tools/perf/util/affinity.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Manage affinity to optimize IPIs inside the kernel perf API. */
   3 #define _GNU_SOURCE 1
   4 #include <sched.h>
   5 #include <stdlib.h>
   6 #include <linux/bitmap.h>
   7 #include <linux/zalloc.h>
   8 #include "perf.h"
   9 #include "cpumap.h"
  10 #include "affinity.h"
  11
  12 static int get_cpu_set_size(void)
  13 {
  14         int sz = cpu__max_cpu().cpu + 8 - 1;
  15         /*
  16          * sched_getaffinity doesn't like masks smaller than the kernel.
  17          * Hopefully that's big enough.
  18          */
  19         if (sz < 4096)
  20                 sz = 4096;
  21         return sz / 8;
  22 }
  23
  24 int affinity__setup(struct affinity *a)
  25 {
  26         int cpu_set_size = get_cpu_set_size();
  27
  28         a->orig_cpus = bitmap_zalloc(cpu_set_size * 8);
  29         if (!a->orig_cpus)
  30                 return -1;
  31         sched_getaffinity(0, cpu_set_size, (cpu_set_t *)a->orig_cpus);
  32         a->sched_cpus = bitmap_zalloc(cpu_set_size * 8);
  33         if (!a->sched_cpus) {
  34                 zfree(&a->orig_cpus);
  35                 return -1;
  36         }
  37         bitmap_zero((unsigned long *)a->sched_cpus, cpu_set_size);
  38         a->changed = false;
  39         return 0;
  40 }
  41
  42 /*
  43  * perf_event_open does an IPI internally to the target CPU.
  44  * It is more efficient to change perf's affinity to the target
  45  * CPU and then set up all events on that CPU, so we amortize
  46  * CPU communication.
  47  */
  48 void affinity__set(struct affinity *a, int cpu)
  49 {
  50         int cpu_set_size = get_cpu_set_size();
  51
  52         /*
  53          * Return:
  54          * - if cpu is -1
  55          * - restrict out of bound access to sched_cpus
  56          */
  57         if (cpu == -1 || ((cpu >= (cpu_set_size * 8))))
  58                 return;
  59
  60         a->changed = true;
  61         __set_bit(cpu, a->sched_cpus);
  62         /*
  63          * We ignore errors because affinity is just an optimization.
  64          * This could happen for example with isolated CPUs or cpusets.
  65          * In this case the IPIs inside the kernel's perf API still work.
  66          */
  67         sched_setaffinity(0, cpu_set_size, (cpu_set_t *)a->sched_cpus);
  68         __clear_bit(cpu, a->sched_cpus);
  69 }
  70
  71 static void __affinity__cleanup(struct affinity *a)
  72 {
  73         int cpu_set_size = get_cpu_set_size();
  74
  75         if (a->changed)
  76                 sched_setaffinity(0, cpu_set_size, (cpu_set_t *)a->orig_cpus);
  77         zfree(&a->sched_cpus);
  78         zfree(&a->orig_cpus);
  79 }
  80
  81 void affinity__cleanup(struct affinity *a)
  82 {
  83         if (a != NULL)
  84                 __affinity__cleanup(a);
  85 }