brintos

brintos / linux-shallow public Read only

0
0
Text · 31.1 KiB · ea510a7 Raw
1217 lines · c
1// SPDX-License-Identifier: GPL-2.0-only2/*3 * intel_pt.c: Intel Processor Trace support4 * Copyright (c) 2013-2015, Intel Corporation.5 */6 7#include <errno.h>8#include <stdbool.h>9#include <linux/kernel.h>10#include <linux/types.h>11#include <linux/bitops.h>12#include <linux/log2.h>13#include <linux/zalloc.h>14#include <linux/err.h>15#include <cpuid.h>16 17#include "../../../util/session.h"18#include "../../../util/event.h"19#include "../../../util/evlist.h"20#include "../../../util/evsel.h"21#include "../../../util/evsel_config.h"22#include "../../../util/cpumap.h"23#include "../../../util/mmap.h"24#include <subcmd/parse-options.h>25#include "../../../util/parse-events.h"26#include "../../../util/pmus.h"27#include "../../../util/debug.h"28#include "../../../util/auxtrace.h"29#include "../../../util/perf_api_probe.h"30#include "../../../util/record.h"31#include "../../../util/target.h"32#include "../../../util/tsc.h"33#include <internal/lib.h> // page_size34#include "../../../util/intel-pt.h"35#include <api/fs/fs.h>36 37#define KiB(x) ((x) * 1024)38#define MiB(x) ((x) * 1024 * 1024)39#define KiB_MASK(x) (KiB(x) - 1)40#define MiB_MASK(x) (MiB(x) - 1)41 42#define INTEL_PT_PSB_PERIOD_NEAR	25643 44struct intel_pt_snapshot_ref {45	void *ref_buf;46	size_t ref_offset;47	bool wrapped;48};49 50struct intel_pt_recording {51	struct auxtrace_record		itr;52	struct perf_pmu			*intel_pt_pmu;53	int				have_sched_switch;54	struct evlist		*evlist;55	bool				snapshot_mode;56	bool				snapshot_init_done;57	size_t				snapshot_size;58	size_t				snapshot_ref_buf_size;59	int				snapshot_ref_cnt;60	struct intel_pt_snapshot_ref	*snapshot_refs;61	size_t				priv_size;62};63 64static int intel_pt_parse_terms_with_default(const struct perf_pmu *pmu,65					     const char *str,66					     u64 *config)67{68	struct parse_events_terms terms;69	struct perf_event_attr attr = { .size = 0, };70	int err;71 72	parse_events_terms__init(&terms);73	err = parse_events_terms(&terms, str, /*input=*/ NULL);74	if (err)75		goto out_free;76 77	attr.config = *config;78	err = perf_pmu__config_terms(pmu, &attr, &terms, /*zero=*/true, /*err=*/NULL);79	if (err)80		goto out_free;81 82	*config = attr.config;83out_free:84	parse_events_terms__exit(&terms);85	return err;86}87 88static int intel_pt_parse_terms(const struct perf_pmu *pmu, const char *str, u64 *config)89{90	*config = 0;91	return intel_pt_parse_terms_with_default(pmu, str, config);92}93 94static u64 intel_pt_masked_bits(u64 mask, u64 bits)95{96	const u64 top_bit = 1ULL << 63;97	u64 res = 0;98	int i;99 100	for (i = 0; i < 64; i++) {101		if (mask & top_bit) {102			res <<= 1;103			if (bits & top_bit)104				res |= 1;105		}106		mask <<= 1;107		bits <<= 1;108	}109 110	return res;111}112 113static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,114				struct evlist *evlist, u64 *res)115{116	struct evsel *evsel;117	u64 mask;118 119	*res = 0;120 121	mask = perf_pmu__format_bits(intel_pt_pmu, str);122	if (!mask)123		return -EINVAL;124 125	evlist__for_each_entry(evlist, evsel) {126		if (evsel->core.attr.type == intel_pt_pmu->type) {127			*res = intel_pt_masked_bits(mask, evsel->core.attr.config);128			return 0;129		}130	}131 132	return -EINVAL;133}134 135static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,136				  struct evlist *evlist)137{138	u64 val;139	int err, topa_multiple_entries;140	size_t psb_period;141 142	if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",143				"%d", &topa_multiple_entries) != 1)144		topa_multiple_entries = 0;145 146	/*147	 * Use caps/topa_multiple_entries to indicate early hardware that had148	 * extra frequent PSBs.149	 */150	if (!topa_multiple_entries) {151		psb_period = 256;152		goto out;153	}154 155	err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);156	if (err)157		val = 0;158 159	psb_period = 1 << (val + 11);160out:161	pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);162	return psb_period;163}164 165static int intel_pt_pick_bit(int bits, int target)166{167	int pos, pick = -1;168 169	for (pos = 0; bits; bits >>= 1, pos++) {170		if (bits & 1) {171			if (pos <= target || pick < 0)172				pick = pos;173			if (pos >= target)174				break;175		}176	}177 178	return pick;179}180 181static u64 intel_pt_default_config(const struct perf_pmu *intel_pt_pmu)182{183	char buf[256];184	int mtc, mtc_periods = 0, mtc_period;185	int psb_cyc, psb_periods, psb_period;186	int pos = 0;187	u64 config;188	char c;189	int dirfd;190 191	dirfd = perf_pmu__event_source_devices_fd();192 193	pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");194 195	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/mtc", "%d",196				   &mtc) != 1)197		mtc = 1;198 199	if (mtc) {200		if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/mtc_periods", "%x",201					   &mtc_periods) != 1)202			mtc_periods = 0;203		if (mtc_periods) {204			mtc_period = intel_pt_pick_bit(mtc_periods, 3);205			pos += scnprintf(buf + pos, sizeof(buf) - pos,206					 ",mtc,mtc_period=%d", mtc_period);207		}208	}209 210	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/psb_cyc", "%d",211				   &psb_cyc) != 1)212		psb_cyc = 1;213 214	if (psb_cyc && mtc_periods) {215		if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "caps/psb_periods", "%x",216					   &psb_periods) != 1)217			psb_periods = 0;218		if (psb_periods) {219			psb_period = intel_pt_pick_bit(psb_periods, 3);220			pos += scnprintf(buf + pos, sizeof(buf) - pos,221					 ",psb_period=%d", psb_period);222		}223	}224 225	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "format/pt", "%c", &c) == 1 &&226	    perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "format/branch", "%c", &c) == 1)227		pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");228 229	pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);230 231	intel_pt_parse_terms(intel_pt_pmu, buf, &config);232 233	close(dirfd);234	return config;235}236 237static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,238					   struct record_opts *opts,239					   const char *str)240{241	struct intel_pt_recording *ptr =242			container_of(itr, struct intel_pt_recording, itr);243	unsigned long long snapshot_size = 0;244	char *endptr;245 246	if (str) {247		snapshot_size = strtoull(str, &endptr, 0);248		if (*endptr || snapshot_size > SIZE_MAX)249			return -1;250	}251 252	opts->auxtrace_snapshot_mode = true;253	opts->auxtrace_snapshot_size = snapshot_size;254 255	ptr->snapshot_size = snapshot_size;256 257	return 0;258}259 260void intel_pt_pmu_default_config(const struct perf_pmu *intel_pt_pmu,261				 struct perf_event_attr *attr)262{263	static u64 config;264	static bool initialized;265 266	if (!initialized) {267		config = intel_pt_default_config(intel_pt_pmu);268		initialized = true;269	}270	attr->config = config;271}272 273static const char *intel_pt_find_filter(struct evlist *evlist,274					struct perf_pmu *intel_pt_pmu)275{276	struct evsel *evsel;277 278	evlist__for_each_entry(evlist, evsel) {279		if (evsel->core.attr.type == intel_pt_pmu->type)280			return evsel->filter;281	}282 283	return NULL;284}285 286static size_t intel_pt_filter_bytes(const char *filter)287{288	size_t len = filter ? strlen(filter) : 0;289 290	return len ? roundup(len + 1, 8) : 0;291}292 293static size_t294intel_pt_info_priv_size(struct auxtrace_record *itr, struct evlist *evlist)295{296	struct intel_pt_recording *ptr =297			container_of(itr, struct intel_pt_recording, itr);298	const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);299 300	ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +301			 intel_pt_filter_bytes(filter);302	ptr->priv_size += sizeof(u64); /* Cap Event Trace */303 304	return ptr->priv_size;305}306 307static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)308{309	unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;310 311	__get_cpuid(0x15, &eax, &ebx, &ecx, &edx);312	*n = ebx;313	*d = eax;314}315 316static int intel_pt_info_fill(struct auxtrace_record *itr,317			      struct perf_session *session,318			      struct perf_record_auxtrace_info *auxtrace_info,319			      size_t priv_size)320{321	struct intel_pt_recording *ptr =322			container_of(itr, struct intel_pt_recording, itr);323	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;324	struct perf_event_mmap_page *pc;325	struct perf_tsc_conversion tc = { .time_mult = 0, };326	bool cap_user_time_zero = false, per_cpu_mmaps;327	u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;328	u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;329	unsigned long max_non_turbo_ratio;330	size_t filter_str_len;331	const char *filter;332	int event_trace;333	__u64 *info;334	int err;335 336	if (priv_size != ptr->priv_size)337		return -EINVAL;338 339	intel_pt_parse_terms(intel_pt_pmu, "tsc", &tsc_bit);340	intel_pt_parse_terms(intel_pt_pmu, "noretcomp", &noretcomp_bit);341	intel_pt_parse_terms(intel_pt_pmu, "mtc", &mtc_bit);342	mtc_freq_bits = perf_pmu__format_bits(intel_pt_pmu, "mtc_period");343	intel_pt_parse_terms(intel_pt_pmu, "cyc", &cyc_bit);344 345	intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);346 347	if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",348				"%lu", &max_non_turbo_ratio) != 1)349		max_non_turbo_ratio = 0;350	if (perf_pmu__scan_file(intel_pt_pmu, "caps/event_trace",351				"%d", &event_trace) != 1)352		event_trace = 0;353 354	filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);355	filter_str_len = filter ? strlen(filter) : 0;356 357	if (!session->evlist->core.nr_mmaps)358		return -EINVAL;359 360	pc = session->evlist->mmap[0].core.base;361	if (pc) {362		err = perf_read_tsc_conversion(pc, &tc);363		if (err) {364			if (err != -EOPNOTSUPP)365				return err;366		} else {367			cap_user_time_zero = tc.time_mult != 0;368		}369		if (!cap_user_time_zero)370			ui__warning("Intel Processor Trace: TSC not available\n");371	}372 373	per_cpu_mmaps = !perf_cpu_map__is_any_cpu_or_is_empty(session->evlist->core.user_requested_cpus);374 375	auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;376	auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;377	auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;378	auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;379	auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;380	auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;381	auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;382	auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;383	auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;384	auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;385	auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;386	auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;387	auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;388	auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;389	auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;390	auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;391	auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;392	auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;393 394	info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;395 396	if (filter_str_len) {397		size_t len = intel_pt_filter_bytes(filter);398 399		strncpy((char *)info, filter, len);400		info += len >> 3;401	}402 403	*info++ = event_trace;404 405	return 0;406}407 408#ifdef HAVE_LIBTRACEEVENT409static int intel_pt_track_switches(struct evlist *evlist)410{411	const char *sched_switch = "sched:sched_switch";412	struct evsel *evsel;413	int err;414 415	if (!evlist__can_select_event(evlist, sched_switch))416		return -EPERM;417 418	evsel = evlist__add_sched_switch(evlist, true);419	if (IS_ERR(evsel)) {420		err = PTR_ERR(evsel);421		pr_debug2("%s: failed to create %s, error = %d\n",422			  __func__, sched_switch, err);423		return err;424	}425 426	evsel->immediate = true;427 428	return 0;429}430#endif431 432static bool intel_pt_exclude_guest(void)433{434	int pt_mode;435 436	if (sysfs__read_int("module/kvm_intel/parameters/pt_mode", &pt_mode))437		pt_mode = 0;438 439	return pt_mode == 1;440}441 442static void intel_pt_valid_str(char *str, size_t len, u64 valid)443{444	unsigned int val, last = 0, state = 1;445	int p = 0;446 447	str[0] = '\0';448 449	for (val = 0; val <= 64; val++, valid >>= 1) {450		if (valid & 1) {451			last = val;452			switch (state) {453			case 0:454				p += scnprintf(str + p, len - p, ",");455				/* Fall through */456			case 1:457				p += scnprintf(str + p, len - p, "%u", val);458				state = 2;459				break;460			case 2:461				state = 3;462				break;463			case 3:464				state = 4;465				break;466			default:467				break;468			}469		} else {470			switch (state) {471			case 3:472				p += scnprintf(str + p, len - p, ",%u", last);473				state = 0;474				break;475			case 4:476				p += scnprintf(str + p, len - p, "-%u", last);477				state = 0;478				break;479			default:480				break;481			}482			if (state != 1)483				state = 0;484		}485	}486}487 488static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu, int dirfd,489				    const char *caps, const char *name,490				    const char *supported, u64 config)491{492	char valid_str[256];493	unsigned int shift;494	unsigned long long valid;495	u64 bits;496	int ok;497 498	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, caps, "%llx", &valid) != 1)499		valid = 0;500 501	if (supported &&502	    perf_pmu__scan_file_at(intel_pt_pmu, dirfd, supported, "%d", &ok) == 1 && !ok)503		valid = 0;504 505	valid |= 1;506 507	bits = perf_pmu__format_bits(intel_pt_pmu, name);508 509	config &= bits;510 511	for (shift = 0; bits && !(bits & 1); shift++)512		bits >>= 1;513 514	config >>= shift;515 516	if (config > 63)517		goto out_err;518 519	if (valid & (1 << config))520		return 0;521out_err:522	intel_pt_valid_str(valid_str, sizeof(valid_str), valid);523	pr_err("Invalid %s for %s. Valid values are: %s\n",524	       name, INTEL_PT_PMU_NAME, valid_str);525	return -EINVAL;526}527 528static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,529				    struct evsel *evsel)530{531	int err, dirfd;532	char c;533 534	if (!evsel)535		return 0;536 537	dirfd = perf_pmu__event_source_devices_fd();538	if (dirfd < 0)539		return dirfd;540 541	/*542	 * If supported, force pass-through config term (pt=1) even if user543	 * sets pt=0, which avoids senseless kernel errors.544	 */545	if (perf_pmu__scan_file_at(intel_pt_pmu, dirfd, "format/pt", "%c", &c) == 1 &&546	    !(evsel->core.attr.config & 1)) {547		pr_warning("pt=0 doesn't make sense, forcing pt=1\n");548		evsel->core.attr.config |= 1;549	}550 551	err = intel_pt_val_config_term(intel_pt_pmu, dirfd, "caps/cycle_thresholds",552				       "cyc_thresh", "caps/psb_cyc",553				       evsel->core.attr.config);554	if (err)555		goto out;556 557	err = intel_pt_val_config_term(intel_pt_pmu, dirfd, "caps/mtc_periods",558				       "mtc_period", "caps/mtc",559				       evsel->core.attr.config);560	if (err)561		goto out;562 563	err = intel_pt_val_config_term(intel_pt_pmu, dirfd, "caps/psb_periods",564					"psb_period", "caps/psb_cyc",565					evsel->core.attr.config);566 567out:568	close(dirfd);569	return err;570}571 572static void intel_pt_min_max_sample_sz(struct evlist *evlist,573				       size_t *min_sz, size_t *max_sz)574{575	struct evsel *evsel;576 577	evlist__for_each_entry(evlist, evsel) {578		size_t sz = evsel->core.attr.aux_sample_size;579 580		if (!sz)581			continue;582		if (min_sz && (sz < *min_sz || !*min_sz))583			*min_sz = sz;584		if (max_sz && sz > *max_sz)585			*max_sz = sz;586	}587}588 589/*590 * Currently, there is not enough information to disambiguate different PEBS591 * events, so only allow one.592 */593static bool intel_pt_too_many_aux_output(struct evlist *evlist)594{595	struct evsel *evsel;596	int aux_output_cnt = 0;597 598	evlist__for_each_entry(evlist, evsel)599		aux_output_cnt += !!evsel->core.attr.aux_output;600 601	if (aux_output_cnt > 1) {602		pr_err(INTEL_PT_PMU_NAME " supports at most one event with aux-output\n");603		return true;604	}605 606	return false;607}608 609static int intel_pt_recording_options(struct auxtrace_record *itr,610				      struct evlist *evlist,611				      struct record_opts *opts)612{613	struct intel_pt_recording *ptr =614			container_of(itr, struct intel_pt_recording, itr);615	struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;616	bool have_timing_info, need_immediate = false;617	struct evsel *evsel, *intel_pt_evsel = NULL;618	const struct perf_cpu_map *cpus = evlist->core.user_requested_cpus;619	bool privileged = perf_event_paranoid_check(-1);620	u64 tsc_bit;621	int err;622 623	ptr->evlist = evlist;624	ptr->snapshot_mode = opts->auxtrace_snapshot_mode;625 626	evlist__for_each_entry(evlist, evsel) {627		if (evsel->core.attr.type == intel_pt_pmu->type) {628			if (intel_pt_evsel) {629				pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");630				return -EINVAL;631			}632			evsel->core.attr.freq = 0;633			evsel->core.attr.sample_period = 1;634			evsel->core.attr.exclude_guest = intel_pt_exclude_guest();635			evsel->no_aux_samples = true;636			evsel->needs_auxtrace_mmap = true;637			intel_pt_evsel = evsel;638			opts->full_auxtrace = true;639		}640	}641 642	if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {643		pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");644		return -EINVAL;645	}646 647	if (opts->auxtrace_snapshot_mode && opts->auxtrace_sample_mode) {648		pr_err("Snapshot mode (" INTEL_PT_PMU_NAME " PMU) and sample trace cannot be used together\n");649		return -EINVAL;650	}651 652	if (opts->use_clockid) {653		pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");654		return -EINVAL;655	}656 657	if (intel_pt_too_many_aux_output(evlist))658		return -EINVAL;659 660	if (!opts->full_auxtrace)661		return 0;662 663	if (opts->auxtrace_sample_mode)664		evsel__set_config_if_unset(intel_pt_pmu, intel_pt_evsel,665					   "psb_period", 0);666 667	err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);668	if (err)669		return err;670 671	/* Set default sizes for snapshot mode */672	if (opts->auxtrace_snapshot_mode) {673		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);674 675		if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {676			if (privileged) {677				opts->auxtrace_mmap_pages = MiB(4) / page_size;678			} else {679				opts->auxtrace_mmap_pages = KiB(128) / page_size;680				if (opts->mmap_pages == UINT_MAX)681					opts->mmap_pages = KiB(256) / page_size;682			}683		} else if (!opts->auxtrace_mmap_pages && !privileged &&684			   opts->mmap_pages == UINT_MAX) {685			opts->mmap_pages = KiB(256) / page_size;686		}687		if (!opts->auxtrace_snapshot_size)688			opts->auxtrace_snapshot_size =689				opts->auxtrace_mmap_pages * (size_t)page_size;690		if (!opts->auxtrace_mmap_pages) {691			size_t sz = opts->auxtrace_snapshot_size;692 693			sz = round_up(sz, page_size) / page_size;694			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);695		}696		if (opts->auxtrace_snapshot_size >697				opts->auxtrace_mmap_pages * (size_t)page_size) {698			pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",699			       opts->auxtrace_snapshot_size,700			       opts->auxtrace_mmap_pages * (size_t)page_size);701			return -EINVAL;702		}703		if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {704			pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");705			return -EINVAL;706		}707		pr_debug2("Intel PT snapshot size: %zu\n",708			  opts->auxtrace_snapshot_size);709		if (psb_period &&710		    opts->auxtrace_snapshot_size <= psb_period +711						  INTEL_PT_PSB_PERIOD_NEAR)712			ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",713				    opts->auxtrace_snapshot_size, psb_period);714	}715 716	/* Set default sizes for sample mode */717	if (opts->auxtrace_sample_mode) {718		size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);719		size_t min_sz = 0, max_sz = 0;720 721		intel_pt_min_max_sample_sz(evlist, &min_sz, &max_sz);722		if (!opts->auxtrace_mmap_pages && !privileged &&723		    opts->mmap_pages == UINT_MAX)724			opts->mmap_pages = KiB(256) / page_size;725		if (!opts->auxtrace_mmap_pages) {726			size_t sz = round_up(max_sz, page_size) / page_size;727 728			opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);729		}730		if (max_sz > opts->auxtrace_mmap_pages * (size_t)page_size) {731			pr_err("Sample size %zu must not be greater than AUX area tracing mmap size %zu\n",732			       max_sz,733			       opts->auxtrace_mmap_pages * (size_t)page_size);734			return -EINVAL;735		}736		pr_debug2("Intel PT min. sample size: %zu max. sample size: %zu\n",737			  min_sz, max_sz);738		if (psb_period &&739		    min_sz <= psb_period + INTEL_PT_PSB_PERIOD_NEAR)740			ui__warning("Intel PT sample size (%zu) may be too small for PSB period (%zu)\n",741				    min_sz, psb_period);742	}743 744	/* Set default sizes for full trace mode */745	if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {746		if (privileged) {747			opts->auxtrace_mmap_pages = MiB(4) / page_size;748		} else {749			opts->auxtrace_mmap_pages = KiB(128) / page_size;750			if (opts->mmap_pages == UINT_MAX)751				opts->mmap_pages = KiB(256) / page_size;752		}753	}754 755	/* Validate auxtrace_mmap_pages */756	if (opts->auxtrace_mmap_pages) {757		size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;758		size_t min_sz;759 760		if (opts->auxtrace_snapshot_mode || opts->auxtrace_sample_mode)761			min_sz = KiB(4);762		else763			min_sz = KiB(8);764 765		if (sz < min_sz || !is_power_of_2(sz)) {766			pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",767			       min_sz / 1024);768			return -EINVAL;769		}770	}771 772	if (!opts->auxtrace_snapshot_mode && !opts->auxtrace_sample_mode) {773		size_t aw = opts->auxtrace_mmap_pages * (size_t)page_size / 4;774		u32 aux_watermark = aw > UINT_MAX ? UINT_MAX : aw;775 776		intel_pt_evsel->core.attr.aux_watermark = aux_watermark;777	}778 779	intel_pt_parse_terms(intel_pt_pmu, "tsc", &tsc_bit);780 781	if (opts->full_auxtrace && (intel_pt_evsel->core.attr.config & tsc_bit))782		have_timing_info = true;783	else784		have_timing_info = false;785 786	/*787	 * Per-cpu recording needs sched_switch events to distinguish different788	 * threads.789	 */790	if (have_timing_info && !perf_cpu_map__is_any_cpu_or_is_empty(cpus) &&791	    !record_opts__no_switch_events(opts)) {792		if (perf_can_record_switch_events()) {793			bool cpu_wide = !target__none(&opts->target) &&794					!target__has_task(&opts->target);795 796			if (!cpu_wide && perf_can_record_cpu_wide()) {797				struct evsel *switch_evsel;798 799				switch_evsel = evlist__add_dummy_on_all_cpus(evlist);800				if (!switch_evsel)801					return -ENOMEM;802 803				switch_evsel->core.attr.context_switch = 1;804				switch_evsel->immediate = true;805 806				evsel__set_sample_bit(switch_evsel, TID);807				evsel__set_sample_bit(switch_evsel, TIME);808				evsel__set_sample_bit(switch_evsel, CPU);809				evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);810 811				opts->record_switch_events = false;812				ptr->have_sched_switch = 3;813			} else {814				opts->record_switch_events = true;815				need_immediate = true;816				if (cpu_wide)817					ptr->have_sched_switch = 3;818				else819					ptr->have_sched_switch = 2;820			}821		} else {822#ifdef HAVE_LIBTRACEEVENT823			err = intel_pt_track_switches(evlist);824			if (err == -EPERM)825				pr_debug2("Unable to select sched:sched_switch\n");826			else if (err)827				return err;828			else829				ptr->have_sched_switch = 1;830#endif831		}832	}833 834	if (have_timing_info && !intel_pt_evsel->core.attr.exclude_kernel &&835	    perf_can_record_text_poke_events() && perf_can_record_cpu_wide())836		opts->text_poke = true;837 838	if (intel_pt_evsel) {839		/*840		 * To obtain the auxtrace buffer file descriptor, the auxtrace841		 * event must come first.842		 */843		evlist__to_front(evlist, intel_pt_evsel);844		/*845		 * In the case of per-cpu mmaps, we need the CPU on the846		 * AUX event.847		 */848		if (!perf_cpu_map__is_any_cpu_or_is_empty(cpus))849			evsel__set_sample_bit(intel_pt_evsel, CPU);850	}851 852	/* Add dummy event to keep tracking */853	if (opts->full_auxtrace) {854		bool need_system_wide_tracking;855		struct evsel *tracking_evsel;856 857		/*858		 * User space tasks can migrate between CPUs, so when tracing859		 * selected CPUs, sideband for all CPUs is still needed.860		 */861		need_system_wide_tracking = opts->target.cpu_list &&862					    !intel_pt_evsel->core.attr.exclude_user;863 864		tracking_evsel = evlist__add_aux_dummy(evlist, need_system_wide_tracking);865		if (!tracking_evsel)866			return -ENOMEM;867 868		evlist__set_tracking_event(evlist, tracking_evsel);869 870		if (need_immediate)871			tracking_evsel->immediate = true;872 873		/* In per-cpu case, always need the time of mmap events etc */874		if (!perf_cpu_map__is_any_cpu_or_is_empty(cpus)) {875			evsel__set_sample_bit(tracking_evsel, TIME);876			/* And the CPU for switch events */877			evsel__set_sample_bit(tracking_evsel, CPU);878		}879		evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);880	}881 882	/*883	 * Warn the user when we do not have enough information to decode i.e.884	 * per-cpu with no sched_switch (except workload-only).885	 */886	if (!ptr->have_sched_switch && !perf_cpu_map__is_any_cpu_or_is_empty(cpus) &&887	    !target__none(&opts->target) &&888	    !intel_pt_evsel->core.attr.exclude_user)889		ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");890 891	return 0;892}893 894static int intel_pt_snapshot_start(struct auxtrace_record *itr)895{896	struct intel_pt_recording *ptr =897			container_of(itr, struct intel_pt_recording, itr);898	struct evsel *evsel;899 900	evlist__for_each_entry(ptr->evlist, evsel) {901		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)902			return evsel__disable(evsel);903	}904	return -EINVAL;905}906 907static int intel_pt_snapshot_finish(struct auxtrace_record *itr)908{909	struct intel_pt_recording *ptr =910			container_of(itr, struct intel_pt_recording, itr);911	struct evsel *evsel;912 913	evlist__for_each_entry(ptr->evlist, evsel) {914		if (evsel->core.attr.type == ptr->intel_pt_pmu->type)915			return evsel__enable(evsel);916	}917	return -EINVAL;918}919 920static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)921{922	const size_t sz = sizeof(struct intel_pt_snapshot_ref);923	int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;924	struct intel_pt_snapshot_ref *refs;925 926	if (!new_cnt)927		new_cnt = 16;928 929	while (new_cnt <= idx)930		new_cnt *= 2;931 932	refs = calloc(new_cnt, sz);933	if (!refs)934		return -ENOMEM;935 936	memcpy(refs, ptr->snapshot_refs, cnt * sz);937 938	ptr->snapshot_refs = refs;939	ptr->snapshot_ref_cnt = new_cnt;940 941	return 0;942}943 944static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)945{946	int i;947 948	for (i = 0; i < ptr->snapshot_ref_cnt; i++)949		zfree(&ptr->snapshot_refs[i].ref_buf);950	zfree(&ptr->snapshot_refs);951}952 953static void intel_pt_recording_free(struct auxtrace_record *itr)954{955	struct intel_pt_recording *ptr =956			container_of(itr, struct intel_pt_recording, itr);957 958	intel_pt_free_snapshot_refs(ptr);959	free(ptr);960}961 962static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,963				       size_t snapshot_buf_size)964{965	size_t ref_buf_size = ptr->snapshot_ref_buf_size;966	void *ref_buf;967 968	ref_buf = zalloc(ref_buf_size);969	if (!ref_buf)970		return -ENOMEM;971 972	ptr->snapshot_refs[idx].ref_buf = ref_buf;973	ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;974 975	return 0;976}977 978static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,979					     size_t snapshot_buf_size)980{981	const size_t max_size = 256 * 1024;982	size_t buf_size = 0, psb_period;983 984	if (ptr->snapshot_size <= 64 * 1024)985		return 0;986 987	psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);988	if (psb_period)989		buf_size = psb_period * 2;990 991	if (!buf_size || buf_size > max_size)992		buf_size = max_size;993 994	if (buf_size >= snapshot_buf_size)995		return 0;996 997	if (buf_size >= ptr->snapshot_size / 2)998		return 0;999 1000	return buf_size;1001}1002 1003static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,1004				  size_t snapshot_buf_size)1005{1006	if (ptr->snapshot_init_done)1007		return 0;1008 1009	ptr->snapshot_init_done = true;1010 1011	ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,1012							snapshot_buf_size);1013 1014	return 0;1015}1016 1017/**1018 * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.1019 * @buf1: first buffer1020 * @compare_size: number of bytes to compare1021 * @buf2: second buffer (a circular buffer)1022 * @offs2: offset in second buffer1023 * @buf2_size: size of second buffer1024 *1025 * The comparison allows for the possibility that the bytes to compare in the1026 * circular buffer are not contiguous.  It is assumed that @compare_size <=1027 * @buf2_size.  This function returns %false if the bytes are identical, %true1028 * otherwise.1029 */1030static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,1031				     void *buf2, size_t offs2, size_t buf2_size)1032{1033	size_t end2 = offs2 + compare_size, part_size;1034 1035	if (end2 <= buf2_size)1036		return memcmp(buf1, buf2 + offs2, compare_size);1037 1038	part_size = end2 - buf2_size;1039	if (memcmp(buf1, buf2 + offs2, part_size))1040		return true;1041 1042	compare_size -= part_size;1043 1044	return memcmp(buf1 + part_size, buf2, compare_size);1045}1046 1047static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,1048				 size_t ref_size, size_t buf_size,1049				 void *data, size_t head)1050{1051	size_t ref_end = ref_offset + ref_size;1052 1053	if (ref_end > buf_size) {1054		if (head > ref_offset || head < ref_end - buf_size)1055			return true;1056	} else if (head > ref_offset && head < ref_end) {1057		return true;1058	}1059 1060	return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,1061					buf_size);1062}1063 1064static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,1065			      void *data, size_t head)1066{1067	if (head >= ref_size) {1068		memcpy(ref_buf, data + head - ref_size, ref_size);1069	} else {1070		memcpy(ref_buf, data, head);1071		ref_size -= head;1072		memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);1073	}1074}1075 1076static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,1077			     struct auxtrace_mmap *mm, unsigned char *data,1078			     u64 head)1079{1080	struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];1081	bool wrapped;1082 1083	wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,1084				       ptr->snapshot_ref_buf_size, mm->len,1085				       data, head);1086 1087	intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,1088			  data, head);1089 1090	return wrapped;1091}1092 1093static bool intel_pt_first_wrap(u64 *data, size_t buf_size)1094{1095	int i, a, b;1096 1097	b = buf_size >> 3;1098	a = b - 512;1099	if (a < 0)1100		a = 0;1101 1102	for (i = a; i < b; i++) {1103		if (data[i])1104			return true;1105	}1106 1107	return false;1108}1109 1110static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,1111				  struct auxtrace_mmap *mm, unsigned char *data,1112				  u64 *head, u64 *old)1113{1114	struct intel_pt_recording *ptr =1115			container_of(itr, struct intel_pt_recording, itr);1116	bool wrapped;1117	int err;1118 1119	pr_debug3("%s: mmap index %d old head %zu new head %zu\n",1120		  __func__, idx, (size_t)*old, (size_t)*head);1121 1122	err = intel_pt_snapshot_init(ptr, mm->len);1123	if (err)1124		goto out_err;1125 1126	if (idx >= ptr->snapshot_ref_cnt) {1127		err = intel_pt_alloc_snapshot_refs(ptr, idx);1128		if (err)1129			goto out_err;1130	}1131 1132	if (ptr->snapshot_ref_buf_size) {1133		if (!ptr->snapshot_refs[idx].ref_buf) {1134			err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);1135			if (err)1136				goto out_err;1137		}1138		wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);1139	} else {1140		wrapped = ptr->snapshot_refs[idx].wrapped;1141		if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {1142			ptr->snapshot_refs[idx].wrapped = true;1143			wrapped = true;1144		}1145	}1146 1147	/*1148	 * In full trace mode 'head' continually increases.  However in snapshot1149	 * mode 'head' is an offset within the buffer.  Here 'old' and 'head'1150	 * are adjusted to match the full trace case which expects that 'old' is1151	 * always less than 'head'.1152	 */1153	if (wrapped) {1154		*old = *head;1155		*head += mm->len;1156	} else {1157		if (mm->mask)1158			*old &= mm->mask;1159		else1160			*old %= mm->len;1161		if (*old > *head)1162			*head += mm->len;1163	}1164 1165	pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",1166		  __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);1167 1168	return 0;1169 1170out_err:1171	pr_err("%s: failed, error %d\n", __func__, err);1172	return err;1173}1174 1175static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)1176{1177	return rdtsc();1178}1179 1180struct auxtrace_record *intel_pt_recording_init(int *err)1181{1182	struct perf_pmu *intel_pt_pmu = perf_pmus__find(INTEL_PT_PMU_NAME);1183	struct intel_pt_recording *ptr;1184 1185	if (!intel_pt_pmu)1186		return NULL;1187 1188	if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {1189		*err = -errno;1190		return NULL;1191	}1192 1193	ptr = zalloc(sizeof(struct intel_pt_recording));1194	if (!ptr) {1195		*err = -ENOMEM;1196		return NULL;1197	}1198 1199	ptr->intel_pt_pmu = intel_pt_pmu;1200	ptr->itr.recording_options = intel_pt_recording_options;1201	ptr->itr.info_priv_size = intel_pt_info_priv_size;1202	ptr->itr.info_fill = intel_pt_info_fill;1203	ptr->itr.free = intel_pt_recording_free;1204	ptr->itr.snapshot_start = intel_pt_snapshot_start;1205	ptr->itr.snapshot_finish = intel_pt_snapshot_finish;1206	ptr->itr.find_snapshot = intel_pt_find_snapshot;1207	ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;1208	ptr->itr.reference = intel_pt_reference;1209	ptr->itr.read_finish = auxtrace_record__read_finish;1210	/*1211	 * Decoding starts at a PSB packet. Minimum PSB period is 2K so 4K1212	 * should give at least 1 PSB per sample.1213	 */1214	ptr->itr.default_aux_sample_size = 4096;1215	return &ptr->itr;1216}1217