| 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/bitfield.h> |
| 8 | #include <linux/coresight.h> |
| 9 | #include <linux/coresight-pmu.h> |
| 10 | #include <linux/cpumask.h> |
| 11 | #include <linux/device.h> |
| 12 | #include <linux/list.h> |
| 13 | #include <linux/mm.h> |
| 14 | #include <linux/init.h> |
| 15 | #include <linux/perf_event.h> |
| 16 | #include <linux/percpu-defs.h> |
| 17 | #include <linux/slab.h> |
| 18 | #include <linux/stringhash.h> |
| 19 | #include <linux/types.h> |
| 20 | #include <linux/workqueue.h> |
| 21 | |
| 22 | #include "coresight-config.h" |
| 23 | #include "coresight-etm-perf.h" |
| 24 | #include "coresight-priv.h" |
| 25 | #include "coresight-syscfg.h" |
| 26 | #include "coresight-trace-id.h" |
| 27 | |
| 28 | static struct pmu etm_pmu; |
| 29 | static bool etm_perf_up; |
| 30 | |
| 31 | /* |
| 32 | * An ETM context for a running event includes the perf aux handle |
| 33 | * and aux_data. For ETM, the aux_data (etm_event_data), consists of |
| 34 | * the trace path and the sink configuration. The event data is accessible |
| 35 | * via perf_get_aux(handle). However, a sink could "end" a perf output |
| 36 | * handle via the IRQ handler. And if the "sink" encounters a failure |
| 37 | * to "begin" another session (e.g due to lack of space in the buffer), |
| 38 | * the handle will be cleared. Thus, the event_data may not be accessible |
| 39 | * from the handle when we get to the etm_event_stop(), which is required |
| 40 | * for stopping the trace path. The event_data is guaranteed to stay alive |
| 41 | * until "free_aux()", which cannot happen as long as the event is active on |
| 42 | * the ETM. Thus the event_data for the session must be part of the ETM context |
| 43 | * to make sure we can disable the trace path. |
| 44 | */ |
| 45 | struct etm_ctxt { |
| 46 | struct perf_output_handle handle; |
| 47 | struct etm_event_data *event_data; |
| 48 | }; |
| 49 | |
| 50 | static DEFINE_PER_CPU(struct etm_ctxt, etm_ctxt); |
| 51 | static DEFINE_PER_CPU(struct coresight_device *, csdev_src); |
| 52 | |
| 53 | /* |
| 54 | * The PMU formats were orignally for ETMv3.5/PTM's ETMCR 'config'; |
| 55 | * now take them as general formats and apply on all ETMs. |
| 56 | */ |
| 57 | PMU_FORMAT_ATTR(branch_broadcast, "config:" __stringify(ETM_OPT_BRANCH_BROADCAST)); |
| 58 | PMU_FORMAT_ATTR(cycacc, "config:" __stringify(ETM_OPT_CYCACC)); |
| 59 | /* contextid1 enables tracing CONTEXTIDR_EL1 for ETMv4 */ |
| 60 | PMU_FORMAT_ATTR(contextid1, "config:" __stringify(ETM_OPT_CTXTID)); |
| 61 | /* contextid2 enables tracing CONTEXTIDR_EL2 for ETMv4 */ |
| 62 | PMU_FORMAT_ATTR(contextid2, "config:" __stringify(ETM_OPT_CTXTID2)); |
| 63 | PMU_FORMAT_ATTR(timestamp, "config:" __stringify(ETM_OPT_TS)); |
| 64 | PMU_FORMAT_ATTR(retstack, "config:" __stringify(ETM_OPT_RETSTK)); |
| 65 | /* preset - if sink ID is used as a configuration selector */ |
| 66 | PMU_FORMAT_ATTR(preset, "config:0-3" ); |
| 67 | /* Sink ID - same for all ETMs */ |
| 68 | PMU_FORMAT_ATTR(sinkid, "config2:0-31" ); |
| 69 | /* config ID - set if a system configuration is selected */ |
| 70 | PMU_FORMAT_ATTR(configid, "config2:32-63" ); |
| 71 | PMU_FORMAT_ATTR(cc_threshold, "config3:0-11" ); |
| 72 | |
| 73 | |
| 74 | /* |
| 75 | * contextid always traces the "PID". The PID is in CONTEXTIDR_EL1 |
| 76 | * when the kernel is running at EL1; when the kernel is at EL2, |
| 77 | * the PID is in CONTEXTIDR_EL2. |
| 78 | */ |
| 79 | static ssize_t format_attr_contextid_show(struct device *dev, |
| 80 | struct device_attribute *attr, |
| 81 | char *page) |
| 82 | { |
| 83 | int pid_fmt = ETM_OPT_CTXTID; |
| 84 | |
| 85 | #if IS_ENABLED(CONFIG_CORESIGHT_SOURCE_ETM4X) |
| 86 | pid_fmt = is_kernel_in_hyp_mode() ? ETM_OPT_CTXTID2 : ETM_OPT_CTXTID; |
| 87 | #endif |
| 88 | return sprintf(buf: page, fmt: "config:%d\n" , pid_fmt); |
| 89 | } |
| 90 | |
| 91 | static struct device_attribute format_attr_contextid = |
| 92 | __ATTR(contextid, 0444, format_attr_contextid_show, NULL); |
| 93 | |
| 94 | static struct attribute *etm_config_formats_attr[] = { |
| 95 | &format_attr_cycacc.attr, |
| 96 | &format_attr_contextid.attr, |
| 97 | &format_attr_contextid1.attr, |
| 98 | &format_attr_contextid2.attr, |
| 99 | &format_attr_timestamp.attr, |
| 100 | &format_attr_retstack.attr, |
| 101 | &format_attr_sinkid.attr, |
| 102 | &format_attr_preset.attr, |
| 103 | &format_attr_configid.attr, |
| 104 | &format_attr_branch_broadcast.attr, |
| 105 | &format_attr_cc_threshold.attr, |
| 106 | NULL, |
| 107 | }; |
| 108 | |
| 109 | static const struct attribute_group etm_pmu_format_group = { |
| 110 | .name = "format" , |
| 111 | .attrs = etm_config_formats_attr, |
| 112 | }; |
| 113 | |
| 114 | static struct attribute *etm_config_sinks_attr[] = { |
| 115 | NULL, |
| 116 | }; |
| 117 | |
| 118 | static const struct attribute_group etm_pmu_sinks_group = { |
| 119 | .name = "sinks" , |
| 120 | .attrs = etm_config_sinks_attr, |
| 121 | }; |
| 122 | |
| 123 | static struct attribute *etm_config_events_attr[] = { |
| 124 | NULL, |
| 125 | }; |
| 126 | |
| 127 | static const struct attribute_group etm_pmu_events_group = { |
| 128 | .name = "events" , |
| 129 | .attrs = etm_config_events_attr, |
| 130 | }; |
| 131 | |
| 132 | static const struct attribute_group *etm_pmu_attr_groups[] = { |
| 133 | &etm_pmu_format_group, |
| 134 | &etm_pmu_sinks_group, |
| 135 | &etm_pmu_events_group, |
| 136 | NULL, |
| 137 | }; |
| 138 | |
| 139 | static inline struct coresight_path ** |
| 140 | etm_event_cpu_path_ptr(struct etm_event_data *data, int cpu) |
| 141 | { |
| 142 | return per_cpu_ptr(data->path, cpu); |
| 143 | } |
| 144 | |
| 145 | static inline struct coresight_path * |
| 146 | etm_event_cpu_path(struct etm_event_data *data, int cpu) |
| 147 | { |
| 148 | return *etm_event_cpu_path_ptr(data, cpu); |
| 149 | } |
| 150 | |
| 151 | static void etm_event_read(struct perf_event *event) {} |
| 152 | |
| 153 | static int etm_addr_filters_alloc(struct perf_event *event) |
| 154 | { |
| 155 | struct etm_filters *filters; |
| 156 | int node = event->cpu == -1 ? -1 : cpu_to_node(cpu: event->cpu); |
| 157 | |
| 158 | filters = kzalloc_node(sizeof(struct etm_filters), GFP_KERNEL, node); |
| 159 | if (!filters) |
| 160 | return -ENOMEM; |
| 161 | |
| 162 | if (event->parent) |
| 163 | memcpy(filters, event->parent->hw.addr_filters, |
| 164 | sizeof(*filters)); |
| 165 | |
| 166 | event->hw.addr_filters = filters; |
| 167 | |
| 168 | return 0; |
| 169 | } |
| 170 | |
| 171 | static void etm_event_destroy(struct perf_event *event) |
| 172 | { |
| 173 | kfree(objp: event->hw.addr_filters); |
| 174 | event->hw.addr_filters = NULL; |
| 175 | } |
| 176 | |
| 177 | static int etm_event_init(struct perf_event *event) |
| 178 | { |
| 179 | int ret = 0; |
| 180 | |
| 181 | if (event->attr.type != etm_pmu.type) { |
| 182 | ret = -ENOENT; |
| 183 | goto out; |
| 184 | } |
| 185 | |
| 186 | ret = etm_addr_filters_alloc(event); |
| 187 | if (ret) |
| 188 | goto out; |
| 189 | |
| 190 | event->destroy = etm_event_destroy; |
| 191 | out: |
| 192 | return ret; |
| 193 | } |
| 194 | |
| 195 | static void free_sink_buffer(struct etm_event_data *event_data) |
| 196 | { |
| 197 | int cpu; |
| 198 | cpumask_t *mask = &event_data->mask; |
| 199 | struct coresight_device *sink; |
| 200 | |
| 201 | if (!event_data->snk_config) |
| 202 | return; |
| 203 | |
| 204 | if (WARN_ON(cpumask_empty(mask))) |
| 205 | return; |
| 206 | |
| 207 | cpu = cpumask_first(srcp: mask); |
| 208 | sink = coresight_get_sink(path: etm_event_cpu_path(data: event_data, cpu)); |
| 209 | sink_ops(sink)->free_buffer(event_data->snk_config); |
| 210 | } |
| 211 | |
| 212 | static void free_event_data(struct work_struct *work) |
| 213 | { |
| 214 | int cpu; |
| 215 | cpumask_t *mask; |
| 216 | struct etm_event_data *event_data; |
| 217 | |
| 218 | event_data = container_of(work, struct etm_event_data, work); |
| 219 | mask = &event_data->mask; |
| 220 | |
| 221 | /* Free the sink buffers, if there are any */ |
| 222 | free_sink_buffer(event_data); |
| 223 | |
| 224 | /* clear any configuration we were using */ |
| 225 | if (event_data->cfg_hash) |
| 226 | cscfg_deactivate_config(cfg_hash: event_data->cfg_hash); |
| 227 | |
| 228 | for_each_cpu(cpu, mask) { |
| 229 | struct coresight_path **ppath; |
| 230 | |
| 231 | ppath = etm_event_cpu_path_ptr(data: event_data, cpu); |
| 232 | if (!(IS_ERR_OR_NULL(ptr: *ppath))) { |
| 233 | struct coresight_device *sink = coresight_get_sink(path: *ppath); |
| 234 | |
| 235 | /* |
| 236 | * Mark perf event as done for trace id allocator, but don't call |
| 237 | * coresight_trace_id_put_cpu_id_map() on individual IDs. Perf sessions |
| 238 | * never free trace IDs to ensure that the ID associated with a CPU |
| 239 | * cannot change during their and other's concurrent sessions. Instead, |
| 240 | * a refcount is used so that the last event to call |
| 241 | * coresight_trace_id_perf_stop() frees all IDs. |
| 242 | */ |
| 243 | coresight_trace_id_perf_stop(id_map: &sink->perf_sink_id_map); |
| 244 | |
| 245 | coresight_release_path(path: *ppath); |
| 246 | } |
| 247 | *ppath = NULL; |
| 248 | } |
| 249 | |
| 250 | free_percpu(pdata: event_data->path); |
| 251 | kfree(objp: event_data); |
| 252 | } |
| 253 | |
| 254 | static void *alloc_event_data(int cpu) |
| 255 | { |
| 256 | cpumask_t *mask; |
| 257 | struct etm_event_data *event_data; |
| 258 | |
| 259 | /* First get memory for the session's data */ |
| 260 | event_data = kzalloc(sizeof(struct etm_event_data), GFP_KERNEL); |
| 261 | if (!event_data) |
| 262 | return NULL; |
| 263 | |
| 264 | |
| 265 | mask = &event_data->mask; |
| 266 | if (cpu != -1) |
| 267 | cpumask_set_cpu(cpu, dstp: mask); |
| 268 | else |
| 269 | cpumask_copy(dstp: mask, cpu_present_mask); |
| 270 | |
| 271 | /* |
| 272 | * Each CPU has a single path between source and destination. As such |
| 273 | * allocate an array using CPU numbers as indexes. That way a path |
| 274 | * for any CPU can easily be accessed at any given time. We proceed |
| 275 | * the same way for sessions involving a single CPU. The cost of |
| 276 | * unused memory when dealing with single CPU trace scenarios is small |
| 277 | * compared to the cost of searching through an optimized array. |
| 278 | */ |
| 279 | event_data->path = alloc_percpu(struct coresight_path *); |
| 280 | |
| 281 | if (!event_data->path) { |
| 282 | kfree(objp: event_data); |
| 283 | return NULL; |
| 284 | } |
| 285 | |
| 286 | return event_data; |
| 287 | } |
| 288 | |
| 289 | static void etm_free_aux(void *data) |
| 290 | { |
| 291 | struct etm_event_data *event_data = data; |
| 292 | |
| 293 | schedule_work(work: &event_data->work); |
| 294 | } |
| 295 | |
| 296 | /* |
| 297 | * Check if two given sinks are compatible with each other, |
| 298 | * so that they can use the same sink buffers, when an event |
| 299 | * moves around. |
| 300 | */ |
| 301 | static bool sinks_compatible(struct coresight_device *a, |
| 302 | struct coresight_device *b) |
| 303 | { |
| 304 | if (!a || !b) |
| 305 | return false; |
| 306 | /* |
| 307 | * If the sinks are of the same subtype and driven |
| 308 | * by the same driver, we can use the same buffer |
| 309 | * on these sinks. |
| 310 | */ |
| 311 | return (a->subtype.sink_subtype == b->subtype.sink_subtype) && |
| 312 | (sink_ops(a) == sink_ops(b)); |
| 313 | } |
| 314 | |
| 315 | static void *etm_setup_aux(struct perf_event *event, void **pages, |
| 316 | int nr_pages, bool overwrite) |
| 317 | { |
| 318 | u32 id, cfg_hash; |
| 319 | int cpu = event->cpu; |
| 320 | cpumask_t *mask; |
| 321 | struct coresight_device *sink = NULL; |
| 322 | struct coresight_device *user_sink = NULL, *last_sink = NULL; |
| 323 | struct etm_event_data *event_data = NULL; |
| 324 | |
| 325 | event_data = alloc_event_data(cpu); |
| 326 | if (!event_data) |
| 327 | return NULL; |
| 328 | INIT_WORK(&event_data->work, free_event_data); |
| 329 | |
| 330 | /* First get the selected sink from user space. */ |
| 331 | if (event->attr.config2 & GENMASK_ULL(31, 0)) { |
| 332 | id = (u32)event->attr.config2; |
| 333 | sink = user_sink = coresight_get_sink_by_id(id); |
| 334 | } |
| 335 | |
| 336 | /* check if user wants a coresight configuration selected */ |
| 337 | cfg_hash = (u32)((event->attr.config2 & GENMASK_ULL(63, 32)) >> 32); |
| 338 | if (cfg_hash) { |
| 339 | if (cscfg_activate_config(cfg_hash)) |
| 340 | goto err; |
| 341 | event_data->cfg_hash = cfg_hash; |
| 342 | } |
| 343 | |
| 344 | mask = &event_data->mask; |
| 345 | |
| 346 | /* |
| 347 | * Setup the path for each CPU in a trace session. We try to build |
| 348 | * trace path for each CPU in the mask. If we don't find an ETM |
| 349 | * for the CPU or fail to build a path, we clear the CPU from the |
| 350 | * mask and continue with the rest. If ever we try to trace on those |
| 351 | * CPUs, we can handle it and fail the session. |
| 352 | */ |
| 353 | for_each_cpu(cpu, mask) { |
| 354 | struct coresight_path *path; |
| 355 | struct coresight_device *csdev; |
| 356 | |
| 357 | csdev = per_cpu(csdev_src, cpu); |
| 358 | /* |
| 359 | * If there is no ETM associated with this CPU clear it from |
| 360 | * the mask and continue with the rest. If ever we try to trace |
| 361 | * on this CPU, we handle it accordingly. |
| 362 | */ |
| 363 | if (!csdev) { |
| 364 | cpumask_clear_cpu(cpu, dstp: mask); |
| 365 | continue; |
| 366 | } |
| 367 | |
| 368 | /* |
| 369 | * If AUX pause feature is enabled but the ETM driver does not |
| 370 | * support the operations, clear this CPU from the mask and |
| 371 | * continue to next one. |
| 372 | */ |
| 373 | if (event->attr.aux_start_paused && |
| 374 | (!source_ops(csdev)->pause_perf || !source_ops(csdev)->resume_perf)) { |
| 375 | dev_err_once(&csdev->dev, "AUX pause is not supported.\n" ); |
| 376 | cpumask_clear_cpu(cpu, dstp: mask); |
| 377 | continue; |
| 378 | } |
| 379 | |
| 380 | /* |
| 381 | * No sink provided - look for a default sink for all the ETMs, |
| 382 | * where this event can be scheduled. |
| 383 | * We allocate the sink specific buffers only once for this |
| 384 | * event. If the ETMs have different default sink devices, we |
| 385 | * can only use a single "type" of sink as the event can carry |
| 386 | * only one sink specific buffer. Thus we have to make sure |
| 387 | * that the sinks are of the same type and driven by the same |
| 388 | * driver, as the one we allocate the buffer for. As such |
| 389 | * we choose the first sink and check if the remaining ETMs |
| 390 | * have a compatible default sink. We don't trace on a CPU |
| 391 | * if the sink is not compatible. |
| 392 | */ |
| 393 | if (!user_sink) { |
| 394 | /* Find the default sink for this ETM */ |
| 395 | sink = coresight_find_default_sink(csdev); |
| 396 | if (!sink) { |
| 397 | cpumask_clear_cpu(cpu, dstp: mask); |
| 398 | continue; |
| 399 | } |
| 400 | |
| 401 | /* Check if this sink compatible with the last sink */ |
| 402 | if (last_sink && !sinks_compatible(a: last_sink, b: sink)) { |
| 403 | cpumask_clear_cpu(cpu, dstp: mask); |
| 404 | continue; |
| 405 | } |
| 406 | last_sink = sink; |
| 407 | } |
| 408 | |
| 409 | /* |
| 410 | * Building a path doesn't enable it, it simply builds a |
| 411 | * list of devices from source to sink that can be |
| 412 | * referenced later when the path is actually needed. |
| 413 | */ |
| 414 | path = coresight_build_path(csdev, sink); |
| 415 | if (IS_ERR(ptr: path)) { |
| 416 | cpumask_clear_cpu(cpu, dstp: mask); |
| 417 | continue; |
| 418 | } |
| 419 | |
| 420 | /* ensure we can allocate a trace ID for this CPU */ |
| 421 | coresight_path_assign_trace_id(path, mode: CS_MODE_PERF); |
| 422 | if (!IS_VALID_CS_TRACE_ID(path->trace_id)) { |
| 423 | cpumask_clear_cpu(cpu, dstp: mask); |
| 424 | coresight_release_path(path); |
| 425 | continue; |
| 426 | } |
| 427 | |
| 428 | coresight_trace_id_perf_start(id_map: &sink->perf_sink_id_map); |
| 429 | *etm_event_cpu_path_ptr(data: event_data, cpu) = path; |
| 430 | } |
| 431 | |
| 432 | /* no sink found for any CPU - cannot trace */ |
| 433 | if (!sink) |
| 434 | goto err; |
| 435 | |
| 436 | /* If we don't have any CPUs ready for tracing, abort */ |
| 437 | cpu = cpumask_first(srcp: mask); |
| 438 | if (cpu >= nr_cpu_ids) |
| 439 | goto err; |
| 440 | |
| 441 | if (!sink_ops(sink)->alloc_buffer || !sink_ops(sink)->free_buffer) |
| 442 | goto err; |
| 443 | |
| 444 | /* |
| 445 | * Allocate the sink buffer for this session. All the sinks |
| 446 | * where this event can be scheduled are ensured to be of the |
| 447 | * same type. Thus the same sink configuration is used by the |
| 448 | * sinks. |
| 449 | */ |
| 450 | event_data->snk_config = |
| 451 | sink_ops(sink)->alloc_buffer(sink, event, pages, |
| 452 | nr_pages, overwrite); |
| 453 | if (!event_data->snk_config) |
| 454 | goto err; |
| 455 | |
| 456 | out: |
| 457 | return event_data; |
| 458 | |
| 459 | err: |
| 460 | etm_free_aux(data: event_data); |
| 461 | event_data = NULL; |
| 462 | goto out; |
| 463 | } |
| 464 | |
| 465 | static int etm_event_resume(struct coresight_device *csdev, |
| 466 | struct etm_ctxt *ctxt) |
| 467 | { |
| 468 | if (!ctxt->event_data) |
| 469 | return 0; |
| 470 | |
| 471 | return coresight_resume_source(csdev); |
| 472 | } |
| 473 | |
| 474 | static void etm_event_start(struct perf_event *event, int flags) |
| 475 | { |
| 476 | int cpu = smp_processor_id(); |
| 477 | struct etm_event_data *event_data; |
| 478 | struct etm_ctxt *ctxt = this_cpu_ptr(&etm_ctxt); |
| 479 | struct perf_output_handle *handle = &ctxt->handle; |
| 480 | struct coresight_device *sink, *csdev = per_cpu(csdev_src, cpu); |
| 481 | struct coresight_path *path; |
| 482 | u64 hw_id; |
| 483 | |
| 484 | if (!csdev) |
| 485 | goto fail; |
| 486 | |
| 487 | if (flags & PERF_EF_RESUME) { |
| 488 | if (etm_event_resume(csdev, ctxt) < 0) { |
| 489 | dev_err(&csdev->dev, "Failed to resume ETM event.\n" ); |
| 490 | goto fail; |
| 491 | } |
| 492 | return; |
| 493 | } |
| 494 | |
| 495 | /* Have we messed up our tracking ? */ |
| 496 | if (WARN_ON(ctxt->event_data)) |
| 497 | goto fail; |
| 498 | |
| 499 | /* |
| 500 | * Deal with the ring buffer API and get a handle on the |
| 501 | * session's information. |
| 502 | */ |
| 503 | event_data = perf_aux_output_begin(handle, event); |
| 504 | if (!event_data) |
| 505 | goto fail; |
| 506 | |
| 507 | /* |
| 508 | * Check if this ETM is allowed to trace, as decided |
| 509 | * at etm_setup_aux(). This could be due to an unreachable |
| 510 | * sink from this ETM. We can't do much in this case if |
| 511 | * the sink was specified or hinted to the driver. For |
| 512 | * now, simply don't record anything on this ETM. |
| 513 | * |
| 514 | * As such we pretend that everything is fine, and let |
| 515 | * it continue without actually tracing. The event could |
| 516 | * continue tracing when it moves to a CPU where it is |
| 517 | * reachable to a sink. |
| 518 | */ |
| 519 | if (!cpumask_test_cpu(cpu, cpumask: &event_data->mask)) |
| 520 | goto out; |
| 521 | |
| 522 | path = etm_event_cpu_path(data: event_data, cpu); |
| 523 | /* We need a sink, no need to continue without one */ |
| 524 | sink = coresight_get_sink(path); |
| 525 | if (WARN_ON_ONCE(!sink)) |
| 526 | goto fail_end_stop; |
| 527 | |
| 528 | /* Nothing will happen without a path */ |
| 529 | if (coresight_enable_path(path, mode: CS_MODE_PERF, sink_data: handle)) |
| 530 | goto fail_end_stop; |
| 531 | |
| 532 | /* Finally enable the tracer */ |
| 533 | if (source_ops(csdev)->enable(csdev, event, CS_MODE_PERF, path)) |
| 534 | goto fail_disable_path; |
| 535 | |
| 536 | /* |
| 537 | * output cpu / trace ID in perf record, once for the lifetime |
| 538 | * of the event. |
| 539 | */ |
| 540 | if (!cpumask_test_cpu(cpu, cpumask: &event_data->aux_hwid_done)) { |
| 541 | cpumask_set_cpu(cpu, dstp: &event_data->aux_hwid_done); |
| 542 | |
| 543 | hw_id = FIELD_PREP(CS_AUX_HW_ID_MAJOR_VERSION_MASK, |
| 544 | CS_AUX_HW_ID_MAJOR_VERSION); |
| 545 | hw_id |= FIELD_PREP(CS_AUX_HW_ID_MINOR_VERSION_MASK, |
| 546 | CS_AUX_HW_ID_MINOR_VERSION); |
| 547 | hw_id |= FIELD_PREP(CS_AUX_HW_ID_TRACE_ID_MASK, path->trace_id); |
| 548 | hw_id |= FIELD_PREP(CS_AUX_HW_ID_SINK_ID_MASK, coresight_get_sink_id(sink)); |
| 549 | |
| 550 | perf_report_aux_output_id(event, hw_id); |
| 551 | } |
| 552 | |
| 553 | out: |
| 554 | /* Tell the perf core the event is alive */ |
| 555 | event->hw.state = 0; |
| 556 | /* Save the event_data for this ETM */ |
| 557 | ctxt->event_data = event_data; |
| 558 | return; |
| 559 | |
| 560 | fail_disable_path: |
| 561 | coresight_disable_path(path); |
| 562 | fail_end_stop: |
| 563 | /* |
| 564 | * Check if the handle is still associated with the event, |
| 565 | * to handle cases where if the sink failed to start the |
| 566 | * trace and TRUNCATED the handle already. |
| 567 | */ |
| 568 | if (READ_ONCE(handle->event)) { |
| 569 | perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED); |
| 570 | perf_aux_output_end(handle, size: 0); |
| 571 | } |
| 572 | fail: |
| 573 | event->hw.state = PERF_HES_STOPPED; |
| 574 | return; |
| 575 | } |
| 576 | |
| 577 | static void etm_event_pause(struct perf_event *event, |
| 578 | struct coresight_device *csdev, |
| 579 | struct etm_ctxt *ctxt) |
| 580 | { |
| 581 | int cpu = smp_processor_id(); |
| 582 | struct coresight_device *sink; |
| 583 | struct perf_output_handle *handle = &ctxt->handle; |
| 584 | struct coresight_path *path; |
| 585 | unsigned long size; |
| 586 | |
| 587 | if (!ctxt->event_data) |
| 588 | return; |
| 589 | |
| 590 | /* Stop tracer */ |
| 591 | coresight_pause_source(csdev); |
| 592 | |
| 593 | path = etm_event_cpu_path(data: ctxt->event_data, cpu); |
| 594 | sink = coresight_get_sink(path); |
| 595 | if (WARN_ON_ONCE(!sink)) |
| 596 | return; |
| 597 | |
| 598 | /* |
| 599 | * The per CPU sink has own interrupt handling, it might have |
| 600 | * race condition with updating buffer on AUX trace pause if |
| 601 | * it is invoked from NMI. To avoid the race condition, |
| 602 | * disallows updating buffer for the per CPU sink case. |
| 603 | */ |
| 604 | if (coresight_is_percpu_sink(csdev: sink)) |
| 605 | return; |
| 606 | |
| 607 | if (WARN_ON_ONCE(handle->event != event)) |
| 608 | return; |
| 609 | |
| 610 | if (!sink_ops(sink)->update_buffer) |
| 611 | return; |
| 612 | |
| 613 | size = sink_ops(sink)->update_buffer(sink, handle, |
| 614 | ctxt->event_data->snk_config); |
| 615 | if (READ_ONCE(handle->event)) { |
| 616 | if (!size) |
| 617 | return; |
| 618 | |
| 619 | perf_aux_output_end(handle, size); |
| 620 | perf_aux_output_begin(handle, event); |
| 621 | } else { |
| 622 | WARN_ON_ONCE(size); |
| 623 | } |
| 624 | } |
| 625 | |
| 626 | static void etm_event_stop(struct perf_event *event, int mode) |
| 627 | { |
| 628 | int cpu = smp_processor_id(); |
| 629 | unsigned long size; |
| 630 | struct coresight_device *sink, *csdev = per_cpu(csdev_src, cpu); |
| 631 | struct etm_ctxt *ctxt = this_cpu_ptr(&etm_ctxt); |
| 632 | struct perf_output_handle *handle = &ctxt->handle; |
| 633 | struct etm_event_data *event_data; |
| 634 | struct coresight_path *path; |
| 635 | |
| 636 | if (mode & PERF_EF_PAUSE) |
| 637 | return etm_event_pause(event, csdev, ctxt); |
| 638 | |
| 639 | /* |
| 640 | * If we still have access to the event_data via handle, |
| 641 | * confirm that we haven't messed up the tracking. |
| 642 | */ |
| 643 | if (handle->event && |
| 644 | WARN_ON(perf_get_aux(handle) != ctxt->event_data)) |
| 645 | return; |
| 646 | |
| 647 | event_data = ctxt->event_data; |
| 648 | /* Clear the event_data as this ETM is stopping the trace. */ |
| 649 | ctxt->event_data = NULL; |
| 650 | |
| 651 | if (event->hw.state == PERF_HES_STOPPED) |
| 652 | return; |
| 653 | |
| 654 | /* We must have a valid event_data for a running event */ |
| 655 | if (WARN_ON(!event_data)) |
| 656 | return; |
| 657 | |
| 658 | /* |
| 659 | * Check if this ETM was allowed to trace, as decided at |
| 660 | * etm_setup_aux(). If it wasn't allowed to trace, then |
| 661 | * nothing needs to be torn down other than outputting a |
| 662 | * zero sized record. |
| 663 | */ |
| 664 | if (handle->event && (mode & PERF_EF_UPDATE) && |
| 665 | !cpumask_test_cpu(cpu, cpumask: &event_data->mask)) { |
| 666 | event->hw.state = PERF_HES_STOPPED; |
| 667 | perf_aux_output_end(handle, size: 0); |
| 668 | return; |
| 669 | } |
| 670 | |
| 671 | if (!csdev) |
| 672 | return; |
| 673 | |
| 674 | path = etm_event_cpu_path(data: event_data, cpu); |
| 675 | if (!path) |
| 676 | return; |
| 677 | |
| 678 | sink = coresight_get_sink(path); |
| 679 | if (!sink) |
| 680 | return; |
| 681 | |
| 682 | /* stop tracer */ |
| 683 | coresight_disable_source(csdev, data: event); |
| 684 | |
| 685 | /* tell the core */ |
| 686 | event->hw.state = PERF_HES_STOPPED; |
| 687 | |
| 688 | /* |
| 689 | * If the handle is not bound to an event anymore |
| 690 | * (e.g, the sink driver was unable to restart the |
| 691 | * handle due to lack of buffer space), we don't |
| 692 | * have to do anything here. |
| 693 | */ |
| 694 | if (handle->event && (mode & PERF_EF_UPDATE)) { |
| 695 | if (WARN_ON_ONCE(handle->event != event)) |
| 696 | return; |
| 697 | |
| 698 | /* update trace information */ |
| 699 | if (!sink_ops(sink)->update_buffer) |
| 700 | return; |
| 701 | |
| 702 | size = sink_ops(sink)->update_buffer(sink, handle, |
| 703 | event_data->snk_config); |
| 704 | /* |
| 705 | * Make sure the handle is still valid as the |
| 706 | * sink could have closed it from an IRQ. |
| 707 | * The sink driver must handle the race with |
| 708 | * update_buffer() and IRQ. Thus either we |
| 709 | * should get a valid handle and valid size |
| 710 | * (which may be 0). |
| 711 | * |
| 712 | * But we should never get a non-zero size with |
| 713 | * an invalid handle. |
| 714 | */ |
| 715 | if (READ_ONCE(handle->event)) |
| 716 | perf_aux_output_end(handle, size); |
| 717 | else |
| 718 | WARN_ON(size); |
| 719 | } |
| 720 | |
| 721 | /* Disabling the path make its elements available to other sessions */ |
| 722 | coresight_disable_path(path); |
| 723 | } |
| 724 | |
| 725 | static int etm_event_add(struct perf_event *event, int mode) |
| 726 | { |
| 727 | int ret = 0; |
| 728 | struct hw_perf_event *hwc = &event->hw; |
| 729 | |
| 730 | if (mode & PERF_EF_START) { |
| 731 | etm_event_start(event, flags: 0); |
| 732 | if (hwc->state & PERF_HES_STOPPED) |
| 733 | ret = -EINVAL; |
| 734 | } else { |
| 735 | hwc->state = PERF_HES_STOPPED; |
| 736 | } |
| 737 | |
| 738 | return ret; |
| 739 | } |
| 740 | |
| 741 | static void etm_event_del(struct perf_event *event, int mode) |
| 742 | { |
| 743 | etm_event_stop(event, PERF_EF_UPDATE); |
| 744 | } |
| 745 | |
| 746 | static int etm_addr_filters_validate(struct list_head *filters) |
| 747 | { |
| 748 | bool range = false, address = false; |
| 749 | int index = 0; |
| 750 | struct perf_addr_filter *filter; |
| 751 | |
| 752 | list_for_each_entry(filter, filters, entry) { |
| 753 | /* |
| 754 | * No need to go further if there's no more |
| 755 | * room for filters. |
| 756 | */ |
| 757 | if (++index > ETM_ADDR_CMP_MAX) |
| 758 | return -EOPNOTSUPP; |
| 759 | |
| 760 | /* filter::size==0 means single address trigger */ |
| 761 | if (filter->size) { |
| 762 | /* |
| 763 | * The existing code relies on START/STOP filters |
| 764 | * being address filters. |
| 765 | */ |
| 766 | if (filter->action == PERF_ADDR_FILTER_ACTION_START || |
| 767 | filter->action == PERF_ADDR_FILTER_ACTION_STOP) |
| 768 | return -EOPNOTSUPP; |
| 769 | |
| 770 | range = true; |
| 771 | } else |
| 772 | address = true; |
| 773 | |
| 774 | /* |
| 775 | * At this time we don't allow range and start/stop filtering |
| 776 | * to cohabitate, they have to be mutually exclusive. |
| 777 | */ |
| 778 | if (range && address) |
| 779 | return -EOPNOTSUPP; |
| 780 | } |
| 781 | |
| 782 | return 0; |
| 783 | } |
| 784 | |
| 785 | static void etm_addr_filters_sync(struct perf_event *event) |
| 786 | { |
| 787 | struct perf_addr_filters_head *head = perf_event_addr_filters(event); |
| 788 | unsigned long start, stop; |
| 789 | struct perf_addr_filter_range *fr = event->addr_filter_ranges; |
| 790 | struct etm_filters *filters = event->hw.addr_filters; |
| 791 | struct etm_filter *etm_filter; |
| 792 | struct perf_addr_filter *filter; |
| 793 | int i = 0; |
| 794 | |
| 795 | list_for_each_entry(filter, &head->list, entry) { |
| 796 | start = fr[i].start; |
| 797 | stop = start + fr[i].size; |
| 798 | etm_filter = &filters->etm_filter[i]; |
| 799 | |
| 800 | switch (filter->action) { |
| 801 | case PERF_ADDR_FILTER_ACTION_FILTER: |
| 802 | etm_filter->start_addr = start; |
| 803 | etm_filter->stop_addr = stop; |
| 804 | etm_filter->type = ETM_ADDR_TYPE_RANGE; |
| 805 | break; |
| 806 | case PERF_ADDR_FILTER_ACTION_START: |
| 807 | etm_filter->start_addr = start; |
| 808 | etm_filter->type = ETM_ADDR_TYPE_START; |
| 809 | break; |
| 810 | case PERF_ADDR_FILTER_ACTION_STOP: |
| 811 | etm_filter->stop_addr = stop; |
| 812 | etm_filter->type = ETM_ADDR_TYPE_STOP; |
| 813 | break; |
| 814 | } |
| 815 | i++; |
| 816 | } |
| 817 | |
| 818 | filters->nr_filters = i; |
| 819 | } |
| 820 | |
| 821 | int etm_perf_symlink(struct coresight_device *csdev, bool link) |
| 822 | { |
| 823 | char entry[sizeof("cpu9999999" )]; |
| 824 | int ret = 0, cpu = source_ops(csdev)->cpu_id(csdev); |
| 825 | struct device *pmu_dev = etm_pmu.dev; |
| 826 | struct device *cs_dev = &csdev->dev; |
| 827 | |
| 828 | sprintf(buf: entry, fmt: "cpu%d" , cpu); |
| 829 | |
| 830 | if (!etm_perf_up) |
| 831 | return -EPROBE_DEFER; |
| 832 | |
| 833 | if (link) { |
| 834 | ret = sysfs_create_link(kobj: &pmu_dev->kobj, target: &cs_dev->kobj, name: entry); |
| 835 | if (ret) |
| 836 | return ret; |
| 837 | per_cpu(csdev_src, cpu) = csdev; |
| 838 | } else { |
| 839 | sysfs_remove_link(kobj: &pmu_dev->kobj, name: entry); |
| 840 | per_cpu(csdev_src, cpu) = NULL; |
| 841 | } |
| 842 | |
| 843 | return 0; |
| 844 | } |
| 845 | EXPORT_SYMBOL_GPL(etm_perf_symlink); |
| 846 | |
| 847 | static ssize_t etm_perf_sink_name_show(struct device *dev, |
| 848 | struct device_attribute *dattr, |
| 849 | char *buf) |
| 850 | { |
| 851 | struct dev_ext_attribute *ea; |
| 852 | |
| 853 | ea = container_of(dattr, struct dev_ext_attribute, attr); |
| 854 | return scnprintf(buf, PAGE_SIZE, fmt: "0x%lx\n" , (unsigned long)(ea->var)); |
| 855 | } |
| 856 | |
| 857 | static struct dev_ext_attribute * |
| 858 | etm_perf_add_symlink_group(struct device *dev, const char *name, const char *group_name) |
| 859 | { |
| 860 | struct dev_ext_attribute *ea; |
| 861 | unsigned long hash; |
| 862 | int ret; |
| 863 | struct device *pmu_dev = etm_pmu.dev; |
| 864 | |
| 865 | if (!etm_perf_up) |
| 866 | return ERR_PTR(error: -EPROBE_DEFER); |
| 867 | |
| 868 | ea = devm_kzalloc(dev, size: sizeof(*ea), GFP_KERNEL); |
| 869 | if (!ea) |
| 870 | return ERR_PTR(error: -ENOMEM); |
| 871 | |
| 872 | /* |
| 873 | * If this function is called adding a sink then the hash is used for |
| 874 | * sink selection - see function coresight_get_sink_by_id(). |
| 875 | * If adding a configuration then the hash is used for selection in |
| 876 | * cscfg_activate_config() |
| 877 | */ |
| 878 | hash = hashlen_hash(hashlen_string(NULL, name)); |
| 879 | |
| 880 | sysfs_attr_init(&ea->attr.attr); |
| 881 | ea->attr.attr.name = devm_kstrdup(dev, s: name, GFP_KERNEL); |
| 882 | if (!ea->attr.attr.name) |
| 883 | return ERR_PTR(error: -ENOMEM); |
| 884 | |
| 885 | ea->attr.attr.mode = 0444; |
| 886 | ea->var = (unsigned long *)hash; |
| 887 | |
| 888 | ret = sysfs_add_file_to_group(kobj: &pmu_dev->kobj, |
| 889 | attr: &ea->attr.attr, group: group_name); |
| 890 | |
| 891 | return ret ? ERR_PTR(error: ret) : ea; |
| 892 | } |
| 893 | |
| 894 | int etm_perf_add_symlink_sink(struct coresight_device *csdev) |
| 895 | { |
| 896 | const char *name; |
| 897 | struct device *dev = &csdev->dev; |
| 898 | int err = 0; |
| 899 | |
| 900 | if (csdev->type != CORESIGHT_DEV_TYPE_SINK && |
| 901 | csdev->type != CORESIGHT_DEV_TYPE_LINKSINK) |
| 902 | return -EINVAL; |
| 903 | |
| 904 | if (csdev->ea != NULL) |
| 905 | return -EINVAL; |
| 906 | |
| 907 | name = dev_name(dev); |
| 908 | csdev->ea = etm_perf_add_symlink_group(dev, name, group_name: "sinks" ); |
| 909 | if (IS_ERR(ptr: csdev->ea)) { |
| 910 | err = PTR_ERR(ptr: csdev->ea); |
| 911 | csdev->ea = NULL; |
| 912 | } else |
| 913 | csdev->ea->attr.show = etm_perf_sink_name_show; |
| 914 | |
| 915 | return err; |
| 916 | } |
| 917 | |
| 918 | static void etm_perf_del_symlink_group(struct dev_ext_attribute *ea, const char *group_name) |
| 919 | { |
| 920 | struct device *pmu_dev = etm_pmu.dev; |
| 921 | |
| 922 | sysfs_remove_file_from_group(kobj: &pmu_dev->kobj, |
| 923 | attr: &ea->attr.attr, group: group_name); |
| 924 | } |
| 925 | |
| 926 | void etm_perf_del_symlink_sink(struct coresight_device *csdev) |
| 927 | { |
| 928 | if (csdev->type != CORESIGHT_DEV_TYPE_SINK && |
| 929 | csdev->type != CORESIGHT_DEV_TYPE_LINKSINK) |
| 930 | return; |
| 931 | |
| 932 | if (!csdev->ea) |
| 933 | return; |
| 934 | |
| 935 | etm_perf_del_symlink_group(ea: csdev->ea, group_name: "sinks" ); |
| 936 | csdev->ea = NULL; |
| 937 | } |
| 938 | |
| 939 | static ssize_t etm_perf_cscfg_event_show(struct device *dev, |
| 940 | struct device_attribute *dattr, |
| 941 | char *buf) |
| 942 | { |
| 943 | struct dev_ext_attribute *ea; |
| 944 | |
| 945 | ea = container_of(dattr, struct dev_ext_attribute, attr); |
| 946 | return scnprintf(buf, PAGE_SIZE, fmt: "configid=0x%lx\n" , (unsigned long)(ea->var)); |
| 947 | } |
| 948 | |
| 949 | int etm_perf_add_symlink_cscfg(struct device *dev, struct cscfg_config_desc *config_desc) |
| 950 | { |
| 951 | int err = 0; |
| 952 | |
| 953 | if (config_desc->event_ea != NULL) |
| 954 | return 0; |
| 955 | |
| 956 | config_desc->event_ea = etm_perf_add_symlink_group(dev, name: config_desc->name, group_name: "events" ); |
| 957 | |
| 958 | /* set the show function to the custom cscfg event */ |
| 959 | if (!IS_ERR(ptr: config_desc->event_ea)) |
| 960 | config_desc->event_ea->attr.show = etm_perf_cscfg_event_show; |
| 961 | else { |
| 962 | err = PTR_ERR(ptr: config_desc->event_ea); |
| 963 | config_desc->event_ea = NULL; |
| 964 | } |
| 965 | |
| 966 | return err; |
| 967 | } |
| 968 | |
| 969 | void etm_perf_del_symlink_cscfg(struct cscfg_config_desc *config_desc) |
| 970 | { |
| 971 | if (!config_desc->event_ea) |
| 972 | return; |
| 973 | |
| 974 | etm_perf_del_symlink_group(ea: config_desc->event_ea, group_name: "events" ); |
| 975 | config_desc->event_ea = NULL; |
| 976 | } |
| 977 | |
| 978 | int __init etm_perf_init(void) |
| 979 | { |
| 980 | int ret; |
| 981 | |
| 982 | etm_pmu.capabilities = (PERF_PMU_CAP_EXCLUSIVE | |
| 983 | PERF_PMU_CAP_ITRACE | |
| 984 | PERF_PMU_CAP_AUX_PAUSE); |
| 985 | |
| 986 | etm_pmu.attr_groups = etm_pmu_attr_groups; |
| 987 | etm_pmu.task_ctx_nr = perf_sw_context; |
| 988 | etm_pmu.read = etm_event_read; |
| 989 | etm_pmu.event_init = etm_event_init; |
| 990 | etm_pmu.setup_aux = etm_setup_aux; |
| 991 | etm_pmu.free_aux = etm_free_aux; |
| 992 | etm_pmu.start = etm_event_start; |
| 993 | etm_pmu.stop = etm_event_stop; |
| 994 | etm_pmu.add = etm_event_add; |
| 995 | etm_pmu.del = etm_event_del; |
| 996 | etm_pmu.addr_filters_sync = etm_addr_filters_sync; |
| 997 | etm_pmu.addr_filters_validate = etm_addr_filters_validate; |
| 998 | etm_pmu.nr_addr_filters = ETM_ADDR_CMP_MAX; |
| 999 | etm_pmu.module = THIS_MODULE; |
| 1000 | |
| 1001 | ret = perf_pmu_register(pmu: &etm_pmu, CORESIGHT_ETM_PMU_NAME, type: -1); |
| 1002 | if (ret == 0) |
| 1003 | etm_perf_up = true; |
| 1004 | |
| 1005 | return ret; |
| 1006 | } |
| 1007 | |
| 1008 | void etm_perf_exit(void) |
| 1009 | { |
| 1010 | perf_pmu_unregister(pmu: &etm_pmu); |
| 1011 | } |
| 1012 | |