drivers/hwtracing/coresight/coresight-etm-perf.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright(C) 2015 Linaro Limited. All rights reserved.
   4  * Author: Mathieu Poirier <mathieu.poirier@linaro.org>
   5  */
   6
   7 #include <linux/coresight.h>
   8 #include <linux/coresight-pmu.h>
   9 #include <linux/cpumask.h>
  10 #include <linux/device.h>
  11 #include <linux/list.h>
  12 #include <linux/mm.h>
  13 #include <linux/init.h>
  14 #include <linux/perf_event.h>
  15 #include <linux/percpu-defs.h>
  16 #include <linux/slab.h>
  17 #include <linux/stringhash.h>
  18 #include <linux/types.h>
  19 #include <linux/workqueue.h>
  20
  21 #include "coresight-etm-perf.h"
  22 #include "coresight-priv.h"
  23
  24 static struct pmu etm_pmu;
  25 static bool etm_perf_up;
  26
  27 static DEFINE_PER_CPU(struct perf_output_handle, ctx_handle);
  28 static DEFINE_PER_CPU(struct coresight_device *, csdev_src);
  29
  30 /*
  31  * The PMU formats were orignally for ETMv3.5/PTM's ETMCR 'config';
  32  * now take them as general formats and apply on all ETMs.
  33  */
  34 PMU_FORMAT_ATTR(cycacc,         "config:" __stringify(ETM_OPT_CYCACC));
  35 PMU_FORMAT_ATTR(contextid,      "config:" __stringify(ETM_OPT_CTXTID));
  36 PMU_FORMAT_ATTR(timestamp,      "config:" __stringify(ETM_OPT_TS));
  37 PMU_FORMAT_ATTR(retstack,       "config:" __stringify(ETM_OPT_RETSTK));
  38 /* Sink ID - same for all ETMs */
  39 PMU_FORMAT_ATTR(sinkid,         "config2:0-31");
  40
  41 static struct attribute *etm_config_formats_attr[] = {
  42         &format_attr_cycacc.attr,
  43         &format_attr_contextid.attr,
  44         &format_attr_timestamp.attr,
  45         &format_attr_retstack.attr,
  46         &format_attr_sinkid.attr,
  47         NULL,
  48 };
  49
  50 static const struct attribute_group etm_pmu_format_group = {
  51         .name   = "format",
  52         .attrs  = etm_config_formats_attr,
  53 };
  54
  55 static struct attribute *etm_config_sinks_attr[] = {
  56         NULL,
  57 };
  58
  59 static const struct attribute_group etm_pmu_sinks_group = {
  60         .name   = "sinks",
  61         .attrs  = etm_config_sinks_attr,
  62 };
  63
  64 static const struct attribute_group *etm_pmu_attr_groups[] = {
  65         &etm_pmu_format_group,
  66         &etm_pmu_sinks_group,
  67         NULL,
  68 };
  69
  70 static inline struct list_head **
  71 etm_event_cpu_path_ptr(struct etm_event_data *data, int cpu)
  72 {
  73         return per_cpu_ptr(data->path, cpu);
  74 }
  75
  76 static inline struct list_head *
  77 etm_event_cpu_path(struct etm_event_data *data, int cpu)
  78 {
  79         return *etm_event_cpu_path_ptr(data, cpu);
  80 }
  81
  82 static void etm_event_read(struct perf_event *event) {}
  83
  84 static int etm_addr_filters_alloc(struct perf_event *event)
  85 {
  86         struct etm_filters *filters;
  87         int node = event->cpu == -1 ? -1 : cpu_to_node(event->cpu);
  88
  89         filters = kzalloc_node(sizeof(struct etm_filters), GFP_KERNEL, node);
  90         if (!filters)
  91                 return -ENOMEM;
  92
  93         if (event->parent)
  94                 memcpy(filters, event->parent->hw.addr_filters,
  95                        sizeof(*filters));
  96
  97         event->hw.addr_filters = filters;
  98
  99         return 0;
 100 }
 101
 102 static void etm_event_destroy(struct perf_event *event)
 103 {
 104         kfree(event->hw.addr_filters);
 105         event->hw.addr_filters = NULL;
 106 }
 107
 108 static int etm_event_init(struct perf_event *event)
 109 {
 110         int ret = 0;
 111
 112         if (event->attr.type != etm_pmu.type) {
 113                 ret = -ENOENT;
 114                 goto out;
 115         }
 116
 117         ret = etm_addr_filters_alloc(event);
 118         if (ret)
 119                 goto out;
 120
 121         event->destroy = etm_event_destroy;
 122 out:
 123         return ret;
 124 }
 125
 126 static void free_sink_buffer(struct etm_event_data *event_data)
 127 {
 128         int cpu;
 129         cpumask_t *mask = &event_data->mask;
 130         struct coresight_device *sink;
 131
 132         if (!event_data->snk_config)
 133                 return;
 134
 135         if (WARN_ON(cpumask_empty(mask)))
 136                 return;
 137
 138         cpu = cpumask_first(mask);
 139         sink = coresight_get_sink(etm_event_cpu_path(event_data, cpu));
 140         sink_ops(sink)->free_buffer(event_data->snk_config);
 141 }
 142
 143 static void free_event_data(struct work_struct *work)
 144 {
 145         int cpu;
 146         cpumask_t *mask;
 147         struct etm_event_data *event_data;
 148
 149         event_data = container_of(work, struct etm_event_data, work);
 150         mask = &event_data->mask;
 151
 152         /* Free the sink buffers, if there are any */
 153         free_sink_buffer(event_data);
 154
 155         for_each_cpu(cpu, mask) {
 156                 struct list_head **ppath;
 157
 158                 ppath = etm_event_cpu_path_ptr(event_data, cpu);
 159                 if (!(IS_ERR_OR_NULL(*ppath)))
 160                         coresight_release_path(*ppath);
 161                 *ppath = NULL;
 162         }
 163
 164         free_percpu(event_data->path);
 165         kfree(event_data);
 166 }
 167
 168 static void *alloc_event_data(int cpu)
 169 {
 170         cpumask_t *mask;
 171         struct etm_event_data *event_data;
 172
 173         /* First get memory for the session's data */
 174         event_data = kzalloc(sizeof(struct etm_event_data), GFP_KERNEL);
 175         if (!event_data)
 176                 return NULL;
 177
 178
 179         mask = &event_data->mask;
 180         if (cpu != -1)
 181                 cpumask_set_cpu(cpu, mask);
 182         else
 183                 cpumask_copy(mask, cpu_present_mask);
 184
 185         /*
 186          * Each CPU has a single path between source and destination.  As such
 187          * allocate an array using CPU numbers as indexes.  That way a path
 188          * for any CPU can easily be accessed at any given time.  We proceed
 189          * the same way for sessions involving a single CPU.  The cost of
 190          * unused memory when dealing with single CPU trace scenarios is small
 191          * compared to the cost of searching through an optimized array.
 192          */
 193         event_data->path = alloc_percpu(struct list_head *);
 194
 195         if (!event_data->path) {
 196                 kfree(event_data);
 197                 return NULL;
 198         }
 199
 200         return event_data;
 201 }
 202
 203 static void etm_free_aux(void *data)
 204 {
 205         struct etm_event_data *event_data = data;
 206
 207         schedule_work(&event_data->work);
 208 }
 209
 210 static void *etm_setup_aux(struct perf_event *event, void **pages,
 211                            int nr_pages, bool overwrite)
 212 {
 213         u32 id;
 214         int cpu = event->cpu;
 215         cpumask_t *mask;
 216         struct coresight_device *sink = NULL;
 217         struct etm_event_data *event_data = NULL;
 218
 219         event_data = alloc_event_data(cpu);
 220         if (!event_data)
 221                 return NULL;
 222         INIT_WORK(&event_data->work, free_event_data);
 223
 224         /* First get the selected sink from user space. */
 225         if (event->attr.config2) {
 226                 id = (u32)event->attr.config2;
 227                 sink = coresight_get_sink_by_id(id);
 228         }
 229
 230         mask = &event_data->mask;
 231
 232         /*
 233          * Setup the path for each CPU in a trace session. We try to build
 234          * trace path for each CPU in the mask. If we don't find an ETM
 235          * for the CPU or fail to build a path, we clear the CPU from the
 236          * mask and continue with the rest. If ever we try to trace on those
 237          * CPUs, we can handle it and fail the session.
 238          */
 239         for_each_cpu(cpu, mask) {
 240                 struct list_head *path;
 241                 struct coresight_device *csdev;
 242
 243                 csdev = per_cpu(csdev_src, cpu);
 244                 /*
 245                  * If there is no ETM associated with this CPU clear it from
 246                  * the mask and continue with the rest. If ever we try to trace
 247                  * on this CPU, we handle it accordingly.
 248                  */
 249                 if (!csdev) {
 250                         cpumask_clear_cpu(cpu, mask);
 251                         continue;
 252                 }
 253
 254                 /*
 255                  * No sink provided - look for a default sink for one of the
 256                  * devices. At present we only support topology where all CPUs
 257                  * use the same sink [N:1], so only need to find one sink. The
 258                  * coresight_build_path later will remove any CPU that does not
 259                  * attach to the sink, or if we have not found a sink.
 260                  */
 261                 if (!sink)
 262                         sink = coresight_find_default_sink(csdev);
 263
 264                 /*
 265                  * Building a path doesn't enable it, it simply builds a
 266                  * list of devices from source to sink that can be
 267                  * referenced later when the path is actually needed.
 268                  */
 269                 path = coresight_build_path(csdev, sink);
 270                 if (IS_ERR(path)) {
 271                         cpumask_clear_cpu(cpu, mask);
 272                         continue;
 273                 }
 274
 275                 *etm_event_cpu_path_ptr(event_data, cpu) = path;
 276         }
 277
 278         /* no sink found for any CPU - cannot trace */
 279         if (!sink)
 280                 goto err;
 281
 282         /* If we don't have any CPUs ready for tracing, abort */
 283         cpu = cpumask_first(mask);
 284         if (cpu >= nr_cpu_ids)
 285                 goto err;
 286
 287         if (!sink_ops(sink)->alloc_buffer || !sink_ops(sink)->free_buffer)
 288                 goto err;
 289
 290         /* Allocate the sink buffer for this session */
 291         event_data->snk_config =
 292                         sink_ops(sink)->alloc_buffer(sink, event, pages,
 293                                                      nr_pages, overwrite);
 294         if (!event_data->snk_config)
 295                 goto err;
 296
 297 out:
 298         return event_data;
 299
 300 err:
 301         etm_free_aux(event_data);
 302         event_data = NULL;
 303         goto out;
 304 }
 305
 306 static void etm_event_start(struct perf_event *event, int flags)
 307 {
 308         int cpu = smp_processor_id();
 309         struct etm_event_data *event_data;
 310         struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle);
 311         struct coresight_device *sink, *csdev = per_cpu(csdev_src, cpu);
 312         struct list_head *path;
 313
 314         if (!csdev)
 315                 goto fail;
 316
 317         /*
 318          * Deal with the ring buffer API and get a handle on the
 319          * session's information.
 320          */
 321         event_data = perf_aux_output_begin(handle, event);
 322         if (!event_data)
 323                 goto fail;
 324
 325         /*
 326          * Check if this ETM is allowed to trace, as decided
 327          * at etm_setup_aux(). This could be due to an unreachable
 328          * sink from this ETM. We can't do much in this case if
 329          * the sink was specified or hinted to the driver. For
 330          * now, simply don't record anything on this ETM.
 331          */
 332         if (!cpumask_test_cpu(cpu, &event_data->mask))
 333                 goto fail_end_stop;
 334
 335         path = etm_event_cpu_path(event_data, cpu);
 336         /* We need a sink, no need to continue without one */
 337         sink = coresight_get_sink(path);
 338         if (WARN_ON_ONCE(!sink))
 339                 goto fail_end_stop;
 340
 341         /* Nothing will happen without a path */
 342         if (coresight_enable_path(path, CS_MODE_PERF, handle))
 343                 goto fail_end_stop;
 344
 345         /* Tell the perf core the event is alive */
 346         event->hw.state = 0;
 347
 348         /* Finally enable the tracer */
 349         if (source_ops(csdev)->enable(csdev, event, CS_MODE_PERF))
 350                 goto fail_disable_path;
 351
 352 out:
 353         return;
 354
 355 fail_disable_path:
 356         coresight_disable_path(path);
 357 fail_end_stop:
 358         perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
 359         perf_aux_output_end(handle, 0);
 360 fail:
 361         event->hw.state = PERF_HES_STOPPED;
 362         goto out;
 363 }
 364
 365 static void etm_event_stop(struct perf_event *event, int mode)
 366 {
 367         int cpu = smp_processor_id();
 368         unsigned long size;
 369         struct coresight_device *sink, *csdev = per_cpu(csdev_src, cpu);
 370         struct perf_output_handle *handle = this_cpu_ptr(&ctx_handle);
 371         struct etm_event_data *event_data = perf_get_aux(handle);
 372         struct list_head *path;
 373
 374         if (event->hw.state == PERF_HES_STOPPED)
 375                 return;
 376
 377         if (!csdev)
 378                 return;
 379
 380         path = etm_event_cpu_path(event_data, cpu);
 381         if (!path)
 382                 return;
 383
 384         sink = coresight_get_sink(path);
 385         if (!sink)
 386                 return;
 387
 388         /* stop tracer */
 389         source_ops(csdev)->disable(csdev, event);
 390
 391         /* tell the core */
 392         event->hw.state = PERF_HES_STOPPED;
 393
 394         if (mode & PERF_EF_UPDATE) {
 395                 if (WARN_ON_ONCE(handle->event != event))
 396                         return;
 397
 398                 /* update trace information */
 399                 if (!sink_ops(sink)->update_buffer)
 400                         return;
 401
 402                 size = sink_ops(sink)->update_buffer(sink, handle,
 403                                               event_data->snk_config);
 404                 perf_aux_output_end(handle, size);
 405         }
 406
 407         /* Disabling the path make its elements available to other sessions */
 408         coresight_disable_path(path);
 409 }
 410
 411 static int etm_event_add(struct perf_event *event, int mode)
 412 {
 413         int ret = 0;
 414         struct hw_perf_event *hwc = &event->hw;
 415
 416         if (mode & PERF_EF_START) {
 417                 etm_event_start(event, 0);
 418                 if (hwc->state & PERF_HES_STOPPED)
 419                         ret = -EINVAL;
 420         } else {
 421                 hwc->state = PERF_HES_STOPPED;
 422         }
 423
 424         return ret;
 425 }
 426
 427 static void etm_event_del(struct perf_event *event, int mode)
 428 {
 429         etm_event_stop(event, PERF_EF_UPDATE);
 430 }
 431
 432 static int etm_addr_filters_validate(struct list_head *filters)
 433 {
 434         bool range = false, address = false;
 435         int index = 0;
 436         struct perf_addr_filter *filter;
 437
 438         list_for_each_entry(filter, filters, entry) {
 439                 /*
 440                  * No need to go further if there's no more
 441                  * room for filters.
 442                  */
 443                 if (++index > ETM_ADDR_CMP_MAX)
 444                         return -EOPNOTSUPP;
 445
 446                 /* filter::size==0 means single address trigger */
 447                 if (filter->size) {
 448                         /*
 449                          * The existing code relies on START/STOP filters
 450                          * being address filters.
 451                          */
 452                         if (filter->action == PERF_ADDR_FILTER_ACTION_START ||
 453                             filter->action == PERF_ADDR_FILTER_ACTION_STOP)
 454                                 return -EOPNOTSUPP;
 455
 456                         range = true;
 457                 } else
 458                         address = true;
 459
 460                 /*
 461                  * At this time we don't allow range and start/stop filtering
 462                  * to cohabitate, they have to be mutually exclusive.
 463                  */
 464                 if (range && address)
 465                         return -EOPNOTSUPP;
 466         }
 467
 468         return 0;
 469 }
 470
 471 static void etm_addr_filters_sync(struct perf_event *event)
 472 {
 473         struct perf_addr_filters_head *head = perf_event_addr_filters(event);
 474         unsigned long start, stop;
 475         struct perf_addr_filter_range *fr = event->addr_filter_ranges;
 476         struct etm_filters *filters = event->hw.addr_filters;
 477         struct etm_filter *etm_filter;
 478         struct perf_addr_filter *filter;
 479         int i = 0;
 480
 481         list_for_each_entry(filter, &head->list, entry) {
 482                 start = fr[i].start;
 483                 stop = start + fr[i].size;
 484                 etm_filter = &filters->etm_filter[i];
 485
 486                 switch (filter->action) {
 487                 case PERF_ADDR_FILTER_ACTION_FILTER:
 488                         etm_filter->start_addr = start;
 489                         etm_filter->stop_addr = stop;
 490                         etm_filter->type = ETM_ADDR_TYPE_RANGE;
 491                         break;
 492                 case PERF_ADDR_FILTER_ACTION_START:
 493                         etm_filter->start_addr = start;
 494                         etm_filter->type = ETM_ADDR_TYPE_START;
 495                         break;
 496                 case PERF_ADDR_FILTER_ACTION_STOP:
 497                         etm_filter->stop_addr = stop;
 498                         etm_filter->type = ETM_ADDR_TYPE_STOP;
 499                         break;
 500                 }
 501                 i++;
 502         }
 503
 504         filters->nr_filters = i;
 505 }
 506
 507 int etm_perf_symlink(struct coresight_device *csdev, bool link)
 508 {
 509         char entry[sizeof("cpu9999999")];
 510         int ret = 0, cpu = source_ops(csdev)->cpu_id(csdev);
 511         struct device *pmu_dev = etm_pmu.dev;
 512         struct device *cs_dev = &csdev->dev;
 513
 514         sprintf(entry, "cpu%d", cpu);
 515
 516         if (!etm_perf_up)
 517                 return -EPROBE_DEFER;
 518
 519         if (link) {
 520                 ret = sysfs_create_link(&pmu_dev->kobj, &cs_dev->kobj, entry);
 521                 if (ret)
 522                         return ret;
 523                 per_cpu(csdev_src, cpu) = csdev;
 524         } else {
 525                 sysfs_remove_link(&pmu_dev->kobj, entry);
 526                 per_cpu(csdev_src, cpu) = NULL;
 527         }
 528
 529         return 0;
 530 }
 531 EXPORT_SYMBOL_GPL(etm_perf_symlink);
 532
 533 static ssize_t etm_perf_sink_name_show(struct device *dev,
 534                                        struct device_attribute *dattr,
 535                                        char *buf)
 536 {
 537         struct dev_ext_attribute *ea;
 538
 539         ea = container_of(dattr, struct dev_ext_attribute, attr);
 540         return scnprintf(buf, PAGE_SIZE, "0x%lx\n", (unsigned long)(ea->var));
 541 }
 542
 543 int etm_perf_add_symlink_sink(struct coresight_device *csdev)
 544 {
 545         int ret;
 546         unsigned long hash;
 547         const char *name;
 548         struct device *pmu_dev = etm_pmu.dev;
 549         struct device *dev = &csdev->dev;
 550         struct dev_ext_attribute *ea;
 551
 552         if (csdev->type != CORESIGHT_DEV_TYPE_SINK &&
 553             csdev->type != CORESIGHT_DEV_TYPE_LINKSINK)
 554                 return -EINVAL;
 555
 556         if (csdev->ea != NULL)
 557                 return -EINVAL;
 558
 559         if (!etm_perf_up)
 560                 return -EPROBE_DEFER;
 561
 562         ea = devm_kzalloc(dev, sizeof(*ea), GFP_KERNEL);
 563         if (!ea)
 564                 return -ENOMEM;
 565
 566         name = dev_name(dev);
 567         /* See function coresight_get_sink_by_id() to know where this is used */
 568         hash = hashlen_hash(hashlen_string(NULL, name));
 569
 570         sysfs_attr_init(&ea->attr.attr);
 571         ea->attr.attr.name = devm_kstrdup(dev, name, GFP_KERNEL);
 572         if (!ea->attr.attr.name)
 573                 return -ENOMEM;
 574
 575         ea->attr.attr.mode = 0444;
 576         ea->attr.show = etm_perf_sink_name_show;
 577         ea->var = (unsigned long *)hash;
 578
 579         ret = sysfs_add_file_to_group(&pmu_dev->kobj,
 580                                       &ea->attr.attr, "sinks");
 581
 582         if (!ret)
 583                 csdev->ea = ea;
 584
 585         return ret;
 586 }
 587
 588 void etm_perf_del_symlink_sink(struct coresight_device *csdev)
 589 {
 590         struct device *pmu_dev = etm_pmu.dev;
 591         struct dev_ext_attribute *ea = csdev->ea;
 592
 593         if (csdev->type != CORESIGHT_DEV_TYPE_SINK &&
 594             csdev->type != CORESIGHT_DEV_TYPE_LINKSINK)
 595                 return;
 596
 597         if (!ea)
 598                 return;
 599
 600         sysfs_remove_file_from_group(&pmu_dev->kobj,
 601                                      &ea->attr.attr, "sinks");
 602         csdev->ea = NULL;
 603 }
 604
 605 int __init etm_perf_init(void)
 606 {
 607         int ret;
 608
 609         etm_pmu.capabilities            = (PERF_PMU_CAP_EXCLUSIVE |
 610                                            PERF_PMU_CAP_ITRACE);
 611
 612         etm_pmu.attr_groups             = etm_pmu_attr_groups;
 613         etm_pmu.task_ctx_nr             = perf_sw_context;
 614         etm_pmu.read                    = etm_event_read;
 615         etm_pmu.event_init              = etm_event_init;
 616         etm_pmu.setup_aux               = etm_setup_aux;
 617         etm_pmu.free_aux                = etm_free_aux;
 618         etm_pmu.start                   = etm_event_start;
 619         etm_pmu.stop                    = etm_event_stop;
 620         etm_pmu.add                     = etm_event_add;
 621         etm_pmu.del                     = etm_event_del;
 622         etm_pmu.addr_filters_sync       = etm_addr_filters_sync;
 623         etm_pmu.addr_filters_validate   = etm_addr_filters_validate;
 624         etm_pmu.nr_addr_filters         = ETM_ADDR_CMP_MAX;
 625
 626         ret = perf_pmu_register(&etm_pmu, CORESIGHT_ETM_PMU_NAME, -1);
 627         if (ret == 0)
 628                 etm_perf_up = true;
 629
 630         return ret;
 631 }
 632
 633 void __exit etm_perf_exit(void)
 634 {
 635         perf_pmu_unregister(&etm_pmu);
 636 }