xref: /openbmc/linux/kernel/sched/cpupri.c (revision 89b15863)
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  */
43 static 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 
68 static 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(&(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 racey, 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(lowest_mask, &p->cpus_mask, vec->mask);
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 
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 
127 /**
128  * cpupri_find_fitness - find the best (lowest-pri) CPU in the system
129  * @cp: The cpupri context
130  * @p: The task
131  * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
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.
134  *
135  * Note: This function returns the recommended CPUs as calculated during the
136  * current invocation.  By the time the call returns, the CPUs may have in
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  *
142  * Return: (int)bool - CPUs were found
143  */
144 int cpupri_find_fitness(struct cpupri *cp, struct task_struct *p,
145 		struct cpumask *lowest_mask,
146 		bool (*fitness_fn)(struct task_struct *p, int cpu))
147 {
148 	int task_pri = convert_prio(p->prio);
149 	int idx, cpu;
150 
151 	BUG_ON(task_pri >= CPUPRI_NR_PRIORITIES);
152 
153 	for (idx = 0; idx < task_pri; idx++) {
154 
155 		if (!__cpupri_find(cp, p, lowest_mask, idx))
156 			continue;
157 
158 		if (!lowest_mask || !fitness_fn)
159 			return 1;
160 
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 		}
166 
167 		/*
168 		 * If no CPU at the current priority can fit the task
169 		 * continue looking
170 		 */
171 		if (cpumask_empty(lowest_mask))
172 			continue;
173 
174 		return 1;
175 	}
176 
177 	/*
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.
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 	 *
189 	 * The main idea here is that if some CPUs were overcommitted, we try
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 	 */
194 	if (fitness_fn)
195 		return cpupri_find(cp, p, lowest_mask);
196 
197 	return 0;
198 }
199 
200 /**
201  * cpupri_set - update the CPU priority setting
202  * @cp: The cpupri context
203  * @cpu: The target CPU
204  * @newpri: The priority (INVALID,NORMAL,RT1-RT99,HIGHER) to assign to this CPU
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 {
212 	int *currpri = &cp->cpu_to_pri[cpu];
213 	int oldpri = *currpri;
214 	int do_mb = 0;
215 
216 	newpri = convert_prio(newpri);
217 
218 	BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
219 
220 	if (newpri == oldpri)
221 		return;
222 
223 	/*
224 	 * If the CPU was currently mapped to a different value, we
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
227 	 * cpu being missed by the priority loop in cpupri_find.
228 	 */
229 	if (likely(newpri != CPUPRI_INVALID)) {
230 		struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
231 
232 		cpumask_set_cpu(cpu, vec->mask);
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 		 */
238 		smp_mb__before_atomic();
239 		atomic_inc(&(vec)->count);
240 		do_mb = 1;
241 	}
242 	if (likely(oldpri != CPUPRI_INVALID)) {
243 		struct cpupri_vec *vec  = &cp->pri_to_cpu[oldpri];
244 
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)
258 			smp_mb__after_atomic();
259 
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);
265 		smp_mb__after_atomic();
266 		cpumask_clear_cpu(cpu, vec->mask);
267 	}
268 
269 	*currpri = newpri;
270 }
271 
272 /**
273  * cpupri_init - initialize the cpupri structure
274  * @cp: The cpupri context
275  *
276  * Return: -ENOMEM on memory allocation failure.
277  */
278 int cpupri_init(struct cpupri *cp)
279 {
280 	int i;
281 
282 	for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
283 		struct cpupri_vec *vec = &cp->pri_to_cpu[i];
284 
285 		atomic_set(&vec->count, 0);
286 		if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
287 			goto cleanup;
288 	}
289 
290 	cp->cpu_to_pri = kcalloc(nr_cpu_ids, sizeof(int), GFP_KERNEL);
291 	if (!cp->cpu_to_pri)
292 		goto cleanup;
293 
294 	for_each_possible_cpu(i)
295 		cp->cpu_to_pri[i] = CPUPRI_INVALID;
296 
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;
303 }
304 
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;
312 
313 	kfree(cp->cpu_to_pri);
314 	for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
315 		free_cpumask_var(cp->pri_to_cpu[i].mask);
316 }
317