1 // SPDX-License-Identifier: GPL-2.0 1 // SPDX-License-Identifier: GPL-2.0
2 /* 2 /*
3 * Deadline Scheduling Class (SCHED_DEADLINE) 3 * Deadline Scheduling Class (SCHED_DEADLINE)
4 * 4 *
5 * Earliest Deadline First (EDF) + Constant Ba 5 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
6 * 6 *
7 * Tasks that periodically executes their inst 7 * Tasks that periodically executes their instances for less than their
8 * runtime won't miss any of their deadlines. 8 * runtime won't miss any of their deadlines.
9 * Tasks that are not periodic or sporadic or 9 * Tasks that are not periodic or sporadic or that tries to execute more
10 * than their reserved bandwidth will be slowe 10 * than their reserved bandwidth will be slowed down (and may potentially
11 * miss some of their deadlines), and won't af 11 * miss some of their deadlines), and won't affect any other task.
12 * 12 *
13 * Copyright (C) 2012 Dario Faggioli <raistlin 13 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
14 * Juri Lelli <juri.lelli@g 14 * Juri Lelli <juri.lelli@gmail.com>,
15 * Michael Trimarchi <micha 15 * Michael Trimarchi <michael@amarulasolutions.com>,
16 * Fabio Checconi <fcheccon 16 * Fabio Checconi <fchecconi@gmail.com>
17 */ 17 */
18 18
19 #include <linux/cpuset.h> 19 #include <linux/cpuset.h>
20 20
21 /* 21 /*
22 * Default limits for DL period; on the top en 22 * Default limits for DL period; on the top end we guard against small util
23 * tasks still getting ridiculously long effec 23 * tasks still getting ridiculously long effective runtimes, on the bottom end we
24 * guard against timer DoS. 24 * guard against timer DoS.
25 */ 25 */
26 static unsigned int sysctl_sched_dl_period_max 26 static unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
27 static unsigned int sysctl_sched_dl_period_min 27 static unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
28 #ifdef CONFIG_SYSCTL 28 #ifdef CONFIG_SYSCTL
29 static struct ctl_table sched_dl_sysctls[] = { 29 static struct ctl_table sched_dl_sysctls[] = {
30 { 30 {
31 .procname = "sched_deadl 31 .procname = "sched_deadline_period_max_us",
32 .data = &sysctl_sche 32 .data = &sysctl_sched_dl_period_max,
33 .maxlen = sizeof(unsig 33 .maxlen = sizeof(unsigned int),
34 .mode = 0644, 34 .mode = 0644,
35 .proc_handler = proc_douintv 35 .proc_handler = proc_douintvec_minmax,
36 .extra1 = (void *)&sys 36 .extra1 = (void *)&sysctl_sched_dl_period_min,
37 }, 37 },
38 { 38 {
39 .procname = "sched_deadl 39 .procname = "sched_deadline_period_min_us",
40 .data = &sysctl_sche 40 .data = &sysctl_sched_dl_period_min,
41 .maxlen = sizeof(unsig 41 .maxlen = sizeof(unsigned int),
42 .mode = 0644, 42 .mode = 0644,
43 .proc_handler = proc_douintv 43 .proc_handler = proc_douintvec_minmax,
44 .extra2 = (void *)&sys 44 .extra2 = (void *)&sysctl_sched_dl_period_max,
45 }, 45 },
>> 46 {}
46 }; 47 };
47 48
48 static int __init sched_dl_sysctl_init(void) 49 static int __init sched_dl_sysctl_init(void)
49 { 50 {
50 register_sysctl_init("kernel", sched_d 51 register_sysctl_init("kernel", sched_dl_sysctls);
51 return 0; 52 return 0;
52 } 53 }
53 late_initcall(sched_dl_sysctl_init); 54 late_initcall(sched_dl_sysctl_init);
54 #endif 55 #endif
55 56
56 static bool dl_server(struct sched_dl_entity * <<
57 { <<
58 return dl_se->dl_server; <<
59 } <<
60 <<
61 static inline struct task_struct *dl_task_of(s 57 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
62 { 58 {
63 BUG_ON(dl_server(dl_se)); <<
64 return container_of(dl_se, struct task 59 return container_of(dl_se, struct task_struct, dl);
65 } 60 }
66 61
67 static inline struct rq *rq_of_dl_rq(struct dl 62 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
68 { 63 {
69 return container_of(dl_rq, struct rq, 64 return container_of(dl_rq, struct rq, dl);
70 } 65 }
71 66
72 static inline struct rq *rq_of_dl_se(struct sc <<
73 { <<
74 struct rq *rq = dl_se->rq; <<
75 <<
76 if (!dl_server(dl_se)) <<
77 rq = task_rq(dl_task_of(dl_se) <<
78 <<
79 return rq; <<
80 } <<
81 <<
82 static inline struct dl_rq *dl_rq_of_se(struct 67 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
83 { 68 {
84 return &rq_of_dl_se(dl_se)->dl; !! 69 struct task_struct *p = dl_task_of(dl_se);
>> 70 struct rq *rq = task_rq(p);
>> 71
>> 72 return &rq->dl;
85 } 73 }
86 74
87 static inline int on_dl_rq(struct sched_dl_ent 75 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
88 { 76 {
89 return !RB_EMPTY_NODE(&dl_se->rb_node) 77 return !RB_EMPTY_NODE(&dl_se->rb_node);
90 } 78 }
91 79
92 #ifdef CONFIG_RT_MUTEXES 80 #ifdef CONFIG_RT_MUTEXES
93 static inline struct sched_dl_entity *pi_of(st 81 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
94 { 82 {
95 return dl_se->pi_se; 83 return dl_se->pi_se;
96 } 84 }
97 85
98 static inline bool is_dl_boosted(struct sched_ 86 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
99 { 87 {
100 return pi_of(dl_se) != dl_se; 88 return pi_of(dl_se) != dl_se;
101 } 89 }
102 #else 90 #else
103 static inline struct sched_dl_entity *pi_of(st 91 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
104 { 92 {
105 return dl_se; 93 return dl_se;
106 } 94 }
107 95
108 static inline bool is_dl_boosted(struct sched_ 96 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
109 { 97 {
110 return false; 98 return false;
111 } 99 }
112 #endif 100 #endif
113 101
114 #ifdef CONFIG_SMP 102 #ifdef CONFIG_SMP
115 static inline struct dl_bw *dl_bw_of(int i) 103 static inline struct dl_bw *dl_bw_of(int i)
116 { 104 {
117 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_ 105 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
118 "sched RCU must be he 106 "sched RCU must be held");
119 return &cpu_rq(i)->rd->dl_bw; 107 return &cpu_rq(i)->rd->dl_bw;
120 } 108 }
121 109
122 static inline int dl_bw_cpus(int i) 110 static inline int dl_bw_cpus(int i)
123 { 111 {
124 struct root_domain *rd = cpu_rq(i)->rd 112 struct root_domain *rd = cpu_rq(i)->rd;
125 int cpus; 113 int cpus;
126 114
127 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_ 115 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
128 "sched RCU must be he 116 "sched RCU must be held");
129 117
130 if (cpumask_subset(rd->span, cpu_activ 118 if (cpumask_subset(rd->span, cpu_active_mask))
131 return cpumask_weight(rd->span 119 return cpumask_weight(rd->span);
132 120
133 cpus = 0; 121 cpus = 0;
134 122
135 for_each_cpu_and(i, rd->span, cpu_acti 123 for_each_cpu_and(i, rd->span, cpu_active_mask)
136 cpus++; 124 cpus++;
137 125
138 return cpus; 126 return cpus;
139 } 127 }
140 128
141 static inline unsigned long __dl_bw_capacity(c 129 static inline unsigned long __dl_bw_capacity(const struct cpumask *mask)
142 { 130 {
143 unsigned long cap = 0; 131 unsigned long cap = 0;
144 int i; 132 int i;
145 133
146 for_each_cpu_and(i, mask, cpu_active_m 134 for_each_cpu_and(i, mask, cpu_active_mask)
147 cap += arch_scale_cpu_capacity !! 135 cap += capacity_orig_of(i);
148 136
149 return cap; 137 return cap;
150 } 138 }
151 139
152 /* 140 /*
153 * XXX Fix: If 'rq->rd == def_root_domain' per 141 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
154 * of the CPU the task is running on rather rd 142 * of the CPU the task is running on rather rd's \Sum CPU capacity.
155 */ 143 */
156 static inline unsigned long dl_bw_capacity(int 144 static inline unsigned long dl_bw_capacity(int i)
157 { 145 {
158 if (!sched_asym_cpucap_active() && 146 if (!sched_asym_cpucap_active() &&
159 arch_scale_cpu_capacity(i) == SCHE !! 147 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
160 return dl_bw_cpus(i) << SCHED_ 148 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
161 } else { 149 } else {
162 RCU_LOCKDEP_WARN(!rcu_read_loc 150 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
163 "sched RCU mu 151 "sched RCU must be held");
164 152
165 return __dl_bw_capacity(cpu_rq 153 return __dl_bw_capacity(cpu_rq(i)->rd->span);
166 } 154 }
167 } 155 }
168 156
169 static inline bool dl_bw_visited(int cpu, u64 157 static inline bool dl_bw_visited(int cpu, u64 gen)
170 { 158 {
171 struct root_domain *rd = cpu_rq(cpu)-> 159 struct root_domain *rd = cpu_rq(cpu)->rd;
172 160
173 if (rd->visit_gen == gen) 161 if (rd->visit_gen == gen)
174 return true; 162 return true;
175 163
176 rd->visit_gen = gen; 164 rd->visit_gen = gen;
177 return false; 165 return false;
178 } 166 }
179 167
180 static inline 168 static inline
181 void __dl_update(struct dl_bw *dl_b, s64 bw) 169 void __dl_update(struct dl_bw *dl_b, s64 bw)
182 { 170 {
183 struct root_domain *rd = container_of( 171 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
184 int i; 172 int i;
185 173
186 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_ 174 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
187 "sched RCU must be he 175 "sched RCU must be held");
188 for_each_cpu_and(i, rd->span, cpu_acti 176 for_each_cpu_and(i, rd->span, cpu_active_mask) {
189 struct rq *rq = cpu_rq(i); 177 struct rq *rq = cpu_rq(i);
190 178
191 rq->dl.extra_bw += bw; 179 rq->dl.extra_bw += bw;
192 } 180 }
193 } 181 }
194 #else 182 #else
195 static inline struct dl_bw *dl_bw_of(int i) 183 static inline struct dl_bw *dl_bw_of(int i)
196 { 184 {
197 return &cpu_rq(i)->dl.dl_bw; 185 return &cpu_rq(i)->dl.dl_bw;
198 } 186 }
199 187
200 static inline int dl_bw_cpus(int i) 188 static inline int dl_bw_cpus(int i)
201 { 189 {
202 return 1; 190 return 1;
203 } 191 }
204 192
205 static inline unsigned long dl_bw_capacity(int 193 static inline unsigned long dl_bw_capacity(int i)
206 { 194 {
207 return SCHED_CAPACITY_SCALE; 195 return SCHED_CAPACITY_SCALE;
208 } 196 }
209 197
210 static inline bool dl_bw_visited(int cpu, u64 198 static inline bool dl_bw_visited(int cpu, u64 gen)
211 { 199 {
212 return false; 200 return false;
213 } 201 }
214 202
215 static inline 203 static inline
216 void __dl_update(struct dl_bw *dl_b, s64 bw) 204 void __dl_update(struct dl_bw *dl_b, s64 bw)
217 { 205 {
218 struct dl_rq *dl = container_of(dl_b, 206 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
219 207
220 dl->extra_bw += bw; 208 dl->extra_bw += bw;
221 } 209 }
222 #endif 210 #endif
223 211
224 static inline 212 static inline
225 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, 213 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
226 { 214 {
227 dl_b->total_bw -= tsk_bw; 215 dl_b->total_bw -= tsk_bw;
228 __dl_update(dl_b, (s32)tsk_bw / cpus); 216 __dl_update(dl_b, (s32)tsk_bw / cpus);
229 } 217 }
230 218
231 static inline 219 static inline
232 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, 220 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
233 { 221 {
234 dl_b->total_bw += tsk_bw; 222 dl_b->total_bw += tsk_bw;
235 __dl_update(dl_b, -((s32)tsk_bw / cpus 223 __dl_update(dl_b, -((s32)tsk_bw / cpus));
236 } 224 }
237 225
238 static inline bool 226 static inline bool
239 __dl_overflow(struct dl_bw *dl_b, unsigned lon 227 __dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
240 { 228 {
241 return dl_b->bw != -1 && 229 return dl_b->bw != -1 &&
242 cap_scale(dl_b->bw, cap) < dl_b 230 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
243 } 231 }
244 232
245 static inline 233 static inline
246 void __add_running_bw(u64 dl_bw, struct dl_rq 234 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
247 { 235 {
248 u64 old = dl_rq->running_bw; 236 u64 old = dl_rq->running_bw;
249 237
250 lockdep_assert_rq_held(rq_of_dl_rq(dl_ 238 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
251 dl_rq->running_bw += dl_bw; 239 dl_rq->running_bw += dl_bw;
252 SCHED_WARN_ON(dl_rq->running_bw < old) 240 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
253 SCHED_WARN_ON(dl_rq->running_bw > dl_r 241 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
254 /* kick cpufreq (see the comment in ke 242 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
255 cpufreq_update_util(rq_of_dl_rq(dl_rq) 243 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
256 } 244 }
257 245
258 static inline 246 static inline
259 void __sub_running_bw(u64 dl_bw, struct dl_rq 247 void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
260 { 248 {
261 u64 old = dl_rq->running_bw; 249 u64 old = dl_rq->running_bw;
262 250
263 lockdep_assert_rq_held(rq_of_dl_rq(dl_ 251 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
264 dl_rq->running_bw -= dl_bw; 252 dl_rq->running_bw -= dl_bw;
265 SCHED_WARN_ON(dl_rq->running_bw > old) 253 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
266 if (dl_rq->running_bw > old) 254 if (dl_rq->running_bw > old)
267 dl_rq->running_bw = 0; 255 dl_rq->running_bw = 0;
268 /* kick cpufreq (see the comment in ke 256 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
269 cpufreq_update_util(rq_of_dl_rq(dl_rq) 257 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
270 } 258 }
271 259
272 static inline 260 static inline
273 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_r 261 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
274 { 262 {
275 u64 old = dl_rq->this_bw; 263 u64 old = dl_rq->this_bw;
276 264
277 lockdep_assert_rq_held(rq_of_dl_rq(dl_ 265 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
278 dl_rq->this_bw += dl_bw; 266 dl_rq->this_bw += dl_bw;
279 SCHED_WARN_ON(dl_rq->this_bw < old); / 267 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
280 } 268 }
281 269
282 static inline 270 static inline
283 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_r 271 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
284 { 272 {
285 u64 old = dl_rq->this_bw; 273 u64 old = dl_rq->this_bw;
286 274
287 lockdep_assert_rq_held(rq_of_dl_rq(dl_ 275 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
288 dl_rq->this_bw -= dl_bw; 276 dl_rq->this_bw -= dl_bw;
289 SCHED_WARN_ON(dl_rq->this_bw > old); / 277 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
290 if (dl_rq->this_bw > old) 278 if (dl_rq->this_bw > old)
291 dl_rq->this_bw = 0; 279 dl_rq->this_bw = 0;
292 SCHED_WARN_ON(dl_rq->running_bw > dl_r 280 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
293 } 281 }
294 282
295 static inline 283 static inline
296 void add_rq_bw(struct sched_dl_entity *dl_se, 284 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
297 { 285 {
298 if (!dl_entity_is_special(dl_se)) 286 if (!dl_entity_is_special(dl_se))
299 __add_rq_bw(dl_se->dl_bw, dl_r 287 __add_rq_bw(dl_se->dl_bw, dl_rq);
300 } 288 }
301 289
302 static inline 290 static inline
303 void sub_rq_bw(struct sched_dl_entity *dl_se, 291 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
304 { 292 {
305 if (!dl_entity_is_special(dl_se)) 293 if (!dl_entity_is_special(dl_se))
306 __sub_rq_bw(dl_se->dl_bw, dl_r 294 __sub_rq_bw(dl_se->dl_bw, dl_rq);
307 } 295 }
308 296
309 static inline 297 static inline
310 void add_running_bw(struct sched_dl_entity *dl 298 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
311 { 299 {
312 if (!dl_entity_is_special(dl_se)) 300 if (!dl_entity_is_special(dl_se))
313 __add_running_bw(dl_se->dl_bw, 301 __add_running_bw(dl_se->dl_bw, dl_rq);
314 } 302 }
315 303
316 static inline 304 static inline
317 void sub_running_bw(struct sched_dl_entity *dl 305 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
318 { 306 {
319 if (!dl_entity_is_special(dl_se)) 307 if (!dl_entity_is_special(dl_se))
320 __sub_running_bw(dl_se->dl_bw, 308 __sub_running_bw(dl_se->dl_bw, dl_rq);
321 } 309 }
322 310
323 static void dl_rq_change_utilization(struct rq !! 311 static void dl_change_utilization(struct task_struct *p, u64 new_bw)
324 { 312 {
325 if (dl_se->dl_non_contending) { !! 313 struct rq *rq;
326 sub_running_bw(dl_se, &rq->dl) !! 314
327 dl_se->dl_non_contending = 0; !! 315 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_SUGOV);
>> 316
>> 317 if (task_on_rq_queued(p))
>> 318 return;
328 319
>> 320 rq = task_rq(p);
>> 321 if (p->dl.dl_non_contending) {
>> 322 sub_running_bw(&p->dl, &rq->dl);
>> 323 p->dl.dl_non_contending = 0;
329 /* 324 /*
330 * If the timer handler is cur 325 * If the timer handler is currently running and the
331 * timer cannot be canceled, i 326 * timer cannot be canceled, inactive_task_timer()
332 * will see that dl_not_conten 327 * will see that dl_not_contending is not set, and
333 * will not touch the rq's act 328 * will not touch the rq's active utilization,
334 * so we are still safe. 329 * so we are still safe.
335 */ 330 */
336 if (hrtimer_try_to_cancel(&dl_ !! 331 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
337 if (!dl_server(dl_se)) !! 332 put_task_struct(p);
338 put_task_struc <<
339 } <<
340 } 333 }
341 __sub_rq_bw(dl_se->dl_bw, &rq->dl); !! 334 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
342 __add_rq_bw(new_bw, &rq->dl); 335 __add_rq_bw(new_bw, &rq->dl);
343 } 336 }
344 337
345 static void dl_change_utilization(struct task_ <<
346 { <<
347 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_ <<
348 <<
349 if (task_on_rq_queued(p)) <<
350 return; <<
351 <<
352 dl_rq_change_utilization(task_rq(p), & <<
353 } <<
354 <<
355 static void __dl_clear_params(struct sched_dl_ <<
356 <<
357 /* 338 /*
358 * The utilization of a task cannot be immedia 339 * The utilization of a task cannot be immediately removed from
359 * the rq active utilization (running_bw) when 340 * the rq active utilization (running_bw) when the task blocks.
360 * Instead, we have to wait for the so called 341 * Instead, we have to wait for the so called "0-lag time".
361 * 342 *
362 * If a task blocks before the "0-lag time", a 343 * If a task blocks before the "0-lag time", a timer (the inactive
363 * timer) is armed, and running_bw is decrease 344 * timer) is armed, and running_bw is decreased when the timer
364 * fires. 345 * fires.
365 * 346 *
366 * If the task wakes up again before the inact 347 * If the task wakes up again before the inactive timer fires,
367 * the timer is canceled, whereas if the task 348 * the timer is canceled, whereas if the task wakes up after the
368 * inactive timer fired (and running_bw has be 349 * inactive timer fired (and running_bw has been decreased) the
369 * task's utilization has to be added to runni 350 * task's utilization has to be added to running_bw again.
370 * A flag in the deadline scheduling entity (d 351 * A flag in the deadline scheduling entity (dl_non_contending)
371 * is used to avoid race conditions between th 352 * is used to avoid race conditions between the inactive timer handler
372 * and task wakeups. 353 * and task wakeups.
373 * 354 *
374 * The following diagram shows how running_bw 355 * The following diagram shows how running_bw is updated. A task is
375 * "ACTIVE" when its utilization contributes t 356 * "ACTIVE" when its utilization contributes to running_bw; an
376 * "ACTIVE contending" task is in the TASK_RUN 357 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
377 * "ACTIVE non contending" task is a blocked t 358 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
378 * has not passed yet. An "INACTIVE" task is a 359 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
379 * time already passed, which does not contrib 360 * time already passed, which does not contribute to running_bw anymore.
380 * +------------- 361 * +------------------+
381 * wakeup | ACTIVE 362 * wakeup | ACTIVE |
382 * +------------------>+ contending 363 * +------------------>+ contending |
383 * | add_running_bw | 364 * | add_running_bw | |
384 * | +----+------+- 365 * | +----+------+------+
385 * | | ^ 366 * | | ^
386 * | dequeue | | 367 * | dequeue | |
387 * +--------+-------+ | | 368 * +--------+-------+ | |
388 * | | t >= 0-lag | | 369 * | | t >= 0-lag | | wakeup
389 * | INACTIVE |<---------------+ | 370 * | INACTIVE |<---------------+ |
390 * | | sub_running_bw | | 371 * | | sub_running_bw | |
391 * +--------+-------+ | | 372 * +--------+-------+ | |
392 * ^ | | 373 * ^ | |
393 * | t < 0-lag | | 374 * | t < 0-lag | |
394 * | | | 375 * | | |
395 * | V | 376 * | V |
396 * | +----+------+- 377 * | +----+------+------+
397 * | sub_running_bw | ACTIVE 378 * | sub_running_bw | ACTIVE |
398 * +-------------------+ 379 * +-------------------+ |
399 * inactive timer | non contend 380 * inactive timer | non contending |
400 * fired +------------- 381 * fired +------------------+
401 * 382 *
402 * The task_non_contending() function is invok 383 * The task_non_contending() function is invoked when a task
403 * blocks, and checks if the 0-lag time alread 384 * blocks, and checks if the 0-lag time already passed or
404 * not (in the first case, it directly updates 385 * not (in the first case, it directly updates running_bw;
405 * in the second case, it arms the inactive ti 386 * in the second case, it arms the inactive timer).
406 * 387 *
407 * The task_contending() function is invoked w 388 * The task_contending() function is invoked when a task wakes
408 * up, and checks if the task is still in the 389 * up, and checks if the task is still in the "ACTIVE non contending"
409 * state or not (in the second case, it update 390 * state or not (in the second case, it updates running_bw).
410 */ 391 */
411 static void task_non_contending(struct sched_d !! 392 static void task_non_contending(struct task_struct *p)
412 { 393 {
>> 394 struct sched_dl_entity *dl_se = &p->dl;
413 struct hrtimer *timer = &dl_se->inacti 395 struct hrtimer *timer = &dl_se->inactive_timer;
414 struct rq *rq = rq_of_dl_se(dl_se); !! 396 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
415 struct dl_rq *dl_rq = &rq->dl; !! 397 struct rq *rq = rq_of_dl_rq(dl_rq);
416 s64 zerolag_time; 398 s64 zerolag_time;
417 399
418 /* 400 /*
419 * If this is a non-deadline task that 401 * If this is a non-deadline task that has been boosted,
420 * do nothing 402 * do nothing
421 */ 403 */
422 if (dl_se->dl_runtime == 0) 404 if (dl_se->dl_runtime == 0)
423 return; 405 return;
424 406
425 if (dl_entity_is_special(dl_se)) 407 if (dl_entity_is_special(dl_se))
426 return; 408 return;
427 409
428 WARN_ON(dl_se->dl_non_contending); 410 WARN_ON(dl_se->dl_non_contending);
429 411
430 zerolag_time = dl_se->deadline - 412 zerolag_time = dl_se->deadline -
431 div64_long((dl_se->runtime * 413 div64_long((dl_se->runtime * dl_se->dl_period),
432 dl_se->dl_runtime); 414 dl_se->dl_runtime);
433 415
434 /* 416 /*
435 * Using relative times instead of the 417 * Using relative times instead of the absolute "0-lag time"
436 * allows to simplify the code 418 * allows to simplify the code
437 */ 419 */
438 zerolag_time -= rq_clock(rq); 420 zerolag_time -= rq_clock(rq);
439 421
440 /* 422 /*
441 * If the "0-lag time" already passed, 423 * If the "0-lag time" already passed, decrease the active
442 * utilization now, instead of startin 424 * utilization now, instead of starting a timer
443 */ 425 */
444 if ((zerolag_time < 0) || hrtimer_acti 426 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
445 if (dl_server(dl_se)) { !! 427 if (dl_task(p))
446 sub_running_bw(dl_se, 428 sub_running_bw(dl_se, dl_rq);
447 } else { !! 429 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
448 struct task_struct *p !! 430 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
449 <<
450 if (dl_task(p)) <<
451 sub_running_bw <<
452 431
453 if (!dl_task(p) || REA !! 432 if (READ_ONCE(p->__state) == TASK_DEAD)
454 struct dl_bw * !! 433 sub_rq_bw(&p->dl, &rq->dl);
455 !! 434 raw_spin_lock(&dl_b->lock);
456 if (READ_ONCE( !! 435 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
457 sub_rq !! 436 raw_spin_unlock(&dl_b->lock);
458 raw_spin_lock( !! 437 __dl_clear_params(p);
459 __dl_sub(dl_b, <<
460 raw_spin_unloc <<
461 __dl_clear_par <<
462 } <<
463 } 438 }
464 439
465 return; 440 return;
466 } 441 }
467 442
468 dl_se->dl_non_contending = 1; 443 dl_se->dl_non_contending = 1;
469 if (!dl_server(dl_se)) !! 444 get_task_struct(p);
470 get_task_struct(dl_task_of(dl_ <<
471 <<
472 hrtimer_start(timer, ns_to_ktime(zerol 445 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
473 } 446 }
474 447
475 static void task_contending(struct sched_dl_en 448 static void task_contending(struct sched_dl_entity *dl_se, int flags)
476 { 449 {
477 struct dl_rq *dl_rq = dl_rq_of_se(dl_s 450 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
478 451
479 /* 452 /*
480 * If this is a non-deadline task that 453 * If this is a non-deadline task that has been boosted,
481 * do nothing 454 * do nothing
482 */ 455 */
483 if (dl_se->dl_runtime == 0) 456 if (dl_se->dl_runtime == 0)
484 return; 457 return;
485 458
486 if (flags & ENQUEUE_MIGRATED) 459 if (flags & ENQUEUE_MIGRATED)
487 add_rq_bw(dl_se, dl_rq); 460 add_rq_bw(dl_se, dl_rq);
488 461
489 if (dl_se->dl_non_contending) { 462 if (dl_se->dl_non_contending) {
490 dl_se->dl_non_contending = 0; 463 dl_se->dl_non_contending = 0;
491 /* 464 /*
492 * If the timer handler is cur 465 * If the timer handler is currently running and the
493 * timer cannot be canceled, i 466 * timer cannot be canceled, inactive_task_timer()
494 * will see that dl_not_conten 467 * will see that dl_not_contending is not set, and
495 * will not touch the rq's act 468 * will not touch the rq's active utilization,
496 * so we are still safe. 469 * so we are still safe.
497 */ 470 */
498 if (hrtimer_try_to_cancel(&dl_ !! 471 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
499 if (!dl_server(dl_se)) !! 472 put_task_struct(dl_task_of(dl_se));
500 put_task_struc <<
501 } <<
502 } else { 473 } else {
503 /* 474 /*
504 * Since "dl_non_contending" i 475 * Since "dl_non_contending" is not set, the
505 * task's utilization has alre 476 * task's utilization has already been removed from
506 * active utilization (either 477 * active utilization (either when the task blocked,
507 * when the "inactive timer" f 478 * when the "inactive timer" fired).
508 * So, add it back. 479 * So, add it back.
509 */ 480 */
510 add_running_bw(dl_se, dl_rq); 481 add_running_bw(dl_se, dl_rq);
511 } 482 }
512 } 483 }
513 484
514 static inline int is_leftmost(struct sched_dl_ !! 485 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
515 { 486 {
>> 487 struct sched_dl_entity *dl_se = &p->dl;
>> 488
516 return rb_first_cached(&dl_rq->root) = 489 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
517 } 490 }
518 491
519 static void init_dl_rq_bw_ratio(struct dl_rq * 492 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
520 493
>> 494 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
>> 495 {
>> 496 raw_spin_lock_init(&dl_b->dl_runtime_lock);
>> 497 dl_b->dl_period = period;
>> 498 dl_b->dl_runtime = runtime;
>> 499 }
>> 500
521 void init_dl_bw(struct dl_bw *dl_b) 501 void init_dl_bw(struct dl_bw *dl_b)
522 { 502 {
523 raw_spin_lock_init(&dl_b->lock); 503 raw_spin_lock_init(&dl_b->lock);
524 if (global_rt_runtime() == RUNTIME_INF 504 if (global_rt_runtime() == RUNTIME_INF)
525 dl_b->bw = -1; 505 dl_b->bw = -1;
526 else 506 else
527 dl_b->bw = to_ratio(global_rt_ 507 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
528 dl_b->total_bw = 0; 508 dl_b->total_bw = 0;
529 } 509 }
530 510
531 void init_dl_rq(struct dl_rq *dl_rq) 511 void init_dl_rq(struct dl_rq *dl_rq)
532 { 512 {
533 dl_rq->root = RB_ROOT_CACHED; 513 dl_rq->root = RB_ROOT_CACHED;
534 514
535 #ifdef CONFIG_SMP 515 #ifdef CONFIG_SMP
536 /* zero means no -deadline tasks */ 516 /* zero means no -deadline tasks */
537 dl_rq->earliest_dl.curr = dl_rq->earli 517 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
538 518
>> 519 dl_rq->dl_nr_migratory = 0;
539 dl_rq->overloaded = 0; 520 dl_rq->overloaded = 0;
540 dl_rq->pushable_dl_tasks_root = RB_ROO 521 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
541 #else 522 #else
542 init_dl_bw(&dl_rq->dl_bw); 523 init_dl_bw(&dl_rq->dl_bw);
543 #endif 524 #endif
544 525
545 dl_rq->running_bw = 0; 526 dl_rq->running_bw = 0;
546 dl_rq->this_bw = 0; 527 dl_rq->this_bw = 0;
547 init_dl_rq_bw_ratio(dl_rq); 528 init_dl_rq_bw_ratio(dl_rq);
548 } 529 }
549 530
550 #ifdef CONFIG_SMP 531 #ifdef CONFIG_SMP
551 532
552 static inline int dl_overloaded(struct rq *rq) 533 static inline int dl_overloaded(struct rq *rq)
553 { 534 {
554 return atomic_read(&rq->rd->dlo_count) 535 return atomic_read(&rq->rd->dlo_count);
555 } 536 }
556 537
557 static inline void dl_set_overload(struct rq * 538 static inline void dl_set_overload(struct rq *rq)
558 { 539 {
559 if (!rq->online) 540 if (!rq->online)
560 return; 541 return;
561 542
562 cpumask_set_cpu(rq->cpu, rq->rd->dlo_m 543 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
563 /* 544 /*
564 * Must be visible before the overload 545 * Must be visible before the overload count is
565 * set (as in sched_rt.c). 546 * set (as in sched_rt.c).
566 * 547 *
567 * Matched by the barrier in pull_dl_t 548 * Matched by the barrier in pull_dl_task().
568 */ 549 */
569 smp_wmb(); 550 smp_wmb();
570 atomic_inc(&rq->rd->dlo_count); 551 atomic_inc(&rq->rd->dlo_count);
571 } 552 }
572 553
573 static inline void dl_clear_overload(struct rq 554 static inline void dl_clear_overload(struct rq *rq)
574 { 555 {
575 if (!rq->online) 556 if (!rq->online)
576 return; 557 return;
577 558
578 atomic_dec(&rq->rd->dlo_count); 559 atomic_dec(&rq->rd->dlo_count);
579 cpumask_clear_cpu(rq->cpu, rq->rd->dlo 560 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
580 } 561 }
581 562
>> 563 static void update_dl_migration(struct dl_rq *dl_rq)
>> 564 {
>> 565 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
>> 566 if (!dl_rq->overloaded) {
>> 567 dl_set_overload(rq_of_dl_rq(dl_rq));
>> 568 dl_rq->overloaded = 1;
>> 569 }
>> 570 } else if (dl_rq->overloaded) {
>> 571 dl_clear_overload(rq_of_dl_rq(dl_rq));
>> 572 dl_rq->overloaded = 0;
>> 573 }
>> 574 }
>> 575
>> 576 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
>> 577 {
>> 578 struct task_struct *p = dl_task_of(dl_se);
>> 579
>> 580 if (p->nr_cpus_allowed > 1)
>> 581 dl_rq->dl_nr_migratory++;
>> 582
>> 583 update_dl_migration(dl_rq);
>> 584 }
>> 585
>> 586 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
>> 587 {
>> 588 struct task_struct *p = dl_task_of(dl_se);
>> 589
>> 590 if (p->nr_cpus_allowed > 1)
>> 591 dl_rq->dl_nr_migratory--;
>> 592
>> 593 update_dl_migration(dl_rq);
>> 594 }
>> 595
582 #define __node_2_pdl(node) \ 596 #define __node_2_pdl(node) \
583 rb_entry((node), struct task_struct, p 597 rb_entry((node), struct task_struct, pushable_dl_tasks)
584 598
585 static inline bool __pushable_less(struct rb_n 599 static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
586 { 600 {
587 return dl_entity_preempt(&__node_2_pdl 601 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
588 } 602 }
589 603
590 static inline int has_pushable_dl_tasks(struct <<
591 { <<
592 return !RB_EMPTY_ROOT(&rq->dl.pushable <<
593 } <<
594 <<
595 /* 604 /*
596 * The list of pushable -deadline task is not 605 * The list of pushable -deadline task is not a plist, like in
597 * sched_rt.c, it is an rb-tree with tasks ord 606 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
598 */ 607 */
599 static void enqueue_pushable_dl_task(struct rq 608 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
600 { 609 {
601 struct rb_node *leftmost; 610 struct rb_node *leftmost;
602 611
603 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushab 612 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
604 613
605 leftmost = rb_add_cached(&p->pushable_ 614 leftmost = rb_add_cached(&p->pushable_dl_tasks,
606 &rq->dl.pusha 615 &rq->dl.pushable_dl_tasks_root,
607 __pushable_le 616 __pushable_less);
608 if (leftmost) 617 if (leftmost)
609 rq->dl.earliest_dl.next = p->d 618 rq->dl.earliest_dl.next = p->dl.deadline;
610 <<
611 if (!rq->dl.overloaded) { <<
612 dl_set_overload(rq); <<
613 rq->dl.overloaded = 1; <<
614 } <<
615 } 619 }
616 620
617 static void dequeue_pushable_dl_task(struct rq 621 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
618 { 622 {
619 struct dl_rq *dl_rq = &rq->dl; 623 struct dl_rq *dl_rq = &rq->dl;
620 struct rb_root_cached *root = &dl_rq-> 624 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
621 struct rb_node *leftmost; 625 struct rb_node *leftmost;
622 626
623 if (RB_EMPTY_NODE(&p->pushable_dl_task 627 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
624 return; 628 return;
625 629
626 leftmost = rb_erase_cached(&p->pushabl 630 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
627 if (leftmost) 631 if (leftmost)
628 dl_rq->earliest_dl.next = __no 632 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
629 633
630 RB_CLEAR_NODE(&p->pushable_dl_tasks); 634 RB_CLEAR_NODE(&p->pushable_dl_tasks);
>> 635 }
631 636
632 if (!has_pushable_dl_tasks(rq) && rq-> !! 637 static inline int has_pushable_dl_tasks(struct rq *rq)
633 dl_clear_overload(rq); !! 638 {
634 rq->dl.overloaded = 0; !! 639 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
635 } <<
636 } 640 }
637 641
638 static int push_dl_task(struct rq *rq); 642 static int push_dl_task(struct rq *rq);
639 643
640 static inline bool need_pull_dl_task(struct rq 644 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
641 { 645 {
642 return rq->online && dl_task(prev); 646 return rq->online && dl_task(prev);
643 } 647 }
644 648
645 static DEFINE_PER_CPU(struct balance_callback, 649 static DEFINE_PER_CPU(struct balance_callback, dl_push_head);
646 static DEFINE_PER_CPU(struct balance_callback, 650 static DEFINE_PER_CPU(struct balance_callback, dl_pull_head);
647 651
648 static void push_dl_tasks(struct rq *); 652 static void push_dl_tasks(struct rq *);
649 static void pull_dl_task(struct rq *); 653 static void pull_dl_task(struct rq *);
650 654
651 static inline void deadline_queue_push_tasks(s 655 static inline void deadline_queue_push_tasks(struct rq *rq)
652 { 656 {
653 if (!has_pushable_dl_tasks(rq)) 657 if (!has_pushable_dl_tasks(rq))
654 return; 658 return;
655 659
656 queue_balance_callback(rq, &per_cpu(dl 660 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
657 } 661 }
658 662
659 static inline void deadline_queue_pull_task(st 663 static inline void deadline_queue_pull_task(struct rq *rq)
660 { 664 {
661 queue_balance_callback(rq, &per_cpu(dl 665 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
662 } 666 }
663 667
664 static struct rq *find_lock_later_rq(struct ta 668 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
665 669
666 static struct rq *dl_task_offline_migration(st 670 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
667 { 671 {
668 struct rq *later_rq = NULL; 672 struct rq *later_rq = NULL;
669 struct dl_bw *dl_b; 673 struct dl_bw *dl_b;
670 674
671 later_rq = find_lock_later_rq(p, rq); 675 later_rq = find_lock_later_rq(p, rq);
672 if (!later_rq) { 676 if (!later_rq) {
673 int cpu; 677 int cpu;
674 678
675 /* 679 /*
676 * If we cannot preempt any rq 680 * If we cannot preempt any rq, fall back to pick any
677 * online CPU: 681 * online CPU:
678 */ 682 */
679 cpu = cpumask_any_and(cpu_acti 683 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
680 if (cpu >= nr_cpu_ids) { 684 if (cpu >= nr_cpu_ids) {
681 /* 685 /*
682 * Failed to find any 686 * Failed to find any suitable CPU.
683 * The task will never 687 * The task will never come back!
684 */ 688 */
685 WARN_ON_ONCE(dl_bandwi 689 WARN_ON_ONCE(dl_bandwidth_enabled());
686 690
687 /* 691 /*
688 * If admission contro 692 * If admission control is disabled we
689 * try a little harder 693 * try a little harder to let the task
690 * run. 694 * run.
691 */ 695 */
692 cpu = cpumask_any(cpu_ 696 cpu = cpumask_any(cpu_active_mask);
693 } 697 }
694 later_rq = cpu_rq(cpu); 698 later_rq = cpu_rq(cpu);
695 double_lock_balance(rq, later_ 699 double_lock_balance(rq, later_rq);
696 } 700 }
697 701
698 if (p->dl.dl_non_contending || p->dl.d 702 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
699 /* 703 /*
700 * Inactive timer is armed (or 704 * Inactive timer is armed (or callback is running, but
701 * waiting for us to release r 705 * waiting for us to release rq locks). In any case, when it
702 * will fire (or continue), it 706 * will fire (or continue), it will see running_bw of this
703 * task migrated to later_rq ( 707 * task migrated to later_rq (and correctly handle it).
704 */ 708 */
705 sub_running_bw(&p->dl, &rq->dl 709 sub_running_bw(&p->dl, &rq->dl);
706 sub_rq_bw(&p->dl, &rq->dl); 710 sub_rq_bw(&p->dl, &rq->dl);
707 711
708 add_rq_bw(&p->dl, &later_rq->d 712 add_rq_bw(&p->dl, &later_rq->dl);
709 add_running_bw(&p->dl, &later_ 713 add_running_bw(&p->dl, &later_rq->dl);
710 } else { 714 } else {
711 sub_rq_bw(&p->dl, &rq->dl); 715 sub_rq_bw(&p->dl, &rq->dl);
712 add_rq_bw(&p->dl, &later_rq->d 716 add_rq_bw(&p->dl, &later_rq->dl);
713 } 717 }
714 718
715 /* 719 /*
716 * And we finally need to fix up root_ !! 720 * And we finally need to fixup root_domain(s) bandwidth accounting,
717 * since p is still hanging out in the 721 * since p is still hanging out in the old (now moved to default) root
718 * domain. 722 * domain.
719 */ 723 */
720 dl_b = &rq->rd->dl_bw; 724 dl_b = &rq->rd->dl_bw;
721 raw_spin_lock(&dl_b->lock); 725 raw_spin_lock(&dl_b->lock);
722 __dl_sub(dl_b, p->dl.dl_bw, cpumask_we 726 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
723 raw_spin_unlock(&dl_b->lock); 727 raw_spin_unlock(&dl_b->lock);
724 728
725 dl_b = &later_rq->rd->dl_bw; 729 dl_b = &later_rq->rd->dl_bw;
726 raw_spin_lock(&dl_b->lock); 730 raw_spin_lock(&dl_b->lock);
727 __dl_add(dl_b, p->dl.dl_bw, cpumask_we 731 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
728 raw_spin_unlock(&dl_b->lock); 732 raw_spin_unlock(&dl_b->lock);
729 733
730 set_task_cpu(p, later_rq->cpu); 734 set_task_cpu(p, later_rq->cpu);
731 double_unlock_balance(later_rq, rq); 735 double_unlock_balance(later_rq, rq);
732 736
733 return later_rq; 737 return later_rq;
734 } 738 }
735 739
736 #else 740 #else
737 741
738 static inline 742 static inline
739 void enqueue_pushable_dl_task(struct rq *rq, s 743 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
740 { 744 {
741 } 745 }
742 746
743 static inline 747 static inline
744 void dequeue_pushable_dl_task(struct rq *rq, s 748 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
745 { 749 {
746 } 750 }
747 751
748 static inline 752 static inline
749 void inc_dl_migration(struct sched_dl_entity * 753 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
750 { 754 {
751 } 755 }
752 756
753 static inline 757 static inline
754 void dec_dl_migration(struct sched_dl_entity * 758 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
755 { 759 {
756 } 760 }
757 761
758 static inline void deadline_queue_push_tasks(s 762 static inline void deadline_queue_push_tasks(struct rq *rq)
759 { 763 {
760 } 764 }
761 765
762 static inline void deadline_queue_pull_task(st 766 static inline void deadline_queue_pull_task(struct rq *rq)
763 { 767 {
764 } 768 }
765 #endif /* CONFIG_SMP */ 769 #endif /* CONFIG_SMP */
766 770
767 static void <<
768 enqueue_dl_entity(struct sched_dl_entity *dl_s <<
769 static void enqueue_task_dl(struct rq *rq, str 771 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
770 static void dequeue_dl_entity(struct sched_dl_ !! 772 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
771 static void wakeup_preempt_dl(struct rq *rq, s !! 773 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
772 774
773 static inline void replenish_dl_new_period(str 775 static inline void replenish_dl_new_period(struct sched_dl_entity *dl_se,
774 st 776 struct rq *rq)
775 { 777 {
776 /* for non-boosted task, pi_of(dl_se) 778 /* for non-boosted task, pi_of(dl_se) == dl_se */
777 dl_se->deadline = rq_clock(rq) + pi_of 779 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
778 dl_se->runtime = pi_of(dl_se)->dl_runt 780 dl_se->runtime = pi_of(dl_se)->dl_runtime;
779 <<
780 /* <<
781 * If it is a deferred reservation, an <<
782 * is not handling an starvation case, <<
783 */ <<
784 if (dl_se->dl_defer & !dl_se->dl_defer <<
785 dl_se->dl_throttled = 1; <<
786 dl_se->dl_defer_armed = 1; <<
787 } <<
788 } 781 }
789 782
790 /* 783 /*
791 * We are being explicitly informed that a new 784 * We are being explicitly informed that a new instance is starting,
792 * and this means that: 785 * and this means that:
793 * - the absolute deadline of the entity has 786 * - the absolute deadline of the entity has to be placed at
794 * current time + relative deadline; 787 * current time + relative deadline;
795 * - the runtime of the entity has to be set 788 * - the runtime of the entity has to be set to the maximum value.
796 * 789 *
797 * The capability of specifying such event is 790 * The capability of specifying such event is useful whenever a -deadline
798 * entity wants to (try to!) synchronize its b 791 * entity wants to (try to!) synchronize its behaviour with the scheduler's
799 * one, and to (try to!) reconcile itself with 792 * one, and to (try to!) reconcile itself with its own scheduling
800 * parameters. 793 * parameters.
801 */ 794 */
802 static inline void setup_new_dl_entity(struct 795 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
803 { 796 {
804 struct dl_rq *dl_rq = dl_rq_of_se(dl_s 797 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
805 struct rq *rq = rq_of_dl_rq(dl_rq); 798 struct rq *rq = rq_of_dl_rq(dl_rq);
806 799
807 WARN_ON(is_dl_boosted(dl_se)); 800 WARN_ON(is_dl_boosted(dl_se));
808 WARN_ON(dl_time_before(rq_clock(rq), d 801 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
809 802
810 /* 803 /*
811 * We are racing with the deadline tim 804 * We are racing with the deadline timer. So, do nothing because
812 * the deadline timer handler will tak 805 * the deadline timer handler will take care of properly recharging
813 * the runtime and postponing the dead 806 * the runtime and postponing the deadline
814 */ 807 */
815 if (dl_se->dl_throttled) 808 if (dl_se->dl_throttled)
816 return; 809 return;
817 810
818 /* 811 /*
819 * We use the regular wall clock time 812 * We use the regular wall clock time to set deadlines in the
820 * future; in fact, we must consider e 813 * future; in fact, we must consider execution overheads (time
821 * spent on hardirq context, etc.). 814 * spent on hardirq context, etc.).
822 */ 815 */
823 replenish_dl_new_period(dl_se, rq); 816 replenish_dl_new_period(dl_se, rq);
824 } 817 }
825 818
826 static int start_dl_timer(struct sched_dl_enti <<
827 static bool dl_entity_overflow(struct sched_dl <<
828 <<
829 /* 819 /*
830 * Pure Earliest Deadline First (EDF) scheduli 820 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
831 * possibility of a entity lasting more than w 821 * possibility of a entity lasting more than what it declared, and thus
832 * exhausting its runtime. 822 * exhausting its runtime.
833 * 823 *
834 * Here we are interested in making runtime ov 824 * Here we are interested in making runtime overrun possible, but we do
835 * not want a entity which is misbehaving to a 825 * not want a entity which is misbehaving to affect the scheduling of all
836 * other entities. 826 * other entities.
837 * Therefore, a budgeting strategy called Cons 827 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
838 * is used, in order to confine each entity wi 828 * is used, in order to confine each entity within its own bandwidth.
839 * 829 *
840 * This function deals exactly with that, and 830 * This function deals exactly with that, and ensures that when the runtime
841 * of a entity is replenished, its deadline is 831 * of a entity is replenished, its deadline is also postponed. That ensures
842 * the overrunning entity can't interfere with 832 * the overrunning entity can't interfere with other entity in the system and
843 * can't make them miss their deadlines. Reaso 833 * can't make them miss their deadlines. Reasons why this kind of overruns
844 * could happen are, typically, a entity volun 834 * could happen are, typically, a entity voluntarily trying to overcome its
845 * runtime, or it just underestimated it durin 835 * runtime, or it just underestimated it during sched_setattr().
846 */ 836 */
847 static void replenish_dl_entity(struct sched_d 837 static void replenish_dl_entity(struct sched_dl_entity *dl_se)
848 { 838 {
849 struct dl_rq *dl_rq = dl_rq_of_se(dl_s 839 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
850 struct rq *rq = rq_of_dl_rq(dl_rq); 840 struct rq *rq = rq_of_dl_rq(dl_rq);
851 841
852 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime 842 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime <= 0);
853 843
854 /* 844 /*
855 * This could be the case for a !-dl t 845 * This could be the case for a !-dl task that is boosted.
856 * Just go with full inherited paramet 846 * Just go with full inherited parameters.
857 * <<
858 * Or, it could be the case of a defer <<
859 * was not able to consume its runtime <<
860 * reached this point with current u > <<
861 * <<
862 * In both cases, set a new period. <<
863 */ 847 */
864 if (dl_se->dl_deadline == 0 || !! 848 if (dl_se->dl_deadline == 0)
865 (dl_se->dl_defer_armed && dl_entit !! 849 replenish_dl_new_period(dl_se, rq);
866 dl_se->deadline = rq_clock(rq) <<
867 dl_se->runtime = pi_of(dl_se)- <<
868 } <<
869 850
870 if (dl_se->dl_yielded && dl_se->runtim 851 if (dl_se->dl_yielded && dl_se->runtime > 0)
871 dl_se->runtime = 0; 852 dl_se->runtime = 0;
872 853
873 /* 854 /*
874 * We keep moving the deadline away un 855 * We keep moving the deadline away until we get some
875 * available runtime for the entity. T 856 * available runtime for the entity. This ensures correct
876 * handling of situations where the ru 857 * handling of situations where the runtime overrun is
877 * arbitrary large. 858 * arbitrary large.
878 */ 859 */
879 while (dl_se->runtime <= 0) { 860 while (dl_se->runtime <= 0) {
880 dl_se->deadline += pi_of(dl_se 861 dl_se->deadline += pi_of(dl_se)->dl_period;
881 dl_se->runtime += pi_of(dl_se) 862 dl_se->runtime += pi_of(dl_se)->dl_runtime;
882 } 863 }
883 864
884 /* 865 /*
885 * At this point, the deadline really 866 * At this point, the deadline really should be "in
886 * the future" with respect to rq->clo 867 * the future" with respect to rq->clock. If it's
887 * not, we are, for some reason, laggi 868 * not, we are, for some reason, lagging too much!
888 * Anyway, after having warn userspace 869 * Anyway, after having warn userspace abut that,
889 * we still try to keep the things run 870 * we still try to keep the things running by
890 * resetting the deadline and the budg 871 * resetting the deadline and the budget of the
891 * entity. 872 * entity.
892 */ 873 */
893 if (dl_time_before(dl_se->deadline, rq 874 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
894 printk_deferred_once("sched: D 875 printk_deferred_once("sched: DL replenish lagged too much\n");
895 replenish_dl_new_period(dl_se, 876 replenish_dl_new_period(dl_se, rq);
896 } 877 }
897 878
898 if (dl_se->dl_yielded) 879 if (dl_se->dl_yielded)
899 dl_se->dl_yielded = 0; 880 dl_se->dl_yielded = 0;
900 if (dl_se->dl_throttled) 881 if (dl_se->dl_throttled)
901 dl_se->dl_throttled = 0; 882 dl_se->dl_throttled = 0;
902 <<
903 /* <<
904 * If this is the replenishment of a d <<
905 * clear the flag and return. <<
906 */ <<
907 if (dl_se->dl_defer_armed) { <<
908 dl_se->dl_defer_armed = 0; <<
909 return; <<
910 } <<
911 <<
912 /* <<
913 * A this point, if the deferred serve <<
914 * is in the future, if it is not runn <<
915 * and arm the defer timer. <<
916 */ <<
917 if (dl_se->dl_defer && !dl_se->dl_defe <<
918 dl_time_before(rq_clock(dl_se->rq) <<
919 if (!is_dl_boosted(dl_se) && d <<
920 <<
921 /* <<
922 * Set dl_se->dl_defer <<
923 * inform the start_dl <<
924 * activation. <<
925 */ <<
926 dl_se->dl_defer_armed <<
927 dl_se->dl_throttled = <<
928 if (!start_dl_timer(dl <<
929 /* <<
930 * If for what <<
931 * queued but <<
932 * deferrable <<
933 */ <<
934 hrtimer_try_to <<
935 dl_se->dl_defe <<
936 dl_se->dl_thro <<
937 } <<
938 } <<
939 } <<
940 } 883 }
941 884
942 /* 885 /*
943 * Here we check if --at time t-- an entity (w 886 * Here we check if --at time t-- an entity (which is probably being
944 * [re]activated or, in general, enqueued) can 887 * [re]activated or, in general, enqueued) can use its remaining runtime
945 * and its current deadline _without_ exceedin 888 * and its current deadline _without_ exceeding the bandwidth it is
946 * assigned (function returns true if it can't 889 * assigned (function returns true if it can't). We are in fact applying
947 * one of the CBS rules: when a task wakes up, 890 * one of the CBS rules: when a task wakes up, if the residual runtime
948 * over residual deadline fits within the allo 891 * over residual deadline fits within the allocated bandwidth, then we
949 * can keep the current (absolute) deadline an 892 * can keep the current (absolute) deadline and residual budget without
950 * disrupting the schedulability of the system 893 * disrupting the schedulability of the system. Otherwise, we should
951 * refill the runtime and set the deadline a p 894 * refill the runtime and set the deadline a period in the future,
952 * because keeping the current (absolute) dead 895 * because keeping the current (absolute) deadline of the task would
953 * result in breaking guarantees promised to o 896 * result in breaking guarantees promised to other tasks (refer to
954 * Documentation/scheduler/sched-deadline.rst 897 * Documentation/scheduler/sched-deadline.rst for more information).
955 * 898 *
956 * This function returns true if: 899 * This function returns true if:
957 * 900 *
958 * runtime / (deadline - t) > dl_runtime / d 901 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
959 * 902 *
960 * IOW we can't recycle current parameters. 903 * IOW we can't recycle current parameters.
961 * 904 *
962 * Notice that the bandwidth check is done aga 905 * Notice that the bandwidth check is done against the deadline. For
963 * task with deadline equal to period this is 906 * task with deadline equal to period this is the same of using
964 * dl_period instead of dl_deadline in the equ 907 * dl_period instead of dl_deadline in the equation above.
965 */ 908 */
966 static bool dl_entity_overflow(struct sched_dl 909 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
967 { 910 {
968 u64 left, right; 911 u64 left, right;
969 912
970 /* 913 /*
971 * left and right are the two sides of 914 * left and right are the two sides of the equation above,
972 * after a bit of shuffling to use mul 915 * after a bit of shuffling to use multiplications instead
973 * of divisions. 916 * of divisions.
974 * 917 *
975 * Note that none of the time values i 918 * Note that none of the time values involved in the two
976 * multiplications are absolute: dl_de 919 * multiplications are absolute: dl_deadline and dl_runtime
977 * are the relative deadline and the m 920 * are the relative deadline and the maximum runtime of each
978 * instance, runtime is the runtime le 921 * instance, runtime is the runtime left for the last instance
979 * and (deadline - t), since t is rq-> 922 * and (deadline - t), since t is rq->clock, is the time left
980 * to the (absolute) deadline. Even if 923 * to the (absolute) deadline. Even if overflowing the u64 type
981 * is very unlikely to occur in both c 924 * is very unlikely to occur in both cases, here we scale down
982 * as we want to avoid that risk at al 925 * as we want to avoid that risk at all. Scaling down by 10
983 * means that we reduce granularity to 926 * means that we reduce granularity to 1us. We are fine with it,
984 * since this is only a true/false che 927 * since this is only a true/false check and, anyway, thinking
985 * of anything below microseconds reso 928 * of anything below microseconds resolution is actually fiction
986 * (but still we want to give the user 929 * (but still we want to give the user that illusion >;).
987 */ 930 */
988 left = (pi_of(dl_se)->dl_deadline >> D 931 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
989 right = ((dl_se->deadline - t) >> DL_S 932 right = ((dl_se->deadline - t) >> DL_SCALE) *
990 (pi_of(dl_se)->dl_runtime >> D 933 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
991 934
992 return dl_time_before(right, left); 935 return dl_time_before(right, left);
993 } 936 }
994 937
995 /* 938 /*
996 * Revised wakeup rule [1]: For self-suspendin 939 * Revised wakeup rule [1]: For self-suspending tasks, rather then
997 * re-initializing task's runtime and deadline 940 * re-initializing task's runtime and deadline, the revised wakeup
998 * rule adjusts the task's runtime to avoid th 941 * rule adjusts the task's runtime to avoid the task to overrun its
999 * density. 942 * density.
1000 * 943 *
1001 * Reasoning: a task may overrun the density 944 * Reasoning: a task may overrun the density if:
1002 * runtime / (deadline - t) > dl_runtime / 945 * runtime / (deadline - t) > dl_runtime / dl_deadline
1003 * 946 *
1004 * Therefore, runtime can be adjusted to: 947 * Therefore, runtime can be adjusted to:
1005 * runtime = (dl_runtime / dl_deadline) * 948 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
1006 * 949 *
1007 * In such way that runtime will be equal to 950 * In such way that runtime will be equal to the maximum density
1008 * the task can use without breaking any rule 951 * the task can use without breaking any rule.
1009 * 952 *
1010 * [1] Luca Abeni, Giuseppe Lipari, and Juri 953 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
1011 * bandwidth server revisited. SIGBED Rev. 11 954 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
1012 */ 955 */
1013 static void 956 static void
1014 update_dl_revised_wakeup(struct sched_dl_enti 957 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
1015 { 958 {
1016 u64 laxity = dl_se->deadline - rq_clo 959 u64 laxity = dl_se->deadline - rq_clock(rq);
1017 960
1018 /* 961 /*
1019 * If the task has deadline < period, 962 * If the task has deadline < period, and the deadline is in the past,
1020 * it should already be throttled bef 963 * it should already be throttled before this check.
1021 * 964 *
1022 * See update_dl_entity() comments fo 965 * See update_dl_entity() comments for further details.
1023 */ 966 */
1024 WARN_ON(dl_time_before(dl_se->deadlin 967 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
1025 968
1026 dl_se->runtime = (dl_se->dl_density * 969 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
1027 } 970 }
1028 971
1029 /* 972 /*
1030 * Regarding the deadline, a task with implic 973 * Regarding the deadline, a task with implicit deadline has a relative
1031 * deadline == relative period. A task with c 974 * deadline == relative period. A task with constrained deadline has a
1032 * relative deadline <= relative period. 975 * relative deadline <= relative period.
1033 * 976 *
1034 * We support constrained deadline tasks. How 977 * We support constrained deadline tasks. However, there are some restrictions
1035 * applied only for tasks which do not have a 978 * applied only for tasks which do not have an implicit deadline. See
1036 * update_dl_entity() to know more about such 979 * update_dl_entity() to know more about such restrictions.
1037 * 980 *
1038 * The dl_is_implicit() returns true if the t 981 * The dl_is_implicit() returns true if the task has an implicit deadline.
1039 */ 982 */
1040 static inline bool dl_is_implicit(struct sche 983 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
1041 { 984 {
1042 return dl_se->dl_deadline == dl_se->d 985 return dl_se->dl_deadline == dl_se->dl_period;
1043 } 986 }
1044 987
1045 /* 988 /*
1046 * When a deadline entity is placed in the ru 989 * When a deadline entity is placed in the runqueue, its runtime and deadline
1047 * might need to be updated. This is done by 990 * might need to be updated. This is done by a CBS wake up rule. There are two
1048 * different rules: 1) the original CBS; and 991 * different rules: 1) the original CBS; and 2) the Revisited CBS.
1049 * 992 *
1050 * When the task is starting a new period, th 993 * When the task is starting a new period, the Original CBS is used. In this
1051 * case, the runtime is replenished and a new 994 * case, the runtime is replenished and a new absolute deadline is set.
1052 * 995 *
1053 * When a task is queued before the begin of 996 * When a task is queued before the begin of the next period, using the
1054 * remaining runtime and deadline could make 997 * remaining runtime and deadline could make the entity to overflow, see
1055 * dl_entity_overflow() to find more about ru 998 * dl_entity_overflow() to find more about runtime overflow. When such case
1056 * is detected, the runtime and deadline need 999 * is detected, the runtime and deadline need to be updated.
1057 * 1000 *
1058 * If the task has an implicit deadline, i.e. 1001 * If the task has an implicit deadline, i.e., deadline == period, the Original
1059 * CBS is applied. The runtime is replenished !! 1002 * CBS is applied. the runtime is replenished and a new absolute deadline is
1060 * set, as in the previous cases. 1003 * set, as in the previous cases.
1061 * 1004 *
1062 * However, the Original CBS does not work pr 1005 * However, the Original CBS does not work properly for tasks with
1063 * deadline < period, which are said to have 1006 * deadline < period, which are said to have a constrained deadline. By
1064 * applying the Original CBS, a constrained d 1007 * applying the Original CBS, a constrained deadline task would be able to run
1065 * runtime/deadline in a period. With deadlin 1008 * runtime/deadline in a period. With deadline < period, the task would
1066 * overrun the runtime/period allowed bandwid 1009 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1067 * 1010 *
1068 * In order to prevent this misbehave, the Re 1011 * In order to prevent this misbehave, the Revisited CBS is used for
1069 * constrained deadline tasks when a runtime 1012 * constrained deadline tasks when a runtime overflow is detected. In the
1070 * Revisited CBS, rather than replenishing & 1013 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1071 * the remaining runtime of the task is reduc 1014 * the remaining runtime of the task is reduced to avoid runtime overflow.
1072 * Please refer to the comments update_dl_rev 1015 * Please refer to the comments update_dl_revised_wakeup() function to find
1073 * more about the Revised CBS rule. 1016 * more about the Revised CBS rule.
1074 */ 1017 */
1075 static void update_dl_entity(struct sched_dl_ 1018 static void update_dl_entity(struct sched_dl_entity *dl_se)
1076 { 1019 {
1077 struct rq *rq = rq_of_dl_se(dl_se); !! 1020 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
>> 1021 struct rq *rq = rq_of_dl_rq(dl_rq);
1078 1022
1079 if (dl_time_before(dl_se->deadline, r 1023 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1080 dl_entity_overflow(dl_se, rq_cloc 1024 dl_entity_overflow(dl_se, rq_clock(rq))) {
1081 1025
1082 if (unlikely(!dl_is_implicit( 1026 if (unlikely(!dl_is_implicit(dl_se) &&
1083 !dl_time_before( 1027 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1084 !is_dl_boosted(d 1028 !is_dl_boosted(dl_se))) {
1085 update_dl_revised_wak 1029 update_dl_revised_wakeup(dl_se, rq);
1086 return; 1030 return;
1087 } 1031 }
1088 1032
1089 replenish_dl_new_period(dl_se 1033 replenish_dl_new_period(dl_se, rq);
1090 } else if (dl_server(dl_se) && dl_se- <<
1091 /* <<
1092 * The server can still use i <<
1093 * it left the dl_defer_runni <<
1094 */ <<
1095 if (!dl_se->dl_defer_running) <<
1096 dl_se->dl_defer_armed <<
1097 dl_se->dl_throttled = <<
1098 } <<
1099 } 1034 }
1100 } 1035 }
1101 1036
1102 static inline u64 dl_next_period(struct sched 1037 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1103 { 1038 {
1104 return dl_se->deadline - dl_se->dl_de 1039 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1105 } 1040 }
1106 1041
1107 /* 1042 /*
1108 * If the entity depleted all its runtime, an 1043 * If the entity depleted all its runtime, and if we want it to sleep
1109 * while waiting for some new execution time 1044 * while waiting for some new execution time to become available, we
1110 * set the bandwidth replenishment timer to t 1045 * set the bandwidth replenishment timer to the replenishment instant
1111 * and try to activate it. 1046 * and try to activate it.
1112 * 1047 *
1113 * Notice that it is important for the caller 1048 * Notice that it is important for the caller to know if the timer
1114 * actually started or not (i.e., the repleni 1049 * actually started or not (i.e., the replenishment instant is in
1115 * the future or in the past). 1050 * the future or in the past).
1116 */ 1051 */
1117 static int start_dl_timer(struct sched_dl_ent !! 1052 static int start_dl_timer(struct task_struct *p)
1118 { 1053 {
>> 1054 struct sched_dl_entity *dl_se = &p->dl;
1119 struct hrtimer *timer = &dl_se->dl_ti 1055 struct hrtimer *timer = &dl_se->dl_timer;
1120 struct dl_rq *dl_rq = dl_rq_of_se(dl_ !! 1056 struct rq *rq = task_rq(p);
1121 struct rq *rq = rq_of_dl_rq(dl_rq); <<
1122 ktime_t now, act; 1057 ktime_t now, act;
1123 s64 delta; 1058 s64 delta;
1124 1059
1125 lockdep_assert_rq_held(rq); 1060 lockdep_assert_rq_held(rq);
1126 1061
1127 /* 1062 /*
1128 * We want the timer to fire at the d 1063 * We want the timer to fire at the deadline, but considering
1129 * that it is actually coming from rq 1064 * that it is actually coming from rq->clock and not from
1130 * hrtimer's time base reading. 1065 * hrtimer's time base reading.
1131 * !! 1066 */
1132 * The deferred reservation will have !! 1067 act = ns_to_ktime(dl_next_period(dl_se));
1133 * (deadline - runtime). At that poin <<
1134 * if the current deadline can be use <<
1135 * required to avoid add too much pre <<
1136 * (current u > U). <<
1137 */ <<
1138 if (dl_se->dl_defer_armed) { <<
1139 WARN_ON_ONCE(!dl_se->dl_throt <<
1140 act = ns_to_ktime(dl_se->dead <<
1141 } else { <<
1142 /* act = deadline - rel-deadl <<
1143 act = ns_to_ktime(dl_next_per <<
1144 } <<
1145 <<
1146 now = hrtimer_cb_get_time(timer); 1068 now = hrtimer_cb_get_time(timer);
1147 delta = ktime_to_ns(now) - rq_clock(r 1069 delta = ktime_to_ns(now) - rq_clock(rq);
1148 act = ktime_add_ns(act, delta); 1070 act = ktime_add_ns(act, delta);
1149 1071
1150 /* 1072 /*
1151 * If the expiry time already passed, 1073 * If the expiry time already passed, e.g., because the value
1152 * chosen as the deadline is too smal 1074 * chosen as the deadline is too small, don't even try to
1153 * start the timer in the past! 1075 * start the timer in the past!
1154 */ 1076 */
1155 if (ktime_us_delta(act, now) < 0) 1077 if (ktime_us_delta(act, now) < 0)
1156 return 0; 1078 return 0;
1157 1079
1158 /* 1080 /*
1159 * !enqueued will guarantee another c 1081 * !enqueued will guarantee another callback; even if one is already in
1160 * progress. This ensures a balanced 1082 * progress. This ensures a balanced {get,put}_task_struct().
1161 * 1083 *
1162 * The race against __run_timer() cle 1084 * The race against __run_timer() clearing the enqueued state is
1163 * harmless because we're holding tas 1085 * harmless because we're holding task_rq()->lock, therefore the timer
1164 * expiring after we've done the chec 1086 * expiring after we've done the check will wait on its task_rq_lock()
1165 * and observe our state. 1087 * and observe our state.
1166 */ 1088 */
1167 if (!hrtimer_is_queued(timer)) { 1089 if (!hrtimer_is_queued(timer)) {
1168 if (!dl_server(dl_se)) !! 1090 get_task_struct(p);
1169 get_task_struct(dl_ta <<
1170 hrtimer_start(timer, act, HRT 1091 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1171 } 1092 }
1172 1093
1173 return 1; 1094 return 1;
1174 } 1095 }
1175 1096
1176 static void __push_dl_task(struct rq *rq, str <<
1177 { <<
1178 #ifdef CONFIG_SMP <<
1179 /* <<
1180 * Queueing this task back might have <<
1181 * to kick someone away. <<
1182 */ <<
1183 if (has_pushable_dl_tasks(rq)) { <<
1184 /* <<
1185 * Nothing relies on rq->lock <<
1186 * rq->lock. <<
1187 */ <<
1188 rq_unpin_lock(rq, rf); <<
1189 push_dl_task(rq); <<
1190 rq_repin_lock(rq, rf); <<
1191 } <<
1192 #endif <<
1193 } <<
1194 <<
1195 /* a defer timer will not be reset if the run <<
1196 static const u64 dl_server_min_res = 1 * NSEC <<
1197 <<
1198 static enum hrtimer_restart dl_server_timer(s <<
1199 { <<
1200 struct rq *rq = rq_of_dl_se(dl_se); <<
1201 u64 fw; <<
1202 <<
1203 scoped_guard (rq_lock, rq) { <<
1204 struct rq_flags *rf = &scope. <<
1205 <<
1206 if (!dl_se->dl_throttled || ! <<
1207 return HRTIMER_NOREST <<
1208 <<
1209 sched_clock_tick(); <<
1210 update_rq_clock(rq); <<
1211 <<
1212 if (!dl_se->dl_runtime) <<
1213 return HRTIMER_NOREST <<
1214 <<
1215 if (!dl_se->server_has_tasks( <<
1216 replenish_dl_entity(d <<
1217 return HRTIMER_NOREST <<
1218 } <<
1219 <<
1220 if (dl_se->dl_defer_armed) { <<
1221 /* <<
1222 * First check if the <<
1223 * If so, it is possi <<
1224 * of time. The dl_se <<
1225 * forwarding the tim <<
1226 */ <<
1227 if (dl_time_before(rq <<
1228 (d <<
1229 <<
1230 /* reset the <<
1231 fw = dl_se->d <<
1232 <<
1233 hrtimer_forwa <<
1234 return HRTIME <<
1235 } <<
1236 <<
1237 dl_se->dl_defer_runni <<
1238 } <<
1239 <<
1240 enqueue_dl_entity(dl_se, ENQU <<
1241 <<
1242 if (!dl_task(dl_se->rq->curr) <<
1243 resched_curr(rq); <<
1244 <<
1245 __push_dl_task(rq, rf); <<
1246 } <<
1247 <<
1248 return HRTIMER_NORESTART; <<
1249 } <<
1250 <<
1251 /* 1097 /*
1252 * This is the bandwidth enforcement timer ca 1098 * This is the bandwidth enforcement timer callback. If here, we know
1253 * a task is not on its dl_rq, since the fact 1099 * a task is not on its dl_rq, since the fact that the timer was running
1254 * means the task is throttled and needs a ru 1100 * means the task is throttled and needs a runtime replenishment.
1255 * 1101 *
1256 * However, what we actually do depends on th 1102 * However, what we actually do depends on the fact the task is active,
1257 * (it is on its rq) or has been removed from 1103 * (it is on its rq) or has been removed from there by a call to
1258 * dequeue_task_dl(). In the former case we m 1104 * dequeue_task_dl(). In the former case we must issue the runtime
1259 * replenishment and add the task back to the 1105 * replenishment and add the task back to the dl_rq; in the latter, we just
1260 * do nothing but clearing dl_throttled, so t 1106 * do nothing but clearing dl_throttled, so that runtime and deadline
1261 * updating (and the queueing back to dl_rq) 1107 * updating (and the queueing back to dl_rq) will be done by the
1262 * next call to enqueue_task_dl(). 1108 * next call to enqueue_task_dl().
1263 */ 1109 */
1264 static enum hrtimer_restart dl_task_timer(str 1110 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1265 { 1111 {
1266 struct sched_dl_entity *dl_se = conta 1112 struct sched_dl_entity *dl_se = container_of(timer,
1267 1113 struct sched_dl_entity,
1268 1114 dl_timer);
1269 struct task_struct *p; !! 1115 struct task_struct *p = dl_task_of(dl_se);
1270 struct rq_flags rf; 1116 struct rq_flags rf;
1271 struct rq *rq; 1117 struct rq *rq;
1272 1118
1273 if (dl_server(dl_se)) <<
1274 return dl_server_timer(timer, <<
1275 <<
1276 p = dl_task_of(dl_se); <<
1277 rq = task_rq_lock(p, &rf); 1119 rq = task_rq_lock(p, &rf);
1278 1120
1279 /* 1121 /*
1280 * The task might have changed its sc 1122 * The task might have changed its scheduling policy to something
1281 * different than SCHED_DEADLINE (thr 1123 * different than SCHED_DEADLINE (through switched_from_dl()).
1282 */ 1124 */
1283 if (!dl_task(p)) 1125 if (!dl_task(p))
1284 goto unlock; 1126 goto unlock;
1285 1127
1286 /* 1128 /*
1287 * The task might have been boosted b 1129 * The task might have been boosted by someone else and might be in the
1288 * boosting/deboosting path, its not 1130 * boosting/deboosting path, its not throttled.
1289 */ 1131 */
1290 if (is_dl_boosted(dl_se)) 1132 if (is_dl_boosted(dl_se))
1291 goto unlock; 1133 goto unlock;
1292 1134
1293 /* 1135 /*
1294 * Spurious timer due to start_dl_tim 1136 * Spurious timer due to start_dl_timer() race; or we already received
1295 * a replenishment from rt_mutex_setp 1137 * a replenishment from rt_mutex_setprio().
1296 */ 1138 */
1297 if (!dl_se->dl_throttled) 1139 if (!dl_se->dl_throttled)
1298 goto unlock; 1140 goto unlock;
1299 1141
1300 sched_clock_tick(); 1142 sched_clock_tick();
1301 update_rq_clock(rq); 1143 update_rq_clock(rq);
1302 1144
1303 /* 1145 /*
1304 * If the throttle happened during sc 1146 * If the throttle happened during sched-out; like:
1305 * 1147 *
1306 * schedule() 1148 * schedule()
1307 * deactivate_task() 1149 * deactivate_task()
1308 * dequeue_task_dl() 1150 * dequeue_task_dl()
1309 * update_curr_dl() 1151 * update_curr_dl()
1310 * start_dl_timer() 1152 * start_dl_timer()
1311 * __dequeue_task_dl() 1153 * __dequeue_task_dl()
1312 * prev->on_rq = 0; 1154 * prev->on_rq = 0;
1313 * 1155 *
1314 * We can be both throttled and !queu 1156 * We can be both throttled and !queued. Replenish the counter
1315 * but do not enqueue -- wait for our 1157 * but do not enqueue -- wait for our wakeup to do that.
1316 */ 1158 */
1317 if (!task_on_rq_queued(p)) { 1159 if (!task_on_rq_queued(p)) {
1318 replenish_dl_entity(dl_se); 1160 replenish_dl_entity(dl_se);
1319 goto unlock; 1161 goto unlock;
1320 } 1162 }
1321 1163
1322 #ifdef CONFIG_SMP 1164 #ifdef CONFIG_SMP
1323 if (unlikely(!rq->online)) { 1165 if (unlikely(!rq->online)) {
1324 /* 1166 /*
1325 * If the runqueue is no long 1167 * If the runqueue is no longer available, migrate the
1326 * task elsewhere. This neces 1168 * task elsewhere. This necessarily changes rq.
1327 */ 1169 */
1328 lockdep_unpin_lock(__rq_lockp 1170 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1329 rq = dl_task_offline_migratio 1171 rq = dl_task_offline_migration(rq, p);
1330 rf.cookie = lockdep_pin_lock( 1172 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1331 update_rq_clock(rq); 1173 update_rq_clock(rq);
1332 1174
1333 /* 1175 /*
1334 * Now that the task has been 1176 * Now that the task has been migrated to the new RQ and we
1335 * have that locked, proceed 1177 * have that locked, proceed as normal and enqueue the task
1336 * there. 1178 * there.
1337 */ 1179 */
1338 } 1180 }
1339 #endif 1181 #endif
1340 1182
1341 enqueue_task_dl(rq, p, ENQUEUE_REPLEN 1183 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1342 if (dl_task(rq->curr)) 1184 if (dl_task(rq->curr))
1343 wakeup_preempt_dl(rq, p, 0); !! 1185 check_preempt_curr_dl(rq, p, 0);
1344 else 1186 else
1345 resched_curr(rq); 1187 resched_curr(rq);
1346 1188
1347 __push_dl_task(rq, &rf); !! 1189 #ifdef CONFIG_SMP
>> 1190 /*
>> 1191 * Queueing this task back might have overloaded rq, check if we need
>> 1192 * to kick someone away.
>> 1193 */
>> 1194 if (has_pushable_dl_tasks(rq)) {
>> 1195 /*
>> 1196 * Nothing relies on rq->lock after this, so its safe to drop
>> 1197 * rq->lock.
>> 1198 */
>> 1199 rq_unpin_lock(rq, &rf);
>> 1200 push_dl_task(rq);
>> 1201 rq_repin_lock(rq, &rf);
>> 1202 }
>> 1203 #endif
1348 1204
1349 unlock: 1205 unlock:
1350 task_rq_unlock(rq, p, &rf); 1206 task_rq_unlock(rq, p, &rf);
1351 1207
1352 /* 1208 /*
1353 * This can free the task_struct, inc 1209 * This can free the task_struct, including this hrtimer, do not touch
1354 * anything related to that after thi 1210 * anything related to that after this.
1355 */ 1211 */
1356 put_task_struct(p); 1212 put_task_struct(p);
1357 1213
1358 return HRTIMER_NORESTART; 1214 return HRTIMER_NORESTART;
1359 } 1215 }
1360 1216
1361 static void init_dl_task_timer(struct sched_d !! 1217 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1362 { 1218 {
1363 struct hrtimer *timer = &dl_se->dl_ti 1219 struct hrtimer *timer = &dl_se->dl_timer;
1364 1220
1365 hrtimer_init(timer, CLOCK_MONOTONIC, 1221 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1366 timer->function = dl_task_timer; 1222 timer->function = dl_task_timer;
1367 } 1223 }
1368 1224
1369 /* 1225 /*
1370 * During the activation, CBS checks if it ca 1226 * During the activation, CBS checks if it can reuse the current task's
1371 * runtime and period. If the deadline of the 1227 * runtime and period. If the deadline of the task is in the past, CBS
1372 * cannot use the runtime, and so it replenis 1228 * cannot use the runtime, and so it replenishes the task. This rule
1373 * works fine for implicit deadline tasks (de 1229 * works fine for implicit deadline tasks (deadline == period), and the
1374 * CBS was designed for implicit deadline tas 1230 * CBS was designed for implicit deadline tasks. However, a task with
1375 * constrained deadline (deadline < period) m 1231 * constrained deadline (deadline < period) might be awakened after the
1376 * deadline, but before the next period. In t 1232 * deadline, but before the next period. In this case, replenishing the
1377 * task would allow it to run for runtime / d 1233 * task would allow it to run for runtime / deadline. As in this case
1378 * deadline < period, CBS enables a task to r 1234 * deadline < period, CBS enables a task to run for more than the
1379 * runtime / period. In a very loaded system, 1235 * runtime / period. In a very loaded system, this can cause a domino
1380 * effect, making other tasks miss their dead 1236 * effect, making other tasks miss their deadlines.
1381 * 1237 *
1382 * To avoid this problem, in the activation o 1238 * To avoid this problem, in the activation of a constrained deadline
1383 * task after the deadline but before the nex 1239 * task after the deadline but before the next period, throttle the
1384 * task and set the replenishing timer to the 1240 * task and set the replenishing timer to the begin of the next period,
1385 * unless it is boosted. 1241 * unless it is boosted.
1386 */ 1242 */
1387 static inline void dl_check_constrained_dl(st 1243 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1388 { 1244 {
1389 struct rq *rq = rq_of_dl_se(dl_se); !! 1245 struct task_struct *p = dl_task_of(dl_se);
>> 1246 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1390 1247
1391 if (dl_time_before(dl_se->deadline, r 1248 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1392 dl_time_before(rq_clock(rq), dl_n 1249 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1393 if (unlikely(is_dl_boosted(dl !! 1250 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(p)))
1394 return; 1251 return;
1395 dl_se->dl_throttled = 1; 1252 dl_se->dl_throttled = 1;
1396 if (dl_se->runtime > 0) 1253 if (dl_se->runtime > 0)
1397 dl_se->runtime = 0; 1254 dl_se->runtime = 0;
1398 } 1255 }
1399 } 1256 }
1400 1257
1401 static 1258 static
1402 int dl_runtime_exceeded(struct sched_dl_entit 1259 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1403 { 1260 {
1404 return (dl_se->runtime <= 0); 1261 return (dl_se->runtime <= 0);
1405 } 1262 }
1406 1263
1407 /* 1264 /*
1408 * This function implements the GRUB accounti !! 1265 * This function implements the GRUB accounting rule:
1409 * GRUB reclaiming algorithm, the runtime is !! 1266 * according to the GRUB reclaiming algorithm, the runtime is
1410 * but as "dq = -(max{u, (Umax - Uinact - Uex !! 1267 * not decreased as "dq = -dt", but as
>> 1268 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1411 * where u is the utilization of the task, Um 1269 * where u is the utilization of the task, Umax is the maximum reclaimable
1412 * utilization, Uinact is the (per-runqueue) 1270 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1413 * as the difference between the "total runqu 1271 * as the difference between the "total runqueue utilization" and the
1414 * "runqueue active utilization", and Uextra !! 1272 * runqueue active utilization, and Uextra is the (per runqueue) extra
1415 * reclaimable utilization. 1273 * reclaimable utilization.
1416 * Since rq->dl.running_bw and rq->dl.this_bw !! 1274 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1417 * by 2^BW_SHIFT, the result has to be shifte !! 1275 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1418 * Since rq->dl.bw_ratio contains 1 / Umax mu !! 1276 * BW_SHIFT.
1419 * is multiplied by rq->dl.bw_ratio and shift !! 1277 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT,
1420 * Since delta is a 64 bit variable, to have !! 1278 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1421 * larger than 2^(64 - 20 - 8), which is more !! 1279 * Since delta is a 64 bit variable, to have an overflow its value
1422 * not an issue here. !! 1280 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
>> 1281 * So, overflow is not an issue here.
1423 */ 1282 */
1424 static u64 grub_reclaim(u64 delta, struct rq 1283 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1425 { 1284 {
1426 u64 u_act; <<
1427 u64 u_inact = rq->dl.this_bw - rq->dl 1285 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
>> 1286 u64 u_act;
>> 1287 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1428 1288
1429 /* 1289 /*
1430 * Instead of computing max{u, (u_max !! 1290 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1431 * compare u_inact + u_extra with u_m !! 1291 * we compare u_inact + rq->dl.extra_bw with
1432 * can be larger than u_max. So, u_ma !! 1292 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1433 * negative leading to wrong results. !! 1293 * u_inact + rq->dl.extra_bw can be larger than
>> 1294 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
>> 1295 * leading to wrong results)
1434 */ 1296 */
1435 if (u_inact + rq->dl.extra_bw > rq->d !! 1297 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1436 u_act = dl_se->dl_bw; !! 1298 u_act = u_act_min;
1437 else 1299 else
1438 u_act = rq->dl.max_bw - u_ina !! 1300 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1439 1301
1440 u_act = (u_act * rq->dl.bw_ratio) >> <<
1441 return (delta * u_act) >> BW_SHIFT; 1302 return (delta * u_act) >> BW_SHIFT;
1442 } 1303 }
1443 1304
1444 s64 dl_scaled_delta_exec(struct rq *rq, struc !! 1305 /*
>> 1306 * Update the current task's runtime statistics (provided it is still
>> 1307 * a -deadline task and has not been removed from the dl_rq).
>> 1308 */
>> 1309 static void update_curr_dl(struct rq *rq)
1445 { 1310 {
1446 s64 scaled_delta_exec; !! 1311 struct task_struct *curr = rq->curr;
>> 1312 struct sched_dl_entity *dl_se = &curr->dl;
>> 1313 u64 delta_exec, scaled_delta_exec;
>> 1314 int cpu = cpu_of(rq);
>> 1315 u64 now;
>> 1316
>> 1317 if (!dl_task(curr) || !on_dl_rq(dl_se))
>> 1318 return;
>> 1319
>> 1320 /*
>> 1321 * Consumed budget is computed considering the time as
>> 1322 * observed by schedulable tasks (excluding time spent
>> 1323 * in hardirq context, etc.). Deadlines are instead
>> 1324 * computed using hard walltime. This seems to be the more
>> 1325 * natural solution, but the full ramifications of this
>> 1326 * approach need further study.
>> 1327 */
>> 1328 now = rq_clock_task(rq);
>> 1329 delta_exec = now - curr->se.exec_start;
>> 1330 if (unlikely((s64)delta_exec <= 0)) {
>> 1331 if (unlikely(dl_se->dl_yielded))
>> 1332 goto throttle;
>> 1333 return;
>> 1334 }
>> 1335
>> 1336 schedstat_set(curr->stats.exec_max,
>> 1337 max(curr->stats.exec_max, delta_exec));
>> 1338
>> 1339 trace_sched_stat_runtime(curr, delta_exec, 0);
>> 1340
>> 1341 update_current_exec_runtime(curr, now, delta_exec);
>> 1342
>> 1343 if (dl_entity_is_special(dl_se))
>> 1344 return;
1447 1345
1448 /* 1346 /*
1449 * For tasks that participate in GRUB 1347 * For tasks that participate in GRUB, we implement GRUB-PA: the
1450 * spare reclaimed bandwidth is used 1348 * spare reclaimed bandwidth is used to clock down frequency.
1451 * 1349 *
1452 * For the others, we still need to s 1350 * For the others, we still need to scale reservation parameters
1453 * according to current frequency and 1351 * according to current frequency and CPU maximum capacity.
1454 */ 1352 */
1455 if (unlikely(dl_se->flags & SCHED_FLA 1353 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1456 scaled_delta_exec = grub_recl !! 1354 scaled_delta_exec = grub_reclaim(delta_exec,
>> 1355 rq,
>> 1356 &curr->dl);
1457 } else { 1357 } else {
1458 int cpu = cpu_of(rq); <<
1459 unsigned long scale_freq = ar 1358 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1460 unsigned long scale_cpu = arc 1359 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1461 1360
1462 scaled_delta_exec = cap_scale 1361 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1463 scaled_delta_exec = cap_scale 1362 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1464 } 1363 }
1465 1364
1466 return scaled_delta_exec; <<
1467 } <<
1468 <<
1469 static inline void <<
1470 update_stats_dequeue_dl(struct dl_rq *dl_rq, <<
1471 int flags); <<
1472 static void update_curr_dl_se(struct rq *rq, <<
1473 { <<
1474 s64 scaled_delta_exec; <<
1475 <<
1476 if (unlikely(delta_exec <= 0)) { <<
1477 if (unlikely(dl_se->dl_yielde <<
1478 goto throttle; <<
1479 return; <<
1480 } <<
1481 <<
1482 if (dl_server(dl_se) && dl_se->dl_thr <<
1483 return; <<
1484 <<
1485 if (dl_entity_is_special(dl_se)) <<
1486 return; <<
1487 <<
1488 scaled_delta_exec = dl_scaled_delta_e <<
1489 <<
1490 dl_se->runtime -= scaled_delta_exec; 1365 dl_se->runtime -= scaled_delta_exec;
1491 1366
1492 /* <<
1493 * The fair server can consume its ru <<
1494 * running as regular CFS). <<
1495 * <<
1496 * If the server consumes its entire <<
1497 * is not required for the current pe <<
1498 * starting a new period, pushing the <<
1499 */ <<
1500 if (dl_se->dl_defer && dl_se->dl_thro <<
1501 /* <<
1502 * If the server was previous <<
1503 * took place, it this point <<
1504 * was able to get runtime in <<
1505 * state. <<
1506 */ <<
1507 dl_se->dl_defer_running = 0; <<
1508 <<
1509 hrtimer_try_to_cancel(&dl_se- <<
1510 <<
1511 replenish_dl_new_period(dl_se <<
1512 <<
1513 /* <<
1514 * Not being able to start th <<
1515 * be started for whatever re <<
1516 * and queue right away. Othe <<
1517 * to what enqueue_dl_entity( <<
1518 */ <<
1519 WARN_ON_ONCE(!start_dl_timer( <<
1520 <<
1521 return; <<
1522 } <<
1523 <<
1524 throttle: 1367 throttle:
1525 if (dl_runtime_exceeded(dl_se) || dl_ 1368 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1526 dl_se->dl_throttled = 1; 1369 dl_se->dl_throttled = 1;
1527 1370
1528 /* If requested, inform the u 1371 /* If requested, inform the user about runtime overruns. */
1529 if (dl_runtime_exceeded(dl_se 1372 if (dl_runtime_exceeded(dl_se) &&
1530 (dl_se->flags & SCHED_FLA 1373 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1531 dl_se->dl_overrun = 1 1374 dl_se->dl_overrun = 1;
1532 1375
1533 dequeue_dl_entity(dl_se, 0); !! 1376 __dequeue_task_dl(rq, curr, 0);
1534 if (!dl_server(dl_se)) { !! 1377 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(curr)))
1535 update_stats_dequeue_ !! 1378 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1536 dequeue_pushable_dl_t <<
1537 } <<
1538 <<
1539 if (unlikely(is_dl_boosted(dl <<
1540 if (dl_server(dl_se)) <<
1541 enqueue_dl_en <<
1542 else <<
1543 enqueue_task_ <<
1544 } <<
1545 1379
1546 if (!is_leftmost(dl_se, &rq-> !! 1380 if (!is_leftmost(curr, &rq->dl))
1547 resched_curr(rq); 1381 resched_curr(rq);
1548 } 1382 }
1549 1383
1550 /* 1384 /*
1551 * The fair server (sole dl_server) d <<
1552 * workload because it is running fai <<
1553 */ <<
1554 if (dl_se == &rq->fair_server) <<
1555 return; <<
1556 <<
1557 #ifdef CONFIG_RT_GROUP_SCHED <<
1558 /* <<
1559 * Because -- for now -- we share the 1385 * Because -- for now -- we share the rt bandwidth, we need to
1560 * account our runtime there too, oth 1386 * account our runtime there too, otherwise actual rt tasks
1561 * would be able to exceed the shared 1387 * would be able to exceed the shared quota.
1562 * 1388 *
1563 * Account to the root rt group for n 1389 * Account to the root rt group for now.
1564 * 1390 *
1565 * The solution we're working towards 1391 * The solution we're working towards is having the RT groups scheduled
1566 * using deadline servers -- however 1392 * using deadline servers -- however there's a few nasties to figure
1567 * out before that can happen. 1393 * out before that can happen.
1568 */ 1394 */
1569 if (rt_bandwidth_enabled()) { 1395 if (rt_bandwidth_enabled()) {
1570 struct rt_rq *rt_rq = &rq->rt 1396 struct rt_rq *rt_rq = &rq->rt;
1571 1397
1572 raw_spin_lock(&rt_rq->rt_runt 1398 raw_spin_lock(&rt_rq->rt_runtime_lock);
1573 /* 1399 /*
1574 * We'll let actual RT tasks 1400 * We'll let actual RT tasks worry about the overflow here, we
1575 * have our own CBS to keep u 1401 * have our own CBS to keep us inline; only account when RT
1576 * bandwidth is relevant. 1402 * bandwidth is relevant.
1577 */ 1403 */
1578 if (sched_rt_bandwidth_accoun 1404 if (sched_rt_bandwidth_account(rt_rq))
1579 rt_rq->rt_time += del 1405 rt_rq->rt_time += delta_exec;
1580 raw_spin_unlock(&rt_rq->rt_ru 1406 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1581 } 1407 }
1582 #endif <<
1583 } <<
1584 <<
1585 /* <<
1586 * In the non-defer mode, the idle time is no <<
1587 * server provides a guarantee. <<
1588 * <<
1589 * If the dl_server is in defer mode, the idl <<
1590 * as time available for the fair server, avo <<
1591 * rt scheduler that did not consumed that ti <<
1592 */ <<
1593 void dl_server_update_idle_time(struct rq *rq <<
1594 { <<
1595 s64 delta_exec, scaled_delta_exec; <<
1596 <<
1597 if (!rq->fair_server.dl_defer) <<
1598 return; <<
1599 <<
1600 /* no need to discount more */ <<
1601 if (rq->fair_server.runtime < 0) <<
1602 return; <<
1603 <<
1604 delta_exec = rq_clock_task(rq) - p->s <<
1605 if (delta_exec < 0) <<
1606 return; <<
1607 <<
1608 scaled_delta_exec = dl_scaled_delta_e <<
1609 <<
1610 rq->fair_server.runtime -= scaled_del <<
1611 <<
1612 if (rq->fair_server.runtime < 0) { <<
1613 rq->fair_server.dl_defer_runn <<
1614 rq->fair_server.runtime = 0; <<
1615 } <<
1616 <<
1617 p->se.exec_start = rq_clock_task(rq); <<
1618 } <<
1619 <<
1620 void dl_server_update(struct sched_dl_entity <<
1621 { <<
1622 /* 0 runtime = fair server disabled * <<
1623 if (dl_se->dl_runtime) <<
1624 update_curr_dl_se(dl_se->rq, <<
1625 } <<
1626 <<
1627 void dl_server_start(struct sched_dl_entity * <<
1628 { <<
1629 struct rq *rq = dl_se->rq; <<
1630 <<
1631 /* <<
1632 * XXX: the apply do not work fine at <<
1633 * fair server because things are not <<
1634 * this before getting generic. <<
1635 */ <<
1636 if (!dl_server(dl_se)) { <<
1637 u64 runtime = 50 * NSEC_PER_ <<
1638 u64 period = 1000 * NSEC_PER_ <<
1639 <<
1640 dl_server_apply_params(dl_se, <<
1641 <<
1642 dl_se->dl_server = 1; <<
1643 dl_se->dl_defer = 1; <<
1644 setup_new_dl_entity(dl_se); <<
1645 } <<
1646 <<
1647 if (!dl_se->dl_runtime) <<
1648 return; <<
1649 <<
1650 enqueue_dl_entity(dl_se, ENQUEUE_WAKE <<
1651 if (!dl_task(dl_se->rq->curr) || dl_e <<
1652 resched_curr(dl_se->rq); <<
1653 } <<
1654 <<
1655 void dl_server_stop(struct sched_dl_entity *d <<
1656 { <<
1657 if (!dl_se->dl_runtime) <<
1658 return; <<
1659 <<
1660 dequeue_dl_entity(dl_se, DEQUEUE_SLEE <<
1661 hrtimer_try_to_cancel(&dl_se->dl_time <<
1662 dl_se->dl_defer_armed = 0; <<
1663 dl_se->dl_throttled = 0; <<
1664 } <<
1665 <<
1666 void dl_server_init(struct sched_dl_entity *d <<
1667 dl_server_has_tasks_f has <<
1668 dl_server_pick_f pick_tas <<
1669 { <<
1670 dl_se->rq = rq; <<
1671 dl_se->server_has_tasks = has_tasks; <<
1672 dl_se->server_pick_task = pick_task; <<
1673 } <<
1674 <<
1675 void __dl_server_attach_root(struct sched_dl_ <<
1676 { <<
1677 u64 new_bw = dl_se->dl_bw; <<
1678 int cpu = cpu_of(rq); <<
1679 struct dl_bw *dl_b; <<
1680 <<
1681 dl_b = dl_bw_of(cpu_of(rq)); <<
1682 guard(raw_spinlock)(&dl_b->lock); <<
1683 <<
1684 if (!dl_bw_cpus(cpu)) <<
1685 return; <<
1686 <<
1687 __dl_add(dl_b, new_bw, dl_bw_cpus(cpu <<
1688 } <<
1689 <<
1690 int dl_server_apply_params(struct sched_dl_en <<
1691 { <<
1692 u64 old_bw = init ? 0 : to_ratio(dl_s <<
1693 u64 new_bw = to_ratio(period, runtime <<
1694 struct rq *rq = dl_se->rq; <<
1695 int cpu = cpu_of(rq); <<
1696 struct dl_bw *dl_b; <<
1697 unsigned long cap; <<
1698 int retval = 0; <<
1699 int cpus; <<
1700 <<
1701 dl_b = dl_bw_of(cpu); <<
1702 guard(raw_spinlock)(&dl_b->lock); <<
1703 <<
1704 cpus = dl_bw_cpus(cpu); <<
1705 cap = dl_bw_capacity(cpu); <<
1706 <<
1707 if (__dl_overflow(dl_b, cap, old_bw, <<
1708 return -EBUSY; <<
1709 <<
1710 if (init) { <<
1711 __add_rq_bw(new_bw, &rq->dl); <<
1712 __dl_add(dl_b, new_bw, cpus); <<
1713 } else { <<
1714 __dl_sub(dl_b, dl_se->dl_bw, <<
1715 __dl_add(dl_b, new_bw, cpus); <<
1716 <<
1717 dl_rq_change_utilization(rq, <<
1718 } <<
1719 <<
1720 dl_se->dl_runtime = runtime; <<
1721 dl_se->dl_deadline = period; <<
1722 dl_se->dl_period = period; <<
1723 <<
1724 dl_se->runtime = 0; <<
1725 dl_se->deadline = 0; <<
1726 <<
1727 dl_se->dl_bw = to_ratio(dl_se->dl_per <<
1728 dl_se->dl_density = to_ratio(dl_se->d <<
1729 <<
1730 return retval; <<
1731 } <<
1732 <<
1733 /* <<
1734 * Update the current task's runtime statisti <<
1735 * a -deadline task and has not been removed <<
1736 */ <<
1737 static void update_curr_dl(struct rq *rq) <<
1738 { <<
1739 struct task_struct *curr = rq->curr; <<
1740 struct sched_dl_entity *dl_se = &curr <<
1741 s64 delta_exec; <<
1742 <<
1743 if (!dl_task(curr) || !on_dl_rq(dl_se <<
1744 return; <<
1745 <<
1746 /* <<
1747 * Consumed budget is computed consid <<
1748 * observed by schedulable tasks (exc <<
1749 * in hardirq context, etc.). Deadlin <<
1750 * computed using hard walltime. This <<
1751 * natural solution, but the full ram <<
1752 * approach need further study. <<
1753 */ <<
1754 delta_exec = update_curr_common(rq); <<
1755 update_curr_dl_se(rq, dl_se, delta_ex <<
1756 } 1408 }
1757 1409
1758 static enum hrtimer_restart inactive_task_tim 1410 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1759 { 1411 {
1760 struct sched_dl_entity *dl_se = conta 1412 struct sched_dl_entity *dl_se = container_of(timer,
1761 1413 struct sched_dl_entity,
1762 1414 inactive_timer);
1763 struct task_struct *p = NULL; !! 1415 struct task_struct *p = dl_task_of(dl_se);
1764 struct rq_flags rf; 1416 struct rq_flags rf;
1765 struct rq *rq; 1417 struct rq *rq;
1766 1418
1767 if (!dl_server(dl_se)) { !! 1419 rq = task_rq_lock(p, &rf);
1768 p = dl_task_of(dl_se); <<
1769 rq = task_rq_lock(p, &rf); <<
1770 } else { <<
1771 rq = dl_se->rq; <<
1772 rq_lock(rq, &rf); <<
1773 } <<
1774 1420
1775 sched_clock_tick(); 1421 sched_clock_tick();
1776 update_rq_clock(rq); 1422 update_rq_clock(rq);
1777 1423
1778 if (dl_server(dl_se)) <<
1779 goto no_task; <<
1780 <<
1781 if (!dl_task(p) || READ_ONCE(p->__sta 1424 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1782 struct dl_bw *dl_b = dl_bw_of 1425 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1783 1426
1784 if (READ_ONCE(p->__state) == 1427 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1785 sub_running_bw(&p->dl 1428 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1786 sub_rq_bw(&p->dl, dl_ 1429 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1787 dl_se->dl_non_contend 1430 dl_se->dl_non_contending = 0;
1788 } 1431 }
1789 1432
1790 raw_spin_lock(&dl_b->lock); 1433 raw_spin_lock(&dl_b->lock);
1791 __dl_sub(dl_b, p->dl.dl_bw, d 1434 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1792 raw_spin_unlock(&dl_b->lock); 1435 raw_spin_unlock(&dl_b->lock);
1793 __dl_clear_params(dl_se); !! 1436 __dl_clear_params(p);
1794 1437
1795 goto unlock; 1438 goto unlock;
1796 } 1439 }
1797 <<
1798 no_task: <<
1799 if (dl_se->dl_non_contending == 0) 1440 if (dl_se->dl_non_contending == 0)
1800 goto unlock; 1441 goto unlock;
1801 1442
1802 sub_running_bw(dl_se, &rq->dl); 1443 sub_running_bw(dl_se, &rq->dl);
1803 dl_se->dl_non_contending = 0; 1444 dl_se->dl_non_contending = 0;
1804 unlock: 1445 unlock:
1805 !! 1446 task_rq_unlock(rq, p, &rf);
1806 if (!dl_server(dl_se)) { !! 1447 put_task_struct(p);
1807 task_rq_unlock(rq, p, &rf); <<
1808 put_task_struct(p); <<
1809 } else { <<
1810 rq_unlock(rq, &rf); <<
1811 } <<
1812 1448
1813 return HRTIMER_NORESTART; 1449 return HRTIMER_NORESTART;
1814 } 1450 }
1815 1451
1816 static void init_dl_inactive_task_timer(struc !! 1452 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1817 { 1453 {
1818 struct hrtimer *timer = &dl_se->inact 1454 struct hrtimer *timer = &dl_se->inactive_timer;
1819 1455
1820 hrtimer_init(timer, CLOCK_MONOTONIC, 1456 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1821 timer->function = inactive_task_timer 1457 timer->function = inactive_task_timer;
1822 } 1458 }
1823 1459
1824 #define __node_2_dle(node) \ 1460 #define __node_2_dle(node) \
1825 rb_entry((node), struct sched_dl_enti 1461 rb_entry((node), struct sched_dl_entity, rb_node)
1826 1462
1827 #ifdef CONFIG_SMP 1463 #ifdef CONFIG_SMP
1828 1464
1829 static void inc_dl_deadline(struct dl_rq *dl_ 1465 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1830 { 1466 {
1831 struct rq *rq = rq_of_dl_rq(dl_rq); 1467 struct rq *rq = rq_of_dl_rq(dl_rq);
1832 1468
1833 if (dl_rq->earliest_dl.curr == 0 || 1469 if (dl_rq->earliest_dl.curr == 0 ||
1834 dl_time_before(deadline, dl_rq->e 1470 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1835 if (dl_rq->earliest_dl.curr = 1471 if (dl_rq->earliest_dl.curr == 0)
1836 cpupri_set(&rq->rd->c 1472 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1837 dl_rq->earliest_dl.curr = dea 1473 dl_rq->earliest_dl.curr = deadline;
1838 cpudl_set(&rq->rd->cpudl, rq- 1474 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1839 } 1475 }
1840 } 1476 }
1841 1477
1842 static void dec_dl_deadline(struct dl_rq *dl_ 1478 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1843 { 1479 {
1844 struct rq *rq = rq_of_dl_rq(dl_rq); 1480 struct rq *rq = rq_of_dl_rq(dl_rq);
1845 1481
1846 /* 1482 /*
1847 * Since we may have removed our earl 1483 * Since we may have removed our earliest (and/or next earliest)
1848 * task we must recompute them. 1484 * task we must recompute them.
1849 */ 1485 */
1850 if (!dl_rq->dl_nr_running) { 1486 if (!dl_rq->dl_nr_running) {
1851 dl_rq->earliest_dl.curr = 0; 1487 dl_rq->earliest_dl.curr = 0;
1852 dl_rq->earliest_dl.next = 0; 1488 dl_rq->earliest_dl.next = 0;
1853 cpudl_clear(&rq->rd->cpudl, r 1489 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1854 cpupri_set(&rq->rd->cpupri, r 1490 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1855 } else { 1491 } else {
1856 struct rb_node *leftmost = rb 1492 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1857 struct sched_dl_entity *entry 1493 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1858 1494
1859 dl_rq->earliest_dl.curr = ent 1495 dl_rq->earliest_dl.curr = entry->deadline;
1860 cpudl_set(&rq->rd->cpudl, rq- 1496 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1861 } 1497 }
1862 } 1498 }
1863 1499
1864 #else 1500 #else
1865 1501
1866 static inline void inc_dl_deadline(struct dl_ 1502 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1867 static inline void dec_dl_deadline(struct dl_ 1503 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1868 1504
1869 #endif /* CONFIG_SMP */ 1505 #endif /* CONFIG_SMP */
1870 1506
1871 static inline 1507 static inline
1872 void inc_dl_tasks(struct sched_dl_entity *dl_ 1508 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1873 { 1509 {
>> 1510 int prio = dl_task_of(dl_se)->prio;
1874 u64 deadline = dl_se->deadline; 1511 u64 deadline = dl_se->deadline;
1875 1512
>> 1513 WARN_ON(!dl_prio(prio));
1876 dl_rq->dl_nr_running++; 1514 dl_rq->dl_nr_running++;
1877 add_nr_running(rq_of_dl_rq(dl_rq), 1) 1515 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1878 1516
1879 inc_dl_deadline(dl_rq, deadline); 1517 inc_dl_deadline(dl_rq, deadline);
>> 1518 inc_dl_migration(dl_se, dl_rq);
1880 } 1519 }
1881 1520
1882 static inline 1521 static inline
1883 void dec_dl_tasks(struct sched_dl_entity *dl_ 1522 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1884 { 1523 {
>> 1524 int prio = dl_task_of(dl_se)->prio;
>> 1525
>> 1526 WARN_ON(!dl_prio(prio));
1885 WARN_ON(!dl_rq->dl_nr_running); 1527 WARN_ON(!dl_rq->dl_nr_running);
1886 dl_rq->dl_nr_running--; 1528 dl_rq->dl_nr_running--;
1887 sub_nr_running(rq_of_dl_rq(dl_rq), 1) 1529 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1888 1530
1889 dec_dl_deadline(dl_rq, dl_se->deadlin 1531 dec_dl_deadline(dl_rq, dl_se->deadline);
>> 1532 dec_dl_migration(dl_se, dl_rq);
1890 } 1533 }
1891 1534
1892 static inline bool __dl_less(struct rb_node * 1535 static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1893 { 1536 {
1894 return dl_time_before(__node_2_dle(a) 1537 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1895 } 1538 }
1896 1539
1897 static __always_inline struct sched_statistic !! 1540 static inline struct sched_statistics *
1898 __schedstats_from_dl_se(struct sched_dl_entit 1541 __schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1899 { 1542 {
1900 if (!schedstat_enabled()) <<
1901 return NULL; <<
1902 <<
1903 if (dl_server(dl_se)) <<
1904 return NULL; <<
1905 <<
1906 return &dl_task_of(dl_se)->stats; 1543 return &dl_task_of(dl_se)->stats;
1907 } 1544 }
1908 1545
1909 static inline void 1546 static inline void
1910 update_stats_wait_start_dl(struct dl_rq *dl_r 1547 update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1911 { 1548 {
1912 struct sched_statistics *stats = __sc !! 1549 struct sched_statistics *stats;
1913 if (stats) !! 1550
1914 __update_stats_wait_start(rq_ !! 1551 if (!schedstat_enabled())
>> 1552 return;
>> 1553
>> 1554 stats = __schedstats_from_dl_se(dl_se);
>> 1555 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1915 } 1556 }
1916 1557
1917 static inline void 1558 static inline void
1918 update_stats_wait_end_dl(struct dl_rq *dl_rq, 1559 update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1919 { 1560 {
1920 struct sched_statistics *stats = __sc !! 1561 struct sched_statistics *stats;
1921 if (stats) !! 1562
1922 __update_stats_wait_end(rq_of !! 1563 if (!schedstat_enabled())
>> 1564 return;
>> 1565
>> 1566 stats = __schedstats_from_dl_se(dl_se);
>> 1567 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1923 } 1568 }
1924 1569
1925 static inline void 1570 static inline void
1926 update_stats_enqueue_sleeper_dl(struct dl_rq 1571 update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1927 { 1572 {
1928 struct sched_statistics *stats = __sc !! 1573 struct sched_statistics *stats;
1929 if (stats) !! 1574
1930 __update_stats_enqueue_sleepe !! 1575 if (!schedstat_enabled())
>> 1576 return;
>> 1577
>> 1578 stats = __schedstats_from_dl_se(dl_se);
>> 1579 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1931 } 1580 }
1932 1581
1933 static inline void 1582 static inline void
1934 update_stats_enqueue_dl(struct dl_rq *dl_rq, 1583 update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1935 int flags) 1584 int flags)
1936 { 1585 {
1937 if (!schedstat_enabled()) 1586 if (!schedstat_enabled())
1938 return; 1587 return;
1939 1588
1940 if (flags & ENQUEUE_WAKEUP) 1589 if (flags & ENQUEUE_WAKEUP)
1941 update_stats_enqueue_sleeper_ 1590 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1942 } 1591 }
1943 1592
1944 static inline void 1593 static inline void
1945 update_stats_dequeue_dl(struct dl_rq *dl_rq, 1594 update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1946 int flags) 1595 int flags)
1947 { 1596 {
1948 struct task_struct *p = dl_task_of(dl 1597 struct task_struct *p = dl_task_of(dl_se);
1949 1598
1950 if (!schedstat_enabled()) 1599 if (!schedstat_enabled())
1951 return; 1600 return;
1952 1601
1953 if ((flags & DEQUEUE_SLEEP)) { 1602 if ((flags & DEQUEUE_SLEEP)) {
1954 unsigned int state; 1603 unsigned int state;
1955 1604
1956 state = READ_ONCE(p->__state) 1605 state = READ_ONCE(p->__state);
1957 if (state & TASK_INTERRUPTIBL 1606 if (state & TASK_INTERRUPTIBLE)
1958 __schedstat_set(p->st 1607 __schedstat_set(p->stats.sleep_start,
1959 rq_cl 1608 rq_clock(rq_of_dl_rq(dl_rq)));
1960 1609
1961 if (state & TASK_UNINTERRUPTI 1610 if (state & TASK_UNINTERRUPTIBLE)
1962 __schedstat_set(p->st 1611 __schedstat_set(p->stats.block_start,
1963 rq_cl 1612 rq_clock(rq_of_dl_rq(dl_rq)));
1964 } 1613 }
1965 } 1614 }
1966 1615
1967 static void __enqueue_dl_entity(struct sched_ 1616 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1968 { 1617 {
1969 struct dl_rq *dl_rq = dl_rq_of_se(dl_ 1618 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1970 1619
1971 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->r 1620 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node));
1972 1621
1973 rb_add_cached(&dl_se->rb_node, &dl_rq 1622 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1974 1623
1975 inc_dl_tasks(dl_se, dl_rq); 1624 inc_dl_tasks(dl_se, dl_rq);
1976 } 1625 }
1977 1626
1978 static void __dequeue_dl_entity(struct sched_ 1627 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1979 { 1628 {
1980 struct dl_rq *dl_rq = dl_rq_of_se(dl_ 1629 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1981 1630
1982 if (RB_EMPTY_NODE(&dl_se->rb_node)) 1631 if (RB_EMPTY_NODE(&dl_se->rb_node))
1983 return; 1632 return;
1984 1633
1985 rb_erase_cached(&dl_se->rb_node, &dl_ 1634 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1986 1635
1987 RB_CLEAR_NODE(&dl_se->rb_node); 1636 RB_CLEAR_NODE(&dl_se->rb_node);
1988 1637
1989 dec_dl_tasks(dl_se, dl_rq); 1638 dec_dl_tasks(dl_se, dl_rq);
1990 } 1639 }
1991 1640
1992 static void 1641 static void
1993 enqueue_dl_entity(struct sched_dl_entity *dl_ 1642 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1994 { 1643 {
1995 WARN_ON_ONCE(on_dl_rq(dl_se)); 1644 WARN_ON_ONCE(on_dl_rq(dl_se));
1996 1645
1997 update_stats_enqueue_dl(dl_rq_of_se(d 1646 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
1998 1647
1999 /* 1648 /*
2000 * Check if a constrained deadline ta <<
2001 * after the deadline but before the <<
2002 * If that is the case, the task will <<
2003 * the replenishment timer will be se <<
2004 */ <<
2005 if (!dl_se->dl_throttled && !dl_is_im <<
2006 dl_check_constrained_dl(dl_se <<
2007 <<
2008 if (flags & (ENQUEUE_RESTORE|ENQUEUE_ <<
2009 struct dl_rq *dl_rq = dl_rq_o <<
2010 <<
2011 add_rq_bw(dl_se, dl_rq); <<
2012 add_running_bw(dl_se, dl_rq); <<
2013 } <<
2014 <<
2015 /* <<
2016 * If p is throttled, we do not enque <<
2017 * its budget it needs a replenishmen <<
2018 * its rq, the bandwidth timer callba <<
2019 * run yet) will take care of this. <<
2020 * However, the active utilization do <<
2021 * that the task is on the runqueue o <<
2022 * task's state - in GRUB parlance, " <<
2023 * In other words, even if a task is <<
2024 * be counted in the active utilizati <<
2025 * add_running_bw(). <<
2026 */ <<
2027 if (!dl_se->dl_defer && dl_se->dl_thr <<
2028 if (flags & ENQUEUE_WAKEUP) <<
2029 task_contending(dl_se <<
2030 <<
2031 return; <<
2032 } <<
2033 <<
2034 /* <<
2035 * If this is a wakeup or a new insta 1649 * If this is a wakeup or a new instance, the scheduling
2036 * parameters of the task might need 1650 * parameters of the task might need updating. Otherwise,
2037 * we want a replenishment of its run 1651 * we want a replenishment of its runtime.
2038 */ 1652 */
2039 if (flags & ENQUEUE_WAKEUP) { 1653 if (flags & ENQUEUE_WAKEUP) {
2040 task_contending(dl_se, flags) 1654 task_contending(dl_se, flags);
2041 update_dl_entity(dl_se); 1655 update_dl_entity(dl_se);
2042 } else if (flags & ENQUEUE_REPLENISH) 1656 } else if (flags & ENQUEUE_REPLENISH) {
2043 replenish_dl_entity(dl_se); 1657 replenish_dl_entity(dl_se);
2044 } else if ((flags & ENQUEUE_RESTORE) 1658 } else if ((flags & ENQUEUE_RESTORE) &&
2045 dl_time_before(dl_se->dead !! 1659 dl_time_before(dl_se->deadline,
>> 1660 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
2046 setup_new_dl_entity(dl_se); 1661 setup_new_dl_entity(dl_se);
2047 } 1662 }
2048 1663
2049 /* <<
2050 * If the reservation is still thrott <<
2051 * deferred task and still got to wai <<
2052 */ <<
2053 if (dl_se->dl_throttled && start_dl_t <<
2054 return; <<
2055 <<
2056 /* <<
2057 * We're about to enqueue, make sure <<
2058 * In case the timer was not started, <<
2059 * has passed, mark as not throttled <<
2060 * Also cancel earlier timers, since <<
2061 */ <<
2062 if (dl_se->dl_throttled) { <<
2063 hrtimer_try_to_cancel(&dl_se- <<
2064 dl_se->dl_defer_armed = 0; <<
2065 dl_se->dl_throttled = 0; <<
2066 } <<
2067 <<
2068 __enqueue_dl_entity(dl_se); 1664 __enqueue_dl_entity(dl_se);
2069 } 1665 }
2070 1666
2071 static void dequeue_dl_entity(struct sched_dl !! 1667 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
2072 { 1668 {
2073 __dequeue_dl_entity(dl_se); 1669 __dequeue_dl_entity(dl_se);
2074 <<
2075 if (flags & (DEQUEUE_SAVE|DEQUEUE_MIG <<
2076 struct dl_rq *dl_rq = dl_rq_o <<
2077 <<
2078 sub_running_bw(dl_se, dl_rq); <<
2079 sub_rq_bw(dl_se, dl_rq); <<
2080 } <<
2081 <<
2082 /* <<
2083 * This check allows to start the ina <<
2084 * decrease the active utilization, i <<
2085 * when the task blocks and when it i <<
2086 * (p->state == TASK_DEAD). We can ha <<
2087 * way, because from GRUB's point of <<
2088 * (the task moves from "active conte <<
2089 * or "inactive") <<
2090 */ <<
2091 if (flags & DEQUEUE_SLEEP) <<
2092 task_non_contending(dl_se); <<
2093 } 1670 }
2094 1671
2095 static void enqueue_task_dl(struct rq *rq, st 1672 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2096 { 1673 {
2097 if (is_dl_boosted(&p->dl)) { 1674 if (is_dl_boosted(&p->dl)) {
2098 /* 1675 /*
2099 * Because of delays in the d 1676 * Because of delays in the detection of the overrun of a
2100 * thread's runtime, it might 1677 * thread's runtime, it might be the case that a thread
2101 * goes to sleep in a rt mute 1678 * goes to sleep in a rt mutex with negative runtime. As
2102 * a consequence, the thread 1679 * a consequence, the thread will be throttled.
2103 * 1680 *
2104 * While waiting for the mute 1681 * While waiting for the mutex, this thread can also be
2105 * boosted via PI, resulting 1682 * boosted via PI, resulting in a thread that is throttled
2106 * and boosted at the same ti 1683 * and boosted at the same time.
2107 * 1684 *
2108 * In this case, the boost ov 1685 * In this case, the boost overrides the throttle.
2109 */ 1686 */
2110 if (p->dl.dl_throttled) { 1687 if (p->dl.dl_throttled) {
2111 /* 1688 /*
2112 * The replenish time 1689 * The replenish timer needs to be canceled. No
2113 * problem if it fire 1690 * problem if it fires concurrently: boosted threads
2114 * are ignored in dl_ 1691 * are ignored in dl_task_timer().
2115 * <<
2116 * If the timer callb <<
2117 * it will eventually <<
2118 */ 1692 */
2119 if (hrtimer_try_to_ca !! 1693 hrtimer_try_to_cancel(&p->dl.dl_timer);
2120 !dl_server(&p->dl <<
2121 put_task_stru <<
2122 p->dl.dl_throttled = 1694 p->dl.dl_throttled = 0;
2123 } 1695 }
2124 } else if (!dl_prio(p->normal_prio)) 1696 } else if (!dl_prio(p->normal_prio)) {
2125 /* 1697 /*
2126 * Special case in which we h 1698 * Special case in which we have a !SCHED_DEADLINE task that is going
2127 * to be deboosted, but excee 1699 * to be deboosted, but exceeds its runtime while doing so. No point in
2128 * replenishing it, as it's g 1700 * replenishing it, as it's going to return back to its original
2129 * scheduling class after thi 1701 * scheduling class after this. If it has been throttled, we need to
2130 * clear the flag, otherwise 1702 * clear the flag, otherwise the task may wake up as throttled after
2131 * being boosted again with n 1703 * being boosted again with no means to replenish the runtime and clear
2132 * the throttle. 1704 * the throttle.
2133 */ 1705 */
2134 p->dl.dl_throttled = 0; 1706 p->dl.dl_throttled = 0;
2135 if (!(flags & ENQUEUE_REPLENI 1707 if (!(flags & ENQUEUE_REPLENISH))
2136 printk_deferred_once( 1708 printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
2137 1709 task_pid_nr(p));
2138 1710
2139 return; 1711 return;
2140 } 1712 }
2141 1713
2142 check_schedstat_required(); !! 1714 /*
2143 update_stats_wait_start_dl(dl_rq_of_s !! 1715 * Check if a constrained deadline task was activated
>> 1716 * after the deadline but before the next period.
>> 1717 * If that is the case, the task will be throttled and
>> 1718 * the replenishment timer will be set to the next period.
>> 1719 */
>> 1720 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
>> 1721 dl_check_constrained_dl(&p->dl);
2144 1722
2145 if (p->on_rq == TASK_ON_RQ_MIGRATING) !! 1723 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
2146 flags |= ENQUEUE_MIGRATING; !! 1724 add_rq_bw(&p->dl, &rq->dl);
>> 1725 add_running_bw(&p->dl, &rq->dl);
>> 1726 }
2147 1727
2148 enqueue_dl_entity(&p->dl, flags); !! 1728 /*
>> 1729 * If p is throttled, we do not enqueue it. In fact, if it exhausted
>> 1730 * its budget it needs a replenishment and, since it now is on
>> 1731 * its rq, the bandwidth timer callback (which clearly has not
>> 1732 * run yet) will take care of this.
>> 1733 * However, the active utilization does not depend on the fact
>> 1734 * that the task is on the runqueue or not (but depends on the
>> 1735 * task's state - in GRUB parlance, "inactive" vs "active contending").
>> 1736 * In other words, even if a task is throttled its utilization must
>> 1737 * be counted in the active utilization; hence, we need to call
>> 1738 * add_running_bw().
>> 1739 */
>> 1740 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
>> 1741 if (flags & ENQUEUE_WAKEUP)
>> 1742 task_contending(&p->dl, flags);
2149 1743
2150 if (dl_server(&p->dl)) <<
2151 return; 1744 return;
>> 1745 }
>> 1746
>> 1747 check_schedstat_required();
>> 1748 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
>> 1749
>> 1750 enqueue_dl_entity(&p->dl, flags);
2152 1751
2153 if (!task_current(rq, p) && !p->dl.dl !! 1752 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
2154 enqueue_pushable_dl_task(rq, 1753 enqueue_pushable_dl_task(rq, p);
2155 } 1754 }
2156 1755
2157 static bool dequeue_task_dl(struct rq *rq, st !! 1756 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2158 { 1757 {
2159 update_curr_dl(rq); !! 1758 update_stats_dequeue_dl(&rq->dl, &p->dl, flags);
>> 1759 dequeue_dl_entity(&p->dl);
>> 1760 dequeue_pushable_dl_task(rq, p);
>> 1761 }
2160 1762
2161 if (p->on_rq == TASK_ON_RQ_MIGRATING) !! 1763 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2162 flags |= DEQUEUE_MIGRATING; !! 1764 {
>> 1765 update_curr_dl(rq);
>> 1766 __dequeue_task_dl(rq, p, flags);
2163 1767
2164 dequeue_dl_entity(&p->dl, flags); !! 1768 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
2165 if (!p->dl.dl_throttled && !dl_server !! 1769 sub_running_bw(&p->dl, &rq->dl);
2166 dequeue_pushable_dl_task(rq, !! 1770 sub_rq_bw(&p->dl, &rq->dl);
>> 1771 }
2167 1772
2168 return true; !! 1773 /*
>> 1774 * This check allows to start the inactive timer (or to immediately
>> 1775 * decrease the active utilization, if needed) in two cases:
>> 1776 * when the task blocks and when it is terminating
>> 1777 * (p->state == TASK_DEAD). We can handle the two cases in the same
>> 1778 * way, because from GRUB's point of view the same thing is happening
>> 1779 * (the task moves from "active contending" to "active non contending"
>> 1780 * or "inactive")
>> 1781 */
>> 1782 if (flags & DEQUEUE_SLEEP)
>> 1783 task_non_contending(p);
2169 } 1784 }
2170 1785
2171 /* 1786 /*
2172 * Yield task semantic for -deadline tasks is 1787 * Yield task semantic for -deadline tasks is:
2173 * 1788 *
2174 * get off from the CPU until our next inst 1789 * get off from the CPU until our next instance, with
2175 * a new runtime. This is of little use now 1790 * a new runtime. This is of little use now, since we
2176 * don't have a bandwidth reclaiming mechan 1791 * don't have a bandwidth reclaiming mechanism. Anyway,
2177 * bandwidth reclaiming is planned for the 1792 * bandwidth reclaiming is planned for the future, and
2178 * yield_task_dl will indicate that some sp 1793 * yield_task_dl will indicate that some spare budget
2179 * is available for other task instances to 1794 * is available for other task instances to use it.
2180 */ 1795 */
2181 static void yield_task_dl(struct rq *rq) 1796 static void yield_task_dl(struct rq *rq)
2182 { 1797 {
2183 /* 1798 /*
2184 * We make the task go to sleep until 1799 * We make the task go to sleep until its current deadline by
2185 * forcing its runtime to zero. This 1800 * forcing its runtime to zero. This way, update_curr_dl() stops
2186 * it and the bandwidth timer will wa 1801 * it and the bandwidth timer will wake it up and will give it
2187 * new scheduling parameters (thanks 1802 * new scheduling parameters (thanks to dl_yielded=1).
2188 */ 1803 */
2189 rq->curr->dl.dl_yielded = 1; 1804 rq->curr->dl.dl_yielded = 1;
2190 1805
2191 update_rq_clock(rq); 1806 update_rq_clock(rq);
2192 update_curr_dl(rq); 1807 update_curr_dl(rq);
2193 /* 1808 /*
2194 * Tell update_rq_clock() that we've 1809 * Tell update_rq_clock() that we've just updated,
2195 * so we don't do microscopic update 1810 * so we don't do microscopic update in schedule()
2196 * and double the fastpath cost. 1811 * and double the fastpath cost.
2197 */ 1812 */
2198 rq_clock_skip_update(rq); 1813 rq_clock_skip_update(rq);
2199 } 1814 }
2200 1815
2201 #ifdef CONFIG_SMP 1816 #ifdef CONFIG_SMP
2202 1817
2203 static inline bool dl_task_is_earliest_deadli 1818 static inline bool dl_task_is_earliest_deadline(struct task_struct *p,
2204 1819 struct rq *rq)
2205 { 1820 {
2206 return (!rq->dl.dl_nr_running || 1821 return (!rq->dl.dl_nr_running ||
2207 dl_time_before(p->dl.deadline 1822 dl_time_before(p->dl.deadline,
2208 rq->dl.earlies 1823 rq->dl.earliest_dl.curr));
2209 } 1824 }
2210 1825
2211 static int find_later_rq(struct task_struct * 1826 static int find_later_rq(struct task_struct *task);
2212 1827
2213 static int 1828 static int
2214 select_task_rq_dl(struct task_struct *p, int 1829 select_task_rq_dl(struct task_struct *p, int cpu, int flags)
2215 { 1830 {
2216 struct task_struct *curr; 1831 struct task_struct *curr;
2217 bool select_rq; 1832 bool select_rq;
2218 struct rq *rq; 1833 struct rq *rq;
2219 1834
2220 if (!(flags & WF_TTWU)) 1835 if (!(flags & WF_TTWU))
2221 goto out; 1836 goto out;
2222 1837
2223 rq = cpu_rq(cpu); 1838 rq = cpu_rq(cpu);
2224 1839
2225 rcu_read_lock(); 1840 rcu_read_lock();
2226 curr = READ_ONCE(rq->curr); /* unlock 1841 curr = READ_ONCE(rq->curr); /* unlocked access */
2227 1842
2228 /* 1843 /*
2229 * If we are dealing with a -deadline 1844 * If we are dealing with a -deadline task, we must
2230 * decide where to wake it up. 1845 * decide where to wake it up.
2231 * If it has a later deadline and the 1846 * If it has a later deadline and the current task
2232 * on this rq can't move (provided th 1847 * on this rq can't move (provided the waking task
2233 * can!) we prefer to send it somewhe 1848 * can!) we prefer to send it somewhere else. On the
2234 * other hand, if it has a shorter de 1849 * other hand, if it has a shorter deadline, we
2235 * try to make it stay here, it might 1850 * try to make it stay here, it might be important.
2236 */ 1851 */
2237 select_rq = unlikely(dl_task(curr)) & 1852 select_rq = unlikely(dl_task(curr)) &&
2238 (curr->nr_cpus_allowed < 1853 (curr->nr_cpus_allowed < 2 ||
2239 !dl_entity_preempt(&p->d 1854 !dl_entity_preempt(&p->dl, &curr->dl)) &&
2240 p->nr_cpus_allowed > 1; 1855 p->nr_cpus_allowed > 1;
2241 1856
2242 /* 1857 /*
2243 * Take the capacity of the CPU into 1858 * Take the capacity of the CPU into account to
2244 * ensure it fits the requirement of 1859 * ensure it fits the requirement of the task.
2245 */ 1860 */
2246 if (sched_asym_cpucap_active()) 1861 if (sched_asym_cpucap_active())
2247 select_rq |= !dl_task_fits_ca 1862 select_rq |= !dl_task_fits_capacity(p, cpu);
2248 1863
2249 if (select_rq) { 1864 if (select_rq) {
2250 int target = find_later_rq(p) 1865 int target = find_later_rq(p);
2251 1866
2252 if (target != -1 && 1867 if (target != -1 &&
2253 dl_task_is_earliest_deadl 1868 dl_task_is_earliest_deadline(p, cpu_rq(target)))
2254 cpu = target; 1869 cpu = target;
2255 } 1870 }
2256 rcu_read_unlock(); 1871 rcu_read_unlock();
2257 1872
2258 out: 1873 out:
2259 return cpu; 1874 return cpu;
2260 } 1875 }
2261 1876
2262 static void migrate_task_rq_dl(struct task_st 1877 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
2263 { 1878 {
2264 struct rq_flags rf; 1879 struct rq_flags rf;
2265 struct rq *rq; 1880 struct rq *rq;
2266 1881
2267 if (READ_ONCE(p->__state) != TASK_WAK 1882 if (READ_ONCE(p->__state) != TASK_WAKING)
2268 return; 1883 return;
2269 1884
2270 rq = task_rq(p); 1885 rq = task_rq(p);
2271 /* 1886 /*
2272 * Since p->state == TASK_WAKING, set 1887 * Since p->state == TASK_WAKING, set_task_cpu() has been called
2273 * from try_to_wake_up(). Hence, p->p 1888 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
2274 * rq->lock is not... So, lock it 1889 * rq->lock is not... So, lock it
2275 */ 1890 */
2276 rq_lock(rq, &rf); 1891 rq_lock(rq, &rf);
2277 if (p->dl.dl_non_contending) { 1892 if (p->dl.dl_non_contending) {
2278 update_rq_clock(rq); 1893 update_rq_clock(rq);
2279 sub_running_bw(&p->dl, &rq->d 1894 sub_running_bw(&p->dl, &rq->dl);
2280 p->dl.dl_non_contending = 0; 1895 p->dl.dl_non_contending = 0;
2281 /* 1896 /*
2282 * If the timer handler is cu 1897 * If the timer handler is currently running and the
2283 * timer cannot be canceled, 1898 * timer cannot be canceled, inactive_task_timer()
2284 * will see that dl_not_conte 1899 * will see that dl_not_contending is not set, and
2285 * will not touch the rq's ac 1900 * will not touch the rq's active utilization,
2286 * so we are still safe. 1901 * so we are still safe.
2287 */ 1902 */
2288 if (hrtimer_try_to_cancel(&p- 1903 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2289 put_task_struct(p); 1904 put_task_struct(p);
2290 } 1905 }
2291 sub_rq_bw(&p->dl, &rq->dl); 1906 sub_rq_bw(&p->dl, &rq->dl);
2292 rq_unlock(rq, &rf); 1907 rq_unlock(rq, &rf);
2293 } 1908 }
2294 1909
2295 static void check_preempt_equal_dl(struct rq 1910 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
2296 { 1911 {
2297 /* 1912 /*
2298 * Current can't be migrated, useless 1913 * Current can't be migrated, useless to reschedule,
2299 * let's hope p can move out. 1914 * let's hope p can move out.
2300 */ 1915 */
2301 if (rq->curr->nr_cpus_allowed == 1 || 1916 if (rq->curr->nr_cpus_allowed == 1 ||
2302 !cpudl_find(&rq->rd->cpudl, rq->c 1917 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
2303 return; 1918 return;
2304 1919
2305 /* 1920 /*
2306 * p is migratable, so let's not sche 1921 * p is migratable, so let's not schedule it and
2307 * see if it is pushed or pulled some 1922 * see if it is pushed or pulled somewhere else.
2308 */ 1923 */
2309 if (p->nr_cpus_allowed != 1 && 1924 if (p->nr_cpus_allowed != 1 &&
2310 cpudl_find(&rq->rd->cpudl, p, NUL 1925 cpudl_find(&rq->rd->cpudl, p, NULL))
2311 return; 1926 return;
2312 1927
2313 resched_curr(rq); 1928 resched_curr(rq);
2314 } 1929 }
2315 1930
2316 static int balance_dl(struct rq *rq, struct t 1931 static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
2317 { 1932 {
2318 if (!on_dl_rq(&p->dl) && need_pull_dl 1933 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
2319 /* 1934 /*
2320 * This is OK, because curren 1935 * This is OK, because current is on_cpu, which avoids it being
2321 * picked for load-balance an 1936 * picked for load-balance and preemption/IRQs are still
2322 * disabled avoiding further 1937 * disabled avoiding further scheduler activity on it and we've
2323 * not yet started the pickin 1938 * not yet started the picking loop.
2324 */ 1939 */
2325 rq_unpin_lock(rq, rf); 1940 rq_unpin_lock(rq, rf);
2326 pull_dl_task(rq); 1941 pull_dl_task(rq);
2327 rq_repin_lock(rq, rf); 1942 rq_repin_lock(rq, rf);
2328 } 1943 }
2329 1944
2330 return sched_stop_runnable(rq) || sch 1945 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
2331 } 1946 }
2332 #endif /* CONFIG_SMP */ 1947 #endif /* CONFIG_SMP */
2333 1948
2334 /* 1949 /*
2335 * Only called when both the current and waki 1950 * Only called when both the current and waking task are -deadline
2336 * tasks. 1951 * tasks.
2337 */ 1952 */
2338 static void wakeup_preempt_dl(struct rq *rq, !! 1953 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
2339 int flags) 1954 int flags)
2340 { 1955 {
2341 if (dl_entity_preempt(&p->dl, &rq->cu 1956 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
2342 resched_curr(rq); 1957 resched_curr(rq);
2343 return; 1958 return;
2344 } 1959 }
2345 1960
2346 #ifdef CONFIG_SMP 1961 #ifdef CONFIG_SMP
2347 /* 1962 /*
2348 * In the unlikely case current and p 1963 * In the unlikely case current and p have the same deadline
2349 * let us try to decide what's the be 1964 * let us try to decide what's the best thing to do...
2350 */ 1965 */
2351 if ((p->dl.deadline == rq->curr->dl.d 1966 if ((p->dl.deadline == rq->curr->dl.deadline) &&
2352 !test_tsk_need_resched(rq->curr)) 1967 !test_tsk_need_resched(rq->curr))
2353 check_preempt_equal_dl(rq, p) 1968 check_preempt_equal_dl(rq, p);
2354 #endif /* CONFIG_SMP */ 1969 #endif /* CONFIG_SMP */
2355 } 1970 }
2356 1971
2357 #ifdef CONFIG_SCHED_HRTICK 1972 #ifdef CONFIG_SCHED_HRTICK
2358 static void start_hrtick_dl(struct rq *rq, st !! 1973 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
2359 { 1974 {
2360 hrtick_start(rq, dl_se->runtime); !! 1975 hrtick_start(rq, p->dl.runtime);
2361 } 1976 }
2362 #else /* !CONFIG_SCHED_HRTICK */ 1977 #else /* !CONFIG_SCHED_HRTICK */
2363 static void start_hrtick_dl(struct rq *rq, st !! 1978 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
2364 { 1979 {
2365 } 1980 }
2366 #endif 1981 #endif
2367 1982
2368 static void set_next_task_dl(struct rq *rq, s 1983 static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
2369 { 1984 {
2370 struct sched_dl_entity *dl_se = &p->d 1985 struct sched_dl_entity *dl_se = &p->dl;
2371 struct dl_rq *dl_rq = &rq->dl; 1986 struct dl_rq *dl_rq = &rq->dl;
2372 1987
2373 p->se.exec_start = rq_clock_task(rq); 1988 p->se.exec_start = rq_clock_task(rq);
2374 if (on_dl_rq(&p->dl)) 1989 if (on_dl_rq(&p->dl))
2375 update_stats_wait_end_dl(dl_r 1990 update_stats_wait_end_dl(dl_rq, dl_se);
2376 1991
2377 /* You can't push away the running ta 1992 /* You can't push away the running task */
2378 dequeue_pushable_dl_task(rq, p); 1993 dequeue_pushable_dl_task(rq, p);
2379 1994
2380 if (!first) 1995 if (!first)
2381 return; 1996 return;
2382 1997
>> 1998 if (hrtick_enabled_dl(rq))
>> 1999 start_hrtick_dl(rq, p);
>> 2000
2383 if (rq->curr->sched_class != &dl_sche 2001 if (rq->curr->sched_class != &dl_sched_class)
2384 update_dl_rq_load_avg(rq_cloc 2002 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2385 2003
2386 deadline_queue_push_tasks(rq); 2004 deadline_queue_push_tasks(rq);
2387 <<
2388 if (hrtick_enabled_dl(rq)) <<
2389 start_hrtick_dl(rq, &p->dl); <<
2390 } 2005 }
2391 2006
2392 static struct sched_dl_entity *pick_next_dl_e 2007 static struct sched_dl_entity *pick_next_dl_entity(struct dl_rq *dl_rq)
2393 { 2008 {
2394 struct rb_node *left = rb_first_cache 2009 struct rb_node *left = rb_first_cached(&dl_rq->root);
2395 2010
2396 if (!left) 2011 if (!left)
2397 return NULL; 2012 return NULL;
2398 2013
2399 return __node_2_dle(left); 2014 return __node_2_dle(left);
2400 } 2015 }
2401 2016
2402 /* !! 2017 static struct task_struct *pick_task_dl(struct rq *rq)
2403 * __pick_next_task_dl - Helper to pick the n <<
2404 * @rq: The runqueue to pick the next task fr <<
2405 */ <<
2406 static struct task_struct *__pick_task_dl(str <<
2407 { 2018 {
2408 struct sched_dl_entity *dl_se; 2019 struct sched_dl_entity *dl_se;
2409 struct dl_rq *dl_rq = &rq->dl; 2020 struct dl_rq *dl_rq = &rq->dl;
2410 struct task_struct *p; 2021 struct task_struct *p;
2411 2022
2412 again: <<
2413 if (!sched_dl_runnable(rq)) 2023 if (!sched_dl_runnable(rq))
2414 return NULL; 2024 return NULL;
2415 2025
2416 dl_se = pick_next_dl_entity(dl_rq); 2026 dl_se = pick_next_dl_entity(dl_rq);
2417 WARN_ON_ONCE(!dl_se); 2027 WARN_ON_ONCE(!dl_se);
2418 !! 2028 p = dl_task_of(dl_se);
2419 if (dl_server(dl_se)) { <<
2420 p = dl_se->server_pick_task(d <<
2421 if (!p) { <<
2422 dl_se->dl_yielded = 1 <<
2423 update_curr_dl_se(rq, <<
2424 goto again; <<
2425 } <<
2426 rq->dl_server = dl_se; <<
2427 } else { <<
2428 p = dl_task_of(dl_se); <<
2429 } <<
2430 2029
2431 return p; 2030 return p;
2432 } 2031 }
2433 2032
2434 static struct task_struct *pick_task_dl(struc !! 2033 static struct task_struct *pick_next_task_dl(struct rq *rq)
2435 { 2034 {
2436 return __pick_task_dl(rq); !! 2035 struct task_struct *p;
>> 2036
>> 2037 p = pick_task_dl(rq);
>> 2038 if (p)
>> 2039 set_next_task_dl(rq, p, true);
>> 2040
>> 2041 return p;
2437 } 2042 }
2438 2043
2439 static void put_prev_task_dl(struct rq *rq, s !! 2044 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
2440 { 2045 {
2441 struct sched_dl_entity *dl_se = &p->d 2046 struct sched_dl_entity *dl_se = &p->dl;
2442 struct dl_rq *dl_rq = &rq->dl; 2047 struct dl_rq *dl_rq = &rq->dl;
2443 2048
2444 if (on_dl_rq(&p->dl)) 2049 if (on_dl_rq(&p->dl))
2445 update_stats_wait_start_dl(dl 2050 update_stats_wait_start_dl(dl_rq, dl_se);
2446 2051
2447 update_curr_dl(rq); 2052 update_curr_dl(rq);
2448 2053
2449 update_dl_rq_load_avg(rq_clock_pelt(r 2054 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2450 if (on_dl_rq(&p->dl) && p->nr_cpus_al 2055 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2451 enqueue_pushable_dl_task(rq, 2056 enqueue_pushable_dl_task(rq, p);
2452 } 2057 }
2453 2058
2454 /* 2059 /*
2455 * scheduler tick hitting a task of our sched 2060 * scheduler tick hitting a task of our scheduling class.
2456 * 2061 *
2457 * NOTE: This function can be called remotely 2062 * NOTE: This function can be called remotely by the tick offload that
2458 * goes along full dynticks. Therefore no loc 2063 * goes along full dynticks. Therefore no local assumption can be made
2459 * and everything must be accessed through th 2064 * and everything must be accessed through the @rq and @curr passed in
2460 * parameters. 2065 * parameters.
2461 */ 2066 */
2462 static void task_tick_dl(struct rq *rq, struc 2067 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2463 { 2068 {
2464 update_curr_dl(rq); 2069 update_curr_dl(rq);
2465 2070
2466 update_dl_rq_load_avg(rq_clock_pelt(r 2071 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2467 /* 2072 /*
2468 * Even when we have runtime, update_ 2073 * Even when we have runtime, update_curr_dl() might have resulted in us
2469 * not being the leftmost task anymor 2074 * not being the leftmost task anymore. In that case NEED_RESCHED will
2470 * be set and schedule() will start a 2075 * be set and schedule() will start a new hrtick for the next task.
2471 */ 2076 */
2472 if (hrtick_enabled_dl(rq) && queued & 2077 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
2473 is_leftmost(&p->dl, &rq->dl)) !! 2078 is_leftmost(p, &rq->dl))
2474 start_hrtick_dl(rq, &p->dl); !! 2079 start_hrtick_dl(rq, p);
2475 } 2080 }
2476 2081
2477 static void task_fork_dl(struct task_struct * 2082 static void task_fork_dl(struct task_struct *p)
2478 { 2083 {
2479 /* 2084 /*
2480 * SCHED_DEADLINE tasks cannot fork a 2085 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2481 * sched_fork() 2086 * sched_fork()
2482 */ 2087 */
2483 } 2088 }
2484 2089
2485 #ifdef CONFIG_SMP 2090 #ifdef CONFIG_SMP
2486 2091
2487 /* Only try algorithms three times */ 2092 /* Only try algorithms three times */
2488 #define DL_MAX_TRIES 3 2093 #define DL_MAX_TRIES 3
2489 2094
2490 static int pick_dl_task(struct rq *rq, struct 2095 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
2491 { 2096 {
2492 if (!task_on_cpu(rq, p) && 2097 if (!task_on_cpu(rq, p) &&
2493 cpumask_test_cpu(cpu, &p->cpus_ma 2098 cpumask_test_cpu(cpu, &p->cpus_mask))
2494 return 1; 2099 return 1;
2495 return 0; 2100 return 0;
2496 } 2101 }
2497 2102
2498 /* 2103 /*
2499 * Return the earliest pushable rq's task, wh 2104 * Return the earliest pushable rq's task, which is suitable to be executed
2500 * on the CPU, NULL otherwise: 2105 * on the CPU, NULL otherwise:
2501 */ 2106 */
2502 static struct task_struct *pick_earliest_push 2107 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2503 { 2108 {
2504 struct task_struct *p = NULL; 2109 struct task_struct *p = NULL;
2505 struct rb_node *next_node; 2110 struct rb_node *next_node;
2506 2111
2507 if (!has_pushable_dl_tasks(rq)) 2112 if (!has_pushable_dl_tasks(rq))
2508 return NULL; 2113 return NULL;
2509 2114
2510 next_node = rb_first_cached(&rq->dl.p 2115 next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2511 2116
2512 next_node: 2117 next_node:
2513 if (next_node) { 2118 if (next_node) {
2514 p = __node_2_pdl(next_node); 2119 p = __node_2_pdl(next_node);
2515 2120
2516 if (pick_dl_task(rq, p, cpu)) 2121 if (pick_dl_task(rq, p, cpu))
2517 return p; 2122 return p;
2518 2123
2519 next_node = rb_next(next_node 2124 next_node = rb_next(next_node);
2520 goto next_node; 2125 goto next_node;
2521 } 2126 }
2522 2127
2523 return NULL; 2128 return NULL;
2524 } 2129 }
2525 2130
2526 static DEFINE_PER_CPU(cpumask_var_t, local_cp 2131 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2527 2132
2528 static int find_later_rq(struct task_struct * 2133 static int find_later_rq(struct task_struct *task)
2529 { 2134 {
2530 struct sched_domain *sd; 2135 struct sched_domain *sd;
2531 struct cpumask *later_mask = this_cpu 2136 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2532 int this_cpu = smp_processor_id(); 2137 int this_cpu = smp_processor_id();
2533 int cpu = task_cpu(task); 2138 int cpu = task_cpu(task);
2534 2139
2535 /* Make sure the mask is initialized 2140 /* Make sure the mask is initialized first */
2536 if (unlikely(!later_mask)) 2141 if (unlikely(!later_mask))
2537 return -1; 2142 return -1;
2538 2143
2539 if (task->nr_cpus_allowed == 1) 2144 if (task->nr_cpus_allowed == 1)
2540 return -1; 2145 return -1;
2541 2146
2542 /* 2147 /*
2543 * We have to consider system topolog 2148 * We have to consider system topology and task affinity
2544 * first, then we can look for a suit 2149 * first, then we can look for a suitable CPU.
2545 */ 2150 */
2546 if (!cpudl_find(&task_rq(task)->rd->c 2151 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2547 return -1; 2152 return -1;
2548 2153
2549 /* 2154 /*
2550 * If we are here, some targets have 2155 * If we are here, some targets have been found, including
2551 * the most suitable which is, among 2156 * the most suitable which is, among the runqueues where the
2552 * current tasks have later deadlines 2157 * current tasks have later deadlines than the task's one, the
2553 * rq with the latest possible one. 2158 * rq with the latest possible one.
2554 * 2159 *
2555 * Now we check how well this matches 2160 * Now we check how well this matches with task's
2556 * affinity and system topology. 2161 * affinity and system topology.
2557 * 2162 *
2558 * The last CPU where the task run is 2163 * The last CPU where the task run is our first
2559 * guess, since it is most likely cac 2164 * guess, since it is most likely cache-hot there.
2560 */ 2165 */
2561 if (cpumask_test_cpu(cpu, later_mask) 2166 if (cpumask_test_cpu(cpu, later_mask))
2562 return cpu; 2167 return cpu;
2563 /* 2168 /*
2564 * Check if this_cpu is to be skipped 2169 * Check if this_cpu is to be skipped (i.e., it is
2565 * not in the mask) or not. 2170 * not in the mask) or not.
2566 */ 2171 */
2567 if (!cpumask_test_cpu(this_cpu, later 2172 if (!cpumask_test_cpu(this_cpu, later_mask))
2568 this_cpu = -1; 2173 this_cpu = -1;
2569 2174
2570 rcu_read_lock(); 2175 rcu_read_lock();
2571 for_each_domain(cpu, sd) { 2176 for_each_domain(cpu, sd) {
2572 if (sd->flags & SD_WAKE_AFFIN 2177 if (sd->flags & SD_WAKE_AFFINE) {
2573 int best_cpu; 2178 int best_cpu;
2574 2179
2575 /* 2180 /*
2576 * If possible, preem 2181 * If possible, preempting this_cpu is
2577 * cheaper than migra 2182 * cheaper than migrating.
2578 */ 2183 */
2579 if (this_cpu != -1 && 2184 if (this_cpu != -1 &&
2580 cpumask_test_cpu( 2185 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2581 rcu_read_unlo 2186 rcu_read_unlock();
2582 return this_c 2187 return this_cpu;
2583 } 2188 }
2584 2189
2585 best_cpu = cpumask_an 2190 best_cpu = cpumask_any_and_distribute(later_mask,
2586 2191 sched_domain_span(sd));
2587 /* 2192 /*
2588 * Last chance: if a 2193 * Last chance: if a CPU being in both later_mask
2589 * and current sd spa 2194 * and current sd span is valid, that becomes our
2590 * choice. Of course, 2195 * choice. Of course, the latest possible CPU is
2591 * already under cons 2196 * already under consideration through later_mask.
2592 */ 2197 */
2593 if (best_cpu < nr_cpu 2198 if (best_cpu < nr_cpu_ids) {
2594 rcu_read_unlo 2199 rcu_read_unlock();
2595 return best_c 2200 return best_cpu;
2596 } 2201 }
2597 } 2202 }
2598 } 2203 }
2599 rcu_read_unlock(); 2204 rcu_read_unlock();
2600 2205
2601 /* 2206 /*
2602 * At this point, all our guesses fai 2207 * At this point, all our guesses failed, we just return
2603 * 'something', and let the caller so 2208 * 'something', and let the caller sort the things out.
2604 */ 2209 */
2605 if (this_cpu != -1) 2210 if (this_cpu != -1)
2606 return this_cpu; 2211 return this_cpu;
2607 2212
2608 cpu = cpumask_any_distribute(later_ma 2213 cpu = cpumask_any_distribute(later_mask);
2609 if (cpu < nr_cpu_ids) 2214 if (cpu < nr_cpu_ids)
2610 return cpu; 2215 return cpu;
2611 2216
2612 return -1; 2217 return -1;
2613 } 2218 }
2614 2219
2615 /* Locks the rq it finds */ 2220 /* Locks the rq it finds */
2616 static struct rq *find_lock_later_rq(struct t 2221 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2617 { 2222 {
2618 struct rq *later_rq = NULL; 2223 struct rq *later_rq = NULL;
2619 int tries; 2224 int tries;
2620 int cpu; 2225 int cpu;
2621 2226
2622 for (tries = 0; tries < DL_MAX_TRIES; 2227 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2623 cpu = find_later_rq(task); 2228 cpu = find_later_rq(task);
2624 2229
2625 if ((cpu == -1) || (cpu == rq 2230 if ((cpu == -1) || (cpu == rq->cpu))
2626 break; 2231 break;
2627 2232
2628 later_rq = cpu_rq(cpu); 2233 later_rq = cpu_rq(cpu);
2629 2234
2630 if (!dl_task_is_earliest_dead 2235 if (!dl_task_is_earliest_deadline(task, later_rq)) {
2631 /* 2236 /*
2632 * Target rq has task 2237 * Target rq has tasks of equal or earlier deadline,
2633 * retrying does not 2238 * retrying does not release any lock and is unlikely
2634 * to yield a differe 2239 * to yield a different result.
2635 */ 2240 */
2636 later_rq = NULL; 2241 later_rq = NULL;
2637 break; 2242 break;
2638 } 2243 }
2639 2244
2640 /* Retry if something changed 2245 /* Retry if something changed. */
2641 if (double_lock_balance(rq, l 2246 if (double_lock_balance(rq, later_rq)) {
2642 if (unlikely(task_rq( 2247 if (unlikely(task_rq(task) != rq ||
2643 !cpumask 2248 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2644 task_on_ 2249 task_on_cpu(rq, task) ||
2645 !dl_task 2250 !dl_task(task) ||
2646 is_migra 2251 is_migration_disabled(task) ||
2647 !task_on 2252 !task_on_rq_queued(task))) {
2648 double_unlock 2253 double_unlock_balance(rq, later_rq);
2649 later_rq = NU 2254 later_rq = NULL;
2650 break; 2255 break;
2651 } 2256 }
2652 } 2257 }
2653 2258
2654 /* 2259 /*
2655 * If the rq we found has no 2260 * If the rq we found has no -deadline task, or
2656 * its earliest one has a lat 2261 * its earliest one has a later deadline than our
2657 * task, the rq is a good one 2262 * task, the rq is a good one.
2658 */ 2263 */
2659 if (dl_task_is_earliest_deadl 2264 if (dl_task_is_earliest_deadline(task, later_rq))
2660 break; 2265 break;
2661 2266
2662 /* Otherwise we try again. */ 2267 /* Otherwise we try again. */
2663 double_unlock_balance(rq, lat 2268 double_unlock_balance(rq, later_rq);
2664 later_rq = NULL; 2269 later_rq = NULL;
2665 } 2270 }
2666 2271
2667 return later_rq; 2272 return later_rq;
2668 } 2273 }
2669 2274
2670 static struct task_struct *pick_next_pushable 2275 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2671 { 2276 {
2672 struct task_struct *p; 2277 struct task_struct *p;
2673 2278
2674 if (!has_pushable_dl_tasks(rq)) 2279 if (!has_pushable_dl_tasks(rq))
2675 return NULL; 2280 return NULL;
2676 2281
2677 p = __node_2_pdl(rb_first_cached(&rq- 2282 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2678 2283
2679 WARN_ON_ONCE(rq->cpu != task_cpu(p)); 2284 WARN_ON_ONCE(rq->cpu != task_cpu(p));
2680 WARN_ON_ONCE(task_current(rq, p)); 2285 WARN_ON_ONCE(task_current(rq, p));
2681 WARN_ON_ONCE(p->nr_cpus_allowed <= 1) 2286 WARN_ON_ONCE(p->nr_cpus_allowed <= 1);
2682 2287
2683 WARN_ON_ONCE(!task_on_rq_queued(p)); 2288 WARN_ON_ONCE(!task_on_rq_queued(p));
2684 WARN_ON_ONCE(!dl_task(p)); 2289 WARN_ON_ONCE(!dl_task(p));
2685 2290
2686 return p; 2291 return p;
2687 } 2292 }
2688 2293
2689 /* 2294 /*
2690 * See if the non running -deadline tasks on 2295 * See if the non running -deadline tasks on this rq
2691 * can be sent to some other CPU where they c 2296 * can be sent to some other CPU where they can preempt
2692 * and start executing. 2297 * and start executing.
2693 */ 2298 */
2694 static int push_dl_task(struct rq *rq) 2299 static int push_dl_task(struct rq *rq)
2695 { 2300 {
2696 struct task_struct *next_task; 2301 struct task_struct *next_task;
2697 struct rq *later_rq; 2302 struct rq *later_rq;
2698 int ret = 0; 2303 int ret = 0;
2699 2304
>> 2305 if (!rq->dl.overloaded)
>> 2306 return 0;
>> 2307
2700 next_task = pick_next_pushable_dl_tas 2308 next_task = pick_next_pushable_dl_task(rq);
2701 if (!next_task) 2309 if (!next_task)
2702 return 0; 2310 return 0;
2703 2311
2704 retry: 2312 retry:
2705 /* 2313 /*
2706 * If next_task preempts rq->curr, an 2314 * If next_task preempts rq->curr, and rq->curr
2707 * can move away, it makes sense to j 2315 * can move away, it makes sense to just reschedule
2708 * without going further in pushing n 2316 * without going further in pushing next_task.
2709 */ 2317 */
2710 if (dl_task(rq->curr) && 2318 if (dl_task(rq->curr) &&
2711 dl_time_before(next_task->dl.dead 2319 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2712 rq->curr->nr_cpus_allowed > 1) { 2320 rq->curr->nr_cpus_allowed > 1) {
2713 resched_curr(rq); 2321 resched_curr(rq);
2714 return 0; 2322 return 0;
2715 } 2323 }
2716 2324
2717 if (is_migration_disabled(next_task)) 2325 if (is_migration_disabled(next_task))
2718 return 0; 2326 return 0;
2719 2327
2720 if (WARN_ON(next_task == rq->curr)) 2328 if (WARN_ON(next_task == rq->curr))
2721 return 0; 2329 return 0;
2722 2330
2723 /* We might release rq lock */ 2331 /* We might release rq lock */
2724 get_task_struct(next_task); 2332 get_task_struct(next_task);
2725 2333
2726 /* Will lock the rq it'll find */ 2334 /* Will lock the rq it'll find */
2727 later_rq = find_lock_later_rq(next_ta 2335 later_rq = find_lock_later_rq(next_task, rq);
2728 if (!later_rq) { 2336 if (!later_rq) {
2729 struct task_struct *task; 2337 struct task_struct *task;
2730 2338
2731 /* 2339 /*
2732 * We must check all this aga 2340 * We must check all this again, since
2733 * find_lock_later_rq release 2341 * find_lock_later_rq releases rq->lock and it is
2734 * then possible that next_ta 2342 * then possible that next_task has migrated.
2735 */ 2343 */
2736 task = pick_next_pushable_dl_ 2344 task = pick_next_pushable_dl_task(rq);
2737 if (task == next_task) { 2345 if (task == next_task) {
2738 /* 2346 /*
2739 * The task is still 2347 * The task is still there. We don't try
2740 * again, some other 2348 * again, some other CPU will pull it when ready.
2741 */ 2349 */
2742 goto out; 2350 goto out;
2743 } 2351 }
2744 2352
2745 if (!task) 2353 if (!task)
2746 /* No more tasks */ 2354 /* No more tasks */
2747 goto out; 2355 goto out;
2748 2356
2749 put_task_struct(next_task); 2357 put_task_struct(next_task);
2750 next_task = task; 2358 next_task = task;
2751 goto retry; 2359 goto retry;
2752 } 2360 }
2753 2361
2754 deactivate_task(rq, next_task, 0); 2362 deactivate_task(rq, next_task, 0);
2755 set_task_cpu(next_task, later_rq->cpu 2363 set_task_cpu(next_task, later_rq->cpu);
2756 activate_task(later_rq, next_task, 0) 2364 activate_task(later_rq, next_task, 0);
2757 ret = 1; 2365 ret = 1;
2758 2366
2759 resched_curr(later_rq); 2367 resched_curr(later_rq);
2760 2368
2761 double_unlock_balance(rq, later_rq); 2369 double_unlock_balance(rq, later_rq);
2762 2370
2763 out: 2371 out:
2764 put_task_struct(next_task); 2372 put_task_struct(next_task);
2765 2373
2766 return ret; 2374 return ret;
2767 } 2375 }
2768 2376
2769 static void push_dl_tasks(struct rq *rq) 2377 static void push_dl_tasks(struct rq *rq)
2770 { 2378 {
2771 /* push_dl_task() will return true if 2379 /* push_dl_task() will return true if it moved a -deadline task */
2772 while (push_dl_task(rq)) 2380 while (push_dl_task(rq))
2773 ; 2381 ;
2774 } 2382 }
2775 2383
2776 static void pull_dl_task(struct rq *this_rq) 2384 static void pull_dl_task(struct rq *this_rq)
2777 { 2385 {
2778 int this_cpu = this_rq->cpu, cpu; 2386 int this_cpu = this_rq->cpu, cpu;
2779 struct task_struct *p, *push_task; 2387 struct task_struct *p, *push_task;
2780 bool resched = false; 2388 bool resched = false;
2781 struct rq *src_rq; 2389 struct rq *src_rq;
2782 u64 dmin = LONG_MAX; 2390 u64 dmin = LONG_MAX;
2783 2391
2784 if (likely(!dl_overloaded(this_rq))) 2392 if (likely(!dl_overloaded(this_rq)))
2785 return; 2393 return;
2786 2394
2787 /* 2395 /*
2788 * Match the barrier from dl_set_over 2396 * Match the barrier from dl_set_overloaded; this guarantees that if we
2789 * see overloaded we must also see th 2397 * see overloaded we must also see the dlo_mask bit.
2790 */ 2398 */
2791 smp_rmb(); 2399 smp_rmb();
2792 2400
2793 for_each_cpu(cpu, this_rq->rd->dlo_ma 2401 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2794 if (this_cpu == cpu) 2402 if (this_cpu == cpu)
2795 continue; 2403 continue;
2796 2404
2797 src_rq = cpu_rq(cpu); 2405 src_rq = cpu_rq(cpu);
2798 2406
2799 /* 2407 /*
2800 * It looks racy, and it is! !! 2408 * It looks racy, abd it is! However, as in sched_rt.c,
2801 * we are fine with this. 2409 * we are fine with this.
2802 */ 2410 */
2803 if (this_rq->dl.dl_nr_running 2411 if (this_rq->dl.dl_nr_running &&
2804 dl_time_before(this_rq->d 2412 dl_time_before(this_rq->dl.earliest_dl.curr,
2805 src_rq->dl 2413 src_rq->dl.earliest_dl.next))
2806 continue; 2414 continue;
2807 2415
2808 /* Might drop this_rq->lock * 2416 /* Might drop this_rq->lock */
2809 push_task = NULL; 2417 push_task = NULL;
2810 double_lock_balance(this_rq, 2418 double_lock_balance(this_rq, src_rq);
2811 2419
2812 /* 2420 /*
2813 * If there are no more pulla 2421 * If there are no more pullable tasks on the
2814 * rq, we're done with it. 2422 * rq, we're done with it.
2815 */ 2423 */
2816 if (src_rq->dl.dl_nr_running 2424 if (src_rq->dl.dl_nr_running <= 1)
2817 goto skip; 2425 goto skip;
2818 2426
2819 p = pick_earliest_pushable_dl 2427 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2820 2428
2821 /* 2429 /*
2822 * We found a task to be pull 2430 * We found a task to be pulled if:
2823 * - it preempts our current 2431 * - it preempts our current (if there's one),
2824 * - it will preempt the las 2432 * - it will preempt the last one we pulled (if any).
2825 */ 2433 */
2826 if (p && dl_time_before(p->dl 2434 if (p && dl_time_before(p->dl.deadline, dmin) &&
2827 dl_task_is_earliest_deadl 2435 dl_task_is_earliest_deadline(p, this_rq)) {
2828 WARN_ON(p == src_rq-> 2436 WARN_ON(p == src_rq->curr);
2829 WARN_ON(!task_on_rq_q 2437 WARN_ON(!task_on_rq_queued(p));
2830 2438
2831 /* 2439 /*
2832 * Then we pull iff p 2440 * Then we pull iff p has actually an earlier
2833 * deadline than the 2441 * deadline than the current task of its runqueue.
2834 */ 2442 */
2835 if (dl_time_before(p- 2443 if (dl_time_before(p->dl.deadline,
2836 sr 2444 src_rq->curr->dl.deadline))
2837 goto skip; 2445 goto skip;
2838 2446
2839 if (is_migration_disa 2447 if (is_migration_disabled(p)) {
2840 push_task = g 2448 push_task = get_push_task(src_rq);
2841 } else { 2449 } else {
2842 deactivate_ta 2450 deactivate_task(src_rq, p, 0);
2843 set_task_cpu( 2451 set_task_cpu(p, this_cpu);
2844 activate_task 2452 activate_task(this_rq, p, 0);
2845 dmin = p->dl. 2453 dmin = p->dl.deadline;
2846 resched = tru 2454 resched = true;
2847 } 2455 }
2848 2456
2849 /* Is there any other 2457 /* Is there any other task even earlier? */
2850 } 2458 }
2851 skip: 2459 skip:
2852 double_unlock_balance(this_rq 2460 double_unlock_balance(this_rq, src_rq);
2853 2461
2854 if (push_task) { 2462 if (push_task) {
2855 preempt_disable(); 2463 preempt_disable();
2856 raw_spin_rq_unlock(th 2464 raw_spin_rq_unlock(this_rq);
2857 stop_one_cpu_nowait(s 2465 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2858 p 2466 push_task, &src_rq->push_work);
2859 preempt_enable(); 2467 preempt_enable();
2860 raw_spin_rq_lock(this 2468 raw_spin_rq_lock(this_rq);
2861 } 2469 }
2862 } 2470 }
2863 2471
2864 if (resched) 2472 if (resched)
2865 resched_curr(this_rq); 2473 resched_curr(this_rq);
2866 } 2474 }
2867 2475
2868 /* 2476 /*
2869 * Since the task is not running and a resche 2477 * Since the task is not running and a reschedule is not going to happen
2870 * anytime soon on its runqueue, we try pushi 2478 * anytime soon on its runqueue, we try pushing it away now.
2871 */ 2479 */
2872 static void task_woken_dl(struct rq *rq, stru 2480 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2873 { 2481 {
2874 if (!task_on_cpu(rq, p) && 2482 if (!task_on_cpu(rq, p) &&
2875 !test_tsk_need_resched(rq->curr) 2483 !test_tsk_need_resched(rq->curr) &&
2876 p->nr_cpus_allowed > 1 && 2484 p->nr_cpus_allowed > 1 &&
2877 dl_task(rq->curr) && 2485 dl_task(rq->curr) &&
2878 (rq->curr->nr_cpus_allowed < 2 || 2486 (rq->curr->nr_cpus_allowed < 2 ||
2879 !dl_entity_preempt(&p->dl, &rq-> 2487 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2880 push_dl_tasks(rq); 2488 push_dl_tasks(rq);
2881 } 2489 }
2882 } 2490 }
2883 2491
2884 static void set_cpus_allowed_dl(struct task_s 2492 static void set_cpus_allowed_dl(struct task_struct *p,
2885 struct affini !! 2493 const struct cpumask *new_mask,
>> 2494 u32 flags)
2886 { 2495 {
2887 struct root_domain *src_rd; 2496 struct root_domain *src_rd;
2888 struct rq *rq; 2497 struct rq *rq;
2889 2498
2890 WARN_ON_ONCE(!dl_task(p)); 2499 WARN_ON_ONCE(!dl_task(p));
2891 2500
2892 rq = task_rq(p); 2501 rq = task_rq(p);
2893 src_rd = rq->rd; 2502 src_rd = rq->rd;
2894 /* 2503 /*
2895 * Migrating a SCHED_DEADLINE task be 2504 * Migrating a SCHED_DEADLINE task between exclusive
2896 * cpusets (different root_domains) e 2505 * cpusets (different root_domains) entails a bandwidth
2897 * update. We already made space for 2506 * update. We already made space for us in the destination
2898 * domain (see cpuset_can_attach()). 2507 * domain (see cpuset_can_attach()).
2899 */ 2508 */
2900 if (!cpumask_intersects(src_rd->span, !! 2509 if (!cpumask_intersects(src_rd->span, new_mask)) {
2901 struct dl_bw *src_dl_b; 2510 struct dl_bw *src_dl_b;
2902 2511
2903 src_dl_b = dl_bw_of(cpu_of(rq 2512 src_dl_b = dl_bw_of(cpu_of(rq));
2904 /* 2513 /*
2905 * We now free resources of t 2514 * We now free resources of the root_domain we are migrating
2906 * off. In the worst case, sc 2515 * off. In the worst case, sched_setattr() may temporary fail
2907 * until we complete the upda 2516 * until we complete the update.
2908 */ 2517 */
2909 raw_spin_lock(&src_dl_b->lock 2518 raw_spin_lock(&src_dl_b->lock);
2910 __dl_sub(src_dl_b, p->dl.dl_b 2519 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2911 raw_spin_unlock(&src_dl_b->lo 2520 raw_spin_unlock(&src_dl_b->lock);
2912 } 2521 }
2913 2522
2914 set_cpus_allowed_common(p, ctx); !! 2523 set_cpus_allowed_common(p, new_mask, flags);
2915 } 2524 }
2916 2525
2917 /* Assumes rq->lock is held */ 2526 /* Assumes rq->lock is held */
2918 static void rq_online_dl(struct rq *rq) 2527 static void rq_online_dl(struct rq *rq)
2919 { 2528 {
2920 if (rq->dl.overloaded) 2529 if (rq->dl.overloaded)
2921 dl_set_overload(rq); 2530 dl_set_overload(rq);
2922 2531
2923 cpudl_set_freecpu(&rq->rd->cpudl, rq- 2532 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2924 if (rq->dl.dl_nr_running > 0) 2533 if (rq->dl.dl_nr_running > 0)
2925 cpudl_set(&rq->rd->cpudl, rq- 2534 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2926 } 2535 }
2927 2536
2928 /* Assumes rq->lock is held */ 2537 /* Assumes rq->lock is held */
2929 static void rq_offline_dl(struct rq *rq) 2538 static void rq_offline_dl(struct rq *rq)
2930 { 2539 {
2931 if (rq->dl.overloaded) 2540 if (rq->dl.overloaded)
2932 dl_clear_overload(rq); 2541 dl_clear_overload(rq);
2933 2542
2934 cpudl_clear(&rq->rd->cpudl, rq->cpu); 2543 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2935 cpudl_clear_freecpu(&rq->rd->cpudl, r 2544 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2936 } 2545 }
2937 2546
2938 void __init init_sched_dl_class(void) 2547 void __init init_sched_dl_class(void)
2939 { 2548 {
2940 unsigned int i; 2549 unsigned int i;
2941 2550
2942 for_each_possible_cpu(i) 2551 for_each_possible_cpu(i)
2943 zalloc_cpumask_var_node(&per_ 2552 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2944 GFP_K 2553 GFP_KERNEL, cpu_to_node(i));
2945 } 2554 }
2946 2555
2947 void dl_add_task_root_domain(struct task_stru 2556 void dl_add_task_root_domain(struct task_struct *p)
2948 { 2557 {
2949 struct rq_flags rf; 2558 struct rq_flags rf;
2950 struct rq *rq; 2559 struct rq *rq;
2951 struct dl_bw *dl_b; 2560 struct dl_bw *dl_b;
2952 2561
2953 raw_spin_lock_irqsave(&p->pi_lock, rf 2562 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2954 if (!dl_task(p)) { 2563 if (!dl_task(p)) {
2955 raw_spin_unlock_irqrestore(&p 2564 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2956 return; 2565 return;
2957 } 2566 }
2958 2567
2959 rq = __task_rq_lock(p, &rf); 2568 rq = __task_rq_lock(p, &rf);
2960 2569
2961 dl_b = &rq->rd->dl_bw; 2570 dl_b = &rq->rd->dl_bw;
2962 raw_spin_lock(&dl_b->lock); 2571 raw_spin_lock(&dl_b->lock);
2963 2572
2964 __dl_add(dl_b, p->dl.dl_bw, cpumask_w 2573 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2965 2574
2966 raw_spin_unlock(&dl_b->lock); 2575 raw_spin_unlock(&dl_b->lock);
2967 2576
2968 task_rq_unlock(rq, p, &rf); 2577 task_rq_unlock(rq, p, &rf);
2969 } 2578 }
2970 2579
2971 void dl_clear_root_domain(struct root_domain 2580 void dl_clear_root_domain(struct root_domain *rd)
2972 { 2581 {
2973 unsigned long flags; 2582 unsigned long flags;
2974 2583
2975 raw_spin_lock_irqsave(&rd->dl_bw.lock 2584 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2976 rd->dl_bw.total_bw = 0; 2585 rd->dl_bw.total_bw = 0;
2977 raw_spin_unlock_irqrestore(&rd->dl_bw 2586 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2978 } 2587 }
2979 2588
2980 #endif /* CONFIG_SMP */ 2589 #endif /* CONFIG_SMP */
2981 2590
2982 static void switched_from_dl(struct rq *rq, s 2591 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2983 { 2592 {
2984 /* 2593 /*
2985 * task_non_contending() can start th 2594 * task_non_contending() can start the "inactive timer" (if the 0-lag
2986 * time is in the future). If the tas 2595 * time is in the future). If the task switches back to dl before
2987 * the "inactive timer" fires, it can 2596 * the "inactive timer" fires, it can continue to consume its current
2988 * runtime using its current deadline 2597 * runtime using its current deadline. If it stays outside of
2989 * SCHED_DEADLINE until the 0-lag tim 2598 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2990 * will reset the task parameters. 2599 * will reset the task parameters.
2991 */ 2600 */
2992 if (task_on_rq_queued(p) && p->dl.dl_ 2601 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2993 task_non_contending(&p->dl); !! 2602 task_non_contending(p);
2994 2603
2995 /* 2604 /*
2996 * In case a task is setscheduled out 2605 * In case a task is setscheduled out from SCHED_DEADLINE we need to
2997 * keep track of that on its cpuset ( 2606 * keep track of that on its cpuset (for correct bandwidth tracking).
2998 */ 2607 */
2999 dec_dl_tasks_cs(p); 2608 dec_dl_tasks_cs(p);
3000 2609
3001 if (!task_on_rq_queued(p)) { 2610 if (!task_on_rq_queued(p)) {
3002 /* 2611 /*
3003 * Inactive timer is armed. H 2612 * Inactive timer is armed. However, p is leaving DEADLINE and
3004 * might migrate away from th 2613 * might migrate away from this rq while continuing to run on
3005 * some other class. We need 2614 * some other class. We need to remove its contribution from
3006 * this rq running_bw now, or 2615 * this rq running_bw now, or sub_rq_bw (below) will complain.
3007 */ 2616 */
3008 if (p->dl.dl_non_contending) 2617 if (p->dl.dl_non_contending)
3009 sub_running_bw(&p->dl 2618 sub_running_bw(&p->dl, &rq->dl);
3010 sub_rq_bw(&p->dl, &rq->dl); 2619 sub_rq_bw(&p->dl, &rq->dl);
3011 } 2620 }
3012 2621
3013 /* 2622 /*
3014 * We cannot use inactive_task_timer( 2623 * We cannot use inactive_task_timer() to invoke sub_running_bw()
3015 * at the 0-lag time, because the tas 2624 * at the 0-lag time, because the task could have been migrated
3016 * while SCHED_OTHER in the meanwhile 2625 * while SCHED_OTHER in the meanwhile.
3017 */ 2626 */
3018 if (p->dl.dl_non_contending) 2627 if (p->dl.dl_non_contending)
3019 p->dl.dl_non_contending = 0; 2628 p->dl.dl_non_contending = 0;
3020 2629
3021 /* 2630 /*
3022 * Since this might be the only -dead 2631 * Since this might be the only -deadline task on the rq,
3023 * this is the right place to try to 2632 * this is the right place to try to pull some other one
3024 * from an overloaded CPU, if any. 2633 * from an overloaded CPU, if any.
3025 */ 2634 */
3026 if (!task_on_rq_queued(p) || rq->dl.d 2635 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
3027 return; 2636 return;
3028 2637
3029 deadline_queue_pull_task(rq); 2638 deadline_queue_pull_task(rq);
3030 } 2639 }
3031 2640
3032 /* 2641 /*
3033 * When switching to -deadline, we may overlo 2642 * When switching to -deadline, we may overload the rq, then
3034 * we try to push someone off, if possible. 2643 * we try to push someone off, if possible.
3035 */ 2644 */
3036 static void switched_to_dl(struct rq *rq, str 2645 static void switched_to_dl(struct rq *rq, struct task_struct *p)
3037 { 2646 {
3038 if (hrtimer_try_to_cancel(&p->dl.inac 2647 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
3039 put_task_struct(p); 2648 put_task_struct(p);
3040 2649
3041 /* 2650 /*
3042 * In case a task is setscheduled to 2651 * In case a task is setscheduled to SCHED_DEADLINE we need to keep
3043 * track of that on its cpuset (for c 2652 * track of that on its cpuset (for correct bandwidth tracking).
3044 */ 2653 */
3045 inc_dl_tasks_cs(p); 2654 inc_dl_tasks_cs(p);
3046 2655
3047 /* If p is not queued we will update 2656 /* If p is not queued we will update its parameters at next wakeup. */
3048 if (!task_on_rq_queued(p)) { 2657 if (!task_on_rq_queued(p)) {
3049 add_rq_bw(&p->dl, &rq->dl); 2658 add_rq_bw(&p->dl, &rq->dl);
3050 2659
3051 return; 2660 return;
3052 } 2661 }
3053 2662
3054 if (rq->curr != p) { 2663 if (rq->curr != p) {
3055 #ifdef CONFIG_SMP 2664 #ifdef CONFIG_SMP
3056 if (p->nr_cpus_allowed > 1 && 2665 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
3057 deadline_queue_push_t 2666 deadline_queue_push_tasks(rq);
3058 #endif 2667 #endif
3059 if (dl_task(rq->curr)) 2668 if (dl_task(rq->curr))
3060 wakeup_preempt_dl(rq, !! 2669 check_preempt_curr_dl(rq, p, 0);
3061 else 2670 else
3062 resched_curr(rq); 2671 resched_curr(rq);
3063 } else { 2672 } else {
3064 update_dl_rq_load_avg(rq_cloc 2673 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
3065 } 2674 }
3066 } 2675 }
3067 2676
3068 /* 2677 /*
3069 * If the scheduling parameters of a -deadlin 2678 * If the scheduling parameters of a -deadline task changed,
3070 * a push or pull operation might be needed. 2679 * a push or pull operation might be needed.
3071 */ 2680 */
3072 static void prio_changed_dl(struct rq *rq, st 2681 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
3073 int oldprio) 2682 int oldprio)
3074 { 2683 {
3075 if (!task_on_rq_queued(p)) !! 2684 if (task_on_rq_queued(p) || task_current(rq, p)) {
3076 return; <<
3077 <<
3078 #ifdef CONFIG_SMP 2685 #ifdef CONFIG_SMP
3079 /* !! 2686 /*
3080 * This might be too much, but unfort !! 2687 * This might be too much, but unfortunately
3081 * we don't have the old deadline val !! 2688 * we don't have the old deadline value, and
3082 * we can't argue if the task is incr !! 2689 * we can't argue if the task is increasing
3083 * or lowering its prio, so... !! 2690 * or lowering its prio, so...
3084 */ !! 2691 */
3085 if (!rq->dl.overloaded) !! 2692 if (!rq->dl.overloaded)
3086 deadline_queue_pull_task(rq); !! 2693 deadline_queue_pull_task(rq);
3087 2694
3088 if (task_current(rq, p)) { <<
3089 /* 2695 /*
3090 * If we now have a earlier d 2696 * If we now have a earlier deadline task than p,
3091 * then reschedule, provided 2697 * then reschedule, provided p is still on this
3092 * runqueue. 2698 * runqueue.
3093 */ 2699 */
3094 if (dl_time_before(rq->dl.ear 2700 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
3095 resched_curr(rq); 2701 resched_curr(rq);
3096 } else { !! 2702 #else
3097 /* 2703 /*
3098 * Current may not be deadlin !! 2704 * Again, we don't know if p has a earlier
3099 * have just replenished it ( !! 2705 * or later deadline, so let's blindly set a
3100 * !! 2706 * (maybe not needed) rescheduling point.
3101 * Otherwise, if p was given <<
3102 */ 2707 */
3103 if (!dl_task(rq->curr) || !! 2708 resched_curr(rq);
3104 dl_time_before(p->dl.dead !! 2709 #endif /* CONFIG_SMP */
3105 resched_curr(rq); <<
3106 } 2710 }
3107 #else <<
3108 /* <<
3109 * We don't know if p has a earlier o <<
3110 * set a (maybe not needed) reschedul <<
3111 */ <<
3112 resched_curr(rq); <<
3113 #endif <<
3114 } <<
3115 <<
3116 #ifdef CONFIG_SCHED_CORE <<
3117 static int task_is_throttled_dl(struct task_s <<
3118 { <<
3119 return p->dl.dl_throttled; <<
3120 } 2711 }
3121 #endif <<
3122 2712
3123 DEFINE_SCHED_CLASS(dl) = { 2713 DEFINE_SCHED_CLASS(dl) = {
3124 2714
3125 .enqueue_task = enqueue_tas 2715 .enqueue_task = enqueue_task_dl,
3126 .dequeue_task = dequeue_tas 2716 .dequeue_task = dequeue_task_dl,
3127 .yield_task = yield_task_ 2717 .yield_task = yield_task_dl,
3128 2718
3129 .wakeup_preempt = wakeup_pree !! 2719 .check_preempt_curr = check_preempt_curr_dl,
3130 2720
3131 .pick_task = pick_task_d !! 2721 .pick_next_task = pick_next_task_dl,
3132 .put_prev_task = put_prev_ta 2722 .put_prev_task = put_prev_task_dl,
3133 .set_next_task = set_next_ta 2723 .set_next_task = set_next_task_dl,
3134 2724
3135 #ifdef CONFIG_SMP 2725 #ifdef CONFIG_SMP
3136 .balance = balance_dl, 2726 .balance = balance_dl,
>> 2727 .pick_task = pick_task_dl,
3137 .select_task_rq = select_task 2728 .select_task_rq = select_task_rq_dl,
3138 .migrate_task_rq = migrate_tas 2729 .migrate_task_rq = migrate_task_rq_dl,
3139 .set_cpus_allowed = set_cpus_al 2730 .set_cpus_allowed = set_cpus_allowed_dl,
3140 .rq_online = rq_online_d 2731 .rq_online = rq_online_dl,
3141 .rq_offline = rq_offline_ 2732 .rq_offline = rq_offline_dl,
3142 .task_woken = task_woken_ 2733 .task_woken = task_woken_dl,
3143 .find_lock_rq = find_lock_l 2734 .find_lock_rq = find_lock_later_rq,
3144 #endif 2735 #endif
3145 2736
3146 .task_tick = task_tick_d 2737 .task_tick = task_tick_dl,
3147 .task_fork = task_fork_d 2738 .task_fork = task_fork_dl,
3148 2739
3149 .prio_changed = prio_change 2740 .prio_changed = prio_changed_dl,
3150 .switched_from = switched_fr 2741 .switched_from = switched_from_dl,
3151 .switched_to = switched_to 2742 .switched_to = switched_to_dl,
3152 2743
3153 .update_curr = update_curr 2744 .update_curr = update_curr_dl,
3154 #ifdef CONFIG_SCHED_CORE <<
3155 .task_is_throttled = task_is_thr <<
3156 #endif <<
3157 }; 2745 };
3158 2746
3159 /* Used for dl_bw check and update, used unde 2747 /* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
3160 static u64 dl_generation; 2748 static u64 dl_generation;
3161 2749
3162 int sched_dl_global_validate(void) 2750 int sched_dl_global_validate(void)
3163 { 2751 {
3164 u64 runtime = global_rt_runtime(); 2752 u64 runtime = global_rt_runtime();
3165 u64 period = global_rt_period(); 2753 u64 period = global_rt_period();
3166 u64 new_bw = to_ratio(period, runtime 2754 u64 new_bw = to_ratio(period, runtime);
3167 u64 gen = ++dl_generation; 2755 u64 gen = ++dl_generation;
3168 struct dl_bw *dl_b; 2756 struct dl_bw *dl_b;
3169 int cpu, cpus, ret = 0; 2757 int cpu, cpus, ret = 0;
3170 unsigned long flags; 2758 unsigned long flags;
3171 2759
3172 /* 2760 /*
3173 * Here we want to check the bandwidt 2761 * Here we want to check the bandwidth not being set to some
3174 * value smaller than the currently a 2762 * value smaller than the currently allocated bandwidth in
3175 * any of the root_domains. 2763 * any of the root_domains.
3176 */ 2764 */
3177 for_each_possible_cpu(cpu) { 2765 for_each_possible_cpu(cpu) {
3178 rcu_read_lock_sched(); 2766 rcu_read_lock_sched();
3179 2767
3180 if (dl_bw_visited(cpu, gen)) 2768 if (dl_bw_visited(cpu, gen))
3181 goto next; 2769 goto next;
3182 2770
3183 dl_b = dl_bw_of(cpu); 2771 dl_b = dl_bw_of(cpu);
3184 cpus = dl_bw_cpus(cpu); 2772 cpus = dl_bw_cpus(cpu);
3185 2773
3186 raw_spin_lock_irqsave(&dl_b-> 2774 raw_spin_lock_irqsave(&dl_b->lock, flags);
3187 if (new_bw * cpus < dl_b->tot 2775 if (new_bw * cpus < dl_b->total_bw)
3188 ret = -EBUSY; 2776 ret = -EBUSY;
3189 raw_spin_unlock_irqrestore(&d 2777 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3190 2778
3191 next: 2779 next:
3192 rcu_read_unlock_sched(); 2780 rcu_read_unlock_sched();
3193 2781
3194 if (ret) 2782 if (ret)
3195 break; 2783 break;
3196 } 2784 }
3197 2785
3198 return ret; 2786 return ret;
3199 } 2787 }
3200 2788
3201 static void init_dl_rq_bw_ratio(struct dl_rq 2789 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
3202 { 2790 {
3203 if (global_rt_runtime() == RUNTIME_IN 2791 if (global_rt_runtime() == RUNTIME_INF) {
3204 dl_rq->bw_ratio = 1 << RATIO_ 2792 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
3205 dl_rq->max_bw = dl_rq->extra_ !! 2793 dl_rq->extra_bw = 1 << BW_SHIFT;
3206 } else { 2794 } else {
3207 dl_rq->bw_ratio = to_ratio(gl 2795 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
3208 global_rt_period()) 2796 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
3209 dl_rq->max_bw = dl_rq->extra_ !! 2797 dl_rq->extra_bw = to_ratio(global_rt_period(),
3210 to_ratio(global_rt_pe !! 2798 global_rt_runtime());
3211 } 2799 }
3212 } 2800 }
3213 2801
3214 void sched_dl_do_global(void) 2802 void sched_dl_do_global(void)
3215 { 2803 {
3216 u64 new_bw = -1; 2804 u64 new_bw = -1;
3217 u64 gen = ++dl_generation; 2805 u64 gen = ++dl_generation;
3218 struct dl_bw *dl_b; 2806 struct dl_bw *dl_b;
3219 int cpu; 2807 int cpu;
3220 unsigned long flags; 2808 unsigned long flags;
3221 2809
3222 if (global_rt_runtime() != RUNTIME_IN 2810 if (global_rt_runtime() != RUNTIME_INF)
3223 new_bw = to_ratio(global_rt_p 2811 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
3224 2812
3225 for_each_possible_cpu(cpu) { 2813 for_each_possible_cpu(cpu) {
3226 rcu_read_lock_sched(); 2814 rcu_read_lock_sched();
3227 2815
3228 if (dl_bw_visited(cpu, gen)) 2816 if (dl_bw_visited(cpu, gen)) {
3229 rcu_read_unlock_sched 2817 rcu_read_unlock_sched();
3230 continue; 2818 continue;
3231 } 2819 }
3232 2820
3233 dl_b = dl_bw_of(cpu); 2821 dl_b = dl_bw_of(cpu);
3234 2822
3235 raw_spin_lock_irqsave(&dl_b-> 2823 raw_spin_lock_irqsave(&dl_b->lock, flags);
3236 dl_b->bw = new_bw; 2824 dl_b->bw = new_bw;
3237 raw_spin_unlock_irqrestore(&d 2825 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3238 2826
3239 rcu_read_unlock_sched(); 2827 rcu_read_unlock_sched();
3240 init_dl_rq_bw_ratio(&cpu_rq(c 2828 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
3241 } 2829 }
3242 } 2830 }
3243 2831
3244 /* 2832 /*
3245 * We must be sure that accepting a new task 2833 * We must be sure that accepting a new task (or allowing changing the
3246 * parameters of an existing one) is consiste 2834 * parameters of an existing one) is consistent with the bandwidth
3247 * constraints. If yes, this function also ac 2835 * constraints. If yes, this function also accordingly updates the currently
3248 * allocated bandwidth to reflect the new sit 2836 * allocated bandwidth to reflect the new situation.
3249 * 2837 *
3250 * This function is called while holding p's 2838 * This function is called while holding p's rq->lock.
3251 */ 2839 */
3252 int sched_dl_overflow(struct task_struct *p, 2840 int sched_dl_overflow(struct task_struct *p, int policy,
3253 const struct sched_attr 2841 const struct sched_attr *attr)
3254 { 2842 {
3255 u64 period = attr->sched_period ?: at 2843 u64 period = attr->sched_period ?: attr->sched_deadline;
3256 u64 runtime = attr->sched_runtime; 2844 u64 runtime = attr->sched_runtime;
3257 u64 new_bw = dl_policy(policy) ? to_r 2845 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
3258 int cpus, err = -1, cpu = task_cpu(p) 2846 int cpus, err = -1, cpu = task_cpu(p);
3259 struct dl_bw *dl_b = dl_bw_of(cpu); 2847 struct dl_bw *dl_b = dl_bw_of(cpu);
3260 unsigned long cap; 2848 unsigned long cap;
3261 2849
3262 if (attr->sched_flags & SCHED_FLAG_SU 2850 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3263 return 0; 2851 return 0;
3264 2852
3265 /* !deadline task may carry old deadl 2853 /* !deadline task may carry old deadline bandwidth */
3266 if (new_bw == p->dl.dl_bw && task_has 2854 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
3267 return 0; 2855 return 0;
3268 2856
3269 /* 2857 /*
3270 * Either if a task, enters, leave, o 2858 * Either if a task, enters, leave, or stays -deadline but changes
3271 * its parameters, we may need to upd 2859 * its parameters, we may need to update accordingly the total
3272 * allocated bandwidth of the contain 2860 * allocated bandwidth of the container.
3273 */ 2861 */
3274 raw_spin_lock(&dl_b->lock); 2862 raw_spin_lock(&dl_b->lock);
3275 cpus = dl_bw_cpus(cpu); 2863 cpus = dl_bw_cpus(cpu);
3276 cap = dl_bw_capacity(cpu); 2864 cap = dl_bw_capacity(cpu);
3277 2865
3278 if (dl_policy(policy) && !task_has_dl 2866 if (dl_policy(policy) && !task_has_dl_policy(p) &&
3279 !__dl_overflow(dl_b, cap, 0, new_ 2867 !__dl_overflow(dl_b, cap, 0, new_bw)) {
3280 if (hrtimer_active(&p->dl.ina 2868 if (hrtimer_active(&p->dl.inactive_timer))
3281 __dl_sub(dl_b, p->dl. 2869 __dl_sub(dl_b, p->dl.dl_bw, cpus);
3282 __dl_add(dl_b, new_bw, cpus); 2870 __dl_add(dl_b, new_bw, cpus);
3283 err = 0; 2871 err = 0;
3284 } else if (dl_policy(policy) && task_ 2872 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
3285 !__dl_overflow(dl_b, cap, 2873 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
3286 /* 2874 /*
3287 * XXX this is slightly incor 2875 * XXX this is slightly incorrect: when the task
3288 * utilization decreases, we 2876 * utilization decreases, we should delay the total
3289 * utilization change until t 2877 * utilization change until the task's 0-lag point.
3290 * But this would require to 2878 * But this would require to set the task's "inactive
3291 * timer" when the task is no 2879 * timer" when the task is not inactive.
3292 */ 2880 */
3293 __dl_sub(dl_b, p->dl.dl_bw, c 2881 __dl_sub(dl_b, p->dl.dl_bw, cpus);
3294 __dl_add(dl_b, new_bw, cpus); 2882 __dl_add(dl_b, new_bw, cpus);
3295 dl_change_utilization(p, new_ 2883 dl_change_utilization(p, new_bw);
3296 err = 0; 2884 err = 0;
3297 } else if (!dl_policy(policy) && task 2885 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
3298 /* 2886 /*
3299 * Do not decrease the total 2887 * Do not decrease the total deadline utilization here,
3300 * switched_from_dl() will ta 2888 * switched_from_dl() will take care to do it at the correct
3301 * (0-lag) time. 2889 * (0-lag) time.
3302 */ 2890 */
3303 err = 0; 2891 err = 0;
3304 } 2892 }
3305 raw_spin_unlock(&dl_b->lock); 2893 raw_spin_unlock(&dl_b->lock);
3306 2894
3307 return err; 2895 return err;
3308 } 2896 }
3309 2897
3310 /* 2898 /*
3311 * This function initializes the sched_dl_ent 2899 * This function initializes the sched_dl_entity of a newly becoming
3312 * SCHED_DEADLINE task. 2900 * SCHED_DEADLINE task.
3313 * 2901 *
3314 * Only the static values are considered here 2902 * Only the static values are considered here, the actual runtime and the
3315 * absolute deadline will be properly calcula 2903 * absolute deadline will be properly calculated when the task is enqueued
3316 * for the first time with its new policy. 2904 * for the first time with its new policy.
3317 */ 2905 */
3318 void __setparam_dl(struct task_struct *p, con 2906 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3319 { 2907 {
3320 struct sched_dl_entity *dl_se = &p->d 2908 struct sched_dl_entity *dl_se = &p->dl;
3321 2909
3322 dl_se->dl_runtime = attr->sched_runti 2910 dl_se->dl_runtime = attr->sched_runtime;
3323 dl_se->dl_deadline = attr->sched_dead 2911 dl_se->dl_deadline = attr->sched_deadline;
3324 dl_se->dl_period = attr->sched_period 2912 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3325 dl_se->flags = attr->sched_flags & SC 2913 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
3326 dl_se->dl_bw = to_ratio(dl_se->dl_per 2914 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3327 dl_se->dl_density = to_ratio(dl_se->d 2915 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
3328 } 2916 }
3329 2917
3330 void __getparam_dl(struct task_struct *p, str 2918 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3331 { 2919 {
3332 struct sched_dl_entity *dl_se = &p->d 2920 struct sched_dl_entity *dl_se = &p->dl;
3333 2921
3334 attr->sched_priority = p->rt_priority 2922 attr->sched_priority = p->rt_priority;
3335 attr->sched_runtime = dl_se->dl_runti 2923 attr->sched_runtime = dl_se->dl_runtime;
3336 attr->sched_deadline = dl_se->dl_dead 2924 attr->sched_deadline = dl_se->dl_deadline;
3337 attr->sched_period = dl_se->dl_period 2925 attr->sched_period = dl_se->dl_period;
3338 attr->sched_flags &= ~SCHED_DL_FLAGS; 2926 attr->sched_flags &= ~SCHED_DL_FLAGS;
3339 attr->sched_flags |= dl_se->flags; 2927 attr->sched_flags |= dl_se->flags;
3340 } 2928 }
3341 2929
3342 /* 2930 /*
3343 * This function validates the new parameters 2931 * This function validates the new parameters of a -deadline task.
3344 * We ask for the deadline not being zero, an 2932 * We ask for the deadline not being zero, and greater or equal
3345 * than the runtime, as well as the period of 2933 * than the runtime, as well as the period of being zero or
3346 * greater than deadline. Furthermore, we hav 2934 * greater than deadline. Furthermore, we have to be sure that
3347 * user parameters are above the internal res 2935 * user parameters are above the internal resolution of 1us (we
3348 * check sched_runtime only since it is alway 2936 * check sched_runtime only since it is always the smaller one) and
3349 * below 2^63 ns (we have to check both sched 2937 * below 2^63 ns (we have to check both sched_deadline and
3350 * sched_period, as the latter can be zero). 2938 * sched_period, as the latter can be zero).
3351 */ 2939 */
3352 bool __checkparam_dl(const struct sched_attr 2940 bool __checkparam_dl(const struct sched_attr *attr)
3353 { 2941 {
3354 u64 period, max, min; 2942 u64 period, max, min;
3355 2943
3356 /* special dl tasks don't actually us 2944 /* special dl tasks don't actually use any parameter */
3357 if (attr->sched_flags & SCHED_FLAG_SU 2945 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3358 return true; 2946 return true;
3359 2947
3360 /* deadline != 0 */ 2948 /* deadline != 0 */
3361 if (attr->sched_deadline == 0) 2949 if (attr->sched_deadline == 0)
3362 return false; 2950 return false;
3363 2951
3364 /* 2952 /*
3365 * Since we truncate DL_SCALE bits, m 2953 * Since we truncate DL_SCALE bits, make sure we're at least
3366 * that big. 2954 * that big.
3367 */ 2955 */
3368 if (attr->sched_runtime < (1ULL << DL 2956 if (attr->sched_runtime < (1ULL << DL_SCALE))
3369 return false; 2957 return false;
3370 2958
3371 /* 2959 /*
3372 * Since we use the MSB for wrap-arou 2960 * Since we use the MSB for wrap-around and sign issues, make
3373 * sure it's not set (mind that perio 2961 * sure it's not set (mind that period can be equal to zero).
3374 */ 2962 */
3375 if (attr->sched_deadline & (1ULL << 6 2963 if (attr->sched_deadline & (1ULL << 63) ||
3376 attr->sched_period & (1ULL << 63) 2964 attr->sched_period & (1ULL << 63))
3377 return false; 2965 return false;
3378 2966
3379 period = attr->sched_period; 2967 period = attr->sched_period;
3380 if (!period) 2968 if (!period)
3381 period = attr->sched_deadline 2969 period = attr->sched_deadline;
3382 2970
3383 /* runtime <= deadline <= period (if 2971 /* runtime <= deadline <= period (if period != 0) */
3384 if (period < attr->sched_deadline || 2972 if (period < attr->sched_deadline ||
3385 attr->sched_deadline < attr->sche 2973 attr->sched_deadline < attr->sched_runtime)
3386 return false; 2974 return false;
3387 2975
3388 max = (u64)READ_ONCE(sysctl_sched_dl_ 2976 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
3389 min = (u64)READ_ONCE(sysctl_sched_dl_ 2977 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
3390 2978
3391 if (period < min || period > max) 2979 if (period < min || period > max)
3392 return false; 2980 return false;
3393 2981
3394 return true; 2982 return true;
3395 } 2983 }
3396 2984
3397 /* 2985 /*
3398 * This function clears the sched_dl_entity s 2986 * This function clears the sched_dl_entity static params.
3399 */ 2987 */
3400 static void __dl_clear_params(struct sched_dl !! 2988 void __dl_clear_params(struct task_struct *p)
3401 { 2989 {
>> 2990 struct sched_dl_entity *dl_se = &p->dl;
>> 2991
3402 dl_se->dl_runtime = 0; 2992 dl_se->dl_runtime = 0;
3403 dl_se->dl_deadline = 0; 2993 dl_se->dl_deadline = 0;
3404 dl_se->dl_period = 0; 2994 dl_se->dl_period = 0;
3405 dl_se->flags = 0; 2995 dl_se->flags = 0;
3406 dl_se->dl_bw = 0; 2996 dl_se->dl_bw = 0;
3407 dl_se->dl_density = 0; 2997 dl_se->dl_density = 0;
3408 2998
3409 dl_se->dl_throttled = 0; 2999 dl_se->dl_throttled = 0;
3410 dl_se->dl_yielded = 0; 3000 dl_se->dl_yielded = 0;
3411 dl_se->dl_non_contending = 0; 3001 dl_se->dl_non_contending = 0;
3412 dl_se->dl_overrun = 0; 3002 dl_se->dl_overrun = 0;
3413 dl_se->dl_server = 0; <<
3414 3003
3415 #ifdef CONFIG_RT_MUTEXES 3004 #ifdef CONFIG_RT_MUTEXES
3416 dl_se->pi_se = dl_ 3005 dl_se->pi_se = dl_se;
3417 #endif 3006 #endif
3418 } <<
3419 <<
3420 void init_dl_entity(struct sched_dl_entity *d <<
3421 { <<
3422 RB_CLEAR_NODE(&dl_se->rb_node); <<
3423 init_dl_task_timer(dl_se); <<
3424 init_dl_inactive_task_timer(dl_se); <<
3425 __dl_clear_params(dl_se); <<
3426 } 3007 }
3427 3008
3428 bool dl_param_changed(struct task_struct *p, 3009 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
3429 { 3010 {
3430 struct sched_dl_entity *dl_se = &p->d 3011 struct sched_dl_entity *dl_se = &p->dl;
3431 3012
3432 if (dl_se->dl_runtime != attr->sched_ 3013 if (dl_se->dl_runtime != attr->sched_runtime ||
3433 dl_se->dl_deadline != attr->sched 3014 dl_se->dl_deadline != attr->sched_deadline ||
3434 dl_se->dl_period != attr->sched_p 3015 dl_se->dl_period != attr->sched_period ||
3435 dl_se->flags != (attr->sched_flag 3016 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
3436 return true; 3017 return true;
3437 3018
3438 return false; 3019 return false;
3439 } 3020 }
3440 3021
3441 #ifdef CONFIG_SMP 3022 #ifdef CONFIG_SMP
3442 int dl_cpuset_cpumask_can_shrink(const struct 3023 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3443 const struct 3024 const struct cpumask *trial)
3444 { 3025 {
3445 unsigned long flags, cap; 3026 unsigned long flags, cap;
3446 struct dl_bw *cur_dl_b; 3027 struct dl_bw *cur_dl_b;
3447 int ret = 1; 3028 int ret = 1;
3448 3029
3449 rcu_read_lock_sched(); 3030 rcu_read_lock_sched();
3450 cur_dl_b = dl_bw_of(cpumask_any(cur)) 3031 cur_dl_b = dl_bw_of(cpumask_any(cur));
3451 cap = __dl_bw_capacity(trial); 3032 cap = __dl_bw_capacity(trial);
3452 raw_spin_lock_irqsave(&cur_dl_b->lock 3033 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3453 if (__dl_overflow(cur_dl_b, cap, 0, 0 3034 if (__dl_overflow(cur_dl_b, cap, 0, 0))
3454 ret = 0; 3035 ret = 0;
3455 raw_spin_unlock_irqrestore(&cur_dl_b- 3036 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3456 rcu_read_unlock_sched(); 3037 rcu_read_unlock_sched();
3457 3038
3458 return ret; 3039 return ret;
3459 } 3040 }
3460 3041
3461 enum dl_bw_request { 3042 enum dl_bw_request {
3462 dl_bw_req_check_overflow = 0, 3043 dl_bw_req_check_overflow = 0,
3463 dl_bw_req_alloc, 3044 dl_bw_req_alloc,
3464 dl_bw_req_free 3045 dl_bw_req_free
3465 }; 3046 };
3466 3047
3467 static int dl_bw_manage(enum dl_bw_request re 3048 static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
3468 { 3049 {
3469 unsigned long flags; 3050 unsigned long flags;
3470 struct dl_bw *dl_b; 3051 struct dl_bw *dl_b;
3471 bool overflow = 0; 3052 bool overflow = 0;
3472 3053
3473 rcu_read_lock_sched(); 3054 rcu_read_lock_sched();
3474 dl_b = dl_bw_of(cpu); 3055 dl_b = dl_bw_of(cpu);
3475 raw_spin_lock_irqsave(&dl_b->lock, fl 3056 raw_spin_lock_irqsave(&dl_b->lock, flags);
3476 3057
3477 if (req == dl_bw_req_free) { 3058 if (req == dl_bw_req_free) {
3478 __dl_sub(dl_b, dl_bw, dl_bw_c 3059 __dl_sub(dl_b, dl_bw, dl_bw_cpus(cpu));
3479 } else { 3060 } else {
3480 unsigned long cap = dl_bw_cap 3061 unsigned long cap = dl_bw_capacity(cpu);
3481 3062
3482 overflow = __dl_overflow(dl_b 3063 overflow = __dl_overflow(dl_b, cap, 0, dl_bw);
3483 3064
3484 if (req == dl_bw_req_alloc && 3065 if (req == dl_bw_req_alloc && !overflow) {
3485 /* 3066 /*
3486 * We reserve space i 3067 * We reserve space in the destination
3487 * root_domain, as we 3068 * root_domain, as we can't fail after this point.
3488 * We will free resou 3069 * We will free resources in the source root_domain
3489 * later on (see set_ 3070 * later on (see set_cpus_allowed_dl()).
3490 */ 3071 */
3491 __dl_add(dl_b, dl_bw, 3072 __dl_add(dl_b, dl_bw, dl_bw_cpus(cpu));
3492 } 3073 }
3493 } 3074 }
3494 3075
3495 raw_spin_unlock_irqrestore(&dl_b->loc 3076 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3496 rcu_read_unlock_sched(); 3077 rcu_read_unlock_sched();
3497 3078
3498 return overflow ? -EBUSY : 0; 3079 return overflow ? -EBUSY : 0;
3499 } 3080 }
3500 3081
3501 int dl_bw_check_overflow(int cpu) 3082 int dl_bw_check_overflow(int cpu)
3502 { 3083 {
3503 return dl_bw_manage(dl_bw_req_check_o 3084 return dl_bw_manage(dl_bw_req_check_overflow, cpu, 0);
3504 } 3085 }
3505 3086
3506 int dl_bw_alloc(int cpu, u64 dl_bw) 3087 int dl_bw_alloc(int cpu, u64 dl_bw)
3507 { 3088 {
3508 return dl_bw_manage(dl_bw_req_alloc, 3089 return dl_bw_manage(dl_bw_req_alloc, cpu, dl_bw);
3509 } 3090 }
3510 3091
3511 void dl_bw_free(int cpu, u64 dl_bw) 3092 void dl_bw_free(int cpu, u64 dl_bw)
3512 { 3093 {
3513 dl_bw_manage(dl_bw_req_free, cpu, dl_ 3094 dl_bw_manage(dl_bw_req_free, cpu, dl_bw);
3514 } 3095 }
3515 #endif 3096 #endif
3516 3097
3517 #ifdef CONFIG_SCHED_DEBUG 3098 #ifdef CONFIG_SCHED_DEBUG
3518 void print_dl_stats(struct seq_file *m, int c 3099 void print_dl_stats(struct seq_file *m, int cpu)
3519 { 3100 {
3520 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl) 3101 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3521 } 3102 }
3522 #endif /* CONFIG_SCHED_DEBUG */ 3103 #endif /* CONFIG_SCHED_DEBUG */
3523 3104
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.