blob: e3c591dd8f19d0b46fe42dc842510f9d220af143 [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 Heoeea8f412015-05-22 17:13:17 -040017#include <linux/blk-cgroup.h>
Tejun Heo6e736be2011-12-14 00:33:38 +010018#include "blk.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
Arianna Avanzinie48453c2015-06-05 23:38:42 +020070/* blkio-related constants */
Tejun Heo3ecca622015-08-18 14:55:35 -070071#define CFQ_WEIGHT_LEGACY_MIN 10
72#define CFQ_WEIGHT_LEGACY_DFL 500
73#define CFQ_WEIGHT_LEGACY_MAX 1000
Arianna Avanzinie48453c2015-06-05 23:38:42 +020074
Tejun Heoc5869802011-12-14 00:33:41 +010075struct cfq_ttime {
76 unsigned long last_end_request;
77
78 unsigned long ttime_total;
79 unsigned long ttime_samples;
80 unsigned long ttime_mean;
81};
82
Jens Axboe22e2c502005-06-27 10:55:12 +020083/*
Jens Axboecc09e292007-04-26 12:53:50 +020084 * Most of our rbtree usage is for sorting with min extraction, so
85 * if we cache the leftmost node we don't have to walk down the tree
86 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
87 * move this into the elevator for the rq sorting as well.
88 */
89struct cfq_rb_root {
90 struct rb_root rb;
91 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010092 unsigned count;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050093 u64 min_vdisktime;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +020094 struct cfq_ttime ttime;
Jens Axboecc09e292007-04-26 12:53:50 +020095};
Shaohua Lif5f2b6c2011-07-12 14:24:55 +020096#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
97 .ttime = {.last_end_request = jiffies,},}
Jens Axboecc09e292007-04-26 12:53:50 +020098
99/*
Jens Axboe6118b702009-06-30 09:34:12 +0200100 * Per process-grouping structure
101 */
102struct cfq_queue {
103 /* reference count */
Shaohua Li30d7b942011-01-07 08:46:59 +0100104 int ref;
Jens Axboe6118b702009-06-30 09:34:12 +0200105 /* various state flags, see below */
106 unsigned int flags;
107 /* parent cfq_data */
108 struct cfq_data *cfqd;
109 /* service_tree member */
110 struct rb_node rb_node;
111 /* service_tree key */
112 unsigned long rb_key;
113 /* prio tree member */
114 struct rb_node p_node;
115 /* prio tree root we belong to, if any */
116 struct rb_root *p_root;
117 /* sorted list of pending requests */
118 struct rb_root sort_list;
119 /* if fifo isn't expired, next request to serve */
120 struct request *next_rq;
121 /* requests queued in sort_list */
122 int queued[2];
123 /* currently allocated requests */
124 int allocated[2];
125 /* fifo list of requests in sort_list */
126 struct list_head fifo;
127
Vivek Goyaldae739e2009-12-03 12:59:45 -0500128 /* time when queue got scheduled in to dispatch first request. */
129 unsigned long dispatch_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500130 unsigned int allocated_slice;
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100131 unsigned int slice_dispatch;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500132 /* time when first request from queue completed and slice started. */
133 unsigned long slice_start;
Jens Axboe6118b702009-06-30 09:34:12 +0200134 unsigned long slice_end;
135 long slice_resid;
Jens Axboe6118b702009-06-30 09:34:12 +0200136
Christoph Hellwig65299a32011-08-23 14:50:29 +0200137 /* pending priority requests */
138 int prio_pending;
Jens Axboe6118b702009-06-30 09:34:12 +0200139 /* number of requests that are on the dispatch list or inside driver */
140 int dispatched;
141
142 /* io prio of this group */
143 unsigned short ioprio, org_ioprio;
Justin TerAvest4aede842011-07-12 08:31:45 +0200144 unsigned short ioprio_class;
Jens Axboe6118b702009-06-30 09:34:12 +0200145
Richard Kennedyc4081ba2010-02-22 13:49:24 +0100146 pid_t pid;
147
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +0100148 u32 seek_history;
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400149 sector_t last_request_pos;
150
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100151 struct cfq_rb_root *service_tree;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400152 struct cfq_queue *new_cfqq;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500153 struct cfq_group *cfqg;
Vivek Goyalc4e78932010-08-23 12:25:03 +0200154 /* Number of sectors dispatched from queue in single dispatch round */
155 unsigned long nr_sectors;
Jens Axboe6118b702009-06-30 09:34:12 +0200156};
157
158/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100159 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100160 * IDLE is handled separately, so it has negative index
161 */
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400162enum wl_class_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100163 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500164 RT_WORKLOAD = 1,
165 IDLE_WORKLOAD = 2,
Vivek Goyalb4627322010-10-22 09:48:43 +0200166 CFQ_PRIO_NR,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100167};
168
169/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100170 * Second index in the service_trees.
171 */
172enum wl_type_t {
173 ASYNC_WORKLOAD = 0,
174 SYNC_NOIDLE_WORKLOAD = 1,
175 SYNC_WORKLOAD = 2
176};
177
Tejun Heo155fead2012-04-01 14:38:44 -0700178struct cfqg_stats {
179#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo155fead2012-04-01 14:38:44 -0700180 /* number of ios merged */
181 struct blkg_rwstat merged;
182 /* total time spent on device in ns, may not be accurate w/ queueing */
183 struct blkg_rwstat service_time;
184 /* total time spent waiting in scheduler queue in ns */
185 struct blkg_rwstat wait_time;
186 /* number of IOs queued up */
187 struct blkg_rwstat queued;
Tejun Heo155fead2012-04-01 14:38:44 -0700188 /* total disk time and nr sectors dispatched by this group */
189 struct blkg_stat time;
190#ifdef CONFIG_DEBUG_BLK_CGROUP
191 /* time not charged to this cgroup */
192 struct blkg_stat unaccounted_time;
193 /* sum of number of ios queued across all samples */
194 struct blkg_stat avg_queue_size_sum;
195 /* count of samples taken for average */
196 struct blkg_stat avg_queue_size_samples;
197 /* how many times this group has been removed from service tree */
198 struct blkg_stat dequeue;
199 /* total time spent waiting for it to be assigned a timeslice. */
200 struct blkg_stat group_wait_time;
Tejun Heo3c798392012-04-16 13:57:25 -0700201 /* time spent idling for this blkcg_gq */
Tejun Heo155fead2012-04-01 14:38:44 -0700202 struct blkg_stat idle_time;
203 /* total time with empty current active q with other requests queued */
204 struct blkg_stat empty_time;
205 /* fields after this shouldn't be cleared on stat reset */
206 uint64_t start_group_wait_time;
207 uint64_t start_idle_time;
208 uint64_t start_empty_time;
209 uint16_t flags;
210#endif /* CONFIG_DEBUG_BLK_CGROUP */
211#endif /* CONFIG_CFQ_GROUP_IOSCHED */
212};
213
Arianna Avanzinie48453c2015-06-05 23:38:42 +0200214/* Per-cgroup data */
215struct cfq_group_data {
216 /* must be the first member */
Tejun Heo81437642015-08-18 14:55:15 -0700217 struct blkcg_policy_data cpd;
Arianna Avanzinie48453c2015-06-05 23:38:42 +0200218
219 unsigned int weight;
220 unsigned int leaf_weight;
221};
222
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500223/* This is per cgroup per device grouping structure */
224struct cfq_group {
Tejun Heof95a04a2012-04-16 13:57:26 -0700225 /* must be the first member */
226 struct blkg_policy_data pd;
227
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500228 /* group service_tree member */
229 struct rb_node rb_node;
230
231 /* group service_tree key */
232 u64 vdisktime;
Tejun Heoe71357e2013-01-09 08:05:10 -0800233
234 /*
Tejun Heo7918ffb2013-01-09 08:05:11 -0800235 * The number of active cfqgs and sum of their weights under this
236 * cfqg. This covers this cfqg's leaf_weight and all children's
237 * weights, but does not cover weights of further descendants.
238 *
239 * If a cfqg is on the service tree, it's active. An active cfqg
240 * also activates its parent and contributes to the children_weight
241 * of the parent.
242 */
243 int nr_active;
244 unsigned int children_weight;
245
246 /*
Tejun Heo1d3650f2013-01-09 08:05:11 -0800247 * vfraction is the fraction of vdisktime that the tasks in this
248 * cfqg are entitled to. This is determined by compounding the
249 * ratios walking up from this cfqg to the root.
250 *
251 * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
252 * vfractions on a service tree is approximately 1. The sum may
253 * deviate a bit due to rounding errors and fluctuations caused by
254 * cfqgs entering and leaving the service tree.
255 */
256 unsigned int vfraction;
257
258 /*
Tejun Heoe71357e2013-01-09 08:05:10 -0800259 * There are two weights - (internal) weight is the weight of this
260 * cfqg against the sibling cfqgs. leaf_weight is the wight of
261 * this cfqg against the child cfqgs. For the root cfqg, both
262 * weights are kept in sync for backward compatibility.
263 */
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500264 unsigned int weight;
Justin TerAvest8184f932011-03-17 16:12:36 +0100265 unsigned int new_weight;
Tejun Heo3381cb82012-04-01 14:38:44 -0700266 unsigned int dev_weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500267
Tejun Heoe71357e2013-01-09 08:05:10 -0800268 unsigned int leaf_weight;
269 unsigned int new_leaf_weight;
270 unsigned int dev_leaf_weight;
271
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500272 /* number of cfqq currently on this group */
273 int nr_cfqq;
274
Jens Axboe22e2c502005-06-27 10:55:12 +0200275 /*
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200276 * Per group busy queues average. Useful for workload slice calc. We
Vivek Goyalb4627322010-10-22 09:48:43 +0200277 * create the array for each prio class but at run time it is used
278 * only for RT and BE class and slot for IDLE class remains unused.
279 * This is primarily done to avoid confusion and a gcc warning.
280 */
281 unsigned int busy_queues_avg[CFQ_PRIO_NR];
282 /*
283 * rr lists of queues with requests. We maintain service trees for
284 * RT and BE classes. These trees are subdivided in subclasses
285 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
286 * class there is no subclassification and all the cfq queues go on
287 * a single tree service_tree_idle.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100288 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200289 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100290 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100291 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500292
Vivek Goyal4d2ceea2012-10-03 16:56:57 -0400293 unsigned long saved_wl_slice;
294 enum wl_type_t saved_wl_type;
295 enum wl_class_t saved_wl_class;
Tejun Heo4eef3042012-03-05 13:15:18 -0800296
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200297 /* number of requests that are on the dispatch list or inside driver */
298 int dispatched;
Shaohua Li7700fc42011-07-12 14:24:56 +0200299 struct cfq_ttime ttime;
Tejun Heo0b399202013-01-09 08:05:13 -0800300 struct cfqg_stats stats; /* stats for this cfqg */
Tejun Heo60a83702015-08-18 14:55:05 -0700301
302 /* async queue for each priority case */
303 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
304 struct cfq_queue *async_idle_cfqq;
305
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500306};
307
Tejun Heoc5869802011-12-14 00:33:41 +0100308struct cfq_io_cq {
309 struct io_cq icq; /* must be the first member */
310 struct cfq_queue *cfqq[2];
311 struct cfq_ttime ttime;
Tejun Heo598971b2012-03-19 15:10:58 -0700312 int ioprio; /* the current ioprio */
313#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heof4da8072014-09-08 08:15:20 +0900314 uint64_t blkcg_serial_nr; /* the current blkcg serial */
Tejun Heo598971b2012-03-19 15:10:58 -0700315#endif
Tejun Heoc5869802011-12-14 00:33:41 +0100316};
317
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500318/*
319 * Per block device queue structure
320 */
321struct cfq_data {
322 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500323 /* Root service tree for cfq_groups */
324 struct cfq_rb_root grp_service_tree;
Tejun Heof51b8022012-03-05 13:15:05 -0800325 struct cfq_group *root_group;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500326
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100327 /*
328 * The priority currently being served
329 */
Vivek Goyal4d2ceea2012-10-03 16:56:57 -0400330 enum wl_class_t serving_wl_class;
331 enum wl_type_t serving_wl_type;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100332 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500333 struct cfq_group *serving_group;
Jens Axboea36e71f2009-04-15 12:15:11 +0200334
335 /*
336 * Each priority tree is sorted by next_request position. These
337 * trees are used when determining if two or more queues are
338 * interleaving requests (see cfq_close_cooperator).
339 */
340 struct rb_root prio_trees[CFQ_PRIO_LISTS];
341
Jens Axboe22e2c502005-06-27 10:55:12 +0200342 unsigned int busy_queues;
Shaohua Lief8a41d2011-03-07 09:26:29 +0100343 unsigned int busy_sync_queues;
Jens Axboe22e2c502005-06-27 10:55:12 +0200344
Corrado Zoccolo53c583d2010-02-28 19:45:05 +0100345 int rq_in_driver;
346 int rq_in_flight[2];
Aaron Carroll45333d52008-08-26 15:52:36 +0200347
348 /*
349 * queue-depth detection
350 */
351 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200352 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100353 /*
354 * hw_tag can be
355 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
356 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
357 * 0 => no NCQ
358 */
359 int hw_tag_est_depth;
360 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200361
362 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200363 * idle window management
364 */
365 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200366 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200367
368 struct cfq_queue *active_queue;
Tejun Heoc5869802011-12-14 00:33:41 +0100369 struct cfq_io_cq *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200370
Jens Axboe6d048f52007-04-25 12:44:27 +0200371 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 /*
374 * tunables, see top of file
375 */
376 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200377 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 unsigned int cfq_back_penalty;
379 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200380 unsigned int cfq_slice[2];
381 unsigned int cfq_slice_async_rq;
382 unsigned int cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +0200383 unsigned int cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200384 unsigned int cfq_latency;
Tao Ma5bf14c02012-04-01 14:33:39 -0700385 unsigned int cfq_target_latency;
Al Virod9ff4182006-03-18 13:51:22 -0500386
Jens Axboe6118b702009-06-30 09:34:12 +0200387 /*
388 * Fallback dummy cfqq for extreme OOM conditions
389 */
390 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200391
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100392 unsigned long last_delayed_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393};
394
Vivek Goyal25fb5162009-12-03 12:59:46 -0500395static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
Tejun Heo60a83702015-08-18 14:55:05 -0700396static void cfq_put_queue(struct cfq_queue *cfqq);
Vivek Goyal25fb5162009-12-03 12:59:46 -0500397
Vivek Goyal34b98d02012-10-03 16:56:58 -0400398static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400399 enum wl_class_t class,
Vivek Goyal65b32a52009-12-16 17:52:59 -0500400 enum wl_type_t type)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100401{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500402 if (!cfqg)
403 return NULL;
404
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400405 if (class == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500406 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100407
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400408 return &cfqg->service_trees[class][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100409}
410
Jens Axboe3b181522005-06-27 10:56:24 +0200411enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100412 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
413 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200414 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100415 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100416 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
417 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
418 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100419 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200420 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400421 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Shaohua Liae54abe2010-02-05 13:11:45 +0100422 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100423 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500424 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
Jens Axboe3b181522005-06-27 10:56:24 +0200425};
426
427#define CFQ_CFQQ_FNS(name) \
428static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
429{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100430 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200431} \
432static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
433{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100434 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200435} \
436static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
437{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100438 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200439}
440
441CFQ_CFQQ_FNS(on_rr);
442CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200443CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200444CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200445CFQ_CFQQ_FNS(fifo_expire);
446CFQ_CFQQ_FNS(idle_window);
447CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100448CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200449CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200450CFQ_CFQQ_FNS(coop);
Shaohua Liae54abe2010-02-05 13:11:45 +0100451CFQ_CFQQ_FNS(split_coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100452CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500453CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200454#undef CFQ_CFQQ_FNS
455
Tejun Heo629ed0b2012-04-01 14:38:44 -0700456#if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700457
Tejun Heo155fead2012-04-01 14:38:44 -0700458/* cfqg stats flags */
459enum cfqg_stats_flags {
460 CFQG_stats_waiting = 0,
461 CFQG_stats_idling,
462 CFQG_stats_empty,
Tejun Heo629ed0b2012-04-01 14:38:44 -0700463};
464
Tejun Heo155fead2012-04-01 14:38:44 -0700465#define CFQG_FLAG_FNS(name) \
466static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700467{ \
Tejun Heo155fead2012-04-01 14:38:44 -0700468 stats->flags |= (1 << CFQG_stats_##name); \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700469} \
Tejun Heo155fead2012-04-01 14:38:44 -0700470static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700471{ \
Tejun Heo155fead2012-04-01 14:38:44 -0700472 stats->flags &= ~(1 << CFQG_stats_##name); \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700473} \
Tejun Heo155fead2012-04-01 14:38:44 -0700474static inline int cfqg_stats_##name(struct cfqg_stats *stats) \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700475{ \
Tejun Heo155fead2012-04-01 14:38:44 -0700476 return (stats->flags & (1 << CFQG_stats_##name)) != 0; \
Tejun Heo629ed0b2012-04-01 14:38:44 -0700477} \
478
Tejun Heo155fead2012-04-01 14:38:44 -0700479CFQG_FLAG_FNS(waiting)
480CFQG_FLAG_FNS(idling)
481CFQG_FLAG_FNS(empty)
482#undef CFQG_FLAG_FNS
Tejun Heo629ed0b2012-04-01 14:38:44 -0700483
484/* This should be called with the queue_lock held. */
Tejun Heo155fead2012-04-01 14:38:44 -0700485static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700486{
487 unsigned long long now;
488
Tejun Heo155fead2012-04-01 14:38:44 -0700489 if (!cfqg_stats_waiting(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700490 return;
491
492 now = sched_clock();
493 if (time_after64(now, stats->start_group_wait_time))
494 blkg_stat_add(&stats->group_wait_time,
495 now - stats->start_group_wait_time);
Tejun Heo155fead2012-04-01 14:38:44 -0700496 cfqg_stats_clear_waiting(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700497}
498
499/* This should be called with the queue_lock held. */
Tejun Heo155fead2012-04-01 14:38:44 -0700500static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
501 struct cfq_group *curr_cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700502{
Tejun Heo155fead2012-04-01 14:38:44 -0700503 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700504
Tejun Heo155fead2012-04-01 14:38:44 -0700505 if (cfqg_stats_waiting(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700506 return;
Tejun Heo155fead2012-04-01 14:38:44 -0700507 if (cfqg == curr_cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700508 return;
Tejun Heo155fead2012-04-01 14:38:44 -0700509 stats->start_group_wait_time = sched_clock();
510 cfqg_stats_mark_waiting(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700511}
512
513/* This should be called with the queue_lock held. */
Tejun Heo155fead2012-04-01 14:38:44 -0700514static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700515{
516 unsigned long long now;
517
Tejun Heo155fead2012-04-01 14:38:44 -0700518 if (!cfqg_stats_empty(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700519 return;
520
521 now = sched_clock();
522 if (time_after64(now, stats->start_empty_time))
523 blkg_stat_add(&stats->empty_time,
524 now - stats->start_empty_time);
Tejun Heo155fead2012-04-01 14:38:44 -0700525 cfqg_stats_clear_empty(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700526}
527
Tejun Heo155fead2012-04-01 14:38:44 -0700528static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700529{
Tejun Heo155fead2012-04-01 14:38:44 -0700530 blkg_stat_add(&cfqg->stats.dequeue, 1);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700531}
532
Tejun Heo155fead2012-04-01 14:38:44 -0700533static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700534{
Tejun Heo155fead2012-04-01 14:38:44 -0700535 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700536
Tejun Heo4d5e80a2013-01-09 08:05:12 -0800537 if (blkg_rwstat_total(&stats->queued))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700538 return;
539
540 /*
541 * group is already marked empty. This can happen if cfqq got new
542 * request in parent group and moved to this group while being added
543 * to service tree. Just ignore the event and move on.
544 */
Tejun Heo155fead2012-04-01 14:38:44 -0700545 if (cfqg_stats_empty(stats))
Tejun Heo629ed0b2012-04-01 14:38:44 -0700546 return;
547
548 stats->start_empty_time = sched_clock();
Tejun Heo155fead2012-04-01 14:38:44 -0700549 cfqg_stats_mark_empty(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700550}
551
Tejun Heo155fead2012-04-01 14:38:44 -0700552static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700553{
Tejun Heo155fead2012-04-01 14:38:44 -0700554 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700555
Tejun Heo155fead2012-04-01 14:38:44 -0700556 if (cfqg_stats_idling(stats)) {
Tejun Heo629ed0b2012-04-01 14:38:44 -0700557 unsigned long long now = sched_clock();
558
559 if (time_after64(now, stats->start_idle_time))
560 blkg_stat_add(&stats->idle_time,
561 now - stats->start_idle_time);
Tejun Heo155fead2012-04-01 14:38:44 -0700562 cfqg_stats_clear_idling(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700563 }
564}
565
Tejun Heo155fead2012-04-01 14:38:44 -0700566static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700567{
Tejun Heo155fead2012-04-01 14:38:44 -0700568 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700569
Tejun Heo155fead2012-04-01 14:38:44 -0700570 BUG_ON(cfqg_stats_idling(stats));
Tejun Heo629ed0b2012-04-01 14:38:44 -0700571
572 stats->start_idle_time = sched_clock();
Tejun Heo155fead2012-04-01 14:38:44 -0700573 cfqg_stats_mark_idling(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700574}
575
Tejun Heo155fead2012-04-01 14:38:44 -0700576static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
Tejun Heo629ed0b2012-04-01 14:38:44 -0700577{
Tejun Heo155fead2012-04-01 14:38:44 -0700578 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700579
580 blkg_stat_add(&stats->avg_queue_size_sum,
Tejun Heo4d5e80a2013-01-09 08:05:12 -0800581 blkg_rwstat_total(&stats->queued));
Tejun Heo629ed0b2012-04-01 14:38:44 -0700582 blkg_stat_add(&stats->avg_queue_size_samples, 1);
Tejun Heo155fead2012-04-01 14:38:44 -0700583 cfqg_stats_update_group_wait_time(stats);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700584}
585
586#else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
587
Tejun Heof48ec1d2012-04-13 13:11:25 -0700588static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
589static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
590static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
591static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
592static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
593static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
594static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
Tejun Heo629ed0b2012-04-01 14:38:44 -0700595
596#endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
597
598#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo2ce4d502012-04-01 14:38:43 -0700599
Jens Axboe4ceab712015-06-19 10:13:01 -0600600static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
601{
602 return pd ? container_of(pd, struct cfq_group, pd) : NULL;
603}
604
605static struct cfq_group_data
606*cpd_to_cfqgd(struct blkcg_policy_data *cpd)
607{
Tejun Heo81437642015-08-18 14:55:15 -0700608 return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
Jens Axboe4ceab712015-06-19 10:13:01 -0600609}
610
611static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
612{
613 return pd_to_blkg(&cfqg->pd);
614}
615
Tejun Heoffea73f2012-06-04 10:02:29 +0200616static struct blkcg_policy blkcg_policy_cfq;
617
618static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
619{
620 return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
621}
622
Arianna Avanzinie48453c2015-06-05 23:38:42 +0200623static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
624{
625 return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
626}
627
Tejun Heod02f7aa2013-01-09 08:05:11 -0800628static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
Tejun Heo7918ffb2013-01-09 08:05:11 -0800629{
Tejun Heod02f7aa2013-01-09 08:05:11 -0800630 struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
Tejun Heo7918ffb2013-01-09 08:05:11 -0800631
Tejun Heod02f7aa2013-01-09 08:05:11 -0800632 return pblkg ? blkg_to_cfqg(pblkg) : NULL;
Tejun Heo7918ffb2013-01-09 08:05:11 -0800633}
634
Jan Kara3984aa52016-01-12 16:24:19 +0100635static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
636 struct cfq_group *ancestor)
637{
638 return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup,
639 cfqg_to_blkg(ancestor)->blkcg->css.cgroup);
640}
641
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100642static inline void cfqg_get(struct cfq_group *cfqg)
643{
644 return blkg_get(cfqg_to_blkg(cfqg));
645}
646
647static inline void cfqg_put(struct cfq_group *cfqg)
648{
649 return blkg_put(cfqg_to_blkg(cfqg));
650}
651
Tejun Heo54e7ed12012-04-16 13:57:23 -0700652#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \
653 char __pbuf[128]; \
654 \
655 blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \
Vivek Goyalb226e5c2012-10-03 16:57:01 -0400656 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
657 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
658 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
Tejun Heo54e7ed12012-04-16 13:57:23 -0700659 __pbuf, ##args); \
660} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500661
Tejun Heo54e7ed12012-04-16 13:57:23 -0700662#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \
663 char __pbuf[128]; \
664 \
665 blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \
666 blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \
667} while (0)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500668
Tejun Heo155fead2012-04-01 14:38:44 -0700669static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
670 struct cfq_group *curr_cfqg, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700671{
Tejun Heo155fead2012-04-01 14:38:44 -0700672 blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
673 cfqg_stats_end_empty_time(&cfqg->stats);
674 cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700675}
676
Tejun Heo155fead2012-04-01 14:38:44 -0700677static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
678 unsigned long time, unsigned long unaccounted_time)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700679{
Tejun Heo155fead2012-04-01 14:38:44 -0700680 blkg_stat_add(&cfqg->stats.time, time);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700681#ifdef CONFIG_DEBUG_BLK_CGROUP
Tejun Heo155fead2012-04-01 14:38:44 -0700682 blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
Tejun Heo629ed0b2012-04-01 14:38:44 -0700683#endif
Tejun Heo2ce4d502012-04-01 14:38:43 -0700684}
685
Tejun Heo155fead2012-04-01 14:38:44 -0700686static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700687{
Tejun Heo155fead2012-04-01 14:38:44 -0700688 blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700689}
690
Tejun Heo155fead2012-04-01 14:38:44 -0700691static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700692{
Tejun Heo155fead2012-04-01 14:38:44 -0700693 blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700694}
695
Tejun Heo155fead2012-04-01 14:38:44 -0700696static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
697 uint64_t start_time, uint64_t io_start_time, int rw)
Tejun Heo2ce4d502012-04-01 14:38:43 -0700698{
Tejun Heo155fead2012-04-01 14:38:44 -0700699 struct cfqg_stats *stats = &cfqg->stats;
Tejun Heo629ed0b2012-04-01 14:38:44 -0700700 unsigned long long now = sched_clock();
Tejun Heo629ed0b2012-04-01 14:38:44 -0700701
702 if (time_after64(now, io_start_time))
703 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
704 if (time_after64(io_start_time, start_time))
705 blkg_rwstat_add(&stats->wait_time, rw,
706 io_start_time - start_time);
Tejun Heo2ce4d502012-04-01 14:38:43 -0700707}
708
Tejun Heo689665a2013-01-09 08:05:13 -0800709/* @stats = 0 */
710static void cfqg_stats_reset(struct cfqg_stats *stats)
Tejun Heo155fead2012-04-01 14:38:44 -0700711{
Tejun Heo155fead2012-04-01 14:38:44 -0700712 /* queued stats shouldn't be cleared */
Tejun Heo155fead2012-04-01 14:38:44 -0700713 blkg_rwstat_reset(&stats->merged);
714 blkg_rwstat_reset(&stats->service_time);
715 blkg_rwstat_reset(&stats->wait_time);
716 blkg_stat_reset(&stats->time);
717#ifdef CONFIG_DEBUG_BLK_CGROUP
718 blkg_stat_reset(&stats->unaccounted_time);
719 blkg_stat_reset(&stats->avg_queue_size_sum);
720 blkg_stat_reset(&stats->avg_queue_size_samples);
721 blkg_stat_reset(&stats->dequeue);
722 blkg_stat_reset(&stats->group_wait_time);
723 blkg_stat_reset(&stats->idle_time);
724 blkg_stat_reset(&stats->empty_time);
725#endif
726}
727
Tejun Heo0b399202013-01-09 08:05:13 -0800728/* @to += @from */
Tejun Heoe6269c42015-08-18 14:55:21 -0700729static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
Tejun Heo0b399202013-01-09 08:05:13 -0800730{
731 /* queued stats shouldn't be cleared */
Tejun Heoe6269c42015-08-18 14:55:21 -0700732 blkg_rwstat_add_aux(&to->merged, &from->merged);
733 blkg_rwstat_add_aux(&to->service_time, &from->service_time);
734 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
735 blkg_stat_add_aux(&from->time, &from->time);
Tejun Heo0b399202013-01-09 08:05:13 -0800736#ifdef CONFIG_DEBUG_BLK_CGROUP
Tejun Heoe6269c42015-08-18 14:55:21 -0700737 blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
738 blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
739 blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
740 blkg_stat_add_aux(&to->dequeue, &from->dequeue);
741 blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
742 blkg_stat_add_aux(&to->idle_time, &from->idle_time);
743 blkg_stat_add_aux(&to->empty_time, &from->empty_time);
Tejun Heo0b399202013-01-09 08:05:13 -0800744#endif
745}
746
747/*
Tejun Heoe6269c42015-08-18 14:55:21 -0700748 * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
Tejun Heo0b399202013-01-09 08:05:13 -0800749 * recursive stats can still account for the amount used by this cfqg after
750 * it's gone.
751 */
752static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
753{
754 struct cfq_group *parent = cfqg_parent(cfqg);
755
756 lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
757
758 if (unlikely(!parent))
759 return;
760
Tejun Heoe6269c42015-08-18 14:55:21 -0700761 cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
Tejun Heo0b399202013-01-09 08:05:13 -0800762 cfqg_stats_reset(&cfqg->stats);
Tejun Heo0b399202013-01-09 08:05:13 -0800763}
764
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100765#else /* CONFIG_CFQ_GROUP_IOSCHED */
766
Tejun Heod02f7aa2013-01-09 08:05:11 -0800767static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
Jan Kara3984aa52016-01-12 16:24:19 +0100768static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
769 struct cfq_group *ancestor)
770{
771 return true;
772}
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100773static inline void cfqg_get(struct cfq_group *cfqg) { }
774static inline void cfqg_put(struct cfq_group *cfqg) { }
775
Jens Axboe7b679132008-05-30 12:23:07 +0200776#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
Vivek Goyalb226e5c2012-10-03 16:57:01 -0400777 blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
778 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
779 cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
780 ##args)
Kyungmin Park4495a7d2011-05-31 10:04:09 +0200781#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100782
Tejun Heo155fead2012-04-01 14:38:44 -0700783static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
784 struct cfq_group *curr_cfqg, int rw) { }
785static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
786 unsigned long time, unsigned long unaccounted_time) { }
787static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
788static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
Tejun Heo155fead2012-04-01 14:38:44 -0700789static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
790 uint64_t start_time, uint64_t io_start_time, int rw) { }
Tejun Heo2ce4d502012-04-01 14:38:43 -0700791
Tejun Heoeb7d8c072012-03-23 14:02:53 +0100792#endif /* CONFIG_CFQ_GROUP_IOSCHED */
793
Jens Axboe7b679132008-05-30 12:23:07 +0200794#define cfq_log(cfqd, fmt, args...) \
795 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
796
Vivek Goyal615f0252009-12-03 12:59:39 -0500797/* Traverses through cfq group service trees */
798#define for_each_cfqg_st(cfqg, i, j, st) \
799 for (i = 0; i <= IDLE_WORKLOAD; i++) \
800 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
801 : &cfqg->service_tree_idle; \
802 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
803 (i == IDLE_WORKLOAD && j == 0); \
804 j++, st = i < IDLE_WORKLOAD ? \
805 &cfqg->service_trees[i][j]: NULL) \
806
Shaohua Lif5f2b6c2011-07-12 14:24:55 +0200807static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
808 struct cfq_ttime *ttime, bool group_idle)
809{
810 unsigned long slice;
811 if (!sample_valid(ttime->ttime_samples))
812 return false;
813 if (group_idle)
814 slice = cfqd->cfq_group_idle;
815 else
816 slice = cfqd->cfq_slice_idle;
817 return ttime->ttime_mean > slice;
818}
Vivek Goyal615f0252009-12-03 12:59:39 -0500819
Vivek Goyal02b35082010-08-23 12:23:53 +0200820static inline bool iops_mode(struct cfq_data *cfqd)
821{
822 /*
823 * If we are not idling on queues and it is a NCQ drive, parallel
824 * execution of requests is on and measuring time is not possible
825 * in most of the cases until and unless we drive shallower queue
826 * depths and that becomes a performance bottleneck. In such cases
827 * switch to start providing fairness in terms of number of IOs.
828 */
829 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
830 return true;
831 else
832 return false;
833}
834
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400835static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100836{
837 if (cfq_class_idle(cfqq))
838 return IDLE_WORKLOAD;
839 if (cfq_class_rt(cfqq))
840 return RT_WORKLOAD;
841 return BE_WORKLOAD;
842}
843
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100844
845static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
846{
847 if (!cfq_cfqq_sync(cfqq))
848 return ASYNC_WORKLOAD;
849 if (!cfq_cfqq_idle_window(cfqq))
850 return SYNC_NOIDLE_WORKLOAD;
851 return SYNC_WORKLOAD;
852}
853
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400854static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500855 struct cfq_data *cfqd,
856 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100857{
Vivek Goyal3bf10fe2012-10-03 16:56:56 -0400858 if (wl_class == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500859 return cfqg->service_tree_idle.count;
860
Vivek Goyal34b98d02012-10-03 16:56:58 -0400861 return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
862 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
863 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100864}
865
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500866static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
867 struct cfq_group *cfqg)
868{
Vivek Goyal34b98d02012-10-03 16:56:58 -0400869 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
870 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500871}
872
Jens Axboe165125e2007-07-24 09:28:11 +0200873static void cfq_dispatch_insert(struct request_queue *, struct request *);
Tejun Heo4f85cb92012-03-05 13:15:28 -0800874static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
Tejun Heo2da8de02015-08-18 14:55:02 -0700875 struct cfq_io_cq *cic, struct bio *bio);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200876
Tejun Heoc5869802011-12-14 00:33:41 +0100877static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
878{
879 /* cic->icq is the first member, %NULL will convert to %NULL */
880 return container_of(icq, struct cfq_io_cq, icq);
881}
882
Tejun Heo47fdd4c2011-12-14 00:33:42 +0100883static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
884 struct io_context *ioc)
885{
886 if (ioc)
887 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
888 return NULL;
889}
890
Tejun Heoc5869802011-12-14 00:33:41 +0100891static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200892{
Jens Axboea6151c32009-10-07 20:02:57 +0200893 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200894}
895
Tejun Heoc5869802011-12-14 00:33:41 +0100896static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
897 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200898{
Jens Axboea6151c32009-10-07 20:02:57 +0200899 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200900}
901
Tejun Heoc5869802011-12-14 00:33:41 +0100902static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400903{
Tejun Heoc5869802011-12-14 00:33:41 +0100904 return cic->icq.q->elevator->elevator_data;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +0400905}
906
Vasily Tarasov91fac312007-04-25 12:29:51 +0200907/*
908 * We regard a request as SYNC, if it's either a read or has the SYNC bit
909 * set (in which case it could also be direct WRITE).
910 */
Jens Axboea6151c32009-10-07 20:02:57 +0200911static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200912{
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +0200913 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200914}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700917 * scheduler run of queue, if there are requests pending and no one in the
918 * driver that will restart queueing
919 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200920static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700921{
Jens Axboe7b679132008-05-30 12:23:07 +0200922 if (cfqd->busy_queues) {
923 cfq_log(cfqd, "schedule dispatch");
Jens Axboe59c3d452014-04-08 09:15:35 -0600924 kblockd_schedule_work(&cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200925 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700926}
927
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100929 * Scale schedule slice based on io priority. Use the sync time slice only
930 * if a queue is marked sync and has sync io queued. A sync queue with async
931 * io only, should not get full sync slice length.
932 */
Jens Axboea6151c32009-10-07 20:02:57 +0200933static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200934 unsigned short prio)
935{
936 const int base_slice = cfqd->cfq_slice[sync];
937
938 WARN_ON(prio >= IOPRIO_BE_NR);
939
940 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
941}
942
Jens Axboe44f7c162007-01-19 11:51:58 +1100943static inline int
944cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
945{
Jens Axboed9e76202007-04-20 14:27:50 +0200946 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100947}
948
Tejun Heo1d3650f2013-01-09 08:05:11 -0800949/**
950 * cfqg_scale_charge - scale disk time charge according to cfqg weight
951 * @charge: disk time being charged
952 * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
953 *
954 * Scale @charge according to @vfraction, which is in range (0, 1]. The
955 * scaling is inversely proportional.
956 *
957 * scaled = charge / vfraction
958 *
959 * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
960 */
961static inline u64 cfqg_scale_charge(unsigned long charge,
962 unsigned int vfraction)
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500963{
Tejun Heo1d3650f2013-01-09 08:05:11 -0800964 u64 c = charge << CFQ_SERVICE_SHIFT; /* make it fixed point */
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500965
Tejun Heo1d3650f2013-01-09 08:05:11 -0800966 /* charge / vfraction */
967 c <<= CFQ_SERVICE_SHIFT;
968 do_div(c, vfraction);
969 return c;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500970}
971
972static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
973{
974 s64 delta = (s64)(vdisktime - min_vdisktime);
975 if (delta > 0)
976 min_vdisktime = vdisktime;
977
978 return min_vdisktime;
979}
980
981static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
982{
983 s64 delta = (s64)(vdisktime - min_vdisktime);
984 if (delta < 0)
985 min_vdisktime = vdisktime;
986
987 return min_vdisktime;
988}
989
990static void update_min_vdisktime(struct cfq_rb_root *st)
991{
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500992 struct cfq_group *cfqg;
993
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500994 if (st->left) {
995 cfqg = rb_entry_cfqg(st->left);
Gui Jianfenga6032712011-03-07 09:28:09 +0100996 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
997 cfqg->vdisktime);
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500998 }
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500999}
1000
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001001/*
1002 * get averaged number of queues of RT/BE priority.
1003 * average is updated, with a formula that gives more weight to higher numbers,
1004 * to quickly follows sudden increases and decrease slowly
1005 */
1006
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001007static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1008 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +01001009{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001010 unsigned min_q, max_q;
1011 unsigned mult = cfq_hist_divisor - 1;
1012 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001013 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001014
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001015 min_q = min(cfqg->busy_queues_avg[rt], busy);
1016 max_q = max(cfqg->busy_queues_avg[rt], busy);
1017 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001018 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001019 return cfqg->busy_queues_avg[rt];
1020}
1021
1022static inline unsigned
1023cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1024{
Tejun Heo41cad6a2013-01-09 08:05:11 -08001025 return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001026}
1027
Shaohua Lic553f8e2011-01-14 08:41:03 +01001028static inline unsigned
Vivek Goyalba5bd522011-01-19 08:25:02 -07001029cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +11001030{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001031 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1032 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001033 /*
1034 * interested queues (we consider only the ones with the same
1035 * priority class in the cfq group)
1036 */
1037 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1038 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001039 unsigned sync_slice = cfqd->cfq_slice[1];
1040 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001041 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1042
1043 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001044 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1045 /* scale low_slice according to IO priority
1046 * and sync vs async */
1047 unsigned low_slice =
1048 min(slice, base_low_slice * slice / sync_slice);
1049 /* the adapted slice value is scaled to fit all iqs
1050 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05001051 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001052 low_slice);
1053 }
1054 }
Shaohua Lic553f8e2011-01-14 08:41:03 +01001055 return slice;
1056}
1057
1058static inline void
1059cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1060{
Vivek Goyalba5bd522011-01-19 08:25:02 -07001061 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01001062
Vivek Goyaldae739e2009-12-03 12:59:45 -05001063 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +01001064 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001065 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +02001066 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +11001067}
1068
1069/*
1070 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1071 * isn't valid until the first request from the dispatch is activated
1072 * and the slice time set.
1073 */
Jens Axboea6151c32009-10-07 20:02:57 +02001074static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +11001075{
1076 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01001077 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +11001078 if (time_before(jiffies, cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01001079 return false;
Jens Axboe44f7c162007-01-19 11:51:58 +11001080
Shaohua Lic1e44752010-11-08 15:01:02 +01001081 return true;
Jens Axboe44f7c162007-01-19 11:51:58 +11001082}
1083
1084/*
Jens Axboe5e705372006-07-13 12:39:25 +02001085 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +02001087 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 */
Jens Axboe5e705372006-07-13 12:39:25 +02001089static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001090cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001092 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +02001094#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
1095#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
1096 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097
Jens Axboe5e705372006-07-13 12:39:25 +02001098 if (rq1 == NULL || rq1 == rq2)
1099 return rq2;
1100 if (rq2 == NULL)
1101 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +02001102
Namhyung Kim229836b2011-05-24 10:23:21 +02001103 if (rq_is_sync(rq1) != rq_is_sync(rq2))
1104 return rq_is_sync(rq1) ? rq1 : rq2;
1105
Christoph Hellwig65299a32011-08-23 14:50:29 +02001106 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1107 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
Jens Axboeb53d1ed2011-08-19 08:34:48 +02001108
Tejun Heo83096eb2009-05-07 22:24:39 +09001109 s1 = blk_rq_pos(rq1);
1110 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 /*
1113 * by definition, 1KiB is 2 sectors
1114 */
1115 back_max = cfqd->cfq_back_max * 2;
1116
1117 /*
1118 * Strict one way elevator _except_ in the case where we allow
1119 * short backward seeks which are biased as twice the cost of a
1120 * similar forward seek.
1121 */
1122 if (s1 >= last)
1123 d1 = s1 - last;
1124 else if (s1 + back_max >= last)
1125 d1 = (last - s1) * cfqd->cfq_back_penalty;
1126 else
Andreas Mohre8a99052006-03-28 08:59:49 +02001127 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128
1129 if (s2 >= last)
1130 d2 = s2 - last;
1131 else if (s2 + back_max >= last)
1132 d2 = (last - s2) * cfqd->cfq_back_penalty;
1133 else
Andreas Mohre8a99052006-03-28 08:59:49 +02001134 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135
1136 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137
Andreas Mohre8a99052006-03-28 08:59:49 +02001138 /*
1139 * By doing switch() on the bit mask "wrap" we avoid having to
1140 * check two variables for all permutations: --> faster!
1141 */
1142 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +02001143 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +02001144 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +02001145 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +02001146 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +02001147 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +02001148 else {
1149 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +02001150 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +02001151 else
Jens Axboe5e705372006-07-13 12:39:25 +02001152 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +02001153 }
1154
1155 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +02001156 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +02001157 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +02001158 return rq2;
1159 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +02001160 default:
1161 /*
1162 * Since both rqs are wrapped,
1163 * start with the one that's further behind head
1164 * (--> only *one* back seek required),
1165 * since back seek takes more time than forward.
1166 */
1167 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +02001168 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 else
Jens Axboe5e705372006-07-13 12:39:25 +02001170 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 }
1172}
1173
Jens Axboe498d3aa22007-04-26 12:54:48 +02001174/*
1175 * The below is leftmost cache rbtree addon
1176 */
Jens Axboe08717142008-01-28 11:38:15 +01001177static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +02001178{
Vivek Goyal615f0252009-12-03 12:59:39 -05001179 /* Service tree is empty */
1180 if (!root->count)
1181 return NULL;
1182
Jens Axboecc09e292007-04-26 12:53:50 +02001183 if (!root->left)
1184 root->left = rb_first(&root->rb);
1185
Jens Axboe08717142008-01-28 11:38:15 +01001186 if (root->left)
1187 return rb_entry(root->left, struct cfq_queue, rb_node);
1188
1189 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +02001190}
1191
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001192static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1193{
1194 if (!root->left)
1195 root->left = rb_first(&root->rb);
1196
1197 if (root->left)
1198 return rb_entry_cfqg(root->left);
1199
1200 return NULL;
1201}
1202
Jens Axboea36e71f2009-04-15 12:15:11 +02001203static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1204{
1205 rb_erase(n, root);
1206 RB_CLEAR_NODE(n);
1207}
1208
Jens Axboecc09e292007-04-26 12:53:50 +02001209static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1210{
1211 if (root->left == n)
1212 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001213 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001214 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +02001215}
1216
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217/*
1218 * would be nice to take fifo expire time into account as well
1219 */
Jens Axboe5e705372006-07-13 12:39:25 +02001220static struct request *
1221cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1222 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223{
Jens Axboe21183b02006-07-13 12:33:14 +02001224 struct rb_node *rbnext = rb_next(&last->rb_node);
1225 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +02001226 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227
Jens Axboe21183b02006-07-13 12:33:14 +02001228 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229
1230 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +02001231 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +02001232
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +02001234 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +02001235 else {
1236 rbnext = rb_first(&cfqq->sort_list);
1237 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +02001238 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +02001239 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001241 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242}
1243
Jens Axboed9e76202007-04-20 14:27:50 +02001244static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1245 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246{
Jens Axboed9e76202007-04-20 14:27:50 +02001247 /*
1248 * just an approximation, should be ok.
1249 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001250 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +01001251 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +02001252}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001254static inline s64
1255cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1256{
1257 return cfqg->vdisktime - st->min_vdisktime;
1258}
1259
1260static void
1261__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1262{
1263 struct rb_node **node = &st->rb.rb_node;
1264 struct rb_node *parent = NULL;
1265 struct cfq_group *__cfqg;
1266 s64 key = cfqg_key(st, cfqg);
1267 int left = 1;
1268
1269 while (*node != NULL) {
1270 parent = *node;
1271 __cfqg = rb_entry_cfqg(parent);
1272
1273 if (key < cfqg_key(st, __cfqg))
1274 node = &parent->rb_left;
1275 else {
1276 node = &parent->rb_right;
1277 left = 0;
1278 }
1279 }
1280
1281 if (left)
1282 st->left = &cfqg->rb_node;
1283
1284 rb_link_node(&cfqg->rb_node, parent, node);
1285 rb_insert_color(&cfqg->rb_node, &st->rb);
1286}
1287
Toshiaki Makita7b5af5c2014-08-28 17:14:58 +09001288/*
1289 * This has to be called only on activation of cfqg
1290 */
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001291static void
Justin TerAvest8184f932011-03-17 16:12:36 +01001292cfq_update_group_weight(struct cfq_group *cfqg)
1293{
Tejun Heo3381cb82012-04-01 14:38:44 -07001294 if (cfqg->new_weight) {
Justin TerAvest8184f932011-03-17 16:12:36 +01001295 cfqg->weight = cfqg->new_weight;
Tejun Heo3381cb82012-04-01 14:38:44 -07001296 cfqg->new_weight = 0;
Justin TerAvest8184f932011-03-17 16:12:36 +01001297 }
Toshiaki Makitae15693e2014-08-26 20:56:36 +09001298}
1299
1300static void
1301cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1302{
1303 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
Tejun Heoe71357e2013-01-09 08:05:10 -08001304
1305 if (cfqg->new_leaf_weight) {
1306 cfqg->leaf_weight = cfqg->new_leaf_weight;
1307 cfqg->new_leaf_weight = 0;
1308 }
Justin TerAvest8184f932011-03-17 16:12:36 +01001309}
1310
1311static void
1312cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1313{
Tejun Heo1d3650f2013-01-09 08:05:11 -08001314 unsigned int vfr = 1 << CFQ_SERVICE_SHIFT; /* start with 1 */
Tejun Heo7918ffb2013-01-09 08:05:11 -08001315 struct cfq_group *pos = cfqg;
Tejun Heo1d3650f2013-01-09 08:05:11 -08001316 struct cfq_group *parent;
Tejun Heo7918ffb2013-01-09 08:05:11 -08001317 bool propagate;
1318
1319 /* add to the service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +01001320 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1321
Toshiaki Makita7b5af5c2014-08-28 17:14:58 +09001322 /*
1323 * Update leaf_weight. We cannot update weight at this point
1324 * because cfqg might already have been activated and is
1325 * contributing its current weight to the parent's child_weight.
1326 */
Toshiaki Makitae15693e2014-08-26 20:56:36 +09001327 cfq_update_group_leaf_weight(cfqg);
Justin TerAvest8184f932011-03-17 16:12:36 +01001328 __cfq_group_service_tree_add(st, cfqg);
Tejun Heo7918ffb2013-01-09 08:05:11 -08001329
1330 /*
Tejun Heo1d3650f2013-01-09 08:05:11 -08001331 * Activate @cfqg and calculate the portion of vfraction @cfqg is
1332 * entitled to. vfraction is calculated by walking the tree
1333 * towards the root calculating the fraction it has at each level.
1334 * The compounded ratio is how much vfraction @cfqg owns.
1335 *
1336 * Start with the proportion tasks in this cfqg has against active
1337 * children cfqgs - its leaf_weight against children_weight.
Tejun Heo7918ffb2013-01-09 08:05:11 -08001338 */
1339 propagate = !pos->nr_active++;
1340 pos->children_weight += pos->leaf_weight;
Tejun Heo1d3650f2013-01-09 08:05:11 -08001341 vfr = vfr * pos->leaf_weight / pos->children_weight;
Tejun Heo7918ffb2013-01-09 08:05:11 -08001342
Tejun Heo1d3650f2013-01-09 08:05:11 -08001343 /*
1344 * Compound ->weight walking up the tree. Both activation and
1345 * vfraction calculation are done in the same loop. Propagation
1346 * stops once an already activated node is met. vfraction
1347 * calculation should always continue to the root.
1348 */
Tejun Heod02f7aa2013-01-09 08:05:11 -08001349 while ((parent = cfqg_parent(pos))) {
Tejun Heo1d3650f2013-01-09 08:05:11 -08001350 if (propagate) {
Toshiaki Makitae15693e2014-08-26 20:56:36 +09001351 cfq_update_group_weight(pos);
Tejun Heo1d3650f2013-01-09 08:05:11 -08001352 propagate = !parent->nr_active++;
1353 parent->children_weight += pos->weight;
1354 }
1355 vfr = vfr * pos->weight / parent->children_weight;
Tejun Heo7918ffb2013-01-09 08:05:11 -08001356 pos = parent;
1357 }
Tejun Heo1d3650f2013-01-09 08:05:11 -08001358
1359 cfqg->vfraction = max_t(unsigned, vfr, 1);
Justin TerAvest8184f932011-03-17 16:12:36 +01001360}
1361
1362static void
1363cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001364{
1365 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1366 struct cfq_group *__cfqg;
1367 struct rb_node *n;
1368
1369 cfqg->nr_cfqq++;
Gui Jianfeng760701b2010-11-30 20:52:47 +01001370 if (!RB_EMPTY_NODE(&cfqg->rb_node))
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001371 return;
1372
1373 /*
1374 * Currently put the group at the end. Later implement something
1375 * so that groups get lesser vtime based on their weights, so that
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001376 * if group does not loose all if it was not continuously backlogged.
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001377 */
1378 n = rb_last(&st->rb);
1379 if (n) {
1380 __cfqg = rb_entry_cfqg(n);
1381 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1382 } else
1383 cfqg->vdisktime = st->min_vdisktime;
Justin TerAvest8184f932011-03-17 16:12:36 +01001384 cfq_group_service_tree_add(st, cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001385}
1386
1387static void
Justin TerAvest8184f932011-03-17 16:12:36 +01001388cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1389{
Tejun Heo7918ffb2013-01-09 08:05:11 -08001390 struct cfq_group *pos = cfqg;
1391 bool propagate;
1392
1393 /*
1394 * Undo activation from cfq_group_service_tree_add(). Deactivate
1395 * @cfqg and propagate deactivation upwards.
1396 */
1397 propagate = !--pos->nr_active;
1398 pos->children_weight -= pos->leaf_weight;
1399
1400 while (propagate) {
Tejun Heod02f7aa2013-01-09 08:05:11 -08001401 struct cfq_group *parent = cfqg_parent(pos);
Tejun Heo7918ffb2013-01-09 08:05:11 -08001402
1403 /* @pos has 0 nr_active at this point */
1404 WARN_ON_ONCE(pos->children_weight);
Tejun Heo1d3650f2013-01-09 08:05:11 -08001405 pos->vfraction = 0;
Tejun Heo7918ffb2013-01-09 08:05:11 -08001406
1407 if (!parent)
1408 break;
1409
1410 propagate = !--parent->nr_active;
1411 parent->children_weight -= pos->weight;
1412 pos = parent;
1413 }
1414
1415 /* remove from the service tree */
Justin TerAvest8184f932011-03-17 16:12:36 +01001416 if (!RB_EMPTY_NODE(&cfqg->rb_node))
1417 cfq_rb_erase(&cfqg->rb_node, st);
1418}
1419
1420static void
1421cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001422{
1423 struct cfq_rb_root *st = &cfqd->grp_service_tree;
1424
1425 BUG_ON(cfqg->nr_cfqq < 1);
1426 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05001427
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001428 /* If there are other cfq queues under this group, don't delete it */
1429 if (cfqg->nr_cfqq)
1430 return;
1431
Vivek Goyal2868ef72009-12-03 12:59:48 -05001432 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Justin TerAvest8184f932011-03-17 16:12:36 +01001433 cfq_group_service_tree_del(st, cfqg);
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04001434 cfqg->saved_wl_slice = 0;
Tejun Heo155fead2012-04-01 14:38:44 -07001435 cfqg_stats_update_dequeue(cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001436}
1437
Justin TerAvest167400d2011-03-12 16:54:00 +01001438static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1439 unsigned int *unaccounted_time)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001440{
Vivek Goyalf75edf22009-12-03 12:59:53 -05001441 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001442
1443 /*
1444 * Queue got expired before even a single request completed or
1445 * got expired immediately after first request completion.
1446 */
1447 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1448 /*
1449 * Also charge the seek time incurred to the group, otherwise
1450 * if there are mutiple queues in the group, each can dispatch
1451 * a single request on seeky media and cause lots of seek time
1452 * and group will never know it.
1453 */
1454 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1455 1);
1456 } else {
1457 slice_used = jiffies - cfqq->slice_start;
Justin TerAvest167400d2011-03-12 16:54:00 +01001458 if (slice_used > cfqq->allocated_slice) {
1459 *unaccounted_time = slice_used - cfqq->allocated_slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001460 slice_used = cfqq->allocated_slice;
Justin TerAvest167400d2011-03-12 16:54:00 +01001461 }
1462 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1463 *unaccounted_time += cfqq->slice_start -
1464 cfqq->dispatch_start;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001465 }
1466
Vivek Goyaldae739e2009-12-03 12:59:45 -05001467 return slice_used;
1468}
1469
1470static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
Vivek Goyale5ff0822010-04-26 19:25:11 +02001471 struct cfq_queue *cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001472{
1473 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Justin TerAvest167400d2011-03-12 16:54:00 +01001474 unsigned int used_sl, charge, unaccounted_sl = 0;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05001475 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1476 - cfqg->service_tree_idle.count;
Tejun Heo1d3650f2013-01-09 08:05:11 -08001477 unsigned int vfr;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001478
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05001479 BUG_ON(nr_sync < 0);
Justin TerAvest167400d2011-03-12 16:54:00 +01001480 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05001481
Vivek Goyal02b35082010-08-23 12:23:53 +02001482 if (iops_mode(cfqd))
1483 charge = cfqq->slice_dispatch;
1484 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1485 charge = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001486
Tejun Heo1d3650f2013-01-09 08:05:11 -08001487 /*
1488 * Can't update vdisktime while on service tree and cfqg->vfraction
1489 * is valid only while on it. Cache vfr, leave the service tree,
1490 * update vdisktime and go back on. The re-addition to the tree
1491 * will also update the weights as necessary.
1492 */
1493 vfr = cfqg->vfraction;
Justin TerAvest8184f932011-03-17 16:12:36 +01001494 cfq_group_service_tree_del(st, cfqg);
Tejun Heo1d3650f2013-01-09 08:05:11 -08001495 cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
Justin TerAvest8184f932011-03-17 16:12:36 +01001496 cfq_group_service_tree_add(st, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -05001497
1498 /* This group is being expired. Save the context */
1499 if (time_after(cfqd->workload_expires, jiffies)) {
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04001500 cfqg->saved_wl_slice = cfqd->workload_expires
Vivek Goyaldae739e2009-12-03 12:59:45 -05001501 - jiffies;
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04001502 cfqg->saved_wl_type = cfqd->serving_wl_type;
1503 cfqg->saved_wl_class = cfqd->serving_wl_class;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001504 } else
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04001505 cfqg->saved_wl_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -05001506
1507 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1508 st->min_vdisktime);
Joe Perchesfd16d262011-06-13 10:42:49 +02001509 cfq_log_cfqq(cfqq->cfqd, cfqq,
1510 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1511 used_sl, cfqq->slice_dispatch, charge,
1512 iops_mode(cfqd), cfqq->nr_sectors);
Tejun Heo155fead2012-04-01 14:38:44 -07001513 cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1514 cfqg_stats_set_start_empty_time(cfqg);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001515}
1516
Tejun Heof51b8022012-03-05 13:15:05 -08001517/**
1518 * cfq_init_cfqg_base - initialize base part of a cfq_group
1519 * @cfqg: cfq_group to initialize
1520 *
1521 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1522 * is enabled or not.
1523 */
1524static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1525{
1526 struct cfq_rb_root *st;
1527 int i, j;
1528
1529 for_each_cfqg_st(cfqg, i, j, st)
1530 *st = CFQ_RB_ROOT;
1531 RB_CLEAR_NODE(&cfqg->rb_node);
1532
1533 cfqg->ttime.last_end_request = jiffies;
1534}
1535
Vivek Goyal25fb5162009-12-03 12:59:46 -05001536#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo69d7fde2015-08-18 14:55:36 -07001537static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1538 bool on_dfl, bool reset_dev, bool is_leaf_weight);
1539
Tejun Heo24bdb8e2015-08-18 14:55:22 -07001540static void cfqg_stats_exit(struct cfqg_stats *stats)
Peter Zijlstra90d38392013-11-12 19:42:14 -08001541{
Tejun Heo24bdb8e2015-08-18 14:55:22 -07001542 blkg_rwstat_exit(&stats->merged);
1543 blkg_rwstat_exit(&stats->service_time);
1544 blkg_rwstat_exit(&stats->wait_time);
1545 blkg_rwstat_exit(&stats->queued);
Tejun Heo24bdb8e2015-08-18 14:55:22 -07001546 blkg_stat_exit(&stats->time);
1547#ifdef CONFIG_DEBUG_BLK_CGROUP
1548 blkg_stat_exit(&stats->unaccounted_time);
1549 blkg_stat_exit(&stats->avg_queue_size_sum);
1550 blkg_stat_exit(&stats->avg_queue_size_samples);
1551 blkg_stat_exit(&stats->dequeue);
1552 blkg_stat_exit(&stats->group_wait_time);
1553 blkg_stat_exit(&stats->idle_time);
1554 blkg_stat_exit(&stats->empty_time);
1555#endif
1556}
1557
1558static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1559{
Tejun Heo77ea7332015-08-18 14:55:24 -07001560 if (blkg_rwstat_init(&stats->merged, gfp) ||
Tejun Heo24bdb8e2015-08-18 14:55:22 -07001561 blkg_rwstat_init(&stats->service_time, gfp) ||
1562 blkg_rwstat_init(&stats->wait_time, gfp) ||
1563 blkg_rwstat_init(&stats->queued, gfp) ||
Tejun Heo24bdb8e2015-08-18 14:55:22 -07001564 blkg_stat_init(&stats->time, gfp))
1565 goto err;
Peter Zijlstra90d38392013-11-12 19:42:14 -08001566
1567#ifdef CONFIG_DEBUG_BLK_CGROUP
Tejun Heo24bdb8e2015-08-18 14:55:22 -07001568 if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1569 blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1570 blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1571 blkg_stat_init(&stats->dequeue, gfp) ||
1572 blkg_stat_init(&stats->group_wait_time, gfp) ||
1573 blkg_stat_init(&stats->idle_time, gfp) ||
1574 blkg_stat_init(&stats->empty_time, gfp))
1575 goto err;
Peter Zijlstra90d38392013-11-12 19:42:14 -08001576#endif
Tejun Heo24bdb8e2015-08-18 14:55:22 -07001577 return 0;
1578err:
1579 cfqg_stats_exit(stats);
1580 return -ENOMEM;
Peter Zijlstra90d38392013-11-12 19:42:14 -08001581}
1582
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001583static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1584{
1585 struct cfq_group_data *cgd;
1586
1587 cgd = kzalloc(sizeof(*cgd), GFP_KERNEL);
1588 if (!cgd)
1589 return NULL;
1590 return &cgd->cpd;
1591}
1592
Tejun Heo81437642015-08-18 14:55:15 -07001593static void cfq_cpd_init(struct blkcg_policy_data *cpd)
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001594{
Tejun Heo81437642015-08-18 14:55:15 -07001595 struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
Tejun Heo9e10a132015-09-18 11:56:28 -04001596 unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
Tejun Heo69d7fde2015-08-18 14:55:36 -07001597 CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001598
Tejun Heo69d7fde2015-08-18 14:55:36 -07001599 if (cpd_to_blkcg(cpd) == &blkcg_root)
1600 weight *= 2;
1601
1602 cgd->weight = weight;
1603 cgd->leaf_weight = weight;
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001604}
1605
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001606static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1607{
1608 kfree(cpd_to_cfqgd(cpd));
1609}
1610
Tejun Heo69d7fde2015-08-18 14:55:36 -07001611static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1612{
1613 struct blkcg *blkcg = cpd_to_blkcg(cpd);
Tejun Heo9e10a132015-09-18 11:56:28 -04001614 bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
Tejun Heo69d7fde2015-08-18 14:55:36 -07001615 unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1616
1617 if (blkcg == &blkcg_root)
1618 weight *= 2;
1619
1620 WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1621 WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1622}
1623
Tejun Heo001bea72015-08-18 14:55:11 -07001624static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1625{
Tejun Heob2ce2642015-08-18 14:55:13 -07001626 struct cfq_group *cfqg;
1627
1628 cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1629 if (!cfqg)
1630 return NULL;
1631
1632 cfq_init_cfqg_base(cfqg);
Tejun Heo24bdb8e2015-08-18 14:55:22 -07001633 if (cfqg_stats_init(&cfqg->stats, gfp)) {
1634 kfree(cfqg);
1635 return NULL;
1636 }
Tejun Heob2ce2642015-08-18 14:55:13 -07001637
1638 return &cfqg->pd;
Tejun Heo001bea72015-08-18 14:55:11 -07001639}
1640
Tejun Heoa9520cd2015-08-18 14:55:14 -07001641static void cfq_pd_init(struct blkg_policy_data *pd)
Vivek Goyalf469a7b2011-05-19 15:38:23 -04001642{
Tejun Heoa9520cd2015-08-18 14:55:14 -07001643 struct cfq_group *cfqg = pd_to_cfqg(pd);
1644 struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001645
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001646 cfqg->weight = cgd->weight;
1647 cfqg->leaf_weight = cgd->leaf_weight;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001648}
1649
Tejun Heoa9520cd2015-08-18 14:55:14 -07001650static void cfq_pd_offline(struct blkg_policy_data *pd)
Tejun Heo0b399202013-01-09 08:05:13 -08001651{
Tejun Heoa9520cd2015-08-18 14:55:14 -07001652 struct cfq_group *cfqg = pd_to_cfqg(pd);
Tejun Heo60a83702015-08-18 14:55:05 -07001653 int i;
1654
1655 for (i = 0; i < IOPRIO_BE_NR; i++) {
1656 if (cfqg->async_cfqq[0][i])
1657 cfq_put_queue(cfqg->async_cfqq[0][i]);
1658 if (cfqg->async_cfqq[1][i])
1659 cfq_put_queue(cfqg->async_cfqq[1][i]);
1660 }
1661
1662 if (cfqg->async_idle_cfqq)
1663 cfq_put_queue(cfqg->async_idle_cfqq);
1664
Tejun Heo0b399202013-01-09 08:05:13 -08001665 /*
1666 * @blkg is going offline and will be ignored by
1667 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
1668 * that they don't get lost. If IOs complete after this point, the
1669 * stats for them will be lost. Oh well...
1670 */
Tejun Heo60a83702015-08-18 14:55:05 -07001671 cfqg_stats_xfer_dead(cfqg);
Tejun Heo0b399202013-01-09 08:05:13 -08001672}
1673
Tejun Heo001bea72015-08-18 14:55:11 -07001674static void cfq_pd_free(struct blkg_policy_data *pd)
1675{
Tejun Heo24bdb8e2015-08-18 14:55:22 -07001676 struct cfq_group *cfqg = pd_to_cfqg(pd);
1677
1678 cfqg_stats_exit(&cfqg->stats);
1679 return kfree(cfqg);
Tejun Heo001bea72015-08-18 14:55:11 -07001680}
1681
Tejun Heoa9520cd2015-08-18 14:55:14 -07001682static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
Tejun Heo689665a2013-01-09 08:05:13 -08001683{
Tejun Heoa9520cd2015-08-18 14:55:14 -07001684 struct cfq_group *cfqg = pd_to_cfqg(pd);
Tejun Heo689665a2013-01-09 08:05:13 -08001685
1686 cfqg_stats_reset(&cfqg->stats);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001687}
1688
Tejun Heoae118892015-08-18 14:55:20 -07001689static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1690 struct blkcg *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05001691{
Tejun Heoae118892015-08-18 14:55:20 -07001692 struct blkcg_gq *blkg;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001693
Tejun Heoae118892015-08-18 14:55:20 -07001694 blkg = blkg_lookup(blkcg, cfqd->queue);
1695 if (likely(blkg))
1696 return blkg_to_cfqg(blkg);
1697 return NULL;
Vivek Goyal25fb5162009-12-03 12:59:46 -05001698}
1699
1700static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1701{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001702 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001703 /* cfqq reference on cfqg */
Tejun Heoeb7d8c072012-03-23 14:02:53 +01001704 cfqg_get(cfqg);
Vivek Goyalb1c35762009-12-03 12:59:47 -05001705}
1706
Tejun Heof95a04a2012-04-16 13:57:26 -07001707static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1708 struct blkg_policy_data *pd, int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001709{
Tejun Heof95a04a2012-04-16 13:57:26 -07001710 struct cfq_group *cfqg = pd_to_cfqg(pd);
Tejun Heo3381cb82012-04-01 14:38:44 -07001711
1712 if (!cfqg->dev_weight)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001713 return 0;
Tejun Heof95a04a2012-04-16 13:57:26 -07001714 return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001715}
1716
Tejun Heo2da8ca82013-12-05 12:28:04 -05001717static int cfqg_print_weight_device(struct seq_file *sf, void *v)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001718{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001719 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1720 cfqg_prfill_weight_device, &blkcg_policy_cfq,
1721 0, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001722 return 0;
1723}
1724
Tejun Heoe71357e2013-01-09 08:05:10 -08001725static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1726 struct blkg_policy_data *pd, int off)
1727{
1728 struct cfq_group *cfqg = pd_to_cfqg(pd);
1729
1730 if (!cfqg->dev_leaf_weight)
1731 return 0;
1732 return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1733}
1734
Tejun Heo2da8ca82013-12-05 12:28:04 -05001735static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
Tejun Heoe71357e2013-01-09 08:05:10 -08001736{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001737 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1738 cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1739 0, false);
Tejun Heoe71357e2013-01-09 08:05:10 -08001740 return 0;
1741}
1742
Tejun Heo2da8ca82013-12-05 12:28:04 -05001743static int cfq_print_weight(struct seq_file *sf, void *v)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001744{
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001745 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
Jens Axboe9470e4a2015-06-19 10:19:36 -06001746 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1747 unsigned int val = 0;
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001748
Jens Axboe9470e4a2015-06-19 10:19:36 -06001749 if (cgd)
1750 val = cgd->weight;
1751
1752 seq_printf(sf, "%u\n", val);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001753 return 0;
1754}
1755
Tejun Heo2da8ca82013-12-05 12:28:04 -05001756static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
Tejun Heoe71357e2013-01-09 08:05:10 -08001757{
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001758 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
Jens Axboe9470e4a2015-06-19 10:19:36 -06001759 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1760 unsigned int val = 0;
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001761
Jens Axboe9470e4a2015-06-19 10:19:36 -06001762 if (cgd)
1763 val = cgd->leaf_weight;
1764
1765 seq_printf(sf, "%u\n", val);
Tejun Heoe71357e2013-01-09 08:05:10 -08001766 return 0;
1767}
1768
Tejun Heo451af502014-05-13 12:16:21 -04001769static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1770 char *buf, size_t nbytes, loff_t off,
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001771 bool on_dfl, bool is_leaf_weight)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001772{
Tejun Heo69d7fde2015-08-18 14:55:36 -07001773 unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1774 unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
Tejun Heo451af502014-05-13 12:16:21 -04001775 struct blkcg *blkcg = css_to_blkcg(of_css(of));
Tejun Heo60c2bc22012-04-01 14:38:43 -07001776 struct blkg_conf_ctx ctx;
Tejun Heo3381cb82012-04-01 14:38:44 -07001777 struct cfq_group *cfqg;
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001778 struct cfq_group_data *cfqgd;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001779 int ret;
Tejun Heo36aa9e52015-08-18 14:55:31 -07001780 u64 v;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001781
Tejun Heo3c798392012-04-16 13:57:25 -07001782 ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001783 if (ret)
1784 return ret;
1785
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001786 if (sscanf(ctx.body, "%llu", &v) == 1) {
1787 /* require "default" on dfl */
1788 ret = -ERANGE;
1789 if (!v && on_dfl)
1790 goto out_finish;
1791 } else if (!strcmp(strim(ctx.body), "default")) {
1792 v = 0;
1793 } else {
1794 ret = -EINVAL;
Tejun Heo36aa9e52015-08-18 14:55:31 -07001795 goto out_finish;
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001796 }
Tejun Heo36aa9e52015-08-18 14:55:31 -07001797
Tejun Heo3381cb82012-04-01 14:38:44 -07001798 cfqg = blkg_to_cfqg(ctx.blkg);
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001799 cfqgd = blkcg_to_cfqgd(blkcg);
Jens Axboeae994ea2015-06-20 10:26:50 -06001800
Tejun Heo20386ce2015-08-18 14:55:28 -07001801 ret = -ERANGE;
Tejun Heo69d7fde2015-08-18 14:55:36 -07001802 if (!v || (v >= min && v <= max)) {
Tejun Heoe71357e2013-01-09 08:05:10 -08001803 if (!is_leaf_weight) {
Tejun Heo36aa9e52015-08-18 14:55:31 -07001804 cfqg->dev_weight = v;
1805 cfqg->new_weight = v ?: cfqgd->weight;
Tejun Heoe71357e2013-01-09 08:05:10 -08001806 } else {
Tejun Heo36aa9e52015-08-18 14:55:31 -07001807 cfqg->dev_leaf_weight = v;
1808 cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
Tejun Heoe71357e2013-01-09 08:05:10 -08001809 }
Tejun Heo60c2bc22012-04-01 14:38:43 -07001810 ret = 0;
1811 }
Tejun Heo36aa9e52015-08-18 14:55:31 -07001812out_finish:
Tejun Heo60c2bc22012-04-01 14:38:43 -07001813 blkg_conf_finish(&ctx);
Tejun Heo451af502014-05-13 12:16:21 -04001814 return ret ?: nbytes;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001815}
1816
Tejun Heo451af502014-05-13 12:16:21 -04001817static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1818 char *buf, size_t nbytes, loff_t off)
Tejun Heoe71357e2013-01-09 08:05:10 -08001819{
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001820 return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
Tejun Heoe71357e2013-01-09 08:05:10 -08001821}
1822
Tejun Heo451af502014-05-13 12:16:21 -04001823static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1824 char *buf, size_t nbytes, loff_t off)
Tejun Heoe71357e2013-01-09 08:05:10 -08001825{
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001826 return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
Tejun Heoe71357e2013-01-09 08:05:10 -08001827}
1828
Tejun Heodd165eb2015-08-18 14:55:33 -07001829static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
Tejun Heo69d7fde2015-08-18 14:55:36 -07001830 bool on_dfl, bool reset_dev, bool is_leaf_weight)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001831{
Tejun Heo69d7fde2015-08-18 14:55:36 -07001832 unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1833 unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
Tejun Heo182446d2013-08-08 20:11:24 -04001834 struct blkcg *blkcg = css_to_blkcg(css);
Tejun Heo3c798392012-04-16 13:57:25 -07001835 struct blkcg_gq *blkg;
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001836 struct cfq_group_data *cfqgd;
Jens Axboeae994ea2015-06-20 10:26:50 -06001837 int ret = 0;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001838
Tejun Heo69d7fde2015-08-18 14:55:36 -07001839 if (val < min || val > max)
1840 return -ERANGE;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001841
1842 spin_lock_irq(&blkcg->lock);
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001843 cfqgd = blkcg_to_cfqgd(blkcg);
Jens Axboeae994ea2015-06-20 10:26:50 -06001844 if (!cfqgd) {
1845 ret = -EINVAL;
1846 goto out;
1847 }
Tejun Heoe71357e2013-01-09 08:05:10 -08001848
1849 if (!is_leaf_weight)
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001850 cfqgd->weight = val;
Tejun Heoe71357e2013-01-09 08:05:10 -08001851 else
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001852 cfqgd->leaf_weight = val;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001853
Sasha Levinb67bfe02013-02-27 17:06:00 -08001854 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
Tejun Heo3381cb82012-04-01 14:38:44 -07001855 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001856
Tejun Heoe71357e2013-01-09 08:05:10 -08001857 if (!cfqg)
1858 continue;
1859
1860 if (!is_leaf_weight) {
Tejun Heo69d7fde2015-08-18 14:55:36 -07001861 if (reset_dev)
1862 cfqg->dev_weight = 0;
Tejun Heoe71357e2013-01-09 08:05:10 -08001863 if (!cfqg->dev_weight)
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001864 cfqg->new_weight = cfqgd->weight;
Tejun Heoe71357e2013-01-09 08:05:10 -08001865 } else {
Tejun Heo69d7fde2015-08-18 14:55:36 -07001866 if (reset_dev)
1867 cfqg->dev_leaf_weight = 0;
Tejun Heoe71357e2013-01-09 08:05:10 -08001868 if (!cfqg->dev_leaf_weight)
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001869 cfqg->new_leaf_weight = cfqgd->leaf_weight;
Tejun Heoe71357e2013-01-09 08:05:10 -08001870 }
Tejun Heo60c2bc22012-04-01 14:38:43 -07001871 }
1872
Jens Axboeae994ea2015-06-20 10:26:50 -06001873out:
Tejun Heo60c2bc22012-04-01 14:38:43 -07001874 spin_unlock_irq(&blkcg->lock);
Jens Axboeae994ea2015-06-20 10:26:50 -06001875 return ret;
Tejun Heo60c2bc22012-04-01 14:38:43 -07001876}
1877
Tejun Heo182446d2013-08-08 20:11:24 -04001878static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1879 u64 val)
Tejun Heoe71357e2013-01-09 08:05:10 -08001880{
Tejun Heo69d7fde2015-08-18 14:55:36 -07001881 return __cfq_set_weight(css, val, false, false, false);
Tejun Heoe71357e2013-01-09 08:05:10 -08001882}
1883
Tejun Heo182446d2013-08-08 20:11:24 -04001884static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1885 struct cftype *cft, u64 val)
Tejun Heoe71357e2013-01-09 08:05:10 -08001886{
Tejun Heo69d7fde2015-08-18 14:55:36 -07001887 return __cfq_set_weight(css, val, false, false, true);
Tejun Heoe71357e2013-01-09 08:05:10 -08001888}
1889
Tejun Heo2da8ca82013-12-05 12:28:04 -05001890static int cfqg_print_stat(struct seq_file *sf, void *v)
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001891{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001892 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1893 &blkcg_policy_cfq, seq_cft(sf)->private, false);
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001894 return 0;
1895}
1896
Tejun Heo2da8ca82013-12-05 12:28:04 -05001897static int cfqg_print_rwstat(struct seq_file *sf, void *v)
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001898{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001899 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1900 &blkcg_policy_cfq, seq_cft(sf)->private, true);
Tejun Heo5bc4afb12012-04-01 14:38:45 -07001901 return 0;
1902}
1903
Tejun Heo43114012013-01-09 08:05:13 -08001904static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1905 struct blkg_policy_data *pd, int off)
1906{
Tejun Heof12c74c2015-08-18 14:55:23 -07001907 u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1908 &blkcg_policy_cfq, off);
Tejun Heo43114012013-01-09 08:05:13 -08001909 return __blkg_prfill_u64(sf, pd, sum);
1910}
1911
1912static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1913 struct blkg_policy_data *pd, int off)
1914{
Tejun Heof12c74c2015-08-18 14:55:23 -07001915 struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1916 &blkcg_policy_cfq, off);
Tejun Heo43114012013-01-09 08:05:13 -08001917 return __blkg_prfill_rwstat(sf, pd, &sum);
1918}
1919
Tejun Heo2da8ca82013-12-05 12:28:04 -05001920static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
Tejun Heo43114012013-01-09 08:05:13 -08001921{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001922 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1923 cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1924 seq_cft(sf)->private, false);
Tejun Heo43114012013-01-09 08:05:13 -08001925 return 0;
1926}
1927
Tejun Heo2da8ca82013-12-05 12:28:04 -05001928static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
Tejun Heo43114012013-01-09 08:05:13 -08001929{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001930 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1931 cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1932 seq_cft(sf)->private, true);
Tejun Heo43114012013-01-09 08:05:13 -08001933 return 0;
1934}
1935
Tejun Heo702747c2015-08-18 14:55:25 -07001936static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1937 int off)
1938{
1939 u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1940
1941 return __blkg_prfill_u64(sf, pd, sum >> 9);
1942}
1943
1944static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1945{
1946 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1947 cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1948 return 0;
1949}
1950
1951static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1952 struct blkg_policy_data *pd, int off)
1953{
1954 struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1955 offsetof(struct blkcg_gq, stat_bytes));
1956 u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1957 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1958
1959 return __blkg_prfill_u64(sf, pd, sum >> 9);
1960}
1961
1962static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1963{
1964 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1965 cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1966 false);
1967 return 0;
1968}
1969
Tejun Heo60c2bc22012-04-01 14:38:43 -07001970#ifdef CONFIG_DEBUG_BLK_CGROUP
Tejun Heof95a04a2012-04-16 13:57:26 -07001971static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1972 struct blkg_policy_data *pd, int off)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001973{
Tejun Heof95a04a2012-04-16 13:57:26 -07001974 struct cfq_group *cfqg = pd_to_cfqg(pd);
Tejun Heo155fead2012-04-01 14:38:44 -07001975 u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001976 u64 v = 0;
1977
1978 if (samples) {
Tejun Heo155fead2012-04-01 14:38:44 -07001979 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
Anatol Pomozovf3cff252013-09-22 12:43:47 -06001980 v = div64_u64(v, samples);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001981 }
Tejun Heof95a04a2012-04-16 13:57:26 -07001982 __blkg_prfill_u64(sf, pd, v);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001983 return 0;
1984}
1985
1986/* print avg_queue_size */
Tejun Heo2da8ca82013-12-05 12:28:04 -05001987static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
Tejun Heo60c2bc22012-04-01 14:38:43 -07001988{
Tejun Heo2da8ca82013-12-05 12:28:04 -05001989 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1990 cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1991 0, false);
Tejun Heo60c2bc22012-04-01 14:38:43 -07001992 return 0;
1993}
1994#endif /* CONFIG_DEBUG_BLK_CGROUP */
1995
Tejun Heo880f50e2015-08-18 14:55:30 -07001996static struct cftype cfq_blkcg_legacy_files[] = {
Tejun Heo1d3650f2013-01-09 08:05:11 -08001997 /* on root, weight is mapped to leaf_weight */
Tejun Heo60c2bc22012-04-01 14:38:43 -07001998 {
1999 .name = "weight_device",
Tejun Heo1d3650f2013-01-09 08:05:11 -08002000 .flags = CFTYPE_ONLY_ON_ROOT,
Tejun Heo2da8ca82013-12-05 12:28:04 -05002001 .seq_show = cfqg_print_leaf_weight_device,
Tejun Heo451af502014-05-13 12:16:21 -04002002 .write = cfqg_set_leaf_weight_device,
Tejun Heo1d3650f2013-01-09 08:05:11 -08002003 },
2004 {
2005 .name = "weight",
2006 .flags = CFTYPE_ONLY_ON_ROOT,
Tejun Heo2da8ca82013-12-05 12:28:04 -05002007 .seq_show = cfq_print_leaf_weight,
Tejun Heo1d3650f2013-01-09 08:05:11 -08002008 .write_u64 = cfq_set_leaf_weight,
2009 },
2010
2011 /* no such mapping necessary for !roots */
2012 {
2013 .name = "weight_device",
2014 .flags = CFTYPE_NOT_ON_ROOT,
Tejun Heo2da8ca82013-12-05 12:28:04 -05002015 .seq_show = cfqg_print_weight_device,
Tejun Heo451af502014-05-13 12:16:21 -04002016 .write = cfqg_set_weight_device,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002017 },
2018 {
2019 .name = "weight",
Tejun Heoe71357e2013-01-09 08:05:10 -08002020 .flags = CFTYPE_NOT_ON_ROOT,
Tejun Heo2da8ca82013-12-05 12:28:04 -05002021 .seq_show = cfq_print_weight,
Tejun Heo3381cb82012-04-01 14:38:44 -07002022 .write_u64 = cfq_set_weight,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002023 },
Tejun Heo1d3650f2013-01-09 08:05:11 -08002024
2025 {
2026 .name = "leaf_weight_device",
Tejun Heo2da8ca82013-12-05 12:28:04 -05002027 .seq_show = cfqg_print_leaf_weight_device,
Tejun Heo451af502014-05-13 12:16:21 -04002028 .write = cfqg_set_leaf_weight_device,
Tejun Heoe71357e2013-01-09 08:05:10 -08002029 },
2030 {
2031 .name = "leaf_weight",
Tejun Heo2da8ca82013-12-05 12:28:04 -05002032 .seq_show = cfq_print_leaf_weight,
Tejun Heoe71357e2013-01-09 08:05:10 -08002033 .write_u64 = cfq_set_leaf_weight,
2034 },
2035
Tejun Heo43114012013-01-09 08:05:13 -08002036 /* statistics, covers only the tasks in the cfqg */
Tejun Heo60c2bc22012-04-01 14:38:43 -07002037 {
2038 .name = "time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002039 .private = offsetof(struct cfq_group, stats.time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002040 .seq_show = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002041 },
2042 {
2043 .name = "sectors",
Tejun Heo702747c2015-08-18 14:55:25 -07002044 .seq_show = cfqg_print_stat_sectors,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002045 },
2046 {
2047 .name = "io_service_bytes",
Tejun Heo77ea7332015-08-18 14:55:24 -07002048 .private = (unsigned long)&blkcg_policy_cfq,
2049 .seq_show = blkg_print_stat_bytes,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002050 },
2051 {
2052 .name = "io_serviced",
Tejun Heo77ea7332015-08-18 14:55:24 -07002053 .private = (unsigned long)&blkcg_policy_cfq,
2054 .seq_show = blkg_print_stat_ios,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002055 },
2056 {
2057 .name = "io_service_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002058 .private = offsetof(struct cfq_group, stats.service_time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002059 .seq_show = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002060 },
2061 {
2062 .name = "io_wait_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002063 .private = offsetof(struct cfq_group, stats.wait_time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002064 .seq_show = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002065 },
2066 {
2067 .name = "io_merged",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002068 .private = offsetof(struct cfq_group, stats.merged),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002069 .seq_show = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002070 },
2071 {
2072 .name = "io_queued",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002073 .private = offsetof(struct cfq_group, stats.queued),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002074 .seq_show = cfqg_print_rwstat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002075 },
Tejun Heo43114012013-01-09 08:05:13 -08002076
2077 /* the same statictics which cover the cfqg and its descendants */
2078 {
2079 .name = "time_recursive",
2080 .private = offsetof(struct cfq_group, stats.time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002081 .seq_show = cfqg_print_stat_recursive,
Tejun Heo43114012013-01-09 08:05:13 -08002082 },
2083 {
2084 .name = "sectors_recursive",
Tejun Heo702747c2015-08-18 14:55:25 -07002085 .seq_show = cfqg_print_stat_sectors_recursive,
Tejun Heo43114012013-01-09 08:05:13 -08002086 },
2087 {
2088 .name = "io_service_bytes_recursive",
Tejun Heo77ea7332015-08-18 14:55:24 -07002089 .private = (unsigned long)&blkcg_policy_cfq,
2090 .seq_show = blkg_print_stat_bytes_recursive,
Tejun Heo43114012013-01-09 08:05:13 -08002091 },
2092 {
2093 .name = "io_serviced_recursive",
Tejun Heo77ea7332015-08-18 14:55:24 -07002094 .private = (unsigned long)&blkcg_policy_cfq,
2095 .seq_show = blkg_print_stat_ios_recursive,
Tejun Heo43114012013-01-09 08:05:13 -08002096 },
2097 {
2098 .name = "io_service_time_recursive",
2099 .private = offsetof(struct cfq_group, stats.service_time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002100 .seq_show = cfqg_print_rwstat_recursive,
Tejun Heo43114012013-01-09 08:05:13 -08002101 },
2102 {
2103 .name = "io_wait_time_recursive",
2104 .private = offsetof(struct cfq_group, stats.wait_time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002105 .seq_show = cfqg_print_rwstat_recursive,
Tejun Heo43114012013-01-09 08:05:13 -08002106 },
2107 {
2108 .name = "io_merged_recursive",
2109 .private = offsetof(struct cfq_group, stats.merged),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002110 .seq_show = cfqg_print_rwstat_recursive,
Tejun Heo43114012013-01-09 08:05:13 -08002111 },
2112 {
2113 .name = "io_queued_recursive",
2114 .private = offsetof(struct cfq_group, stats.queued),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002115 .seq_show = cfqg_print_rwstat_recursive,
Tejun Heo43114012013-01-09 08:05:13 -08002116 },
Tejun Heo60c2bc22012-04-01 14:38:43 -07002117#ifdef CONFIG_DEBUG_BLK_CGROUP
2118 {
2119 .name = "avg_queue_size",
Tejun Heo2da8ca82013-12-05 12:28:04 -05002120 .seq_show = cfqg_print_avg_queue_size,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002121 },
2122 {
2123 .name = "group_wait_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002124 .private = offsetof(struct cfq_group, stats.group_wait_time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002125 .seq_show = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002126 },
2127 {
2128 .name = "idle_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002129 .private = offsetof(struct cfq_group, stats.idle_time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002130 .seq_show = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002131 },
2132 {
2133 .name = "empty_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002134 .private = offsetof(struct cfq_group, stats.empty_time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002135 .seq_show = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002136 },
2137 {
2138 .name = "dequeue",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002139 .private = offsetof(struct cfq_group, stats.dequeue),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002140 .seq_show = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002141 },
2142 {
2143 .name = "unaccounted_time",
Tejun Heo5bc4afb12012-04-01 14:38:45 -07002144 .private = offsetof(struct cfq_group, stats.unaccounted_time),
Tejun Heo2da8ca82013-12-05 12:28:04 -05002145 .seq_show = cfqg_print_stat,
Tejun Heo60c2bc22012-04-01 14:38:43 -07002146 },
2147#endif /* CONFIG_DEBUG_BLK_CGROUP */
2148 { } /* terminate */
2149};
Tejun Heo2ee867dc2015-08-18 14:55:34 -07002150
2151static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2152{
2153 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2154 struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2155
2156 seq_printf(sf, "default %u\n", cgd->weight);
2157 blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2158 &blkcg_policy_cfq, 0, false);
2159 return 0;
2160}
2161
2162static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2163 char *buf, size_t nbytes, loff_t off)
2164{
2165 char *endp;
2166 int ret;
2167 u64 v;
2168
2169 buf = strim(buf);
2170
2171 /* "WEIGHT" or "default WEIGHT" sets the default weight */
2172 v = simple_strtoull(buf, &endp, 0);
2173 if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
Tejun Heo69d7fde2015-08-18 14:55:36 -07002174 ret = __cfq_set_weight(of_css(of), v, true, false, false);
Tejun Heo2ee867dc2015-08-18 14:55:34 -07002175 return ret ?: nbytes;
2176 }
2177
2178 /* "MAJ:MIN WEIGHT" */
2179 return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2180}
2181
2182static struct cftype cfq_blkcg_files[] = {
2183 {
2184 .name = "weight",
2185 .flags = CFTYPE_NOT_ON_ROOT,
2186 .seq_show = cfq_print_weight_on_dfl,
2187 .write = cfq_set_weight_on_dfl,
2188 },
2189 { } /* terminate */
2190};
2191
Vivek Goyal25fb5162009-12-03 12:59:46 -05002192#else /* GROUP_IOSCHED */
Tejun Heoae118892015-08-18 14:55:20 -07002193static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2194 struct blkcg *blkcg)
Vivek Goyal25fb5162009-12-03 12:59:46 -05002195{
Tejun Heof51b8022012-03-05 13:15:05 -08002196 return cfqd->root_group;
Vivek Goyal25fb5162009-12-03 12:59:46 -05002197}
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02002198
Vivek Goyal25fb5162009-12-03 12:59:46 -05002199static inline void
2200cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2201 cfqq->cfqg = cfqg;
2202}
2203
2204#endif /* GROUP_IOSCHED */
2205
Jens Axboe498d3aa22007-04-26 12:54:48 +02002206/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002207 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02002208 * requests waiting to be processed. It is sorted in the order that
2209 * we will service the queues.
2210 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002211static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002212 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02002213{
Jens Axboe08717142008-01-28 11:38:15 +01002214 struct rb_node **p, *parent;
2215 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002216 unsigned long rb_key;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002217 struct cfq_rb_root *st;
Jens Axboe498d3aa22007-04-26 12:54:48 +02002218 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002219 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05002220
Vivek Goyal34b98d02012-10-03 16:56:58 -04002221 st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
Jens Axboe08717142008-01-28 11:38:15 +01002222 if (cfq_class_idle(cfqq)) {
2223 rb_key = CFQ_IDLE_DELAY;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002224 parent = rb_last(&st->rb);
Jens Axboe08717142008-01-28 11:38:15 +01002225 if (parent && parent != &cfqq->rb_node) {
2226 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2227 rb_key += __cfqq->rb_key;
2228 } else
2229 rb_key += jiffies;
2230 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02002231 /*
2232 * Get our rb key offset. Subtract any residual slice
2233 * value carried from last service. A negative resid
2234 * count indicates slice overrun, and this should position
2235 * the next service time further away in the tree.
2236 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02002237 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02002238 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02002239 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02002240 } else {
2241 rb_key = -HZ;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002242 __cfqq = cfq_rb_first(st);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02002243 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2244 }
Jens Axboed9e76202007-04-20 14:27:50 +02002245
2246 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05002247 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01002248 /*
Jens Axboed9e76202007-04-20 14:27:50 +02002249 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01002250 */
Vivek Goyal34b98d02012-10-03 16:56:58 -04002251 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
Jens Axboed9e76202007-04-20 14:27:50 +02002252 return;
Jens Axboe53b037442006-07-28 09:48:51 +02002253
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01002254 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2255 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002256 }
Jens Axboed9e76202007-04-20 14:27:50 +02002257
Jens Axboe498d3aa22007-04-26 12:54:48 +02002258 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01002259 parent = NULL;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002260 cfqq->service_tree = st;
2261 p = &st->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02002262 while (*p) {
2263 parent = *p;
2264 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2265
Jens Axboe0c534e02007-04-18 20:01:57 +02002266 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002267 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02002268 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002269 if (time_before(rb_key, __cfqq->rb_key))
Vivek Goyal1f23f122012-10-03 16:57:00 -04002270 p = &parent->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002271 else {
Vivek Goyal1f23f122012-10-03 16:57:00 -04002272 p = &parent->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02002273 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002274 }
Jens Axboed9e76202007-04-20 14:27:50 +02002275 }
2276
Jens Axboecc09e292007-04-26 12:53:50 +02002277 if (left)
Vivek Goyal34b98d02012-10-03 16:56:58 -04002278 st->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02002279
Jens Axboed9e76202007-04-20 14:27:50 +02002280 cfqq->rb_key = rb_key;
2281 rb_link_node(&cfqq->rb_node, parent, p);
Vivek Goyal34b98d02012-10-03 16:56:58 -04002282 rb_insert_color(&cfqq->rb_node, &st->rb);
2283 st->count++;
Namhyung Kim20359f22011-05-24 10:23:22 +02002284 if (add_front || !new_cfqq)
Vivek Goyaldae739e2009-12-03 12:59:45 -05002285 return;
Justin TerAvest8184f932011-03-17 16:12:36 +01002286 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287}
2288
Jens Axboea36e71f2009-04-15 12:15:11 +02002289static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002290cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2291 sector_t sector, struct rb_node **ret_parent,
2292 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02002293{
Jens Axboea36e71f2009-04-15 12:15:11 +02002294 struct rb_node **p, *parent;
2295 struct cfq_queue *cfqq = NULL;
2296
2297 parent = NULL;
2298 p = &root->rb_node;
2299 while (*p) {
2300 struct rb_node **n;
2301
2302 parent = *p;
2303 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2304
2305 /*
2306 * Sort strictly based on sector. Smallest to the left,
2307 * largest to the right.
2308 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09002309 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002310 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09002311 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002312 n = &(*p)->rb_left;
2313 else
2314 break;
2315 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02002316 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02002317 }
2318
2319 *ret_parent = parent;
2320 if (rb_link)
2321 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02002322 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02002323}
2324
2325static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2326{
Jens Axboea36e71f2009-04-15 12:15:11 +02002327 struct rb_node **p, *parent;
2328 struct cfq_queue *__cfqq;
2329
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002330 if (cfqq->p_root) {
2331 rb_erase(&cfqq->p_node, cfqq->p_root);
2332 cfqq->p_root = NULL;
2333 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002334
2335 if (cfq_class_idle(cfqq))
2336 return;
2337 if (!cfqq->next_rq)
2338 return;
2339
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002340 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09002341 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2342 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02002343 if (!__cfqq) {
2344 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002345 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2346 } else
2347 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02002348}
2349
Jens Axboe498d3aa22007-04-26 12:54:48 +02002350/*
2351 * Update cfqq's position in the service tree.
2352 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02002353static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002354{
Jens Axboe6d048f52007-04-25 12:44:27 +02002355 /*
2356 * Resorting requires the cfqq to be on the RR list already.
2357 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002358 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02002359 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02002360 cfq_prio_tree_add(cfqd, cfqq);
2361 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002362}
2363
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364/*
2365 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02002366 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 */
Jens Axboefebffd62008-01-28 13:19:43 +01002368static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369{
Jens Axboe7b679132008-05-30 12:23:07 +02002370 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02002371 BUG_ON(cfq_cfqq_on_rr(cfqq));
2372 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 cfqd->busy_queues++;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002374 if (cfq_cfqq_sync(cfqq))
2375 cfqd->busy_sync_queues++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376
Jens Axboeedd75ff2007-04-19 12:03:34 +02002377 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378}
2379
Jens Axboe498d3aa22007-04-26 12:54:48 +02002380/*
2381 * Called when the cfqq no longer has requests pending, remove it from
2382 * the service tree.
2383 */
Jens Axboefebffd62008-01-28 13:19:43 +01002384static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385{
Jens Axboe7b679132008-05-30 12:23:07 +02002386 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02002387 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2388 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01002390 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2391 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2392 cfqq->service_tree = NULL;
2393 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002394 if (cfqq->p_root) {
2395 rb_erase(&cfqq->p_node, cfqq->p_root);
2396 cfqq->p_root = NULL;
2397 }
Jens Axboed9e76202007-04-20 14:27:50 +02002398
Justin TerAvest8184f932011-03-17 16:12:36 +01002399 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 BUG_ON(!cfqd->busy_queues);
2401 cfqd->busy_queues--;
Shaohua Lief8a41d2011-03-07 09:26:29 +01002402 if (cfq_cfqq_sync(cfqq))
2403 cfqd->busy_sync_queues--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404}
2405
2406/*
2407 * rb tree support functions
2408 */
Jens Axboefebffd62008-01-28 13:19:43 +01002409static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410{
Jens Axboe5e705372006-07-13 12:39:25 +02002411 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02002412 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413
Jens Axboeb4878f22005-10-20 16:42:29 +02002414 BUG_ON(!cfqq->queued[sync]);
2415 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416
Jens Axboe5e705372006-07-13 12:39:25 +02002417 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418
Vivek Goyalf04a6422009-12-03 12:59:40 -05002419 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2420 /*
2421 * Queue will be deleted from service tree when we actually
2422 * expire it later. Right now just remove it from prio tree
2423 * as it is empty.
2424 */
2425 if (cfqq->p_root) {
2426 rb_erase(&cfqq->p_node, cfqq->p_root);
2427 cfqq->p_root = NULL;
2428 }
2429 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430}
2431
Jens Axboe5e705372006-07-13 12:39:25 +02002432static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433{
Jens Axboe5e705372006-07-13 12:39:25 +02002434 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 struct cfq_data *cfqd = cfqq->cfqd;
Jeff Moyer796d5112011-06-02 21:19:05 +02002436 struct request *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437
Jens Axboe5380a102006-07-13 12:37:56 +02002438 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439
Jeff Moyer796d5112011-06-02 21:19:05 +02002440 elv_rb_add(&cfqq->sort_list, rq);
Jens Axboe5fccbf62006-10-31 14:21:55 +01002441
2442 if (!cfq_cfqq_on_rr(cfqq))
2443 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02002444
2445 /*
2446 * check if this request is a better next-serve candidate
2447 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002448 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01002449 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02002450
2451 /*
2452 * adjust priority tree position, if ->next_rq changes
2453 */
2454 if (prev != cfqq->next_rq)
2455 cfq_prio_tree_add(cfqd, cfqq);
2456
Jens Axboe5044eed2007-04-25 11:53:48 +02002457 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458}
2459
Jens Axboefebffd62008-01-28 13:19:43 +01002460static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461{
Jens Axboe5380a102006-07-13 12:37:56 +02002462 elv_rb_del(&cfqq->sort_list, rq);
2463 cfqq->queued[rq_is_sync(rq)]--;
Tejun Heo155fead2012-04-01 14:38:44 -07002464 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
Jens Axboe5e705372006-07-13 12:39:25 +02002465 cfq_add_rq_rb(rq);
Tejun Heo155fead2012-04-01 14:38:44 -07002466 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2467 rq->cmd_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468}
2469
Jens Axboe206dc692006-03-28 13:03:44 +02002470static struct request *
2471cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472{
Jens Axboe206dc692006-03-28 13:03:44 +02002473 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01002474 struct cfq_io_cq *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02002475 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476
Jens Axboe4ac845a2008-01-24 08:44:49 +01002477 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002478 if (!cic)
2479 return NULL;
2480
2481 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Kent Overstreetf73a1c72012-09-25 15:05:12 -07002482 if (cfqq)
2483 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 return NULL;
2486}
2487
Jens Axboe165125e2007-07-24 09:28:11 +02002488static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02002489{
2490 struct cfq_data *cfqd = q->elevator->elevator_data;
2491
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002492 cfqd->rq_in_driver++;
Jens Axboe7b679132008-05-30 12:23:07 +02002493 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002494 cfqd->rq_in_driver);
Jens Axboe25776e32006-06-01 10:12:26 +02002495
Tejun Heo5b936292009-05-07 22:24:38 +09002496 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02002497}
2498
Jens Axboe165125e2007-07-24 09:28:11 +02002499static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500{
Jens Axboe22e2c502005-06-27 10:55:12 +02002501 struct cfq_data *cfqd = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002503 WARN_ON(!cfqd->rq_in_driver);
2504 cfqd->rq_in_driver--;
Jens Axboe7b679132008-05-30 12:23:07 +02002505 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01002506 cfqd->rq_in_driver);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507}
2508
Jens Axboeb4878f22005-10-20 16:42:29 +02002509static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510{
Jens Axboe5e705372006-07-13 12:39:25 +02002511 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02002512
Jens Axboe5e705372006-07-13 12:39:25 +02002513 if (cfqq->next_rq == rq)
2514 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515
Jens Axboeb4878f22005-10-20 16:42:29 +02002516 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02002517 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02002518
Aaron Carroll45333d52008-08-26 15:52:36 +02002519 cfqq->cfqd->rq_queued--;
Tejun Heo155fead2012-04-01 14:38:44 -07002520 cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
Christoph Hellwig65299a32011-08-23 14:50:29 +02002521 if (rq->cmd_flags & REQ_PRIO) {
2522 WARN_ON(!cfqq->prio_pending);
2523 cfqq->prio_pending--;
Jens Axboeb53d1ed2011-08-19 08:34:48 +02002524 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525}
2526
Jens Axboe165125e2007-07-24 09:28:11 +02002527static int cfq_merge(struct request_queue *q, struct request **req,
2528 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529{
2530 struct cfq_data *cfqd = q->elevator->elevator_data;
2531 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532
Jens Axboe206dc692006-03-28 13:03:44 +02002533 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002534 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02002535 *req = __rq;
2536 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 }
2538
2539 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540}
2541
Jens Axboe165125e2007-07-24 09:28:11 +02002542static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02002543 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544{
Jens Axboe21183b02006-07-13 12:33:14 +02002545 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02002546 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547
Jens Axboe5e705372006-07-13 12:39:25 +02002548 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550}
2551
Divyesh Shah812d4022010-04-08 21:14:23 -07002552static void cfq_bio_merged(struct request_queue *q, struct request *req,
2553 struct bio *bio)
2554{
Tejun Heo155fead2012-04-01 14:38:44 -07002555 cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
Divyesh Shah812d4022010-04-08 21:14:23 -07002556}
2557
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558static void
Jens Axboe165125e2007-07-24 09:28:11 +02002559cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 struct request *next)
2561{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01002562 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Shaohua Li4a0b75c2011-12-16 14:00:22 +01002563 struct cfq_data *cfqd = q->elevator->elevator_data;
2564
Jens Axboe22e2c502005-06-27 10:55:12 +02002565 /*
2566 * reposition in fifo if next is older than rq
2567 */
2568 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jan Kara8b4922d2014-02-24 16:39:52 +01002569 time_before(next->fifo_time, rq->fifo_time) &&
Shaohua Li3d106fba2012-11-06 12:39:51 +01002570 cfqq == RQ_CFQQ(next)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002571 list_move(&rq->queuelist, &next->queuelist);
Jan Kara8b4922d2014-02-24 16:39:52 +01002572 rq->fifo_time = next->fifo_time;
Jens Axboe30996f42009-10-05 11:03:39 +02002573 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002574
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01002575 if (cfqq->next_rq == next)
2576 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02002577 cfq_remove_request(next);
Tejun Heo155fead2012-04-01 14:38:44 -07002578 cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
Shaohua Li4a0b75c2011-12-16 14:00:22 +01002579
2580 cfqq = RQ_CFQQ(next);
2581 /*
2582 * all requests of this queue are merged to other queues, delete it
2583 * from the service tree. If it's the active_queue,
2584 * cfq_dispatch_requests() will choose to expire it or do idle
2585 */
2586 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2587 cfqq != cfqd->active_queue)
2588 cfq_del_cfqq_rr(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002589}
2590
Jens Axboe165125e2007-07-24 09:28:11 +02002591static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01002592 struct bio *bio)
2593{
2594 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heoc5869802011-12-14 00:33:41 +01002595 struct cfq_io_cq *cic;
Jens Axboeda775262006-12-20 11:04:12 +01002596 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01002597
2598 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01002599 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01002600 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02002601 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02002602 return false;
Jens Axboeda775262006-12-20 11:04:12 +01002603
2604 /*
Tejun Heof1a4f4d2011-12-14 00:33:39 +01002605 * Lookup the cfqq that this bio will be queued with and allow
Tejun Heo07c2bd32012-02-08 09:19:42 +01002606 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01002607 */
Tejun Heo07c2bd32012-02-08 09:19:42 +01002608 cic = cfq_cic_lookup(cfqd, current->io_context);
2609 if (!cic)
2610 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01002611
Vasily Tarasov91fac312007-04-25 12:29:51 +02002612 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02002613 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01002614}
2615
Divyesh Shah812df482010-04-08 21:15:35 -07002616static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2617{
2618 del_timer(&cfqd->idle_slice_timer);
Tejun Heo155fead2012-04-01 14:38:44 -07002619 cfqg_stats_update_idle_time(cfqq->cfqg);
Divyesh Shah812df482010-04-08 21:15:35 -07002620}
2621
Jens Axboefebffd62008-01-28 13:19:43 +01002622static void __cfq_set_active_queue(struct cfq_data *cfqd,
2623 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02002624{
2625 if (cfqq) {
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04002626 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04002627 cfqd->serving_wl_class, cfqd->serving_wl_type);
Tejun Heo155fead2012-04-01 14:38:44 -07002628 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
Justin TerAvest62a37f62011-03-23 08:25:44 +01002629 cfqq->slice_start = 0;
2630 cfqq->dispatch_start = jiffies;
2631 cfqq->allocated_slice = 0;
2632 cfqq->slice_end = 0;
2633 cfqq->slice_dispatch = 0;
2634 cfqq->nr_sectors = 0;
2635
2636 cfq_clear_cfqq_wait_request(cfqq);
2637 cfq_clear_cfqq_must_dispatch(cfqq);
2638 cfq_clear_cfqq_must_alloc_slice(cfqq);
2639 cfq_clear_cfqq_fifo_expire(cfqq);
2640 cfq_mark_cfqq_slice_new(cfqq);
2641
2642 cfq_del_timer(cfqd, cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002643 }
2644
2645 cfqd->active_queue = cfqq;
2646}
2647
2648/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002649 * current cfqq expired its slice (or was too idle), select new one
2650 */
2651static void
2652__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Vivek Goyale5ff0822010-04-26 19:25:11 +02002653 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002654{
Jens Axboe7b679132008-05-30 12:23:07 +02002655 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2656
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002657 if (cfq_cfqq_wait_request(cfqq))
Divyesh Shah812df482010-04-08 21:15:35 -07002658 cfq_del_timer(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002659
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002660 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05002661 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002662
2663 /*
Shaohua Liae54abe2010-02-05 13:11:45 +01002664 * If this cfqq is shared between multiple processes, check to
2665 * make sure that those processes are still issuing I/Os within
2666 * the mean seek distance. If not, it may be time to break the
2667 * queues apart again.
2668 */
2669 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2670 cfq_mark_cfqq_split_coop(cfqq);
2671
2672 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02002673 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002674 */
Shaohua Lic553f8e2011-01-14 08:41:03 +01002675 if (timed_out) {
2676 if (cfq_cfqq_slice_new(cfqq))
Vivek Goyalba5bd522011-01-19 08:25:02 -07002677 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
Shaohua Lic553f8e2011-01-14 08:41:03 +01002678 else
2679 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02002680 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2681 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002682
Vivek Goyale5ff0822010-04-26 19:25:11 +02002683 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
Vivek Goyaldae739e2009-12-03 12:59:45 -05002684
Vivek Goyalf04a6422009-12-03 12:59:40 -05002685 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2686 cfq_del_cfqq_rr(cfqd, cfqq);
2687
Jens Axboeedd75ff2007-04-19 12:03:34 +02002688 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002689
2690 if (cfqq == cfqd->active_queue)
2691 cfqd->active_queue = NULL;
2692
2693 if (cfqd->active_cic) {
Tejun Heo11a31222012-02-07 07:51:30 +01002694 put_io_context(cfqd->active_cic->icq.ioc);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002695 cfqd->active_cic = NULL;
2696 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002697}
2698
Vivek Goyale5ff0822010-04-26 19:25:11 +02002699static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002700{
2701 struct cfq_queue *cfqq = cfqd->active_queue;
2702
2703 if (cfqq)
Vivek Goyale5ff0822010-04-26 19:25:11 +02002704 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002705}
2706
Jens Axboe498d3aa22007-04-26 12:54:48 +02002707/*
2708 * Get next queue for service. Unless we have a queue preemption,
2709 * we'll simply select the first cfqq in the service tree.
2710 */
Jens Axboe6d048f52007-04-25 12:44:27 +02002711static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002712{
Vivek Goyal34b98d02012-10-03 16:56:58 -04002713 struct cfq_rb_root *st = st_for(cfqd->serving_group,
2714 cfqd->serving_wl_class, cfqd->serving_wl_type);
Jens Axboeedd75ff2007-04-19 12:03:34 +02002715
Vivek Goyalf04a6422009-12-03 12:59:40 -05002716 if (!cfqd->rq_queued)
2717 return NULL;
2718
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002719 /* There is nothing to dispatch */
Vivek Goyal34b98d02012-10-03 16:56:58 -04002720 if (!st)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002721 return NULL;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002722 if (RB_EMPTY_ROOT(&st->rb))
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002723 return NULL;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002724 return cfq_rb_first(st);
Jens Axboe6d048f52007-04-25 12:44:27 +02002725}
2726
Vivek Goyalf04a6422009-12-03 12:59:40 -05002727static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2728{
Vivek Goyal25fb5162009-12-03 12:59:46 -05002729 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05002730 struct cfq_queue *cfqq;
2731 int i, j;
2732 struct cfq_rb_root *st;
2733
2734 if (!cfqd->rq_queued)
2735 return NULL;
2736
Vivek Goyal25fb5162009-12-03 12:59:46 -05002737 cfqg = cfq_get_next_cfqg(cfqd);
2738 if (!cfqg)
2739 return NULL;
2740
Vivek Goyalf04a6422009-12-03 12:59:40 -05002741 for_each_cfqg_st(cfqg, i, j, st)
2742 if ((cfqq = cfq_rb_first(st)) != NULL)
2743 return cfqq;
2744 return NULL;
2745}
2746
Jens Axboe498d3aa22007-04-26 12:54:48 +02002747/*
2748 * Get and set a new active queue for service.
2749 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002750static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2751 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002752{
Jens Axboee00ef792009-11-04 08:54:55 +01002753 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02002754 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02002755
Jens Axboe22e2c502005-06-27 10:55:12 +02002756 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02002757 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002758}
2759
Jens Axboed9e76202007-04-20 14:27:50 +02002760static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2761 struct request *rq)
2762{
Tejun Heo83096eb2009-05-07 22:24:39 +09002763 if (blk_rq_pos(rq) >= cfqd->last_position)
2764 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02002765 else
Tejun Heo83096eb2009-05-07 22:24:39 +09002766 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02002767}
2768
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002769static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Shaohua Lie9ce3352010-03-19 08:03:04 +01002770 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002771{
Shaohua Lie9ce3352010-03-19 08:03:04 +01002772 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02002773}
2774
Jens Axboea36e71f2009-04-15 12:15:11 +02002775static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2776 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02002777{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002778 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02002779 struct rb_node *parent, *node;
2780 struct cfq_queue *__cfqq;
2781 sector_t sector = cfqd->last_position;
2782
2783 if (RB_EMPTY_ROOT(root))
2784 return NULL;
2785
2786 /*
2787 * First, if we find a request starting at the end of the last
2788 * request, choose it.
2789 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02002790 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02002791 if (__cfqq)
2792 return __cfqq;
2793
2794 /*
2795 * If the exact sector wasn't found, the parent of the NULL leaf
2796 * will contain the closest sector.
2797 */
2798 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01002799 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002800 return __cfqq;
2801
Tejun Heo2e46e8b2009-05-07 22:24:41 +09002802 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02002803 node = rb_next(&__cfqq->p_node);
2804 else
2805 node = rb_prev(&__cfqq->p_node);
2806 if (!node)
2807 return NULL;
2808
2809 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Shaohua Lie9ce3352010-03-19 08:03:04 +01002810 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02002811 return __cfqq;
2812
2813 return NULL;
2814}
2815
2816/*
2817 * cfqd - obvious
2818 * cur_cfqq - passed in so that we don't decide that the current queue is
2819 * closely cooperating with itself.
2820 *
2821 * So, basically we're assuming that that cur_cfqq has dispatched at least
2822 * one request, and that cfqd->last_position reflects a position on the disk
2823 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
2824 * assumption.
2825 */
2826static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002827 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02002828{
2829 struct cfq_queue *cfqq;
2830
Divyesh Shah39c01b22010-03-25 15:45:57 +01002831 if (cfq_class_idle(cur_cfqq))
2832 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002833 if (!cfq_cfqq_sync(cur_cfqq))
2834 return NULL;
2835 if (CFQQ_SEEKY(cur_cfqq))
2836 return NULL;
2837
Jens Axboea36e71f2009-04-15 12:15:11 +02002838 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01002839 * Don't search priority tree if it's the only queue in the group.
2840 */
2841 if (cur_cfqq->cfqg->nr_cfqq == 1)
2842 return NULL;
2843
2844 /*
Jens Axboed9e76202007-04-20 14:27:50 +02002845 * We should notice if some of the queues are cooperating, eg
2846 * working closely on the same area of the disk. In that case,
2847 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02002848 */
Jens Axboea36e71f2009-04-15 12:15:11 +02002849 cfqq = cfqq_close(cfqd, cur_cfqq);
2850 if (!cfqq)
2851 return NULL;
2852
Vivek Goyal8682e1f2009-12-03 12:59:50 -05002853 /* If new queue belongs to different cfq_group, don't choose it */
2854 if (cur_cfqq->cfqg != cfqq->cfqg)
2855 return NULL;
2856
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002857 /*
2858 * It only makes sense to merge sync queues.
2859 */
2860 if (!cfq_cfqq_sync(cfqq))
2861 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04002862 if (CFQQ_SEEKY(cfqq))
2863 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002864
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01002865 /*
2866 * Do not merge queues of different priority classes
2867 */
2868 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2869 return NULL;
2870
Jens Axboea36e71f2009-04-15 12:15:11 +02002871 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02002872}
2873
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002874/*
2875 * Determine whether we should enforce idle window for this queue.
2876 */
2877
2878static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2879{
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04002880 enum wl_class_t wl_class = cfqq_class(cfqq);
Vivek Goyal34b98d02012-10-03 16:56:58 -04002881 struct cfq_rb_root *st = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002882
Vivek Goyal34b98d02012-10-03 16:56:58 -04002883 BUG_ON(!st);
2884 BUG_ON(!st->count);
Vivek Goyalf04a6422009-12-03 12:59:40 -05002885
Vivek Goyalb6508c12010-08-23 12:23:33 +02002886 if (!cfqd->cfq_slice_idle)
2887 return false;
2888
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002889 /* We never do for idle class queues. */
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04002890 if (wl_class == IDLE_WORKLOAD)
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002891 return false;
2892
2893 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01002894 if (cfq_cfqq_idle_window(cfqq) &&
2895 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002896 return true;
2897
2898 /*
2899 * Otherwise, we do only if they are the last ones
2900 * in their service tree.
2901 */
Vivek Goyal34b98d02012-10-03 16:56:58 -04002902 if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2903 !cfq_io_thinktime_big(cfqd, &st->ttime, false))
Shaohua Lic1e44752010-11-08 15:01:02 +01002904 return true;
Vivek Goyal34b98d02012-10-03 16:56:58 -04002905 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
Shaohua Lic1e44752010-11-08 15:01:02 +01002906 return false;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002907}
2908
Jens Axboe6d048f52007-04-25 12:44:27 +02002909static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002910{
Jens Axboe17926692007-01-19 11:59:30 +11002911 struct cfq_queue *cfqq = cfqd->active_queue;
Jan Karae7954212016-01-12 16:24:15 +01002912 struct cfq_rb_root *st = cfqq->service_tree;
Tejun Heoc5869802011-12-14 00:33:41 +01002913 struct cfq_io_cq *cic;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002914 unsigned long sl, group_idle = 0;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002915
Jens Axboea68bbdd2008-09-24 13:03:33 +02002916 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002917 * SSD device without seek penalty, disable idling. But only do so
2918 * for devices that support queuing, otherwise we still have a problem
2919 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02002920 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02002921 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02002922 return;
2923
Jens Axboedd67d052006-06-21 09:36:18 +02002924 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02002925 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02002926
2927 /*
2928 * idle is disabled, either manually or by past process history
2929 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002930 if (!cfq_should_idle(cfqd, cfqq)) {
2931 /* no queue idling. Check for group idling */
2932 if (cfqd->cfq_group_idle)
2933 group_idle = cfqd->cfq_group_idle;
2934 else
2935 return;
2936 }
Jens Axboe6d048f52007-04-25 12:44:27 +02002937
Jens Axboe22e2c502005-06-27 10:55:12 +02002938 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002939 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02002940 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002941 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02002942 return;
2943
2944 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02002945 * task has exited, don't wait
2946 */
Jens Axboe206dc692006-03-28 13:03:44 +02002947 cic = cfqd->active_cic;
Tejun Heof6e8d012012-03-05 13:15:26 -08002948 if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
Jens Axboe6d048f52007-04-25 12:44:27 +02002949 return;
2950
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002951 /*
2952 * If our average think time is larger than the remaining time
2953 * slice, then don't idle. This avoids overrunning the allotted
2954 * time slice.
2955 */
Shaohua Li383cd722011-07-12 14:24:35 +02002956 if (sample_valid(cic->ttime.ttime_samples) &&
2957 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
Joe Perchesfd16d262011-06-13 10:42:49 +02002958 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
Shaohua Li383cd722011-07-12 14:24:35 +02002959 cic->ttime.ttime_mean);
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002960 return;
Divyesh Shahb1ffe732010-03-25 15:45:03 +01002961 }
Corrado Zoccolo355b6592009-10-08 08:43:32 +02002962
Jan Karae7954212016-01-12 16:24:15 +01002963 /*
2964 * There are other queues in the group or this is the only group and
2965 * it has too big thinktime, don't do group idle.
2966 */
2967 if (group_idle &&
2968 (cfqq->cfqg->nr_cfqq > 1 ||
2969 cfq_io_thinktime_big(cfqd, &st->ttime, true)))
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002970 return;
2971
Jens Axboe3b181522005-06-27 10:56:24 +02002972 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002973
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002974 if (group_idle)
2975 sl = cfqd->cfq_group_idle;
2976 else
2977 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02002978
Jens Axboe7b14e3b2006-02-28 09:35:11 +01002979 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Tejun Heo155fead2012-04-01 14:38:44 -07002980 cfqg_stats_set_start_idle_time(cfqq->cfqg);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002981 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2982 group_idle ? 1 : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983}
2984
Jens Axboe498d3aa22007-04-26 12:54:48 +02002985/*
2986 * Move request from internal lists to the request queue dispatch list.
2987 */
Jens Axboe165125e2007-07-24 09:28:11 +02002988static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989{
Jens Axboe3ed9a292007-04-23 08:33:33 +02002990 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02002991 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002992
Jens Axboe7b679132008-05-30 12:23:07 +02002993 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2994
Jeff Moyer06d21882009-09-11 17:08:59 +02002995 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02002996 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02002997 cfqq->dispatched++;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02002998 (RQ_CFQG(rq))->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02002999 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02003000
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003001 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
Vivek Goyalc4e78932010-08-23 12:25:03 +02003002 cfqq->nr_sectors += blk_rq_sectors(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003}
3004
3005/*
3006 * return expired entry, or NULL to just start from scratch in rbtree
3007 */
Jens Axboefebffd62008-01-28 13:19:43 +01003008static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009{
Jens Axboe30996f42009-10-05 11:03:39 +02003010 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011
Jens Axboe3b181522005-06-27 10:56:24 +02003012 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11003014
3015 cfq_mark_cfqq_fifo_expire(cfqq);
3016
Jens Axboe89850f72006-07-22 16:48:31 +02003017 if (list_empty(&cfqq->fifo))
3018 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019
Jens Axboe89850f72006-07-22 16:48:31 +02003020 rq = rq_entry_fifo(cfqq->fifo.next);
Jan Kara8b4922d2014-02-24 16:39:52 +01003021 if (time_before(jiffies, rq->fifo_time))
Jens Axboe7b679132008-05-30 12:23:07 +02003022 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023
Jens Axboe30996f42009-10-05 11:03:39 +02003024 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02003025 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026}
3027
Jens Axboe22e2c502005-06-27 10:55:12 +02003028static inline int
3029cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3030{
3031 const int base_rq = cfqd->cfq_slice_async_rq;
3032
3033 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
3034
Namhyung Kimb9f8ce02011-05-24 10:23:21 +02003035 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02003036}
3037
3038/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003039 * Must be called with the queue_lock held.
3040 */
3041static int cfqq_process_refs(struct cfq_queue *cfqq)
3042{
3043 int process_refs, io_refs;
3044
3045 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
Shaohua Li30d7b942011-01-07 08:46:59 +01003046 process_refs = cfqq->ref - io_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003047 BUG_ON(process_refs < 0);
3048 return process_refs;
3049}
3050
3051static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3052{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003053 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003054 struct cfq_queue *__cfqq;
3055
Jeff Moyerc10b61f2010-06-17 10:19:11 -04003056 /*
3057 * If there are no process references on the new_cfqq, then it is
3058 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3059 * chain may have dropped their last reference (not just their
3060 * last process reference).
3061 */
3062 if (!cfqq_process_refs(new_cfqq))
3063 return;
3064
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003065 /* Avoid a circular list and skip interim queue merges */
3066 while ((__cfqq = new_cfqq->new_cfqq)) {
3067 if (__cfqq == cfqq)
3068 return;
3069 new_cfqq = __cfqq;
3070 }
3071
3072 process_refs = cfqq_process_refs(cfqq);
Jeff Moyerc10b61f2010-06-17 10:19:11 -04003073 new_process_refs = cfqq_process_refs(new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003074 /*
3075 * If the process for the cfqq has gone away, there is no
3076 * sense in merging the queues.
3077 */
Jeff Moyerc10b61f2010-06-17 10:19:11 -04003078 if (process_refs == 0 || new_process_refs == 0)
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003079 return;
3080
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003081 /*
3082 * Merge in the direction of the lesser amount of work.
3083 */
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003084 if (new_process_refs >= process_refs) {
3085 cfqq->new_cfqq = new_cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01003086 new_cfqq->ref += process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003087 } else {
3088 new_cfqq->new_cfqq = cfqq;
Shaohua Li30d7b942011-01-07 08:46:59 +01003089 cfqq->ref += new_process_refs;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003090 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003091}
3092
Vivek Goyal6d816ec2012-10-03 16:56:59 -04003093static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04003094 struct cfq_group *cfqg, enum wl_class_t wl_class)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003095{
3096 struct cfq_queue *queue;
3097 int i;
3098 bool key_valid = false;
3099 unsigned long lowest_key = 0;
3100 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3101
Vivek Goyal65b32a52009-12-16 17:52:59 -05003102 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3103 /* select the one with lowest rb_key */
Vivek Goyal34b98d02012-10-03 16:56:58 -04003104 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003105 if (queue &&
3106 (!key_valid || time_before(queue->rb_key, lowest_key))) {
3107 lowest_key = queue->rb_key;
3108 cur_best = i;
3109 key_valid = true;
3110 }
3111 }
3112
3113 return cur_best;
3114}
3115
Vivek Goyal6d816ec2012-10-03 16:56:59 -04003116static void
3117choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003118{
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003119 unsigned slice;
3120 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003121 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05003122 unsigned group_slice;
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003123 enum wl_class_t original_class = cfqd->serving_wl_class;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003124
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003125 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05003126 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003127 cfqd->serving_wl_class = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05003128 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003129 cfqd->serving_wl_class = BE_WORKLOAD;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003130 else {
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003131 cfqd->serving_wl_class = IDLE_WORKLOAD;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003132 cfqd->workload_expires = jiffies + 1;
3133 return;
3134 }
3135
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003136 if (original_class != cfqd->serving_wl_class)
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01003137 goto new_workload;
3138
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003139 /*
3140 * For RT and BE, we have to choose also the type
3141 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3142 * expiration time
3143 */
Vivek Goyal34b98d02012-10-03 16:56:58 -04003144 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003145 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003146
3147 /*
Vivek Goyal65b32a52009-12-16 17:52:59 -05003148 * check workload expiration, and that we still have other queues ready
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003149 */
Vivek Goyal65b32a52009-12-16 17:52:59 -05003150 if (count && !time_after(jiffies, cfqd->workload_expires))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003151 return;
3152
Shaohua Li writese4ea0c12010-12-13 14:32:22 +01003153new_workload:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003154 /* otherwise select new workload type */
Vivek Goyal6d816ec2012-10-03 16:56:59 -04003155 cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003156 cfqd->serving_wl_class);
Vivek Goyal34b98d02012-10-03 16:56:58 -04003157 st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003158 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003159
3160 /*
3161 * the workload slice is computed as a fraction of target latency
3162 * proportional to the number of queues in that workload, over
3163 * all the queues in the same priority class
3164 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05003165 group_slice = cfq_group_slice(cfqd, cfqg);
3166
3167 slice = group_slice * count /
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003168 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3169 cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
Vivek Goyal3bf10fe2012-10-03 16:56:56 -04003170 cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003171
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003172 if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05003173 unsigned int tmp;
3174
3175 /*
3176 * Async queues are currently system wide. Just taking
3177 * proportion of queues with-in same group will lead to higher
3178 * async ratio system wide as generally root group is going
3179 * to have higher weight. A more accurate thing would be to
3180 * calculate system wide asnc/sync ratio.
3181 */
Tao Ma5bf14c02012-04-01 14:33:39 -07003182 tmp = cfqd->cfq_target_latency *
3183 cfqg_busy_async_queues(cfqd, cfqg);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05003184 tmp = tmp/cfqd->busy_queues;
3185 slice = min_t(unsigned, slice, tmp);
3186
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003187 /* async workload slice is scaled down according to
3188 * the sync/async slice ratio. */
3189 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05003190 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003191 /* sync workload slice is at least 2 * cfq_slice_idle */
3192 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3193
3194 slice = max_t(unsigned, slice, CFQ_MIN_TT);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01003195 cfq_log(cfqd, "workload slice:%d", slice);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003196 cfqd->workload_expires = jiffies + slice;
3197}
3198
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003199static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3200{
3201 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003202 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003203
3204 if (RB_EMPTY_ROOT(&st->rb))
3205 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003206 cfqg = cfq_rb_first_group(st);
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003207 update_min_vdisktime(st);
3208 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003209}
3210
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003211static void cfq_choose_cfqg(struct cfq_data *cfqd)
3212{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003213 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3214
3215 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05003216
3217 /* Restore the workload type data */
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003218 if (cfqg->saved_wl_slice) {
3219 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3220 cfqd->serving_wl_type = cfqg->saved_wl_type;
3221 cfqd->serving_wl_class = cfqg->saved_wl_class;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01003222 } else
3223 cfqd->workload_expires = jiffies - 1;
3224
Vivek Goyal6d816ec2012-10-03 16:56:59 -04003225 choose_wl_class_and_type(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003226}
3227
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003228/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02003229 * Select a queue for service. If we have a current active queue,
3230 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02003231 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003232static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02003233{
Jens Axboea36e71f2009-04-15 12:15:11 +02003234 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02003235
3236 cfqq = cfqd->active_queue;
3237 if (!cfqq)
3238 goto new_queue;
3239
Vivek Goyalf04a6422009-12-03 12:59:40 -05003240 if (!cfqd->rq_queued)
3241 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05003242
3243 /*
3244 * We were waiting for group to get backlogged. Expire the queue
3245 */
3246 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3247 goto expire;
3248
Jens Axboe22e2c502005-06-27 10:55:12 +02003249 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02003250 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02003251 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003252 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3253 /*
3254 * If slice had not expired at the completion of last request
3255 * we might not have turned on wait_busy flag. Don't expire
3256 * the queue yet. Allow the group to get backlogged.
3257 *
3258 * The very fact that we have used the slice, that means we
3259 * have been idling all along on this queue and it should be
3260 * ok to wait for this request to complete.
3261 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01003262 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3263 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3264 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05003265 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01003266 } else
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003267 goto check_group_idle;
Vivek Goyal7667aa02009-12-08 17:52:58 -05003268 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003269
3270 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02003271 * The active queue has requests and isn't expired, allow it to
3272 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02003273 */
Jens Axboedd67d052006-06-21 09:36:18 +02003274 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003275 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02003276
3277 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02003278 * If another queue has a request waiting within our mean seek
3279 * distance, let it run. The expire code will check for close
3280 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003281 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02003282 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003283 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003284 if (new_cfqq) {
3285 if (!cfqq->new_cfqq)
3286 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02003287 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003288 }
Jens Axboea36e71f2009-04-15 12:15:11 +02003289
3290 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02003291 * No requests pending. If the active queue still has requests in
3292 * flight or is idling for a new request, allow either of these
3293 * conditions to happen (or time out) before selecting a new queue.
3294 */
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003295 if (timer_pending(&cfqd->idle_slice_timer)) {
3296 cfqq = NULL;
3297 goto keep_queue;
3298 }
3299
Shaohua Li8e1ac662010-11-08 15:01:04 +01003300 /*
3301 * This is a deep seek queue, but the device is much faster than
3302 * the queue can deliver, don't idle
3303 **/
3304 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3305 (cfq_cfqq_slice_new(cfqq) ||
3306 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3307 cfq_clear_cfqq_deep(cfqq);
3308 cfq_clear_cfqq_idle_window(cfqq);
3309 }
3310
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02003311 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3312 cfqq = NULL;
3313 goto keep_queue;
3314 }
3315
3316 /*
3317 * If group idle is enabled and there are requests dispatched from
3318 * this group, wait for requests to complete.
3319 */
3320check_group_idle:
Shaohua Li7700fc42011-07-12 14:24:56 +02003321 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3322 cfqq->cfqg->dispatched &&
3323 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02003324 cfqq = NULL;
3325 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003326 }
3327
Jens Axboe3b181522005-06-27 10:56:24 +02003328expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02003329 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02003330new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003331 /*
3332 * Current queue expired. Check if we have to switch to a new
3333 * service tree
3334 */
3335 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003336 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003337
Jens Axboea36e71f2009-04-15 12:15:11 +02003338 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003339keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02003340 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003341}
3342
Jens Axboefebffd62008-01-28 13:19:43 +01003343static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02003344{
3345 int dispatched = 0;
3346
3347 while (cfqq->next_rq) {
3348 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3349 dispatched++;
3350 }
3351
3352 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05003353
3354 /* By default cfqq is not expired if it is empty. Do it explicitly */
Vivek Goyale5ff0822010-04-26 19:25:11 +02003355 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02003356 return dispatched;
3357}
3358
Jens Axboe498d3aa22007-04-26 12:54:48 +02003359/*
3360 * Drain our current requests. Used for barriers and when switching
3361 * io schedulers on-the-fly.
3362 */
Jens Axboed9e76202007-04-20 14:27:50 +02003363static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003364{
Jens Axboe08717142008-01-28 11:38:15 +01003365 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02003366 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003367
Divyesh Shah3440c492010-04-09 09:29:57 +02003368 /* Expire the timeslice of the current active queue first */
Vivek Goyale5ff0822010-04-26 19:25:11 +02003369 cfq_slice_expired(cfqd, 0);
Divyesh Shah3440c492010-04-09 09:29:57 +02003370 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3371 __cfq_set_active_queue(cfqd, cfqq);
Vivek Goyalf04a6422009-12-03 12:59:40 -05003372 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Divyesh Shah3440c492010-04-09 09:29:57 +02003373 }
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003374
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003375 BUG_ON(cfqd->busy_queues);
3376
Jeff Moyer69237152009-06-12 15:29:30 +02003377 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01003378 return dispatched;
3379}
3380
Shaohua Liabc3c742010-03-01 09:20:54 +01003381static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3382 struct cfq_queue *cfqq)
3383{
3384 /* the queue hasn't finished any request, can't estimate */
3385 if (cfq_cfqq_slice_new(cfqq))
Shaohua Lic1e44752010-11-08 15:01:02 +01003386 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01003387 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3388 cfqq->slice_end))
Shaohua Lic1e44752010-11-08 15:01:02 +01003389 return true;
Shaohua Liabc3c742010-03-01 09:20:54 +01003390
Shaohua Lic1e44752010-11-08 15:01:02 +01003391 return false;
Shaohua Liabc3c742010-03-01 09:20:54 +01003392}
3393
Jens Axboe0b182d62009-10-06 20:49:37 +02003394static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02003395{
Jens Axboe2f5cb732009-04-07 08:51:19 +02003396 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397
Jens Axboe2f5cb732009-04-07 08:51:19 +02003398 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02003399 * Drain async requests before we start sync IO
3400 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003401 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02003402 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02003403
3404 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02003405 * If this is an async queue and we have sync IO in flight, let it wait
3406 */
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01003407 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02003408 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02003409
Shaohua Liabc3c742010-03-01 09:20:54 +01003410 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
Jens Axboe2f5cb732009-04-07 08:51:19 +02003411 if (cfq_class_idle(cfqq))
3412 max_dispatch = 1;
3413
3414 /*
3415 * Does this cfqq already have too much IO in flight?
3416 */
3417 if (cfqq->dispatched >= max_dispatch) {
Shaohua Lief8a41d2011-03-07 09:26:29 +01003418 bool promote_sync = false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02003419 /*
3420 * idle queue must always only have a single IO in flight
3421 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02003422 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02003423 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02003424
Jens Axboe2f5cb732009-04-07 08:51:19 +02003425 /*
Li, Shaohuac4ade942011-03-23 08:30:34 +01003426 * If there is only one sync queue
3427 * we can ignore async queue here and give the sync
Shaohua Lief8a41d2011-03-07 09:26:29 +01003428 * queue no dispatch limit. The reason is a sync queue can
3429 * preempt async queue, limiting the sync queue doesn't make
3430 * sense. This is useful for aiostress test.
3431 */
Li, Shaohuac4ade942011-03-23 08:30:34 +01003432 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3433 promote_sync = true;
Shaohua Lief8a41d2011-03-07 09:26:29 +01003434
3435 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02003436 * We have other queues, don't allow more IO from this one
3437 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01003438 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3439 !promote_sync)
Jens Axboe0b182d62009-10-06 20:49:37 +02003440 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11003441
Jens Axboe2f5cb732009-04-07 08:51:19 +02003442 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01003443 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02003444 */
Shaohua Lief8a41d2011-03-07 09:26:29 +01003445 if (cfqd->busy_queues == 1 || promote_sync)
Shaohua Liabc3c742010-03-01 09:20:54 +01003446 max_dispatch = -1;
3447 else
3448 /*
3449 * Normally we start throttling cfqq when cfq_quantum/2
3450 * requests have been dispatched. But we can drive
3451 * deeper queue depths at the beginning of slice
3452 * subjected to upper limit of cfq_quantum.
3453 * */
3454 max_dispatch = cfqd->cfq_quantum;
Jens Axboe8e296752009-10-03 16:26:03 +02003455 }
3456
3457 /*
3458 * Async queues must wait a bit before being allowed dispatch.
3459 * We also ramp up the dispatch depth gradually for async IO,
3460 * based on the last sync IO we serviced
3461 */
Jens Axboe963b72f2009-10-03 19:42:18 +02003462 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003463 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02003464 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02003465
Jens Axboe61f0c1d2009-10-03 19:46:03 +02003466 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02003467 if (!depth && !cfqq->dispatched)
3468 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02003469 if (depth < max_dispatch)
3470 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471 }
3472
Jens Axboe0b182d62009-10-06 20:49:37 +02003473 /*
3474 * If we're below the current max, allow a dispatch
3475 */
3476 return cfqq->dispatched < max_dispatch;
3477}
3478
3479/*
3480 * Dispatch a request from cfqq, moving them to the request queue
3481 * dispatch list.
3482 */
3483static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3484{
3485 struct request *rq;
3486
3487 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3488
3489 if (!cfq_may_dispatch(cfqd, cfqq))
3490 return false;
3491
3492 /*
3493 * follow expired path, else get first next available
3494 */
3495 rq = cfq_check_fifo(cfqq);
3496 if (!rq)
3497 rq = cfqq->next_rq;
3498
3499 /*
3500 * insert request into driver dispatch list
3501 */
3502 cfq_dispatch_insert(cfqd->queue, rq);
3503
3504 if (!cfqd->active_cic) {
Tejun Heoc5869802011-12-14 00:33:41 +01003505 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe0b182d62009-10-06 20:49:37 +02003506
Tejun Heoc5869802011-12-14 00:33:41 +01003507 atomic_long_inc(&cic->icq.ioc->refcount);
Jens Axboe0b182d62009-10-06 20:49:37 +02003508 cfqd->active_cic = cic;
3509 }
3510
3511 return true;
3512}
3513
3514/*
3515 * Find the cfqq that we need to service and move a request from that to the
3516 * dispatch list
3517 */
3518static int cfq_dispatch_requests(struct request_queue *q, int force)
3519{
3520 struct cfq_data *cfqd = q->elevator->elevator_data;
3521 struct cfq_queue *cfqq;
3522
3523 if (!cfqd->busy_queues)
3524 return 0;
3525
3526 if (unlikely(force))
3527 return cfq_forced_dispatch(cfqd);
3528
3529 cfqq = cfq_select_queue(cfqd);
3530 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02003531 return 0;
3532
Jens Axboe2f5cb732009-04-07 08:51:19 +02003533 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02003534 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02003535 */
Jens Axboe0b182d62009-10-06 20:49:37 +02003536 if (!cfq_dispatch_request(cfqd, cfqq))
3537 return 0;
3538
Jens Axboe2f5cb732009-04-07 08:51:19 +02003539 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02003540 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02003541
3542 /*
3543 * expire an async queue immediately if it has used up its slice. idle
3544 * queue always expire after 1 dispatch round.
3545 */
3546 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3547 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3548 cfq_class_idle(cfqq))) {
3549 cfqq->slice_end = jiffies + 1;
Vivek Goyale5ff0822010-04-26 19:25:11 +02003550 cfq_slice_expired(cfqd, 0);
Jens Axboe2f5cb732009-04-07 08:51:19 +02003551 }
3552
Shan Weib217a902009-09-01 10:06:42 +02003553 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02003554 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555}
3556
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557/*
Jens Axboe5e705372006-07-13 12:39:25 +02003558 * task holds one reference to the queue, dropped when task exits. each rq
3559 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05003561 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562 * queue lock must be held here.
3563 */
3564static void cfq_put_queue(struct cfq_queue *cfqq)
3565{
Jens Axboe22e2c502005-06-27 10:55:12 +02003566 struct cfq_data *cfqd = cfqq->cfqd;
Justin TerAvest0bbfeb82011-03-01 15:05:08 -05003567 struct cfq_group *cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02003568
Shaohua Li30d7b942011-01-07 08:46:59 +01003569 BUG_ON(cfqq->ref <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570
Shaohua Li30d7b942011-01-07 08:46:59 +01003571 cfqq->ref--;
3572 if (cfqq->ref)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 return;
3574
Jens Axboe7b679132008-05-30 12:23:07 +02003575 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02003577 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003578 cfqg = cfqq->cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579
Jens Axboe28f95cbc2007-01-19 12:09:53 +11003580 if (unlikely(cfqd->active_queue == cfqq)) {
Vivek Goyale5ff0822010-04-26 19:25:11 +02003581 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02003582 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11003583 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003584
Vivek Goyalf04a6422009-12-03 12:59:40 -05003585 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 kmem_cache_free(cfq_pool, cfqq);
Tejun Heoeb7d8c072012-03-23 14:02:53 +01003587 cfqg_put(cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588}
3589
Shaohua Lid02a2c02010-05-25 10:16:53 +02003590static void cfq_put_cooperator(struct cfq_queue *cfqq)
Jens Axboe89850f72006-07-22 16:48:31 +02003591{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003592 struct cfq_queue *__cfqq, *next;
3593
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003594 /*
3595 * If this queue was scheduled to merge with another queue, be
3596 * sure to drop the reference taken on that queue (and others in
3597 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
3598 */
3599 __cfqq = cfqq->new_cfqq;
3600 while (__cfqq) {
3601 if (__cfqq == cfqq) {
3602 WARN(1, "cfqq->new_cfqq loop detected\n");
3603 break;
3604 }
3605 next = __cfqq->new_cfqq;
3606 cfq_put_queue(__cfqq);
3607 __cfqq = next;
3608 }
Shaohua Lid02a2c02010-05-25 10:16:53 +02003609}
3610
3611static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3612{
3613 if (unlikely(cfqq == cfqd->active_queue)) {
3614 __cfq_slice_expired(cfqd, cfqq, 0);
3615 cfq_schedule_dispatch(cfqd);
3616 }
3617
3618 cfq_put_cooperator(cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003619
Jens Axboe89850f72006-07-22 16:48:31 +02003620 cfq_put_queue(cfqq);
3621}
3622
Tejun Heo9b84cac2011-12-14 00:33:42 +01003623static void cfq_init_icq(struct io_cq *icq)
3624{
3625 struct cfq_io_cq *cic = icq_to_cic(icq);
3626
3627 cic->ttime.last_end_request = jiffies;
3628}
3629
Tejun Heoc5869802011-12-14 00:33:41 +01003630static void cfq_exit_icq(struct io_cq *icq)
Jens Axboe89850f72006-07-22 16:48:31 +02003631{
Tejun Heoc5869802011-12-14 00:33:41 +01003632 struct cfq_io_cq *cic = icq_to_cic(icq);
Tejun Heo283287a2011-12-14 00:33:38 +01003633 struct cfq_data *cfqd = cic_to_cfqd(cic);
Fabio Checconi4faa3c82008-04-10 08:28:01 +02003634
Tejun Heo563180a2015-08-18 14:55:00 -07003635 if (cic_to_cfqq(cic, false)) {
3636 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3637 cic_set_cfqq(cic, NULL, false);
Jens Axboe89850f72006-07-22 16:48:31 +02003638 }
3639
Tejun Heo563180a2015-08-18 14:55:00 -07003640 if (cic_to_cfqq(cic, true)) {
3641 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3642 cic_set_cfqq(cic, NULL, true);
Jens Axboe89850f72006-07-22 16:48:31 +02003643 }
Jens Axboe89850f72006-07-22 16:48:31 +02003644}
3645
Tejun Heoabede6d2012-03-19 15:10:57 -07003646static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02003647{
3648 struct task_struct *tsk = current;
3649 int ioprio_class;
3650
Jens Axboe3b181522005-06-27 10:56:24 +02003651 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003652 return;
3653
Tejun Heo598971b2012-03-19 15:10:58 -07003654 ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02003655 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01003656 default:
3657 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3658 case IOPRIO_CLASS_NONE:
3659 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02003660 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01003661 */
3662 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02003663 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01003664 break;
3665 case IOPRIO_CLASS_RT:
Tejun Heo598971b2012-03-19 15:10:58 -07003666 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
Jens Axboefe094d92008-01-31 13:08:54 +01003667 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3668 break;
3669 case IOPRIO_CLASS_BE:
Tejun Heo598971b2012-03-19 15:10:58 -07003670 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
Jens Axboefe094d92008-01-31 13:08:54 +01003671 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3672 break;
3673 case IOPRIO_CLASS_IDLE:
3674 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3675 cfqq->ioprio = 7;
3676 cfq_clear_cfqq_idle_window(cfqq);
3677 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02003678 }
3679
3680 /*
3681 * keep track of original prio settings in case we have to temporarily
3682 * elevate the priority of this queue
3683 */
3684 cfqq->org_ioprio = cfqq->ioprio;
Jens Axboe3b181522005-06-27 10:56:24 +02003685 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003686}
3687
Tejun Heo598971b2012-03-19 15:10:58 -07003688static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
Jens Axboe22e2c502005-06-27 10:55:12 +02003689{
Tejun Heo598971b2012-03-19 15:10:58 -07003690 int ioprio = cic->icq.ioc->ioprio;
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003691 struct cfq_data *cfqd = cic_to_cfqd(cic);
Al Viro478a82b2006-03-18 13:25:24 -05003692 struct cfq_queue *cfqq;
Jens Axboe35e60772006-06-14 09:10:45 +02003693
Tejun Heo598971b2012-03-19 15:10:58 -07003694 /*
3695 * Check whether ioprio has changed. The condition may trigger
3696 * spuriously on a newly created cic but there's no harm.
3697 */
3698 if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
Jens Axboecaaa5f92006-06-16 11:23:00 +02003699 return;
3700
Tejun Heo563180a2015-08-18 14:55:00 -07003701 cfqq = cic_to_cfqq(cic, false);
Jens Axboecaaa5f92006-06-16 11:23:00 +02003702 if (cfqq) {
Tejun Heo563180a2015-08-18 14:55:00 -07003703 cfq_put_queue(cfqq);
Tejun Heo2da8de02015-08-18 14:55:02 -07003704 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
Tejun Heo563180a2015-08-18 14:55:00 -07003705 cic_set_cfqq(cic, cfqq, false);
Jens Axboe22e2c502005-06-27 10:55:12 +02003706 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003707
Tejun Heo563180a2015-08-18 14:55:00 -07003708 cfqq = cic_to_cfqq(cic, true);
Jens Axboecaaa5f92006-06-16 11:23:00 +02003709 if (cfqq)
3710 cfq_mark_cfqq_prio_changed(cfqq);
Tejun Heo598971b2012-03-19 15:10:58 -07003711
3712 cic->ioprio = ioprio;
Jens Axboe22e2c502005-06-27 10:55:12 +02003713}
3714
Jens Axboed5036d72009-06-26 10:44:34 +02003715static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02003716 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02003717{
3718 RB_CLEAR_NODE(&cfqq->rb_node);
3719 RB_CLEAR_NODE(&cfqq->p_node);
3720 INIT_LIST_HEAD(&cfqq->fifo);
3721
Shaohua Li30d7b942011-01-07 08:46:59 +01003722 cfqq->ref = 0;
Jens Axboed5036d72009-06-26 10:44:34 +02003723 cfqq->cfqd = cfqd;
3724
3725 cfq_mark_cfqq_prio_changed(cfqq);
3726
3727 if (is_sync) {
3728 if (!cfq_class_idle(cfqq))
3729 cfq_mark_cfqq_idle_window(cfqq);
3730 cfq_mark_cfqq_sync(cfqq);
3731 }
3732 cfqq->pid = pid;
3733}
3734
Vivek Goyal246103332009-12-03 12:59:51 -05003735#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo598971b2012-03-19 15:10:58 -07003736static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
Vivek Goyal246103332009-12-03 12:59:51 -05003737{
Konstantin Khlebnikovbca4b912010-05-20 23:21:34 +04003738 struct cfq_data *cfqd = cic_to_cfqd(cic);
Tejun Heo60a83702015-08-18 14:55:05 -07003739 struct cfq_queue *cfqq;
Tejun Heof4da8072014-09-08 08:15:20 +09003740 uint64_t serial_nr;
Vivek Goyal246103332009-12-03 12:59:51 -05003741
Tejun Heo598971b2012-03-19 15:10:58 -07003742 rcu_read_lock();
Tejun Heof4da8072014-09-08 08:15:20 +09003743 serial_nr = bio_blkcg(bio)->css.serial_nr;
Tejun Heo598971b2012-03-19 15:10:58 -07003744 rcu_read_unlock();
3745
3746 /*
3747 * Check whether blkcg has changed. The condition may trigger
3748 * spuriously on a newly created cic but there's no harm.
3749 */
Tejun Heof4da8072014-09-08 08:15:20 +09003750 if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
Vivek Goyal246103332009-12-03 12:59:51 -05003751 return;
3752
Tejun Heo60a83702015-08-18 14:55:05 -07003753 /*
3754 * Drop reference to queues. New queues will be assigned in new
3755 * group upon arrival of fresh requests.
3756 */
3757 cfqq = cic_to_cfqq(cic, false);
3758 if (cfqq) {
3759 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3760 cic_set_cfqq(cic, NULL, false);
3761 cfq_put_queue(cfqq);
3762 }
3763
3764 cfqq = cic_to_cfqq(cic, true);
3765 if (cfqq) {
3766 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3767 cic_set_cfqq(cic, NULL, true);
3768 cfq_put_queue(cfqq);
Vivek Goyal246103332009-12-03 12:59:51 -05003769 }
Tejun Heo598971b2012-03-19 15:10:58 -07003770
Tejun Heof4da8072014-09-08 08:15:20 +09003771 cic->blkcg_serial_nr = serial_nr;
Vivek Goyal246103332009-12-03 12:59:51 -05003772}
Tejun Heo598971b2012-03-19 15:10:58 -07003773#else
3774static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
Vivek Goyal246103332009-12-03 12:59:51 -05003775#endif /* CONFIG_CFQ_GROUP_IOSCHED */
3776
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003777static struct cfq_queue **
Tejun Heo60a83702015-08-18 14:55:05 -07003778cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003779{
Jens Axboefe094d92008-01-31 13:08:54 +01003780 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003781 case IOPRIO_CLASS_RT:
Tejun Heo60a83702015-08-18 14:55:05 -07003782 return &cfqg->async_cfqq[0][ioprio];
Tejun Heo598971b2012-03-19 15:10:58 -07003783 case IOPRIO_CLASS_NONE:
3784 ioprio = IOPRIO_NORM;
3785 /* fall through */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003786 case IOPRIO_CLASS_BE:
Tejun Heo60a83702015-08-18 14:55:05 -07003787 return &cfqg->async_cfqq[1][ioprio];
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003788 case IOPRIO_CLASS_IDLE:
Tejun Heo60a83702015-08-18 14:55:05 -07003789 return &cfqg->async_idle_cfqq;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003790 default:
3791 BUG();
3792 }
3793}
3794
Jens Axboe15c31be2007-07-10 13:43:25 +02003795static struct cfq_queue *
Tejun Heoabede6d2012-03-19 15:10:57 -07003796cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
Tejun Heo2da8de02015-08-18 14:55:02 -07003797 struct bio *bio)
Jens Axboe15c31be2007-07-10 13:43:25 +02003798{
Jeff Moyerc6ce1942015-01-12 15:21:01 -05003799 int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3800 int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
Tejun Heod4aad7f2015-08-18 14:55:04 -07003801 struct cfq_queue **async_cfqq = NULL;
Tejun Heo4ebc1c62015-08-18 14:54:57 -07003802 struct cfq_queue *cfqq;
Tejun Heo322731e2015-08-18 14:55:03 -07003803 struct cfq_group *cfqg;
3804
3805 rcu_read_lock();
Tejun Heoae118892015-08-18 14:55:20 -07003806 cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
Tejun Heo322731e2015-08-18 14:55:03 -07003807 if (!cfqg) {
3808 cfqq = &cfqd->oom_cfqq;
3809 goto out;
3810 }
Jens Axboe15c31be2007-07-10 13:43:25 +02003811
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003812 if (!is_sync) {
Jeff Moyerc6ce1942015-01-12 15:21:01 -05003813 if (!ioprio_valid(cic->ioprio)) {
3814 struct task_struct *tsk = current;
3815 ioprio = task_nice_ioprio(tsk);
3816 ioprio_class = task_nice_ioclass(tsk);
3817 }
Tejun Heo60a83702015-08-18 14:55:05 -07003818 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003819 cfqq = *async_cfqq;
Tejun Heo4ebc1c62015-08-18 14:54:57 -07003820 if (cfqq)
3821 goto out;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003822 }
3823
Tejun Heod4aad7f2015-08-18 14:55:04 -07003824 cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3825 cfqd->queue->node);
3826 if (!cfqq) {
3827 cfqq = &cfqd->oom_cfqq;
3828 goto out;
3829 }
Jens Axboe15c31be2007-07-10 13:43:25 +02003830
Tejun Heod4aad7f2015-08-18 14:55:04 -07003831 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3832 cfq_init_prio_data(cfqq, cic);
3833 cfq_link_cfqq_cfqg(cfqq, cfqg);
3834 cfq_log_cfqq(cfqd, cfqq, "alloced");
3835
3836 if (async_cfqq) {
3837 /* a new async queue is created, pin and remember */
Shaohua Li30d7b942011-01-07 08:46:59 +01003838 cfqq->ref++;
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003839 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02003840 }
Tejun Heo4ebc1c62015-08-18 14:54:57 -07003841out:
Shaohua Li30d7b942011-01-07 08:46:59 +01003842 cfqq->ref++;
Tejun Heo322731e2015-08-18 14:55:03 -07003843 rcu_read_unlock();
Jens Axboe15c31be2007-07-10 13:43:25 +02003844 return cfqq;
3845}
3846
Jens Axboe22e2c502005-06-27 10:55:12 +02003847static void
Shaohua Li383cd722011-07-12 14:24:35 +02003848__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003849{
Shaohua Li383cd722011-07-12 14:24:35 +02003850 unsigned long elapsed = jiffies - ttime->last_end_request;
3851 elapsed = min(elapsed, 2UL * slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02003852
Shaohua Li383cd722011-07-12 14:24:35 +02003853 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3854 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3855 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3856}
3857
3858static void
3859cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01003860 struct cfq_io_cq *cic)
Shaohua Li383cd722011-07-12 14:24:35 +02003861{
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003862 if (cfq_cfqq_sync(cfqq)) {
Shaohua Li383cd722011-07-12 14:24:35 +02003863 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02003864 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3865 cfqd->cfq_slice_idle);
3866 }
Shaohua Li7700fc42011-07-12 14:24:56 +02003867#ifdef CONFIG_CFQ_GROUP_IOSCHED
3868 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3869#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02003870}
3871
Jens Axboe206dc692006-03-28 13:03:44 +02003872static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003873cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003874 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003875{
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003876 sector_t sdist = 0;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003877 sector_t n_sec = blk_rq_sectors(rq);
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003878 if (cfqq->last_request_pos) {
3879 if (cfqq->last_request_pos < blk_rq_pos(rq))
3880 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3881 else
3882 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3883 }
Jens Axboe206dc692006-03-28 13:03:44 +02003884
Corrado Zoccolo3dde36d2010-02-27 19:45:39 +01003885 cfqq->seek_history <<= 1;
Corrado Zoccolo41647e72010-02-27 19:45:40 +01003886 if (blk_queue_nonrot(cfqd->queue))
3887 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3888 else
3889 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
Jens Axboe206dc692006-03-28 13:03:44 +02003890}
Jens Axboe22e2c502005-06-27 10:55:12 +02003891
3892/*
3893 * Disable idle window if the process thinks too long or seeks so much that
3894 * it doesn't matter
3895 */
3896static void
3897cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Tejun Heoc5869802011-12-14 00:33:41 +01003898 struct cfq_io_cq *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02003899{
Jens Axboe7b679132008-05-30 12:23:07 +02003900 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003901
Jens Axboe08717142008-01-28 11:38:15 +01003902 /*
3903 * Don't idle for async or idle io prio class
3904 */
3905 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003906 return;
3907
Jens Axboec265a7f2008-06-26 13:49:33 +02003908 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003909
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003910 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3911 cfq_mark_cfqq_deep(cfqq);
3912
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02003913 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3914 enable_idle = 0;
Tejun Heof6e8d012012-03-05 13:15:26 -08003915 else if (!atomic_read(&cic->icq.ioc->active_ref) ||
Tejun Heoc5869802011-12-14 00:33:41 +01003916 !cfqd->cfq_slice_idle ||
3917 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003918 enable_idle = 0;
Shaohua Li383cd722011-07-12 14:24:35 +02003919 else if (sample_valid(cic->ttime.ttime_samples)) {
3920 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003921 enable_idle = 0;
3922 else
3923 enable_idle = 1;
3924 }
3925
Jens Axboe7b679132008-05-30 12:23:07 +02003926 if (old_idle != enable_idle) {
3927 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3928 if (enable_idle)
3929 cfq_mark_cfqq_idle_window(cfqq);
3930 else
3931 cfq_clear_cfqq_idle_window(cfqq);
3932 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003933}
3934
Jens Axboe22e2c502005-06-27 10:55:12 +02003935/*
3936 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3937 * no or if we aren't sure, a 1 will cause a preempt.
3938 */
Jens Axboea6151c32009-10-07 20:02:57 +02003939static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003940cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003941 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003942{
Jens Axboe6d048f52007-04-25 12:44:27 +02003943 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003944
Jens Axboe6d048f52007-04-25 12:44:27 +02003945 cfqq = cfqd->active_queue;
3946 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003947 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003948
Jens Axboe6d048f52007-04-25 12:44:27 +02003949 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003950 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003951
3952 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003953 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003954
Jens Axboe22e2c502005-06-27 10:55:12 +02003955 /*
Divyesh Shah875feb62010-01-06 18:58:20 -08003956 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3957 */
3958 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3959 return false;
3960
3961 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003962 * if the new request is sync, but the currently running queue is
3963 * not, let the sync request have priority.
3964 */
Jens Axboe5e705372006-07-13 12:39:25 +02003965 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003966 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003967
Jan Kara3984aa52016-01-12 16:24:19 +01003968 /*
3969 * Treat ancestors of current cgroup the same way as current cgroup.
3970 * For anybody else we disallow preemption to guarantee service
3971 * fairness among cgroups.
3972 */
3973 if (!cfqg_is_descendant(cfqq->cfqg, new_cfqq->cfqg))
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003974 return false;
3975
3976 if (cfq_slice_used(cfqq))
3977 return true;
3978
Jan Kara6c80731c2016-01-12 16:24:16 +01003979 /*
3980 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3981 */
3982 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3983 return true;
3984
3985 WARN_ON_ONCE(cfqq->ioprio_class != new_cfqq->ioprio_class);
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003986 /* Allow preemption only if we are idling on sync-noidle tree */
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04003987 if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003988 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003989 RB_EMPTY_ROOT(&cfqq->sort_list))
3990 return true;
3991
Jens Axboe374f84a2006-07-23 01:42:19 +02003992 /*
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003993 * So both queues are sync. Let the new request get disk time if
3994 * it's a metadata request and the current queue is doing regular IO.
3995 */
Christoph Hellwig65299a32011-08-23 14:50:29 +02003996 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
Jens Axboeb53d1ed2011-08-19 08:34:48 +02003997 return true;
3998
Shaohua Lid2d59e12010-11-08 15:01:03 +01003999 /* An idle queue should not be idle now for some reason */
4000 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
4001 return true;
4002
Jens Axboe1e3335d2007-02-14 19:59:49 +01004003 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02004004 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01004005
4006 /*
4007 * if this request is as-good as one we would expect from the
4008 * current cfqq, let it preempt
4009 */
Shaohua Lie9ce3352010-03-19 08:03:04 +01004010 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02004011 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01004012
Jens Axboea6151c32009-10-07 20:02:57 +02004013 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02004014}
4015
4016/*
4017 * cfqq preempts the active queue. if we allowed preempt with no slice left,
4018 * let it have half of its nominal slice.
4019 */
4020static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4021{
Shaohua Lidf0793a2012-01-19 09:20:09 +01004022 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
4023
Jens Axboe7b679132008-05-30 12:23:07 +02004024 cfq_log_cfqq(cfqd, cfqq, "preempt");
Shaohua Lidf0793a2012-01-19 09:20:09 +01004025 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004026
Jens Axboebf572252006-07-19 20:29:12 +02004027 /*
Shaohua Lif8ae6e32011-01-14 08:41:02 +01004028 * workload type is changed, don't save slice, otherwise preempt
4029 * doesn't happen
4030 */
Shaohua Lidf0793a2012-01-19 09:20:09 +01004031 if (old_type != cfqq_type(cfqq))
Vivek Goyal4d2ceea2012-10-03 16:56:57 -04004032 cfqq->cfqg->saved_wl_slice = 0;
Shaohua Lif8ae6e32011-01-14 08:41:02 +01004033
4034 /*
Jens Axboebf572252006-07-19 20:29:12 +02004035 * Put the new queue at the front of the of the current list,
4036 * so we know that it will be selected next.
4037 */
4038 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02004039
4040 cfq_service_tree_add(cfqd, cfqq, 1);
Justin TerAvesteda5e0c2011-03-22 21:26:49 +01004041
Justin TerAvest62a37f62011-03-23 08:25:44 +01004042 cfqq->slice_end = 0;
4043 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004044}
4045
4046/*
Jens Axboe5e705372006-07-13 12:39:25 +02004047 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02004048 * something we should do about it
4049 */
4050static void
Jens Axboe5e705372006-07-13 12:39:25 +02004051cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
4052 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02004053{
Tejun Heoc5869802011-12-14 00:33:41 +01004054 struct cfq_io_cq *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02004055
Aaron Carroll45333d52008-08-26 15:52:36 +02004056 cfqd->rq_queued++;
Christoph Hellwig65299a32011-08-23 14:50:29 +02004057 if (rq->cmd_flags & REQ_PRIO)
4058 cfqq->prio_pending++;
Jens Axboe374f84a2006-07-23 01:42:19 +02004059
Shaohua Li383cd722011-07-12 14:24:35 +02004060 cfq_update_io_thinktime(cfqd, cfqq, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04004061 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02004062 cfq_update_idle_window(cfqd, cfqq, cic);
4063
Jeff Moyerb2c18e12009-10-23 17:14:49 -04004064 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004065
4066 if (cfqq == cfqd->active_queue) {
4067 /*
Jens Axboeb0291952009-04-07 11:38:31 +02004068 * Remember that we saw a request from this process, but
4069 * don't start queuing just yet. Otherwise we risk seeing lots
4070 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02004071 * and merging. If the request is already larger than a single
4072 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02004073 * merging is already done. Ditto for a busy system that
4074 * has other work pending, don't risk delaying until the
4075 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02004076 */
Jens Axboed6ceb252009-04-14 14:18:16 +02004077 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02004078 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
4079 cfqd->busy_queues > 1) {
Divyesh Shah812df482010-04-08 21:15:35 -07004080 cfq_del_timer(cfqd, cfqq);
Gui Jianfeng554554f2009-12-10 09:38:39 +01004081 cfq_clear_cfqq_wait_request(cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02004082 __blk_run_queue(cfqd->queue);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02004083 } else {
Tejun Heo155fead2012-04-01 14:38:44 -07004084 cfqg_stats_update_idle_time(cfqq->cfqg);
Vivek Goyalbf7919372009-12-03 12:59:37 -05004085 cfq_mark_cfqq_must_dispatch(cfqq);
Divyesh Shaha11cdaa2010-04-13 19:59:17 +02004086 }
Jens Axboed6ceb252009-04-14 14:18:16 +02004087 }
Jens Axboe5e705372006-07-13 12:39:25 +02004088 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02004089 /*
4090 * not the active queue - expire current slice if it is
4091 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01004092 * has some old slice time left and is of higher priority or
4093 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02004094 */
4095 cfq_preempt_queue(cfqd, cfqq);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02004096 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02004097 }
4098}
4099
Jens Axboe165125e2007-07-24 09:28:11 +02004100static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02004101{
Jens Axboeb4878f22005-10-20 16:42:29 +02004102 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02004103 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004104
Jens Axboe7b679132008-05-30 12:23:07 +02004105 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Tejun Heoabede6d2012-03-19 15:10:57 -07004106 cfq_init_prio_data(cfqq, RQ_CIC(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107
Jan Kara8b4922d2014-02-24 16:39:52 +01004108 rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
Jens Axboe22e2c502005-06-27 10:55:12 +02004109 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01004110 cfq_add_rq_rb(rq);
Tejun Heo155fead2012-04-01 14:38:44 -07004111 cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4112 rq->cmd_flags);
Jens Axboe5e705372006-07-13 12:39:25 +02004113 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114}
4115
Aaron Carroll45333d52008-08-26 15:52:36 +02004116/*
4117 * Update hw_tag based on peak queue depth over 50 samples under
4118 * sufficient load.
4119 */
4120static void cfq_update_hw_tag(struct cfq_data *cfqd)
4121{
Shaohua Li1a1238a2009-10-27 08:46:23 +01004122 struct cfq_queue *cfqq = cfqd->active_queue;
4123
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004124 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4125 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01004126
4127 if (cfqd->hw_tag == 1)
4128 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02004129
4130 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004131 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02004132 return;
4133
Shaohua Li1a1238a2009-10-27 08:46:23 +01004134 /*
4135 * If active queue hasn't enough requests and can idle, cfq might not
4136 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4137 * case
4138 */
4139 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4140 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004141 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
Shaohua Li1a1238a2009-10-27 08:46:23 +01004142 return;
4143
Aaron Carroll45333d52008-08-26 15:52:36 +02004144 if (cfqd->hw_tag_samples++ < 50)
4145 return;
4146
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01004147 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02004148 cfqd->hw_tag = 1;
4149 else
4150 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02004151}
4152
Vivek Goyal7667aa02009-12-08 17:52:58 -05004153static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4154{
Tejun Heoc5869802011-12-14 00:33:41 +01004155 struct cfq_io_cq *cic = cfqd->active_cic;
Vivek Goyal7667aa02009-12-08 17:52:58 -05004156
Justin TerAvest02a8f012011-02-09 14:20:03 +01004157 /* If the queue already has requests, don't wait */
4158 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4159 return false;
4160
Vivek Goyal7667aa02009-12-08 17:52:58 -05004161 /* If there are other queues in the group, don't wait */
4162 if (cfqq->cfqg->nr_cfqq > 1)
4163 return false;
4164
Shaohua Li7700fc42011-07-12 14:24:56 +02004165 /* the only queue in the group, but think time is big */
4166 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4167 return false;
4168
Vivek Goyal7667aa02009-12-08 17:52:58 -05004169 if (cfq_slice_used(cfqq))
4170 return true;
4171
4172 /* if slice left is less than think time, wait busy */
Shaohua Li383cd722011-07-12 14:24:35 +02004173 if (cic && sample_valid(cic->ttime.ttime_samples)
4174 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
Vivek Goyal7667aa02009-12-08 17:52:58 -05004175 return true;
4176
4177 /*
4178 * If think times is less than a jiffy than ttime_mean=0 and above
4179 * will not be true. It might happen that slice has not expired yet
4180 * but will expire soon (4-5 ns) during select_queue(). To cover the
4181 * case where think time is less than a jiffy, mark the queue wait
4182 * busy if only 1 jiffy is left in the slice.
4183 */
4184 if (cfqq->slice_end - jiffies == 1)
4185 return true;
4186
4187 return false;
4188}
4189
Jens Axboe165125e2007-07-24 09:28:11 +02004190static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004191{
Jens Axboe5e705372006-07-13 12:39:25 +02004192 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02004193 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02004194 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02004195 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196
Jens Axboeb4878f22005-10-20 16:42:29 +02004197 now = jiffies;
Christoph Hellwig33659eb2010-08-07 18:17:56 +02004198 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4199 !!(rq->cmd_flags & REQ_NOIDLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200
Aaron Carroll45333d52008-08-26 15:52:36 +02004201 cfq_update_hw_tag(cfqd);
4202
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004203 WARN_ON(!cfqd->rq_in_driver);
Jens Axboe6d048f52007-04-25 12:44:27 +02004204 WARN_ON(!cfqq->dispatched);
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004205 cfqd->rq_in_driver--;
Jens Axboe6d048f52007-04-25 12:44:27 +02004206 cfqq->dispatched--;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004207 (RQ_CFQG(rq))->dispatched--;
Tejun Heo155fead2012-04-01 14:38:44 -07004208 cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4209 rq_io_start_time_ns(rq), rq->cmd_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004211 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
Jens Axboe3ed9a292007-04-23 08:33:33 +02004212
Vivek Goyal365722b2009-10-03 15:21:27 +02004213 if (sync) {
Vivek Goyal34b98d02012-10-03 16:56:58 -04004214 struct cfq_rb_root *st;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02004215
Shaohua Li383cd722011-07-12 14:24:35 +02004216 RQ_CIC(rq)->ttime.last_end_request = now;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02004217
4218 if (cfq_cfqq_on_rr(cfqq))
Vivek Goyal34b98d02012-10-03 16:56:58 -04004219 st = cfqq->service_tree;
Shaohua Lif5f2b6c2011-07-12 14:24:55 +02004220 else
Vivek Goyal34b98d02012-10-03 16:56:58 -04004221 st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4222 cfqq_type(cfqq));
4223
4224 st->ttime.last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01004225 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4226 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02004227 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02004228
Shaohua Li7700fc42011-07-12 14:24:56 +02004229#ifdef CONFIG_CFQ_GROUP_IOSCHED
4230 cfqq->cfqg->ttime.last_end_request = now;
4231#endif
4232
Jens Axboecaaa5f92006-06-16 11:23:00 +02004233 /*
4234 * If this is the active queue, check if it needs to be expired,
4235 * or if we want to idle in case it has no pending requests.
4236 */
4237 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02004238 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4239
Jens Axboe44f7c162007-01-19 11:51:58 +11004240 if (cfq_cfqq_slice_new(cfqq)) {
4241 cfq_set_prio_slice(cfqd, cfqq);
4242 cfq_clear_cfqq_slice_new(cfqq);
4243 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05004244
4245 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05004246 * Should we wait for next request to come in before we expire
4247 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05004248 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05004249 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004250 unsigned long extend_sl = cfqd->cfq_slice_idle;
4251 if (!cfqd->cfq_slice_idle)
4252 extend_sl = cfqd->cfq_group_idle;
4253 cfqq->slice_end = jiffies + extend_sl;
Vivek Goyalf75edf22009-12-03 12:59:53 -05004254 cfq_mark_cfqq_wait_busy(cfqq);
Divyesh Shahb1ffe732010-03-25 15:45:03 +01004255 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
Vivek Goyalf75edf22009-12-03 12:59:53 -05004256 }
4257
Jens Axboea36e71f2009-04-15 12:15:11 +02004258 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01004259 * Idling is not enabled on:
4260 * - expired queues
4261 * - idle-priority queues
4262 * - async queues
4263 * - queues with still some requests queued
4264 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02004265 */
Jens Axboe08717142008-01-28 11:38:15 +01004266 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Vivek Goyale5ff0822010-04-26 19:25:11 +02004267 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01004268 else if (sync && cfqq_empty &&
4269 !cfq_close_cooperator(cfqd, cfqq)) {
Corrado Zoccolo749ef9f2010-09-20 15:24:50 +02004270 cfq_arm_slice_timer(cfqd);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01004271 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02004272 }
Jens Axboe6d048f52007-04-25 12:44:27 +02004273
Corrado Zoccolo53c583d2010-02-28 19:45:05 +01004274 if (!cfqd->rq_in_driver)
Jens Axboe23e018a2009-10-05 08:52:35 +02004275 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004276}
4277
Jens Axboe89850f72006-07-22 16:48:31 +02004278static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02004279{
Jens Axboe1b379d82009-08-11 08:26:11 +02004280 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02004281 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004282 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02004283 }
Jens Axboe22e2c502005-06-27 10:55:12 +02004284
4285 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02004286}
4287
Jens Axboe165125e2007-07-24 09:28:11 +02004288static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02004289{
4290 struct cfq_data *cfqd = q->elevator->elevator_data;
4291 struct task_struct *tsk = current;
Tejun Heoc5869802011-12-14 00:33:41 +01004292 struct cfq_io_cq *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02004293 struct cfq_queue *cfqq;
4294
4295 /*
4296 * don't force setup of a queue from here, as a call to may_queue
4297 * does not necessarily imply that a request actually will be queued.
4298 * so just lookup a possibly existing queue, or return 'may queue'
4299 * if that fails
4300 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01004301 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02004302 if (!cic)
4303 return ELV_MQUEUE_MAY;
4304
Jens Axboeb0b78f82009-04-08 10:56:08 +02004305 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02004306 if (cfqq) {
Tejun Heoabede6d2012-03-19 15:10:57 -07004307 cfq_init_prio_data(cfqq, cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02004308
Jens Axboe89850f72006-07-22 16:48:31 +02004309 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004310 }
4311
4312 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004313}
4314
Linus Torvalds1da177e2005-04-16 15:20:36 -07004315/*
4316 * queue lock held here
4317 */
Jens Axboebb37b942006-12-01 10:42:33 +01004318static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004319{
Jens Axboe5e705372006-07-13 12:39:25 +02004320 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004321
Jens Axboe5e705372006-07-13 12:39:25 +02004322 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02004323 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004324
Jens Axboe22e2c502005-06-27 10:55:12 +02004325 BUG_ON(!cfqq->allocated[rw]);
4326 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004327
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02004328 /* Put down rq reference on cfqg */
Tejun Heoeb7d8c072012-03-23 14:02:53 +01004329 cfqg_put(RQ_CFQG(rq));
Tejun Heoa612fdd2011-12-14 00:33:41 +01004330 rq->elv.priv[0] = NULL;
4331 rq->elv.priv[1] = NULL;
Vivek Goyal7f1dc8a2010-04-21 17:44:16 +02004332
Linus Torvalds1da177e2005-04-16 15:20:36 -07004333 cfq_put_queue(cfqq);
4334 }
4335}
4336
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004337static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01004338cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004339 struct cfq_queue *cfqq)
4340{
4341 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4342 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04004343 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004344 cfq_put_queue(cfqq);
4345 return cic_to_cfqq(cic, 1);
4346}
4347
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004348/*
4349 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4350 * was the last process referring to said cfqq.
4351 */
4352static struct cfq_queue *
Tejun Heoc5869802011-12-14 00:33:41 +01004353split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004354{
4355 if (cfqq_process_refs(cfqq) == 1) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004356 cfqq->pid = current->pid;
4357 cfq_clear_cfqq_coop(cfqq);
Shaohua Liae54abe2010-02-05 13:11:45 +01004358 cfq_clear_cfqq_split_coop(cfqq);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004359 return cfqq;
4360 }
4361
4362 cic_set_cfqq(cic, NULL, 1);
Shaohua Lid02a2c02010-05-25 10:16:53 +02004363
4364 cfq_put_cooperator(cfqq);
4365
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004366 cfq_put_queue(cfqq);
4367 return NULL;
4368}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004369/*
Jens Axboe22e2c502005-06-27 10:55:12 +02004370 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004371 */
Jens Axboe22e2c502005-06-27 10:55:12 +02004372static int
Tejun Heo852c7882012-03-05 13:15:27 -08004373cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4374 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004375{
4376 struct cfq_data *cfqd = q->elevator->elevator_data;
Tejun Heof1f8cc92011-12-14 00:33:42 +01004377 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004378 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02004379 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004380 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381
Tejun Heo216284c2011-12-14 00:33:38 +01004382 spin_lock_irq(q->queue_lock);
Tejun Heof1f8cc92011-12-14 00:33:42 +01004383
Tejun Heo598971b2012-03-19 15:10:58 -07004384 check_ioprio_changed(cic, bio);
4385 check_blkcg_changed(cic, bio);
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004386new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02004387 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02004388 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Tejun Heobce61332015-08-18 14:54:59 -07004389 if (cfqq)
4390 cfq_put_queue(cfqq);
Tejun Heo2da8de02015-08-18 14:55:02 -07004391 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
Vasily Tarasov91fac312007-04-25 12:29:51 +02004392 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004393 } else {
4394 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004395 * If the queue was seeky for too long, break it apart.
4396 */
Shaohua Liae54abe2010-02-05 13:11:45 +01004397 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
Jeff Moyere6c5bc72009-10-23 17:14:52 -04004398 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4399 cfqq = split_cfqq(cic, cfqq);
4400 if (!cfqq)
4401 goto new_queue;
4402 }
4403
4404 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04004405 * Check to see if this queue is scheduled to merge with
4406 * another, closely cooperating queue. The merging of
4407 * queues happens here as it must be done in process context.
4408 * The reference on new_cfqq was taken in merge_cfqqs.
4409 */
4410 if (cfqq->new_cfqq)
4411 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02004412 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004413
4414 cfqq->allocated[rw]++;
Jens Axboe5e705372006-07-13 12:39:25 +02004415
Jens Axboe6fae9c22011-03-01 15:04:39 -05004416 cfqq->ref++;
Tejun Heoeb7d8c072012-03-23 14:02:53 +01004417 cfqg_get(cfqq->cfqg);
Tejun Heoa612fdd2011-12-14 00:33:41 +01004418 rq->elv.priv[0] = cfqq;
Tejun Heo1adaf3d2012-03-05 13:15:15 -08004419 rq->elv.priv[1] = cfqq->cfqg;
Tejun Heo216284c2011-12-14 00:33:38 +01004420 spin_unlock_irq(q->queue_lock);
Jens Axboe5e705372006-07-13 12:39:25 +02004421 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004422}
4423
David Howells65f27f32006-11-22 14:55:48 +00004424static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02004425{
David Howells65f27f32006-11-22 14:55:48 +00004426 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02004427 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02004428 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02004429
Jens Axboe40bb54d2009-04-15 12:11:10 +02004430 spin_lock_irq(q->queue_lock);
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +02004431 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02004432 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02004433}
4434
4435/*
4436 * Timer running if the active_queue is currently idling inside its time slice
4437 */
4438static void cfq_idle_slice_timer(unsigned long data)
4439{
4440 struct cfq_data *cfqd = (struct cfq_data *) data;
4441 struct cfq_queue *cfqq;
4442 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11004443 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02004444
Jens Axboe7b679132008-05-30 12:23:07 +02004445 cfq_log(cfqd, "idle timer fired");
4446
Jens Axboe22e2c502005-06-27 10:55:12 +02004447 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4448
Jens Axboefe094d92008-01-31 13:08:54 +01004449 cfqq = cfqd->active_queue;
4450 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11004451 timed_out = 0;
4452
Jens Axboe22e2c502005-06-27 10:55:12 +02004453 /*
Jens Axboeb0291952009-04-07 11:38:31 +02004454 * We saw a request before the queue expired, let it through
4455 */
4456 if (cfq_cfqq_must_dispatch(cfqq))
4457 goto out_kick;
4458
4459 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02004460 * expired
4461 */
Jens Axboe44f7c162007-01-19 11:51:58 +11004462 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02004463 goto expire;
4464
4465 /*
4466 * only expire and reinvoke request handler, if there are
4467 * other queues with pending requests
4468 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02004469 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02004470 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02004471
4472 /*
4473 * not expired and it has a request pending, let it dispatch
4474 */
Jens Axboe75e50982009-04-07 08:56:14 +02004475 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02004476 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01004477
4478 /*
4479 * Queue depth flag is reset only when the idle didn't succeed
4480 */
4481 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02004482 }
4483expire:
Vivek Goyale5ff0822010-04-26 19:25:11 +02004484 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02004485out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02004486 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02004487out_cont:
4488 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4489}
4490
Jens Axboe3b181522005-06-27 10:56:24 +02004491static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4492{
4493 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02004494 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02004495}
Jens Axboe22e2c502005-06-27 10:55:12 +02004496
Jens Axboeb374d182008-10-31 10:05:07 +01004497static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004498{
Jens Axboe22e2c502005-06-27 10:55:12 +02004499 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02004500 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02004501
Jens Axboe3b181522005-06-27 10:56:24 +02004502 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02004503
Al Virod9ff4182006-03-18 13:51:22 -05004504 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02004505
Al Virod9ff4182006-03-18 13:51:22 -05004506 if (cfqd->active_queue)
Vivek Goyale5ff0822010-04-26 19:25:11 +02004507 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02004508
Tejun Heo03aa2642012-03-05 13:15:19 -08004509 spin_unlock_irq(q->queue_lock);
4510
Al Viroa90d7422006-03-18 12:05:37 -05004511 cfq_shutdown_timer_wq(cfqd);
4512
Tejun Heoffea73f2012-06-04 10:02:29 +02004513#ifdef CONFIG_CFQ_GROUP_IOSCHED
4514 blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4515#else
Tejun Heof51b8022012-03-05 13:15:05 -08004516 kfree(cfqd->root_group);
Vivek Goyal2abae552011-05-23 10:02:19 +02004517#endif
Vivek Goyal56edf7d2011-05-19 15:38:22 -04004518 kfree(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004519}
4520
Jianpeng Mad50235b2013-07-03 13:25:24 +02004521static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004522{
4523 struct cfq_data *cfqd;
Tejun Heo3c798392012-04-16 13:57:25 -07004524 struct blkcg_gq *blkg __maybe_unused;
Tejun Heoa2b16932012-04-13 13:11:33 -07004525 int i, ret;
Jianpeng Mad50235b2013-07-03 13:25:24 +02004526 struct elevator_queue *eq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004527
Jianpeng Mad50235b2013-07-03 13:25:24 +02004528 eq = elevator_alloc(q, e);
4529 if (!eq)
Tejun Heob2fab5a2012-03-05 13:14:57 -08004530 return -ENOMEM;
Konstantin Khlebnikov80b15c72010-05-20 23:21:41 +04004531
Joe Perchesc1b511e2013-08-29 15:21:42 -07004532 cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
Jianpeng Mad50235b2013-07-03 13:25:24 +02004533 if (!cfqd) {
4534 kobject_put(&eq->kobj);
4535 return -ENOMEM;
4536 }
4537 eq->elevator_data = cfqd;
4538
Tejun Heof51b8022012-03-05 13:15:05 -08004539 cfqd->queue = q;
Jianpeng Mad50235b2013-07-03 13:25:24 +02004540 spin_lock_irq(q->queue_lock);
4541 q->elevator = eq;
4542 spin_unlock_irq(q->queue_lock);
Tejun Heof51b8022012-03-05 13:15:05 -08004543
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05004544 /* Init root service tree */
4545 cfqd->grp_service_tree = CFQ_RB_ROOT;
4546
Tejun Heof51b8022012-03-05 13:15:05 -08004547 /* Init root group and prefer root group over other groups by default */
Vivek Goyal25fb5162009-12-03 12:59:46 -05004548#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004549 ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
Tejun Heoa2b16932012-04-13 13:11:33 -07004550 if (ret)
4551 goto out_free;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04004552
Tejun Heoa2b16932012-04-13 13:11:33 -07004553 cfqd->root_group = blkg_to_cfqg(q->root_blkg);
Tejun Heof51b8022012-03-05 13:15:05 -08004554#else
Tejun Heoa2b16932012-04-13 13:11:33 -07004555 ret = -ENOMEM;
Tejun Heof51b8022012-03-05 13:15:05 -08004556 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4557 GFP_KERNEL, cfqd->queue->node);
Tejun Heoa2b16932012-04-13 13:11:33 -07004558 if (!cfqd->root_group)
4559 goto out_free;
Vivek Goyal5624a4e2011-05-19 15:38:28 -04004560
Tejun Heoa2b16932012-04-13 13:11:33 -07004561 cfq_init_cfqg_base(cfqd->root_group);
Tejun Heo3ecca622015-08-18 14:55:35 -07004562 cfqd->root_group->weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4563 cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
Tejun Heo69d7fde2015-08-18 14:55:36 -07004564#endif
Vivek Goyal5624a4e2011-05-19 15:38:28 -04004565
Jens Axboe26a2ac02009-04-23 12:13:27 +02004566 /*
4567 * Not strictly needed (since RB_ROOT just clears the node and we
4568 * zeroed cfqd on alloc), but better be safe in case someone decides
4569 * to add magic to the rb code
4570 */
4571 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4572 cfqd->prio_trees[i] = RB_ROOT;
4573
Jens Axboe6118b702009-06-30 09:34:12 +02004574 /*
Tejun Heod4aad7f2015-08-18 14:55:04 -07004575 * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
Jens Axboe6118b702009-06-30 09:34:12 +02004576 * Grab a permanent reference to it, so that the normal code flow
Tejun Heof51b8022012-03-05 13:15:05 -08004577 * will not attempt to free it. oom_cfqq is linked to root_group
4578 * but shouldn't hold a reference as it'll never be unlinked. Lose
4579 * the reference from linking right away.
Jens Axboe6118b702009-06-30 09:34:12 +02004580 */
4581 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
Shaohua Li30d7b942011-01-07 08:46:59 +01004582 cfqd->oom_cfqq.ref++;
Tejun Heo1adaf3d2012-03-05 13:15:15 -08004583
4584 spin_lock_irq(q->queue_lock);
Tejun Heof51b8022012-03-05 13:15:05 -08004585 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
Tejun Heoeb7d8c072012-03-23 14:02:53 +01004586 cfqg_put(cfqd->root_group);
Tejun Heo1adaf3d2012-03-05 13:15:15 -08004587 spin_unlock_irq(q->queue_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004588
Jens Axboe22e2c502005-06-27 10:55:12 +02004589 init_timer(&cfqd->idle_slice_timer);
4590 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4591 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4592
Jens Axboe23e018a2009-10-05 08:52:35 +02004593 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02004594
Linus Torvalds1da177e2005-04-16 15:20:36 -07004595 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02004596 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4597 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004598 cfqd->cfq_back_max = cfq_back_max;
4599 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02004600 cfqd->cfq_slice[0] = cfq_slice_async;
4601 cfqd->cfq_slice[1] = cfq_slice_sync;
Tao Ma5bf14c02012-04-01 14:33:39 -07004602 cfqd->cfq_target_latency = cfq_target_latency;
Jens Axboe22e2c502005-06-27 10:55:12 +02004603 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
Jens Axboe0bb97942015-06-10 08:01:20 -06004604 cfqd->cfq_slice_idle = cfq_slice_idle;
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004605 cfqd->cfq_group_idle = cfq_group_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02004606 cfqd->cfq_latency = 1;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01004607 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01004608 /*
4609 * we optimistically start assuming sync ops weren't delayed in last
4610 * second, in order to have larger depth for async operations.
4611 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01004612 cfqd->last_delayed_sync = jiffies - HZ;
Tejun Heob2fab5a2012-03-05 13:14:57 -08004613 return 0;
Tejun Heoa2b16932012-04-13 13:11:33 -07004614
4615out_free:
4616 kfree(cfqd);
Jianpeng Mad50235b2013-07-03 13:25:24 +02004617 kobject_put(&eq->kobj);
Tejun Heoa2b16932012-04-13 13:11:33 -07004618 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004619}
4620
Jens Axboe0bb97942015-06-10 08:01:20 -06004621static void cfq_registered_queue(struct request_queue *q)
4622{
4623 struct elevator_queue *e = q->elevator;
4624 struct cfq_data *cfqd = e->elevator_data;
4625
4626 /*
4627 * Default to IOPS mode with no idling for SSDs
4628 */
4629 if (blk_queue_nonrot(q))
4630 cfqd->cfq_slice_idle = 0;
4631}
4632
Linus Torvalds1da177e2005-04-16 15:20:36 -07004633/*
4634 * sysfs parts below -->
4635 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004636static ssize_t
4637cfq_var_show(unsigned int var, char *page)
4638{
Masanari Iida176167a2014-04-28 12:38:34 +09004639 return sprintf(page, "%u\n", var);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004640}
4641
4642static ssize_t
4643cfq_var_store(unsigned int *var, const char *page, size_t count)
4644{
4645 char *p = (char *) page;
4646
4647 *var = simple_strtoul(p, &p, 10);
4648 return count;
4649}
4650
Linus Torvalds1da177e2005-04-16 15:20:36 -07004651#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004652static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004653{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004654 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004655 unsigned int __data = __VAR; \
4656 if (__CONV) \
4657 __data = jiffies_to_msecs(__data); \
4658 return cfq_var_show(__data, (page)); \
4659}
4660SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004661SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4662SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05004663SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4664SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004665SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004666SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004667SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4668SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4669SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004670SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07004671SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004672#undef SHOW_FUNCTION
4673
4674#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01004675static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004676{ \
Al Viro3d1ab402006-03-18 18:35:43 -05004677 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07004678 unsigned int __data; \
4679 int ret = cfq_var_store(&__data, (page), count); \
4680 if (__data < (MIN)) \
4681 __data = (MIN); \
4682 else if (__data > (MAX)) \
4683 __data = (MAX); \
4684 if (__CONV) \
4685 *(__PTR) = msecs_to_jiffies(__data); \
4686 else \
4687 *(__PTR) = __data; \
4688 return ret; \
4689}
4690STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004691STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4692 UINT_MAX, 1);
4693STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4694 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05004695STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01004696STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4697 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02004698STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004699STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02004700STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4701STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01004702STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4703 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02004704STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Tao Ma5bf14c02012-04-01 14:33:39 -07004705STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004706#undef STORE_FUNCTION
4707
Al Viroe572ec72006-03-18 22:27:18 -05004708#define CFQ_ATTR(name) \
4709 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02004710
Al Viroe572ec72006-03-18 22:27:18 -05004711static struct elv_fs_entry cfq_attrs[] = {
4712 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05004713 CFQ_ATTR(fifo_expire_sync),
4714 CFQ_ATTR(fifo_expire_async),
4715 CFQ_ATTR(back_seek_max),
4716 CFQ_ATTR(back_seek_penalty),
4717 CFQ_ATTR(slice_sync),
4718 CFQ_ATTR(slice_async),
4719 CFQ_ATTR(slice_async_rq),
4720 CFQ_ATTR(slice_idle),
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004721 CFQ_ATTR(group_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02004722 CFQ_ATTR(low_latency),
Tao Ma5bf14c02012-04-01 14:33:39 -07004723 CFQ_ATTR(target_latency),
Al Viroe572ec72006-03-18 22:27:18 -05004724 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07004725};
4726
Linus Torvalds1da177e2005-04-16 15:20:36 -07004727static struct elevator_type iosched_cfq = {
4728 .ops = {
4729 .elevator_merge_fn = cfq_merge,
4730 .elevator_merged_fn = cfq_merged_request,
4731 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01004732 .elevator_allow_merge_fn = cfq_allow_merge,
Divyesh Shah812d4022010-04-08 21:14:23 -07004733 .elevator_bio_merged_fn = cfq_bio_merged,
Jens Axboeb4878f22005-10-20 16:42:29 +02004734 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004735 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02004736 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004737 .elevator_deactivate_req_fn = cfq_deactivate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004738 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02004739 .elevator_former_req_fn = elv_rb_former_request,
4740 .elevator_latter_req_fn = elv_rb_latter_request,
Tejun Heo9b84cac2011-12-14 00:33:42 +01004741 .elevator_init_icq_fn = cfq_init_icq,
Tejun Heo7e5a8792011-12-14 00:33:42 +01004742 .elevator_exit_icq_fn = cfq_exit_icq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004743 .elevator_set_req_fn = cfq_set_request,
4744 .elevator_put_req_fn = cfq_put_request,
4745 .elevator_may_queue_fn = cfq_may_queue,
4746 .elevator_init_fn = cfq_init_queue,
4747 .elevator_exit_fn = cfq_exit_queue,
Jens Axboe0bb97942015-06-10 08:01:20 -06004748 .elevator_registered_fn = cfq_registered_queue,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004749 },
Tejun Heo3d3c2372011-12-14 00:33:42 +01004750 .icq_size = sizeof(struct cfq_io_cq),
4751 .icq_align = __alignof__(struct cfq_io_cq),
Al Viro3d1ab402006-03-18 18:35:43 -05004752 .elevator_attrs = cfq_attrs,
Tejun Heo3d3c2372011-12-14 00:33:42 +01004753 .elevator_name = "cfq",
Linus Torvalds1da177e2005-04-16 15:20:36 -07004754 .elevator_owner = THIS_MODULE,
4755};
4756
Vivek Goyal3e252062009-12-04 10:36:42 -05004757#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004758static struct blkcg_policy blkcg_policy_cfq = {
Tejun Heo2ee867dc2015-08-18 14:55:34 -07004759 .dfl_cftypes = cfq_blkcg_files,
Tejun Heo880f50e2015-08-18 14:55:30 -07004760 .legacy_cftypes = cfq_blkcg_legacy_files,
Tejun Heof9fcc2d2012-04-16 13:57:27 -07004761
Tejun Heoe4a9bde2015-08-18 14:55:16 -07004762 .cpd_alloc_fn = cfq_cpd_alloc,
Arianna Avanzinie48453c2015-06-05 23:38:42 +02004763 .cpd_init_fn = cfq_cpd_init,
Tejun Heoe4a9bde2015-08-18 14:55:16 -07004764 .cpd_free_fn = cfq_cpd_free,
Tejun Heo69d7fde2015-08-18 14:55:36 -07004765 .cpd_bind_fn = cfq_cpd_bind,
Tejun Heoe4a9bde2015-08-18 14:55:16 -07004766
Tejun Heo001bea72015-08-18 14:55:11 -07004767 .pd_alloc_fn = cfq_pd_alloc,
Tejun Heof9fcc2d2012-04-16 13:57:27 -07004768 .pd_init_fn = cfq_pd_init,
Tejun Heo0b399202013-01-09 08:05:13 -08004769 .pd_offline_fn = cfq_pd_offline,
Tejun Heo001bea72015-08-18 14:55:11 -07004770 .pd_free_fn = cfq_pd_free,
Tejun Heof9fcc2d2012-04-16 13:57:27 -07004771 .pd_reset_stats_fn = cfq_pd_reset_stats,
Vivek Goyal3e252062009-12-04 10:36:42 -05004772};
Vivek Goyal3e252062009-12-04 10:36:42 -05004773#endif
4774
Linus Torvalds1da177e2005-04-16 15:20:36 -07004775static int __init cfq_init(void)
4776{
Tejun Heo3d3c2372011-12-14 00:33:42 +01004777 int ret;
4778
Jens Axboe22e2c502005-06-27 10:55:12 +02004779 /*
4780 * could be 0 on HZ < 1000 setups
4781 */
4782 if (!cfq_slice_async)
4783 cfq_slice_async = 1;
4784 if (!cfq_slice_idle)
4785 cfq_slice_idle = 1;
4786
Vivek Goyal80bdf0c2010-08-23 12:24:26 +02004787#ifdef CONFIG_CFQ_GROUP_IOSCHED
4788 if (!cfq_group_idle)
4789 cfq_group_idle = 1;
Tejun Heo8bd435b2012-04-13 13:11:28 -07004790
Tejun Heo3c798392012-04-16 13:57:25 -07004791 ret = blkcg_policy_register(&blkcg_policy_cfq);
Tejun Heo8bd435b2012-04-13 13:11:28 -07004792 if (ret)
4793 return ret;
Tejun Heoffea73f2012-06-04 10:02:29 +02004794#else
4795 cfq_group_idle = 0;
4796#endif
Tejun Heo8bd435b2012-04-13 13:11:28 -07004797
Tejun Heofd794952012-06-04 10:01:38 +02004798 ret = -ENOMEM;
Tejun Heo3d3c2372011-12-14 00:33:42 +01004799 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4800 if (!cfq_pool)
Tejun Heo8bd435b2012-04-13 13:11:28 -07004801 goto err_pol_unreg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004802
Tejun Heo3d3c2372011-12-14 00:33:42 +01004803 ret = elv_register(&iosched_cfq);
Tejun Heo8bd435b2012-04-13 13:11:28 -07004804 if (ret)
4805 goto err_free_pool;
Tejun Heo3d3c2372011-12-14 00:33:42 +01004806
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01004807 return 0;
Tejun Heo8bd435b2012-04-13 13:11:28 -07004808
4809err_free_pool:
4810 kmem_cache_destroy(cfq_pool);
4811err_pol_unreg:
Tejun Heoffea73f2012-06-04 10:02:29 +02004812#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004813 blkcg_policy_unregister(&blkcg_policy_cfq);
Tejun Heoffea73f2012-06-04 10:02:29 +02004814#endif
Tejun Heo8bd435b2012-04-13 13:11:28 -07004815 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004816}
4817
4818static void __exit cfq_exit(void)
4819{
Tejun Heoffea73f2012-06-04 10:02:29 +02004820#ifdef CONFIG_CFQ_GROUP_IOSCHED
Tejun Heo3c798392012-04-16 13:57:25 -07004821 blkcg_policy_unregister(&blkcg_policy_cfq);
Tejun Heoffea73f2012-06-04 10:02:29 +02004822#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004823 elv_unregister(&iosched_cfq);
Tejun Heo3d3c2372011-12-14 00:33:42 +01004824 kmem_cache_destroy(cfq_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004825}
4826
4827module_init(cfq_init);
4828module_exit(cfq_exit);
4829
4830MODULE_AUTHOR("Jens Axboe");
4831MODULE_LICENSE("GPL");
4832MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");