xref: /linux/kernel/sched/cpupri.c (revision a1ce39288e6fbefdd8d607021d02384eb4a20b99)
1 /*
2  *  kernel/sched/cpupri.c
3  *
4  *  CPU priority management
5  *
6  *  Copyright (C) 2007-2008 Novell
7  *
8  *  Author: Gregory Haskins <ghaskins@novell.com>
9  *
10  *  This code tracks the priority of each CPU so that global migration
11  *  decisions are easy to calculate.  Each CPU can be in a state as follows:
12  *
13  *                 (INVALID), IDLE, NORMAL, RT1, ... RT99
14  *
15  *  going from the lowest priority to the highest.  CPUs in the INVALID state
16  *  are not eligible for routing.  The system maintains this state with
17  *  a 2 dimensional bitmap (the first for priority class, the second for cpus
18  *  in that class).  Therefore a typical application without affinity
19  *  restrictions can find a suitable CPU with O(1) complexity (e.g. two bit
20  *  searches).  For tasks with affinity restrictions, the algorithm has a
21  *  worst case complexity of O(min(102, nr_domcpus)), though the scenario that
22  *  yields the worst case search is fairly contrived.
23  *
24  *  This program is free software; you can redistribute it and/or
25  *  modify it under the terms of the GNU General Public License
26  *  as published by the Free Software Foundation; version 2
27  *  of the License.
28  */
29 
30 #include <linux/gfp.h>
31 #include "cpupri.h"
32 
33 /* Convert between a 140 based task->prio, and our 102 based cpupri */
34 static int convert_prio(int prio)
35 {
36 	int cpupri;
37 
38 	if (prio == CPUPRI_INVALID)
39 		cpupri = CPUPRI_INVALID;
40 	else if (prio == MAX_PRIO)
41 		cpupri = CPUPRI_IDLE;
42 	else if (prio >= MAX_RT_PRIO)
43 		cpupri = CPUPRI_NORMAL;
44 	else
45 		cpupri = MAX_RT_PRIO - prio + 1;
46 
47 	return cpupri;
48 }
49 
50 /**
51  * cpupri_find - find the best (lowest-pri) CPU in the system
52  * @cp: The cpupri context
53  * @p: The task
54  * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
55  *
56  * Note: This function returns the recommended CPUs as calculated during the
57  * current invocation.  By the time the call returns, the CPUs may have in
58  * fact changed priorities any number of times.  While not ideal, it is not
59  * an issue of correctness since the normal rebalancer logic will correct
60  * any discrepancies created by racing against the uncertainty of the current
61  * priority configuration.
62  *
63  * Returns: (int)bool - CPUs were found
64  */
65 int cpupri_find(struct cpupri *cp, struct task_struct *p,
66 		struct cpumask *lowest_mask)
67 {
68 	int idx = 0;
69 	int task_pri = convert_prio(p->prio);
70 
71 	if (task_pri >= MAX_RT_PRIO)
72 		return 0;
73 
74 	for (idx = 0; idx < task_pri; idx++) {
75 		struct cpupri_vec *vec  = &cp->pri_to_cpu[idx];
76 		int skip = 0;
77 
78 		if (!atomic_read(&(vec)->count))
79 			skip = 1;
80 		/*
81 		 * When looking at the vector, we need to read the counter,
82 		 * do a memory barrier, then read the mask.
83 		 *
84 		 * Note: This is still all racey, but we can deal with it.
85 		 *  Ideally, we only want to look at masks that are set.
86 		 *
87 		 *  If a mask is not set, then the only thing wrong is that we
88 		 *  did a little more work than necessary.
89 		 *
90 		 *  If we read a zero count but the mask is set, because of the
91 		 *  memory barriers, that can only happen when the highest prio
92 		 *  task for a run queue has left the run queue, in which case,
93 		 *  it will be followed by a pull. If the task we are processing
94 		 *  fails to find a proper place to go, that pull request will
95 		 *  pull this task if the run queue is running at a lower
96 		 *  priority.
97 		 */
98 		smp_rmb();
99 
100 		/* Need to do the rmb for every iteration */
101 		if (skip)
102 			continue;
103 
104 		if (cpumask_any_and(&p->cpus_allowed, vec->mask) >= nr_cpu_ids)
105 			continue;
106 
107 		if (lowest_mask) {
108 			cpumask_and(lowest_mask, &p->cpus_allowed, vec->mask);
109 
110 			/*
111 			 * We have to ensure that we have at least one bit
112 			 * still set in the array, since the map could have
113 			 * been concurrently emptied between the first and
114 			 * second reads of vec->mask.  If we hit this
115 			 * condition, simply act as though we never hit this
116 			 * priority level and continue on.
117 			 */
118 			if (cpumask_any(lowest_mask) >= nr_cpu_ids)
119 				continue;
120 		}
121 
122 		return 1;
123 	}
124 
125 	return 0;
126 }
127 
128 /**
129  * cpupri_set - update the cpu priority setting
130  * @cp: The cpupri context
131  * @cpu: The target cpu
132  * @newpri: The priority (INVALID-RT99) to assign to this CPU
133  *
134  * Note: Assumes cpu_rq(cpu)->lock is locked
135  *
136  * Returns: (void)
137  */
138 void cpupri_set(struct cpupri *cp, int cpu, int newpri)
139 {
140 	int *currpri = &cp->cpu_to_pri[cpu];
141 	int oldpri = *currpri;
142 	int do_mb = 0;
143 
144 	newpri = convert_prio(newpri);
145 
146 	BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
147 
148 	if (newpri == oldpri)
149 		return;
150 
151 	/*
152 	 * If the cpu was currently mapped to a different value, we
153 	 * need to map it to the new value then remove the old value.
154 	 * Note, we must add the new value first, otherwise we risk the
155 	 * cpu being missed by the priority loop in cpupri_find.
156 	 */
157 	if (likely(newpri != CPUPRI_INVALID)) {
158 		struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
159 
160 		cpumask_set_cpu(cpu, vec->mask);
161 		/*
162 		 * When adding a new vector, we update the mask first,
163 		 * do a write memory barrier, and then update the count, to
164 		 * make sure the vector is visible when count is set.
165 		 */
166 		smp_mb__before_atomic_inc();
167 		atomic_inc(&(vec)->count);
168 		do_mb = 1;
169 	}
170 	if (likely(oldpri != CPUPRI_INVALID)) {
171 		struct cpupri_vec *vec  = &cp->pri_to_cpu[oldpri];
172 
173 		/*
174 		 * Because the order of modification of the vec->count
175 		 * is important, we must make sure that the update
176 		 * of the new prio is seen before we decrement the
177 		 * old prio. This makes sure that the loop sees
178 		 * one or the other when we raise the priority of
179 		 * the run queue. We don't care about when we lower the
180 		 * priority, as that will trigger an rt pull anyway.
181 		 *
182 		 * We only need to do a memory barrier if we updated
183 		 * the new priority vec.
184 		 */
185 		if (do_mb)
186 			smp_mb__after_atomic_inc();
187 
188 		/*
189 		 * When removing from the vector, we decrement the counter first
190 		 * do a memory barrier and then clear the mask.
191 		 */
192 		atomic_dec(&(vec)->count);
193 		smp_mb__after_atomic_inc();
194 		cpumask_clear_cpu(cpu, vec->mask);
195 	}
196 
197 	*currpri = newpri;
198 }
199 
200 /**
201  * cpupri_init - initialize the cpupri structure
202  * @cp: The cpupri context
203  *
204  * Returns: -ENOMEM if memory fails.
205  */
206 int cpupri_init(struct cpupri *cp)
207 {
208 	int i;
209 
210 	memset(cp, 0, sizeof(*cp));
211 
212 	for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
213 		struct cpupri_vec *vec = &cp->pri_to_cpu[i];
214 
215 		atomic_set(&vec->count, 0);
216 		if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
217 			goto cleanup;
218 	}
219 
220 	for_each_possible_cpu(i)
221 		cp->cpu_to_pri[i] = CPUPRI_INVALID;
222 	return 0;
223 
224 cleanup:
225 	for (i--; i >= 0; i--)
226 		free_cpumask_var(cp->pri_to_cpu[i].mask);
227 	return -ENOMEM;
228 }
229 
230 /**
231  * cpupri_cleanup - clean up the cpupri structure
232  * @cp: The cpupri context
233  */
234 void cpupri_cleanup(struct cpupri *cp)
235 {
236 	int i;
237 
238 	for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
239 		free_cpumask_var(cp->pri_to_cpu[i].mask);
240 }
241