|
| 1 | +// SPDX-License-Identifier: GPL-2.0 |
| 2 | +#include "util/bpf_counter.h" |
| 3 | +#include "util/debug.h" |
| 4 | +#include "util/evsel.h" |
| 5 | +#include "util/evlist.h" |
| 6 | +#include "util/off_cpu.h" |
| 7 | +#include "util/perf-hooks.h" |
| 8 | +#include "util/session.h" |
| 9 | +#include <bpf/bpf.h> |
| 10 | + |
| 11 | +#include "bpf_skel/off_cpu.skel.h" |
| 12 | + |
| 13 | +#define MAX_STACKS 32 |
| 14 | +/* we don't need actual timestamp, just want to put the samples at last */ |
| 15 | +#define OFF_CPU_TIMESTAMP (~0ull << 32) |
| 16 | + |
| 17 | +static struct off_cpu_bpf *skel; |
| 18 | + |
| 19 | +struct off_cpu_key { |
| 20 | + u32 pid; |
| 21 | + u32 tgid; |
| 22 | + u32 stack_id; |
| 23 | + u32 state; |
| 24 | +}; |
| 25 | + |
| 26 | +union off_cpu_data { |
| 27 | + struct perf_event_header hdr; |
| 28 | + u64 array[1024 / sizeof(u64)]; |
| 29 | +}; |
| 30 | + |
| 31 | +static int off_cpu_config(struct evlist *evlist) |
| 32 | +{ |
| 33 | + struct evsel *evsel; |
| 34 | + struct perf_event_attr attr = { |
| 35 | + .type = PERF_TYPE_SOFTWARE, |
| 36 | + .config = PERF_COUNT_SW_BPF_OUTPUT, |
| 37 | + .size = sizeof(attr), /* to capture ABI version */ |
| 38 | + }; |
| 39 | + char *evname = strdup(OFFCPU_EVENT); |
| 40 | + |
| 41 | + if (evname == NULL) |
| 42 | + return -ENOMEM; |
| 43 | + |
| 44 | + evsel = evsel__new(&attr); |
| 45 | + if (!evsel) { |
| 46 | + free(evname); |
| 47 | + return -ENOMEM; |
| 48 | + } |
| 49 | + |
| 50 | + evsel->core.attr.freq = 1; |
| 51 | + evsel->core.attr.sample_period = 1; |
| 52 | + /* off-cpu analysis depends on stack trace */ |
| 53 | + evsel->core.attr.sample_type = PERF_SAMPLE_CALLCHAIN; |
| 54 | + |
| 55 | + evlist__add(evlist, evsel); |
| 56 | + |
| 57 | + free(evsel->name); |
| 58 | + evsel->name = evname; |
| 59 | + |
| 60 | + return 0; |
| 61 | +} |
| 62 | + |
| 63 | +static void off_cpu_start(void *arg __maybe_unused) |
| 64 | +{ |
| 65 | + skel->bss->enabled = 1; |
| 66 | +} |
| 67 | + |
| 68 | +static void off_cpu_finish(void *arg __maybe_unused) |
| 69 | +{ |
| 70 | + skel->bss->enabled = 0; |
| 71 | + off_cpu_bpf__destroy(skel); |
| 72 | +} |
| 73 | + |
| 74 | +int off_cpu_prepare(struct evlist *evlist) |
| 75 | +{ |
| 76 | + int err; |
| 77 | + |
| 78 | + if (off_cpu_config(evlist) < 0) { |
| 79 | + pr_err("Failed to config off-cpu BPF event\n"); |
| 80 | + return -1; |
| 81 | + } |
| 82 | + |
| 83 | + set_max_rlimit(); |
| 84 | + |
| 85 | + skel = off_cpu_bpf__open_and_load(); |
| 86 | + if (!skel) { |
| 87 | + pr_err("Failed to open off-cpu BPF skeleton\n"); |
| 88 | + return -1; |
| 89 | + } |
| 90 | + |
| 91 | + err = off_cpu_bpf__attach(skel); |
| 92 | + if (err) { |
| 93 | + pr_err("Failed to attach off-cpu BPF skeleton\n"); |
| 94 | + goto out; |
| 95 | + } |
| 96 | + |
| 97 | + if (perf_hooks__set_hook("record_start", off_cpu_start, NULL) || |
| 98 | + perf_hooks__set_hook("record_end", off_cpu_finish, NULL)) { |
| 99 | + pr_err("Failed to attach off-cpu skeleton\n"); |
| 100 | + goto out; |
| 101 | + } |
| 102 | + |
| 103 | + return 0; |
| 104 | + |
| 105 | +out: |
| 106 | + off_cpu_bpf__destroy(skel); |
| 107 | + return -1; |
| 108 | +} |
| 109 | + |
| 110 | +int off_cpu_write(struct perf_session *session) |
| 111 | +{ |
| 112 | + int bytes = 0, size; |
| 113 | + int fd, stack; |
| 114 | + u64 sample_type, val, sid = 0; |
| 115 | + struct evsel *evsel; |
| 116 | + struct perf_data_file *file = &session->data->file; |
| 117 | + struct off_cpu_key prev, key; |
| 118 | + union off_cpu_data data = { |
| 119 | + .hdr = { |
| 120 | + .type = PERF_RECORD_SAMPLE, |
| 121 | + .misc = PERF_RECORD_MISC_USER, |
| 122 | + }, |
| 123 | + }; |
| 124 | + u64 tstamp = OFF_CPU_TIMESTAMP; |
| 125 | + |
| 126 | + skel->bss->enabled = 0; |
| 127 | + |
| 128 | + evsel = evlist__find_evsel_by_str(session->evlist, OFFCPU_EVENT); |
| 129 | + if (evsel == NULL) { |
| 130 | + pr_err("%s evsel not found\n", OFFCPU_EVENT); |
| 131 | + return 0; |
| 132 | + } |
| 133 | + |
| 134 | + sample_type = evsel->core.attr.sample_type; |
| 135 | + |
| 136 | + if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) { |
| 137 | + if (evsel->core.id) |
| 138 | + sid = evsel->core.id[0]; |
| 139 | + } |
| 140 | + |
| 141 | + fd = bpf_map__fd(skel->maps.off_cpu); |
| 142 | + stack = bpf_map__fd(skel->maps.stacks); |
| 143 | + memset(&prev, 0, sizeof(prev)); |
| 144 | + |
| 145 | + while (!bpf_map_get_next_key(fd, &prev, &key)) { |
| 146 | + int n = 1; /* start from perf_event_header */ |
| 147 | + int ip_pos = -1; |
| 148 | + |
| 149 | + bpf_map_lookup_elem(fd, &key, &val); |
| 150 | + |
| 151 | + if (sample_type & PERF_SAMPLE_IDENTIFIER) |
| 152 | + data.array[n++] = sid; |
| 153 | + if (sample_type & PERF_SAMPLE_IP) { |
| 154 | + ip_pos = n; |
| 155 | + data.array[n++] = 0; /* will be updated */ |
| 156 | + } |
| 157 | + if (sample_type & PERF_SAMPLE_TID) |
| 158 | + data.array[n++] = (u64)key.pid << 32 | key.tgid; |
| 159 | + if (sample_type & PERF_SAMPLE_TIME) |
| 160 | + data.array[n++] = tstamp; |
| 161 | + if (sample_type & PERF_SAMPLE_ID) |
| 162 | + data.array[n++] = sid; |
| 163 | + if (sample_type & PERF_SAMPLE_CPU) |
| 164 | + data.array[n++] = 0; |
| 165 | + if (sample_type & PERF_SAMPLE_PERIOD) |
| 166 | + data.array[n++] = val; |
| 167 | + if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
| 168 | + int len = 0; |
| 169 | + |
| 170 | + /* data.array[n] is callchain->nr (updated later) */ |
| 171 | + data.array[n + 1] = PERF_CONTEXT_USER; |
| 172 | + data.array[n + 2] = 0; |
| 173 | + |
| 174 | + bpf_map_lookup_elem(stack, &key.stack_id, &data.array[n + 2]); |
| 175 | + while (data.array[n + 2 + len]) |
| 176 | + len++; |
| 177 | + |
| 178 | + /* update length of callchain */ |
| 179 | + data.array[n] = len + 1; |
| 180 | + |
| 181 | + /* update sample ip with the first callchain entry */ |
| 182 | + if (ip_pos >= 0) |
| 183 | + data.array[ip_pos] = data.array[n + 2]; |
| 184 | + |
| 185 | + /* calculate sample callchain data array length */ |
| 186 | + n += len + 2; |
| 187 | + } |
| 188 | + /* TODO: handle more sample types */ |
| 189 | + |
| 190 | + size = n * sizeof(u64); |
| 191 | + data.hdr.size = size; |
| 192 | + bytes += size; |
| 193 | + |
| 194 | + if (perf_data_file__write(file, &data, size) < 0) { |
| 195 | + pr_err("failed to write perf data, error: %m\n"); |
| 196 | + return bytes; |
| 197 | + } |
| 198 | + |
| 199 | + prev = key; |
| 200 | + /* increase dummy timestamp to sort later samples */ |
| 201 | + tstamp++; |
| 202 | + } |
| 203 | + return bytes; |
| 204 | +} |
0 commit comments