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 "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
 50#define for_each_cpupri_active(array, idx)                    \
 51	for_each_set_bit(idx, array, CPUPRI_NR_PRIORITIES)
 52
 53/**
 54 * cpupri_find - find the best (lowest-pri) CPU in the system
 55 * @cp: The cpupri context
 56 * @p: The task
 57 * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
 58 *
 59 * Note: This function returns the recommended CPUs as calculated during the
 60 * current invocation.  By the time the call returns, the CPUs may have in
 61 * fact changed priorities any number of times.  While not ideal, it is not
 62 * an issue of correctness since the normal rebalancer logic will correct
 63 * any discrepancies created by racing against the uncertainty of the current
 64 * priority configuration.
 65 *
 66 * Returns: (int)bool - CPUs were found
 67 */
 68int cpupri_find(struct cpupri *cp, struct task_struct *p,
 69		struct cpumask *lowest_mask)
 70{
 71	int                  idx      = 0;
 72	int                  task_pri = convert_prio(p->prio);
 73
 74	for_each_cpupri_active(cp->pri_active, idx) {
 75		struct cpupri_vec *vec  = &cp->pri_to_cpu[idx];
 76
 77		if (idx >= task_pri)
 78			break;
 79
 80		if (cpumask_any_and(&p->cpus_allowed, vec->mask) >= nr_cpu_ids)
 81			continue;
 82
 83		if (lowest_mask) {
 84			cpumask_and(lowest_mask, &p->cpus_allowed, vec->mask);
 85
 86			/*
 87			 * We have to ensure that we have at least one bit
 88			 * still set in the array, since the map could have
 89			 * been concurrently emptied between the first and
 90			 * second reads of vec->mask.  If we hit this
 91			 * condition, simply act as though we never hit this
 92			 * priority level and continue on.
 93			 */
 94			if (cpumask_any(lowest_mask) >= nr_cpu_ids)
 95				continue;
 96		}
 97
 98		return 1;
 99	}
100
101	return 0;
102}
103
104/**
105 * cpupri_set - update the cpu priority setting
106 * @cp: The cpupri context
107 * @cpu: The target cpu
108 * @pri: The priority (INVALID-RT99) to assign to this CPU
109 *
110 * Note: Assumes cpu_rq(cpu)->lock is locked
111 *
112 * Returns: (void)
113 */
114void cpupri_set(struct cpupri *cp, int cpu, int newpri)
115{
116	int                 *currpri = &cp->cpu_to_pri[cpu];
117	int                  oldpri  = *currpri;
118	unsigned long        flags;
119
120	newpri = convert_prio(newpri);
121
122	BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
123
124	if (newpri == oldpri)
125		return;
126
127	/*
128	 * If the cpu was currently mapped to a different value, we
129	 * need to map it to the new value then remove the old value.
130	 * Note, we must add the new value first, otherwise we risk the
131	 * cpu being cleared from pri_active, and this cpu could be
132	 * missed for a push or pull.
133	 */
134	if (likely(newpri != CPUPRI_INVALID)) {
135		struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
136
137		raw_spin_lock_irqsave(&vec->lock, flags);
138
139		cpumask_set_cpu(cpu, vec->mask);
140		vec->count++;
141		if (vec->count == 1)
142			set_bit(newpri, cp->pri_active);
143
144		raw_spin_unlock_irqrestore(&vec->lock, flags);
145	}
146	if (likely(oldpri != CPUPRI_INVALID)) {
147		struct cpupri_vec *vec  = &cp->pri_to_cpu[oldpri];
148
149		raw_spin_lock_irqsave(&vec->lock, flags);
150
151		vec->count--;
152		if (!vec->count)
153			clear_bit(oldpri, cp->pri_active);
154		cpumask_clear_cpu(cpu, vec->mask);
155
156		raw_spin_unlock_irqrestore(&vec->lock, flags);
157	}
158
159	*currpri = newpri;
160}
161
162/**
163 * cpupri_init - initialize the cpupri structure
164 * @cp: The cpupri context
165 * @bootmem: true if allocations need to use bootmem
166 *
167 * Returns: -ENOMEM if memory fails.
168 */
169int cpupri_init(struct cpupri *cp)
170{
171	int i;
172
173	memset(cp, 0, sizeof(*cp));
174
175	for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
176		struct cpupri_vec *vec = &cp->pri_to_cpu[i];
177
178		raw_spin_lock_init(&vec->lock);
179		vec->count = 0;
180		if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
181			goto cleanup;
182	}
183
184	for_each_possible_cpu(i)
185		cp->cpu_to_pri[i] = CPUPRI_INVALID;
186	return 0;
187
188cleanup:
189	for (i--; i >= 0; i--)
190		free_cpumask_var(cp->pri_to_cpu[i].mask);
191	return -ENOMEM;
192}
193
194/**
195 * cpupri_cleanup - clean up the cpupri structure
196 * @cp: The cpupri context
197 */
198void cpupri_cleanup(struct cpupri *cp)
199{
200	int i;
201
202	for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
203		free_cpumask_var(cp->pri_to_cpu[i].mask);
204}