blob: d5bbdcfd0dab5f56fa76f9ca08dc979fe16d1a35 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * CFQ, or complete fairness queueing, disk scheduler.
3 *
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6 *
Jens Axboe0fe23472006-09-04 15:41:16 +02007 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090010#include <linux/slab.h>
Al Viro1cc9be62006-03-18 12:29:52 -050011#include <linux/blkdev.h>
12#include <linux/elevator.h>
Randy Dunlapad5ebd22009-11-11 13:47:45 +010013#include <linux/jiffies.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020015#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020016#include <linux/blktrace_api.h>
Tejun Heo6e736be2011-12-14 00:33:38 +010017#include "blk.h"
Tejun Heo629ed0b2012-04-01 14:38:44 -070018#include "blk-cgroup.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070019
20/*
21 * tunables
22 */
Jens Axboefe094d92008-01-31 13:08:54 +010023/* max queue in one round of service */
Shaohua Liabc3c742010-03-01 09:20:54 +010024static const int cfq_quantum = 8;
Arjan van de Ven64100092006-01-06 09:46:02 +010025static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010026/* maximum backwards seek, in KiB */
27static const int cfq_back_max = 16 * 1024;
28/* penalty of a backwards seek */
29static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010030static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020031static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010032static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020033static int cfq_slice_idle = HZ / 125;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +020034static int cfq_group_idle = HZ / 125;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +010035static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36static const int cfq_hist_divisor = 4;
Jens Axboe22e2c502005-06-27 10:55:12 +020037
Jens Axboed9e76202007-04-20 14:27:50 +020038/*
Jens Axboe08717142008-01-28 11:38:15 +010039 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020040 */
Jens Axboe08717142008-01-28 11:38:15 +010041#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020042
43/*
44 * below this threshold, we consider thinktime immediate
45 */
46#define CFQ_MIN_TT (2)
47
Jens Axboe22e2c502005-06-27 10:55:12 +020048#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020049#define CFQ_HW_QUEUE_MIN (5)
Vivek Goyal25bc6b02009-12-03 12:59:43 -050050#define CFQ_SERVICE_SHIFT 12
Jens Axboe22e2c502005-06-27 10:55:12 +020051
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010052#define CFQQ_SEEK_THR (sector_t)(8 * 100)
Shaohua Lie9ce3352010-03-19 08:03:04 +010053#define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
Corrado Zoccolo41647e72010-02-27 19:45:40 +010054#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +010055#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
Shaohua Liae54abe2010-02-05 13:11:45 +010056
Tejun Heoa612fdd2011-12-14 00:33:41 +010057#define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Christoph Lametere18b8902006-12-06 20:33:20 -080061static struct kmem_cache *cfq_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
Jens Axboe22e2c502005-06-27 10:55:12 +020063#define CFQ_PRIO_LISTS IOPRIO_BE_NR
64#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020065#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
66
Jens Axboe206dc692006-03-28 13:03:44 +020067#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050068#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020069
Tejun Heoc5869802011-12-14 00:33:41 +010070struct cfq_ttime {
71 unsigned long last_end_request;
72
73 unsigned long ttime_total;
74 unsigned long ttime_samples;
75 unsigned long ttime_mean;
76};
77
Jens Axboe22e2c502005-06-27 10:55:12 +020078/*
Jens Axboecc09e292007-04-26 12:53:50 +020079 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
83 */
84struct cfq_rb_root {
85 struct rb_root rb;
86 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010087 unsigned count;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050088 u64 min_vdisktime;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +020089 struct cfq_ttime ttime;
Jens Axboecc09e292007-04-26 12:53:50 +020090};
Shaohua Lif5f2b6c2011-07-12 14:24:55 +020091#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
92 .ttime = {.last_end_request = jiffies,},}
Jens Axboecc09e292007-04-26 12:53:50 +020093
94/*
Jens Axboe6118b702009-06-30 09:34:12 +020095 * Per process-grouping structure
96 */
97struct cfq_queue {
98 /* reference count */
Shaohua Li30d7b942011-01-07 08:46:59 +010099 int ref;
Jens Axboe6118b702009-06-30 09:34:12 +0200100 /* various state flags, see below */
101 unsigned int flags;
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
117 int queued[2];
118 /* currently allocated requests */
119 int allocated[2];
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
122
Vivek Goyaldae739e2009-12-03 12:59:45 -0500123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500125 unsigned int allocated_slice;
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100126 unsigned int slice_dispatch;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
Jens Axboe6118b702009-06-30 09:34:12 +0200129 unsigned long slice_end;
130 long slice_resid;
Jens Axboe6118b702009-06-30 09:34:12 +0200131
Christoph Hellwig65299a32011-08-23 14:50:29 +0200132 /* pending priority requests */
133 int prio_pending;
Jens Axboe6118b702009-06-30 09:34:12 +0200134 /* number of requests that are on the dispatch list or inside driver */
135 int dispatched;
136
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
Justin TerAvest4aede842011-07-12 08:31:45 +0200139 unsigned short ioprio_class;
Jens Axboe6118b702009-06-30 09:34:12 +0200140
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100141 pid_t pid;
142
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +0100143 u32 seek_history;
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400144 sector_t last_request_pos;
145
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100146 struct cfq_rb_root *service_tree;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400147 struct cfq_queue *new_cfqq;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500148 struct cfq_group *cfqg;
Vivek Goyalc4e78932010-08-23 12:25:03 +0200149 /* Number of sectors dispatched from queue in single dispatch round */
150 unsigned long nr_sectors;
Jens Axboe6118b702009-06-30 09:34:12 +0200151};
152
153/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100154 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100155 * IDLE is handled separately, so it has negative index
156 */
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400157enum wl_class_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100158 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500159 RT_WORKLOAD = 1,
160 IDLE_WORKLOAD = 2,
Vivek Goyalb4627322010-10-22 09:48:43 +0200161 CFQ_PRIO_NR,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100162};
163
164/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100165 * Second index in the service_trees.
166 */
167enum wl_type_t {
168 ASYNC_WORKLOAD = 0,
169 SYNC_NOIDLE_WORKLOAD = 1,
170 SYNC_WORKLOAD = 2
171};
172
Tejun Heo155fead2012-04-01 14:38:44 -0700173struct cfqg_stats {
174#ifdef CONFIG_CFQ_GROUP_IOSCHED
175 /* total bytes transferred */
176 struct blkg_rwstat service_bytes;
177 /* total IOs serviced, post merge */
178 struct blkg_rwstat serviced;
179 /* number of ios merged */
180 struct blkg_rwstat merged;
181 /* total time spent on device in ns, may not be accurate w/ queueing */
182 struct blkg_rwstat service_time;
183 /* total time spent waiting in scheduler queue in ns */
184 struct blkg_rwstat wait_time;
185 /* number of IOs queued up */
186 struct blkg_rwstat queued;
187 /* total sectors transferred */
188 struct blkg_stat sectors;
189 /* total disk time and nr sectors dispatched by this group */
190 struct blkg_stat time;
191#ifdef CONFIG_DEBUG_BLK_CGROUP
192 /* time not charged to this cgroup */
193 struct blkg_stat unaccounted_time;
194 /* sum of number of ios queued across all samples */
195 struct blkg_stat avg_queue_size_sum;
196 /* count of samples taken for average */
197 struct blkg_stat avg_queue_size_samples;
198 /* how many times this group has been removed from service tree */
199 struct blkg_stat dequeue;
200 /* total time spent waiting for it to be assigned a timeslice. */
201 struct blkg_stat group_wait_time;
Tejun Heo3c798392012-04-16 13:57:25 -0700202 /* time spent idling for this blkcg_gq */
Tejun Heo155fead2012-04-01 14:38:44 -0700203 struct blkg_stat idle_time;
204 /* total time with empty current active q with other requests queued */
205 struct blkg_stat empty_time;
206 /* fields after this shouldn't be cleared on stat reset */
207 uint64_t start_group_wait_time;
208 uint64_t start_idle_time;
209 uint64_t start_empty_time;
210 uint16_t flags;
211#endif /* CONFIG_DEBUG_BLK_CGROUP */
212#endif /* CONFIG_CFQ_GROUP_IOSCHED */
213};
214
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500215/* This is per cgroup per device grouping structure */
216struct cfq_group {
Tejun Heof95a04a2012-04-16 13:57:26 -0700217 /* must be the first member */
218 struct blkg_policy_data pd;
219
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500220 /* group service_tree member */
221 struct rb_node rb_node;
222
223 /* group service_tree key */
224 u64 vdisktime;
Tejun Heoe71357e2013-01-09 08:05:10 -0800225
226 /*
Tejun Heo7918ffb2013-01-09 08:05:11 -0800227 * The number of active cfqgs and sum of their weights under this
228 * cfqg. This covers this cfqg's leaf_weight and all children's
229 * weights, but does not cover weights of further descendants.
230 *
231 * If a cfqg is on the service tree, it's active. An active cfqg
232 * also activates its parent and contributes to the children_weight
233 * of the parent.
234 */
235 int nr_active;
236 unsigned int children_weight;
237
238 /*
Tejun Heo1d3650f2013-01-09 08:05:11 -0800239 * vfraction is the fraction of vdisktime that the tasks in this
240 * cfqg are entitled to. This is determined by compounding the
241 * ratios walking up from this cfqg to the root.
242 *
243 * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
244 * vfractions on a service tree is approximately 1. The sum may
245 * deviate a bit due to rounding errors and fluctuations caused by
246 * cfqgs entering and leaving the service tree.
247 */
248 unsigned int vfraction;
249
250 /*
Tejun Heoe71357e2013-01-09 08:05:10 -0800251 * There are two weights - (internal) weight is the weight of this
252 * cfqg against the sibling cfqgs. leaf_weight is the wight of
253 * this cfqg against the child cfqgs. For the root cfqg, both
254 * weights are kept in sync for backward compatibility.
255 */
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500256 unsigned int weight;
Justin TerAvest8184f932011-03-17 16:12:36 +0100257 unsigned int new_weight;
Tejun Heo3381cb82012-04-01 14:38:44 -0700258 unsigned int dev_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500259
Tejun Heoe71357e2013-01-09 08:05:10 -0800260 unsigned int leaf_weight;
261 unsigned int new_leaf_weight;
262 unsigned int dev_leaf_weight;
263
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500264 /* number of cfqq currently on this group */
265 int nr_cfqq;
266
Jens Axboe22e2c502005-06-27 10:55:12 +0200267 /*
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200268 * Per group busy queues average. Useful for workload slice calc. We
Vivek Goyalb4627322010-10-22 09:48:43 +0200269 * create the array for each prio class but at run time it is used
270 * only for RT and BE class and slot for IDLE class remains unused.
271 * This is primarily done to avoid confusion and a gcc warning.
272 */
273 unsigned int busy_queues_avg[CFQ_PRIO_NR];
274 /*
275 * rr lists of queues with requests. We maintain service trees for
276 * RT and BE classes. These trees are subdivided in subclasses
277 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
278 * class there is no subclassification and all the cfq queues go on
279 * a single tree service_tree_idle.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100280 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200281 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100282 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100283 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500284
Vivek Goyal4d2ceea2012-10-03 16:56:57 -0400285 unsigned long saved_wl_slice;
286 enum wl_type_t saved_wl_type;
287 enum wl_class_t saved_wl_class;
Tejun Heo4eef3042012-03-05 13:15:18 -0800288
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200289 /* number of requests that are on the dispatch list or inside driver */
290 int dispatched;
Shaohua Li7700fc42011-07-12 14:24:56 +0200291 struct cfq_ttime ttime;
Tejun Heo0b399202013-01-09 08:05:13 -0800292 struct cfqg_stats stats; /* stats for this cfqg */
293 struct cfqg_stats dead_stats; /* stats pushed from dead children */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500294};
295
Tejun Heoc5869802011-12-14 00:33:41 +0100296struct cfq_io_cq {
297 struct io_cq icq; /* must be the first member */
298 struct cfq_queue *cfqq[2];
299 struct cfq_ttime ttime;
Tejun Heo598971b2012-03-19 15:10:58 -0700300 int ioprio; /* the current ioprio */
301#ifdef CONFIG_CFQ_GROUP_IOSCHED
302 uint64_t blkcg_id; /* the current blkcg ID */
303#endif
Tejun Heoc5869802011-12-14 00:33:41 +0100304};
305
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500306/*
307 * Per block device queue structure
308 */
309struct cfq_data {
310 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500311 /* Root service tree for cfq_groups */
312 struct cfq_rb_root grp_service_tree;
Tejun Heof51b8022012-03-05 13:15:05 -0800313 struct cfq_group *root_group;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500314
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100315 /*
316 * The priority currently being served
317 */
Vivek Goyal4d2ceea2012-10-03 16:56:57 -0400318 enum wl_class_t serving_wl_class;
319 enum wl_type_t serving_wl_type;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100320 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500321 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200322
323 /*
324 * Each priority tree is sorted by next_request position. These
325 * trees are used when determining if two or more queues are
326 * interleaving requests (see cfq_close_cooperator).
327 */
328 struct rb_root prio_trees[CFQ_PRIO_LISTS];
329
Jens Axboe22e2c502005-06-27 10:55:12 +0200330 unsigned int busy_queues;
Shaohua Lief8a41d2011-03-07 09:26:29 +0100331 unsigned int busy_sync_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +0200332
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100333 int rq_in_driver;
334 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200335
336 /*
337 * queue-depth detection
338 */
339 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200340 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100341 /*
342 * hw_tag can be
343 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
344 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
345 * 0 => no NCQ
346 */
347 int hw_tag_est_depth;
348 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200349
350 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200351 * idle window management
352 */
353 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200354 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200355
356 struct cfq_queue *active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +0100357 struct cfq_io_cq *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200358
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200359 /*
360 * async queue for each priority case
361 */
362 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
363 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200364
Jens Axboe6d048f52007-04-25 12:44:27 +0200365 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 /*
368 * tunables, see top of file
369 */
370 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200371 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 unsigned int cfq_back_penalty;
373 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200374 unsigned int cfq_slice[2];
375 unsigned int cfq_slice_async_rq;
376 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200377 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200378 unsigned int cfq_latency;
Tao Ma5bf14c02012-04-01 14:33:39 -0700379 unsigned int cfq_target_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500380
Jens Axboe6118b702009-06-30 09:34:12 +0200381 /*
382 * Fallback dummy cfqq for extreme OOM conditions
383 */
384 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200385
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100386 unsigned long last_delayed_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387};
388
Vivek Goyal25fb5162009-12-03 12:59:46 -0500389static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
390
Vivek Goyal34b98d02012-10-03 16:56:58 -0400391static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400392 enum wl_class_t class,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500393 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100394{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500395 if (!cfqg)
396 return NULL;
397
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400398 if (class == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500399 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100400
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400401 return &cfqg->service_trees[class][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100402}
403
Jens Axboe3b181522005-06-27 10:56:24 +0200404enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100405 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
406 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200407 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100408 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100409 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
410 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
411 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100412 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200413 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400414 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100415 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100416 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500417 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200418};
419
420#define CFQ_CFQQ_FNS(name) \
421static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
422{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100423 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200424} \
425static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
426{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100427 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200428} \
429static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
430{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100431 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200432}
433
434CFQ_CFQQ_FNS(on_rr);
435CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200436CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200437CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200438CFQ_CFQQ_FNS(fifo_expire);
439CFQ_CFQQ_FNS(idle_window);
440CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100441CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200442CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200443CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100444CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100445CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500446CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200447#undef CFQ_CFQQ_FNS
448
Tejun Heof95a04a2012-04-16 13:57:26 -0700449static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
450{
451 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
452}
453
Tejun Heof95a04a2012-04-16 13:57:26 -0700454static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
455{
456 return pd_to_blkg(&cfqg->pd);
457}
458
Tejun Heo629ed0b2012-04-01 14:38:44 -0700459#if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700460
Tejun Heo155fead2012-04-01 14:38:44 -0700461/* cfqg stats flags */
462enum cfqg_stats_flags {
463 CFQG_stats_waiting = 0,
464 CFQG_stats_idling,
465 CFQG_stats_empty,
Tejun Heo629ed0b2012-04-01 14:38:44 -0700466};
467
Tejun Heo155fead2012-04-01 14:38:44 -0700468#define CFQG_FLAG_FNS(name) \
469static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700470{ \
Tejun Heo155fead2012-04-01 14:38:44 -0700471 stats->flags |= (1 << CFQG_stats_##name); \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700472} \
Tejun Heo155fead2012-04-01 14:38:44 -0700473static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700474{ \
Tejun Heo155fead2012-04-01 14:38:44 -0700475 stats->flags &= ~(1 << CFQG_stats_##name); \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700476} \
Tejun Heo155fead2012-04-01 14:38:44 -0700477static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700478{ \
Tejun Heo155fead2012-04-01 14:38:44 -0700479 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700480} \
481
Tejun Heo155fead2012-04-01 14:38:44 -0700482CFQG_FLAG_FNS(waiting)
483CFQG_FLAG_FNS(idling)
484CFQG_FLAG_FNS(empty)
485#undef CFQG_FLAG_FNS
Tejun Heo629ed0b2012-04-01 14:38:44 -0700486
487/* This should be called with the queue_lock held. */
Tejun Heo155fead2012-04-01 14:38:44 -0700488static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700489{
490 unsigned long long now;
491
Tejun Heo155fead2012-04-01 14:38:44 -0700492 if (!cfqg_stats_waiting(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700493 return;
494
495 now = sched_clock();
496 if (time_after64(now, stats->start_group_wait_time))
497 blkg_stat_add(&stats->group_wait_time,
498 now - stats->start_group_wait_time);
Tejun Heo155fead2012-04-01 14:38:44 -0700499 cfqg_stats_clear_waiting(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700500}
501
502/* This should be called with the queue_lock held. */
Tejun Heo155fead2012-04-01 14:38:44 -0700503static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
504 struct cfq_group *curr_cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700505{
Tejun Heo155fead2012-04-01 14:38:44 -0700506 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700507
Tejun Heo155fead2012-04-01 14:38:44 -0700508 if (cfqg_stats_waiting(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700509 return;
Tejun Heo155fead2012-04-01 14:38:44 -0700510 if (cfqg == curr_cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700511 return;
Tejun Heo155fead2012-04-01 14:38:44 -0700512 stats->start_group_wait_time = sched_clock();
513 cfqg_stats_mark_waiting(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700514}
515
516/* This should be called with the queue_lock held. */
Tejun Heo155fead2012-04-01 14:38:44 -0700517static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700518{
519 unsigned long long now;
520
Tejun Heo155fead2012-04-01 14:38:44 -0700521 if (!cfqg_stats_empty(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700522 return;
523
524 now = sched_clock();
525 if (time_after64(now, stats->start_empty_time))
526 blkg_stat_add(&stats->empty_time,
527 now - stats->start_empty_time);
Tejun Heo155fead2012-04-01 14:38:44 -0700528 cfqg_stats_clear_empty(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700529}
530
Tejun Heo155fead2012-04-01 14:38:44 -0700531static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700532{
Tejun Heo155fead2012-04-01 14:38:44 -0700533 blkg_stat_add(&cfqg->stats.dequeue, 1);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700534}
535
Tejun Heo155fead2012-04-01 14:38:44 -0700536static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700537{
Tejun Heo155fead2012-04-01 14:38:44 -0700538 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700539
Tejun Heo4d5e80a2013-01-09 08:05:12 -0800540 if (blkg_rwstat_total(&stats->queued))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700541 return;
542
543 /*
544 * group is already marked empty. This can happen if cfqq got new
545 * request in parent group and moved to this group while being added
546 * to service tree. Just ignore the event and move on.
547 */
Tejun Heo155fead2012-04-01 14:38:44 -0700548 if (cfqg_stats_empty(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700549 return;
550
551 stats->start_empty_time = sched_clock();
Tejun Heo155fead2012-04-01 14:38:44 -0700552 cfqg_stats_mark_empty(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700553}
554
Tejun Heo155fead2012-04-01 14:38:44 -0700555static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700556{
Tejun Heo155fead2012-04-01 14:38:44 -0700557 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700558
Tejun Heo155fead2012-04-01 14:38:44 -0700559 if (cfqg_stats_idling(stats)) {
Tejun Heo629ed0b2012-04-01 14:38:44 -0700560 unsigned long long now = sched_clock();
561
562 if (time_after64(now, stats->start_idle_time))
563 blkg_stat_add(&stats->idle_time,
564 now - stats->start_idle_time);
Tejun Heo155fead2012-04-01 14:38:44 -0700565 cfqg_stats_clear_idling(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700566 }
567}
568
Tejun Heo155fead2012-04-01 14:38:44 -0700569static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700570{
Tejun Heo155fead2012-04-01 14:38:44 -0700571 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700572
Tejun Heo155fead2012-04-01 14:38:44 -0700573 BUG_ON(cfqg_stats_idling(stats));
Tejun Heo629ed0b2012-04-01 14:38:44 -0700574
575 stats->start_idle_time = sched_clock();
Tejun Heo155fead2012-04-01 14:38:44 -0700576 cfqg_stats_mark_idling(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700577}
578
Tejun Heo155fead2012-04-01 14:38:44 -0700579static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700580{
Tejun Heo155fead2012-04-01 14:38:44 -0700581 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700582
583 blkg_stat_add(&stats->avg_queue_size_sum,
Tejun Heo4d5e80a2013-01-09 08:05:12 -0800584 blkg_rwstat_total(&stats->queued));
Tejun Heo629ed0b2012-04-01 14:38:44 -0700585 blkg_stat_add(&stats->avg_queue_size_samples, 1);
Tejun Heo155fead2012-04-01 14:38:44 -0700586 cfqg_stats_update_group_wait_time(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700587}
588
589#else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
590
Tejun Heof48ec1d2012-04-13 13:11:25 -0700591static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
592static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
593static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
594static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
595static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
596static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
597static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
Tejun Heo629ed0b2012-04-01 14:38:44 -0700598
599#endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
600
601#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo2ce4d502012-04-01 14:38:43 -0700602
Tejun Heoffea73f2012-06-04 10:02:29 +0200603static struct blkcg_policy blkcg_policy_cfq;
604
605static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
606{
607 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
608}
609
Tejun Heod02f7aa2013-01-09 08:05:11 -0800610static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
Tejun Heo7918ffb2013-01-09 08:05:11 -0800611{
Tejun Heod02f7aa2013-01-09 08:05:11 -0800612 struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
Tejun Heo7918ffb2013-01-09 08:05:11 -0800613
Tejun Heod02f7aa2013-01-09 08:05:11 -0800614 return pblkg ? blkg_to_cfqg(pblkg) : NULL;
Tejun Heo7918ffb2013-01-09 08:05:11 -0800615}
616
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100617static inline void cfqg_get(struct cfq_group *cfqg)
618{
619 return blkg_get(cfqg_to_blkg(cfqg));
620}
621
622static inline void cfqg_put(struct cfq_group *cfqg)
623{
624 return blkg_put(cfqg_to_blkg(cfqg));
625}
626
Tejun Heo54e7ed12012-04-16 13:57:23 -0700627#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
628 char __pbuf[128]; \
629 \
630 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
Vivek Goyalb226e5c2012-10-03 16:57:01 -0400631 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
632 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
633 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
Tejun Heo54e7ed12012-04-16 13:57:23 -0700634 __pbuf, ##args); \
635} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500636
Tejun Heo54e7ed12012-04-16 13:57:23 -0700637#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
638 char __pbuf[128]; \
639 \
640 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
641 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
642} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500643
Tejun Heo155fead2012-04-01 14:38:44 -0700644static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
645 struct cfq_group *curr_cfqg, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700646{
Tejun Heo155fead2012-04-01 14:38:44 -0700647 blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
648 cfqg_stats_end_empty_time(&cfqg->stats);
649 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700650}
651
Tejun Heo155fead2012-04-01 14:38:44 -0700652static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
653 unsigned long time, unsigned long unaccounted_time)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700654{
Tejun Heo155fead2012-04-01 14:38:44 -0700655 blkg_stat_add(&cfqg->stats.time, time);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700656#ifdef CONFIG_DEBUG_BLK_CGROUP
Tejun Heo155fead2012-04-01 14:38:44 -0700657 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700658#endif
Tejun Heo2ce4d502012-04-01 14:38:43 -0700659}
660
Tejun Heo155fead2012-04-01 14:38:44 -0700661static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700662{
Tejun Heo155fead2012-04-01 14:38:44 -0700663 blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700664}
665
Tejun Heo155fead2012-04-01 14:38:44 -0700666static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700667{
Tejun Heo155fead2012-04-01 14:38:44 -0700668 blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700669}
670
Tejun Heo155fead2012-04-01 14:38:44 -0700671static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
672 uint64_t bytes, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700673{
Tejun Heo155fead2012-04-01 14:38:44 -0700674 blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
675 blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
676 blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700677}
678
Tejun Heo155fead2012-04-01 14:38:44 -0700679static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
680 uint64_t start_time, uint64_t io_start_time, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700681{
Tejun Heo155fead2012-04-01 14:38:44 -0700682 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700683 unsigned long long now = sched_clock();
Tejun Heo629ed0b2012-04-01 14:38:44 -0700684
685 if (time_after64(now, io_start_time))
686 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
687 if (time_after64(io_start_time, start_time))
688 blkg_rwstat_add(&stats->wait_time, rw,
689 io_start_time - start_time);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700690}
691
Tejun Heo689665a2013-01-09 08:05:13 -0800692/* @stats = 0 */
693static void cfqg_stats_reset(struct cfqg_stats *stats)
Tejun Heo155fead2012-04-01 14:38:44 -0700694{
Tejun Heo155fead2012-04-01 14:38:44 -0700695 /* queued stats shouldn't be cleared */
696 blkg_rwstat_reset(&stats->service_bytes);
697 blkg_rwstat_reset(&stats->serviced);
698 blkg_rwstat_reset(&stats->merged);
699 blkg_rwstat_reset(&stats->service_time);
700 blkg_rwstat_reset(&stats->wait_time);
701 blkg_stat_reset(&stats->time);
702#ifdef CONFIG_DEBUG_BLK_CGROUP
703 blkg_stat_reset(&stats->unaccounted_time);
704 blkg_stat_reset(&stats->avg_queue_size_sum);
705 blkg_stat_reset(&stats->avg_queue_size_samples);
706 blkg_stat_reset(&stats->dequeue);
707 blkg_stat_reset(&stats->group_wait_time);
708 blkg_stat_reset(&stats->idle_time);
709 blkg_stat_reset(&stats->empty_time);
710#endif
711}
712
Tejun Heo0b399202013-01-09 08:05:13 -0800713/* @to += @from */
714static void cfqg_stats_merge(struct cfqg_stats *to, struct cfqg_stats *from)
715{
716 /* queued stats shouldn't be cleared */
717 blkg_rwstat_merge(&to->service_bytes, &from->service_bytes);
718 blkg_rwstat_merge(&to->serviced, &from->serviced);
719 blkg_rwstat_merge(&to->merged, &from->merged);
720 blkg_rwstat_merge(&to->service_time, &from->service_time);
721 blkg_rwstat_merge(&to->wait_time, &from->wait_time);
722 blkg_stat_merge(&from->time, &from->time);
723#ifdef CONFIG_DEBUG_BLK_CGROUP
724 blkg_stat_merge(&to->unaccounted_time, &from->unaccounted_time);
725 blkg_stat_merge(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
726 blkg_stat_merge(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
727 blkg_stat_merge(&to->dequeue, &from->dequeue);
728 blkg_stat_merge(&to->group_wait_time, &from->group_wait_time);
729 blkg_stat_merge(&to->idle_time, &from->idle_time);
730 blkg_stat_merge(&to->empty_time, &from->empty_time);
731#endif
732}
733
734/*
735 * Transfer @cfqg's stats to its parent's dead_stats so that the ancestors'
736 * recursive stats can still account for the amount used by this cfqg after
737 * it's gone.
738 */
739static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
740{
741 struct cfq_group *parent = cfqg_parent(cfqg);
742
743 lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
744
745 if (unlikely(!parent))
746 return;
747
748 cfqg_stats_merge(&parent->dead_stats, &cfqg->stats);
749 cfqg_stats_merge(&parent->dead_stats, &cfqg->dead_stats);
750 cfqg_stats_reset(&cfqg->stats);
751 cfqg_stats_reset(&cfqg->dead_stats);
752}
753
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100754#else /* CONFIG_CFQ_GROUP_IOSCHED */
755
Tejun Heod02f7aa2013-01-09 08:05:11 -0800756static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100757static inline void cfqg_get(struct cfq_group *cfqg) { }
758static inline void cfqg_put(struct cfq_group *cfqg) { }
759
Jens Axboe7b679132008-05-30 12:23:07 +0200760#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
Vivek Goyalb226e5c2012-10-03 16:57:01 -0400761 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
762 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
763 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
764 ##args)
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200765#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100766
Tejun Heo155fead2012-04-01 14:38:44 -0700767static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
768 struct cfq_group *curr_cfqg, int rw) { }
769static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
770 unsigned long time, unsigned long unaccounted_time) { }
771static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
772static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
773static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
774 uint64_t bytes, int rw) { }
775static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
776 uint64_t start_time, uint64_t io_start_time, int rw) { }
Tejun Heo2ce4d502012-04-01 14:38:43 -0700777
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100778#endif /* CONFIG_CFQ_GROUP_IOSCHED */
779
Jens Axboe7b679132008-05-30 12:23:07 +0200780#define cfq_log(cfqd, fmt, args...) \
781 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
782
Vivek Goyal615f0252009-12-03 12:59:39 -0500783/* Traverses through cfq group service trees */
784#define for_each_cfqg_st(cfqg, i, j, st) \
785 for (i = 0; i <= IDLE_WORKLOAD; i++) \
786 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
787 : &cfqg->service_tree_idle; \
788 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
789 (i == IDLE_WORKLOAD && j == 0); \
790 j++, st = i < IDLE_WORKLOAD ? \
791 &cfqg->service_trees[i][j]: NULL) \
792
Shaohua Lif5f2b6c2011-07-12 14:24:55 +0200793static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
794 struct cfq_ttime *ttime, bool group_idle)
795{
796 unsigned long slice;
797 if (!sample_valid(ttime->ttime_samples))
798 return false;
799 if (group_idle)
800 slice = cfqd->cfq_group_idle;
801 else
802 slice = cfqd->cfq_slice_idle;
803 return ttime->ttime_mean > slice;
804}
Vivek Goyal615f0252009-12-03 12:59:39 -0500805
Vivek Goyal02b35082010-08-23 12:23:53 +0200806static inline bool iops_mode(struct cfq_data *cfqd)
807{
808 /*
809 * If we are not idling on queues and it is a NCQ drive, parallel
810 * execution of requests is on and measuring time is not possible
811 * in most of the cases until and unless we drive shallower queue
812 * depths and that becomes a performance bottleneck. In such cases
813 * switch to start providing fairness in terms of number of IOs.
814 */
815 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
816 return true;
817 else
818 return false;
819}
820
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400821static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100822{
823 if (cfq_class_idle(cfqq))
824 return IDLE_WORKLOAD;
825 if (cfq_class_rt(cfqq))
826 return RT_WORKLOAD;
827 return BE_WORKLOAD;
828}
829
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100830
831static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
832{
833 if (!cfq_cfqq_sync(cfqq))
834 return ASYNC_WORKLOAD;
835 if (!cfq_cfqq_idle_window(cfqq))
836 return SYNC_NOIDLE_WORKLOAD;
837 return SYNC_WORKLOAD;
838}
839
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400840static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500841 struct cfq_data *cfqd,
842 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100843{
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400844 if (wl_class == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500845 return cfqg->service_tree_idle.count;
846
Vivek Goyal34b98d02012-10-03 16:56:58 -0400847 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
848 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
849 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100850}
851
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500852static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
853 struct cfq_group *cfqg)
854{
Vivek Goyal34b98d02012-10-03 16:56:58 -0400855 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
856 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500857}
858
Jens Axboe165125e2007-07-24 09:28:11 +0200859static void cfq_dispatch_insert(struct request_queue *, struct request *);
Tejun Heo4f85cb92012-03-05 13:15:28 -0800860static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
Tejun Heoabede6d2012-03-19 15:10:57 -0700861 struct cfq_io_cq *cic, struct bio *bio,
Tejun Heo4f85cb92012-03-05 13:15:28 -0800862 gfp_t gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200863
Tejun Heoc5869802011-12-14 00:33:41 +0100864static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
865{
866 /* cic->icq is the first member, %NULL will convert to %NULL */
867 return container_of(icq, struct cfq_io_cq, icq);
868}
869
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100870static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
871 struct io_context *ioc)
872{
873 if (ioc)
874 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
875 return NULL;
876}
877
Tejun Heoc5869802011-12-14 00:33:41 +0100878static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200879{
Jens Axboea6151c32009-10-07 20:02:57 +0200880 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200881}
882
Tejun Heoc5869802011-12-14 00:33:41 +0100883static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
884 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200885{
Jens Axboea6151c32009-10-07 20:02:57 +0200886 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200887}
888
Tejun Heoc5869802011-12-14 00:33:41 +0100889static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400890{
Tejun Heoc5869802011-12-14 00:33:41 +0100891 return cic->icq.q->elevator->elevator_data;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400892}
893
Vasily Tarasov91fac312007-04-25 12:29:51 +0200894/*
895 * We regard a request as SYNC, if it's either a read or has the SYNC bit
896 * set (in which case it could also be direct WRITE).
897 */
Jens Axboea6151c32009-10-07 20:02:57 +0200898static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200899{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200900 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200901}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700904 * scheduler run of queue, if there are requests pending and no one in the
905 * driver that will restart queueing
906 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200907static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700908{
Jens Axboe7b679132008-05-30 12:23:07 +0200909 if (cfqd->busy_queues) {
910 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200911 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200912 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700913}
914
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100916 * Scale schedule slice based on io priority. Use the sync time slice only
917 * if a queue is marked sync and has sync io queued. A sync queue with async
918 * io only, should not get full sync slice length.
919 */
Jens Axboea6151c32009-10-07 20:02:57 +0200920static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200921 unsigned short prio)
922{
923 const int base_slice = cfqd->cfq_slice[sync];
924
925 WARN_ON(prio >= IOPRIO_BE_NR);
926
927 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
928}
929
Jens Axboe44f7c162007-01-19 11:51:58 +1100930static inline int
931cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
932{
Jens Axboed9e76202007-04-20 14:27:50 +0200933 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100934}
935
Tejun Heo1d3650f2013-01-09 08:05:11 -0800936/**
937 * cfqg_scale_charge - scale disk time charge according to cfqg weight
938 * @charge: disk time being charged
939 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
940 *
941 * Scale @charge according to @vfraction, which is in range (0, 1]. The
942 * scaling is inversely proportional.
943 *
944 * scaled = charge / vfraction
945 *
946 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
947 */
948static inline u64 cfqg_scale_charge(unsigned long charge,
949 unsigned int vfraction)
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500950{
Tejun Heo1d3650f2013-01-09 08:05:11 -0800951 u64 c = charge << CFQ_SERVICE_SHIFT; /* make it fixed point */
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500952
Tejun Heo1d3650f2013-01-09 08:05:11 -0800953 /* charge / vfraction */
954 c <<= CFQ_SERVICE_SHIFT;
955 do_div(c, vfraction);
956 return c;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500957}
958
959static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
960{
961 s64 delta = (s64)(vdisktime - min_vdisktime);
962 if (delta > 0)
963 min_vdisktime = vdisktime;
964
965 return min_vdisktime;
966}
967
968static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
969{
970 s64 delta = (s64)(vdisktime - min_vdisktime);
971 if (delta < 0)
972 min_vdisktime = vdisktime;
973
974 return min_vdisktime;
975}
976
977static void update_min_vdisktime(struct cfq_rb_root *st)
978{
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500979 struct cfq_group *cfqg;
980
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500981 if (st->left) {
982 cfqg = rb_entry_cfqg(st->left);
Gui Jianfenga6032712011-03-07 09:28:09 +0100983 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
984 cfqg->vdisktime);
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500985 }
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500986}
987
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100988/*
989 * get averaged number of queues of RT/BE priority.
990 * average is updated, with a formula that gives more weight to higher numbers,
991 * to quickly follows sudden increases and decrease slowly
992 */
993
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500994static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
995 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100996{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100997 unsigned min_q, max_q;
998 unsigned mult = cfq_hist_divisor - 1;
999 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001000 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001001
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001002 min_q = min(cfqg->busy_queues_avg[rt], busy);
1003 max_q = max(cfqg->busy_queues_avg[rt], busy);
1004 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001005 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001006 return cfqg->busy_queues_avg[rt];
1007}
1008
1009static inline unsigned
1010cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1011{
Tejun Heo41cad6a2013-01-09 08:05:11 -08001012 return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001013}
1014
Shaohua Lic553f8e2011-01-14 08:41:03 +01001015static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -07001016cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +11001017{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001018 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1019 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001020 /*
1021 * interested queues (we consider only the ones with the same
1022 * priority class in the cfq group)
1023 */
1024 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1025 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001026 unsigned sync_slice = cfqd->cfq_slice[1];
1027 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001028 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1029
1030 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001031 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1032 /* scale low_slice according to IO priority
1033 * and sync vs async */
1034 unsigned low_slice =
1035 min(slice, base_low_slice * slice / sync_slice);
1036 /* the adapted slice value is scaled to fit all iqs
1037 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001038 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001039 low_slice);
1040 }
1041 }
Shaohua Lic553f8e2011-01-14 08:41:03 +01001042 return slice;
1043}
1044
1045static inline void
1046cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1047{
Vivek Goyalba5bd522011-01-19 08:25:02 -07001048 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001049
Vivek Goyaldae739e2009-12-03 12:59:45 -05001050 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001051 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001052 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +02001053 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +11001054}
1055
1056/*
1057 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1058 * isn't valid until the first request from the dispatch is activated
1059 * and the slice time set.
1060 */
Jens Axboea6151c32009-10-07 20:02:57 +02001061static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +11001062{
1063 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01001064 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +11001065 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01001066 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +11001067
Shaohua Lic1e44752010-11-08 15:01:02 +01001068 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +11001069}
1070
1071/*
Jens Axboe5e705372006-07-13 12:39:25 +02001072 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +02001074 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 */
Jens Axboe5e705372006-07-13 12:39:25 +02001076static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001077cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001079 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +02001081#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1082#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1083 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084
Jens Axboe5e705372006-07-13 12:39:25 +02001085 if (rq1 == NULL || rq1 == rq2)
1086 return rq2;
1087 if (rq2 == NULL)
1088 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001089
Namhyung Kim229836b2011-05-24 10:23:21 +02001090 if (rq_is_sync(rq1) != rq_is_sync(rq2))
1091 return rq_is_sync(rq1) ? rq1 : rq2;
1092
Christoph Hellwig65299a32011-08-23 14:50:29 +02001093 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1094 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
Jens Axboeb53d1ed2011-08-19 08:34:48 +02001095
Tejun Heo83096eb2009-05-07 22:24:39 +09001096 s1 = blk_rq_pos(rq1);
1097 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 /*
1100 * by definition, 1KiB is 2 sectors
1101 */
1102 back_max = cfqd->cfq_back_max * 2;
1103
1104 /*
1105 * Strict one way elevator _except_ in the case where we allow
1106 * short backward seeks which are biased as twice the cost of a
1107 * similar forward seek.
1108 */
1109 if (s1 >= last)
1110 d1 = s1 - last;
1111 else if (s1 + back_max >= last)
1112 d1 = (last - s1) * cfqd->cfq_back_penalty;
1113 else
Andreas Mohre8a99052006-03-28 08:59:49 +02001114 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115
1116 if (s2 >= last)
1117 d2 = s2 - last;
1118 else if (s2 + back_max >= last)
1119 d2 = (last - s2) * cfqd->cfq_back_penalty;
1120 else
Andreas Mohre8a99052006-03-28 08:59:49 +02001121 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122
1123 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124
Andreas Mohre8a99052006-03-28 08:59:49 +02001125 /*
1126 * By doing switch() on the bit mask "wrap" we avoid having to
1127 * check two variables for all permutations: --> faster!
1128 */
1129 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +02001130 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +02001131 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +02001132 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +02001133 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +02001134 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +02001135 else {
1136 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +02001137 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +02001138 else
Jens Axboe5e705372006-07-13 12:39:25 +02001139 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +02001140 }
1141
1142 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +02001143 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +02001144 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +02001145 return rq2;
1146 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +02001147 default:
1148 /*
1149 * Since both rqs are wrapped,
1150 * start with the one that's further behind head
1151 * (--> only *one* back seek required),
1152 * since back seek takes more time than forward.
1153 */
1154 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +02001155 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 else
Jens Axboe5e705372006-07-13 12:39:25 +02001157 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 }
1159}
1160
Jens Axboe498d3aa22007-04-26 12:54:48 +02001161/*
1162 * The below is leftmost cache rbtree addon
1163 */
Jens Axboe08717142008-01-28 11:38:15 +01001164static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +02001165{
Vivek Goyal615f0252009-12-03 12:59:39 -05001166 /* Service tree is empty */
1167 if (!root->count)
1168 return NULL;
1169
Jens Axboecc09e292007-04-26 12:53:50 +02001170 if (!root->left)
1171 root->left = rb_first(&root->rb);
1172
Jens Axboe08717142008-01-28 11:38:15 +01001173 if (root->left)
1174 return rb_entry(root->left, struct cfq_queue, rb_node);
1175
1176 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +02001177}
1178
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001179static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1180{
1181 if (!root->left)
1182 root->left = rb_first(&root->rb);
1183
1184 if (root->left)
1185 return rb_entry_cfqg(root->left);
1186
1187 return NULL;
1188}
1189
Jens Axboea36e71f2009-04-15 12:15:11 +02001190static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1191{
1192 rb_erase(n, root);
1193 RB_CLEAR_NODE(n);
1194}
1195
Jens Axboecc09e292007-04-26 12:53:50 +02001196static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1197{
1198 if (root->left == n)
1199 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001200 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001201 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +02001202}
1203
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204/*
1205 * would be nice to take fifo expire time into account as well
1206 */
Jens Axboe5e705372006-07-13 12:39:25 +02001207static struct request *
1208cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1209 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210{
Jens Axboe21183b02006-07-13 12:33:14 +02001211 struct rb_node *rbnext = rb_next(&last->rb_node);
1212 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +02001213 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214
Jens Axboe21183b02006-07-13 12:33:14 +02001215 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216
1217 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +02001218 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +02001219
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +02001221 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +02001222 else {
1223 rbnext = rb_first(&cfqq->sort_list);
1224 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +02001225 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +02001226 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001228 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229}
1230
Jens Axboed9e76202007-04-20 14:27:50 +02001231static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1232 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233{
Jens Axboed9e76202007-04-20 14:27:50 +02001234 /*
1235 * just an approximation, should be ok.
1236 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001237 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +01001238 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +02001239}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001241static inline s64
1242cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1243{
1244 return cfqg->vdisktime - st->min_vdisktime;
1245}
1246
1247static void
1248__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1249{
1250 struct rb_node **node = &st->rb.rb_node;
1251 struct rb_node *parent = NULL;
1252 struct cfq_group *__cfqg;
1253 s64 key = cfqg_key(st, cfqg);
1254 int left = 1;
1255
1256 while (*node != NULL) {
1257 parent = *node;
1258 __cfqg = rb_entry_cfqg(parent);
1259
1260 if (key < cfqg_key(st, __cfqg))
1261 node = &parent->rb_left;
1262 else {
1263 node = &parent->rb_right;
1264 left = 0;
1265 }
1266 }
1267
1268 if (left)
1269 st->left = &cfqg->rb_node;
1270
1271 rb_link_node(&cfqg->rb_node, parent, node);
1272 rb_insert_color(&cfqg->rb_node, &st->rb);
1273}
1274
1275static void
Justin TerAvest8184f932011-03-17 16:12:36 +01001276cfq_update_group_weight(struct cfq_group *cfqg)
1277{
1278 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
Tejun Heoe71357e2013-01-09 08:05:10 -08001279
Tejun Heo3381cb82012-04-01 14:38:44 -07001280 if (cfqg->new_weight) {
Justin TerAvest8184f932011-03-17 16:12:36 +01001281 cfqg->weight = cfqg->new_weight;
Tejun Heo3381cb82012-04-01 14:38:44 -07001282 cfqg->new_weight = 0;
Justin TerAvest8184f932011-03-17 16:12:36 +01001283 }
Tejun Heoe71357e2013-01-09 08:05:10 -08001284
1285 if (cfqg->new_leaf_weight) {
1286 cfqg->leaf_weight = cfqg->new_leaf_weight;
1287 cfqg->new_leaf_weight = 0;
1288 }
Justin TerAvest8184f932011-03-17 16:12:36 +01001289}
1290
1291static void
1292cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1293{
Tejun Heo1d3650f2013-01-09 08:05:11 -08001294 unsigned int vfr = 1 << CFQ_SERVICE_SHIFT; /* start with 1 */
Tejun Heo7918ffb2013-01-09 08:05:11 -08001295 struct cfq_group *pos = cfqg;
Tejun Heo1d3650f2013-01-09 08:05:11 -08001296 struct cfq_group *parent;
Tejun Heo7918ffb2013-01-09 08:05:11 -08001297 bool propagate;
1298
1299 /* add to the service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +01001300 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1301
1302 cfq_update_group_weight(cfqg);
1303 __cfq_group_service_tree_add(st, cfqg);
Tejun Heo7918ffb2013-01-09 08:05:11 -08001304
1305 /*
Tejun Heo1d3650f2013-01-09 08:05:11 -08001306 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1307 * entitled to. vfraction is calculated by walking the tree
1308 * towards the root calculating the fraction it has at each level.
1309 * The compounded ratio is how much vfraction @cfqg owns.
1310 *
1311 * Start with the proportion tasks in this cfqg has against active
1312 * children cfqgs - its leaf_weight against children_weight.
Tejun Heo7918ffb2013-01-09 08:05:11 -08001313 */
1314 propagate = !pos->nr_active++;
1315 pos->children_weight += pos->leaf_weight;
Tejun Heo1d3650f2013-01-09 08:05:11 -08001316 vfr = vfr * pos->leaf_weight / pos->children_weight;
Tejun Heo7918ffb2013-01-09 08:05:11 -08001317
Tejun Heo1d3650f2013-01-09 08:05:11 -08001318 /*
1319 * Compound ->weight walking up the tree. Both activation and
1320 * vfraction calculation are done in the same loop. Propagation
1321 * stops once an already activated node is met. vfraction
1322 * calculation should always continue to the root.
1323 */
Tejun Heod02f7aa2013-01-09 08:05:11 -08001324 while ((parent = cfqg_parent(pos))) {
Tejun Heo1d3650f2013-01-09 08:05:11 -08001325 if (propagate) {
1326 propagate = !parent->nr_active++;
1327 parent->children_weight += pos->weight;
1328 }
1329 vfr = vfr * pos->weight / parent->children_weight;
Tejun Heo7918ffb2013-01-09 08:05:11 -08001330 pos = parent;
1331 }
Tejun Heo1d3650f2013-01-09 08:05:11 -08001332
1333 cfqg->vfraction = max_t(unsigned, vfr, 1);
Justin TerAvest8184f932011-03-17 16:12:36 +01001334}
1335
1336static void
1337cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001338{
1339 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1340 struct cfq_group *__cfqg;
1341 struct rb_node *n;
1342
1343 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +01001344 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001345 return;
1346
1347 /*
1348 * Currently put the group at the end. Later implement something
1349 * so that groups get lesser vtime based on their weights, so that
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001350 * if group does not loose all if it was not continuously backlogged.
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001351 */
1352 n = rb_last(&st->rb);
1353 if (n) {
1354 __cfqg = rb_entry_cfqg(n);
1355 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1356 } else
1357 cfqg->vdisktime = st->min_vdisktime;
Justin TerAvest8184f932011-03-17 16:12:36 +01001358 cfq_group_service_tree_add(st, cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001359}
1360
1361static void
Justin TerAvest8184f932011-03-17 16:12:36 +01001362cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1363{
Tejun Heo7918ffb2013-01-09 08:05:11 -08001364 struct cfq_group *pos = cfqg;
1365 bool propagate;
1366
1367 /*
1368 * Undo activation from cfq_group_service_tree_add(). Deactivate
1369 * @cfqg and propagate deactivation upwards.
1370 */
1371 propagate = !--pos->nr_active;
1372 pos->children_weight -= pos->leaf_weight;
1373
1374 while (propagate) {
Tejun Heod02f7aa2013-01-09 08:05:11 -08001375 struct cfq_group *parent = cfqg_parent(pos);
Tejun Heo7918ffb2013-01-09 08:05:11 -08001376
1377 /* @pos has 0 nr_active at this point */
1378 WARN_ON_ONCE(pos->children_weight);
Tejun Heo1d3650f2013-01-09 08:05:11 -08001379 pos->vfraction = 0;
Tejun Heo7918ffb2013-01-09 08:05:11 -08001380
1381 if (!parent)
1382 break;
1383
1384 propagate = !--parent->nr_active;
1385 parent->children_weight -= pos->weight;
1386 pos = parent;
1387 }
1388
1389 /* remove from the service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +01001390 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1391 cfq_rb_erase(&cfqg->rb_node, st);
1392}
1393
1394static void
1395cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001396{
1397 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1398
1399 BUG_ON(cfqg->nr_cfqq < 1);
1400 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05001401
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001402 /* If there are other cfq queues under this group, don't delete it */
1403 if (cfqg->nr_cfqq)
1404 return;
1405
Vivek Goyal2868ef72009-12-03 12:59:48 -05001406 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Justin TerAvest8184f932011-03-17 16:12:36 +01001407 cfq_group_service_tree_del(st, cfqg);
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04001408 cfqg->saved_wl_slice = 0;
Tejun Heo155fead2012-04-01 14:38:44 -07001409 cfqg_stats_update_dequeue(cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001410}
1411
Justin TerAvest167400d2011-03-12 16:54:00 +01001412static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1413 unsigned int *unaccounted_time)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001414{
Vivek Goyalf75edf22009-12-03 12:59:53 -05001415 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001416
1417 /*
1418 * Queue got expired before even a single request completed or
1419 * got expired immediately after first request completion.
1420 */
1421 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1422 /*
1423 * Also charge the seek time incurred to the group, otherwise
1424 * if there are mutiple queues in the group, each can dispatch
1425 * a single request on seeky media and cause lots of seek time
1426 * and group will never know it.
1427 */
1428 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1429 1);
1430 } else {
1431 slice_used = jiffies - cfqq->slice_start;
Justin TerAvest167400d2011-03-12 16:54:00 +01001432 if (slice_used > cfqq->allocated_slice) {
1433 *unaccounted_time = slice_used - cfqq->allocated_slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001434 slice_used = cfqq->allocated_slice;
Justin TerAvest167400d2011-03-12 16:54:00 +01001435 }
1436 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1437 *unaccounted_time += cfqq->slice_start -
1438 cfqq->dispatch_start;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001439 }
1440
Vivek Goyaldae739e2009-12-03 12:59:45 -05001441 return slice_used;
1442}
1443
1444static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001445 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001446{
1447 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Justin TerAvest167400d2011-03-12 16:54:00 +01001448 unsigned int used_sl, charge, unaccounted_sl = 0;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05001449 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1450 - cfqg->service_tree_idle.count;
Tejun Heo1d3650f2013-01-09 08:05:11 -08001451 unsigned int vfr;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001452
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05001453 BUG_ON(nr_sync < 0);
Justin TerAvest167400d2011-03-12 16:54:00 +01001454 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05001455
Vivek Goyal02b35082010-08-23 12:23:53 +02001456 if (iops_mode(cfqd))
1457 charge = cfqq->slice_dispatch;
1458 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1459 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001460
Tejun Heo1d3650f2013-01-09 08:05:11 -08001461 /*
1462 * Can't update vdisktime while on service tree and cfqg->vfraction
1463 * is valid only while on it. Cache vfr, leave the service tree,
1464 * update vdisktime and go back on. The re-addition to the tree
1465 * will also update the weights as necessary.
1466 */
1467 vfr = cfqg->vfraction;
Justin TerAvest8184f932011-03-17 16:12:36 +01001468 cfq_group_service_tree_del(st, cfqg);
Tejun Heo1d3650f2013-01-09 08:05:11 -08001469 cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
Justin TerAvest8184f932011-03-17 16:12:36 +01001470 cfq_group_service_tree_add(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001471
1472 /* This group is being expired. Save the context */
1473 if (time_after(cfqd->workload_expires, jiffies)) {
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04001474 cfqg->saved_wl_slice = cfqd->workload_expires
Vivek Goyaldae739e2009-12-03 12:59:45 -05001475 - jiffies;
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04001476 cfqg->saved_wl_type = cfqd->serving_wl_type;
1477 cfqg->saved_wl_class = cfqd->serving_wl_class;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001478 } else
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04001479 cfqg->saved_wl_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -05001480
1481 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1482 st->min_vdisktime);
Joe Perchesfd16d262011-06-13 10:42:49 +02001483 cfq_log_cfqq(cfqq->cfqd, cfqq,
1484 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1485 used_sl, cfqq->slice_dispatch, charge,
1486 iops_mode(cfqd), cfqq->nr_sectors);
Tejun Heo155fead2012-04-01 14:38:44 -07001487 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1488 cfqg_stats_set_start_empty_time(cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001489}
1490
Tejun Heof51b8022012-03-05 13:15:05 -08001491/**
1492 * cfq_init_cfqg_base - initialize base part of a cfq_group
1493 * @cfqg: cfq_group to initialize
1494 *
1495 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1496 * is enabled or not.
1497 */
1498static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1499{
1500 struct cfq_rb_root *st;
1501 int i, j;
1502
1503 for_each_cfqg_st(cfqg, i, j, st)
1504 *st = CFQ_RB_ROOT;
1505 RB_CLEAR_NODE(&cfqg->rb_node);
1506
1507 cfqg->ttime.last_end_request = jiffies;
1508}
1509
Vivek Goyal25fb5162009-12-03 12:59:46 -05001510#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07001511static void cfq_pd_init(struct blkcg_gq *blkg)
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001512{
Tejun Heo03814112012-03-05 13:15:14 -08001513 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001514
Tejun Heof51b8022012-03-05 13:15:05 -08001515 cfq_init_cfqg_base(cfqg);
Tejun Heo3381cb82012-04-01 14:38:44 -07001516 cfqg->weight = blkg->blkcg->cfq_weight;
Tejun Heoe71357e2013-01-09 08:05:10 -08001517 cfqg->leaf_weight = blkg->blkcg->cfq_leaf_weight;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001518}
1519
Tejun Heo0b399202013-01-09 08:05:13 -08001520static void cfq_pd_offline(struct blkcg_gq *blkg)
1521{
1522 /*
1523 * @blkg is going offline and will be ignored by
1524 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
1525 * that they don't get lost. If IOs complete after this point, the
1526 * stats for them will be lost. Oh well...
1527 */
1528 cfqg_stats_xfer_dead(blkg_to_cfqg(blkg));
1529}
1530
Tejun Heo43114012013-01-09 08:05:13 -08001531/* offset delta from cfqg->stats to cfqg->dead_stats */
1532static const int dead_stats_off_delta = offsetof(struct cfq_group, dead_stats) -
1533 offsetof(struct cfq_group, stats);
1534
1535/* to be used by recursive prfill, sums live and dead stats recursively */
1536static u64 cfqg_stat_pd_recursive_sum(struct blkg_policy_data *pd, int off)
1537{
1538 u64 sum = 0;
1539
1540 sum += blkg_stat_recursive_sum(pd, off);
1541 sum += blkg_stat_recursive_sum(pd, off + dead_stats_off_delta);
1542 return sum;
1543}
1544
1545/* to be used by recursive prfill, sums live and dead rwstats recursively */
1546static struct blkg_rwstat cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data *pd,
1547 int off)
1548{
1549 struct blkg_rwstat a, b;
1550
1551 a = blkg_rwstat_recursive_sum(pd, off);
1552 b = blkg_rwstat_recursive_sum(pd, off + dead_stats_off_delta);
1553 blkg_rwstat_merge(&a, &b);
1554 return a;
1555}
1556
Tejun Heo689665a2013-01-09 08:05:13 -08001557static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
1558{
1559 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1560
1561 cfqg_stats_reset(&cfqg->stats);
Tejun Heo0b399202013-01-09 08:05:13 -08001562 cfqg_stats_reset(&cfqg->dead_stats);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001563}
1564
1565/*
Vivek Goyal3e59cf92011-05-19 15:38:21 -04001566 * Search for the cfq group current task belongs to. request_queue lock must
1567 * be held.
Vivek Goyal25fb5162009-12-03 12:59:46 -05001568 */
Tejun Heocd1604f2012-03-05 13:15:06 -08001569static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
Tejun Heo3c798392012-04-16 13:57:25 -07001570 struct blkcg *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001571{
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001572 struct request_queue *q = cfqd->queue;
Tejun Heocd1604f2012-03-05 13:15:06 -08001573 struct cfq_group *cfqg = NULL;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001574
Tejun Heo3c798392012-04-16 13:57:25 -07001575 /* avoid lookup for the common case where there's no blkcg */
1576 if (blkcg == &blkcg_root) {
Tejun Heocd1604f2012-03-05 13:15:06 -08001577 cfqg = cfqd->root_group;
1578 } else {
Tejun Heo3c798392012-04-16 13:57:25 -07001579 struct blkcg_gq *blkg;
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001580
Tejun Heo3c96cb32012-04-13 13:11:34 -07001581 blkg = blkg_lookup_create(blkcg, q);
Tejun Heocd1604f2012-03-05 13:15:06 -08001582 if (!IS_ERR(blkg))
Tejun Heo03814112012-03-05 13:15:14 -08001583 cfqg = blkg_to_cfqg(blkg);
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001584 }
1585
Vivek Goyal25fb5162009-12-03 12:59:46 -05001586 return cfqg;
1587}
1588
1589static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1590{
1591 /* Currently, all async queues are mapped to root group */
1592 if (!cfq_cfqq_sync(cfqq))
Tejun Heof51b8022012-03-05 13:15:05 -08001593 cfqg = cfqq->cfqd->root_group;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001594
1595 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001596 /* cfqq reference on cfqg */
Tejun Heoeb7d8c072012-03-23 14:02:53 +01001597 cfqg_get(cfqg);
Vivek Goyalb1c35762009-12-03 12:59:47 -05001598}
1599
Tejun Heof95a04a2012-04-16 13:57:26 -07001600static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1601 struct blkg_policy_data *pd, int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001602{
Tejun Heof95a04a2012-04-16 13:57:26 -07001603 struct cfq_group *cfqg = pd_to_cfqg(pd);
Tejun Heo3381cb82012-04-01 14:38:44 -07001604
1605 if (!cfqg->dev_weight)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001606 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -07001607 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001608}
1609
Tejun Heo3381cb82012-04-01 14:38:44 -07001610static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1611 struct seq_file *sf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001612{
Tejun Heo3c798392012-04-16 13:57:25 -07001613 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1614 cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001615 false);
1616 return 0;
1617}
1618
Tejun Heoe71357e2013-01-09 08:05:10 -08001619static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1620 struct blkg_policy_data *pd, int off)
1621{
1622 struct cfq_group *cfqg = pd_to_cfqg(pd);
1623
1624 if (!cfqg->dev_leaf_weight)
1625 return 0;
1626 return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1627}
1628
1629static int cfqg_print_leaf_weight_device(struct cgroup *cgrp,
1630 struct cftype *cft,
1631 struct seq_file *sf)
1632{
1633 blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1634 cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq, 0,
1635 false);
1636 return 0;
1637}
1638
Tejun Heo3381cb82012-04-01 14:38:44 -07001639static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1640 struct seq_file *sf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001641{
Tejun Heo3c798392012-04-16 13:57:25 -07001642 seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001643 return 0;
1644}
1645
Tejun Heoe71357e2013-01-09 08:05:10 -08001646static int cfq_print_leaf_weight(struct cgroup *cgrp, struct cftype *cft,
1647 struct seq_file *sf)
1648{
1649 seq_printf(sf, "%u\n",
1650 cgroup_to_blkcg(cgrp)->cfq_leaf_weight);
1651 return 0;
1652}
1653
1654static int __cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1655 const char *buf, bool is_leaf_weight)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001656{
Tejun Heo3c798392012-04-16 13:57:25 -07001657 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001658 struct blkg_conf_ctx ctx;
Tejun Heo3381cb82012-04-01 14:38:44 -07001659 struct cfq_group *cfqg;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001660 int ret;
1661
Tejun Heo3c798392012-04-16 13:57:25 -07001662 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001663 if (ret)
1664 return ret;
1665
1666 ret = -EINVAL;
Tejun Heo3381cb82012-04-01 14:38:44 -07001667 cfqg = blkg_to_cfqg(ctx.blkg);
Tejun Heoa2b16932012-04-13 13:11:33 -07001668 if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
Tejun Heoe71357e2013-01-09 08:05:10 -08001669 if (!is_leaf_weight) {
1670 cfqg->dev_weight = ctx.v;
1671 cfqg->new_weight = ctx.v ?: blkcg->cfq_weight;
1672 } else {
1673 cfqg->dev_leaf_weight = ctx.v;
1674 cfqg->new_leaf_weight = ctx.v ?: blkcg->cfq_leaf_weight;
1675 }
Tejun Heo60c2bc22012-04-01 14:38:43 -07001676 ret = 0;
1677 }
1678
1679 blkg_conf_finish(&ctx);
1680 return ret;
1681}
1682
Tejun Heoe71357e2013-01-09 08:05:10 -08001683static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1684 const char *buf)
1685{
1686 return __cfqg_set_weight_device(cgrp, cft, buf, false);
1687}
1688
1689static int cfqg_set_leaf_weight_device(struct cgroup *cgrp, struct cftype *cft,
1690 const char *buf)
1691{
1692 return __cfqg_set_weight_device(cgrp, cft, buf, true);
1693}
1694
1695static int __cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val,
1696 bool is_leaf_weight)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001697{
Tejun Heo3c798392012-04-16 13:57:25 -07001698 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1699 struct blkcg_gq *blkg;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001700
Tejun Heo3381cb82012-04-01 14:38:44 -07001701 if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001702 return -EINVAL;
1703
1704 spin_lock_irq(&blkcg->lock);
Tejun Heoe71357e2013-01-09 08:05:10 -08001705
1706 if (!is_leaf_weight)
1707 blkcg->cfq_weight = val;
1708 else
1709 blkcg->cfq_leaf_weight = val;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001710
Sasha Levinb67bfe02013-02-27 17:06:00 -08001711 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
Tejun Heo3381cb82012-04-01 14:38:44 -07001712 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001713
Tejun Heoe71357e2013-01-09 08:05:10 -08001714 if (!cfqg)
1715 continue;
1716
1717 if (!is_leaf_weight) {
1718 if (!cfqg->dev_weight)
1719 cfqg->new_weight = blkcg->cfq_weight;
1720 } else {
1721 if (!cfqg->dev_leaf_weight)
1722 cfqg->new_leaf_weight = blkcg->cfq_leaf_weight;
1723 }
Tejun Heo60c2bc22012-04-01 14:38:43 -07001724 }
1725
1726 spin_unlock_irq(&blkcg->lock);
1727 return 0;
1728}
1729
Tejun Heoe71357e2013-01-09 08:05:10 -08001730static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1731{
1732 return __cfq_set_weight(cgrp, cft, val, false);
1733}
1734
1735static int cfq_set_leaf_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1736{
1737 return __cfq_set_weight(cgrp, cft, val, true);
1738}
1739
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001740static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1741 struct seq_file *sf)
1742{
Tejun Heo3c798392012-04-16 13:57:25 -07001743 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001744
Tejun Heo3c798392012-04-16 13:57:25 -07001745 blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001746 cft->private, false);
1747 return 0;
1748}
1749
1750static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1751 struct seq_file *sf)
1752{
Tejun Heo3c798392012-04-16 13:57:25 -07001753 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001754
Tejun Heo3c798392012-04-16 13:57:25 -07001755 blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001756 cft->private, true);
1757 return 0;
1758}
1759
Tejun Heo43114012013-01-09 08:05:13 -08001760static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1761 struct blkg_policy_data *pd, int off)
1762{
1763 u64 sum = cfqg_stat_pd_recursive_sum(pd, off);
1764
1765 return __blkg_prfill_u64(sf, pd, sum);
1766}
1767
1768static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1769 struct blkg_policy_data *pd, int off)
1770{
1771 struct blkg_rwstat sum = cfqg_rwstat_pd_recursive_sum(pd, off);
1772
1773 return __blkg_prfill_rwstat(sf, pd, &sum);
1774}
1775
1776static int cfqg_print_stat_recursive(struct cgroup *cgrp, struct cftype *cft,
1777 struct seq_file *sf)
1778{
1779 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1780
1781 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_stat_recursive,
1782 &blkcg_policy_cfq, cft->private, false);
1783 return 0;
1784}
1785
1786static int cfqg_print_rwstat_recursive(struct cgroup *cgrp, struct cftype *cft,
1787 struct seq_file *sf)
1788{
1789 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1790
1791 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_rwstat_recursive,
1792 &blkcg_policy_cfq, cft->private, true);
1793 return 0;
1794}
1795
Tejun Heo60c2bc22012-04-01 14:38:43 -07001796#ifdef CONFIG_DEBUG_BLK_CGROUP
Tejun Heof95a04a2012-04-16 13:57:26 -07001797static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1798 struct blkg_policy_data *pd, int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001799{
Tejun Heof95a04a2012-04-16 13:57:26 -07001800 struct cfq_group *cfqg = pd_to_cfqg(pd);
Tejun Heo155fead2012-04-01 14:38:44 -07001801 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001802 u64 v = 0;
1803
1804 if (samples) {
Tejun Heo155fead2012-04-01 14:38:44 -07001805 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001806 do_div(v, samples);
1807 }
Tejun Heof95a04a2012-04-16 13:57:26 -07001808 __blkg_prfill_u64(sf, pd, v);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001809 return 0;
1810}
1811
1812/* print avg_queue_size */
Tejun Heo155fead2012-04-01 14:38:44 -07001813static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1814 struct seq_file *sf)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001815{
Tejun Heo3c798392012-04-16 13:57:25 -07001816 struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001817
Tejun Heo155fead2012-04-01 14:38:44 -07001818 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
Tejun Heo3c798392012-04-16 13:57:25 -07001819 &blkcg_policy_cfq, 0, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001820 return 0;
1821}
1822#endif /* CONFIG_DEBUG_BLK_CGROUP */
1823
1824static struct cftype cfq_blkcg_files[] = {
Tejun Heo1d3650f2013-01-09 08:05:11 -08001825 /* on root, weight is mapped to leaf_weight */
Tejun Heo60c2bc22012-04-01 14:38:43 -07001826 {
1827 .name = "weight_device",
Tejun Heo1d3650f2013-01-09 08:05:11 -08001828 .flags = CFTYPE_ONLY_ON_ROOT,
1829 .read_seq_string = cfqg_print_leaf_weight_device,
1830 .write_string = cfqg_set_leaf_weight_device,
1831 .max_write_len = 256,
1832 },
1833 {
1834 .name = "weight",
1835 .flags = CFTYPE_ONLY_ON_ROOT,
1836 .read_seq_string = cfq_print_leaf_weight,
1837 .write_u64 = cfq_set_leaf_weight,
1838 },
1839
1840 /* no such mapping necessary for !roots */
1841 {
1842 .name = "weight_device",
1843 .flags = CFTYPE_NOT_ON_ROOT,
Tejun Heo3381cb82012-04-01 14:38:44 -07001844 .read_seq_string = cfqg_print_weight_device,
1845 .write_string = cfqg_set_weight_device,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001846 .max_write_len = 256,
1847 },
1848 {
1849 .name = "weight",
Tejun Heoe71357e2013-01-09 08:05:10 -08001850 .flags = CFTYPE_NOT_ON_ROOT,
Tejun Heo3381cb82012-04-01 14:38:44 -07001851 .read_seq_string = cfq_print_weight,
1852 .write_u64 = cfq_set_weight,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001853 },
Tejun Heo1d3650f2013-01-09 08:05:11 -08001854
1855 {
1856 .name = "leaf_weight_device",
Tejun Heoe71357e2013-01-09 08:05:10 -08001857 .read_seq_string = cfqg_print_leaf_weight_device,
1858 .write_string = cfqg_set_leaf_weight_device,
1859 .max_write_len = 256,
1860 },
1861 {
1862 .name = "leaf_weight",
Tejun Heoe71357e2013-01-09 08:05:10 -08001863 .read_seq_string = cfq_print_leaf_weight,
1864 .write_u64 = cfq_set_leaf_weight,
1865 },
1866
Tejun Heo43114012013-01-09 08:05:13 -08001867 /* statistics, covers only the tasks in the cfqg */
Tejun Heo60c2bc22012-04-01 14:38:43 -07001868 {
1869 .name = "time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001870 .private = offsetof(struct cfq_group, stats.time),
1871 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001872 },
1873 {
1874 .name = "sectors",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001875 .private = offsetof(struct cfq_group, stats.sectors),
1876 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001877 },
1878 {
1879 .name = "io_service_bytes",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001880 .private = offsetof(struct cfq_group, stats.service_bytes),
1881 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001882 },
1883 {
1884 .name = "io_serviced",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001885 .private = offsetof(struct cfq_group, stats.serviced),
1886 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001887 },
1888 {
1889 .name = "io_service_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001890 .private = offsetof(struct cfq_group, stats.service_time),
1891 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001892 },
1893 {
1894 .name = "io_wait_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001895 .private = offsetof(struct cfq_group, stats.wait_time),
1896 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001897 },
1898 {
1899 .name = "io_merged",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001900 .private = offsetof(struct cfq_group, stats.merged),
1901 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001902 },
1903 {
1904 .name = "io_queued",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001905 .private = offsetof(struct cfq_group, stats.queued),
1906 .read_seq_string = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001907 },
Tejun Heo43114012013-01-09 08:05:13 -08001908
1909 /* the same statictics which cover the cfqg and its descendants */
1910 {
1911 .name = "time_recursive",
1912 .private = offsetof(struct cfq_group, stats.time),
1913 .read_seq_string = cfqg_print_stat_recursive,
1914 },
1915 {
1916 .name = "sectors_recursive",
1917 .private = offsetof(struct cfq_group, stats.sectors),
1918 .read_seq_string = cfqg_print_stat_recursive,
1919 },
1920 {
1921 .name = "io_service_bytes_recursive",
1922 .private = offsetof(struct cfq_group, stats.service_bytes),
1923 .read_seq_string = cfqg_print_rwstat_recursive,
1924 },
1925 {
1926 .name = "io_serviced_recursive",
1927 .private = offsetof(struct cfq_group, stats.serviced),
1928 .read_seq_string = cfqg_print_rwstat_recursive,
1929 },
1930 {
1931 .name = "io_service_time_recursive",
1932 .private = offsetof(struct cfq_group, stats.service_time),
1933 .read_seq_string = cfqg_print_rwstat_recursive,
1934 },
1935 {
1936 .name = "io_wait_time_recursive",
1937 .private = offsetof(struct cfq_group, stats.wait_time),
1938 .read_seq_string = cfqg_print_rwstat_recursive,
1939 },
1940 {
1941 .name = "io_merged_recursive",
1942 .private = offsetof(struct cfq_group, stats.merged),
1943 .read_seq_string = cfqg_print_rwstat_recursive,
1944 },
1945 {
1946 .name = "io_queued_recursive",
1947 .private = offsetof(struct cfq_group, stats.queued),
1948 .read_seq_string = cfqg_print_rwstat_recursive,
1949 },
Tejun Heo60c2bc22012-04-01 14:38:43 -07001950#ifdef CONFIG_DEBUG_BLK_CGROUP
1951 {
1952 .name = "avg_queue_size",
Tejun Heo155fead2012-04-01 14:38:44 -07001953 .read_seq_string = cfqg_print_avg_queue_size,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001954 },
1955 {
1956 .name = "group_wait_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001957 .private = offsetof(struct cfq_group, stats.group_wait_time),
1958 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001959 },
1960 {
1961 .name = "idle_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001962 .private = offsetof(struct cfq_group, stats.idle_time),
1963 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001964 },
1965 {
1966 .name = "empty_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001967 .private = offsetof(struct cfq_group, stats.empty_time),
1968 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001969 },
1970 {
1971 .name = "dequeue",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001972 .private = offsetof(struct cfq_group, stats.dequeue),
1973 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001974 },
1975 {
1976 .name = "unaccounted_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001977 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1978 .read_seq_string = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07001979 },
1980#endif /* CONFIG_DEBUG_BLK_CGROUP */
1981 { } /* terminate */
1982};
Vivek Goyal25fb5162009-12-03 12:59:46 -05001983#else /* GROUP_IOSCHED */
Tejun Heocd1604f2012-03-05 13:15:06 -08001984static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
Tejun Heo3c798392012-04-16 13:57:25 -07001985 struct blkcg *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001986{
Tejun Heof51b8022012-03-05 13:15:05 -08001987 return cfqd->root_group;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001988}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02001989
Vivek Goyal25fb5162009-12-03 12:59:46 -05001990static inline void
1991cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1992 cfqq->cfqg = cfqg;
1993}
1994
1995#endif /* GROUP_IOSCHED */
1996
Jens Axboe498d3aa22007-04-26 12:54:48 +02001997/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001998 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001999 * requests waiting to be processed. It is sorted in the order that
2000 * we will service the queues.
2001 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002002static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002003 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02002004{
Jens Axboe08717142008-01-28 11:38:15 +01002005 struct rb_node **p, *parent;
2006 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002007 unsigned long rb_key;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002008 struct cfq_rb_root *st;
Jens Axboe498d3aa22007-04-26 12:54:48 +02002009 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002010 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05002011
Vivek Goyal34b98d02012-10-03 16:56:58 -04002012 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01002013 if (cfq_class_idle(cfqq)) {
2014 rb_key = CFQ_IDLE_DELAY;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002015 parent = rb_last(&st->rb);
Jens Axboe08717142008-01-28 11:38:15 +01002016 if (parent && parent != &cfqq->rb_node) {
2017 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2018 rb_key += __cfqq->rb_key;
2019 } else
2020 rb_key += jiffies;
2021 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02002022 /*
2023 * Get our rb key offset. Subtract any residual slice
2024 * value carried from last service. A negative resid
2025 * count indicates slice overrun, and this should position
2026 * the next service time further away in the tree.
2027 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02002028 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02002029 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02002030 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02002031 } else {
2032 rb_key = -HZ;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002033 __cfqq = cfq_rb_first(st);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02002034 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2035 }
Jens Axboed9e76202007-04-20 14:27:50 +02002036
2037 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05002038 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01002039 /*
Jens Axboed9e76202007-04-20 14:27:50 +02002040 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01002041 */
Vivek Goyal34b98d02012-10-03 16:56:58 -04002042 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
Jens Axboed9e76202007-04-20 14:27:50 +02002043 return;
Jens Axboe53b037442006-07-28 09:48:51 +02002044
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01002045 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2046 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002047 }
Jens Axboed9e76202007-04-20 14:27:50 +02002048
Jens Axboe498d3aa22007-04-26 12:54:48 +02002049 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01002050 parent = NULL;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002051 cfqq->service_tree = st;
2052 p = &st->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02002053 while (*p) {
2054 parent = *p;
2055 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2056
Jens Axboe0c534e02007-04-18 20:01:57 +02002057 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002058 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02002059 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002060 if (time_before(rb_key, __cfqq->rb_key))
Vivek Goyal1f23f122012-10-03 16:57:00 -04002061 p = &parent->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002062 else {
Vivek Goyal1f23f122012-10-03 16:57:00 -04002063 p = &parent->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02002064 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002065 }
Jens Axboed9e76202007-04-20 14:27:50 +02002066 }
2067
Jens Axboecc09e292007-04-26 12:53:50 +02002068 if (left)
Vivek Goyal34b98d02012-10-03 16:56:58 -04002069 st->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02002070
Jens Axboed9e76202007-04-20 14:27:50 +02002071 cfqq->rb_key = rb_key;
2072 rb_link_node(&cfqq->rb_node, parent, p);
Vivek Goyal34b98d02012-10-03 16:56:58 -04002073 rb_insert_color(&cfqq->rb_node, &st->rb);
2074 st->count++;
Namhyung Kim20359f22011-05-24 10:23:22 +02002075 if (add_front || !new_cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05002076 return;
Justin TerAvest8184f932011-03-17 16:12:36 +01002077 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078}
2079
Jens Axboea36e71f2009-04-15 12:15:11 +02002080static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002081cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2082 sector_t sector, struct rb_node **ret_parent,
2083 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02002084{
Jens Axboea36e71f2009-04-15 12:15:11 +02002085 struct rb_node **p, *parent;
2086 struct cfq_queue *cfqq = NULL;
2087
2088 parent = NULL;
2089 p = &root->rb_node;
2090 while (*p) {
2091 struct rb_node **n;
2092
2093 parent = *p;
2094 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2095
2096 /*
2097 * Sort strictly based on sector. Smallest to the left,
2098 * largest to the right.
2099 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09002100 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002101 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09002102 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002103 n = &(*p)->rb_left;
2104 else
2105 break;
2106 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02002107 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02002108 }
2109
2110 *ret_parent = parent;
2111 if (rb_link)
2112 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02002113 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02002114}
2115
2116static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2117{
Jens Axboea36e71f2009-04-15 12:15:11 +02002118 struct rb_node **p, *parent;
2119 struct cfq_queue *__cfqq;
2120
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002121 if (cfqq->p_root) {
2122 rb_erase(&cfqq->p_node, cfqq->p_root);
2123 cfqq->p_root = NULL;
2124 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002125
2126 if (cfq_class_idle(cfqq))
2127 return;
2128 if (!cfqq->next_rq)
2129 return;
2130
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002131 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09002132 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2133 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02002134 if (!__cfqq) {
2135 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002136 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2137 } else
2138 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02002139}
2140
Jens Axboe498d3aa22007-04-26 12:54:48 +02002141/*
2142 * Update cfqq's position in the service tree.
2143 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02002144static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002145{
Jens Axboe6d048f52007-04-25 12:44:27 +02002146 /*
2147 * Resorting requires the cfqq to be on the RR list already.
2148 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002149 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02002150 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02002151 cfq_prio_tree_add(cfqd, cfqq);
2152 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002153}
2154
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155/*
2156 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02002157 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 */
Jens Axboefebffd62008-01-28 13:19:43 +01002159static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160{
Jens Axboe7b679132008-05-30 12:23:07 +02002161 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02002162 BUG_ON(cfq_cfqq_on_rr(cfqq));
2163 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 cfqd->busy_queues++;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002165 if (cfq_cfqq_sync(cfqq))
2166 cfqd->busy_sync_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167
Jens Axboeedd75ff2007-04-19 12:03:34 +02002168 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169}
2170
Jens Axboe498d3aa22007-04-26 12:54:48 +02002171/*
2172 * Called when the cfqq no longer has requests pending, remove it from
2173 * the service tree.
2174 */
Jens Axboefebffd62008-01-28 13:19:43 +01002175static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176{
Jens Axboe7b679132008-05-30 12:23:07 +02002177 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02002178 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2179 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01002181 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2182 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2183 cfqq->service_tree = NULL;
2184 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002185 if (cfqq->p_root) {
2186 rb_erase(&cfqq->p_node, cfqq->p_root);
2187 cfqq->p_root = NULL;
2188 }
Jens Axboed9e76202007-04-20 14:27:50 +02002189
Justin TerAvest8184f932011-03-17 16:12:36 +01002190 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 BUG_ON(!cfqd->busy_queues);
2192 cfqd->busy_queues--;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002193 if (cfq_cfqq_sync(cfqq))
2194 cfqd->busy_sync_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195}
2196
2197/*
2198 * rb tree support functions
2199 */
Jens Axboefebffd62008-01-28 13:19:43 +01002200static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201{
Jens Axboe5e705372006-07-13 12:39:25 +02002202 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02002203 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204
Jens Axboeb4878f22005-10-20 16:42:29 +02002205 BUG_ON(!cfqq->queued[sync]);
2206 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207
Jens Axboe5e705372006-07-13 12:39:25 +02002208 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209
Vivek Goyalf04a6422009-12-03 12:59:40 -05002210 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2211 /*
2212 * Queue will be deleted from service tree when we actually
2213 * expire it later. Right now just remove it from prio tree
2214 * as it is empty.
2215 */
2216 if (cfqq->p_root) {
2217 rb_erase(&cfqq->p_node, cfqq->p_root);
2218 cfqq->p_root = NULL;
2219 }
2220 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221}
2222
Jens Axboe5e705372006-07-13 12:39:25 +02002223static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224{
Jens Axboe5e705372006-07-13 12:39:25 +02002225 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226 struct cfq_data *cfqd = cfqq->cfqd;
Jeff Moyer796d5112011-06-02 21:19:05 +02002227 struct request *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228
Jens Axboe5380a102006-07-13 12:37:56 +02002229 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230
Jeff Moyer796d5112011-06-02 21:19:05 +02002231 elv_rb_add(&cfqq->sort_list, rq);
Jens Axboe5fccbf62006-10-31 14:21:55 +01002232
2233 if (!cfq_cfqq_on_rr(cfqq))
2234 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02002235
2236 /*
2237 * check if this request is a better next-serve candidate
2238 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002239 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01002240 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02002241
2242 /*
2243 * adjust priority tree position, if ->next_rq changes
2244 */
2245 if (prev != cfqq->next_rq)
2246 cfq_prio_tree_add(cfqd, cfqq);
2247
Jens Axboe5044eed2007-04-25 11:53:48 +02002248 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249}
2250
Jens Axboefebffd62008-01-28 13:19:43 +01002251static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252{
Jens Axboe5380a102006-07-13 12:37:56 +02002253 elv_rb_del(&cfqq->sort_list, rq);
2254 cfqq->queued[rq_is_sync(rq)]--;
Tejun Heo155fead2012-04-01 14:38:44 -07002255 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
Jens Axboe5e705372006-07-13 12:39:25 +02002256 cfq_add_rq_rb(rq);
Tejun Heo155fead2012-04-01 14:38:44 -07002257 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2258 rq->cmd_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259}
2260
Jens Axboe206dc692006-03-28 13:03:44 +02002261static struct request *
2262cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263{
Jens Axboe206dc692006-03-28 13:03:44 +02002264 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01002265 struct cfq_io_cq *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02002266 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267
Jens Axboe4ac845a2008-01-24 08:44:49 +01002268 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002269 if (!cic)
2270 return NULL;
2271
2272 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Kent Overstreetf73a1c72012-09-25 15:05:12 -07002273 if (cfqq)
2274 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 return NULL;
2277}
2278
Jens Axboe165125e2007-07-24 09:28:11 +02002279static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02002280{
2281 struct cfq_data *cfqd = q->elevator->elevator_data;
2282
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002283 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02002284 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002285 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02002286
Tejun Heo5b936292009-05-07 22:24:38 +09002287 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002288}
2289
Jens Axboe165125e2007-07-24 09:28:11 +02002290static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291{
Jens Axboe22e2c502005-06-27 10:55:12 +02002292 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002294 WARN_ON(!cfqd->rq_in_driver);
2295 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02002296 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002297 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298}
2299
Jens Axboeb4878f22005-10-20 16:42:29 +02002300static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301{
Jens Axboe5e705372006-07-13 12:39:25 +02002302 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02002303
Jens Axboe5e705372006-07-13 12:39:25 +02002304 if (cfqq->next_rq == rq)
2305 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306
Jens Axboeb4878f22005-10-20 16:42:29 +02002307 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02002308 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02002309
Aaron Carroll45333d52008-08-26 15:52:36 +02002310 cfqq->cfqd->rq_queued--;
Tejun Heo155fead2012-04-01 14:38:44 -07002311 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
Christoph Hellwig65299a32011-08-23 14:50:29 +02002312 if (rq->cmd_flags & REQ_PRIO) {
2313 WARN_ON(!cfqq->prio_pending);
2314 cfqq->prio_pending--;
Jens Axboeb53d1ed2011-08-19 08:34:48 +02002315 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316}
2317
Jens Axboe165125e2007-07-24 09:28:11 +02002318static int cfq_merge(struct request_queue *q, struct request **req,
2319 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320{
2321 struct cfq_data *cfqd = q->elevator->elevator_data;
2322 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323
Jens Axboe206dc692006-03-28 13:03:44 +02002324 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002325 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02002326 *req = __rq;
2327 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 }
2329
2330 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331}
2332
Jens Axboe165125e2007-07-24 09:28:11 +02002333static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02002334 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335{
Jens Axboe21183b02006-07-13 12:33:14 +02002336 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02002337 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338
Jens Axboe5e705372006-07-13 12:39:25 +02002339 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341}
2342
Divyesh Shah812d4022010-04-08 21:14:23 -07002343static void cfq_bio_merged(struct request_queue *q, struct request *req,
2344 struct bio *bio)
2345{
Tejun Heo155fead2012-04-01 14:38:44 -07002346 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
Divyesh Shah812d4022010-04-08 21:14:23 -07002347}
2348
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349static void
Jens Axboe165125e2007-07-24 09:28:11 +02002350cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 struct request *next)
2352{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01002353 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Shaohua Li4a0b75c2011-12-16 14:00:22 +01002354 struct cfq_data *cfqd = q->elevator->elevator_data;
2355
Jens Axboe22e2c502005-06-27 10:55:12 +02002356 /*
2357 * reposition in fifo if next is older than rq
2358 */
2359 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Shaohua Li3d106fba2012-11-06 12:39:51 +01002360 time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
2361 cfqq == RQ_CFQQ(next)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002362 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02002363 rq_set_fifo_time(rq, rq_fifo_time(next));
2364 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002365
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01002366 if (cfqq->next_rq == next)
2367 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02002368 cfq_remove_request(next);
Tejun Heo155fead2012-04-01 14:38:44 -07002369 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
Shaohua Li4a0b75c2011-12-16 14:00:22 +01002370
2371 cfqq = RQ_CFQQ(next);
2372 /*
2373 * all requests of this queue are merged to other queues, delete it
2374 * from the service tree. If it's the active_queue,
2375 * cfq_dispatch_requests() will choose to expire it or do idle
2376 */
2377 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2378 cfqq != cfqd->active_queue)
2379 cfq_del_cfqq_rr(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002380}
2381
Jens Axboe165125e2007-07-24 09:28:11 +02002382static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01002383 struct bio *bio)
2384{
2385 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heoc5869802011-12-14 00:33:41 +01002386 struct cfq_io_cq *cic;
Jens Axboeda775262006-12-20 11:04:12 +01002387 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01002388
2389 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01002390 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01002391 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02002392 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02002393 return false;
Jens Axboeda775262006-12-20 11:04:12 +01002394
2395 /*
Tejun Heof1a4f4d2011-12-14 00:33:39 +01002396 * Lookup the cfqq that this bio will be queued with and allow
Tejun Heo07c2bd32012-02-08 09:19:42 +01002397 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01002398 */
Tejun Heo07c2bd32012-02-08 09:19:42 +01002399 cic = cfq_cic_lookup(cfqd, current->io_context);
2400 if (!cic)
2401 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01002402
Vasily Tarasov91fac312007-04-25 12:29:51 +02002403 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02002404 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01002405}
2406
Divyesh Shah812df482010-04-08 21:15:35 -07002407static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2408{
2409 del_timer(&cfqd->idle_slice_timer);
Tejun Heo155fead2012-04-01 14:38:44 -07002410 cfqg_stats_update_idle_time(cfqq->cfqg);
Divyesh Shah812df482010-04-08 21:15:35 -07002411}
2412
Jens Axboefebffd62008-01-28 13:19:43 +01002413static void __cfq_set_active_queue(struct cfq_data *cfqd,
2414 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002415{
2416 if (cfqq) {
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04002417 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002418 cfqd->serving_wl_class, cfqd->serving_wl_type);
Tejun Heo155fead2012-04-01 14:38:44 -07002419 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
Justin TerAvest62a37f62011-03-23 08:25:44 +01002420 cfqq->slice_start = 0;
2421 cfqq->dispatch_start = jiffies;
2422 cfqq->allocated_slice = 0;
2423 cfqq->slice_end = 0;
2424 cfqq->slice_dispatch = 0;
2425 cfqq->nr_sectors = 0;
2426
2427 cfq_clear_cfqq_wait_request(cfqq);
2428 cfq_clear_cfqq_must_dispatch(cfqq);
2429 cfq_clear_cfqq_must_alloc_slice(cfqq);
2430 cfq_clear_cfqq_fifo_expire(cfqq);
2431 cfq_mark_cfqq_slice_new(cfqq);
2432
2433 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002434 }
2435
2436 cfqd->active_queue = cfqq;
2437}
2438
2439/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002440 * current cfqq expired its slice (or was too idle), select new one
2441 */
2442static void
2443__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02002444 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002445{
Jens Axboe7b679132008-05-30 12:23:07 +02002446 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2447
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002448 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07002449 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002450
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002451 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05002452 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002453
2454 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01002455 * If this cfqq is shared between multiple processes, check to
2456 * make sure that those processes are still issuing I/Os within
2457 * the mean seek distance. If not, it may be time to break the
2458 * queues apart again.
2459 */
2460 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2461 cfq_mark_cfqq_split_coop(cfqq);
2462
2463 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02002464 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002465 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01002466 if (timed_out) {
2467 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07002468 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01002469 else
2470 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02002471 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2472 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002473
Vivek Goyale5ff0822010-04-26 19:25:11 +02002474 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05002475
Vivek Goyalf04a6422009-12-03 12:59:40 -05002476 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2477 cfq_del_cfqq_rr(cfqd, cfqq);
2478
Jens Axboeedd75ff2007-04-19 12:03:34 +02002479 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002480
2481 if (cfqq == cfqd->active_queue)
2482 cfqd->active_queue = NULL;
2483
2484 if (cfqd->active_cic) {
Tejun Heo11a31222012-02-07 07:51:30 +01002485 put_io_context(cfqd->active_cic->icq.ioc);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002486 cfqd->active_cic = NULL;
2487 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002488}
2489
Vivek Goyale5ff0822010-04-26 19:25:11 +02002490static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002491{
2492 struct cfq_queue *cfqq = cfqd->active_queue;
2493
2494 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02002495 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002496}
2497
Jens Axboe498d3aa22007-04-26 12:54:48 +02002498/*
2499 * Get next queue for service. Unless we have a queue preemption,
2500 * we'll simply select the first cfqq in the service tree.
2501 */
Jens Axboe6d048f52007-04-25 12:44:27 +02002502static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002503{
Vivek Goyal34b98d02012-10-03 16:56:58 -04002504 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2505 cfqd->serving_wl_class, cfqd->serving_wl_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02002506
Vivek Goyalf04a6422009-12-03 12:59:40 -05002507 if (!cfqd->rq_queued)
2508 return NULL;
2509
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002510 /* There is nothing to dispatch */
Vivek Goyal34b98d02012-10-03 16:56:58 -04002511 if (!st)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002512 return NULL;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002513 if (RB_EMPTY_ROOT(&st->rb))
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002514 return NULL;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002515 return cfq_rb_first(st);
Jens Axboe6d048f52007-04-25 12:44:27 +02002516}
2517
Vivek Goyalf04a6422009-12-03 12:59:40 -05002518static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2519{
Vivek Goyal25fb5162009-12-03 12:59:46 -05002520 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05002521 struct cfq_queue *cfqq;
2522 int i, j;
2523 struct cfq_rb_root *st;
2524
2525 if (!cfqd->rq_queued)
2526 return NULL;
2527
Vivek Goyal25fb5162009-12-03 12:59:46 -05002528 cfqg = cfq_get_next_cfqg(cfqd);
2529 if (!cfqg)
2530 return NULL;
2531
Vivek Goyalf04a6422009-12-03 12:59:40 -05002532 for_each_cfqg_st(cfqg, i, j, st)
2533 if ((cfqq = cfq_rb_first(st)) != NULL)
2534 return cfqq;
2535 return NULL;
2536}
2537
Jens Axboe498d3aa22007-04-26 12:54:48 +02002538/*
2539 * Get and set a new active queue for service.
2540 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002541static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2542 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002543{
Jens Axboee00ef792009-11-04 08:54:55 +01002544 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02002545 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02002546
Jens Axboe22e2c502005-06-27 10:55:12 +02002547 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02002548 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002549}
2550
Jens Axboed9e76202007-04-20 14:27:50 +02002551static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2552 struct request *rq)
2553{
Tejun Heo83096eb2009-05-07 22:24:39 +09002554 if (blk_rq_pos(rq) >= cfqd->last_position)
2555 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02002556 else
Tejun Heo83096eb2009-05-07 22:24:39 +09002557 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02002558}
2559
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002560static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01002561 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002562{
Shaohua Lie9ce3352010-03-19 08:03:04 +01002563 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02002564}
2565
Jens Axboea36e71f2009-04-15 12:15:11 +02002566static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2567 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002568{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002569 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02002570 struct rb_node *parent, *node;
2571 struct cfq_queue *__cfqq;
2572 sector_t sector = cfqd->last_position;
2573
2574 if (RB_EMPTY_ROOT(root))
2575 return NULL;
2576
2577 /*
2578 * First, if we find a request starting at the end of the last
2579 * request, choose it.
2580 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002581 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02002582 if (__cfqq)
2583 return __cfqq;
2584
2585 /*
2586 * If the exact sector wasn't found, the parent of the NULL leaf
2587 * will contain the closest sector.
2588 */
2589 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01002590 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002591 return __cfqq;
2592
Tejun Heo2e46e8b2009-05-07 22:24:41 +09002593 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02002594 node = rb_next(&__cfqq->p_node);
2595 else
2596 node = rb_prev(&__cfqq->p_node);
2597 if (!node)
2598 return NULL;
2599
2600 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01002601 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002602 return __cfqq;
2603
2604 return NULL;
2605}
2606
2607/*
2608 * cfqd - obvious
2609 * cur_cfqq - passed in so that we don't decide that the current queue is
2610 * closely cooperating with itself.
2611 *
2612 * So, basically we're assuming that that cur_cfqq has dispatched at least
2613 * one request, and that cfqd->last_position reflects a position on the disk
2614 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2615 * assumption.
2616 */
2617static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002618 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02002619{
2620 struct cfq_queue *cfqq;
2621
Divyesh Shah39c01b22010-03-25 15:45:57 +01002622 if (cfq_class_idle(cur_cfqq))
2623 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002624 if (!cfq_cfqq_sync(cur_cfqq))
2625 return NULL;
2626 if (CFQQ_SEEKY(cur_cfqq))
2627 return NULL;
2628
Jens Axboea36e71f2009-04-15 12:15:11 +02002629 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01002630 * Don't search priority tree if it's the only queue in the group.
2631 */
2632 if (cur_cfqq->cfqg->nr_cfqq == 1)
2633 return NULL;
2634
2635 /*
Jens Axboed9e76202007-04-20 14:27:50 +02002636 * We should notice if some of the queues are cooperating, eg
2637 * working closely on the same area of the disk. In that case,
2638 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02002639 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002640 cfqq = cfqq_close(cfqd, cur_cfqq);
2641 if (!cfqq)
2642 return NULL;
2643
Vivek Goyal8682e1f2009-12-03 12:59:50 -05002644 /* If new queue belongs to different cfq_group, don't choose it */
2645 if (cur_cfqq->cfqg != cfqq->cfqg)
2646 return NULL;
2647
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002648 /*
2649 * It only makes sense to merge sync queues.
2650 */
2651 if (!cfq_cfqq_sync(cfqq))
2652 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002653 if (CFQQ_SEEKY(cfqq))
2654 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002655
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002656 /*
2657 * Do not merge queues of different priority classes
2658 */
2659 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2660 return NULL;
2661
Jens Axboea36e71f2009-04-15 12:15:11 +02002662 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02002663}
2664
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002665/*
2666 * Determine whether we should enforce idle window for this queue.
2667 */
2668
2669static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2670{
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04002671 enum wl_class_t wl_class = cfqq_class(cfqq);
Vivek Goyal34b98d02012-10-03 16:56:58 -04002672 struct cfq_rb_root *st = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002673
Vivek Goyal34b98d02012-10-03 16:56:58 -04002674 BUG_ON(!st);
2675 BUG_ON(!st->count);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002676
Vivek Goyalb6508c12010-08-23 12:23:33 +02002677 if (!cfqd->cfq_slice_idle)
2678 return false;
2679
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002680 /* We never do for idle class queues. */
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04002681 if (wl_class == IDLE_WORKLOAD)
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002682 return false;
2683
2684 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01002685 if (cfq_cfqq_idle_window(cfqq) &&
2686 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002687 return true;
2688
2689 /*
2690 * Otherwise, we do only if they are the last ones
2691 * in their service tree.
2692 */
Vivek Goyal34b98d02012-10-03 16:56:58 -04002693 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2694 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
Shaohua Lic1e44752010-11-08 15:01:02 +01002695 return true;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002696 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01002697 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002698}
2699
Jens Axboe6d048f52007-04-25 12:44:27 +02002700static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002701{
Jens Axboe17926692007-01-19 11:59:30 +11002702 struct cfq_queue *cfqq = cfqd->active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +01002703 struct cfq_io_cq *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002704 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002705
Jens Axboea68bbdd2008-09-24 13:03:33 +02002706 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002707 * SSD device without seek penalty, disable idling. But only do so
2708 * for devices that support queuing, otherwise we still have a problem
2709 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02002710 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002711 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02002712 return;
2713
Jens Axboedd67d052006-06-21 09:36:18 +02002714 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02002715 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02002716
2717 /*
2718 * idle is disabled, either manually or by past process history
2719 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002720 if (!cfq_should_idle(cfqd, cfqq)) {
2721 /* no queue idling. Check for group idling */
2722 if (cfqd->cfq_group_idle)
2723 group_idle = cfqd->cfq_group_idle;
2724 else
2725 return;
2726 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002727
Jens Axboe22e2c502005-06-27 10:55:12 +02002728 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002729 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02002730 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002731 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02002732 return;
2733
2734 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002735 * task has exited, don't wait
2736 */
Jens Axboe206dc692006-03-28 13:03:44 +02002737 cic = cfqd->active_cic;
Tejun Heof6e8d012012-03-05 13:15:26 -08002738 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
Jens Axboe6d048f52007-04-25 12:44:27 +02002739 return;
2740
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002741 /*
2742 * If our average think time is larger than the remaining time
2743 * slice, then don't idle. This avoids overrunning the allotted
2744 * time slice.
2745 */
Shaohua Li383cd722011-07-12 14:24:35 +02002746 if (sample_valid(cic->ttime.ttime_samples) &&
2747 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
Joe Perchesfd16d262011-06-13 10:42:49 +02002748 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
Shaohua Li383cd722011-07-12 14:24:35 +02002749 cic->ttime.ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002750 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002751 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002752
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002753 /* There are other queues in the group, don't do group idle */
2754 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2755 return;
2756
Jens Axboe3b181522005-06-27 10:56:24 +02002757 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002758
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002759 if (group_idle)
2760 sl = cfqd->cfq_group_idle;
2761 else
2762 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02002763
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002764 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Tejun Heo155fead2012-04-01 14:38:44 -07002765 cfqg_stats_set_start_idle_time(cfqq->cfqg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002766 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2767 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768}
2769
Jens Axboe498d3aa22007-04-26 12:54:48 +02002770/*
2771 * Move request from internal lists to the request queue dispatch list.
2772 */
Jens Axboe165125e2007-07-24 09:28:11 +02002773static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774{
Jens Axboe3ed9a292007-04-23 08:33:33 +02002775 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002776 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002777
Jens Axboe7b679132008-05-30 12:23:07 +02002778 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2779
Jeff Moyer06d21882009-09-11 17:08:59 +02002780 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02002781 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002782 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002783 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02002784 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02002785
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002786 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02002787 cfqq->nr_sectors += blk_rq_sectors(rq);
Tejun Heo155fead2012-04-01 14:38:44 -07002788 cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789}
2790
2791/*
2792 * return expired entry, or NULL to just start from scratch in rbtree
2793 */
Jens Axboefebffd62008-01-28 13:19:43 +01002794static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795{
Jens Axboe30996f42009-10-05 11:03:39 +02002796 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797
Jens Axboe3b181522005-06-27 10:56:24 +02002798 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11002800
2801 cfq_mark_cfqq_fifo_expire(cfqq);
2802
Jens Axboe89850f72006-07-22 16:48:31 +02002803 if (list_empty(&cfqq->fifo))
2804 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805
Jens Axboe89850f72006-07-22 16:48:31 +02002806 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02002807 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02002808 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809
Jens Axboe30996f42009-10-05 11:03:39 +02002810 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002811 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812}
2813
Jens Axboe22e2c502005-06-27 10:55:12 +02002814static inline int
2815cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2816{
2817 const int base_rq = cfqd->cfq_slice_async_rq;
2818
2819 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2820
Namhyung Kimb9f8ce02011-05-24 10:23:21 +02002821 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002822}
2823
2824/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002825 * Must be called with the queue_lock held.
2826 */
2827static int cfqq_process_refs(struct cfq_queue *cfqq)
2828{
2829 int process_refs, io_refs;
2830
2831 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01002832 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002833 BUG_ON(process_refs < 0);
2834 return process_refs;
2835}
2836
2837static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2838{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002839 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002840 struct cfq_queue *__cfqq;
2841
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002842 /*
2843 * If there are no process references on the new_cfqq, then it is
2844 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2845 * chain may have dropped their last reference (not just their
2846 * last process reference).
2847 */
2848 if (!cfqq_process_refs(new_cfqq))
2849 return;
2850
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002851 /* Avoid a circular list and skip interim queue merges */
2852 while ((__cfqq = new_cfqq->new_cfqq)) {
2853 if (__cfqq == cfqq)
2854 return;
2855 new_cfqq = __cfqq;
2856 }
2857
2858 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002859 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002860 /*
2861 * If the process for the cfqq has gone away, there is no
2862 * sense in merging the queues.
2863 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04002864 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002865 return;
2866
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002867 /*
2868 * Merge in the direction of the lesser amount of work.
2869 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002870 if (new_process_refs >= process_refs) {
2871 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002872 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002873 } else {
2874 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01002875 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002876 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002877}
2878
Vivek Goyal6d816ec2012-10-03 16:56:59 -04002879static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04002880 struct cfq_group *cfqg, enum wl_class_t wl_class)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002881{
2882 struct cfq_queue *queue;
2883 int i;
2884 bool key_valid = false;
2885 unsigned long lowest_key = 0;
2886 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2887
Vivek Goyal65b32a52009-12-16 17:52:59 -05002888 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2889 /* select the one with lowest rb_key */
Vivek Goyal34b98d02012-10-03 16:56:58 -04002890 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002891 if (queue &&
2892 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2893 lowest_key = queue->rb_key;
2894 cur_best = i;
2895 key_valid = true;
2896 }
2897 }
2898
2899 return cur_best;
2900}
2901
Vivek Goyal6d816ec2012-10-03 16:56:59 -04002902static void
2903choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002904{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002905 unsigned slice;
2906 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002907 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002908 unsigned group_slice;
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002909 enum wl_class_t original_class = cfqd->serving_wl_class;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002910
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002911 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002912 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002913 cfqd->serving_wl_class = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002914 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002915 cfqd->serving_wl_class = BE_WORKLOAD;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002916 else {
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002917 cfqd->serving_wl_class = IDLE_WORKLOAD;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002918 cfqd->workload_expires = jiffies + 1;
2919 return;
2920 }
2921
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002922 if (original_class != cfqd->serving_wl_class)
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002923 goto new_workload;
2924
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002925 /*
2926 * For RT and BE, we have to choose also the type
2927 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2928 * expiration time
2929 */
Vivek Goyal34b98d02012-10-03 16:56:58 -04002930 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002931 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002932
2933 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05002934 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002935 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05002936 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002937 return;
2938
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01002939new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002940 /* otherwise select new workload type */
Vivek Goyal6d816ec2012-10-03 16:56:59 -04002941 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002942 cfqd->serving_wl_class);
Vivek Goyal34b98d02012-10-03 16:56:58 -04002943 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002944 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002945
2946 /*
2947 * the workload slice is computed as a fraction of target latency
2948 * proportional to the number of queues in that workload, over
2949 * all the queues in the same priority class
2950 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002951 group_slice = cfq_group_slice(cfqd, cfqg);
2952
2953 slice = group_slice * count /
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002954 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2955 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04002956 cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002957
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002958 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002959 unsigned int tmp;
2960
2961 /*
2962 * Async queues are currently system wide. Just taking
2963 * proportion of queues with-in same group will lead to higher
2964 * async ratio system wide as generally root group is going
2965 * to have higher weight. A more accurate thing would be to
2966 * calculate system wide asnc/sync ratio.
2967 */
Tao Ma5bf14c02012-04-01 14:33:39 -07002968 tmp = cfqd->cfq_target_latency *
2969 cfqg_busy_async_queues(cfqd, cfqg);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002970 tmp = tmp/cfqd->busy_queues;
2971 slice = min_t(unsigned, slice, tmp);
2972
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002973 /* async workload slice is scaled down according to
2974 * the sync/async slice ratio. */
2975 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002976 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002977 /* sync workload slice is at least 2 * cfq_slice_idle */
2978 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2979
2980 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002981 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002982 cfqd->workload_expires = jiffies + slice;
2983}
2984
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002985static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2986{
2987 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002988 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002989
2990 if (RB_EMPTY_ROOT(&st->rb))
2991 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002992 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002993 update_min_vdisktime(st);
2994 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002995}
2996
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002997static void cfq_choose_cfqg(struct cfq_data *cfqd)
2998{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002999 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3000
3001 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05003002
3003 /* Restore the workload type data */
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003004 if (cfqg->saved_wl_slice) {
3005 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3006 cfqd->serving_wl_type = cfqg->saved_wl_type;
3007 cfqd->serving_wl_class = cfqg->saved_wl_class;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01003008 } else
3009 cfqd->workload_expires = jiffies - 1;
3010
Vivek Goyal6d816ec2012-10-03 16:56:59 -04003011 choose_wl_class_and_type(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003012}
3013
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003014/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02003015 * Select a queue for service. If we have a current active queue,
3016 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02003017 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003018static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02003019{
Jens Axboea36e71f2009-04-15 12:15:11 +02003020 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02003021
3022 cfqq = cfqd->active_queue;
3023 if (!cfqq)
3024 goto new_queue;
3025
Vivek Goyalf04a6422009-12-03 12:59:40 -05003026 if (!cfqd->rq_queued)
3027 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05003028
3029 /*
3030 * We were waiting for group to get backlogged. Expire the queue
3031 */
3032 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3033 goto expire;
3034
Jens Axboe22e2c502005-06-27 10:55:12 +02003035 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02003036 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02003037 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003038 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3039 /*
3040 * If slice had not expired at the completion of last request
3041 * we might not have turned on wait_busy flag. Don't expire
3042 * the queue yet. Allow the group to get backlogged.
3043 *
3044 * The very fact that we have used the slice, that means we
3045 * have been idling all along on this queue and it should be
3046 * ok to wait for this request to complete.
3047 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01003048 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3049 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3050 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05003051 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01003052 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003053 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05003054 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003055
3056 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02003057 * The active queue has requests and isn't expired, allow it to
3058 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02003059 */
Jens Axboedd67d052006-06-21 09:36:18 +02003060 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003061 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02003062
3063 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02003064 * If another queue has a request waiting within our mean seek
3065 * distance, let it run. The expire code will check for close
3066 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003067 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02003068 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003069 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003070 if (new_cfqq) {
3071 if (!cfqq->new_cfqq)
3072 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02003073 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003074 }
Jens Axboea36e71f2009-04-15 12:15:11 +02003075
3076 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02003077 * No requests pending. If the active queue still has requests in
3078 * flight or is idling for a new request, allow either of these
3079 * conditions to happen (or time out) before selecting a new queue.
3080 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003081 if (timer_pending(&cfqd->idle_slice_timer)) {
3082 cfqq = NULL;
3083 goto keep_queue;
3084 }
3085
Shaohua Li8e1ac662010-11-08 15:01:04 +01003086 /*
3087 * This is a deep seek queue, but the device is much faster than
3088 * the queue can deliver, don't idle
3089 **/
3090 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3091 (cfq_cfqq_slice_new(cfqq) ||
3092 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3093 cfq_clear_cfqq_deep(cfqq);
3094 cfq_clear_cfqq_idle_window(cfqq);
3095 }
3096
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003097 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3098 cfqq = NULL;
3099 goto keep_queue;
3100 }
3101
3102 /*
3103 * If group idle is enabled and there are requests dispatched from
3104 * this group, wait for requests to complete.
3105 */
3106check_group_idle:
Shaohua Li7700fc42011-07-12 14:24:56 +02003107 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3108 cfqq->cfqg->dispatched &&
3109 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02003110 cfqq = NULL;
3111 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003112 }
3113
Jens Axboe3b181522005-06-27 10:56:24 +02003114expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003115 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02003116new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003117 /*
3118 * Current queue expired. Check if we have to switch to a new
3119 * service tree
3120 */
3121 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003122 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003123
Jens Axboea36e71f2009-04-15 12:15:11 +02003124 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003125keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02003126 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003127}
3128
Jens Axboefebffd62008-01-28 13:19:43 +01003129static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02003130{
3131 int dispatched = 0;
3132
3133 while (cfqq->next_rq) {
3134 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3135 dispatched++;
3136 }
3137
3138 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05003139
3140 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02003141 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02003142 return dispatched;
3143}
3144
Jens Axboe498d3aa22007-04-26 12:54:48 +02003145/*
3146 * Drain our current requests. Used for barriers and when switching
3147 * io schedulers on-the-fly.
3148 */
Jens Axboed9e76202007-04-20 14:27:50 +02003149static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003150{
Jens Axboe08717142008-01-28 11:38:15 +01003151 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02003152 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003153
Divyesh Shah3440c492010-04-09 09:29:57 +02003154 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02003155 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02003156 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3157 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05003158 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02003159 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003160
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003161 BUG_ON(cfqd->busy_queues);
3162
Jeff Moyer69237152009-06-12 15:29:30 +02003163 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003164 return dispatched;
3165}
3166
Shaohua Liabc3c742010-03-01 09:20:54 +01003167static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3168 struct cfq_queue *cfqq)
3169{
3170 /* the queue hasn't finished any request, can't estimate */
3171 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01003172 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01003173 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3174 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01003175 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01003176
Shaohua Lic1e44752010-11-08 15:01:02 +01003177 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01003178}
3179
Jens Axboe0b182d62009-10-06 20:49:37 +02003180static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02003181{
Jens Axboe2f5cb732009-04-07 08:51:19 +02003182 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183
Jens Axboe2f5cb732009-04-07 08:51:19 +02003184 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02003185 * Drain async requests before we start sync IO
3186 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003187 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02003188 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02003189
3190 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02003191 * If this is an async queue and we have sync IO in flight, let it wait
3192 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003193 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02003194 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02003195
Shaohua Liabc3c742010-03-01 09:20:54 +01003196 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02003197 if (cfq_class_idle(cfqq))
3198 max_dispatch = 1;
3199
3200 /*
3201 * Does this cfqq already have too much IO in flight?
3202 */
3203 if (cfqq->dispatched >= max_dispatch) {
Shaohua Lief8a41d2011-03-07 09:26:29 +01003204 bool promote_sync = false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02003205 /*
3206 * idle queue must always only have a single IO in flight
3207 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02003208 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02003209 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003210
Jens Axboe2f5cb732009-04-07 08:51:19 +02003211 /*
Li, Shaohuac4ade942011-03-23 08:30:34 +01003212 * If there is only one sync queue
3213 * we can ignore async queue here and give the sync
Shaohua Lief8a41d2011-03-07 09:26:29 +01003214 * queue no dispatch limit. The reason is a sync queue can
3215 * preempt async queue, limiting the sync queue doesn't make
3216 * sense. This is useful for aiostress test.
3217 */
Li, Shaohuac4ade942011-03-23 08:30:34 +01003218 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3219 promote_sync = true;
Shaohua Lief8a41d2011-03-07 09:26:29 +01003220
3221 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02003222 * We have other queues, don't allow more IO from this one
3223 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01003224 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3225 !promote_sync)
Jens Axboe0b182d62009-10-06 20:49:37 +02003226 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11003227
Jens Axboe2f5cb732009-04-07 08:51:19 +02003228 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01003229 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02003230 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01003231 if (cfqd->busy_queues == 1 || promote_sync)
Shaohua Liabc3c742010-03-01 09:20:54 +01003232 max_dispatch = -1;
3233 else
3234 /*
3235 * Normally we start throttling cfqq when cfq_quantum/2
3236 * requests have been dispatched. But we can drive
3237 * deeper queue depths at the beginning of slice
3238 * subjected to upper limit of cfq_quantum.
3239 * */
3240 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02003241 }
3242
3243 /*
3244 * Async queues must wait a bit before being allowed dispatch.
3245 * We also ramp up the dispatch depth gradually for async IO,
3246 * based on the last sync IO we serviced
3247 */
Jens Axboe963b72f2009-10-03 19:42:18 +02003248 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003249 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02003250 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02003251
Jens Axboe61f0c1d2009-10-03 19:46:03 +02003252 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02003253 if (!depth && !cfqq->dispatched)
3254 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02003255 if (depth < max_dispatch)
3256 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257 }
3258
Jens Axboe0b182d62009-10-06 20:49:37 +02003259 /*
3260 * If we're below the current max, allow a dispatch
3261 */
3262 return cfqq->dispatched < max_dispatch;
3263}
3264
3265/*
3266 * Dispatch a request from cfqq, moving them to the request queue
3267 * dispatch list.
3268 */
3269static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3270{
3271 struct request *rq;
3272
3273 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3274
3275 if (!cfq_may_dispatch(cfqd, cfqq))
3276 return false;
3277
3278 /*
3279 * follow expired path, else get first next available
3280 */
3281 rq = cfq_check_fifo(cfqq);
3282 if (!rq)
3283 rq = cfqq->next_rq;
3284
3285 /*
3286 * insert request into driver dispatch list
3287 */
3288 cfq_dispatch_insert(cfqd->queue, rq);
3289
3290 if (!cfqd->active_cic) {
Tejun Heoc5869802011-12-14 00:33:41 +01003291 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe0b182d62009-10-06 20:49:37 +02003292
Tejun Heoc5869802011-12-14 00:33:41 +01003293 atomic_long_inc(&cic->icq.ioc->refcount);
Jens Axboe0b182d62009-10-06 20:49:37 +02003294 cfqd->active_cic = cic;
3295 }
3296
3297 return true;
3298}
3299
3300/*
3301 * Find the cfqq that we need to service and move a request from that to the
3302 * dispatch list
3303 */
3304static int cfq_dispatch_requests(struct request_queue *q, int force)
3305{
3306 struct cfq_data *cfqd = q->elevator->elevator_data;
3307 struct cfq_queue *cfqq;
3308
3309 if (!cfqd->busy_queues)
3310 return 0;
3311
3312 if (unlikely(force))
3313 return cfq_forced_dispatch(cfqd);
3314
3315 cfqq = cfq_select_queue(cfqd);
3316 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02003317 return 0;
3318
Jens Axboe2f5cb732009-04-07 08:51:19 +02003319 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02003320 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02003321 */
Jens Axboe0b182d62009-10-06 20:49:37 +02003322 if (!cfq_dispatch_request(cfqd, cfqq))
3323 return 0;
3324
Jens Axboe2f5cb732009-04-07 08:51:19 +02003325 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02003326 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02003327
3328 /*
3329 * expire an async queue immediately if it has used up its slice. idle
3330 * queue always expire after 1 dispatch round.
3331 */
3332 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3333 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3334 cfq_class_idle(cfqq))) {
3335 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02003336 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02003337 }
3338
Shan Weib217a902009-09-01 10:06:42 +02003339 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02003340 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003341}
3342
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343/*
Jens Axboe5e705372006-07-13 12:39:25 +02003344 * task holds one reference to the queue, dropped when task exits. each rq
3345 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05003347 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348 * queue lock must be held here.
3349 */
3350static void cfq_put_queue(struct cfq_queue *cfqq)
3351{
Jens Axboe22e2c502005-06-27 10:55:12 +02003352 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05003353 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02003354
Shaohua Li30d7b942011-01-07 08:46:59 +01003355 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003356
Shaohua Li30d7b942011-01-07 08:46:59 +01003357 cfqq->ref--;
3358 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359 return;
3360
Jens Axboe7b679132008-05-30 12:23:07 +02003361 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02003363 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003364 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365
Jens Axboe28f95cbc2007-01-19 12:09:53 +11003366 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02003367 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02003368 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11003369 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003370
Vivek Goyalf04a6422009-12-03 12:59:40 -05003371 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003372 kmem_cache_free(cfq_pool, cfqq);
Tejun Heoeb7d8c072012-03-23 14:02:53 +01003373 cfqg_put(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003374}
3375
Shaohua Lid02a2c02010-05-25 10:16:53 +02003376static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02003377{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003378 struct cfq_queue *__cfqq, *next;
3379
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003380 /*
3381 * If this queue was scheduled to merge with another queue, be
3382 * sure to drop the reference taken on that queue (and others in
3383 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3384 */
3385 __cfqq = cfqq->new_cfqq;
3386 while (__cfqq) {
3387 if (__cfqq == cfqq) {
3388 WARN(1, "cfqq->new_cfqq loop detected\n");
3389 break;
3390 }
3391 next = __cfqq->new_cfqq;
3392 cfq_put_queue(__cfqq);
3393 __cfqq = next;
3394 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02003395}
3396
3397static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3398{
3399 if (unlikely(cfqq == cfqd->active_queue)) {
3400 __cfq_slice_expired(cfqd, cfqq, 0);
3401 cfq_schedule_dispatch(cfqd);
3402 }
3403
3404 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003405
Jens Axboe89850f72006-07-22 16:48:31 +02003406 cfq_put_queue(cfqq);
3407}
3408
Tejun Heo9b84cac2011-12-14 00:33:42 +01003409static void cfq_init_icq(struct io_cq *icq)
3410{
3411 struct cfq_io_cq *cic = icq_to_cic(icq);
3412
3413 cic->ttime.last_end_request = jiffies;
3414}
3415
Tejun Heoc5869802011-12-14 00:33:41 +01003416static void cfq_exit_icq(struct io_cq *icq)
Jens Axboe89850f72006-07-22 16:48:31 +02003417{
Tejun Heoc5869802011-12-14 00:33:41 +01003418 struct cfq_io_cq *cic = icq_to_cic(icq);
Tejun Heo283287a2011-12-14 00:33:38 +01003419 struct cfq_data *cfqd = cic_to_cfqd(cic);
Fabio Checconi4faa3c82008-04-10 08:28:01 +02003420
Jens Axboeff6657c2009-04-08 10:58:57 +02003421 if (cic->cfqq[BLK_RW_ASYNC]) {
3422 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3423 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02003424 }
3425
Jens Axboeff6657c2009-04-08 10:58:57 +02003426 if (cic->cfqq[BLK_RW_SYNC]) {
3427 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3428 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02003429 }
Jens Axboe89850f72006-07-22 16:48:31 +02003430}
3431
Tejun Heoabede6d2012-03-19 15:10:57 -07003432static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02003433{
3434 struct task_struct *tsk = current;
3435 int ioprio_class;
3436
Jens Axboe3b181522005-06-27 10:56:24 +02003437 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003438 return;
3439
Tejun Heo598971b2012-03-19 15:10:58 -07003440 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02003441 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01003442 default:
3443 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3444 case IOPRIO_CLASS_NONE:
3445 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02003446 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01003447 */
3448 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02003449 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01003450 break;
3451 case IOPRIO_CLASS_RT:
Tejun Heo598971b2012-03-19 15:10:58 -07003452 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
Jens Axboefe094d92008-01-31 13:08:54 +01003453 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3454 break;
3455 case IOPRIO_CLASS_BE:
Tejun Heo598971b2012-03-19 15:10:58 -07003456 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
Jens Axboefe094d92008-01-31 13:08:54 +01003457 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3458 break;
3459 case IOPRIO_CLASS_IDLE:
3460 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3461 cfqq->ioprio = 7;
3462 cfq_clear_cfqq_idle_window(cfqq);
3463 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02003464 }
3465
3466 /*
3467 * keep track of original prio settings in case we have to temporarily
3468 * elevate the priority of this queue
3469 */
3470 cfqq->org_ioprio = cfqq->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02003471 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003472}
3473
Tejun Heo598971b2012-03-19 15:10:58 -07003474static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
Jens Axboe22e2c502005-06-27 10:55:12 +02003475{
Tejun Heo598971b2012-03-19 15:10:58 -07003476 int ioprio = cic->icq.ioc->ioprio;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003477 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05003478 struct cfq_queue *cfqq;
Jens Axboe35e60772006-06-14 09:10:45 +02003479
Tejun Heo598971b2012-03-19 15:10:58 -07003480 /*
3481 * Check whether ioprio has changed. The condition may trigger
3482 * spuriously on a newly created cic but there's no harm.
3483 */
3484 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
Jens Axboecaaa5f92006-06-16 11:23:00 +02003485 return;
3486
Jens Axboeff6657c2009-04-08 10:58:57 +02003487 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02003488 if (cfqq) {
3489 struct cfq_queue *new_cfqq;
Tejun Heoabede6d2012-03-19 15:10:57 -07003490 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3491 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02003492 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02003493 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02003494 cfq_put_queue(cfqq);
3495 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003496 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003497
Jens Axboeff6657c2009-04-08 10:58:57 +02003498 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02003499 if (cfqq)
3500 cfq_mark_cfqq_prio_changed(cfqq);
Tejun Heo598971b2012-03-19 15:10:58 -07003501
3502 cic->ioprio = ioprio;
Jens Axboe22e2c502005-06-27 10:55:12 +02003503}
3504
Jens Axboed5036d72009-06-26 10:44:34 +02003505static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02003506 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02003507{
3508 RB_CLEAR_NODE(&cfqq->rb_node);
3509 RB_CLEAR_NODE(&cfqq->p_node);
3510 INIT_LIST_HEAD(&cfqq->fifo);
3511
Shaohua Li30d7b942011-01-07 08:46:59 +01003512 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02003513 cfqq->cfqd = cfqd;
3514
3515 cfq_mark_cfqq_prio_changed(cfqq);
3516
3517 if (is_sync) {
3518 if (!cfq_class_idle(cfqq))
3519 cfq_mark_cfqq_idle_window(cfqq);
3520 cfq_mark_cfqq_sync(cfqq);
3521 }
3522 cfqq->pid = pid;
3523}
3524
Vivek Goyal246103332009-12-03 12:59:51 -05003525#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo598971b2012-03-19 15:10:58 -07003526static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
Vivek Goyal246103332009-12-03 12:59:51 -05003527{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003528 struct cfq_data *cfqd = cic_to_cfqd(cic);
Tejun Heo598971b2012-03-19 15:10:58 -07003529 struct cfq_queue *sync_cfqq;
3530 uint64_t id;
Vivek Goyal246103332009-12-03 12:59:51 -05003531
Tejun Heo598971b2012-03-19 15:10:58 -07003532 rcu_read_lock();
Tejun Heo3c798392012-04-16 13:57:25 -07003533 id = bio_blkcg(bio)->id;
Tejun Heo598971b2012-03-19 15:10:58 -07003534 rcu_read_unlock();
3535
3536 /*
3537 * Check whether blkcg has changed. The condition may trigger
3538 * spuriously on a newly created cic but there's no harm.
3539 */
3540 if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
Vivek Goyal246103332009-12-03 12:59:51 -05003541 return;
3542
Tejun Heo598971b2012-03-19 15:10:58 -07003543 sync_cfqq = cic_to_cfqq(cic, 1);
Vivek Goyal246103332009-12-03 12:59:51 -05003544 if (sync_cfqq) {
3545 /*
3546 * Drop reference to sync queue. A new sync queue will be
3547 * assigned in new group upon arrival of a fresh request.
3548 */
3549 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3550 cic_set_cfqq(cic, NULL, 1);
3551 cfq_put_queue(sync_cfqq);
3552 }
Tejun Heo598971b2012-03-19 15:10:58 -07003553
3554 cic->blkcg_id = id;
Vivek Goyal246103332009-12-03 12:59:51 -05003555}
Tejun Heo598971b2012-03-19 15:10:58 -07003556#else
3557static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
Vivek Goyal246103332009-12-03 12:59:51 -05003558#endif /* CONFIG_CFQ_GROUP_IOSCHED */
3559
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560static struct cfq_queue *
Tejun Heoabede6d2012-03-19 15:10:57 -07003561cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3562 struct bio *bio, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563{
Tejun Heo3c798392012-04-16 13:57:25 -07003564 struct blkcg *blkcg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003566 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567
3568retry:
Tejun Heo2a7f1242012-03-05 13:15:01 -08003569 rcu_read_lock();
3570
Tejun Heo3c798392012-04-16 13:57:25 -07003571 blkcg = bio_blkcg(bio);
Tejun Heocd1604f2012-03-05 13:15:06 -08003572 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003573 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574
Jens Axboe6118b702009-06-30 09:34:12 +02003575 /*
3576 * Always try a new alloc if we fell back to the OOM cfqq
3577 * originally, since it should just be a temporary situation.
3578 */
3579 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3580 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 if (new_cfqq) {
3582 cfqq = new_cfqq;
3583 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02003584 } else if (gfp_mask & __GFP_WAIT) {
Tejun Heo2a7f1242012-03-05 13:15:01 -08003585 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07003587 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02003588 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07003589 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02003591 if (new_cfqq)
3592 goto retry;
Glauber Costaa3cc86c2013-02-21 15:16:41 -08003593 else
3594 return &cfqd->oom_cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003595 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07003596 cfqq = kmem_cache_alloc_node(cfq_pool,
3597 gfp_mask | __GFP_ZERO,
3598 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02003599 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600
Jens Axboe6118b702009-06-30 09:34:12 +02003601 if (cfqq) {
3602 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
Tejun Heoabede6d2012-03-19 15:10:57 -07003603 cfq_init_prio_data(cfqq, cic);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003604 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02003605 cfq_log_cfqq(cfqd, cfqq, "alloced");
3606 } else
3607 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608 }
3609
3610 if (new_cfqq)
3611 kmem_cache_free(cfq_pool, new_cfqq);
3612
Tejun Heo2a7f1242012-03-05 13:15:01 -08003613 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614 return cfqq;
3615}
3616
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003617static struct cfq_queue **
3618cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3619{
Jens Axboefe094d92008-01-31 13:08:54 +01003620 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003621 case IOPRIO_CLASS_RT:
3622 return &cfqd->async_cfqq[0][ioprio];
Tejun Heo598971b2012-03-19 15:10:58 -07003623 case IOPRIO_CLASS_NONE:
3624 ioprio = IOPRIO_NORM;
3625 /* fall through */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003626 case IOPRIO_CLASS_BE:
3627 return &cfqd->async_cfqq[1][ioprio];
3628 case IOPRIO_CLASS_IDLE:
3629 return &cfqd->async_idle_cfqq;
3630 default:
3631 BUG();
3632 }
3633}
3634
Jens Axboe15c31be2007-07-10 13:43:25 +02003635static struct cfq_queue *
Tejun Heoabede6d2012-03-19 15:10:57 -07003636cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
Tejun Heo4f85cb92012-03-05 13:15:28 -08003637 struct bio *bio, gfp_t gfp_mask)
Jens Axboe15c31be2007-07-10 13:43:25 +02003638{
Tejun Heo598971b2012-03-19 15:10:58 -07003639 const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3640 const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003641 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02003642 struct cfq_queue *cfqq = NULL;
3643
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003644 if (!is_sync) {
3645 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3646 cfqq = *async_cfqq;
3647 }
3648
Jens Axboe6118b702009-06-30 09:34:12 +02003649 if (!cfqq)
Tejun Heoabede6d2012-03-19 15:10:57 -07003650 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02003651
3652 /*
3653 * pin the queue now that it's allocated, scheduler exit will prune it
3654 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003655 if (!is_sync && !(*async_cfqq)) {
Shaohua Li30d7b942011-01-07 08:46:59 +01003656 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003657 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02003658 }
3659
Shaohua Li30d7b942011-01-07 08:46:59 +01003660 cfqq->ref++;
Jens Axboe15c31be2007-07-10 13:43:25 +02003661 return cfqq;
3662}
3663
Jens Axboe22e2c502005-06-27 10:55:12 +02003664static void
Shaohua Li383cd722011-07-12 14:24:35 +02003665__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003666{
Shaohua Li383cd722011-07-12 14:24:35 +02003667 unsigned long elapsed = jiffies - ttime->last_end_request;
3668 elapsed = min(elapsed, 2UL * slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003669
Shaohua Li383cd722011-07-12 14:24:35 +02003670 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3671 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3672 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3673}
3674
3675static void
3676cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01003677 struct cfq_io_cq *cic)
Shaohua Li383cd722011-07-12 14:24:35 +02003678{
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003679 if (cfq_cfqq_sync(cfqq)) {
Shaohua Li383cd722011-07-12 14:24:35 +02003680 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003681 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3682 cfqd->cfq_slice_idle);
3683 }
Shaohua Li7700fc42011-07-12 14:24:56 +02003684#ifdef CONFIG_CFQ_GROUP_IOSCHED
3685 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3686#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003687}
3688
Jens Axboe206dc692006-03-28 13:03:44 +02003689static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003690cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003691 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003692{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003693 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003694 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003695 if (cfqq->last_request_pos) {
3696 if (cfqq->last_request_pos < blk_rq_pos(rq))
3697 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3698 else
3699 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3700 }
Jens Axboe206dc692006-03-28 13:03:44 +02003701
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003702 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003703 if (blk_queue_nonrot(cfqd->queue))
3704 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3705 else
3706 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003707}
Jens Axboe22e2c502005-06-27 10:55:12 +02003708
3709/*
3710 * Disable idle window if the process thinks too long or seeks so much that
3711 * it doesn't matter
3712 */
3713static void
3714cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01003715 struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02003716{
Jens Axboe7b679132008-05-30 12:23:07 +02003717 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003718
Jens Axboe08717142008-01-28 11:38:15 +01003719 /*
3720 * Don't idle for async or idle io prio class
3721 */
3722 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003723 return;
3724
Jens Axboec265a7f2008-06-26 13:49:33 +02003725 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003726
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003727 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3728 cfq_mark_cfqq_deep(cfqq);
3729
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003730 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3731 enable_idle = 0;
Tejun Heof6e8d012012-03-05 13:15:26 -08003732 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
Tejun Heoc5869802011-12-14 00:33:41 +01003733 !cfqd->cfq_slice_idle ||
3734 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003735 enable_idle = 0;
Shaohua Li383cd722011-07-12 14:24:35 +02003736 else if (sample_valid(cic->ttime.ttime_samples)) {
3737 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003738 enable_idle = 0;
3739 else
3740 enable_idle = 1;
3741 }
3742
Jens Axboe7b679132008-05-30 12:23:07 +02003743 if (old_idle != enable_idle) {
3744 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3745 if (enable_idle)
3746 cfq_mark_cfqq_idle_window(cfqq);
3747 else
3748 cfq_clear_cfqq_idle_window(cfqq);
3749 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003750}
3751
Jens Axboe22e2c502005-06-27 10:55:12 +02003752/*
3753 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3754 * no or if we aren't sure, a 1 will cause a preempt.
3755 */
Jens Axboea6151c32009-10-07 20:02:57 +02003756static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003757cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003758 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003759{
Jens Axboe6d048f52007-04-25 12:44:27 +02003760 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003761
Jens Axboe6d048f52007-04-25 12:44:27 +02003762 cfqq = cfqd->active_queue;
3763 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003764 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003765
Jens Axboe6d048f52007-04-25 12:44:27 +02003766 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003767 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003768
3769 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003770 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003771
Jens Axboe22e2c502005-06-27 10:55:12 +02003772 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003773 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3774 */
3775 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3776 return false;
3777
3778 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003779 * if the new request is sync, but the currently running queue is
3780 * not, let the sync request have priority.
3781 */
Jens Axboe5e705372006-07-13 12:39:25 +02003782 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003783 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003784
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003785 if (new_cfqq->cfqg != cfqq->cfqg)
3786 return false;
3787
3788 if (cfq_slice_used(cfqq))
3789 return true;
3790
3791 /* Allow preemption only if we are idling on sync-noidle tree */
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003792 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003793 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3794 new_cfqq->service_tree->count == 2 &&
3795 RB_EMPTY_ROOT(&cfqq->sort_list))
3796 return true;
3797
Jens Axboe374f84a2006-07-23 01:42:19 +02003798 /*
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003799 * So both queues are sync. Let the new request get disk time if
3800 * it's a metadata request and the current queue is doing regular IO.
3801 */
Christoph Hellwig65299a32011-08-23 14:50:29 +02003802 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003803 return true;
3804
3805 /*
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003806 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3807 */
3808 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003809 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003810
Shaohua Lid2d59e12010-11-08 15:01:03 +01003811 /* An idle queue should not be idle now for some reason */
3812 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3813 return true;
3814
Jens Axboe1e3335d2007-02-14 19:59:49 +01003815 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003816 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003817
3818 /*
3819 * if this request is as-good as one we would expect from the
3820 * current cfqq, let it preempt
3821 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01003822 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003823 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003824
Jens Axboea6151c32009-10-07 20:02:57 +02003825 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003826}
3827
3828/*
3829 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3830 * let it have half of its nominal slice.
3831 */
3832static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3833{
Shaohua Lidf0793a2012-01-19 09:20:09 +01003834 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3835
Jens Axboe7b679132008-05-30 12:23:07 +02003836 cfq_log_cfqq(cfqd, cfqq, "preempt");
Shaohua Lidf0793a2012-01-19 09:20:09 +01003837 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003838
Jens Axboebf572252006-07-19 20:29:12 +02003839 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003840 * workload type is changed, don't save slice, otherwise preempt
3841 * doesn't happen
3842 */
Shaohua Lidf0793a2012-01-19 09:20:09 +01003843 if (old_type != cfqq_type(cfqq))
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003844 cfqq->cfqg->saved_wl_slice = 0;
Shaohua Lif8ae6e32011-01-14 08:41:02 +01003845
3846 /*
Jens Axboebf572252006-07-19 20:29:12 +02003847 * Put the new queue at the front of the of the current list,
3848 * so we know that it will be selected next.
3849 */
3850 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003851
3852 cfq_service_tree_add(cfqd, cfqq, 1);
Justin TerAvesteda5e0c2011-03-22 21:26:49 +01003853
Justin TerAvest62a37f62011-03-23 08:25:44 +01003854 cfqq->slice_end = 0;
3855 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003856}
3857
3858/*
Jens Axboe5e705372006-07-13 12:39:25 +02003859 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003860 * something we should do about it
3861 */
3862static void
Jens Axboe5e705372006-07-13 12:39:25 +02003863cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3864 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003865{
Tejun Heoc5869802011-12-14 00:33:41 +01003866 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003867
Aaron Carroll45333d52008-08-26 15:52:36 +02003868 cfqd->rq_queued++;
Christoph Hellwig65299a32011-08-23 14:50:29 +02003869 if (rq->cmd_flags & REQ_PRIO)
3870 cfqq->prio_pending++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003871
Shaohua Li383cd722011-07-12 14:24:35 +02003872 cfq_update_io_thinktime(cfqd, cfqq, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003873 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003874 cfq_update_idle_window(cfqd, cfqq, cic);
3875
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003876 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003877
3878 if (cfqq == cfqd->active_queue) {
3879 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003880 * Remember that we saw a request from this process, but
3881 * don't start queuing just yet. Otherwise we risk seeing lots
3882 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003883 * and merging. If the request is already larger than a single
3884 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003885 * merging is already done. Ditto for a busy system that
3886 * has other work pending, don't risk delaying until the
3887 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003888 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003889 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003890 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3891 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07003892 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003893 cfq_clear_cfqq_wait_request(cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003894 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003895 } else {
Tejun Heo155fead2012-04-01 14:38:44 -07003896 cfqg_stats_update_idle_time(cfqq->cfqg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003897 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02003898 }
Jens Axboed6ceb252009-04-14 14:18:16 +02003899 }
Jens Axboe5e705372006-07-13 12:39:25 +02003900 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003901 /*
3902 * not the active queue - expire current slice if it is
3903 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003904 * has some old slice time left and is of higher priority or
3905 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003906 */
3907 cfq_preempt_queue(cfqd, cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02003908 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003909 }
3910}
3911
Jens Axboe165125e2007-07-24 09:28:11 +02003912static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003913{
Jens Axboeb4878f22005-10-20 16:42:29 +02003914 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003915 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003916
Jens Axboe7b679132008-05-30 12:23:07 +02003917 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Tejun Heoabede6d2012-03-19 15:10:57 -07003918 cfq_init_prio_data(cfqq, RQ_CIC(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919
Jens Axboe30996f42009-10-05 11:03:39 +02003920 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003921 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003922 cfq_add_rq_rb(rq);
Tejun Heo155fead2012-04-01 14:38:44 -07003923 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3924 rq->cmd_flags);
Jens Axboe5e705372006-07-13 12:39:25 +02003925 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926}
3927
Aaron Carroll45333d52008-08-26 15:52:36 +02003928/*
3929 * Update hw_tag based on peak queue depth over 50 samples under
3930 * sufficient load.
3931 */
3932static void cfq_update_hw_tag(struct cfq_data *cfqd)
3933{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003934 struct cfq_queue *cfqq = cfqd->active_queue;
3935
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003936 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3937 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003938
3939 if (cfqd->hw_tag == 1)
3940 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003941
3942 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003943 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003944 return;
3945
Shaohua Li1a1238a2009-10-27 08:46:23 +01003946 /*
3947 * If active queue hasn't enough requests and can idle, cfq might not
3948 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3949 * case
3950 */
3951 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3952 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003953 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01003954 return;
3955
Aaron Carroll45333d52008-08-26 15:52:36 +02003956 if (cfqd->hw_tag_samples++ < 50)
3957 return;
3958
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003959 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003960 cfqd->hw_tag = 1;
3961 else
3962 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003963}
3964
Vivek Goyal7667aa02009-12-08 17:52:58 -05003965static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3966{
Tejun Heoc5869802011-12-14 00:33:41 +01003967 struct cfq_io_cq *cic = cfqd->active_cic;
Vivek Goyal7667aa02009-12-08 17:52:58 -05003968
Justin TerAvest02a8f012011-02-09 14:20:03 +01003969 /* If the queue already has requests, don't wait */
3970 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3971 return false;
3972
Vivek Goyal7667aa02009-12-08 17:52:58 -05003973 /* If there are other queues in the group, don't wait */
3974 if (cfqq->cfqg->nr_cfqq > 1)
3975 return false;
3976
Shaohua Li7700fc42011-07-12 14:24:56 +02003977 /* the only queue in the group, but think time is big */
3978 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3979 return false;
3980
Vivek Goyal7667aa02009-12-08 17:52:58 -05003981 if (cfq_slice_used(cfqq))
3982 return true;
3983
3984 /* if slice left is less than think time, wait busy */
Shaohua Li383cd722011-07-12 14:24:35 +02003985 if (cic && sample_valid(cic->ttime.ttime_samples)
3986 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
Vivek Goyal7667aa02009-12-08 17:52:58 -05003987 return true;
3988
3989 /*
3990 * If think times is less than a jiffy than ttime_mean=0 and above
3991 * will not be true. It might happen that slice has not expired yet
3992 * but will expire soon (4-5 ns) during select_queue(). To cover the
3993 * case where think time is less than a jiffy, mark the queue wait
3994 * busy if only 1 jiffy is left in the slice.
3995 */
3996 if (cfqq->slice_end - jiffies == 1)
3997 return true;
3998
3999 return false;
4000}
4001
Jens Axboe165125e2007-07-24 09:28:11 +02004002static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003{
Jens Axboe5e705372006-07-13 12:39:25 +02004004 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02004005 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02004006 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02004007 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008
Jens Axboeb4878f22005-10-20 16:42:29 +02004009 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02004010 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4011 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012
Aaron Carroll45333d52008-08-26 15:52:36 +02004013 cfq_update_hw_tag(cfqd);
4014
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004015 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02004016 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004017 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02004018 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004019 (RQ_CFQG(rq))->dispatched--;
Tejun Heo155fead2012-04-01 14:38:44 -07004020 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4021 rq_io_start_time_ns(rq), rq->cmd_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004023 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02004024
Vivek Goyal365722b2009-10-03 15:21:27 +02004025 if (sync) {
Vivek Goyal34b98d02012-10-03 16:56:58 -04004026 struct cfq_rb_root *st;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02004027
Shaohua Li383cd722011-07-12 14:24:35 +02004028 RQ_CIC(rq)->ttime.last_end_request = now;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02004029
4030 if (cfq_cfqq_on_rr(cfqq))
Vivek Goyal34b98d02012-10-03 16:56:58 -04004031 st = cfqq->service_tree;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02004032 else
Vivek Goyal34b98d02012-10-03 16:56:58 -04004033 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4034 cfqq_type(cfqq));
4035
4036 st->ttime.last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01004037 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4038 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02004039 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02004040
Shaohua Li7700fc42011-07-12 14:24:56 +02004041#ifdef CONFIG_CFQ_GROUP_IOSCHED
4042 cfqq->cfqg->ttime.last_end_request = now;
4043#endif
4044
Jens Axboecaaa5f92006-06-16 11:23:00 +02004045 /*
4046 * If this is the active queue, check if it needs to be expired,
4047 * or if we want to idle in case it has no pending requests.
4048 */
4049 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02004050 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4051
Jens Axboe44f7c162007-01-19 11:51:58 +11004052 if (cfq_cfqq_slice_new(cfqq)) {
4053 cfq_set_prio_slice(cfqd, cfqq);
4054 cfq_clear_cfqq_slice_new(cfqq);
4055 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05004056
4057 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05004058 * Should we wait for next request to come in before we expire
4059 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05004060 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05004061 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004062 unsigned long extend_sl = cfqd->cfq_slice_idle;
4063 if (!cfqd->cfq_slice_idle)
4064 extend_sl = cfqd->cfq_group_idle;
4065 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05004066 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01004067 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05004068 }
4069
Jens Axboea36e71f2009-04-15 12:15:11 +02004070 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01004071 * Idling is not enabled on:
4072 * - expired queues
4073 * - idle-priority queues
4074 * - async queues
4075 * - queues with still some requests queued
4076 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02004077 */
Jens Axboe08717142008-01-28 11:38:15 +01004078 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02004079 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01004080 else if (sync && cfqq_empty &&
4081 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02004082 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01004083 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02004084 }
Jens Axboe6d048f52007-04-25 12:44:27 +02004085
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004086 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02004087 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088}
4089
Jens Axboe89850f72006-07-22 16:48:31 +02004090static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02004091{
Jens Axboe1b379d82009-08-11 08:26:11 +02004092 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02004093 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004094 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02004095 }
Jens Axboe22e2c502005-06-27 10:55:12 +02004096
4097 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02004098}
4099
Jens Axboe165125e2007-07-24 09:28:11 +02004100static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02004101{
4102 struct cfq_data *cfqd = q->elevator->elevator_data;
4103 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01004104 struct cfq_io_cq *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02004105 struct cfq_queue *cfqq;
4106
4107 /*
4108 * don't force setup of a queue from here, as a call to may_queue
4109 * does not necessarily imply that a request actually will be queued.
4110 * so just lookup a possibly existing queue, or return 'may queue'
4111 * if that fails
4112 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01004113 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02004114 if (!cic)
4115 return ELV_MQUEUE_MAY;
4116
Jens Axboeb0b78f82009-04-08 10:56:08 +02004117 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02004118 if (cfqq) {
Tejun Heoabede6d2012-03-19 15:10:57 -07004119 cfq_init_prio_data(cfqq, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02004120
Jens Axboe89850f72006-07-22 16:48:31 +02004121 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004122 }
4123
4124 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125}
4126
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127/*
4128 * queue lock held here
4129 */
Jens Axboebb37b942006-12-01 10:42:33 +01004130static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131{
Jens Axboe5e705372006-07-13 12:39:25 +02004132 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133
Jens Axboe5e705372006-07-13 12:39:25 +02004134 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02004135 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004136
Jens Axboe22e2c502005-06-27 10:55:12 +02004137 BUG_ON(!cfqq->allocated[rw]);
4138 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004139
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02004140 /* Put down rq reference on cfqg */
Tejun Heoeb7d8c072012-03-23 14:02:53 +01004141 cfqg_put(RQ_CFQG(rq));
Tejun Heoa612fdd2011-12-14 00:33:41 +01004142 rq->elv.priv[0] = NULL;
4143 rq->elv.priv[1] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02004144
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145 cfq_put_queue(cfqq);
4146 }
4147}
4148
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004149static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01004150cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004151 struct cfq_queue *cfqq)
4152{
4153 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4154 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04004155 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004156 cfq_put_queue(cfqq);
4157 return cic_to_cfqq(cic, 1);
4158}
4159
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004160/*
4161 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4162 * was the last process referring to said cfqq.
4163 */
4164static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01004165split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004166{
4167 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004168 cfqq->pid = current->pid;
4169 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01004170 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004171 return cfqq;
4172 }
4173
4174 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02004175
4176 cfq_put_cooperator(cfqq);
4177
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004178 cfq_put_queue(cfqq);
4179 return NULL;
4180}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181/*
Jens Axboe22e2c502005-06-27 10:55:12 +02004182 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004183 */
Jens Axboe22e2c502005-06-27 10:55:12 +02004184static int
Tejun Heo852c7882012-03-05 13:15:27 -08004185cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4186 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187{
4188 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heof1f8cc92011-12-14 00:33:42 +01004189 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02004191 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004192 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193
4194 might_sleep_if(gfp_mask & __GFP_WAIT);
4195
Tejun Heo216284c2011-12-14 00:33:38 +01004196 spin_lock_irq(q->queue_lock);
Tejun Heof1f8cc92011-12-14 00:33:42 +01004197
Tejun Heo598971b2012-03-19 15:10:58 -07004198 check_ioprio_changed(cic, bio);
4199 check_blkcg_changed(cic, bio);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004200new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02004201 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02004202 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Tejun Heoabede6d2012-03-19 15:10:57 -07004203 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02004204 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004205 } else {
4206 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004207 * If the queue was seeky for too long, break it apart.
4208 */
Shaohua Liae54abe2010-02-05 13:11:45 +01004209 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004210 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4211 cfqq = split_cfqq(cic, cfqq);
4212 if (!cfqq)
4213 goto new_queue;
4214 }
4215
4216 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004217 * Check to see if this queue is scheduled to merge with
4218 * another, closely cooperating queue. The merging of
4219 * queues happens here as it must be done in process context.
4220 * The reference on new_cfqq was taken in merge_cfqqs.
4221 */
4222 if (cfqq->new_cfqq)
4223 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02004224 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225
4226 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02004227
Jens Axboe6fae9c22011-03-01 15:04:39 -05004228 cfqq->ref++;
Tejun Heoeb7d8c072012-03-23 14:02:53 +01004229 cfqg_get(cfqq->cfqg);
Tejun Heoa612fdd2011-12-14 00:33:41 +01004230 rq->elv.priv[0] = cfqq;
Tejun Heo1adaf3d2012-03-05 13:15:15 -08004231 rq->elv.priv[1] = cfqq->cfqg;
Tejun Heo216284c2011-12-14 00:33:38 +01004232 spin_unlock_irq(q->queue_lock);
Jens Axboe5e705372006-07-13 12:39:25 +02004233 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004234}
4235
David Howells65f27f32006-11-22 14:55:48 +00004236static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02004237{
David Howells65f27f32006-11-22 14:55:48 +00004238 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02004239 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02004240 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02004241
Jens Axboe40bb54d2009-04-15 12:11:10 +02004242 spin_lock_irq(q->queue_lock);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02004243 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02004244 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02004245}
4246
4247/*
4248 * Timer running if the active_queue is currently idling inside its time slice
4249 */
4250static void cfq_idle_slice_timer(unsigned long data)
4251{
4252 struct cfq_data *cfqd = (struct cfq_data *) data;
4253 struct cfq_queue *cfqq;
4254 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11004255 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02004256
Jens Axboe7b679132008-05-30 12:23:07 +02004257 cfq_log(cfqd, "idle timer fired");
4258
Jens Axboe22e2c502005-06-27 10:55:12 +02004259 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4260
Jens Axboefe094d92008-01-31 13:08:54 +01004261 cfqq = cfqd->active_queue;
4262 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11004263 timed_out = 0;
4264
Jens Axboe22e2c502005-06-27 10:55:12 +02004265 /*
Jens Axboeb0291952009-04-07 11:38:31 +02004266 * We saw a request before the queue expired, let it through
4267 */
4268 if (cfq_cfqq_must_dispatch(cfqq))
4269 goto out_kick;
4270
4271 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02004272 * expired
4273 */
Jens Axboe44f7c162007-01-19 11:51:58 +11004274 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02004275 goto expire;
4276
4277 /*
4278 * only expire and reinvoke request handler, if there are
4279 * other queues with pending requests
4280 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02004281 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02004282 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02004283
4284 /*
4285 * not expired and it has a request pending, let it dispatch
4286 */
Jens Axboe75e50982009-04-07 08:56:14 +02004287 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02004288 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01004289
4290 /*
4291 * Queue depth flag is reset only when the idle didn't succeed
4292 */
4293 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004294 }
4295expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02004296 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02004297out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02004298 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02004299out_cont:
4300 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4301}
4302
Jens Axboe3b181522005-06-27 10:56:24 +02004303static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4304{
4305 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02004306 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02004307}
Jens Axboe22e2c502005-06-27 10:55:12 +02004308
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02004309static void cfq_put_async_queues(struct cfq_data *cfqd)
4310{
4311 int i;
4312
4313 for (i = 0; i < IOPRIO_BE_NR; i++) {
4314 if (cfqd->async_cfqq[0][i])
4315 cfq_put_queue(cfqd->async_cfqq[0][i]);
4316 if (cfqd->async_cfqq[1][i])
4317 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02004318 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01004319
4320 if (cfqd->async_idle_cfqq)
4321 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02004322}
4323
Jens Axboeb374d182008-10-31 10:05:07 +01004324static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004325{
Jens Axboe22e2c502005-06-27 10:55:12 +02004326 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02004327 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02004328
Jens Axboe3b181522005-06-27 10:56:24 +02004329 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02004330
Al Virod9ff4182006-03-18 13:51:22 -05004331 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02004332
Al Virod9ff4182006-03-18 13:51:22 -05004333 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02004334 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02004335
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02004336 cfq_put_async_queues(cfqd);
Tejun Heo03aa2642012-03-05 13:15:19 -08004337
4338 spin_unlock_irq(q->queue_lock);
4339
Al Viroa90d7422006-03-18 12:05:37 -05004340 cfq_shutdown_timer_wq(cfqd);
4341
Tejun Heoffea73f2012-06-04 10:02:29 +02004342#ifdef CONFIG_CFQ_GROUP_IOSCHED
4343 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4344#else
Tejun Heof51b8022012-03-05 13:15:05 -08004345 kfree(cfqd->root_group);
Vivek Goyal2abae552011-05-23 10:02:19 +02004346#endif
Vivek Goyal56edf7d2011-05-19 15:38:22 -04004347 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004348}
4349
Jianpeng Mad50235b2013-07-03 13:25:24 +02004350static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004351{
4352 struct cfq_data *cfqd;
Tejun Heo3c798392012-04-16 13:57:25 -07004353 struct blkcg_gq *blkg __maybe_unused;
Tejun Heoa2b16932012-04-13 13:11:33 -07004354 int i, ret;
Jianpeng Mad50235b2013-07-03 13:25:24 +02004355 struct elevator_queue *eq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004356
Jianpeng Mad50235b2013-07-03 13:25:24 +02004357 eq = elevator_alloc(q, e);
4358 if (!eq)
Tejun Heob2fab5a2012-03-05 13:14:57 -08004359 return -ENOMEM;
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004360
Jianpeng Mad50235b2013-07-03 13:25:24 +02004361 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4362 if (!cfqd) {
4363 kobject_put(&eq->kobj);
4364 return -ENOMEM;
4365 }
4366 eq->elevator_data = cfqd;
4367
Tejun Heof51b8022012-03-05 13:15:05 -08004368 cfqd->queue = q;
Jianpeng Mad50235b2013-07-03 13:25:24 +02004369 spin_lock_irq(q->queue_lock);
4370 q->elevator = eq;
4371 spin_unlock_irq(q->queue_lock);
Tejun Heof51b8022012-03-05 13:15:05 -08004372
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05004373 /* Init root service tree */
4374 cfqd->grp_service_tree = CFQ_RB_ROOT;
4375
Tejun Heof51b8022012-03-05 13:15:05 -08004376 /* Init root group and prefer root group over other groups by default */
Vivek Goyal25fb5162009-12-03 12:59:46 -05004377#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004378 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
Tejun Heoa2b16932012-04-13 13:11:33 -07004379 if (ret)
4380 goto out_free;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04004381
Tejun Heoa2b16932012-04-13 13:11:33 -07004382 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
Tejun Heof51b8022012-03-05 13:15:05 -08004383#else
Tejun Heoa2b16932012-04-13 13:11:33 -07004384 ret = -ENOMEM;
Tejun Heof51b8022012-03-05 13:15:05 -08004385 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4386 GFP_KERNEL, cfqd->queue->node);
Tejun Heoa2b16932012-04-13 13:11:33 -07004387 if (!cfqd->root_group)
4388 goto out_free;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04004389
Tejun Heoa2b16932012-04-13 13:11:33 -07004390 cfq_init_cfqg_base(cfqd->root_group);
4391#endif
Tejun Heo3381cb82012-04-01 14:38:44 -07004392 cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
Tejun Heoe71357e2013-01-09 08:05:10 -08004393 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04004394
Jens Axboe26a2ac02009-04-23 12:13:27 +02004395 /*
4396 * Not strictly needed (since RB_ROOT just clears the node and we
4397 * zeroed cfqd on alloc), but better be safe in case someone decides
4398 * to add magic to the rb code
4399 */
4400 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4401 cfqd->prio_trees[i] = RB_ROOT;
4402
Jens Axboe6118b702009-06-30 09:34:12 +02004403 /*
4404 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4405 * Grab a permanent reference to it, so that the normal code flow
Tejun Heof51b8022012-03-05 13:15:05 -08004406 * will not attempt to free it. oom_cfqq is linked to root_group
4407 * but shouldn't hold a reference as it'll never be unlinked. Lose
4408 * the reference from linking right away.
Jens Axboe6118b702009-06-30 09:34:12 +02004409 */
4410 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01004411 cfqd->oom_cfqq.ref++;
Tejun Heo1adaf3d2012-03-05 13:15:15 -08004412
4413 spin_lock_irq(q->queue_lock);
Tejun Heof51b8022012-03-05 13:15:05 -08004414 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
Tejun Heoeb7d8c072012-03-23 14:02:53 +01004415 cfqg_put(cfqd->root_group);
Tejun Heo1adaf3d2012-03-05 13:15:15 -08004416 spin_unlock_irq(q->queue_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004417
Jens Axboe22e2c502005-06-27 10:55:12 +02004418 init_timer(&cfqd->idle_slice_timer);
4419 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4420 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4421
Jens Axboe23e018a2009-10-05 08:52:35 +02004422 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02004423
Linus Torvalds1da177e2005-04-16 15:20:36 -07004424 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02004425 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4426 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004427 cfqd->cfq_back_max = cfq_back_max;
4428 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02004429 cfqd->cfq_slice[0] = cfq_slice_async;
4430 cfqd->cfq_slice[1] = cfq_slice_sync;
Tao Ma5bf14c02012-04-01 14:33:39 -07004431 cfqd->cfq_target_latency = cfq_target_latency;
Jens Axboe22e2c502005-06-27 10:55:12 +02004432 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4433 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004434 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02004435 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01004436 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01004437 /*
4438 * we optimistically start assuming sync ops weren't delayed in last
4439 * second, in order to have larger depth for async operations.
4440 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01004441 cfqd->last_delayed_sync = jiffies - HZ;
Tejun Heob2fab5a2012-03-05 13:14:57 -08004442 return 0;
Tejun Heoa2b16932012-04-13 13:11:33 -07004443
4444out_free:
4445 kfree(cfqd);
Jianpeng Mad50235b2013-07-03 13:25:24 +02004446 kobject_put(&eq->kobj);
Tejun Heoa2b16932012-04-13 13:11:33 -07004447 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004448}
4449
Linus Torvalds1da177e2005-04-16 15:20:36 -07004450/*
4451 * sysfs parts below -->
4452 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004453static ssize_t
4454cfq_var_show(unsigned int var, char *page)
4455{
4456 return sprintf(page, "%d\n", var);
4457}
4458
4459static ssize_t
4460cfq_var_store(unsigned int *var, const char *page, size_t count)
4461{
4462 char *p = (char *) page;
4463
4464 *var = simple_strtoul(p, &p, 10);
4465 return count;
4466}
4467
Linus Torvalds1da177e2005-04-16 15:20:36 -07004468#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004469static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004470{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004471 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004472 unsigned int __data = __VAR; \
4473 if (__CONV) \
4474 __data = jiffies_to_msecs(__data); \
4475 return cfq_var_show(__data, (page)); \
4476}
4477SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004478SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4479SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05004480SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4481SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004482SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004483SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004484SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4485SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4486SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004487SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07004488SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004489#undef SHOW_FUNCTION
4490
4491#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004492static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004493{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004494 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004495 unsigned int __data; \
4496 int ret = cfq_var_store(&__data, (page), count); \
4497 if (__data < (MIN)) \
4498 __data = (MIN); \
4499 else if (__data > (MAX)) \
4500 __data = (MAX); \
4501 if (__CONV) \
4502 *(__PTR) = msecs_to_jiffies(__data); \
4503 else \
4504 *(__PTR) = __data; \
4505 return ret; \
4506}
4507STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004508STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4509 UINT_MAX, 1);
4510STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4511 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004512STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004513STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4514 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004515STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004516STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004517STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4518STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004519STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4520 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004521STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07004522STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004523#undef STORE_FUNCTION
4524
Al Viroe572ec72006-03-18 22:27:18 -05004525#define CFQ_ATTR(name) \
4526 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004527
Al Viroe572ec72006-03-18 22:27:18 -05004528static struct elv_fs_entry cfq_attrs[] = {
4529 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004530 CFQ_ATTR(fifo_expire_sync),
4531 CFQ_ATTR(fifo_expire_async),
4532 CFQ_ATTR(back_seek_max),
4533 CFQ_ATTR(back_seek_penalty),
4534 CFQ_ATTR(slice_sync),
4535 CFQ_ATTR(slice_async),
4536 CFQ_ATTR(slice_async_rq),
4537 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004538 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004539 CFQ_ATTR(low_latency),
Tao Ma5bf14c02012-04-01 14:33:39 -07004540 CFQ_ATTR(target_latency),
Al Viroe572ec72006-03-18 22:27:18 -05004541 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004542};
4543
Linus Torvalds1da177e2005-04-16 15:20:36 -07004544static struct elevator_type iosched_cfq = {
4545 .ops = {
4546 .elevator_merge_fn = cfq_merge,
4547 .elevator_merged_fn = cfq_merged_request,
4548 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004549 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004550 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004551 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004552 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004553 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004554 .elevator_deactivate_req_fn = cfq_deactivate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004555 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004556 .elevator_former_req_fn = elv_rb_former_request,
4557 .elevator_latter_req_fn = elv_rb_latter_request,
Tejun Heo9b84cac2011-12-14 00:33:42 +01004558 .elevator_init_icq_fn = cfq_init_icq,
Tejun Heo7e5a8792011-12-14 00:33:42 +01004559 .elevator_exit_icq_fn = cfq_exit_icq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004560 .elevator_set_req_fn = cfq_set_request,
4561 .elevator_put_req_fn = cfq_put_request,
4562 .elevator_may_queue_fn = cfq_may_queue,
4563 .elevator_init_fn = cfq_init_queue,
4564 .elevator_exit_fn = cfq_exit_queue,
4565 },
Tejun Heo3d3c2372011-12-14 00:33:42 +01004566 .icq_size = sizeof(struct cfq_io_cq),
4567 .icq_align = __alignof__(struct cfq_io_cq),
Al Viro3d1ab402006-03-18 18:35:43 -05004568 .elevator_attrs = cfq_attrs,
Tejun Heo3d3c2372011-12-14 00:33:42 +01004569 .elevator_name = "cfq",
Linus Torvalds1da177e2005-04-16 15:20:36 -07004570 .elevator_owner = THIS_MODULE,
4571};
4572
Vivek Goyal3e252062009-12-04 10:36:42 -05004573#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004574static struct blkcg_policy blkcg_policy_cfq = {
Tejun Heof9fcc2d2012-04-16 13:57:27 -07004575 .pd_size = sizeof(struct cfq_group),
4576 .cftypes = cfq_blkcg_files,
4577
4578 .pd_init_fn = cfq_pd_init,
Tejun Heo0b399202013-01-09 08:05:13 -08004579 .pd_offline_fn = cfq_pd_offline,
Tejun Heof9fcc2d2012-04-16 13:57:27 -07004580 .pd_reset_stats_fn = cfq_pd_reset_stats,
Vivek Goyal3e252062009-12-04 10:36:42 -05004581};
Vivek Goyal3e252062009-12-04 10:36:42 -05004582#endif
4583
Linus Torvalds1da177e2005-04-16 15:20:36 -07004584static int __init cfq_init(void)
4585{
Tejun Heo3d3c2372011-12-14 00:33:42 +01004586 int ret;
4587
Jens Axboe22e2c502005-06-27 10:55:12 +02004588 /*
4589 * could be 0 on HZ < 1000 setups
4590 */
4591 if (!cfq_slice_async)
4592 cfq_slice_async = 1;
4593 if (!cfq_slice_idle)
4594 cfq_slice_idle = 1;
4595
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004596#ifdef CONFIG_CFQ_GROUP_IOSCHED
4597 if (!cfq_group_idle)
4598 cfq_group_idle = 1;
Tejun Heo8bd435b2012-04-13 13:11:28 -07004599
Tejun Heo3c798392012-04-16 13:57:25 -07004600 ret = blkcg_policy_register(&blkcg_policy_cfq);
Tejun Heo8bd435b2012-04-13 13:11:28 -07004601 if (ret)
4602 return ret;
Tejun Heoffea73f2012-06-04 10:02:29 +02004603#else
4604 cfq_group_idle = 0;
4605#endif
Tejun Heo8bd435b2012-04-13 13:11:28 -07004606
Tejun Heofd794952012-06-04 10:01:38 +02004607 ret = -ENOMEM;
Tejun Heo3d3c2372011-12-14 00:33:42 +01004608 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4609 if (!cfq_pool)
Tejun Heo8bd435b2012-04-13 13:11:28 -07004610 goto err_pol_unreg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004611
Tejun Heo3d3c2372011-12-14 00:33:42 +01004612 ret = elv_register(&iosched_cfq);
Tejun Heo8bd435b2012-04-13 13:11:28 -07004613 if (ret)
4614 goto err_free_pool;
Tejun Heo3d3c2372011-12-14 00:33:42 +01004615
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004616 return 0;
Tejun Heo8bd435b2012-04-13 13:11:28 -07004617
4618err_free_pool:
4619 kmem_cache_destroy(cfq_pool);
4620err_pol_unreg:
Tejun Heoffea73f2012-06-04 10:02:29 +02004621#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004622 blkcg_policy_unregister(&blkcg_policy_cfq);
Tejun Heoffea73f2012-06-04 10:02:29 +02004623#endif
Tejun Heo8bd435b2012-04-13 13:11:28 -07004624 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004625}
4626
4627static void __exit cfq_exit(void)
4628{
Tejun Heoffea73f2012-06-04 10:02:29 +02004629#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004630 blkcg_policy_unregister(&blkcg_policy_cfq);
Tejun Heoffea73f2012-06-04 10:02:29 +02004631#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004632 elv_unregister(&iosched_cfq);
Tejun Heo3d3c2372011-12-14 00:33:42 +01004633 kmem_cache_destroy(cfq_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004634}
4635
4636module_init(cfq_init);
4637module_exit(cfq_exit);
4638
4639MODULE_AUTHOR("Jens Axboe");
4640MODULE_LICENSE("GPL");
4641MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");