Commit | Line | Data |
---|---|---|
6e0534f2 GH |
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 | ||
5a0e3ad6 | 30 | #include <linux/gfp.h> |
6e0534f2 GH |
31 | #include "sched_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 | ||
6e0534f2 GH |
50 | /** |
51 | * cpupri_find - find the best (lowest-pri) CPU in the system | |
52 | * @cp: The cpupri context | |
53 | * @p: The task | |
13b8bd0a | 54 | * @lowest_mask: A mask to fill in with selected CPUs (or NULL) |
6e0534f2 GH |
55 | * |
56 | * Note: This function returns the recommended CPUs as calculated during the | |
2a61aa40 | 57 | * current invocation. By the time the call returns, the CPUs may have in |
6e0534f2 GH |
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, | |
68e74568 | 66 | struct cpumask *lowest_mask) |
6e0534f2 GH |
67 | { |
68 | int idx = 0; | |
69 | int task_pri = convert_prio(p->prio); | |
70 | ||
c92211d9 SR |
71 | if (task_pri >= MAX_RT_PRIO) |
72 | return 0; | |
73 | ||
74 | for (idx = 0; idx < task_pri; idx++) { | |
6e0534f2 | 75 | struct cpupri_vec *vec = &cp->pri_to_cpu[idx]; |
d473750b | 76 | int skip = 0; |
6e0534f2 | 77 | |
c92211d9 | 78 | if (!atomic_read(&(vec)->count)) |
d473750b | 79 | skip = 1; |
c92211d9 SR |
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(); | |
6e0534f2 | 99 | |
d473750b SR |
100 | /* Need to do the rmb for every iteration */ |
101 | if (skip) | |
102 | continue; | |
103 | ||
68e74568 | 104 | if (cpumask_any_and(&p->cpus_allowed, vec->mask) >= nr_cpu_ids) |
6e0534f2 GH |
105 | continue; |
106 | ||
07903af1 | 107 | if (lowest_mask) { |
13b8bd0a | 108 | cpumask_and(lowest_mask, &p->cpus_allowed, vec->mask); |
07903af1 GH |
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 | ||
6e0534f2 GH |
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 | * @pri: 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; | |
d473750b | 142 | int do_mb = 0; |
6e0534f2 GH |
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 | |
c3a2ae3d SR |
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 | |
5710f15b | 155 | * cpu being missed by the priority loop in cpupri_find. |
6e0534f2 | 156 | */ |
6e0534f2 GH |
157 | if (likely(newpri != CPUPRI_INVALID)) { |
158 | struct cpupri_vec *vec = &cp->pri_to_cpu[newpri]; | |
159 | ||
68e74568 | 160 | cpumask_set_cpu(cpu, vec->mask); |
c92211d9 SR |
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 | */ | |
d473750b | 166 | smp_mb__before_atomic_inc(); |
c92211d9 | 167 | atomic_inc(&(vec)->count); |
d473750b | 168 | do_mb = 1; |
6e0534f2 | 169 | } |
c3a2ae3d SR |
170 | if (likely(oldpri != CPUPRI_INVALID)) { |
171 | struct cpupri_vec *vec = &cp->pri_to_cpu[oldpri]; | |
172 | ||
d473750b SR |
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 | ||
c92211d9 SR |
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); | |
d473750b | 193 | smp_mb__after_atomic_inc(); |
c3a2ae3d | 194 | cpumask_clear_cpu(cpu, vec->mask); |
c3a2ae3d | 195 | } |
6e0534f2 GH |
196 | |
197 | *currpri = newpri; | |
198 | } | |
199 | ||
200 | /** | |
201 | * cpupri_init - initialize the cpupri structure | |
202 | * @cp: The cpupri context | |
68e74568 | 203 | * @bootmem: true if allocations need to use bootmem |
6e0534f2 | 204 | * |
68e74568 | 205 | * Returns: -ENOMEM if memory fails. |
6e0534f2 | 206 | */ |
68c38fc3 | 207 | int cpupri_init(struct cpupri *cp) |
6e0534f2 GH |
208 | { |
209 | int i; | |
210 | ||
211 | memset(cp, 0, sizeof(*cp)); | |
212 | ||
213 | for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) { | |
214 | struct cpupri_vec *vec = &cp->pri_to_cpu[i]; | |
215 | ||
c92211d9 | 216 | atomic_set(&vec->count, 0); |
68c38fc3 | 217 | if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL)) |
68e74568 | 218 | goto cleanup; |
6e0534f2 GH |
219 | } |
220 | ||
221 | for_each_possible_cpu(i) | |
222 | cp->cpu_to_pri[i] = CPUPRI_INVALID; | |
68e74568 RR |
223 | return 0; |
224 | ||
225 | cleanup: | |
226 | for (i--; i >= 0; i--) | |
227 | free_cpumask_var(cp->pri_to_cpu[i].mask); | |
228 | return -ENOMEM; | |
6e0534f2 GH |
229 | } |
230 | ||
68e74568 RR |
231 | /** |
232 | * cpupri_cleanup - clean up the cpupri structure | |
233 | * @cp: The cpupri context | |
234 | */ | |
235 | void cpupri_cleanup(struct cpupri *cp) | |
236 | { | |
237 | int i; | |
6e0534f2 | 238 | |
68e74568 RR |
239 | for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) |
240 | free_cpumask_var(cp->pri_to_cpu[i].mask); | |
241 | } |