blob: a33bd4377c615e789509431c5f8a7583939149a4 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * CFQ, or complete fairness queueing, disk scheduler.
3 *
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6 *
Jens Axboe0fe23472006-09-04 15:41:16 +02007 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090010#include <linux/slab.h>
Al Viro1cc9be62006-03-18 12:29:52 -050011#include <linux/blkdev.h>
12#include <linux/elevator.h>
Randy Dunlapad5ebd22009-11-11 13:47:45 +010013#include <linux/jiffies.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020015#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020016#include <linux/blktrace_api.h>
Vivek Goyale98ef892010-06-18 10:39:47 -040017#include "cfq.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070018
19/*
20 * tunables
21 */
Jens Axboefe094d92008-01-31 13:08:54 +010022/* max queue in one round of service */
Shaohua Liabc3c742010-03-01 09:20:54 +010023static const int cfq_quantum = 8;
Arjan van de Ven64100092006-01-06 09:46:02 +010024static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010025/* maximum backwards seek, in KiB */
26static const int cfq_back_max = 16 * 1024;
27/* penalty of a backwards seek */
28static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010029static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020030static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010031static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020032static int cfq_slice_idle = HZ / 125;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +020033static int cfq_group_idle = HZ / 125;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +010034static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35static const int cfq_hist_divisor = 4;
Jens Axboe22e2c502005-06-27 10:55:12 +020036
Jens Axboed9e76202007-04-20 14:27:50 +020037/*
Jens Axboe08717142008-01-28 11:38:15 +010038 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020039 */
Jens Axboe08717142008-01-28 11:38:15 +010040#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020041
42/*
43 * below this threshold, we consider thinktime immediate
44 */
45#define CFQ_MIN_TT (2)
46
Jens Axboe22e2c502005-06-27 10:55:12 +020047#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020048#define CFQ_HW_QUEUE_MIN (5)
Vivek Goyal25bc6b02009-12-03 12:59:43 -050049#define CFQ_SERVICE_SHIFT 12
Jens Axboe22e2c502005-06-27 10:55:12 +020050
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010051#define CFQQ_SEEK_THR (sector_t)(8 * 100)
Shaohua Lie9ce3352010-03-19 08:03:04 +010052#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
Corrado Zoccolo41647e72010-02-27 19:45:40 +010053#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010054#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
Shaohua Liae54abe2010-02-05 13:11:45 +010055
Jens Axboefe094d92008-01-31 13:08:54 +010056#define RQ_CIC(rq) \
Mike Snitzerc1867942011-02-11 11:08:00 +010057 ((struct cfq_io_context *) (rq)->elevator_private[0])
58#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1])
59#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2])
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Christoph Lametere18b8902006-12-06 20:33:20 -080061static struct kmem_cache *cfq_pool;
62static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
Tejun Heo245b2e72009-06-24 15:13:48 +090064static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050065static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020066static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050067
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +040068static DEFINE_SPINLOCK(cic_index_lock);
69static DEFINE_IDA(cic_index_ida);
70
Jens Axboe22e2c502005-06-27 10:55:12 +020071#define CFQ_PRIO_LISTS IOPRIO_BE_NR
72#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020073#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
Jens Axboe206dc692006-03-28 13:03:44 +020075#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050076#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020077
Jens Axboe22e2c502005-06-27 10:55:12 +020078/*
Jens Axboecc09e292007-04-26 12:53:50 +020079 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
83 */
84struct cfq_rb_root {
85 struct rb_root rb;
86 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010087 unsigned count;
Richard Kennedy73e9ffd2010-03-01 10:50:20 +010088 unsigned total_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050089 u64 min_vdisktime;
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
Jens Axboeb53d1ed2011-08-19 08:34:48 +0200133 /* pending metadata requests */
134 int meta_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
219/*
220 * Per block device queue structure
221 */
222struct cfq_data {
223 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500224 /* Root service tree for cfq_groups */
225 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500226 struct cfq_group root_group;
227
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100228 /*
229 * The priority currently being served
230 */
231 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100232 enum wl_type_t serving_type;
233 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500234 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200235
236 /*
237 * Each priority tree is sorted by next_request position. These
238 * trees are used when determining if two or more queues are
239 * interleaving requests (see cfq_close_cooperator).
240 */
241 struct rb_root prio_trees[CFQ_PRIO_LISTS];
242
Jens Axboe22e2c502005-06-27 10:55:12 +0200243 unsigned int busy_queues;
Shaohua Lief8a41d2011-03-07 09:26:29 +0100244 unsigned int busy_sync_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +0200245
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100246 int rq_in_driver;
247 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200248
249 /*
250 * queue-depth detection
251 */
252 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200253 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100254 /*
255 * hw_tag can be
256 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
257 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
258 * 0 => no NCQ
259 */
260 int hw_tag_est_depth;
261 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200262
263 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200264 * idle window management
265 */
266 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200267 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200268
269 struct cfq_queue *active_queue;
270 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200271
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200272 /*
273 * async queue for each priority case
274 */
275 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
276 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200277
Jens Axboe6d048f52007-04-25 12:44:27 +0200278 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 /*
281 * tunables, see top of file
282 */
283 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200284 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285 unsigned int cfq_back_penalty;
286 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200287 unsigned int cfq_slice[2];
288 unsigned int cfq_slice_async_rq;
289 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200290 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200291 unsigned int cfq_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500292
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400293 unsigned int cic_index;
Al Virod9ff4182006-03-18 13:51:22 -0500294 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295
Jens Axboe6118b702009-06-30 09:34:12 +0200296 /*
297 * Fallback dummy cfqq for extreme OOM conditions
298 */
299 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200300
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100301 unsigned long last_delayed_sync;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500302
303 /* List of cfq groups being managed on this device*/
304 struct hlist_head cfqg_list;
Vivek Goyal56edf7d2011-05-19 15:38:22 -0400305
306 /* Number of groups which are on blkcg->blkg_list */
307 unsigned int nr_blkcg_linked_grps;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308};
309
Vivek Goyal25fb5162009-12-03 12:59:46 -0500310static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
311
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500312static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
313 enum wl_prio_t prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500314 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100315{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500316 if (!cfqg)
317 return NULL;
318
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100319 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500320 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100321
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500322 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100323}
324
Jens Axboe3b181522005-06-27 10:56:24 +0200325enum cfqq_state_flags {
Jens Axboeb0b8d742007-01-19 11:35:30 +1100326 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
327 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200328 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100329 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d742007-01-19 11:35:30 +1100330 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
331 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
332 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100333 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200334 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400335 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100336 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100337 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500338 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200339};
340
341#define CFQ_CFQQ_FNS(name) \
342static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
343{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100344 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200345} \
346static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
347{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100348 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200349} \
350static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
351{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100352 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200353}
354
355CFQ_CFQQ_FNS(on_rr);
356CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200357CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200358CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200359CFQ_CFQQ_FNS(fifo_expire);
360CFQ_CFQQ_FNS(idle_window);
361CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100362CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200363CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200364CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100365CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100366CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500367CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200368#undef CFQ_CFQQ_FNS
369
Vivek Goyalafc24d42010-04-26 19:27:56 +0200370#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyal2868ef72009-12-03 12:59:48 -0500371#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
372 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
373 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200374 blkg_path(&(cfqq)->cfqg->blkg), ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500375
376#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
377 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200378 blkg_path(&(cfqg)->blkg), ##args) \
Vivek Goyal2868ef72009-12-03 12:59:48 -0500379
380#else
Jens Axboe7b679132008-05-30 12:23:07 +0200381#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
382 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200383#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500384#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200385#define cfq_log(cfqd, fmt, args...) \
386 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
387
Vivek Goyal615f0252009-12-03 12:59:39 -0500388/* Traverses through cfq group service trees */
389#define for_each_cfqg_st(cfqg, i, j, st) \
390 for (i = 0; i <= IDLE_WORKLOAD; i++) \
391 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
392 : &cfqg->service_tree_idle; \
393 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
394 (i == IDLE_WORKLOAD && j == 0); \
395 j++, st = i < IDLE_WORKLOAD ? \
396 &cfqg->service_trees[i][j]: NULL) \
397
Shaohua Lif5f2b6c2011-07-12 14:24:55 +0200398static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
399 struct cfq_ttime *ttime, bool group_idle)
400{
401 unsigned long slice;
402 if (!sample_valid(ttime->ttime_samples))
403 return false;
404 if (group_idle)
405 slice = cfqd->cfq_group_idle;
406 else
407 slice = cfqd->cfq_slice_idle;
408 return ttime->ttime_mean > slice;
409}
Vivek Goyal615f0252009-12-03 12:59:39 -0500410
Vivek Goyal02b35082010-08-23 12:23:53 +0200411static inline bool iops_mode(struct cfq_data *cfqd)
412{
413 /*
414 * If we are not idling on queues and it is a NCQ drive, parallel
415 * execution of requests is on and measuring time is not possible
416 * in most of the cases until and unless we drive shallower queue
417 * depths and that becomes a performance bottleneck. In such cases
418 * switch to start providing fairness in terms of number of IOs.
419 */
420 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
421 return true;
422 else
423 return false;
424}
425
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100426static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
427{
428 if (cfq_class_idle(cfqq))
429 return IDLE_WORKLOAD;
430 if (cfq_class_rt(cfqq))
431 return RT_WORKLOAD;
432 return BE_WORKLOAD;
433}
434
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100435
436static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
437{
438 if (!cfq_cfqq_sync(cfqq))
439 return ASYNC_WORKLOAD;
440 if (!cfq_cfqq_idle_window(cfqq))
441 return SYNC_NOIDLE_WORKLOAD;
442 return SYNC_WORKLOAD;
443}
444
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500445static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
446 struct cfq_data *cfqd,
447 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100448{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500449 if (wl == IDLE_WORKLOAD)
450 return cfqg->service_tree_idle.count;
451
452 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
453 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
454 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100455}
456
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500457static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
458 struct cfq_group *cfqg)
459{
460 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
461 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
462}
463
Jens Axboe165125e2007-07-24 09:28:11 +0200464static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200465static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100466 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100467static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200468 struct io_context *);
469
470static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200471 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200472{
Jens Axboea6151c32009-10-07 20:02:57 +0200473 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200474}
475
476static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200477 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200478{
Jens Axboea6151c32009-10-07 20:02:57 +0200479 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200480}
481
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400482#define CIC_DEAD_KEY 1ul
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400483#define CIC_DEAD_INDEX_SHIFT 1
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400484
485static inline void *cfqd_dead_key(struct cfq_data *cfqd)
486{
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +0400487 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400488}
489
490static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
491{
492 struct cfq_data *cfqd = cic->key;
493
494 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
495 return NULL;
496
497 return cfqd;
498}
499
Vasily Tarasov91fac312007-04-25 12:29:51 +0200500/*
501 * We regard a request as SYNC, if it's either a read or has the SYNC bit
502 * set (in which case it could also be direct WRITE).
503 */
Jens Axboea6151c32009-10-07 20:02:57 +0200504static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200505{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200506 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200507}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700510 * scheduler run of queue, if there are requests pending and no one in the
511 * driver that will restart queueing
512 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200513static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700514{
Jens Axboe7b679132008-05-30 12:23:07 +0200515 if (cfqd->busy_queues) {
516 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200517 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200518 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700519}
520
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100522 * Scale schedule slice based on io priority. Use the sync time slice only
523 * if a queue is marked sync and has sync io queued. A sync queue with async
524 * io only, should not get full sync slice length.
525 */
Jens Axboea6151c32009-10-07 20:02:57 +0200526static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200527 unsigned short prio)
528{
529 const int base_slice = cfqd->cfq_slice[sync];
530
531 WARN_ON(prio >= IOPRIO_BE_NR);
532
533 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
534}
535
Jens Axboe44f7c162007-01-19 11:51:58 +1100536static inline int
537cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
538{
Jens Axboed9e76202007-04-20 14:27:50 +0200539 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100540}
541
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500542static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
543{
544 u64 d = delta << CFQ_SERVICE_SHIFT;
545
546 d = d * BLKIO_WEIGHT_DEFAULT;
547 do_div(d, cfqg->weight);
548 return d;
549}
550
551static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
552{
553 s64 delta = (s64)(vdisktime - min_vdisktime);
554 if (delta > 0)
555 min_vdisktime = vdisktime;
556
557 return min_vdisktime;
558}
559
560static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
561{
562 s64 delta = (s64)(vdisktime - min_vdisktime);
563 if (delta < 0)
564 min_vdisktime = vdisktime;
565
566 return min_vdisktime;
567}
568
569static void update_min_vdisktime(struct cfq_rb_root *st)
570{
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500571 struct cfq_group *cfqg;
572
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500573 if (st->left) {
574 cfqg = rb_entry_cfqg(st->left);
Gui Jianfenga6032712011-03-07 09:28:09 +0100575 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
576 cfqg->vdisktime);
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500577 }
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500578}
579
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100580/*
581 * get averaged number of queues of RT/BE priority.
582 * average is updated, with a formula that gives more weight to higher numbers,
583 * to quickly follows sudden increases and decrease slowly
584 */
585
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500586static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
587 struct cfq_group *cfqg, bool rt)
Jens Axboe58696192009-10-28 09:27:07 +0100588{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100589 unsigned min_q, max_q;
590 unsigned mult = cfq_hist_divisor - 1;
591 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500592 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100593
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500594 min_q = min(cfqg->busy_queues_avg[rt], busy);
595 max_q = max(cfqg->busy_queues_avg[rt], busy);
596 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100597 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500598 return cfqg->busy_queues_avg[rt];
599}
600
601static inline unsigned
602cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
603{
604 struct cfq_rb_root *st = &cfqd->grp_service_tree;
605
606 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100607}
608
Shaohua Lic553f8e2011-01-14 08:41:03 +0100609static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -0700610cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100611{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100612 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
613 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500614 /*
615 * interested queues (we consider only the ones with the same
616 * priority class in the cfq group)
617 */
618 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
619 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100620 unsigned sync_slice = cfqd->cfq_slice[1];
621 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500622 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
623
624 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100625 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
626 /* scale low_slice according to IO priority
627 * and sync vs async */
628 unsigned low_slice =
629 min(slice, base_low_slice * slice / sync_slice);
630 /* the adapted slice value is scaled to fit all iqs
631 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500632 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100633 low_slice);
634 }
635 }
Shaohua Lic553f8e2011-01-14 08:41:03 +0100636 return slice;
637}
638
639static inline void
640cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
641{
Vivek Goyalba5bd522011-01-19 08:25:02 -0700642 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +0100643
Vivek Goyaldae739e2009-12-03 12:59:45 -0500644 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100645 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500646 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200647 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100648}
649
650/*
651 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
652 * isn't valid until the first request from the dispatch is activated
653 * and the slice time set.
654 */
Jens Axboea6151c32009-10-07 20:02:57 +0200655static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100656{
657 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +0100658 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100659 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +0100660 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +1100661
Shaohua Lic1e44752010-11-08 15:01:02 +0100662 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +1100663}
664
665/*
Jens Axboe5e705372006-07-13 12:39:25 +0200666 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200668 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 */
Jens Axboe5e705372006-07-13 12:39:25 +0200670static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100671cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100673 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200675#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
676#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
677 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678
Jens Axboe5e705372006-07-13 12:39:25 +0200679 if (rq1 == NULL || rq1 == rq2)
680 return rq2;
681 if (rq2 == NULL)
682 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200683
Namhyung Kim229836b2011-05-24 10:23:21 +0200684 if (rq_is_sync(rq1) != rq_is_sync(rq2))
685 return rq_is_sync(rq1) ? rq1 : rq2;
686
Jens Axboeb53d1ed2011-08-19 08:34:48 +0200687 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_META)
688 return rq1->cmd_flags & REQ_META ? rq1 : rq2;
689
Tejun Heo83096eb2009-05-07 22:24:39 +0900690 s1 = blk_rq_pos(rq1);
691 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 /*
694 * by definition, 1KiB is 2 sectors
695 */
696 back_max = cfqd->cfq_back_max * 2;
697
698 /*
699 * Strict one way elevator _except_ in the case where we allow
700 * short backward seeks which are biased as twice the cost of a
701 * similar forward seek.
702 */
703 if (s1 >= last)
704 d1 = s1 - last;
705 else if (s1 + back_max >= last)
706 d1 = (last - s1) * cfqd->cfq_back_penalty;
707 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200708 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
710 if (s2 >= last)
711 d2 = s2 - last;
712 else if (s2 + back_max >= last)
713 d2 = (last - s2) * cfqd->cfq_back_penalty;
714 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200715 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716
717 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718
Andreas Mohre8a99052006-03-28 08:59:49 +0200719 /*
720 * By doing switch() on the bit mask "wrap" we avoid having to
721 * check two variables for all permutations: --> faster!
722 */
723 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200724 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200725 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200726 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200727 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200728 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200729 else {
730 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200731 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200732 else
Jens Axboe5e705372006-07-13 12:39:25 +0200733 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200734 }
735
736 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200737 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200738 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200739 return rq2;
740 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200741 default:
742 /*
743 * Since both rqs are wrapped,
744 * start with the one that's further behind head
745 * (--> only *one* back seek required),
746 * since back seek takes more time than forward.
747 */
748 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200749 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 else
Jens Axboe5e705372006-07-13 12:39:25 +0200751 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 }
753}
754
Jens Axboe498d3aa2007-04-26 12:54:48 +0200755/*
756 * The below is leftmost cache rbtree addon
757 */
Jens Axboe08717142008-01-28 11:38:15 +0100758static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200759{
Vivek Goyal615f0252009-12-03 12:59:39 -0500760 /* Service tree is empty */
761 if (!root->count)
762 return NULL;
763
Jens Axboecc09e292007-04-26 12:53:50 +0200764 if (!root->left)
765 root->left = rb_first(&root->rb);
766
Jens Axboe08717142008-01-28 11:38:15 +0100767 if (root->left)
768 return rb_entry(root->left, struct cfq_queue, rb_node);
769
770 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200771}
772
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500773static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
774{
775 if (!root->left)
776 root->left = rb_first(&root->rb);
777
778 if (root->left)
779 return rb_entry_cfqg(root->left);
780
781 return NULL;
782}
783
Jens Axboea36e71f2009-04-15 12:15:11 +0200784static void rb_erase_init(struct rb_node *n, struct rb_root *root)
785{
786 rb_erase(n, root);
787 RB_CLEAR_NODE(n);
788}
789
Jens Axboecc09e292007-04-26 12:53:50 +0200790static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
791{
792 if (root->left == n)
793 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200794 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100795 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200796}
797
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798/*
799 * would be nice to take fifo expire time into account as well
800 */
Jens Axboe5e705372006-07-13 12:39:25 +0200801static struct request *
802cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
803 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804{
Jens Axboe21183b02006-07-13 12:33:14 +0200805 struct rb_node *rbnext = rb_next(&last->rb_node);
806 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200807 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808
Jens Axboe21183b02006-07-13 12:33:14 +0200809 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810
811 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200812 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200813
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200815 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200816 else {
817 rbnext = rb_first(&cfqq->sort_list);
818 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200819 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200820 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100822 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823}
824
Jens Axboed9e76202007-04-20 14:27:50 +0200825static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
826 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827{
Jens Axboed9e76202007-04-20 14:27:50 +0200828 /*
829 * just an approximation, should be ok.
830 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500831 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100832 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200833}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500835static inline s64
836cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
837{
838 return cfqg->vdisktime - st->min_vdisktime;
839}
840
841static void
842__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
843{
844 struct rb_node **node = &st->rb.rb_node;
845 struct rb_node *parent = NULL;
846 struct cfq_group *__cfqg;
847 s64 key = cfqg_key(st, cfqg);
848 int left = 1;
849
850 while (*node != NULL) {
851 parent = *node;
852 __cfqg = rb_entry_cfqg(parent);
853
854 if (key < cfqg_key(st, __cfqg))
855 node = &parent->rb_left;
856 else {
857 node = &parent->rb_right;
858 left = 0;
859 }
860 }
861
862 if (left)
863 st->left = &cfqg->rb_node;
864
865 rb_link_node(&cfqg->rb_node, parent, node);
866 rb_insert_color(&cfqg->rb_node, &st->rb);
867}
868
869static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100870cfq_update_group_weight(struct cfq_group *cfqg)
871{
872 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
873 if (cfqg->needs_update) {
874 cfqg->weight = cfqg->new_weight;
875 cfqg->needs_update = false;
876 }
877}
878
879static void
880cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
881{
882 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
883
884 cfq_update_group_weight(cfqg);
885 __cfq_group_service_tree_add(st, cfqg);
886 st->total_weight += cfqg->weight;
887}
888
889static void
890cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500891{
892 struct cfq_rb_root *st = &cfqd->grp_service_tree;
893 struct cfq_group *__cfqg;
894 struct rb_node *n;
895
896 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +0100897 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500898 return;
899
900 /*
901 * Currently put the group at the end. Later implement something
902 * so that groups get lesser vtime based on their weights, so that
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300903 * if group does not loose all if it was not continuously backlogged.
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500904 */
905 n = rb_last(&st->rb);
906 if (n) {
907 __cfqg = rb_entry_cfqg(n);
908 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
909 } else
910 cfqg->vdisktime = st->min_vdisktime;
Justin TerAvest8184f932011-03-17 16:12:36 +0100911 cfq_group_service_tree_add(st, cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500912}
913
914static void
Justin TerAvest8184f932011-03-17 16:12:36 +0100915cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
916{
917 st->total_weight -= cfqg->weight;
918 if (!RB_EMPTY_NODE(&cfqg->rb_node))
919 cfq_rb_erase(&cfqg->rb_node, st);
920}
921
922static void
923cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500924{
925 struct cfq_rb_root *st = &cfqd->grp_service_tree;
926
927 BUG_ON(cfqg->nr_cfqq < 1);
928 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500929
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500930 /* If there are other cfq queues under this group, don't delete it */
931 if (cfqg->nr_cfqq)
932 return;
933
Vivek Goyal2868ef72009-12-03 12:59:48 -0500934 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Justin TerAvest8184f932011-03-17 16:12:36 +0100935 cfq_group_service_tree_del(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500936 cfqg->saved_workload_slice = 0;
Vivek Goyale98ef892010-06-18 10:39:47 -0400937 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500938}
939
Justin TerAvest167400d2011-03-12 16:54:00 +0100940static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
941 unsigned int *unaccounted_time)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500942{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500943 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500944
945 /*
946 * Queue got expired before even a single request completed or
947 * got expired immediately after first request completion.
948 */
949 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
950 /*
951 * Also charge the seek time incurred to the group, otherwise
952 * if there are mutiple queues in the group, each can dispatch
953 * a single request on seeky media and cause lots of seek time
954 * and group will never know it.
955 */
956 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
957 1);
958 } else {
959 slice_used = jiffies - cfqq->slice_start;
Justin TerAvest167400d2011-03-12 16:54:00 +0100960 if (slice_used > cfqq->allocated_slice) {
961 *unaccounted_time = slice_used - cfqq->allocated_slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500962 slice_used = cfqq->allocated_slice;
Justin TerAvest167400d2011-03-12 16:54:00 +0100963 }
964 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
965 *unaccounted_time += cfqq->slice_start -
966 cfqq->dispatch_start;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500967 }
968
Vivek Goyaldae739e2009-12-03 12:59:45 -0500969 return slice_used;
970}
971
972static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +0200973 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -0500974{
975 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Justin TerAvest167400d2011-03-12 16:54:00 +0100976 unsigned int used_sl, charge, unaccounted_sl = 0;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500977 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
978 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500979
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500980 BUG_ON(nr_sync < 0);
Justin TerAvest167400d2011-03-12 16:54:00 +0100981 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500982
Vivek Goyal02b35082010-08-23 12:23:53 +0200983 if (iops_mode(cfqd))
984 charge = cfqq->slice_dispatch;
985 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
986 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500987
988 /* Can't update vdisktime while group is on service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +0100989 cfq_group_service_tree_del(st, cfqg);
Vivek Goyal02b35082010-08-23 12:23:53 +0200990 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
Justin TerAvest8184f932011-03-17 16:12:36 +0100991 /* If a new weight was requested, update now, off tree */
992 cfq_group_service_tree_add(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500993
994 /* This group is being expired. Save the context */
995 if (time_after(cfqd->workload_expires, jiffies)) {
996 cfqg->saved_workload_slice = cfqd->workload_expires
997 - jiffies;
998 cfqg->saved_workload = cfqd->serving_type;
999 cfqg->saved_serving_prio = cfqd->serving_prio;
1000 } else
1001 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -05001002
1003 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1004 st->min_vdisktime);
Joe Perchesfd16d262011-06-13 10:42:49 +02001005 cfq_log_cfqq(cfqq->cfqd, cfqq,
1006 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1007 used_sl, cfqq->slice_dispatch, charge,
1008 iops_mode(cfqd), cfqq->nr_sectors);
Justin TerAvest167400d2011-03-12 16:54:00 +01001009 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1010 unaccounted_sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04001011 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001012}
1013
Vivek Goyal25fb5162009-12-03 12:59:46 -05001014#ifdef CONFIG_CFQ_GROUP_IOSCHED
1015static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1016{
1017 if (blkg)
1018 return container_of(blkg, struct cfq_group, blkg);
1019 return NULL;
1020}
1021
Paul Bolle8aea4542011-06-06 05:07:54 +02001022static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1023 unsigned int weight)
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001024{
Justin TerAvest8184f932011-03-17 16:12:36 +01001025 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1026 cfqg->new_weight = weight;
1027 cfqg->needs_update = true;
Vivek Goyalf8d461d2009-12-03 12:59:52 -05001028}
1029
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001030static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1031 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001032{
Vivek Goyal22084192009-12-03 12:59:49 -05001033 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1034 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001035
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001036 /*
1037 * Add group onto cgroup list. It might happen that bdi->dev is
1038 * not initialized yet. Initialize this new group without major
1039 * and minor info and this info will be filled in once a new thread
1040 * comes for IO.
1041 */
1042 if (bdi->dev) {
Ricky Beniteza74b2ad2010-04-05 18:22:17 +02001043 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001044 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1045 (void *)cfqd, MKDEV(major, minor));
1046 } else
1047 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1048 (void *)cfqd, 0);
1049
1050 cfqd->nr_blkcg_linked_grps++;
1051 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1052
1053 /* Add group on cfqd list */
1054 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1055}
1056
1057/*
1058 * Should be called from sleepable context. No request queue lock as per
1059 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1060 * from sleepable context.
1061 */
1062static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1063{
1064 struct cfq_group *cfqg = NULL;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001065 int i, j, ret;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001066 struct cfq_rb_root *st;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001067
1068 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1069 if (!cfqg)
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001070 return NULL;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001071
Vivek Goyal25fb5162009-12-03 12:59:46 -05001072 for_each_cfqg_st(cfqg, i, j, st)
1073 *st = CFQ_RB_ROOT;
1074 RB_CLEAR_NODE(&cfqg->rb_node);
1075
Shaohua Li7700fc42011-07-12 14:24:56 +02001076 cfqg->ttime.last_end_request = jiffies;
1077
Vivek Goyalb1c35762009-12-03 12:59:47 -05001078 /*
1079 * Take the initial reference that will be released on destroy
1080 * This can be thought of a joint reference by cgroup and
1081 * elevator which will be dropped by either elevator exit
1082 * or cgroup deletion path depending on who is exiting first.
1083 */
Shaohua Li329a6782011-01-07 08:48:28 +01001084 cfqg->ref = 1;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001085
1086 ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1087 if (ret) {
1088 kfree(cfqg);
1089 return NULL;
1090 }
1091
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001092 return cfqg;
1093}
1094
1095static struct cfq_group *
1096cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1097{
1098 struct cfq_group *cfqg = NULL;
1099 void *key = cfqd;
1100 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1101 unsigned int major, minor;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001102
Vivek Goyal180be2a2010-09-14 08:47:11 +02001103 /*
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001104 * This is the common case when there are no blkio cgroups.
1105 * Avoid lookup in this case
Vivek Goyal180be2a2010-09-14 08:47:11 +02001106 */
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001107 if (blkcg == &blkio_root_cgroup)
1108 cfqg = &cfqd->root_group;
1109 else
1110 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1111
1112 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
Vivek Goyal180be2a2010-09-14 08:47:11 +02001113 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001114 cfqg->blkg.dev = MKDEV(major, minor);
1115 }
Vivek Goyal180be2a2010-09-14 08:47:11 +02001116
Vivek Goyal25fb5162009-12-03 12:59:46 -05001117 return cfqg;
1118}
1119
1120/*
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001121 * Search for the cfq group current task belongs to. request_queue lock must
1122 * be held.
Vivek Goyal25fb5162009-12-03 12:59:46 -05001123 */
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001124static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001125{
Vivek Goyal70087dc2011-05-16 15:24:08 +02001126 struct blkio_cgroup *blkcg;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001127 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1128 struct request_queue *q = cfqd->queue;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001129
1130 rcu_read_lock();
Vivek Goyal70087dc2011-05-16 15:24:08 +02001131 blkcg = task_blkio_cgroup(current);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001132 cfqg = cfq_find_cfqg(cfqd, blkcg);
1133 if (cfqg) {
1134 rcu_read_unlock();
1135 return cfqg;
1136 }
1137
1138 /*
1139 * Need to allocate a group. Allocation of group also needs allocation
1140 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1141 * we need to drop rcu lock and queue_lock before we call alloc.
1142 *
1143 * Not taking any queue reference here and assuming that queue is
1144 * around by the time we return. CFQ queue allocation code does
1145 * the same. It might be racy though.
1146 */
1147
1148 rcu_read_unlock();
1149 spin_unlock_irq(q->queue_lock);
1150
1151 cfqg = cfq_alloc_cfqg(cfqd);
1152
1153 spin_lock_irq(q->queue_lock);
1154
1155 rcu_read_lock();
1156 blkcg = task_blkio_cgroup(current);
1157
1158 /*
1159 * If some other thread already allocated the group while we were
1160 * not holding queue lock, free up the group
1161 */
1162 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1163
1164 if (__cfqg) {
1165 kfree(cfqg);
1166 rcu_read_unlock();
1167 return __cfqg;
1168 }
1169
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001170 if (!cfqg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001171 cfqg = &cfqd->root_group;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001172
1173 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001174 rcu_read_unlock();
1175 return cfqg;
1176}
1177
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001178static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1179{
Shaohua Li329a6782011-01-07 08:48:28 +01001180 cfqg->ref++;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001181 return cfqg;
1182}
1183
Vivek Goyal25fb5162009-12-03 12:59:46 -05001184static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1185{
1186 /* Currently, all async queues are mapped to root group */
1187 if (!cfq_cfqq_sync(cfqq))
1188 cfqg = &cfqq->cfqd->root_group;
1189
1190 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001191 /* cfqq reference on cfqg */
Shaohua Li329a6782011-01-07 08:48:28 +01001192 cfqq->cfqg->ref++;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001193}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001194
1195static void cfq_put_cfqg(struct cfq_group *cfqg)
1196{
1197 struct cfq_rb_root *st;
1198 int i, j;
1199
Shaohua Li329a6782011-01-07 08:48:28 +01001200 BUG_ON(cfqg->ref <= 0);
1201 cfqg->ref--;
1202 if (cfqg->ref)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001203 return;
1204 for_each_cfqg_st(cfqg, i, j, st)
Gui Jianfengb54ce602010-11-30 20:52:46 +01001205 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
Vivek Goyal5624a4e2011-05-19 15:38:28 -04001206 free_percpu(cfqg->blkg.stats_cpu);
Vivek Goyalb1c35762009-12-03 12:59:47 -05001207 kfree(cfqg);
1208}
1209
1210static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1211{
1212 /* Something wrong if we are trying to remove same group twice */
1213 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1214
1215 hlist_del_init(&cfqg->cfqd_node);
1216
Vivek Goyala5395b82011-08-02 09:24:09 +02001217 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1218 cfqd->nr_blkcg_linked_grps--;
1219
Vivek Goyalb1c35762009-12-03 12:59:47 -05001220 /*
1221 * Put the reference taken at the time of creation so that when all
1222 * queues are gone, group can be destroyed.
1223 */
1224 cfq_put_cfqg(cfqg);
1225}
1226
1227static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1228{
1229 struct hlist_node *pos, *n;
1230 struct cfq_group *cfqg;
1231
1232 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1233 /*
1234 * If cgroup removal path got to blk_group first and removed
1235 * it from cgroup list, then it will take care of destroying
1236 * cfqg also.
1237 */
Vivek Goyale98ef892010-06-18 10:39:47 -04001238 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
Vivek Goyalb1c35762009-12-03 12:59:47 -05001239 cfq_destroy_cfqg(cfqd, cfqg);
1240 }
1241}
1242
1243/*
1244 * Blk cgroup controller notification saying that blkio_group object is being
1245 * delinked as associated cgroup object is going away. That also means that
1246 * no new IO will come in this group. So get rid of this group as soon as
1247 * any pending IO in the group is finished.
1248 *
1249 * This function is called under rcu_read_lock(). key is the rcu protected
1250 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1251 * read lock.
1252 *
1253 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1254 * it should not be NULL as even if elevator was exiting, cgroup deltion
1255 * path got to it first.
1256 */
Paul Bolle8aea4542011-06-06 05:07:54 +02001257static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
Vivek Goyalb1c35762009-12-03 12:59:47 -05001258{
1259 unsigned long flags;
1260 struct cfq_data *cfqd = key;
1261
1262 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1263 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1264 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1265}
1266
Vivek Goyal25fb5162009-12-03 12:59:46 -05001267#else /* GROUP_IOSCHED */
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001268static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001269{
1270 return &cfqd->root_group;
1271}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001272
1273static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1274{
Dmitry Monakhov50eaeb32010-04-28 19:50:33 +02001275 return cfqg;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001276}
1277
Vivek Goyal25fb5162009-12-03 12:59:46 -05001278static inline void
1279cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1280 cfqq->cfqg = cfqg;
1281}
1282
Vivek Goyalb1c35762009-12-03 12:59:47 -05001283static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1284static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1285
Vivek Goyal25fb5162009-12-03 12:59:46 -05001286#endif /* GROUP_IOSCHED */
1287
Jens Axboe498d3aa2007-04-26 12:54:48 +02001288/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001289 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa2007-04-26 12:54:48 +02001290 * requests waiting to be processed. It is sorted in the order that
1291 * we will service the queues.
1292 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001293static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001294 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001295{
Jens Axboe08717142008-01-28 11:38:15 +01001296 struct rb_node **p, *parent;
1297 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001298 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001299 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa2007-04-26 12:54:48 +02001300 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001301 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001302
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001303 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
Vivek Goyal65b32a52009-12-16 17:52:59 -05001304 cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01001305 if (cfq_class_idle(cfqq)) {
1306 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001307 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001308 if (parent && parent != &cfqq->rb_node) {
1309 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1310 rb_key += __cfqq->rb_key;
1311 } else
1312 rb_key += jiffies;
1313 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001314 /*
1315 * Get our rb key offset. Subtract any residual slice
1316 * value carried from last service. A negative resid
1317 * count indicates slice overrun, and this should position
1318 * the next service time further away in the tree.
1319 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001320 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001321 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001322 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001323 } else {
1324 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001325 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001326 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1327 }
Jens Axboed9e76202007-04-20 14:27:50 +02001328
1329 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001330 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001331 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001332 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001333 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001334 if (rb_key == cfqq->rb_key &&
1335 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001336 return;
Jens Axboe53b03742006-07-28 09:48:51 +02001337
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001338 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1339 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001340 }
Jens Axboed9e76202007-04-20 14:27:50 +02001341
Jens Axboe498d3aa2007-04-26 12:54:48 +02001342 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001343 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001344 cfqq->service_tree = service_tree;
1345 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001346 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001347 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001348
Jens Axboed9e76202007-04-20 14:27:50 +02001349 parent = *p;
1350 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1351
Jens Axboe0c534e02007-04-18 20:01:57 +02001352 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001353 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001354 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001355 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001356 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001357 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001358 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001359 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001360 }
Jens Axboe67060e32007-04-18 20:13:32 +02001361
1362 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001363 }
1364
Jens Axboecc09e292007-04-26 12:53:50 +02001365 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001366 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001367
Jens Axboed9e76202007-04-20 14:27:50 +02001368 cfqq->rb_key = rb_key;
1369 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001370 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1371 service_tree->count++;
Namhyung Kim20359f22011-05-24 10:23:22 +02001372 if (add_front || !new_cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001373 return;
Justin TerAvest8184f932011-03-17 16:12:36 +01001374 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375}
1376
Jens Axboea36e71f2009-04-15 12:15:11 +02001377static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001378cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1379 sector_t sector, struct rb_node **ret_parent,
1380 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001381{
Jens Axboea36e71f2009-04-15 12:15:11 +02001382 struct rb_node **p, *parent;
1383 struct cfq_queue *cfqq = NULL;
1384
1385 parent = NULL;
1386 p = &root->rb_node;
1387 while (*p) {
1388 struct rb_node **n;
1389
1390 parent = *p;
1391 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1392
1393 /*
1394 * Sort strictly based on sector. Smallest to the left,
1395 * largest to the right.
1396 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001397 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001398 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001399 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001400 n = &(*p)->rb_left;
1401 else
1402 break;
1403 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001404 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001405 }
1406
1407 *ret_parent = parent;
1408 if (rb_link)
1409 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001410 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001411}
1412
1413static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1414{
Jens Axboea36e71f2009-04-15 12:15:11 +02001415 struct rb_node **p, *parent;
1416 struct cfq_queue *__cfqq;
1417
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001418 if (cfqq->p_root) {
1419 rb_erase(&cfqq->p_node, cfqq->p_root);
1420 cfqq->p_root = NULL;
1421 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001422
1423 if (cfq_class_idle(cfqq))
1424 return;
1425 if (!cfqq->next_rq)
1426 return;
1427
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001428 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001429 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1430 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001431 if (!__cfqq) {
1432 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001433 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1434 } else
1435 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001436}
1437
Jens Axboe498d3aa2007-04-26 12:54:48 +02001438/*
1439 * Update cfqq's position in the service tree.
1440 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001441static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001442{
Jens Axboe6d048f52007-04-25 12:44:27 +02001443 /*
1444 * Resorting requires the cfqq to be on the RR list already.
1445 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001446 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001447 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001448 cfq_prio_tree_add(cfqd, cfqq);
1449 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001450}
1451
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452/*
1453 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001454 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455 */
Jens Axboefebffd62008-01-28 13:19:43 +01001456static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457{
Jens Axboe7b679132008-05-30 12:23:07 +02001458 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001459 BUG_ON(cfq_cfqq_on_rr(cfqq));
1460 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 cfqd->busy_queues++;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001462 if (cfq_cfqq_sync(cfqq))
1463 cfqd->busy_sync_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464
Jens Axboeedd75ff2007-04-19 12:03:34 +02001465 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466}
1467
Jens Axboe498d3aa2007-04-26 12:54:48 +02001468/*
1469 * Called when the cfqq no longer has requests pending, remove it from
1470 * the service tree.
1471 */
Jens Axboefebffd62008-01-28 13:19:43 +01001472static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473{
Jens Axboe7b679132008-05-30 12:23:07 +02001474 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001475 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1476 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001478 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1479 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1480 cfqq->service_tree = NULL;
1481 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001482 if (cfqq->p_root) {
1483 rb_erase(&cfqq->p_node, cfqq->p_root);
1484 cfqq->p_root = NULL;
1485 }
Jens Axboed9e76202007-04-20 14:27:50 +02001486
Justin TerAvest8184f932011-03-17 16:12:36 +01001487 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 BUG_ON(!cfqd->busy_queues);
1489 cfqd->busy_queues--;
Shaohua Lief8a41d2011-03-07 09:26:29 +01001490 if (cfq_cfqq_sync(cfqq))
1491 cfqd->busy_sync_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492}
1493
1494/*
1495 * rb tree support functions
1496 */
Jens Axboefebffd62008-01-28 13:19:43 +01001497static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498{
Jens Axboe5e705372006-07-13 12:39:25 +02001499 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001500 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501
Jens Axboeb4878f22005-10-20 16:42:29 +02001502 BUG_ON(!cfqq->queued[sync]);
1503 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504
Jens Axboe5e705372006-07-13 12:39:25 +02001505 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506
Vivek Goyalf04a6422009-12-03 12:59:40 -05001507 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1508 /*
1509 * Queue will be deleted from service tree when we actually
1510 * expire it later. Right now just remove it from prio tree
1511 * as it is empty.
1512 */
1513 if (cfqq->p_root) {
1514 rb_erase(&cfqq->p_node, cfqq->p_root);
1515 cfqq->p_root = NULL;
1516 }
1517 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518}
1519
Jens Axboe5e705372006-07-13 12:39:25 +02001520static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521{
Jens Axboe5e705372006-07-13 12:39:25 +02001522 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 struct cfq_data *cfqd = cfqq->cfqd;
Jeff Moyer796d5112011-06-02 21:19:05 +02001524 struct request *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525
Jens Axboe5380a102006-07-13 12:37:56 +02001526 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527
Jeff Moyer796d5112011-06-02 21:19:05 +02001528 elv_rb_add(&cfqq->sort_list, rq);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001529
1530 if (!cfq_cfqq_on_rr(cfqq))
1531 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001532
1533 /*
1534 * check if this request is a better next-serve candidate
1535 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001536 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001537 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001538
1539 /*
1540 * adjust priority tree position, if ->next_rq changes
1541 */
1542 if (prev != cfqq->next_rq)
1543 cfq_prio_tree_add(cfqd, cfqq);
1544
Jens Axboe5044eed2007-04-25 11:53:48 +02001545 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546}
1547
Jens Axboefebffd62008-01-28 13:19:43 +01001548static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549{
Jens Axboe5380a102006-07-13 12:37:56 +02001550 elv_rb_del(&cfqq->sort_list, rq);
1551 cfqq->queued[rq_is_sync(rq)]--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001552 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1553 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02001554 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04001555 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001556 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1557 rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558}
1559
Jens Axboe206dc692006-03-28 13:03:44 +02001560static struct request *
1561cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562{
Jens Axboe206dc692006-03-28 13:03:44 +02001563 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001564 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001565 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566
Jens Axboe4ac845a2008-01-24 08:44:49 +01001567 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001568 if (!cic)
1569 return NULL;
1570
1571 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001572 if (cfqq) {
1573 sector_t sector = bio->bi_sector + bio_sectors(bio);
1574
Jens Axboe21183b02006-07-13 12:33:14 +02001575 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001576 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 return NULL;
1579}
1580
Jens Axboe165125e2007-07-24 09:28:11 +02001581static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001582{
1583 struct cfq_data *cfqd = q->elevator->elevator_data;
1584
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001585 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02001586 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001587 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02001588
Tejun Heo5b936292009-05-07 22:24:38 +09001589 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001590}
1591
Jens Axboe165125e2007-07-24 09:28:11 +02001592static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593{
Jens Axboe22e2c502005-06-27 10:55:12 +02001594 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001596 WARN_ON(!cfqd->rq_in_driver);
1597 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02001598 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01001599 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600}
1601
Jens Axboeb4878f22005-10-20 16:42:29 +02001602static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603{
Jens Axboe5e705372006-07-13 12:39:25 +02001604 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001605
Jens Axboe5e705372006-07-13 12:39:25 +02001606 if (cfqq->next_rq == rq)
1607 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608
Jens Axboeb4878f22005-10-20 16:42:29 +02001609 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001610 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001611
Aaron Carroll45333d52008-08-26 15:52:36 +02001612 cfqq->cfqd->rq_queued--;
Vivek Goyale98ef892010-06-18 10:39:47 -04001613 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1614 rq_data_dir(rq), rq_is_sync(rq));
Jens Axboeb53d1ed2011-08-19 08:34:48 +02001615 if (rq->cmd_flags & REQ_META) {
1616 WARN_ON(!cfqq->meta_pending);
1617 cfqq->meta_pending--;
1618 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619}
1620
Jens Axboe165125e2007-07-24 09:28:11 +02001621static int cfq_merge(struct request_queue *q, struct request **req,
1622 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623{
1624 struct cfq_data *cfqd = q->elevator->elevator_data;
1625 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626
Jens Axboe206dc692006-03-28 13:03:44 +02001627 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001628 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001629 *req = __rq;
1630 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 }
1632
1633 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634}
1635
Jens Axboe165125e2007-07-24 09:28:11 +02001636static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001637 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638{
Jens Axboe21183b02006-07-13 12:33:14 +02001639 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001640 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641
Jens Axboe5e705372006-07-13 12:39:25 +02001642 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644}
1645
Divyesh Shah812d4022010-04-08 21:14:23 -07001646static void cfq_bio_merged(struct request_queue *q, struct request *req,
1647 struct bio *bio)
1648{
Vivek Goyale98ef892010-06-18 10:39:47 -04001649 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1650 bio_data_dir(bio), cfq_bio_sync(bio));
Divyesh Shah812d4022010-04-08 21:14:23 -07001651}
1652
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653static void
Jens Axboe165125e2007-07-24 09:28:11 +02001654cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 struct request *next)
1656{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001657 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001658 /*
1659 * reposition in fifo if next is older than rq
1660 */
1661 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001662 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001663 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001664 rq_set_fifo_time(rq, rq_fifo_time(next));
1665 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001666
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001667 if (cfqq->next_rq == next)
1668 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001669 cfq_remove_request(next);
Vivek Goyale98ef892010-06-18 10:39:47 -04001670 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1671 rq_data_dir(next), rq_is_sync(next));
Jens Axboe22e2c502005-06-27 10:55:12 +02001672}
1673
Jens Axboe165125e2007-07-24 09:28:11 +02001674static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001675 struct bio *bio)
1676{
1677 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001678 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001679 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001680
1681 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001682 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001683 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001684 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001685 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001686
1687 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001688 * Lookup the cfqq that this bio will be queued with. Allow
1689 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001690 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001691 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001692 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001693 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001694
Vasily Tarasov91fac312007-04-25 12:29:51 +02001695 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001696 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001697}
1698
Divyesh Shah812df482010-04-08 21:15:35 -07001699static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1700{
1701 del_timer(&cfqd->idle_slice_timer);
Vivek Goyale98ef892010-06-18 10:39:47 -04001702 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
Divyesh Shah812df482010-04-08 21:15:35 -07001703}
1704
Jens Axboefebffd62008-01-28 13:19:43 +01001705static void __cfq_set_active_queue(struct cfq_data *cfqd,
1706 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001707{
1708 if (cfqq) {
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001709 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1710 cfqd->serving_prio, cfqd->serving_type);
Justin TerAvest62a37f62011-03-23 08:25:44 +01001711 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1712 cfqq->slice_start = 0;
1713 cfqq->dispatch_start = jiffies;
1714 cfqq->allocated_slice = 0;
1715 cfqq->slice_end = 0;
1716 cfqq->slice_dispatch = 0;
1717 cfqq->nr_sectors = 0;
1718
1719 cfq_clear_cfqq_wait_request(cfqq);
1720 cfq_clear_cfqq_must_dispatch(cfqq);
1721 cfq_clear_cfqq_must_alloc_slice(cfqq);
1722 cfq_clear_cfqq_fifo_expire(cfqq);
1723 cfq_mark_cfqq_slice_new(cfqq);
1724
1725 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001726 }
1727
1728 cfqd->active_queue = cfqq;
1729}
1730
1731/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001732 * current cfqq expired its slice (or was too idle), select new one
1733 */
1734static void
1735__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001736 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001737{
Jens Axboe7b679132008-05-30 12:23:07 +02001738 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1739
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001740 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07001741 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001742
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001743 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001744 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001745
1746 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01001747 * If this cfqq is shared between multiple processes, check to
1748 * make sure that those processes are still issuing I/Os within
1749 * the mean seek distance. If not, it may be time to break the
1750 * queues apart again.
1751 */
1752 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1753 cfq_mark_cfqq_split_coop(cfqq);
1754
1755 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001756 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001757 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01001758 if (timed_out) {
1759 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07001760 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001761 else
1762 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001763 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1764 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001765
Vivek Goyale5ff0822010-04-26 19:25:11 +02001766 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001767
Vivek Goyalf04a6422009-12-03 12:59:40 -05001768 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1769 cfq_del_cfqq_rr(cfqd, cfqq);
1770
Jens Axboeedd75ff2007-04-19 12:03:34 +02001771 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001772
1773 if (cfqq == cfqd->active_queue)
1774 cfqd->active_queue = NULL;
1775
1776 if (cfqd->active_cic) {
1777 put_io_context(cfqd->active_cic->ioc);
1778 cfqd->active_cic = NULL;
1779 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001780}
1781
Vivek Goyale5ff0822010-04-26 19:25:11 +02001782static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001783{
1784 struct cfq_queue *cfqq = cfqd->active_queue;
1785
1786 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02001787 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001788}
1789
Jens Axboe498d3aa2007-04-26 12:54:48 +02001790/*
1791 * Get next queue for service. Unless we have a queue preemption,
1792 * we'll simply select the first cfqq in the service tree.
1793 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001794static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001795{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001796 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001797 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
Vivek Goyal65b32a52009-12-16 17:52:59 -05001798 cfqd->serving_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001799
Vivek Goyalf04a6422009-12-03 12:59:40 -05001800 if (!cfqd->rq_queued)
1801 return NULL;
1802
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001803 /* There is nothing to dispatch */
1804 if (!service_tree)
1805 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001806 if (RB_EMPTY_ROOT(&service_tree->rb))
1807 return NULL;
1808 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001809}
1810
Vivek Goyalf04a6422009-12-03 12:59:40 -05001811static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1812{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001813 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001814 struct cfq_queue *cfqq;
1815 int i, j;
1816 struct cfq_rb_root *st;
1817
1818 if (!cfqd->rq_queued)
1819 return NULL;
1820
Vivek Goyal25fb5162009-12-03 12:59:46 -05001821 cfqg = cfq_get_next_cfqg(cfqd);
1822 if (!cfqg)
1823 return NULL;
1824
Vivek Goyalf04a6422009-12-03 12:59:40 -05001825 for_each_cfqg_st(cfqg, i, j, st)
1826 if ((cfqq = cfq_rb_first(st)) != NULL)
1827 return cfqq;
1828 return NULL;
1829}
1830
Jens Axboe498d3aa2007-04-26 12:54:48 +02001831/*
1832 * Get and set a new active queue for service.
1833 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001834static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1835 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001836{
Jens Axboee00ef792009-11-04 08:54:55 +01001837 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001838 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001839
Jens Axboe22e2c502005-06-27 10:55:12 +02001840 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001841 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001842}
1843
Jens Axboed9e76202007-04-20 14:27:50 +02001844static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1845 struct request *rq)
1846{
Tejun Heo83096eb2009-05-07 22:24:39 +09001847 if (blk_rq_pos(rq) >= cfqd->last_position)
1848 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001849 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001850 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001851}
1852
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001853static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01001854 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001855{
Shaohua Lie9ce3352010-03-19 08:03:04 +01001856 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001857}
1858
Jens Axboea36e71f2009-04-15 12:15:11 +02001859static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1860 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001861{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001862 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001863 struct rb_node *parent, *node;
1864 struct cfq_queue *__cfqq;
1865 sector_t sector = cfqd->last_position;
1866
1867 if (RB_EMPTY_ROOT(root))
1868 return NULL;
1869
1870 /*
1871 * First, if we find a request starting at the end of the last
1872 * request, choose it.
1873 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001874 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001875 if (__cfqq)
1876 return __cfqq;
1877
1878 /*
1879 * If the exact sector wasn't found, the parent of the NULL leaf
1880 * will contain the closest sector.
1881 */
1882 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001883 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001884 return __cfqq;
1885
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001886 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001887 node = rb_next(&__cfqq->p_node);
1888 else
1889 node = rb_prev(&__cfqq->p_node);
1890 if (!node)
1891 return NULL;
1892
1893 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01001894 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001895 return __cfqq;
1896
1897 return NULL;
1898}
1899
1900/*
1901 * cfqd - obvious
1902 * cur_cfqq - passed in so that we don't decide that the current queue is
1903 * closely cooperating with itself.
1904 *
1905 * So, basically we're assuming that that cur_cfqq has dispatched at least
1906 * one request, and that cfqd->last_position reflects a position on the disk
1907 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1908 * assumption.
1909 */
1910static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001911 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001912{
1913 struct cfq_queue *cfqq;
1914
Divyesh Shah39c01b22010-03-25 15:45:57 +01001915 if (cfq_class_idle(cur_cfqq))
1916 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001917 if (!cfq_cfqq_sync(cur_cfqq))
1918 return NULL;
1919 if (CFQQ_SEEKY(cur_cfqq))
1920 return NULL;
1921
Jens Axboea36e71f2009-04-15 12:15:11 +02001922 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001923 * Don't search priority tree if it's the only queue in the group.
1924 */
1925 if (cur_cfqq->cfqg->nr_cfqq == 1)
1926 return NULL;
1927
1928 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001929 * We should notice if some of the queues are cooperating, eg
1930 * working closely on the same area of the disk. In that case,
1931 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001932 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001933 cfqq = cfqq_close(cfqd, cur_cfqq);
1934 if (!cfqq)
1935 return NULL;
1936
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001937 /* If new queue belongs to different cfq_group, don't choose it */
1938 if (cur_cfqq->cfqg != cfqq->cfqg)
1939 return NULL;
1940
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001941 /*
1942 * It only makes sense to merge sync queues.
1943 */
1944 if (!cfq_cfqq_sync(cfqq))
1945 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001946 if (CFQQ_SEEKY(cfqq))
1947 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001948
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001949 /*
1950 * Do not merge queues of different priority classes
1951 */
1952 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1953 return NULL;
1954
Jens Axboea36e71f2009-04-15 12:15:11 +02001955 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001956}
1957
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001958/*
1959 * Determine whether we should enforce idle window for this queue.
1960 */
1961
1962static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1963{
1964 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001965 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001966
Vivek Goyalf04a6422009-12-03 12:59:40 -05001967 BUG_ON(!service_tree);
1968 BUG_ON(!service_tree->count);
1969
Vivek Goyalb6508c12010-08-23 12:23:33 +02001970 if (!cfqd->cfq_slice_idle)
1971 return false;
1972
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001973 /* We never do for idle class queues. */
1974 if (prio == IDLE_WORKLOAD)
1975 return false;
1976
1977 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001978 if (cfq_cfqq_idle_window(cfqq) &&
1979 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001980 return true;
1981
1982 /*
1983 * Otherwise, we do only if they are the last ones
1984 * in their service tree.
1985 */
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02001986 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1987 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
Shaohua Lic1e44752010-11-08 15:01:02 +01001988 return true;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01001989 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1990 service_tree->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01001991 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001992}
1993
Jens Axboe6d048f52007-04-25 12:44:27 +02001994static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001995{
Jens Axboe17926692007-01-19 11:59:30 +11001996 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001997 struct cfq_io_context *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02001998 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001999
Jens Axboea68bbdd2008-09-24 13:03:33 +02002000 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002001 * SSD device without seek penalty, disable idling. But only do so
2002 * for devices that support queuing, otherwise we still have a problem
2003 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02002004 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002005 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02002006 return;
2007
Jens Axboedd67d052006-06-21 09:36:18 +02002008 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02002009 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02002010
2011 /*
2012 * idle is disabled, either manually or by past process history
2013 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002014 if (!cfq_should_idle(cfqd, cfqq)) {
2015 /* no queue idling. Check for group idling */
2016 if (cfqd->cfq_group_idle)
2017 group_idle = cfqd->cfq_group_idle;
2018 else
2019 return;
2020 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002021
Jens Axboe22e2c502005-06-27 10:55:12 +02002022 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002023 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02002024 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002025 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02002026 return;
2027
2028 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002029 * task has exited, don't wait
2030 */
Jens Axboe206dc692006-03-28 13:03:44 +02002031 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01002032 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02002033 return;
2034
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002035 /*
2036 * If our average think time is larger than the remaining time
2037 * slice, then don't idle. This avoids overrunning the allotted
2038 * time slice.
2039 */
Shaohua Li383cd722011-07-12 14:24:35 +02002040 if (sample_valid(cic->ttime.ttime_samples) &&
2041 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
Joe Perchesfd16d262011-06-13 10:42:49 +02002042 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
Shaohua Li383cd722011-07-12 14:24:35 +02002043 cic->ttime.ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002044 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002045 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002046
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002047 /* There are other queues in the group, don't do group idle */
2048 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2049 return;
2050
Jens Axboe3b181522005-06-27 10:56:24 +02002051 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002052
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002053 if (group_idle)
2054 sl = cfqd->cfq_group_idle;
2055 else
2056 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02002057
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002058 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Vivek Goyale98ef892010-06-18 10:39:47 -04002059 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002060 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2061 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062}
2063
Jens Axboe498d3aa2007-04-26 12:54:48 +02002064/*
2065 * Move request from internal lists to the request queue dispatch list.
2066 */
Jens Axboe165125e2007-07-24 09:28:11 +02002067static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068{
Jens Axboe3ed9a292007-04-23 08:33:33 +02002069 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002070 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002071
Jens Axboe7b679132008-05-30 12:23:07 +02002072 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2073
Jeff Moyer06d21882009-09-11 17:08:59 +02002074 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02002075 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002076 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002077 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02002078 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02002079
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002080 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02002081 cfqq->nr_sectors += blk_rq_sectors(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04002082 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
Divyesh Shah84c124d2010-04-09 08:31:19 +02002083 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084}
2085
2086/*
2087 * return expired entry, or NULL to just start from scratch in rbtree
2088 */
Jens Axboefebffd62008-01-28 13:19:43 +01002089static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090{
Jens Axboe30996f42009-10-05 11:03:39 +02002091 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092
Jens Axboe3b181522005-06-27 10:56:24 +02002093 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002095
2096 cfq_mark_cfqq_fifo_expire(cfqq);
2097
Jens Axboe89850f72006-07-22 16:48:31 +02002098 if (list_empty(&cfqq->fifo))
2099 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100
Jens Axboe89850f72006-07-22 16:48:31 +02002101 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002102 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002103 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104
Jens Axboe30996f42009-10-05 11:03:39 +02002105 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002106 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107}
2108
Jens Axboe22e2c502005-06-27 10:55:12 +02002109static inline int
2110cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2111{
2112 const int base_rq = cfqd->cfq_slice_async_rq;
2113
2114 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2115
Namhyung Kimb9f8ce02011-05-24 10:23:21 +02002116 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002117}
2118
2119/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002120 * Must be called with the queue_lock held.
2121 */
2122static int cfqq_process_refs(struct cfq_queue *cfqq)
2123{
2124 int process_refs, io_refs;
2125
2126 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002127 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002128 BUG_ON(process_refs < 0);
2129 return process_refs;
2130}
2131
2132static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2133{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002134 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002135 struct cfq_queue *__cfqq;
2136
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002137 /*
2138 * If there are no process references on the new_cfqq, then it is
2139 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2140 * chain may have dropped their last reference (not just their
2141 * last process reference).
2142 */
2143 if (!cfqq_process_refs(new_cfqq))
2144 return;
2145
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002146 /* Avoid a circular list and skip interim queue merges */
2147 while ((__cfqq = new_cfqq->new_cfqq)) {
2148 if (__cfqq == cfqq)
2149 return;
2150 new_cfqq = __cfqq;
2151 }
2152
2153 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002154 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002155 /*
2156 * If the process for the cfqq has gone away, there is no
2157 * sense in merging the queues.
2158 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002159 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002160 return;
2161
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002162 /*
2163 * Merge in the direction of the lesser amount of work.
2164 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002165 if (new_process_refs >= process_refs) {
2166 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002167 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002168 } else {
2169 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002170 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002171 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002172}
2173
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002174static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
Vivek Goyal65b32a52009-12-16 17:52:59 -05002175 struct cfq_group *cfqg, enum wl_prio_t prio)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002176{
2177 struct cfq_queue *queue;
2178 int i;
2179 bool key_valid = false;
2180 unsigned long lowest_key = 0;
2181 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2182
Vivek Goyal65b32a52009-12-16 17:52:59 -05002183 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2184 /* select the one with lowest rb_key */
2185 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002186 if (queue &&
2187 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2188 lowest_key = queue->rb_key;
2189 cur_best = i;
2190 key_valid = true;
2191 }
2192 }
2193
2194 return cur_best;
2195}
2196
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002197static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002198{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002199 unsigned slice;
2200 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002201 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002202 unsigned group_slice;
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002203 enum wl_prio_t original_prio = cfqd->serving_prio;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002204
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002205 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002206 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002207 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002208 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002209 cfqd->serving_prio = BE_WORKLOAD;
2210 else {
2211 cfqd->serving_prio = IDLE_WORKLOAD;
2212 cfqd->workload_expires = jiffies + 1;
2213 return;
2214 }
2215
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002216 if (original_prio != cfqd->serving_prio)
2217 goto new_workload;
2218
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002219 /*
2220 * For RT and BE, we have to choose also the type
2221 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2222 * expiration time
2223 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002224 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002225 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002226
2227 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002228 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002229 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002230 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002231 return;
2232
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002233new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002234 /* otherwise select new workload type */
2235 cfqd->serving_type =
Vivek Goyal65b32a52009-12-16 17:52:59 -05002236 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2237 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002238 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002239
2240 /*
2241 * the workload slice is computed as a fraction of target latency
2242 * proportional to the number of queues in that workload, over
2243 * all the queues in the same priority class
2244 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002245 group_slice = cfq_group_slice(cfqd, cfqg);
2246
2247 slice = group_slice * count /
2248 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2249 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002250
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002251 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2252 unsigned int tmp;
2253
2254 /*
2255 * Async queues are currently system wide. Just taking
2256 * proportion of queues with-in same group will lead to higher
2257 * async ratio system wide as generally root group is going
2258 * to have higher weight. A more accurate thing would be to
2259 * calculate system wide asnc/sync ratio.
2260 */
2261 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2262 tmp = tmp/cfqd->busy_queues;
2263 slice = min_t(unsigned, slice, tmp);
2264
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002265 /* async workload slice is scaled down according to
2266 * the sync/async slice ratio. */
2267 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002268 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002269 /* sync workload slice is at least 2 * cfq_slice_idle */
2270 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2271
2272 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002273 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002274 cfqd->workload_expires = jiffies + slice;
2275}
2276
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002277static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2278{
2279 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002280 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002281
2282 if (RB_EMPTY_ROOT(&st->rb))
2283 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002284 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002285 update_min_vdisktime(st);
2286 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002287}
2288
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002289static void cfq_choose_cfqg(struct cfq_data *cfqd)
2290{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002291 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2292
2293 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002294
2295 /* Restore the workload type data */
2296 if (cfqg->saved_workload_slice) {
2297 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2298 cfqd->serving_type = cfqg->saved_workload;
2299 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002300 } else
2301 cfqd->workload_expires = jiffies - 1;
2302
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002303 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002304}
2305
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002306/*
Jens Axboe498d3aa2007-04-26 12:54:48 +02002307 * Select a queue for service. If we have a current active queue,
2308 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002309 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002310static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002311{
Jens Axboea36e71f2009-04-15 12:15:11 +02002312 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002313
2314 cfqq = cfqd->active_queue;
2315 if (!cfqq)
2316 goto new_queue;
2317
Vivek Goyalf04a6422009-12-03 12:59:40 -05002318 if (!cfqd->rq_queued)
2319 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002320
2321 /*
2322 * We were waiting for group to get backlogged. Expire the queue
2323 */
2324 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2325 goto expire;
2326
Jens Axboe22e2c502005-06-27 10:55:12 +02002327 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002328 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002329 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002330 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2331 /*
2332 * If slice had not expired at the completion of last request
2333 * we might not have turned on wait_busy flag. Don't expire
2334 * the queue yet. Allow the group to get backlogged.
2335 *
2336 * The very fact that we have used the slice, that means we
2337 * have been idling all along on this queue and it should be
2338 * ok to wait for this request to complete.
2339 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002340 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2341 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2342 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002343 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002344 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002345 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002346 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002347
2348 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002349 * The active queue has requests and isn't expired, allow it to
2350 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002351 */
Jens Axboedd67d052006-06-21 09:36:18 +02002352 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002353 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002354
2355 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002356 * If another queue has a request waiting within our mean seek
2357 * distance, let it run. The expire code will check for close
2358 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002359 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002360 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002361 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002362 if (new_cfqq) {
2363 if (!cfqq->new_cfqq)
2364 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002365 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002366 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002367
2368 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002369 * No requests pending. If the active queue still has requests in
2370 * flight or is idling for a new request, allow either of these
2371 * conditions to happen (or time out) before selecting a new queue.
2372 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002373 if (timer_pending(&cfqd->idle_slice_timer)) {
2374 cfqq = NULL;
2375 goto keep_queue;
2376 }
2377
Shaohua Li8e1ac662010-11-08 15:01:04 +01002378 /*
2379 * This is a deep seek queue, but the device is much faster than
2380 * the queue can deliver, don't idle
2381 **/
2382 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2383 (cfq_cfqq_slice_new(cfqq) ||
2384 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2385 cfq_clear_cfqq_deep(cfqq);
2386 cfq_clear_cfqq_idle_window(cfqq);
2387 }
2388
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002389 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2390 cfqq = NULL;
2391 goto keep_queue;
2392 }
2393
2394 /*
2395 * If group idle is enabled and there are requests dispatched from
2396 * this group, wait for requests to complete.
2397 */
2398check_group_idle:
Shaohua Li7700fc42011-07-12 14:24:56 +02002399 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2400 cfqq->cfqg->dispatched &&
2401 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002402 cfqq = NULL;
2403 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002404 }
2405
Jens Axboe3b181522005-06-27 10:56:24 +02002406expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02002407 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002408new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002409 /*
2410 * Current queue expired. Check if we have to switch to a new
2411 * service tree
2412 */
2413 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002414 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002415
Jens Axboea36e71f2009-04-15 12:15:11 +02002416 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002417keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002418 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002419}
2420
Jens Axboefebffd62008-01-28 13:19:43 +01002421static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002422{
2423 int dispatched = 0;
2424
2425 while (cfqq->next_rq) {
2426 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2427 dispatched++;
2428 }
2429
2430 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002431
2432 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002433 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002434 return dispatched;
2435}
2436
Jens Axboe498d3aa2007-04-26 12:54:48 +02002437/*
2438 * Drain our current requests. Used for barriers and when switching
2439 * io schedulers on-the-fly.
2440 */
Jens Axboed9e76202007-04-20 14:27:50 +02002441static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002442{
Jens Axboe08717142008-01-28 11:38:15 +01002443 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002444 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002445
Divyesh Shah3440c492010-04-09 09:29:57 +02002446 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02002447 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02002448 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2449 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002450 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02002451 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002452
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002453 BUG_ON(cfqd->busy_queues);
2454
Jeff Moyer69237152009-06-12 15:29:30 +02002455 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002456 return dispatched;
2457}
2458
Shaohua Liabc3c742010-03-01 09:20:54 +01002459static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2460 struct cfq_queue *cfqq)
2461{
2462 /* the queue hasn't finished any request, can't estimate */
2463 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01002464 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002465 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2466 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01002467 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01002468
Shaohua Lic1e44752010-11-08 15:01:02 +01002469 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01002470}
2471
Jens Axboe0b182d62009-10-06 20:49:37 +02002472static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002473{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002474 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475
Jens Axboe2f5cb732009-04-07 08:51:19 +02002476 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002477 * Drain async requests before we start sync IO
2478 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002479 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002480 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002481
2482 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002483 * If this is an async queue and we have sync IO in flight, let it wait
2484 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002485 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002486 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002487
Shaohua Liabc3c742010-03-01 09:20:54 +01002488 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002489 if (cfq_class_idle(cfqq))
2490 max_dispatch = 1;
2491
2492 /*
2493 * Does this cfqq already have too much IO in flight?
2494 */
2495 if (cfqq->dispatched >= max_dispatch) {
Shaohua Lief8a41d2011-03-07 09:26:29 +01002496 bool promote_sync = false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002497 /*
2498 * idle queue must always only have a single IO in flight
2499 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002500 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002501 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002502
Jens Axboe2f5cb732009-04-07 08:51:19 +02002503 /*
Li, Shaohuac4ade942011-03-23 08:30:34 +01002504 * If there is only one sync queue
2505 * we can ignore async queue here and give the sync
Shaohua Lief8a41d2011-03-07 09:26:29 +01002506 * queue no dispatch limit. The reason is a sync queue can
2507 * preempt async queue, limiting the sync queue doesn't make
2508 * sense. This is useful for aiostress test.
2509 */
Li, Shaohuac4ade942011-03-23 08:30:34 +01002510 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2511 promote_sync = true;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002512
2513 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002514 * We have other queues, don't allow more IO from this one
2515 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002516 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2517 !promote_sync)
Jens Axboe0b182d62009-10-06 20:49:37 +02002518 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002519
Jens Axboe2f5cb732009-04-07 08:51:19 +02002520 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002521 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002522 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01002523 if (cfqd->busy_queues == 1 || promote_sync)
Shaohua Liabc3c742010-03-01 09:20:54 +01002524 max_dispatch = -1;
2525 else
2526 /*
2527 * Normally we start throttling cfqq when cfq_quantum/2
2528 * requests have been dispatched. But we can drive
2529 * deeper queue depths at the beginning of slice
2530 * subjected to upper limit of cfq_quantum.
2531 * */
2532 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02002533 }
2534
2535 /*
2536 * Async queues must wait a bit before being allowed dispatch.
2537 * We also ramp up the dispatch depth gradually for async IO,
2538 * based on the last sync IO we serviced
2539 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002540 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002541 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002542 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002543
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002544 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002545 if (!depth && !cfqq->dispatched)
2546 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002547 if (depth < max_dispatch)
2548 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 }
2550
Jens Axboe0b182d62009-10-06 20:49:37 +02002551 /*
2552 * If we're below the current max, allow a dispatch
2553 */
2554 return cfqq->dispatched < max_dispatch;
2555}
2556
2557/*
2558 * Dispatch a request from cfqq, moving them to the request queue
2559 * dispatch list.
2560 */
2561static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2562{
2563 struct request *rq;
2564
2565 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2566
2567 if (!cfq_may_dispatch(cfqd, cfqq))
2568 return false;
2569
2570 /*
2571 * follow expired path, else get first next available
2572 */
2573 rq = cfq_check_fifo(cfqq);
2574 if (!rq)
2575 rq = cfqq->next_rq;
2576
2577 /*
2578 * insert request into driver dispatch list
2579 */
2580 cfq_dispatch_insert(cfqd->queue, rq);
2581
2582 if (!cfqd->active_cic) {
2583 struct cfq_io_context *cic = RQ_CIC(rq);
2584
2585 atomic_long_inc(&cic->ioc->refcount);
2586 cfqd->active_cic = cic;
2587 }
2588
2589 return true;
2590}
2591
2592/*
2593 * Find the cfqq that we need to service and move a request from that to the
2594 * dispatch list
2595 */
2596static int cfq_dispatch_requests(struct request_queue *q, int force)
2597{
2598 struct cfq_data *cfqd = q->elevator->elevator_data;
2599 struct cfq_queue *cfqq;
2600
2601 if (!cfqd->busy_queues)
2602 return 0;
2603
2604 if (unlikely(force))
2605 return cfq_forced_dispatch(cfqd);
2606
2607 cfqq = cfq_select_queue(cfqd);
2608 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002609 return 0;
2610
Jens Axboe2f5cb732009-04-07 08:51:19 +02002611 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002612 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002613 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002614 if (!cfq_dispatch_request(cfqd, cfqq))
2615 return 0;
2616
Jens Axboe2f5cb732009-04-07 08:51:19 +02002617 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002618 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002619
2620 /*
2621 * expire an async queue immediately if it has used up its slice. idle
2622 * queue always expire after 1 dispatch round.
2623 */
2624 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2625 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2626 cfq_class_idle(cfqq))) {
2627 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02002628 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002629 }
2630
Shan Weib217a902009-09-01 10:06:42 +02002631 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002632 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633}
2634
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635/*
Jens Axboe5e705372006-07-13 12:39:25 +02002636 * task holds one reference to the queue, dropped when task exits. each rq
2637 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002639 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640 * queue lock must be held here.
2641 */
2642static void cfq_put_queue(struct cfq_queue *cfqq)
2643{
Jens Axboe22e2c502005-06-27 10:55:12 +02002644 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05002645 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002646
Shaohua Li30d7b942011-01-07 08:46:59 +01002647 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648
Shaohua Li30d7b942011-01-07 08:46:59 +01002649 cfqq->ref--;
2650 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 return;
2652
Jens Axboe7b679132008-05-30 12:23:07 +02002653 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002655 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002656 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002658 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02002659 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002660 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002661 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002662
Vivek Goyalf04a6422009-12-03 12:59:40 -05002663 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002665 cfq_put_cfqg(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666}
2667
Jens Axboed6de8be2008-05-28 14:46:59 +02002668/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002669 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002670 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002671static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002672call_for_each_cic(struct io_context *ioc,
2673 void (*func)(struct io_context *, struct cfq_io_context *))
2674{
Jens Axboe5f45c692011-04-19 09:10:35 +02002675 struct cfq_io_context *cic;
2676 struct hlist_node *n;
2677
Jens Axboe4ac845a2008-01-24 08:44:49 +01002678 rcu_read_lock();
Jens Axboe5f45c692011-04-19 09:10:35 +02002679
2680 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2681 func(ioc, cic);
2682
Jens Axboe4ac845a2008-01-24 08:44:49 +01002683 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002684}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002685
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002686static void cfq_cic_free_rcu(struct rcu_head *head)
2687{
2688 struct cfq_io_context *cic;
2689
2690 cic = container_of(head, struct cfq_io_context, rcu_head);
2691
2692 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002693 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002694
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002695 if (ioc_gone) {
2696 /*
2697 * CFQ scheduler is exiting, grab exit lock and check
2698 * the pending io context count. If it hits zero,
2699 * complete ioc_gone and set it back to NULL
2700 */
2701 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002702 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002703 complete(ioc_gone);
2704 ioc_gone = NULL;
2705 }
2706 spin_unlock(&ioc_gone_lock);
2707 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002708}
2709
2710static void cfq_cic_free(struct cfq_io_context *cic)
2711{
2712 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002713}
2714
2715static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2716{
2717 unsigned long flags;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002718 unsigned long dead_key = (unsigned long) cic->key;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002719
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002720 BUG_ON(!(dead_key & CIC_DEAD_KEY));
Jens Axboe4ac845a2008-01-24 08:44:49 +01002721
2722 spin_lock_irqsave(&ioc->lock, flags);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04002723 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
Jens Axboeffc4e752008-02-19 10:02:29 +01002724 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002725 spin_unlock_irqrestore(&ioc->lock, flags);
2726
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002727 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002728}
2729
Jens Axboed6de8be2008-05-28 14:46:59 +02002730/*
2731 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2732 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2733 * and ->trim() which is called with the task lock held
2734 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002735static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002737 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002738 * ioc->refcount is zero here, or we are called from elv_unregister(),
2739 * so no more cic's are allowed to be linked into this ioc. So it
2740 * should be ok to iterate over the known list, we will see all cic's
2741 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002742 */
Jens Axboe5f45c692011-04-19 09:10:35 +02002743 call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744}
2745
Shaohua Lid02a2c02010-05-25 10:16:53 +02002746static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02002747{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002748 struct cfq_queue *__cfqq, *next;
2749
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002750 /*
2751 * If this queue was scheduled to merge with another queue, be
2752 * sure to drop the reference taken on that queue (and others in
2753 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2754 */
2755 __cfqq = cfqq->new_cfqq;
2756 while (__cfqq) {
2757 if (__cfqq == cfqq) {
2758 WARN(1, "cfqq->new_cfqq loop detected\n");
2759 break;
2760 }
2761 next = __cfqq->new_cfqq;
2762 cfq_put_queue(__cfqq);
2763 __cfqq = next;
2764 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02002765}
2766
2767static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2768{
2769 if (unlikely(cfqq == cfqd->active_queue)) {
2770 __cfq_slice_expired(cfqd, cfqq, 0);
2771 cfq_schedule_dispatch(cfqd);
2772 }
2773
2774 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002775
Jens Axboe89850f72006-07-22 16:48:31 +02002776 cfq_put_queue(cfqq);
2777}
2778
2779static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2780 struct cfq_io_context *cic)
2781{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002782 struct io_context *ioc = cic->ioc;
2783
Jens Axboefc463792006-08-29 09:05:44 +02002784 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002785
2786 /*
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002787 * Make sure dead mark is seen for dead queues
Jens Axboe4ac845a2008-01-24 08:44:49 +01002788 */
Jens Axboefc463792006-08-29 09:05:44 +02002789 smp_wmb();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002790 cic->key = cfqd_dead_key(cfqd);
Jens Axboefc463792006-08-29 09:05:44 +02002791
Shaohua Li3181faa2011-06-27 09:03:47 +02002792 rcu_read_lock();
Jens Axboe9b509022011-06-05 06:01:13 +02002793 if (rcu_dereference(ioc->ioc_data) == cic) {
Shaohua Li3181faa2011-06-27 09:03:47 +02002794 rcu_read_unlock();
Jens Axboe9b509022011-06-05 06:01:13 +02002795 spin_lock(&ioc->lock);
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002796 rcu_assign_pointer(ioc->ioc_data, NULL);
Jens Axboe9b509022011-06-05 06:01:13 +02002797 spin_unlock(&ioc->lock);
Shaohua Li3181faa2011-06-27 09:03:47 +02002798 } else
2799 rcu_read_unlock();
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002800
Jens Axboeff6657c2009-04-08 10:58:57 +02002801 if (cic->cfqq[BLK_RW_ASYNC]) {
2802 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2803 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002804 }
2805
Jens Axboeff6657c2009-04-08 10:58:57 +02002806 if (cic->cfqq[BLK_RW_SYNC]) {
2807 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2808 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002809 }
Jens Axboe89850f72006-07-22 16:48:31 +02002810}
2811
Jens Axboe4ac845a2008-01-24 08:44:49 +01002812static void cfq_exit_single_io_context(struct io_context *ioc,
2813 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002814{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002815 struct cfq_data *cfqd = cic_to_cfqd(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002816
Jens Axboe89850f72006-07-22 16:48:31 +02002817 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002818 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002819 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002820
Jens Axboe4ac845a2008-01-24 08:44:49 +01002821 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002822
2823 /*
2824 * Ensure we get a fresh copy of the ->key to prevent
2825 * race between exiting task and queue
2826 */
2827 smp_read_barrier_depends();
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002828 if (cic->key == cfqd)
Jens Axboe62c1fe92008-12-15 21:19:25 +01002829 __cfq_exit_single_io_context(cfqd, cic);
2830
Jens Axboe4ac845a2008-01-24 08:44:49 +01002831 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002832 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002833}
2834
Jens Axboe498d3aa2007-04-26 12:54:48 +02002835/*
2836 * The process that ioc belongs to has exited, we need to clean up
2837 * and put the internal structures we have that belongs to that process.
2838 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002839static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002841 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842}
2843
Jens Axboe22e2c502005-06-27 10:55:12 +02002844static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002845cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846{
Jens Axboeb5deef92006-07-19 23:39:40 +02002847 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848
Christoph Lameter94f60302007-07-17 04:03:29 -07002849 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2850 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 if (cic) {
Shaohua Li383cd722011-07-12 14:24:35 +02002852 cic->ttime.last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002853 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002854 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002855 cic->dtor = cfq_free_io_context;
2856 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002857 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 }
2859
2860 return cic;
2861}
2862
Jens Axboefd0928d2008-01-24 08:52:45 +01002863static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002864{
2865 struct task_struct *tsk = current;
2866 int ioprio_class;
2867
Jens Axboe3b181522005-06-27 10:56:24 +02002868 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002869 return;
2870
Jens Axboefd0928d2008-01-24 08:52:45 +01002871 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002872 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002873 default:
2874 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2875 case IOPRIO_CLASS_NONE:
2876 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002877 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002878 */
2879 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002880 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002881 break;
2882 case IOPRIO_CLASS_RT:
2883 cfqq->ioprio = task_ioprio(ioc);
2884 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2885 break;
2886 case IOPRIO_CLASS_BE:
2887 cfqq->ioprio = task_ioprio(ioc);
2888 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2889 break;
2890 case IOPRIO_CLASS_IDLE:
2891 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2892 cfqq->ioprio = 7;
2893 cfq_clear_cfqq_idle_window(cfqq);
2894 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002895 }
2896
2897 /*
2898 * keep track of original prio settings in case we have to temporarily
2899 * elevate the priority of this queue
2900 */
2901 cfqq->org_ioprio = cfqq->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02002902 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002903}
2904
Jens Axboefebffd62008-01-28 13:19:43 +01002905static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002906{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002907 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05002908 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002909 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002910
Jens Axboecaaa5f92006-06-16 11:23:00 +02002911 if (unlikely(!cfqd))
2912 return;
2913
Jens Axboec1b707d2006-10-30 19:54:23 +01002914 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002915
Jens Axboeff6657c2009-04-08 10:58:57 +02002916 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002917 if (cfqq) {
2918 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002919 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2920 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002921 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002922 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002923 cfq_put_queue(cfqq);
2924 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002925 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002926
Jens Axboeff6657c2009-04-08 10:58:57 +02002927 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002928 if (cfqq)
2929 cfq_mark_cfqq_prio_changed(cfqq);
2930
Jens Axboec1b707d2006-10-30 19:54:23 +01002931 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002932}
2933
Jens Axboefc463792006-08-29 09:05:44 +02002934static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002936 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002937 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938}
2939
Jens Axboed5036d72009-06-26 10:44:34 +02002940static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002941 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002942{
2943 RB_CLEAR_NODE(&cfqq->rb_node);
2944 RB_CLEAR_NODE(&cfqq->p_node);
2945 INIT_LIST_HEAD(&cfqq->fifo);
2946
Shaohua Li30d7b942011-01-07 08:46:59 +01002947 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02002948 cfqq->cfqd = cfqd;
2949
2950 cfq_mark_cfqq_prio_changed(cfqq);
2951
2952 if (is_sync) {
2953 if (!cfq_class_idle(cfqq))
2954 cfq_mark_cfqq_idle_window(cfqq);
2955 cfq_mark_cfqq_sync(cfqq);
2956 }
2957 cfqq->pid = pid;
2958}
2959
Vivek Goyal24610332009-12-03 12:59:51 -05002960#ifdef CONFIG_CFQ_GROUP_IOSCHED
2961static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2962{
2963 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04002964 struct cfq_data *cfqd = cic_to_cfqd(cic);
Vivek Goyal24610332009-12-03 12:59:51 -05002965 unsigned long flags;
2966 struct request_queue *q;
2967
2968 if (unlikely(!cfqd))
2969 return;
2970
2971 q = cfqd->queue;
2972
2973 spin_lock_irqsave(q->queue_lock, flags);
2974
2975 if (sync_cfqq) {
2976 /*
2977 * Drop reference to sync queue. A new sync queue will be
2978 * assigned in new group upon arrival of a fresh request.
2979 */
2980 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2981 cic_set_cfqq(cic, NULL, 1);
2982 cfq_put_queue(sync_cfqq);
2983 }
2984
2985 spin_unlock_irqrestore(q->queue_lock, flags);
2986}
2987
2988static void cfq_ioc_set_cgroup(struct io_context *ioc)
2989{
2990 call_for_each_cic(ioc, changed_cgroup);
2991 ioc->cgroup_changed = 0;
2992}
2993#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2994
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002996cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002997 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003000 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003001 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002
3003retry:
Vivek Goyal3e59cf92011-05-19 15:38:21 -04003004 cfqg = cfq_get_cfqg(cfqd);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003005 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003006 /* cic always exists here */
3007 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008
Jens Axboe6118b702009-06-30 09:34:12 +02003009 /*
3010 * Always try a new alloc if we fell back to the OOM cfqq
3011 * originally, since it should just be a temporary situation.
3012 */
3013 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3014 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 if (new_cfqq) {
3016 cfqq = new_cfqq;
3017 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02003018 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07003020 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02003021 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07003022 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02003024 if (new_cfqq)
3025 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02003026 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07003027 cfqq = kmem_cache_alloc_node(cfq_pool,
3028 gfp_mask | __GFP_ZERO,
3029 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02003030 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031
Jens Axboe6118b702009-06-30 09:34:12 +02003032 if (cfqq) {
3033 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3034 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003035 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02003036 cfq_log_cfqq(cfqd, cfqq, "alloced");
3037 } else
3038 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039 }
3040
3041 if (new_cfqq)
3042 kmem_cache_free(cfq_pool, new_cfqq);
3043
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 return cfqq;
3045}
3046
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003047static struct cfq_queue **
3048cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3049{
Jens Axboefe094d92008-01-31 13:08:54 +01003050 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003051 case IOPRIO_CLASS_RT:
3052 return &cfqd->async_cfqq[0][ioprio];
3053 case IOPRIO_CLASS_BE:
3054 return &cfqd->async_cfqq[1][ioprio];
3055 case IOPRIO_CLASS_IDLE:
3056 return &cfqd->async_idle_cfqq;
3057 default:
3058 BUG();
3059 }
3060}
3061
Jens Axboe15c31be2007-07-10 13:43:25 +02003062static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02003063cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02003064 gfp_t gfp_mask)
3065{
Jens Axboefd0928d2008-01-24 08:52:45 +01003066 const int ioprio = task_ioprio(ioc);
3067 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003068 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02003069 struct cfq_queue *cfqq = NULL;
3070
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003071 if (!is_sync) {
3072 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3073 cfqq = *async_cfqq;
3074 }
3075
Jens Axboe6118b702009-06-30 09:34:12 +02003076 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01003077 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02003078
3079 /*
3080 * pin the queue now that it's allocated, scheduler exit will prune it
3081 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003082 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01003083 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003084 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02003085 }
3086
Shaohua Li30d7b942011-01-07 08:46:59 +01003087 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02003088 return cfqq;
3089}
3090
Jens Axboe498d3aa2007-04-26 12:54:48 +02003091/*
3092 * We drop cfq io contexts lazily, so we may find a dead one.
3093 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003094static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01003095cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3096 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003097{
Jens Axboe4ac845a2008-01-24 08:44:49 +01003098 unsigned long flags;
3099
Jens Axboefc463792006-08-29 09:05:44 +02003100 WARN_ON(!list_empty(&cic->queue_list));
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003101 BUG_ON(cic->key != cfqd_dead_key(cfqd));
Jens Axboe597bc482007-04-24 21:23:53 +02003102
Jens Axboe4ac845a2008-01-24 08:44:49 +01003103 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02003104
Shaohua Li726e99a2011-06-27 09:03:48 +02003105 BUG_ON(rcu_dereference_check(ioc->ioc_data,
3106 lockdep_is_held(&ioc->lock)) == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003107
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003108 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
Jens Axboeffc4e752008-02-19 10:02:29 +01003109 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003110 spin_unlock_irqrestore(&ioc->lock, flags);
3111
3112 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003113}
3114
Jens Axboee2d74ac2006-03-28 08:59:01 +02003115static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01003116cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003117{
Jens Axboee2d74ac2006-03-28 08:59:01 +02003118 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003119 unsigned long flags;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003120
Vasily Tarasov91fac312007-04-25 12:29:51 +02003121 if (unlikely(!ioc))
3122 return NULL;
3123
Jens Axboed6de8be2008-05-28 14:46:59 +02003124 rcu_read_lock();
3125
Jens Axboe597bc482007-04-24 21:23:53 +02003126 /*
3127 * we maintain a last-hit cache, to avoid browsing over the tree
3128 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003129 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02003130 if (cic && cic->key == cfqd) {
3131 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02003132 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02003133 }
Jens Axboe597bc482007-04-24 21:23:53 +02003134
Jens Axboe4ac845a2008-01-24 08:44:49 +01003135 do {
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003136 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003137 rcu_read_unlock();
3138 if (!cic)
3139 break;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003140 if (unlikely(cic->key != cfqd)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01003141 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003142 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01003143 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003144 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003145
Jens Axboed6de8be2008-05-28 14:46:59 +02003146 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003147 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02003148 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003149 break;
3150 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003151
Jens Axboe4ac845a2008-01-24 08:44:49 +01003152 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003153}
3154
Jens Axboe4ac845a2008-01-24 08:44:49 +01003155/*
3156 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3157 * the process specific cfq io context when entered from the block layer.
3158 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3159 */
Jens Axboefebffd62008-01-28 13:19:43 +01003160static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3161 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02003162{
Jens Axboe0261d682006-10-30 19:07:48 +01003163 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003164 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003165
Jens Axboe4ac845a2008-01-24 08:44:49 +01003166 ret = radix_tree_preload(gfp_mask);
3167 if (!ret) {
3168 cic->ioc = ioc;
3169 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003170
Jens Axboe4ac845a2008-01-24 08:44:49 +01003171 spin_lock_irqsave(&ioc->lock, flags);
3172 ret = radix_tree_insert(&ioc->radix_root,
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003173 cfqd->cic_index, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01003174 if (!ret)
3175 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01003176 spin_unlock_irqrestore(&ioc->lock, flags);
3177
3178 radix_tree_preload_end();
3179
3180 if (!ret) {
3181 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3182 list_add(&cic->queue_list, &cfqd->cic_list);
3183 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02003184 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003185 }
3186
Jens Axboe4ac845a2008-01-24 08:44:49 +01003187 if (ret)
3188 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02003189
Jens Axboe4ac845a2008-01-24 08:44:49 +01003190 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02003191}
3192
Jens Axboe22e2c502005-06-27 10:55:12 +02003193/*
3194 * Setup general io context and cfq io context. There can be several cfq
3195 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02003196 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02003197 */
3198static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02003199cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200{
Jens Axboe22e2c502005-06-27 10:55:12 +02003201 struct io_context *ioc = NULL;
3202 struct cfq_io_context *cic;
3203
3204 might_sleep_if(gfp_mask & __GFP_WAIT);
3205
Jens Axboeb5deef92006-07-19 23:39:40 +02003206 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02003207 if (!ioc)
3208 return NULL;
3209
Jens Axboe4ac845a2008-01-24 08:44:49 +01003210 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003211 if (cic)
3212 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02003213
Jens Axboee2d74ac2006-03-28 08:59:01 +02003214 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3215 if (cic == NULL)
3216 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02003217
Jens Axboe4ac845a2008-01-24 08:44:49 +01003218 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3219 goto err_free;
3220
Jens Axboe22e2c502005-06-27 10:55:12 +02003221out:
Jens Axboefc463792006-08-29 09:05:44 +02003222 smp_read_barrier_depends();
3223 if (unlikely(ioc->ioprio_changed))
3224 cfq_ioc_set_ioprio(ioc);
3225
Vivek Goyal24610332009-12-03 12:59:51 -05003226#ifdef CONFIG_CFQ_GROUP_IOSCHED
3227 if (unlikely(ioc->cgroup_changed))
3228 cfq_ioc_set_cgroup(ioc);
3229#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003230 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01003231err_free:
3232 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02003233err:
3234 put_io_context(ioc);
3235 return NULL;
3236}
3237
3238static void
Shaohua Li383cd722011-07-12 14:24:35 +02003239__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003240{
Shaohua Li383cd722011-07-12 14:24:35 +02003241 unsigned long elapsed = jiffies - ttime->last_end_request;
3242 elapsed = min(elapsed, 2UL * slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003243
Shaohua Li383cd722011-07-12 14:24:35 +02003244 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3245 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3246 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3247}
3248
3249static void
3250cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3251 struct cfq_io_context *cic)
3252{
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003253 if (cfq_cfqq_sync(cfqq)) {
Shaohua Li383cd722011-07-12 14:24:35 +02003254 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003255 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3256 cfqd->cfq_slice_idle);
3257 }
Shaohua Li7700fc42011-07-12 14:24:56 +02003258#ifdef CONFIG_CFQ_GROUP_IOSCHED
3259 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3260#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003261}
3262
Jens Axboe206dc692006-03-28 13:03:44 +02003263static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003264cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003265 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003266{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003267 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003268 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003269 if (cfqq->last_request_pos) {
3270 if (cfqq->last_request_pos < blk_rq_pos(rq))
3271 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3272 else
3273 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3274 }
Jens Axboe206dc692006-03-28 13:03:44 +02003275
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003276 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003277 if (blk_queue_nonrot(cfqd->queue))
3278 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3279 else
3280 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003281}
Jens Axboe22e2c502005-06-27 10:55:12 +02003282
3283/*
3284 * Disable idle window if the process thinks too long or seeks so much that
3285 * it doesn't matter
3286 */
3287static void
3288cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3289 struct cfq_io_context *cic)
3290{
Jens Axboe7b679132008-05-30 12:23:07 +02003291 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003292
Jens Axboe08717142008-01-28 11:38:15 +01003293 /*
3294 * Don't idle for async or idle io prio class
3295 */
3296 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003297 return;
3298
Jens Axboec265a7f2008-06-26 13:49:33 +02003299 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003300
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003301 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3302 cfq_mark_cfqq_deep(cfqq);
3303
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003304 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3305 enable_idle = 0;
3306 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003307 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003308 enable_idle = 0;
Shaohua Li383cd722011-07-12 14:24:35 +02003309 else if (sample_valid(cic->ttime.ttime_samples)) {
3310 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003311 enable_idle = 0;
3312 else
3313 enable_idle = 1;
3314 }
3315
Jens Axboe7b679132008-05-30 12:23:07 +02003316 if (old_idle != enable_idle) {
3317 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3318 if (enable_idle)
3319 cfq_mark_cfqq_idle_window(cfqq);
3320 else
3321 cfq_clear_cfqq_idle_window(cfqq);
3322 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003323}
3324
Jens Axboe22e2c502005-06-27 10:55:12 +02003325/*
3326 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3327 * no or if we aren't sure, a 1 will cause a preempt.
3328 */
Jens Axboea6151c32009-10-07 20:02:57 +02003329static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003330cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003331 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003332{
Jens Axboe6d048f52007-04-25 12:44:27 +02003333 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003334
Jens Axboe6d048f52007-04-25 12:44:27 +02003335 cfqq = cfqd->active_queue;
3336 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003337 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003338
Jens Axboe6d048f52007-04-25 12:44:27 +02003339 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003340 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003341
3342 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003343 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003344
Jens Axboe22e2c502005-06-27 10:55:12 +02003345 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003346 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3347 */
3348 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3349 return false;
3350
3351 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003352 * if the new request is sync, but the currently running queue is
3353 * not, let the sync request have priority.
3354 */
Jens Axboe5e705372006-07-13 12:39:25 +02003355 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003356 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003357
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003358 if (new_cfqq->cfqg != cfqq->cfqg)
3359 return false;
3360
3361 if (cfq_slice_used(cfqq))
3362 return true;
3363
3364 /* Allow preemption only if we are idling on sync-noidle tree */
3365 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3366 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3367 new_cfqq->service_tree->count == 2 &&
3368 RB_EMPTY_ROOT(&cfqq->sort_list))
3369 return true;
3370
Jens Axboe374f84a2006-07-23 01:42:19 +02003371 /*
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003372 * So both queues are sync. Let the new request get disk time if
3373 * it's a metadata request and the current queue is doing regular IO.
3374 */
3375 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3376 return true;
3377
3378 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003379 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3380 */
3381 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003382 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003383
Shaohua Lid2d59e12010-11-08 15:01:03 +01003384 /* An idle queue should not be idle now for some reason */
3385 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3386 return true;
3387
Jens Axboe1e3335d2007-02-14 19:59:49 +01003388 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003389 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003390
3391 /*
3392 * if this request is as-good as one we would expect from the
3393 * current cfqq, let it preempt
3394 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003395 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003396 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003397
Jens Axboea6151c32009-10-07 20:02:57 +02003398 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003399}
3400
3401/*
3402 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3403 * let it have half of its nominal slice.
3404 */
3405static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3406{
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003407 struct cfq_queue *old_cfqq = cfqd->active_queue;
3408
Jens Axboe7b679132008-05-30 12:23:07 +02003409 cfq_log_cfqq(cfqd, cfqq, "preempt");
Vivek Goyale5ff0822010-04-26 19:25:11 +02003410 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003411
Jens Axboebf572252006-07-19 20:29:12 +02003412 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003413 * workload type is changed, don't save slice, otherwise preempt
3414 * doesn't happen
3415 */
3416 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3417 cfqq->cfqg->saved_workload_slice = 0;
3418
3419 /*
Jens Axboebf572252006-07-19 20:29:12 +02003420 * Put the new queue at the front of the of the current list,
3421 * so we know that it will be selected next.
3422 */
3423 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003424
3425 cfq_service_tree_add(cfqd, cfqq, 1);
Justin TerAvesteda5e0c2011-03-22 21:26:49 +01003426
Justin TerAvest62a37f62011-03-23 08:25:44 +01003427 cfqq->slice_end = 0;
3428 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003429}
3430
3431/*
Jens Axboe5e705372006-07-13 12:39:25 +02003432 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003433 * something we should do about it
3434 */
3435static void
Jens Axboe5e705372006-07-13 12:39:25 +02003436cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3437 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003438{
Jens Axboe5e705372006-07-13 12:39:25 +02003439 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003440
Aaron Carroll45333d52008-08-26 15:52:36 +02003441 cfqd->rq_queued++;
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003442 if (rq->cmd_flags & REQ_META)
3443 cfqq->meta_pending++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003444
Shaohua Li383cd722011-07-12 14:24:35 +02003445 cfq_update_io_thinktime(cfqd, cfqq, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003446 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003447 cfq_update_idle_window(cfqd, cfqq, cic);
3448
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003449 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003450
3451 if (cfqq == cfqd->active_queue) {
3452 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003453 * Remember that we saw a request from this process, but
3454 * don't start queuing just yet. Otherwise we risk seeing lots
3455 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003456 * and merging. If the request is already larger than a single
3457 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003458 * merging is already done. Ditto for a busy system that
3459 * has other work pending, don't risk delaying until the
3460 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003461 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003462 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003463 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3464 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003465 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003466 cfq_clear_cfqq_wait_request(cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003467 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003468 } else {
Vivek Goyale98ef892010-06-18 10:39:47 -04003469 cfq_blkiocg_update_idle_time_stats(
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003470 &cfqq->cfqg->blkg);
Vivek Goyalbf791932009-12-03 12:59:37 -05003471 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003472 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003473 }
Jens Axboe5e705372006-07-13 12:39:25 +02003474 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003475 /*
3476 * not the active queue - expire current slice if it is
3477 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003478 * has some old slice time left and is of higher priority or
3479 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003480 */
3481 cfq_preempt_queue(cfqd, cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003482 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003483 }
3484}
3485
Jens Axboe165125e2007-07-24 09:28:11 +02003486static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003487{
Jens Axboeb4878f22005-10-20 16:42:29 +02003488 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003489 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003490
Jens Axboe7b679132008-05-30 12:23:07 +02003491 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003492 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493
Jens Axboe30996f42009-10-05 11:03:39 +02003494 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003495 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003496 cfq_add_rq_rb(rq);
Vivek Goyale98ef892010-06-18 10:39:47 -04003497 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
Divyesh Shahcdc11842010-04-08 21:15:10 -07003498 &cfqd->serving_group->blkg, rq_data_dir(rq),
3499 rq_is_sync(rq));
Jens Axboe5e705372006-07-13 12:39:25 +02003500 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501}
3502
Aaron Carroll45333d52008-08-26 15:52:36 +02003503/*
3504 * Update hw_tag based on peak queue depth over 50 samples under
3505 * sufficient load.
3506 */
3507static void cfq_update_hw_tag(struct cfq_data *cfqd)
3508{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003509 struct cfq_queue *cfqq = cfqd->active_queue;
3510
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003511 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3512 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003513
3514 if (cfqd->hw_tag == 1)
3515 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003516
3517 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003518 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003519 return;
3520
Shaohua Li1a1238a2009-10-27 08:46:23 +01003521 /*
3522 * If active queue hasn't enough requests and can idle, cfq might not
3523 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3524 * case
3525 */
3526 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3527 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003528 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003529 return;
3530
Aaron Carroll45333d52008-08-26 15:52:36 +02003531 if (cfqd->hw_tag_samples++ < 50)
3532 return;
3533
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003534 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003535 cfqd->hw_tag = 1;
3536 else
3537 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003538}
3539
Vivek Goyal7667aa02009-12-08 17:52:58 -05003540static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3541{
3542 struct cfq_io_context *cic = cfqd->active_cic;
3543
Justin TerAvest02a8f012011-02-09 14:20:03 +01003544 /* If the queue already has requests, don't wait */
3545 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3546 return false;
3547
Vivek Goyal7667aa02009-12-08 17:52:58 -05003548 /* If there are other queues in the group, don't wait */
3549 if (cfqq->cfqg->nr_cfqq > 1)
3550 return false;
3551
Shaohua Li7700fc42011-07-12 14:24:56 +02003552 /* the only queue in the group, but think time is big */
3553 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3554 return false;
3555
Vivek Goyal7667aa02009-12-08 17:52:58 -05003556 if (cfq_slice_used(cfqq))
3557 return true;
3558
3559 /* if slice left is less than think time, wait busy */
Shaohua Li383cd722011-07-12 14:24:35 +02003560 if (cic && sample_valid(cic->ttime.ttime_samples)
3561 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
Vivek Goyal7667aa02009-12-08 17:52:58 -05003562 return true;
3563
3564 /*
3565 * If think times is less than a jiffy than ttime_mean=0 and above
3566 * will not be true. It might happen that slice has not expired yet
3567 * but will expire soon (4-5 ns) during select_queue(). To cover the
3568 * case where think time is less than a jiffy, mark the queue wait
3569 * busy if only 1 jiffy is left in the slice.
3570 */
3571 if (cfqq->slice_end - jiffies == 1)
3572 return true;
3573
3574 return false;
3575}
3576
Jens Axboe165125e2007-07-24 09:28:11 +02003577static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578{
Jens Axboe5e705372006-07-13 12:39:25 +02003579 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003580 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003581 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003582 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583
Jens Axboeb4878f22005-10-20 16:42:29 +02003584 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02003585 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3586 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587
Aaron Carroll45333d52008-08-26 15:52:36 +02003588 cfq_update_hw_tag(cfqd);
3589
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003590 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02003591 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003592 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003593 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003594 (RQ_CFQG(rq))->dispatched--;
Vivek Goyale98ef892010-06-18 10:39:47 -04003595 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3596 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3597 rq_data_dir(rq), rq_is_sync(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003599 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003600
Vivek Goyal365722b2009-10-03 15:21:27 +02003601 if (sync) {
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003602 struct cfq_rb_root *service_tree;
3603
Shaohua Li383cd722011-07-12 14:24:35 +02003604 RQ_CIC(rq)->ttime.last_end_request = now;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003605
3606 if (cfq_cfqq_on_rr(cfqq))
3607 service_tree = cfqq->service_tree;
3608 else
3609 service_tree = service_tree_for(cfqq->cfqg,
3610 cfqq_prio(cfqq), cfqq_type(cfqq));
3611 service_tree->ttime.last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003612 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3613 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003614 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003615
Shaohua Li7700fc42011-07-12 14:24:56 +02003616#ifdef CONFIG_CFQ_GROUP_IOSCHED
3617 cfqq->cfqg->ttime.last_end_request = now;
3618#endif
3619
Jens Axboecaaa5f92006-06-16 11:23:00 +02003620 /*
3621 * If this is the active queue, check if it needs to be expired,
3622 * or if we want to idle in case it has no pending requests.
3623 */
3624 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003625 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3626
Jens Axboe44f7c162007-01-19 11:51:58 +11003627 if (cfq_cfqq_slice_new(cfqq)) {
3628 cfq_set_prio_slice(cfqd, cfqq);
3629 cfq_clear_cfqq_slice_new(cfqq);
3630 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003631
3632 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003633 * Should we wait for next request to come in before we expire
3634 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003635 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003636 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003637 unsigned long extend_sl = cfqd->cfq_slice_idle;
3638 if (!cfqd->cfq_slice_idle)
3639 extend_sl = cfqd->cfq_group_idle;
3640 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05003641 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003642 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05003643 }
3644
Jens Axboea36e71f2009-04-15 12:15:11 +02003645 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003646 * Idling is not enabled on:
3647 * - expired queues
3648 * - idle-priority queues
3649 * - async queues
3650 * - queues with still some requests queued
3651 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003652 */
Jens Axboe08717142008-01-28 11:38:15 +01003653 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02003654 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003655 else if (sync && cfqq_empty &&
3656 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003657 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003658 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003659 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003660
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003661 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02003662 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003663}
3664
Jens Axboe89850f72006-07-22 16:48:31 +02003665static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003666{
Jens Axboe1b379d82009-08-11 08:26:11 +02003667 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003668 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003669 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003670 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003671
3672 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003673}
3674
Jens Axboe165125e2007-07-24 09:28:11 +02003675static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003676{
3677 struct cfq_data *cfqd = q->elevator->elevator_data;
3678 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003679 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003680 struct cfq_queue *cfqq;
3681
3682 /*
3683 * don't force setup of a queue from here, as a call to may_queue
3684 * does not necessarily imply that a request actually will be queued.
3685 * so just lookup a possibly existing queue, or return 'may queue'
3686 * if that fails
3687 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003688 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003689 if (!cic)
3690 return ELV_MQUEUE_MAY;
3691
Jens Axboeb0b78f82009-04-08 10:56:08 +02003692 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003693 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003694 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003695
Jens Axboe89850f72006-07-22 16:48:31 +02003696 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003697 }
3698
3699 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700}
3701
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702/*
3703 * queue lock held here
3704 */
Jens Axboebb37b942006-12-01 10:42:33 +01003705static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706{
Jens Axboe5e705372006-07-13 12:39:25 +02003707 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708
Jens Axboe5e705372006-07-13 12:39:25 +02003709 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003710 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711
Jens Axboe22e2c502005-06-27 10:55:12 +02003712 BUG_ON(!cfqq->allocated[rw]);
3713 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003714
Jens Axboe5e705372006-07-13 12:39:25 +02003715 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716
Mike Snitzerc1867942011-02-11 11:08:00 +01003717 rq->elevator_private[0] = NULL;
3718 rq->elevator_private[1] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003720 /* Put down rq reference on cfqg */
3721 cfq_put_cfqg(RQ_CFQG(rq));
Mike Snitzerc1867942011-02-11 11:08:00 +01003722 rq->elevator_private[2] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02003723
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724 cfq_put_queue(cfqq);
3725 }
3726}
3727
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003728static struct cfq_queue *
3729cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3730 struct cfq_queue *cfqq)
3731{
3732 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3733 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003734 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003735 cfq_put_queue(cfqq);
3736 return cic_to_cfqq(cic, 1);
3737}
3738
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003739/*
3740 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3741 * was the last process referring to said cfqq.
3742 */
3743static struct cfq_queue *
3744split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3745{
3746 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003747 cfqq->pid = current->pid;
3748 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01003749 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003750 return cfqq;
3751 }
3752
3753 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02003754
3755 cfq_put_cooperator(cfqq);
3756
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003757 cfq_put_queue(cfqq);
3758 return NULL;
3759}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003761 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003763static int
Jens Axboe165125e2007-07-24 09:28:11 +02003764cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765{
3766 struct cfq_data *cfqd = q->elevator->elevator_data;
3767 struct cfq_io_context *cic;
3768 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003769 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003770 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771 unsigned long flags;
3772
3773 might_sleep_if(gfp_mask & __GFP_WAIT);
3774
Jens Axboee2d74ac2006-03-28 08:59:01 +02003775 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003776
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777 spin_lock_irqsave(q->queue_lock, flags);
3778
Jens Axboe22e2c502005-06-27 10:55:12 +02003779 if (!cic)
3780 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003781
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003782new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003783 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003784 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003785 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003786 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003787 } else {
3788 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003789 * If the queue was seeky for too long, break it apart.
3790 */
Shaohua Liae54abe2010-02-05 13:11:45 +01003791 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003792 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3793 cfqq = split_cfqq(cic, cfqq);
3794 if (!cfqq)
3795 goto new_queue;
3796 }
3797
3798 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003799 * Check to see if this queue is scheduled to merge with
3800 * another, closely cooperating queue. The merging of
3801 * queues happens here as it must be done in process context.
3802 * The reference on new_cfqq was taken in merge_cfqqs.
3803 */
3804 if (cfqq->new_cfqq)
3805 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003806 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807
3808 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02003809
Jens Axboe6fae9c22011-03-01 15:04:39 -05003810 cfqq->ref++;
Mike Snitzerc1867942011-02-11 11:08:00 +01003811 rq->elevator_private[0] = cic;
3812 rq->elevator_private[1] = cfqq;
3813 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
Jens Axboe93803e02011-03-07 08:59:06 +01003814 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe5e705372006-07-13 12:39:25 +02003815 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003816
Jens Axboe22e2c502005-06-27 10:55:12 +02003817queue_fail:
Jens Axboe23e018a2009-10-05 08:52:35 +02003818 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003820 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821 return 1;
3822}
3823
David Howells65f27f32006-11-22 14:55:48 +00003824static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003825{
David Howells65f27f32006-11-22 14:55:48 +00003826 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003827 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003828 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003829
Jens Axboe40bb54d2009-04-15 12:11:10 +02003830 spin_lock_irq(q->queue_lock);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003831 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003832 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003833}
3834
3835/*
3836 * Timer running if the active_queue is currently idling inside its time slice
3837 */
3838static void cfq_idle_slice_timer(unsigned long data)
3839{
3840 struct cfq_data *cfqd = (struct cfq_data *) data;
3841 struct cfq_queue *cfqq;
3842 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003843 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003844
Jens Axboe7b679132008-05-30 12:23:07 +02003845 cfq_log(cfqd, "idle timer fired");
3846
Jens Axboe22e2c502005-06-27 10:55:12 +02003847 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3848
Jens Axboefe094d92008-01-31 13:08:54 +01003849 cfqq = cfqd->active_queue;
3850 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003851 timed_out = 0;
3852
Jens Axboe22e2c502005-06-27 10:55:12 +02003853 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003854 * We saw a request before the queue expired, let it through
3855 */
3856 if (cfq_cfqq_must_dispatch(cfqq))
3857 goto out_kick;
3858
3859 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003860 * expired
3861 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003862 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003863 goto expire;
3864
3865 /*
3866 * only expire and reinvoke request handler, if there are
3867 * other queues with pending requests
3868 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003869 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003870 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003871
3872 /*
3873 * not expired and it has a request pending, let it dispatch
3874 */
Jens Axboe75e50982009-04-07 08:56:14 +02003875 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003876 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003877
3878 /*
3879 * Queue depth flag is reset only when the idle didn't succeed
3880 */
3881 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003882 }
3883expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003884 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003885out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003886 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003887out_cont:
3888 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3889}
3890
Jens Axboe3b181522005-06-27 10:56:24 +02003891static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3892{
3893 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003894 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003895}
Jens Axboe22e2c502005-06-27 10:55:12 +02003896
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003897static void cfq_put_async_queues(struct cfq_data *cfqd)
3898{
3899 int i;
3900
3901 for (i = 0; i < IOPRIO_BE_NR; i++) {
3902 if (cfqd->async_cfqq[0][i])
3903 cfq_put_queue(cfqd->async_cfqq[0][i]);
3904 if (cfqd->async_cfqq[1][i])
3905 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003906 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003907
3908 if (cfqd->async_idle_cfqq)
3909 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003910}
3911
Jens Axboeb374d182008-10-31 10:05:07 +01003912static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913{
Jens Axboe22e2c502005-06-27 10:55:12 +02003914 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003915 struct request_queue *q = cfqd->queue;
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003916 bool wait = false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003917
Jens Axboe3b181522005-06-27 10:56:24 +02003918 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003919
Al Virod9ff4182006-03-18 13:51:22 -05003920 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003921
Al Virod9ff4182006-03-18 13:51:22 -05003922 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02003923 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003924
3925 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003926 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3927 struct cfq_io_context,
3928 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003929
3930 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003931 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003932
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003933 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003934 cfq_release_cfq_groups(cfqd);
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003935
3936 /*
3937 * If there are groups which we could not unlink from blkcg list,
3938 * wait for a rcu period for them to be freed.
3939 */
3940 if (cfqd->nr_blkcg_linked_grps)
3941 wait = true;
Jens Axboe15c31be2007-07-10 13:43:25 +02003942
Al Virod9ff4182006-03-18 13:51:22 -05003943 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003944
3945 cfq_shutdown_timer_wq(cfqd);
3946
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003947 spin_lock(&cic_index_lock);
3948 ida_remove(&cic_index_ida, cfqd->cic_index);
3949 spin_unlock(&cic_index_lock);
3950
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003951 /*
3952 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3953 * Do this wait only if there are other unlinked groups out
3954 * there. This can happen if cgroup deletion path claimed the
3955 * responsibility of cleaning up a group before queue cleanup code
3956 * get to the group.
3957 *
3958 * Do not call synchronize_rcu() unconditionally as there are drivers
3959 * which create/delete request queue hundreds of times during scan/boot
3960 * and synchronize_rcu() can take significant time and slow down boot.
3961 */
3962 if (wait)
3963 synchronize_rcu();
Vivek Goyal2abae552011-05-23 10:02:19 +02003964
3965#ifdef CONFIG_CFQ_GROUP_IOSCHED
3966 /* Free up per cpu stats for root group */
3967 free_percpu(cfqd->root_group.blkg.stats_cpu);
3968#endif
Vivek Goyal56edf7d2011-05-19 15:38:22 -04003969 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970}
3971
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003972static int cfq_alloc_cic_index(void)
3973{
3974 int index, error;
3975
3976 do {
3977 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3978 return -ENOMEM;
3979
3980 spin_lock(&cic_index_lock);
3981 error = ida_get_new(&cic_index_ida, &index);
3982 spin_unlock(&cic_index_lock);
3983 if (error && error != -EAGAIN)
3984 return error;
3985 } while (error);
3986
3987 return index;
3988}
3989
Jens Axboe165125e2007-07-24 09:28:11 +02003990static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991{
3992 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003993 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003994 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003995 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04003997 i = cfq_alloc_cic_index();
3998 if (i < 0)
3999 return NULL;
4000
Christoph Lameter94f60302007-07-17 04:03:29 -07004001 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Namhyung Kim15470102011-05-24 10:23:22 +02004002 if (!cfqd) {
4003 spin_lock(&cic_index_lock);
4004 ida_remove(&cic_index_ida, i);
4005 spin_unlock(&cic_index_lock);
Jens Axboebc1c1162006-06-08 08:49:06 +02004006 return NULL;
Namhyung Kim15470102011-05-24 10:23:22 +02004007 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008
Shaohua Li30d7b942011-01-07 08:46:59 +01004009 /*
4010 * Don't need take queue_lock in the routine, since we are
4011 * initializing the ioscheduler, and nobody is using cfqd
4012 */
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004013 cfqd->cic_index = i;
4014
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05004015 /* Init root service tree */
4016 cfqd->grp_service_tree = CFQ_RB_ROOT;
4017
Vivek Goyalcdb16e82009-12-03 12:59:38 -05004018 /* Init root group */
4019 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05004020 for_each_cfqg_st(cfqg, i, j, st)
4021 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05004022 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02004023
Vivek Goyal25bc6b02009-12-03 12:59:43 -05004024 /* Give preference to root group over other groups */
4025 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
4026
Vivek Goyal25fb5162009-12-03 12:59:46 -05004027#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05004028 /*
Vivek Goyal56edf7d2011-05-19 15:38:22 -04004029 * Set root group reference to 2. One reference will be dropped when
4030 * all groups on cfqd->cfqg_list are being deleted during queue exit.
4031 * Other reference will remain there as we don't want to delete this
4032 * group as it is statically allocated and gets destroyed when
4033 * throtl_data goes away.
Vivek Goyalb1c35762009-12-03 12:59:47 -05004034 */
Vivek Goyal56edf7d2011-05-19 15:38:22 -04004035 cfqg->ref = 2;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04004036
4037 if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4038 kfree(cfqg);
4039 kfree(cfqd);
4040 return NULL;
4041 }
4042
Vivek Goyaldcf097b2010-04-22 11:54:52 -04004043 rcu_read_lock();
Vivek Goyal5624a4e2011-05-19 15:38:28 -04004044
Vivek Goyale98ef892010-06-18 10:39:47 -04004045 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4046 (void *)cfqd, 0);
Vivek Goyaldcf097b2010-04-22 11:54:52 -04004047 rcu_read_unlock();
Vivek Goyal56edf7d2011-05-19 15:38:22 -04004048 cfqd->nr_blkcg_linked_grps++;
4049
4050 /* Add group on cfqd->cfqg_list */
4051 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
Vivek Goyal25fb5162009-12-03 12:59:46 -05004052#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02004053 /*
4054 * Not strictly needed (since RB_ROOT just clears the node and we
4055 * zeroed cfqd on alloc), but better be safe in case someone decides
4056 * to add magic to the rb code
4057 */
4058 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4059 cfqd->prio_trees[i] = RB_ROOT;
4060
Jens Axboe6118b702009-06-30 09:34:12 +02004061 /*
4062 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4063 * Grab a permanent reference to it, so that the normal code flow
4064 * will not attempt to free it.
4065 */
4066 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01004067 cfqd->oom_cfqq.ref++;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05004068 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02004069
Al Virod9ff4182006-03-18 13:51:22 -05004070 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071
Linus Torvalds1da177e2005-04-16 15:20:36 -07004072 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073
Jens Axboe22e2c502005-06-27 10:55:12 +02004074 init_timer(&cfqd->idle_slice_timer);
4075 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4076 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4077
Jens Axboe23e018a2009-10-05 08:52:35 +02004078 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02004079
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02004081 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4082 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083 cfqd->cfq_back_max = cfq_back_max;
4084 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02004085 cfqd->cfq_slice[0] = cfq_slice_async;
4086 cfqd->cfq_slice[1] = cfq_slice_sync;
4087 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4088 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004089 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02004090 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01004091 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01004092 /*
4093 * we optimistically start assuming sync ops weren't delayed in last
4094 * second, in order to have larger depth for async operations.
4095 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01004096 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebc1c1162006-06-08 08:49:06 +02004097 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098}
4099
4100static void cfq_slab_kill(void)
4101{
Jens Axboed6de8be2008-05-28 14:46:59 +02004102 /*
4103 * Caller already ensured that pending RCU callbacks are completed,
4104 * so we should have no busy allocations at this point.
4105 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106 if (cfq_pool)
4107 kmem_cache_destroy(cfq_pool);
4108 if (cfq_ioc_pool)
4109 kmem_cache_destroy(cfq_ioc_pool);
4110}
4111
4112static int __init cfq_slab_setup(void)
4113{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07004114 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115 if (!cfq_pool)
4116 goto fail;
4117
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02004118 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119 if (!cfq_ioc_pool)
4120 goto fail;
4121
4122 return 0;
4123fail:
4124 cfq_slab_kill();
4125 return -ENOMEM;
4126}
4127
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128/*
4129 * sysfs parts below -->
4130 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131static ssize_t
4132cfq_var_show(unsigned int var, char *page)
4133{
4134 return sprintf(page, "%d\n", var);
4135}
4136
4137static ssize_t
4138cfq_var_store(unsigned int *var, const char *page, size_t count)
4139{
4140 char *p = (char *) page;
4141
4142 *var = simple_strtoul(p, &p, 10);
4143 return count;
4144}
4145
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004147static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004149 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150 unsigned int __data = __VAR; \
4151 if (__CONV) \
4152 __data = jiffies_to_msecs(__data); \
4153 return cfq_var_show(__data, (page)); \
4154}
4155SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004156SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4157SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05004158SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4159SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004160SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004161SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004162SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4163SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4164SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004165SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004166#undef SHOW_FUNCTION
4167
4168#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004169static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004171 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172 unsigned int __data; \
4173 int ret = cfq_var_store(&__data, (page), count); \
4174 if (__data < (MIN)) \
4175 __data = (MIN); \
4176 else if (__data > (MAX)) \
4177 __data = (MAX); \
4178 if (__CONV) \
4179 *(__PTR) = msecs_to_jiffies(__data); \
4180 else \
4181 *(__PTR) = __data; \
4182 return ret; \
4183}
4184STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004185STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4186 UINT_MAX, 1);
4187STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4188 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004189STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004190STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4191 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004192STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004193STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004194STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4195STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004196STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4197 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004198STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004199#undef STORE_FUNCTION
4200
Al Viroe572ec72006-03-18 22:27:18 -05004201#define CFQ_ATTR(name) \
4202 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004203
Al Viroe572ec72006-03-18 22:27:18 -05004204static struct elv_fs_entry cfq_attrs[] = {
4205 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004206 CFQ_ATTR(fifo_expire_sync),
4207 CFQ_ATTR(fifo_expire_async),
4208 CFQ_ATTR(back_seek_max),
4209 CFQ_ATTR(back_seek_penalty),
4210 CFQ_ATTR(slice_sync),
4211 CFQ_ATTR(slice_async),
4212 CFQ_ATTR(slice_async_rq),
4213 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004214 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004215 CFQ_ATTR(low_latency),
Al Viroe572ec72006-03-18 22:27:18 -05004216 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004217};
4218
Linus Torvalds1da177e2005-04-16 15:20:36 -07004219static struct elevator_type iosched_cfq = {
4220 .ops = {
4221 .elevator_merge_fn = cfq_merge,
4222 .elevator_merged_fn = cfq_merged_request,
4223 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004224 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004225 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004226 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004227 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004228 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229 .elevator_deactivate_req_fn = cfq_deactivate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004231 .elevator_former_req_fn = elv_rb_former_request,
4232 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233 .elevator_set_req_fn = cfq_set_request,
4234 .elevator_put_req_fn = cfq_put_request,
4235 .elevator_may_queue_fn = cfq_may_queue,
4236 .elevator_init_fn = cfq_init_queue,
4237 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02004238 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239 },
Al Viro3d1ab402006-03-18 18:35:43 -05004240 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004241 .elevator_name = "cfq",
4242 .elevator_owner = THIS_MODULE,
4243};
4244
Vivek Goyal3e252062009-12-04 10:36:42 -05004245#ifdef CONFIG_CFQ_GROUP_IOSCHED
4246static struct blkio_policy_type blkio_policy_cfq = {
4247 .ops = {
4248 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4249 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4250 },
Vivek Goyal062a6442010-09-15 17:06:33 -04004251 .plid = BLKIO_POLICY_PROP,
Vivek Goyal3e252062009-12-04 10:36:42 -05004252};
4253#else
4254static struct blkio_policy_type blkio_policy_cfq;
4255#endif
4256
Linus Torvalds1da177e2005-04-16 15:20:36 -07004257static int __init cfq_init(void)
4258{
Jens Axboe22e2c502005-06-27 10:55:12 +02004259 /*
4260 * could be 0 on HZ < 1000 setups
4261 */
4262 if (!cfq_slice_async)
4263 cfq_slice_async = 1;
4264 if (!cfq_slice_idle)
4265 cfq_slice_idle = 1;
4266
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004267#ifdef CONFIG_CFQ_GROUP_IOSCHED
4268 if (!cfq_group_idle)
4269 cfq_group_idle = 1;
4270#else
4271 cfq_group_idle = 0;
4272#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273 if (cfq_slab_setup())
4274 return -ENOMEM;
4275
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004276 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05004277 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004278
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004279 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004280}
4281
4282static void __exit cfq_exit(void)
4283{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07004284 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05004285 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05004287 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02004288 /* ioc_gone's update must be visible before reading ioc_count */
4289 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02004290
4291 /*
4292 * this also protects us from entering cfq_slab_kill() with
4293 * pending RCU callbacks
4294 */
Tejun Heo245b2e72009-06-24 15:13:48 +09004295 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02004296 wait_for_completion(&all_gone);
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004297 ida_destroy(&cic_index_ida);
Christoph Hellwig83521d32005-10-30 15:01:39 -08004298 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004299}
4300
4301module_init(cfq_init);
4302module_exit(cfq_exit);
4303
4304MODULE_AUTHOR("Jens Axboe");
4305MODULE_LICENSE("GPL");
4306MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");