Commit | Line | Data |
---|---|---|
b886d83c | 1 | // SPDX-License-Identifier: GPL-2.0-only |
6e0534f2 | 2 | /* |
391e43da | 3 | * kernel/sched/cpupri.c |
6e0534f2 GH |
4 | * |
5 | * CPU priority management | |
6 | * | |
7 | * Copyright (C) 2007-2008 Novell | |
8 | * | |
9 | * Author: Gregory Haskins <ghaskins@novell.com> | |
10 | * | |
11 | * This code tracks the priority of each CPU so that global migration | |
12 | * decisions are easy to calculate. Each CPU can be in a state as follows: | |
13 | * | |
b13772f8 | 14 | * (INVALID), NORMAL, RT1, ... RT99, HIGHER |
6e0534f2 GH |
15 | * |
16 | * going from the lowest priority to the highest. CPUs in the INVALID state | |
17 | * are not eligible for routing. The system maintains this state with | |
97fb7a0a | 18 | * a 2 dimensional bitmap (the first for priority class, the second for CPUs |
6e0534f2 GH |
19 | * in that class). Therefore a typical application without affinity |
20 | * restrictions can find a suitable CPU with O(1) complexity (e.g. two bit | |
21 | * searches). For tasks with affinity restrictions, the algorithm has a | |
b13772f8 | 22 | * worst case complexity of O(min(101, nr_domcpus)), though the scenario that |
6e0534f2 | 23 | * yields the worst case search is fairly contrived. |
6e0534f2 | 24 | */ |
6e0534f2 | 25 | |
934fc331 PZ |
26 | /* |
27 | * p->rt_priority p->prio newpri cpupri | |
28 | * | |
29 | * -1 -1 (CPUPRI_INVALID) | |
30 | * | |
31 | * 99 0 (CPUPRI_NORMAL) | |
32 | * | |
33 | * 1 98 98 1 | |
34 | * ... | |
35 | * 49 50 50 49 | |
36 | * 50 49 49 50 | |
37 | * ... | |
38 | * 99 0 0 99 | |
b13772f8 PZ |
39 | * |
40 | * 100 100 (CPUPRI_HIGHER) | |
934fc331 | 41 | */ |
6e0534f2 GH |
42 | static int convert_prio(int prio) |
43 | { | |
44 | int cpupri; | |
45 | ||
934fc331 PZ |
46 | switch (prio) { |
47 | case CPUPRI_INVALID: | |
48 | cpupri = CPUPRI_INVALID; /* -1 */ | |
49 | break; | |
50 | ||
51 | case 0 ... 98: | |
52 | cpupri = MAX_RT_PRIO-1 - prio; /* 1 ... 99 */ | |
53 | break; | |
54 | ||
55 | case MAX_RT_PRIO-1: | |
56 | cpupri = CPUPRI_NORMAL; /* 0 */ | |
57 | break; | |
b13772f8 PZ |
58 | |
59 | case MAX_RT_PRIO: | |
60 | cpupri = CPUPRI_HIGHER; /* 100 */ | |
61 | break; | |
934fc331 | 62 | } |
6e0534f2 GH |
63 | |
64 | return cpupri; | |
65 | } | |
66 | ||
d9cb236b QY |
67 | static inline int __cpupri_find(struct cpupri *cp, struct task_struct *p, |
68 | struct cpumask *lowest_mask, int idx) | |
69 | { | |
70 | struct cpupri_vec *vec = &cp->pri_to_cpu[idx]; | |
71 | int skip = 0; | |
72 | ||
73 | if (!atomic_read(&(vec)->count)) | |
74 | skip = 1; | |
75 | /* | |
76 | * When looking at the vector, we need to read the counter, | |
77 | * do a memory barrier, then read the mask. | |
78 | * | |
3b03706f | 79 | * Note: This is still all racy, but we can deal with it. |
d9cb236b QY |
80 | * Ideally, we only want to look at masks that are set. |
81 | * | |
82 | * If a mask is not set, then the only thing wrong is that we | |
83 | * did a little more work than necessary. | |
84 | * | |
85 | * If we read a zero count but the mask is set, because of the | |
86 | * memory barriers, that can only happen when the highest prio | |
87 | * task for a run queue has left the run queue, in which case, | |
88 | * it will be followed by a pull. If the task we are processing | |
89 | * fails to find a proper place to go, that pull request will | |
90 | * pull this task if the run queue is running at a lower | |
91 | * priority. | |
92 | */ | |
93 | smp_rmb(); | |
94 | ||
95 | /* Need to do the rmb for every iteration */ | |
96 | if (skip) | |
97 | return 0; | |
98 | ||
95158a89 | 99 | if (cpumask_any_and(&p->cpus_mask, vec->mask) >= nr_cpu_ids) |
d9cb236b QY |
100 | return 0; |
101 | ||
102 | if (lowest_mask) { | |
95158a89 | 103 | cpumask_and(lowest_mask, &p->cpus_mask, vec->mask); |
fc090277 | 104 | cpumask_and(lowest_mask, lowest_mask, cpu_active_mask); |
d9cb236b QY |
105 | |
106 | /* | |
107 | * We have to ensure that we have at least one bit | |
108 | * still set in the array, since the map could have | |
109 | * been concurrently emptied between the first and | |
110 | * second reads of vec->mask. If we hit this | |
111 | * condition, simply act as though we never hit this | |
112 | * priority level and continue on. | |
113 | */ | |
114 | if (cpumask_empty(lowest_mask)) | |
115 | return 0; | |
116 | } | |
117 | ||
118 | return 1; | |
119 | } | |
120 | ||
a1bd02e1 QY |
121 | int cpupri_find(struct cpupri *cp, struct task_struct *p, |
122 | struct cpumask *lowest_mask) | |
123 | { | |
124 | return cpupri_find_fitness(cp, p, lowest_mask, NULL); | |
125 | } | |
126 | ||
6e0534f2 | 127 | /** |
a1bd02e1 | 128 | * cpupri_find_fitness - find the best (lowest-pri) CPU in the system |
6e0534f2 GH |
129 | * @cp: The cpupri context |
130 | * @p: The task | |
13b8bd0a | 131 | * @lowest_mask: A mask to fill in with selected CPUs (or NULL) |
804d402f QY |
132 | * @fitness_fn: A pointer to a function to do custom checks whether the CPU |
133 | * fits a specific criteria so that we only return those CPUs. | |
6e0534f2 GH |
134 | * |
135 | * Note: This function returns the recommended CPUs as calculated during the | |
2a61aa40 | 136 | * current invocation. By the time the call returns, the CPUs may have in |
6e0534f2 GH |
137 | * fact changed priorities any number of times. While not ideal, it is not |
138 | * an issue of correctness since the normal rebalancer logic will correct | |
139 | * any discrepancies created by racing against the uncertainty of the current | |
140 | * priority configuration. | |
141 | * | |
e69f6186 | 142 | * Return: (int)bool - CPUs were found |
6e0534f2 | 143 | */ |
a1bd02e1 | 144 | int cpupri_find_fitness(struct cpupri *cp, struct task_struct *p, |
804d402f QY |
145 | struct cpumask *lowest_mask, |
146 | bool (*fitness_fn)(struct task_struct *p, int cpu)) | |
6e0534f2 | 147 | { |
014acbf0 | 148 | int task_pri = convert_prio(p->prio); |
e94f80f6 | 149 | int idx, cpu; |
6e0534f2 | 150 | |
09348d75 | 151 | WARN_ON_ONCE(task_pri >= CPUPRI_NR_PRIORITIES); |
c92211d9 SR |
152 | |
153 | for (idx = 0; idx < task_pri; idx++) { | |
d473750b | 154 | |
d9cb236b | 155 | if (!__cpupri_find(cp, p, lowest_mask, idx)) |
6e0534f2 GH |
156 | continue; |
157 | ||
d9cb236b QY |
158 | if (!lowest_mask || !fitness_fn) |
159 | return 1; | |
804d402f | 160 | |
d9cb236b QY |
161 | /* Ensure the capacity of the CPUs fit the task */ |
162 | for_each_cpu(cpu, lowest_mask) { | |
163 | if (!fitness_fn(p, cpu)) | |
164 | cpumask_clear_cpu(cpu, lowest_mask); | |
165 | } | |
07903af1 | 166 | |
d9cb236b QY |
167 | /* |
168 | * If no CPU at the current priority can fit the task | |
169 | * continue looking | |
170 | */ | |
e94f80f6 | 171 | if (cpumask_empty(lowest_mask)) |
d9cb236b | 172 | continue; |
07903af1 | 173 | |
6e0534f2 GH |
174 | return 1; |
175 | } | |
176 | ||
d9cb236b | 177 | /* |
e94f80f6 QY |
178 | * If we failed to find a fitting lowest_mask, kick off a new search |
179 | * but without taking into account any fitness criteria this time. | |
d9cb236b QY |
180 | * |
181 | * This rule favours honouring priority over fitting the task in the | |
182 | * correct CPU (Capacity Awareness being the only user now). | |
183 | * The idea is that if a higher priority task can run, then it should | |
184 | * run even if this ends up being on unfitting CPU. | |
185 | * | |
186 | * The cost of this trade-off is not entirely clear and will probably | |
187 | * be good for some workloads and bad for others. | |
188 | * | |
3b03706f | 189 | * The main idea here is that if some CPUs were over-committed, we try |
d9cb236b QY |
190 | * to spread which is what the scheduler traditionally did. Sys admins |
191 | * must do proper RT planning to avoid overloading the system if they | |
192 | * really care. | |
193 | */ | |
e94f80f6 QY |
194 | if (fitness_fn) |
195 | return cpupri_find(cp, p, lowest_mask); | |
d9cb236b | 196 | |
6e0534f2 GH |
197 | return 0; |
198 | } | |
199 | ||
200 | /** | |
97fb7a0a | 201 | * cpupri_set - update the CPU priority setting |
6e0534f2 | 202 | * @cp: The cpupri context |
97fb7a0a | 203 | * @cpu: The target CPU |
b13772f8 | 204 | * @newpri: The priority (INVALID,NORMAL,RT1-RT99,HIGHER) to assign to this CPU |
6e0534f2 GH |
205 | * |
206 | * Note: Assumes cpu_rq(cpu)->lock is locked | |
207 | * | |
208 | * Returns: (void) | |
209 | */ | |
210 | void cpupri_set(struct cpupri *cp, int cpu, int newpri) | |
211 | { | |
014acbf0 YX |
212 | int *currpri = &cp->cpu_to_pri[cpu]; |
213 | int oldpri = *currpri; | |
214 | int do_mb = 0; | |
6e0534f2 GH |
215 | |
216 | newpri = convert_prio(newpri); | |
217 | ||
218 | BUG_ON(newpri >= CPUPRI_NR_PRIORITIES); | |
219 | ||
220 | if (newpri == oldpri) | |
221 | return; | |
222 | ||
223 | /* | |
97fb7a0a | 224 | * If the CPU was currently mapped to a different value, we |
c3a2ae3d SR |
225 | * need to map it to the new value then remove the old value. |
226 | * Note, we must add the new value first, otherwise we risk the | |
5710f15b | 227 | * cpu being missed by the priority loop in cpupri_find. |
6e0534f2 | 228 | */ |
6e0534f2 GH |
229 | if (likely(newpri != CPUPRI_INVALID)) { |
230 | struct cpupri_vec *vec = &cp->pri_to_cpu[newpri]; | |
231 | ||
68e74568 | 232 | cpumask_set_cpu(cpu, vec->mask); |
c92211d9 SR |
233 | /* |
234 | * When adding a new vector, we update the mask first, | |
235 | * do a write memory barrier, and then update the count, to | |
236 | * make sure the vector is visible when count is set. | |
237 | */ | |
4e857c58 | 238 | smp_mb__before_atomic(); |
c92211d9 | 239 | atomic_inc(&(vec)->count); |
d473750b | 240 | do_mb = 1; |
6e0534f2 | 241 | } |
c3a2ae3d SR |
242 | if (likely(oldpri != CPUPRI_INVALID)) { |
243 | struct cpupri_vec *vec = &cp->pri_to_cpu[oldpri]; | |
244 | ||
d473750b SR |
245 | /* |
246 | * Because the order of modification of the vec->count | |
247 | * is important, we must make sure that the update | |
248 | * of the new prio is seen before we decrement the | |
249 | * old prio. This makes sure that the loop sees | |
250 | * one or the other when we raise the priority of | |
251 | * the run queue. We don't care about when we lower the | |
252 | * priority, as that will trigger an rt pull anyway. | |
253 | * | |
254 | * We only need to do a memory barrier if we updated | |
255 | * the new priority vec. | |
256 | */ | |
257 | if (do_mb) | |
4e857c58 | 258 | smp_mb__after_atomic(); |
d473750b | 259 | |
c92211d9 SR |
260 | /* |
261 | * When removing from the vector, we decrement the counter first | |
262 | * do a memory barrier and then clear the mask. | |
263 | */ | |
264 | atomic_dec(&(vec)->count); | |
4e857c58 | 265 | smp_mb__after_atomic(); |
c3a2ae3d | 266 | cpumask_clear_cpu(cpu, vec->mask); |
c3a2ae3d | 267 | } |
6e0534f2 GH |
268 | |
269 | *currpri = newpri; | |
270 | } | |
271 | ||
272 | /** | |
273 | * cpupri_init - initialize the cpupri structure | |
274 | * @cp: The cpupri context | |
275 | * | |
e69f6186 | 276 | * Return: -ENOMEM on memory allocation failure. |
6e0534f2 | 277 | */ |
68c38fc3 | 278 | int cpupri_init(struct cpupri *cp) |
6e0534f2 GH |
279 | { |
280 | int i; | |
281 | ||
6e0534f2 GH |
282 | for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) { |
283 | struct cpupri_vec *vec = &cp->pri_to_cpu[i]; | |
284 | ||
c92211d9 | 285 | atomic_set(&vec->count, 0); |
68c38fc3 | 286 | if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL)) |
68e74568 | 287 | goto cleanup; |
6e0534f2 GH |
288 | } |
289 | ||
4dac0b63 PZ |
290 | cp->cpu_to_pri = kcalloc(nr_cpu_ids, sizeof(int), GFP_KERNEL); |
291 | if (!cp->cpu_to_pri) | |
292 | goto cleanup; | |
293 | ||
6e0534f2 GH |
294 | for_each_possible_cpu(i) |
295 | cp->cpu_to_pri[i] = CPUPRI_INVALID; | |
4dac0b63 | 296 | |
68e74568 RR |
297 | return 0; |
298 | ||
299 | cleanup: | |
300 | for (i--; i >= 0; i--) | |
301 | free_cpumask_var(cp->pri_to_cpu[i].mask); | |
302 | return -ENOMEM; | |
6e0534f2 GH |
303 | } |
304 | ||
68e74568 RR |
305 | /** |
306 | * cpupri_cleanup - clean up the cpupri structure | |
307 | * @cp: The cpupri context | |
308 | */ | |
309 | void cpupri_cleanup(struct cpupri *cp) | |
310 | { | |
311 | int i; | |
6e0534f2 | 312 | |
4dac0b63 | 313 | kfree(cp->cpu_to_pri); |
68e74568 RR |
314 | for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) |
315 | free_cpumask_var(cp->pri_to_cpu[i].mask); | |
316 | } |