300 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002301
2302 if (RB_EMPTY_ROOT(&st->rb))
2303 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002304 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002305 update_min_vdisktime(st);
2306 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002307}
2308
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002309static void cfq_choose_cfqg(struct cfq_data *cfqd)
2310{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002311 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2312
2313 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002314
2315 /* Restore the workload type data */
2316 if (cfqg->saved_workload_slice) {
2317 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2318 cfqd->serving_type = cfqg->saved_workload;
2319 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002320 } else
2321 cfqd->workload_expires = jiffies - 1;
2322
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002323 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002324}
2325
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002326/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002327 * Select a queue for service. If we have a current active queue,
2328 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002329 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002330static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002331{
Jens Axboea36e71f2009-04-15 12:15:11 +02002332 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002333
2334 cfqq = cfqd->active_queue;
2335 if (!cfqq)
2336 goto new_queue;
2337
Vivek Goyalf04a6422009-12-03 12:59:40 -05002338 if (!cfqd->rq_queued)
2339 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002340
2341 /*
2342 * We were waiting for group to get backlogged. Expire the queue
2343 */
2344 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2345 goto expire;
2346
Jens Axboe22e2c502005-06-27 10:55:12 +02002347 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002348 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002349 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002350 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2351 /*
2352 * If slice had not expired at the completion of last request
2353 * we might not have turned on wait_busy flag. Don't expire
2354 * the queue yet. Allow the group to get backlogged.
2355 *
2356 * The very fact that we have used the slice, that means we
2357 * have been idling all along on this queue and it should be
2358 * ok to wait for this request to complete.
2359 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002360 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2361 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2362 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002363 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002364 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002365 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002366 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002367
2368 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002369 * The active queue has requests and isn't expired, allow it to
2370 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002371 */
Jens Axboedd67d052006-06-21 09:36:18 +02002372 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002373 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002374
2375 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002376 * If another queue has a request waiting within our mean seek
2377 * distance, let it run. The expire code will check for close
2378 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002379 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002380 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002381 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002382 if (new_cfqq) {
2383 if (!cfqq->new_cfqq)
2384 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002385 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002386 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002387
2388 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002389 * No requests pending. If the active queue still has requests in
2390 * flight or is idling for a new request, allow either of these
2391 * conditions to happen (or time out) before selecting a new queue.
2392 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002393 if (timer_pending(&cfqd->idle_slice_timer)) {
2394 cfqq = NULL;
2395 goto keep_queue;
2396 }
2397
Shaohua Li8e1ac662010-11-08 15:01:04 +01002398 /*
2399 * This is a deep seek queue, but the device is much faster than
2400 * the queue can deliver, don't idle
2401 **/
2402 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2403 (cfq_cfqq_slice_new(cfqq) ||
2404 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2405 cfq_clear_cfqq_deep(cfqq);
2406 cfq_clear_cfqq_idle_window(cfqq);
2407 }
2408
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002409 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2410 cfqq = NULL;
2411 goto keep_queue;
2412 }
2413
2414 /*
2415 * If group idle is enabled and there are requests dispatched from
2416 * this group, wait for requests to complete.
2417 */
2418check_group_idle:
Shaohua Li7700fc42011-07-12 14:24:56 +02002419 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2420 cfqq->cfqg->dispatched &&
2421 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002422 cfqq = NULL;
2423 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002424 }
2425
Jens Axboe3b181522005-06-27 10:56:24 +02002426expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002427 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002428new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002429 /*
2430 * Current queue expired. Check if we have to switch to a new
2431 * service tree
2432 */
2433 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002434 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002435
Jens Axboea36e71f2009-04-15 12:15:11 +02002436 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002437keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002438 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002439}
2440
Jens Axboefebffd62008-01-28 13:19:43 +01002441static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002442{
2443 int dispatched = 0;
2444
2445 while (cfqq->next_rq) {
2446 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2447 dispatched++;
2448 }
2449
2450 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002451
2452 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002453 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002454 return dispatched;
2455}
2456
Jens Axboe498d3aa22007-04-26 12:54:48 +02002457/*
2458 * Drain our current requests. Used for barriers and when switching
2459 * io schedulers on-the-fly.
2460 */
Jens Axboed9e76202007-04-20 14:27:50 +02002461static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002462{
Jens Axboe08717142008-01-28 11:38:15 +01002463 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002464 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002465
Divyesh Shah3440c492010-04-09 09:29:57 +02002466 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002467 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002468 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2469 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002470 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002471 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002472
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002473 BUG_ON(cfqd->busy_queues);
2474
Jeff Moyer69237152009-06-12 15:29:30 +02002475 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002476 return dispatched;
2477}
2478
Shaohua Liabc3c742010-03-01 09:20:54 +01002479static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2480 struct cfq_queue *cfqq)
2481{
2482 /* the queue hasn't finished any request, can't estimate */
2483 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01002484 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002485 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2486 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01002487 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002488
Shaohua Lic1e44752010-11-08 15:01:02 +01002489 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01002490}
2491
Jens Axboe0b182d62009-10-06 20:49:37 +02002492static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002493{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002494 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495
Jens Axboe2f5cb732009-04-07 08:51:19 +02002496 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002497 * Drain async requests before we start sync IO
2498 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002499 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002500 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002501
2502 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002503 * If this is an async queue and we have sync IO in flight, let it wait
2504 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002505 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002506 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002507
Shaohua Liabc3c742010-03-01 09:20:54 +01002508 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002509 if (cfq_class_idle(cfqq))
2510 max_dispatch = 1;
2511
2512 /*
2513 * Does this cfqq already have too much IO in flight?
2514 */
2515 if (cfqq->dispatched >= max_dispatch) {
Shaohua Lief8a41d2011-03-07 09:26:29 +01002516 bool promote_sync = false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002517 /*
2518 * idle queue must always only have a single IO in flight
2519 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002520 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002521 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002522
Jens Axboe2f5cb732009-04-07 08:51:19 +02002523 /*
Li, Shaohuac4ade942011-03-23 08:30:34 +01002524 * If there is only one sync queue
2525 * we can ignore async queue here and give the sync
Shaohua Lief8a41d2011-03-07 09:26:29 +01002526 * queue no dispatch limit. The reason is a sync queue can
2527 * preempt async queue, limiting the sync queue doesn't make
2528 * sense. This is useful for aiostress test.
2529 */
Li, Shaohuac4ade942011-03-23 08:30:34 +01002530 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2531 promote_sync = true;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002532
2533 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002534 * We have other queues, don't allow more IO from this one
2535 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002536 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2537 !promote_sync)
Jens Axboe0b182d62009-10-06 20:49:37 +02002538 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002539
Jens Axboe2f5cb732009-04-07 08:51:19 +02002540 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002541 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002542 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002543 if (cfqd->busy_queues == 1 || promote_sync)
Shaohua Liabc3c742010-03-01 09:20:54 +01002544 max_dispatch = -1;
2545 else
2546 /*
2547 * Normally we start throttling cfqq when cfq_quantum/2
2548 * requests have been dispatched. But we can drive
2549 * deeper queue depths at the beginning of slice
2550 * subjected to upper limit of cfq_quantum.
2551 * */
2552 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002553 }
2554
2555 /*
2556 * Async queues must wait a bit before being allowed dispatch.
2557 * We also ramp up the dispatch depth gradually for async IO,
2558 * based on the last sync IO we serviced
2559 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002560 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002561 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002562 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002563
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002564 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002565 if (!depth && !cfqq->dispatched)
2566 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002567 if (depth < max_dispatch)
2568 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 }
2570
Jens Axboe0b182d62009-10-06 20:49:37 +02002571 /*
2572 * If we're below the current max, allow a dispatch
2573 */
2574 return cfqq->dispatched < max_dispatch;
2575}
2576
2577/*
2578 * Dispatch a request from cfqq, moving them to the request queue
2579 * dispatch list.
2580 */
2581static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2582{
2583 struct request *rq;
2584
2585 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2586
2587 if (!cfq_may_dispatch(cfqd, cfqq))
2588 return false;
2589
2590 /*
2591 * follow expired path, else get first next available
2592 */
2593 rq = cfq_check_fifo(cfqq);
2594 if (!rq)
2595 rq = cfqq->next_rq;
2596
2597 /*
2598 * insert request into driver dispatch list
2599 */
2600 cfq_dispatch_insert(cfqd->queue, rq);
2601
2602 if (!cfqd->active_cic) {
Tejun Heoc5869802011-12-14 00:33:41 +01002603 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe0b182d62009-10-06 20:49:37 +02002604
Tejun Heoc5869802011-12-14 00:33:41 +01002605 atomic_long_inc(&cic->icq.ioc->refcount);
Jens Axboe0b182d62009-10-06 20:49:37 +02002606 cfqd->active_cic = cic;
2607 }
2608
2609 return true;
2610}
2611
2612/*
2613 * Find the cfqq that we need to service and move a request from that to the
2614 * dispatch list
2615 */
2616static int cfq_dispatch_requests(struct request_queue *q, int force)
2617{
2618 struct cfq_data *cfqd = q->elevator->elevator_data;
2619 struct cfq_queue *cfqq;
2620
2621 if (!cfqd->busy_queues)
2622 return 0;
2623
2624 if (unlikely(force))
2625 return cfq_forced_dispatch(cfqd);
2626
2627 cfqq = cfq_select_queue(cfqd);
2628 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002629 return 0;
2630
Jens Axboe2f5cb732009-04-07 08:51:19 +02002631 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002632 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002633 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002634 if (!cfq_dispatch_request(cfqd, cfqq))
2635 return 0;
2636
Jens Axboe2f5cb732009-04-07 08:51:19 +02002637 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002638 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002639
2640 /*
2641 * expire an async queue immediately if it has used up its slice. idle
2642 * queue always expire after 1 dispatch round.
2643 */
2644 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2645 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2646 cfq_class_idle(cfqq))) {
2647 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002648 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002649 }
2650
Shan Weib217a902009-09-01 10:06:42 +02002651 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002652 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653}
2654
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655/*
Jens Axboe5e705372006-07-13 12:39:25 +02002656 * task holds one reference to the queue, dropped when task exits. each rq
2657 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002659 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 * queue lock must be held here.
2661 */
2662static void cfq_put_queue(struct cfq_queue *cfqq)
2663{
Jens Axboe22e2c502005-06-27 10:55:12 +02002664 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05002665 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002666
Shaohua Li30d7b942011-01-07 08:46:59 +01002667 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668
Shaohua Li30d7b942011-01-07 08:46:59 +01002669 cfqq->ref--;
2670 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 return;
2672
Jens Axboe7b679132008-05-30 12:23:07 +02002673 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002675 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002676 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002678 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002679 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002680 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002681 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002682
Vivek Goyalf04a6422009-12-03 12:59:40 -05002683 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002685 cfq_put_cfqg(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686}
2687
Shaohua Lid02a2c02010-05-25 10:16:53 +02002688static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002689{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002690 struct cfq_queue *__cfqq, *next;
2691
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002692 /*
2693 * If this queue was scheduled to merge with another queue, be
2694 * sure to drop the reference taken on that queue (and others in
2695 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2696 */
2697 __cfqq = cfqq->new_cfqq;
2698 while (__cfqq) {
2699 if (__cfqq == cfqq) {
2700 WARN(1, "cfqq->new_cfqq loop detected\n");
2701 break;
2702 }
2703 next = __cfqq->new_cfqq;
2704 cfq_put_queue(__cfqq);
2705 __cfqq = next;
2706 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002707}
2708
2709static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2710{
2711 if (unlikely(cfqq == cfqd->active_queue)) {
2712 __cfq_slice_expired(cfqd, cfqq, 0);
2713 cfq_schedule_dispatch(cfqd);
2714 }
2715
2716 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002717
Jens Axboe89850f72006-07-22 16:48:31 +02002718 cfq_put_queue(cfqq);
2719}
2720
Tejun Heo9b84cac2011-12-14 00:33:42 +01002721static void cfq_init_icq(struct io_cq *icq)
2722{
2723 struct cfq_io_cq *cic = icq_to_cic(icq);
2724
2725 cic->ttime.last_end_request = jiffies;
2726}
2727
Tejun Heoc5869802011-12-14 00:33:41 +01002728static void cfq_exit_icq(struct io_cq *icq)
Jens Axboe89850f72006-07-22 16:48:31 +02002729{
Tejun Heoc5869802011-12-14 00:33:41 +01002730 struct cfq_io_cq *cic = icq_to_cic(icq);
Tejun Heo283287a2011-12-14 00:33:38 +01002731 struct cfq_data *cfqd = cic_to_cfqd(cic);
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002732
Jens Axboeff6657c2009-04-08 10:58:57 +02002733 if (cic->cfqq[BLK_RW_ASYNC]) {
2734 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2735 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002736 }
2737
Jens Axboeff6657c2009-04-08 10:58:57 +02002738 if (cic->cfqq[BLK_RW_SYNC]) {
2739 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2740 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002741 }
Jens Axboe89850f72006-07-22 16:48:31 +02002742}
2743
Jens Axboefd0928d2008-01-24 08:52:45 +01002744static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002745{
2746 struct task_struct *tsk = current;
2747 int ioprio_class;
2748
Jens Axboe3b181522005-06-27 10:56:24 +02002749 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002750 return;
2751
Jens Axboefd0928d2008-01-24 08:52:45 +01002752 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002753 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002754 default:
2755 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2756 case IOPRIO_CLASS_NONE:
2757 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002758 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002759 */
2760 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002761 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002762 break;
2763 case IOPRIO_CLASS_RT:
2764 cfqq->ioprio = task_ioprio(ioc);
2765 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2766 break;
2767 case IOPRIO_CLASS_BE:
2768 cfqq->ioprio = task_ioprio(ioc);
2769 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2770 break;
2771 case IOPRIO_CLASS_IDLE:
2772 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2773 cfqq->ioprio = 7;
2774 cfq_clear_cfqq_idle_window(cfqq);
2775 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002776 }
2777
2778 /*
2779 * keep track of original prio settings in case we have to temporarily
2780 * elevate the priority of this queue
2781 */
2782 cfqq->org_ioprio = cfqq->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02002783 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002784}
2785
Tejun Heoc5869802011-12-14 00:33:41 +01002786static void changed_ioprio(struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002787{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002788 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002789 struct cfq_queue *cfqq;
Jens Axboe35e60772006-06-14 09:10:45 +02002790
Jens Axboecaaa5f92006-06-16 11:23:00 +02002791 if (unlikely(!cfqd))
2792 return;
2793
Jens Axboeff6657c2009-04-08 10:58:57 +02002794 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002795 if (cfqq) {
2796 struct cfq_queue *new_cfqq;
Tejun Heoc5869802011-12-14 00:33:41 +01002797 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
Jens Axboeff6657c2009-04-08 10:58:57 +02002798 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002799 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002800 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002801 cfq_put_queue(cfqq);
2802 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002803 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002804
Jens Axboeff6657c2009-04-08 10:58:57 +02002805 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002806 if (cfqq)
2807 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002808}
2809
Jens Axboed5036d72009-06-26 10:44:34 +02002810static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002811 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002812{
2813 RB_CLEAR_NODE(&cfqq->rb_node);
2814 RB_CLEAR_NODE(&cfqq->p_node);
2815 INIT_LIST_HEAD(&cfqq->fifo);
2816
Shaohua Li30d7b942011-01-07 08:46:59 +01002817 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02002818 cfqq->cfqd = cfqd;
2819
2820 cfq_mark_cfqq_prio_changed(cfqq);
2821
2822 if (is_sync) {
2823 if (!cfq_class_idle(cfqq))
2824 cfq_mark_cfqq_idle_window(cfqq);
2825 cfq_mark_cfqq_sync(cfqq);
2826 }
2827 cfqq->pid = pid;
2828}
2829
Vivek Goyal246103332009-12-03 12:59:51 -05002830#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heoc5869802011-12-14 00:33:41 +01002831static void changed_cgroup(struct cfq_io_cq *cic)
Vivek Goyal246103332009-12-03 12:59:51 -05002832{
2833 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002834 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal246103332009-12-03 12:59:51 -05002835 struct request_queue *q;
2836
2837 if (unlikely(!cfqd))
2838 return;
2839
2840 q = cfqd->queue;
2841
Vivek Goyal246103332009-12-03 12:59:51 -05002842 if (sync_cfqq) {
2843 /*
2844 * Drop reference to sync queue. A new sync queue will be
2845 * assigned in new group upon arrival of a fresh request.
2846 */
2847 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2848 cic_set_cfqq(cic, NULL, 1);
2849 cfq_put_queue(sync_cfqq);
2850 }
Vivek Goyal246103332009-12-03 12:59:51 -05002851}
Vivek Goyal246103332009-12-03 12:59:51 -05002852#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2853
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002855cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002856 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 struct cfq_queue *cfqq, *new_cfqq = NULL;
Tejun Heoc5869802011-12-14 00:33:41 +01002859 struct cfq_io_cq *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002860 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861
2862retry:
Vivek Goyal3e59cf92011-05-19 15:38:21 -04002863 cfqg = cfq_get_cfqg(cfqd);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002864 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002865 /* cic always exists here */
2866 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867
Jens Axboe6118b702009-06-30 09:34:12 +02002868 /*
2869 * Always try a new alloc if we fell back to the OOM cfqq
2870 * originally, since it should just be a temporary situation.
2871 */
2872 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2873 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 if (new_cfqq) {
2875 cfqq = new_cfqq;
2876 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002877 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002879 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002880 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002881 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002883 if (new_cfqq)
2884 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002885 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002886 cfqq = kmem_cache_alloc_node(cfq_pool,
2887 gfp_mask | __GFP_ZERO,
2888 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002889 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890
Jens Axboe6118b702009-06-30 09:34:12 +02002891 if (cfqq) {
2892 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2893 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002894 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002895 cfq_log_cfqq(cfqd, cfqq, "alloced");
2896 } else
2897 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 }
2899
2900 if (new_cfqq)
2901 kmem_cache_free(cfq_pool, new_cfqq);
2902
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 return cfqq;
2904}
2905
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002906static struct cfq_queue **
2907cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2908{
Jens Axboefe094d92008-01-31 13:08:54 +01002909 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002910 case IOPRIO_CLASS_RT:
2911 return &cfqd->async_cfqq[0][ioprio];
2912 case IOPRIO_CLASS_BE:
2913 return &cfqd->async_cfqq[1][ioprio];
2914 case IOPRIO_CLASS_IDLE:
2915 return &cfqd->async_idle_cfqq;
2916 default:
2917 BUG();
2918 }
2919}
2920
Jens Axboe15c31be2007-07-10 13:43:25 +02002921static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002922cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002923 gfp_t gfp_mask)
2924{
Jens Axboefd0928d2008-01-24 08:52:45 +01002925 const int ioprio = task_ioprio(ioc);
2926 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002927 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002928 struct cfq_queue *cfqq = NULL;
2929
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002930 if (!is_sync) {
2931 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2932 cfqq = *async_cfqq;
2933 }
2934
Jens Axboe6118b702009-06-30 09:34:12 +02002935 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002936 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002937
2938 /*
2939 * pin the queue now that it's allocated, scheduler exit will prune it
2940 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002941 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01002942 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002943 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002944 }
2945
Shaohua Li30d7b942011-01-07 08:46:59 +01002946 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02002947 return cfqq;
2948}
2949
Jens Axboe22e2c502005-06-27 10:55:12 +02002950static void
Shaohua Li383cd722011-07-12 14:24:35 +02002951__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02002952{
Shaohua Li383cd722011-07-12 14:24:35 +02002953 unsigned long elapsed = jiffies - ttime->last_end_request;
2954 elapsed = min(elapsed, 2UL * slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02002955
Shaohua Li383cd722011-07-12 14:24:35 +02002956 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
2957 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
2958 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
2959}
2960
2961static void
2962cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01002963 struct cfq_io_cq *cic)
Shaohua Li383cd722011-07-12 14:24:35 +02002964{
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002965 if (cfq_cfqq_sync(cfqq)) {
Shaohua Li383cd722011-07-12 14:24:35 +02002966 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002967 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
2968 cfqd->cfq_slice_idle);
2969 }
Shaohua Li7700fc42011-07-12 14:24:56 +02002970#ifdef CONFIG_CFQ_GROUP_IOSCHED
2971 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
2972#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02002973}
2974
Jens Axboe206dc692006-03-28 13:03:44 +02002975static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002976cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02002977 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02002978{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002979 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01002980 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002981 if (cfqq->last_request_pos) {
2982 if (cfqq->last_request_pos < blk_rq_pos(rq))
2983 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2984 else
2985 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2986 }
Jens Axboe206dc692006-03-28 13:03:44 +02002987
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002988 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01002989 if (blk_queue_nonrot(cfqd->queue))
2990 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
2991 else
2992 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02002993}
Jens Axboe22e2c502005-06-27 10:55:12 +02002994
2995/*
2996 * Disable idle window if the process thinks too long or seeks so much that
2997 * it doesn't matter
2998 */
2999static void
3000cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01003001 struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02003002{
Jens Axboe7b679132008-05-30 12:23:07 +02003003 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003004
Jens Axboe08717142008-01-28 11:38:15 +01003005 /*
3006 * Don't idle for async or idle io prio class
3007 */
3008 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003009 return;
3010
Jens Axboec265a7f2008-06-26 13:49:33 +02003011 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003012
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003013 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3014 cfq_mark_cfqq_deep(cfqq);
3015
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003016 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3017 enable_idle = 0;
Tejun Heoc5869802011-12-14 00:33:41 +01003018 else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
3019 !cfqd->cfq_slice_idle ||
3020 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003021 enable_idle = 0;
Shaohua Li383cd722011-07-12 14:24:35 +02003022 else if (sample_valid(cic->ttime.ttime_samples)) {
3023 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003024 enable_idle = 0;
3025 else
3026 enable_idle = 1;
3027 }
3028
Jens Axboe7b679132008-05-30 12:23:07 +02003029 if (old_idle != enable_idle) {
3030 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3031 if (enable_idle)
3032 cfq_mark_cfqq_idle_window(cfqq);
3033 else
3034 cfq_clear_cfqq_idle_window(cfqq);
3035 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003036}
3037
Jens Axboe22e2c502005-06-27 10:55:12 +02003038/*
3039 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3040 * no or if we aren't sure, a 1 will cause a preempt.
3041 */
Jens Axboea6151c32009-10-07 20:02:57 +02003042static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003043cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003044 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003045{
Jens Axboe6d048f52007-04-25 12:44:27 +02003046 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003047
Jens Axboe6d048f52007-04-25 12:44:27 +02003048 cfqq = cfqd->active_queue;
3049 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003050 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003051
Jens Axboe6d048f52007-04-25 12:44:27 +02003052 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003053 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003054
3055 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003056 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003057
Jens Axboe22e2c502005-06-27 10:55:12 +02003058 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003059 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3060 */
3061 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3062 return false;
3063
3064 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003065 * if the new request is sync, but the currently running queue is
3066 * not, let the sync request have priority.
3067 */
Jens Axboe5e705372006-07-13 12:39:25 +02003068 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003069 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003070
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003071 if (new_cfqq->cfqg != cfqq->cfqg)
3072 return false;
3073
3074 if (cfq_slice_used(cfqq))
3075 return true;
3076
3077 /* Allow preemption only if we are idling on sync-noidle tree */
3078 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3079 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3080 new_cfqq->service_tree->count == 2 &&
3081 RB_EMPTY_ROOT(&cfqq->sort_list))
3082 return true;
3083
Jens Axboe374f84a2006-07-23 01:42:19 +02003084 /*
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003085 * So both queues are sync. Let the new request get disk time if
3086 * it's a metadata request and the current queue is doing regular IO.
3087 */
Christoph Hellwig65299a32011-08-23 14:50:29 +02003088 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003089 return true;
3090
3091 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003092 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3093 */
3094 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003095 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003096
Shaohua Lid2d59e12010-11-08 15:01:03 +01003097 /* An idle queue should not be idle now for some reason */
3098 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3099 return true;
3100
Jens Axboe1e3335d2007-02-14 19:59:49 +01003101 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003102 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003103
3104 /*
3105 * if this request is as-good as one we would expect from the
3106 * current cfqq, let it preempt
3107 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003108 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003109 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003110
Jens Axboea6151c32009-10-07 20:02:57 +02003111 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003112}
3113
3114/*
3115 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3116 * let it have half of its nominal slice.
3117 */
3118static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3119{
Jens Axboe7b679132008-05-30 12:23:07 +02003120 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe22e2c502005-06-27 10:55:12 +02003121
Jens Axboebf572252006-07-19 20:29:12 +02003122 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003123 * workload type is changed, don't save slice, otherwise preempt
3124 * doesn't happen
3125 */
Jens Axboe54b466e2012-01-17 21:26:11 +01003126 if (cfqq_type(cfqd->active_queue) != cfqq_type(cfqq))
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003127 cfqq->cfqg->saved_workload_slice = 0;
3128
Jens Axboe54b466e2012-01-17 21:26:11 +01003129 cfq_slice_expired(cfqd, 1);
3130
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003131 /*
Jens Axboebf572252006-07-19 20:29:12 +02003132 * Put the new queue at the front of the of the current list,
3133 * so we know that it will be selected next.
3134 */
3135 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003136
3137 cfq_service_tree_add(cfqd, cfqq, 1);
Justin TerAvesteda5e0c2011-03-22 21:26:49 +01003138
Justin TerAvest62a37f62011-03-23 08:25:44 +01003139 cfqq->slice_end = 0;
3140 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003141}
3142
3143/*
Jens Axboe5e705372006-07-13 12:39:25 +02003144 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003145 * something we should do about it
3146 */
3147static void
Jens Axboe5e705372006-07-13 12:39:25 +02003148cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3149 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003150{
Tejun Heoc5869802011-12-14 00:33:41 +01003151 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003152
Aaron Carroll45333d52008-08-26 15:52:36 +02003153 cfqd->rq_queued++;
Christoph Hellwig65299a32011-08-23 14:50:29 +02003154 if (rq->cmd_flags & REQ_PRIO)
3155 cfqq->prio_pending++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003156
Shaohua Li383cd722011-07-12 14:24:35 +02003157 cfq_update_io_thinktime(cfqd, cfqq, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003158 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003159 cfq_update_idle_window(cfqd, cfqq, cic);
3160
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003161 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003162
3163 if (cfqq == cfqd->active_queue) {
3164 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003165 * Remember that we saw a request from this process, but
3166 * don't start queuing just yet. Otherwise we risk seeing lots
3167 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003168 * and merging. If the request is already larger than a single
3169 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003170 * merging is already done. Ditto for a busy system that
3171 * has other work pending, don't risk delaying until the
3172 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003173 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003174 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003175 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3176 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003177 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003178 cfq_clear_cfqq_wait_request(cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003179 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003180 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003181 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003182 &cfqq->cfqg->blkg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003183 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003184 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003185 }
Jens Axboe5e705372006-07-13 12:39:25 +02003186 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003187 /*
3188 * not the active queue - expire current slice if it is
3189 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003190 * has some old slice time left and is of higher priority or
3191 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003192 */
3193 cfq_preempt_queue(cfqd, cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003194 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003195 }
3196}
3197
Jens Axboe165125e2007-07-24 09:28:11 +02003198static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003199{
Jens Axboeb4878f22005-10-20 16:42:29 +02003200 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003201 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003202
Jens Axboe7b679132008-05-30 12:23:07 +02003203 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Tejun Heoc5869802011-12-14 00:33:41 +01003204 cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003205
Jens Axboe30996f42009-10-05 11:03:39 +02003206 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003207 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003208 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003209 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003210 &cfqd->serving_group->blkg, rq_data_dir(rq),
3211 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003212 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003213}
3214
Aaron Carroll45333d52008-08-26 15:52:36 +02003215/*
3216 * Update hw_tag based on peak queue depth over 50 samples under
3217 * sufficient load.
3218 */
3219static void cfq_update_hw_tag(struct cfq_data *cfqd)
3220{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003221 struct cfq_queue *cfqq = cfqd->active_queue;
3222
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003223 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3224 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003225
3226 if (cfqd->hw_tag == 1)
3227 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003228
3229 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003230 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003231 return;
3232
Shaohua Li1a1238a2009-10-27 08:46:23 +01003233 /*
3234 * If active queue hasn't enough requests and can idle, cfq might not
3235 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3236 * case
3237 */
3238 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3239 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003240 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003241 return;
3242
Aaron Carroll45333d52008-08-26 15:52:36 +02003243 if (cfqd->hw_tag_samples++ < 50)
3244 return;
3245
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003246 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003247 cfqd->hw_tag = 1;
3248 else
3249 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003250}
3251
Vivek Goyal7667aa02009-12-08 17:52:58 -05003252static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3253{
Tejun Heoc5869802011-12-14 00:33:41 +01003254 struct cfq_io_cq *cic = cfqd->active_cic;
Vivek Goyal7667aa02009-12-08 17:52:58 -05003255
Justin TerAvest02a8f012011-02-09 14:20:03 +01003256 /* If the queue already has requests, don't wait */
3257 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3258 return false;
3259
Vivek Goyal7667aa02009-12-08 17:52:58 -05003260 /* If there are other queues in the group, don't wait */
3261 if (cfqq->cfqg->nr_cfqq > 1)
3262 return false;
3263
Shaohua Li7700fc42011-07-12 14:24:56 +02003264 /* the only queue in the group, but think time is big */
3265 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3266 return false;
3267
Vivek Goyal7667aa02009-12-08 17:52:58 -05003268 if (cfq_slice_used(cfqq))
3269 return true;
3270
3271 /* if slice left is less than think time, wait busy */
Shaohua Li383cd722011-07-12 14:24:35 +02003272 if (cic && sample_valid(cic->ttime.ttime_samples)
3273 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
Vivek Goyal7667aa02009-12-08 17:52:58 -05003274 return true;
3275
3276 /*
3277 * If think times is less than a jiffy than ttime_mean=0 and above
3278 * will not be true. It might happen that slice has not expired yet
3279 * but will expire soon (4-5 ns) during select_queue(). To cover the
3280 * case where think time is less than a jiffy, mark the queue wait
3281 * busy if only 1 jiffy is left in the slice.
3282 */
3283 if (cfqq->slice_end - jiffies == 1)
3284 return true;
3285
3286 return false;
3287}
3288
Jens Axboe165125e2007-07-24 09:28:11 +02003289static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290{
Jens Axboe5e705372006-07-13 12:39:25 +02003291 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003292 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003293 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003294 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295
Jens Axboeb4878f22005-10-20 16:42:29 +02003296 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003297 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3298 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299
Aaron Carroll45333d52008-08-26 15:52:36 +02003300 cfq_update_hw_tag(cfqd);
3301
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003302 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003303 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003304 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003305 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003306 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003307 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3308 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3309 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003311 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003312
Vivek Goyal365722b2009-10-03 15:21:27 +02003313 if (sync) {
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003314 struct cfq_rb_root *service_tree;
3315
Shaohua Li383cd722011-07-12 14:24:35 +02003316 RQ_CIC(rq)->ttime.last_end_request = now;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003317
3318 if (cfq_cfqq_on_rr(cfqq))
3319 service_tree = cfqq->service_tree;
3320 else
3321 service_tree = service_tree_for(cfqq->cfqg,
3322 cfqq_prio(cfqq), cfqq_type(cfqq));
3323 service_tree->ttime.last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003324 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3325 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003326 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003327
Shaohua Li7700fc42011-07-12 14:24:56 +02003328#ifdef CONFIG_CFQ_GROUP_IOSCHED
3329 cfqq->cfqg->ttime.last_end_request = now;
3330#endif
3331
Jens Axboecaaa5f92006-06-16 11:23:00 +02003332 /*
3333 * If this is the active queue, check if it needs to be expired,
3334 * or if we want to idle in case it has no pending requests.
3335 */
3336 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003337 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3338
Jens Axboe44f7c162007-01-19 11:51:58 +11003339 if (cfq_cfqq_slice_new(cfqq)) {
3340 cfq_set_prio_slice(cfqd, cfqq);
3341 cfq_clear_cfqq_slice_new(cfqq);
3342 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003343
3344 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003345 * Should we wait for next request to come in before we expire
3346 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003347 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003348 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003349 unsigned long extend_sl = cfqd->cfq_slice_idle;
3350 if (!cfqd->cfq_slice_idle)
3351 extend_sl = cfqd->cfq_group_idle;
3352 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003353 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003354 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003355 }
3356
Jens Axboea36e71f2009-04-15 12:15:11 +02003357 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003358 * Idling is not enabled on:
3359 * - expired queues
3360 * - idle-priority queues
3361 * - async queues
3362 * - queues with still some requests queued
3363 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003364 */
Jens Axboe08717142008-01-28 11:38:15 +01003365 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003366 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003367 else if (sync && cfqq_empty &&
3368 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003369 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003370 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003371 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003372
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003373 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003374 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375}
3376
Jens Axboe89850f72006-07-22 16:48:31 +02003377static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003378{
Jens Axboe1b379d82009-08-11 08:26:11 +02003379 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003380 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003381 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003382 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003383
3384 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003385}
3386
Jens Axboe165125e2007-07-24 09:28:11 +02003387static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003388{
3389 struct cfq_data *cfqd = q->elevator->elevator_data;
3390 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01003391 struct cfq_io_cq *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003392 struct cfq_queue *cfqq;
3393
3394 /*
3395 * don't force setup of a queue from here, as a call to may_queue
3396 * does not necessarily imply that a request actually will be queued.
3397 * so just lookup a possibly existing queue, or return 'may queue'
3398 * if that fails
3399 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003400 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003401 if (!cic)
3402 return ELV_MQUEUE_MAY;
3403
Jens Axboeb0b78f82009-04-08 10:56:08 +02003404 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003405 if (cfqq) {
Tejun Heoc5869802011-12-14 00:33:41 +01003406 cfq_init_prio_data(cfqq, cic->icq.ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003407
Jens Axboe89850f72006-07-22 16:48:31 +02003408 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003409 }
3410
3411 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412}
3413
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414/*
3415 * queue lock held here
3416 */
Jens Axboebb37b942006-12-01 10:42:33 +01003417static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418{
Jens Axboe5e705372006-07-13 12:39:25 +02003419 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420
Jens Axboe5e705372006-07-13 12:39:25 +02003421 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003422 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423
Jens Axboe22e2c502005-06-27 10:55:12 +02003424 BUG_ON(!cfqq->allocated[rw]);
3425 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003427 /* Put down rq reference on cfqg */
3428 cfq_put_cfqg(RQ_CFQG(rq));
Tejun Heoa612fdd2011-12-14 00:33:41 +01003429 rq->elv.priv[0] = NULL;
3430 rq->elv.priv[1] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003431
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432 cfq_put_queue(cfqq);
3433 }
3434}
3435
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003436static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01003437cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003438 struct cfq_queue *cfqq)
3439{
3440 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3441 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003442 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003443 cfq_put_queue(cfqq);
3444 return cic_to_cfqq(cic, 1);
3445}
3446
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003447/*
3448 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3449 * was the last process referring to said cfqq.
3450 */
3451static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01003452split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003453{
3454 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003455 cfqq->pid = current->pid;
3456 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003457 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003458 return cfqq;
3459 }
3460
3461 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003462
3463 cfq_put_cooperator(cfqq);
3464
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003465 cfq_put_queue(cfqq);
3466 return NULL;
3467}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003469 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003471static int
Jens Axboe165125e2007-07-24 09:28:11 +02003472cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473{
3474 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heof1f8cc92011-12-14 00:33:42 +01003475 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003477 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003478 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479
3480 might_sleep_if(gfp_mask & __GFP_WAIT);
3481
Tejun Heo216284c2011-12-14 00:33:38 +01003482 spin_lock_irq(q->queue_lock);
Tejun Heof1f8cc92011-12-14 00:33:42 +01003483
3484 /* handle changed notifications */
3485 if (unlikely(cic->icq.changed)) {
3486 if (test_and_clear_bit(ICQ_IOPRIO_CHANGED, &cic->icq.changed))
3487 changed_ioprio(cic);
3488#ifdef CONFIG_CFQ_GROUP_IOSCHED
3489 if (test_and_clear_bit(ICQ_CGROUP_CHANGED, &cic->icq.changed))
3490 changed_cgroup(cic);
3491#endif
3492 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003494new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003495 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003496 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Tejun Heoc5869802011-12-14 00:33:41 +01003497 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003498 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003499 } else {
3500 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003501 * If the queue was seeky for too long, break it apart.
3502 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003503 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003504 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3505 cfqq = split_cfqq(cic, cfqq);
3506 if (!cfqq)
3507 goto new_queue;
3508 }
3509
3510 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003511 * Check to see if this queue is scheduled to merge with
3512 * another, closely cooperating queue. The merging of
3513 * queues happens here as it must be done in process context.
3514 * The reference on new_cfqq was taken in merge_cfqqs.
3515 */
3516 if (cfqq->new_cfqq)
3517 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003518 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519
3520 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02003521
Jens Axboe6fae9c22011-03-01 15:04:39 -05003522 cfqq->ref++;
Tejun Heoa612fdd2011-12-14 00:33:41 +01003523 rq->elv.priv[0] = cfqq;
3524 rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
Tejun Heo216284c2011-12-14 00:33:38 +01003525 spin_unlock_irq(q->queue_lock);
Jens Axboe5e705372006-07-13 12:39:25 +02003526 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527}
3528
David Howells65f27f32006-11-22 14:55:48 +00003529static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003530{
David Howells65f27f32006-11-22 14:55:48 +00003531 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003532 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003533 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003534
Jens Axboe40bb54d2009-04-15 12:11:10 +02003535 spin_lock_irq(q->queue_lock);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003536 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003537 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003538}
3539
3540/*
3541 * Timer running if the active_queue is currently idling inside its time slice
3542 */
3543static void cfq_idle_slice_timer(unsigned long data)
3544{
3545 struct cfq_data *cfqd = (struct cfq_data *) data;
3546 struct cfq_queue *cfqq;
3547 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003548 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003549
Jens Axboe7b679132008-05-30 12:23:07 +02003550 cfq_log(cfqd, "idle timer fired");
3551
Jens Axboe22e2c502005-06-27 10:55:12 +02003552 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3553
Jens Axboefe094d92008-01-31 13:08:54 +01003554 cfqq = cfqd->active_queue;
3555 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003556 timed_out = 0;
3557
Jens Axboe22e2c502005-06-27 10:55:12 +02003558 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003559 * We saw a request before the queue expired, let it through
3560 */
3561 if (cfq_cfqq_must_dispatch(cfqq))
3562 goto out_kick;
3563
3564 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003565 * expired
3566 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003567 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003568 goto expire;
3569
3570 /*
3571 * only expire and reinvoke request handler, if there are
3572 * other queues with pending requests
3573 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003574 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003575 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003576
3577 /*
3578 * not expired and it has a request pending, let it dispatch
3579 */
Jens Axboe75e50982009-04-07 08:56:14 +02003580 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003581 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003582
3583 /*
3584 * Queue depth flag is reset only when the idle didn't succeed
3585 */
3586 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003587 }
3588expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003589 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003590out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003591 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003592out_cont:
3593 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3594}
3595
Jens Axboe3b181522005-06-27 10:56:24 +02003596static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3597{
3598 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003599 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003600}
Jens Axboe22e2c502005-06-27 10:55:12 +02003601
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003602static void cfq_put_async_queues(struct cfq_data *cfqd)
3603{
3604 int i;
3605
3606 for (i = 0; i < IOPRIO_BE_NR; i++) {
3607 if (cfqd->async_cfqq[0][i])
3608 cfq_put_queue(cfqd->async_cfqq[0][i]);
3609 if (cfqd->async_cfqq[1][i])
3610 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003611 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003612
3613 if (cfqd->async_idle_cfqq)
3614 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003615}
3616
Jens Axboeb374d182008-10-31 10:05:07 +01003617static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618{
Jens Axboe22e2c502005-06-27 10:55:12 +02003619 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003620 struct request_queue *q = cfqd->queue;
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003621 bool wait = false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003622
Jens Axboe3b181522005-06-27 10:56:24 +02003623 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003624
Al Virod9ff4182006-03-18 13:51:22 -05003625 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003626
Al Virod9ff4182006-03-18 13:51:22 -05003627 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003628 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003629
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003630 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003631 cfq_release_cfq_groups(cfqd);
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003632
3633 /*
3634 * If there are groups which we could not unlink from blkcg list,
3635 * wait for a rcu period for them to be freed.
3636 */
3637 if (cfqd->nr_blkcg_linked_grps)
3638 wait = true;
Jens Axboe15c31be2007-07-10 13:43:25 +02003639
Al Virod9ff4182006-03-18 13:51:22 -05003640 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003641
3642 cfq_shutdown_timer_wq(cfqd);
3643
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003644 /*
3645 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3646 * Do this wait only if there are other unlinked groups out
3647 * there. This can happen if cgroup deletion path claimed the
3648 * responsibility of cleaning up a group before queue cleanup code
3649 * get to the group.
3650 *
3651 * Do not call synchronize_rcu() unconditionally as there are drivers
3652 * which create/delete request queue hundreds of times during scan/boot
3653 * and synchronize_rcu() can take significant time and slow down boot.
3654 */
3655 if (wait)
3656 synchronize_rcu();
Vivek Goyal2abae552011-05-23 10:02:19 +02003657
3658#ifdef CONFIG_CFQ_GROUP_IOSCHED
3659 /* Free up per cpu stats for root group */
3660 free_percpu(cfqd->root_group.blkg.stats_cpu);
3661#endif
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003662 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003663}
3664
Jens Axboe165125e2007-07-24 09:28:11 +02003665static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666{
3667 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003668 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003669 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003670 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671
Christoph Lameter94f60302007-07-17 04:03:29 -07003672 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Tejun Heoa73f7302011-12-14 00:33:37 +01003673 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003674 return NULL;
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003675
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003676 /* Init root service tree */
3677 cfqd->grp_service_tree = CFQ_RB_ROOT;
3678
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003679 /* Init root group */
3680 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003681 for_each_cfqg_st(cfqg, i, j, st)
3682 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003683 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003684
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003685 /* Give preference to root group over other groups */
3686 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3687
Vivek Goyal25fb5162009-12-03 12:59:46 -05003688#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003689 /*
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003690 * Set root group reference to 2. One reference will be dropped when
3691 * all groups on cfqd->cfqg_list are being deleted during queue exit.
3692 * Other reference will remain there as we don't want to delete this
3693 * group as it is statically allocated and gets destroyed when
3694 * throtl_data goes away.
Vivek Goyalb1c35762009-12-03 12:59:47 -05003695 */
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003696 cfqg->ref = 2;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04003697
3698 if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3699 kfree(cfqg);
3700 kfree(cfqd);
3701 return NULL;
3702 }
3703
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003704 rcu_read_lock();
Vivek Goyal5624a4e2011-05-19 15:38:28 -04003705
Vivek Goyale98ef892010-06-18 10:39:47 -04003706 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3707 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003708 rcu_read_unlock();
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003709 cfqd->nr_blkcg_linked_grps++;
3710
3711 /* Add group on cfqd->cfqg_list */
3712 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
Vivek Goyal25fb5162009-12-03 12:59:46 -05003713#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003714 /*
3715 * Not strictly needed (since RB_ROOT just clears the node and we
3716 * zeroed cfqd on alloc), but better be safe in case someone decides
3717 * to add magic to the rb code
3718 */
3719 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3720 cfqd->prio_trees[i] = RB_ROOT;
3721
Jens Axboe6118b702009-06-30 09:34:12 +02003722 /*
3723 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3724 * Grab a permanent reference to it, so that the normal code flow
3725 * will not attempt to free it.
3726 */
3727 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01003728 cfqd->oom_cfqq.ref++;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003729 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003730
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732
Jens Axboe22e2c502005-06-27 10:55:12 +02003733 init_timer(&cfqd->idle_slice_timer);
3734 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3735 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3736
Jens Axboe23e018a2009-10-05 08:52:35 +02003737 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003738
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003740 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3741 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003742 cfqd->cfq_back_max = cfq_back_max;
3743 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003744 cfqd->cfq_slice[0] = cfq_slice_async;
3745 cfqd->cfq_slice[1] = cfq_slice_sync;
3746 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3747 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003748 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003749 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003750 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003751 /*
3752 * we optimistically start assuming sync ops weren't delayed in last
3753 * second, in order to have larger depth for async operations.
3754 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003755 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003756 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757}
3758
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759/*
3760 * sysfs parts below -->
3761 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762static ssize_t
3763cfq_var_show(unsigned int var, char *page)
3764{
3765 return sprintf(page, "%d\n", var);
3766}
3767
3768static ssize_t
3769cfq_var_store(unsigned int *var, const char *page, size_t count)
3770{
3771 char *p = (char *) page;
3772
3773 *var = simple_strtoul(p, &p, 10);
3774 return count;
3775}
3776
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003778static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003780 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003781 unsigned int __data = __VAR; \
3782 if (__CONV) \
3783 __data = jiffies_to_msecs(__data); \
3784 return cfq_var_show(__data, (page)); \
3785}
3786SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003787SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3788SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003789SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3790SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003791SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003792SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003793SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3794SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3795SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003796SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797#undef SHOW_FUNCTION
3798
3799#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003800static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003802 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803 unsigned int __data; \
3804 int ret = cfq_var_store(&__data, (page), count); \
3805 if (__data < (MIN)) \
3806 __data = (MIN); \
3807 else if (__data > (MAX)) \
3808 __data = (MAX); \
3809 if (__CONV) \
3810 *(__PTR) = msecs_to_jiffies(__data); \
3811 else \
3812 *(__PTR) = __data; \
3813 return ret; \
3814}
3815STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003816STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3817 UINT_MAX, 1);
3818STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3819 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05003820STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003821STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3822 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003823STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003824STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003825STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3826STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01003827STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3828 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003829STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830#undef STORE_FUNCTION
3831
Al Viroe572ec72006-03-18 22:27:18 -05003832#define CFQ_ATTR(name) \
3833 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02003834
Al Viroe572ec72006-03-18 22:27:18 -05003835static struct elv_fs_entry cfq_attrs[] = {
3836 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05003837 CFQ_ATTR(fifo_expire_sync),
3838 CFQ_ATTR(fifo_expire_async),
3839 CFQ_ATTR(back_seek_max),
3840 CFQ_ATTR(back_seek_penalty),
3841 CFQ_ATTR(slice_sync),
3842 CFQ_ATTR(slice_async),
3843 CFQ_ATTR(slice_async_rq),
3844 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003845 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02003846 CFQ_ATTR(low_latency),
Al Viroe572ec72006-03-18 22:27:18 -05003847 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848};
3849
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850static struct elevator_type iosched_cfq = {
3851 .ops = {
3852 .elevator_merge_fn = cfq_merge,
3853 .elevator_merged_fn = cfq_merged_request,
3854 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01003855 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07003856 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02003857 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02003859 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860 .elevator_deactivate_req_fn = cfq_deactivate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02003862 .elevator_former_req_fn = elv_rb_former_request,
3863 .elevator_latter_req_fn = elv_rb_latter_request,
Tejun Heo9b84cac2011-12-14 00:33:42 +01003864 .elevator_init_icq_fn = cfq_init_icq,
Tejun Heo7e5a8792011-12-14 00:33:42 +01003865 .elevator_exit_icq_fn = cfq_exit_icq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866 .elevator_set_req_fn = cfq_set_request,
3867 .elevator_put_req_fn = cfq_put_request,
3868 .elevator_may_queue_fn = cfq_may_queue,
3869 .elevator_init_fn = cfq_init_queue,
3870 .elevator_exit_fn = cfq_exit_queue,
3871 },
Tejun Heo3d3c2372011-12-14 00:33:42 +01003872 .icq_size = sizeof(struct cfq_io_cq),
3873 .icq_align = __alignof__(struct cfq_io_cq),
Al Viro3d1ab402006-03-18 18:35:43 -05003874 .elevator_attrs = cfq_attrs,
Tejun Heo3d3c2372011-12-14 00:33:42 +01003875 .elevator_name = "cfq",
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876 .elevator_owner = THIS_MODULE,
3877};
3878
Vivek Goyal3e252062009-12-04 10:36:42 -05003879#ifdef CONFIG_CFQ_GROUP_IOSCHED
3880static struct blkio_policy_type blkio_policy_cfq = {
3881 .ops = {
3882 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3883 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3884 },
Vivek Goyal062a6442010-09-15 17:06:33 -04003885 .plid = BLKIO_POLICY_PROP,
Vivek Goyal3e252062009-12-04 10:36:42 -05003886};
3887#else
3888static struct blkio_policy_type blkio_policy_cfq;
3889#endif
3890
Linus Torvalds1da177e2005-04-16 15:20:36 -07003891static int __init cfq_init(void)
3892{
Tejun Heo3d3c2372011-12-14 00:33:42 +01003893 int ret;
3894
Jens Axboe22e2c502005-06-27 10:55:12 +02003895 /*
3896 * could be 0 on HZ < 1000 setups
3897 */
3898 if (!cfq_slice_async)
3899 cfq_slice_async = 1;
3900 if (!cfq_slice_idle)
3901 cfq_slice_idle = 1;
3902
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003903#ifdef CONFIG_CFQ_GROUP_IOSCHED
3904 if (!cfq_group_idle)
3905 cfq_group_idle = 1;
3906#else
3907 cfq_group_idle = 0;
3908#endif
Tejun Heo3d3c2372011-12-14 00:33:42 +01003909 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3910 if (!cfq_pool)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 return -ENOMEM;
3912
Tejun Heo3d3c2372011-12-14 00:33:42 +01003913 ret = elv_register(&iosched_cfq);
3914 if (ret) {
3915 kmem_cache_destroy(cfq_pool);
3916 return ret;
3917 }
Tejun Heo3d3c2372011-12-14 00:33:42 +01003918
Vivek Goyal3e252062009-12-04 10:36:42 -05003919 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01003921 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922}
3923
3924static void __exit cfq_exit(void)
3925{
Vivek Goyal3e252062009-12-04 10:36:42 -05003926 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003927 elv_unregister(&iosched_cfq);
Tejun Heo3d3c2372011-12-14 00:33:42 +01003928 kmem_cache_destroy(cfq_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003929}
3930
3931module_init(cfq_init);
3932module_exit(cfq_exit);
3933
3934MODULE_AUTHOR("Jens Axboe");
3935MODULE_LICENSE("GPL");
3936MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");