1 // SPDX-License-Identifier: GPL-2.0-only
3 * kernel/sched/cpupri.c
5 * CPU priority management
7 * Copyright (C) 2007-2008 Novell
9 * Author: Gregory Haskins <ghaskins@novell.com>
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:
14 * (INVALID), NORMAL, RT1, ... RT99
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(100, nr_domcpus)), though the scenario that
23 * yields the worst case search is fairly contrived.
27 /* Convert between a 140 based task->prio, and our 100 based cpupri */
28 static int convert_prio(int prio)
32 if (prio == CPUPRI_INVALID)
33 cpupri = CPUPRI_INVALID;
34 else if (prio >= MAX_RT_PRIO)
35 cpupri = CPUPRI_NORMAL;
37 cpupri = MAX_RT_PRIO - prio - 1;
42 static inline int __cpupri_find(struct cpupri *cp, struct task_struct *p,
43 struct cpumask *lowest_mask, int idx)
45 struct cpupri_vec *vec = &cp->pri_to_cpu[idx];
48 if (!atomic_read(&(vec)->count))
51 * When looking at the vector, we need to read the counter,
52 * do a memory barrier, then read the mask.
54 * Note: This is still all racey, but we can deal with it.
55 * Ideally, we only want to look at masks that are set.
57 * If a mask is not set, then the only thing wrong is that we
58 * did a little more work than necessary.
60 * If we read a zero count but the mask is set, because of the
61 * memory barriers, that can only happen when the highest prio
62 * task for a run queue has left the run queue, in which case,
63 * it will be followed by a pull. If the task we are processing
64 * fails to find a proper place to go, that pull request will
65 * pull this task if the run queue is running at a lower
70 /* Need to do the rmb for every iteration */
74 if (cpumask_any_and(p->cpus_ptr, vec->mask) >= nr_cpu_ids)
78 cpumask_and(lowest_mask, p->cpus_ptr, vec->mask);
81 * We have to ensure that we have at least one bit
82 * still set in the array, since the map could have
83 * been concurrently emptied between the first and
84 * second reads of vec->mask. If we hit this
85 * condition, simply act as though we never hit this
86 * priority level and continue on.
88 if (cpumask_empty(lowest_mask))
95 int cpupri_find(struct cpupri *cp, struct task_struct *p,
96 struct cpumask *lowest_mask)
98 return cpupri_find_fitness(cp, p, lowest_mask, NULL);
102 * cpupri_find_fitness - find the best (lowest-pri) CPU in the system
103 * @cp: The cpupri context
105 * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
106 * @fitness_fn: A pointer to a function to do custom checks whether the CPU
107 * fits a specific criteria so that we only return those CPUs.
109 * Note: This function returns the recommended CPUs as calculated during the
110 * current invocation. By the time the call returns, the CPUs may have in
111 * fact changed priorities any number of times. While not ideal, it is not
112 * an issue of correctness since the normal rebalancer logic will correct
113 * any discrepancies created by racing against the uncertainty of the current
114 * priority configuration.
116 * Return: (int)bool - CPUs were found
118 int cpupri_find_fitness(struct cpupri *cp, struct task_struct *p,
119 struct cpumask *lowest_mask,
120 bool (*fitness_fn)(struct task_struct *p, int cpu))
122 int task_pri = convert_prio(p->prio);
125 BUG_ON(task_pri >= CPUPRI_NR_PRIORITIES);
127 for (idx = 0; idx < task_pri; idx++) {
129 if (!__cpupri_find(cp, p, lowest_mask, idx))
132 if (!lowest_mask || !fitness_fn)
135 /* Ensure the capacity of the CPUs fit the task */
136 for_each_cpu(cpu, lowest_mask) {
137 if (!fitness_fn(p, cpu))
138 cpumask_clear_cpu(cpu, lowest_mask);
142 * If no CPU at the current priority can fit the task
145 if (cpumask_empty(lowest_mask))
152 * If we failed to find a fitting lowest_mask, kick off a new search
153 * but without taking into account any fitness criteria this time.
155 * This rule favours honouring priority over fitting the task in the
156 * correct CPU (Capacity Awareness being the only user now).
157 * The idea is that if a higher priority task can run, then it should
158 * run even if this ends up being on unfitting CPU.
160 * The cost of this trade-off is not entirely clear and will probably
161 * be good for some workloads and bad for others.
163 * The main idea here is that if some CPUs were overcommitted, we try
164 * to spread which is what the scheduler traditionally did. Sys admins
165 * must do proper RT planning to avoid overloading the system if they
169 return cpupri_find(cp, p, lowest_mask);
175 * cpupri_set - update the CPU priority setting
176 * @cp: The cpupri context
177 * @cpu: The target CPU
178 * @newpri: The priority (INVALID-RT99) to assign to this CPU
180 * Note: Assumes cpu_rq(cpu)->lock is locked
184 void cpupri_set(struct cpupri *cp, int cpu, int newpri)
186 int *currpri = &cp->cpu_to_pri[cpu];
187 int oldpri = *currpri;
190 newpri = convert_prio(newpri);
192 BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
194 if (newpri == oldpri)
198 * If the CPU was currently mapped to a different value, we
199 * need to map it to the new value then remove the old value.
200 * Note, we must add the new value first, otherwise we risk the
201 * cpu being missed by the priority loop in cpupri_find.
203 if (likely(newpri != CPUPRI_INVALID)) {
204 struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
206 cpumask_set_cpu(cpu, vec->mask);
208 * When adding a new vector, we update the mask first,
209 * do a write memory barrier, and then update the count, to
210 * make sure the vector is visible when count is set.
212 smp_mb__before_atomic();
213 atomic_inc(&(vec)->count);
216 if (likely(oldpri != CPUPRI_INVALID)) {
217 struct cpupri_vec *vec = &cp->pri_to_cpu[oldpri];
220 * Because the order of modification of the vec->count
221 * is important, we must make sure that the update
222 * of the new prio is seen before we decrement the
223 * old prio. This makes sure that the loop sees
224 * one or the other when we raise the priority of
225 * the run queue. We don't care about when we lower the
226 * priority, as that will trigger an rt pull anyway.
228 * We only need to do a memory barrier if we updated
229 * the new priority vec.
232 smp_mb__after_atomic();
235 * When removing from the vector, we decrement the counter first
236 * do a memory barrier and then clear the mask.
238 atomic_dec(&(vec)->count);
239 smp_mb__after_atomic();
240 cpumask_clear_cpu(cpu, vec->mask);
247 * cpupri_init - initialize the cpupri structure
248 * @cp: The cpupri context
250 * Return: -ENOMEM on memory allocation failure.
252 int cpupri_init(struct cpupri *cp)
256 for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
257 struct cpupri_vec *vec = &cp->pri_to_cpu[i];
259 atomic_set(&vec->count, 0);
260 if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
264 cp->cpu_to_pri = kcalloc(nr_cpu_ids, sizeof(int), GFP_KERNEL);
268 for_each_possible_cpu(i)
269 cp->cpu_to_pri[i] = CPUPRI_INVALID;
274 for (i--; i >= 0; i--)
275 free_cpumask_var(cp->pri_to_cpu[i].mask);
280 * cpupri_cleanup - clean up the cpupri structure
281 * @cp: The cpupri context
283 void cpupri_cleanup(struct cpupri *cp)
287 kfree(cp->cpu_to_pri);
288 for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
289 free_cpumask_var(cp->pri_to_cpu[i].mask);