kernel/sched/cpupri.c

   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 <linux/sched.h>
  32 #include <linux/sched/rt.h>
  33 #include "cpupri.h"
  34
  35 /* Convert between a 140 based task->prio, and our 102 based cpupri */
  36 static int convert_prio(int prio)
  37 {
  38         int cpupri;
  39
  40         if (prio == CPUPRI_INVALID)
  41                 cpupri = CPUPRI_INVALID;
  42         else if (prio == MAX_PRIO)
  43                 cpupri = CPUPRI_IDLE;
  44         else if (prio >= MAX_RT_PRIO)
  45                 cpupri = CPUPRI_NORMAL;
  46         else
  47                 cpupri = MAX_RT_PRIO - prio + 1;
  48
  49         return cpupri;
  50 }
  51
  52 /**
  53  * cpupri_find - find the best (lowest-pri) CPU in the system
  54  * @cp: The cpupri context
  55  * @p: The task
  56  * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
  57  *
  58  * Note: This function returns the recommended CPUs as calculated during the
  59  * current invocation.  By the time the call returns, the CPUs may have in
  60  * fact changed priorities any number of times.  While not ideal, it is not
  61  * an issue of correctness since the normal rebalancer logic will correct
  62  * any discrepancies created by racing against the uncertainty of the current
  63  * priority configuration.
  64  *
  65  * Returns: (int)bool - CPUs were found
  66  */
  67 int cpupri_find(struct cpupri *cp, struct task_struct *p,
  68                 struct cpumask *lowest_mask)
  69 {
  70         int idx = 0;
  71         int task_pri = convert_prio(p->prio);
  72
  73         BUG_ON(task_pri >= CPUPRI_NR_PRIORITIES);
  74
  75         for (idx = 0; idx < task_pri; idx++) {
  76                 struct cpupri_vec *vec  = &cp->pri_to_cpu[idx];
  77                 int skip = 0;
  78
  79                 if (!atomic_read(&(vec)->count))
  80                         skip = 1;
  81                 /*
  82                  * When looking at the vector, we need to read the counter,
  83                  * do a memory barrier, then read the mask.
  84                  *
  85                  * Note: This is still all racey, but we can deal with it.
  86                  *  Ideally, we only want to look at masks that are set.
  87                  *
  88                  *  If a mask is not set, then the only thing wrong is that we
  89                  *  did a little more work than necessary.
  90                  *
  91                  *  If we read a zero count but the mask is set, because of the
  92                  *  memory barriers, that can only happen when the highest prio
  93                  *  task for a run queue has left the run queue, in which case,
  94                  *  it will be followed by a pull. If the task we are processing
  95                  *  fails to find a proper place to go, that pull request will
  96                  *  pull this task if the run queue is running at a lower
  97                  *  priority.
  98                  */
  99                 smp_rmb();
 100
 101                 /* Need to do the rmb for every iteration */
 102                 if (skip)
 103                         continue;
 104
 105                 if (cpumask_any_and(&p->cpus_allowed, vec->mask) >= nr_cpu_ids)
 106                         continue;
 107
 108                 if (lowest_mask) {
 109                         cpumask_and(lowest_mask, &p->cpus_allowed, vec->mask);
 110
 111                         /*
 112                          * We have to ensure that we have at least one bit
 113                          * still set in the array, since the map could have
 114                          * been concurrently emptied between the first and
 115                          * second reads of vec->mask.  If we hit this
 116                          * condition, simply act as though we never hit this
 117                          * priority level and continue on.
 118                          */
 119                         if (cpumask_any(lowest_mask) >= nr_cpu_ids)
 120                                 continue;
 121                 }
 122
 123                 return 1;
 124         }
 125
 126         return 0;
 127 }
 128
 129 /**
 130  * cpupri_set - update the cpu priority setting
 131  * @cp: The cpupri context
 132  * @cpu: The target cpu
 133  * @newpri: The priority (INVALID-RT99) to assign to this CPU
 134  *
 135  * Note: Assumes cpu_rq(cpu)->lock is locked
 136  *
 137  * Returns: (void)
 138  */
 139 void cpupri_set(struct cpupri *cp, int cpu, int newpri)
 140 {
 141         int *currpri = &cp->cpu_to_pri[cpu];
 142         int oldpri = *currpri;
 143         int do_mb = 0;
 144
 145         newpri = convert_prio(newpri);
 146
 147         BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
 148
 149         if (newpri == oldpri)
 150                 return;
 151
 152         /*
 153          * If the cpu was currently mapped to a different value, we
 154          * need to map it to the new value then remove the old value.
 155          * Note, we must add the new value first, otherwise we risk the
 156          * cpu being missed by the priority loop in cpupri_find.
 157          */
 158         if (likely(newpri != CPUPRI_INVALID)) {
 159                 struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
 160
 161                 cpumask_set_cpu(cpu, vec->mask);
 162                 /*
 163                  * When adding a new vector, we update the mask first,
 164                  * do a write memory barrier, and then update the count, to
 165                  * make sure the vector is visible when count is set.
 166                  */
 167                 smp_mb__before_atomic_inc();
 168                 atomic_inc(&(vec)->count);
 169                 do_mb = 1;
 170         }
 171         if (likely(oldpri != CPUPRI_INVALID)) {
 172                 struct cpupri_vec *vec  = &cp->pri_to_cpu[oldpri];
 173
 174                 /*
 175                  * Because the order of modification of the vec->count
 176                  * is important, we must make sure that the update
 177                  * of the new prio is seen before we decrement the
 178                  * old prio. This makes sure that the loop sees
 179                  * one or the other when we raise the priority of
 180                  * the run queue. We don't care about when we lower the
 181                  * priority, as that will trigger an rt pull anyway.
 182                  *
 183                  * We only need to do a memory barrier if we updated
 184                  * the new priority vec.
 185                  */
 186                 if (do_mb)
 187                         smp_mb__after_atomic_inc();
 188
 189                 /*
 190                  * When removing from the vector, we decrement the counter first
 191                  * do a memory barrier and then clear the mask.
 192                  */
 193                 atomic_dec(&(vec)->count);
 194                 smp_mb__after_atomic_inc();
 195                 cpumask_clear_cpu(cpu, vec->mask);
 196         }
 197
 198         *currpri = newpri;
 199 }
 200
 201 /**
 202  * cpupri_init - initialize the cpupri structure
 203  * @cp: The cpupri context
 204  *
 205  * Returns: -ENOMEM if memory fails.
 206  */
 207 int cpupri_init(struct cpupri *cp)
 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
 216                 atomic_set(&vec->count, 0);
 217                 if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
 218                         goto cleanup;
 219         }
 220
 221         for_each_possible_cpu(i)
 222                 cp->cpu_to_pri[i] = CPUPRI_INVALID;
 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;
 229 }
 230
 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;
 238
 239         for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
 240                 free_cpumask_var(cp->pri_to_cpu[i].mask);
 241 }