tools/perf/arch/arm64/util/arm-spe.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Arm Statistical Profiling Extensions (SPE) support
   4  * Copyright (c) 2017-2018, Arm Ltd.
   5  */
   6
   7 #include <linux/kernel.h>
   8 #include <linux/types.h>
   9 #include <linux/bitops.h>
  10 #include <linux/log2.h>
  11 #include <linux/zalloc.h>
  12 #include <time.h>
  13
  14 #include "../../../util/cpumap.h"
  15 #include "../../../util/event.h"
  16 #include "../../../util/evsel.h"
  17 #include "../../../util/evsel_config.h"
  18 #include "../../../util/evlist.h"
  19 #include "../../../util/session.h"
  20 #include <internal/lib.h> // page_size
  21 #include "../../../util/pmu.h"
  22 #include "../../../util/debug.h"
  23 #include "../../../util/auxtrace.h"
  24 #include "../../../util/record.h"
  25 #include "../../../util/arm-spe.h"
  26
  27 #define KiB(x) ((x) * 1024)
  28 #define MiB(x) ((x) * 1024 * 1024)
  29
  30 struct arm_spe_recording {
  31         struct auxtrace_record          itr;
  32         struct perf_pmu                 *arm_spe_pmu;
  33         struct evlist           *evlist;
  34 };
  35
  36 static void arm_spe_set_timestamp(struct auxtrace_record *itr,
  37                                   struct evsel *evsel)
  38 {
  39         struct arm_spe_recording *ptr;
  40         struct perf_pmu *arm_spe_pmu;
  41         struct evsel_config_term *term = evsel__get_config_term(evsel, CFG_CHG);
  42         u64 user_bits = 0, bit;
  43
  44         ptr = container_of(itr, struct arm_spe_recording, itr);
  45         arm_spe_pmu = ptr->arm_spe_pmu;
  46
  47         if (term)
  48                 user_bits = term->val.cfg_chg;
  49
  50         bit = perf_pmu__format_bits(&arm_spe_pmu->format, "ts_enable");
  51
  52         /* Skip if user has set it */
  53         if (bit & user_bits)
  54                 return;
  55
  56         evsel->core.attr.config |= bit;
  57 }
  58
  59 static size_t
  60 arm_spe_info_priv_size(struct auxtrace_record *itr __maybe_unused,
  61                        struct evlist *evlist __maybe_unused)
  62 {
  63         return ARM_SPE_AUXTRACE_PRIV_SIZE;
  64 }
  65
  66 static int arm_spe_info_fill(struct auxtrace_record *itr,
  67                              struct perf_session *session,
  68                              struct perf_record_auxtrace_info *auxtrace_info,
  69                              size_t priv_size)
  70 {
  71         struct arm_spe_recording *sper =
  72                         container_of(itr, struct arm_spe_recording, itr);
  73         struct perf_pmu *arm_spe_pmu = sper->arm_spe_pmu;
  74
  75         if (priv_size != ARM_SPE_AUXTRACE_PRIV_SIZE)
  76                 return -EINVAL;
  77
  78         if (!session->evlist->core.nr_mmaps)
  79                 return -EINVAL;
  80
  81         auxtrace_info->type = PERF_AUXTRACE_ARM_SPE;
  82         auxtrace_info->priv[ARM_SPE_PMU_TYPE] = arm_spe_pmu->type;
  83
  84         return 0;
  85 }
  86
  87 static int arm_spe_recording_options(struct auxtrace_record *itr,
  88                                      struct evlist *evlist,
  89                                      struct record_opts *opts)
  90 {
  91         struct arm_spe_recording *sper =
  92                         container_of(itr, struct arm_spe_recording, itr);
  93         struct perf_pmu *arm_spe_pmu = sper->arm_spe_pmu;
  94         struct evsel *evsel, *arm_spe_evsel = NULL;
  95         struct perf_cpu_map *cpus = evlist->core.cpus;
  96         bool privileged = perf_event_paranoid_check(-1);
  97         struct evsel *tracking_evsel;
  98         int err;
  99
 100         sper->evlist = evlist;
 101
 102         evlist__for_each_entry(evlist, evsel) {
 103                 if (evsel->core.attr.type == arm_spe_pmu->type) {
 104                         if (arm_spe_evsel) {
 105                                 pr_err("There may be only one " ARM_SPE_PMU_NAME "x event\n");
 106                                 return -EINVAL;
 107                         }
 108                         evsel->core.attr.freq = 0;
 109                         evsel->core.attr.sample_period = 1;
 110                         arm_spe_evsel = evsel;
 111                         opts->full_auxtrace = true;
 112                 }
 113         }
 114
 115         if (!opts->full_auxtrace)
 116                 return 0;
 117
 118         /* We are in full trace mode but '-m,xyz' wasn't specified */
 119         if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
 120                 if (privileged) {
 121                         opts->auxtrace_mmap_pages = MiB(4) / page_size;
 122                 } else {
 123                         opts->auxtrace_mmap_pages = KiB(128) / page_size;
 124                         if (opts->mmap_pages == UINT_MAX)
 125                                 opts->mmap_pages = KiB(256) / page_size;
 126                 }
 127         }
 128
 129         /* Validate auxtrace_mmap_pages */
 130         if (opts->auxtrace_mmap_pages) {
 131                 size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
 132                 size_t min_sz = KiB(8);
 133
 134                 if (sz < min_sz || !is_power_of_2(sz)) {
 135                         pr_err("Invalid mmap size for ARM SPE: must be at least %zuKiB and a power of 2\n",
 136                                min_sz / 1024);
 137                         return -EINVAL;
 138                 }
 139         }
 140
 141
 142         /*
 143          * To obtain the auxtrace buffer file descriptor, the auxtrace event
 144          * must come first.
 145          */
 146         evlist__to_front(evlist, arm_spe_evsel);
 147
 148         /*
 149          * In the case of per-cpu mmaps, sample CPU for AUX event;
 150          * also enable the timestamp tracing for samples correlation.
 151          */
 152         if (!perf_cpu_map__empty(cpus)) {
 153                 evsel__set_sample_bit(arm_spe_evsel, CPU);
 154                 arm_spe_set_timestamp(itr, arm_spe_evsel);
 155         }
 156
 157         /* Add dummy event to keep tracking */
 158         err = parse_events(evlist, "dummy:u", NULL);
 159         if (err)
 160                 return err;
 161
 162         tracking_evsel = evlist__last(evlist);
 163         evlist__set_tracking_event(evlist, tracking_evsel);
 164
 165         tracking_evsel->core.attr.freq = 0;
 166         tracking_evsel->core.attr.sample_period = 1;
 167
 168         /* In per-cpu case, always need the time of mmap events etc */
 169         if (!perf_cpu_map__empty(cpus))
 170                 evsel__set_sample_bit(tracking_evsel, TIME);
 171
 172         return 0;
 173 }
 174
 175 static u64 arm_spe_reference(struct auxtrace_record *itr __maybe_unused)
 176 {
 177         struct timespec ts;
 178
 179         clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
 180
 181         return ts.tv_sec ^ ts.tv_nsec;
 182 }
 183
 184 static void arm_spe_recording_free(struct auxtrace_record *itr)
 185 {
 186         struct arm_spe_recording *sper =
 187                         container_of(itr, struct arm_spe_recording, itr);
 188
 189         free(sper);
 190 }
 191
 192 struct auxtrace_record *arm_spe_recording_init(int *err,
 193                                                struct perf_pmu *arm_spe_pmu)
 194 {
 195         struct arm_spe_recording *sper;
 196
 197         if (!arm_spe_pmu) {
 198                 *err = -ENODEV;
 199                 return NULL;
 200         }
 201
 202         sper = zalloc(sizeof(struct arm_spe_recording));
 203         if (!sper) {
 204                 *err = -ENOMEM;
 205                 return NULL;
 206         }
 207
 208         sper->arm_spe_pmu = arm_spe_pmu;
 209         sper->itr.pmu = arm_spe_pmu;
 210         sper->itr.recording_options = arm_spe_recording_options;
 211         sper->itr.info_priv_size = arm_spe_info_priv_size;
 212         sper->itr.info_fill = arm_spe_info_fill;
 213         sper->itr.free = arm_spe_recording_free;
 214         sper->itr.reference = arm_spe_reference;
 215         sper->itr.read_finish = auxtrace_record__read_finish;
 216         sper->itr.alignment = 0;
 217
 218         *err = 0;
 219         return &sper->itr;
 220 }
 221
 222 struct perf_event_attr
 223 *arm_spe_pmu_default_config(struct perf_pmu *arm_spe_pmu)
 224 {
 225         struct perf_event_attr *attr;
 226
 227         attr = zalloc(sizeof(struct perf_event_attr));
 228         if (!attr) {
 229                 pr_err("arm_spe default config cannot allocate a perf_event_attr\n");
 230                 return NULL;
 231         }
 232
 233         /*
 234          * If kernel driver doesn't advertise a minimum,
 235          * use max allowable by PMSIDR_EL1.INTERVAL
 236          */
 237         if (perf_pmu__scan_file(arm_spe_pmu, "caps/min_interval", "%llu",
 238                                   &attr->sample_period) != 1) {
 239                 pr_debug("arm_spe driver doesn't advertise a min. interval. Using 4096\n");
 240                 attr->sample_period = 4096;
 241         }
 242
 243         arm_spe_pmu->selectable = true;
 244         arm_spe_pmu->is_uncore = false;
 245
 246         return attr;
 247 }