| 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/cpumap.h> |
| 9 | #include "perf.h" |
| 10 | #include "cpumap.h" |
| 11 | #include "affinity.h" |
| 12 | |
| 13 | static int get_cpu_set_size(void) |
| 14 | { |
| 15 | int sz = cpu__max_cpu().cpu + 8 - 1; |
| 16 | /* |
| 17 | * sched_getaffinity doesn't like masks smaller than the kernel. |
| 18 | * Hopefully that's big enough. |
| 19 | */ |
| 20 | if (sz < 4096) |
| 21 | sz = 4096; |
| 22 | return sz / 8; |
| 23 | } |
| 24 | |
| 25 | int affinity__setup(struct affinity *a) |
| 26 | { |
| 27 | int cpu_set_size = get_cpu_set_size(); |
| 28 | |
| 29 | a->orig_cpus = bitmap_zalloc(cpu_set_size * 8); |
| 30 | if (!a->orig_cpus) |
| 31 | return -1; |
| 32 | sched_getaffinity(0, cpu_set_size, (cpu_set_t *)a->orig_cpus); |
| 33 | a->sched_cpus = bitmap_zalloc(cpu_set_size * 8); |
| 34 | if (!a->sched_cpus) { |
| 35 | zfree(&a->orig_cpus); |
| 36 | return -1; |
| 37 | } |
| 38 | bitmap_zero(dst: (unsigned long *)a->sched_cpus, nbits: cpu_set_size); |
| 39 | a->changed = false; |
| 40 | return 0; |
| 41 | } |
| 42 | |
| 43 | /* |
| 44 | * perf_event_open does an IPI internally to the target CPU. |
| 45 | * It is more efficient to change perf's affinity to the target |
| 46 | * CPU and then set up all events on that CPU, so we amortize |
| 47 | * CPU communication. |
| 48 | */ |
| 49 | void affinity__set(struct affinity *a, int cpu) |
| 50 | { |
| 51 | int cpu_set_size = get_cpu_set_size(); |
| 52 | |
| 53 | /* |
| 54 | * Return: |
| 55 | * - if cpu is -1 |
| 56 | * - restrict out of bound access to sched_cpus |
| 57 | */ |
| 58 | if (cpu == -1 || ((cpu >= (cpu_set_size * 8)))) |
| 59 | return; |
| 60 | |
| 61 | a->changed = true; |
| 62 | __set_bit(cpu, a->sched_cpus); |
| 63 | /* |
| 64 | * We ignore errors because affinity is just an optimization. |
| 65 | * This could happen for example with isolated CPUs or cpusets. |
| 66 | * In this case the IPIs inside the kernel's perf API still work. |
| 67 | */ |
| 68 | sched_setaffinity(0, cpu_set_size, (cpu_set_t *)a->sched_cpus); |
| 69 | __clear_bit(cpu, a->sched_cpus); |
| 70 | } |
| 71 | |
| 72 | static void __affinity__cleanup(struct affinity *a) |
| 73 | { |
| 74 | int cpu_set_size = get_cpu_set_size(); |
| 75 | |
| 76 | if (a->changed) |
| 77 | sched_setaffinity(0, cpu_set_size, (cpu_set_t *)a->orig_cpus); |
| 78 | zfree(&a->sched_cpus); |
| 79 | zfree(&a->orig_cpus); |
| 80 | } |
| 81 | |
| 82 | void affinity__cleanup(struct affinity *a) |
| 83 | { |
| 84 | if (a != NULL) |
| 85 | __affinity__cleanup(a); |
| 86 | } |
| 87 | |
| 88 | void cpu_map__set_affinity(const struct perf_cpu_map *cpumap) |
| 89 | { |
| 90 | int cpu_set_size = get_cpu_set_size(); |
| 91 | unsigned long *cpuset = bitmap_zalloc(cpu_set_size * 8); |
| 92 | struct perf_cpu cpu; |
| 93 | int idx; |
| 94 | |
| 95 | if (!cpuset) |
| 96 | return; |
| 97 | |
| 98 | perf_cpu_map__for_each_cpu_skip_any(cpu, idx, cpumap) |
| 99 | __set_bit(cpu.cpu, cpuset); |
| 100 | |
| 101 | sched_setaffinity(0, cpu_set_size, (cpu_set_t *)cpuset); |
| 102 | zfree(&cpuset); |
| 103 | } |
| 104 | |