1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kernel/sched/cpupri.c
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 *
14 * (INVALID), NORMAL, RT1, ... RT99, HIGHER
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
18 * a 2 dimensional bitmap (the first for priority class, the second for CPUs
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
22 * worst case complexity of O(min(101, nr_domcpus)), though the scenario that
23 * yields the worst case search is fairly contrived.
24 */
25#include "sched.h"
26
27/*
28 * p->rt_priority p->prio newpri cpupri
29 *
30 * -1 -1 (CPUPRI_INVALID)
31 *
32 * 99 0 (CPUPRI_NORMAL)
33 *
34 * 1 98 98 1
35 * ...
36 * 49 50 50 49
37 * 50 49 49 50
38 * ...
39 * 99 0 0 99
40 *
41 * 100 100 (CPUPRI_HIGHER)
42 */
43static int convert_prio(int prio)
44{
45 int cpupri;
46
47 switch (prio) {
48 case CPUPRI_INVALID:
49 cpupri = CPUPRI_INVALID; /* -1 */
50 break;
51
52 case 0 ... 98:
53 cpupri = MAX_RT_PRIO-1 - prio; /* 1 ... 99 */
54 break;
55
56 case MAX_RT_PRIO-1:
57 cpupri = CPUPRI_NORMAL; /* 0 */
58 break;
59
60 case MAX_RT_PRIO:
61 cpupri = CPUPRI_HIGHER; /* 100 */
62 break;
63 }
64
65 return cpupri;
66}
67
68static inline int __cpupri_find(struct cpupri *cp, struct task_struct *p,
69 struct cpumask *lowest_mask, int idx)
70{
71 struct cpupri_vec *vec = &cp->pri_to_cpu[idx];
72 int skip = 0;
73
74 if (!atomic_read(v: &(vec)->count))
75 skip = 1;
76 /*
77 * When looking at the vector, we need to read the counter,
78 * do a memory barrier, then read the mask.
79 *
80 * Note: This is still all racy, but we can deal with it.
81 * Ideally, we only want to look at masks that are set.
82 *
83 * If a mask is not set, then the only thing wrong is that we
84 * did a little more work than necessary.
85 *
86 * If we read a zero count but the mask is set, because of the
87 * memory barriers, that can only happen when the highest prio
88 * task for a run queue has left the run queue, in which case,
89 * it will be followed by a pull. If the task we are processing
90 * fails to find a proper place to go, that pull request will
91 * pull this task if the run queue is running at a lower
92 * priority.
93 */
94 smp_rmb();
95
96 /* Need to do the rmb for every iteration */
97 if (skip)
98 return 0;
99
100 if (cpumask_any_and(&p->cpus_mask, vec->mask) >= nr_cpu_ids)
101 return 0;
102
103 if (lowest_mask) {
104 cpumask_and(dstp: lowest_mask, src1p: &p->cpus_mask, src2p: vec->mask);
105 cpumask_and(dstp: lowest_mask, src1p: lowest_mask, cpu_active_mask);
106
107 /*
108 * We have to ensure that we have at least one bit
109 * still set in the array, since the map could have
110 * been concurrently emptied between the first and
111 * second reads of vec->mask. If we hit this
112 * condition, simply act as though we never hit this
113 * priority level and continue on.
114 */
115 if (cpumask_empty(srcp: lowest_mask))
116 return 0;
117 }
118
119 return 1;
120}
121
122int cpupri_find(struct cpupri *cp, struct task_struct *p,
123 struct cpumask *lowest_mask)
124{
125 return cpupri_find_fitness(cp, p, lowest_mask, NULL);
126}
127
128/**
129 * cpupri_find_fitness - find the best (lowest-pri) CPU in the system
130 * @cp: The cpupri context
131 * @p: The task
132 * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
133 * @fitness_fn: A pointer to a function to do custom checks whether the CPU
134 * fits a specific criteria so that we only return those CPUs.
135 *
136 * Note: This function returns the recommended CPUs as calculated during the
137 * current invocation. By the time the call returns, the CPUs may have in
138 * fact changed priorities any number of times. While not ideal, it is not
139 * an issue of correctness since the normal rebalancer logic will correct
140 * any discrepancies created by racing against the uncertainty of the current
141 * priority configuration.
142 *
143 * Return: (int)bool - CPUs were found
144 */
145int cpupri_find_fitness(struct cpupri *cp, struct task_struct *p,
146 struct cpumask *lowest_mask,
147 bool (*fitness_fn)(struct task_struct *p, int cpu))
148{
149 int task_pri = convert_prio(prio: p->prio);
150 int idx, cpu;
151
152 WARN_ON_ONCE(task_pri >= CPUPRI_NR_PRIORITIES);
153
154 for (idx = 0; idx < task_pri; idx++) {
155
156 if (!__cpupri_find(cp, p, lowest_mask, idx))
157 continue;
158
159 if (!lowest_mask || !fitness_fn)
160 return 1;
161
162 /* Ensure the capacity of the CPUs fit the task */
163 for_each_cpu(cpu, lowest_mask) {
164 if (!fitness_fn(p, cpu))
165 cpumask_clear_cpu(cpu, dstp: lowest_mask);
166 }
167
168 /*
169 * If no CPU at the current priority can fit the task
170 * continue looking
171 */
172 if (cpumask_empty(srcp: lowest_mask))
173 continue;
174
175 return 1;
176 }
177
178 /*
179 * If we failed to find a fitting lowest_mask, kick off a new search
180 * but without taking into account any fitness criteria this time.
181 *
182 * This rule favours honouring priority over fitting the task in the
183 * correct CPU (Capacity Awareness being the only user now).
184 * The idea is that if a higher priority task can run, then it should
185 * run even if this ends up being on unfitting CPU.
186 *
187 * The cost of this trade-off is not entirely clear and will probably
188 * be good for some workloads and bad for others.
189 *
190 * The main idea here is that if some CPUs were over-committed, we try
191 * to spread which is what the scheduler traditionally did. Sys admins
192 * must do proper RT planning to avoid overloading the system if they
193 * really care.
194 */
195 if (fitness_fn)
196 return cpupri_find(cp, p, lowest_mask);
197
198 return 0;
199}
200
201/**
202 * cpupri_set - update the CPU priority setting
203 * @cp: The cpupri context
204 * @cpu: The target CPU
205 * @newpri: The priority (INVALID,NORMAL,RT1-RT99,HIGHER) to assign to this CPU
206 *
207 * Note: Assumes cpu_rq(cpu)->lock is locked
208 *
209 * Returns: (void)
210 */
211void cpupri_set(struct cpupri *cp, int cpu, int newpri)
212{
213 int *currpri = &cp->cpu_to_pri[cpu];
214 int oldpri = *currpri;
215 int do_mb = 0;
216
217 newpri = convert_prio(prio: newpri);
218
219 BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
220
221 if (newpri == oldpri)
222 return;
223
224 /*
225 * If the CPU was currently mapped to a different value, we
226 * need to map it to the new value then remove the old value.
227 * Note, we must add the new value first, otherwise we risk the
228 * cpu being missed by the priority loop in cpupri_find.
229 */
230 if (likely(newpri != CPUPRI_INVALID)) {
231 struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
232
233 cpumask_set_cpu(cpu, dstp: vec->mask);
234 /*
235 * When adding a new vector, we update the mask first,
236 * do a write memory barrier, and then update the count, to
237 * make sure the vector is visible when count is set.
238 */
239 smp_mb__before_atomic();
240 atomic_inc(v: &(vec)->count);
241 do_mb = 1;
242 }
243 if (likely(oldpri != CPUPRI_INVALID)) {
244 struct cpupri_vec *vec = &cp->pri_to_cpu[oldpri];
245
246 /*
247 * Because the order of modification of the vec->count
248 * is important, we must make sure that the update
249 * of the new prio is seen before we decrement the
250 * old prio. This makes sure that the loop sees
251 * one or the other when we raise the priority of
252 * the run queue. We don't care about when we lower the
253 * priority, as that will trigger an rt pull anyway.
254 *
255 * We only need to do a memory barrier if we updated
256 * the new priority vec.
257 */
258 if (do_mb)
259 smp_mb__after_atomic();
260
261 /*
262 * When removing from the vector, we decrement the counter first
263 * do a memory barrier and then clear the mask.
264 */
265 atomic_dec(v: &(vec)->count);
266 smp_mb__after_atomic();
267 cpumask_clear_cpu(cpu, dstp: vec->mask);
268 }
269
270 *currpri = newpri;
271}
272
273/**
274 * cpupri_init - initialize the cpupri structure
275 * @cp: The cpupri context
276 *
277 * Return: -ENOMEM on memory allocation failure.
278 */
279int cpupri_init(struct cpupri *cp)
280{
281 int i;
282
283 for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
284 struct cpupri_vec *vec = &cp->pri_to_cpu[i];
285
286 atomic_set(v: &vec->count, i: 0);
287 if (!zalloc_cpumask_var(mask: &vec->mask, GFP_KERNEL))
288 goto cleanup;
289 }
290
291 cp->cpu_to_pri = kcalloc(nr_cpu_ids, sizeof(int), GFP_KERNEL);
292 if (!cp->cpu_to_pri)
293 goto cleanup;
294
295 for_each_possible_cpu(i)
296 cp->cpu_to_pri[i] = CPUPRI_INVALID;
297
298 return 0;
299
300cleanup:
301 for (i--; i >= 0; i--)
302 free_cpumask_var(mask: cp->pri_to_cpu[i].mask);
303 return -ENOMEM;
304}
305
306/**
307 * cpupri_cleanup - clean up the cpupri structure
308 * @cp: The cpupri context
309 */
310void cpupri_cleanup(struct cpupri *cp)
311{
312 int i;
313
314 kfree(objp: cp->cpu_to_pri);
315 for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
316 free_cpumask_var(mask: cp->pri_to_cpu[i].mask);
317}
318