| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * intel-bts.c: Intel Processor Trace support |
| 4 | * Copyright (c) 2013-2015, Intel Corporation. |
| 5 | */ |
| 6 | |
| 7 | #include <errno.h> |
| 8 | #include <linux/kernel.h> |
| 9 | #include <linux/types.h> |
| 10 | #include <linux/bitops.h> |
| 11 | #include <linux/log2.h> |
| 12 | |
| 13 | #include "../../util/cpumap.h" |
| 14 | #include "../../util/evsel.h" |
| 15 | #include "../../util/evlist.h" |
| 16 | #include "../../util/session.h" |
| 17 | #include "../../util/util.h" |
| 18 | #include "../../util/pmu.h" |
| 19 | #include "../../util/debug.h" |
| 20 | #include "../../util/tsc.h" |
| 21 | #include "../../util/auxtrace.h" |
| 22 | #include "../../util/intel-bts.h" |
| 23 | |
| 24 | #define KiB(x) ((x) * 1024) |
| 25 | #define MiB(x) ((x) * 1024 * 1024) |
| 26 | #define KiB_MASK(x) (KiB(x) - 1) |
| 27 | #define MiB_MASK(x) (MiB(x) - 1) |
| 28 | |
| 29 | struct intel_bts_snapshot_ref { |
| 30 | void *ref_buf; |
| 31 | size_t ref_offset; |
| 32 | bool wrapped; |
| 33 | }; |
| 34 | |
| 35 | struct intel_bts_recording { |
| 36 | struct auxtrace_record itr; |
| 37 | struct perf_pmu *intel_bts_pmu; |
| 38 | struct perf_evlist *evlist; |
| 39 | bool snapshot_mode; |
| 40 | size_t snapshot_size; |
| 41 | int snapshot_ref_cnt; |
| 42 | struct intel_bts_snapshot_ref *snapshot_refs; |
| 43 | }; |
| 44 | |
| 45 | struct branch { |
| 46 | u64 from; |
| 47 | u64 to; |
| 48 | u64 misc; |
| 49 | }; |
| 50 | |
| 51 | static size_t |
| 52 | intel_bts_info_priv_size(struct auxtrace_record *itr __maybe_unused, |
| 53 | struct perf_evlist *evlist __maybe_unused) |
| 54 | { |
| 55 | return INTEL_BTS_AUXTRACE_PRIV_SIZE; |
| 56 | } |
| 57 | |
| 58 | static int intel_bts_info_fill(struct auxtrace_record *itr, |
| 59 | struct perf_session *session, |
| 60 | struct auxtrace_info_event *auxtrace_info, |
| 61 | size_t priv_size) |
| 62 | { |
| 63 | struct intel_bts_recording *btsr = |
| 64 | container_of(itr, struct intel_bts_recording, itr); |
| 65 | struct perf_pmu *intel_bts_pmu = btsr->intel_bts_pmu; |
| 66 | struct perf_event_mmap_page *pc; |
| 67 | struct perf_tsc_conversion tc = { .time_mult = 0, }; |
| 68 | bool cap_user_time_zero = false; |
| 69 | int err; |
| 70 | |
| 71 | if (priv_size != INTEL_BTS_AUXTRACE_PRIV_SIZE) |
| 72 | return -EINVAL; |
| 73 | |
| 74 | if (!session->evlist->nr_mmaps) |
| 75 | return -EINVAL; |
| 76 | |
| 77 | pc = session->evlist->mmap[0].base; |
| 78 | if (pc) { |
| 79 | err = perf_read_tsc_conversion(pc, &tc); |
| 80 | if (err) { |
| 81 | if (err != -EOPNOTSUPP) |
| 82 | return err; |
| 83 | } else { |
| 84 | cap_user_time_zero = tc.time_mult != 0; |
| 85 | } |
| 86 | if (!cap_user_time_zero) |
| 87 | ui__warning("Intel BTS: TSC not available\n"); |
| 88 | } |
| 89 | |
| 90 | auxtrace_info->type = PERF_AUXTRACE_INTEL_BTS; |
| 91 | auxtrace_info->priv[INTEL_BTS_PMU_TYPE] = intel_bts_pmu->type; |
| 92 | auxtrace_info->priv[INTEL_BTS_TIME_SHIFT] = tc.time_shift; |
| 93 | auxtrace_info->priv[INTEL_BTS_TIME_MULT] = tc.time_mult; |
| 94 | auxtrace_info->priv[INTEL_BTS_TIME_ZERO] = tc.time_zero; |
| 95 | auxtrace_info->priv[INTEL_BTS_CAP_USER_TIME_ZERO] = cap_user_time_zero; |
| 96 | auxtrace_info->priv[INTEL_BTS_SNAPSHOT_MODE] = btsr->snapshot_mode; |
| 97 | |
| 98 | return 0; |
| 99 | } |
| 100 | |
| 101 | static int intel_bts_recording_options(struct auxtrace_record *itr, |
| 102 | struct perf_evlist *evlist, |
| 103 | struct record_opts *opts) |
| 104 | { |
| 105 | struct intel_bts_recording *btsr = |
| 106 | container_of(itr, struct intel_bts_recording, itr); |
| 107 | struct perf_pmu *intel_bts_pmu = btsr->intel_bts_pmu; |
| 108 | struct perf_evsel *evsel, *intel_bts_evsel = NULL; |
| 109 | const struct cpu_map *cpus = evlist->cpus; |
| 110 | bool privileged = geteuid() == 0 || perf_event_paranoid() < 0; |
| 111 | |
| 112 | btsr->evlist = evlist; |
| 113 | btsr->snapshot_mode = opts->auxtrace_snapshot_mode; |
| 114 | |
| 115 | evlist__for_each_entry(evlist, evsel) { |
| 116 | if (evsel->attr.type == intel_bts_pmu->type) { |
| 117 | if (intel_bts_evsel) { |
| 118 | pr_err("There may be only one " INTEL_BTS_PMU_NAME " event\n"); |
| 119 | return -EINVAL; |
| 120 | } |
| 121 | evsel->attr.freq = 0; |
| 122 | evsel->attr.sample_period = 1; |
| 123 | intel_bts_evsel = evsel; |
| 124 | opts->full_auxtrace = true; |
| 125 | } |
| 126 | } |
| 127 | |
| 128 | if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) { |
| 129 | pr_err("Snapshot mode (-S option) requires " INTEL_BTS_PMU_NAME " PMU event (-e " INTEL_BTS_PMU_NAME ")\n"); |
| 130 | return -EINVAL; |
| 131 | } |
| 132 | |
| 133 | if (!opts->full_auxtrace) |
| 134 | return 0; |
| 135 | |
| 136 | if (opts->full_auxtrace && !cpu_map__empty(cpus)) { |
| 137 | pr_err(INTEL_BTS_PMU_NAME " does not support per-cpu recording\n"); |
| 138 | return -EINVAL; |
| 139 | } |
| 140 | |
| 141 | /* Set default sizes for snapshot mode */ |
| 142 | if (opts->auxtrace_snapshot_mode) { |
| 143 | if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) { |
| 144 | if (privileged) { |
| 145 | opts->auxtrace_mmap_pages = MiB(4) / page_size; |
| 146 | } else { |
| 147 | opts->auxtrace_mmap_pages = KiB(128) / page_size; |
| 148 | if (opts->mmap_pages == UINT_MAX) |
| 149 | opts->mmap_pages = KiB(256) / page_size; |
| 150 | } |
| 151 | } else if (!opts->auxtrace_mmap_pages && !privileged && |
| 152 | opts->mmap_pages == UINT_MAX) { |
| 153 | opts->mmap_pages = KiB(256) / page_size; |
| 154 | } |
| 155 | if (!opts->auxtrace_snapshot_size) |
| 156 | opts->auxtrace_snapshot_size = |
| 157 | opts->auxtrace_mmap_pages * (size_t)page_size; |
| 158 | if (!opts->auxtrace_mmap_pages) { |
| 159 | size_t sz = opts->auxtrace_snapshot_size; |
| 160 | |
| 161 | sz = round_up(sz, page_size) / page_size; |
| 162 | opts->auxtrace_mmap_pages = roundup_pow_of_two(sz); |
| 163 | } |
| 164 | if (opts->auxtrace_snapshot_size > |
| 165 | opts->auxtrace_mmap_pages * (size_t)page_size) { |
| 166 | pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n", |
| 167 | opts->auxtrace_snapshot_size, |
| 168 | opts->auxtrace_mmap_pages * (size_t)page_size); |
| 169 | return -EINVAL; |
| 170 | } |
| 171 | if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) { |
| 172 | pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n"); |
| 173 | return -EINVAL; |
| 174 | } |
| 175 | pr_debug2("Intel BTS snapshot size: %zu\n", |
| 176 | opts->auxtrace_snapshot_size); |
| 177 | } |
| 178 | |
| 179 | /* Set default sizes for full trace mode */ |
| 180 | if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) { |
| 181 | if (privileged) { |
| 182 | opts->auxtrace_mmap_pages = MiB(4) / page_size; |
| 183 | } else { |
| 184 | opts->auxtrace_mmap_pages = KiB(128) / page_size; |
| 185 | if (opts->mmap_pages == UINT_MAX) |
| 186 | opts->mmap_pages = KiB(256) / page_size; |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | /* Validate auxtrace_mmap_pages */ |
| 191 | if (opts->auxtrace_mmap_pages) { |
| 192 | size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size; |
| 193 | size_t min_sz; |
| 194 | |
| 195 | if (opts->auxtrace_snapshot_mode) |
| 196 | min_sz = KiB(4); |
| 197 | else |
| 198 | min_sz = KiB(8); |
| 199 | |
| 200 | if (sz < min_sz || !is_power_of_2(sz)) { |
| 201 | pr_err("Invalid mmap size for Intel BTS: must be at least %zuKiB and a power of 2\n", |
| 202 | min_sz / 1024); |
| 203 | return -EINVAL; |
| 204 | } |
| 205 | } |
| 206 | |
| 207 | if (intel_bts_evsel) { |
| 208 | /* |
| 209 | * To obtain the auxtrace buffer file descriptor, the auxtrace event |
| 210 | * must come first. |
| 211 | */ |
| 212 | perf_evlist__to_front(evlist, intel_bts_evsel); |
| 213 | /* |
| 214 | * In the case of per-cpu mmaps, we need the CPU on the |
| 215 | * AUX event. |
| 216 | */ |
| 217 | if (!cpu_map__empty(cpus)) |
| 218 | perf_evsel__set_sample_bit(intel_bts_evsel, CPU); |
| 219 | } |
| 220 | |
| 221 | /* Add dummy event to keep tracking */ |
| 222 | if (opts->full_auxtrace) { |
| 223 | struct perf_evsel *tracking_evsel; |
| 224 | int err; |
| 225 | |
| 226 | err = parse_events(evlist, "dummy:u", NULL); |
| 227 | if (err) |
| 228 | return err; |
| 229 | |
| 230 | tracking_evsel = perf_evlist__last(evlist); |
| 231 | |
| 232 | perf_evlist__set_tracking_event(evlist, tracking_evsel); |
| 233 | |
| 234 | tracking_evsel->attr.freq = 0; |
| 235 | tracking_evsel->attr.sample_period = 1; |
| 236 | } |
| 237 | |
| 238 | return 0; |
| 239 | } |
| 240 | |
| 241 | static int intel_bts_parse_snapshot_options(struct auxtrace_record *itr, |
| 242 | struct record_opts *opts, |
| 243 | const char *str) |
| 244 | { |
| 245 | struct intel_bts_recording *btsr = |
| 246 | container_of(itr, struct intel_bts_recording, itr); |
| 247 | unsigned long long snapshot_size = 0; |
| 248 | char *endptr; |
| 249 | |
| 250 | if (str) { |
| 251 | snapshot_size = strtoull(str, &endptr, 0); |
| 252 | if (*endptr || snapshot_size > SIZE_MAX) |
| 253 | return -1; |
| 254 | } |
| 255 | |
| 256 | opts->auxtrace_snapshot_mode = true; |
| 257 | opts->auxtrace_snapshot_size = snapshot_size; |
| 258 | |
| 259 | btsr->snapshot_size = snapshot_size; |
| 260 | |
| 261 | return 0; |
| 262 | } |
| 263 | |
| 264 | static u64 intel_bts_reference(struct auxtrace_record *itr __maybe_unused) |
| 265 | { |
| 266 | return rdtsc(); |
| 267 | } |
| 268 | |
| 269 | static int intel_bts_alloc_snapshot_refs(struct intel_bts_recording *btsr, |
| 270 | int idx) |
| 271 | { |
| 272 | const size_t sz = sizeof(struct intel_bts_snapshot_ref); |
| 273 | int cnt = btsr->snapshot_ref_cnt, new_cnt = cnt * 2; |
| 274 | struct intel_bts_snapshot_ref *refs; |
| 275 | |
| 276 | if (!new_cnt) |
| 277 | new_cnt = 16; |
| 278 | |
| 279 | while (new_cnt <= idx) |
| 280 | new_cnt *= 2; |
| 281 | |
| 282 | refs = calloc(new_cnt, sz); |
| 283 | if (!refs) |
| 284 | return -ENOMEM; |
| 285 | |
| 286 | memcpy(refs, btsr->snapshot_refs, cnt * sz); |
| 287 | |
| 288 | btsr->snapshot_refs = refs; |
| 289 | btsr->snapshot_ref_cnt = new_cnt; |
| 290 | |
| 291 | return 0; |
| 292 | } |
| 293 | |
| 294 | static void intel_bts_free_snapshot_refs(struct intel_bts_recording *btsr) |
| 295 | { |
| 296 | int i; |
| 297 | |
| 298 | for (i = 0; i < btsr->snapshot_ref_cnt; i++) |
| 299 | zfree(&btsr->snapshot_refs[i].ref_buf); |
| 300 | zfree(&btsr->snapshot_refs); |
| 301 | } |
| 302 | |
| 303 | static void intel_bts_recording_free(struct auxtrace_record *itr) |
| 304 | { |
| 305 | struct intel_bts_recording *btsr = |
| 306 | container_of(itr, struct intel_bts_recording, itr); |
| 307 | |
| 308 | intel_bts_free_snapshot_refs(btsr); |
| 309 | free(btsr); |
| 310 | } |
| 311 | |
| 312 | static int intel_bts_snapshot_start(struct auxtrace_record *itr) |
| 313 | { |
| 314 | struct intel_bts_recording *btsr = |
| 315 | container_of(itr, struct intel_bts_recording, itr); |
| 316 | struct perf_evsel *evsel; |
| 317 | |
| 318 | evlist__for_each_entry(btsr->evlist, evsel) { |
| 319 | if (evsel->attr.type == btsr->intel_bts_pmu->type) |
| 320 | return perf_evsel__disable(evsel); |
| 321 | } |
| 322 | return -EINVAL; |
| 323 | } |
| 324 | |
| 325 | static int intel_bts_snapshot_finish(struct auxtrace_record *itr) |
| 326 | { |
| 327 | struct intel_bts_recording *btsr = |
| 328 | container_of(itr, struct intel_bts_recording, itr); |
| 329 | struct perf_evsel *evsel; |
| 330 | |
| 331 | evlist__for_each_entry(btsr->evlist, evsel) { |
| 332 | if (evsel->attr.type == btsr->intel_bts_pmu->type) |
| 333 | return perf_evsel__enable(evsel); |
| 334 | } |
| 335 | return -EINVAL; |
| 336 | } |
| 337 | |
| 338 | static bool intel_bts_first_wrap(u64 *data, size_t buf_size) |
| 339 | { |
| 340 | int i, a, b; |
| 341 | |
| 342 | b = buf_size >> 3; |
| 343 | a = b - 512; |
| 344 | if (a < 0) |
| 345 | a = 0; |
| 346 | |
| 347 | for (i = a; i < b; i++) { |
| 348 | if (data[i]) |
| 349 | return true; |
| 350 | } |
| 351 | |
| 352 | return false; |
| 353 | } |
| 354 | |
| 355 | static int intel_bts_find_snapshot(struct auxtrace_record *itr, int idx, |
| 356 | struct auxtrace_mmap *mm, unsigned char *data, |
| 357 | u64 *head, u64 *old) |
| 358 | { |
| 359 | struct intel_bts_recording *btsr = |
| 360 | container_of(itr, struct intel_bts_recording, itr); |
| 361 | bool wrapped; |
| 362 | int err; |
| 363 | |
| 364 | pr_debug3("%s: mmap index %d old head %zu new head %zu\n", |
| 365 | __func__, idx, (size_t)*old, (size_t)*head); |
| 366 | |
| 367 | if (idx >= btsr->snapshot_ref_cnt) { |
| 368 | err = intel_bts_alloc_snapshot_refs(btsr, idx); |
| 369 | if (err) |
| 370 | goto out_err; |
| 371 | } |
| 372 | |
| 373 | wrapped = btsr->snapshot_refs[idx].wrapped; |
| 374 | if (!wrapped && intel_bts_first_wrap((u64 *)data, mm->len)) { |
| 375 | btsr->snapshot_refs[idx].wrapped = true; |
| 376 | wrapped = true; |
| 377 | } |
| 378 | |
| 379 | /* |
| 380 | * In full trace mode 'head' continually increases. However in snapshot |
| 381 | * mode 'head' is an offset within the buffer. Here 'old' and 'head' |
| 382 | * are adjusted to match the full trace case which expects that 'old' is |
| 383 | * always less than 'head'. |
| 384 | */ |
| 385 | if (wrapped) { |
| 386 | *old = *head; |
| 387 | *head += mm->len; |
| 388 | } else { |
| 389 | if (mm->mask) |
| 390 | *old &= mm->mask; |
| 391 | else |
| 392 | *old %= mm->len; |
| 393 | if (*old > *head) |
| 394 | *head += mm->len; |
| 395 | } |
| 396 | |
| 397 | pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n", |
| 398 | __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head); |
| 399 | |
| 400 | return 0; |
| 401 | |
| 402 | out_err: |
| 403 | pr_err("%s: failed, error %d\n", __func__, err); |
| 404 | return err; |
| 405 | } |
| 406 | |
| 407 | static int intel_bts_read_finish(struct auxtrace_record *itr, int idx) |
| 408 | { |
| 409 | struct intel_bts_recording *btsr = |
| 410 | container_of(itr, struct intel_bts_recording, itr); |
| 411 | struct perf_evsel *evsel; |
| 412 | |
| 413 | evlist__for_each_entry(btsr->evlist, evsel) { |
| 414 | if (evsel->attr.type == btsr->intel_bts_pmu->type) |
| 415 | return perf_evlist__enable_event_idx(btsr->evlist, |
| 416 | evsel, idx); |
| 417 | } |
| 418 | return -EINVAL; |
| 419 | } |
| 420 | |
| 421 | struct auxtrace_record *intel_bts_recording_init(int *err) |
| 422 | { |
| 423 | struct perf_pmu *intel_bts_pmu = perf_pmu__find(INTEL_BTS_PMU_NAME); |
| 424 | struct intel_bts_recording *btsr; |
| 425 | |
| 426 | if (!intel_bts_pmu) |
| 427 | return NULL; |
| 428 | |
| 429 | if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) { |
| 430 | *err = -errno; |
| 431 | return NULL; |
| 432 | } |
| 433 | |
| 434 | btsr = zalloc(sizeof(struct intel_bts_recording)); |
| 435 | if (!btsr) { |
| 436 | *err = -ENOMEM; |
| 437 | return NULL; |
| 438 | } |
| 439 | |
| 440 | btsr->intel_bts_pmu = intel_bts_pmu; |
| 441 | btsr->itr.recording_options = intel_bts_recording_options; |
| 442 | btsr->itr.info_priv_size = intel_bts_info_priv_size; |
| 443 | btsr->itr.info_fill = intel_bts_info_fill; |
| 444 | btsr->itr.free = intel_bts_recording_free; |
| 445 | btsr->itr.snapshot_start = intel_bts_snapshot_start; |
| 446 | btsr->itr.snapshot_finish = intel_bts_snapshot_finish; |
| 447 | btsr->itr.find_snapshot = intel_bts_find_snapshot; |
| 448 | btsr->itr.parse_snapshot_options = intel_bts_parse_snapshot_options; |
| 449 | btsr->itr.reference = intel_bts_reference; |
| 450 | btsr->itr.read_finish = intel_bts_read_finish; |
| 451 | btsr->itr.alignment = sizeof(struct branch); |
| 452 | return &btsr->itr; |
| 453 | } |