blob: 7761a2669fff717ee8afb64841639a0259fff015 [file] [log] [blame]
Gregory Haskins6e0534f2008-05-12 21:21:01 +02001/*
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
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090030#include <linux/gfp.h>
Gregory Haskins6e0534f2008-05-12 21:21:01 +020031#include "sched_cpupri.h"
32
33/* Convert between a 140 based task->prio, and our 102 based cpupri */
34static 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
Gregory Haskins6e0534f2008-05-12 21:21:01 +020050/**
51 * cpupri_find - find the best (lowest-pri) CPU in the system
52 * @cp: The cpupri context
53 * @p: The task
Rusty Russell13b8bd02009-03-25 15:01:22 +103054 * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
Gregory Haskins6e0534f2008-05-12 21:21:01 +020055 *
56 * Note: This function returns the recommended CPUs as calculated during the
Adam Buchbinder2a61aa42009-12-11 16:35:40 -050057 * current invocation. By the time the call returns, the CPUs may have in
Gregory Haskins6e0534f2008-05-12 21:21:01 +020058 * 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 */
65int cpupri_find(struct cpupri *cp, struct task_struct *p,
Rusty Russell68e74562008-11-25 02:35:13 +103066 struct cpumask *lowest_mask)
Gregory Haskins6e0534f2008-05-12 21:21:01 +020067{
68 int idx = 0;
69 int task_pri = convert_prio(p->prio);
70
Steven Rostedtc92211d2011-08-02 16:36:12 -040071 if (task_pri >= MAX_RT_PRIO)
72 return 0;
73
74 for (idx = 0; idx < task_pri; idx++) {
Gregory Haskins6e0534f2008-05-12 21:21:01 +020075 struct cpupri_vec *vec = &cp->pri_to_cpu[idx];
Gregory Haskins6e0534f2008-05-12 21:21:01 +020076
Steven Rostedtc92211d2011-08-02 16:36:12 -040077 if (!atomic_read(&(vec)->count))
78 continue;
79 /*
80 * When looking at the vector, we need to read the counter,
81 * do a memory barrier, then read the mask.
82 *
83 * Note: This is still all racey, but we can deal with it.
84 * Ideally, we only want to look at masks that are set.
85 *
86 * If a mask is not set, then the only thing wrong is that we
87 * did a little more work than necessary.
88 *
89 * If we read a zero count but the mask is set, because of the
90 * memory barriers, that can only happen when the highest prio
91 * task for a run queue has left the run queue, in which case,
92 * it will be followed by a pull. If the task we are processing
93 * fails to find a proper place to go, that pull request will
94 * pull this task if the run queue is running at a lower
95 * priority.
96 */
97 smp_rmb();
Gregory Haskins6e0534f2008-05-12 21:21:01 +020098
Rusty Russell68e74562008-11-25 02:35:13 +103099 if (cpumask_any_and(&p->cpus_allowed, vec->mask) >= nr_cpu_ids)
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200100 continue;
101
Gregory Haskins07903af2009-07-30 10:57:28 -0400102 if (lowest_mask) {
Rusty Russell13b8bd02009-03-25 15:01:22 +1030103 cpumask_and(lowest_mask, &p->cpus_allowed, vec->mask);
Gregory Haskins07903af2009-07-30 10:57:28 -0400104
105 /*
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.
112 */
113 if (cpumask_any(lowest_mask) >= nr_cpu_ids)
114 continue;
115 }
116
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200117 return 1;
118 }
119
120 return 0;
121}
122
123/**
124 * cpupri_set - update the cpu priority setting
125 * @cp: The cpupri context
126 * @cpu: The target cpu
127 * @pri: The priority (INVALID-RT99) to assign to this CPU
128 *
129 * Note: Assumes cpu_rq(cpu)->lock is locked
130 *
131 * Returns: (void)
132 */
133void cpupri_set(struct cpupri *cp, int cpu, int newpri)
134{
135 int *currpri = &cp->cpu_to_pri[cpu];
136 int oldpri = *currpri;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200137
138 newpri = convert_prio(newpri);
139
140 BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
141
142 if (newpri == oldpri)
143 return;
144
145 /*
146 * If the cpu was currently mapped to a different value, we
Steven Rostedtc3a2ae32009-07-29 00:21:23 -0400147 * need to map it to the new value then remove the old value.
148 * Note, we must add the new value first, otherwise we risk the
149 * cpu being cleared from pri_active, and this cpu could be
150 * missed for a push or pull.
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200151 */
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200152 if (likely(newpri != CPUPRI_INVALID)) {
153 struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
154
Rusty Russell68e74562008-11-25 02:35:13 +1030155 cpumask_set_cpu(cpu, vec->mask);
Steven Rostedtc92211d2011-08-02 16:36:12 -0400156 /*
157 * When adding a new vector, we update the mask first,
158 * do a write memory barrier, and then update the count, to
159 * make sure the vector is visible when count is set.
160 */
161 smp_wmb();
162 atomic_inc(&(vec)->count);
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200163 }
Steven Rostedtc3a2ae32009-07-29 00:21:23 -0400164 if (likely(oldpri != CPUPRI_INVALID)) {
165 struct cpupri_vec *vec = &cp->pri_to_cpu[oldpri];
166
Steven Rostedtc92211d2011-08-02 16:36:12 -0400167 /*
168 * When removing from the vector, we decrement the counter first
169 * do a memory barrier and then clear the mask.
170 */
171 atomic_dec(&(vec)->count);
172 smp_wmb();
Steven Rostedtc3a2ae32009-07-29 00:21:23 -0400173 cpumask_clear_cpu(cpu, vec->mask);
Steven Rostedtc3a2ae32009-07-29 00:21:23 -0400174 }
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200175
176 *currpri = newpri;
177}
178
179/**
180 * cpupri_init - initialize the cpupri structure
181 * @cp: The cpupri context
Rusty Russell68e74562008-11-25 02:35:13 +1030182 * @bootmem: true if allocations need to use bootmem
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200183 *
Rusty Russell68e74562008-11-25 02:35:13 +1030184 * Returns: -ENOMEM if memory fails.
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200185 */
Pekka Enberg68c38fc2010-07-15 23:18:22 +0300186int cpupri_init(struct cpupri *cp)
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200187{
188 int i;
189
190 memset(cp, 0, sizeof(*cp));
191
192 for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
193 struct cpupri_vec *vec = &cp->pri_to_cpu[i];
194
Steven Rostedtc92211d2011-08-02 16:36:12 -0400195 atomic_set(&vec->count, 0);
Pekka Enberg68c38fc2010-07-15 23:18:22 +0300196 if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
Rusty Russell68e74562008-11-25 02:35:13 +1030197 goto cleanup;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200198 }
199
200 for_each_possible_cpu(i)
201 cp->cpu_to_pri[i] = CPUPRI_INVALID;
Rusty Russell68e74562008-11-25 02:35:13 +1030202 return 0;
203
204cleanup:
205 for (i--; i >= 0; i--)
206 free_cpumask_var(cp->pri_to_cpu[i].mask);
207 return -ENOMEM;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200208}
209
Rusty Russell68e74562008-11-25 02:35:13 +1030210/**
211 * cpupri_cleanup - clean up the cpupri structure
212 * @cp: The cpupri context
213 */
214void cpupri_cleanup(struct cpupri *cp)
215{
216 int i;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200217
Rusty Russell68e74562008-11-25 02:35:13 +1030218 for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
219 free_cpumask_var(cp->pri_to_cpu[i].mask);
220}