Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1 | /* |
| 2 | * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR |
| 3 | * policies) |
| 4 | */ |
| 5 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 6 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) |
| 7 | { |
| 8 | return container_of(rt_se, struct task_struct, rt); |
| 9 | } |
| 10 | |
| 11 | #ifdef CONFIG_RT_GROUP_SCHED |
| 12 | |
| 13 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) |
| 14 | { |
| 15 | return rt_rq->rq; |
| 16 | } |
| 17 | |
| 18 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) |
| 19 | { |
| 20 | return rt_se->rt_rq; |
| 21 | } |
| 22 | |
| 23 | #else /* CONFIG_RT_GROUP_SCHED */ |
| 24 | |
| 25 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) |
| 26 | { |
| 27 | return container_of(rt_rq, struct rq, rt); |
| 28 | } |
| 29 | |
| 30 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) |
| 31 | { |
| 32 | struct task_struct *p = rt_task_of(rt_se); |
| 33 | struct rq *rq = task_rq(p); |
| 34 | |
| 35 | return &rq->rt; |
| 36 | } |
| 37 | |
| 38 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 39 | |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 40 | #ifdef CONFIG_SMP |
Ingo Molnar | 84de427 | 2008-01-25 21:08:15 +0100 | [diff] [blame] | 41 | |
Gregory Haskins | 637f508 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 42 | static inline int rt_overloaded(struct rq *rq) |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 43 | { |
Gregory Haskins | 637f508 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 44 | return atomic_read(&rq->rd->rto_count); |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 45 | } |
Ingo Molnar | 84de427 | 2008-01-25 21:08:15 +0100 | [diff] [blame] | 46 | |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 47 | static inline void rt_set_overload(struct rq *rq) |
| 48 | { |
Gregory Haskins | 1f11eb6 | 2008-06-04 15:04:05 -0400 | [diff] [blame] | 49 | if (!rq->online) |
| 50 | return; |
| 51 | |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 52 | cpumask_set_cpu(rq->cpu, rq->rd->rto_mask); |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 53 | /* |
| 54 | * Make sure the mask is visible before we set |
| 55 | * the overload count. That is checked to determine |
| 56 | * if we should look at the mask. It would be a shame |
| 57 | * if we looked at the mask, but the mask was not |
| 58 | * updated yet. |
| 59 | */ |
| 60 | wmb(); |
Gregory Haskins | 637f508 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 61 | atomic_inc(&rq->rd->rto_count); |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 62 | } |
Ingo Molnar | 84de427 | 2008-01-25 21:08:15 +0100 | [diff] [blame] | 63 | |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 64 | static inline void rt_clear_overload(struct rq *rq) |
| 65 | { |
Gregory Haskins | 1f11eb6 | 2008-06-04 15:04:05 -0400 | [diff] [blame] | 66 | if (!rq->online) |
| 67 | return; |
| 68 | |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 69 | /* the order here really doesn't matter */ |
Gregory Haskins | 637f508 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 70 | atomic_dec(&rq->rd->rto_count); |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 71 | cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask); |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 72 | } |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 73 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 74 | static void update_rt_migration(struct rt_rq *rt_rq) |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 75 | { |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 76 | if (rt_rq->rt_nr_migratory && (rt_rq->rt_nr_running > 1)) { |
| 77 | if (!rt_rq->overloaded) { |
| 78 | rt_set_overload(rq_of_rt_rq(rt_rq)); |
| 79 | rt_rq->overloaded = 1; |
Gregory Haskins | cdc8eb9 | 2008-01-25 21:08:23 +0100 | [diff] [blame] | 80 | } |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 81 | } else if (rt_rq->overloaded) { |
| 82 | rt_clear_overload(rq_of_rt_rq(rt_rq)); |
| 83 | rt_rq->overloaded = 0; |
Gregory Haskins | 637f508 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 84 | } |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 85 | } |
Steven Rostedt | 4fd2917 | 2008-01-25 21:08:06 +0100 | [diff] [blame] | 86 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 87 | static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
Peter Zijlstra | fa85ae2 | 2008-01-25 21:08:29 +0100 | [diff] [blame] | 88 | { |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 89 | if (rt_se->nr_cpus_allowed > 1) |
| 90 | rt_rq->rt_nr_migratory++; |
| 91 | |
| 92 | update_rt_migration(rt_rq); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 93 | } |
| 94 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 95 | static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
| 96 | { |
| 97 | if (rt_se->nr_cpus_allowed > 1) |
| 98 | rt_rq->rt_nr_migratory--; |
| 99 | |
| 100 | update_rt_migration(rt_rq); |
| 101 | } |
| 102 | |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 103 | static void enqueue_pushable_task(struct rq *rq, struct task_struct *p) |
| 104 | { |
| 105 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); |
| 106 | plist_node_init(&p->pushable_tasks, p->prio); |
| 107 | plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks); |
| 108 | } |
| 109 | |
| 110 | static void dequeue_pushable_task(struct rq *rq, struct task_struct *p) |
| 111 | { |
| 112 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); |
| 113 | } |
| 114 | |
| 115 | #else |
| 116 | |
Peter Zijlstra | ceacc2c | 2009-01-16 14:46:40 +0100 | [diff] [blame] | 117 | static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p) |
| 118 | { |
| 119 | } |
| 120 | |
| 121 | static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p) |
| 122 | { |
| 123 | } |
| 124 | |
Gregory Haskins | b07430a | 2009-01-14 08:55:39 -0500 | [diff] [blame] | 125 | static inline |
Peter Zijlstra | ceacc2c | 2009-01-16 14:46:40 +0100 | [diff] [blame] | 126 | void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
| 127 | { |
| 128 | } |
| 129 | |
Gregory Haskins | b07430a | 2009-01-14 08:55:39 -0500 | [diff] [blame] | 130 | static inline |
Peter Zijlstra | ceacc2c | 2009-01-16 14:46:40 +0100 | [diff] [blame] | 131 | void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
| 132 | { |
| 133 | } |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 134 | |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 135 | #endif /* CONFIG_SMP */ |
| 136 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 137 | static inline int on_rt_rq(struct sched_rt_entity *rt_se) |
| 138 | { |
| 139 | return !list_empty(&rt_se->run_list); |
| 140 | } |
| 141 | |
Peter Zijlstra | 052f1dc | 2008-02-13 15:45:40 +0100 | [diff] [blame] | 142 | #ifdef CONFIG_RT_GROUP_SCHED |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 143 | |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 144 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 145 | { |
| 146 | if (!rt_rq->tg) |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 147 | return RUNTIME_INF; |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 148 | |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 149 | return rt_rq->rt_runtime; |
| 150 | } |
| 151 | |
| 152 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) |
| 153 | { |
| 154 | return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 155 | } |
| 156 | |
| 157 | #define for_each_leaf_rt_rq(rt_rq, rq) \ |
Bharata B Rao | 80f40ee | 2008-12-15 11:56:48 +0530 | [diff] [blame] | 158 | list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 159 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 160 | #define for_each_sched_rt_entity(rt_se) \ |
| 161 | for (; rt_se; rt_se = rt_se->parent) |
| 162 | |
| 163 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) |
| 164 | { |
| 165 | return rt_se->my_q; |
| 166 | } |
| 167 | |
| 168 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se); |
| 169 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se); |
| 170 | |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 171 | static void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 172 | { |
Dario Faggioli | f6121f4 | 2008-10-03 17:40:46 +0200 | [diff] [blame] | 173 | struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr; |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 174 | struct sched_rt_entity *rt_se = rt_rq->rt_se; |
| 175 | |
Dario Faggioli | f6121f4 | 2008-10-03 17:40:46 +0200 | [diff] [blame] | 176 | if (rt_rq->rt_nr_running) { |
| 177 | if (rt_se && !on_rt_rq(rt_se)) |
| 178 | enqueue_rt_entity(rt_se); |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 179 | if (rt_rq->highest_prio.curr < curr->prio) |
Peter Zijlstra | 1020387 | 2008-01-25 21:08:32 +0100 | [diff] [blame] | 180 | resched_task(curr); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 181 | } |
| 182 | } |
| 183 | |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 184 | static void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 185 | { |
| 186 | struct sched_rt_entity *rt_se = rt_rq->rt_se; |
| 187 | |
| 188 | if (rt_se && on_rt_rq(rt_se)) |
| 189 | dequeue_rt_entity(rt_se); |
| 190 | } |
| 191 | |
Peter Zijlstra | 23b0fdf | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 192 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
| 193 | { |
| 194 | return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted; |
| 195 | } |
| 196 | |
| 197 | static int rt_se_boosted(struct sched_rt_entity *rt_se) |
| 198 | { |
| 199 | struct rt_rq *rt_rq = group_rt_rq(rt_se); |
| 200 | struct task_struct *p; |
| 201 | |
| 202 | if (rt_rq) |
| 203 | return !!rt_rq->rt_nr_boosted; |
| 204 | |
| 205 | p = rt_task_of(rt_se); |
| 206 | return p->prio != p->normal_prio; |
| 207 | } |
| 208 | |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 209 | #ifdef CONFIG_SMP |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 210 | static inline const struct cpumask *sched_rt_period_mask(void) |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 211 | { |
| 212 | return cpu_rq(smp_processor_id())->rd->span; |
| 213 | } |
| 214 | #else |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 215 | static inline const struct cpumask *sched_rt_period_mask(void) |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 216 | { |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 217 | return cpu_online_mask; |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 218 | } |
| 219 | #endif |
| 220 | |
| 221 | static inline |
| 222 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) |
| 223 | { |
| 224 | return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu]; |
| 225 | } |
| 226 | |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 227 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
| 228 | { |
| 229 | return &rt_rq->tg->rt_bandwidth; |
| 230 | } |
| 231 | |
Dhaval Giani | 55e12e5 | 2008-06-24 23:39:43 +0530 | [diff] [blame] | 232 | #else /* !CONFIG_RT_GROUP_SCHED */ |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 233 | |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 234 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 235 | { |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 236 | return rt_rq->rt_runtime; |
| 237 | } |
| 238 | |
| 239 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) |
| 240 | { |
| 241 | return ktime_to_ns(def_rt_bandwidth.rt_period); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 242 | } |
| 243 | |
| 244 | #define for_each_leaf_rt_rq(rt_rq, rq) \ |
| 245 | for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL) |
| 246 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 247 | #define for_each_sched_rt_entity(rt_se) \ |
| 248 | for (; rt_se; rt_se = NULL) |
| 249 | |
| 250 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) |
| 251 | { |
| 252 | return NULL; |
| 253 | } |
| 254 | |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 255 | static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 256 | { |
John Blackwood | f3ade83 | 2008-08-26 15:09:43 -0400 | [diff] [blame] | 257 | if (rt_rq->rt_nr_running) |
| 258 | resched_task(rq_of_rt_rq(rt_rq)->curr); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 259 | } |
| 260 | |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 261 | static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 262 | { |
| 263 | } |
| 264 | |
Peter Zijlstra | 23b0fdf | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 265 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
| 266 | { |
| 267 | return rt_rq->rt_throttled; |
| 268 | } |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 269 | |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 270 | static inline const struct cpumask *sched_rt_period_mask(void) |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 271 | { |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 272 | return cpu_online_mask; |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 273 | } |
| 274 | |
| 275 | static inline |
| 276 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) |
| 277 | { |
| 278 | return &cpu_rq(cpu)->rt; |
| 279 | } |
| 280 | |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 281 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
| 282 | { |
| 283 | return &def_rt_bandwidth; |
| 284 | } |
| 285 | |
Dhaval Giani | 55e12e5 | 2008-06-24 23:39:43 +0530 | [diff] [blame] | 286 | #endif /* CONFIG_RT_GROUP_SCHED */ |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 287 | |
Peter Zijlstra | b79f383 | 2008-06-19 14:22:25 +0200 | [diff] [blame] | 288 | #ifdef CONFIG_SMP |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 289 | /* |
| 290 | * We ran out of runtime, see if we can borrow some from our neighbours. |
| 291 | */ |
Peter Zijlstra | b79f383 | 2008-06-19 14:22:25 +0200 | [diff] [blame] | 292 | static int do_balance_runtime(struct rt_rq *rt_rq) |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 293 | { |
| 294 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
| 295 | struct root_domain *rd = cpu_rq(smp_processor_id())->rd; |
| 296 | int i, weight, more = 0; |
| 297 | u64 rt_period; |
| 298 | |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 299 | weight = cpumask_weight(rd->span); |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 300 | |
| 301 | spin_lock(&rt_b->rt_runtime_lock); |
| 302 | rt_period = ktime_to_ns(rt_b->rt_period); |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 303 | for_each_cpu(i, rd->span) { |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 304 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); |
| 305 | s64 diff; |
| 306 | |
| 307 | if (iter == rt_rq) |
| 308 | continue; |
| 309 | |
| 310 | spin_lock(&iter->rt_runtime_lock); |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 311 | /* |
| 312 | * Either all rqs have inf runtime and there's nothing to steal |
| 313 | * or __disable_runtime() below sets a specific rq to inf to |
| 314 | * indicate its been disabled and disalow stealing. |
| 315 | */ |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 316 | if (iter->rt_runtime == RUNTIME_INF) |
| 317 | goto next; |
| 318 | |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 319 | /* |
| 320 | * From runqueues with spare time, take 1/n part of their |
| 321 | * spare time, but no more than our period. |
| 322 | */ |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 323 | diff = iter->rt_runtime - iter->rt_time; |
| 324 | if (diff > 0) { |
Peter Zijlstra | 58838cf | 2008-07-24 12:43:13 +0200 | [diff] [blame] | 325 | diff = div_u64((u64)diff, weight); |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 326 | if (rt_rq->rt_runtime + diff > rt_period) |
| 327 | diff = rt_period - rt_rq->rt_runtime; |
| 328 | iter->rt_runtime -= diff; |
| 329 | rt_rq->rt_runtime += diff; |
| 330 | more = 1; |
| 331 | if (rt_rq->rt_runtime == rt_period) { |
| 332 | spin_unlock(&iter->rt_runtime_lock); |
| 333 | break; |
| 334 | } |
| 335 | } |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 336 | next: |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 337 | spin_unlock(&iter->rt_runtime_lock); |
| 338 | } |
| 339 | spin_unlock(&rt_b->rt_runtime_lock); |
| 340 | |
| 341 | return more; |
| 342 | } |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 343 | |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 344 | /* |
| 345 | * Ensure this RQ takes back all the runtime it lend to its neighbours. |
| 346 | */ |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 347 | static void __disable_runtime(struct rq *rq) |
| 348 | { |
| 349 | struct root_domain *rd = rq->rd; |
| 350 | struct rt_rq *rt_rq; |
| 351 | |
| 352 | if (unlikely(!scheduler_running)) |
| 353 | return; |
| 354 | |
| 355 | for_each_leaf_rt_rq(rt_rq, rq) { |
| 356 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
| 357 | s64 want; |
| 358 | int i; |
| 359 | |
| 360 | spin_lock(&rt_b->rt_runtime_lock); |
| 361 | spin_lock(&rt_rq->rt_runtime_lock); |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 362 | /* |
| 363 | * Either we're all inf and nobody needs to borrow, or we're |
| 364 | * already disabled and thus have nothing to do, or we have |
| 365 | * exactly the right amount of runtime to take out. |
| 366 | */ |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 367 | if (rt_rq->rt_runtime == RUNTIME_INF || |
| 368 | rt_rq->rt_runtime == rt_b->rt_runtime) |
| 369 | goto balanced; |
| 370 | spin_unlock(&rt_rq->rt_runtime_lock); |
| 371 | |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 372 | /* |
| 373 | * Calculate the difference between what we started out with |
| 374 | * and what we current have, that's the amount of runtime |
| 375 | * we lend and now have to reclaim. |
| 376 | */ |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 377 | want = rt_b->rt_runtime - rt_rq->rt_runtime; |
| 378 | |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 379 | /* |
| 380 | * Greedy reclaim, take back as much as we can. |
| 381 | */ |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 382 | for_each_cpu(i, rd->span) { |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 383 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); |
| 384 | s64 diff; |
| 385 | |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 386 | /* |
| 387 | * Can't reclaim from ourselves or disabled runqueues. |
| 388 | */ |
Peter Zijlstra | f1679d0 | 2008-08-14 15:49:00 +0200 | [diff] [blame] | 389 | if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF) |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 390 | continue; |
| 391 | |
| 392 | spin_lock(&iter->rt_runtime_lock); |
| 393 | if (want > 0) { |
| 394 | diff = min_t(s64, iter->rt_runtime, want); |
| 395 | iter->rt_runtime -= diff; |
| 396 | want -= diff; |
| 397 | } else { |
| 398 | iter->rt_runtime -= want; |
| 399 | want -= want; |
| 400 | } |
| 401 | spin_unlock(&iter->rt_runtime_lock); |
| 402 | |
| 403 | if (!want) |
| 404 | break; |
| 405 | } |
| 406 | |
| 407 | spin_lock(&rt_rq->rt_runtime_lock); |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 408 | /* |
| 409 | * We cannot be left wanting - that would mean some runtime |
| 410 | * leaked out of the system. |
| 411 | */ |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 412 | BUG_ON(want); |
| 413 | balanced: |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 414 | /* |
| 415 | * Disable all the borrow logic by pretending we have inf |
| 416 | * runtime - in which case borrowing doesn't make sense. |
| 417 | */ |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 418 | rt_rq->rt_runtime = RUNTIME_INF; |
| 419 | spin_unlock(&rt_rq->rt_runtime_lock); |
| 420 | spin_unlock(&rt_b->rt_runtime_lock); |
| 421 | } |
| 422 | } |
| 423 | |
| 424 | static void disable_runtime(struct rq *rq) |
| 425 | { |
| 426 | unsigned long flags; |
| 427 | |
| 428 | spin_lock_irqsave(&rq->lock, flags); |
| 429 | __disable_runtime(rq); |
| 430 | spin_unlock_irqrestore(&rq->lock, flags); |
| 431 | } |
| 432 | |
| 433 | static void __enable_runtime(struct rq *rq) |
| 434 | { |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 435 | struct rt_rq *rt_rq; |
| 436 | |
| 437 | if (unlikely(!scheduler_running)) |
| 438 | return; |
| 439 | |
Peter Zijlstra | 78333cd | 2008-09-23 15:33:43 +0200 | [diff] [blame] | 440 | /* |
| 441 | * Reset each runqueue's bandwidth settings |
| 442 | */ |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 443 | for_each_leaf_rt_rq(rt_rq, rq) { |
| 444 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
| 445 | |
| 446 | spin_lock(&rt_b->rt_runtime_lock); |
| 447 | spin_lock(&rt_rq->rt_runtime_lock); |
| 448 | rt_rq->rt_runtime = rt_b->rt_runtime; |
| 449 | rt_rq->rt_time = 0; |
Zhang, Yanmin | baf2573 | 2008-09-09 11:26:33 +0800 | [diff] [blame] | 450 | rt_rq->rt_throttled = 0; |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 451 | spin_unlock(&rt_rq->rt_runtime_lock); |
| 452 | spin_unlock(&rt_b->rt_runtime_lock); |
| 453 | } |
| 454 | } |
| 455 | |
| 456 | static void enable_runtime(struct rq *rq) |
| 457 | { |
| 458 | unsigned long flags; |
| 459 | |
| 460 | spin_lock_irqsave(&rq->lock, flags); |
| 461 | __enable_runtime(rq); |
| 462 | spin_unlock_irqrestore(&rq->lock, flags); |
| 463 | } |
| 464 | |
Peter Zijlstra | eff6549 | 2008-06-19 14:22:26 +0200 | [diff] [blame] | 465 | static int balance_runtime(struct rt_rq *rt_rq) |
| 466 | { |
| 467 | int more = 0; |
| 468 | |
| 469 | if (rt_rq->rt_time > rt_rq->rt_runtime) { |
| 470 | spin_unlock(&rt_rq->rt_runtime_lock); |
| 471 | more = do_balance_runtime(rt_rq); |
| 472 | spin_lock(&rt_rq->rt_runtime_lock); |
| 473 | } |
| 474 | |
| 475 | return more; |
| 476 | } |
Dhaval Giani | 55e12e5 | 2008-06-24 23:39:43 +0530 | [diff] [blame] | 477 | #else /* !CONFIG_SMP */ |
Peter Zijlstra | eff6549 | 2008-06-19 14:22:26 +0200 | [diff] [blame] | 478 | static inline int balance_runtime(struct rt_rq *rt_rq) |
| 479 | { |
| 480 | return 0; |
| 481 | } |
Dhaval Giani | 55e12e5 | 2008-06-24 23:39:43 +0530 | [diff] [blame] | 482 | #endif /* CONFIG_SMP */ |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 483 | |
| 484 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun) |
| 485 | { |
| 486 | int i, idle = 1; |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 487 | const struct cpumask *span; |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 488 | |
Peter Zijlstra | 0b148fa | 2008-08-19 12:33:04 +0200 | [diff] [blame] | 489 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 490 | return 1; |
| 491 | |
| 492 | span = sched_rt_period_mask(); |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 493 | for_each_cpu(i, span) { |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 494 | int enqueue = 0; |
| 495 | struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i); |
| 496 | struct rq *rq = rq_of_rt_rq(rt_rq); |
| 497 | |
| 498 | spin_lock(&rq->lock); |
| 499 | if (rt_rq->rt_time) { |
| 500 | u64 runtime; |
| 501 | |
| 502 | spin_lock(&rt_rq->rt_runtime_lock); |
Peter Zijlstra | eff6549 | 2008-06-19 14:22:26 +0200 | [diff] [blame] | 503 | if (rt_rq->rt_throttled) |
| 504 | balance_runtime(rt_rq); |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 505 | runtime = rt_rq->rt_runtime; |
| 506 | rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime); |
| 507 | if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) { |
| 508 | rt_rq->rt_throttled = 0; |
| 509 | enqueue = 1; |
| 510 | } |
| 511 | if (rt_rq->rt_time || rt_rq->rt_nr_running) |
| 512 | idle = 0; |
| 513 | spin_unlock(&rt_rq->rt_runtime_lock); |
Peter Zijlstra | 8a8cde1 | 2008-06-19 14:22:28 +0200 | [diff] [blame] | 514 | } else if (rt_rq->rt_nr_running) |
| 515 | idle = 0; |
Peter Zijlstra | d0b27fa | 2008-04-19 19:44:57 +0200 | [diff] [blame] | 516 | |
| 517 | if (enqueue) |
| 518 | sched_rt_rq_enqueue(rt_rq); |
| 519 | spin_unlock(&rq->lock); |
| 520 | } |
| 521 | |
| 522 | return idle; |
| 523 | } |
| 524 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 525 | static inline int rt_se_prio(struct sched_rt_entity *rt_se) |
| 526 | { |
Peter Zijlstra | 052f1dc | 2008-02-13 15:45:40 +0100 | [diff] [blame] | 527 | #ifdef CONFIG_RT_GROUP_SCHED |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 528 | struct rt_rq *rt_rq = group_rt_rq(rt_se); |
| 529 | |
| 530 | if (rt_rq) |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 531 | return rt_rq->highest_prio.curr; |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 532 | #endif |
| 533 | |
| 534 | return rt_task_of(rt_se)->prio; |
| 535 | } |
| 536 | |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 537 | static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 538 | { |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 539 | u64 runtime = sched_rt_runtime(rt_rq); |
Peter Zijlstra | fa85ae2 | 2008-01-25 21:08:29 +0100 | [diff] [blame] | 540 | |
Peter Zijlstra | fa85ae2 | 2008-01-25 21:08:29 +0100 | [diff] [blame] | 541 | if (rt_rq->rt_throttled) |
Peter Zijlstra | 23b0fdf | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 542 | return rt_rq_throttled(rt_rq); |
Peter Zijlstra | fa85ae2 | 2008-01-25 21:08:29 +0100 | [diff] [blame] | 543 | |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 544 | if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq)) |
| 545 | return 0; |
| 546 | |
Peter Zijlstra | b79f383 | 2008-06-19 14:22:25 +0200 | [diff] [blame] | 547 | balance_runtime(rt_rq); |
| 548 | runtime = sched_rt_runtime(rt_rq); |
| 549 | if (runtime == RUNTIME_INF) |
| 550 | return 0; |
Peter Zijlstra | ac086bc | 2008-04-19 19:44:58 +0200 | [diff] [blame] | 551 | |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 552 | if (rt_rq->rt_time > runtime) { |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 553 | rt_rq->rt_throttled = 1; |
Peter Zijlstra | 23b0fdf | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 554 | if (rt_rq_throttled(rt_rq)) { |
Peter Zijlstra | 9f0c1e5 | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 555 | sched_rt_rq_dequeue(rt_rq); |
Peter Zijlstra | 23b0fdf | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 556 | return 1; |
| 557 | } |
Peter Zijlstra | fa85ae2 | 2008-01-25 21:08:29 +0100 | [diff] [blame] | 558 | } |
| 559 | |
| 560 | return 0; |
| 561 | } |
| 562 | |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 563 | /* |
| 564 | * Update the current task's runtime statistics. Skip current tasks that |
| 565 | * are not in our scheduling class. |
| 566 | */ |
Alexey Dobriyan | a995744 | 2007-10-15 17:00:13 +0200 | [diff] [blame] | 567 | static void update_curr_rt(struct rq *rq) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 568 | { |
| 569 | struct task_struct *curr = rq->curr; |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 570 | struct sched_rt_entity *rt_se = &curr->rt; |
| 571 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 572 | u64 delta_exec; |
| 573 | |
| 574 | if (!task_has_rt_policy(curr)) |
| 575 | return; |
| 576 | |
Ingo Molnar | d281918 | 2007-08-09 11:16:47 +0200 | [diff] [blame] | 577 | delta_exec = rq->clock - curr->se.exec_start; |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 578 | if (unlikely((s64)delta_exec < 0)) |
| 579 | delta_exec = 0; |
Ingo Molnar | 6cfb0d5 | 2007-08-02 17:41:40 +0200 | [diff] [blame] | 580 | |
| 581 | schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec)); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 582 | |
| 583 | curr->se.sum_exec_runtime += delta_exec; |
Frank Mayhar | f06febc | 2008-09-12 09:54:39 -0700 | [diff] [blame] | 584 | account_group_exec_runtime(curr, delta_exec); |
| 585 | |
Ingo Molnar | d281918 | 2007-08-09 11:16:47 +0200 | [diff] [blame] | 586 | curr->se.exec_start = rq->clock; |
Srivatsa Vaddagiri | d842de8 | 2007-12-02 20:04:49 +0100 | [diff] [blame] | 587 | cpuacct_charge(curr, delta_exec); |
Peter Zijlstra | fa85ae2 | 2008-01-25 21:08:29 +0100 | [diff] [blame] | 588 | |
Peter Zijlstra | 0b148fa | 2008-08-19 12:33:04 +0200 | [diff] [blame] | 589 | if (!rt_bandwidth_enabled()) |
| 590 | return; |
| 591 | |
Dhaval Giani | 354d60c | 2008-04-19 19:44:59 +0200 | [diff] [blame] | 592 | for_each_sched_rt_entity(rt_se) { |
| 593 | rt_rq = rt_rq_of_se(rt_se); |
| 594 | |
Peter Zijlstra | cc2991c | 2008-08-19 12:33:03 +0200 | [diff] [blame] | 595 | if (sched_rt_runtime(rt_rq) != RUNTIME_INF) { |
Dimitri Sivanich | e113a74 | 2008-10-31 08:03:41 -0500 | [diff] [blame] | 596 | spin_lock(&rt_rq->rt_runtime_lock); |
Peter Zijlstra | cc2991c | 2008-08-19 12:33:03 +0200 | [diff] [blame] | 597 | rt_rq->rt_time += delta_exec; |
| 598 | if (sched_rt_runtime_exceeded(rt_rq)) |
| 599 | resched_task(curr); |
Dimitri Sivanich | e113a74 | 2008-10-31 08:03:41 -0500 | [diff] [blame] | 600 | spin_unlock(&rt_rq->rt_runtime_lock); |
Peter Zijlstra | cc2991c | 2008-08-19 12:33:03 +0200 | [diff] [blame] | 601 | } |
Dhaval Giani | 354d60c | 2008-04-19 19:44:59 +0200 | [diff] [blame] | 602 | } |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 603 | } |
| 604 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 605 | #if defined CONFIG_SMP |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 606 | |
| 607 | static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu); |
| 608 | |
| 609 | static inline int next_prio(struct rq *rq) |
Steven Rostedt | 63489e4 | 2008-01-25 21:08:03 +0100 | [diff] [blame] | 610 | { |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 611 | struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu); |
Gregory Haskins | 1f11eb6 | 2008-06-04 15:04:05 -0400 | [diff] [blame] | 612 | |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 613 | if (next && rt_prio(next->prio)) |
| 614 | return next->prio; |
| 615 | else |
| 616 | return MAX_RT_PRIO; |
| 617 | } |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 618 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 619 | static void |
| 620 | inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 621 | { |
Gregory Haskins | 4d98427 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 622 | struct rq *rq = rq_of_rt_rq(rt_rq); |
Gregory Haskins | 4d98427 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 623 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 624 | if (prio < prev_prio) { |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 625 | |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 626 | /* |
| 627 | * If the new task is higher in priority than anything on the |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 628 | * run-queue, we know that the previous high becomes our |
| 629 | * next-highest. |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 630 | */ |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 631 | rt_rq->highest_prio.next = prev_prio; |
| 632 | |
Gregory Haskins | 1f11eb6 | 2008-06-04 15:04:05 -0400 | [diff] [blame] | 633 | if (rq->online) |
Gregory Haskins | 4d98427 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 634 | cpupri_set(&rq->rd->cpupri, rq->cpu, prio); |
Ingo Molnar | 1100ac9 | 2008-06-05 12:25:37 +0200 | [diff] [blame] | 635 | |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 636 | } else if (prio == rt_rq->highest_prio.curr) |
| 637 | /* |
| 638 | * If the next task is equal in priority to the highest on |
| 639 | * the run-queue, then we implicitly know that the next highest |
| 640 | * task cannot be any lower than current |
| 641 | */ |
| 642 | rt_rq->highest_prio.next = prio; |
| 643 | else if (prio < rt_rq->highest_prio.next) |
| 644 | /* |
| 645 | * Otherwise, we need to recompute next-highest |
| 646 | */ |
| 647 | rt_rq->highest_prio.next = next_prio(rq); |
Steven Rostedt | 63489e4 | 2008-01-25 21:08:03 +0100 | [diff] [blame] | 648 | } |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 649 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 650 | static void |
| 651 | dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) |
Steven Rostedt | 63489e4 | 2008-01-25 21:08:03 +0100 | [diff] [blame] | 652 | { |
Gregory Haskins | 4d98427 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 653 | struct rq *rq = rq_of_rt_rq(rt_rq); |
Gregory Haskins | 6e0534f | 2008-05-12 21:21:01 +0200 | [diff] [blame] | 654 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 655 | if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next)) |
| 656 | rt_rq->highest_prio.next = next_prio(rq); |
| 657 | |
| 658 | if (rq->online && rt_rq->highest_prio.curr != prev_prio) |
| 659 | cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr); |
| 660 | } |
| 661 | |
| 662 | #else /* CONFIG_SMP */ |
| 663 | |
| 664 | static inline |
| 665 | void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} |
| 666 | static inline |
| 667 | void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} |
| 668 | |
| 669 | #endif /* CONFIG_SMP */ |
| 670 | |
Steven Rostedt | 63489e4 | 2008-01-25 21:08:03 +0100 | [diff] [blame] | 671 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 672 | static void |
| 673 | inc_rt_prio(struct rt_rq *rt_rq, int prio) |
| 674 | { |
| 675 | int prev_prio = rt_rq->highest_prio.curr; |
Steven Rostedt | 63489e4 | 2008-01-25 21:08:03 +0100 | [diff] [blame] | 676 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 677 | if (prio < prev_prio) |
| 678 | rt_rq->highest_prio.curr = prio; |
| 679 | |
| 680 | inc_rt_prio_smp(rt_rq, prio, prev_prio); |
| 681 | } |
| 682 | |
| 683 | static void |
| 684 | dec_rt_prio(struct rt_rq *rt_rq, int prio) |
| 685 | { |
| 686 | int prev_prio = rt_rq->highest_prio.curr; |
| 687 | |
| 688 | if (rt_rq->rt_nr_running) { |
| 689 | |
| 690 | WARN_ON(prio < prev_prio); |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 691 | |
| 692 | /* |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 693 | * This may have been our highest task, and therefore |
| 694 | * we may have some recomputation to do |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 695 | */ |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 696 | if (prio == prev_prio) { |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 697 | struct rt_prio_array *array = &rt_rq->active; |
| 698 | |
| 699 | rt_rq->highest_prio.curr = |
Steven Rostedt | 764a9d6 | 2008-01-25 21:08:04 +0100 | [diff] [blame] | 700 | sched_find_first_bit(array->bitmap); |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 701 | } |
| 702 | |
Steven Rostedt | 764a9d6 | 2008-01-25 21:08:04 +0100 | [diff] [blame] | 703 | } else |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 704 | rt_rq->highest_prio.curr = MAX_RT_PRIO; |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 705 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 706 | dec_rt_prio_smp(rt_rq, prio, prev_prio); |
| 707 | } |
Gregory Haskins | 1f11eb6 | 2008-06-04 15:04:05 -0400 | [diff] [blame] | 708 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 709 | #else |
| 710 | |
| 711 | static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {} |
| 712 | static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {} |
| 713 | |
| 714 | #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */ |
| 715 | |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 716 | #ifdef CONFIG_RT_GROUP_SCHED |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 717 | |
| 718 | static void |
| 719 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
| 720 | { |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 721 | if (rt_se_boosted(rt_se)) |
Steven Rostedt | 764a9d6 | 2008-01-25 21:08:04 +0100 | [diff] [blame] | 722 | rt_rq->rt_nr_boosted++; |
Peter Zijlstra | 052f1dc | 2008-02-13 15:45:40 +0100 | [diff] [blame] | 723 | |
Peter Zijlstra | 23b0fdf | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 724 | if (rt_rq->tg) |
| 725 | start_rt_bandwidth(&rt_rq->tg->rt_bandwidth); |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 726 | } |
| 727 | |
| 728 | static void |
| 729 | dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
| 730 | { |
Peter Zijlstra | 23b0fdf | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 731 | if (rt_se_boosted(rt_se)) |
| 732 | rt_rq->rt_nr_boosted--; |
| 733 | |
| 734 | WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted); |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 735 | } |
| 736 | |
| 737 | #else /* CONFIG_RT_GROUP_SCHED */ |
| 738 | |
| 739 | static void |
| 740 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
| 741 | { |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 742 | start_rt_bandwidth(&def_rt_bandwidth); |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 743 | } |
| 744 | |
| 745 | static inline |
| 746 | void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {} |
| 747 | |
| 748 | #endif /* CONFIG_RT_GROUP_SCHED */ |
| 749 | |
| 750 | static inline |
| 751 | void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
| 752 | { |
| 753 | int prio = rt_se_prio(rt_se); |
| 754 | |
| 755 | WARN_ON(!rt_prio(prio)); |
| 756 | rt_rq->rt_nr_running++; |
| 757 | |
| 758 | inc_rt_prio(rt_rq, prio); |
| 759 | inc_rt_migration(rt_se, rt_rq); |
| 760 | inc_rt_group(rt_se, rt_rq); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 761 | } |
| 762 | |
Peter Zijlstra | 23b0fdf | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 763 | static inline |
| 764 | void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
| 765 | { |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 766 | WARN_ON(!rt_prio(rt_se_prio(rt_se))); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 767 | WARN_ON(!rt_rq->rt_nr_running); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 768 | rt_rq->rt_nr_running--; |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 769 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 770 | dec_rt_prio(rt_rq, rt_se_prio(rt_se)); |
| 771 | dec_rt_migration(rt_se, rt_rq); |
| 772 | dec_rt_group(rt_se, rt_rq); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 773 | } |
| 774 | |
Peter Zijlstra | ad2a3f1 | 2008-06-19 09:06:57 +0200 | [diff] [blame] | 775 | static void __enqueue_rt_entity(struct sched_rt_entity *rt_se) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 776 | { |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 777 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
| 778 | struct rt_prio_array *array = &rt_rq->active; |
| 779 | struct rt_rq *group_rq = group_rt_rq(rt_se); |
Dmitry Adamushko | 20b6331 | 2008-06-11 00:58:30 +0200 | [diff] [blame] | 780 | struct list_head *queue = array->queue + rt_se_prio(rt_se); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 781 | |
Peter Zijlstra | ad2a3f1 | 2008-06-19 09:06:57 +0200 | [diff] [blame] | 782 | /* |
| 783 | * Don't enqueue the group if its throttled, or when empty. |
| 784 | * The latter is a consequence of the former when a child group |
| 785 | * get throttled and the current group doesn't have any other |
| 786 | * active members. |
| 787 | */ |
| 788 | if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 789 | return; |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 790 | |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 791 | list_add_tail(&rt_se->run_list, queue); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 792 | __set_bit(rt_se_prio(rt_se), array->bitmap); |
Peter Zijlstra | 78f2c7d | 2008-01-25 21:08:27 +0100 | [diff] [blame] | 793 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 794 | inc_rt_tasks(rt_se, rt_rq); |
| 795 | } |
| 796 | |
Peter Zijlstra | ad2a3f1 | 2008-06-19 09:06:57 +0200 | [diff] [blame] | 797 | static void __dequeue_rt_entity(struct sched_rt_entity *rt_se) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 798 | { |
| 799 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
| 800 | struct rt_prio_array *array = &rt_rq->active; |
| 801 | |
| 802 | list_del_init(&rt_se->run_list); |
| 803 | if (list_empty(array->queue + rt_se_prio(rt_se))) |
| 804 | __clear_bit(rt_se_prio(rt_se), array->bitmap); |
| 805 | |
| 806 | dec_rt_tasks(rt_se, rt_rq); |
| 807 | } |
| 808 | |
| 809 | /* |
| 810 | * Because the prio of an upper entry depends on the lower |
| 811 | * entries, we must remove entries top - down. |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 812 | */ |
Peter Zijlstra | ad2a3f1 | 2008-06-19 09:06:57 +0200 | [diff] [blame] | 813 | static void dequeue_rt_stack(struct sched_rt_entity *rt_se) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 814 | { |
Peter Zijlstra | ad2a3f1 | 2008-06-19 09:06:57 +0200 | [diff] [blame] | 815 | struct sched_rt_entity *back = NULL; |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 816 | |
Peter Zijlstra | 58d6c2d | 2008-04-19 19:45:00 +0200 | [diff] [blame] | 817 | for_each_sched_rt_entity(rt_se) { |
| 818 | rt_se->back = back; |
| 819 | back = rt_se; |
| 820 | } |
| 821 | |
| 822 | for (rt_se = back; rt_se; rt_se = rt_se->back) { |
| 823 | if (on_rt_rq(rt_se)) |
Peter Zijlstra | ad2a3f1 | 2008-06-19 09:06:57 +0200 | [diff] [blame] | 824 | __dequeue_rt_entity(rt_se); |
| 825 | } |
| 826 | } |
| 827 | |
| 828 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se) |
| 829 | { |
| 830 | dequeue_rt_stack(rt_se); |
| 831 | for_each_sched_rt_entity(rt_se) |
| 832 | __enqueue_rt_entity(rt_se); |
| 833 | } |
| 834 | |
| 835 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se) |
| 836 | { |
| 837 | dequeue_rt_stack(rt_se); |
| 838 | |
| 839 | for_each_sched_rt_entity(rt_se) { |
| 840 | struct rt_rq *rt_rq = group_rt_rq(rt_se); |
| 841 | |
| 842 | if (rt_rq && rt_rq->rt_nr_running) |
| 843 | __enqueue_rt_entity(rt_se); |
Peter Zijlstra | 58d6c2d | 2008-04-19 19:45:00 +0200 | [diff] [blame] | 844 | } |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 845 | } |
| 846 | |
| 847 | /* |
| 848 | * Adding/removing a task to/from a priority array: |
| 849 | */ |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 850 | static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup) |
| 851 | { |
| 852 | struct sched_rt_entity *rt_se = &p->rt; |
| 853 | |
| 854 | if (wakeup) |
| 855 | rt_se->timeout = 0; |
| 856 | |
Peter Zijlstra | ad2a3f1 | 2008-06-19 09:06:57 +0200 | [diff] [blame] | 857 | enqueue_rt_entity(rt_se); |
Peter Zijlstra | c09595f | 2008-06-27 13:41:14 +0200 | [diff] [blame] | 858 | |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 859 | if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1) |
| 860 | enqueue_pushable_task(rq, p); |
| 861 | |
Peter Zijlstra | c09595f | 2008-06-27 13:41:14 +0200 | [diff] [blame] | 862 | inc_cpu_load(rq, p->se.load.weight); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 863 | } |
| 864 | |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 865 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep) |
| 866 | { |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 867 | struct sched_rt_entity *rt_se = &p->rt; |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 868 | |
| 869 | update_curr_rt(rq); |
Peter Zijlstra | ad2a3f1 | 2008-06-19 09:06:57 +0200 | [diff] [blame] | 870 | dequeue_rt_entity(rt_se); |
Peter Zijlstra | c09595f | 2008-06-27 13:41:14 +0200 | [diff] [blame] | 871 | |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 872 | dequeue_pushable_task(rq, p); |
| 873 | |
Peter Zijlstra | c09595f | 2008-06-27 13:41:14 +0200 | [diff] [blame] | 874 | dec_cpu_load(rq, p->se.load.weight); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 875 | } |
| 876 | |
| 877 | /* |
| 878 | * Put task to the end of the run list without the overhead of dequeue |
| 879 | * followed by enqueue. |
| 880 | */ |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 881 | static void |
| 882 | requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 883 | { |
Ingo Molnar | 1cdad71 | 2008-06-19 09:09:15 +0200 | [diff] [blame] | 884 | if (on_rt_rq(rt_se)) { |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 885 | struct rt_prio_array *array = &rt_rq->active; |
| 886 | struct list_head *queue = array->queue + rt_se_prio(rt_se); |
| 887 | |
| 888 | if (head) |
| 889 | list_move(&rt_se->run_list, queue); |
| 890 | else |
| 891 | list_move_tail(&rt_se->run_list, queue); |
Ingo Molnar | 1cdad71 | 2008-06-19 09:09:15 +0200 | [diff] [blame] | 892 | } |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 893 | } |
| 894 | |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 895 | static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 896 | { |
| 897 | struct sched_rt_entity *rt_se = &p->rt; |
| 898 | struct rt_rq *rt_rq; |
| 899 | |
| 900 | for_each_sched_rt_entity(rt_se) { |
| 901 | rt_rq = rt_rq_of_se(rt_se); |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 902 | requeue_rt_entity(rt_rq, rt_se, head); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 903 | } |
| 904 | } |
| 905 | |
| 906 | static void yield_task_rt(struct rq *rq) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 907 | { |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 908 | requeue_task_rt(rq, rq->curr, 0); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 909 | } |
| 910 | |
Gregory Haskins | e7693a3 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 911 | #ifdef CONFIG_SMP |
Gregory Haskins | 318e089 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 912 | static int find_lowest_rq(struct task_struct *task); |
| 913 | |
Gregory Haskins | e7693a3 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 914 | static int select_task_rq_rt(struct task_struct *p, int sync) |
| 915 | { |
Gregory Haskins | 318e089 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 916 | struct rq *rq = task_rq(p); |
| 917 | |
| 918 | /* |
Steven Rostedt | e1f47d8 | 2008-01-25 21:08:12 +0100 | [diff] [blame] | 919 | * If the current task is an RT task, then |
| 920 | * try to see if we can wake this RT task up on another |
| 921 | * runqueue. Otherwise simply start this RT task |
| 922 | * on its current runqueue. |
| 923 | * |
| 924 | * We want to avoid overloading runqueues. Even if |
| 925 | * the RT task is of higher priority than the current RT task. |
| 926 | * RT tasks behave differently than other tasks. If |
| 927 | * one gets preempted, we try to push it off to another queue. |
| 928 | * So trying to keep a preempting RT task on the same |
| 929 | * cache hot CPU will force the running RT task to |
| 930 | * a cold CPU. So we waste all the cache for the lower |
| 931 | * RT task in hopes of saving some of a RT task |
| 932 | * that is just being woken and probably will have |
| 933 | * cold cache anyway. |
Gregory Haskins | 318e089 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 934 | */ |
Gregory Haskins | 17b3279 | 2008-01-25 21:08:13 +0100 | [diff] [blame] | 935 | if (unlikely(rt_task(rq->curr)) && |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 936 | (p->rt.nr_cpus_allowed > 1)) { |
Gregory Haskins | 318e089 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 937 | int cpu = find_lowest_rq(p); |
| 938 | |
| 939 | return (cpu == -1) ? task_cpu(p) : cpu; |
| 940 | } |
| 941 | |
| 942 | /* |
| 943 | * Otherwise, just let it ride on the affined RQ and the |
| 944 | * post-schedule router will push the preempted task away |
| 945 | */ |
Gregory Haskins | e7693a3 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 946 | return task_cpu(p); |
| 947 | } |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 948 | |
| 949 | static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p) |
| 950 | { |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 951 | if (rq->curr->rt.nr_cpus_allowed == 1) |
| 952 | return; |
| 953 | |
Rusty Russell | 13b8bd0 | 2009-03-25 15:01:22 +1030 | [diff] [blame] | 954 | if (p->rt.nr_cpus_allowed != 1 |
| 955 | && cpupri_find(&rq->rd->cpupri, p, NULL)) |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 956 | return; |
| 957 | |
Rusty Russell | 13b8bd0 | 2009-03-25 15:01:22 +1030 | [diff] [blame] | 958 | if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL)) |
| 959 | return; |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 960 | |
| 961 | /* |
| 962 | * There appears to be other cpus that can accept |
| 963 | * current and none to run 'p', so lets reschedule |
| 964 | * to try and push current away: |
| 965 | */ |
| 966 | requeue_task_rt(rq, p, 1); |
| 967 | resched_task(rq->curr); |
| 968 | } |
| 969 | |
Gregory Haskins | e7693a3 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 970 | #endif /* CONFIG_SMP */ |
| 971 | |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 972 | /* |
| 973 | * Preempt the current task with a newly woken task if needed: |
| 974 | */ |
Peter Zijlstra | 15afe09 | 2008-09-20 23:38:02 +0200 | [diff] [blame] | 975 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 976 | { |
Gregory Haskins | 45c01e8 | 2008-05-12 21:20:41 +0200 | [diff] [blame] | 977 | if (p->prio < rq->curr->prio) { |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 978 | resched_task(rq->curr); |
Gregory Haskins | 45c01e8 | 2008-05-12 21:20:41 +0200 | [diff] [blame] | 979 | return; |
| 980 | } |
| 981 | |
| 982 | #ifdef CONFIG_SMP |
| 983 | /* |
| 984 | * If: |
| 985 | * |
| 986 | * - the newly woken task is of equal priority to the current task |
| 987 | * - the newly woken task is non-migratable while current is migratable |
| 988 | * - current will be preempted on the next reschedule |
| 989 | * |
| 990 | * we should check to see if current can readily move to a different |
| 991 | * cpu. If so, we will reschedule to allow the push logic to try |
| 992 | * to move current somewhere else, making room for our non-migratable |
| 993 | * task. |
| 994 | */ |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 995 | if (p->prio == rq->curr->prio && !need_resched()) |
| 996 | check_preempt_equal_prio(rq, p); |
Gregory Haskins | 45c01e8 | 2008-05-12 21:20:41 +0200 | [diff] [blame] | 997 | #endif |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 998 | } |
| 999 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1000 | static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq, |
| 1001 | struct rt_rq *rt_rq) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1002 | { |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1003 | struct rt_prio_array *array = &rt_rq->active; |
| 1004 | struct sched_rt_entity *next = NULL; |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1005 | struct list_head *queue; |
| 1006 | int idx; |
| 1007 | |
| 1008 | idx = sched_find_first_bit(array->bitmap); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1009 | BUG_ON(idx >= MAX_RT_PRIO); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1010 | |
| 1011 | queue = array->queue + idx; |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1012 | next = list_entry(queue->next, struct sched_rt_entity, run_list); |
Dmitry Adamushko | 326587b | 2008-01-25 21:08:34 +0100 | [diff] [blame] | 1013 | |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1014 | return next; |
| 1015 | } |
| 1016 | |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1017 | static struct task_struct *_pick_next_task_rt(struct rq *rq) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1018 | { |
| 1019 | struct sched_rt_entity *rt_se; |
| 1020 | struct task_struct *p; |
| 1021 | struct rt_rq *rt_rq; |
| 1022 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1023 | rt_rq = &rq->rt; |
| 1024 | |
| 1025 | if (unlikely(!rt_rq->rt_nr_running)) |
| 1026 | return NULL; |
| 1027 | |
Peter Zijlstra | 23b0fdf | 2008-02-13 15:45:39 +0100 | [diff] [blame] | 1028 | if (rt_rq_throttled(rt_rq)) |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1029 | return NULL; |
| 1030 | |
| 1031 | do { |
| 1032 | rt_se = pick_next_rt_entity(rq, rt_rq); |
Dmitry Adamushko | 326587b | 2008-01-25 21:08:34 +0100 | [diff] [blame] | 1033 | BUG_ON(!rt_se); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1034 | rt_rq = group_rt_rq(rt_se); |
| 1035 | } while (rt_rq); |
| 1036 | |
| 1037 | p = rt_task_of(rt_se); |
| 1038 | p->se.exec_start = rq->clock; |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1039 | |
| 1040 | return p; |
| 1041 | } |
| 1042 | |
| 1043 | static struct task_struct *pick_next_task_rt(struct rq *rq) |
| 1044 | { |
| 1045 | struct task_struct *p = _pick_next_task_rt(rq); |
| 1046 | |
| 1047 | /* The running task is never eligible for pushing */ |
| 1048 | if (p) |
| 1049 | dequeue_pushable_task(rq, p); |
| 1050 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1051 | return p; |
| 1052 | } |
| 1053 | |
Ingo Molnar | 31ee529 | 2007-08-09 11:16:49 +0200 | [diff] [blame] | 1054 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1055 | { |
Ingo Molnar | f1e14ef | 2007-08-09 11:16:48 +0200 | [diff] [blame] | 1056 | update_curr_rt(rq); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1057 | p->se.exec_start = 0; |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1058 | |
| 1059 | /* |
| 1060 | * The previous task needs to be made eligible for pushing |
| 1061 | * if it is still active |
| 1062 | */ |
| 1063 | if (p->se.on_rq && p->rt.nr_cpus_allowed > 1) |
| 1064 | enqueue_pushable_task(rq, p); |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1065 | } |
| 1066 | |
Peter Williams | 681f3e6 | 2007-10-24 18:23:51 +0200 | [diff] [blame] | 1067 | #ifdef CONFIG_SMP |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1068 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1069 | /* Only try algorithms three times */ |
| 1070 | #define RT_MAX_TRIES 3 |
| 1071 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1072 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep); |
| 1073 | |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1074 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
| 1075 | { |
| 1076 | if (!task_running(rq, p) && |
Rusty Russell | 96f874e | 2008-11-25 02:35:14 +1030 | [diff] [blame] | 1077 | (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) && |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1078 | (p->rt.nr_cpus_allowed > 1)) |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1079 | return 1; |
| 1080 | return 0; |
| 1081 | } |
| 1082 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1083 | /* Return the second highest RT task, NULL otherwise */ |
Ingo Molnar | 79064fb | 2008-01-25 21:08:14 +0100 | [diff] [blame] | 1084 | static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu) |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1085 | { |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1086 | struct task_struct *next = NULL; |
| 1087 | struct sched_rt_entity *rt_se; |
| 1088 | struct rt_prio_array *array; |
| 1089 | struct rt_rq *rt_rq; |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1090 | int idx; |
| 1091 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1092 | for_each_leaf_rt_rq(rt_rq, rq) { |
| 1093 | array = &rt_rq->active; |
| 1094 | idx = sched_find_first_bit(array->bitmap); |
| 1095 | next_idx: |
| 1096 | if (idx >= MAX_RT_PRIO) |
| 1097 | continue; |
| 1098 | if (next && next->prio < idx) |
| 1099 | continue; |
| 1100 | list_for_each_entry(rt_se, array->queue + idx, run_list) { |
| 1101 | struct task_struct *p = rt_task_of(rt_se); |
| 1102 | if (pick_rt_task(rq, p, cpu)) { |
| 1103 | next = p; |
| 1104 | break; |
| 1105 | } |
| 1106 | } |
| 1107 | if (!next) { |
| 1108 | idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1); |
| 1109 | goto next_idx; |
| 1110 | } |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1111 | } |
| 1112 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1113 | return next; |
| 1114 | } |
| 1115 | |
Rusty Russell | 0e3900e | 2008-11-25 02:35:13 +1030 | [diff] [blame] | 1116 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1117 | |
Mike Travis | d38b223 | 2009-01-10 21:58:11 -0800 | [diff] [blame] | 1118 | static inline int pick_optimal_cpu(int this_cpu, |
| 1119 | const struct cpumask *mask) |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1120 | { |
| 1121 | int first; |
| 1122 | |
| 1123 | /* "this_cpu" is cheaper to preempt than a remote processor */ |
Mike Travis | d38b223 | 2009-01-10 21:58:11 -0800 | [diff] [blame] | 1124 | if ((this_cpu != -1) && cpumask_test_cpu(this_cpu, mask)) |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1125 | return this_cpu; |
| 1126 | |
Rusty Russell | 3d39870 | 2009-01-31 23:21:24 +1030 | [diff] [blame] | 1127 | first = cpumask_first(mask); |
| 1128 | if (first < nr_cpu_ids) |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1129 | return first; |
| 1130 | |
| 1131 | return -1; |
| 1132 | } |
| 1133 | |
| 1134 | static int find_lowest_rq(struct task_struct *task) |
| 1135 | { |
| 1136 | struct sched_domain *sd; |
Rusty Russell | 96f874e | 2008-11-25 02:35:14 +1030 | [diff] [blame] | 1137 | struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask); |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1138 | int this_cpu = smp_processor_id(); |
| 1139 | int cpu = task_cpu(task); |
Mike Travis | d38b223 | 2009-01-10 21:58:11 -0800 | [diff] [blame] | 1140 | cpumask_var_t domain_mask; |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1141 | |
Gregory Haskins | 6e0534f | 2008-05-12 21:21:01 +0200 | [diff] [blame] | 1142 | if (task->rt.nr_cpus_allowed == 1) |
| 1143 | return -1; /* No other targets possible */ |
| 1144 | |
| 1145 | if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask)) |
Gregory Haskins | 06f90db | 2008-01-25 21:08:13 +0100 | [diff] [blame] | 1146 | return -1; /* No targets found */ |
| 1147 | |
| 1148 | /* |
Max Krasnyansky | e761b77 | 2008-07-15 04:43:49 -0700 | [diff] [blame] | 1149 | * Only consider CPUs that are usable for migration. |
| 1150 | * I guess we might want to change cpupri_find() to ignore those |
| 1151 | * in the first place. |
| 1152 | */ |
Rusty Russell | 96f874e | 2008-11-25 02:35:14 +1030 | [diff] [blame] | 1153 | cpumask_and(lowest_mask, lowest_mask, cpu_active_mask); |
Max Krasnyansky | e761b77 | 2008-07-15 04:43:49 -0700 | [diff] [blame] | 1154 | |
| 1155 | /* |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1156 | * At this point we have built a mask of cpus representing the |
| 1157 | * lowest priority tasks in the system. Now we want to elect |
| 1158 | * the best one based on our affinity and topology. |
| 1159 | * |
| 1160 | * We prioritize the last cpu that the task executed on since |
| 1161 | * it is most likely cache-hot in that location. |
| 1162 | */ |
Rusty Russell | 96f874e | 2008-11-25 02:35:14 +1030 | [diff] [blame] | 1163 | if (cpumask_test_cpu(cpu, lowest_mask)) |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1164 | return cpu; |
| 1165 | |
| 1166 | /* |
| 1167 | * Otherwise, we consult the sched_domains span maps to figure |
| 1168 | * out which cpu is logically closest to our hot cache data. |
| 1169 | */ |
| 1170 | if (this_cpu == cpu) |
| 1171 | this_cpu = -1; /* Skip this_cpu opt if the same */ |
| 1172 | |
Mike Travis | d38b223 | 2009-01-10 21:58:11 -0800 | [diff] [blame] | 1173 | if (alloc_cpumask_var(&domain_mask, GFP_ATOMIC)) { |
| 1174 | for_each_domain(cpu, sd) { |
| 1175 | if (sd->flags & SD_WAKE_AFFINE) { |
| 1176 | int best_cpu; |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1177 | |
Mike Travis | d38b223 | 2009-01-10 21:58:11 -0800 | [diff] [blame] | 1178 | cpumask_and(domain_mask, |
| 1179 | sched_domain_span(sd), |
| 1180 | lowest_mask); |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1181 | |
Mike Travis | d38b223 | 2009-01-10 21:58:11 -0800 | [diff] [blame] | 1182 | best_cpu = pick_optimal_cpu(this_cpu, |
| 1183 | domain_mask); |
| 1184 | |
| 1185 | if (best_cpu != -1) { |
| 1186 | free_cpumask_var(domain_mask); |
| 1187 | return best_cpu; |
| 1188 | } |
| 1189 | } |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1190 | } |
Mike Travis | d38b223 | 2009-01-10 21:58:11 -0800 | [diff] [blame] | 1191 | free_cpumask_var(domain_mask); |
Gregory Haskins | 6e1254d | 2008-01-25 21:08:11 +0100 | [diff] [blame] | 1192 | } |
| 1193 | |
| 1194 | /* |
| 1195 | * And finally, if there were no matches within the domains |
| 1196 | * just give the caller *something* to work with from the compatible |
| 1197 | * locations. |
| 1198 | */ |
| 1199 | return pick_optimal_cpu(this_cpu, lowest_mask); |
Gregory Haskins | 07b4032 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 1200 | } |
| 1201 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1202 | /* Will lock the rq it finds */ |
Ingo Molnar | 4df64c0 | 2008-01-25 21:08:15 +0100 | [diff] [blame] | 1203 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1204 | { |
| 1205 | struct rq *lowest_rq = NULL; |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1206 | int tries; |
Ingo Molnar | 4df64c0 | 2008-01-25 21:08:15 +0100 | [diff] [blame] | 1207 | int cpu; |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1208 | |
| 1209 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
Gregory Haskins | 07b4032 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 1210 | cpu = find_lowest_rq(task); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1211 | |
Gregory Haskins | 2de0b46 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 1212 | if ((cpu == -1) || (cpu == rq->cpu)) |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1213 | break; |
| 1214 | |
Gregory Haskins | 07b4032 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 1215 | lowest_rq = cpu_rq(cpu); |
| 1216 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1217 | /* if the prio of this runqueue changed, try again */ |
Gregory Haskins | 07b4032 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 1218 | if (double_lock_balance(rq, lowest_rq)) { |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1219 | /* |
| 1220 | * We had to unlock the run queue. In |
| 1221 | * the mean time, task could have |
| 1222 | * migrated already or had its affinity changed. |
| 1223 | * Also make sure that it wasn't scheduled on its rq. |
| 1224 | */ |
Gregory Haskins | 07b4032 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 1225 | if (unlikely(task_rq(task) != rq || |
Rusty Russell | 96f874e | 2008-11-25 02:35:14 +1030 | [diff] [blame] | 1226 | !cpumask_test_cpu(lowest_rq->cpu, |
| 1227 | &task->cpus_allowed) || |
Gregory Haskins | 07b4032 | 2008-01-25 21:08:10 +0100 | [diff] [blame] | 1228 | task_running(rq, task) || |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1229 | !task->se.on_rq)) { |
Ingo Molnar | 4df64c0 | 2008-01-25 21:08:15 +0100 | [diff] [blame] | 1230 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1231 | spin_unlock(&lowest_rq->lock); |
| 1232 | lowest_rq = NULL; |
| 1233 | break; |
| 1234 | } |
| 1235 | } |
| 1236 | |
| 1237 | /* If this rq is still suitable use it. */ |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1238 | if (lowest_rq->rt.highest_prio.curr > task->prio) |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1239 | break; |
| 1240 | |
| 1241 | /* try again */ |
Peter Zijlstra | 1b12bbc | 2008-08-11 09:30:22 +0200 | [diff] [blame] | 1242 | double_unlock_balance(rq, lowest_rq); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1243 | lowest_rq = NULL; |
| 1244 | } |
| 1245 | |
| 1246 | return lowest_rq; |
| 1247 | } |
| 1248 | |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1249 | static inline int has_pushable_tasks(struct rq *rq) |
| 1250 | { |
| 1251 | return !plist_head_empty(&rq->rt.pushable_tasks); |
| 1252 | } |
| 1253 | |
| 1254 | static struct task_struct *pick_next_pushable_task(struct rq *rq) |
| 1255 | { |
| 1256 | struct task_struct *p; |
| 1257 | |
| 1258 | if (!has_pushable_tasks(rq)) |
| 1259 | return NULL; |
| 1260 | |
| 1261 | p = plist_first_entry(&rq->rt.pushable_tasks, |
| 1262 | struct task_struct, pushable_tasks); |
| 1263 | |
| 1264 | BUG_ON(rq->cpu != task_cpu(p)); |
| 1265 | BUG_ON(task_current(rq, p)); |
| 1266 | BUG_ON(p->rt.nr_cpus_allowed <= 1); |
| 1267 | |
| 1268 | BUG_ON(!p->se.on_rq); |
| 1269 | BUG_ON(!rt_task(p)); |
| 1270 | |
| 1271 | return p; |
| 1272 | } |
| 1273 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1274 | /* |
| 1275 | * If the current CPU has more than one RT task, see if the non |
| 1276 | * running task can migrate over to a CPU that is running a task |
| 1277 | * of lesser priority. |
| 1278 | */ |
Gregory Haskins | 697f0a4 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 1279 | static int push_rt_task(struct rq *rq) |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1280 | { |
| 1281 | struct task_struct *next_task; |
| 1282 | struct rq *lowest_rq; |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1283 | |
Gregory Haskins | a22d7fc | 2008-01-25 21:08:12 +0100 | [diff] [blame] | 1284 | if (!rq->rt.overloaded) |
| 1285 | return 0; |
| 1286 | |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1287 | next_task = pick_next_pushable_task(rq); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1288 | if (!next_task) |
| 1289 | return 0; |
| 1290 | |
| 1291 | retry: |
Gregory Haskins | 697f0a4 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 1292 | if (unlikely(next_task == rq->curr)) { |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1293 | WARN_ON(1); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1294 | return 0; |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1295 | } |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1296 | |
| 1297 | /* |
| 1298 | * It's possible that the next_task slipped in of |
| 1299 | * higher priority than current. If that's the case |
| 1300 | * just reschedule current. |
| 1301 | */ |
Gregory Haskins | 697f0a4 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 1302 | if (unlikely(next_task->prio < rq->curr->prio)) { |
| 1303 | resched_task(rq->curr); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1304 | return 0; |
| 1305 | } |
| 1306 | |
Gregory Haskins | 697f0a4 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 1307 | /* We might release rq lock */ |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1308 | get_task_struct(next_task); |
| 1309 | |
| 1310 | /* find_lock_lowest_rq locks the rq if found */ |
Gregory Haskins | 697f0a4 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 1311 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1312 | if (!lowest_rq) { |
| 1313 | struct task_struct *task; |
| 1314 | /* |
Gregory Haskins | 697f0a4 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 1315 | * find lock_lowest_rq releases rq->lock |
Gregory Haskins | 1563513 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1316 | * so it is possible that next_task has migrated. |
| 1317 | * |
| 1318 | * We need to make sure that the task is still on the same |
| 1319 | * run-queue and is also still the next task eligible for |
| 1320 | * pushing. |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1321 | */ |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1322 | task = pick_next_pushable_task(rq); |
Gregory Haskins | 1563513 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1323 | if (task_cpu(next_task) == rq->cpu && task == next_task) { |
| 1324 | /* |
| 1325 | * If we get here, the task hasnt moved at all, but |
| 1326 | * it has failed to push. We will not try again, |
| 1327 | * since the other cpus will pull from us when they |
| 1328 | * are ready. |
| 1329 | */ |
| 1330 | dequeue_pushable_task(rq, next_task); |
| 1331 | goto out; |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1332 | } |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1333 | |
Gregory Haskins | 1563513 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1334 | if (!task) |
| 1335 | /* No more tasks, just exit */ |
| 1336 | goto out; |
| 1337 | |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1338 | /* |
Gregory Haskins | 1563513 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1339 | * Something has shifted, try again. |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1340 | */ |
Gregory Haskins | 1563513 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1341 | put_task_struct(next_task); |
| 1342 | next_task = task; |
| 1343 | goto retry; |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1344 | } |
| 1345 | |
Gregory Haskins | 697f0a4 | 2008-01-25 21:08:09 +0100 | [diff] [blame] | 1346 | deactivate_task(rq, next_task, 0); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1347 | set_task_cpu(next_task, lowest_rq->cpu); |
| 1348 | activate_task(lowest_rq, next_task, 0); |
| 1349 | |
| 1350 | resched_task(lowest_rq->curr); |
| 1351 | |
Peter Zijlstra | 1b12bbc | 2008-08-11 09:30:22 +0200 | [diff] [blame] | 1352 | double_unlock_balance(rq, lowest_rq); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1353 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1354 | out: |
| 1355 | put_task_struct(next_task); |
| 1356 | |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1357 | return 1; |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1358 | } |
| 1359 | |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1360 | static void push_rt_tasks(struct rq *rq) |
| 1361 | { |
| 1362 | /* push_rt_task will return true if it moved an RT */ |
| 1363 | while (push_rt_task(rq)) |
| 1364 | ; |
| 1365 | } |
| 1366 | |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1367 | static int pull_rt_task(struct rq *this_rq) |
| 1368 | { |
Ingo Molnar | 80bf317 | 2008-01-25 21:08:17 +0100 | [diff] [blame] | 1369 | int this_cpu = this_rq->cpu, ret = 0, cpu; |
Gregory Haskins | a872894 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1370 | struct task_struct *p; |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1371 | struct rq *src_rq; |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1372 | |
Gregory Haskins | 637f508 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 1373 | if (likely(!rt_overloaded(this_rq))) |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1374 | return 0; |
| 1375 | |
Rusty Russell | c6c4927 | 2008-11-25 02:35:05 +1030 | [diff] [blame] | 1376 | for_each_cpu(cpu, this_rq->rd->rto_mask) { |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1377 | if (this_cpu == cpu) |
| 1378 | continue; |
| 1379 | |
| 1380 | src_rq = cpu_rq(cpu); |
Gregory Haskins | 74ab8e4 | 2008-12-29 09:39:50 -0500 | [diff] [blame] | 1381 | |
| 1382 | /* |
| 1383 | * Don't bother taking the src_rq->lock if the next highest |
| 1384 | * task is known to be lower-priority than our current task. |
| 1385 | * This may look racy, but if this value is about to go |
| 1386 | * logically higher, the src_rq will push this task away. |
| 1387 | * And if its going logically lower, we do not care |
| 1388 | */ |
| 1389 | if (src_rq->rt.highest_prio.next >= |
| 1390 | this_rq->rt.highest_prio.curr) |
| 1391 | continue; |
| 1392 | |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1393 | /* |
| 1394 | * We can potentially drop this_rq's lock in |
| 1395 | * double_lock_balance, and another CPU could |
Gregory Haskins | a872894 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1396 | * alter this_rq |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1397 | */ |
Gregory Haskins | a872894 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1398 | double_lock_balance(this_rq, src_rq); |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1399 | |
| 1400 | /* |
| 1401 | * Are there still pullable RT tasks? |
| 1402 | */ |
Mike Galbraith | 614ee1f | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1403 | if (src_rq->rt.rt_nr_running <= 1) |
| 1404 | goto skip; |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1405 | |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1406 | p = pick_next_highest_task_rt(src_rq, this_cpu); |
| 1407 | |
| 1408 | /* |
| 1409 | * Do we have an RT task that preempts |
| 1410 | * the to-be-scheduled task? |
| 1411 | */ |
Gregory Haskins | a872894 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1412 | if (p && (p->prio < this_rq->rt.highest_prio.curr)) { |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1413 | WARN_ON(p == src_rq->curr); |
| 1414 | WARN_ON(!p->se.on_rq); |
| 1415 | |
| 1416 | /* |
| 1417 | * There's a chance that p is higher in priority |
| 1418 | * than what's currently running on its cpu. |
| 1419 | * This is just that p is wakeing up and hasn't |
| 1420 | * had a chance to schedule. We only pull |
| 1421 | * p if it is lower in priority than the |
Gregory Haskins | a872894 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1422 | * current task on the run queue |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1423 | */ |
Gregory Haskins | a872894 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1424 | if (p->prio < src_rq->curr->prio) |
Mike Galbraith | 614ee1f | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1425 | goto skip; |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1426 | |
| 1427 | ret = 1; |
| 1428 | |
| 1429 | deactivate_task(src_rq, p, 0); |
| 1430 | set_task_cpu(p, this_cpu); |
| 1431 | activate_task(this_rq, p, 0); |
| 1432 | /* |
| 1433 | * We continue with the search, just in |
| 1434 | * case there's an even higher prio task |
| 1435 | * in another runqueue. (low likelyhood |
| 1436 | * but possible) |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1437 | */ |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1438 | } |
Mike Galbraith | 614ee1f | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1439 | skip: |
Peter Zijlstra | 1b12bbc | 2008-08-11 09:30:22 +0200 | [diff] [blame] | 1440 | double_unlock_balance(this_rq, src_rq); |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1441 | } |
| 1442 | |
| 1443 | return ret; |
| 1444 | } |
| 1445 | |
Steven Rostedt | 9a897c5 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1446 | static void pre_schedule_rt(struct rq *rq, struct task_struct *prev) |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1447 | { |
| 1448 | /* Try to pull RT tasks here if we lower this rq's prio */ |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1449 | if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio) |
Steven Rostedt | f65eda4 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1450 | pull_rt_task(rq); |
| 1451 | } |
| 1452 | |
Gregory Haskins | 967fc04 | 2008-12-29 09:39:52 -0500 | [diff] [blame] | 1453 | /* |
| 1454 | * assumes rq->lock is held |
| 1455 | */ |
| 1456 | static int needs_post_schedule_rt(struct rq *rq) |
| 1457 | { |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1458 | return has_pushable_tasks(rq); |
Gregory Haskins | 967fc04 | 2008-12-29 09:39:52 -0500 | [diff] [blame] | 1459 | } |
| 1460 | |
Steven Rostedt | 9a897c5 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1461 | static void post_schedule_rt(struct rq *rq) |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1462 | { |
| 1463 | /* |
Gregory Haskins | 967fc04 | 2008-12-29 09:39:52 -0500 | [diff] [blame] | 1464 | * This is only called if needs_post_schedule_rt() indicates that |
| 1465 | * we need to push tasks away |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1466 | */ |
Gregory Haskins | 967fc04 | 2008-12-29 09:39:52 -0500 | [diff] [blame] | 1467 | spin_lock_irq(&rq->lock); |
| 1468 | push_rt_tasks(rq); |
| 1469 | spin_unlock_irq(&rq->lock); |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1470 | } |
| 1471 | |
Gregory Haskins | 8ae121a | 2008-04-23 07:13:29 -0400 | [diff] [blame] | 1472 | /* |
| 1473 | * If we are not running and we are not going to reschedule soon, we should |
| 1474 | * try to push tasks away now |
| 1475 | */ |
Steven Rostedt | 9a897c5 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1476 | static void task_wake_up_rt(struct rq *rq, struct task_struct *p) |
Steven Rostedt | 4642daf | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1477 | { |
Steven Rostedt | 9a897c5 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1478 | if (!task_running(rq, p) && |
Gregory Haskins | 8ae121a | 2008-04-23 07:13:29 -0400 | [diff] [blame] | 1479 | !test_tsk_need_resched(rq->curr) && |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1480 | has_pushable_tasks(rq) && |
Gregory Haskins | 777c2f3 | 2008-12-29 09:39:50 -0500 | [diff] [blame] | 1481 | p->rt.nr_cpus_allowed > 1) |
Steven Rostedt | 4642daf | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1482 | push_rt_tasks(rq); |
| 1483 | } |
| 1484 | |
Peter Williams | 4301065 | 2007-08-09 11:16:46 +0200 | [diff] [blame] | 1485 | static unsigned long |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1486 | load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, |
Peter Williams | e1d1484 | 2007-10-24 18:23:51 +0200 | [diff] [blame] | 1487 | unsigned long max_load_move, |
| 1488 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 1489 | int *all_pinned, int *this_best_prio) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1490 | { |
Steven Rostedt | c7a1e46 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1491 | /* don't touch RT tasks */ |
| 1492 | return 0; |
Peter Williams | e1d1484 | 2007-10-24 18:23:51 +0200 | [diff] [blame] | 1493 | } |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1494 | |
Peter Williams | e1d1484 | 2007-10-24 18:23:51 +0200 | [diff] [blame] | 1495 | static int |
| 1496 | move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 1497 | struct sched_domain *sd, enum cpu_idle_type idle) |
| 1498 | { |
Steven Rostedt | c7a1e46 | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1499 | /* don't touch RT tasks */ |
| 1500 | return 0; |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1501 | } |
Ingo Molnar | deeeccd | 2008-01-25 21:08:15 +0100 | [diff] [blame] | 1502 | |
Mike Travis | cd8ba7c | 2008-03-26 14:23:49 -0700 | [diff] [blame] | 1503 | static void set_cpus_allowed_rt(struct task_struct *p, |
Rusty Russell | 96f874e | 2008-11-25 02:35:14 +1030 | [diff] [blame] | 1504 | const struct cpumask *new_mask) |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1505 | { |
Rusty Russell | 96f874e | 2008-11-25 02:35:14 +1030 | [diff] [blame] | 1506 | int weight = cpumask_weight(new_mask); |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1507 | |
| 1508 | BUG_ON(!rt_task(p)); |
| 1509 | |
| 1510 | /* |
| 1511 | * Update the migration status of the RQ if we have an RT task |
| 1512 | * which is running AND changing its weight value. |
| 1513 | */ |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1514 | if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) { |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1515 | struct rq *rq = task_rq(p); |
| 1516 | |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1517 | if (!task_current(rq, p)) { |
| 1518 | /* |
| 1519 | * Make sure we dequeue this task from the pushable list |
| 1520 | * before going further. It will either remain off of |
| 1521 | * the list because we are no longer pushable, or it |
| 1522 | * will be requeued. |
| 1523 | */ |
| 1524 | if (p->rt.nr_cpus_allowed > 1) |
| 1525 | dequeue_pushable_task(rq, p); |
| 1526 | |
| 1527 | /* |
| 1528 | * Requeue if our weight is changing and still > 1 |
| 1529 | */ |
| 1530 | if (weight > 1) |
| 1531 | enqueue_pushable_task(rq, p); |
| 1532 | |
| 1533 | } |
| 1534 | |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1535 | if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) { |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1536 | rq->rt.rt_nr_migratory++; |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1537 | } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) { |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1538 | BUG_ON(!rq->rt.rt_nr_migratory); |
| 1539 | rq->rt.rt_nr_migratory--; |
| 1540 | } |
| 1541 | |
Gregory Haskins | 398a153 | 2009-01-14 09:10:04 -0500 | [diff] [blame] | 1542 | update_rt_migration(&rq->rt); |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1543 | } |
| 1544 | |
Rusty Russell | 96f874e | 2008-11-25 02:35:14 +1030 | [diff] [blame] | 1545 | cpumask_copy(&p->cpus_allowed, new_mask); |
Peter Zijlstra | 6f505b1 | 2008-01-25 21:08:30 +0100 | [diff] [blame] | 1546 | p->rt.nr_cpus_allowed = weight; |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1547 | } |
Ingo Molnar | deeeccd | 2008-01-25 21:08:15 +0100 | [diff] [blame] | 1548 | |
Ingo Molnar | bdd7c81 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 1549 | /* Assumes rq->lock is held */ |
Gregory Haskins | 1f11eb6 | 2008-06-04 15:04:05 -0400 | [diff] [blame] | 1550 | static void rq_online_rt(struct rq *rq) |
Ingo Molnar | bdd7c81 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 1551 | { |
| 1552 | if (rq->rt.overloaded) |
| 1553 | rt_set_overload(rq); |
Gregory Haskins | 6e0534f | 2008-05-12 21:21:01 +0200 | [diff] [blame] | 1554 | |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 1555 | __enable_runtime(rq); |
| 1556 | |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1557 | cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr); |
Ingo Molnar | bdd7c81 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 1558 | } |
| 1559 | |
| 1560 | /* Assumes rq->lock is held */ |
Gregory Haskins | 1f11eb6 | 2008-06-04 15:04:05 -0400 | [diff] [blame] | 1561 | static void rq_offline_rt(struct rq *rq) |
Ingo Molnar | bdd7c81 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 1562 | { |
| 1563 | if (rq->rt.overloaded) |
| 1564 | rt_clear_overload(rq); |
Gregory Haskins | 6e0534f | 2008-05-12 21:21:01 +0200 | [diff] [blame] | 1565 | |
Peter Zijlstra | 7def2be | 2008-06-05 14:49:58 +0200 | [diff] [blame] | 1566 | __disable_runtime(rq); |
| 1567 | |
Gregory Haskins | 6e0534f | 2008-05-12 21:21:01 +0200 | [diff] [blame] | 1568 | cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID); |
Ingo Molnar | bdd7c81 | 2008-01-25 21:08:18 +0100 | [diff] [blame] | 1569 | } |
Steven Rostedt | cb46984 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1570 | |
| 1571 | /* |
| 1572 | * When switch from the rt queue, we bring ourselves to a position |
| 1573 | * that we might want to pull RT tasks from other runqueues. |
| 1574 | */ |
| 1575 | static void switched_from_rt(struct rq *rq, struct task_struct *p, |
| 1576 | int running) |
| 1577 | { |
| 1578 | /* |
| 1579 | * If there are other RT tasks then we will reschedule |
| 1580 | * and the scheduling of the other RT tasks will handle |
| 1581 | * the balancing. But if we are the last RT task |
| 1582 | * we may need to handle the pulling of RT tasks |
| 1583 | * now. |
| 1584 | */ |
| 1585 | if (!rq->rt.rt_nr_running) |
| 1586 | pull_rt_task(rq); |
| 1587 | } |
Rusty Russell | 3d8cbdf | 2008-11-25 09:58:41 +1030 | [diff] [blame] | 1588 | |
| 1589 | static inline void init_sched_rt_class(void) |
| 1590 | { |
| 1591 | unsigned int i; |
| 1592 | |
| 1593 | for_each_possible_cpu(i) |
Mike Travis | 6ca09df | 2008-12-31 18:08:45 -0800 | [diff] [blame] | 1594 | alloc_cpumask_var_node(&per_cpu(local_cpu_mask, i), |
| 1595 | GFP_KERNEL, cpu_to_node(i)); |
Rusty Russell | 3d8cbdf | 2008-11-25 09:58:41 +1030 | [diff] [blame] | 1596 | } |
Steven Rostedt | e8fa136 | 2008-01-25 21:08:05 +0100 | [diff] [blame] | 1597 | #endif /* CONFIG_SMP */ |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1598 | |
Steven Rostedt | cb46984 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1599 | /* |
| 1600 | * When switching a task to RT, we may overload the runqueue |
| 1601 | * with RT tasks. In this case we try to push them off to |
| 1602 | * other runqueues. |
| 1603 | */ |
| 1604 | static void switched_to_rt(struct rq *rq, struct task_struct *p, |
| 1605 | int running) |
| 1606 | { |
| 1607 | int check_resched = 1; |
| 1608 | |
| 1609 | /* |
| 1610 | * If we are already running, then there's nothing |
| 1611 | * that needs to be done. But if we are not running |
| 1612 | * we may need to preempt the current running task. |
| 1613 | * If that current running task is also an RT task |
| 1614 | * then see if we can move to another run queue. |
| 1615 | */ |
| 1616 | if (!running) { |
| 1617 | #ifdef CONFIG_SMP |
| 1618 | if (rq->rt.overloaded && push_rt_task(rq) && |
| 1619 | /* Don't resched if we changed runqueues */ |
| 1620 | rq != task_rq(p)) |
| 1621 | check_resched = 0; |
| 1622 | #endif /* CONFIG_SMP */ |
| 1623 | if (check_resched && p->prio < rq->curr->prio) |
| 1624 | resched_task(rq->curr); |
| 1625 | } |
| 1626 | } |
| 1627 | |
| 1628 | /* |
| 1629 | * Priority of the task has changed. This may cause |
| 1630 | * us to initiate a push or pull. |
| 1631 | */ |
| 1632 | static void prio_changed_rt(struct rq *rq, struct task_struct *p, |
| 1633 | int oldprio, int running) |
| 1634 | { |
| 1635 | if (running) { |
| 1636 | #ifdef CONFIG_SMP |
| 1637 | /* |
| 1638 | * If our priority decreases while running, we |
| 1639 | * may need to pull tasks to this runqueue. |
| 1640 | */ |
| 1641 | if (oldprio < p->prio) |
| 1642 | pull_rt_task(rq); |
| 1643 | /* |
| 1644 | * If there's a higher priority task waiting to run |
Steven Rostedt | 6fa46fa | 2008-03-05 10:00:12 -0500 | [diff] [blame] | 1645 | * then reschedule. Note, the above pull_rt_task |
| 1646 | * can release the rq lock and p could migrate. |
| 1647 | * Only reschedule if p is still on the same runqueue. |
Steven Rostedt | cb46984 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1648 | */ |
Gregory Haskins | e864c49 | 2008-12-29 09:39:49 -0500 | [diff] [blame] | 1649 | if (p->prio > rq->rt.highest_prio.curr && rq->curr == p) |
Steven Rostedt | cb46984 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1650 | resched_task(p); |
| 1651 | #else |
| 1652 | /* For UP simply resched on drop of prio */ |
| 1653 | if (oldprio < p->prio) |
| 1654 | resched_task(p); |
| 1655 | #endif /* CONFIG_SMP */ |
| 1656 | } else { |
| 1657 | /* |
| 1658 | * This task is not running, but if it is |
| 1659 | * greater than the current running task |
| 1660 | * then reschedule. |
| 1661 | */ |
| 1662 | if (p->prio < rq->curr->prio) |
| 1663 | resched_task(rq->curr); |
| 1664 | } |
| 1665 | } |
| 1666 | |
Peter Zijlstra | 78f2c7d | 2008-01-25 21:08:27 +0100 | [diff] [blame] | 1667 | static void watchdog(struct rq *rq, struct task_struct *p) |
| 1668 | { |
| 1669 | unsigned long soft, hard; |
| 1670 | |
| 1671 | if (!p->signal) |
| 1672 | return; |
| 1673 | |
| 1674 | soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur; |
| 1675 | hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max; |
| 1676 | |
| 1677 | if (soft != RLIM_INFINITY) { |
| 1678 | unsigned long next; |
| 1679 | |
| 1680 | p->rt.timeout++; |
| 1681 | next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ); |
Peter Zijlstra | 5a52dd5 | 2008-01-25 21:08:32 +0100 | [diff] [blame] | 1682 | if (p->rt.timeout > next) |
Frank Mayhar | f06febc | 2008-09-12 09:54:39 -0700 | [diff] [blame] | 1683 | p->cputime_expires.sched_exp = p->se.sum_exec_runtime; |
Peter Zijlstra | 78f2c7d | 2008-01-25 21:08:27 +0100 | [diff] [blame] | 1684 | } |
| 1685 | } |
Steven Rostedt | cb46984 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1686 | |
Peter Zijlstra | 8f4d37e | 2008-01-25 21:08:29 +0100 | [diff] [blame] | 1687 | static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1688 | { |
Peter Zijlstra | 67e2be0 | 2007-12-20 15:01:17 +0100 | [diff] [blame] | 1689 | update_curr_rt(rq); |
| 1690 | |
Peter Zijlstra | 78f2c7d | 2008-01-25 21:08:27 +0100 | [diff] [blame] | 1691 | watchdog(rq, p); |
| 1692 | |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1693 | /* |
| 1694 | * RR tasks need a special form of timeslice management. |
| 1695 | * FIFO tasks have no timeslices. |
| 1696 | */ |
| 1697 | if (p->policy != SCHED_RR) |
| 1698 | return; |
| 1699 | |
Peter Zijlstra | fa71706 | 2008-01-25 21:08:27 +0100 | [diff] [blame] | 1700 | if (--p->rt.time_slice) |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1701 | return; |
| 1702 | |
Peter Zijlstra | fa71706 | 2008-01-25 21:08:27 +0100 | [diff] [blame] | 1703 | p->rt.time_slice = DEF_TIMESLICE; |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1704 | |
Dmitry Adamushko | 98fbc79 | 2007-08-24 20:39:10 +0200 | [diff] [blame] | 1705 | /* |
| 1706 | * Requeue to the end of queue if we are not the only element |
| 1707 | * on the queue: |
| 1708 | */ |
Peter Zijlstra | fa71706 | 2008-01-25 21:08:27 +0100 | [diff] [blame] | 1709 | if (p->rt.run_list.prev != p->rt.run_list.next) { |
Dmitry Adamushko | 7ebefa8 | 2008-07-01 23:32:15 +0200 | [diff] [blame] | 1710 | requeue_task_rt(rq, p, 0); |
Dmitry Adamushko | 98fbc79 | 2007-08-24 20:39:10 +0200 | [diff] [blame] | 1711 | set_tsk_need_resched(p); |
| 1712 | } |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1713 | } |
| 1714 | |
Srivatsa Vaddagiri | 83b699e | 2007-10-15 17:00:08 +0200 | [diff] [blame] | 1715 | static void set_curr_task_rt(struct rq *rq) |
| 1716 | { |
| 1717 | struct task_struct *p = rq->curr; |
| 1718 | |
| 1719 | p->se.exec_start = rq->clock; |
Gregory Haskins | 917b627 | 2008-12-29 09:39:53 -0500 | [diff] [blame] | 1720 | |
| 1721 | /* The running task is never eligible for pushing */ |
| 1722 | dequeue_pushable_task(rq, p); |
Srivatsa Vaddagiri | 83b699e | 2007-10-15 17:00:08 +0200 | [diff] [blame] | 1723 | } |
| 1724 | |
Harvey Harrison | 2abdad0 | 2008-04-25 10:53:13 -0700 | [diff] [blame] | 1725 | static const struct sched_class rt_sched_class = { |
Ingo Molnar | 5522d5d | 2007-10-15 17:00:12 +0200 | [diff] [blame] | 1726 | .next = &fair_sched_class, |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1727 | .enqueue_task = enqueue_task_rt, |
| 1728 | .dequeue_task = dequeue_task_rt, |
| 1729 | .yield_task = yield_task_rt, |
| 1730 | |
| 1731 | .check_preempt_curr = check_preempt_curr_rt, |
| 1732 | |
| 1733 | .pick_next_task = pick_next_task_rt, |
| 1734 | .put_prev_task = put_prev_task_rt, |
| 1735 | |
Peter Williams | 681f3e6 | 2007-10-24 18:23:51 +0200 | [diff] [blame] | 1736 | #ifdef CONFIG_SMP |
Li Zefan | 4ce72a2 | 2008-10-22 15:25:26 +0800 | [diff] [blame] | 1737 | .select_task_rq = select_task_rq_rt, |
| 1738 | |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1739 | .load_balance = load_balance_rt, |
Peter Williams | e1d1484 | 2007-10-24 18:23:51 +0200 | [diff] [blame] | 1740 | .move_one_task = move_one_task_rt, |
Gregory Haskins | 73fe6aa | 2008-01-25 21:08:07 +0100 | [diff] [blame] | 1741 | .set_cpus_allowed = set_cpus_allowed_rt, |
Gregory Haskins | 1f11eb6 | 2008-06-04 15:04:05 -0400 | [diff] [blame] | 1742 | .rq_online = rq_online_rt, |
| 1743 | .rq_offline = rq_offline_rt, |
Steven Rostedt | 9a897c5 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1744 | .pre_schedule = pre_schedule_rt, |
Gregory Haskins | 967fc04 | 2008-12-29 09:39:52 -0500 | [diff] [blame] | 1745 | .needs_post_schedule = needs_post_schedule_rt, |
Steven Rostedt | 9a897c5 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1746 | .post_schedule = post_schedule_rt, |
| 1747 | .task_wake_up = task_wake_up_rt, |
Steven Rostedt | cb46984 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1748 | .switched_from = switched_from_rt, |
Peter Williams | 681f3e6 | 2007-10-24 18:23:51 +0200 | [diff] [blame] | 1749 | #endif |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1750 | |
Srivatsa Vaddagiri | 83b699e | 2007-10-15 17:00:08 +0200 | [diff] [blame] | 1751 | .set_curr_task = set_curr_task_rt, |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1752 | .task_tick = task_tick_rt, |
Steven Rostedt | cb46984 | 2008-01-25 21:08:22 +0100 | [diff] [blame] | 1753 | |
| 1754 | .prio_changed = prio_changed_rt, |
| 1755 | .switched_to = switched_to_rt, |
Ingo Molnar | bb44e5d | 2007-07-09 18:51:58 +0200 | [diff] [blame] | 1756 | }; |
Peter Zijlstra | ada18de | 2008-06-19 14:22:24 +0200 | [diff] [blame] | 1757 | |
| 1758 | #ifdef CONFIG_SCHED_DEBUG |
| 1759 | extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq); |
| 1760 | |
| 1761 | static void print_rt_stats(struct seq_file *m, int cpu) |
| 1762 | { |
| 1763 | struct rt_rq *rt_rq; |
| 1764 | |
| 1765 | rcu_read_lock(); |
| 1766 | for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu)) |
| 1767 | print_rt_rq(m, cpu, rt_rq); |
| 1768 | rcu_read_unlock(); |
| 1769 | } |
Dhaval Giani | 55e12e5 | 2008-06-24 23:39:43 +0530 | [diff] [blame] | 1770 | #endif /* CONFIG_SCHED_DEBUG */ |
Rusty Russell | 0e3900e | 2008-11-25 02:35:13 +1030 | [diff] [blame] | 1771 | |