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), NORMAL, RT1, ... RT99, HIGHER
  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(101, nr_domcpus)), though the scenario that
  23  *  yields the worst case search is fairly contrived.
  24  */
  25
  26 /*
  27  * p->rt_priority   p->prio   newpri   cpupri
  28  *
  29  *                                -1       -1 (CPUPRI_INVALID)
  30  *
  31  *                                99        0 (CPUPRI_NORMAL)
  32  *
  33  *              1        98       98        1
  34  *            ...
  35  *             49        50       50       49
  36  *             50        49       49       50
  37  *            ...
  38  *             99         0        0       99
  39  *
  40  *                               100      100 (CPUPRI_HIGHER)
  41  */
  42 static int convert_prio(int prio)
  43 {
  44         int cpupri;
  45
  46         switch (prio) {
  47         case CPUPRI_INVALID:
  48                 cpupri = CPUPRI_INVALID;        /* -1 */
  49                 break;
  50
  51         case 0 ... 98:
  52                 cpupri = MAX_RT_PRIO-1 - prio;  /* 1 ... 99 */
  53                 break;
  54
  55         case MAX_RT_PRIO-1:
  56                 cpupri = CPUPRI_NORMAL;         /*  0 */
  57                 break;
  58
  59         case MAX_RT_PRIO:
  60                 cpupri = CPUPRI_HIGHER;         /* 100 */
  61                 break;
  62         }
  63
  64         return cpupri;
  65 }
  66
  67 static inline int __cpupri_find(struct cpupri *cp, struct task_struct *p,
  68                                 struct cpumask *lowest_mask, int idx)
  69 {
  70         struct cpupri_vec *vec  = &cp->pri_to_cpu[idx];
  71         int skip = 0;
  72
  73         if (!atomic_read(&(vec)->count))
  74                 skip = 1;
  75         /*
  76          * When looking at the vector, we need to read the counter,
  77          * do a memory barrier, then read the mask.
  78          *
  79          * Note: This is still all racy, but we can deal with it.
  80          *  Ideally, we only want to look at masks that are set.
  81          *
  82          *  If a mask is not set, then the only thing wrong is that we
  83          *  did a little more work than necessary.
  84          *
  85          *  If we read a zero count but the mask is set, because of the
  86          *  memory barriers, that can only happen when the highest prio
  87          *  task for a run queue has left the run queue, in which case,
  88          *  it will be followed by a pull. If the task we are processing
  89          *  fails to find a proper place to go, that pull request will
  90          *  pull this task if the run queue is running at a lower
  91          *  priority.
  92          */
  93         smp_rmb();
  94
  95         /* Need to do the rmb for every iteration */
  96         if (skip)
  97                 return 0;
  98
  99         if (cpumask_any_and(&p->cpus_mask, vec->mask) >= nr_cpu_ids)
 100                 return 0;
 101
 102         if (lowest_mask) {
 103                 cpumask_and(lowest_mask, &p->cpus_mask, vec->mask);
 104                 cpumask_and(lowest_mask, lowest_mask, cpu_active_mask);
 105
 106                 /*
 107                  * We have to ensure that we have at least one bit
 108                  * still set in the array, since the map could have
 109                  * been concurrently emptied between the first and
 110                  * second reads of vec->mask.  If we hit this
 111                  * condition, simply act as though we never hit this
 112                  * priority level and continue on.
 113                  */
 114                 if (cpumask_empty(lowest_mask))
 115                         return 0;
 116         }
 117
 118         return 1;
 119 }
 120
 121 int cpupri_find(struct cpupri *cp, struct task_struct *p,
 122                 struct cpumask *lowest_mask)
 123 {
 124         return cpupri_find_fitness(cp, p, lowest_mask, NULL);
 125 }
 126
 127 /**
 128  * cpupri_find_fitness - find the best (lowest-pri) CPU in the system
 129  * @cp: The cpupri context
 130  * @p: The task
 131  * @lowest_mask: A mask to fill in with selected CPUs (or NULL)
 132  * @fitness_fn: A pointer to a function to do custom checks whether the CPU
 133  *              fits a specific criteria so that we only return those CPUs.
 134  *
 135  * Note: This function returns the recommended CPUs as calculated during the
 136  * current invocation.  By the time the call returns, the CPUs may have in
 137  * fact changed priorities any number of times.  While not ideal, it is not
 138  * an issue of correctness since the normal rebalancer logic will correct
 139  * any discrepancies created by racing against the uncertainty of the current
 140  * priority configuration.
 141  *
 142  * Return: (int)bool - CPUs were found
 143  */
 144 int cpupri_find_fitness(struct cpupri *cp, struct task_struct *p,
 145                 struct cpumask *lowest_mask,
 146                 bool (*fitness_fn)(struct task_struct *p, int cpu))
 147 {
 148         int task_pri = convert_prio(p->prio);
 149         int idx, cpu;
 150
 151         WARN_ON_ONCE(task_pri >= CPUPRI_NR_PRIORITIES);
 152
 153         for (idx = 0; idx < task_pri; idx++) {
 154
 155                 if (!__cpupri_find(cp, p, lowest_mask, idx))
 156                         continue;
 157
 158                 if (!lowest_mask || !fitness_fn)
 159                         return 1;
 160
 161                 /* Ensure the capacity of the CPUs fit the task */
 162                 for_each_cpu(cpu, lowest_mask) {
 163                         if (!fitness_fn(p, cpu))
 164                                 cpumask_clear_cpu(cpu, lowest_mask);
 165                 }
 166
 167                 /*
 168                  * If no CPU at the current priority can fit the task
 169                  * continue looking
 170                  */
 171                 if (cpumask_empty(lowest_mask))
 172                         continue;
 173
 174                 return 1;
 175         }
 176
 177         /*
 178          * If we failed to find a fitting lowest_mask, kick off a new search
 179          * but without taking into account any fitness criteria this time.
 180          *
 181          * This rule favours honouring priority over fitting the task in the
 182          * correct CPU (Capacity Awareness being the only user now).
 183          * The idea is that if a higher priority task can run, then it should
 184          * run even if this ends up being on unfitting CPU.
 185          *
 186          * The cost of this trade-off is not entirely clear and will probably
 187          * be good for some workloads and bad for others.
 188          *
 189          * The main idea here is that if some CPUs were over-committed, we try
 190          * to spread which is what the scheduler traditionally did. Sys admins
 191          * must do proper RT planning to avoid overloading the system if they
 192          * really care.
 193          */
 194         if (fitness_fn)
 195                 return cpupri_find(cp, p, lowest_mask);
 196
 197         return 0;
 198 }
 199
 200 /**
 201  * cpupri_set - update the CPU priority setting
 202  * @cp: The cpupri context
 203  * @cpu: The target CPU
 204  * @newpri: The priority (INVALID,NORMAL,RT1-RT99,HIGHER) to assign to this CPU
 205  *
 206  * Note: Assumes cpu_rq(cpu)->lock is locked
 207  *
 208  * Returns: (void)
 209  */
 210 void cpupri_set(struct cpupri *cp, int cpu, int newpri)
 211 {
 212         int *currpri = &cp->cpu_to_pri[cpu];
 213         int oldpri = *currpri;
 214         int do_mb = 0;
 215
 216         newpri = convert_prio(newpri);
 217
 218         BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
 219
 220         if (newpri == oldpri)
 221                 return;
 222
 223         /*
 224          * If the CPU was currently mapped to a different value, we
 225          * need to map it to the new value then remove the old value.
 226          * Note, we must add the new value first, otherwise we risk the
 227          * cpu being missed by the priority loop in cpupri_find.
 228          */
 229         if (likely(newpri != CPUPRI_INVALID)) {
 230                 struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
 231
 232                 cpumask_set_cpu(cpu, vec->mask);
 233                 /*
 234                  * When adding a new vector, we update the mask first,
 235                  * do a write memory barrier, and then update the count, to
 236                  * make sure the vector is visible when count is set.
 237                  */
 238                 smp_mb__before_atomic();
 239                 atomic_inc(&(vec)->count);
 240                 do_mb = 1;
 241         }
 242         if (likely(oldpri != CPUPRI_INVALID)) {
 243                 struct cpupri_vec *vec  = &cp->pri_to_cpu[oldpri];
 244
 245                 /*
 246                  * Because the order of modification of the vec->count
 247                  * is important, we must make sure that the update
 248                  * of the new prio is seen before we decrement the
 249                  * old prio. This makes sure that the loop sees
 250                  * one or the other when we raise the priority of
 251                  * the run queue. We don't care about when we lower the
 252                  * priority, as that will trigger an rt pull anyway.
 253                  *
 254                  * We only need to do a memory barrier if we updated
 255                  * the new priority vec.
 256                  */
 257                 if (do_mb)
 258                         smp_mb__after_atomic();
 259
 260                 /*
 261                  * When removing from the vector, we decrement the counter first
 262                  * do a memory barrier and then clear the mask.
 263                  */
 264                 atomic_dec(&(vec)->count);
 265                 smp_mb__after_atomic();
 266                 cpumask_clear_cpu(cpu, vec->mask);
 267         }
 268
 269         *currpri = newpri;
 270 }
 271
 272 /**
 273  * cpupri_init - initialize the cpupri structure
 274  * @cp: The cpupri context
 275  *
 276  * Return: -ENOMEM on memory allocation failure.
 277  */
 278 int cpupri_init(struct cpupri *cp)
 279 {
 280         int i;
 281
 282         for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
 283                 struct cpupri_vec *vec = &cp->pri_to_cpu[i];
 284
 285                 atomic_set(&vec->count, 0);
 286                 if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
 287                         goto cleanup;
 288         }
 289
 290         cp->cpu_to_pri = kcalloc(nr_cpu_ids, sizeof(int), GFP_KERNEL);
 291         if (!cp->cpu_to_pri)
 292                 goto cleanup;
 293
 294         for_each_possible_cpu(i)
 295                 cp->cpu_to_pri[i] = CPUPRI_INVALID;
 296
 297         return 0;
 298
 299 cleanup:
 300         for (i--; i >= 0; i--)
 301                 free_cpumask_var(cp->pri_to_cpu[i].mask);
 302         return -ENOMEM;
 303 }
 304
 305 /**
 306  * cpupri_cleanup - clean up the cpupri structure
 307  * @cp: The cpupri context
 308  */
 309 void cpupri_cleanup(struct cpupri *cp)
 310 {
 311         int i;
 312
 313         kfree(cp->cpu_to_pri);
 314         for (i = 0; i < CPUPRI_NR_PRIORITIES; i++)
 315                 free_cpumask_var(cp->pri_to_cpu[i].mask);
 316 }