blob: cfb0b2f5f63de6d75213b16373f84a248fbd6421 [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>
Al Viro1cc9be62006-03-18 12:29:52 -050010#include <linux/blkdev.h>
11#include <linux/elevator.h>
Randy Dunlapad5ebd22009-11-11 13:47:45 +010012#include <linux/jiffies.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/rbtree.h>
Jens Axboe22e2c502005-06-27 10:55:12 +020014#include <linux/ioprio.h>
Jens Axboe7b679132008-05-30 12:23:07 +020015#include <linux/blktrace_api.h>
Vivek Goyal25bc6b02009-12-03 12:59:43 -050016#include "blk-cgroup.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070017
18/*
19 * tunables
20 */
Jens Axboefe094d92008-01-31 13:08:54 +010021/* max queue in one round of service */
22static const int cfq_quantum = 4;
Arjan van de Ven64100092006-01-06 09:46:02 +010023static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
Jens Axboefe094d92008-01-31 13:08:54 +010024/* maximum backwards seek, in KiB */
25static const int cfq_back_max = 16 * 1024;
26/* penalty of a backwards seek */
27static const int cfq_back_penalty = 2;
Arjan van de Ven64100092006-01-06 09:46:02 +010028static const int cfq_slice_sync = HZ / 10;
Jens Axboe3b181522005-06-27 10:56:24 +020029static int cfq_slice_async = HZ / 25;
Arjan van de Ven64100092006-01-06 09:46:02 +010030static const int cfq_slice_async_rq = 2;
Jens Axboecaaa5f92006-06-16 11:23:00 +020031static int cfq_slice_idle = HZ / 125;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +010032static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33static const int cfq_hist_divisor = 4;
Jens Axboe22e2c502005-06-27 10:55:12 +020034
Jens Axboed9e76202007-04-20 14:27:50 +020035/*
Jens Axboe08717142008-01-28 11:38:15 +010036 * offset from end of service tree
Jens Axboed9e76202007-04-20 14:27:50 +020037 */
Jens Axboe08717142008-01-28 11:38:15 +010038#define CFQ_IDLE_DELAY (HZ / 5)
Jens Axboed9e76202007-04-20 14:27:50 +020039
40/*
41 * below this threshold, we consider thinktime immediate
42 */
43#define CFQ_MIN_TT (2)
44
Jeff Moyere6c5bc72009-10-23 17:14:52 -040045/*
46 * Allow merged cfqqs to perform this amount of seeky I/O before
47 * deciding to break the queues up again.
48 */
49#define CFQQ_COOP_TOUT (HZ)
50
Jens Axboe22e2c502005-06-27 10:55:12 +020051#define CFQ_SLICE_SCALE (5)
Aaron Carroll45333d52008-08-26 15:52:36 +020052#define CFQ_HW_QUEUE_MIN (5)
Vivek Goyal25bc6b02009-12-03 12:59:43 -050053#define CFQ_SERVICE_SHIFT 12
Jens Axboe22e2c502005-06-27 10:55:12 +020054
Jens Axboefe094d92008-01-31 13:08:54 +010055#define RQ_CIC(rq) \
56 ((struct cfq_io_context *) (rq)->elevator_private)
Jens Axboe7b679132008-05-30 12:23:07 +020057#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
Linus Torvalds1da177e2005-04-16 15:20:36 -070058
Christoph Lametere18b8902006-12-06 20:33:20 -080059static struct kmem_cache *cfq_pool;
60static struct kmem_cache *cfq_ioc_pool;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
Tejun Heo245b2e72009-06-24 15:13:48 +090062static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -050063static struct completion *ioc_gone;
Jens Axboe9a11b4e2008-05-29 09:32:08 +020064static DEFINE_SPINLOCK(ioc_gone_lock);
Al Viro334e94d2006-03-18 15:05:53 -050065
Jens Axboe22e2c502005-06-27 10:55:12 +020066#define CFQ_PRIO_LISTS IOPRIO_BE_NR
67#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
Jens Axboe22e2c502005-06-27 10:55:12 +020068#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
69
Jens Axboe206dc692006-03-28 13:03:44 +020070#define sample_valid(samples) ((samples) > 80)
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050071#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
Jens Axboe206dc692006-03-28 13:03:44 +020072
Jens Axboe22e2c502005-06-27 10:55:12 +020073/*
Jens Axboecc09e292007-04-26 12:53:50 +020074 * Most of our rbtree usage is for sorting with min extraction, so
75 * if we cache the leftmost node we don't have to walk down the tree
76 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
77 * move this into the elevator for the rq sorting as well.
78 */
79struct cfq_rb_root {
80 struct rb_root rb;
81 struct rb_node *left;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +010082 unsigned count;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050083 u64 min_vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -050084 struct rb_node *active;
Vivek Goyal58ff82f2009-12-03 12:59:44 -050085 unsigned total_weight;
Jens Axboecc09e292007-04-26 12:53:50 +020086};
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -050087#define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
Jens Axboecc09e292007-04-26 12:53:50 +020088
89/*
Jens Axboe6118b702009-06-30 09:34:12 +020090 * Per process-grouping structure
91 */
92struct cfq_queue {
93 /* reference count */
94 atomic_t ref;
95 /* various state flags, see below */
96 unsigned int flags;
97 /* parent cfq_data */
98 struct cfq_data *cfqd;
99 /* service_tree member */
100 struct rb_node rb_node;
101 /* service_tree key */
102 unsigned long rb_key;
103 /* prio tree member */
104 struct rb_node p_node;
105 /* prio tree root we belong to, if any */
106 struct rb_root *p_root;
107 /* sorted list of pending requests */
108 struct rb_root sort_list;
109 /* if fifo isn't expired, next request to serve */
110 struct request *next_rq;
111 /* requests queued in sort_list */
112 int queued[2];
113 /* currently allocated requests */
114 int allocated[2];
115 /* fifo list of requests in sort_list */
116 struct list_head fifo;
117
Vivek Goyaldae739e2009-12-03 12:59:45 -0500118 /* time when queue got scheduled in to dispatch first request. */
119 unsigned long dispatch_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500120 unsigned int allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500121 /* time when first request from queue completed and slice started. */
122 unsigned long slice_start;
Jens Axboe6118b702009-06-30 09:34:12 +0200123 unsigned long slice_end;
124 long slice_resid;
125 unsigned int slice_dispatch;
126
127 /* pending metadata requests */
128 int meta_pending;
129 /* number of requests that are on the dispatch list or inside driver */
130 int dispatched;
131
132 /* io prio of this group */
133 unsigned short ioprio, org_ioprio;
134 unsigned short ioprio_class, org_ioprio_class;
135
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400136 unsigned int seek_samples;
137 u64 seek_total;
138 sector_t seek_mean;
139 sector_t last_request_pos;
Jeff Moyere6c5bc72009-10-23 17:14:52 -0400140 unsigned long seeky_start;
Jeff Moyerb2c18e12009-10-23 17:14:49 -0400141
Jens Axboe6118b702009-06-30 09:34:12 +0200142 pid_t pid;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400143
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100144 struct cfq_rb_root *service_tree;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -0400145 struct cfq_queue *new_cfqq;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500146 struct cfq_group *cfqg;
Vivek Goyalae30c282009-12-03 12:59:55 -0500147 struct cfq_group *orig_cfqg;
Vivek Goyal22084192009-12-03 12:59:49 -0500148 /* Sectors dispatched in current dispatch round */
149 unsigned long nr_sectors;
Jens Axboe6118b702009-06-30 09:34:12 +0200150};
151
152/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100153 * First index in the service_trees.
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100154 * IDLE is handled separately, so it has negative index
155 */
156enum wl_prio_t {
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100157 BE_WORKLOAD = 0,
Vivek Goyal615f0252009-12-03 12:59:39 -0500158 RT_WORKLOAD = 1,
159 IDLE_WORKLOAD = 2,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100160};
161
162/*
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100163 * Second index in the service_trees.
164 */
165enum wl_type_t {
166 ASYNC_WORKLOAD = 0,
167 SYNC_NOIDLE_WORKLOAD = 1,
168 SYNC_WORKLOAD = 2
169};
170
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500171/* This is per cgroup per device grouping structure */
172struct cfq_group {
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500173 /* group service_tree member */
174 struct rb_node rb_node;
175
176 /* group service_tree key */
177 u64 vdisktime;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500178 unsigned int weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500179 bool on_st;
180
181 /* number of cfqq currently on this group */
182 int nr_cfqq;
183
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500184 /* Per group busy queus average. Useful for workload slice calc. */
185 unsigned int busy_queues_avg[2];
Jens Axboe22e2c502005-06-27 10:55:12 +0200186 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100187 * rr lists of queues with requests, onle rr for each priority class.
188 * Counts are embedded in the cfq_rb_root
Jens Axboe22e2c502005-06-27 10:55:12 +0200189 */
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100190 struct cfq_rb_root service_trees[2][3];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100191 struct cfq_rb_root service_tree_idle;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500192
193 unsigned long saved_workload_slice;
194 enum wl_type_t saved_workload;
195 enum wl_prio_t saved_serving_prio;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500196 struct blkio_group blkg;
197#ifdef CONFIG_CFQ_GROUP_IOSCHED
198 struct hlist_node cfqd_node;
Vivek Goyalb1c35762009-12-03 12:59:47 -0500199 atomic_t ref;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500200#endif
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500201};
202
203/*
204 * Per block device queue structure
205 */
206struct cfq_data {
207 struct request_queue *queue;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500208 /* Root service tree for cfq_groups */
209 struct cfq_rb_root grp_service_tree;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500210 struct cfq_group root_group;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500211 /* Number of active cfq groups on group service tree */
212 int nr_groups;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500213
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100214 /*
215 * The priority currently being served
216 */
217 enum wl_prio_t serving_prio;
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100218 enum wl_type_t serving_type;
219 unsigned long workload_expires;
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500220 struct cfq_group *serving_group;
Corrado Zoccolo8e550632009-11-26 10:02:58 +0100221 bool noidle_tree_requires_idle;
Jens Axboea36e71f2009-04-15 12:15:11 +0200222
223 /*
224 * Each priority tree is sorted by next_request position. These
225 * trees are used when determining if two or more queues are
226 * interleaving requests (see cfq_close_cooperator).
227 */
228 struct rb_root prio_trees[CFQ_PRIO_LISTS];
229
Jens Axboe22e2c502005-06-27 10:55:12 +0200230 unsigned int busy_queues;
231
Jens Axboe5ad531d2009-07-03 12:57:48 +0200232 int rq_in_driver[2];
Jens Axboe3ed9a292007-04-23 08:33:33 +0200233 int sync_flight;
Aaron Carroll45333d52008-08-26 15:52:36 +0200234
235 /*
236 * queue-depth detection
237 */
238 int rq_queued;
Jens Axboe25776e32006-06-01 10:12:26 +0200239 int hw_tag;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +0100240 /*
241 * hw_tag can be
242 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
243 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
244 * 0 => no NCQ
245 */
246 int hw_tag_est_depth;
247 unsigned int hw_tag_samples;
Jens Axboe22e2c502005-06-27 10:55:12 +0200248
249 /*
Jens Axboe22e2c502005-06-27 10:55:12 +0200250 * idle window management
251 */
252 struct timer_list idle_slice_timer;
Jens Axboe23e018a2009-10-05 08:52:35 +0200253 struct work_struct unplug_work;
Jens Axboe22e2c502005-06-27 10:55:12 +0200254
255 struct cfq_queue *active_queue;
256 struct cfq_io_context *active_cic;
Jens Axboe22e2c502005-06-27 10:55:12 +0200257
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +0200258 /*
259 * async queue for each priority case
260 */
261 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
262 struct cfq_queue *async_idle_cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +0200263
Jens Axboe6d048f52007-04-25 12:44:27 +0200264 sector_t last_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 /*
267 * tunables, see top of file
268 */
269 unsigned int cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +0200270 unsigned int cfq_fifo_expire[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 unsigned int cfq_back_penalty;
272 unsigned int cfq_back_max;
Jens Axboe22e2c502005-06-27 10:55:12 +0200273 unsigned int cfq_slice[2];
274 unsigned int cfq_slice_async_rq;
275 unsigned int cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +0200276 unsigned int cfq_latency;
Vivek Goyalae30c282009-12-03 12:59:55 -0500277 unsigned int cfq_group_isolation;
Al Virod9ff4182006-03-18 13:51:22 -0500278
279 struct list_head cic_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280
Jens Axboe6118b702009-06-30 09:34:12 +0200281 /*
282 * Fallback dummy cfqq for extreme OOM conditions
283 */
284 struct cfq_queue oom_cfqq;
Vivek Goyal365722b2009-10-03 15:21:27 +0200285
286 unsigned long last_end_sync_rq;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500287
288 /* List of cfq groups being managed on this device*/
289 struct hlist_head cfqg_list;
Jens Axboebb729bc2009-12-06 09:54:19 +0100290 struct rcu_head rcu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291};
292
Vivek Goyal25fb5162009-12-03 12:59:46 -0500293static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
294
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500295static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
296 enum wl_prio_t prio,
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100297 enum wl_type_t type,
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100298 struct cfq_data *cfqd)
299{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500300 if (!cfqg)
301 return NULL;
302
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100303 if (prio == IDLE_WORKLOAD)
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500304 return &cfqg->service_tree_idle;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100305
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500306 return &cfqg->service_trees[prio][type];
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100307}
308
Jens Axboe3b181522005-06-27 10:56:24 +0200309enum cfqq_state_flags {
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100310 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
311 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
Jens Axboeb0291952009-04-07 11:38:31 +0200312 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100313 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
Jens Axboeb0b8d7492007-01-19 11:35:30 +1100314 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
315 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
316 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
Jens Axboe44f7c162007-01-19 11:51:58 +1100317 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
Vasily Tarasov91fac312007-04-25 12:29:51 +0200318 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
Jeff Moyerb3b6d042009-10-23 17:14:51 -0400319 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100320 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
Vivek Goyalf75edf22009-12-03 12:59:53 -0500321 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
322 CFQ_CFQQ_FLAG_wait_busy_done, /* Got new request. Expire the queue */
Jens Axboe3b181522005-06-27 10:56:24 +0200323};
324
325#define CFQ_CFQQ_FNS(name) \
326static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
327{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100328 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200329} \
330static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
331{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100332 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200333} \
334static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
335{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100336 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200337}
338
339CFQ_CFQQ_FNS(on_rr);
340CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200341CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200342CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200343CFQ_CFQQ_FNS(fifo_expire);
344CFQ_CFQQ_FNS(idle_window);
345CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100346CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200347CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200348CFQ_CFQQ_FNS(coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100349CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500350CFQ_CFQQ_FNS(wait_busy);
351CFQ_CFQQ_FNS(wait_busy_done);
Jens Axboe3b181522005-06-27 10:56:24 +0200352#undef CFQ_CFQQ_FNS
353
Vivek Goyal2868ef72009-12-03 12:59:48 -0500354#ifdef CONFIG_DEBUG_CFQ_IOSCHED
355#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
356 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
357 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
358 blkg_path(&(cfqq)->cfqg->blkg), ##args);
359
360#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
361 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
362 blkg_path(&(cfqg)->blkg), ##args); \
363
364#else
Jens Axboe7b679132008-05-30 12:23:07 +0200365#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
366 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500367#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
368#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200369#define cfq_log(cfqd, fmt, args...) \
370 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
371
Vivek Goyal615f0252009-12-03 12:59:39 -0500372/* Traverses through cfq group service trees */
373#define for_each_cfqg_st(cfqg, i, j, st) \
374 for (i = 0; i <= IDLE_WORKLOAD; i++) \
375 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
376 : &cfqg->service_tree_idle; \
377 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
378 (i == IDLE_WORKLOAD && j == 0); \
379 j++, st = i < IDLE_WORKLOAD ? \
380 &cfqg->service_trees[i][j]: NULL) \
381
382
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100383static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
384{
385 if (cfq_class_idle(cfqq))
386 return IDLE_WORKLOAD;
387 if (cfq_class_rt(cfqq))
388 return RT_WORKLOAD;
389 return BE_WORKLOAD;
390}
391
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100392
393static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
394{
395 if (!cfq_cfqq_sync(cfqq))
396 return ASYNC_WORKLOAD;
397 if (!cfq_cfqq_idle_window(cfqq))
398 return SYNC_NOIDLE_WORKLOAD;
399 return SYNC_WORKLOAD;
400}
401
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500402static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
403 struct cfq_data *cfqd,
404 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100405{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500406 if (wl == IDLE_WORKLOAD)
407 return cfqg->service_tree_idle.count;
408
409 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
410 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
411 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100412}
413
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500414static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
415 struct cfq_group *cfqg)
416{
417 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
418 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
419}
420
Jens Axboe165125e2007-07-24 09:28:11 +0200421static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200422static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100423 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100424static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200425 struct io_context *);
426
Jens Axboe5ad531d2009-07-03 12:57:48 +0200427static inline int rq_in_driver(struct cfq_data *cfqd)
428{
429 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
430}
431
Vasily Tarasov91fac312007-04-25 12:29:51 +0200432static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200433 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200434{
Jens Axboea6151c32009-10-07 20:02:57 +0200435 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200436}
437
438static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200439 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200440{
Jens Axboea6151c32009-10-07 20:02:57 +0200441 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200442}
443
444/*
445 * We regard a request as SYNC, if it's either a read or has the SYNC bit
446 * set (in which case it could also be direct WRITE).
447 */
Jens Axboea6151c32009-10-07 20:02:57 +0200448static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200449{
Jens Axboea6151c32009-10-07 20:02:57 +0200450 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200451}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700454 * scheduler run of queue, if there are requests pending and no one in the
455 * driver that will restart queueing
456 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200457static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700458{
Jens Axboe7b679132008-05-30 12:23:07 +0200459 if (cfqd->busy_queues) {
460 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200461 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200462 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700463}
464
Jens Axboe165125e2007-07-24 09:28:11 +0200465static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700466{
467 struct cfq_data *cfqd = q->elevator->elevator_data;
468
Vivek Goyalf04a6422009-12-03 12:59:40 -0500469 return !cfqd->rq_queued;
Andrew Morton99f95e52005-06-27 20:14:05 -0700470}
471
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100473 * Scale schedule slice based on io priority. Use the sync time slice only
474 * if a queue is marked sync and has sync io queued. A sync queue with async
475 * io only, should not get full sync slice length.
476 */
Jens Axboea6151c32009-10-07 20:02:57 +0200477static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200478 unsigned short prio)
479{
480 const int base_slice = cfqd->cfq_slice[sync];
481
482 WARN_ON(prio >= IOPRIO_BE_NR);
483
484 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
485}
486
Jens Axboe44f7c162007-01-19 11:51:58 +1100487static inline int
488cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
489{
Jens Axboed9e76202007-04-20 14:27:50 +0200490 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100491}
492
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500493static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
494{
495 u64 d = delta << CFQ_SERVICE_SHIFT;
496
497 d = d * BLKIO_WEIGHT_DEFAULT;
498 do_div(d, cfqg->weight);
499 return d;
500}
501
502static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
503{
504 s64 delta = (s64)(vdisktime - min_vdisktime);
505 if (delta > 0)
506 min_vdisktime = vdisktime;
507
508 return min_vdisktime;
509}
510
511static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
512{
513 s64 delta = (s64)(vdisktime - min_vdisktime);
514 if (delta < 0)
515 min_vdisktime = vdisktime;
516
517 return min_vdisktime;
518}
519
520static void update_min_vdisktime(struct cfq_rb_root *st)
521{
522 u64 vdisktime = st->min_vdisktime;
523 struct cfq_group *cfqg;
524
525 if (st->active) {
526 cfqg = rb_entry_cfqg(st->active);
527 vdisktime = cfqg->vdisktime;
528 }
529
530 if (st->left) {
531 cfqg = rb_entry_cfqg(st->left);
532 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
533 }
534
535 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
536}
537
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100538/*
539 * get averaged number of queues of RT/BE priority.
540 * average is updated, with a formula that gives more weight to higher numbers,
541 * to quickly follows sudden increases and decrease slowly
542 */
543
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500544static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
545 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100546{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100547 unsigned min_q, max_q;
548 unsigned mult = cfq_hist_divisor - 1;
549 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500550 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100551
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500552 min_q = min(cfqg->busy_queues_avg[rt], busy);
553 max_q = max(cfqg->busy_queues_avg[rt], busy);
554 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100555 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500556 return cfqg->busy_queues_avg[rt];
557}
558
559static inline unsigned
560cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
561{
562 struct cfq_rb_root *st = &cfqd->grp_service_tree;
563
564 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100565}
566
Jens Axboe44f7c162007-01-19 11:51:58 +1100567static inline void
568cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
569{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100570 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
571 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500572 /*
573 * interested queues (we consider only the ones with the same
574 * priority class in the cfq group)
575 */
576 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
577 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100578 unsigned sync_slice = cfqd->cfq_slice[1];
579 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500580 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
581
582 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100583 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
584 /* scale low_slice according to IO priority
585 * and sync vs async */
586 unsigned low_slice =
587 min(slice, base_low_slice * slice / sync_slice);
588 /* the adapted slice value is scaled to fit all iqs
589 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500590 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100591 low_slice);
592 }
593 }
Vivek Goyaldae739e2009-12-03 12:59:45 -0500594 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100595 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500596 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200597 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100598}
599
600/*
601 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
602 * isn't valid until the first request from the dispatch is activated
603 * and the slice time set.
604 */
Jens Axboea6151c32009-10-07 20:02:57 +0200605static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100606{
607 if (cfq_cfqq_slice_new(cfqq))
608 return 0;
609 if (time_before(jiffies, cfqq->slice_end))
610 return 0;
611
612 return 1;
613}
614
615/*
Jens Axboe5e705372006-07-13 12:39:25 +0200616 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200618 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 */
Jens Axboe5e705372006-07-13 12:39:25 +0200620static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100621cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100623 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200625#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
626#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
627 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628
Jens Axboe5e705372006-07-13 12:39:25 +0200629 if (rq1 == NULL || rq1 == rq2)
630 return rq2;
631 if (rq2 == NULL)
632 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200633
Jens Axboe5e705372006-07-13 12:39:25 +0200634 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
635 return rq1;
636 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
637 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200638 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
639 return rq1;
640 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
641 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642
Tejun Heo83096eb2009-05-07 22:24:39 +0900643 s1 = blk_rq_pos(rq1);
644 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 /*
647 * by definition, 1KiB is 2 sectors
648 */
649 back_max = cfqd->cfq_back_max * 2;
650
651 /*
652 * Strict one way elevator _except_ in the case where we allow
653 * short backward seeks which are biased as twice the cost of a
654 * similar forward seek.
655 */
656 if (s1 >= last)
657 d1 = s1 - last;
658 else if (s1 + back_max >= last)
659 d1 = (last - s1) * cfqd->cfq_back_penalty;
660 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200661 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
663 if (s2 >= last)
664 d2 = s2 - last;
665 else if (s2 + back_max >= last)
666 d2 = (last - s2) * cfqd->cfq_back_penalty;
667 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200668 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
670 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
Andreas Mohre8a99052006-03-28 08:59:49 +0200672 /*
673 * By doing switch() on the bit mask "wrap" we avoid having to
674 * check two variables for all permutations: --> faster!
675 */
676 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200677 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200678 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200679 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200680 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200681 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200682 else {
683 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200684 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200685 else
Jens Axboe5e705372006-07-13 12:39:25 +0200686 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200687 }
688
689 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200690 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200691 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200692 return rq2;
693 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200694 default:
695 /*
696 * Since both rqs are wrapped,
697 * start with the one that's further behind head
698 * (--> only *one* back seek required),
699 * since back seek takes more time than forward.
700 */
701 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200702 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 else
Jens Axboe5e705372006-07-13 12:39:25 +0200704 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 }
706}
707
Jens Axboe498d3aa22007-04-26 12:54:48 +0200708/*
709 * The below is leftmost cache rbtree addon
710 */
Jens Axboe08717142008-01-28 11:38:15 +0100711static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200712{
Vivek Goyal615f0252009-12-03 12:59:39 -0500713 /* Service tree is empty */
714 if (!root->count)
715 return NULL;
716
Jens Axboecc09e292007-04-26 12:53:50 +0200717 if (!root->left)
718 root->left = rb_first(&root->rb);
719
Jens Axboe08717142008-01-28 11:38:15 +0100720 if (root->left)
721 return rb_entry(root->left, struct cfq_queue, rb_node);
722
723 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200724}
725
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500726static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
727{
728 if (!root->left)
729 root->left = rb_first(&root->rb);
730
731 if (root->left)
732 return rb_entry_cfqg(root->left);
733
734 return NULL;
735}
736
Jens Axboea36e71f2009-04-15 12:15:11 +0200737static void rb_erase_init(struct rb_node *n, struct rb_root *root)
738{
739 rb_erase(n, root);
740 RB_CLEAR_NODE(n);
741}
742
Jens Axboecc09e292007-04-26 12:53:50 +0200743static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
744{
745 if (root->left == n)
746 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200747 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100748 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200749}
750
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751/*
752 * would be nice to take fifo expire time into account as well
753 */
Jens Axboe5e705372006-07-13 12:39:25 +0200754static struct request *
755cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
756 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757{
Jens Axboe21183b02006-07-13 12:33:14 +0200758 struct rb_node *rbnext = rb_next(&last->rb_node);
759 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200760 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761
Jens Axboe21183b02006-07-13 12:33:14 +0200762 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763
764 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200765 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200766
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200768 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200769 else {
770 rbnext = rb_first(&cfqq->sort_list);
771 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200772 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200773 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100775 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776}
777
Jens Axboed9e76202007-04-20 14:27:50 +0200778static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
779 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780{
Jens Axboed9e76202007-04-20 14:27:50 +0200781 /*
782 * just an approximation, should be ok.
783 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500784 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100785 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200786}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500788static inline s64
789cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
790{
791 return cfqg->vdisktime - st->min_vdisktime;
792}
793
794static void
795__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
796{
797 struct rb_node **node = &st->rb.rb_node;
798 struct rb_node *parent = NULL;
799 struct cfq_group *__cfqg;
800 s64 key = cfqg_key(st, cfqg);
801 int left = 1;
802
803 while (*node != NULL) {
804 parent = *node;
805 __cfqg = rb_entry_cfqg(parent);
806
807 if (key < cfqg_key(st, __cfqg))
808 node = &parent->rb_left;
809 else {
810 node = &parent->rb_right;
811 left = 0;
812 }
813 }
814
815 if (left)
816 st->left = &cfqg->rb_node;
817
818 rb_link_node(&cfqg->rb_node, parent, node);
819 rb_insert_color(&cfqg->rb_node, &st->rb);
820}
821
822static void
823cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
824{
825 struct cfq_rb_root *st = &cfqd->grp_service_tree;
826 struct cfq_group *__cfqg;
827 struct rb_node *n;
828
829 cfqg->nr_cfqq++;
830 if (cfqg->on_st)
831 return;
832
833 /*
834 * Currently put the group at the end. Later implement something
835 * so that groups get lesser vtime based on their weights, so that
836 * if group does not loose all if it was not continously backlogged.
837 */
838 n = rb_last(&st->rb);
839 if (n) {
840 __cfqg = rb_entry_cfqg(n);
841 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
842 } else
843 cfqg->vdisktime = st->min_vdisktime;
844
845 __cfq_group_service_tree_add(st, cfqg);
846 cfqg->on_st = true;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500847 cfqd->nr_groups++;
848 st->total_weight += cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500849}
850
851static void
852cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
853{
854 struct cfq_rb_root *st = &cfqd->grp_service_tree;
855
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500856 if (st->active == &cfqg->rb_node)
857 st->active = NULL;
858
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500859 BUG_ON(cfqg->nr_cfqq < 1);
860 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500861
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500862 /* If there are other cfq queues under this group, don't delete it */
863 if (cfqg->nr_cfqq)
864 return;
865
Vivek Goyal2868ef72009-12-03 12:59:48 -0500866 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500867 cfqg->on_st = false;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500868 cfqd->nr_groups--;
869 st->total_weight -= cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500870 if (!RB_EMPTY_NODE(&cfqg->rb_node))
871 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500872 cfqg->saved_workload_slice = 0;
Vivek Goyal22084192009-12-03 12:59:49 -0500873 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500874}
875
876static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
877{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500878 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500879
880 /*
881 * Queue got expired before even a single request completed or
882 * got expired immediately after first request completion.
883 */
884 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
885 /*
886 * Also charge the seek time incurred to the group, otherwise
887 * if there are mutiple queues in the group, each can dispatch
888 * a single request on seeky media and cause lots of seek time
889 * and group will never know it.
890 */
891 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
892 1);
893 } else {
894 slice_used = jiffies - cfqq->slice_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500895 if (slice_used > cfqq->allocated_slice)
896 slice_used = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500897 }
898
Vivek Goyal22084192009-12-03 12:59:49 -0500899 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
900 cfqq->nr_sectors);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500901 return slice_used;
902}
903
904static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
905 struct cfq_queue *cfqq)
906{
907 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500908 unsigned int used_sl, charge_sl;
909 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
910 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500911
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500912 BUG_ON(nr_sync < 0);
913 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
914
915 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
916 charge_sl = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500917
918 /* Can't update vdisktime while group is on service tree */
919 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500920 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500921 __cfq_group_service_tree_add(st, cfqg);
922
923 /* This group is being expired. Save the context */
924 if (time_after(cfqd->workload_expires, jiffies)) {
925 cfqg->saved_workload_slice = cfqd->workload_expires
926 - jiffies;
927 cfqg->saved_workload = cfqd->serving_type;
928 cfqg->saved_serving_prio = cfqd->serving_prio;
929 } else
930 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500931
932 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
933 st->min_vdisktime);
Vivek Goyal22084192009-12-03 12:59:49 -0500934 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
935 cfqq->nr_sectors);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500936}
937
Vivek Goyal25fb5162009-12-03 12:59:46 -0500938#ifdef CONFIG_CFQ_GROUP_IOSCHED
939static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
940{
941 if (blkg)
942 return container_of(blkg, struct cfq_group, blkg);
943 return NULL;
944}
945
Vivek Goyalf8d461d2009-12-03 12:59:52 -0500946void
947cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
948{
949 cfqg_of_blkg(blkg)->weight = weight;
950}
951
Vivek Goyal25fb5162009-12-03 12:59:46 -0500952static struct cfq_group *
953cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
954{
955 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
956 struct cfq_group *cfqg = NULL;
957 void *key = cfqd;
958 int i, j;
959 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -0500960 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
961 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500962
963 /* Do we need to take this reference */
Vivek Goyal9d6a9862009-12-04 10:36:41 -0500964 if (!blkiocg_css_tryget(blkcg))
Vivek Goyal25fb5162009-12-03 12:59:46 -0500965 return NULL;;
966
967 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
968 if (cfqg || !create)
969 goto done;
970
971 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
972 if (!cfqg)
973 goto done;
974
975 cfqg->weight = blkcg->weight;
976 for_each_cfqg_st(cfqg, i, j, st)
977 *st = CFQ_RB_ROOT;
978 RB_CLEAR_NODE(&cfqg->rb_node);
979
Vivek Goyalb1c35762009-12-03 12:59:47 -0500980 /*
981 * Take the initial reference that will be released on destroy
982 * This can be thought of a joint reference by cgroup and
983 * elevator which will be dropped by either elevator exit
984 * or cgroup deletion path depending on who is exiting first.
985 */
986 atomic_set(&cfqg->ref, 1);
987
Vivek Goyal25fb5162009-12-03 12:59:46 -0500988 /* Add group onto cgroup list */
Vivek Goyal22084192009-12-03 12:59:49 -0500989 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
990 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
991 MKDEV(major, minor));
Vivek Goyal25fb5162009-12-03 12:59:46 -0500992
993 /* Add group on cfqd list */
994 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
995
996done:
Vivek Goyal9d6a9862009-12-04 10:36:41 -0500997 blkiocg_css_put(blkcg);
Vivek Goyal25fb5162009-12-03 12:59:46 -0500998 return cfqg;
999}
1000
1001/*
1002 * Search for the cfq group current task belongs to. If create = 1, then also
1003 * create the cfq group if it does not exist. request_queue lock must be held.
1004 */
1005static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1006{
1007 struct cgroup *cgroup;
1008 struct cfq_group *cfqg = NULL;
1009
1010 rcu_read_lock();
1011 cgroup = task_cgroup(current, blkio_subsys_id);
1012 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1013 if (!cfqg && create)
1014 cfqg = &cfqd->root_group;
1015 rcu_read_unlock();
1016 return cfqg;
1017}
1018
1019static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1020{
1021 /* Currently, all async queues are mapped to root group */
1022 if (!cfq_cfqq_sync(cfqq))
1023 cfqg = &cfqq->cfqd->root_group;
1024
1025 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001026 /* cfqq reference on cfqg */
1027 atomic_inc(&cfqq->cfqg->ref);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001028}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001029
1030static void cfq_put_cfqg(struct cfq_group *cfqg)
1031{
1032 struct cfq_rb_root *st;
1033 int i, j;
1034
1035 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1036 if (!atomic_dec_and_test(&cfqg->ref))
1037 return;
1038 for_each_cfqg_st(cfqg, i, j, st)
1039 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1040 kfree(cfqg);
1041}
1042
1043static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1044{
1045 /* Something wrong if we are trying to remove same group twice */
1046 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1047
1048 hlist_del_init(&cfqg->cfqd_node);
1049
1050 /*
1051 * Put the reference taken at the time of creation so that when all
1052 * queues are gone, group can be destroyed.
1053 */
1054 cfq_put_cfqg(cfqg);
1055}
1056
1057static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1058{
1059 struct hlist_node *pos, *n;
1060 struct cfq_group *cfqg;
1061
1062 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1063 /*
1064 * If cgroup removal path got to blk_group first and removed
1065 * it from cgroup list, then it will take care of destroying
1066 * cfqg also.
1067 */
1068 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1069 cfq_destroy_cfqg(cfqd, cfqg);
1070 }
1071}
1072
1073/*
1074 * Blk cgroup controller notification saying that blkio_group object is being
1075 * delinked as associated cgroup object is going away. That also means that
1076 * no new IO will come in this group. So get rid of this group as soon as
1077 * any pending IO in the group is finished.
1078 *
1079 * This function is called under rcu_read_lock(). key is the rcu protected
1080 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1081 * read lock.
1082 *
1083 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1084 * it should not be NULL as even if elevator was exiting, cgroup deltion
1085 * path got to it first.
1086 */
1087void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1088{
1089 unsigned long flags;
1090 struct cfq_data *cfqd = key;
1091
1092 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1093 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1094 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1095}
1096
Vivek Goyal25fb5162009-12-03 12:59:46 -05001097#else /* GROUP_IOSCHED */
1098static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1099{
1100 return &cfqd->root_group;
1101}
1102static inline void
1103cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1104 cfqq->cfqg = cfqg;
1105}
1106
Vivek Goyalb1c35762009-12-03 12:59:47 -05001107static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1108static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1109
Vivek Goyal25fb5162009-12-03 12:59:46 -05001110#endif /* GROUP_IOSCHED */
1111
Jens Axboe498d3aa22007-04-26 12:54:48 +02001112/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001113 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001114 * requests waiting to be processed. It is sorted in the order that
1115 * we will service the queues.
1116 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001117static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001118 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001119{
Jens Axboe08717142008-01-28 11:38:15 +01001120 struct rb_node **p, *parent;
1121 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001122 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001123 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001124 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001125 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001126 int group_changed = 0;
1127
1128#ifdef CONFIG_CFQ_GROUP_IOSCHED
1129 if (!cfqd->cfq_group_isolation
1130 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1131 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1132 /* Move this cfq to root group */
1133 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1134 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1135 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1136 cfqq->orig_cfqg = cfqq->cfqg;
1137 cfqq->cfqg = &cfqd->root_group;
1138 atomic_inc(&cfqd->root_group.ref);
1139 group_changed = 1;
1140 } else if (!cfqd->cfq_group_isolation
1141 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1142 /* cfqq is sequential now needs to go to its original group */
1143 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1144 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1145 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1146 cfq_put_cfqg(cfqq->cfqg);
1147 cfqq->cfqg = cfqq->orig_cfqg;
1148 cfqq->orig_cfqg = NULL;
1149 group_changed = 1;
1150 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1151 }
1152#endif
Jens Axboed9e76202007-04-20 14:27:50 +02001153
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001154 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1155 cfqq_type(cfqq), cfqd);
Jens Axboe08717142008-01-28 11:38:15 +01001156 if (cfq_class_idle(cfqq)) {
1157 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001158 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001159 if (parent && parent != &cfqq->rb_node) {
1160 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1161 rb_key += __cfqq->rb_key;
1162 } else
1163 rb_key += jiffies;
1164 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001165 /*
1166 * Get our rb key offset. Subtract any residual slice
1167 * value carried from last service. A negative resid
1168 * count indicates slice overrun, and this should position
1169 * the next service time further away in the tree.
1170 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001171 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001172 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001173 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001174 } else {
1175 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001176 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001177 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1178 }
Jens Axboed9e76202007-04-20 14:27:50 +02001179
1180 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001181 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001182 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001183 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001184 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001185 if (rb_key == cfqq->rb_key &&
1186 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001187 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001188
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001189 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1190 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001191 }
Jens Axboed9e76202007-04-20 14:27:50 +02001192
Jens Axboe498d3aa22007-04-26 12:54:48 +02001193 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001194 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001195 cfqq->service_tree = service_tree;
1196 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001197 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001198 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001199
Jens Axboed9e76202007-04-20 14:27:50 +02001200 parent = *p;
1201 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1202
Jens Axboe0c534e02007-04-18 20:01:57 +02001203 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001204 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001205 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001206 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001207 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001208 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001209 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001210 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001211 }
Jens Axboe67060e32007-04-18 20:13:32 +02001212
1213 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001214 }
1215
Jens Axboecc09e292007-04-26 12:53:50 +02001216 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001217 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001218
Jens Axboed9e76202007-04-20 14:27:50 +02001219 cfqq->rb_key = rb_key;
1220 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001221 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1222 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001223 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001224 return;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001225 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226}
1227
Jens Axboea36e71f2009-04-15 12:15:11 +02001228static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001229cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1230 sector_t sector, struct rb_node **ret_parent,
1231 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001232{
Jens Axboea36e71f2009-04-15 12:15:11 +02001233 struct rb_node **p, *parent;
1234 struct cfq_queue *cfqq = NULL;
1235
1236 parent = NULL;
1237 p = &root->rb_node;
1238 while (*p) {
1239 struct rb_node **n;
1240
1241 parent = *p;
1242 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1243
1244 /*
1245 * Sort strictly based on sector. Smallest to the left,
1246 * largest to the right.
1247 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001248 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001249 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001250 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001251 n = &(*p)->rb_left;
1252 else
1253 break;
1254 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001255 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001256 }
1257
1258 *ret_parent = parent;
1259 if (rb_link)
1260 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001261 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001262}
1263
1264static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1265{
Jens Axboea36e71f2009-04-15 12:15:11 +02001266 struct rb_node **p, *parent;
1267 struct cfq_queue *__cfqq;
1268
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001269 if (cfqq->p_root) {
1270 rb_erase(&cfqq->p_node, cfqq->p_root);
1271 cfqq->p_root = NULL;
1272 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001273
1274 if (cfq_class_idle(cfqq))
1275 return;
1276 if (!cfqq->next_rq)
1277 return;
1278
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001279 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001280 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1281 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001282 if (!__cfqq) {
1283 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001284 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1285 } else
1286 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001287}
1288
Jens Axboe498d3aa22007-04-26 12:54:48 +02001289/*
1290 * Update cfqq's position in the service tree.
1291 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001292static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001293{
Jens Axboe6d048f52007-04-25 12:44:27 +02001294 /*
1295 * Resorting requires the cfqq to be on the RR list already.
1296 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001297 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001298 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001299 cfq_prio_tree_add(cfqd, cfqq);
1300 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001301}
1302
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303/*
1304 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001305 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306 */
Jens Axboefebffd62008-01-28 13:19:43 +01001307static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308{
Jens Axboe7b679132008-05-30 12:23:07 +02001309 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001310 BUG_ON(cfq_cfqq_on_rr(cfqq));
1311 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 cfqd->busy_queues++;
1313
Jens Axboeedd75ff2007-04-19 12:03:34 +02001314 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315}
1316
Jens Axboe498d3aa22007-04-26 12:54:48 +02001317/*
1318 * Called when the cfqq no longer has requests pending, remove it from
1319 * the service tree.
1320 */
Jens Axboefebffd62008-01-28 13:19:43 +01001321static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322{
Jens Axboe7b679132008-05-30 12:23:07 +02001323 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001324 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1325 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001327 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1328 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1329 cfqq->service_tree = NULL;
1330 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001331 if (cfqq->p_root) {
1332 rb_erase(&cfqq->p_node, cfqq->p_root);
1333 cfqq->p_root = NULL;
1334 }
Jens Axboed9e76202007-04-20 14:27:50 +02001335
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001336 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 BUG_ON(!cfqd->busy_queues);
1338 cfqd->busy_queues--;
1339}
1340
1341/*
1342 * rb tree support functions
1343 */
Jens Axboefebffd62008-01-28 13:19:43 +01001344static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345{
Jens Axboe5e705372006-07-13 12:39:25 +02001346 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001347 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
Jens Axboeb4878f22005-10-20 16:42:29 +02001349 BUG_ON(!cfqq->queued[sync]);
1350 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351
Jens Axboe5e705372006-07-13 12:39:25 +02001352 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353
Vivek Goyalf04a6422009-12-03 12:59:40 -05001354 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1355 /*
1356 * Queue will be deleted from service tree when we actually
1357 * expire it later. Right now just remove it from prio tree
1358 * as it is empty.
1359 */
1360 if (cfqq->p_root) {
1361 rb_erase(&cfqq->p_node, cfqq->p_root);
1362 cfqq->p_root = NULL;
1363 }
1364 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365}
1366
Jens Axboe5e705372006-07-13 12:39:25 +02001367static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368{
Jens Axboe5e705372006-07-13 12:39:25 +02001369 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001371 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
Jens Axboe5380a102006-07-13 12:37:56 +02001373 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374
1375 /*
1376 * looks a little odd, but the first insert might return an alias.
1377 * if that happens, put the alias on the dispatch list
1378 */
Jens Axboe21183b02006-07-13 12:33:14 +02001379 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001380 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001381
1382 if (!cfq_cfqq_on_rr(cfqq))
1383 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001384
1385 /*
1386 * check if this request is a better next-serve candidate
1387 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001388 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001389 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001390
1391 /*
1392 * adjust priority tree position, if ->next_rq changes
1393 */
1394 if (prev != cfqq->next_rq)
1395 cfq_prio_tree_add(cfqd, cfqq);
1396
Jens Axboe5044eed2007-04-25 11:53:48 +02001397 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398}
1399
Jens Axboefebffd62008-01-28 13:19:43 +01001400static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401{
Jens Axboe5380a102006-07-13 12:37:56 +02001402 elv_rb_del(&cfqq->sort_list, rq);
1403 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +02001404 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405}
1406
Jens Axboe206dc692006-03-28 13:03:44 +02001407static struct request *
1408cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409{
Jens Axboe206dc692006-03-28 13:03:44 +02001410 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001411 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001412 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413
Jens Axboe4ac845a2008-01-24 08:44:49 +01001414 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001415 if (!cic)
1416 return NULL;
1417
1418 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001419 if (cfqq) {
1420 sector_t sector = bio->bi_sector + bio_sectors(bio);
1421
Jens Axboe21183b02006-07-13 12:33:14 +02001422 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001423 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425 return NULL;
1426}
1427
Jens Axboe165125e2007-07-24 09:28:11 +02001428static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001429{
1430 struct cfq_data *cfqd = q->elevator->elevator_data;
1431
Jens Axboe5ad531d2009-07-03 12:57:48 +02001432 cfqd->rq_in_driver[rq_is_sync(rq)]++;
Jens Axboe7b679132008-05-30 12:23:07 +02001433 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +02001434 rq_in_driver(cfqd));
Jens Axboe25776e32006-06-01 10:12:26 +02001435
Tejun Heo5b936292009-05-07 22:24:38 +09001436 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001437}
1438
Jens Axboe165125e2007-07-24 09:28:11 +02001439static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440{
Jens Axboe22e2c502005-06-27 10:55:12 +02001441 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5ad531d2009-07-03 12:57:48 +02001442 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443
Jens Axboe5ad531d2009-07-03 12:57:48 +02001444 WARN_ON(!cfqd->rq_in_driver[sync]);
1445 cfqd->rq_in_driver[sync]--;
Jens Axboe7b679132008-05-30 12:23:07 +02001446 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +02001447 rq_in_driver(cfqd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448}
1449
Jens Axboeb4878f22005-10-20 16:42:29 +02001450static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451{
Jens Axboe5e705372006-07-13 12:39:25 +02001452 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001453
Jens Axboe5e705372006-07-13 12:39:25 +02001454 if (cfqq->next_rq == rq)
1455 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456
Jens Axboeb4878f22005-10-20 16:42:29 +02001457 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001458 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001459
Aaron Carroll45333d52008-08-26 15:52:36 +02001460 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +02001461 if (rq_is_meta(rq)) {
1462 WARN_ON(!cfqq->meta_pending);
1463 cfqq->meta_pending--;
1464 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465}
1466
Jens Axboe165125e2007-07-24 09:28:11 +02001467static int cfq_merge(struct request_queue *q, struct request **req,
1468 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469{
1470 struct cfq_data *cfqd = q->elevator->elevator_data;
1471 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472
Jens Axboe206dc692006-03-28 13:03:44 +02001473 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001474 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001475 *req = __rq;
1476 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 }
1478
1479 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480}
1481
Jens Axboe165125e2007-07-24 09:28:11 +02001482static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001483 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484{
Jens Axboe21183b02006-07-13 12:33:14 +02001485 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001486 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487
Jens Axboe5e705372006-07-13 12:39:25 +02001488 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490}
1491
1492static void
Jens Axboe165125e2007-07-24 09:28:11 +02001493cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 struct request *next)
1495{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001496 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001497 /*
1498 * reposition in fifo if next is older than rq
1499 */
1500 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001501 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001502 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001503 rq_set_fifo_time(rq, rq_fifo_time(next));
1504 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001505
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001506 if (cfqq->next_rq == next)
1507 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001508 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +02001509}
1510
Jens Axboe165125e2007-07-24 09:28:11 +02001511static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001512 struct bio *bio)
1513{
1514 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001515 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001516 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001517
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001518 /* Deny merge if bio and rq don't belong to same cfq group */
1519 if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0))
1520 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001521 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001522 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001523 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001524 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001525 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001526
1527 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001528 * Lookup the cfqq that this bio will be queued with. Allow
1529 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001530 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001531 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001532 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001533 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001534
Vasily Tarasov91fac312007-04-25 12:29:51 +02001535 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001536 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001537}
1538
Jens Axboefebffd62008-01-28 13:19:43 +01001539static void __cfq_set_active_queue(struct cfq_data *cfqd,
1540 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001541{
1542 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +02001543 cfq_log_cfqq(cfqd, cfqq, "set_active");
Vivek Goyaldae739e2009-12-03 12:59:45 -05001544 cfqq->slice_start = 0;
1545 cfqq->dispatch_start = jiffies;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001546 cfqq->allocated_slice = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001547 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001548 cfqq->slice_dispatch = 0;
Vivek Goyal22084192009-12-03 12:59:49 -05001549 cfqq->nr_sectors = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001550
Jens Axboe2f5cb732009-04-07 08:51:19 +02001551 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +02001552 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001553 cfq_clear_cfqq_must_alloc_slice(cfqq);
1554 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +11001555 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001556
1557 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +02001558 }
1559
1560 cfqd->active_queue = cfqq;
1561}
1562
1563/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001564 * current cfqq expired its slice (or was too idle), select new one
1565 */
1566static void
1567__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001568 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001569{
Jens Axboe7b679132008-05-30 12:23:07 +02001570 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1571
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001572 if (cfq_cfqq_wait_request(cfqq))
1573 del_timer(&cfqd->idle_slice_timer);
1574
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001575 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001576 cfq_clear_cfqq_wait_busy(cfqq);
1577 cfq_clear_cfqq_wait_busy_done(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001578
1579 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001580 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001581 */
Jens Axboe7b679132008-05-30 12:23:07 +02001582 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +11001583 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001584 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1585 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001586
Vivek Goyaldae739e2009-12-03 12:59:45 -05001587 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1588
Vivek Goyalf04a6422009-12-03 12:59:40 -05001589 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1590 cfq_del_cfqq_rr(cfqd, cfqq);
1591
Jens Axboeedd75ff2007-04-19 12:03:34 +02001592 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001593
1594 if (cfqq == cfqd->active_queue)
1595 cfqd->active_queue = NULL;
1596
Vivek Goyaldae739e2009-12-03 12:59:45 -05001597 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1598 cfqd->grp_service_tree.active = NULL;
1599
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001600 if (cfqd->active_cic) {
1601 put_io_context(cfqd->active_cic->ioc);
1602 cfqd->active_cic = NULL;
1603 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001604}
1605
Jens Axboea6151c32009-10-07 20:02:57 +02001606static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001607{
1608 struct cfq_queue *cfqq = cfqd->active_queue;
1609
1610 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +02001611 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001612}
1613
Jens Axboe498d3aa22007-04-26 12:54:48 +02001614/*
1615 * Get next queue for service. Unless we have a queue preemption,
1616 * we'll simply select the first cfqq in the service tree.
1617 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001618static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001619{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001620 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001621 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1622 cfqd->serving_type, cfqd);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001623
Vivek Goyalf04a6422009-12-03 12:59:40 -05001624 if (!cfqd->rq_queued)
1625 return NULL;
1626
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001627 /* There is nothing to dispatch */
1628 if (!service_tree)
1629 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001630 if (RB_EMPTY_ROOT(&service_tree->rb))
1631 return NULL;
1632 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001633}
1634
Vivek Goyalf04a6422009-12-03 12:59:40 -05001635static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1636{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001637 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001638 struct cfq_queue *cfqq;
1639 int i, j;
1640 struct cfq_rb_root *st;
1641
1642 if (!cfqd->rq_queued)
1643 return NULL;
1644
Vivek Goyal25fb5162009-12-03 12:59:46 -05001645 cfqg = cfq_get_next_cfqg(cfqd);
1646 if (!cfqg)
1647 return NULL;
1648
Vivek Goyalf04a6422009-12-03 12:59:40 -05001649 for_each_cfqg_st(cfqg, i, j, st)
1650 if ((cfqq = cfq_rb_first(st)) != NULL)
1651 return cfqq;
1652 return NULL;
1653}
1654
Jens Axboe498d3aa22007-04-26 12:54:48 +02001655/*
1656 * Get and set a new active queue for service.
1657 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001658static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1659 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001660{
Jens Axboee00ef792009-11-04 08:54:55 +01001661 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001662 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001663
Jens Axboe22e2c502005-06-27 10:55:12 +02001664 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001665 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001666}
1667
Jens Axboed9e76202007-04-20 14:27:50 +02001668static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1669 struct request *rq)
1670{
Tejun Heo83096eb2009-05-07 22:24:39 +09001671 if (blk_rq_pos(rq) >= cfqd->last_position)
1672 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001673 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001674 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001675}
1676
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001677#define CFQQ_SEEK_THR 8 * 1024
1678#define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
Jeff Moyer04dc6e72009-04-21 07:31:56 +02001679
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001680static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1681 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001682{
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001683 sector_t sdist = cfqq->seek_mean;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001684
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001685 if (!sample_valid(cfqq->seek_samples))
1686 sdist = CFQQ_SEEK_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001687
Jeff Moyer04dc6e72009-04-21 07:31:56 +02001688 return cfq_dist_from_last(cfqd, rq) <= sdist;
Jens Axboe6d048f52007-04-25 12:44:27 +02001689}
1690
Jens Axboea36e71f2009-04-15 12:15:11 +02001691static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1692 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001693{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001694 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001695 struct rb_node *parent, *node;
1696 struct cfq_queue *__cfqq;
1697 sector_t sector = cfqd->last_position;
1698
1699 if (RB_EMPTY_ROOT(root))
1700 return NULL;
1701
1702 /*
1703 * First, if we find a request starting at the end of the last
1704 * request, choose it.
1705 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001706 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001707 if (__cfqq)
1708 return __cfqq;
1709
1710 /*
1711 * If the exact sector wasn't found, the parent of the NULL leaf
1712 * will contain the closest sector.
1713 */
1714 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001715 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001716 return __cfqq;
1717
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001718 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001719 node = rb_next(&__cfqq->p_node);
1720 else
1721 node = rb_prev(&__cfqq->p_node);
1722 if (!node)
1723 return NULL;
1724
1725 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001726 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001727 return __cfqq;
1728
1729 return NULL;
1730}
1731
1732/*
1733 * cfqd - obvious
1734 * cur_cfqq - passed in so that we don't decide that the current queue is
1735 * closely cooperating with itself.
1736 *
1737 * So, basically we're assuming that that cur_cfqq has dispatched at least
1738 * one request, and that cfqd->last_position reflects a position on the disk
1739 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1740 * assumption.
1741 */
1742static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001743 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001744{
1745 struct cfq_queue *cfqq;
1746
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001747 if (!cfq_cfqq_sync(cur_cfqq))
1748 return NULL;
1749 if (CFQQ_SEEKY(cur_cfqq))
1750 return NULL;
1751
Jens Axboea36e71f2009-04-15 12:15:11 +02001752 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001753 * We should notice if some of the queues are cooperating, eg
1754 * working closely on the same area of the disk. In that case,
1755 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001756 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001757 cfqq = cfqq_close(cfqd, cur_cfqq);
1758 if (!cfqq)
1759 return NULL;
1760
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001761 /* If new queue belongs to different cfq_group, don't choose it */
1762 if (cur_cfqq->cfqg != cfqq->cfqg)
1763 return NULL;
1764
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001765 /*
1766 * It only makes sense to merge sync queues.
1767 */
1768 if (!cfq_cfqq_sync(cfqq))
1769 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001770 if (CFQQ_SEEKY(cfqq))
1771 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001772
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001773 /*
1774 * Do not merge queues of different priority classes
1775 */
1776 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1777 return NULL;
1778
Jens Axboea36e71f2009-04-15 12:15:11 +02001779 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001780}
1781
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001782/*
1783 * Determine whether we should enforce idle window for this queue.
1784 */
1785
1786static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1787{
1788 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001789 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001790
Vivek Goyalf04a6422009-12-03 12:59:40 -05001791 BUG_ON(!service_tree);
1792 BUG_ON(!service_tree->count);
1793
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001794 /* We never do for idle class queues. */
1795 if (prio == IDLE_WORKLOAD)
1796 return false;
1797
1798 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001799 if (cfq_cfqq_idle_window(cfqq) &&
1800 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001801 return true;
1802
1803 /*
1804 * Otherwise, we do only if they are the last ones
1805 * in their service tree.
1806 */
Vivek Goyalf04a6422009-12-03 12:59:40 -05001807 return service_tree->count == 1;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001808}
1809
Jens Axboe6d048f52007-04-25 12:44:27 +02001810static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001811{
Jens Axboe17926692007-01-19 11:59:30 +11001812 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001813 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001814 unsigned long sl;
1815
Jens Axboea68bbdd2008-09-24 13:03:33 +02001816 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001817 * SSD device without seek penalty, disable idling. But only do so
1818 * for devices that support queuing, otherwise we still have a problem
1819 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001820 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001821 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001822 return;
1823
Jens Axboedd67d052006-06-21 09:36:18 +02001824 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001825 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001826
1827 /*
1828 * idle is disabled, either manually or by past process history
1829 */
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001830 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
Jens Axboe6d048f52007-04-25 12:44:27 +02001831 return;
1832
Jens Axboe22e2c502005-06-27 10:55:12 +02001833 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001834 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001835 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001836 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001837 return;
1838
1839 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001840 * task has exited, don't wait
1841 */
Jens Axboe206dc692006-03-28 13:03:44 +02001842 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001843 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001844 return;
1845
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001846 /*
1847 * If our average think time is larger than the remaining time
1848 * slice, then don't idle. This avoids overrunning the allotted
1849 * time slice.
1850 */
1851 if (sample_valid(cic->ttime_samples) &&
1852 (cfqq->slice_end - jiffies < cic->ttime_mean))
1853 return;
1854
Jens Axboe3b181522005-06-27 10:56:24 +02001855 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001856
Jens Axboe6d048f52007-04-25 12:44:27 +02001857 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001858
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001859 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001860 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861}
1862
Jens Axboe498d3aa22007-04-26 12:54:48 +02001863/*
1864 * Move request from internal lists to the request queue dispatch list.
1865 */
Jens Axboe165125e2007-07-24 09:28:11 +02001866static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001868 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001869 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001870
Jens Axboe7b679132008-05-30 12:23:07 +02001871 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1872
Jeff Moyer06d21882009-09-11 17:08:59 +02001873 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001874 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001875 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001876 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001877
1878 if (cfq_cfqq_sync(cfqq))
1879 cfqd->sync_flight++;
Vivek Goyal22084192009-12-03 12:59:49 -05001880 cfqq->nr_sectors += blk_rq_sectors(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881}
1882
1883/*
1884 * return expired entry, or NULL to just start from scratch in rbtree
1885 */
Jens Axboefebffd62008-01-28 13:19:43 +01001886static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887{
Jens Axboe30996f42009-10-05 11:03:39 +02001888 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889
Jens Axboe3b181522005-06-27 10:56:24 +02001890 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001892
1893 cfq_mark_cfqq_fifo_expire(cfqq);
1894
Jens Axboe89850f72006-07-22 16:48:31 +02001895 if (list_empty(&cfqq->fifo))
1896 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897
Jens Axboe89850f72006-07-22 16:48:31 +02001898 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02001899 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02001900 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901
Jens Axboe30996f42009-10-05 11:03:39 +02001902 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001903 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904}
1905
Jens Axboe22e2c502005-06-27 10:55:12 +02001906static inline int
1907cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1908{
1909 const int base_rq = cfqd->cfq_slice_async_rq;
1910
1911 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1912
1913 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1914}
1915
1916/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001917 * Must be called with the queue_lock held.
1918 */
1919static int cfqq_process_refs(struct cfq_queue *cfqq)
1920{
1921 int process_refs, io_refs;
1922
1923 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1924 process_refs = atomic_read(&cfqq->ref) - io_refs;
1925 BUG_ON(process_refs < 0);
1926 return process_refs;
1927}
1928
1929static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1930{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001931 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001932 struct cfq_queue *__cfqq;
1933
1934 /* Avoid a circular list and skip interim queue merges */
1935 while ((__cfqq = new_cfqq->new_cfqq)) {
1936 if (__cfqq == cfqq)
1937 return;
1938 new_cfqq = __cfqq;
1939 }
1940
1941 process_refs = cfqq_process_refs(cfqq);
1942 /*
1943 * If the process for the cfqq has gone away, there is no
1944 * sense in merging the queues.
1945 */
1946 if (process_refs == 0)
1947 return;
1948
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001949 /*
1950 * Merge in the direction of the lesser amount of work.
1951 */
1952 new_process_refs = cfqq_process_refs(new_cfqq);
1953 if (new_process_refs >= process_refs) {
1954 cfqq->new_cfqq = new_cfqq;
1955 atomic_add(process_refs, &new_cfqq->ref);
1956 } else {
1957 new_cfqq->new_cfqq = cfqq;
1958 atomic_add(new_process_refs, &cfqq->ref);
1959 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001960}
1961
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001962static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1963 struct cfq_group *cfqg, enum wl_prio_t prio,
1964 bool prio_changed)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001965{
1966 struct cfq_queue *queue;
1967 int i;
1968 bool key_valid = false;
1969 unsigned long lowest_key = 0;
1970 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1971
1972 if (prio_changed) {
1973 /*
1974 * When priorities switched, we prefer starting
1975 * from SYNC_NOIDLE (first choice), or just SYNC
1976 * over ASYNC
1977 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001978 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001979 return cur_best;
1980 cur_best = SYNC_WORKLOAD;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001981 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001982 return cur_best;
1983
1984 return ASYNC_WORKLOAD;
1985 }
1986
1987 for (i = 0; i < 3; ++i) {
1988 /* otherwise, select the one with lowest rb_key */
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001989 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001990 if (queue &&
1991 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1992 lowest_key = queue->rb_key;
1993 cur_best = i;
1994 key_valid = true;
1995 }
1996 }
1997
1998 return cur_best;
1999}
2000
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002001static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002002{
2003 enum wl_prio_t previous_prio = cfqd->serving_prio;
2004 bool prio_changed;
2005 unsigned slice;
2006 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002007 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002008 unsigned group_slice;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002009
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002010 if (!cfqg) {
2011 cfqd->serving_prio = IDLE_WORKLOAD;
2012 cfqd->workload_expires = jiffies + 1;
2013 return;
2014 }
2015
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002016 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002017 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002018 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002019 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002020 cfqd->serving_prio = BE_WORKLOAD;
2021 else {
2022 cfqd->serving_prio = IDLE_WORKLOAD;
2023 cfqd->workload_expires = jiffies + 1;
2024 return;
2025 }
2026
2027 /*
2028 * For RT and BE, we have to choose also the type
2029 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2030 * expiration time
2031 */
2032 prio_changed = (cfqd->serving_prio != previous_prio);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002033 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2034 cfqd);
2035 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002036
2037 /*
2038 * If priority didn't change, check workload expiration,
2039 * and that we still have other queues ready
2040 */
2041 if (!prio_changed && count &&
2042 !time_after(jiffies, cfqd->workload_expires))
2043 return;
2044
2045 /* otherwise select new workload type */
2046 cfqd->serving_type =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002047 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
2048 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2049 cfqd);
2050 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002051
2052 /*
2053 * the workload slice is computed as a fraction of target latency
2054 * proportional to the number of queues in that workload, over
2055 * all the queues in the same priority class
2056 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002057 group_slice = cfq_group_slice(cfqd, cfqg);
2058
2059 slice = group_slice * count /
2060 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2061 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002062
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002063 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2064 unsigned int tmp;
2065
2066 /*
2067 * Async queues are currently system wide. Just taking
2068 * proportion of queues with-in same group will lead to higher
2069 * async ratio system wide as generally root group is going
2070 * to have higher weight. A more accurate thing would be to
2071 * calculate system wide asnc/sync ratio.
2072 */
2073 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2074 tmp = tmp/cfqd->busy_queues;
2075 slice = min_t(unsigned, slice, tmp);
2076
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002077 /* async workload slice is scaled down according to
2078 * the sync/async slice ratio. */
2079 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002080 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002081 /* sync workload slice is at least 2 * cfq_slice_idle */
2082 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2083
2084 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2085 cfqd->workload_expires = jiffies + slice;
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002086 cfqd->noidle_tree_requires_idle = false;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002087}
2088
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002089static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2090{
2091 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002092 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002093
2094 if (RB_EMPTY_ROOT(&st->rb))
2095 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002096 cfqg = cfq_rb_first_group(st);
2097 st->active = &cfqg->rb_node;
2098 update_min_vdisktime(st);
2099 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002100}
2101
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002102static void cfq_choose_cfqg(struct cfq_data *cfqd)
2103{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002104 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2105
2106 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002107
2108 /* Restore the workload type data */
2109 if (cfqg->saved_workload_slice) {
2110 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2111 cfqd->serving_type = cfqg->saved_workload;
2112 cfqd->serving_prio = cfqg->saved_serving_prio;
2113 }
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002114 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002115}
2116
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002117/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002118 * Select a queue for service. If we have a current active queue,
2119 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002120 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002121static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002122{
Jens Axboea36e71f2009-04-15 12:15:11 +02002123 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002124
2125 cfqq = cfqd->active_queue;
2126 if (!cfqq)
2127 goto new_queue;
2128
Vivek Goyalf04a6422009-12-03 12:59:40 -05002129 if (!cfqd->rq_queued)
2130 return NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002131 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002132 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002133 */
Vivek Goyalf75edf22009-12-03 12:59:53 -05002134 if ((cfq_slice_used(cfqq) || cfq_cfqq_wait_busy_done(cfqq))
2135 && !cfq_cfqq_must_dispatch(cfqq))
Jens Axboe3b181522005-06-27 10:56:24 +02002136 goto expire;
Jens Axboe22e2c502005-06-27 10:55:12 +02002137
2138 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002139 * The active queue has requests and isn't expired, allow it to
2140 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002141 */
Jens Axboedd67d052006-06-21 09:36:18 +02002142 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002143 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002144
2145 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002146 * If another queue has a request waiting within our mean seek
2147 * distance, let it run. The expire code will check for close
2148 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002149 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002150 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002151 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002152 if (new_cfqq) {
2153 if (!cfqq->new_cfqq)
2154 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002155 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002156 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002157
2158 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002159 * No requests pending. If the active queue still has requests in
2160 * flight or is idling for a new request, allow either of these
2161 * conditions to happen (or time out) before selecting a new queue.
2162 */
Jens Axboecc197472007-04-20 20:45:39 +02002163 if (timer_pending(&cfqd->idle_slice_timer) ||
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002164 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002165 cfqq = NULL;
2166 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002167 }
2168
Jens Axboe3b181522005-06-27 10:56:24 +02002169expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002170 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002171new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002172 /*
2173 * Current queue expired. Check if we have to switch to a new
2174 * service tree
2175 */
2176 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002177 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002178
Jens Axboea36e71f2009-04-15 12:15:11 +02002179 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002180keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002181 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002182}
2183
Jens Axboefebffd62008-01-28 13:19:43 +01002184static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002185{
2186 int dispatched = 0;
2187
2188 while (cfqq->next_rq) {
2189 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2190 dispatched++;
2191 }
2192
2193 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002194
2195 /* By default cfqq is not expired if it is empty. Do it explicitly */
2196 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002197 return dispatched;
2198}
2199
Jens Axboe498d3aa22007-04-26 12:54:48 +02002200/*
2201 * Drain our current requests. Used for barriers and when switching
2202 * io schedulers on-the-fly.
2203 */
Jens Axboed9e76202007-04-20 14:27:50 +02002204static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002205{
Jens Axboe08717142008-01-28 11:38:15 +01002206 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002207 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002208
Vivek Goyalf04a6422009-12-03 12:59:40 -05002209 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2210 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002211
Jens Axboe6084cdd2007-04-23 08:25:00 +02002212 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002213 BUG_ON(cfqd->busy_queues);
2214
Jeff Moyer69237152009-06-12 15:29:30 +02002215 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002216 return dispatched;
2217}
2218
Jens Axboe0b182d62009-10-06 20:49:37 +02002219static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002220{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002221 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222
Jens Axboe2f5cb732009-04-07 08:51:19 +02002223 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002224 * Drain async requests before we start sync IO
2225 */
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002226 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002227 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002228
2229 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002230 * If this is an async queue and we have sync IO in flight, let it wait
2231 */
2232 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002233 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002234
2235 max_dispatch = cfqd->cfq_quantum;
2236 if (cfq_class_idle(cfqq))
2237 max_dispatch = 1;
2238
2239 /*
2240 * Does this cfqq already have too much IO in flight?
2241 */
2242 if (cfqq->dispatched >= max_dispatch) {
2243 /*
2244 * idle queue must always only have a single IO in flight
2245 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002246 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002247 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002248
Jens Axboe2f5cb732009-04-07 08:51:19 +02002249 /*
2250 * We have other queues, don't allow more IO from this one
2251 */
2252 if (cfqd->busy_queues > 1)
Jens Axboe0b182d62009-10-06 20:49:37 +02002253 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002254
Jens Axboe2f5cb732009-04-07 08:51:19 +02002255 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002256 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002257 */
Shaohua Li474b18c2009-12-03 12:58:05 +01002258 max_dispatch = -1;
Jens Axboe8e296752009-10-03 16:26:03 +02002259 }
2260
2261 /*
2262 * Async queues must wait a bit before being allowed dispatch.
2263 * We also ramp up the dispatch depth gradually for async IO,
2264 * based on the last sync IO we serviced
2265 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002266 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Jens Axboe8e296752009-10-03 16:26:03 +02002267 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
2268 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002269
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002270 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002271 if (!depth && !cfqq->dispatched)
2272 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002273 if (depth < max_dispatch)
2274 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 }
2276
Jens Axboe0b182d62009-10-06 20:49:37 +02002277 /*
2278 * If we're below the current max, allow a dispatch
2279 */
2280 return cfqq->dispatched < max_dispatch;
2281}
2282
2283/*
2284 * Dispatch a request from cfqq, moving them to the request queue
2285 * dispatch list.
2286 */
2287static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2288{
2289 struct request *rq;
2290
2291 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2292
2293 if (!cfq_may_dispatch(cfqd, cfqq))
2294 return false;
2295
2296 /*
2297 * follow expired path, else get first next available
2298 */
2299 rq = cfq_check_fifo(cfqq);
2300 if (!rq)
2301 rq = cfqq->next_rq;
2302
2303 /*
2304 * insert request into driver dispatch list
2305 */
2306 cfq_dispatch_insert(cfqd->queue, rq);
2307
2308 if (!cfqd->active_cic) {
2309 struct cfq_io_context *cic = RQ_CIC(rq);
2310
2311 atomic_long_inc(&cic->ioc->refcount);
2312 cfqd->active_cic = cic;
2313 }
2314
2315 return true;
2316}
2317
2318/*
2319 * Find the cfqq that we need to service and move a request from that to the
2320 * dispatch list
2321 */
2322static int cfq_dispatch_requests(struct request_queue *q, int force)
2323{
2324 struct cfq_data *cfqd = q->elevator->elevator_data;
2325 struct cfq_queue *cfqq;
2326
2327 if (!cfqd->busy_queues)
2328 return 0;
2329
2330 if (unlikely(force))
2331 return cfq_forced_dispatch(cfqd);
2332
2333 cfqq = cfq_select_queue(cfqd);
2334 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002335 return 0;
2336
Jens Axboe2f5cb732009-04-07 08:51:19 +02002337 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002338 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002339 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002340 if (!cfq_dispatch_request(cfqd, cfqq))
2341 return 0;
2342
Jens Axboe2f5cb732009-04-07 08:51:19 +02002343 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002344 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002345
2346 /*
2347 * expire an async queue immediately if it has used up its slice. idle
2348 * queue always expire after 1 dispatch round.
2349 */
2350 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2351 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2352 cfq_class_idle(cfqq))) {
2353 cfqq->slice_end = jiffies + 1;
2354 cfq_slice_expired(cfqd, 0);
2355 }
2356
Shan Weib217a902009-09-01 10:06:42 +02002357 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002358 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359}
2360
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361/*
Jens Axboe5e705372006-07-13 12:39:25 +02002362 * task holds one reference to the queue, dropped when task exits. each rq
2363 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002365 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 * queue lock must be held here.
2367 */
2368static void cfq_put_queue(struct cfq_queue *cfqq)
2369{
Jens Axboe22e2c502005-06-27 10:55:12 +02002370 struct cfq_data *cfqd = cfqq->cfqd;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002371 struct cfq_group *cfqg, *orig_cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002372
2373 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374
2375 if (!atomic_dec_and_test(&cfqq->ref))
2376 return;
2377
Jens Axboe7b679132008-05-30 12:23:07 +02002378 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002380 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002381 cfqg = cfqq->cfqg;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002382 orig_cfqg = cfqq->orig_cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002384 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02002385 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002386 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002387 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002388
Vivek Goyalf04a6422009-12-03 12:59:40 -05002389 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002391 cfq_put_cfqg(cfqg);
Vivek Goyal878eadd2009-12-07 19:37:15 +01002392 if (orig_cfqg)
2393 cfq_put_cfqg(orig_cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394}
2395
Jens Axboed6de8be2008-05-28 14:46:59 +02002396/*
2397 * Must always be called with the rcu_read_lock() held
2398 */
Jens Axboe07416d22008-05-07 09:17:12 +02002399static void
2400__call_for_each_cic(struct io_context *ioc,
2401 void (*func)(struct io_context *, struct cfq_io_context *))
2402{
2403 struct cfq_io_context *cic;
2404 struct hlist_node *n;
2405
2406 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2407 func(ioc, cic);
2408}
2409
Jens Axboe4ac845a2008-01-24 08:44:49 +01002410/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002411 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002412 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002413static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002414call_for_each_cic(struct io_context *ioc,
2415 void (*func)(struct io_context *, struct cfq_io_context *))
2416{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002417 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002418 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002419 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002420}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002421
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002422static void cfq_cic_free_rcu(struct rcu_head *head)
2423{
2424 struct cfq_io_context *cic;
2425
2426 cic = container_of(head, struct cfq_io_context, rcu_head);
2427
2428 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002429 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002430
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002431 if (ioc_gone) {
2432 /*
2433 * CFQ scheduler is exiting, grab exit lock and check
2434 * the pending io context count. If it hits zero,
2435 * complete ioc_gone and set it back to NULL
2436 */
2437 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002438 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002439 complete(ioc_gone);
2440 ioc_gone = NULL;
2441 }
2442 spin_unlock(&ioc_gone_lock);
2443 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002444}
2445
2446static void cfq_cic_free(struct cfq_io_context *cic)
2447{
2448 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002449}
2450
2451static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2452{
2453 unsigned long flags;
2454
2455 BUG_ON(!cic->dead_key);
2456
2457 spin_lock_irqsave(&ioc->lock, flags);
2458 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01002459 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002460 spin_unlock_irqrestore(&ioc->lock, flags);
2461
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002462 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002463}
2464
Jens Axboed6de8be2008-05-28 14:46:59 +02002465/*
2466 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2467 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2468 * and ->trim() which is called with the task lock held
2469 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002470static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002472 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002473 * ioc->refcount is zero here, or we are called from elv_unregister(),
2474 * so no more cic's are allowed to be linked into this ioc. So it
2475 * should be ok to iterate over the known list, we will see all cic's
2476 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002477 */
Jens Axboe07416d22008-05-07 09:17:12 +02002478 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479}
2480
Jens Axboe89850f72006-07-22 16:48:31 +02002481static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2482{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002483 struct cfq_queue *__cfqq, *next;
2484
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002485 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02002486 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002487 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002488 }
Jens Axboe89850f72006-07-22 16:48:31 +02002489
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002490 /*
2491 * If this queue was scheduled to merge with another queue, be
2492 * sure to drop the reference taken on that queue (and others in
2493 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2494 */
2495 __cfqq = cfqq->new_cfqq;
2496 while (__cfqq) {
2497 if (__cfqq == cfqq) {
2498 WARN(1, "cfqq->new_cfqq loop detected\n");
2499 break;
2500 }
2501 next = __cfqq->new_cfqq;
2502 cfq_put_queue(__cfqq);
2503 __cfqq = next;
2504 }
2505
Jens Axboe89850f72006-07-22 16:48:31 +02002506 cfq_put_queue(cfqq);
2507}
2508
2509static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2510 struct cfq_io_context *cic)
2511{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002512 struct io_context *ioc = cic->ioc;
2513
Jens Axboefc463792006-08-29 09:05:44 +02002514 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002515
2516 /*
2517 * Make sure key == NULL is seen for dead queues
2518 */
Jens Axboefc463792006-08-29 09:05:44 +02002519 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01002520 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02002521 cic->key = NULL;
2522
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002523 if (ioc->ioc_data == cic)
2524 rcu_assign_pointer(ioc->ioc_data, NULL);
2525
Jens Axboeff6657c2009-04-08 10:58:57 +02002526 if (cic->cfqq[BLK_RW_ASYNC]) {
2527 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2528 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002529 }
2530
Jens Axboeff6657c2009-04-08 10:58:57 +02002531 if (cic->cfqq[BLK_RW_SYNC]) {
2532 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2533 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002534 }
Jens Axboe89850f72006-07-22 16:48:31 +02002535}
2536
Jens Axboe4ac845a2008-01-24 08:44:49 +01002537static void cfq_exit_single_io_context(struct io_context *ioc,
2538 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002539{
Al Viro478a82b2006-03-18 13:25:24 -05002540 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02002541
Jens Axboe89850f72006-07-22 16:48:31 +02002542 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002543 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002544 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002545
Jens Axboe4ac845a2008-01-24 08:44:49 +01002546 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002547
2548 /*
2549 * Ensure we get a fresh copy of the ->key to prevent
2550 * race between exiting task and queue
2551 */
2552 smp_read_barrier_depends();
2553 if (cic->key)
2554 __cfq_exit_single_io_context(cfqd, cic);
2555
Jens Axboe4ac845a2008-01-24 08:44:49 +01002556 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002557 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002558}
2559
Jens Axboe498d3aa22007-04-26 12:54:48 +02002560/*
2561 * The process that ioc belongs to has exited, we need to clean up
2562 * and put the internal structures we have that belongs to that process.
2563 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002564static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002566 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567}
2568
Jens Axboe22e2c502005-06-27 10:55:12 +02002569static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002570cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571{
Jens Axboeb5deef92006-07-19 23:39:40 +02002572 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573
Christoph Lameter94f60302007-07-17 04:03:29 -07002574 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2575 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002577 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002578 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002579 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002580 cic->dtor = cfq_free_io_context;
2581 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002582 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583 }
2584
2585 return cic;
2586}
2587
Jens Axboefd0928d2008-01-24 08:52:45 +01002588static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002589{
2590 struct task_struct *tsk = current;
2591 int ioprio_class;
2592
Jens Axboe3b181522005-06-27 10:56:24 +02002593 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002594 return;
2595
Jens Axboefd0928d2008-01-24 08:52:45 +01002596 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002597 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002598 default:
2599 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2600 case IOPRIO_CLASS_NONE:
2601 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002602 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002603 */
2604 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002605 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002606 break;
2607 case IOPRIO_CLASS_RT:
2608 cfqq->ioprio = task_ioprio(ioc);
2609 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2610 break;
2611 case IOPRIO_CLASS_BE:
2612 cfqq->ioprio = task_ioprio(ioc);
2613 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2614 break;
2615 case IOPRIO_CLASS_IDLE:
2616 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2617 cfqq->ioprio = 7;
2618 cfq_clear_cfqq_idle_window(cfqq);
2619 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002620 }
2621
2622 /*
2623 * keep track of original prio settings in case we have to temporarily
2624 * elevate the priority of this queue
2625 */
2626 cfqq->org_ioprio = cfqq->ioprio;
2627 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002628 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002629}
2630
Jens Axboefebffd62008-01-28 13:19:43 +01002631static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002632{
Al Viro478a82b2006-03-18 13:25:24 -05002633 struct cfq_data *cfqd = cic->key;
2634 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002635 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002636
Jens Axboecaaa5f92006-06-16 11:23:00 +02002637 if (unlikely(!cfqd))
2638 return;
2639
Jens Axboec1b707d2006-10-30 19:54:23 +01002640 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002641
Jens Axboeff6657c2009-04-08 10:58:57 +02002642 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002643 if (cfqq) {
2644 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002645 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2646 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002647 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002648 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002649 cfq_put_queue(cfqq);
2650 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002651 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002652
Jens Axboeff6657c2009-04-08 10:58:57 +02002653 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002654 if (cfqq)
2655 cfq_mark_cfqq_prio_changed(cfqq);
2656
Jens Axboec1b707d2006-10-30 19:54:23 +01002657 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002658}
2659
Jens Axboefc463792006-08-29 09:05:44 +02002660static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002662 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002663 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664}
2665
Jens Axboed5036d72009-06-26 10:44:34 +02002666static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002667 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002668{
2669 RB_CLEAR_NODE(&cfqq->rb_node);
2670 RB_CLEAR_NODE(&cfqq->p_node);
2671 INIT_LIST_HEAD(&cfqq->fifo);
2672
2673 atomic_set(&cfqq->ref, 0);
2674 cfqq->cfqd = cfqd;
2675
2676 cfq_mark_cfqq_prio_changed(cfqq);
2677
2678 if (is_sync) {
2679 if (!cfq_class_idle(cfqq))
2680 cfq_mark_cfqq_idle_window(cfqq);
2681 cfq_mark_cfqq_sync(cfqq);
2682 }
2683 cfqq->pid = pid;
2684}
2685
Vivek Goyal246103332009-12-03 12:59:51 -05002686#ifdef CONFIG_CFQ_GROUP_IOSCHED
2687static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2688{
2689 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2690 struct cfq_data *cfqd = cic->key;
2691 unsigned long flags;
2692 struct request_queue *q;
2693
2694 if (unlikely(!cfqd))
2695 return;
2696
2697 q = cfqd->queue;
2698
2699 spin_lock_irqsave(q->queue_lock, flags);
2700
2701 if (sync_cfqq) {
2702 /*
2703 * Drop reference to sync queue. A new sync queue will be
2704 * assigned in new group upon arrival of a fresh request.
2705 */
2706 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2707 cic_set_cfqq(cic, NULL, 1);
2708 cfq_put_queue(sync_cfqq);
2709 }
2710
2711 spin_unlock_irqrestore(q->queue_lock, flags);
2712}
2713
2714static void cfq_ioc_set_cgroup(struct io_context *ioc)
2715{
2716 call_for_each_cic(ioc, changed_cgroup);
2717 ioc->cgroup_changed = 0;
2718}
2719#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2720
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002722cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002723 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002726 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002727 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728
2729retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002730 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002731 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002732 /* cic always exists here */
2733 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734
Jens Axboe6118b702009-06-30 09:34:12 +02002735 /*
2736 * Always try a new alloc if we fell back to the OOM cfqq
2737 * originally, since it should just be a temporary situation.
2738 */
2739 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2740 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741 if (new_cfqq) {
2742 cfqq = new_cfqq;
2743 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002744 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002746 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002747 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002748 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002750 if (new_cfqq)
2751 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002752 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002753 cfqq = kmem_cache_alloc_node(cfq_pool,
2754 gfp_mask | __GFP_ZERO,
2755 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002756 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757
Jens Axboe6118b702009-06-30 09:34:12 +02002758 if (cfqq) {
2759 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2760 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002761 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002762 cfq_log_cfqq(cfqd, cfqq, "alloced");
2763 } else
2764 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 }
2766
2767 if (new_cfqq)
2768 kmem_cache_free(cfq_pool, new_cfqq);
2769
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770 return cfqq;
2771}
2772
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002773static struct cfq_queue **
2774cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2775{
Jens Axboefe094d92008-01-31 13:08:54 +01002776 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002777 case IOPRIO_CLASS_RT:
2778 return &cfqd->async_cfqq[0][ioprio];
2779 case IOPRIO_CLASS_BE:
2780 return &cfqd->async_cfqq[1][ioprio];
2781 case IOPRIO_CLASS_IDLE:
2782 return &cfqd->async_idle_cfqq;
2783 default:
2784 BUG();
2785 }
2786}
2787
Jens Axboe15c31be2007-07-10 13:43:25 +02002788static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002789cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002790 gfp_t gfp_mask)
2791{
Jens Axboefd0928d2008-01-24 08:52:45 +01002792 const int ioprio = task_ioprio(ioc);
2793 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002794 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002795 struct cfq_queue *cfqq = NULL;
2796
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002797 if (!is_sync) {
2798 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2799 cfqq = *async_cfqq;
2800 }
2801
Jens Axboe6118b702009-06-30 09:34:12 +02002802 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002803 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002804
2805 /*
2806 * pin the queue now that it's allocated, scheduler exit will prune it
2807 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002808 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02002809 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002810 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002811 }
2812
2813 atomic_inc(&cfqq->ref);
2814 return cfqq;
2815}
2816
Jens Axboe498d3aa22007-04-26 12:54:48 +02002817/*
2818 * We drop cfq io contexts lazily, so we may find a dead one.
2819 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002820static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002821cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2822 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002823{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002824 unsigned long flags;
2825
Jens Axboefc463792006-08-29 09:05:44 +02002826 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02002827
Jens Axboe4ac845a2008-01-24 08:44:49 +01002828 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02002829
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002830 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002831
2832 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01002833 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002834 spin_unlock_irqrestore(&ioc->lock, flags);
2835
2836 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002837}
2838
Jens Axboee2d74ac2006-03-28 08:59:01 +02002839static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01002840cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002841{
Jens Axboee2d74ac2006-03-28 08:59:01 +02002842 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002843 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002844 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002845
Vasily Tarasov91fac312007-04-25 12:29:51 +02002846 if (unlikely(!ioc))
2847 return NULL;
2848
Jens Axboed6de8be2008-05-28 14:46:59 +02002849 rcu_read_lock();
2850
Jens Axboe597bc482007-04-24 21:23:53 +02002851 /*
2852 * we maintain a last-hit cache, to avoid browsing over the tree
2853 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002854 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02002855 if (cic && cic->key == cfqd) {
2856 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02002857 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002858 }
Jens Axboe597bc482007-04-24 21:23:53 +02002859
Jens Axboe4ac845a2008-01-24 08:44:49 +01002860 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01002861 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2862 rcu_read_unlock();
2863 if (!cic)
2864 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02002865 /* ->key must be copied to avoid race with cfq_exit_queue() */
2866 k = cic->key;
2867 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01002868 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002869 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01002870 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002871 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002872
Jens Axboed6de8be2008-05-28 14:46:59 +02002873 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002874 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002875 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002876 break;
2877 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002878
Jens Axboe4ac845a2008-01-24 08:44:49 +01002879 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002880}
2881
Jens Axboe4ac845a2008-01-24 08:44:49 +01002882/*
2883 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2884 * the process specific cfq io context when entered from the block layer.
2885 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2886 */
Jens Axboefebffd62008-01-28 13:19:43 +01002887static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2888 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002889{
Jens Axboe0261d682006-10-30 19:07:48 +01002890 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002891 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002892
Jens Axboe4ac845a2008-01-24 08:44:49 +01002893 ret = radix_tree_preload(gfp_mask);
2894 if (!ret) {
2895 cic->ioc = ioc;
2896 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002897
Jens Axboe4ac845a2008-01-24 08:44:49 +01002898 spin_lock_irqsave(&ioc->lock, flags);
2899 ret = radix_tree_insert(&ioc->radix_root,
2900 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01002901 if (!ret)
2902 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002903 spin_unlock_irqrestore(&ioc->lock, flags);
2904
2905 radix_tree_preload_end();
2906
2907 if (!ret) {
2908 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2909 list_add(&cic->queue_list, &cfqd->cic_list);
2910 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002911 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002912 }
2913
Jens Axboe4ac845a2008-01-24 08:44:49 +01002914 if (ret)
2915 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02002916
Jens Axboe4ac845a2008-01-24 08:44:49 +01002917 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002918}
2919
Jens Axboe22e2c502005-06-27 10:55:12 +02002920/*
2921 * Setup general io context and cfq io context. There can be several cfq
2922 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02002923 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02002924 */
2925static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02002926cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927{
Jens Axboe22e2c502005-06-27 10:55:12 +02002928 struct io_context *ioc = NULL;
2929 struct cfq_io_context *cic;
2930
2931 might_sleep_if(gfp_mask & __GFP_WAIT);
2932
Jens Axboeb5deef92006-07-19 23:39:40 +02002933 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02002934 if (!ioc)
2935 return NULL;
2936
Jens Axboe4ac845a2008-01-24 08:44:49 +01002937 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002938 if (cic)
2939 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02002940
Jens Axboee2d74ac2006-03-28 08:59:01 +02002941 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2942 if (cic == NULL)
2943 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02002944
Jens Axboe4ac845a2008-01-24 08:44:49 +01002945 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2946 goto err_free;
2947
Jens Axboe22e2c502005-06-27 10:55:12 +02002948out:
Jens Axboefc463792006-08-29 09:05:44 +02002949 smp_read_barrier_depends();
2950 if (unlikely(ioc->ioprio_changed))
2951 cfq_ioc_set_ioprio(ioc);
2952
Vivek Goyal246103332009-12-03 12:59:51 -05002953#ifdef CONFIG_CFQ_GROUP_IOSCHED
2954 if (unlikely(ioc->cgroup_changed))
2955 cfq_ioc_set_cgroup(ioc);
2956#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02002957 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002958err_free:
2959 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002960err:
2961 put_io_context(ioc);
2962 return NULL;
2963}
2964
2965static void
2966cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2967{
Jens Axboeaaf12282007-01-19 11:30:16 +11002968 unsigned long elapsed = jiffies - cic->last_end_request;
2969 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02002970
2971 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2972 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2973 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2974}
2975
Jens Axboe206dc692006-03-28 13:03:44 +02002976static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002977cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02002978 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02002979{
2980 sector_t sdist;
2981 u64 total;
2982
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002983 if (!cfqq->last_request_pos)
Jeff Moyer4d00aa42009-04-21 07:25:04 +02002984 sdist = 0;
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002985 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2986 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02002987 else
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002988 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
Jens Axboe206dc692006-03-28 13:03:44 +02002989
2990 /*
2991 * Don't allow the seek distance to get too large from the
2992 * odd fragment, pagein, etc
2993 */
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002994 if (cfqq->seek_samples <= 60) /* second&third seek */
2995 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
Jens Axboe206dc692006-03-28 13:03:44 +02002996 else
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002997 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
Jens Axboe206dc692006-03-28 13:03:44 +02002998
Jeff Moyerb2c18e12009-10-23 17:14:49 -04002999 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
3000 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
3001 total = cfqq->seek_total + (cfqq->seek_samples/2);
3002 do_div(total, cfqq->seek_samples);
3003 cfqq->seek_mean = (sector_t)total;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003004
3005 /*
3006 * If this cfqq is shared between multiple processes, check to
3007 * make sure that those processes are still issuing I/Os within
3008 * the mean seek distance. If not, it may be time to break the
3009 * queues apart again.
3010 */
3011 if (cfq_cfqq_coop(cfqq)) {
3012 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
3013 cfqq->seeky_start = jiffies;
3014 else if (!CFQQ_SEEKY(cfqq))
3015 cfqq->seeky_start = 0;
3016 }
Jens Axboe206dc692006-03-28 13:03:44 +02003017}
Jens Axboe22e2c502005-06-27 10:55:12 +02003018
3019/*
3020 * Disable idle window if the process thinks too long or seeks so much that
3021 * it doesn't matter
3022 */
3023static void
3024cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3025 struct cfq_io_context *cic)
3026{
Jens Axboe7b679132008-05-30 12:23:07 +02003027 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003028
Jens Axboe08717142008-01-28 11:38:15 +01003029 /*
3030 * Don't idle for async or idle io prio class
3031 */
3032 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003033 return;
3034
Jens Axboec265a7f2008-06-26 13:49:33 +02003035 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003036
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003037 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3038 cfq_mark_cfqq_deep(cfqq);
3039
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01003040 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003041 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
3042 && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003043 enable_idle = 0;
3044 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003045 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003046 enable_idle = 0;
3047 else
3048 enable_idle = 1;
3049 }
3050
Jens Axboe7b679132008-05-30 12:23:07 +02003051 if (old_idle != enable_idle) {
3052 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3053 if (enable_idle)
3054 cfq_mark_cfqq_idle_window(cfqq);
3055 else
3056 cfq_clear_cfqq_idle_window(cfqq);
3057 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003058}
3059
Jens Axboe22e2c502005-06-27 10:55:12 +02003060/*
3061 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3062 * no or if we aren't sure, a 1 will cause a preempt.
3063 */
Jens Axboea6151c32009-10-07 20:02:57 +02003064static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003065cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003066 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003067{
Jens Axboe6d048f52007-04-25 12:44:27 +02003068 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003069
Jens Axboe6d048f52007-04-25 12:44:27 +02003070 cfqq = cfqd->active_queue;
3071 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003072 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003073
Jens Axboe6d048f52007-04-25 12:44:27 +02003074 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003075 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003076
3077 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003078 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003079
Jens Axboe22e2c502005-06-27 10:55:12 +02003080 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003081 * if the new request is sync, but the currently running queue is
3082 * not, let the sync request have priority.
3083 */
Jens Axboe5e705372006-07-13 12:39:25 +02003084 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003085 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003086
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003087 if (new_cfqq->cfqg != cfqq->cfqg)
3088 return false;
3089
3090 if (cfq_slice_used(cfqq))
3091 return true;
3092
3093 /* Allow preemption only if we are idling on sync-noidle tree */
3094 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3095 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3096 new_cfqq->service_tree->count == 2 &&
3097 RB_EMPTY_ROOT(&cfqq->sort_list))
3098 return true;
3099
Jens Axboe374f84a2006-07-23 01:42:19 +02003100 /*
3101 * So both queues are sync. Let the new request get disk time if
3102 * it's a metadata request and the current queue is doing regular IO.
3103 */
3104 if (rq_is_meta(rq) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003105 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003106
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003107 /*
3108 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3109 */
3110 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003111 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003112
Jens Axboe1e3335d2007-02-14 19:59:49 +01003113 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003114 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003115
3116 /*
3117 * if this request is as-good as one we would expect from the
3118 * current cfqq, let it preempt
3119 */
Jens Axboee00ef792009-11-04 08:54:55 +01003120 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003121 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003122
Jens Axboea6151c32009-10-07 20:02:57 +02003123 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003124}
3125
3126/*
3127 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3128 * let it have half of its nominal slice.
3129 */
3130static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3131{
Jens Axboe7b679132008-05-30 12:23:07 +02003132 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02003133 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003134
Jens Axboebf572252006-07-19 20:29:12 +02003135 /*
3136 * Put the new queue at the front of the of the current list,
3137 * so we know that it will be selected next.
3138 */
3139 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003140
3141 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02003142
Jens Axboe44f7c162007-01-19 11:51:58 +11003143 cfqq->slice_end = 0;
3144 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003145}
3146
3147/*
Jens Axboe5e705372006-07-13 12:39:25 +02003148 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003149 * something we should do about it
3150 */
3151static void
Jens Axboe5e705372006-07-13 12:39:25 +02003152cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3153 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003154{
Jens Axboe5e705372006-07-13 12:39:25 +02003155 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003156
Aaron Carroll45333d52008-08-26 15:52:36 +02003157 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003158 if (rq_is_meta(rq))
3159 cfqq->meta_pending++;
3160
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003161 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003162 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003163 cfq_update_idle_window(cfqd, cfqq, cic);
3164
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003165 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003166
3167 if (cfqq == cfqd->active_queue) {
Vivek Goyalf75edf22009-12-03 12:59:53 -05003168 if (cfq_cfqq_wait_busy(cfqq)) {
3169 cfq_clear_cfqq_wait_busy(cfqq);
3170 cfq_mark_cfqq_wait_busy_done(cfqq);
3171 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003172 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003173 * Remember that we saw a request from this process, but
3174 * don't start queuing just yet. Otherwise we risk seeing lots
3175 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003176 * and merging. If the request is already larger than a single
3177 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003178 * merging is already done. Ditto for a busy system that
3179 * has other work pending, don't risk delaying until the
3180 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003181 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003182 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003183 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3184 cfqd->busy_queues > 1) {
Jens Axboed6ceb252009-04-14 14:18:16 +02003185 del_timer(&cfqd->idle_slice_timer);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003186 __blk_run_queue(cfqd->queue);
3187 } else
3188 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboed6ceb252009-04-14 14:18:16 +02003189 }
Jens Axboe5e705372006-07-13 12:39:25 +02003190 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003191 /*
3192 * not the active queue - expire current slice if it is
3193 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003194 * has some old slice time left and is of higher priority or
3195 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003196 */
3197 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09003198 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003199 }
3200}
3201
Jens Axboe165125e2007-07-24 09:28:11 +02003202static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003203{
Jens Axboeb4878f22005-10-20 16:42:29 +02003204 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003205 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003206
Jens Axboe7b679132008-05-30 12:23:07 +02003207 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003208 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209
Jens Axboe30996f42009-10-05 11:03:39 +02003210 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003211 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003212 cfq_add_rq_rb(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003213
Jens Axboe5e705372006-07-13 12:39:25 +02003214 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215}
3216
Aaron Carroll45333d52008-08-26 15:52:36 +02003217/*
3218 * Update hw_tag based on peak queue depth over 50 samples under
3219 * sufficient load.
3220 */
3221static void cfq_update_hw_tag(struct cfq_data *cfqd)
3222{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003223 struct cfq_queue *cfqq = cfqd->active_queue;
3224
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003225 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3226 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3227
3228 if (cfqd->hw_tag == 1)
3229 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003230
3231 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Jens Axboe5ad531d2009-07-03 12:57:48 +02003232 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003233 return;
3234
Shaohua Li1a1238a2009-10-27 08:46:23 +01003235 /*
3236 * If active queue hasn't enough requests and can idle, cfq might not
3237 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3238 * case
3239 */
3240 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3241 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3242 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3243 return;
3244
Aaron Carroll45333d52008-08-26 15:52:36 +02003245 if (cfqd->hw_tag_samples++ < 50)
3246 return;
3247
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003248 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003249 cfqd->hw_tag = 1;
3250 else
3251 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003252}
3253
Jens Axboe165125e2007-07-24 09:28:11 +02003254static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255{
Jens Axboe5e705372006-07-13 12:39:25 +02003256 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003257 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003258 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003259 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260
Jens Axboeb4878f22005-10-20 16:42:29 +02003261 now = jiffies;
Vivek Goyal2868ef72009-12-03 12:59:48 -05003262 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263
Aaron Carroll45333d52008-08-26 15:52:36 +02003264 cfq_update_hw_tag(cfqd);
3265
Jens Axboe5ad531d2009-07-03 12:57:48 +02003266 WARN_ON(!cfqd->rq_in_driver[sync]);
Jens Axboe6d048f52007-04-25 12:44:27 +02003267 WARN_ON(!cfqq->dispatched);
Jens Axboe5ad531d2009-07-03 12:57:48 +02003268 cfqd->rq_in_driver[sync]--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003269 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270
Jens Axboe3ed9a292007-04-23 08:33:33 +02003271 if (cfq_cfqq_sync(cfqq))
3272 cfqd->sync_flight--;
3273
Vivek Goyal365722b2009-10-03 15:21:27 +02003274 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003275 RQ_CIC(rq)->last_end_request = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003276 cfqd->last_end_sync_rq = now;
3277 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003278
3279 /*
3280 * If this is the active queue, check if it needs to be expired,
3281 * or if we want to idle in case it has no pending requests.
3282 */
3283 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003284 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3285
Jens Axboe44f7c162007-01-19 11:51:58 +11003286 if (cfq_cfqq_slice_new(cfqq)) {
3287 cfq_set_prio_slice(cfqd, cfqq);
3288 cfq_clear_cfqq_slice_new(cfqq);
3289 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003290
3291 /*
3292 * If this queue consumed its slice and this is last queue
3293 * in the group, wait for next request before we expire
3294 * the queue
3295 */
3296 if (cfq_slice_used(cfqq) && cfqq->cfqg->nr_cfqq == 1) {
3297 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3298 cfq_mark_cfqq_wait_busy(cfqq);
3299 }
3300
Jens Axboea36e71f2009-04-15 12:15:11 +02003301 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003302 * Idling is not enabled on:
3303 * - expired queues
3304 * - idle-priority queues
3305 * - async queues
3306 * - queues with still some requests queued
3307 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003308 */
Jens Axboe08717142008-01-28 11:38:15 +01003309 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02003310 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003311 else if (sync && cfqq_empty &&
3312 !cfq_close_cooperator(cfqd, cfqq)) {
3313 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3314 /*
3315 * Idling is enabled for SYNC_WORKLOAD.
3316 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3317 * only if we processed at least one !rq_noidle request
3318 */
3319 if (cfqd->serving_type == SYNC_WORKLOAD
Vivek Goyalc04645e2009-12-03 12:59:56 -05003320 || cfqd->noidle_tree_requires_idle
3321 || cfqq->cfqg->nr_cfqq == 1)
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003322 cfq_arm_slice_timer(cfqd);
3323 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003324 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003325
Jens Axboe5ad531d2009-07-03 12:57:48 +02003326 if (!rq_in_driver(cfqd))
Jens Axboe23e018a2009-10-05 08:52:35 +02003327 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328}
3329
Jens Axboe22e2c502005-06-27 10:55:12 +02003330/*
3331 * we temporarily boost lower priority queues if they are holding fs exclusive
3332 * resources. they are boosted to normal prio (CLASS_BE/4)
3333 */
3334static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335{
Jens Axboe22e2c502005-06-27 10:55:12 +02003336 if (has_fs_excl()) {
3337 /*
3338 * boost idle prio on transactions that would lock out other
3339 * users of the filesystem
3340 */
3341 if (cfq_class_idle(cfqq))
3342 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3343 if (cfqq->ioprio > IOPRIO_NORM)
3344 cfqq->ioprio = IOPRIO_NORM;
3345 } else {
3346 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003347 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003348 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003349 cfqq->ioprio_class = cfqq->org_ioprio_class;
3350 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003352}
3353
Jens Axboe89850f72006-07-22 16:48:31 +02003354static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003355{
Jens Axboe1b379d82009-08-11 08:26:11 +02003356 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003357 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003358 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003359 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003360
3361 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003362}
3363
Jens Axboe165125e2007-07-24 09:28:11 +02003364static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003365{
3366 struct cfq_data *cfqd = q->elevator->elevator_data;
3367 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003368 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003369 struct cfq_queue *cfqq;
3370
3371 /*
3372 * don't force setup of a queue from here, as a call to may_queue
3373 * does not necessarily imply that a request actually will be queued.
3374 * so just lookup a possibly existing queue, or return 'may queue'
3375 * if that fails
3376 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003377 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003378 if (!cic)
3379 return ELV_MQUEUE_MAY;
3380
Jens Axboeb0b78f82009-04-08 10:56:08 +02003381 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003382 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003383 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003384 cfq_prio_boost(cfqq);
3385
Jens Axboe89850f72006-07-22 16:48:31 +02003386 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003387 }
3388
3389 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390}
3391
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392/*
3393 * queue lock held here
3394 */
Jens Axboebb37b942006-12-01 10:42:33 +01003395static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396{
Jens Axboe5e705372006-07-13 12:39:25 +02003397 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398
Jens Axboe5e705372006-07-13 12:39:25 +02003399 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003400 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401
Jens Axboe22e2c502005-06-27 10:55:12 +02003402 BUG_ON(!cfqq->allocated[rw]);
3403 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404
Jens Axboe5e705372006-07-13 12:39:25 +02003405 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02003408 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 cfq_put_queue(cfqq);
3411 }
3412}
3413
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003414static struct cfq_queue *
3415cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3416 struct cfq_queue *cfqq)
3417{
3418 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3419 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003420 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003421 cfq_put_queue(cfqq);
3422 return cic_to_cfqq(cic, 1);
3423}
3424
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003425static int should_split_cfqq(struct cfq_queue *cfqq)
3426{
3427 if (cfqq->seeky_start &&
3428 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3429 return 1;
3430 return 0;
3431}
3432
3433/*
3434 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3435 * was the last process referring to said cfqq.
3436 */
3437static struct cfq_queue *
3438split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3439{
3440 if (cfqq_process_refs(cfqq) == 1) {
3441 cfqq->seeky_start = 0;
3442 cfqq->pid = current->pid;
3443 cfq_clear_cfqq_coop(cfqq);
3444 return cfqq;
3445 }
3446
3447 cic_set_cfqq(cic, NULL, 1);
3448 cfq_put_queue(cfqq);
3449 return NULL;
3450}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003452 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003454static int
Jens Axboe165125e2007-07-24 09:28:11 +02003455cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456{
3457 struct cfq_data *cfqd = q->elevator->elevator_data;
3458 struct cfq_io_context *cic;
3459 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003460 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003461 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462 unsigned long flags;
3463
3464 might_sleep_if(gfp_mask & __GFP_WAIT);
3465
Jens Axboee2d74ac2006-03-28 08:59:01 +02003466 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003467
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468 spin_lock_irqsave(q->queue_lock, flags);
3469
Jens Axboe22e2c502005-06-27 10:55:12 +02003470 if (!cic)
3471 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003473new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003474 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003475 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003476 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003477 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003478 } else {
3479 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003480 * If the queue was seeky for too long, break it apart.
3481 */
3482 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3483 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3484 cfqq = split_cfqq(cic, cfqq);
3485 if (!cfqq)
3486 goto new_queue;
3487 }
3488
3489 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003490 * Check to see if this queue is scheduled to merge with
3491 * another, closely cooperating queue. The merging of
3492 * queues happens here as it must be done in process context.
3493 * The reference on new_cfqq was taken in merge_cfqqs.
3494 */
3495 if (cfqq->new_cfqq)
3496 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003497 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498
3499 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02003500 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02003501
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502 spin_unlock_irqrestore(q->queue_lock, flags);
3503
Jens Axboe5e705372006-07-13 12:39:25 +02003504 rq->elevator_private = cic;
3505 rq->elevator_private2 = cfqq;
3506 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003507
Jens Axboe22e2c502005-06-27 10:55:12 +02003508queue_fail:
3509 if (cic)
3510 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003511
Jens Axboe23e018a2009-10-05 08:52:35 +02003512 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003514 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515 return 1;
3516}
3517
David Howells65f27f32006-11-22 14:55:48 +00003518static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003519{
David Howells65f27f32006-11-22 14:55:48 +00003520 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003521 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003522 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003523
Jens Axboe40bb54d2009-04-15 12:11:10 +02003524 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09003525 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003526 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003527}
3528
3529/*
3530 * Timer running if the active_queue is currently idling inside its time slice
3531 */
3532static void cfq_idle_slice_timer(unsigned long data)
3533{
3534 struct cfq_data *cfqd = (struct cfq_data *) data;
3535 struct cfq_queue *cfqq;
3536 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003537 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003538
Jens Axboe7b679132008-05-30 12:23:07 +02003539 cfq_log(cfqd, "idle timer fired");
3540
Jens Axboe22e2c502005-06-27 10:55:12 +02003541 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3542
Jens Axboefe094d92008-01-31 13:08:54 +01003543 cfqq = cfqd->active_queue;
3544 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003545 timed_out = 0;
3546
Jens Axboe22e2c502005-06-27 10:55:12 +02003547 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003548 * We saw a request before the queue expired, let it through
3549 */
3550 if (cfq_cfqq_must_dispatch(cfqq))
3551 goto out_kick;
3552
3553 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003554 * expired
3555 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003556 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003557 goto expire;
3558
3559 /*
3560 * only expire and reinvoke request handler, if there are
3561 * other queues with pending requests
3562 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003563 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003564 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003565
3566 /*
3567 * not expired and it has a request pending, let it dispatch
3568 */
Jens Axboe75e50982009-04-07 08:56:14 +02003569 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003570 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003571
3572 /*
3573 * Queue depth flag is reset only when the idle didn't succeed
3574 */
3575 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003576 }
3577expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02003578 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003579out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003580 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003581out_cont:
3582 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3583}
3584
Jens Axboe3b181522005-06-27 10:56:24 +02003585static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3586{
3587 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003588 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003589}
Jens Axboe22e2c502005-06-27 10:55:12 +02003590
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003591static void cfq_put_async_queues(struct cfq_data *cfqd)
3592{
3593 int i;
3594
3595 for (i = 0; i < IOPRIO_BE_NR; i++) {
3596 if (cfqd->async_cfqq[0][i])
3597 cfq_put_queue(cfqd->async_cfqq[0][i]);
3598 if (cfqd->async_cfqq[1][i])
3599 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003600 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003601
3602 if (cfqd->async_idle_cfqq)
3603 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003604}
3605
Jens Axboebb729bc2009-12-06 09:54:19 +01003606static void cfq_cfqd_free(struct rcu_head *head)
3607{
3608 kfree(container_of(head, struct cfq_data, rcu));
3609}
3610
Jens Axboeb374d182008-10-31 10:05:07 +01003611static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612{
Jens Axboe22e2c502005-06-27 10:55:12 +02003613 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003614 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003615
Jens Axboe3b181522005-06-27 10:56:24 +02003616 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003617
Al Virod9ff4182006-03-18 13:51:22 -05003618 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003619
Al Virod9ff4182006-03-18 13:51:22 -05003620 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02003621 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003622
3623 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003624 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3625 struct cfq_io_context,
3626 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003627
3628 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003629 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003630
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003631 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003632 cfq_release_cfq_groups(cfqd);
3633 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003634
Al Virod9ff4182006-03-18 13:51:22 -05003635 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003636
3637 cfq_shutdown_timer_wq(cfqd);
3638
Vivek Goyalb1c35762009-12-03 12:59:47 -05003639 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003640 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003641}
3642
Jens Axboe165125e2007-07-24 09:28:11 +02003643static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644{
3645 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003646 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003647 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003648 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649
Christoph Lameter94f60302007-07-17 04:03:29 -07003650 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003651 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003652 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003654 /* Init root service tree */
3655 cfqd->grp_service_tree = CFQ_RB_ROOT;
3656
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003657 /* Init root group */
3658 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003659 for_each_cfqg_st(cfqg, i, j, st)
3660 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003661 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003662
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003663 /* Give preference to root group over other groups */
3664 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3665
Vivek Goyal25fb5162009-12-03 12:59:46 -05003666#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003667 /*
3668 * Take a reference to root group which we never drop. This is just
3669 * to make sure that cfq_put_cfqg() does not try to kfree root group
3670 */
3671 atomic_set(&cfqg->ref, 1);
Vivek Goyal22084192009-12-03 12:59:49 -05003672 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3673 0);
Vivek Goyal25fb5162009-12-03 12:59:46 -05003674#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003675 /*
3676 * Not strictly needed (since RB_ROOT just clears the node and we
3677 * zeroed cfqd on alloc), but better be safe in case someone decides
3678 * to add magic to the rb code
3679 */
3680 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3681 cfqd->prio_trees[i] = RB_ROOT;
3682
Jens Axboe6118b702009-06-30 09:34:12 +02003683 /*
3684 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3685 * Grab a permanent reference to it, so that the normal code flow
3686 * will not attempt to free it.
3687 */
3688 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3689 atomic_inc(&cfqd->oom_cfqq.ref);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003690 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003691
Al Virod9ff4182006-03-18 13:51:22 -05003692 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695
Jens Axboe22e2c502005-06-27 10:55:12 +02003696 init_timer(&cfqd->idle_slice_timer);
3697 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3698 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3699
Jens Axboe23e018a2009-10-05 08:52:35 +02003700 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003701
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003703 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3704 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 cfqd->cfq_back_max = cfq_back_max;
3706 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003707 cfqd->cfq_slice[0] = cfq_slice_async;
3708 cfqd->cfq_slice[1] = cfq_slice_sync;
3709 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3710 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003711 cfqd->cfq_latency = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05003712 cfqd->cfq_group_isolation = 0;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003713 cfqd->hw_tag = -1;
Vivek Goyal365722b2009-10-03 15:21:27 +02003714 cfqd->last_end_sync_rq = jiffies;
Jens Axboebb729bc2009-12-06 09:54:19 +01003715 INIT_RCU_HEAD(&cfqd->rcu);
Jens Axboebc1c1162006-06-08 08:49:06 +02003716 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717}
3718
3719static void cfq_slab_kill(void)
3720{
Jens Axboed6de8be2008-05-28 14:46:59 +02003721 /*
3722 * Caller already ensured that pending RCU callbacks are completed,
3723 * so we should have no busy allocations at this point.
3724 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725 if (cfq_pool)
3726 kmem_cache_destroy(cfq_pool);
3727 if (cfq_ioc_pool)
3728 kmem_cache_destroy(cfq_ioc_pool);
3729}
3730
3731static int __init cfq_slab_setup(void)
3732{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003733 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 if (!cfq_pool)
3735 goto fail;
3736
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003737 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738 if (!cfq_ioc_pool)
3739 goto fail;
3740
3741 return 0;
3742fail:
3743 cfq_slab_kill();
3744 return -ENOMEM;
3745}
3746
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747/*
3748 * sysfs parts below -->
3749 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003750static ssize_t
3751cfq_var_show(unsigned int var, char *page)
3752{
3753 return sprintf(page, "%d\n", var);
3754}
3755
3756static ssize_t
3757cfq_var_store(unsigned int *var, const char *page, size_t count)
3758{
3759 char *p = (char *) page;
3760
3761 *var = simple_strtoul(p, &p, 10);
3762 return count;
3763}
3764
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003766static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003768 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769 unsigned int __data = __VAR; \
3770 if (__CONV) \
3771 __data = jiffies_to_msecs(__data); \
3772 return cfq_var_show(__data, (page)); \
3773}
3774SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003775SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3776SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003777SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3778SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003779SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3780SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3781SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3782SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003783SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003784SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003785#undef SHOW_FUNCTION
3786
3787#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003788static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003790 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791 unsigned int __data; \
3792 int ret = cfq_var_store(&__data, (page), count); \
3793 if (__data < (MIN)) \
3794 __data = (MIN); \
3795 else if (__data > (MAX)) \
3796 __data = (MAX); \
3797 if (__CONV) \
3798 *(__PTR) = msecs_to_jiffies(__data); \
3799 else \
3800 *(__PTR) = __data; \
3801 return ret; \
3802}
3803STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003804STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3805 UINT_MAX, 1);
3806STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3807 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05003808STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003809STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3810 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003811STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3812STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3813STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01003814STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3815 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003816STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003817STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818#undef STORE_FUNCTION
3819
Al Viroe572ec72006-03-18 22:27:18 -05003820#define CFQ_ATTR(name) \
3821 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02003822
Al Viroe572ec72006-03-18 22:27:18 -05003823static struct elv_fs_entry cfq_attrs[] = {
3824 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05003825 CFQ_ATTR(fifo_expire_sync),
3826 CFQ_ATTR(fifo_expire_async),
3827 CFQ_ATTR(back_seek_max),
3828 CFQ_ATTR(back_seek_penalty),
3829 CFQ_ATTR(slice_sync),
3830 CFQ_ATTR(slice_async),
3831 CFQ_ATTR(slice_async_rq),
3832 CFQ_ATTR(slice_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02003833 CFQ_ATTR(low_latency),
Vivek Goyalae30c282009-12-03 12:59:55 -05003834 CFQ_ATTR(group_isolation),
Al Viroe572ec72006-03-18 22:27:18 -05003835 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836};
3837
Linus Torvalds1da177e2005-04-16 15:20:36 -07003838static struct elevator_type iosched_cfq = {
3839 .ops = {
3840 .elevator_merge_fn = cfq_merge,
3841 .elevator_merged_fn = cfq_merged_request,
3842 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01003843 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02003844 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02003846 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847 .elevator_deactivate_req_fn = cfq_deactivate_request,
3848 .elevator_queue_empty_fn = cfq_queue_empty,
3849 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02003850 .elevator_former_req_fn = elv_rb_former_request,
3851 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852 .elevator_set_req_fn = cfq_set_request,
3853 .elevator_put_req_fn = cfq_put_request,
3854 .elevator_may_queue_fn = cfq_may_queue,
3855 .elevator_init_fn = cfq_init_queue,
3856 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02003857 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858 },
Al Viro3d1ab402006-03-18 18:35:43 -05003859 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860 .elevator_name = "cfq",
3861 .elevator_owner = THIS_MODULE,
3862};
3863
Vivek Goyal3e252062009-12-04 10:36:42 -05003864#ifdef CONFIG_CFQ_GROUP_IOSCHED
3865static struct blkio_policy_type blkio_policy_cfq = {
3866 .ops = {
3867 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3868 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3869 },
3870};
3871#else
3872static struct blkio_policy_type blkio_policy_cfq;
3873#endif
3874
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875static int __init cfq_init(void)
3876{
Jens Axboe22e2c502005-06-27 10:55:12 +02003877 /*
3878 * could be 0 on HZ < 1000 setups
3879 */
3880 if (!cfq_slice_async)
3881 cfq_slice_async = 1;
3882 if (!cfq_slice_idle)
3883 cfq_slice_idle = 1;
3884
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885 if (cfq_slab_setup())
3886 return -ENOMEM;
3887
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01003888 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05003889 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01003891 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892}
3893
3894static void __exit cfq_exit(void)
3895{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07003896 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05003897 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05003899 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02003900 /* ioc_gone's update must be visible before reading ioc_count */
3901 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02003902
3903 /*
3904 * this also protects us from entering cfq_slab_kill() with
3905 * pending RCU callbacks
3906 */
Tejun Heo245b2e72009-06-24 15:13:48 +09003907 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02003908 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08003909 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910}
3911
3912module_init(cfq_init);
3913module_exit(cfq_exit);
3914
3915MODULE_AUTHOR("Jens Axboe");
3916MODULE_LICENSE("GPL");
3917MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");