blob: b4711cbfb4932fe3be0bddd53c6e1220b90be012 [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
Tatyana Brokhmanb410a822013-07-07 17:33:05 +0300242 unsigned int nr_urgent_pending;
243 unsigned int nr_urgent_in_flight;
244
Jens Axboea36e71f2009-04-15 12:15:11 +0200245 /*
246 * Each priority tree is sorted by next_request position. These
247 * trees are used when determining if two or more queues are
248 * interleaving requests (see cfq_close_cooperator).
249 */
250 struct rb_root prio_trees[CFQ_PRIO_LISTS];
251
Jens Axboe22e2c502005-06-27 10:55:12 +0200252 unsigned int busy_queues;
Shaohua Lief8a41d2011-03-07 09:26:29 +0100253 unsigned int busy_sync_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +0200254
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100255 int rq_in_driver;
256 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200257
258 /*
259 * queue-depth detection
260 */
261 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200262 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100263 /*
264 * hw_tag can be
265 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
266 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
267 * 0 => no NCQ
268 */
269 int hw_tag_est_depth;
270 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200271
272 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200273 * idle window management
274 */
275 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200276 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200277
278 struct cfq_queue *active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +0100279 struct cfq_io_cq *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200280
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200281 /*
282 * async queue for each priority case
283 */
284 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
285 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200286
Jens Axboe6d048f52007-04-25 12:44:27 +0200287 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 /*
290 * tunables, see top of file
291 */
292 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200293 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 unsigned int cfq_back_penalty;
295 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200296 unsigned int cfq_slice[2];
297 unsigned int cfq_slice_async_rq;
298 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200299 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200300 unsigned int cfq_latency;
Tao Ma5bf14c02012-04-01 14:33:39 -0700301 unsigned int cfq_target_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500302
Jens Axboe6118b702009-06-30 09:34:12 +0200303 /*
304 * Fallback dummy cfqq for extreme OOM conditions
305 */
306 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200307
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100308 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500309
310 /* List of cfq groups being managed on this device*/
311 struct hlist_head cfqg_list;
Vivek Goyal56edf7d2011-05-19 15:38:22 -0400312
313 /* Number of groups which are on blkcg->blkg_list */
314 unsigned int nr_blkcg_linked_grps;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315};
316
Vivek Goyal25fb5162009-12-03 12:59:46 -0500317static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
318
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500319static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
320 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500321 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100322{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500323 if (!cfqg)
324 return NULL;
325
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100326 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500327 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100328
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500329 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100330}
331
Jens Axboe3b181522005-06-27 10:56:24 +0200332enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100333 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
334 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200335 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100336 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100337 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
338 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
339 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100340 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200341 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400342 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100343 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100344 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500345 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200346};
347
348#define CFQ_CFQQ_FNS(name) \
349static inline void cfq_mark_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 void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
354{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100355 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200356} \
357static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
358{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100359 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200360}
361
362CFQ_CFQQ_FNS(on_rr);
363CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200364CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200365CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200366CFQ_CFQQ_FNS(fifo_expire);
367CFQ_CFQQ_FNS(idle_window);
368CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100369CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200370CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200371CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100372CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100373CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500374CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200375#undef CFQ_CFQQ_FNS
376
Vivek Goyalafc24d42010-04-26 19:27:56 +0200377#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500378#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
379 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
380 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200381 blkg_path(&(cfqq)->cfqg->blkg), ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500382
383#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
384 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200385 blkg_path(&(cfqg)->blkg), ##args) \
Vivek Goyal2868ef72009-12-03 12:59:48 -0500386
387#else
Jens Axboe7b679132008-05-30 12:23:07 +0200388#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
389 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200390#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500391#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200392#define cfq_log(cfqd, fmt, args...) \
393 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
394
Vivek Goyal615f0252009-12-03 12:59:39 -0500395/* Traverses through cfq group service trees */
396#define for_each_cfqg_st(cfqg, i, j, st) \
397 for (i = 0; i <= IDLE_WORKLOAD; i++) \
398 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
399 : &cfqg->service_tree_idle; \
400 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
401 (i == IDLE_WORKLOAD && j == 0); \
402 j++, st = i < IDLE_WORKLOAD ? \
403 &cfqg->service_trees[i][j]: NULL) \
404
Shaohua Lif5f2b6c2011-07-12 14:24:55 +0200405static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
406 struct cfq_ttime *ttime, bool group_idle)
407{
408 unsigned long slice;
409 if (!sample_valid(ttime->ttime_samples))
410 return false;
411 if (group_idle)
412 slice = cfqd->cfq_group_idle;
413 else
414 slice = cfqd->cfq_slice_idle;
415 return ttime->ttime_mean > slice;
416}
Vivek Goyal615f0252009-12-03 12:59:39 -0500417
Vivek Goyal02b35082010-08-23 12:23:53 +0200418static inline bool iops_mode(struct cfq_data *cfqd)
419{
420 /*
421 * If we are not idling on queues and it is a NCQ drive, parallel
422 * execution of requests is on and measuring time is not possible
423 * in most of the cases until and unless we drive shallower queue
424 * depths and that becomes a performance bottleneck. In such cases
425 * switch to start providing fairness in terms of number of IOs.
426 */
427 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
428 return true;
429 else
430 return false;
431}
432
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100433static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
434{
435 if (cfq_class_idle(cfqq))
436 return IDLE_WORKLOAD;
437 if (cfq_class_rt(cfqq))
438 return RT_WORKLOAD;
439 return BE_WORKLOAD;
440}
441
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100442
443static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
444{
445 if (!cfq_cfqq_sync(cfqq))
446 return ASYNC_WORKLOAD;
447 if (!cfq_cfqq_idle_window(cfqq))
448 return SYNC_NOIDLE_WORKLOAD;
449 return SYNC_WORKLOAD;
450}
451
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500452static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
453 struct cfq_data *cfqd,
454 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100455{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500456 if (wl == IDLE_WORKLOAD)
457 return cfqg->service_tree_idle.count;
458
459 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
460 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
461 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100462}
463
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500464static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
465 struct cfq_group *cfqg)
466{
467 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
468 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
469}
470
Jens Axboe165125e2007-07-24 09:28:11 +0200471static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200472static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100473 struct io_context *, gfp_t);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200474
Tejun Heoc5869802011-12-14 00:33:41 +0100475static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
476{
477 /* cic->icq is the first member, %NULL will convert to %NULL */
478 return container_of(icq, struct cfq_io_cq, icq);
479}
480
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100481static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
482 struct io_context *ioc)
483{
484 if (ioc)
485 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
486 return NULL;
487}
488
Tejun Heoc5869802011-12-14 00:33:41 +0100489static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200490{
Jens Axboea6151c32009-10-07 20:02:57 +0200491 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200492}
493
Tejun Heoc5869802011-12-14 00:33:41 +0100494static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
495 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200496{
Jens Axboea6151c32009-10-07 20:02:57 +0200497 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200498}
499
Tejun Heoc5869802011-12-14 00:33:41 +0100500static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400501{
Tejun Heoc5869802011-12-14 00:33:41 +0100502 return cic->icq.q->elevator->elevator_data;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400503}
504
Vasily Tarasov91fac312007-04-25 12:29:51 +0200505/*
506 * We regard a request as SYNC, if it's either a read or has the SYNC bit
507 * set (in which case it could also be direct WRITE).
508 */
Jens Axboea6151c32009-10-07 20:02:57 +0200509static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200510{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200511 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200512}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700515 * scheduler run of queue, if there are requests pending and no one in the
516 * driver that will restart queueing
517 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200518static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700519{
Jens Axboe7b679132008-05-30 12:23:07 +0200520 if (cfqd->busy_queues) {
521 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200522 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200523 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700524}
525
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100527 * Scale schedule slice based on io priority. Use the sync time slice only
528 * if a queue is marked sync and has sync io queued. A sync queue with async
529 * io only, should not get full sync slice length.
530 */
Jens Axboea6151c32009-10-07 20:02:57 +0200531static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200532 unsigned short prio)
533{
534 const int base_slice = cfqd->cfq_slice[sync];
535
536 WARN_ON(prio >= IOPRIO_BE_NR);
537
538 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
539}
540
Jens Axboe44f7c162007-01-19 11:51:58 +1100541static inline int
542cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
543{
Jens Axboed9e76202007-04-20 14:27:50 +0200544 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100545}
546
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500547static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
548{
549 u64 d = delta << CFQ_SERVICE_SHIFT;
550
551 d = d * BLKIO_WEIGHT_DEFAULT;
552 do_div(d, cfqg->weight);
553 return d;
554}
555
556static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
557{
558 s64 delta = (s64)(vdisktime - min_vdisktime);
559 if (delta > 0)
560 min_vdisktime = vdisktime;
561
562 return min_vdisktime;
563}
564
565static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
566{
567 s64 delta = (s64)(vdisktime - min_vdisktime);
568 if (delta < 0)
569 min_vdisktime = vdisktime;
570
571 return min_vdisktime;
572}
573
574static void update_min_vdisktime(struct cfq_rb_root *st)
575{
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500576 struct cfq_group *cfqg;
577
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500578 if (st->left) {
579 cfqg = rb_entry_cfqg(st->left);
Gui Jianfenga6032712011-03-07 09:28:09 +0100580 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
581 cfqg->vdisktime);
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500582 }
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500583}
584
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100585/*
586 * get averaged number of queues of RT/BE priority.
587 * average is updated, with a formula that gives more weight to higher numbers,
588 * to quickly follows sudden increases and decrease slowly
589 */
590
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500591static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
592 struct cfq_group *cfqg, bool rt)
Jens Axboe58696192009-10-28 09:27:07 +0100593{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100594 unsigned min_q, max_q;
595 unsigned mult = cfq_hist_divisor - 1;
596 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500597 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100598
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500599 min_q = min(cfqg->busy_queues_avg[rt], busy);
600 max_q = max(cfqg->busy_queues_avg[rt], busy);
601 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100602 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500603 return cfqg->busy_queues_avg[rt];
604}
605
606static inline unsigned
607cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
608{
609 struct cfq_rb_root *st = &cfqd->grp_service_tree;
610
Tao Ma5bf14c02012-04-01 14:33:39 -0700611 return cfqd->cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100612}
613
Shaohua Lic553f8e2011-01-14 08:41:03 +0100614static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -0700615cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100616{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100617 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
618 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500619 /*
620 * interested queues (we consider only the ones with the same
621 * priority class in the cfq group)
622 */
623 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
624 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100625 unsigned sync_slice = cfqd->cfq_slice[1];
626 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500627 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
628
629 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100630 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
631 /* scale low_slice according to IO priority
632 * and sync vs async */
633 unsigned low_slice =
634 min(slice, base_low_slice * slice / sync_slice);
635 /* the adapted slice value is scaled to fit all iqs
636 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500637 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100638 low_slice);
639 }
640 }
Shaohua Lic553f8e2011-01-14 08:41:03 +0100641 return slice;
642}
643
644static inline void
645cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
646{
Vivek Goyalba5bd522011-01-19 08:25:02 -0700647 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +0100648
Vivek Goyaldae739e2009-12-03 12:59:45 -0500649 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100650 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500651 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200652 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100653}
654
655/*
656 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
657 * isn't valid until the first request from the dispatch is activated
658 * and the slice time set.
659 */
Jens Axboea6151c32009-10-07 20:02:57 +0200660static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100661{
662 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +0100663 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100664 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +0100665 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100666
Shaohua Lic1e44752010-11-08 15:01:02 +0100667 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +1100668}
669
670/*
Jens Axboe5e705372006-07-13 12:39:25 +0200671 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200673 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 */
Jens Axboe5e705372006-07-13 12:39:25 +0200675static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100676cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100678 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200680#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
681#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
682 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
Jens Axboe5e705372006-07-13 12:39:25 +0200684 if (rq1 == NULL || rq1 == rq2)
685 return rq2;
686 if (rq2 == NULL)
687 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200688
Namhyung Kim229836b2011-05-24 10:23:21 +0200689 if (rq_is_sync(rq1) != rq_is_sync(rq2))
690 return rq_is_sync(rq1) ? rq1 : rq2;
691
Christoph Hellwig65299a32011-08-23 14:50:29 +0200692 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
693 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
Jens Axboeb53d1ed2011-08-19 08:34:48 +0200694
Tejun Heo83096eb2009-05-07 22:24:39 +0900695 s1 = blk_rq_pos(rq1);
696 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 /*
699 * by definition, 1KiB is 2 sectors
700 */
701 back_max = cfqd->cfq_back_max * 2;
702
703 /*
704 * Strict one way elevator _except_ in the case where we allow
705 * short backward seeks which are biased as twice the cost of a
706 * similar forward seek.
707 */
708 if (s1 >= last)
709 d1 = s1 - last;
710 else if (s1 + back_max >= last)
711 d1 = (last - s1) * cfqd->cfq_back_penalty;
712 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200713 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714
715 if (s2 >= last)
716 d2 = s2 - last;
717 else if (s2 + back_max >= last)
718 d2 = (last - s2) * cfqd->cfq_back_penalty;
719 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200720 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721
722 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723
Andreas Mohre8a99052006-03-28 08:59:49 +0200724 /*
725 * By doing switch() on the bit mask "wrap" we avoid having to
726 * check two variables for all permutations: --> faster!
727 */
728 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200729 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200730 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200731 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200732 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200733 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200734 else {
735 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200736 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200737 else
Jens Axboe5e705372006-07-13 12:39:25 +0200738 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200739 }
740
741 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200742 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200743 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200744 return rq2;
745 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200746 default:
747 /*
748 * Since both rqs are wrapped,
749 * start with the one that's further behind head
750 * (--> only *one* back seek required),
751 * since back seek takes more time than forward.
752 */
753 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200754 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 else
Jens Axboe5e705372006-07-13 12:39:25 +0200756 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 }
758}
759
Jens Axboe498d3aa2007-04-26 12:54:48 +0200760/*
761 * The below is leftmost cache rbtree addon
762 */
Jens Axboe08717142008-01-28 11:38:15 +0100763static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200764{
Vivek Goyal615f0252009-12-03 12:59:39 -0500765 /* Service tree is empty */
766 if (!root->count)
767 return NULL;
768
Jens Axboecc09e292007-04-26 12:53:50 +0200769 if (!root->left)
770 root->left = rb_first(&root->rb);
771
Jens Axboe08717142008-01-28 11:38:15 +0100772 if (root->left)
773 return rb_entry(root->left, struct cfq_queue, rb_node);
774
775 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200776}
777
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500778static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
779{
780 if (!root->left)
781 root->left = rb_first(&root->rb);
782
783 if (root->left)
784 return rb_entry_cfqg(root->left);
785
786 return NULL;
787}
788
Jens Axboea36e71f2009-04-15 12:15:11 +0200789static void rb_erase_init(struct rb_node *n, struct rb_root *root)
790{
791 rb_erase(n, root);
792 RB_CLEAR_NODE(n);
793}
794
Jens Axboecc09e292007-04-26 12:53:50 +0200795static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
796{
797 if (root->left == n)
798 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200799 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100800 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200801}
802
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803/*
804 * would be nice to take fifo expire time into account as well
805 */
Jens Axboe5e705372006-07-13 12:39:25 +0200806static struct request *
807cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
808 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809{
Jens Axboe21183b02006-07-13 12:33:14 +0200810 struct rb_node *rbnext = rb_next(&last->rb_node);
811 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200812 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813
Jens Axboe21183b02006-07-13 12:33:14 +0200814 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815
816 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200817 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200818
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200820 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200821 else {
822 rbnext = rb_first(&cfqq->sort_list);
823 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200824 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200825 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100827 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828}
829
Jens Axboed9e76202007-04-20 14:27:50 +0200830static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
831 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832{
Jens Axboed9e76202007-04-20 14:27:50 +0200833 /*
834 * just an approximation, should be ok.
835 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500836 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100837 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200838}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500840static inline s64
841cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
842{
843 return cfqg->vdisktime - st->min_vdisktime;
844}
845
846static void
847__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
848{
849 struct rb_node **node = &st->rb.rb_node;
850 struct rb_node *parent = NULL;
851 struct cfq_group *__cfqg;
852 s64 key = cfqg_key(st, cfqg);
853 int left = 1;
854
855 while (*node != NULL) {
856 parent = *node;
857 __cfqg = rb_entry_cfqg(parent);
858
859 if (key < cfqg_key(st, __cfqg))
860 node = &parent->rb_left;
861 else {
862 node = &parent->rb_right;
863 left = 0;
864 }
865 }
866
867 if (left)
868 st->left = &cfqg->rb_node;
869
870 rb_link_node(&cfqg->rb_node, parent, node);
871 rb_insert_color(&cfqg->rb_node, &st->rb);
872}
873
874static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100875cfq_update_group_weight(struct cfq_group *cfqg)
876{
877 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
878 if (cfqg->needs_update) {
879 cfqg->weight = cfqg->new_weight;
880 cfqg->needs_update = false;
881 }
882}
883
884static void
885cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
886{
887 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
888
889 cfq_update_group_weight(cfqg);
890 __cfq_group_service_tree_add(st, cfqg);
891 st->total_weight += cfqg->weight;
892}
893
894static void
895cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500896{
897 struct cfq_rb_root *st = &cfqd->grp_service_tree;
898 struct cfq_group *__cfqg;
899 struct rb_node *n;
900
901 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +0100902 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500903 return;
904
905 /*
906 * Currently put the group at the end. Later implement something
907 * so that groups get lesser vtime based on their weights, so that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300908 * if group does not loose all if it was not continuously backlogged.
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500909 */
910 n = rb_last(&st->rb);
911 if (n) {
912 __cfqg = rb_entry_cfqg(n);
913 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
914 } else
915 cfqg->vdisktime = st->min_vdisktime;
Justin TerAvest8184f932011-03-17 16:12:36 +0100916 cfq_group_service_tree_add(st, cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500917}
918
919static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100920cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
921{
922 st->total_weight -= cfqg->weight;
923 if (!RB_EMPTY_NODE(&cfqg->rb_node))
924 cfq_rb_erase(&cfqg->rb_node, st);
925}
926
927static void
928cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500929{
930 struct cfq_rb_root *st = &cfqd->grp_service_tree;
931
932 BUG_ON(cfqg->nr_cfqq < 1);
933 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500934
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500935 /* If there are other cfq queues under this group, don't delete it */
936 if (cfqg->nr_cfqq)
937 return;
938
Vivek Goyal2868ef72009-12-03 12:59:48 -0500939 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Justin TerAvest8184f932011-03-17 16:12:36 +0100940 cfq_group_service_tree_del(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500941 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400942 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500943}
944
Justin TerAvest167400d2011-03-12 16:54:00 +0100945static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
946 unsigned int *unaccounted_time)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500947{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500948 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500949
950 /*
951 * Queue got expired before even a single request completed or
952 * got expired immediately after first request completion.
953 */
954 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
955 /*
956 * Also charge the seek time incurred to the group, otherwise
957 * if there are mutiple queues in the group, each can dispatch
958 * a single request on seeky media and cause lots of seek time
959 * and group will never know it.
960 */
961 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
962 1);
963 } else {
964 slice_used = jiffies - cfqq->slice_start;
Justin TerAvest167400d2011-03-12 16:54:00 +0100965 if (slice_used > cfqq->allocated_slice) {
966 *unaccounted_time = slice_used - cfqq->allocated_slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500967 slice_used = cfqq->allocated_slice;
Justin TerAvest167400d2011-03-12 16:54:00 +0100968 }
969 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
970 *unaccounted_time += cfqq->slice_start -
971 cfqq->dispatch_start;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500972 }
973
Vivek Goyaldae739e2009-12-03 12:59:45 -0500974 return slice_used;
975}
976
977static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200978 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500979{
980 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Justin TerAvest167400d2011-03-12 16:54:00 +0100981 unsigned int used_sl, charge, unaccounted_sl = 0;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500982 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
983 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500984
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500985 BUG_ON(nr_sync < 0);
Justin TerAvest167400d2011-03-12 16:54:00 +0100986 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500987
Vivek Goyal02b35082010-08-23 12:23:53 +0200988 if (iops_mode(cfqd))
989 charge = cfqq->slice_dispatch;
990 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
991 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500992
993 /* Can't update vdisktime while group is on service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +0100994 cfq_group_service_tree_del(st, cfqg);
Vivek Goyal02b35082010-08-23 12:23:53 +0200995 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Justin TerAvest8184f932011-03-17 16:12:36 +0100996 /* If a new weight was requested, update now, off tree */
997 cfq_group_service_tree_add(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500998
999 /* This group is being expired. Save the context */
1000 if (time_after(cfqd->workload_expires, jiffies)) {
1001 cfqg->saved_workload_slice = cfqd->workload_expires
1002 - jiffies;
1003 cfqg->saved_workload = cfqd->serving_type;
1004 cfqg->saved_serving_prio = cfqd->serving_prio;
1005 } else
1006 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -05001007
1008 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1009 st->min_vdisktime);
Joe Perchesfd16d262011-06-13 10:42:49 +02001010 cfq_log_cfqq(cfqq->cfqd, cfqq,
1011 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1012 used_sl, cfqq->slice_dispatch, charge,
1013 iops_mode(cfqd), cfqq->nr_sectors);
Justin TerAvest167400d2011-03-12 16:54:00 +01001014 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1015 unaccounted_sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001016 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001017}
1018
Vivek Goyal25fb5162009-12-03 12:59:46 -05001019#ifdef CONFIG_CFQ_GROUP_IOSCHED
1020static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1021{
1022 if (blkg)
1023 return container_of(blkg, struct cfq_group, blkg);
1024 return NULL;
1025}
1026
Paul Bolle8aea4542011-06-06 05:07:54 +02001027static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1028 unsigned int weight)
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001029{
Justin TerAvest8184f932011-03-17 16:12:36 +01001030 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1031 cfqg->new_weight = weight;
1032 cfqg->needs_update = true;
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001033}
1034
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001035static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1036 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001037{
Vivek Goyal22084192009-12-03 12:59:49 -05001038 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1039 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001040
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001041 /*
1042 * Add group onto cgroup list. It might happen that bdi->dev is
1043 * not initialized yet. Initialize this new group without major
1044 * and minor info and this info will be filled in once a new thread
1045 * comes for IO.
1046 */
1047 if (bdi->dev) {
Ricky Beniteza74b2ad2010-04-05 18:22:17 +02001048 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001049 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1050 (void *)cfqd, MKDEV(major, minor));
1051 } else
1052 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1053 (void *)cfqd, 0);
1054
1055 cfqd->nr_blkcg_linked_grps++;
1056 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1057
1058 /* Add group on cfqd list */
1059 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1060}
1061
1062/*
1063 * Should be called from sleepable context. No request queue lock as per
1064 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1065 * from sleepable context.
1066 */
1067static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1068{
1069 struct cfq_group *cfqg = NULL;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001070 int i, j, ret;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001071 struct cfq_rb_root *st;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001072
1073 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1074 if (!cfqg)
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001075 return NULL;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001076
Vivek Goyal25fb5162009-12-03 12:59:46 -05001077 for_each_cfqg_st(cfqg, i, j, st)
1078 *st = CFQ_RB_ROOT;
1079 RB_CLEAR_NODE(&cfqg->rb_node);
1080
Shaohua Li7700fc42011-07-12 14:24:56 +02001081 cfqg->ttime.last_end_request = jiffies;
1082
Vivek Goyalb1c35762009-12-03 12:59:47 -05001083 /*
1084 * Take the initial reference that will be released on destroy
1085 * This can be thought of a joint reference by cgroup and
1086 * elevator which will be dropped by either elevator exit
1087 * or cgroup deletion path depending on who is exiting first.
1088 */
Shaohua Li329a6782011-01-07 08:48:28 +01001089 cfqg->ref = 1;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001090
1091 ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1092 if (ret) {
1093 kfree(cfqg);
1094 return NULL;
1095 }
1096
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001097 return cfqg;
1098}
1099
1100static struct cfq_group *
1101cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1102{
1103 struct cfq_group *cfqg = NULL;
1104 void *key = cfqd;
1105 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1106 unsigned int major, minor;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001107
Vivek Goyal180be2a2010-09-14 08:47:11 +02001108 /*
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001109 * This is the common case when there are no blkio cgroups.
1110 * Avoid lookup in this case
Vivek Goyal180be2a2010-09-14 08:47:11 +02001111 */
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001112 if (blkcg == &blkio_root_cgroup)
1113 cfqg = &cfqd->root_group;
1114 else
1115 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1116
1117 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
Vivek Goyal180be2a2010-09-14 08:47:11 +02001118 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001119 cfqg->blkg.dev = MKDEV(major, minor);
1120 }
Vivek Goyal180be2a2010-09-14 08:47:11 +02001121
Vivek Goyal25fb5162009-12-03 12:59:46 -05001122 return cfqg;
1123}
1124
1125/*
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001126 * Search for the cfq group current task belongs to. request_queue lock must
1127 * be held.
Vivek Goyal25fb5162009-12-03 12:59:46 -05001128 */
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001129static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001130{
Vivek Goyal70087dc2011-05-16 15:24:08 +02001131 struct blkio_cgroup *blkcg;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001132 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1133 struct request_queue *q = cfqd->queue;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001134
1135 rcu_read_lock();
Vivek Goyal70087dc2011-05-16 15:24:08 +02001136 blkcg = task_blkio_cgroup(current);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001137 cfqg = cfq_find_cfqg(cfqd, blkcg);
1138 if (cfqg) {
1139 rcu_read_unlock();
1140 return cfqg;
1141 }
1142
1143 /*
1144 * Need to allocate a group. Allocation of group also needs allocation
1145 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1146 * we need to drop rcu lock and queue_lock before we call alloc.
1147 *
1148 * Not taking any queue reference here and assuming that queue is
1149 * around by the time we return. CFQ queue allocation code does
1150 * the same. It might be racy though.
1151 */
1152
1153 rcu_read_unlock();
1154 spin_unlock_irq(q->queue_lock);
1155
1156 cfqg = cfq_alloc_cfqg(cfqd);
1157
1158 spin_lock_irq(q->queue_lock);
1159
1160 rcu_read_lock();
1161 blkcg = task_blkio_cgroup(current);
1162
1163 /*
1164 * If some other thread already allocated the group while we were
1165 * not holding queue lock, free up the group
1166 */
1167 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1168
1169 if (__cfqg) {
1170 kfree(cfqg);
1171 rcu_read_unlock();
1172 return __cfqg;
1173 }
1174
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001175 if (!cfqg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001176 cfqg = &cfqd->root_group;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001177
1178 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001179 rcu_read_unlock();
1180 return cfqg;
1181}
1182
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001183static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1184{
Shaohua Li329a6782011-01-07 08:48:28 +01001185 cfqg->ref++;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001186 return cfqg;
1187}
1188
Vivek Goyal25fb5162009-12-03 12:59:46 -05001189static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1190{
1191 /* Currently, all async queues are mapped to root group */
1192 if (!cfq_cfqq_sync(cfqq))
1193 cfqg = &cfqq->cfqd->root_group;
1194
1195 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001196 /* cfqq reference on cfqg */
Shaohua Li329a6782011-01-07 08:48:28 +01001197 cfqq->cfqg->ref++;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001198}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001199
1200static void cfq_put_cfqg(struct cfq_group *cfqg)
1201{
1202 struct cfq_rb_root *st;
1203 int i, j;
1204
Shaohua Li329a6782011-01-07 08:48:28 +01001205 BUG_ON(cfqg->ref <= 0);
1206 cfqg->ref--;
1207 if (cfqg->ref)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001208 return;
1209 for_each_cfqg_st(cfqg, i, j, st)
Gui Jianfengb54ce602010-11-30 20:52:46 +01001210 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001211 free_percpu(cfqg->blkg.stats_cpu);
Vivek Goyalb1c35762009-12-03 12:59:47 -05001212 kfree(cfqg);
1213}
1214
1215static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1216{
1217 /* Something wrong if we are trying to remove same group twice */
1218 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1219
1220 hlist_del_init(&cfqg->cfqd_node);
1221
Vivek Goyala5395b82011-08-02 09:24:09 +02001222 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1223 cfqd->nr_blkcg_linked_grps--;
1224
Vivek Goyalb1c35762009-12-03 12:59:47 -05001225 /*
1226 * Put the reference taken at the time of creation so that when all
1227 * queues are gone, group can be destroyed.
1228 */
1229 cfq_put_cfqg(cfqg);
1230}
1231
1232static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1233{
1234 struct hlist_node *pos, *n;
1235 struct cfq_group *cfqg;
1236
1237 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1238 /*
1239 * If cgroup removal path got to blk_group first and removed
1240 * it from cgroup list, then it will take care of destroying
1241 * cfqg also.
1242 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001243 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001244 cfq_destroy_cfqg(cfqd, cfqg);
1245 }
1246}
1247
1248/*
1249 * Blk cgroup controller notification saying that blkio_group object is being
1250 * delinked as associated cgroup object is going away. That also means that
1251 * no new IO will come in this group. So get rid of this group as soon as
1252 * any pending IO in the group is finished.
1253 *
1254 * This function is called under rcu_read_lock(). key is the rcu protected
1255 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1256 * read lock.
1257 *
1258 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1259 * it should not be NULL as even if elevator was exiting, cgroup deltion
1260 * path got to it first.
1261 */
Paul Bolle8aea4542011-06-06 05:07:54 +02001262static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001263{
1264 unsigned long flags;
1265 struct cfq_data *cfqd = key;
1266
1267 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1268 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1269 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1270}
1271
Vivek Goyal25fb5162009-12-03 12:59:46 -05001272#else /* GROUP_IOSCHED */
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001273static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001274{
1275 return &cfqd->root_group;
1276}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001277
1278static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1279{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001280 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001281}
1282
Vivek Goyal25fb5162009-12-03 12:59:46 -05001283static inline void
1284cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1285 cfqq->cfqg = cfqg;
1286}
1287
Vivek Goyalb1c35762009-12-03 12:59:47 -05001288static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1289static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1290
Vivek Goyal25fb5162009-12-03 12:59:46 -05001291#endif /* GROUP_IOSCHED */
1292
Jens Axboe498d3aa2007-04-26 12:54:48 +02001293/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001294 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa2007-04-26 12:54:48 +02001295 * requests waiting to be processed. It is sorted in the order that
1296 * we will service the queues.
1297 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001298static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001299 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001300{
Jens Axboe08717142008-01-28 11:38:15 +01001301 struct rb_node **p, *parent;
1302 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001303 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001304 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa2007-04-26 12:54:48 +02001305 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001306 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001307
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001308 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001309 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001310 if (cfq_class_idle(cfqq)) {
1311 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001312 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001313 if (parent && parent != &cfqq->rb_node) {
1314 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1315 rb_key += __cfqq->rb_key;
1316 } else
1317 rb_key += jiffies;
1318 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001319 /*
1320 * Get our rb key offset. Subtract any residual slice
1321 * value carried from last service. A negative resid
1322 * count indicates slice overrun, and this should position
1323 * the next service time further away in the tree.
1324 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001325 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001326 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001327 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001328 } else {
1329 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001330 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001331 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1332 }
Jens Axboed9e76202007-04-20 14:27:50 +02001333
1334 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001335 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001336 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001337 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001338 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001339 if (rb_key == cfqq->rb_key &&
1340 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001341 return;
Jens Axboe53b03742006-07-28 09:48:51 +02001342
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001343 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1344 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001345 }
Jens Axboed9e76202007-04-20 14:27:50 +02001346
Jens Axboe498d3aa2007-04-26 12:54:48 +02001347 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001348 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001349 cfqq->service_tree = service_tree;
1350 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001351 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001352 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001353
Jens Axboed9e76202007-04-20 14:27:50 +02001354 parent = *p;
1355 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1356
Jens Axboe0c534e02007-04-18 20:01:57 +02001357 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001358 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001359 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001360 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001361 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001362 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001363 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001364 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001365 }
Jens Axboe67060e32007-04-18 20:13:32 +02001366
1367 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001368 }
1369
Jens Axboecc09e292007-04-26 12:53:50 +02001370 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001371 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001372
Jens Axboed9e76202007-04-20 14:27:50 +02001373 cfqq->rb_key = rb_key;
1374 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001375 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1376 service_tree->count++;
Namhyung Kim20359f22011-05-24 10:23:22 +02001377 if (add_front || !new_cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001378 return;
Justin TerAvest8184f932011-03-17 16:12:36 +01001379 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380}
1381
Jens Axboea36e71f2009-04-15 12:15:11 +02001382static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001383cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1384 sector_t sector, struct rb_node **ret_parent,
1385 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001386{
Jens Axboea36e71f2009-04-15 12:15:11 +02001387 struct rb_node **p, *parent;
1388 struct cfq_queue *cfqq = NULL;
1389
1390 parent = NULL;
1391 p = &root->rb_node;
1392 while (*p) {
1393 struct rb_node **n;
1394
1395 parent = *p;
1396 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1397
1398 /*
1399 * Sort strictly based on sector. Smallest to the left,
1400 * largest to the right.
1401 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001402 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001403 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001404 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001405 n = &(*p)->rb_left;
1406 else
1407 break;
1408 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001409 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001410 }
1411
1412 *ret_parent = parent;
1413 if (rb_link)
1414 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001415 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001416}
1417
1418static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1419{
Jens Axboea36e71f2009-04-15 12:15:11 +02001420 struct rb_node **p, *parent;
1421 struct cfq_queue *__cfqq;
1422
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001423 if (cfqq->p_root) {
1424 rb_erase(&cfqq->p_node, cfqq->p_root);
1425 cfqq->p_root = NULL;
1426 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001427
1428 if (cfq_class_idle(cfqq))
1429 return;
1430 if (!cfqq->next_rq)
1431 return;
1432
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001433 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001434 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1435 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001436 if (!__cfqq) {
1437 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001438 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1439 } else
1440 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001441}
1442
Jens Axboe498d3aa2007-04-26 12:54:48 +02001443/*
1444 * Update cfqq's position in the service tree.
1445 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001446static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001447{
Jens Axboe6d048f52007-04-25 12:44:27 +02001448 /*
1449 * Resorting requires the cfqq to be on the RR list already.
1450 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001451 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001452 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001453 cfq_prio_tree_add(cfqd, cfqq);
1454 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001455}
1456
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457/*
1458 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001459 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460 */
Jens Axboefebffd62008-01-28 13:19:43 +01001461static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462{
Jens Axboe7b679132008-05-30 12:23:07 +02001463 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001464 BUG_ON(cfq_cfqq_on_rr(cfqq));
1465 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 cfqd->busy_queues++;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001467 if (cfq_cfqq_sync(cfqq))
1468 cfqd->busy_sync_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469
Jens Axboeedd75ff2007-04-19 12:03:34 +02001470 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471}
1472
Jens Axboe498d3aa2007-04-26 12:54:48 +02001473/*
1474 * Called when the cfqq no longer has requests pending, remove it from
1475 * the service tree.
1476 */
Jens Axboefebffd62008-01-28 13:19:43 +01001477static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478{
Jens Axboe7b679132008-05-30 12:23:07 +02001479 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001480 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1481 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001483 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1484 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1485 cfqq->service_tree = NULL;
1486 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001487 if (cfqq->p_root) {
1488 rb_erase(&cfqq->p_node, cfqq->p_root);
1489 cfqq->p_root = NULL;
1490 }
Jens Axboed9e76202007-04-20 14:27:50 +02001491
Justin TerAvest8184f932011-03-17 16:12:36 +01001492 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 BUG_ON(!cfqd->busy_queues);
1494 cfqd->busy_queues--;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001495 if (cfq_cfqq_sync(cfqq))
1496 cfqd->busy_sync_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497}
1498
1499/*
1500 * rb tree support functions
1501 */
Jens Axboefebffd62008-01-28 13:19:43 +01001502static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503{
Jens Axboe5e705372006-07-13 12:39:25 +02001504 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001505 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506
Jens Axboeb4878f22005-10-20 16:42:29 +02001507 BUG_ON(!cfqq->queued[sync]);
1508 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509
Jens Axboe5e705372006-07-13 12:39:25 +02001510 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511
Vivek Goyalf04a6422009-12-03 12:59:40 -05001512 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1513 /*
1514 * Queue will be deleted from service tree when we actually
1515 * expire it later. Right now just remove it from prio tree
1516 * as it is empty.
1517 */
1518 if (cfqq->p_root) {
1519 rb_erase(&cfqq->p_node, cfqq->p_root);
1520 cfqq->p_root = NULL;
1521 }
1522 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523}
1524
Jens Axboe5e705372006-07-13 12:39:25 +02001525static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526{
Jens Axboe5e705372006-07-13 12:39:25 +02001527 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 struct cfq_data *cfqd = cfqq->cfqd;
Jeff Moyer796d5112011-06-02 21:19:05 +02001529 struct request *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530
Jens Axboe5380a102006-07-13 12:37:56 +02001531 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532
Jeff Moyer796d5112011-06-02 21:19:05 +02001533 elv_rb_add(&cfqq->sort_list, rq);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001534
1535 if (!cfq_cfqq_on_rr(cfqq))
1536 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001537
1538 /*
1539 * check if this request is a better next-serve candidate
1540 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001541 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001542 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001543
1544 /*
1545 * adjust priority tree position, if ->next_rq changes
1546 */
1547 if (prev != cfqq->next_rq)
1548 cfq_prio_tree_add(cfqd, cfqq);
1549
Jens Axboe5044eed2007-04-25 11:53:48 +02001550 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551}
1552
Jens Axboefebffd62008-01-28 13:19:43 +01001553static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554{
Jens Axboe5380a102006-07-13 12:37:56 +02001555 elv_rb_del(&cfqq->sort_list, rq);
1556 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001557 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1558 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001559 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001560 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001561 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1562 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563}
1564
Jens Axboe206dc692006-03-28 13:03:44 +02001565static struct request *
1566cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567{
Jens Axboe206dc692006-03-28 13:03:44 +02001568 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01001569 struct cfq_io_cq *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001570 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571
Jens Axboe4ac845a2008-01-24 08:44:49 +01001572 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001573 if (!cic)
1574 return NULL;
1575
1576 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001577 if (cfqq) {
1578 sector_t sector = bio->bi_sector + bio_sectors(bio);
1579
Jens Axboe21183b02006-07-13 12:33:14 +02001580 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001581 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 return NULL;
1584}
1585
Jens Axboe165125e2007-07-24 09:28:11 +02001586static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001587{
1588 struct cfq_data *cfqd = q->elevator->elevator_data;
1589
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001590 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001591 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001592 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001593
Tejun Heo5b936292009-05-07 22:24:38 +09001594 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001595}
1596
Jens Axboe165125e2007-07-24 09:28:11 +02001597static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598{
Jens Axboe22e2c502005-06-27 10:55:12 +02001599 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001601 WARN_ON(!cfqd->rq_in_driver);
1602 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001603 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001604 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605}
1606
Jens Axboeb4878f22005-10-20 16:42:29 +02001607static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608{
Jens Axboe5e705372006-07-13 12:39:25 +02001609 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001610
Jens Axboe5e705372006-07-13 12:39:25 +02001611 if (cfqq->next_rq == rq)
1612 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613
Jens Axboeb4878f22005-10-20 16:42:29 +02001614 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001615 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001616
Aaron Carroll45333d52008-08-26 15:52:36 +02001617 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001618 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1619 rq_data_dir(rq), rq_is_sync(rq));
Christoph Hellwig65299a32011-08-23 14:50:29 +02001620 if (rq->cmd_flags & REQ_PRIO) {
1621 WARN_ON(!cfqq->prio_pending);
1622 cfqq->prio_pending--;
Jens Axboeb53d1ed2011-08-19 08:34:48 +02001623 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624}
1625
Jens Axboe165125e2007-07-24 09:28:11 +02001626static int cfq_merge(struct request_queue *q, struct request **req,
1627 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628{
1629 struct cfq_data *cfqd = q->elevator->elevator_data;
1630 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631
Jens Axboe206dc692006-03-28 13:03:44 +02001632 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001633 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001634 *req = __rq;
1635 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 }
1637
1638 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639}
1640
Jens Axboe165125e2007-07-24 09:28:11 +02001641static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001642 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643{
Jens Axboe21183b02006-07-13 12:33:14 +02001644 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001645 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646
Jens Axboe5e705372006-07-13 12:39:25 +02001647 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649}
1650
Divyesh Shah812d4022010-04-08 21:14:23 -07001651static void cfq_bio_merged(struct request_queue *q, struct request *req,
1652 struct bio *bio)
1653{
Vivek Goyale98ef892010-06-18 10:39:47 -04001654 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1655 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001656}
1657
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658static void
Jens Axboe165125e2007-07-24 09:28:11 +02001659cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 struct request *next)
1661{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001662 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Shaohua Li4a0b75c2011-12-16 14:00:22 +01001663 struct cfq_data *cfqd = q->elevator->elevator_data;
1664
Jens Axboe22e2c502005-06-27 10:55:12 +02001665 /*
1666 * reposition in fifo if next is older than rq
1667 */
1668 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001669 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001670 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001671 rq_set_fifo_time(rq, rq_fifo_time(next));
1672 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001673
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001674 if (cfqq->next_rq == next)
1675 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001676 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001677 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1678 rq_data_dir(next), rq_is_sync(next));
Shaohua Li4a0b75c2011-12-16 14:00:22 +01001679
1680 cfqq = RQ_CFQQ(next);
1681 /*
1682 * all requests of this queue are merged to other queues, delete it
1683 * from the service tree. If it's the active_queue,
1684 * cfq_dispatch_requests() will choose to expire it or do idle
1685 */
1686 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1687 cfqq != cfqd->active_queue)
1688 cfq_del_cfqq_rr(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001689}
1690
Jens Axboe165125e2007-07-24 09:28:11 +02001691static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001692 struct bio *bio)
1693{
1694 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heoc5869802011-12-14 00:33:41 +01001695 struct cfq_io_cq *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001696 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001697
1698 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001699 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001700 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001701 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001702 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001703
1704 /*
Tejun Heof1a4f4d2011-12-14 00:33:39 +01001705 * Lookup the cfqq that this bio will be queued with and allow
Tejun Heo07c2bd32012-02-08 09:19:42 +01001706 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001707 */
Tejun Heo07c2bd32012-02-08 09:19:42 +01001708 cic = cfq_cic_lookup(cfqd, current->io_context);
1709 if (!cic)
1710 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001711
Vasily Tarasov91fac312007-04-25 12:29:51 +02001712 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001713 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001714}
1715
Divyesh Shah812df482010-04-08 21:15:35 -07001716static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1717{
1718 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001719 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001720}
1721
Jens Axboefebffd62008-01-28 13:19:43 +01001722static void __cfq_set_active_queue(struct cfq_data *cfqd,
1723 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001724{
1725 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001726 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1727 cfqd->serving_prio, cfqd->serving_type);
Justin TerAvest62a37f62011-03-23 08:25:44 +01001728 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1729 cfqq->slice_start = 0;
1730 cfqq->dispatch_start = jiffies;
1731 cfqq->allocated_slice = 0;
1732 cfqq->slice_end = 0;
1733 cfqq->slice_dispatch = 0;
1734 cfqq->nr_sectors = 0;
1735
1736 cfq_clear_cfqq_wait_request(cfqq);
1737 cfq_clear_cfqq_must_dispatch(cfqq);
1738 cfq_clear_cfqq_must_alloc_slice(cfqq);
1739 cfq_clear_cfqq_fifo_expire(cfqq);
1740 cfq_mark_cfqq_slice_new(cfqq);
1741
1742 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001743 }
1744
1745 cfqd->active_queue = cfqq;
1746}
1747
1748/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001749 * current cfqq expired its slice (or was too idle), select new one
1750 */
1751static void
1752__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001753 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001754{
Jens Axboe7b679132008-05-30 12:23:07 +02001755 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1756
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001757 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001758 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001759
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001760 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001761 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001762
1763 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001764 * If this cfqq is shared between multiple processes, check to
1765 * make sure that those processes are still issuing I/Os within
1766 * the mean seek distance. If not, it may be time to break the
1767 * queues apart again.
1768 */
1769 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1770 cfq_mark_cfqq_split_coop(cfqq);
1771
1772 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001773 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001774 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01001775 if (timed_out) {
1776 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07001777 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001778 else
1779 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001780 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1781 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001782
Vivek Goyale5ff0822010-04-26 19:25:11 +02001783 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001784
Vivek Goyalf04a6422009-12-03 12:59:40 -05001785 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1786 cfq_del_cfqq_rr(cfqd, cfqq);
1787
Jens Axboeedd75ff2007-04-19 12:03:34 +02001788 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001789
1790 if (cfqq == cfqd->active_queue)
1791 cfqd->active_queue = NULL;
1792
1793 if (cfqd->active_cic) {
Tejun Heo11a31222012-02-07 07:51:30 +01001794 put_io_context(cfqd->active_cic->icq.ioc);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001795 cfqd->active_cic = NULL;
1796 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001797}
1798
Vivek Goyale5ff0822010-04-26 19:25:11 +02001799static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001800{
1801 struct cfq_queue *cfqq = cfqd->active_queue;
1802
1803 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001804 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001805}
1806
Jens Axboe498d3aa2007-04-26 12:54:48 +02001807/*
1808 * Get next queue for service. Unless we have a queue preemption,
1809 * we'll simply select the first cfqq in the service tree.
1810 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001811static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001812{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001813 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001814 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001815 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001816
Vivek Goyalf04a6422009-12-03 12:59:40 -05001817 if (!cfqd->rq_queued)
1818 return NULL;
1819
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001820 /* There is nothing to dispatch */
1821 if (!service_tree)
1822 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001823 if (RB_EMPTY_ROOT(&service_tree->rb))
1824 return NULL;
1825 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001826}
1827
Vivek Goyalf04a6422009-12-03 12:59:40 -05001828static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1829{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001830 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001831 struct cfq_queue *cfqq;
1832 int i, j;
1833 struct cfq_rb_root *st;
1834
1835 if (!cfqd->rq_queued)
1836 return NULL;
1837
Vivek Goyal25fb5162009-12-03 12:59:46 -05001838 cfqg = cfq_get_next_cfqg(cfqd);
1839 if (!cfqg)
1840 return NULL;
1841
Vivek Goyalf04a6422009-12-03 12:59:40 -05001842 for_each_cfqg_st(cfqg, i, j, st)
1843 if ((cfqq = cfq_rb_first(st)) != NULL)
1844 return cfqq;
1845 return NULL;
1846}
1847
Jens Axboe498d3aa2007-04-26 12:54:48 +02001848/*
1849 * Get and set a new active queue for service.
1850 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001851static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1852 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001853{
Jens Axboee00ef792009-11-04 08:54:55 +01001854 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001855 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001856
Jens Axboe22e2c502005-06-27 10:55:12 +02001857 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001858 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001859}
1860
Jens Axboed9e76202007-04-20 14:27:50 +02001861static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1862 struct request *rq)
1863{
Tejun Heo83096eb2009-05-07 22:24:39 +09001864 if (blk_rq_pos(rq) >= cfqd->last_position)
1865 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001866 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001867 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001868}
1869
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001870static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001871 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001872{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001873 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001874}
1875
Jens Axboea36e71f2009-04-15 12:15:11 +02001876static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1877 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001878{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001879 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001880 struct rb_node *parent, *node;
1881 struct cfq_queue *__cfqq;
1882 sector_t sector = cfqd->last_position;
1883
1884 if (RB_EMPTY_ROOT(root))
1885 return NULL;
1886
1887 /*
1888 * First, if we find a request starting at the end of the last
1889 * request, choose it.
1890 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001891 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001892 if (__cfqq)
1893 return __cfqq;
1894
1895 /*
1896 * If the exact sector wasn't found, the parent of the NULL leaf
1897 * will contain the closest sector.
1898 */
1899 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001900 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001901 return __cfqq;
1902
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001903 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001904 node = rb_next(&__cfqq->p_node);
1905 else
1906 node = rb_prev(&__cfqq->p_node);
1907 if (!node)
1908 return NULL;
1909
1910 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001911 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001912 return __cfqq;
1913
1914 return NULL;
1915}
1916
1917/*
1918 * cfqd - obvious
1919 * cur_cfqq - passed in so that we don't decide that the current queue is
1920 * closely cooperating with itself.
1921 *
1922 * So, basically we're assuming that that cur_cfqq has dispatched at least
1923 * one request, and that cfqd->last_position reflects a position on the disk
1924 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1925 * assumption.
1926 */
1927static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001928 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001929{
1930 struct cfq_queue *cfqq;
1931
Divyesh Shah39c01b22010-03-25 15:45:57 +01001932 if (cfq_class_idle(cur_cfqq))
1933 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001934 if (!cfq_cfqq_sync(cur_cfqq))
1935 return NULL;
1936 if (CFQQ_SEEKY(cur_cfqq))
1937 return NULL;
1938
Jens Axboea36e71f2009-04-15 12:15:11 +02001939 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001940 * Don't search priority tree if it's the only queue in the group.
1941 */
1942 if (cur_cfqq->cfqg->nr_cfqq == 1)
1943 return NULL;
1944
1945 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001946 * We should notice if some of the queues are cooperating, eg
1947 * working closely on the same area of the disk. In that case,
1948 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001949 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001950 cfqq = cfqq_close(cfqd, cur_cfqq);
1951 if (!cfqq)
1952 return NULL;
1953
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001954 /* If new queue belongs to different cfq_group, don't choose it */
1955 if (cur_cfqq->cfqg != cfqq->cfqg)
1956 return NULL;
1957
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001958 /*
1959 * It only makes sense to merge sync queues.
1960 */
1961 if (!cfq_cfqq_sync(cfqq))
1962 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001963 if (CFQQ_SEEKY(cfqq))
1964 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001965
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001966 /*
1967 * Do not merge queues of different priority classes
1968 */
1969 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1970 return NULL;
1971
Jens Axboea36e71f2009-04-15 12:15:11 +02001972 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001973}
1974
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001975/*
1976 * Determine whether we should enforce idle window for this queue.
1977 */
1978
1979static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1980{
1981 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001982 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001983
Vivek Goyalf04a6422009-12-03 12:59:40 -05001984 BUG_ON(!service_tree);
1985 BUG_ON(!service_tree->count);
1986
Vivek Goyalb6508c12010-08-23 12:23:33 +02001987 if (!cfqd->cfq_slice_idle)
1988 return false;
1989
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001990 /* We never do for idle class queues. */
1991 if (prio == IDLE_WORKLOAD)
1992 return false;
1993
1994 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001995 if (cfq_cfqq_idle_window(cfqq) &&
1996 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001997 return true;
1998
1999 /*
2000 * Otherwise, we do only if they are the last ones
2001 * in their service tree.
2002 */
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002003 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2004 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
Shaohua Lic1e44752010-11-08 15:01:02 +01002005 return true;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002006 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2007 service_tree->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01002008 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002009}
2010
Jens Axboe6d048f52007-04-25 12:44:27 +02002011static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002012{
Jens Axboe17926692007-01-19 11:59:30 +11002013 struct cfq_queue *cfqq = cfqd->active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +01002014 struct cfq_io_cq *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002015 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002016
Jens Axboea68bbdd2008-09-24 13:03:33 +02002017 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002018 * SSD device without seek penalty, disable idling. But only do so
2019 * for devices that support queuing, otherwise we still have a problem
2020 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02002021 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002022 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02002023 return;
2024
Jens Axboedd67d052006-06-21 09:36:18 +02002025 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02002026 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02002027
2028 /*
2029 * idle is disabled, either manually or by past process history
2030 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002031 if (!cfq_should_idle(cfqd, cfqq)) {
2032 /* no queue idling. Check for group idling */
2033 if (cfqd->cfq_group_idle)
2034 group_idle = cfqd->cfq_group_idle;
2035 else
2036 return;
2037 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002038
Jens Axboe22e2c502005-06-27 10:55:12 +02002039 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002040 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02002041 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002042 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02002043 return;
2044
2045 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002046 * task has exited, don't wait
2047 */
Jens Axboe206dc692006-03-28 13:03:44 +02002048 cic = cfqd->active_cic;
Tejun Heoc5869802011-12-14 00:33:41 +01002049 if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02002050 return;
2051
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002052 /*
2053 * If our average think time is larger than the remaining time
2054 * slice, then don't idle. This avoids overrunning the allotted
2055 * time slice.
2056 */
Shaohua Li383cd722011-07-12 14:24:35 +02002057 if (sample_valid(cic->ttime.ttime_samples) &&
2058 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
Joe Perchesfd16d262011-06-13 10:42:49 +02002059 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
Shaohua Li383cd722011-07-12 14:24:35 +02002060 cic->ttime.ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002061 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002062 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002063
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002064 /* There are other queues in the group, don't do group idle */
2065 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2066 return;
2067
Jens Axboe3b181522005-06-27 10:56:24 +02002068 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002069
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002070 if (group_idle)
2071 sl = cfqd->cfq_group_idle;
2072 else
2073 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02002074
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002075 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04002076 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002077 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2078 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079}
2080
Jens Axboe498d3aa2007-04-26 12:54:48 +02002081/*
2082 * Move request from internal lists to the request queue dispatch list.
2083 */
Jens Axboe165125e2007-07-24 09:28:11 +02002084static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085{
Jens Axboe3ed9a292007-04-23 08:33:33 +02002086 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002087 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002088
Jens Axboe7b679132008-05-30 12:23:07 +02002089 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2090
Jeff Moyer06d21882009-09-11 17:08:59 +02002091 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02002092 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002093 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002094 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02002095 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02002096
Tatyana Brokhmanb410a822013-07-07 17:33:05 +03002097 if (rq->cmd_flags & REQ_URGENT) {
2098 if (!cfqd->nr_urgent_pending)
2099 WARN_ON(1);
2100 else
2101 cfqd->nr_urgent_pending--;
2102 cfqd->nr_urgent_in_flight++;
2103 }
2104
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002105 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02002106 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04002107 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02002108 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109}
2110
2111/*
2112 * return expired entry, or NULL to just start from scratch in rbtree
2113 */
Jens Axboefebffd62008-01-28 13:19:43 +01002114static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115{
Jens Axboe30996f42009-10-05 11:03:39 +02002116 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
Jens Axboe3b181522005-06-27 10:56:24 +02002118 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002120
2121 cfq_mark_cfqq_fifo_expire(cfqq);
2122
Jens Axboe89850f72006-07-22 16:48:31 +02002123 if (list_empty(&cfqq->fifo))
2124 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125
Jens Axboe89850f72006-07-22 16:48:31 +02002126 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002127 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002128 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
Jens Axboe30996f42009-10-05 11:03:39 +02002130 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002131 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132}
2133
Jens Axboe22e2c502005-06-27 10:55:12 +02002134static inline int
2135cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2136{
2137 const int base_rq = cfqd->cfq_slice_async_rq;
2138
2139 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2140
Namhyung Kimb9f8ce02011-05-24 10:23:21 +02002141 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002142}
2143
2144/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002145 * Must be called with the queue_lock held.
2146 */
2147static int cfqq_process_refs(struct cfq_queue *cfqq)
2148{
2149 int process_refs, io_refs;
2150
2151 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002152 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002153 BUG_ON(process_refs < 0);
2154 return process_refs;
2155}
2156
2157static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2158{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002159 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002160 struct cfq_queue *__cfqq;
2161
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002162 /*
2163 * If there are no process references on the new_cfqq, then it is
2164 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2165 * chain may have dropped their last reference (not just their
2166 * last process reference).
2167 */
2168 if (!cfqq_process_refs(new_cfqq))
2169 return;
2170
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002171 /* Avoid a circular list and skip interim queue merges */
2172 while ((__cfqq = new_cfqq->new_cfqq)) {
2173 if (__cfqq == cfqq)
2174 return;
2175 new_cfqq = __cfqq;
2176 }
2177
2178 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002179 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002180 /*
2181 * If the process for the cfqq has gone away, there is no
2182 * sense in merging the queues.
2183 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002184 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002185 return;
2186
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002187 /*
2188 * Merge in the direction of the lesser amount of work.
2189 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002190 if (new_process_refs >= process_refs) {
2191 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002192 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002193 } else {
2194 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002195 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002196 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002197}
2198
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002199static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002200 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002201{
2202 struct cfq_queue *queue;
2203 int i;
2204 bool key_valid = false;
2205 unsigned long lowest_key = 0;
2206 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2207
Vivek Goyal65b32a52009-12-16 17:52:59 -05002208 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2209 /* select the one with lowest rb_key */
2210 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002211 if (queue &&
2212 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2213 lowest_key = queue->rb_key;
2214 cur_best = i;
2215 key_valid = true;
2216 }
2217 }
2218
2219 return cur_best;
2220}
2221
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002222static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002223{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002224 unsigned slice;
2225 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002226 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002227 unsigned group_slice;
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002228 enum wl_prio_t original_prio = cfqd->serving_prio;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002229
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002230 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002231 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002232 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002233 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002234 cfqd->serving_prio = BE_WORKLOAD;
2235 else {
2236 cfqd->serving_prio = IDLE_WORKLOAD;
2237 cfqd->workload_expires = jiffies + 1;
2238 return;
2239 }
2240
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002241 if (original_prio != cfqd->serving_prio)
2242 goto new_workload;
2243
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002244 /*
2245 * For RT and BE, we have to choose also the type
2246 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2247 * expiration time
2248 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002249 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002250 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002251
2252 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002253 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002254 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002255 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002256 return;
2257
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002258new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002259 /* otherwise select new workload type */
2260 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002261 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2262 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002263 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002264
2265 /*
2266 * the workload slice is computed as a fraction of target latency
2267 * proportional to the number of queues in that workload, over
2268 * all the queues in the same priority class
2269 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002270 group_slice = cfq_group_slice(cfqd, cfqg);
2271
2272 slice = group_slice * count /
2273 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2274 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002275
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002276 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2277 unsigned int tmp;
2278
2279 /*
2280 * Async queues are currently system wide. Just taking
2281 * proportion of queues with-in same group will lead to higher
2282 * async ratio system wide as generally root group is going
2283 * to have higher weight. A more accurate thing would be to
2284 * calculate system wide asnc/sync ratio.
2285 */
Tao Ma5bf14c02012-04-01 14:33:39 -07002286 tmp = cfqd->cfq_target_latency *
2287 cfqg_busy_async_queues(cfqd, cfqg);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002288 tmp = tmp/cfqd->busy_queues;
2289 slice = min_t(unsigned, slice, tmp);
2290
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002291 /* async workload slice is scaled down according to
2292 * the sync/async slice ratio. */
2293 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002294 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002295 /* sync workload slice is at least 2 * cfq_slice_idle */
2296 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2297
2298 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002299 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002300 cfqd->workload_expires = jiffies + slice;
2301}
2302
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002303static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2304{
2305 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002306 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002307
2308 if (RB_EMPTY_ROOT(&st->rb))
2309 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002310 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002311 update_min_vdisktime(st);
2312 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002313}
2314
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002315static void cfq_choose_cfqg(struct cfq_data *cfqd)
2316{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002317 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2318
Sujit Reddy Thumma5746a1c2012-12-14 19:36:58 +05302319 if (!cfqg)
2320 return;
2321
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002322 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002323
2324 /* Restore the workload type data */
2325 if (cfqg->saved_workload_slice) {
2326 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2327 cfqd->serving_type = cfqg->saved_workload;
2328 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002329 } else
2330 cfqd->workload_expires = jiffies - 1;
2331
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002332 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002333}
2334
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002335/*
Jens Axboe498d3aa2007-04-26 12:54:48 +02002336 * Select a queue for service. If we have a current active queue,
2337 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002338 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002339static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002340{
Jens Axboea36e71f2009-04-15 12:15:11 +02002341 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002342
2343 cfqq = cfqd->active_queue;
2344 if (!cfqq)
2345 goto new_queue;
2346
Vivek Goyalf04a6422009-12-03 12:59:40 -05002347 if (!cfqd->rq_queued)
2348 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002349
2350 /*
2351 * We were waiting for group to get backlogged. Expire the queue
2352 */
2353 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2354 goto expire;
2355
Jens Axboe22e2c502005-06-27 10:55:12 +02002356 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002357 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002358 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002359 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2360 /*
2361 * If slice had not expired at the completion of last request
2362 * we might not have turned on wait_busy flag. Don't expire
2363 * the queue yet. Allow the group to get backlogged.
2364 *
2365 * The very fact that we have used the slice, that means we
2366 * have been idling all along on this queue and it should be
2367 * ok to wait for this request to complete.
2368 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002369 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2370 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2371 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002372 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002373 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002374 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002375 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002376
2377 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002378 * The active queue has requests and isn't expired, allow it to
2379 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002380 */
Jens Axboedd67d052006-06-21 09:36:18 +02002381 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002382 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002383
2384 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002385 * If another queue has a request waiting within our mean seek
2386 * distance, let it run. The expire code will check for close
2387 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002388 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002389 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002390 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002391 if (new_cfqq) {
2392 if (!cfqq->new_cfqq)
2393 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002394 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002395 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002396
2397 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002398 * No requests pending. If the active queue still has requests in
2399 * flight or is idling for a new request, allow either of these
2400 * conditions to happen (or time out) before selecting a new queue.
2401 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002402 if (timer_pending(&cfqd->idle_slice_timer)) {
2403 cfqq = NULL;
2404 goto keep_queue;
2405 }
2406
Shaohua Li8e1ac662010-11-08 15:01:04 +01002407 /*
2408 * This is a deep seek queue, but the device is much faster than
2409 * the queue can deliver, don't idle
2410 **/
2411 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2412 (cfq_cfqq_slice_new(cfqq) ||
2413 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2414 cfq_clear_cfqq_deep(cfqq);
2415 cfq_clear_cfqq_idle_window(cfqq);
2416 }
2417
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002418 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2419 cfqq = NULL;
2420 goto keep_queue;
2421 }
2422
2423 /*
2424 * If group idle is enabled and there are requests dispatched from
2425 * this group, wait for requests to complete.
2426 */
2427check_group_idle:
Shaohua Li7700fc42011-07-12 14:24:56 +02002428 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2429 cfqq->cfqg->dispatched &&
2430 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002431 cfqq = NULL;
2432 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002433 }
2434
Jens Axboe3b181522005-06-27 10:56:24 +02002435expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002436 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002437new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002438 /*
2439 * Current queue expired. Check if we have to switch to a new
2440 * service tree
2441 */
2442 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002443 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002444
Jens Axboea36e71f2009-04-15 12:15:11 +02002445 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002446keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002447 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002448}
2449
Jens Axboefebffd62008-01-28 13:19:43 +01002450static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002451{
2452 int dispatched = 0;
2453
2454 while (cfqq->next_rq) {
2455 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2456 dispatched++;
2457 }
2458
2459 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002460
2461 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002462 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002463 return dispatched;
2464}
2465
Jens Axboe498d3aa2007-04-26 12:54:48 +02002466/*
2467 * Drain our current requests. Used for barriers and when switching
2468 * io schedulers on-the-fly.
2469 */
Jens Axboed9e76202007-04-20 14:27:50 +02002470static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002471{
Jens Axboe08717142008-01-28 11:38:15 +01002472 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002473 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002474
Divyesh Shah3440c492010-04-09 09:29:57 +02002475 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002476 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002477 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2478 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002479 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002480 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002481
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002482 BUG_ON(cfqd->busy_queues);
2483
Jeff Moyer69237152009-06-12 15:29:30 +02002484 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002485 return dispatched;
2486}
2487
Shaohua Liabc3c742010-03-01 09:20:54 +01002488static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2489 struct cfq_queue *cfqq)
2490{
2491 /* the queue hasn't finished any request, can't estimate */
2492 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01002493 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002494 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2495 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01002496 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002497
Shaohua Lic1e44752010-11-08 15:01:02 +01002498 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01002499}
2500
Jens Axboe0b182d62009-10-06 20:49:37 +02002501static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002502{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002503 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504
Jens Axboe2f5cb732009-04-07 08:51:19 +02002505 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002506 * Drain async requests before we start sync IO
2507 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002508 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002509 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002510
2511 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002512 * If this is an async queue and we have sync IO in flight, let it wait
2513 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002514 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002515 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002516
Shaohua Liabc3c742010-03-01 09:20:54 +01002517 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002518 if (cfq_class_idle(cfqq))
2519 max_dispatch = 1;
2520
2521 /*
2522 * Does this cfqq already have too much IO in flight?
2523 */
2524 if (cfqq->dispatched >= max_dispatch) {
Shaohua Lief8a41d2011-03-07 09:26:29 +01002525 bool promote_sync = false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002526 /*
2527 * idle queue must always only have a single IO in flight
2528 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002529 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002530 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002531
Jens Axboe2f5cb732009-04-07 08:51:19 +02002532 /*
Li, Shaohuac4ade942011-03-23 08:30:34 +01002533 * If there is only one sync queue
2534 * we can ignore async queue here and give the sync
Shaohua Lief8a41d2011-03-07 09:26:29 +01002535 * queue no dispatch limit. The reason is a sync queue can
2536 * preempt async queue, limiting the sync queue doesn't make
2537 * sense. This is useful for aiostress test.
2538 */
Li, Shaohuac4ade942011-03-23 08:30:34 +01002539 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2540 promote_sync = true;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002541
2542 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002543 * We have other queues, don't allow more IO from this one
2544 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002545 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2546 !promote_sync)
Jens Axboe0b182d62009-10-06 20:49:37 +02002547 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002548
Jens Axboe2f5cb732009-04-07 08:51:19 +02002549 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002550 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002551 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002552 if (cfqd->busy_queues == 1 || promote_sync)
Shaohua Liabc3c742010-03-01 09:20:54 +01002553 max_dispatch = -1;
2554 else
2555 /*
2556 * Normally we start throttling cfqq when cfq_quantum/2
2557 * requests have been dispatched. But we can drive
2558 * deeper queue depths at the beginning of slice
2559 * subjected to upper limit of cfq_quantum.
2560 * */
2561 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002562 }
2563
2564 /*
2565 * Async queues must wait a bit before being allowed dispatch.
2566 * We also ramp up the dispatch depth gradually for async IO,
2567 * based on the last sync IO we serviced
2568 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002569 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002570 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002571 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002572
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002573 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002574 if (!depth && !cfqq->dispatched)
2575 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002576 if (depth < max_dispatch)
2577 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 }
2579
Jens Axboe0b182d62009-10-06 20:49:37 +02002580 /*
2581 * If we're below the current max, allow a dispatch
2582 */
2583 return cfqq->dispatched < max_dispatch;
2584}
2585
2586/*
2587 * Dispatch a request from cfqq, moving them to the request queue
2588 * dispatch list.
2589 */
2590static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2591{
2592 struct request *rq;
2593
2594 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2595
2596 if (!cfq_may_dispatch(cfqd, cfqq))
2597 return false;
2598
2599 /*
2600 * follow expired path, else get first next available
2601 */
2602 rq = cfq_check_fifo(cfqq);
2603 if (!rq)
2604 rq = cfqq->next_rq;
2605
2606 /*
2607 * insert request into driver dispatch list
2608 */
2609 cfq_dispatch_insert(cfqd->queue, rq);
2610
2611 if (!cfqd->active_cic) {
Tejun Heoc5869802011-12-14 00:33:41 +01002612 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe0b182d62009-10-06 20:49:37 +02002613
Tejun Heoc5869802011-12-14 00:33:41 +01002614 atomic_long_inc(&cic->icq.ioc->refcount);
Jens Axboe0b182d62009-10-06 20:49:37 +02002615 cfqd->active_cic = cic;
2616 }
2617
2618 return true;
2619}
2620
2621/*
2622 * Find the cfqq that we need to service and move a request from that to the
2623 * dispatch list
2624 */
2625static int cfq_dispatch_requests(struct request_queue *q, int force)
2626{
2627 struct cfq_data *cfqd = q->elevator->elevator_data;
2628 struct cfq_queue *cfqq;
2629
2630 if (!cfqd->busy_queues)
2631 return 0;
2632
2633 if (unlikely(force))
2634 return cfq_forced_dispatch(cfqd);
2635
2636 cfqq = cfq_select_queue(cfqd);
2637 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002638 return 0;
2639
Jens Axboe2f5cb732009-04-07 08:51:19 +02002640 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002641 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002642 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002643 if (!cfq_dispatch_request(cfqd, cfqq))
2644 return 0;
2645
Jens Axboe2f5cb732009-04-07 08:51:19 +02002646 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002647 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002648
2649 /*
2650 * expire an async queue immediately if it has used up its slice. idle
2651 * queue always expire after 1 dispatch round.
2652 */
2653 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2654 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2655 cfq_class_idle(cfqq))) {
2656 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002657 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002658 }
2659
Shan Weib217a902009-09-01 10:06:42 +02002660 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002661 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662}
2663
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664/*
Jens Axboe5e705372006-07-13 12:39:25 +02002665 * task holds one reference to the queue, dropped when task exits. each rq
2666 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002668 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 * queue lock must be held here.
2670 */
2671static void cfq_put_queue(struct cfq_queue *cfqq)
2672{
Jens Axboe22e2c502005-06-27 10:55:12 +02002673 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05002674 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002675
Shaohua Li30d7b942011-01-07 08:46:59 +01002676 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677
Shaohua Li30d7b942011-01-07 08:46:59 +01002678 cfqq->ref--;
2679 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 return;
2681
Jens Axboe7b679132008-05-30 12:23:07 +02002682 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002684 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002685 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002687 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002688 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002689 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002690 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002691
Vivek Goyalf04a6422009-12-03 12:59:40 -05002692 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002694 cfq_put_cfqg(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695}
2696
Shaohua Lid02a2c02010-05-25 10:16:53 +02002697static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002698{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002699 struct cfq_queue *__cfqq, *next;
2700
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002701 /*
2702 * If this queue was scheduled to merge with another queue, be
2703 * sure to drop the reference taken on that queue (and others in
2704 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2705 */
2706 __cfqq = cfqq->new_cfqq;
2707 while (__cfqq) {
2708 if (__cfqq == cfqq) {
2709 WARN(1, "cfqq->new_cfqq loop detected\n");
2710 break;
2711 }
2712 next = __cfqq->new_cfqq;
2713 cfq_put_queue(__cfqq);
2714 __cfqq = next;
2715 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002716}
2717
2718static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2719{
2720 if (unlikely(cfqq == cfqd->active_queue)) {
2721 __cfq_slice_expired(cfqd, cfqq, 0);
2722 cfq_schedule_dispatch(cfqd);
2723 }
2724
2725 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002726
Jens Axboe89850f72006-07-22 16:48:31 +02002727 cfq_put_queue(cfqq);
2728}
2729
Tejun Heo9b84cac2011-12-14 00:33:42 +01002730static void cfq_init_icq(struct io_cq *icq)
2731{
2732 struct cfq_io_cq *cic = icq_to_cic(icq);
2733
2734 cic->ttime.last_end_request = jiffies;
2735}
2736
Tejun Heoc5869802011-12-14 00:33:41 +01002737static void cfq_exit_icq(struct io_cq *icq)
Jens Axboe89850f72006-07-22 16:48:31 +02002738{
Tejun Heoc5869802011-12-14 00:33:41 +01002739 struct cfq_io_cq *cic = icq_to_cic(icq);
Tejun Heo283287a2011-12-14 00:33:38 +01002740 struct cfq_data *cfqd = cic_to_cfqd(cic);
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002741
Jens Axboeff6657c2009-04-08 10:58:57 +02002742 if (cic->cfqq[BLK_RW_ASYNC]) {
2743 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2744 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002745 }
2746
Jens Axboeff6657c2009-04-08 10:58:57 +02002747 if (cic->cfqq[BLK_RW_SYNC]) {
2748 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2749 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002750 }
Jens Axboe89850f72006-07-22 16:48:31 +02002751}
2752
Jens Axboefd0928d2008-01-24 08:52:45 +01002753static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002754{
2755 struct task_struct *tsk = current;
2756 int ioprio_class;
2757
Jens Axboe3b181522005-06-27 10:56:24 +02002758 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002759 return;
2760
Jens Axboefd0928d2008-01-24 08:52:45 +01002761 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002762 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002763 default:
2764 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2765 case IOPRIO_CLASS_NONE:
2766 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002767 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002768 */
2769 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002770 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002771 break;
2772 case IOPRIO_CLASS_RT:
2773 cfqq->ioprio = task_ioprio(ioc);
2774 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2775 break;
2776 case IOPRIO_CLASS_BE:
2777 cfqq->ioprio = task_ioprio(ioc);
2778 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2779 break;
2780 case IOPRIO_CLASS_IDLE:
2781 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2782 cfqq->ioprio = 7;
2783 cfq_clear_cfqq_idle_window(cfqq);
2784 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002785 }
2786
2787 /*
2788 * keep track of original prio settings in case we have to temporarily
2789 * elevate the priority of this queue
2790 */
2791 cfqq->org_ioprio = cfqq->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02002792 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002793}
2794
Tejun Heoc5869802011-12-14 00:33:41 +01002795static void changed_ioprio(struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002796{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002797 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002798 struct cfq_queue *cfqq;
Jens Axboe35e60772006-06-14 09:10:45 +02002799
Jens Axboecaaa5f92006-06-16 11:23:00 +02002800 if (unlikely(!cfqd))
2801 return;
2802
Jens Axboeff6657c2009-04-08 10:58:57 +02002803 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002804 if (cfqq) {
2805 struct cfq_queue *new_cfqq;
Tejun Heoc5869802011-12-14 00:33:41 +01002806 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
Jens Axboeff6657c2009-04-08 10:58:57 +02002807 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002808 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002809 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002810 cfq_put_queue(cfqq);
2811 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002812 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002813
Jens Axboeff6657c2009-04-08 10:58:57 +02002814 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002815 if (cfqq)
2816 cfq_mark_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002817}
2818
Jens Axboed5036d72009-06-26 10:44:34 +02002819static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002820 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002821{
2822 RB_CLEAR_NODE(&cfqq->rb_node);
2823 RB_CLEAR_NODE(&cfqq->p_node);
2824 INIT_LIST_HEAD(&cfqq->fifo);
2825
Shaohua Li30d7b942011-01-07 08:46:59 +01002826 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02002827 cfqq->cfqd = cfqd;
2828
2829 cfq_mark_cfqq_prio_changed(cfqq);
2830
2831 if (is_sync) {
2832 if (!cfq_class_idle(cfqq))
2833 cfq_mark_cfqq_idle_window(cfqq);
2834 cfq_mark_cfqq_sync(cfqq);
2835 }
2836 cfqq->pid = pid;
2837}
2838
Vivek Goyal24610332009-12-03 12:59:51 -05002839#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heoc5869802011-12-14 00:33:41 +01002840static void changed_cgroup(struct cfq_io_cq *cic)
Vivek Goyal24610332009-12-03 12:59:51 -05002841{
2842 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002843 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal24610332009-12-03 12:59:51 -05002844 struct request_queue *q;
2845
2846 if (unlikely(!cfqd))
2847 return;
2848
2849 q = cfqd->queue;
2850
Vivek Goyal24610332009-12-03 12:59:51 -05002851 if (sync_cfqq) {
2852 /*
2853 * Drop reference to sync queue. A new sync queue will be
2854 * assigned in new group upon arrival of a fresh request.
2855 */
2856 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2857 cic_set_cfqq(cic, NULL, 1);
2858 cfq_put_queue(sync_cfqq);
2859 }
Vivek Goyal24610332009-12-03 12:59:51 -05002860}
Vivek Goyal24610332009-12-03 12:59:51 -05002861#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2862
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002864cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002865 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 struct cfq_queue *cfqq, *new_cfqq = NULL;
Tejun Heoc5869802011-12-14 00:33:41 +01002868 struct cfq_io_cq *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002869 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870
2871retry:
Vivek Goyal3e59cf92011-05-19 15:38:21 -04002872 cfqg = cfq_get_cfqg(cfqd);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002873 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002874 /* cic always exists here */
2875 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876
Jens Axboe6118b702009-06-30 09:34:12 +02002877 /*
2878 * Always try a new alloc if we fell back to the OOM cfqq
2879 * originally, since it should just be a temporary situation.
2880 */
2881 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2882 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 if (new_cfqq) {
2884 cfqq = new_cfqq;
2885 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002886 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002888 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002889 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002890 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002892 if (new_cfqq)
2893 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002894 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002895 cfqq = kmem_cache_alloc_node(cfq_pool,
2896 gfp_mask | __GFP_ZERO,
2897 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002898 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899
Jens Axboe6118b702009-06-30 09:34:12 +02002900 if (cfqq) {
2901 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2902 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002903 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002904 cfq_log_cfqq(cfqd, cfqq, "alloced");
2905 } else
2906 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907 }
2908
2909 if (new_cfqq)
2910 kmem_cache_free(cfq_pool, new_cfqq);
2911
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 return cfqq;
2913}
2914
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002915static struct cfq_queue **
2916cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2917{
Jens Axboefe094d92008-01-31 13:08:54 +01002918 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002919 case IOPRIO_CLASS_RT:
2920 return &cfqd->async_cfqq[0][ioprio];
2921 case IOPRIO_CLASS_BE:
2922 return &cfqd->async_cfqq[1][ioprio];
2923 case IOPRIO_CLASS_IDLE:
2924 return &cfqd->async_idle_cfqq;
2925 default:
2926 BUG();
2927 }
2928}
2929
Jens Axboe15c31be2007-07-10 13:43:25 +02002930static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002931cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002932 gfp_t gfp_mask)
2933{
Jens Axboefd0928d2008-01-24 08:52:45 +01002934 const int ioprio = task_ioprio(ioc);
2935 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002936 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002937 struct cfq_queue *cfqq = NULL;
2938
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002939 if (!is_sync) {
2940 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2941 cfqq = *async_cfqq;
2942 }
2943
Jens Axboe6118b702009-06-30 09:34:12 +02002944 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002945 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002946
2947 /*
2948 * pin the queue now that it's allocated, scheduler exit will prune it
2949 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002950 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01002951 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002952 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002953 }
2954
Shaohua Li30d7b942011-01-07 08:46:59 +01002955 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02002956 return cfqq;
2957}
2958
Jens Axboe22e2c502005-06-27 10:55:12 +02002959static void
Shaohua Li383cd722011-07-12 14:24:35 +02002960__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02002961{
Shaohua Li383cd722011-07-12 14:24:35 +02002962 unsigned long elapsed = jiffies - ttime->last_end_request;
2963 elapsed = min(elapsed, 2UL * slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02002964
Shaohua Li383cd722011-07-12 14:24:35 +02002965 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
2966 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
2967 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
2968}
2969
2970static void
2971cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01002972 struct cfq_io_cq *cic)
Shaohua Li383cd722011-07-12 14:24:35 +02002973{
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002974 if (cfq_cfqq_sync(cfqq)) {
Shaohua Li383cd722011-07-12 14:24:35 +02002975 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02002976 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
2977 cfqd->cfq_slice_idle);
2978 }
Shaohua Li7700fc42011-07-12 14:24:56 +02002979#ifdef CONFIG_CFQ_GROUP_IOSCHED
2980 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
2981#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02002982}
2983
Jens Axboe206dc692006-03-28 13:03:44 +02002984static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002985cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02002986 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02002987{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002988 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01002989 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002990 if (cfqq->last_request_pos) {
2991 if (cfqq->last_request_pos < blk_rq_pos(rq))
2992 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2993 else
2994 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2995 }
Jens Axboe206dc692006-03-28 13:03:44 +02002996
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01002997 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01002998 if (blk_queue_nonrot(cfqd->queue))
2999 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3000 else
3001 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003002}
Jens Axboe22e2c502005-06-27 10:55:12 +02003003
3004/*
3005 * Disable idle window if the process thinks too long or seeks so much that
3006 * it doesn't matter
3007 */
3008static void
3009cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01003010 struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02003011{
Jens Axboe7b679132008-05-30 12:23:07 +02003012 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003013
Jens Axboe08717142008-01-28 11:38:15 +01003014 /*
3015 * Don't idle for async or idle io prio class
3016 */
3017 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003018 return;
3019
Jens Axboec265a7f2008-06-26 13:49:33 +02003020 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003021
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003022 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3023 cfq_mark_cfqq_deep(cfqq);
3024
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003025 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3026 enable_idle = 0;
Tejun Heoc5869802011-12-14 00:33:41 +01003027 else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
3028 !cfqd->cfq_slice_idle ||
3029 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003030 enable_idle = 0;
Shaohua Li383cd722011-07-12 14:24:35 +02003031 else if (sample_valid(cic->ttime.ttime_samples)) {
3032 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003033 enable_idle = 0;
3034 else
3035 enable_idle = 1;
3036 }
3037
Jens Axboe7b679132008-05-30 12:23:07 +02003038 if (old_idle != enable_idle) {
3039 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3040 if (enable_idle)
3041 cfq_mark_cfqq_idle_window(cfqq);
3042 else
3043 cfq_clear_cfqq_idle_window(cfqq);
3044 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003045}
3046
Jens Axboe22e2c502005-06-27 10:55:12 +02003047/*
3048 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3049 * no or if we aren't sure, a 1 will cause a preempt.
3050 */
Jens Axboea6151c32009-10-07 20:02:57 +02003051static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003052cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003053 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003054{
Jens Axboe6d048f52007-04-25 12:44:27 +02003055 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003056
Jens Axboe6d048f52007-04-25 12:44:27 +02003057 cfqq = cfqd->active_queue;
3058 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003059 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003060
Jens Axboe6d048f52007-04-25 12:44:27 +02003061 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003062 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003063
3064 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003065 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003066
Jens Axboe22e2c502005-06-27 10:55:12 +02003067 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003068 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3069 */
3070 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3071 return false;
3072
3073 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003074 * if the new request is sync, but the currently running queue is
3075 * not, let the sync request have priority.
3076 */
Jens Axboe5e705372006-07-13 12:39:25 +02003077 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003078 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003079
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003080 if (new_cfqq->cfqg != cfqq->cfqg)
3081 return false;
3082
3083 if (cfq_slice_used(cfqq))
3084 return true;
3085
3086 /* Allow preemption only if we are idling on sync-noidle tree */
3087 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3088 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3089 new_cfqq->service_tree->count == 2 &&
3090 RB_EMPTY_ROOT(&cfqq->sort_list))
3091 return true;
3092
Jens Axboe374f84a2006-07-23 01:42:19 +02003093 /*
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003094 * So both queues are sync. Let the new request get disk time if
3095 * it's a metadata request and the current queue is doing regular IO.
3096 */
Christoph Hellwig65299a32011-08-23 14:50:29 +02003097 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003098 return true;
3099
3100 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003101 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3102 */
3103 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003104 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003105
Shaohua Lid2d59e12010-11-08 15:01:03 +01003106 /* An idle queue should not be idle now for some reason */
3107 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3108 return true;
3109
Jens Axboe1e3335d2007-02-14 19:59:49 +01003110 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003111 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003112
3113 /*
3114 * if this request is as-good as one we would expect from the
3115 * current cfqq, let it preempt
3116 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003117 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003118 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003119
Jens Axboea6151c32009-10-07 20:02:57 +02003120 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003121}
3122
3123/*
3124 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3125 * let it have half of its nominal slice.
3126 */
3127static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3128{
Shaohua Lidf0793a2012-01-19 09:20:09 +01003129 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3130
Jens Axboe7b679132008-05-30 12:23:07 +02003131 cfq_log_cfqq(cfqd, cfqq, "preempt");
Shaohua Lidf0793a2012-01-19 09:20:09 +01003132 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003133
Jens Axboebf572252006-07-19 20:29:12 +02003134 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003135 * workload type is changed, don't save slice, otherwise preempt
3136 * doesn't happen
3137 */
Shaohua Lidf0793a2012-01-19 09:20:09 +01003138 if (old_type != cfqq_type(cfqq))
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003139 cfqq->cfqg->saved_workload_slice = 0;
3140
3141 /*
Jens Axboebf572252006-07-19 20:29:12 +02003142 * Put the new queue at the front of the of the current list,
3143 * so we know that it will be selected next.
3144 */
3145 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003146
3147 cfq_service_tree_add(cfqd, cfqq, 1);
Justin TerAvesteda5e0c2011-03-22 21:26:49 +01003148
Justin TerAvest62a37f62011-03-23 08:25:44 +01003149 cfqq->slice_end = 0;
3150 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003151}
3152
3153/*
Jens Axboe5e705372006-07-13 12:39:25 +02003154 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003155 * something we should do about it
3156 */
3157static void
Jens Axboe5e705372006-07-13 12:39:25 +02003158cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3159 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003160{
Tejun Heoc5869802011-12-14 00:33:41 +01003161 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003162
Aaron Carroll45333d52008-08-26 15:52:36 +02003163 cfqd->rq_queued++;
Christoph Hellwig65299a32011-08-23 14:50:29 +02003164 if (rq->cmd_flags & REQ_PRIO)
3165 cfqq->prio_pending++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003166
Shaohua Li383cd722011-07-12 14:24:35 +02003167 cfq_update_io_thinktime(cfqd, cfqq, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003168 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003169 cfq_update_idle_window(cfqd, cfqq, cic);
3170
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003171 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003172
3173 if (cfqq == cfqd->active_queue) {
3174 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003175 * Remember that we saw a request from this process, but
3176 * don't start queuing just yet. Otherwise we risk seeing lots
3177 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003178 * and merging. If the request is already larger than a single
3179 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003180 * merging is already done. Ditto for a busy system that
3181 * has other work pending, don't risk delaying until the
3182 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003183 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003184 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003185 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3186 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003187 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003188 cfq_clear_cfqq_wait_request(cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003189 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003190 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003191 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003192 &cfqq->cfqg->blkg);
Vivek Goyalbf791932009-12-03 12:59:37 -05003193 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003194 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003195 }
Jens Axboe5e705372006-07-13 12:39:25 +02003196 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003197 /*
3198 * not the active queue - expire current slice if it is
3199 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003200 * has some old slice time left and is of higher priority or
3201 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003202 */
3203 cfq_preempt_queue(cfqd, cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003204 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003205 }
3206}
3207
Tatyana Brokhmanb410a822013-07-07 17:33:05 +03003208/*
3209 * Called when a request (rq) is reinserted (to cfqq). Check if there's
3210 * something we should do about it
3211 */
3212static void
3213cfq_rq_requeued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3214 struct request *rq)
3215{
3216 struct cfq_io_cq *cic = RQ_CIC(rq);
3217
3218 cfqd->rq_queued++;
3219 if (rq->cmd_flags & REQ_PRIO)
3220 cfqq->prio_pending++;
3221
3222 cfqq->dispatched--;
3223 (RQ_CFQG(rq))->dispatched--;
3224
3225 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3226
3227 cfq_update_io_thinktime(cfqd, cfqq, cic);
3228 cfq_update_io_seektime(cfqd, cfqq, rq);
3229 cfq_update_idle_window(cfqd, cfqq, cic);
3230
3231 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3232
3233 if (cfqq == cfqd->active_queue) {
3234 if (cfq_cfqq_wait_request(cfqq)) {
3235 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3236 cfqd->busy_queues > 1) {
3237 cfq_del_timer(cfqd, cfqq);
3238 cfq_clear_cfqq_wait_request(cfqq);
3239 } else {
3240 cfq_blkiocg_update_idle_time_stats(
3241 &cfqq->cfqg->blkg);
3242 cfq_mark_cfqq_must_dispatch(cfqq);
3243 }
3244 }
3245 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3246 cfq_preempt_queue(cfqd, cfqq);
3247 }
3248}
3249
3250static int cfq_reinsert_request(struct request_queue *q, struct request *rq)
3251{
3252 struct cfq_data *cfqd = q->elevator->elevator_data;
3253 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3254
3255 if (!cfqq || cfqq->cfqd != cfqd)
3256 return -EIO;
3257
3258 cfq_log_cfqq(cfqd, cfqq, "re-insert_request");
3259 list_add(&rq->queuelist, &cfqq->fifo);
3260 cfq_add_rq_rb(rq);
3261
3262 cfq_rq_requeued(cfqd, cfqq, rq);
3263 if (rq->cmd_flags & REQ_URGENT) {
3264 if (cfqd->nr_urgent_in_flight)
3265 cfqd->nr_urgent_in_flight--;
3266 cfqd->nr_urgent_pending++;
3267 }
3268 return 0;
3269}
3270
Jens Axboe165125e2007-07-24 09:28:11 +02003271static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003272{
Jens Axboeb4878f22005-10-20 16:42:29 +02003273 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003274 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003275
Jens Axboe7b679132008-05-30 12:23:07 +02003276 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Tejun Heoc5869802011-12-14 00:33:41 +01003277 cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278
Jens Axboe30996f42009-10-05 11:03:39 +02003279 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003280 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003281 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003282 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003283 &cfqd->serving_group->blkg, rq_data_dir(rq),
3284 rq_is_sync(rq));
Tatyana Brokhmanb410a822013-07-07 17:33:05 +03003285
Jens Axboe5e705372006-07-13 12:39:25 +02003286 cfq_rq_enqueued(cfqd, cfqq, rq);
Tatyana Brokhmanb410a822013-07-07 17:33:05 +03003287
3288 if (rq->cmd_flags & REQ_URGENT) {
3289 WARN_ON(1);
3290 blk_dump_rq_flags(rq, "");
3291 rq->cmd_flags &= ~REQ_URGENT;
3292 }
3293
3294 /* Request is considered URGENT if:
3295 * 1. The queue being served is of a lower IO priority then the new
3296 * request
3297 * OR:
3298 * 2. The workload being performed is ASYNC
3299 * Only READ requests may be considered as URGENT
3300 */
3301 if ((cfqd->active_queue &&
3302 cfqq->ioprio_class < cfqd->active_queue->ioprio_class) ||
3303 (cfqd->serving_type == ASYNC_WORKLOAD &&
3304 rq_data_dir(rq) == READ)) {
3305 rq->cmd_flags |= REQ_URGENT;
3306 cfqd->nr_urgent_pending++;
3307 }
3308}
3309
3310
3311/**
3312 * cfq_urgent_pending() - Return TRUE if there is an urgent
3313 * request on scheduler
3314 * @q: requests queue
3315 */
3316static bool cfq_urgent_pending(struct request_queue *q)
3317{
3318 struct cfq_data *cfqd = q->elevator->elevator_data;
3319
3320 if (cfqd->nr_urgent_pending && !cfqd->nr_urgent_in_flight)
3321 return true;
3322
3323 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324}
3325
Aaron Carroll45333d52008-08-26 15:52:36 +02003326/*
3327 * Update hw_tag based on peak queue depth over 50 samples under
3328 * sufficient load.
3329 */
3330static void cfq_update_hw_tag(struct cfq_data *cfqd)
3331{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003332 struct cfq_queue *cfqq = cfqd->active_queue;
3333
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003334 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3335 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003336
3337 if (cfqd->hw_tag == 1)
3338 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003339
3340 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003341 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003342 return;
3343
Shaohua Li1a1238a2009-10-27 08:46:23 +01003344 /*
3345 * If active queue hasn't enough requests and can idle, cfq might not
3346 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3347 * case
3348 */
3349 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3350 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003351 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003352 return;
3353
Aaron Carroll45333d52008-08-26 15:52:36 +02003354 if (cfqd->hw_tag_samples++ < 50)
3355 return;
3356
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003357 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003358 cfqd->hw_tag = 1;
3359 else
3360 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003361}
3362
Vivek Goyal7667aa02009-12-08 17:52:58 -05003363static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3364{
Tejun Heoc5869802011-12-14 00:33:41 +01003365 struct cfq_io_cq *cic = cfqd->active_cic;
Vivek Goyal7667aa02009-12-08 17:52:58 -05003366
Justin TerAvest02a8f012011-02-09 14:20:03 +01003367 /* If the queue already has requests, don't wait */
3368 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3369 return false;
3370
Vivek Goyal7667aa02009-12-08 17:52:58 -05003371 /* If there are other queues in the group, don't wait */
3372 if (cfqq->cfqg->nr_cfqq > 1)
3373 return false;
3374
Shaohua Li7700fc42011-07-12 14:24:56 +02003375 /* the only queue in the group, but think time is big */
3376 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3377 return false;
3378
Vivek Goyal7667aa02009-12-08 17:52:58 -05003379 if (cfq_slice_used(cfqq))
3380 return true;
3381
3382 /* if slice left is less than think time, wait busy */
Shaohua Li383cd722011-07-12 14:24:35 +02003383 if (cic && sample_valid(cic->ttime.ttime_samples)
3384 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
Vivek Goyal7667aa02009-12-08 17:52:58 -05003385 return true;
3386
3387 /*
3388 * If think times is less than a jiffy than ttime_mean=0 and above
3389 * will not be true. It might happen that slice has not expired yet
3390 * but will expire soon (4-5 ns) during select_queue(). To cover the
3391 * case where think time is less than a jiffy, mark the queue wait
3392 * busy if only 1 jiffy is left in the slice.
3393 */
3394 if (cfqq->slice_end - jiffies == 1)
3395 return true;
3396
3397 return false;
3398}
3399
Jens Axboe165125e2007-07-24 09:28:11 +02003400static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401{
Jens Axboe5e705372006-07-13 12:39:25 +02003402 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003403 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003404 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003405 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406
Tatyana Brokhmanb410a822013-07-07 17:33:05 +03003407 if (rq->cmd_flags & REQ_URGENT) {
3408 if (!cfqd->nr_urgent_in_flight)
3409 WARN_ON(1);
3410 else
3411 cfqd->nr_urgent_in_flight--;
3412 rq->cmd_flags &= ~REQ_URGENT;
3413 }
3414
Jens Axboeb4878f22005-10-20 16:42:29 +02003415 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003416 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3417 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418
Aaron Carroll45333d52008-08-26 15:52:36 +02003419 cfq_update_hw_tag(cfqd);
3420
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003421 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003422 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003423 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003424 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003425 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003426 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3427 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3428 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003430 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003431
Vivek Goyal365722b2009-10-03 15:21:27 +02003432 if (sync) {
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003433 struct cfq_rb_root *service_tree;
3434
Shaohua Li383cd722011-07-12 14:24:35 +02003435 RQ_CIC(rq)->ttime.last_end_request = now;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003436
3437 if (cfq_cfqq_on_rr(cfqq))
3438 service_tree = cfqq->service_tree;
3439 else
3440 service_tree = service_tree_for(cfqq->cfqg,
3441 cfqq_prio(cfqq), cfqq_type(cfqq));
3442 service_tree->ttime.last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003443 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3444 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003445 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003446
Shaohua Li7700fc42011-07-12 14:24:56 +02003447#ifdef CONFIG_CFQ_GROUP_IOSCHED
3448 cfqq->cfqg->ttime.last_end_request = now;
3449#endif
3450
Jens Axboecaaa5f92006-06-16 11:23:00 +02003451 /*
3452 * If this is the active queue, check if it needs to be expired,
3453 * or if we want to idle in case it has no pending requests.
3454 */
3455 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003456 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3457
Jens Axboe44f7c162007-01-19 11:51:58 +11003458 if (cfq_cfqq_slice_new(cfqq)) {
3459 cfq_set_prio_slice(cfqd, cfqq);
3460 cfq_clear_cfqq_slice_new(cfqq);
3461 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003462
3463 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003464 * Should we wait for next request to come in before we expire
3465 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003466 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003467 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003468 unsigned long extend_sl = cfqd->cfq_slice_idle;
3469 if (!cfqd->cfq_slice_idle)
3470 extend_sl = cfqd->cfq_group_idle;
3471 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003472 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003473 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003474 }
3475
Jens Axboea36e71f2009-04-15 12:15:11 +02003476 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003477 * Idling is not enabled on:
3478 * - expired queues
3479 * - idle-priority queues
3480 * - async queues
3481 * - queues with still some requests queued
3482 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003483 */
Jens Axboe08717142008-01-28 11:38:15 +01003484 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003485 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003486 else if (sync && cfqq_empty &&
3487 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003488 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003489 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003490 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003491
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003492 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003493 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494}
3495
Jens Axboe89850f72006-07-22 16:48:31 +02003496static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003497{
Jens Axboe1b379d82009-08-11 08:26:11 +02003498 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003499 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003500 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003501 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003502
3503 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003504}
3505
Jens Axboe165125e2007-07-24 09:28:11 +02003506static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003507{
3508 struct cfq_data *cfqd = q->elevator->elevator_data;
3509 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01003510 struct cfq_io_cq *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003511 struct cfq_queue *cfqq;
3512
3513 /*
3514 * don't force setup of a queue from here, as a call to may_queue
3515 * does not necessarily imply that a request actually will be queued.
3516 * so just lookup a possibly existing queue, or return 'may queue'
3517 * if that fails
3518 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003519 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003520 if (!cic)
3521 return ELV_MQUEUE_MAY;
3522
Jens Axboeb0b78f82009-04-08 10:56:08 +02003523 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003524 if (cfqq) {
Tejun Heoc5869802011-12-14 00:33:41 +01003525 cfq_init_prio_data(cfqq, cic->icq.ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003526
Jens Axboe89850f72006-07-22 16:48:31 +02003527 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003528 }
3529
3530 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531}
3532
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533/*
3534 * queue lock held here
3535 */
Jens Axboebb37b942006-12-01 10:42:33 +01003536static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537{
Jens Axboe5e705372006-07-13 12:39:25 +02003538 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539
Jens Axboe5e705372006-07-13 12:39:25 +02003540 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003541 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542
Jens Axboe22e2c502005-06-27 10:55:12 +02003543 BUG_ON(!cfqq->allocated[rw]);
3544 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003546 /* Put down rq reference on cfqg */
3547 cfq_put_cfqg(RQ_CFQG(rq));
Tejun Heoa612fdd2011-12-14 00:33:41 +01003548 rq->elv.priv[0] = NULL;
3549 rq->elv.priv[1] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003550
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551 cfq_put_queue(cfqq);
3552 }
3553}
3554
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003555static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01003556cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003557 struct cfq_queue *cfqq)
3558{
3559 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3560 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003561 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003562 cfq_put_queue(cfqq);
3563 return cic_to_cfqq(cic, 1);
3564}
3565
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003566/*
3567 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3568 * was the last process referring to said cfqq.
3569 */
3570static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01003571split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003572{
3573 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003574 cfqq->pid = current->pid;
3575 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003576 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003577 return cfqq;
3578 }
3579
3580 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003581
3582 cfq_put_cooperator(cfqq);
3583
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003584 cfq_put_queue(cfqq);
3585 return NULL;
3586}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003588 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003590static int
Jens Axboe165125e2007-07-24 09:28:11 +02003591cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592{
3593 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heof1f8cc92011-12-14 00:33:42 +01003594 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003596 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003597 struct cfq_queue *cfqq;
Tejun Heod705ae62012-02-15 09:45:49 +01003598 unsigned int changed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599
3600 might_sleep_if(gfp_mask & __GFP_WAIT);
3601
Tejun Heo216284c2011-12-14 00:33:38 +01003602 spin_lock_irq(q->queue_lock);
Tejun Heof1f8cc92011-12-14 00:33:42 +01003603
3604 /* handle changed notifications */
Tejun Heod705ae62012-02-15 09:45:49 +01003605 changed = icq_get_changed(&cic->icq);
3606 if (unlikely(changed & ICQ_IOPRIO_CHANGED))
3607 changed_ioprio(cic);
Tejun Heof1f8cc92011-12-14 00:33:42 +01003608#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heod705ae62012-02-15 09:45:49 +01003609 if (unlikely(changed & ICQ_CGROUP_CHANGED))
3610 changed_cgroup(cic);
Tejun Heof1f8cc92011-12-14 00:33:42 +01003611#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003613new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003614 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003615 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Tejun Heoc5869802011-12-14 00:33:41 +01003616 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003617 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003618 } else {
3619 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003620 * If the queue was seeky for too long, break it apart.
3621 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003622 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003623 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3624 cfqq = split_cfqq(cic, cfqq);
3625 if (!cfqq)
3626 goto new_queue;
3627 }
3628
3629 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003630 * Check to see if this queue is scheduled to merge with
3631 * another, closely cooperating queue. The merging of
3632 * queues happens here as it must be done in process context.
3633 * The reference on new_cfqq was taken in merge_cfqqs.
3634 */
3635 if (cfqq->new_cfqq)
3636 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003637 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638
3639 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02003640
Jens Axboe6fae9c22011-03-01 15:04:39 -05003641 cfqq->ref++;
Tejun Heoa612fdd2011-12-14 00:33:41 +01003642 rq->elv.priv[0] = cfqq;
3643 rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
Tejun Heo216284c2011-12-14 00:33:38 +01003644 spin_unlock_irq(q->queue_lock);
Jens Axboe5e705372006-07-13 12:39:25 +02003645 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646}
3647
David Howells65f27f32006-11-22 14:55:48 +00003648static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003649{
David Howells65f27f32006-11-22 14:55:48 +00003650 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003651 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003652 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003653
Jens Axboe40bb54d2009-04-15 12:11:10 +02003654 spin_lock_irq(q->queue_lock);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003655 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003656 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003657}
3658
3659/*
3660 * Timer running if the active_queue is currently idling inside its time slice
3661 */
3662static void cfq_idle_slice_timer(unsigned long data)
3663{
3664 struct cfq_data *cfqd = (struct cfq_data *) data;
3665 struct cfq_queue *cfqq;
3666 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003667 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003668
Jens Axboe7b679132008-05-30 12:23:07 +02003669 cfq_log(cfqd, "idle timer fired");
3670
Jens Axboe22e2c502005-06-27 10:55:12 +02003671 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3672
Jens Axboefe094d92008-01-31 13:08:54 +01003673 cfqq = cfqd->active_queue;
3674 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003675 timed_out = 0;
3676
Jens Axboe22e2c502005-06-27 10:55:12 +02003677 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003678 * We saw a request before the queue expired, let it through
3679 */
3680 if (cfq_cfqq_must_dispatch(cfqq))
3681 goto out_kick;
3682
3683 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003684 * expired
3685 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003686 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003687 goto expire;
3688
3689 /*
3690 * only expire and reinvoke request handler, if there are
3691 * other queues with pending requests
3692 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003693 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003694 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003695
3696 /*
3697 * not expired and it has a request pending, let it dispatch
3698 */
Jens Axboe75e50982009-04-07 08:56:14 +02003699 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003700 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003701
3702 /*
3703 * Queue depth flag is reset only when the idle didn't succeed
3704 */
3705 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003706 }
3707expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003708 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003709out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003710 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003711out_cont:
3712 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3713}
3714
Jens Axboe3b181522005-06-27 10:56:24 +02003715static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3716{
3717 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003718 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003719}
Jens Axboe22e2c502005-06-27 10:55:12 +02003720
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003721static void cfq_put_async_queues(struct cfq_data *cfqd)
3722{
3723 int i;
3724
3725 for (i = 0; i < IOPRIO_BE_NR; i++) {
3726 if (cfqd->async_cfqq[0][i])
3727 cfq_put_queue(cfqd->async_cfqq[0][i]);
3728 if (cfqd->async_cfqq[1][i])
3729 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003730 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003731
3732 if (cfqd->async_idle_cfqq)
3733 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003734}
3735
Jens Axboeb374d182008-10-31 10:05:07 +01003736static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003737{
Jens Axboe22e2c502005-06-27 10:55:12 +02003738 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003739 struct request_queue *q = cfqd->queue;
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003740 bool wait = false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003741
Jens Axboe3b181522005-06-27 10:56:24 +02003742 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003743
Al Virod9ff4182006-03-18 13:51:22 -05003744 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003745
Al Virod9ff4182006-03-18 13:51:22 -05003746 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003747 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003748
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003749 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003750 cfq_release_cfq_groups(cfqd);
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003751
3752 /*
3753 * If there are groups which we could not unlink from blkcg list,
3754 * wait for a rcu period for them to be freed.
3755 */
3756 if (cfqd->nr_blkcg_linked_grps)
3757 wait = true;
Jens Axboe15c31be2007-07-10 13:43:25 +02003758
Al Virod9ff4182006-03-18 13:51:22 -05003759 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003760
3761 cfq_shutdown_timer_wq(cfqd);
3762
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003763 /*
3764 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3765 * Do this wait only if there are other unlinked groups out
3766 * there. This can happen if cgroup deletion path claimed the
3767 * responsibility of cleaning up a group before queue cleanup code
3768 * get to the group.
3769 *
3770 * Do not call synchronize_rcu() unconditionally as there are drivers
3771 * which create/delete request queue hundreds of times during scan/boot
3772 * and synchronize_rcu() can take significant time and slow down boot.
3773 */
3774 if (wait)
3775 synchronize_rcu();
Vivek Goyal2abae552011-05-23 10:02:19 +02003776
3777#ifdef CONFIG_CFQ_GROUP_IOSCHED
3778 /* Free up per cpu stats for root group */
3779 free_percpu(cfqd->root_group.blkg.stats_cpu);
3780#endif
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003781 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003782}
3783
Jens Axboe165125e2007-07-24 09:28:11 +02003784static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003785{
3786 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003787 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003788 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003789 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790
Christoph Lameter94f60302007-07-17 04:03:29 -07003791 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Tejun Heoa73f7302011-12-14 00:33:37 +01003792 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003793 return NULL;
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003794
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003795 /* Init root service tree */
3796 cfqd->grp_service_tree = CFQ_RB_ROOT;
3797
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003798 /* Init root group */
3799 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003800 for_each_cfqg_st(cfqg, i, j, st)
3801 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003802 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003803
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003804 /* Give preference to root group over other groups */
3805 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3806
Vivek Goyal25fb5162009-12-03 12:59:46 -05003807#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003808 /*
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003809 * Set root group reference to 2. One reference will be dropped when
3810 * all groups on cfqd->cfqg_list are being deleted during queue exit.
3811 * Other reference will remain there as we don't want to delete this
3812 * group as it is statically allocated and gets destroyed when
3813 * throtl_data goes away.
Vivek Goyalb1c35762009-12-03 12:59:47 -05003814 */
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003815 cfqg->ref = 2;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04003816
3817 if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3818 kfree(cfqg);
3819 kfree(cfqd);
3820 return NULL;
3821 }
3822
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003823 rcu_read_lock();
Vivek Goyal5624a4e2011-05-19 15:38:28 -04003824
Vivek Goyale98ef892010-06-18 10:39:47 -04003825 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3826 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04003827 rcu_read_unlock();
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003828 cfqd->nr_blkcg_linked_grps++;
3829
3830 /* Add group on cfqd->cfqg_list */
3831 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
Vivek Goyal25fb5162009-12-03 12:59:46 -05003832#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003833 /*
3834 * Not strictly needed (since RB_ROOT just clears the node and we
3835 * zeroed cfqd on alloc), but better be safe in case someone decides
3836 * to add magic to the rb code
3837 */
3838 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3839 cfqd->prio_trees[i] = RB_ROOT;
3840
Jens Axboe6118b702009-06-30 09:34:12 +02003841 /*
3842 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3843 * Grab a permanent reference to it, so that the normal code flow
3844 * will not attempt to free it.
3845 */
3846 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01003847 cfqd->oom_cfqq.ref++;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003848 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003849
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851
Jens Axboe22e2c502005-06-27 10:55:12 +02003852 init_timer(&cfqd->idle_slice_timer);
3853 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3854 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3855
Jens Axboe23e018a2009-10-05 08:52:35 +02003856 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003857
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003859 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3860 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 cfqd->cfq_back_max = cfq_back_max;
3862 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003863 cfqd->cfq_slice[0] = cfq_slice_async;
3864 cfqd->cfq_slice[1] = cfq_slice_sync;
Tao Ma5bf14c02012-04-01 14:33:39 -07003865 cfqd->cfq_target_latency = cfq_target_latency;
Jens Axboe22e2c502005-06-27 10:55:12 +02003866 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3867 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003868 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003869 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003870 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003871 /*
3872 * we optimistically start assuming sync ops weren't delayed in last
3873 * second, in order to have larger depth for async operations.
3874 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003875 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02003876 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877}
3878
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879/*
3880 * sysfs parts below -->
3881 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003882static ssize_t
3883cfq_var_show(unsigned int var, char *page)
3884{
3885 return sprintf(page, "%d\n", var);
3886}
3887
3888static ssize_t
3889cfq_var_store(unsigned int *var, const char *page, size_t count)
3890{
3891 char *p = (char *) page;
3892
3893 *var = simple_strtoul(p, &p, 10);
3894 return count;
3895}
3896
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003898static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003900 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901 unsigned int __data = __VAR; \
3902 if (__CONV) \
3903 __data = jiffies_to_msecs(__data); \
3904 return cfq_var_show(__data, (page)); \
3905}
3906SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003907SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3908SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003909SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3910SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003911SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003912SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003913SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3914SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3915SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003916SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07003917SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918#undef SHOW_FUNCTION
3919
3920#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003921static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003923 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924 unsigned int __data; \
3925 int ret = cfq_var_store(&__data, (page), count); \
3926 if (__data < (MIN)) \
3927 __data = (MIN); \
3928 else if (__data > (MAX)) \
3929 __data = (MAX); \
3930 if (__CONV) \
3931 *(__PTR) = msecs_to_jiffies(__data); \
3932 else \
3933 *(__PTR) = __data; \
3934 return ret; \
3935}
3936STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003937STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3938 UINT_MAX, 1);
3939STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3940 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05003941STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003942STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3943 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003944STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003945STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003946STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3947STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01003948STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3949 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003950STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07003951STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952#undef STORE_FUNCTION
3953
Al Viroe572ec72006-03-18 22:27:18 -05003954#define CFQ_ATTR(name) \
3955 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02003956
Al Viroe572ec72006-03-18 22:27:18 -05003957static struct elv_fs_entry cfq_attrs[] = {
3958 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05003959 CFQ_ATTR(fifo_expire_sync),
3960 CFQ_ATTR(fifo_expire_async),
3961 CFQ_ATTR(back_seek_max),
3962 CFQ_ATTR(back_seek_penalty),
3963 CFQ_ATTR(slice_sync),
3964 CFQ_ATTR(slice_async),
3965 CFQ_ATTR(slice_async_rq),
3966 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003967 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02003968 CFQ_ATTR(low_latency),
Tao Ma5bf14c02012-04-01 14:33:39 -07003969 CFQ_ATTR(target_latency),
Al Viroe572ec72006-03-18 22:27:18 -05003970 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971};
3972
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973static struct elevator_type iosched_cfq = {
3974 .ops = {
3975 .elevator_merge_fn = cfq_merge,
3976 .elevator_merged_fn = cfq_merged_request,
3977 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01003978 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07003979 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02003980 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981 .elevator_add_req_fn = cfq_insert_request,
Tatyana Brokhmanb410a822013-07-07 17:33:05 +03003982 .elevator_reinsert_req_fn = cfq_reinsert_request,
3983 .elevator_is_urgent_fn = cfq_urgent_pending,
Jens Axboeb4878f22005-10-20 16:42:29 +02003984 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985 .elevator_deactivate_req_fn = cfq_deactivate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02003987 .elevator_former_req_fn = elv_rb_former_request,
3988 .elevator_latter_req_fn = elv_rb_latter_request,
Tejun Heo9b84cac2011-12-14 00:33:42 +01003989 .elevator_init_icq_fn = cfq_init_icq,
Tejun Heo7e5a8792011-12-14 00:33:42 +01003990 .elevator_exit_icq_fn = cfq_exit_icq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991 .elevator_set_req_fn = cfq_set_request,
3992 .elevator_put_req_fn = cfq_put_request,
3993 .elevator_may_queue_fn = cfq_may_queue,
3994 .elevator_init_fn = cfq_init_queue,
3995 .elevator_exit_fn = cfq_exit_queue,
3996 },
Tejun Heo3d3c2372011-12-14 00:33:42 +01003997 .icq_size = sizeof(struct cfq_io_cq),
3998 .icq_align = __alignof__(struct cfq_io_cq),
Al Viro3d1ab402006-03-18 18:35:43 -05003999 .elevator_attrs = cfq_attrs,
Tejun Heo3d3c2372011-12-14 00:33:42 +01004000 .elevator_name = "cfq",
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001 .elevator_owner = THIS_MODULE,
4002};
4003
Vivek Goyal3e252062009-12-04 10:36:42 -05004004#ifdef CONFIG_CFQ_GROUP_IOSCHED
4005static struct blkio_policy_type blkio_policy_cfq = {
4006 .ops = {
4007 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4008 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4009 },
Vivek Goyal062a6442010-09-15 17:06:33 -04004010 .plid = BLKIO_POLICY_PROP,
Vivek Goyal3e252062009-12-04 10:36:42 -05004011};
4012#else
4013static struct blkio_policy_type blkio_policy_cfq;
4014#endif
4015
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016static int __init cfq_init(void)
4017{
Tejun Heo3d3c2372011-12-14 00:33:42 +01004018 int ret;
4019
Jens Axboe22e2c502005-06-27 10:55:12 +02004020 /*
4021 * could be 0 on HZ < 1000 setups
4022 */
4023 if (!cfq_slice_async)
4024 cfq_slice_async = 1;
4025 if (!cfq_slice_idle)
4026 cfq_slice_idle = 1;
4027
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004028#ifdef CONFIG_CFQ_GROUP_IOSCHED
4029 if (!cfq_group_idle)
4030 cfq_group_idle = 1;
4031#else
4032 cfq_group_idle = 0;
4033#endif
Tejun Heo3d3c2372011-12-14 00:33:42 +01004034 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4035 if (!cfq_pool)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036 return -ENOMEM;
4037
Tejun Heo3d3c2372011-12-14 00:33:42 +01004038 ret = elv_register(&iosched_cfq);
4039 if (ret) {
4040 kmem_cache_destroy(cfq_pool);
4041 return ret;
4042 }
Tejun Heo3d3c2372011-12-14 00:33:42 +01004043
Vivek Goyal3e252062009-12-04 10:36:42 -05004044 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004045
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004046 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047}
4048
4049static void __exit cfq_exit(void)
4050{
Vivek Goyal3e252062009-12-04 10:36:42 -05004051 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052 elv_unregister(&iosched_cfq);
Tejun Heo3d3c2372011-12-14 00:33:42 +01004053 kmem_cache_destroy(cfq_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054}
4055
4056module_init(cfq_init);
4057module_exit(cfq_exit);
4058
4059MODULE_AUTHOR("Jens Axboe");
4060MODULE_LICENSE("GPL");
4061MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");