kernel/sched/cpupri.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  kernel/sched/cpupri.c
   4  *
   5  *  CPU priority management
   6  *
   7  *  Copyright (C) 2007-2008 Novell
   8  *
   9  *  Author: Gregory Haskins <ghaskins@novell.com>
  10  *
  11  *  This code tracks the priority of each CPU so that global migration
  12  *  decisions are easy to calculate.  Each CPU can be in a state as follows:
  13  *
  14  *                 (INVALID), IDLE, NORMAL, RT1, ... RT99
  15  *
  16  *  going from the lowest priority to the highest.  CPUs in the INVALID state
  17  *  are not eligible for routing.  The system maintains this state with
  18  *  a 2 dimensional bitmap (the first for priority class, the second for CPUs
  19  *  in that class).  Therefore a typical application without affinity
  20  *  restrictions can find a suitable CPU with O(1) complexity (e.g. two bit
  21  *  searches).  For tasks with affinity restrictions, the algorithm has a
  22  *  worst case complexity of O(min(102, nr_domcpus)), though the scenario that
  23  *  yields the worst case search is fairly contrived.
  24  */
  25 #include "sched.h"
  26
  27 /* Convert between a 140 based task->prio, and our 102 based cpupri */
  28 static int convert_prio(int prio)
  29 {
  30         int cpupri;
  31
  32         if (prio == CPUPRI_INVALID)
  33                 cpupri = CPUPRI_INVALID;
  34         else if (prio == MAX_PRIO)
  35                 cpupri = CPUPRI_IDLE;
  36         else if (prio >= MAX_RT_PRIO)
  37                 cpupri = CPUPRI_NORMAL;
  38         else
  39                 cpupri = MAX_RT_PRIO - prio + 1;
  40
  41         return cpupri;
  42 }
  43
  44 /**
  45  * cpupri_find - find the best (lowest-pri) CPU in the system
  46  * @cp: The cpupri context
  47  * @p: The task
  48  * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
  49  * @fitness_fn: A pointer to a function to do custom checks whether the CPU
  50  *              fits a specific criteria so that we only return those CPUs.
  51  *
  52  * Note: This function returns the recommended CPUs as calculated during the
  53  * current invocation.  By the time the call returns, the CPUs may have in
  54  * fact changed priorities any number of times.  While not ideal, it is not
  55  * an issue of correctness since the normal rebalancer logic will correct
  56  * any discrepancies created by racing against the uncertainty of the current
  57  * priority configuration.
  58  *
  59  * Return: (int)bool - CPUs were found
  60  */
  61 int cpupri_find(struct cpupri *cp, struct task_struct *p,
  62                 struct cpumask *lowest_mask,
  63                 bool (*fitness_fn)(struct task_struct *p, int cpu))
  64 {
  65         int idx = 0;
  66         int task_pri = convert_prio(p->prio);
  67
  68         BUG_ON(task_pri >= CPUPRI_NR_PRIORITIES);
  69
  70         for (idx = 0; idx < task_pri; idx++) {
  71                 struct cpupri_vec *vec  = &cp->pri_to_cpu[idx];
  72                 int skip = 0;
  73
  74                 if (!atomic_read(&(vec)->count))
  75                         skip = 1;
  76                 /*
  77                  * When looking at the vector, we need to read the counter,
  78                  * do a memory barrier, then read the mask.
  79                  *
  80                  * Note: This is still all racey, but we can deal with it.
  81                  *  Ideally, we only want to look at masks that are set.
  82                  *
  83                  *  If a mask is not set, then the only thing wrong is that we
  84                  *  did a little more work than necessary.
  85                  *
  86                  *  If we read a zero count but the mask is set, because of the
  87                  *  memory barriers, that can only happen when the highest prio
  88                  *  task for a run queue has left the run queue, in which case,
  89                  *  it will be followed by a pull. If the task we are processing
  90                  *  fails to find a proper place to go, that pull request will
  91                  *  pull this task if the run queue is running at a lower
  92                  *  priority.
  93                  */
  94                 smp_rmb();
  95
  96                 /* Need to do the rmb for every iteration */
  97                 if (skip)
  98                         continue;
  99
 100                 if (cpumask_any_and(p->cpus_ptr, vec->mask) >= nr_cpu_ids)
 101                         continue;
 102
 103                 if (lowest_mask) {
 104                         int cpu;
 105
 106                         cpumask_and(lowest_mask, p->cpus_ptr, vec->mask);
 107
 108                         /*
 109                          * We have to ensure that we have at least one bit
 110                          * still set in the array, since the map could have
 111                          * been concurrently emptied between the first and
 112                          * second reads of vec->mask.  If we hit this
 113                          * condition, simply act as though we never hit this
 114                          * priority level and continue on.
 115                          */
 116                         if (cpumask_empty(lowest_mask))
 117                                 continue;
 118
 119                         if (!fitness_fn)
 120                                 return 1;
 121
 122                         /* Ensure the capacity of the CPUs fit the task */
 123                         for_each_cpu(cpu, lowest_mask) {
 124                                 if (!fitness_fn(p, cpu))
 125                                         cpumask_clear_cpu(cpu, lowest_mask);
 126                         }
 127
 128                         /*
 129                          * If no CPU at the current priority can fit the task
 130                          * continue looking
 131                          */
 132                         if (cpumask_empty(lowest_mask))
 133                                 continue;
 134                 }
 135
 136                 return 1;
 137         }
 138
 139         return 0;
 140 }
 141
 142 /**
 143  * cpupri_set - update the CPU priority setting
 144  * @cp: The cpupri context
 145  * @cpu: The target CPU
 146  * @newpri: The priority (INVALID-RT99) to assign to this CPU
 147  *
 148  * Note: Assumes cpu_rq(cpu)->lock is locked
 149  *
 150  * Returns: (void)
 151  */
 152 void cpupri_set(struct cpupri *cp, int cpu, int newpri)
 153 {
 154         int *currpri = &cp->cpu_to_pri[cpu];
 155         int oldpri = *currpri;
 156         int do_mb = 0;
 157
 158         newpri = convert_prio(newpri);
 159
 160         BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
 161
 162         if (newpri == oldpri)
 163                 return;
 164
 165         /*
 166          * If the CPU was currently mapped to a different value, we
 167          * need to map it to the new value then remove the old value.
 168          * Note, we must add the new value first, otherwise we risk the
 169          * cpu being missed by the priority loop in cpupri_find.
 170          */
 171         if (likely(newpri != CPUPRI_INVALID)) {
 172                 struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
 173
 174                 cpumask_set_cpu(cpu, vec->mask);
 175                 /*
 176                  * When adding a new vector, we update the mask first,
 177                  * do a write memory barrier, and then update the count, to
 178                  * make sure the vector is visible when count is set.
 179                  */
 180                 smp_mb__before_atomic();
 181                 atomic_inc(&(vec)->count);
 182                 do_mb = 1;
 183         }
 184         if (likely(oldpri != CPUPRI_INVALID)) {
 185                 struct cpupri_vec *vec  = &cp->pri_to_cpu[oldpri];
 186
 187                 /*
 188                  * Because the order of modification of the vec->count
 189                  * is important, we must make sure that the update
 190                  * of the new prio is seen before we decrement the
 191                  * old prio. This makes sure that the loop sees
 192                  * one or the other when we raise the priority of
 193                  * the run queue. We don't care about when we lower the
 194                  * priority, as that will trigger an rt pull anyway.
 195                  *
 196                  * We only need to do a memory barrier if we updated
 197                  * the new priority vec.
 198                  */
 199                 if (do_mb)
 200                         smp_mb__after_atomic();
 201
 202                 /*
 203                  * When removing from the vector, we decrement the counter first
 204                  * do a memory barrier and then clear the mask.
 205                  */
 206                 atomic_dec(&(vec)->count);
 207                 smp_mb__after_atomic();
 208                 cpumask_clear_cpu(cpu, vec->mask);
 209         }
 210
 211         *currpri = newpri;
 212 }
 213
 214 /**
 215  * cpupri_init - initialize the cpupri structure
 216  * @cp: The cpupri context
 217  *
 218  * Return: -ENOMEM on memory allocation failure.
 219  */
 220 int cpupri_init(struct cpupri *cp)
 221 {
 222         int i;
 223
 224         for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
 225                 struct cpupri_vec *vec = &cp->pri_to_cpu[i];
 226
 227                 atomic_set(&vec->count, 0);
 228                 if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
 229                         goto cleanup;
 230         }
 231
 232         cp->cpu_to_pri = kcalloc(nr_cpu_ids, sizeof(int), GFP_KERNEL);
 233         if (!cp->cpu_to_pri)
 234                 goto cleanup;
 235
 236         for_each_possible_cpu(i)
 237                 cp->cpu_to_pri[i] = CPUPRI_INVALID;
 238
 239         return 0;
 240
 241 cleanup:
 242         for (i--; i >= 0; i--)
 243                 free_cpumask_var(cp->pri_to_cpu[i].mask);
 244         return -ENOMEM;
 245 }
 246
 247 /**
 248  * cpupri_cleanup - clean up the cpupri structure
 249  * @cp: The cpupri context
 250  */
 251 void cpupri_cleanup(struct cpupri *cp)
 252 {
 253         int i;
 254
 255         kfree(cp->cpu_to_pri);
 256         for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
 257                 free_cpumask_var(cp->pri_to_cpu[i].mask);
 258 }