blob: e2f80463ed0d621df96bef311543f714462d7f7d [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
Corrado Zoccolo573412b2009-12-06 11:48:52 +0100286 unsigned long last_delayed_sync;
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 */
Jens Axboe3b181522005-06-27 10:56:24 +0200322};
323
324#define CFQ_CFQQ_FNS(name) \
325static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
326{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100327 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200328} \
329static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
330{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100331 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
Jens Axboe3b181522005-06-27 10:56:24 +0200332} \
333static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
334{ \
Jens Axboefe094d92008-01-31 13:08:54 +0100335 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
Jens Axboe3b181522005-06-27 10:56:24 +0200336}
337
338CFQ_CFQQ_FNS(on_rr);
339CFQ_CFQQ_FNS(wait_request);
Jens Axboeb0291952009-04-07 11:38:31 +0200340CFQ_CFQQ_FNS(must_dispatch);
Jens Axboe3b181522005-06-27 10:56:24 +0200341CFQ_CFQQ_FNS(must_alloc_slice);
Jens Axboe3b181522005-06-27 10:56:24 +0200342CFQ_CFQQ_FNS(fifo_expire);
343CFQ_CFQQ_FNS(idle_window);
344CFQ_CFQQ_FNS(prio_changed);
Jens Axboe44f7c162007-01-19 11:51:58 +1100345CFQ_CFQQ_FNS(slice_new);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200346CFQ_CFQQ_FNS(sync);
Jens Axboea36e71f2009-04-15 12:15:11 +0200347CFQ_CFQQ_FNS(coop);
Corrado Zoccolo76280af2009-11-26 10:02:58 +0100348CFQ_CFQQ_FNS(deep);
Vivek Goyalf75edf22009-12-03 12:59:53 -0500349CFQ_CFQQ_FNS(wait_busy);
Jens Axboe3b181522005-06-27 10:56:24 +0200350#undef CFQ_CFQQ_FNS
351
Vivek Goyal2868ef72009-12-03 12:59:48 -0500352#ifdef CONFIG_DEBUG_CFQ_IOSCHED
353#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
354 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
355 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
356 blkg_path(&(cfqq)->cfqg->blkg), ##args);
357
358#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
359 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
360 blkg_path(&(cfqg)->blkg), ##args); \
361
362#else
Jens Axboe7b679132008-05-30 12:23:07 +0200363#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
364 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
Vivek Goyal2868ef72009-12-03 12:59:48 -0500365#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
366#endif
Jens Axboe7b679132008-05-30 12:23:07 +0200367#define cfq_log(cfqd, fmt, args...) \
368 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
369
Vivek Goyal615f0252009-12-03 12:59:39 -0500370/* Traverses through cfq group service trees */
371#define for_each_cfqg_st(cfqg, i, j, st) \
372 for (i = 0; i <= IDLE_WORKLOAD; i++) \
373 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
374 : &cfqg->service_tree_idle; \
375 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
376 (i == IDLE_WORKLOAD && j == 0); \
377 j++, st = i < IDLE_WORKLOAD ? \
378 &cfqg->service_trees[i][j]: NULL) \
379
380
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100381static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
382{
383 if (cfq_class_idle(cfqq))
384 return IDLE_WORKLOAD;
385 if (cfq_class_rt(cfqq))
386 return RT_WORKLOAD;
387 return BE_WORKLOAD;
388}
389
Corrado Zoccolo718eee02009-10-26 22:45:29 +0100390
391static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
392{
393 if (!cfq_cfqq_sync(cfqq))
394 return ASYNC_WORKLOAD;
395 if (!cfq_cfqq_idle_window(cfqq))
396 return SYNC_NOIDLE_WORKLOAD;
397 return SYNC_WORKLOAD;
398}
399
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500400static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
401 struct cfq_data *cfqd,
402 struct cfq_group *cfqg)
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100403{
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500404 if (wl == IDLE_WORKLOAD)
405 return cfqg->service_tree_idle.count;
406
407 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
408 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
409 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +0100410}
411
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500412static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
413 struct cfq_group *cfqg)
414{
415 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
416 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
417}
418
Jens Axboe165125e2007-07-24 09:28:11 +0200419static void cfq_dispatch_insert(struct request_queue *, struct request *);
Jens Axboea6151c32009-10-07 20:02:57 +0200420static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
Jens Axboefd0928d2008-01-24 08:52:45 +0100421 struct io_context *, gfp_t);
Jens Axboe4ac845a2008-01-24 08:44:49 +0100422static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
Vasily Tarasov91fac312007-04-25 12:29:51 +0200423 struct io_context *);
424
Jens Axboe5ad531d2009-07-03 12:57:48 +0200425static inline int rq_in_driver(struct cfq_data *cfqd)
426{
427 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
428}
429
Vasily Tarasov91fac312007-04-25 12:29:51 +0200430static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200431 bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200432{
Jens Axboea6151c32009-10-07 20:02:57 +0200433 return cic->cfqq[is_sync];
Vasily Tarasov91fac312007-04-25 12:29:51 +0200434}
435
436static inline void cic_set_cfqq(struct cfq_io_context *cic,
Jens Axboea6151c32009-10-07 20:02:57 +0200437 struct cfq_queue *cfqq, bool is_sync)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200438{
Jens Axboea6151c32009-10-07 20:02:57 +0200439 cic->cfqq[is_sync] = cfqq;
Vasily Tarasov91fac312007-04-25 12:29:51 +0200440}
441
442/*
443 * We regard a request as SYNC, if it's either a read or has the SYNC bit
444 * set (in which case it could also be direct WRITE).
445 */
Jens Axboea6151c32009-10-07 20:02:57 +0200446static inline bool cfq_bio_sync(struct bio *bio)
Vasily Tarasov91fac312007-04-25 12:29:51 +0200447{
Jens Axboea6151c32009-10-07 20:02:57 +0200448 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
Vasily Tarasov91fac312007-04-25 12:29:51 +0200449}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451/*
Andrew Morton99f95e52005-06-27 20:14:05 -0700452 * scheduler run of queue, if there are requests pending and no one in the
453 * driver that will restart queueing
454 */
Jens Axboe23e018a2009-10-05 08:52:35 +0200455static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
Andrew Morton99f95e52005-06-27 20:14:05 -0700456{
Jens Axboe7b679132008-05-30 12:23:07 +0200457 if (cfqd->busy_queues) {
458 cfq_log(cfqd, "schedule dispatch");
Jens Axboe23e018a2009-10-05 08:52:35 +0200459 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
Jens Axboe7b679132008-05-30 12:23:07 +0200460 }
Andrew Morton99f95e52005-06-27 20:14:05 -0700461}
462
Jens Axboe165125e2007-07-24 09:28:11 +0200463static int cfq_queue_empty(struct request_queue *q)
Andrew Morton99f95e52005-06-27 20:14:05 -0700464{
465 struct cfq_data *cfqd = q->elevator->elevator_data;
466
Vivek Goyalf04a6422009-12-03 12:59:40 -0500467 return !cfqd->rq_queued;
Andrew Morton99f95e52005-06-27 20:14:05 -0700468}
469
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470/*
Jens Axboe44f7c162007-01-19 11:51:58 +1100471 * Scale schedule slice based on io priority. Use the sync time slice only
472 * if a queue is marked sync and has sync io queued. A sync queue with async
473 * io only, should not get full sync slice length.
474 */
Jens Axboea6151c32009-10-07 20:02:57 +0200475static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
Jens Axboed9e76202007-04-20 14:27:50 +0200476 unsigned short prio)
477{
478 const int base_slice = cfqd->cfq_slice[sync];
479
480 WARN_ON(prio >= IOPRIO_BE_NR);
481
482 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
483}
484
Jens Axboe44f7c162007-01-19 11:51:58 +1100485static inline int
486cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
487{
Jens Axboed9e76202007-04-20 14:27:50 +0200488 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
Jens Axboe44f7c162007-01-19 11:51:58 +1100489}
490
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500491static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
492{
493 u64 d = delta << CFQ_SERVICE_SHIFT;
494
495 d = d * BLKIO_WEIGHT_DEFAULT;
496 do_div(d, cfqg->weight);
497 return d;
498}
499
500static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
501{
502 s64 delta = (s64)(vdisktime - min_vdisktime);
503 if (delta > 0)
504 min_vdisktime = vdisktime;
505
506 return min_vdisktime;
507}
508
509static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
510{
511 s64 delta = (s64)(vdisktime - min_vdisktime);
512 if (delta < 0)
513 min_vdisktime = vdisktime;
514
515 return min_vdisktime;
516}
517
518static void update_min_vdisktime(struct cfq_rb_root *st)
519{
520 u64 vdisktime = st->min_vdisktime;
521 struct cfq_group *cfqg;
522
523 if (st->active) {
524 cfqg = rb_entry_cfqg(st->active);
525 vdisktime = cfqg->vdisktime;
526 }
527
528 if (st->left) {
529 cfqg = rb_entry_cfqg(st->left);
530 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
531 }
532
533 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
534}
535
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100536/*
537 * get averaged number of queues of RT/BE priority.
538 * average is updated, with a formula that gives more weight to higher numbers,
539 * to quickly follows sudden increases and decrease slowly
540 */
541
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500542static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
543 struct cfq_group *cfqg, bool rt)
Jens Axboe5869619c2009-10-28 09:27:07 +0100544{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100545 unsigned min_q, max_q;
546 unsigned mult = cfq_hist_divisor - 1;
547 unsigned round = cfq_hist_divisor / 2;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500548 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100549
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500550 min_q = min(cfqg->busy_queues_avg[rt], busy);
551 max_q = max(cfqg->busy_queues_avg[rt], busy);
552 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100553 cfq_hist_divisor;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500554 return cfqg->busy_queues_avg[rt];
555}
556
557static inline unsigned
558cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
559{
560 struct cfq_rb_root *st = &cfqd->grp_service_tree;
561
562 return cfq_target_latency * cfqg->weight / st->total_weight;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100563}
564
Jens Axboe44f7c162007-01-19 11:51:58 +1100565static inline void
566cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
567{
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100568 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
569 if (cfqd->cfq_latency) {
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500570 /*
571 * interested queues (we consider only the ones with the same
572 * priority class in the cfq group)
573 */
574 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
575 cfq_class_rt(cfqq));
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100576 unsigned sync_slice = cfqd->cfq_slice[1];
577 unsigned expect_latency = sync_slice * iq;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500578 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
579
580 if (expect_latency > group_slice) {
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100581 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
582 /* scale low_slice according to IO priority
583 * and sync vs async */
584 unsigned low_slice =
585 min(slice, base_low_slice * slice / sync_slice);
586 /* the adapted slice value is scaled to fit all iqs
587 * into the target latency */
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500588 slice = max(slice * group_slice / expect_latency,
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100589 low_slice);
590 }
591 }
Vivek Goyaldae739e2009-12-03 12:59:45 -0500592 cfqq->slice_start = jiffies;
Corrado Zoccolo5db5d642009-10-26 22:44:04 +0100593 cfqq->slice_end = jiffies + slice;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500594 cfqq->allocated_slice = slice;
Jens Axboe7b679132008-05-30 12:23:07 +0200595 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
Jens Axboe44f7c162007-01-19 11:51:58 +1100596}
597
598/*
599 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
600 * isn't valid until the first request from the dispatch is activated
601 * and the slice time set.
602 */
Jens Axboea6151c32009-10-07 20:02:57 +0200603static inline bool cfq_slice_used(struct cfq_queue *cfqq)
Jens Axboe44f7c162007-01-19 11:51:58 +1100604{
605 if (cfq_cfqq_slice_new(cfqq))
606 return 0;
607 if (time_before(jiffies, cfqq->slice_end))
608 return 0;
609
610 return 1;
611}
612
613/*
Jens Axboe5e705372006-07-13 12:39:25 +0200614 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 * We choose the request that is closest to the head right now. Distance
Andreas Mohre8a99052006-03-28 08:59:49 +0200616 * behind the head is penalized and only allowed to a certain extent.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 */
Jens Axboe5e705372006-07-13 12:39:25 +0200618static struct request *
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100619cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100621 sector_t s1, s2, d1 = 0, d2 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 unsigned long back_max;
Andreas Mohre8a99052006-03-28 08:59:49 +0200623#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
624#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
625 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626
Jens Axboe5e705372006-07-13 12:39:25 +0200627 if (rq1 == NULL || rq1 == rq2)
628 return rq2;
629 if (rq2 == NULL)
630 return rq1;
Jens Axboe9c2c38a2005-08-24 14:57:54 +0200631
Jens Axboe5e705372006-07-13 12:39:25 +0200632 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
633 return rq1;
634 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
635 return rq2;
Jens Axboe374f84a2006-07-23 01:42:19 +0200636 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
637 return rq1;
638 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
639 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640
Tejun Heo83096eb2009-05-07 22:24:39 +0900641 s1 = blk_rq_pos(rq1);
642 s2 = blk_rq_pos(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 /*
645 * by definition, 1KiB is 2 sectors
646 */
647 back_max = cfqd->cfq_back_max * 2;
648
649 /*
650 * Strict one way elevator _except_ in the case where we allow
651 * short backward seeks which are biased as twice the cost of a
652 * similar forward seek.
653 */
654 if (s1 >= last)
655 d1 = s1 - last;
656 else if (s1 + back_max >= last)
657 d1 = (last - s1) * cfqd->cfq_back_penalty;
658 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200659 wrap |= CFQ_RQ1_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
661 if (s2 >= last)
662 d2 = s2 - last;
663 else if (s2 + back_max >= last)
664 d2 = (last - s2) * cfqd->cfq_back_penalty;
665 else
Andreas Mohre8a99052006-03-28 08:59:49 +0200666 wrap |= CFQ_RQ2_WRAP;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667
668 /* Found required data */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
Andreas Mohre8a99052006-03-28 08:59:49 +0200670 /*
671 * By doing switch() on the bit mask "wrap" we avoid having to
672 * check two variables for all permutations: --> faster!
673 */
674 switch (wrap) {
Jens Axboe5e705372006-07-13 12:39:25 +0200675 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200676 if (d1 < d2)
Jens Axboe5e705372006-07-13 12:39:25 +0200677 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200678 else if (d2 < d1)
Jens Axboe5e705372006-07-13 12:39:25 +0200679 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200680 else {
681 if (s1 >= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200682 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200683 else
Jens Axboe5e705372006-07-13 12:39:25 +0200684 return rq2;
Andreas Mohre8a99052006-03-28 08:59:49 +0200685 }
686
687 case CFQ_RQ2_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200688 return rq1;
Andreas Mohre8a99052006-03-28 08:59:49 +0200689 case CFQ_RQ1_WRAP:
Jens Axboe5e705372006-07-13 12:39:25 +0200690 return rq2;
691 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
Andreas Mohre8a99052006-03-28 08:59:49 +0200692 default:
693 /*
694 * Since both rqs are wrapped,
695 * start with the one that's further behind head
696 * (--> only *one* back seek required),
697 * since back seek takes more time than forward.
698 */
699 if (s1 <= s2)
Jens Axboe5e705372006-07-13 12:39:25 +0200700 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701 else
Jens Axboe5e705372006-07-13 12:39:25 +0200702 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 }
704}
705
Jens Axboe498d3aa22007-04-26 12:54:48 +0200706/*
707 * The below is leftmost cache rbtree addon
708 */
Jens Axboe08717142008-01-28 11:38:15 +0100709static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
Jens Axboecc09e292007-04-26 12:53:50 +0200710{
Vivek Goyal615f0252009-12-03 12:59:39 -0500711 /* Service tree is empty */
712 if (!root->count)
713 return NULL;
714
Jens Axboecc09e292007-04-26 12:53:50 +0200715 if (!root->left)
716 root->left = rb_first(&root->rb);
717
Jens Axboe08717142008-01-28 11:38:15 +0100718 if (root->left)
719 return rb_entry(root->left, struct cfq_queue, rb_node);
720
721 return NULL;
Jens Axboecc09e292007-04-26 12:53:50 +0200722}
723
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500724static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
725{
726 if (!root->left)
727 root->left = rb_first(&root->rb);
728
729 if (root->left)
730 return rb_entry_cfqg(root->left);
731
732 return NULL;
733}
734
Jens Axboea36e71f2009-04-15 12:15:11 +0200735static void rb_erase_init(struct rb_node *n, struct rb_root *root)
736{
737 rb_erase(n, root);
738 RB_CLEAR_NODE(n);
739}
740
Jens Axboecc09e292007-04-26 12:53:50 +0200741static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
742{
743 if (root->left == n)
744 root->left = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +0200745 rb_erase_init(n, &root->rb);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +0100746 --root->count;
Jens Axboecc09e292007-04-26 12:53:50 +0200747}
748
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749/*
750 * would be nice to take fifo expire time into account as well
751 */
Jens Axboe5e705372006-07-13 12:39:25 +0200752static struct request *
753cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
754 struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755{
Jens Axboe21183b02006-07-13 12:33:14 +0200756 struct rb_node *rbnext = rb_next(&last->rb_node);
757 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe5e705372006-07-13 12:39:25 +0200758 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
Jens Axboe21183b02006-07-13 12:33:14 +0200760 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761
762 if (rbprev)
Jens Axboe5e705372006-07-13 12:39:25 +0200763 prev = rb_entry_rq(rbprev);
Jens Axboe21183b02006-07-13 12:33:14 +0200764
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 if (rbnext)
Jens Axboe5e705372006-07-13 12:39:25 +0200766 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200767 else {
768 rbnext = rb_first(&cfqq->sort_list);
769 if (rbnext && rbnext != &last->rb_node)
Jens Axboe5e705372006-07-13 12:39:25 +0200770 next = rb_entry_rq(rbnext);
Jens Axboe21183b02006-07-13 12:33:14 +0200771 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +0100773 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774}
775
Jens Axboed9e76202007-04-20 14:27:50 +0200776static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
777 struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778{
Jens Axboed9e76202007-04-20 14:27:50 +0200779 /*
780 * just an approximation, should be ok.
781 */
Vivek Goyalcdb16e82009-12-03 12:59:38 -0500782 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
Jens Axboe464191c2009-11-30 09:38:13 +0100783 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
Jens Axboed9e76202007-04-20 14:27:50 +0200784}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500786static inline s64
787cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
788{
789 return cfqg->vdisktime - st->min_vdisktime;
790}
791
792static void
793__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
794{
795 struct rb_node **node = &st->rb.rb_node;
796 struct rb_node *parent = NULL;
797 struct cfq_group *__cfqg;
798 s64 key = cfqg_key(st, cfqg);
799 int left = 1;
800
801 while (*node != NULL) {
802 parent = *node;
803 __cfqg = rb_entry_cfqg(parent);
804
805 if (key < cfqg_key(st, __cfqg))
806 node = &parent->rb_left;
807 else {
808 node = &parent->rb_right;
809 left = 0;
810 }
811 }
812
813 if (left)
814 st->left = &cfqg->rb_node;
815
816 rb_link_node(&cfqg->rb_node, parent, node);
817 rb_insert_color(&cfqg->rb_node, &st->rb);
818}
819
820static void
821cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
822{
823 struct cfq_rb_root *st = &cfqd->grp_service_tree;
824 struct cfq_group *__cfqg;
825 struct rb_node *n;
826
827 cfqg->nr_cfqq++;
828 if (cfqg->on_st)
829 return;
830
831 /*
832 * Currently put the group at the end. Later implement something
833 * so that groups get lesser vtime based on their weights, so that
834 * if group does not loose all if it was not continously backlogged.
835 */
836 n = rb_last(&st->rb);
837 if (n) {
838 __cfqg = rb_entry_cfqg(n);
839 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
840 } else
841 cfqg->vdisktime = st->min_vdisktime;
842
843 __cfq_group_service_tree_add(st, cfqg);
844 cfqg->on_st = true;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500845 cfqd->nr_groups++;
846 st->total_weight += cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500847}
848
849static void
850cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
851{
852 struct cfq_rb_root *st = &cfqd->grp_service_tree;
853
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500854 if (st->active == &cfqg->rb_node)
855 st->active = NULL;
856
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500857 BUG_ON(cfqg->nr_cfqq < 1);
858 cfqg->nr_cfqq--;
Vivek Goyal25bc6b02009-12-03 12:59:43 -0500859
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500860 /* If there are other cfq queues under this group, don't delete it */
861 if (cfqg->nr_cfqq)
862 return;
863
Vivek Goyal2868ef72009-12-03 12:59:48 -0500864 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500865 cfqg->on_st = false;
Vivek Goyal58ff82f2009-12-03 12:59:44 -0500866 cfqd->nr_groups--;
867 st->total_weight -= cfqg->weight;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500868 if (!RB_EMPTY_NODE(&cfqg->rb_node))
869 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500870 cfqg->saved_workload_slice = 0;
Vivek Goyal22084192009-12-03 12:59:49 -0500871 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500872}
873
874static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
875{
Vivek Goyalf75edf22009-12-03 12:59:53 -0500876 unsigned int slice_used;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500877
878 /*
879 * Queue got expired before even a single request completed or
880 * got expired immediately after first request completion.
881 */
882 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
883 /*
884 * Also charge the seek time incurred to the group, otherwise
885 * if there are mutiple queues in the group, each can dispatch
886 * a single request on seeky media and cause lots of seek time
887 * and group will never know it.
888 */
889 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
890 1);
891 } else {
892 slice_used = jiffies - cfqq->slice_start;
Vivek Goyalf75edf22009-12-03 12:59:53 -0500893 if (slice_used > cfqq->allocated_slice)
894 slice_used = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500895 }
896
Vivek Goyal22084192009-12-03 12:59:49 -0500897 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
898 cfqq->nr_sectors);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500899 return slice_used;
900}
901
902static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
903 struct cfq_queue *cfqq)
904{
905 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500906 unsigned int used_sl, charge_sl;
907 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
908 - cfqg->service_tree_idle.count;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500909
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500910 BUG_ON(nr_sync < 0);
911 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
912
913 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
914 charge_sl = cfqq->allocated_slice;
Vivek Goyaldae739e2009-12-03 12:59:45 -0500915
916 /* Can't update vdisktime while group is on service tree */
917 cfq_rb_erase(&cfqg->rb_node, st);
Vivek Goyalf26bd1f2009-12-03 12:59:54 -0500918 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
Vivek Goyaldae739e2009-12-03 12:59:45 -0500919 __cfq_group_service_tree_add(st, cfqg);
920
921 /* This group is being expired. Save the context */
922 if (time_after(cfqd->workload_expires, jiffies)) {
923 cfqg->saved_workload_slice = cfqd->workload_expires
924 - jiffies;
925 cfqg->saved_workload = cfqd->serving_type;
926 cfqg->saved_serving_prio = cfqd->serving_prio;
927 } else
928 cfqg->saved_workload_slice = 0;
Vivek Goyal2868ef72009-12-03 12:59:48 -0500929
930 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
931 st->min_vdisktime);
Vivek Goyal22084192009-12-03 12:59:49 -0500932 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
933 cfqq->nr_sectors);
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -0500934}
935
Vivek Goyal25fb5162009-12-03 12:59:46 -0500936#ifdef CONFIG_CFQ_GROUP_IOSCHED
937static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
938{
939 if (blkg)
940 return container_of(blkg, struct cfq_group, blkg);
941 return NULL;
942}
943
Vivek Goyalf8d461d2009-12-03 12:59:52 -0500944void
945cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
946{
947 cfqg_of_blkg(blkg)->weight = weight;
948}
949
Vivek Goyal25fb5162009-12-03 12:59:46 -0500950static struct cfq_group *
951cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
952{
953 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
954 struct cfq_group *cfqg = NULL;
955 void *key = cfqd;
956 int i, j;
957 struct cfq_rb_root *st;
Vivek Goyal22084192009-12-03 12:59:49 -0500958 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
959 unsigned int major, minor;
Vivek Goyal25fb5162009-12-03 12:59:46 -0500960
961 /* Do we need to take this reference */
Vivek Goyal9d6a9862009-12-04 10:36:41 -0500962 if (!blkiocg_css_tryget(blkcg))
Vivek Goyal25fb5162009-12-03 12:59:46 -0500963 return NULL;;
964
965 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
966 if (cfqg || !create)
967 goto done;
968
969 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
970 if (!cfqg)
971 goto done;
972
973 cfqg->weight = blkcg->weight;
974 for_each_cfqg_st(cfqg, i, j, st)
975 *st = CFQ_RB_ROOT;
976 RB_CLEAR_NODE(&cfqg->rb_node);
977
Vivek Goyalb1c35762009-12-03 12:59:47 -0500978 /*
979 * Take the initial reference that will be released on destroy
980 * This can be thought of a joint reference by cgroup and
981 * elevator which will be dropped by either elevator exit
982 * or cgroup deletion path depending on who is exiting first.
983 */
984 atomic_set(&cfqg->ref, 1);
985
Vivek Goyal25fb5162009-12-03 12:59:46 -0500986 /* Add group onto cgroup list */
Vivek Goyal22084192009-12-03 12:59:49 -0500987 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
988 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
989 MKDEV(major, minor));
Vivek Goyal25fb5162009-12-03 12:59:46 -0500990
991 /* Add group on cfqd list */
992 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
993
994done:
Vivek Goyal9d6a9862009-12-04 10:36:41 -0500995 blkiocg_css_put(blkcg);
Vivek Goyal25fb5162009-12-03 12:59:46 -0500996 return cfqg;
997}
998
999/*
1000 * Search for the cfq group current task belongs to. If create = 1, then also
1001 * create the cfq group if it does not exist. request_queue lock must be held.
1002 */
1003static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1004{
1005 struct cgroup *cgroup;
1006 struct cfq_group *cfqg = NULL;
1007
1008 rcu_read_lock();
1009 cgroup = task_cgroup(current, blkio_subsys_id);
1010 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1011 if (!cfqg && create)
1012 cfqg = &cfqd->root_group;
1013 rcu_read_unlock();
1014 return cfqg;
1015}
1016
1017static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1018{
1019 /* Currently, all async queues are mapped to root group */
1020 if (!cfq_cfqq_sync(cfqq))
1021 cfqg = &cfqq->cfqd->root_group;
1022
1023 cfqq->cfqg = cfqg;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001024 /* cfqq reference on cfqg */
1025 atomic_inc(&cfqq->cfqg->ref);
Vivek Goyal25fb5162009-12-03 12:59:46 -05001026}
Vivek Goyalb1c35762009-12-03 12:59:47 -05001027
1028static void cfq_put_cfqg(struct cfq_group *cfqg)
1029{
1030 struct cfq_rb_root *st;
1031 int i, j;
1032
1033 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1034 if (!atomic_dec_and_test(&cfqg->ref))
1035 return;
1036 for_each_cfqg_st(cfqg, i, j, st)
1037 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1038 kfree(cfqg);
1039}
1040
1041static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1042{
1043 /* Something wrong if we are trying to remove same group twice */
1044 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1045
1046 hlist_del_init(&cfqg->cfqd_node);
1047
1048 /*
1049 * Put the reference taken at the time of creation so that when all
1050 * queues are gone, group can be destroyed.
1051 */
1052 cfq_put_cfqg(cfqg);
1053}
1054
1055static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1056{
1057 struct hlist_node *pos, *n;
1058 struct cfq_group *cfqg;
1059
1060 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1061 /*
1062 * If cgroup removal path got to blk_group first and removed
1063 * it from cgroup list, then it will take care of destroying
1064 * cfqg also.
1065 */
1066 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1067 cfq_destroy_cfqg(cfqd, cfqg);
1068 }
1069}
1070
1071/*
1072 * Blk cgroup controller notification saying that blkio_group object is being
1073 * delinked as associated cgroup object is going away. That also means that
1074 * no new IO will come in this group. So get rid of this group as soon as
1075 * any pending IO in the group is finished.
1076 *
1077 * This function is called under rcu_read_lock(). key is the rcu protected
1078 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1079 * read lock.
1080 *
1081 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1082 * it should not be NULL as even if elevator was exiting, cgroup deltion
1083 * path got to it first.
1084 */
1085void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1086{
1087 unsigned long flags;
1088 struct cfq_data *cfqd = key;
1089
1090 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1091 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1092 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1093}
1094
Vivek Goyal25fb5162009-12-03 12:59:46 -05001095#else /* GROUP_IOSCHED */
1096static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1097{
1098 return &cfqd->root_group;
1099}
1100static inline void
1101cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1102 cfqq->cfqg = cfqg;
1103}
1104
Vivek Goyalb1c35762009-12-03 12:59:47 -05001105static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1106static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1107
Vivek Goyal25fb5162009-12-03 12:59:46 -05001108#endif /* GROUP_IOSCHED */
1109
Jens Axboe498d3aa22007-04-26 12:54:48 +02001110/*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001111 * The cfqd->service_trees holds all pending cfq_queue's that have
Jens Axboe498d3aa22007-04-26 12:54:48 +02001112 * requests waiting to be processed. It is sorted in the order that
1113 * we will service the queues.
1114 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001115static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001116 bool add_front)
Jens Axboed9e76202007-04-20 14:27:50 +02001117{
Jens Axboe08717142008-01-28 11:38:15 +01001118 struct rb_node **p, *parent;
1119 struct cfq_queue *__cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02001120 unsigned long rb_key;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001121 struct cfq_rb_root *service_tree;
Jens Axboe498d3aa22007-04-26 12:54:48 +02001122 int left;
Vivek Goyaldae739e2009-12-03 12:59:45 -05001123 int new_cfqq = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05001124 int group_changed = 0;
1125
1126#ifdef CONFIG_CFQ_GROUP_IOSCHED
1127 if (!cfqd->cfq_group_isolation
1128 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1129 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1130 /* Move this cfq to root group */
1131 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1132 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1133 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1134 cfqq->orig_cfqg = cfqq->cfqg;
1135 cfqq->cfqg = &cfqd->root_group;
1136 atomic_inc(&cfqd->root_group.ref);
1137 group_changed = 1;
1138 } else if (!cfqd->cfq_group_isolation
1139 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1140 /* cfqq is sequential now needs to go to its original group */
1141 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1142 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1143 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1144 cfq_put_cfqg(cfqq->cfqg);
1145 cfqq->cfqg = cfqq->orig_cfqg;
1146 cfqq->orig_cfqg = NULL;
1147 group_changed = 1;
1148 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1149 }
1150#endif
Jens Axboed9e76202007-04-20 14:27:50 +02001151
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001152 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1153 cfqq_type(cfqq), cfqd);
Jens Axboe08717142008-01-28 11:38:15 +01001154 if (cfq_class_idle(cfqq)) {
1155 rb_key = CFQ_IDLE_DELAY;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001156 parent = rb_last(&service_tree->rb);
Jens Axboe08717142008-01-28 11:38:15 +01001157 if (parent && parent != &cfqq->rb_node) {
1158 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1159 rb_key += __cfqq->rb_key;
1160 } else
1161 rb_key += jiffies;
1162 } else if (!add_front) {
Jens Axboeb9c89462009-10-06 20:53:44 +02001163 /*
1164 * Get our rb key offset. Subtract any residual slice
1165 * value carried from last service. A negative resid
1166 * count indicates slice overrun, and this should position
1167 * the next service time further away in the tree.
1168 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001169 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
Jens Axboeb9c89462009-10-06 20:53:44 +02001170 rb_key -= cfqq->slice_resid;
Jens Axboeedd75ff2007-04-19 12:03:34 +02001171 cfqq->slice_resid = 0;
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001172 } else {
1173 rb_key = -HZ;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001174 __cfqq = cfq_rb_first(service_tree);
Corrado Zoccolo48e025e2009-10-05 08:49:23 +02001175 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1176 }
Jens Axboed9e76202007-04-20 14:27:50 +02001177
1178 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
Vivek Goyaldae739e2009-12-03 12:59:45 -05001179 new_cfqq = 0;
Jens Axboe99f96282007-02-05 11:56:25 +01001180 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001181 * same position, nothing more to do
Jens Axboe99f96282007-02-05 11:56:25 +01001182 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001183 if (rb_key == cfqq->rb_key &&
1184 cfqq->service_tree == service_tree)
Jens Axboed9e76202007-04-20 14:27:50 +02001185 return;
Jens Axboe53b037442006-07-28 09:48:51 +02001186
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001187 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1188 cfqq->service_tree = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02001189 }
Jens Axboed9e76202007-04-20 14:27:50 +02001190
Jens Axboe498d3aa22007-04-26 12:54:48 +02001191 left = 1;
Jens Axboe08717142008-01-28 11:38:15 +01001192 parent = NULL;
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001193 cfqq->service_tree = service_tree;
1194 p = &service_tree->rb.rb_node;
Jens Axboed9e76202007-04-20 14:27:50 +02001195 while (*p) {
Jens Axboe67060e32007-04-18 20:13:32 +02001196 struct rb_node **n;
Jens Axboecc09e292007-04-26 12:53:50 +02001197
Jens Axboed9e76202007-04-20 14:27:50 +02001198 parent = *p;
1199 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1200
Jens Axboe0c534e02007-04-18 20:01:57 +02001201 /*
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001202 * sort by key, that represents service time.
Jens Axboe0c534e02007-04-18 20:01:57 +02001203 */
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001204 if (time_before(rb_key, __cfqq->rb_key))
Jens Axboe67060e32007-04-18 20:13:32 +02001205 n = &(*p)->rb_left;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001206 else {
Jens Axboe67060e32007-04-18 20:13:32 +02001207 n = &(*p)->rb_right;
Jens Axboecc09e292007-04-26 12:53:50 +02001208 left = 0;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001209 }
Jens Axboe67060e32007-04-18 20:13:32 +02001210
1211 p = n;
Jens Axboed9e76202007-04-20 14:27:50 +02001212 }
1213
Jens Axboecc09e292007-04-26 12:53:50 +02001214 if (left)
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001215 service_tree->left = &cfqq->rb_node;
Jens Axboecc09e292007-04-26 12:53:50 +02001216
Jens Axboed9e76202007-04-20 14:27:50 +02001217 cfqq->rb_key = rb_key;
1218 rb_link_node(&cfqq->rb_node, parent, p);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001219 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1220 service_tree->count++;
Vivek Goyalae30c282009-12-03 12:59:55 -05001221 if ((add_front || !new_cfqq) && !group_changed)
Vivek Goyaldae739e2009-12-03 12:59:45 -05001222 return;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001223 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224}
1225
Jens Axboea36e71f2009-04-15 12:15:11 +02001226static struct cfq_queue *
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001227cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1228 sector_t sector, struct rb_node **ret_parent,
1229 struct rb_node ***rb_link)
Jens Axboea36e71f2009-04-15 12:15:11 +02001230{
Jens Axboea36e71f2009-04-15 12:15:11 +02001231 struct rb_node **p, *parent;
1232 struct cfq_queue *cfqq = NULL;
1233
1234 parent = NULL;
1235 p = &root->rb_node;
1236 while (*p) {
1237 struct rb_node **n;
1238
1239 parent = *p;
1240 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1241
1242 /*
1243 * Sort strictly based on sector. Smallest to the left,
1244 * largest to the right.
1245 */
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001246 if (sector > blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001247 n = &(*p)->rb_right;
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001248 else if (sector < blk_rq_pos(cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001249 n = &(*p)->rb_left;
1250 else
1251 break;
1252 p = n;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001253 cfqq = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001254 }
1255
1256 *ret_parent = parent;
1257 if (rb_link)
1258 *rb_link = p;
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001259 return cfqq;
Jens Axboea36e71f2009-04-15 12:15:11 +02001260}
1261
1262static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1263{
Jens Axboea36e71f2009-04-15 12:15:11 +02001264 struct rb_node **p, *parent;
1265 struct cfq_queue *__cfqq;
1266
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001267 if (cfqq->p_root) {
1268 rb_erase(&cfqq->p_node, cfqq->p_root);
1269 cfqq->p_root = NULL;
1270 }
Jens Axboea36e71f2009-04-15 12:15:11 +02001271
1272 if (cfq_class_idle(cfqq))
1273 return;
1274 if (!cfqq->next_rq)
1275 return;
1276
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001277 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001278 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1279 blk_rq_pos(cfqq->next_rq), &parent, &p);
Jens Axboe3ac6c9f2009-04-23 12:14:56 +02001280 if (!__cfqq) {
1281 rb_link_node(&cfqq->p_node, parent, p);
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001282 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1283 } else
1284 cfqq->p_root = NULL;
Jens Axboea36e71f2009-04-15 12:15:11 +02001285}
1286
Jens Axboe498d3aa22007-04-26 12:54:48 +02001287/*
1288 * Update cfqq's position in the service tree.
1289 */
Jens Axboeedd75ff2007-04-19 12:03:34 +02001290static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001291{
Jens Axboe6d048f52007-04-25 12:44:27 +02001292 /*
1293 * Resorting requires the cfqq to be on the RR list already.
1294 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001295 if (cfq_cfqq_on_rr(cfqq)) {
Jens Axboeedd75ff2007-04-19 12:03:34 +02001296 cfq_service_tree_add(cfqd, cfqq, 0);
Jens Axboea36e71f2009-04-15 12:15:11 +02001297 cfq_prio_tree_add(cfqd, cfqq);
1298 }
Jens Axboe6d048f52007-04-25 12:44:27 +02001299}
1300
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301/*
1302 * add to busy list of queues for service, trying to be fair in ordering
Jens Axboe22e2c502005-06-27 10:55:12 +02001303 * the pending list according to last request service
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 */
Jens Axboefebffd62008-01-28 13:19:43 +01001305static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306{
Jens Axboe7b679132008-05-30 12:23:07 +02001307 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001308 BUG_ON(cfq_cfqq_on_rr(cfqq));
1309 cfq_mark_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 cfqd->busy_queues++;
1311
Jens Axboeedd75ff2007-04-19 12:03:34 +02001312 cfq_resort_rr_list(cfqd, cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313}
1314
Jens Axboe498d3aa22007-04-26 12:54:48 +02001315/*
1316 * Called when the cfqq no longer has requests pending, remove it from
1317 * the service tree.
1318 */
Jens Axboefebffd62008-01-28 13:19:43 +01001319static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320{
Jens Axboe7b679132008-05-30 12:23:07 +02001321 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
Jens Axboe3b181522005-06-27 10:56:24 +02001322 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1323 cfq_clear_cfqq_on_rr(cfqq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01001325 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1326 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1327 cfqq->service_tree = NULL;
1328 }
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001329 if (cfqq->p_root) {
1330 rb_erase(&cfqq->p_node, cfqq->p_root);
1331 cfqq->p_root = NULL;
1332 }
Jens Axboed9e76202007-04-20 14:27:50 +02001333
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001334 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 BUG_ON(!cfqd->busy_queues);
1336 cfqd->busy_queues--;
1337}
1338
1339/*
1340 * rb tree support functions
1341 */
Jens Axboefebffd62008-01-28 13:19:43 +01001342static void cfq_del_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343{
Jens Axboe5e705372006-07-13 12:39:25 +02001344 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe5e705372006-07-13 12:39:25 +02001345 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346
Jens Axboeb4878f22005-10-20 16:42:29 +02001347 BUG_ON(!cfqq->queued[sync]);
1348 cfqq->queued[sync]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349
Jens Axboe5e705372006-07-13 12:39:25 +02001350 elv_rb_del(&cfqq->sort_list, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351
Vivek Goyalf04a6422009-12-03 12:59:40 -05001352 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1353 /*
1354 * Queue will be deleted from service tree when we actually
1355 * expire it later. Right now just remove it from prio tree
1356 * as it is empty.
1357 */
1358 if (cfqq->p_root) {
1359 rb_erase(&cfqq->p_node, cfqq->p_root);
1360 cfqq->p_root = NULL;
1361 }
1362 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363}
1364
Jens Axboe5e705372006-07-13 12:39:25 +02001365static void cfq_add_rq_rb(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366{
Jens Axboe5e705372006-07-13 12:39:25 +02001367 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboea36e71f2009-04-15 12:15:11 +02001369 struct request *__alias, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370
Jens Axboe5380a102006-07-13 12:37:56 +02001371 cfqq->queued[rq_is_sync(rq)]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
1373 /*
1374 * looks a little odd, but the first insert might return an alias.
1375 * if that happens, put the alias on the dispatch list
1376 */
Jens Axboe21183b02006-07-13 12:33:14 +02001377 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
Jens Axboe5e705372006-07-13 12:39:25 +02001378 cfq_dispatch_insert(cfqd->queue, __alias);
Jens Axboe5fccbf62006-10-31 14:21:55 +01001379
1380 if (!cfq_cfqq_on_rr(cfqq))
1381 cfq_add_cfqq_rr(cfqd, cfqq);
Jens Axboe5044eed2007-04-25 11:53:48 +02001382
1383 /*
1384 * check if this request is a better next-serve candidate
1385 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001386 prev = cfqq->next_rq;
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001387 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
Jens Axboea36e71f2009-04-15 12:15:11 +02001388
1389 /*
1390 * adjust priority tree position, if ->next_rq changes
1391 */
1392 if (prev != cfqq->next_rq)
1393 cfq_prio_tree_add(cfqd, cfqq);
1394
Jens Axboe5044eed2007-04-25 11:53:48 +02001395 BUG_ON(!cfqq->next_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396}
1397
Jens Axboefebffd62008-01-28 13:19:43 +01001398static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399{
Jens Axboe5380a102006-07-13 12:37:56 +02001400 elv_rb_del(&cfqq->sort_list, rq);
1401 cfqq->queued[rq_is_sync(rq)]--;
Jens Axboe5e705372006-07-13 12:39:25 +02001402 cfq_add_rq_rb(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403}
1404
Jens Axboe206dc692006-03-28 13:03:44 +02001405static struct request *
1406cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407{
Jens Axboe206dc692006-03-28 13:03:44 +02001408 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001409 struct cfq_io_context *cic;
Jens Axboe206dc692006-03-28 13:03:44 +02001410 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411
Jens Axboe4ac845a2008-01-24 08:44:49 +01001412 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001413 if (!cic)
1414 return NULL;
1415
1416 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboe89850f72006-07-22 16:48:31 +02001417 if (cfqq) {
1418 sector_t sector = bio->bi_sector + bio_sectors(bio);
1419
Jens Axboe21183b02006-07-13 12:33:14 +02001420 return elv_rb_find(&cfqq->sort_list, sector);
Jens Axboe89850f72006-07-22 16:48:31 +02001421 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 return NULL;
1424}
1425
Jens Axboe165125e2007-07-24 09:28:11 +02001426static void cfq_activate_request(struct request_queue *q, struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +02001427{
1428 struct cfq_data *cfqd = q->elevator->elevator_data;
1429
Jens Axboe5ad531d2009-07-03 12:57:48 +02001430 cfqd->rq_in_driver[rq_is_sync(rq)]++;
Jens Axboe7b679132008-05-30 12:23:07 +02001431 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +02001432 rq_in_driver(cfqd));
Jens Axboe25776e32006-06-01 10:12:26 +02001433
Tejun Heo5b936292009-05-07 22:24:38 +09001434 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02001435}
1436
Jens Axboe165125e2007-07-24 09:28:11 +02001437static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438{
Jens Axboe22e2c502005-06-27 10:55:12 +02001439 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5ad531d2009-07-03 12:57:48 +02001440 const int sync = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441
Jens Axboe5ad531d2009-07-03 12:57:48 +02001442 WARN_ON(!cfqd->rq_in_driver[sync]);
1443 cfqd->rq_in_driver[sync]--;
Jens Axboe7b679132008-05-30 12:23:07 +02001444 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
Jens Axboe5ad531d2009-07-03 12:57:48 +02001445 rq_in_driver(cfqd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446}
1447
Jens Axboeb4878f22005-10-20 16:42:29 +02001448static void cfq_remove_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449{
Jens Axboe5e705372006-07-13 12:39:25 +02001450 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe21183b02006-07-13 12:33:14 +02001451
Jens Axboe5e705372006-07-13 12:39:25 +02001452 if (cfqq->next_rq == rq)
1453 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454
Jens Axboeb4878f22005-10-20 16:42:29 +02001455 list_del_init(&rq->queuelist);
Jens Axboe5e705372006-07-13 12:39:25 +02001456 cfq_del_rq_rb(rq);
Jens Axboe374f84a2006-07-23 01:42:19 +02001457
Aaron Carroll45333d52008-08-26 15:52:36 +02001458 cfqq->cfqd->rq_queued--;
Jens Axboe374f84a2006-07-23 01:42:19 +02001459 if (rq_is_meta(rq)) {
1460 WARN_ON(!cfqq->meta_pending);
1461 cfqq->meta_pending--;
1462 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463}
1464
Jens Axboe165125e2007-07-24 09:28:11 +02001465static int cfq_merge(struct request_queue *q, struct request **req,
1466 struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467{
1468 struct cfq_data *cfqd = q->elevator->elevator_data;
1469 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470
Jens Axboe206dc692006-03-28 13:03:44 +02001471 __rq = cfq_find_rq_fmerge(cfqd, bio);
Jens Axboe22e2c502005-06-27 10:55:12 +02001472 if (__rq && elv_rq_merge_ok(__rq, bio)) {
Jens Axboe98170642006-07-28 09:23:08 +02001473 *req = __rq;
1474 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 }
1476
1477 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478}
1479
Jens Axboe165125e2007-07-24 09:28:11 +02001480static void cfq_merged_request(struct request_queue *q, struct request *req,
Jens Axboe21183b02006-07-13 12:33:14 +02001481 int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482{
Jens Axboe21183b02006-07-13 12:33:14 +02001483 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe5e705372006-07-13 12:39:25 +02001484 struct cfq_queue *cfqq = RQ_CFQQ(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485
Jens Axboe5e705372006-07-13 12:39:25 +02001486 cfq_reposition_rq_rb(cfqq, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488}
1489
1490static void
Jens Axboe165125e2007-07-24 09:28:11 +02001491cfq_merged_requests(struct request_queue *q, struct request *rq,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 struct request *next)
1493{
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001494 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001495 /*
1496 * reposition in fifo if next is older than rq
1497 */
1498 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
Jens Axboe30996f42009-10-05 11:03:39 +02001499 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
Jens Axboe22e2c502005-06-27 10:55:12 +02001500 list_move(&rq->queuelist, &next->queuelist);
Jens Axboe30996f42009-10-05 11:03:39 +02001501 rq_set_fifo_time(rq, rq_fifo_time(next));
1502 }
Jens Axboe22e2c502005-06-27 10:55:12 +02001503
Corrado Zoccolocf7c25c2009-11-08 17:16:46 +01001504 if (cfqq->next_rq == next)
1505 cfqq->next_rq = rq;
Jens Axboeb4878f22005-10-20 16:42:29 +02001506 cfq_remove_request(next);
Jens Axboe22e2c502005-06-27 10:55:12 +02001507}
1508
Jens Axboe165125e2007-07-24 09:28:11 +02001509static int cfq_allow_merge(struct request_queue *q, struct request *rq,
Jens Axboeda775262006-12-20 11:04:12 +01001510 struct bio *bio)
1511{
1512 struct cfq_data *cfqd = q->elevator->elevator_data;
Vasily Tarasov91fac312007-04-25 12:29:51 +02001513 struct cfq_io_context *cic;
Jens Axboeda775262006-12-20 11:04:12 +01001514 struct cfq_queue *cfqq;
Jens Axboeda775262006-12-20 11:04:12 +01001515
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001516 /* Deny merge if bio and rq don't belong to same cfq group */
1517 if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0))
1518 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001519 /*
Jens Axboeec8acb62007-01-02 18:32:11 +01001520 * Disallow merge of a sync bio into an async request.
Jens Axboeda775262006-12-20 11:04:12 +01001521 */
Vasily Tarasov91fac312007-04-25 12:29:51 +02001522 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
Jens Axboea6151c32009-10-07 20:02:57 +02001523 return false;
Jens Axboeda775262006-12-20 11:04:12 +01001524
1525 /*
Jens Axboe719d3402006-12-22 09:38:53 +01001526 * Lookup the cfqq that this bio will be queued with. Allow
1527 * merge only if rq is queued there.
Jens Axboeda775262006-12-20 11:04:12 +01001528 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01001529 cic = cfq_cic_lookup(cfqd, current->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02001530 if (!cic)
Jens Axboea6151c32009-10-07 20:02:57 +02001531 return false;
Jens Axboe719d3402006-12-22 09:38:53 +01001532
Vasily Tarasov91fac312007-04-25 12:29:51 +02001533 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
Jens Axboea6151c32009-10-07 20:02:57 +02001534 return cfqq == RQ_CFQQ(rq);
Jens Axboeda775262006-12-20 11:04:12 +01001535}
1536
Jens Axboefebffd62008-01-28 13:19:43 +01001537static void __cfq_set_active_queue(struct cfq_data *cfqd,
1538 struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02001539{
1540 if (cfqq) {
Jens Axboe7b679132008-05-30 12:23:07 +02001541 cfq_log_cfqq(cfqd, cfqq, "set_active");
Vivek Goyaldae739e2009-12-03 12:59:45 -05001542 cfqq->slice_start = 0;
1543 cfqq->dispatch_start = jiffies;
Vivek Goyalf75edf22009-12-03 12:59:53 -05001544 cfqq->allocated_slice = 0;
Jens Axboe22e2c502005-06-27 10:55:12 +02001545 cfqq->slice_end = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001546 cfqq->slice_dispatch = 0;
Vivek Goyal22084192009-12-03 12:59:49 -05001547 cfqq->nr_sectors = 0;
Jens Axboe2f5cb732009-04-07 08:51:19 +02001548
Jens Axboe2f5cb732009-04-07 08:51:19 +02001549 cfq_clear_cfqq_wait_request(cfqq);
Jens Axboeb0291952009-04-07 11:38:31 +02001550 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001551 cfq_clear_cfqq_must_alloc_slice(cfqq);
1552 cfq_clear_cfqq_fifo_expire(cfqq);
Jens Axboe44f7c162007-01-19 11:51:58 +11001553 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02001554
1555 del_timer(&cfqd->idle_slice_timer);
Jens Axboe22e2c502005-06-27 10:55:12 +02001556 }
1557
1558 cfqd->active_queue = cfqq;
1559}
1560
1561/*
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001562 * current cfqq expired its slice (or was too idle), select new one
1563 */
1564static void
1565__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02001566 bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001567{
Jens Axboe7b679132008-05-30 12:23:07 +02001568 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1569
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001570 if (cfq_cfqq_wait_request(cfqq))
1571 del_timer(&cfqd->idle_slice_timer);
1572
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001573 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalf75edf22009-12-03 12:59:53 -05001574 cfq_clear_cfqq_wait_busy(cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001575
1576 /*
Jens Axboe6084cdd2007-04-23 08:25:00 +02001577 * store what was left of this slice, if the queue idled/timed out
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001578 */
Jens Axboe7b679132008-05-30 12:23:07 +02001579 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
Jens Axboec5b680f2007-01-19 11:56:49 +11001580 cfqq->slice_resid = cfqq->slice_end - jiffies;
Jens Axboe7b679132008-05-30 12:23:07 +02001581 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1582 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001583
Vivek Goyaldae739e2009-12-03 12:59:45 -05001584 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1585
Vivek Goyalf04a6422009-12-03 12:59:40 -05001586 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1587 cfq_del_cfqq_rr(cfqd, cfqq);
1588
Jens Axboeedd75ff2007-04-19 12:03:34 +02001589 cfq_resort_rr_list(cfqd, cfqq);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001590
1591 if (cfqq == cfqd->active_queue)
1592 cfqd->active_queue = NULL;
1593
Vivek Goyaldae739e2009-12-03 12:59:45 -05001594 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1595 cfqd->grp_service_tree.active = NULL;
1596
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001597 if (cfqd->active_cic) {
1598 put_io_context(cfqd->active_cic->ioc);
1599 cfqd->active_cic = NULL;
1600 }
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001601}
1602
Jens Axboea6151c32009-10-07 20:02:57 +02001603static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001604{
1605 struct cfq_queue *cfqq = cfqd->active_queue;
1606
1607 if (cfqq)
Jens Axboe6084cdd2007-04-23 08:25:00 +02001608 __cfq_slice_expired(cfqd, cfqq, timed_out);
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001609}
1610
Jens Axboe498d3aa22007-04-26 12:54:48 +02001611/*
1612 * Get next queue for service. Unless we have a queue preemption,
1613 * we'll simply select the first cfqq in the service tree.
1614 */
Jens Axboe6d048f52007-04-25 12:44:27 +02001615static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001616{
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001617 struct cfq_rb_root *service_tree =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001618 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1619 cfqd->serving_type, cfqd);
Jens Axboeedd75ff2007-04-19 12:03:34 +02001620
Vivek Goyalf04a6422009-12-03 12:59:40 -05001621 if (!cfqd->rq_queued)
1622 return NULL;
1623
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05001624 /* There is nothing to dispatch */
1625 if (!service_tree)
1626 return NULL;
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001627 if (RB_EMPTY_ROOT(&service_tree->rb))
1628 return NULL;
1629 return cfq_rb_first(service_tree);
Jens Axboe6d048f52007-04-25 12:44:27 +02001630}
1631
Vivek Goyalf04a6422009-12-03 12:59:40 -05001632static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1633{
Vivek Goyal25fb5162009-12-03 12:59:46 -05001634 struct cfq_group *cfqg;
Vivek Goyalf04a6422009-12-03 12:59:40 -05001635 struct cfq_queue *cfqq;
1636 int i, j;
1637 struct cfq_rb_root *st;
1638
1639 if (!cfqd->rq_queued)
1640 return NULL;
1641
Vivek Goyal25fb5162009-12-03 12:59:46 -05001642 cfqg = cfq_get_next_cfqg(cfqd);
1643 if (!cfqg)
1644 return NULL;
1645
Vivek Goyalf04a6422009-12-03 12:59:40 -05001646 for_each_cfqg_st(cfqg, i, j, st)
1647 if ((cfqq = cfq_rb_first(st)) != NULL)
1648 return cfqq;
1649 return NULL;
1650}
1651
Jens Axboe498d3aa22007-04-26 12:54:48 +02001652/*
1653 * Get and set a new active queue for service.
1654 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001655static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1656 struct cfq_queue *cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001657{
Jens Axboee00ef792009-11-04 08:54:55 +01001658 if (!cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001659 cfqq = cfq_get_next_queue(cfqd);
Jens Axboe6d048f52007-04-25 12:44:27 +02001660
Jens Axboe22e2c502005-06-27 10:55:12 +02001661 __cfq_set_active_queue(cfqd, cfqq);
Jens Axboe3b181522005-06-27 10:56:24 +02001662 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02001663}
1664
Jens Axboed9e76202007-04-20 14:27:50 +02001665static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1666 struct request *rq)
1667{
Tejun Heo83096eb2009-05-07 22:24:39 +09001668 if (blk_rq_pos(rq) >= cfqd->last_position)
1669 return blk_rq_pos(rq) - cfqd->last_position;
Jens Axboed9e76202007-04-20 14:27:50 +02001670 else
Tejun Heo83096eb2009-05-07 22:24:39 +09001671 return cfqd->last_position - blk_rq_pos(rq);
Jens Axboed9e76202007-04-20 14:27:50 +02001672}
1673
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001674#define CFQQ_SEEK_THR 8 * 1024
1675#define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
Jeff Moyer04dc6e72009-04-21 07:31:56 +02001676
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001677static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1678 struct request *rq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001679{
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001680 sector_t sdist = cfqq->seek_mean;
Jens Axboecaaa5f92006-06-16 11:23:00 +02001681
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001682 if (!sample_valid(cfqq->seek_samples))
1683 sdist = CFQQ_SEEK_THR;
Jens Axboe6d048f52007-04-25 12:44:27 +02001684
Jeff Moyer04dc6e72009-04-21 07:31:56 +02001685 return cfq_dist_from_last(cfqd, rq) <= sdist;
Jens Axboe6d048f52007-04-25 12:44:27 +02001686}
1687
Jens Axboea36e71f2009-04-15 12:15:11 +02001688static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1689 struct cfq_queue *cur_cfqq)
Jens Axboe6d048f52007-04-25 12:44:27 +02001690{
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001691 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
Jens Axboea36e71f2009-04-15 12:15:11 +02001692 struct rb_node *parent, *node;
1693 struct cfq_queue *__cfqq;
1694 sector_t sector = cfqd->last_position;
1695
1696 if (RB_EMPTY_ROOT(root))
1697 return NULL;
1698
1699 /*
1700 * First, if we find a request starting at the end of the last
1701 * request, choose it.
1702 */
Jens Axboef2d1f0a2009-04-23 12:19:38 +02001703 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
Jens Axboea36e71f2009-04-15 12:15:11 +02001704 if (__cfqq)
1705 return __cfqq;
1706
1707 /*
1708 * If the exact sector wasn't found, the parent of the NULL leaf
1709 * will contain the closest sector.
1710 */
1711 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001712 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001713 return __cfqq;
1714
Tejun Heo2e46e8b2009-05-07 22:24:41 +09001715 if (blk_rq_pos(__cfqq->next_rq) < sector)
Jens Axboea36e71f2009-04-15 12:15:11 +02001716 node = rb_next(&__cfqq->p_node);
1717 else
1718 node = rb_prev(&__cfqq->p_node);
1719 if (!node)
1720 return NULL;
1721
1722 __cfqq = rb_entry(node, struct cfq_queue, p_node);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04001723 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
Jens Axboea36e71f2009-04-15 12:15:11 +02001724 return __cfqq;
1725
1726 return NULL;
1727}
1728
1729/*
1730 * cfqd - obvious
1731 * cur_cfqq - passed in so that we don't decide that the current queue is
1732 * closely cooperating with itself.
1733 *
1734 * So, basically we're assuming that that cur_cfqq has dispatched at least
1735 * one request, and that cfqd->last_position reflects a position on the disk
1736 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1737 * assumption.
1738 */
1739static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
Jeff Moyerb3b6d042009-10-23 17:14:51 -04001740 struct cfq_queue *cur_cfqq)
Jens Axboea36e71f2009-04-15 12:15:11 +02001741{
1742 struct cfq_queue *cfqq;
1743
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001744 if (!cfq_cfqq_sync(cur_cfqq))
1745 return NULL;
1746 if (CFQQ_SEEKY(cur_cfqq))
1747 return NULL;
1748
Jens Axboea36e71f2009-04-15 12:15:11 +02001749 /*
Gui Jianfengb9d8f4c2009-12-08 08:54:17 +01001750 * Don't search priority tree if it's the only queue in the group.
1751 */
1752 if (cur_cfqq->cfqg->nr_cfqq == 1)
1753 return NULL;
1754
1755 /*
Jens Axboed9e76202007-04-20 14:27:50 +02001756 * We should notice if some of the queues are cooperating, eg
1757 * working closely on the same area of the disk. In that case,
1758 * we can group them together and don't waste time idling.
Jens Axboe6d048f52007-04-25 12:44:27 +02001759 */
Jens Axboea36e71f2009-04-15 12:15:11 +02001760 cfqq = cfqq_close(cfqd, cur_cfqq);
1761 if (!cfqq)
1762 return NULL;
1763
Vivek Goyal8682e1f2009-12-03 12:59:50 -05001764 /* If new queue belongs to different cfq_group, don't choose it */
1765 if (cur_cfqq->cfqg != cfqq->cfqg)
1766 return NULL;
1767
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001768 /*
1769 * It only makes sense to merge sync queues.
1770 */
1771 if (!cfq_cfqq_sync(cfqq))
1772 return NULL;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001773 if (CFQQ_SEEKY(cfqq))
1774 return NULL;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001775
Corrado Zoccoloc0324a02009-10-27 19:16:03 +01001776 /*
1777 * Do not merge queues of different priority classes
1778 */
1779 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1780 return NULL;
1781
Jens Axboea36e71f2009-04-15 12:15:11 +02001782 return cfqq;
Jens Axboe6d048f52007-04-25 12:44:27 +02001783}
1784
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001785/*
1786 * Determine whether we should enforce idle window for this queue.
1787 */
1788
1789static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1790{
1791 enum wl_prio_t prio = cfqq_prio(cfqq);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001792 struct cfq_rb_root *service_tree = cfqq->service_tree;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001793
Vivek Goyalf04a6422009-12-03 12:59:40 -05001794 BUG_ON(!service_tree);
1795 BUG_ON(!service_tree->count);
1796
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001797 /* We never do for idle class queues. */
1798 if (prio == IDLE_WORKLOAD)
1799 return false;
1800
1801 /* We do for queues that were marked with idle window flag. */
Shaohua Li3c764b72009-12-04 13:12:06 +01001802 if (cfq_cfqq_idle_window(cfqq) &&
1803 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001804 return true;
1805
1806 /*
1807 * Otherwise, we do only if they are the last ones
1808 * in their service tree.
1809 */
Vivek Goyalf04a6422009-12-03 12:59:40 -05001810 return service_tree->count == 1;
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001811}
1812
Jens Axboe6d048f52007-04-25 12:44:27 +02001813static void cfq_arm_slice_timer(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02001814{
Jens Axboe17926692007-01-19 11:59:30 +11001815 struct cfq_queue *cfqq = cfqd->active_queue;
Jens Axboe206dc692006-03-28 13:03:44 +02001816 struct cfq_io_context *cic;
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001817 unsigned long sl;
1818
Jens Axboea68bbdd2008-09-24 13:03:33 +02001819 /*
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001820 * SSD device without seek penalty, disable idling. But only do so
1821 * for devices that support queuing, otherwise we still have a problem
1822 * with sync vs async workloads.
Jens Axboea68bbdd2008-09-24 13:03:33 +02001823 */
Jens Axboef7d7b7a2008-09-25 11:37:50 +02001824 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
Jens Axboea68bbdd2008-09-24 13:03:33 +02001825 return;
1826
Jens Axboedd67d052006-06-21 09:36:18 +02001827 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
Jens Axboe6d048f52007-04-25 12:44:27 +02001828 WARN_ON(cfq_cfqq_slice_new(cfqq));
Jens Axboe22e2c502005-06-27 10:55:12 +02001829
1830 /*
1831 * idle is disabled, either manually or by past process history
1832 */
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01001833 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
Jens Axboe6d048f52007-04-25 12:44:27 +02001834 return;
1835
Jens Axboe22e2c502005-06-27 10:55:12 +02001836 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001837 * still active requests from this queue, don't idle
Jens Axboe7b679132008-05-30 12:23:07 +02001838 */
Corrado Zoccolo8e550632009-11-26 10:02:58 +01001839 if (cfqq->dispatched)
Jens Axboe7b679132008-05-30 12:23:07 +02001840 return;
1841
1842 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02001843 * task has exited, don't wait
1844 */
Jens Axboe206dc692006-03-28 13:03:44 +02001845 cic = cfqd->active_cic;
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01001846 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
Jens Axboe6d048f52007-04-25 12:44:27 +02001847 return;
1848
Corrado Zoccolo355b6592009-10-08 08:43:32 +02001849 /*
1850 * If our average think time is larger than the remaining time
1851 * slice, then don't idle. This avoids overrunning the allotted
1852 * time slice.
1853 */
1854 if (sample_valid(cic->ttime_samples) &&
1855 (cfqq->slice_end - jiffies < cic->ttime_mean))
1856 return;
1857
Jens Axboe3b181522005-06-27 10:56:24 +02001858 cfq_mark_cfqq_wait_request(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001859
Jens Axboe6d048f52007-04-25 12:44:27 +02001860 sl = cfqd->cfq_slice_idle;
Jens Axboe206dc692006-03-28 13:03:44 +02001861
Jens Axboe7b14e3b2006-02-28 09:35:11 +01001862 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
Jens Axboe9481ffd2009-04-15 12:14:13 +02001863 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864}
1865
Jens Axboe498d3aa22007-04-26 12:54:48 +02001866/*
1867 * Move request from internal lists to the request queue dispatch list.
1868 */
Jens Axboe165125e2007-07-24 09:28:11 +02001869static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870{
Jens Axboe3ed9a292007-04-23 08:33:33 +02001871 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02001872 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02001873
Jens Axboe7b679132008-05-30 12:23:07 +02001874 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1875
Jeff Moyer06d21882009-09-11 17:08:59 +02001876 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
Jens Axboe5380a102006-07-13 12:37:56 +02001877 cfq_remove_request(rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001878 cfqq->dispatched++;
Jens Axboe5380a102006-07-13 12:37:56 +02001879 elv_dispatch_sort(q, rq);
Jens Axboe3ed9a292007-04-23 08:33:33 +02001880
1881 if (cfq_cfqq_sync(cfqq))
1882 cfqd->sync_flight++;
Vivek Goyal22084192009-12-03 12:59:49 -05001883 cfqq->nr_sectors += blk_rq_sectors(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884}
1885
1886/*
1887 * return expired entry, or NULL to just start from scratch in rbtree
1888 */
Jens Axboefebffd62008-01-28 13:19:43 +01001889static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890{
Jens Axboe30996f42009-10-05 11:03:39 +02001891 struct request *rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892
Jens Axboe3b181522005-06-27 10:56:24 +02001893 if (cfq_cfqq_fifo_expire(cfqq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 return NULL;
Jens Axboecb887412007-01-19 12:01:16 +11001895
1896 cfq_mark_cfqq_fifo_expire(cfqq);
1897
Jens Axboe89850f72006-07-22 16:48:31 +02001898 if (list_empty(&cfqq->fifo))
1899 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900
Jens Axboe89850f72006-07-22 16:48:31 +02001901 rq = rq_entry_fifo(cfqq->fifo.next);
Jens Axboe30996f42009-10-05 11:03:39 +02001902 if (time_before(jiffies, rq_fifo_time(rq)))
Jens Axboe7b679132008-05-30 12:23:07 +02001903 rq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904
Jens Axboe30996f42009-10-05 11:03:39 +02001905 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
Jens Axboe6d048f52007-04-25 12:44:27 +02001906 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907}
1908
Jens Axboe22e2c502005-06-27 10:55:12 +02001909static inline int
1910cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1911{
1912 const int base_rq = cfqd->cfq_slice_async_rq;
1913
1914 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1915
1916 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1917}
1918
1919/*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001920 * Must be called with the queue_lock held.
1921 */
1922static int cfqq_process_refs(struct cfq_queue *cfqq)
1923{
1924 int process_refs, io_refs;
1925
1926 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1927 process_refs = atomic_read(&cfqq->ref) - io_refs;
1928 BUG_ON(process_refs < 0);
1929 return process_refs;
1930}
1931
1932static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1933{
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001934 int process_refs, new_process_refs;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001935 struct cfq_queue *__cfqq;
1936
1937 /* Avoid a circular list and skip interim queue merges */
1938 while ((__cfqq = new_cfqq->new_cfqq)) {
1939 if (__cfqq == cfqq)
1940 return;
1941 new_cfqq = __cfqq;
1942 }
1943
1944 process_refs = cfqq_process_refs(cfqq);
1945 /*
1946 * If the process for the cfqq has gone away, there is no
1947 * sense in merging the queues.
1948 */
1949 if (process_refs == 0)
1950 return;
1951
Jeff Moyere6c5bc72009-10-23 17:14:52 -04001952 /*
1953 * Merge in the direction of the lesser amount of work.
1954 */
1955 new_process_refs = cfqq_process_refs(new_cfqq);
1956 if (new_process_refs >= process_refs) {
1957 cfqq->new_cfqq = new_cfqq;
1958 atomic_add(process_refs, &new_cfqq->ref);
1959 } else {
1960 new_cfqq->new_cfqq = cfqq;
1961 atomic_add(new_process_refs, &cfqq->ref);
1962 }
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04001963}
1964
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001965static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1966 struct cfq_group *cfqg, enum wl_prio_t prio,
1967 bool prio_changed)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001968{
1969 struct cfq_queue *queue;
1970 int i;
1971 bool key_valid = false;
1972 unsigned long lowest_key = 0;
1973 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1974
1975 if (prio_changed) {
1976 /*
1977 * When priorities switched, we prefer starting
1978 * from SYNC_NOIDLE (first choice), or just SYNC
1979 * over ASYNC
1980 */
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 cur_best = SYNC_WORKLOAD;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001984 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001985 return cur_best;
1986
1987 return ASYNC_WORKLOAD;
1988 }
1989
1990 for (i = 0; i < 3; ++i) {
1991 /* otherwise, select the one with lowest rb_key */
Vivek Goyalcdb16e82009-12-03 12:59:38 -05001992 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01001993 if (queue &&
1994 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1995 lowest_key = queue->rb_key;
1996 cur_best = i;
1997 key_valid = true;
1998 }
1999 }
2000
2001 return cur_best;
2002}
2003
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002004static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002005{
2006 enum wl_prio_t previous_prio = cfqd->serving_prio;
2007 bool prio_changed;
2008 unsigned slice;
2009 unsigned count;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002010 struct cfq_rb_root *st;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002011 unsigned group_slice;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002012
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002013 if (!cfqg) {
2014 cfqd->serving_prio = IDLE_WORKLOAD;
2015 cfqd->workload_expires = jiffies + 1;
2016 return;
2017 }
2018
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002019 /* Choose next priority. RT > BE > IDLE */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002020 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002021 cfqd->serving_prio = RT_WORKLOAD;
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002022 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002023 cfqd->serving_prio = BE_WORKLOAD;
2024 else {
2025 cfqd->serving_prio = IDLE_WORKLOAD;
2026 cfqd->workload_expires = jiffies + 1;
2027 return;
2028 }
2029
2030 /*
2031 * For RT and BE, we have to choose also the type
2032 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2033 * expiration time
2034 */
2035 prio_changed = (cfqd->serving_prio != previous_prio);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002036 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2037 cfqd);
2038 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002039
2040 /*
2041 * If priority didn't change, check workload expiration,
2042 * and that we still have other queues ready
2043 */
2044 if (!prio_changed && count &&
2045 !time_after(jiffies, cfqd->workload_expires))
2046 return;
2047
2048 /* otherwise select new workload type */
2049 cfqd->serving_type =
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002050 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
2051 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2052 cfqd);
2053 count = st->count;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002054
2055 /*
2056 * the workload slice is computed as a fraction of target latency
2057 * proportional to the number of queues in that workload, over
2058 * all the queues in the same priority class
2059 */
Vivek Goyal58ff82f2009-12-03 12:59:44 -05002060 group_slice = cfq_group_slice(cfqd, cfqg);
2061
2062 slice = group_slice * count /
2063 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2064 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002065
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002066 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2067 unsigned int tmp;
2068
2069 /*
2070 * Async queues are currently system wide. Just taking
2071 * proportion of queues with-in same group will lead to higher
2072 * async ratio system wide as generally root group is going
2073 * to have higher weight. A more accurate thing would be to
2074 * calculate system wide asnc/sync ratio.
2075 */
2076 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2077 tmp = tmp/cfqd->busy_queues;
2078 slice = min_t(unsigned, slice, tmp);
2079
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002080 /* async workload slice is scaled down according to
2081 * the sync/async slice ratio. */
2082 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
Vivek Goyalf26bd1f2009-12-03 12:59:54 -05002083 } else
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002084 /* sync workload slice is at least 2 * cfq_slice_idle */
2085 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2086
2087 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2088 cfqd->workload_expires = jiffies + slice;
Corrado Zoccolo8e550632009-11-26 10:02:58 +01002089 cfqd->noidle_tree_requires_idle = false;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002090}
2091
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002092static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2093{
2094 struct cfq_rb_root *st = &cfqd->grp_service_tree;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002095 struct cfq_group *cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002096
2097 if (RB_EMPTY_ROOT(&st->rb))
2098 return NULL;
Vivek Goyal25bc6b02009-12-03 12:59:43 -05002099 cfqg = cfq_rb_first_group(st);
2100 st->active = &cfqg->rb_node;
2101 update_min_vdisktime(st);
2102 return cfqg;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002103}
2104
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002105static void cfq_choose_cfqg(struct cfq_data *cfqd)
2106{
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002107 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2108
2109 cfqd->serving_group = cfqg;
Vivek Goyaldae739e2009-12-03 12:59:45 -05002110
2111 /* Restore the workload type data */
2112 if (cfqg->saved_workload_slice) {
2113 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2114 cfqd->serving_type = cfqg->saved_workload;
2115 cfqd->serving_prio = cfqg->saved_serving_prio;
Gui Jianfeng66ae2912009-12-15 10:08:45 +01002116 } else
2117 cfqd->workload_expires = jiffies - 1;
2118
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05002119 choose_service_tree(cfqd, cfqg);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002120}
2121
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002122/*
Jens Axboe498d3aa22007-04-26 12:54:48 +02002123 * Select a queue for service. If we have a current active queue,
2124 * check whether to continue servicing it, or retrieve and set a new one.
Jens Axboe22e2c502005-06-27 10:55:12 +02002125 */
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002126static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
Jens Axboe22e2c502005-06-27 10:55:12 +02002127{
Jens Axboea36e71f2009-04-15 12:15:11 +02002128 struct cfq_queue *cfqq, *new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002129
2130 cfqq = cfqd->active_queue;
2131 if (!cfqq)
2132 goto new_queue;
2133
Vivek Goyalf04a6422009-12-03 12:59:40 -05002134 if (!cfqd->rq_queued)
2135 return NULL;
Vivek Goyalc244bb52009-12-08 17:52:57 -05002136
2137 /*
2138 * We were waiting for group to get backlogged. Expire the queue
2139 */
2140 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2141 goto expire;
2142
Jens Axboe22e2c502005-06-27 10:55:12 +02002143 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002144 * The active queue has run out of time, expire it and select new.
Jens Axboe22e2c502005-06-27 10:55:12 +02002145 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05002146 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2147 /*
2148 * If slice had not expired at the completion of last request
2149 * we might not have turned on wait_busy flag. Don't expire
2150 * the queue yet. Allow the group to get backlogged.
2151 *
2152 * The very fact that we have used the slice, that means we
2153 * have been idling all along on this queue and it should be
2154 * ok to wait for this request to complete.
2155 */
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002156 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2157 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2158 cfqq = NULL;
Vivek Goyal7667aa02009-12-08 17:52:58 -05002159 goto keep_queue;
Vivek Goyal82bbbf22009-12-10 19:25:41 +01002160 } else
Vivek Goyal7667aa02009-12-08 17:52:58 -05002161 goto expire;
2162 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002163
2164 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002165 * The active queue has requests and isn't expired, allow it to
2166 * dispatch.
Jens Axboe22e2c502005-06-27 10:55:12 +02002167 */
Jens Axboedd67d052006-06-21 09:36:18 +02002168 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02002169 goto keep_queue;
Jens Axboe6d048f52007-04-25 12:44:27 +02002170
2171 /*
Jens Axboea36e71f2009-04-15 12:15:11 +02002172 * If another queue has a request waiting within our mean seek
2173 * distance, let it run. The expire code will check for close
2174 * cooperators and put the close queue at the front of the service
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002175 * tree. If possible, merge the expiring queue with the new cfqq.
Jens Axboea36e71f2009-04-15 12:15:11 +02002176 */
Jeff Moyerb3b6d042009-10-23 17:14:51 -04002177 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002178 if (new_cfqq) {
2179 if (!cfqq->new_cfqq)
2180 cfq_setup_merge(cfqq, new_cfqq);
Jens Axboea36e71f2009-04-15 12:15:11 +02002181 goto expire;
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002182 }
Jens Axboea36e71f2009-04-15 12:15:11 +02002183
2184 /*
Jens Axboe6d048f52007-04-25 12:44:27 +02002185 * No requests pending. If the active queue still has requests in
2186 * flight or is idling for a new request, allow either of these
2187 * conditions to happen (or time out) before selecting a new queue.
2188 */
Jens Axboecc197472007-04-20 20:45:39 +02002189 if (timer_pending(&cfqd->idle_slice_timer) ||
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002190 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
Jens Axboecaaa5f92006-06-16 11:23:00 +02002191 cfqq = NULL;
2192 goto keep_queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02002193 }
2194
Jens Axboe3b181522005-06-27 10:56:24 +02002195expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02002196 cfq_slice_expired(cfqd, 0);
Jens Axboe3b181522005-06-27 10:56:24 +02002197new_queue:
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002198 /*
2199 * Current queue expired. Check if we have to switch to a new
2200 * service tree
2201 */
2202 if (!new_cfqq)
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002203 cfq_choose_cfqg(cfqd);
Corrado Zoccolo718eee02009-10-26 22:45:29 +01002204
Jens Axboea36e71f2009-04-15 12:15:11 +02002205 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002206keep_queue:
Jens Axboe3b181522005-06-27 10:56:24 +02002207 return cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02002208}
2209
Jens Axboefebffd62008-01-28 13:19:43 +01002210static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
Jens Axboed9e76202007-04-20 14:27:50 +02002211{
2212 int dispatched = 0;
2213
2214 while (cfqq->next_rq) {
2215 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2216 dispatched++;
2217 }
2218
2219 BUG_ON(!list_empty(&cfqq->fifo));
Vivek Goyalf04a6422009-12-03 12:59:40 -05002220
2221 /* By default cfqq is not expired if it is empty. Do it explicitly */
2222 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
Jens Axboed9e76202007-04-20 14:27:50 +02002223 return dispatched;
2224}
2225
Jens Axboe498d3aa22007-04-26 12:54:48 +02002226/*
2227 * Drain our current requests. Used for barriers and when switching
2228 * io schedulers on-the-fly.
2229 */
Jens Axboed9e76202007-04-20 14:27:50 +02002230static int cfq_forced_dispatch(struct cfq_data *cfqd)
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002231{
Jens Axboe08717142008-01-28 11:38:15 +01002232 struct cfq_queue *cfqq;
Jens Axboed9e76202007-04-20 14:27:50 +02002233 int dispatched = 0;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002234
Vivek Goyalf04a6422009-12-03 12:59:40 -05002235 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2236 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002237
Jens Axboe6084cdd2007-04-23 08:25:00 +02002238 cfq_slice_expired(cfqd, 0);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002239 BUG_ON(cfqd->busy_queues);
2240
Jeff Moyer69237152009-06-12 15:29:30 +02002241 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
Tejun Heo1b5ed5e12005-11-10 08:49:19 +01002242 return dispatched;
2243}
2244
Jens Axboe0b182d62009-10-06 20:49:37 +02002245static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
Jens Axboe2f5cb732009-04-07 08:51:19 +02002246{
Jens Axboe2f5cb732009-04-07 08:51:19 +02002247 unsigned int max_dispatch;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248
Jens Axboe2f5cb732009-04-07 08:51:19 +02002249 /*
Jens Axboe5ad531d2009-07-03 12:57:48 +02002250 * Drain async requests before we start sync IO
2251 */
Corrado Zoccoloa6d44e92009-10-26 22:45:11 +01002252 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
Jens Axboe0b182d62009-10-06 20:49:37 +02002253 return false;
Jens Axboe5ad531d2009-07-03 12:57:48 +02002254
2255 /*
Jens Axboe2f5cb732009-04-07 08:51:19 +02002256 * If this is an async queue and we have sync IO in flight, let it wait
2257 */
2258 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002259 return false;
Jens Axboe2f5cb732009-04-07 08:51:19 +02002260
2261 max_dispatch = cfqd->cfq_quantum;
2262 if (cfq_class_idle(cfqq))
2263 max_dispatch = 1;
2264
2265 /*
2266 * Does this cfqq already have too much IO in flight?
2267 */
2268 if (cfqq->dispatched >= max_dispatch) {
2269 /*
2270 * idle queue must always only have a single IO in flight
2271 */
Jens Axboe3ed9a292007-04-23 08:33:33 +02002272 if (cfq_class_idle(cfqq))
Jens Axboe0b182d62009-10-06 20:49:37 +02002273 return false;
Jens Axboe3ed9a292007-04-23 08:33:33 +02002274
Jens Axboe2f5cb732009-04-07 08:51:19 +02002275 /*
2276 * We have other queues, don't allow more IO from this one
2277 */
2278 if (cfqd->busy_queues > 1)
Jens Axboe0b182d62009-10-06 20:49:37 +02002279 return false;
Jens Axboe9ede2092007-01-19 12:11:44 +11002280
Jens Axboe2f5cb732009-04-07 08:51:19 +02002281 /*
Shaohua Li474b18c2009-12-03 12:58:05 +01002282 * Sole queue user, no limit
Vivek Goyal365722b2009-10-03 15:21:27 +02002283 */
Shaohua Li474b18c2009-12-03 12:58:05 +01002284 max_dispatch = -1;
Jens Axboe8e296752009-10-03 16:26:03 +02002285 }
2286
2287 /*
2288 * Async queues must wait a bit before being allowed dispatch.
2289 * We also ramp up the dispatch depth gradually for async IO,
2290 * based on the last sync IO we serviced
2291 */
Jens Axboe963b72f2009-10-03 19:42:18 +02002292 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
Corrado Zoccolo573412b2009-12-06 11:48:52 +01002293 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
Jens Axboe8e296752009-10-03 16:26:03 +02002294 unsigned int depth;
Vivek Goyal365722b2009-10-03 15:21:27 +02002295
Jens Axboe61f0c1d2009-10-03 19:46:03 +02002296 depth = last_sync / cfqd->cfq_slice[1];
Jens Axboee00c54c2009-10-04 20:36:19 +02002297 if (!depth && !cfqq->dispatched)
2298 depth = 1;
Jens Axboe8e296752009-10-03 16:26:03 +02002299 if (depth < max_dispatch)
2300 max_dispatch = depth;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301 }
2302
Jens Axboe0b182d62009-10-06 20:49:37 +02002303 /*
2304 * If we're below the current max, allow a dispatch
2305 */
2306 return cfqq->dispatched < max_dispatch;
2307}
2308
2309/*
2310 * Dispatch a request from cfqq, moving them to the request queue
2311 * dispatch list.
2312 */
2313static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2314{
2315 struct request *rq;
2316
2317 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2318
2319 if (!cfq_may_dispatch(cfqd, cfqq))
2320 return false;
2321
2322 /*
2323 * follow expired path, else get first next available
2324 */
2325 rq = cfq_check_fifo(cfqq);
2326 if (!rq)
2327 rq = cfqq->next_rq;
2328
2329 /*
2330 * insert request into driver dispatch list
2331 */
2332 cfq_dispatch_insert(cfqd->queue, rq);
2333
2334 if (!cfqd->active_cic) {
2335 struct cfq_io_context *cic = RQ_CIC(rq);
2336
2337 atomic_long_inc(&cic->ioc->refcount);
2338 cfqd->active_cic = cic;
2339 }
2340
2341 return true;
2342}
2343
2344/*
2345 * Find the cfqq that we need to service and move a request from that to the
2346 * dispatch list
2347 */
2348static int cfq_dispatch_requests(struct request_queue *q, int force)
2349{
2350 struct cfq_data *cfqd = q->elevator->elevator_data;
2351 struct cfq_queue *cfqq;
2352
2353 if (!cfqd->busy_queues)
2354 return 0;
2355
2356 if (unlikely(force))
2357 return cfq_forced_dispatch(cfqd);
2358
2359 cfqq = cfq_select_queue(cfqd);
2360 if (!cfqq)
Jens Axboe8e296752009-10-03 16:26:03 +02002361 return 0;
2362
Jens Axboe2f5cb732009-04-07 08:51:19 +02002363 /*
Jens Axboe0b182d62009-10-06 20:49:37 +02002364 * Dispatch a request from this cfqq, if it is allowed
Jens Axboe2f5cb732009-04-07 08:51:19 +02002365 */
Jens Axboe0b182d62009-10-06 20:49:37 +02002366 if (!cfq_dispatch_request(cfqd, cfqq))
2367 return 0;
2368
Jens Axboe2f5cb732009-04-07 08:51:19 +02002369 cfqq->slice_dispatch++;
Jens Axboeb0291952009-04-07 11:38:31 +02002370 cfq_clear_cfqq_must_dispatch(cfqq);
Jens Axboe2f5cb732009-04-07 08:51:19 +02002371
2372 /*
2373 * expire an async queue immediately if it has used up its slice. idle
2374 * queue always expire after 1 dispatch round.
2375 */
2376 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2377 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2378 cfq_class_idle(cfqq))) {
2379 cfqq->slice_end = jiffies + 1;
2380 cfq_slice_expired(cfqd, 0);
2381 }
2382
Shan Weib217a902009-09-01 10:06:42 +02002383 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
Jens Axboe2f5cb732009-04-07 08:51:19 +02002384 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385}
2386
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387/*
Jens Axboe5e705372006-07-13 12:39:25 +02002388 * task holds one reference to the queue, dropped when task exits. each rq
2389 * in-flight on this queue also holds a reference, dropped when rq is freed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 *
Vivek Goyalb1c35762009-12-03 12:59:47 -05002391 * Each cfq queue took a reference on the parent group. Drop it now.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 * queue lock must be held here.
2393 */
2394static void cfq_put_queue(struct cfq_queue *cfqq)
2395{
Jens Axboe22e2c502005-06-27 10:55:12 +02002396 struct cfq_data *cfqd = cfqq->cfqd;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002397 struct cfq_group *cfqg, *orig_cfqg;
Jens Axboe22e2c502005-06-27 10:55:12 +02002398
2399 BUG_ON(atomic_read(&cfqq->ref) <= 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400
2401 if (!atomic_dec_and_test(&cfqq->ref))
2402 return;
2403
Jens Axboe7b679132008-05-30 12:23:07 +02002404 cfq_log_cfqq(cfqd, cfqq, "put_queue");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 BUG_ON(rb_first(&cfqq->sort_list));
Jens Axboe22e2c502005-06-27 10:55:12 +02002406 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002407 cfqg = cfqq->cfqg;
Vivek Goyal878eadd2009-12-07 19:37:15 +01002408 orig_cfqg = cfqq->orig_cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002410 if (unlikely(cfqd->active_queue == cfqq)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02002411 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002412 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002413 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002414
Vivek Goyalf04a6422009-12-03 12:59:40 -05002415 BUG_ON(cfq_cfqq_on_rr(cfqq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416 kmem_cache_free(cfq_pool, cfqq);
Vivek Goyalb1c35762009-12-03 12:59:47 -05002417 cfq_put_cfqg(cfqg);
Vivek Goyal878eadd2009-12-07 19:37:15 +01002418 if (orig_cfqg)
2419 cfq_put_cfqg(orig_cfqg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420}
2421
Jens Axboed6de8be2008-05-28 14:46:59 +02002422/*
2423 * Must always be called with the rcu_read_lock() held
2424 */
Jens Axboe07416d22008-05-07 09:17:12 +02002425static void
2426__call_for_each_cic(struct io_context *ioc,
2427 void (*func)(struct io_context *, struct cfq_io_context *))
2428{
2429 struct cfq_io_context *cic;
2430 struct hlist_node *n;
2431
2432 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2433 func(ioc, cic);
2434}
2435
Jens Axboe4ac845a2008-01-24 08:44:49 +01002436/*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002437 * Call func for each cic attached to this ioc.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002438 */
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002439static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002440call_for_each_cic(struct io_context *ioc,
2441 void (*func)(struct io_context *, struct cfq_io_context *))
2442{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002443 rcu_read_lock();
Jens Axboe07416d22008-05-07 09:17:12 +02002444 __call_for_each_cic(ioc, func);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002445 rcu_read_unlock();
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002446}
Jens Axboe4ac845a2008-01-24 08:44:49 +01002447
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002448static void cfq_cic_free_rcu(struct rcu_head *head)
2449{
2450 struct cfq_io_context *cic;
2451
2452 cic = container_of(head, struct cfq_io_context, rcu_head);
2453
2454 kmem_cache_free(cfq_ioc_pool, cic);
Tejun Heo245b2e72009-06-24 15:13:48 +09002455 elv_ioc_count_dec(cfq_ioc_count);
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002456
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002457 if (ioc_gone) {
2458 /*
2459 * CFQ scheduler is exiting, grab exit lock and check
2460 * the pending io context count. If it hits zero,
2461 * complete ioc_gone and set it back to NULL
2462 */
2463 spin_lock(&ioc_gone_lock);
Tejun Heo245b2e72009-06-24 15:13:48 +09002464 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
Jens Axboe9a11b4e2008-05-29 09:32:08 +02002465 complete(ioc_gone);
2466 ioc_gone = NULL;
2467 }
2468 spin_unlock(&ioc_gone_lock);
2469 }
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002470}
2471
2472static void cfq_cic_free(struct cfq_io_context *cic)
2473{
2474 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002475}
2476
2477static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2478{
2479 unsigned long flags;
2480
2481 BUG_ON(!cic->dead_key);
2482
2483 spin_lock_irqsave(&ioc->lock, flags);
2484 radix_tree_delete(&ioc->radix_root, cic->dead_key);
Jens Axboeffc4e752008-02-19 10:02:29 +01002485 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002486 spin_unlock_irqrestore(&ioc->lock, flags);
2487
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002488 cfq_cic_free(cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002489}
2490
Jens Axboed6de8be2008-05-28 14:46:59 +02002491/*
2492 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2493 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2494 * and ->trim() which is called with the task lock held
2495 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002496static void cfq_free_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002498 /*
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02002499 * ioc->refcount is zero here, or we are called from elv_unregister(),
2500 * so no more cic's are allowed to be linked into this ioc. So it
2501 * should be ok to iterate over the known list, we will see all cic's
2502 * since no new ones are added.
Jens Axboe4ac845a2008-01-24 08:44:49 +01002503 */
Jens Axboe07416d22008-05-07 09:17:12 +02002504 __call_for_each_cic(ioc, cic_free_func);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505}
2506
Jens Axboe89850f72006-07-22 16:48:31 +02002507static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2508{
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002509 struct cfq_queue *__cfqq, *next;
2510
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002511 if (unlikely(cfqq == cfqd->active_queue)) {
Jens Axboe6084cdd2007-04-23 08:25:00 +02002512 __cfq_slice_expired(cfqd, cfqq, 0);
Jens Axboe23e018a2009-10-05 08:52:35 +02002513 cfq_schedule_dispatch(cfqd);
Jens Axboe28f95cbc2007-01-19 12:09:53 +11002514 }
Jens Axboe89850f72006-07-22 16:48:31 +02002515
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04002516 /*
2517 * If this queue was scheduled to merge with another queue, be
2518 * sure to drop the reference taken on that queue (and others in
2519 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2520 */
2521 __cfqq = cfqq->new_cfqq;
2522 while (__cfqq) {
2523 if (__cfqq == cfqq) {
2524 WARN(1, "cfqq->new_cfqq loop detected\n");
2525 break;
2526 }
2527 next = __cfqq->new_cfqq;
2528 cfq_put_queue(__cfqq);
2529 __cfqq = next;
2530 }
2531
Jens Axboe89850f72006-07-22 16:48:31 +02002532 cfq_put_queue(cfqq);
2533}
2534
2535static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2536 struct cfq_io_context *cic)
2537{
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002538 struct io_context *ioc = cic->ioc;
2539
Jens Axboefc463792006-08-29 09:05:44 +02002540 list_del_init(&cic->queue_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002541
2542 /*
2543 * Make sure key == NULL is seen for dead queues
2544 */
Jens Axboefc463792006-08-29 09:05:44 +02002545 smp_wmb();
Jens Axboe4ac845a2008-01-24 08:44:49 +01002546 cic->dead_key = (unsigned long) cic->key;
Jens Axboefc463792006-08-29 09:05:44 +02002547 cic->key = NULL;
2548
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002549 if (ioc->ioc_data == cic)
2550 rcu_assign_pointer(ioc->ioc_data, NULL);
2551
Jens Axboeff6657c2009-04-08 10:58:57 +02002552 if (cic->cfqq[BLK_RW_ASYNC]) {
2553 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2554 cic->cfqq[BLK_RW_ASYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002555 }
2556
Jens Axboeff6657c2009-04-08 10:58:57 +02002557 if (cic->cfqq[BLK_RW_SYNC]) {
2558 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2559 cic->cfqq[BLK_RW_SYNC] = NULL;
Jens Axboe89850f72006-07-22 16:48:31 +02002560 }
Jens Axboe89850f72006-07-22 16:48:31 +02002561}
2562
Jens Axboe4ac845a2008-01-24 08:44:49 +01002563static void cfq_exit_single_io_context(struct io_context *ioc,
2564 struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002565{
Al Viro478a82b2006-03-18 13:25:24 -05002566 struct cfq_data *cfqd = cic->key;
Jens Axboe22e2c502005-06-27 10:55:12 +02002567
Jens Axboe89850f72006-07-22 16:48:31 +02002568 if (cfqd) {
Jens Axboe165125e2007-07-24 09:28:11 +02002569 struct request_queue *q = cfqd->queue;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002570 unsigned long flags;
Jens Axboe22e2c502005-06-27 10:55:12 +02002571
Jens Axboe4ac845a2008-01-24 08:44:49 +01002572 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboe62c1fe92008-12-15 21:19:25 +01002573
2574 /*
2575 * Ensure we get a fresh copy of the ->key to prevent
2576 * race between exiting task and queue
2577 */
2578 smp_read_barrier_depends();
2579 if (cic->key)
2580 __cfq_exit_single_io_context(cfqd, cic);
2581
Jens Axboe4ac845a2008-01-24 08:44:49 +01002582 spin_unlock_irqrestore(q->queue_lock, flags);
Al Viro12a05732006-03-18 13:38:01 -05002583 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002584}
2585
Jens Axboe498d3aa22007-04-26 12:54:48 +02002586/*
2587 * The process that ioc belongs to has exited, we need to clean up
2588 * and put the internal structures we have that belongs to that process.
2589 */
Jens Axboee2d74ac2006-03-28 08:59:01 +02002590static void cfq_exit_io_context(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002592 call_for_each_cic(ioc, cfq_exit_single_io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593}
2594
Jens Axboe22e2c502005-06-27 10:55:12 +02002595static struct cfq_io_context *
Al Viro8267e262005-10-21 03:20:53 -04002596cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597{
Jens Axboeb5deef92006-07-19 23:39:40 +02002598 struct cfq_io_context *cic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599
Christoph Lameter94f60302007-07-17 04:03:29 -07002600 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2601 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 if (cic) {
Jens Axboe22e2c502005-06-27 10:55:12 +02002603 cic->last_end_request = jiffies;
Jens Axboe553698f2006-06-14 19:11:57 +02002604 INIT_LIST_HEAD(&cic->queue_list);
Jens Axboeffc4e752008-02-19 10:02:29 +01002605 INIT_HLIST_NODE(&cic->cic_list);
Jens Axboe22e2c502005-06-27 10:55:12 +02002606 cic->dtor = cfq_free_io_context;
2607 cic->exit = cfq_exit_io_context;
Tejun Heo245b2e72009-06-24 15:13:48 +09002608 elv_ioc_count_inc(cfq_ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609 }
2610
2611 return cic;
2612}
2613
Jens Axboefd0928d2008-01-24 08:52:45 +01002614static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
Jens Axboe22e2c502005-06-27 10:55:12 +02002615{
2616 struct task_struct *tsk = current;
2617 int ioprio_class;
2618
Jens Axboe3b181522005-06-27 10:56:24 +02002619 if (!cfq_cfqq_prio_changed(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02002620 return;
2621
Jens Axboefd0928d2008-01-24 08:52:45 +01002622 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
Jens Axboe22e2c502005-06-27 10:55:12 +02002623 switch (ioprio_class) {
Jens Axboefe094d92008-01-31 13:08:54 +01002624 default:
2625 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2626 case IOPRIO_CLASS_NONE:
2627 /*
Jens Axboe6d63c272008-05-07 09:51:23 +02002628 * no prio set, inherit CPU scheduling settings
Jens Axboefe094d92008-01-31 13:08:54 +01002629 */
2630 cfqq->ioprio = task_nice_ioprio(tsk);
Jens Axboe6d63c272008-05-07 09:51:23 +02002631 cfqq->ioprio_class = task_nice_ioclass(tsk);
Jens Axboefe094d92008-01-31 13:08:54 +01002632 break;
2633 case IOPRIO_CLASS_RT:
2634 cfqq->ioprio = task_ioprio(ioc);
2635 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2636 break;
2637 case IOPRIO_CLASS_BE:
2638 cfqq->ioprio = task_ioprio(ioc);
2639 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2640 break;
2641 case IOPRIO_CLASS_IDLE:
2642 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2643 cfqq->ioprio = 7;
2644 cfq_clear_cfqq_idle_window(cfqq);
2645 break;
Jens Axboe22e2c502005-06-27 10:55:12 +02002646 }
2647
2648 /*
2649 * keep track of original prio settings in case we have to temporarily
2650 * elevate the priority of this queue
2651 */
2652 cfqq->org_ioprio = cfqq->ioprio;
2653 cfqq->org_ioprio_class = cfqq->ioprio_class;
Jens Axboe3b181522005-06-27 10:56:24 +02002654 cfq_clear_cfqq_prio_changed(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02002655}
2656
Jens Axboefebffd62008-01-28 13:19:43 +01002657static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
Jens Axboe22e2c502005-06-27 10:55:12 +02002658{
Al Viro478a82b2006-03-18 13:25:24 -05002659 struct cfq_data *cfqd = cic->key;
2660 struct cfq_queue *cfqq;
Jens Axboec1b707d2006-10-30 19:54:23 +01002661 unsigned long flags;
Jens Axboe35e60772006-06-14 09:10:45 +02002662
Jens Axboecaaa5f92006-06-16 11:23:00 +02002663 if (unlikely(!cfqd))
2664 return;
2665
Jens Axboec1b707d2006-10-30 19:54:23 +01002666 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002667
Jens Axboeff6657c2009-04-08 10:58:57 +02002668 cfqq = cic->cfqq[BLK_RW_ASYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002669 if (cfqq) {
2670 struct cfq_queue *new_cfqq;
Jens Axboeff6657c2009-04-08 10:58:57 +02002671 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2672 GFP_ATOMIC);
Jens Axboecaaa5f92006-06-16 11:23:00 +02002673 if (new_cfqq) {
Jens Axboeff6657c2009-04-08 10:58:57 +02002674 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
Jens Axboecaaa5f92006-06-16 11:23:00 +02002675 cfq_put_queue(cfqq);
2676 }
Jens Axboe22e2c502005-06-27 10:55:12 +02002677 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02002678
Jens Axboeff6657c2009-04-08 10:58:57 +02002679 cfqq = cic->cfqq[BLK_RW_SYNC];
Jens Axboecaaa5f92006-06-16 11:23:00 +02002680 if (cfqq)
2681 cfq_mark_cfqq_prio_changed(cfqq);
2682
Jens Axboec1b707d2006-10-30 19:54:23 +01002683 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
Jens Axboe22e2c502005-06-27 10:55:12 +02002684}
2685
Jens Axboefc463792006-08-29 09:05:44 +02002686static void cfq_ioc_set_ioprio(struct io_context *ioc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002688 call_for_each_cic(ioc, changed_ioprio);
Jens Axboefc463792006-08-29 09:05:44 +02002689 ioc->ioprio_changed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690}
2691
Jens Axboed5036d72009-06-26 10:44:34 +02002692static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboea6151c32009-10-07 20:02:57 +02002693 pid_t pid, bool is_sync)
Jens Axboed5036d72009-06-26 10:44:34 +02002694{
2695 RB_CLEAR_NODE(&cfqq->rb_node);
2696 RB_CLEAR_NODE(&cfqq->p_node);
2697 INIT_LIST_HEAD(&cfqq->fifo);
2698
2699 atomic_set(&cfqq->ref, 0);
2700 cfqq->cfqd = cfqd;
2701
2702 cfq_mark_cfqq_prio_changed(cfqq);
2703
2704 if (is_sync) {
2705 if (!cfq_class_idle(cfqq))
2706 cfq_mark_cfqq_idle_window(cfqq);
2707 cfq_mark_cfqq_sync(cfqq);
2708 }
2709 cfqq->pid = pid;
2710}
2711
Vivek Goyal246103332009-12-03 12:59:51 -05002712#ifdef CONFIG_CFQ_GROUP_IOSCHED
2713static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2714{
2715 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2716 struct cfq_data *cfqd = cic->key;
2717 unsigned long flags;
2718 struct request_queue *q;
2719
2720 if (unlikely(!cfqd))
2721 return;
2722
2723 q = cfqd->queue;
2724
2725 spin_lock_irqsave(q->queue_lock, flags);
2726
2727 if (sync_cfqq) {
2728 /*
2729 * Drop reference to sync queue. A new sync queue will be
2730 * assigned in new group upon arrival of a fresh request.
2731 */
2732 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2733 cic_set_cfqq(cic, NULL, 1);
2734 cfq_put_queue(sync_cfqq);
2735 }
2736
2737 spin_unlock_irqrestore(q->queue_lock, flags);
2738}
2739
2740static void cfq_ioc_set_cgroup(struct io_context *ioc)
2741{
2742 call_for_each_cic(ioc, changed_cgroup);
2743 ioc->cgroup_changed = 0;
2744}
2745#endif /* CONFIG_CFQ_GROUP_IOSCHED */
2746
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002748cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
Jens Axboefd0928d2008-01-24 08:52:45 +01002749 struct io_context *ioc, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751 struct cfq_queue *cfqq, *new_cfqq = NULL;
Vasily Tarasov91fac312007-04-25 12:29:51 +02002752 struct cfq_io_context *cic;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002753 struct cfq_group *cfqg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754
2755retry:
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002756 cfqg = cfq_get_cfqg(cfqd, 1);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002757 cic = cfq_cic_lookup(cfqd, ioc);
Vasily Tarasov91fac312007-04-25 12:29:51 +02002758 /* cic always exists here */
2759 cfqq = cic_to_cfqq(cic, is_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760
Jens Axboe6118b702009-06-30 09:34:12 +02002761 /*
2762 * Always try a new alloc if we fell back to the OOM cfqq
2763 * originally, since it should just be a temporary situation.
2764 */
2765 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2766 cfqq = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 if (new_cfqq) {
2768 cfqq = new_cfqq;
2769 new_cfqq = NULL;
Jens Axboe22e2c502005-06-27 10:55:12 +02002770 } else if (gfp_mask & __GFP_WAIT) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 spin_unlock_irq(cfqd->queue->queue_lock);
Christoph Lameter94f60302007-07-17 04:03:29 -07002772 new_cfqq = kmem_cache_alloc_node(cfq_pool,
Jens Axboe6118b702009-06-30 09:34:12 +02002773 gfp_mask | __GFP_ZERO,
Christoph Lameter94f60302007-07-17 04:03:29 -07002774 cfqd->queue->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 spin_lock_irq(cfqd->queue->queue_lock);
Jens Axboe6118b702009-06-30 09:34:12 +02002776 if (new_cfqq)
2777 goto retry;
Jens Axboe22e2c502005-06-27 10:55:12 +02002778 } else {
Christoph Lameter94f60302007-07-17 04:03:29 -07002779 cfqq = kmem_cache_alloc_node(cfq_pool,
2780 gfp_mask | __GFP_ZERO,
2781 cfqd->queue->node);
Kiyoshi Ueda db3b5842005-06-17 16:15:10 +02002782 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783
Jens Axboe6118b702009-06-30 09:34:12 +02002784 if (cfqq) {
2785 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2786 cfq_init_prio_data(cfqq, ioc);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05002787 cfq_link_cfqq_cfqg(cfqq, cfqg);
Jens Axboe6118b702009-06-30 09:34:12 +02002788 cfq_log_cfqq(cfqd, cfqq, "alloced");
2789 } else
2790 cfqq = &cfqd->oom_cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 }
2792
2793 if (new_cfqq)
2794 kmem_cache_free(cfq_pool, new_cfqq);
2795
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 return cfqq;
2797}
2798
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002799static struct cfq_queue **
2800cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2801{
Jens Axboefe094d92008-01-31 13:08:54 +01002802 switch (ioprio_class) {
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002803 case IOPRIO_CLASS_RT:
2804 return &cfqd->async_cfqq[0][ioprio];
2805 case IOPRIO_CLASS_BE:
2806 return &cfqd->async_cfqq[1][ioprio];
2807 case IOPRIO_CLASS_IDLE:
2808 return &cfqd->async_idle_cfqq;
2809 default:
2810 BUG();
2811 }
2812}
2813
Jens Axboe15c31be2007-07-10 13:43:25 +02002814static struct cfq_queue *
Jens Axboea6151c32009-10-07 20:02:57 +02002815cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
Jens Axboe15c31be2007-07-10 13:43:25 +02002816 gfp_t gfp_mask)
2817{
Jens Axboefd0928d2008-01-24 08:52:45 +01002818 const int ioprio = task_ioprio(ioc);
2819 const int ioprio_class = task_ioprio_class(ioc);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002820 struct cfq_queue **async_cfqq = NULL;
Jens Axboe15c31be2007-07-10 13:43:25 +02002821 struct cfq_queue *cfqq = NULL;
2822
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002823 if (!is_sync) {
2824 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2825 cfqq = *async_cfqq;
2826 }
2827
Jens Axboe6118b702009-06-30 09:34:12 +02002828 if (!cfqq)
Jens Axboefd0928d2008-01-24 08:52:45 +01002829 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
Jens Axboe15c31be2007-07-10 13:43:25 +02002830
2831 /*
2832 * pin the queue now that it's allocated, scheduler exit will prune it
2833 */
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002834 if (!is_sync && !(*async_cfqq)) {
Jens Axboe15c31be2007-07-10 13:43:25 +02002835 atomic_inc(&cfqq->ref);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02002836 *async_cfqq = cfqq;
Jens Axboe15c31be2007-07-10 13:43:25 +02002837 }
2838
2839 atomic_inc(&cfqq->ref);
2840 return cfqq;
2841}
2842
Jens Axboe498d3aa22007-04-26 12:54:48 +02002843/*
2844 * We drop cfq io contexts lazily, so we may find a dead one.
2845 */
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002846static void
Jens Axboe4ac845a2008-01-24 08:44:49 +01002847cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2848 struct cfq_io_context *cic)
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002849{
Jens Axboe4ac845a2008-01-24 08:44:49 +01002850 unsigned long flags;
2851
Jens Axboefc463792006-08-29 09:05:44 +02002852 WARN_ON(!list_empty(&cic->queue_list));
Jens Axboe597bc482007-04-24 21:23:53 +02002853
Jens Axboe4ac845a2008-01-24 08:44:49 +01002854 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe597bc482007-04-24 21:23:53 +02002855
Fabio Checconi4faa3c82008-04-10 08:28:01 +02002856 BUG_ON(ioc->ioc_data == cic);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002857
2858 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
Jens Axboeffc4e752008-02-19 10:02:29 +01002859 hlist_del_rcu(&cic->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002860 spin_unlock_irqrestore(&ioc->lock, flags);
2861
2862 cfq_cic_free(cic);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002863}
2864
Jens Axboee2d74ac2006-03-28 08:59:01 +02002865static struct cfq_io_context *
Jens Axboe4ac845a2008-01-24 08:44:49 +01002866cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002867{
Jens Axboee2d74ac2006-03-28 08:59:01 +02002868 struct cfq_io_context *cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002869 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002870 void *k;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002871
Vasily Tarasov91fac312007-04-25 12:29:51 +02002872 if (unlikely(!ioc))
2873 return NULL;
2874
Jens Axboed6de8be2008-05-28 14:46:59 +02002875 rcu_read_lock();
2876
Jens Axboe597bc482007-04-24 21:23:53 +02002877 /*
2878 * we maintain a last-hit cache, to avoid browsing over the tree
2879 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01002880 cic = rcu_dereference(ioc->ioc_data);
Jens Axboed6de8be2008-05-28 14:46:59 +02002881 if (cic && cic->key == cfqd) {
2882 rcu_read_unlock();
Jens Axboe597bc482007-04-24 21:23:53 +02002883 return cic;
Jens Axboed6de8be2008-05-28 14:46:59 +02002884 }
Jens Axboe597bc482007-04-24 21:23:53 +02002885
Jens Axboe4ac845a2008-01-24 08:44:49 +01002886 do {
Jens Axboe4ac845a2008-01-24 08:44:49 +01002887 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2888 rcu_read_unlock();
2889 if (!cic)
2890 break;
OGAWA Hirofumibe3b0752006-04-18 19:18:31 +02002891 /* ->key must be copied to avoid race with cfq_exit_queue() */
2892 k = cic->key;
2893 if (unlikely(!k)) {
Jens Axboe4ac845a2008-01-24 08:44:49 +01002894 cfq_drop_dead_cic(cfqd, ioc, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002895 rcu_read_lock();
Jens Axboe4ac845a2008-01-24 08:44:49 +01002896 continue;
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002897 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002898
Jens Axboed6de8be2008-05-28 14:46:59 +02002899 spin_lock_irqsave(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002900 rcu_assign_pointer(ioc->ioc_data, cic);
Jens Axboed6de8be2008-05-28 14:46:59 +02002901 spin_unlock_irqrestore(&ioc->lock, flags);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002902 break;
2903 } while (1);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002904
Jens Axboe4ac845a2008-01-24 08:44:49 +01002905 return cic;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002906}
2907
Jens Axboe4ac845a2008-01-24 08:44:49 +01002908/*
2909 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2910 * the process specific cfq io context when entered from the block layer.
2911 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2912 */
Jens Axboefebffd62008-01-28 13:19:43 +01002913static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2914 struct cfq_io_context *cic, gfp_t gfp_mask)
Jens Axboee2d74ac2006-03-28 08:59:01 +02002915{
Jens Axboe0261d682006-10-30 19:07:48 +01002916 unsigned long flags;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002917 int ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002918
Jens Axboe4ac845a2008-01-24 08:44:49 +01002919 ret = radix_tree_preload(gfp_mask);
2920 if (!ret) {
2921 cic->ioc = ioc;
2922 cic->key = cfqd;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002923
Jens Axboe4ac845a2008-01-24 08:44:49 +01002924 spin_lock_irqsave(&ioc->lock, flags);
2925 ret = radix_tree_insert(&ioc->radix_root,
2926 (unsigned long) cfqd, cic);
Jens Axboeffc4e752008-02-19 10:02:29 +01002927 if (!ret)
2928 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
Jens Axboe4ac845a2008-01-24 08:44:49 +01002929 spin_unlock_irqrestore(&ioc->lock, flags);
2930
2931 radix_tree_preload_end();
2932
2933 if (!ret) {
2934 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2935 list_add(&cic->queue_list, &cfqd->cic_list);
2936 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
OGAWA Hirofumidbecf3a2006-04-18 09:45:18 +02002937 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02002938 }
2939
Jens Axboe4ac845a2008-01-24 08:44:49 +01002940 if (ret)
2941 printk(KERN_ERR "cfq: cic link failed!\n");
Jens Axboefc463792006-08-29 09:05:44 +02002942
Jens Axboe4ac845a2008-01-24 08:44:49 +01002943 return ret;
Jens Axboee2d74ac2006-03-28 08:59:01 +02002944}
2945
Jens Axboe22e2c502005-06-27 10:55:12 +02002946/*
2947 * Setup general io context and cfq io context. There can be several cfq
2948 * io contexts per general io context, if this process is doing io to more
Jens Axboee2d74ac2006-03-28 08:59:01 +02002949 * than one device managed by cfq.
Jens Axboe22e2c502005-06-27 10:55:12 +02002950 */
2951static struct cfq_io_context *
Jens Axboee2d74ac2006-03-28 08:59:01 +02002952cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953{
Jens Axboe22e2c502005-06-27 10:55:12 +02002954 struct io_context *ioc = NULL;
2955 struct cfq_io_context *cic;
2956
2957 might_sleep_if(gfp_mask & __GFP_WAIT);
2958
Jens Axboeb5deef92006-07-19 23:39:40 +02002959 ioc = get_io_context(gfp_mask, cfqd->queue->node);
Jens Axboe22e2c502005-06-27 10:55:12 +02002960 if (!ioc)
2961 return NULL;
2962
Jens Axboe4ac845a2008-01-24 08:44:49 +01002963 cic = cfq_cic_lookup(cfqd, ioc);
Jens Axboee2d74ac2006-03-28 08:59:01 +02002964 if (cic)
2965 goto out;
Jens Axboe22e2c502005-06-27 10:55:12 +02002966
Jens Axboee2d74ac2006-03-28 08:59:01 +02002967 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2968 if (cic == NULL)
2969 goto err;
Jens Axboe22e2c502005-06-27 10:55:12 +02002970
Jens Axboe4ac845a2008-01-24 08:44:49 +01002971 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2972 goto err_free;
2973
Jens Axboe22e2c502005-06-27 10:55:12 +02002974out:
Jens Axboefc463792006-08-29 09:05:44 +02002975 smp_read_barrier_depends();
2976 if (unlikely(ioc->ioprio_changed))
2977 cfq_ioc_set_ioprio(ioc);
2978
Vivek Goyal246103332009-12-03 12:59:51 -05002979#ifdef CONFIG_CFQ_GROUP_IOSCHED
2980 if (unlikely(ioc->cgroup_changed))
2981 cfq_ioc_set_cgroup(ioc);
2982#endif
Jens Axboe22e2c502005-06-27 10:55:12 +02002983 return cic;
Jens Axboe4ac845a2008-01-24 08:44:49 +01002984err_free:
2985 cfq_cic_free(cic);
Jens Axboe22e2c502005-06-27 10:55:12 +02002986err:
2987 put_io_context(ioc);
2988 return NULL;
2989}
2990
2991static void
2992cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2993{
Jens Axboeaaf12282007-01-19 11:30:16 +11002994 unsigned long elapsed = jiffies - cic->last_end_request;
2995 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
Jens Axboe22e2c502005-06-27 10:55:12 +02002996
2997 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2998 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2999 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3000}
3001
Jens Axboe206dc692006-03-28 13:03:44 +02003002static void
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003003cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
Jens Axboe6d048f52007-04-25 12:44:27 +02003004 struct request *rq)
Jens Axboe206dc692006-03-28 13:03:44 +02003005{
3006 sector_t sdist;
3007 u64 total;
3008
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003009 if (!cfqq->last_request_pos)
Jeff Moyer4d00aa42009-04-21 07:25:04 +02003010 sdist = 0;
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003011 else if (cfqq->last_request_pos < blk_rq_pos(rq))
3012 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
Jens Axboe206dc692006-03-28 13:03:44 +02003013 else
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003014 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
Jens Axboe206dc692006-03-28 13:03:44 +02003015
3016 /*
3017 * Don't allow the seek distance to get too large from the
3018 * odd fragment, pagein, etc
3019 */
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003020 if (cfqq->seek_samples <= 60) /* second&third seek */
3021 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
Jens Axboe206dc692006-03-28 13:03:44 +02003022 else
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003023 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
Jens Axboe206dc692006-03-28 13:03:44 +02003024
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003025 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
3026 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
3027 total = cfqq->seek_total + (cfqq->seek_samples/2);
3028 do_div(total, cfqq->seek_samples);
3029 cfqq->seek_mean = (sector_t)total;
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003030
3031 /*
3032 * If this cfqq is shared between multiple processes, check to
3033 * make sure that those processes are still issuing I/Os within
3034 * the mean seek distance. If not, it may be time to break the
3035 * queues apart again.
3036 */
3037 if (cfq_cfqq_coop(cfqq)) {
3038 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
3039 cfqq->seeky_start = jiffies;
3040 else if (!CFQQ_SEEKY(cfqq))
3041 cfqq->seeky_start = 0;
3042 }
Jens Axboe206dc692006-03-28 13:03:44 +02003043}
Jens Axboe22e2c502005-06-27 10:55:12 +02003044
3045/*
3046 * Disable idle window if the process thinks too long or seeks so much that
3047 * it doesn't matter
3048 */
3049static void
3050cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3051 struct cfq_io_context *cic)
3052{
Jens Axboe7b679132008-05-30 12:23:07 +02003053 int old_idle, enable_idle;
Jens Axboe1be92f22007-04-19 14:32:26 +02003054
Jens Axboe08717142008-01-28 11:38:15 +01003055 /*
3056 * Don't idle for async or idle io prio class
3057 */
3058 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
Jens Axboe1be92f22007-04-19 14:32:26 +02003059 return;
3060
Jens Axboec265a7f2008-06-26 13:49:33 +02003061 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003062
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003063 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3064 cfq_mark_cfqq_deep(cfqq);
3065
Nikanth Karthikesan66dac982007-11-27 12:47:04 +01003066 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003067 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
3068 && CFQQ_SEEKY(cfqq)))
Jens Axboe22e2c502005-06-27 10:55:12 +02003069 enable_idle = 0;
3070 else if (sample_valid(cic->ttime_samples)) {
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003071 if (cic->ttime_mean > cfqd->cfq_slice_idle)
Jens Axboe22e2c502005-06-27 10:55:12 +02003072 enable_idle = 0;
3073 else
3074 enable_idle = 1;
3075 }
3076
Jens Axboe7b679132008-05-30 12:23:07 +02003077 if (old_idle != enable_idle) {
3078 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3079 if (enable_idle)
3080 cfq_mark_cfqq_idle_window(cfqq);
3081 else
3082 cfq_clear_cfqq_idle_window(cfqq);
3083 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003084}
3085
Jens Axboe22e2c502005-06-27 10:55:12 +02003086/*
3087 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3088 * no or if we aren't sure, a 1 will cause a preempt.
3089 */
Jens Axboea6151c32009-10-07 20:02:57 +02003090static bool
Jens Axboe22e2c502005-06-27 10:55:12 +02003091cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
Jens Axboe5e705372006-07-13 12:39:25 +02003092 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003093{
Jens Axboe6d048f52007-04-25 12:44:27 +02003094 struct cfq_queue *cfqq;
Jens Axboe22e2c502005-06-27 10:55:12 +02003095
Jens Axboe6d048f52007-04-25 12:44:27 +02003096 cfqq = cfqd->active_queue;
3097 if (!cfqq)
Jens Axboea6151c32009-10-07 20:02:57 +02003098 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003099
Jens Axboe6d048f52007-04-25 12:44:27 +02003100 if (cfq_class_idle(new_cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003101 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003102
3103 if (cfq_class_idle(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003104 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003105
Jens Axboe22e2c502005-06-27 10:55:12 +02003106 /*
Jens Axboe374f84a2006-07-23 01:42:19 +02003107 * if the new request is sync, but the currently running queue is
3108 * not, let the sync request have priority.
3109 */
Jens Axboe5e705372006-07-13 12:39:25 +02003110 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003111 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003112
Vivek Goyal8682e1f2009-12-03 12:59:50 -05003113 if (new_cfqq->cfqg != cfqq->cfqg)
3114 return false;
3115
3116 if (cfq_slice_used(cfqq))
3117 return true;
3118
3119 /* Allow preemption only if we are idling on sync-noidle tree */
3120 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3121 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3122 new_cfqq->service_tree->count == 2 &&
3123 RB_EMPTY_ROOT(&cfqq->sort_list))
3124 return true;
3125
Jens Axboe374f84a2006-07-23 01:42:19 +02003126 /*
3127 * So both queues are sync. Let the new request get disk time if
3128 * it's a metadata request and the current queue is doing regular IO.
3129 */
3130 if (rq_is_meta(rq) && !cfqq->meta_pending)
Jens Axboee6ec4fe2009-11-03 20:21:35 +01003131 return true;
Jens Axboe22e2c502005-06-27 10:55:12 +02003132
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003133 /*
3134 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3135 */
3136 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003137 return true;
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003138
Jens Axboe1e3335d2007-02-14 19:59:49 +01003139 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
Jens Axboea6151c32009-10-07 20:02:57 +02003140 return false;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003141
3142 /*
3143 * if this request is as-good as one we would expect from the
3144 * current cfqq, let it preempt
3145 */
Jens Axboee00ef792009-11-04 08:54:55 +01003146 if (cfq_rq_close(cfqd, cfqq, rq))
Jens Axboea6151c32009-10-07 20:02:57 +02003147 return true;
Jens Axboe1e3335d2007-02-14 19:59:49 +01003148
Jens Axboea6151c32009-10-07 20:02:57 +02003149 return false;
Jens Axboe22e2c502005-06-27 10:55:12 +02003150}
3151
3152/*
3153 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3154 * let it have half of its nominal slice.
3155 */
3156static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3157{
Jens Axboe7b679132008-05-30 12:23:07 +02003158 cfq_log_cfqq(cfqd, cfqq, "preempt");
Jens Axboe6084cdd2007-04-23 08:25:00 +02003159 cfq_slice_expired(cfqd, 1);
Jens Axboe22e2c502005-06-27 10:55:12 +02003160
Jens Axboebf572252006-07-19 20:29:12 +02003161 /*
3162 * Put the new queue at the front of the of the current list,
3163 * so we know that it will be selected next.
3164 */
3165 BUG_ON(!cfq_cfqq_on_rr(cfqq));
Jens Axboeedd75ff2007-04-19 12:03:34 +02003166
3167 cfq_service_tree_add(cfqd, cfqq, 1);
Jens Axboebf572252006-07-19 20:29:12 +02003168
Jens Axboe44f7c162007-01-19 11:51:58 +11003169 cfqq->slice_end = 0;
3170 cfq_mark_cfqq_slice_new(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003171}
3172
3173/*
Jens Axboe5e705372006-07-13 12:39:25 +02003174 * Called when a new fs request (rq) is added (to cfqq). Check if there's
Jens Axboe22e2c502005-06-27 10:55:12 +02003175 * something we should do about it
3176 */
3177static void
Jens Axboe5e705372006-07-13 12:39:25 +02003178cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3179 struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003180{
Jens Axboe5e705372006-07-13 12:39:25 +02003181 struct cfq_io_context *cic = RQ_CIC(rq);
Jens Axboe12e9fdd2006-06-01 10:09:56 +02003182
Aaron Carroll45333d52008-08-26 15:52:36 +02003183 cfqd->rq_queued++;
Jens Axboe374f84a2006-07-23 01:42:19 +02003184 if (rq_is_meta(rq))
3185 cfqq->meta_pending++;
3186
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003187 cfq_update_io_thinktime(cfqd, cic);
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003188 cfq_update_io_seektime(cfqd, cfqq, rq);
Jens Axboe9c2c38a2005-08-24 14:57:54 +02003189 cfq_update_idle_window(cfqd, cfqq, cic);
3190
Jeff Moyerb2c18e12009-10-23 17:14:49 -04003191 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003192
3193 if (cfqq == cfqd->active_queue) {
3194 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003195 * Remember that we saw a request from this process, but
3196 * don't start queuing just yet. Otherwise we risk seeing lots
3197 * of tiny requests, because we disrupt the normal plugging
Jens Axboed6ceb252009-04-14 14:18:16 +02003198 * and merging. If the request is already larger than a single
3199 * page, let it rip immediately. For that case we assume that
Jens Axboe2d870722009-04-15 12:12:46 +02003200 * merging is already done. Ditto for a busy system that
3201 * has other work pending, don't risk delaying until the
3202 * idle timer unplug to continue working.
Jens Axboe22e2c502005-06-27 10:55:12 +02003203 */
Jens Axboed6ceb252009-04-14 14:18:16 +02003204 if (cfq_cfqq_wait_request(cfqq)) {
Jens Axboe2d870722009-04-15 12:12:46 +02003205 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3206 cfqd->busy_queues > 1) {
Jens Axboed6ceb252009-04-14 14:18:16 +02003207 del_timer(&cfqd->idle_slice_timer);
Gui Jianfeng554554f2009-12-10 09:38:39 +01003208 cfq_clear_cfqq_wait_request(cfqq);
Vivek Goyalbf7919372009-12-03 12:59:37 -05003209 __blk_run_queue(cfqd->queue);
3210 } else
3211 cfq_mark_cfqq_must_dispatch(cfqq);
Jens Axboed6ceb252009-04-14 14:18:16 +02003212 }
Jens Axboe5e705372006-07-13 12:39:25 +02003213 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003214 /*
3215 * not the active queue - expire current slice if it is
3216 * idle and has expired it's mean thinktime or this new queue
Divyesh Shah3a9a3f62009-01-30 12:46:41 +01003217 * has some old slice time left and is of higher priority or
3218 * this new queue is RT and the current one is BE
Jens Axboe22e2c502005-06-27 10:55:12 +02003219 */
3220 cfq_preempt_queue(cfqd, cfqq);
Tejun Heoa7f55792009-04-23 11:05:17 +09003221 __blk_run_queue(cfqd->queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003222 }
3223}
3224
Jens Axboe165125e2007-07-24 09:28:11 +02003225static void cfq_insert_request(struct request_queue *q, struct request *rq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003226{
Jens Axboeb4878f22005-10-20 16:42:29 +02003227 struct cfq_data *cfqd = q->elevator->elevator_data;
Jens Axboe5e705372006-07-13 12:39:25 +02003228 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003229
Jens Axboe7b679132008-05-30 12:23:07 +02003230 cfq_log_cfqq(cfqd, cfqq, "insert_request");
Jens Axboefd0928d2008-01-24 08:52:45 +01003231 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232
Jens Axboe30996f42009-10-05 11:03:39 +02003233 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
Jens Axboe22e2c502005-06-27 10:55:12 +02003234 list_add_tail(&rq->queuelist, &cfqq->fifo);
Corrado Zoccoloaa6f6a32009-10-26 22:44:33 +01003235 cfq_add_rq_rb(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003236
Jens Axboe5e705372006-07-13 12:39:25 +02003237 cfq_rq_enqueued(cfqd, cfqq, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238}
3239
Aaron Carroll45333d52008-08-26 15:52:36 +02003240/*
3241 * Update hw_tag based on peak queue depth over 50 samples under
3242 * sufficient load.
3243 */
3244static void cfq_update_hw_tag(struct cfq_data *cfqd)
3245{
Shaohua Li1a1238a2009-10-27 08:46:23 +01003246 struct cfq_queue *cfqq = cfqd->active_queue;
3247
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003248 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3249 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3250
3251 if (cfqd->hw_tag == 1)
3252 return;
Aaron Carroll45333d52008-08-26 15:52:36 +02003253
3254 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
Jens Axboe5ad531d2009-07-03 12:57:48 +02003255 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003256 return;
3257
Shaohua Li1a1238a2009-10-27 08:46:23 +01003258 /*
3259 * If active queue hasn't enough requests and can idle, cfq might not
3260 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3261 * case
3262 */
3263 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3264 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3265 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3266 return;
3267
Aaron Carroll45333d52008-08-26 15:52:36 +02003268 if (cfqd->hw_tag_samples++ < 50)
3269 return;
3270
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003271 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
Aaron Carroll45333d52008-08-26 15:52:36 +02003272 cfqd->hw_tag = 1;
3273 else
3274 cfqd->hw_tag = 0;
Aaron Carroll45333d52008-08-26 15:52:36 +02003275}
3276
Vivek Goyal7667aa02009-12-08 17:52:58 -05003277static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3278{
3279 struct cfq_io_context *cic = cfqd->active_cic;
3280
3281 /* If there are other queues in the group, don't wait */
3282 if (cfqq->cfqg->nr_cfqq > 1)
3283 return false;
3284
3285 if (cfq_slice_used(cfqq))
3286 return true;
3287
3288 /* if slice left is less than think time, wait busy */
3289 if (cic && sample_valid(cic->ttime_samples)
3290 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3291 return true;
3292
3293 /*
3294 * If think times is less than a jiffy than ttime_mean=0 and above
3295 * will not be true. It might happen that slice has not expired yet
3296 * but will expire soon (4-5 ns) during select_queue(). To cover the
3297 * case where think time is less than a jiffy, mark the queue wait
3298 * busy if only 1 jiffy is left in the slice.
3299 */
3300 if (cfqq->slice_end - jiffies == 1)
3301 return true;
3302
3303 return false;
3304}
3305
Jens Axboe165125e2007-07-24 09:28:11 +02003306static void cfq_completed_request(struct request_queue *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003307{
Jens Axboe5e705372006-07-13 12:39:25 +02003308 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003309 struct cfq_data *cfqd = cfqq->cfqd;
Jens Axboe5380a102006-07-13 12:37:56 +02003310 const int sync = rq_is_sync(rq);
Jens Axboeb4878f22005-10-20 16:42:29 +02003311 unsigned long now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312
Jens Axboeb4878f22005-10-20 16:42:29 +02003313 now = jiffies;
Vivek Goyal2868ef72009-12-03 12:59:48 -05003314 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315
Aaron Carroll45333d52008-08-26 15:52:36 +02003316 cfq_update_hw_tag(cfqd);
3317
Jens Axboe5ad531d2009-07-03 12:57:48 +02003318 WARN_ON(!cfqd->rq_in_driver[sync]);
Jens Axboe6d048f52007-04-25 12:44:27 +02003319 WARN_ON(!cfqq->dispatched);
Jens Axboe5ad531d2009-07-03 12:57:48 +02003320 cfqd->rq_in_driver[sync]--;
Jens Axboe6d048f52007-04-25 12:44:27 +02003321 cfqq->dispatched--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322
Jens Axboe3ed9a292007-04-23 08:33:33 +02003323 if (cfq_cfqq_sync(cfqq))
3324 cfqd->sync_flight--;
3325
Vivek Goyal365722b2009-10-03 15:21:27 +02003326 if (sync) {
Jens Axboe5e705372006-07-13 12:39:25 +02003327 RQ_CIC(rq)->last_end_request = now;
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003328 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3329 cfqd->last_delayed_sync = now;
Vivek Goyal365722b2009-10-03 15:21:27 +02003330 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003331
3332 /*
3333 * If this is the active queue, check if it needs to be expired,
3334 * or if we want to idle in case it has no pending requests.
3335 */
3336 if (cfqd->active_queue == cfqq) {
Jens Axboea36e71f2009-04-15 12:15:11 +02003337 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3338
Jens Axboe44f7c162007-01-19 11:51:58 +11003339 if (cfq_cfqq_slice_new(cfqq)) {
3340 cfq_set_prio_slice(cfqd, cfqq);
3341 cfq_clear_cfqq_slice_new(cfqq);
3342 }
Vivek Goyalf75edf22009-12-03 12:59:53 -05003343
3344 /*
Vivek Goyal7667aa02009-12-08 17:52:58 -05003345 * Should we wait for next request to come in before we expire
3346 * the queue.
Vivek Goyalf75edf22009-12-03 12:59:53 -05003347 */
Vivek Goyal7667aa02009-12-08 17:52:58 -05003348 if (cfq_should_wait_busy(cfqd, cfqq)) {
Vivek Goyalf75edf22009-12-03 12:59:53 -05003349 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3350 cfq_mark_cfqq_wait_busy(cfqq);
3351 }
3352
Jens Axboea36e71f2009-04-15 12:15:11 +02003353 /*
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003354 * Idling is not enabled on:
3355 * - expired queues
3356 * - idle-priority queues
3357 * - async queues
3358 * - queues with still some requests queued
3359 * - when there is a close cooperator
Jens Axboea36e71f2009-04-15 12:15:11 +02003360 */
Jens Axboe08717142008-01-28 11:38:15 +01003361 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
Jens Axboe6084cdd2007-04-23 08:25:00 +02003362 cfq_slice_expired(cfqd, 1);
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003363 else if (sync && cfqq_empty &&
3364 !cfq_close_cooperator(cfqd, cfqq)) {
3365 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3366 /*
3367 * Idling is enabled for SYNC_WORKLOAD.
3368 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3369 * only if we processed at least one !rq_noidle request
3370 */
3371 if (cfqd->serving_type == SYNC_WORKLOAD
Vivek Goyalc04645e2009-12-03 12:59:56 -05003372 || cfqd->noidle_tree_requires_idle
3373 || cfqq->cfqg->nr_cfqq == 1)
Corrado Zoccolo8e550632009-11-26 10:02:58 +01003374 cfq_arm_slice_timer(cfqd);
3375 }
Jens Axboecaaa5f92006-06-16 11:23:00 +02003376 }
Jens Axboe6d048f52007-04-25 12:44:27 +02003377
Jens Axboe5ad531d2009-07-03 12:57:48 +02003378 if (!rq_in_driver(cfqd))
Jens Axboe23e018a2009-10-05 08:52:35 +02003379 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380}
3381
Jens Axboe22e2c502005-06-27 10:55:12 +02003382/*
3383 * we temporarily boost lower priority queues if they are holding fs exclusive
3384 * resources. they are boosted to normal prio (CLASS_BE/4)
3385 */
3386static void cfq_prio_boost(struct cfq_queue *cfqq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387{
Jens Axboe22e2c502005-06-27 10:55:12 +02003388 if (has_fs_excl()) {
3389 /*
3390 * boost idle prio on transactions that would lock out other
3391 * users of the filesystem
3392 */
3393 if (cfq_class_idle(cfqq))
3394 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3395 if (cfqq->ioprio > IOPRIO_NORM)
3396 cfqq->ioprio = IOPRIO_NORM;
3397 } else {
3398 /*
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003399 * unboost the queue (if needed)
Jens Axboe22e2c502005-06-27 10:55:12 +02003400 */
Corrado Zoccolodddb7452009-11-02 10:40:37 +01003401 cfqq->ioprio_class = cfqq->org_ioprio_class;
3402 cfqq->ioprio = cfqq->org_ioprio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003404}
3405
Jens Axboe89850f72006-07-22 16:48:31 +02003406static inline int __cfq_may_queue(struct cfq_queue *cfqq)
Jens Axboe22e2c502005-06-27 10:55:12 +02003407{
Jens Axboe1b379d82009-08-11 08:26:11 +02003408 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
Jens Axboe3b181522005-06-27 10:56:24 +02003409 cfq_mark_cfqq_must_alloc_slice(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003410 return ELV_MQUEUE_MUST;
Jens Axboe3b181522005-06-27 10:56:24 +02003411 }
Jens Axboe22e2c502005-06-27 10:55:12 +02003412
3413 return ELV_MQUEUE_MAY;
Jens Axboe22e2c502005-06-27 10:55:12 +02003414}
3415
Jens Axboe165125e2007-07-24 09:28:11 +02003416static int cfq_may_queue(struct request_queue *q, int rw)
Jens Axboe22e2c502005-06-27 10:55:12 +02003417{
3418 struct cfq_data *cfqd = q->elevator->elevator_data;
3419 struct task_struct *tsk = current;
Vasily Tarasov91fac312007-04-25 12:29:51 +02003420 struct cfq_io_context *cic;
Jens Axboe22e2c502005-06-27 10:55:12 +02003421 struct cfq_queue *cfqq;
3422
3423 /*
3424 * don't force setup of a queue from here, as a call to may_queue
3425 * does not necessarily imply that a request actually will be queued.
3426 * so just lookup a possibly existing queue, or return 'may queue'
3427 * if that fails
3428 */
Jens Axboe4ac845a2008-01-24 08:44:49 +01003429 cic = cfq_cic_lookup(cfqd, tsk->io_context);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003430 if (!cic)
3431 return ELV_MQUEUE_MAY;
3432
Jens Axboeb0b78f82009-04-08 10:56:08 +02003433 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
Jens Axboe22e2c502005-06-27 10:55:12 +02003434 if (cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003435 cfq_init_prio_data(cfqq, cic->ioc);
Jens Axboe22e2c502005-06-27 10:55:12 +02003436 cfq_prio_boost(cfqq);
3437
Jens Axboe89850f72006-07-22 16:48:31 +02003438 return __cfq_may_queue(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003439 }
3440
3441 return ELV_MQUEUE_MAY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442}
3443
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444/*
3445 * queue lock held here
3446 */
Jens Axboebb37b942006-12-01 10:42:33 +01003447static void cfq_put_request(struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448{
Jens Axboe5e705372006-07-13 12:39:25 +02003449 struct cfq_queue *cfqq = RQ_CFQQ(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450
Jens Axboe5e705372006-07-13 12:39:25 +02003451 if (cfqq) {
Jens Axboe22e2c502005-06-27 10:55:12 +02003452 const int rw = rq_data_dir(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453
Jens Axboe22e2c502005-06-27 10:55:12 +02003454 BUG_ON(!cfqq->allocated[rw]);
3455 cfqq->allocated[rw]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456
Jens Axboe5e705372006-07-13 12:39:25 +02003457 put_io_context(RQ_CIC(rq)->ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459 rq->elevator_private = NULL;
Jens Axboe5e705372006-07-13 12:39:25 +02003460 rq->elevator_private2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462 cfq_put_queue(cfqq);
3463 }
3464}
3465
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003466static struct cfq_queue *
3467cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3468 struct cfq_queue *cfqq)
3469{
3470 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3471 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
Jeff Moyerb3b6d042009-10-23 17:14:51 -04003472 cfq_mark_cfqq_coop(cfqq->new_cfqq);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003473 cfq_put_queue(cfqq);
3474 return cic_to_cfqq(cic, 1);
3475}
3476
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003477static int should_split_cfqq(struct cfq_queue *cfqq)
3478{
3479 if (cfqq->seeky_start &&
3480 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3481 return 1;
3482 return 0;
3483}
3484
3485/*
3486 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3487 * was the last process referring to said cfqq.
3488 */
3489static struct cfq_queue *
3490split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3491{
3492 if (cfqq_process_refs(cfqq) == 1) {
3493 cfqq->seeky_start = 0;
3494 cfqq->pid = current->pid;
3495 cfq_clear_cfqq_coop(cfqq);
3496 return cfqq;
3497 }
3498
3499 cic_set_cfqq(cic, NULL, 1);
3500 cfq_put_queue(cfqq);
3501 return NULL;
3502}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503/*
Jens Axboe22e2c502005-06-27 10:55:12 +02003504 * Allocate cfq data structures associated with this request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 */
Jens Axboe22e2c502005-06-27 10:55:12 +02003506static int
Jens Axboe165125e2007-07-24 09:28:11 +02003507cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508{
3509 struct cfq_data *cfqd = q->elevator->elevator_data;
3510 struct cfq_io_context *cic;
3511 const int rw = rq_data_dir(rq);
Jens Axboea6151c32009-10-07 20:02:57 +02003512 const bool is_sync = rq_is_sync(rq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003513 struct cfq_queue *cfqq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 unsigned long flags;
3515
3516 might_sleep_if(gfp_mask & __GFP_WAIT);
3517
Jens Axboee2d74ac2006-03-28 08:59:01 +02003518 cic = cfq_get_io_context(cfqd, gfp_mask);
Jens Axboe22e2c502005-06-27 10:55:12 +02003519
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 spin_lock_irqsave(q->queue_lock, flags);
3521
Jens Axboe22e2c502005-06-27 10:55:12 +02003522 if (!cic)
3523 goto queue_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003525new_queue:
Vasily Tarasov91fac312007-04-25 12:29:51 +02003526 cfqq = cic_to_cfqq(cic, is_sync);
Vivek Goyal32f2e802009-07-09 22:13:16 +02003527 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
Jens Axboefd0928d2008-01-24 08:52:45 +01003528 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003529 cic_set_cfqq(cic, cfqq, is_sync);
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003530 } else {
3531 /*
Jeff Moyere6c5bc72009-10-23 17:14:52 -04003532 * If the queue was seeky for too long, break it apart.
3533 */
3534 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3535 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3536 cfqq = split_cfqq(cic, cfqq);
3537 if (!cfqq)
3538 goto new_queue;
3539 }
3540
3541 /*
Jeff Moyerdf5fe3e2009-10-23 17:14:50 -04003542 * Check to see if this queue is scheduled to merge with
3543 * another, closely cooperating queue. The merging of
3544 * queues happens here as it must be done in process context.
3545 * The reference on new_cfqq was taken in merge_cfqqs.
3546 */
3547 if (cfqq->new_cfqq)
3548 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
Vasily Tarasov91fac312007-04-25 12:29:51 +02003549 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550
3551 cfqq->allocated[rw]++;
Jens Axboe22e2c502005-06-27 10:55:12 +02003552 atomic_inc(&cfqq->ref);
Jens Axboe5e705372006-07-13 12:39:25 +02003553
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554 spin_unlock_irqrestore(q->queue_lock, flags);
3555
Jens Axboe5e705372006-07-13 12:39:25 +02003556 rq->elevator_private = cic;
3557 rq->elevator_private2 = cfqq;
3558 return 0;
Jens Axboe3b181522005-06-27 10:56:24 +02003559
Jens Axboe22e2c502005-06-27 10:55:12 +02003560queue_fail:
3561 if (cic)
3562 put_io_context(cic->ioc);
Jens Axboe89850f72006-07-22 16:48:31 +02003563
Jens Axboe23e018a2009-10-05 08:52:35 +02003564 cfq_schedule_dispatch(cfqd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565 spin_unlock_irqrestore(q->queue_lock, flags);
Jens Axboe7b679132008-05-30 12:23:07 +02003566 cfq_log(cfqd, "set_request fail");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 return 1;
3568}
3569
David Howells65f27f32006-11-22 14:55:48 +00003570static void cfq_kick_queue(struct work_struct *work)
Jens Axboe22e2c502005-06-27 10:55:12 +02003571{
David Howells65f27f32006-11-22 14:55:48 +00003572 struct cfq_data *cfqd =
Jens Axboe23e018a2009-10-05 08:52:35 +02003573 container_of(work, struct cfq_data, unplug_work);
Jens Axboe165125e2007-07-24 09:28:11 +02003574 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003575
Jens Axboe40bb54d2009-04-15 12:11:10 +02003576 spin_lock_irq(q->queue_lock);
Tejun Heoa7f55792009-04-23 11:05:17 +09003577 __blk_run_queue(cfqd->queue);
Jens Axboe40bb54d2009-04-15 12:11:10 +02003578 spin_unlock_irq(q->queue_lock);
Jens Axboe22e2c502005-06-27 10:55:12 +02003579}
3580
3581/*
3582 * Timer running if the active_queue is currently idling inside its time slice
3583 */
3584static void cfq_idle_slice_timer(unsigned long data)
3585{
3586 struct cfq_data *cfqd = (struct cfq_data *) data;
3587 struct cfq_queue *cfqq;
3588 unsigned long flags;
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003589 int timed_out = 1;
Jens Axboe22e2c502005-06-27 10:55:12 +02003590
Jens Axboe7b679132008-05-30 12:23:07 +02003591 cfq_log(cfqd, "idle timer fired");
3592
Jens Axboe22e2c502005-06-27 10:55:12 +02003593 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3594
Jens Axboefe094d92008-01-31 13:08:54 +01003595 cfqq = cfqd->active_queue;
3596 if (cfqq) {
Jens Axboe3c6bd2f2007-01-19 12:06:33 +11003597 timed_out = 0;
3598
Jens Axboe22e2c502005-06-27 10:55:12 +02003599 /*
Jens Axboeb0291952009-04-07 11:38:31 +02003600 * We saw a request before the queue expired, let it through
3601 */
3602 if (cfq_cfqq_must_dispatch(cfqq))
3603 goto out_kick;
3604
3605 /*
Jens Axboe22e2c502005-06-27 10:55:12 +02003606 * expired
3607 */
Jens Axboe44f7c162007-01-19 11:51:58 +11003608 if (cfq_slice_used(cfqq))
Jens Axboe22e2c502005-06-27 10:55:12 +02003609 goto expire;
3610
3611 /*
3612 * only expire and reinvoke request handler, if there are
3613 * other queues with pending requests
3614 */
Jens Axboecaaa5f92006-06-16 11:23:00 +02003615 if (!cfqd->busy_queues)
Jens Axboe22e2c502005-06-27 10:55:12 +02003616 goto out_cont;
Jens Axboe22e2c502005-06-27 10:55:12 +02003617
3618 /*
3619 * not expired and it has a request pending, let it dispatch
3620 */
Jens Axboe75e50982009-04-07 08:56:14 +02003621 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
Jens Axboe22e2c502005-06-27 10:55:12 +02003622 goto out_kick;
Corrado Zoccolo76280af2009-11-26 10:02:58 +01003623
3624 /*
3625 * Queue depth flag is reset only when the idle didn't succeed
3626 */
3627 cfq_clear_cfqq_deep(cfqq);
Jens Axboe22e2c502005-06-27 10:55:12 +02003628 }
3629expire:
Jens Axboe6084cdd2007-04-23 08:25:00 +02003630 cfq_slice_expired(cfqd, timed_out);
Jens Axboe22e2c502005-06-27 10:55:12 +02003631out_kick:
Jens Axboe23e018a2009-10-05 08:52:35 +02003632 cfq_schedule_dispatch(cfqd);
Jens Axboe22e2c502005-06-27 10:55:12 +02003633out_cont:
3634 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3635}
3636
Jens Axboe3b181522005-06-27 10:56:24 +02003637static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3638{
3639 del_timer_sync(&cfqd->idle_slice_timer);
Jens Axboe23e018a2009-10-05 08:52:35 +02003640 cancel_work_sync(&cfqd->unplug_work);
Jens Axboe3b181522005-06-27 10:56:24 +02003641}
Jens Axboe22e2c502005-06-27 10:55:12 +02003642
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003643static void cfq_put_async_queues(struct cfq_data *cfqd)
3644{
3645 int i;
3646
3647 for (i = 0; i < IOPRIO_BE_NR; i++) {
3648 if (cfqd->async_cfqq[0][i])
3649 cfq_put_queue(cfqd->async_cfqq[0][i]);
3650 if (cfqd->async_cfqq[1][i])
3651 cfq_put_queue(cfqd->async_cfqq[1][i]);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003652 }
Oleg Nesterov2389d1e2007-11-05 08:58:05 +01003653
3654 if (cfqd->async_idle_cfqq)
3655 cfq_put_queue(cfqd->async_idle_cfqq);
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003656}
3657
Jens Axboebb729bc2009-12-06 09:54:19 +01003658static void cfq_cfqd_free(struct rcu_head *head)
3659{
3660 kfree(container_of(head, struct cfq_data, rcu));
3661}
3662
Jens Axboeb374d182008-10-31 10:05:07 +01003663static void cfq_exit_queue(struct elevator_queue *e)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664{
Jens Axboe22e2c502005-06-27 10:55:12 +02003665 struct cfq_data *cfqd = e->elevator_data;
Jens Axboe165125e2007-07-24 09:28:11 +02003666 struct request_queue *q = cfqd->queue;
Jens Axboe22e2c502005-06-27 10:55:12 +02003667
Jens Axboe3b181522005-06-27 10:56:24 +02003668 cfq_shutdown_timer_wq(cfqd);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003669
Al Virod9ff4182006-03-18 13:51:22 -05003670 spin_lock_irq(q->queue_lock);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003671
Al Virod9ff4182006-03-18 13:51:22 -05003672 if (cfqd->active_queue)
Jens Axboe6084cdd2007-04-23 08:25:00 +02003673 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
Jens Axboee2d74ac2006-03-28 08:59:01 +02003674
3675 while (!list_empty(&cfqd->cic_list)) {
Al Virod9ff4182006-03-18 13:51:22 -05003676 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3677 struct cfq_io_context,
3678 queue_list);
Jens Axboe89850f72006-07-22 16:48:31 +02003679
3680 __cfq_exit_single_io_context(cfqd, cic);
Al Virod9ff4182006-03-18 13:51:22 -05003681 }
Jens Axboee2d74ac2006-03-28 08:59:01 +02003682
Vasily Tarasovc2dea2d2007-07-20 10:06:38 +02003683 cfq_put_async_queues(cfqd);
Vivek Goyalb1c35762009-12-03 12:59:47 -05003684 cfq_release_cfq_groups(cfqd);
3685 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
Jens Axboe15c31be2007-07-10 13:43:25 +02003686
Al Virod9ff4182006-03-18 13:51:22 -05003687 spin_unlock_irq(q->queue_lock);
Al Viroa90d7422006-03-18 12:05:37 -05003688
3689 cfq_shutdown_timer_wq(cfqd);
3690
Vivek Goyalb1c35762009-12-03 12:59:47 -05003691 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
Jens Axboebb729bc2009-12-06 09:54:19 +01003692 call_rcu(&cfqd->rcu, cfq_cfqd_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693}
3694
Jens Axboe165125e2007-07-24 09:28:11 +02003695static void *cfq_init_queue(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696{
3697 struct cfq_data *cfqd;
Corrado Zoccolo718eee02009-10-26 22:45:29 +01003698 int i, j;
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003699 struct cfq_group *cfqg;
Vivek Goyal615f0252009-12-03 12:59:39 -05003700 struct cfq_rb_root *st;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701
Christoph Lameter94f60302007-07-17 04:03:29 -07003702 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703 if (!cfqd)
Jens Axboebc1c1162006-06-08 08:49:06 +02003704 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003706 /* Init root service tree */
3707 cfqd->grp_service_tree = CFQ_RB_ROOT;
3708
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003709 /* Init root group */
3710 cfqg = &cfqd->root_group;
Vivek Goyal615f0252009-12-03 12:59:39 -05003711 for_each_cfqg_st(cfqg, i, j, st)
3712 *st = CFQ_RB_ROOT;
Vivek Goyal1fa8f6d2009-12-03 12:59:41 -05003713 RB_CLEAR_NODE(&cfqg->rb_node);
Jens Axboe26a2ac02009-04-23 12:13:27 +02003714
Vivek Goyal25bc6b02009-12-03 12:59:43 -05003715 /* Give preference to root group over other groups */
3716 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3717
Vivek Goyal25fb5162009-12-03 12:59:46 -05003718#ifdef CONFIG_CFQ_GROUP_IOSCHED
Vivek Goyalb1c35762009-12-03 12:59:47 -05003719 /*
3720 * Take a reference to root group which we never drop. This is just
3721 * to make sure that cfq_put_cfqg() does not try to kfree root group
3722 */
3723 atomic_set(&cfqg->ref, 1);
Vivek Goyal22084192009-12-03 12:59:49 -05003724 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3725 0);
Vivek Goyal25fb5162009-12-03 12:59:46 -05003726#endif
Jens Axboe26a2ac02009-04-23 12:13:27 +02003727 /*
3728 * Not strictly needed (since RB_ROOT just clears the node and we
3729 * zeroed cfqd on alloc), but better be safe in case someone decides
3730 * to add magic to the rb code
3731 */
3732 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3733 cfqd->prio_trees[i] = RB_ROOT;
3734
Jens Axboe6118b702009-06-30 09:34:12 +02003735 /*
3736 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3737 * Grab a permanent reference to it, so that the normal code flow
3738 * will not attempt to free it.
3739 */
3740 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3741 atomic_inc(&cfqd->oom_cfqq.ref);
Vivek Goyalcdb16e82009-12-03 12:59:38 -05003742 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
Jens Axboe6118b702009-06-30 09:34:12 +02003743
Al Virod9ff4182006-03-18 13:51:22 -05003744 INIT_LIST_HEAD(&cfqd->cic_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746 cfqd->queue = q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747
Jens Axboe22e2c502005-06-27 10:55:12 +02003748 init_timer(&cfqd->idle_slice_timer);
3749 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3750 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3751
Jens Axboe23e018a2009-10-05 08:52:35 +02003752 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
Jens Axboe22e2c502005-06-27 10:55:12 +02003753
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754 cfqd->cfq_quantum = cfq_quantum;
Jens Axboe22e2c502005-06-27 10:55:12 +02003755 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3756 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757 cfqd->cfq_back_max = cfq_back_max;
3758 cfqd->cfq_back_penalty = cfq_back_penalty;
Jens Axboe22e2c502005-06-27 10:55:12 +02003759 cfqd->cfq_slice[0] = cfq_slice_async;
3760 cfqd->cfq_slice[1] = cfq_slice_sync;
3761 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3762 cfqd->cfq_slice_idle = cfq_slice_idle;
Jens Axboe963b72f2009-10-03 19:42:18 +02003763 cfqd->cfq_latency = 1;
Vivek Goyalae30c282009-12-03 12:59:55 -05003764 cfqd->cfq_group_isolation = 0;
Corrado Zoccoloe459dd02009-11-26 10:02:57 +01003765 cfqd->hw_tag = -1;
Corrado Zoccoloedc71132009-12-09 20:56:04 +01003766 /*
3767 * we optimistically start assuming sync ops weren't delayed in last
3768 * second, in order to have larger depth for async operations.
3769 */
Corrado Zoccolo573412b2009-12-06 11:48:52 +01003770 cfqd->last_delayed_sync = jiffies - HZ;
Jens Axboebb729bc2009-12-06 09:54:19 +01003771 INIT_RCU_HEAD(&cfqd->rcu);
Jens Axboebc1c1162006-06-08 08:49:06 +02003772 return cfqd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773}
3774
3775static void cfq_slab_kill(void)
3776{
Jens Axboed6de8be2008-05-28 14:46:59 +02003777 /*
3778 * Caller already ensured that pending RCU callbacks are completed,
3779 * so we should have no busy allocations at this point.
3780 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003781 if (cfq_pool)
3782 kmem_cache_destroy(cfq_pool);
3783 if (cfq_ioc_pool)
3784 kmem_cache_destroy(cfq_ioc_pool);
3785}
3786
3787static int __init cfq_slab_setup(void)
3788{
Christoph Lameter0a31bd52007-05-06 14:49:57 -07003789 cfq_pool = KMEM_CACHE(cfq_queue, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790 if (!cfq_pool)
3791 goto fail;
3792
Fabio Checconi34e6bbf2008-04-02 14:31:02 +02003793 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 if (!cfq_ioc_pool)
3795 goto fail;
3796
3797 return 0;
3798fail:
3799 cfq_slab_kill();
3800 return -ENOMEM;
3801}
3802
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803/*
3804 * sysfs parts below -->
3805 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003806static ssize_t
3807cfq_var_show(unsigned int var, char *page)
3808{
3809 return sprintf(page, "%d\n", var);
3810}
3811
3812static ssize_t
3813cfq_var_store(unsigned int *var, const char *page, size_t count)
3814{
3815 char *p = (char *) page;
3816
3817 *var = simple_strtoul(p, &p, 10);
3818 return count;
3819}
3820
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003822static ssize_t __FUNC(struct elevator_queue *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003824 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825 unsigned int __data = __VAR; \
3826 if (__CONV) \
3827 __data = jiffies_to_msecs(__data); \
3828 return cfq_var_show(__data, (page)); \
3829}
3830SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003831SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3832SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
Al Viroe572ec72006-03-18 22:27:18 -05003833SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3834SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003835SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3836SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3837SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3838SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003839SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003840SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841#undef SHOW_FUNCTION
3842
3843#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
Jens Axboeb374d182008-10-31 10:05:07 +01003844static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845{ \
Al Viro3d1ab402006-03-18 18:35:43 -05003846 struct cfq_data *cfqd = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847 unsigned int __data; \
3848 int ret = cfq_var_store(&__data, (page), count); \
3849 if (__data < (MIN)) \
3850 __data = (MIN); \
3851 else if (__data > (MAX)) \
3852 __data = (MAX); \
3853 if (__CONV) \
3854 *(__PTR) = msecs_to_jiffies(__data); \
3855 else \
3856 *(__PTR) = __data; \
3857 return ret; \
3858}
3859STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003860STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3861 UINT_MAX, 1);
3862STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3863 UINT_MAX, 1);
Al Viroe572ec72006-03-18 22:27:18 -05003864STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
Jens Axboefe094d92008-01-31 13:08:54 +01003865STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3866 UINT_MAX, 0);
Jens Axboe22e2c502005-06-27 10:55:12 +02003867STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3868STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3869STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
Jens Axboefe094d92008-01-31 13:08:54 +01003870STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3871 UINT_MAX, 0);
Jens Axboe963b72f2009-10-03 19:42:18 +02003872STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
Vivek Goyalae30c282009-12-03 12:59:55 -05003873STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003874#undef STORE_FUNCTION
3875
Al Viroe572ec72006-03-18 22:27:18 -05003876#define CFQ_ATTR(name) \
3877 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
Jens Axboe3b181522005-06-27 10:56:24 +02003878
Al Viroe572ec72006-03-18 22:27:18 -05003879static struct elv_fs_entry cfq_attrs[] = {
3880 CFQ_ATTR(quantum),
Al Viroe572ec72006-03-18 22:27:18 -05003881 CFQ_ATTR(fifo_expire_sync),
3882 CFQ_ATTR(fifo_expire_async),
3883 CFQ_ATTR(back_seek_max),
3884 CFQ_ATTR(back_seek_penalty),
3885 CFQ_ATTR(slice_sync),
3886 CFQ_ATTR(slice_async),
3887 CFQ_ATTR(slice_async_rq),
3888 CFQ_ATTR(slice_idle),
Jens Axboe963b72f2009-10-03 19:42:18 +02003889 CFQ_ATTR(low_latency),
Vivek Goyalae30c282009-12-03 12:59:55 -05003890 CFQ_ATTR(group_isolation),
Al Viroe572ec72006-03-18 22:27:18 -05003891 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892};
3893
Linus Torvalds1da177e2005-04-16 15:20:36 -07003894static struct elevator_type iosched_cfq = {
3895 .ops = {
3896 .elevator_merge_fn = cfq_merge,
3897 .elevator_merged_fn = cfq_merged_request,
3898 .elevator_merge_req_fn = cfq_merged_requests,
Jens Axboeda775262006-12-20 11:04:12 +01003899 .elevator_allow_merge_fn = cfq_allow_merge,
Jens Axboeb4878f22005-10-20 16:42:29 +02003900 .elevator_dispatch_fn = cfq_dispatch_requests,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901 .elevator_add_req_fn = cfq_insert_request,
Jens Axboeb4878f22005-10-20 16:42:29 +02003902 .elevator_activate_req_fn = cfq_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003903 .elevator_deactivate_req_fn = cfq_deactivate_request,
3904 .elevator_queue_empty_fn = cfq_queue_empty,
3905 .elevator_completed_req_fn = cfq_completed_request,
Jens Axboe21183b02006-07-13 12:33:14 +02003906 .elevator_former_req_fn = elv_rb_former_request,
3907 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908 .elevator_set_req_fn = cfq_set_request,
3909 .elevator_put_req_fn = cfq_put_request,
3910 .elevator_may_queue_fn = cfq_may_queue,
3911 .elevator_init_fn = cfq_init_queue,
3912 .elevator_exit_fn = cfq_exit_queue,
Jens Axboefc463792006-08-29 09:05:44 +02003913 .trim = cfq_free_io_context,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914 },
Al Viro3d1ab402006-03-18 18:35:43 -05003915 .elevator_attrs = cfq_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916 .elevator_name = "cfq",
3917 .elevator_owner = THIS_MODULE,
3918};
3919
Vivek Goyal3e252062009-12-04 10:36:42 -05003920#ifdef CONFIG_CFQ_GROUP_IOSCHED
3921static struct blkio_policy_type blkio_policy_cfq = {
3922 .ops = {
3923 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3924 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3925 },
3926};
3927#else
3928static struct blkio_policy_type blkio_policy_cfq;
3929#endif
3930
Linus Torvalds1da177e2005-04-16 15:20:36 -07003931static int __init cfq_init(void)
3932{
Jens Axboe22e2c502005-06-27 10:55:12 +02003933 /*
3934 * could be 0 on HZ < 1000 setups
3935 */
3936 if (!cfq_slice_async)
3937 cfq_slice_async = 1;
3938 if (!cfq_slice_idle)
3939 cfq_slice_idle = 1;
3940
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941 if (cfq_slab_setup())
3942 return -ENOMEM;
3943
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01003944 elv_register(&iosched_cfq);
Vivek Goyal3e252062009-12-04 10:36:42 -05003945 blkio_policy_register(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003946
Adrian Bunk2fdd82b2007-12-12 18:51:56 +01003947 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948}
3949
3950static void __exit cfq_exit(void)
3951{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07003952 DECLARE_COMPLETION_ONSTACK(all_gone);
Vivek Goyal3e252062009-12-04 10:36:42 -05003953 blkio_policy_unregister(&blkio_policy_cfq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954 elv_unregister(&iosched_cfq);
Al Viro334e94d2006-03-18 15:05:53 -05003955 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02003956 /* ioc_gone's update must be visible before reading ioc_count */
3957 smp_wmb();
Jens Axboed6de8be2008-05-28 14:46:59 +02003958
3959 /*
3960 * this also protects us from entering cfq_slab_kill() with
3961 * pending RCU callbacks
3962 */
Tejun Heo245b2e72009-06-24 15:13:48 +09003963 if (elv_ioc_count_read(cfq_ioc_count))
Jens Axboe9a11b4e2008-05-29 09:32:08 +02003964 wait_for_completion(&all_gone);
Christoph Hellwig83521d32005-10-30 15:01:39 -08003965 cfq_slab_kill();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966}
3967
3968module_init(cfq_init);
3969module_exit(cfq_exit);
3970
3971MODULE_AUTHOR("Jens Axboe");
3972MODULE_LICENSE("GPL");
3973MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");