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, HIGHER
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.
27 * p->rt_priority p->prio newpri cpupri
29 * -1 -1 (CPUPRI_INVALID)
31 * 99 0 (CPUPRI_NORMAL)
40 * 100 100 (CPUPRI_HIGHER)
42 static int convert_prio(int prio)
48 cpupri = CPUPRI_INVALID; /* -1 */
52 cpupri = MAX_RT_PRIO-1 - prio; /* 1 ... 99 */
56 cpupri = CPUPRI_NORMAL; /* 0 */
60 cpupri = CPUPRI_HIGHER; /* 100 */
67 static inline int __cpupri_find(struct cpupri *cp, struct task_struct *p,
68 struct cpumask *lowest_mask, int idx)
70 struct cpupri_vec *vec = &cp->pri_to_cpu[idx];
73 if (!atomic_read(&(vec)->count))
76 * When looking at the vector, we need to read the counter,
77 * do a memory barrier, then read the mask.
79 * Note: This is still all racy, but we can deal with it.
80 * Ideally, we only want to look at masks that are set.
82 * If a mask is not set, then the only thing wrong is that we
83 * did a little more work than necessary.
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
95 /* Need to do the rmb for every iteration */
99 if (cpumask_any_and(&p->cpus_mask, vec->mask) >= nr_cpu_ids)
103 cpumask_and(lowest_mask, &p->cpus_mask, vec->mask);
106 * We have to ensure that we have at least one bit
107 * still set in the array, since the map could have
108 * been concurrently emptied between the first and
109 * second reads of vec->mask. If we hit this
110 * condition, simply act as though we never hit this
111 * priority level and continue on.
113 if (cpumask_empty(lowest_mask))
120 int cpupri_find(struct cpupri *cp, struct task_struct *p,
121 struct cpumask *lowest_mask)
123 return cpupri_find_fitness(cp, p, lowest_mask, NULL);
127 * cpupri_find_fitness - find the best (lowest-pri) CPU in the system
128 * @cp: The cpupri context
130 * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
131 * @fitness_fn: A pointer to a function to do custom checks whether the CPU
132 * fits a specific criteria so that we only return those CPUs.
134 * Note: This function returns the recommended CPUs as calculated during the
135 * current invocation. By the time the call returns, the CPUs may have in
136 * fact changed priorities any number of times. While not ideal, it is not
137 * an issue of correctness since the normal rebalancer logic will correct
138 * any discrepancies created by racing against the uncertainty of the current
139 * priority configuration.
141 * Return: (int)bool - CPUs were found
143 int cpupri_find_fitness(struct cpupri *cp, struct task_struct *p,
144 struct cpumask *lowest_mask,
145 bool (*fitness_fn)(struct task_struct *p, int cpu))
147 int task_pri = convert_prio(p->prio);
150 WARN_ON_ONCE(task_pri >= CPUPRI_NR_PRIORITIES);
152 for (idx = 0; idx < task_pri; idx++) {
154 if (!__cpupri_find(cp, p, lowest_mask, idx))
157 if (!lowest_mask || !fitness_fn)
160 /* Ensure the capacity of the CPUs fit the task */
161 for_each_cpu(cpu, lowest_mask) {
162 if (!fitness_fn(p, cpu))
163 cpumask_clear_cpu(cpu, lowest_mask);
167 * If no CPU at the current priority can fit the task
170 if (cpumask_empty(lowest_mask))
177 * If we failed to find a fitting lowest_mask, kick off a new search
178 * but without taking into account any fitness criteria this time.
180 * This rule favours honouring priority over fitting the task in the
181 * correct CPU (Capacity Awareness being the only user now).
182 * The idea is that if a higher priority task can run, then it should
183 * run even if this ends up being on unfitting CPU.
185 * The cost of this trade-off is not entirely clear and will probably
186 * be good for some workloads and bad for others.
188 * The main idea here is that if some CPUs were over-committed, we try
189 * to spread which is what the scheduler traditionally did. Sys admins
190 * must do proper RT planning to avoid overloading the system if they
194 return cpupri_find(cp, p, lowest_mask);
200 * cpupri_set - update the CPU priority setting
201 * @cp: The cpupri context
202 * @cpu: The target CPU
203 * @newpri: The priority (INVALID,NORMAL,RT1-RT99,HIGHER) to assign to this CPU
205 * Note: Assumes cpu_rq(cpu)->lock is locked
209 void cpupri_set(struct cpupri *cp, int cpu, int newpri)
211 int *currpri = &cp->cpu_to_pri[cpu];
212 int oldpri = *currpri;
215 newpri = convert_prio(newpri);
217 BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
219 if (newpri == oldpri)
223 * If the CPU was currently mapped to a different value, we
224 * need to map it to the new value then remove the old value.
225 * Note, we must add the new value first, otherwise we risk the
226 * cpu being missed by the priority loop in cpupri_find.
228 if (likely(newpri != CPUPRI_INVALID)) {
229 struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
231 cpumask_set_cpu(cpu, vec->mask);
233 * When adding a new vector, we update the mask first,
234 * do a write memory barrier, and then update the count, to
235 * make sure the vector is visible when count is set.
237 smp_mb__before_atomic();
238 atomic_inc(&(vec)->count);
241 if (likely(oldpri != CPUPRI_INVALID)) {
242 struct cpupri_vec *vec = &cp->pri_to_cpu[oldpri];
245 * Because the order of modification of the vec->count
246 * is important, we must make sure that the update
247 * of the new prio is seen before we decrement the
248 * old prio. This makes sure that the loop sees
249 * one or the other when we raise the priority of
250 * the run queue. We don't care about when we lower the
251 * priority, as that will trigger an rt pull anyway.
253 * We only need to do a memory barrier if we updated
254 * the new priority vec.
257 smp_mb__after_atomic();
260 * When removing from the vector, we decrement the counter first
261 * do a memory barrier and then clear the mask.
263 atomic_dec(&(vec)->count);
264 smp_mb__after_atomic();
265 cpumask_clear_cpu(cpu, vec->mask);
272 * cpupri_init - initialize the cpupri structure
273 * @cp: The cpupri context
275 * Return: -ENOMEM on memory allocation failure.
277 int cpupri_init(struct cpupri *cp)
281 for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
282 struct cpupri_vec *vec = &cp->pri_to_cpu[i];
284 atomic_set(&vec->count, 0);
285 if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
289 cp->cpu_to_pri = kcalloc(nr_cpu_ids, sizeof(int), GFP_KERNEL);
293 for_each_possible_cpu(i)
294 cp->cpu_to_pri[i] = CPUPRI_INVALID;
299 for (i--; i >= 0; i--)
300 free_cpumask_var(cp->pri_to_cpu[i].mask);
305 * cpupri_cleanup - clean up the cpupri structure
306 * @cp: The cpupri context
308 void cpupri_cleanup(struct cpupri *cp)
312 kfree(cp->cpu_to_pri);
313 for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
314 free_cpumask_var(cp->pri_to_cpu[i].mask);