| 1 | /* |
| 2 | * Performance events core code: |
| 3 | * |
| 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
| 5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
| 6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> |
| 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
| 8 | * |
| 9 | * For licensing details see kernel-base/COPYING |
| 10 | */ |
| 11 | |
| 12 | #include <linux/fs.h> |
| 13 | #include <linux/mm.h> |
| 14 | #include <linux/cpu.h> |
| 15 | #include <linux/smp.h> |
| 16 | #include <linux/idr.h> |
| 17 | #include <linux/file.h> |
| 18 | #include <linux/poll.h> |
| 19 | #include <linux/slab.h> |
| 20 | #include <linux/hash.h> |
| 21 | #include <linux/tick.h> |
| 22 | #include <linux/sysfs.h> |
| 23 | #include <linux/dcache.h> |
| 24 | #include <linux/percpu.h> |
| 25 | #include <linux/ptrace.h> |
| 26 | #include <linux/reboot.h> |
| 27 | #include <linux/vmstat.h> |
| 28 | #include <linux/device.h> |
| 29 | #include <linux/export.h> |
| 30 | #include <linux/vmalloc.h> |
| 31 | #include <linux/hardirq.h> |
| 32 | #include <linux/rculist.h> |
| 33 | #include <linux/uaccess.h> |
| 34 | #include <linux/syscalls.h> |
| 35 | #include <linux/anon_inodes.h> |
| 36 | #include <linux/kernel_stat.h> |
| 37 | #include <linux/cgroup.h> |
| 38 | #include <linux/perf_event.h> |
| 39 | #include <linux/ftrace_event.h> |
| 40 | #include <linux/hw_breakpoint.h> |
| 41 | #include <linux/mm_types.h> |
| 42 | #include <linux/module.h> |
| 43 | #include <linux/mman.h> |
| 44 | #include <linux/compat.h> |
| 45 | #include <linux/bpf.h> |
| 46 | #include <linux/filter.h> |
| 47 | |
| 48 | #include "internal.h" |
| 49 | |
| 50 | #include <asm/irq_regs.h> |
| 51 | |
| 52 | static struct workqueue_struct *perf_wq; |
| 53 | |
| 54 | struct remote_function_call { |
| 55 | struct task_struct *p; |
| 56 | int (*func)(void *info); |
| 57 | void *info; |
| 58 | int ret; |
| 59 | }; |
| 60 | |
| 61 | static void remote_function(void *data) |
| 62 | { |
| 63 | struct remote_function_call *tfc = data; |
| 64 | struct task_struct *p = tfc->p; |
| 65 | |
| 66 | if (p) { |
| 67 | tfc->ret = -EAGAIN; |
| 68 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) |
| 69 | return; |
| 70 | } |
| 71 | |
| 72 | tfc->ret = tfc->func(tfc->info); |
| 73 | } |
| 74 | |
| 75 | /** |
| 76 | * task_function_call - call a function on the cpu on which a task runs |
| 77 | * @p: the task to evaluate |
| 78 | * @func: the function to be called |
| 79 | * @info: the function call argument |
| 80 | * |
| 81 | * Calls the function @func when the task is currently running. This might |
| 82 | * be on the current CPU, which just calls the function directly |
| 83 | * |
| 84 | * returns: @func return value, or |
| 85 | * -ESRCH - when the process isn't running |
| 86 | * -EAGAIN - when the process moved away |
| 87 | */ |
| 88 | static int |
| 89 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) |
| 90 | { |
| 91 | struct remote_function_call data = { |
| 92 | .p = p, |
| 93 | .func = func, |
| 94 | .info = info, |
| 95 | .ret = -ESRCH, /* No such (running) process */ |
| 96 | }; |
| 97 | |
| 98 | if (task_curr(p)) |
| 99 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); |
| 100 | |
| 101 | return data.ret; |
| 102 | } |
| 103 | |
| 104 | /** |
| 105 | * cpu_function_call - call a function on the cpu |
| 106 | * @func: the function to be called |
| 107 | * @info: the function call argument |
| 108 | * |
| 109 | * Calls the function @func on the remote cpu. |
| 110 | * |
| 111 | * returns: @func return value or -ENXIO when the cpu is offline |
| 112 | */ |
| 113 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) |
| 114 | { |
| 115 | struct remote_function_call data = { |
| 116 | .p = NULL, |
| 117 | .func = func, |
| 118 | .info = info, |
| 119 | .ret = -ENXIO, /* No such CPU */ |
| 120 | }; |
| 121 | |
| 122 | smp_call_function_single(cpu, remote_function, &data, 1); |
| 123 | |
| 124 | return data.ret; |
| 125 | } |
| 126 | |
| 127 | #define EVENT_OWNER_KERNEL ((void *) -1) |
| 128 | |
| 129 | static bool is_kernel_event(struct perf_event *event) |
| 130 | { |
| 131 | return event->owner == EVENT_OWNER_KERNEL; |
| 132 | } |
| 133 | |
| 134 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
| 135 | PERF_FLAG_FD_OUTPUT |\ |
| 136 | PERF_FLAG_PID_CGROUP |\ |
| 137 | PERF_FLAG_FD_CLOEXEC) |
| 138 | |
| 139 | /* |
| 140 | * branch priv levels that need permission checks |
| 141 | */ |
| 142 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ |
| 143 | (PERF_SAMPLE_BRANCH_KERNEL |\ |
| 144 | PERF_SAMPLE_BRANCH_HV) |
| 145 | |
| 146 | enum event_type_t { |
| 147 | EVENT_FLEXIBLE = 0x1, |
| 148 | EVENT_PINNED = 0x2, |
| 149 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, |
| 150 | }; |
| 151 | |
| 152 | /* |
| 153 | * perf_sched_events : >0 events exist |
| 154 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu |
| 155 | */ |
| 156 | struct static_key_deferred perf_sched_events __read_mostly; |
| 157 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
| 158 | static DEFINE_PER_CPU(int, perf_sched_cb_usages); |
| 159 | |
| 160 | static atomic_t nr_mmap_events __read_mostly; |
| 161 | static atomic_t nr_comm_events __read_mostly; |
| 162 | static atomic_t nr_task_events __read_mostly; |
| 163 | static atomic_t nr_freq_events __read_mostly; |
| 164 | |
| 165 | static LIST_HEAD(pmus); |
| 166 | static DEFINE_MUTEX(pmus_lock); |
| 167 | static struct srcu_struct pmus_srcu; |
| 168 | |
| 169 | /* |
| 170 | * perf event paranoia level: |
| 171 | * -1 - not paranoid at all |
| 172 | * 0 - disallow raw tracepoint access for unpriv |
| 173 | * 1 - disallow cpu events for unpriv |
| 174 | * 2 - disallow kernel profiling for unpriv |
| 175 | */ |
| 176 | int sysctl_perf_event_paranoid __read_mostly = 1; |
| 177 | |
| 178 | /* Minimum for 512 kiB + 1 user control page */ |
| 179 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ |
| 180 | |
| 181 | /* |
| 182 | * max perf event sample rate |
| 183 | */ |
| 184 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
| 185 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) |
| 186 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 |
| 187 | |
| 188 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; |
| 189 | |
| 190 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); |
| 191 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; |
| 192 | |
| 193 | static int perf_sample_allowed_ns __read_mostly = |
| 194 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; |
| 195 | |
| 196 | void update_perf_cpu_limits(void) |
| 197 | { |
| 198 | u64 tmp = perf_sample_period_ns; |
| 199 | |
| 200 | tmp *= sysctl_perf_cpu_time_max_percent; |
| 201 | do_div(tmp, 100); |
| 202 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
| 203 | } |
| 204 | |
| 205 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
| 206 | |
| 207 | int perf_proc_update_handler(struct ctl_table *table, int write, |
| 208 | void __user *buffer, size_t *lenp, |
| 209 | loff_t *ppos) |
| 210 | { |
| 211 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
| 212 | |
| 213 | if (ret || !write) |
| 214 | return ret; |
| 215 | |
| 216 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); |
| 217 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
| 218 | update_perf_cpu_limits(); |
| 219 | |
| 220 | return 0; |
| 221 | } |
| 222 | |
| 223 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; |
| 224 | |
| 225 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, |
| 226 | void __user *buffer, size_t *lenp, |
| 227 | loff_t *ppos) |
| 228 | { |
| 229 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); |
| 230 | |
| 231 | if (ret || !write) |
| 232 | return ret; |
| 233 | |
| 234 | update_perf_cpu_limits(); |
| 235 | |
| 236 | return 0; |
| 237 | } |
| 238 | |
| 239 | /* |
| 240 | * perf samples are done in some very critical code paths (NMIs). |
| 241 | * If they take too much CPU time, the system can lock up and not |
| 242 | * get any real work done. This will drop the sample rate when |
| 243 | * we detect that events are taking too long. |
| 244 | */ |
| 245 | #define NR_ACCUMULATED_SAMPLES 128 |
| 246 | static DEFINE_PER_CPU(u64, running_sample_length); |
| 247 | |
| 248 | static void perf_duration_warn(struct irq_work *w) |
| 249 | { |
| 250 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
| 251 | u64 avg_local_sample_len; |
| 252 | u64 local_samples_len; |
| 253 | |
| 254 | local_samples_len = __this_cpu_read(running_sample_length); |
| 255 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; |
| 256 | |
| 257 | printk_ratelimited(KERN_WARNING |
| 258 | "perf interrupt took too long (%lld > %lld), lowering " |
| 259 | "kernel.perf_event_max_sample_rate to %d\n", |
| 260 | avg_local_sample_len, allowed_ns >> 1, |
| 261 | sysctl_perf_event_sample_rate); |
| 262 | } |
| 263 | |
| 264 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); |
| 265 | |
| 266 | void perf_sample_event_took(u64 sample_len_ns) |
| 267 | { |
| 268 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
| 269 | u64 avg_local_sample_len; |
| 270 | u64 local_samples_len; |
| 271 | |
| 272 | if (allowed_ns == 0) |
| 273 | return; |
| 274 | |
| 275 | /* decay the counter by 1 average sample */ |
| 276 | local_samples_len = __this_cpu_read(running_sample_length); |
| 277 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; |
| 278 | local_samples_len += sample_len_ns; |
| 279 | __this_cpu_write(running_sample_length, local_samples_len); |
| 280 | |
| 281 | /* |
| 282 | * note: this will be biased artifically low until we have |
| 283 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us |
| 284 | * from having to maintain a count. |
| 285 | */ |
| 286 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; |
| 287 | |
| 288 | if (avg_local_sample_len <= allowed_ns) |
| 289 | return; |
| 290 | |
| 291 | if (max_samples_per_tick <= 1) |
| 292 | return; |
| 293 | |
| 294 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); |
| 295 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; |
| 296 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
| 297 | |
| 298 | update_perf_cpu_limits(); |
| 299 | |
| 300 | if (!irq_work_queue(&perf_duration_work)) { |
| 301 | early_printk("perf interrupt took too long (%lld > %lld), lowering " |
| 302 | "kernel.perf_event_max_sample_rate to %d\n", |
| 303 | avg_local_sample_len, allowed_ns >> 1, |
| 304 | sysctl_perf_event_sample_rate); |
| 305 | } |
| 306 | } |
| 307 | |
| 308 | static atomic64_t perf_event_id; |
| 309 | |
| 310 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
| 311 | enum event_type_t event_type); |
| 312 | |
| 313 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
| 314 | enum event_type_t event_type, |
| 315 | struct task_struct *task); |
| 316 | |
| 317 | static void update_context_time(struct perf_event_context *ctx); |
| 318 | static u64 perf_event_time(struct perf_event *event); |
| 319 | |
| 320 | void __weak perf_event_print_debug(void) { } |
| 321 | |
| 322 | extern __weak const char *perf_pmu_name(void) |
| 323 | { |
| 324 | return "pmu"; |
| 325 | } |
| 326 | |
| 327 | static inline u64 perf_clock(void) |
| 328 | { |
| 329 | return local_clock(); |
| 330 | } |
| 331 | |
| 332 | static inline u64 perf_event_clock(struct perf_event *event) |
| 333 | { |
| 334 | return event->clock(); |
| 335 | } |
| 336 | |
| 337 | static inline struct perf_cpu_context * |
| 338 | __get_cpu_context(struct perf_event_context *ctx) |
| 339 | { |
| 340 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); |
| 341 | } |
| 342 | |
| 343 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
| 344 | struct perf_event_context *ctx) |
| 345 | { |
| 346 | raw_spin_lock(&cpuctx->ctx.lock); |
| 347 | if (ctx) |
| 348 | raw_spin_lock(&ctx->lock); |
| 349 | } |
| 350 | |
| 351 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, |
| 352 | struct perf_event_context *ctx) |
| 353 | { |
| 354 | if (ctx) |
| 355 | raw_spin_unlock(&ctx->lock); |
| 356 | raw_spin_unlock(&cpuctx->ctx.lock); |
| 357 | } |
| 358 | |
| 359 | #ifdef CONFIG_CGROUP_PERF |
| 360 | |
| 361 | static inline bool |
| 362 | perf_cgroup_match(struct perf_event *event) |
| 363 | { |
| 364 | struct perf_event_context *ctx = event->ctx; |
| 365 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
| 366 | |
| 367 | /* @event doesn't care about cgroup */ |
| 368 | if (!event->cgrp) |
| 369 | return true; |
| 370 | |
| 371 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ |
| 372 | if (!cpuctx->cgrp) |
| 373 | return false; |
| 374 | |
| 375 | /* |
| 376 | * Cgroup scoping is recursive. An event enabled for a cgroup is |
| 377 | * also enabled for all its descendant cgroups. If @cpuctx's |
| 378 | * cgroup is a descendant of @event's (the test covers identity |
| 379 | * case), it's a match. |
| 380 | */ |
| 381 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, |
| 382 | event->cgrp->css.cgroup); |
| 383 | } |
| 384 | |
| 385 | static inline void perf_detach_cgroup(struct perf_event *event) |
| 386 | { |
| 387 | css_put(&event->cgrp->css); |
| 388 | event->cgrp = NULL; |
| 389 | } |
| 390 | |
| 391 | static inline int is_cgroup_event(struct perf_event *event) |
| 392 | { |
| 393 | return event->cgrp != NULL; |
| 394 | } |
| 395 | |
| 396 | static inline u64 perf_cgroup_event_time(struct perf_event *event) |
| 397 | { |
| 398 | struct perf_cgroup_info *t; |
| 399 | |
| 400 | t = per_cpu_ptr(event->cgrp->info, event->cpu); |
| 401 | return t->time; |
| 402 | } |
| 403 | |
| 404 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) |
| 405 | { |
| 406 | struct perf_cgroup_info *info; |
| 407 | u64 now; |
| 408 | |
| 409 | now = perf_clock(); |
| 410 | |
| 411 | info = this_cpu_ptr(cgrp->info); |
| 412 | |
| 413 | info->time += now - info->timestamp; |
| 414 | info->timestamp = now; |
| 415 | } |
| 416 | |
| 417 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) |
| 418 | { |
| 419 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; |
| 420 | if (cgrp_out) |
| 421 | __update_cgrp_time(cgrp_out); |
| 422 | } |
| 423 | |
| 424 | static inline void update_cgrp_time_from_event(struct perf_event *event) |
| 425 | { |
| 426 | struct perf_cgroup *cgrp; |
| 427 | |
| 428 | /* |
| 429 | * ensure we access cgroup data only when needed and |
| 430 | * when we know the cgroup is pinned (css_get) |
| 431 | */ |
| 432 | if (!is_cgroup_event(event)) |
| 433 | return; |
| 434 | |
| 435 | cgrp = perf_cgroup_from_task(current); |
| 436 | /* |
| 437 | * Do not update time when cgroup is not active |
| 438 | */ |
| 439 | if (cgrp == event->cgrp) |
| 440 | __update_cgrp_time(event->cgrp); |
| 441 | } |
| 442 | |
| 443 | static inline void |
| 444 | perf_cgroup_set_timestamp(struct task_struct *task, |
| 445 | struct perf_event_context *ctx) |
| 446 | { |
| 447 | struct perf_cgroup *cgrp; |
| 448 | struct perf_cgroup_info *info; |
| 449 | |
| 450 | /* |
| 451 | * ctx->lock held by caller |
| 452 | * ensure we do not access cgroup data |
| 453 | * unless we have the cgroup pinned (css_get) |
| 454 | */ |
| 455 | if (!task || !ctx->nr_cgroups) |
| 456 | return; |
| 457 | |
| 458 | cgrp = perf_cgroup_from_task(task); |
| 459 | info = this_cpu_ptr(cgrp->info); |
| 460 | info->timestamp = ctx->timestamp; |
| 461 | } |
| 462 | |
| 463 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ |
| 464 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ |
| 465 | |
| 466 | /* |
| 467 | * reschedule events based on the cgroup constraint of task. |
| 468 | * |
| 469 | * mode SWOUT : schedule out everything |
| 470 | * mode SWIN : schedule in based on cgroup for next |
| 471 | */ |
| 472 | void perf_cgroup_switch(struct task_struct *task, int mode) |
| 473 | { |
| 474 | struct perf_cpu_context *cpuctx; |
| 475 | struct pmu *pmu; |
| 476 | unsigned long flags; |
| 477 | |
| 478 | /* |
| 479 | * disable interrupts to avoid geting nr_cgroup |
| 480 | * changes via __perf_event_disable(). Also |
| 481 | * avoids preemption. |
| 482 | */ |
| 483 | local_irq_save(flags); |
| 484 | |
| 485 | /* |
| 486 | * we reschedule only in the presence of cgroup |
| 487 | * constrained events. |
| 488 | */ |
| 489 | rcu_read_lock(); |
| 490 | |
| 491 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
| 492 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
| 493 | if (cpuctx->unique_pmu != pmu) |
| 494 | continue; /* ensure we process each cpuctx once */ |
| 495 | |
| 496 | /* |
| 497 | * perf_cgroup_events says at least one |
| 498 | * context on this CPU has cgroup events. |
| 499 | * |
| 500 | * ctx->nr_cgroups reports the number of cgroup |
| 501 | * events for a context. |
| 502 | */ |
| 503 | if (cpuctx->ctx.nr_cgroups > 0) { |
| 504 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
| 505 | perf_pmu_disable(cpuctx->ctx.pmu); |
| 506 | |
| 507 | if (mode & PERF_CGROUP_SWOUT) { |
| 508 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
| 509 | /* |
| 510 | * must not be done before ctxswout due |
| 511 | * to event_filter_match() in event_sched_out() |
| 512 | */ |
| 513 | cpuctx->cgrp = NULL; |
| 514 | } |
| 515 | |
| 516 | if (mode & PERF_CGROUP_SWIN) { |
| 517 | WARN_ON_ONCE(cpuctx->cgrp); |
| 518 | /* |
| 519 | * set cgrp before ctxsw in to allow |
| 520 | * event_filter_match() to not have to pass |
| 521 | * task around |
| 522 | */ |
| 523 | cpuctx->cgrp = perf_cgroup_from_task(task); |
| 524 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); |
| 525 | } |
| 526 | perf_pmu_enable(cpuctx->ctx.pmu); |
| 527 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); |
| 528 | } |
| 529 | } |
| 530 | |
| 531 | rcu_read_unlock(); |
| 532 | |
| 533 | local_irq_restore(flags); |
| 534 | } |
| 535 | |
| 536 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
| 537 | struct task_struct *next) |
| 538 | { |
| 539 | struct perf_cgroup *cgrp1; |
| 540 | struct perf_cgroup *cgrp2 = NULL; |
| 541 | |
| 542 | /* |
| 543 | * we come here when we know perf_cgroup_events > 0 |
| 544 | */ |
| 545 | cgrp1 = perf_cgroup_from_task(task); |
| 546 | |
| 547 | /* |
| 548 | * next is NULL when called from perf_event_enable_on_exec() |
| 549 | * that will systematically cause a cgroup_switch() |
| 550 | */ |
| 551 | if (next) |
| 552 | cgrp2 = perf_cgroup_from_task(next); |
| 553 | |
| 554 | /* |
| 555 | * only schedule out current cgroup events if we know |
| 556 | * that we are switching to a different cgroup. Otherwise, |
| 557 | * do no touch the cgroup events. |
| 558 | */ |
| 559 | if (cgrp1 != cgrp2) |
| 560 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); |
| 561 | } |
| 562 | |
| 563 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
| 564 | struct task_struct *task) |
| 565 | { |
| 566 | struct perf_cgroup *cgrp1; |
| 567 | struct perf_cgroup *cgrp2 = NULL; |
| 568 | |
| 569 | /* |
| 570 | * we come here when we know perf_cgroup_events > 0 |
| 571 | */ |
| 572 | cgrp1 = perf_cgroup_from_task(task); |
| 573 | |
| 574 | /* prev can never be NULL */ |
| 575 | cgrp2 = perf_cgroup_from_task(prev); |
| 576 | |
| 577 | /* |
| 578 | * only need to schedule in cgroup events if we are changing |
| 579 | * cgroup during ctxsw. Cgroup events were not scheduled |
| 580 | * out of ctxsw out if that was not the case. |
| 581 | */ |
| 582 | if (cgrp1 != cgrp2) |
| 583 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); |
| 584 | } |
| 585 | |
| 586 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, |
| 587 | struct perf_event_attr *attr, |
| 588 | struct perf_event *group_leader) |
| 589 | { |
| 590 | struct perf_cgroup *cgrp; |
| 591 | struct cgroup_subsys_state *css; |
| 592 | struct fd f = fdget(fd); |
| 593 | int ret = 0; |
| 594 | |
| 595 | if (!f.file) |
| 596 | return -EBADF; |
| 597 | |
| 598 | css = css_tryget_online_from_dir(f.file->f_path.dentry, |
| 599 | &perf_event_cgrp_subsys); |
| 600 | if (IS_ERR(css)) { |
| 601 | ret = PTR_ERR(css); |
| 602 | goto out; |
| 603 | } |
| 604 | |
| 605 | cgrp = container_of(css, struct perf_cgroup, css); |
| 606 | event->cgrp = cgrp; |
| 607 | |
| 608 | /* |
| 609 | * all events in a group must monitor |
| 610 | * the same cgroup because a task belongs |
| 611 | * to only one perf cgroup at a time |
| 612 | */ |
| 613 | if (group_leader && group_leader->cgrp != cgrp) { |
| 614 | perf_detach_cgroup(event); |
| 615 | ret = -EINVAL; |
| 616 | } |
| 617 | out: |
| 618 | fdput(f); |
| 619 | return ret; |
| 620 | } |
| 621 | |
| 622 | static inline void |
| 623 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) |
| 624 | { |
| 625 | struct perf_cgroup_info *t; |
| 626 | t = per_cpu_ptr(event->cgrp->info, event->cpu); |
| 627 | event->shadow_ctx_time = now - t->timestamp; |
| 628 | } |
| 629 | |
| 630 | static inline void |
| 631 | perf_cgroup_defer_enabled(struct perf_event *event) |
| 632 | { |
| 633 | /* |
| 634 | * when the current task's perf cgroup does not match |
| 635 | * the event's, we need to remember to call the |
| 636 | * perf_mark_enable() function the first time a task with |
| 637 | * a matching perf cgroup is scheduled in. |
| 638 | */ |
| 639 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) |
| 640 | event->cgrp_defer_enabled = 1; |
| 641 | } |
| 642 | |
| 643 | static inline void |
| 644 | perf_cgroup_mark_enabled(struct perf_event *event, |
| 645 | struct perf_event_context *ctx) |
| 646 | { |
| 647 | struct perf_event *sub; |
| 648 | u64 tstamp = perf_event_time(event); |
| 649 | |
| 650 | if (!event->cgrp_defer_enabled) |
| 651 | return; |
| 652 | |
| 653 | event->cgrp_defer_enabled = 0; |
| 654 | |
| 655 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
| 656 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
| 657 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { |
| 658 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; |
| 659 | sub->cgrp_defer_enabled = 0; |
| 660 | } |
| 661 | } |
| 662 | } |
| 663 | #else /* !CONFIG_CGROUP_PERF */ |
| 664 | |
| 665 | static inline bool |
| 666 | perf_cgroup_match(struct perf_event *event) |
| 667 | { |
| 668 | return true; |
| 669 | } |
| 670 | |
| 671 | static inline void perf_detach_cgroup(struct perf_event *event) |
| 672 | {} |
| 673 | |
| 674 | static inline int is_cgroup_event(struct perf_event *event) |
| 675 | { |
| 676 | return 0; |
| 677 | } |
| 678 | |
| 679 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) |
| 680 | { |
| 681 | return 0; |
| 682 | } |
| 683 | |
| 684 | static inline void update_cgrp_time_from_event(struct perf_event *event) |
| 685 | { |
| 686 | } |
| 687 | |
| 688 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) |
| 689 | { |
| 690 | } |
| 691 | |
| 692 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
| 693 | struct task_struct *next) |
| 694 | { |
| 695 | } |
| 696 | |
| 697 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
| 698 | struct task_struct *task) |
| 699 | { |
| 700 | } |
| 701 | |
| 702 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, |
| 703 | struct perf_event_attr *attr, |
| 704 | struct perf_event *group_leader) |
| 705 | { |
| 706 | return -EINVAL; |
| 707 | } |
| 708 | |
| 709 | static inline void |
| 710 | perf_cgroup_set_timestamp(struct task_struct *task, |
| 711 | struct perf_event_context *ctx) |
| 712 | { |
| 713 | } |
| 714 | |
| 715 | void |
| 716 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) |
| 717 | { |
| 718 | } |
| 719 | |
| 720 | static inline void |
| 721 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) |
| 722 | { |
| 723 | } |
| 724 | |
| 725 | static inline u64 perf_cgroup_event_time(struct perf_event *event) |
| 726 | { |
| 727 | return 0; |
| 728 | } |
| 729 | |
| 730 | static inline void |
| 731 | perf_cgroup_defer_enabled(struct perf_event *event) |
| 732 | { |
| 733 | } |
| 734 | |
| 735 | static inline void |
| 736 | perf_cgroup_mark_enabled(struct perf_event *event, |
| 737 | struct perf_event_context *ctx) |
| 738 | { |
| 739 | } |
| 740 | #endif |
| 741 | |
| 742 | /* |
| 743 | * set default to be dependent on timer tick just |
| 744 | * like original code |
| 745 | */ |
| 746 | #define PERF_CPU_HRTIMER (1000 / HZ) |
| 747 | /* |
| 748 | * function must be called with interrupts disbled |
| 749 | */ |
| 750 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) |
| 751 | { |
| 752 | struct perf_cpu_context *cpuctx; |
| 753 | enum hrtimer_restart ret = HRTIMER_NORESTART; |
| 754 | int rotations = 0; |
| 755 | |
| 756 | WARN_ON(!irqs_disabled()); |
| 757 | |
| 758 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); |
| 759 | |
| 760 | rotations = perf_rotate_context(cpuctx); |
| 761 | |
| 762 | /* |
| 763 | * arm timer if needed |
| 764 | */ |
| 765 | if (rotations) { |
| 766 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); |
| 767 | ret = HRTIMER_RESTART; |
| 768 | } |
| 769 | |
| 770 | return ret; |
| 771 | } |
| 772 | |
| 773 | /* CPU is going down */ |
| 774 | void perf_cpu_hrtimer_cancel(int cpu) |
| 775 | { |
| 776 | struct perf_cpu_context *cpuctx; |
| 777 | struct pmu *pmu; |
| 778 | unsigned long flags; |
| 779 | |
| 780 | if (WARN_ON(cpu != smp_processor_id())) |
| 781 | return; |
| 782 | |
| 783 | local_irq_save(flags); |
| 784 | |
| 785 | rcu_read_lock(); |
| 786 | |
| 787 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
| 788 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
| 789 | |
| 790 | if (pmu->task_ctx_nr == perf_sw_context) |
| 791 | continue; |
| 792 | |
| 793 | hrtimer_cancel(&cpuctx->hrtimer); |
| 794 | } |
| 795 | |
| 796 | rcu_read_unlock(); |
| 797 | |
| 798 | local_irq_restore(flags); |
| 799 | } |
| 800 | |
| 801 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) |
| 802 | { |
| 803 | struct hrtimer *hr = &cpuctx->hrtimer; |
| 804 | struct pmu *pmu = cpuctx->ctx.pmu; |
| 805 | int timer; |
| 806 | |
| 807 | /* no multiplexing needed for SW PMU */ |
| 808 | if (pmu->task_ctx_nr == perf_sw_context) |
| 809 | return; |
| 810 | |
| 811 | /* |
| 812 | * check default is sane, if not set then force to |
| 813 | * default interval (1/tick) |
| 814 | */ |
| 815 | timer = pmu->hrtimer_interval_ms; |
| 816 | if (timer < 1) |
| 817 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; |
| 818 | |
| 819 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); |
| 820 | |
| 821 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); |
| 822 | hr->function = perf_cpu_hrtimer_handler; |
| 823 | } |
| 824 | |
| 825 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) |
| 826 | { |
| 827 | struct hrtimer *hr = &cpuctx->hrtimer; |
| 828 | struct pmu *pmu = cpuctx->ctx.pmu; |
| 829 | |
| 830 | /* not for SW PMU */ |
| 831 | if (pmu->task_ctx_nr == perf_sw_context) |
| 832 | return; |
| 833 | |
| 834 | if (hrtimer_active(hr)) |
| 835 | return; |
| 836 | |
| 837 | if (!hrtimer_callback_running(hr)) |
| 838 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, |
| 839 | 0, HRTIMER_MODE_REL_PINNED, 0); |
| 840 | } |
| 841 | |
| 842 | void perf_pmu_disable(struct pmu *pmu) |
| 843 | { |
| 844 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
| 845 | if (!(*count)++) |
| 846 | pmu->pmu_disable(pmu); |
| 847 | } |
| 848 | |
| 849 | void perf_pmu_enable(struct pmu *pmu) |
| 850 | { |
| 851 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
| 852 | if (!--(*count)) |
| 853 | pmu->pmu_enable(pmu); |
| 854 | } |
| 855 | |
| 856 | static DEFINE_PER_CPU(struct list_head, active_ctx_list); |
| 857 | |
| 858 | /* |
| 859 | * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and |
| 860 | * perf_event_task_tick() are fully serialized because they're strictly cpu |
| 861 | * affine and perf_event_ctx{activate,deactivate} are called with IRQs |
| 862 | * disabled, while perf_event_task_tick is called from IRQ context. |
| 863 | */ |
| 864 | static void perf_event_ctx_activate(struct perf_event_context *ctx) |
| 865 | { |
| 866 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
| 867 | |
| 868 | WARN_ON(!irqs_disabled()); |
| 869 | |
| 870 | WARN_ON(!list_empty(&ctx->active_ctx_list)); |
| 871 | |
| 872 | list_add(&ctx->active_ctx_list, head); |
| 873 | } |
| 874 | |
| 875 | static void perf_event_ctx_deactivate(struct perf_event_context *ctx) |
| 876 | { |
| 877 | WARN_ON(!irqs_disabled()); |
| 878 | |
| 879 | WARN_ON(list_empty(&ctx->active_ctx_list)); |
| 880 | |
| 881 | list_del_init(&ctx->active_ctx_list); |
| 882 | } |
| 883 | |
| 884 | static void get_ctx(struct perf_event_context *ctx) |
| 885 | { |
| 886 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
| 887 | } |
| 888 | |
| 889 | static void free_ctx(struct rcu_head *head) |
| 890 | { |
| 891 | struct perf_event_context *ctx; |
| 892 | |
| 893 | ctx = container_of(head, struct perf_event_context, rcu_head); |
| 894 | kfree(ctx->task_ctx_data); |
| 895 | kfree(ctx); |
| 896 | } |
| 897 | |
| 898 | static void put_ctx(struct perf_event_context *ctx) |
| 899 | { |
| 900 | if (atomic_dec_and_test(&ctx->refcount)) { |
| 901 | if (ctx->parent_ctx) |
| 902 | put_ctx(ctx->parent_ctx); |
| 903 | if (ctx->task) |
| 904 | put_task_struct(ctx->task); |
| 905 | call_rcu(&ctx->rcu_head, free_ctx); |
| 906 | } |
| 907 | } |
| 908 | |
| 909 | /* |
| 910 | * Because of perf_event::ctx migration in sys_perf_event_open::move_group and |
| 911 | * perf_pmu_migrate_context() we need some magic. |
| 912 | * |
| 913 | * Those places that change perf_event::ctx will hold both |
| 914 | * perf_event_ctx::mutex of the 'old' and 'new' ctx value. |
| 915 | * |
| 916 | * Lock ordering is by mutex address. There is one other site where |
| 917 | * perf_event_context::mutex nests and that is put_event(). But remember that |
| 918 | * that is a parent<->child context relation, and migration does not affect |
| 919 | * children, therefore these two orderings should not interact. |
| 920 | * |
| 921 | * The change in perf_event::ctx does not affect children (as claimed above) |
| 922 | * because the sys_perf_event_open() case will install a new event and break |
| 923 | * the ctx parent<->child relation, and perf_pmu_migrate_context() is only |
| 924 | * concerned with cpuctx and that doesn't have children. |
| 925 | * |
| 926 | * The places that change perf_event::ctx will issue: |
| 927 | * |
| 928 | * perf_remove_from_context(); |
| 929 | * synchronize_rcu(); |
| 930 | * perf_install_in_context(); |
| 931 | * |
| 932 | * to affect the change. The remove_from_context() + synchronize_rcu() should |
| 933 | * quiesce the event, after which we can install it in the new location. This |
| 934 | * means that only external vectors (perf_fops, prctl) can perturb the event |
| 935 | * while in transit. Therefore all such accessors should also acquire |
| 936 | * perf_event_context::mutex to serialize against this. |
| 937 | * |
| 938 | * However; because event->ctx can change while we're waiting to acquire |
| 939 | * ctx->mutex we must be careful and use the below perf_event_ctx_lock() |
| 940 | * function. |
| 941 | * |
| 942 | * Lock order: |
| 943 | * task_struct::perf_event_mutex |
| 944 | * perf_event_context::mutex |
| 945 | * perf_event_context::lock |
| 946 | * perf_event::child_mutex; |
| 947 | * perf_event::mmap_mutex |
| 948 | * mmap_sem |
| 949 | */ |
| 950 | static struct perf_event_context * |
| 951 | perf_event_ctx_lock_nested(struct perf_event *event, int nesting) |
| 952 | { |
| 953 | struct perf_event_context *ctx; |
| 954 | |
| 955 | again: |
| 956 | rcu_read_lock(); |
| 957 | ctx = ACCESS_ONCE(event->ctx); |
| 958 | if (!atomic_inc_not_zero(&ctx->refcount)) { |
| 959 | rcu_read_unlock(); |
| 960 | goto again; |
| 961 | } |
| 962 | rcu_read_unlock(); |
| 963 | |
| 964 | mutex_lock_nested(&ctx->mutex, nesting); |
| 965 | if (event->ctx != ctx) { |
| 966 | mutex_unlock(&ctx->mutex); |
| 967 | put_ctx(ctx); |
| 968 | goto again; |
| 969 | } |
| 970 | |
| 971 | return ctx; |
| 972 | } |
| 973 | |
| 974 | static inline struct perf_event_context * |
| 975 | perf_event_ctx_lock(struct perf_event *event) |
| 976 | { |
| 977 | return perf_event_ctx_lock_nested(event, 0); |
| 978 | } |
| 979 | |
| 980 | static void perf_event_ctx_unlock(struct perf_event *event, |
| 981 | struct perf_event_context *ctx) |
| 982 | { |
| 983 | mutex_unlock(&ctx->mutex); |
| 984 | put_ctx(ctx); |
| 985 | } |
| 986 | |
| 987 | /* |
| 988 | * This must be done under the ctx->lock, such as to serialize against |
| 989 | * context_equiv(), therefore we cannot call put_ctx() since that might end up |
| 990 | * calling scheduler related locks and ctx->lock nests inside those. |
| 991 | */ |
| 992 | static __must_check struct perf_event_context * |
| 993 | unclone_ctx(struct perf_event_context *ctx) |
| 994 | { |
| 995 | struct perf_event_context *parent_ctx = ctx->parent_ctx; |
| 996 | |
| 997 | lockdep_assert_held(&ctx->lock); |
| 998 | |
| 999 | if (parent_ctx) |
| 1000 | ctx->parent_ctx = NULL; |
| 1001 | ctx->generation++; |
| 1002 | |
| 1003 | return parent_ctx; |
| 1004 | } |
| 1005 | |
| 1006 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
| 1007 | { |
| 1008 | /* |
| 1009 | * only top level events have the pid namespace they were created in |
| 1010 | */ |
| 1011 | if (event->parent) |
| 1012 | event = event->parent; |
| 1013 | |
| 1014 | return task_tgid_nr_ns(p, event->ns); |
| 1015 | } |
| 1016 | |
| 1017 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) |
| 1018 | { |
| 1019 | /* |
| 1020 | * only top level events have the pid namespace they were created in |
| 1021 | */ |
| 1022 | if (event->parent) |
| 1023 | event = event->parent; |
| 1024 | |
| 1025 | return task_pid_nr_ns(p, event->ns); |
| 1026 | } |
| 1027 | |
| 1028 | /* |
| 1029 | * If we inherit events we want to return the parent event id |
| 1030 | * to userspace. |
| 1031 | */ |
| 1032 | static u64 primary_event_id(struct perf_event *event) |
| 1033 | { |
| 1034 | u64 id = event->id; |
| 1035 | |
| 1036 | if (event->parent) |
| 1037 | id = event->parent->id; |
| 1038 | |
| 1039 | return id; |
| 1040 | } |
| 1041 | |
| 1042 | /* |
| 1043 | * Get the perf_event_context for a task and lock it. |
| 1044 | * This has to cope with with the fact that until it is locked, |
| 1045 | * the context could get moved to another task. |
| 1046 | */ |
| 1047 | static struct perf_event_context * |
| 1048 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
| 1049 | { |
| 1050 | struct perf_event_context *ctx; |
| 1051 | |
| 1052 | retry: |
| 1053 | /* |
| 1054 | * One of the few rules of preemptible RCU is that one cannot do |
| 1055 | * rcu_read_unlock() while holding a scheduler (or nested) lock when |
| 1056 | * part of the read side critical section was preemptible -- see |
| 1057 | * rcu_read_unlock_special(). |
| 1058 | * |
| 1059 | * Since ctx->lock nests under rq->lock we must ensure the entire read |
| 1060 | * side critical section is non-preemptible. |
| 1061 | */ |
| 1062 | preempt_disable(); |
| 1063 | rcu_read_lock(); |
| 1064 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
| 1065 | if (ctx) { |
| 1066 | /* |
| 1067 | * If this context is a clone of another, it might |
| 1068 | * get swapped for another underneath us by |
| 1069 | * perf_event_task_sched_out, though the |
| 1070 | * rcu_read_lock() protects us from any context |
| 1071 | * getting freed. Lock the context and check if it |
| 1072 | * got swapped before we could get the lock, and retry |
| 1073 | * if so. If we locked the right context, then it |
| 1074 | * can't get swapped on us any more. |
| 1075 | */ |
| 1076 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
| 1077 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
| 1078 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
| 1079 | rcu_read_unlock(); |
| 1080 | preempt_enable(); |
| 1081 | goto retry; |
| 1082 | } |
| 1083 | |
| 1084 | if (!atomic_inc_not_zero(&ctx->refcount)) { |
| 1085 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
| 1086 | ctx = NULL; |
| 1087 | } |
| 1088 | } |
| 1089 | rcu_read_unlock(); |
| 1090 | preempt_enable(); |
| 1091 | return ctx; |
| 1092 | } |
| 1093 | |
| 1094 | /* |
| 1095 | * Get the context for a task and increment its pin_count so it |
| 1096 | * can't get swapped to another task. This also increments its |
| 1097 | * reference count so that the context can't get freed. |
| 1098 | */ |
| 1099 | static struct perf_event_context * |
| 1100 | perf_pin_task_context(struct task_struct *task, int ctxn) |
| 1101 | { |
| 1102 | struct perf_event_context *ctx; |
| 1103 | unsigned long flags; |
| 1104 | |
| 1105 | ctx = perf_lock_task_context(task, ctxn, &flags); |
| 1106 | if (ctx) { |
| 1107 | ++ctx->pin_count; |
| 1108 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
| 1109 | } |
| 1110 | return ctx; |
| 1111 | } |
| 1112 | |
| 1113 | static void perf_unpin_context(struct perf_event_context *ctx) |
| 1114 | { |
| 1115 | unsigned long flags; |
| 1116 | |
| 1117 | raw_spin_lock_irqsave(&ctx->lock, flags); |
| 1118 | --ctx->pin_count; |
| 1119 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
| 1120 | } |
| 1121 | |
| 1122 | /* |
| 1123 | * Update the record of the current time in a context. |
| 1124 | */ |
| 1125 | static void update_context_time(struct perf_event_context *ctx) |
| 1126 | { |
| 1127 | u64 now = perf_clock(); |
| 1128 | |
| 1129 | ctx->time += now - ctx->timestamp; |
| 1130 | ctx->timestamp = now; |
| 1131 | } |
| 1132 | |
| 1133 | static u64 perf_event_time(struct perf_event *event) |
| 1134 | { |
| 1135 | struct perf_event_context *ctx = event->ctx; |
| 1136 | |
| 1137 | if (is_cgroup_event(event)) |
| 1138 | return perf_cgroup_event_time(event); |
| 1139 | |
| 1140 | return ctx ? ctx->time : 0; |
| 1141 | } |
| 1142 | |
| 1143 | /* |
| 1144 | * Update the total_time_enabled and total_time_running fields for a event. |
| 1145 | * The caller of this function needs to hold the ctx->lock. |
| 1146 | */ |
| 1147 | static void update_event_times(struct perf_event *event) |
| 1148 | { |
| 1149 | struct perf_event_context *ctx = event->ctx; |
| 1150 | u64 run_end; |
| 1151 | |
| 1152 | if (event->state < PERF_EVENT_STATE_INACTIVE || |
| 1153 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) |
| 1154 | return; |
| 1155 | /* |
| 1156 | * in cgroup mode, time_enabled represents |
| 1157 | * the time the event was enabled AND active |
| 1158 | * tasks were in the monitored cgroup. This is |
| 1159 | * independent of the activity of the context as |
| 1160 | * there may be a mix of cgroup and non-cgroup events. |
| 1161 | * |
| 1162 | * That is why we treat cgroup events differently |
| 1163 | * here. |
| 1164 | */ |
| 1165 | if (is_cgroup_event(event)) |
| 1166 | run_end = perf_cgroup_event_time(event); |
| 1167 | else if (ctx->is_active) |
| 1168 | run_end = ctx->time; |
| 1169 | else |
| 1170 | run_end = event->tstamp_stopped; |
| 1171 | |
| 1172 | event->total_time_enabled = run_end - event->tstamp_enabled; |
| 1173 | |
| 1174 | if (event->state == PERF_EVENT_STATE_INACTIVE) |
| 1175 | run_end = event->tstamp_stopped; |
| 1176 | else |
| 1177 | run_end = perf_event_time(event); |
| 1178 | |
| 1179 | event->total_time_running = run_end - event->tstamp_running; |
| 1180 | |
| 1181 | } |
| 1182 | |
| 1183 | /* |
| 1184 | * Update total_time_enabled and total_time_running for all events in a group. |
| 1185 | */ |
| 1186 | static void update_group_times(struct perf_event *leader) |
| 1187 | { |
| 1188 | struct perf_event *event; |
| 1189 | |
| 1190 | update_event_times(leader); |
| 1191 | list_for_each_entry(event, &leader->sibling_list, group_entry) |
| 1192 | update_event_times(event); |
| 1193 | } |
| 1194 | |
| 1195 | static struct list_head * |
| 1196 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) |
| 1197 | { |
| 1198 | if (event->attr.pinned) |
| 1199 | return &ctx->pinned_groups; |
| 1200 | else |
| 1201 | return &ctx->flexible_groups; |
| 1202 | } |
| 1203 | |
| 1204 | /* |
| 1205 | * Add a event from the lists for its context. |
| 1206 | * Must be called with ctx->mutex and ctx->lock held. |
| 1207 | */ |
| 1208 | static void |
| 1209 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
| 1210 | { |
| 1211 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
| 1212 | event->attach_state |= PERF_ATTACH_CONTEXT; |
| 1213 | |
| 1214 | /* |
| 1215 | * If we're a stand alone event or group leader, we go to the context |
| 1216 | * list, group events are kept attached to the group so that |
| 1217 | * perf_group_detach can, at all times, locate all siblings. |
| 1218 | */ |
| 1219 | if (event->group_leader == event) { |
| 1220 | struct list_head *list; |
| 1221 | |
| 1222 | if (is_software_event(event)) |
| 1223 | event->group_flags |= PERF_GROUP_SOFTWARE; |
| 1224 | |
| 1225 | list = ctx_group_list(event, ctx); |
| 1226 | list_add_tail(&event->group_entry, list); |
| 1227 | } |
| 1228 | |
| 1229 | if (is_cgroup_event(event)) |
| 1230 | ctx->nr_cgroups++; |
| 1231 | |
| 1232 | list_add_rcu(&event->event_entry, &ctx->event_list); |
| 1233 | ctx->nr_events++; |
| 1234 | if (event->attr.inherit_stat) |
| 1235 | ctx->nr_stat++; |
| 1236 | |
| 1237 | ctx->generation++; |
| 1238 | } |
| 1239 | |
| 1240 | /* |
| 1241 | * Initialize event state based on the perf_event_attr::disabled. |
| 1242 | */ |
| 1243 | static inline void perf_event__state_init(struct perf_event *event) |
| 1244 | { |
| 1245 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : |
| 1246 | PERF_EVENT_STATE_INACTIVE; |
| 1247 | } |
| 1248 | |
| 1249 | /* |
| 1250 | * Called at perf_event creation and when events are attached/detached from a |
| 1251 | * group. |
| 1252 | */ |
| 1253 | static void perf_event__read_size(struct perf_event *event) |
| 1254 | { |
| 1255 | int entry = sizeof(u64); /* value */ |
| 1256 | int size = 0; |
| 1257 | int nr = 1; |
| 1258 | |
| 1259 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
| 1260 | size += sizeof(u64); |
| 1261 | |
| 1262 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
| 1263 | size += sizeof(u64); |
| 1264 | |
| 1265 | if (event->attr.read_format & PERF_FORMAT_ID) |
| 1266 | entry += sizeof(u64); |
| 1267 | |
| 1268 | if (event->attr.read_format & PERF_FORMAT_GROUP) { |
| 1269 | nr += event->group_leader->nr_siblings; |
| 1270 | size += sizeof(u64); |
| 1271 | } |
| 1272 | |
| 1273 | size += entry * nr; |
| 1274 | event->read_size = size; |
| 1275 | } |
| 1276 | |
| 1277 | static void perf_event__header_size(struct perf_event *event) |
| 1278 | { |
| 1279 | struct perf_sample_data *data; |
| 1280 | u64 sample_type = event->attr.sample_type; |
| 1281 | u16 size = 0; |
| 1282 | |
| 1283 | perf_event__read_size(event); |
| 1284 | |
| 1285 | if (sample_type & PERF_SAMPLE_IP) |
| 1286 | size += sizeof(data->ip); |
| 1287 | |
| 1288 | if (sample_type & PERF_SAMPLE_ADDR) |
| 1289 | size += sizeof(data->addr); |
| 1290 | |
| 1291 | if (sample_type & PERF_SAMPLE_PERIOD) |
| 1292 | size += sizeof(data->period); |
| 1293 | |
| 1294 | if (sample_type & PERF_SAMPLE_WEIGHT) |
| 1295 | size += sizeof(data->weight); |
| 1296 | |
| 1297 | if (sample_type & PERF_SAMPLE_READ) |
| 1298 | size += event->read_size; |
| 1299 | |
| 1300 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
| 1301 | size += sizeof(data->data_src.val); |
| 1302 | |
| 1303 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
| 1304 | size += sizeof(data->txn); |
| 1305 | |
| 1306 | event->header_size = size; |
| 1307 | } |
| 1308 | |
| 1309 | static void perf_event__id_header_size(struct perf_event *event) |
| 1310 | { |
| 1311 | struct perf_sample_data *data; |
| 1312 | u64 sample_type = event->attr.sample_type; |
| 1313 | u16 size = 0; |
| 1314 | |
| 1315 | if (sample_type & PERF_SAMPLE_TID) |
| 1316 | size += sizeof(data->tid_entry); |
| 1317 | |
| 1318 | if (sample_type & PERF_SAMPLE_TIME) |
| 1319 | size += sizeof(data->time); |
| 1320 | |
| 1321 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
| 1322 | size += sizeof(data->id); |
| 1323 | |
| 1324 | if (sample_type & PERF_SAMPLE_ID) |
| 1325 | size += sizeof(data->id); |
| 1326 | |
| 1327 | if (sample_type & PERF_SAMPLE_STREAM_ID) |
| 1328 | size += sizeof(data->stream_id); |
| 1329 | |
| 1330 | if (sample_type & PERF_SAMPLE_CPU) |
| 1331 | size += sizeof(data->cpu_entry); |
| 1332 | |
| 1333 | event->id_header_size = size; |
| 1334 | } |
| 1335 | |
| 1336 | static void perf_group_attach(struct perf_event *event) |
| 1337 | { |
| 1338 | struct perf_event *group_leader = event->group_leader, *pos; |
| 1339 | |
| 1340 | /* |
| 1341 | * We can have double attach due to group movement in perf_event_open. |
| 1342 | */ |
| 1343 | if (event->attach_state & PERF_ATTACH_GROUP) |
| 1344 | return; |
| 1345 | |
| 1346 | event->attach_state |= PERF_ATTACH_GROUP; |
| 1347 | |
| 1348 | if (group_leader == event) |
| 1349 | return; |
| 1350 | |
| 1351 | WARN_ON_ONCE(group_leader->ctx != event->ctx); |
| 1352 | |
| 1353 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && |
| 1354 | !is_software_event(event)) |
| 1355 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; |
| 1356 | |
| 1357 | list_add_tail(&event->group_entry, &group_leader->sibling_list); |
| 1358 | group_leader->nr_siblings++; |
| 1359 | |
| 1360 | perf_event__header_size(group_leader); |
| 1361 | |
| 1362 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) |
| 1363 | perf_event__header_size(pos); |
| 1364 | } |
| 1365 | |
| 1366 | /* |
| 1367 | * Remove a event from the lists for its context. |
| 1368 | * Must be called with ctx->mutex and ctx->lock held. |
| 1369 | */ |
| 1370 | static void |
| 1371 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
| 1372 | { |
| 1373 | struct perf_cpu_context *cpuctx; |
| 1374 | |
| 1375 | WARN_ON_ONCE(event->ctx != ctx); |
| 1376 | lockdep_assert_held(&ctx->lock); |
| 1377 | |
| 1378 | /* |
| 1379 | * We can have double detach due to exit/hot-unplug + close. |
| 1380 | */ |
| 1381 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) |
| 1382 | return; |
| 1383 | |
| 1384 | event->attach_state &= ~PERF_ATTACH_CONTEXT; |
| 1385 | |
| 1386 | if (is_cgroup_event(event)) { |
| 1387 | ctx->nr_cgroups--; |
| 1388 | cpuctx = __get_cpu_context(ctx); |
| 1389 | /* |
| 1390 | * if there are no more cgroup events |
| 1391 | * then cler cgrp to avoid stale pointer |
| 1392 | * in update_cgrp_time_from_cpuctx() |
| 1393 | */ |
| 1394 | if (!ctx->nr_cgroups) |
| 1395 | cpuctx->cgrp = NULL; |
| 1396 | } |
| 1397 | |
| 1398 | ctx->nr_events--; |
| 1399 | if (event->attr.inherit_stat) |
| 1400 | ctx->nr_stat--; |
| 1401 | |
| 1402 | list_del_rcu(&event->event_entry); |
| 1403 | |
| 1404 | if (event->group_leader == event) |
| 1405 | list_del_init(&event->group_entry); |
| 1406 | |
| 1407 | update_group_times(event); |
| 1408 | |
| 1409 | /* |
| 1410 | * If event was in error state, then keep it |
| 1411 | * that way, otherwise bogus counts will be |
| 1412 | * returned on read(). The only way to get out |
| 1413 | * of error state is by explicit re-enabling |
| 1414 | * of the event |
| 1415 | */ |
| 1416 | if (event->state > PERF_EVENT_STATE_OFF) |
| 1417 | event->state = PERF_EVENT_STATE_OFF; |
| 1418 | |
| 1419 | ctx->generation++; |
| 1420 | } |
| 1421 | |
| 1422 | static void perf_group_detach(struct perf_event *event) |
| 1423 | { |
| 1424 | struct perf_event *sibling, *tmp; |
| 1425 | struct list_head *list = NULL; |
| 1426 | |
| 1427 | /* |
| 1428 | * We can have double detach due to exit/hot-unplug + close. |
| 1429 | */ |
| 1430 | if (!(event->attach_state & PERF_ATTACH_GROUP)) |
| 1431 | return; |
| 1432 | |
| 1433 | event->attach_state &= ~PERF_ATTACH_GROUP; |
| 1434 | |
| 1435 | /* |
| 1436 | * If this is a sibling, remove it from its group. |
| 1437 | */ |
| 1438 | if (event->group_leader != event) { |
| 1439 | list_del_init(&event->group_entry); |
| 1440 | event->group_leader->nr_siblings--; |
| 1441 | goto out; |
| 1442 | } |
| 1443 | |
| 1444 | if (!list_empty(&event->group_entry)) |
| 1445 | list = &event->group_entry; |
| 1446 | |
| 1447 | /* |
| 1448 | * If this was a group event with sibling events then |
| 1449 | * upgrade the siblings to singleton events by adding them |
| 1450 | * to whatever list we are on. |
| 1451 | */ |
| 1452 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
| 1453 | if (list) |
| 1454 | list_move_tail(&sibling->group_entry, list); |
| 1455 | sibling->group_leader = sibling; |
| 1456 | |
| 1457 | /* Inherit group flags from the previous leader */ |
| 1458 | sibling->group_flags = event->group_flags; |
| 1459 | |
| 1460 | WARN_ON_ONCE(sibling->ctx != event->ctx); |
| 1461 | } |
| 1462 | |
| 1463 | out: |
| 1464 | perf_event__header_size(event->group_leader); |
| 1465 | |
| 1466 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) |
| 1467 | perf_event__header_size(tmp); |
| 1468 | } |
| 1469 | |
| 1470 | /* |
| 1471 | * User event without the task. |
| 1472 | */ |
| 1473 | static bool is_orphaned_event(struct perf_event *event) |
| 1474 | { |
| 1475 | return event && !is_kernel_event(event) && !event->owner; |
| 1476 | } |
| 1477 | |
| 1478 | /* |
| 1479 | * Event has a parent but parent's task finished and it's |
| 1480 | * alive only because of children holding refference. |
| 1481 | */ |
| 1482 | static bool is_orphaned_child(struct perf_event *event) |
| 1483 | { |
| 1484 | return is_orphaned_event(event->parent); |
| 1485 | } |
| 1486 | |
| 1487 | static void orphans_remove_work(struct work_struct *work); |
| 1488 | |
| 1489 | static void schedule_orphans_remove(struct perf_event_context *ctx) |
| 1490 | { |
| 1491 | if (!ctx->task || ctx->orphans_remove_sched || !perf_wq) |
| 1492 | return; |
| 1493 | |
| 1494 | if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) { |
| 1495 | get_ctx(ctx); |
| 1496 | ctx->orphans_remove_sched = true; |
| 1497 | } |
| 1498 | } |
| 1499 | |
| 1500 | static int __init perf_workqueue_init(void) |
| 1501 | { |
| 1502 | perf_wq = create_singlethread_workqueue("perf"); |
| 1503 | WARN(!perf_wq, "failed to create perf workqueue\n"); |
| 1504 | return perf_wq ? 0 : -1; |
| 1505 | } |
| 1506 | |
| 1507 | core_initcall(perf_workqueue_init); |
| 1508 | |
| 1509 | static inline int |
| 1510 | event_filter_match(struct perf_event *event) |
| 1511 | { |
| 1512 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
| 1513 | && perf_cgroup_match(event); |
| 1514 | } |
| 1515 | |
| 1516 | static void |
| 1517 | event_sched_out(struct perf_event *event, |
| 1518 | struct perf_cpu_context *cpuctx, |
| 1519 | struct perf_event_context *ctx) |
| 1520 | { |
| 1521 | u64 tstamp = perf_event_time(event); |
| 1522 | u64 delta; |
| 1523 | |
| 1524 | WARN_ON_ONCE(event->ctx != ctx); |
| 1525 | lockdep_assert_held(&ctx->lock); |
| 1526 | |
| 1527 | /* |
| 1528 | * An event which could not be activated because of |
| 1529 | * filter mismatch still needs to have its timings |
| 1530 | * maintained, otherwise bogus information is return |
| 1531 | * via read() for time_enabled, time_running: |
| 1532 | */ |
| 1533 | if (event->state == PERF_EVENT_STATE_INACTIVE |
| 1534 | && !event_filter_match(event)) { |
| 1535 | delta = tstamp - event->tstamp_stopped; |
| 1536 | event->tstamp_running += delta; |
| 1537 | event->tstamp_stopped = tstamp; |
| 1538 | } |
| 1539 | |
| 1540 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
| 1541 | return; |
| 1542 | |
| 1543 | perf_pmu_disable(event->pmu); |
| 1544 | |
| 1545 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 1546 | if (event->pending_disable) { |
| 1547 | event->pending_disable = 0; |
| 1548 | event->state = PERF_EVENT_STATE_OFF; |
| 1549 | } |
| 1550 | event->tstamp_stopped = tstamp; |
| 1551 | event->pmu->del(event, 0); |
| 1552 | event->oncpu = -1; |
| 1553 | |
| 1554 | if (!is_software_event(event)) |
| 1555 | cpuctx->active_oncpu--; |
| 1556 | if (!--ctx->nr_active) |
| 1557 | perf_event_ctx_deactivate(ctx); |
| 1558 | if (event->attr.freq && event->attr.sample_freq) |
| 1559 | ctx->nr_freq--; |
| 1560 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
| 1561 | cpuctx->exclusive = 0; |
| 1562 | |
| 1563 | if (is_orphaned_child(event)) |
| 1564 | schedule_orphans_remove(ctx); |
| 1565 | |
| 1566 | perf_pmu_enable(event->pmu); |
| 1567 | } |
| 1568 | |
| 1569 | static void |
| 1570 | group_sched_out(struct perf_event *group_event, |
| 1571 | struct perf_cpu_context *cpuctx, |
| 1572 | struct perf_event_context *ctx) |
| 1573 | { |
| 1574 | struct perf_event *event; |
| 1575 | int state = group_event->state; |
| 1576 | |
| 1577 | event_sched_out(group_event, cpuctx, ctx); |
| 1578 | |
| 1579 | /* |
| 1580 | * Schedule out siblings (if any): |
| 1581 | */ |
| 1582 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
| 1583 | event_sched_out(event, cpuctx, ctx); |
| 1584 | |
| 1585 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
| 1586 | cpuctx->exclusive = 0; |
| 1587 | } |
| 1588 | |
| 1589 | struct remove_event { |
| 1590 | struct perf_event *event; |
| 1591 | bool detach_group; |
| 1592 | }; |
| 1593 | |
| 1594 | /* |
| 1595 | * Cross CPU call to remove a performance event |
| 1596 | * |
| 1597 | * We disable the event on the hardware level first. After that we |
| 1598 | * remove it from the context list. |
| 1599 | */ |
| 1600 | static int __perf_remove_from_context(void *info) |
| 1601 | { |
| 1602 | struct remove_event *re = info; |
| 1603 | struct perf_event *event = re->event; |
| 1604 | struct perf_event_context *ctx = event->ctx; |
| 1605 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
| 1606 | |
| 1607 | raw_spin_lock(&ctx->lock); |
| 1608 | event_sched_out(event, cpuctx, ctx); |
| 1609 | if (re->detach_group) |
| 1610 | perf_group_detach(event); |
| 1611 | list_del_event(event, ctx); |
| 1612 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
| 1613 | ctx->is_active = 0; |
| 1614 | cpuctx->task_ctx = NULL; |
| 1615 | } |
| 1616 | raw_spin_unlock(&ctx->lock); |
| 1617 | |
| 1618 | return 0; |
| 1619 | } |
| 1620 | |
| 1621 | |
| 1622 | /* |
| 1623 | * Remove the event from a task's (or a CPU's) list of events. |
| 1624 | * |
| 1625 | * CPU events are removed with a smp call. For task events we only |
| 1626 | * call when the task is on a CPU. |
| 1627 | * |
| 1628 | * If event->ctx is a cloned context, callers must make sure that |
| 1629 | * every task struct that event->ctx->task could possibly point to |
| 1630 | * remains valid. This is OK when called from perf_release since |
| 1631 | * that only calls us on the top-level context, which can't be a clone. |
| 1632 | * When called from perf_event_exit_task, it's OK because the |
| 1633 | * context has been detached from its task. |
| 1634 | */ |
| 1635 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
| 1636 | { |
| 1637 | struct perf_event_context *ctx = event->ctx; |
| 1638 | struct task_struct *task = ctx->task; |
| 1639 | struct remove_event re = { |
| 1640 | .event = event, |
| 1641 | .detach_group = detach_group, |
| 1642 | }; |
| 1643 | |
| 1644 | lockdep_assert_held(&ctx->mutex); |
| 1645 | |
| 1646 | if (!task) { |
| 1647 | /* |
| 1648 | * Per cpu events are removed via an smp call. The removal can |
| 1649 | * fail if the CPU is currently offline, but in that case we |
| 1650 | * already called __perf_remove_from_context from |
| 1651 | * perf_event_exit_cpu. |
| 1652 | */ |
| 1653 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
| 1654 | return; |
| 1655 | } |
| 1656 | |
| 1657 | retry: |
| 1658 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
| 1659 | return; |
| 1660 | |
| 1661 | raw_spin_lock_irq(&ctx->lock); |
| 1662 | /* |
| 1663 | * If we failed to find a running task, but find the context active now |
| 1664 | * that we've acquired the ctx->lock, retry. |
| 1665 | */ |
| 1666 | if (ctx->is_active) { |
| 1667 | raw_spin_unlock_irq(&ctx->lock); |
| 1668 | /* |
| 1669 | * Reload the task pointer, it might have been changed by |
| 1670 | * a concurrent perf_event_context_sched_out(). |
| 1671 | */ |
| 1672 | task = ctx->task; |
| 1673 | goto retry; |
| 1674 | } |
| 1675 | |
| 1676 | /* |
| 1677 | * Since the task isn't running, its safe to remove the event, us |
| 1678 | * holding the ctx->lock ensures the task won't get scheduled in. |
| 1679 | */ |
| 1680 | if (detach_group) |
| 1681 | perf_group_detach(event); |
| 1682 | list_del_event(event, ctx); |
| 1683 | raw_spin_unlock_irq(&ctx->lock); |
| 1684 | } |
| 1685 | |
| 1686 | /* |
| 1687 | * Cross CPU call to disable a performance event |
| 1688 | */ |
| 1689 | int __perf_event_disable(void *info) |
| 1690 | { |
| 1691 | struct perf_event *event = info; |
| 1692 | struct perf_event_context *ctx = event->ctx; |
| 1693 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
| 1694 | |
| 1695 | /* |
| 1696 | * If this is a per-task event, need to check whether this |
| 1697 | * event's task is the current task on this cpu. |
| 1698 | * |
| 1699 | * Can trigger due to concurrent perf_event_context_sched_out() |
| 1700 | * flipping contexts around. |
| 1701 | */ |
| 1702 | if (ctx->task && cpuctx->task_ctx != ctx) |
| 1703 | return -EINVAL; |
| 1704 | |
| 1705 | raw_spin_lock(&ctx->lock); |
| 1706 | |
| 1707 | /* |
| 1708 | * If the event is on, turn it off. |
| 1709 | * If it is in error state, leave it in error state. |
| 1710 | */ |
| 1711 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
| 1712 | update_context_time(ctx); |
| 1713 | update_cgrp_time_from_event(event); |
| 1714 | update_group_times(event); |
| 1715 | if (event == event->group_leader) |
| 1716 | group_sched_out(event, cpuctx, ctx); |
| 1717 | else |
| 1718 | event_sched_out(event, cpuctx, ctx); |
| 1719 | event->state = PERF_EVENT_STATE_OFF; |
| 1720 | } |
| 1721 | |
| 1722 | raw_spin_unlock(&ctx->lock); |
| 1723 | |
| 1724 | return 0; |
| 1725 | } |
| 1726 | |
| 1727 | /* |
| 1728 | * Disable a event. |
| 1729 | * |
| 1730 | * If event->ctx is a cloned context, callers must make sure that |
| 1731 | * every task struct that event->ctx->task could possibly point to |
| 1732 | * remains valid. This condition is satisifed when called through |
| 1733 | * perf_event_for_each_child or perf_event_for_each because they |
| 1734 | * hold the top-level event's child_mutex, so any descendant that |
| 1735 | * goes to exit will block in sync_child_event. |
| 1736 | * When called from perf_pending_event it's OK because event->ctx |
| 1737 | * is the current context on this CPU and preemption is disabled, |
| 1738 | * hence we can't get into perf_event_task_sched_out for this context. |
| 1739 | */ |
| 1740 | static void _perf_event_disable(struct perf_event *event) |
| 1741 | { |
| 1742 | struct perf_event_context *ctx = event->ctx; |
| 1743 | struct task_struct *task = ctx->task; |
| 1744 | |
| 1745 | if (!task) { |
| 1746 | /* |
| 1747 | * Disable the event on the cpu that it's on |
| 1748 | */ |
| 1749 | cpu_function_call(event->cpu, __perf_event_disable, event); |
| 1750 | return; |
| 1751 | } |
| 1752 | |
| 1753 | retry: |
| 1754 | if (!task_function_call(task, __perf_event_disable, event)) |
| 1755 | return; |
| 1756 | |
| 1757 | raw_spin_lock_irq(&ctx->lock); |
| 1758 | /* |
| 1759 | * If the event is still active, we need to retry the cross-call. |
| 1760 | */ |
| 1761 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
| 1762 | raw_spin_unlock_irq(&ctx->lock); |
| 1763 | /* |
| 1764 | * Reload the task pointer, it might have been changed by |
| 1765 | * a concurrent perf_event_context_sched_out(). |
| 1766 | */ |
| 1767 | task = ctx->task; |
| 1768 | goto retry; |
| 1769 | } |
| 1770 | |
| 1771 | /* |
| 1772 | * Since we have the lock this context can't be scheduled |
| 1773 | * in, so we can change the state safely. |
| 1774 | */ |
| 1775 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
| 1776 | update_group_times(event); |
| 1777 | event->state = PERF_EVENT_STATE_OFF; |
| 1778 | } |
| 1779 | raw_spin_unlock_irq(&ctx->lock); |
| 1780 | } |
| 1781 | |
| 1782 | /* |
| 1783 | * Strictly speaking kernel users cannot create groups and therefore this |
| 1784 | * interface does not need the perf_event_ctx_lock() magic. |
| 1785 | */ |
| 1786 | void perf_event_disable(struct perf_event *event) |
| 1787 | { |
| 1788 | struct perf_event_context *ctx; |
| 1789 | |
| 1790 | ctx = perf_event_ctx_lock(event); |
| 1791 | _perf_event_disable(event); |
| 1792 | perf_event_ctx_unlock(event, ctx); |
| 1793 | } |
| 1794 | EXPORT_SYMBOL_GPL(perf_event_disable); |
| 1795 | |
| 1796 | static void perf_set_shadow_time(struct perf_event *event, |
| 1797 | struct perf_event_context *ctx, |
| 1798 | u64 tstamp) |
| 1799 | { |
| 1800 | /* |
| 1801 | * use the correct time source for the time snapshot |
| 1802 | * |
| 1803 | * We could get by without this by leveraging the |
| 1804 | * fact that to get to this function, the caller |
| 1805 | * has most likely already called update_context_time() |
| 1806 | * and update_cgrp_time_xx() and thus both timestamp |
| 1807 | * are identical (or very close). Given that tstamp is, |
| 1808 | * already adjusted for cgroup, we could say that: |
| 1809 | * tstamp - ctx->timestamp |
| 1810 | * is equivalent to |
| 1811 | * tstamp - cgrp->timestamp. |
| 1812 | * |
| 1813 | * Then, in perf_output_read(), the calculation would |
| 1814 | * work with no changes because: |
| 1815 | * - event is guaranteed scheduled in |
| 1816 | * - no scheduled out in between |
| 1817 | * - thus the timestamp would be the same |
| 1818 | * |
| 1819 | * But this is a bit hairy. |
| 1820 | * |
| 1821 | * So instead, we have an explicit cgroup call to remain |
| 1822 | * within the time time source all along. We believe it |
| 1823 | * is cleaner and simpler to understand. |
| 1824 | */ |
| 1825 | if (is_cgroup_event(event)) |
| 1826 | perf_cgroup_set_shadow_time(event, tstamp); |
| 1827 | else |
| 1828 | event->shadow_ctx_time = tstamp - ctx->timestamp; |
| 1829 | } |
| 1830 | |
| 1831 | #define MAX_INTERRUPTS (~0ULL) |
| 1832 | |
| 1833 | static void perf_log_throttle(struct perf_event *event, int enable); |
| 1834 | static void perf_log_itrace_start(struct perf_event *event); |
| 1835 | |
| 1836 | static int |
| 1837 | event_sched_in(struct perf_event *event, |
| 1838 | struct perf_cpu_context *cpuctx, |
| 1839 | struct perf_event_context *ctx) |
| 1840 | { |
| 1841 | u64 tstamp = perf_event_time(event); |
| 1842 | int ret = 0; |
| 1843 | |
| 1844 | lockdep_assert_held(&ctx->lock); |
| 1845 | |
| 1846 | if (event->state <= PERF_EVENT_STATE_OFF) |
| 1847 | return 0; |
| 1848 | |
| 1849 | event->state = PERF_EVENT_STATE_ACTIVE; |
| 1850 | event->oncpu = smp_processor_id(); |
| 1851 | |
| 1852 | /* |
| 1853 | * Unthrottle events, since we scheduled we might have missed several |
| 1854 | * ticks already, also for a heavily scheduling task there is little |
| 1855 | * guarantee it'll get a tick in a timely manner. |
| 1856 | */ |
| 1857 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { |
| 1858 | perf_log_throttle(event, 1); |
| 1859 | event->hw.interrupts = 0; |
| 1860 | } |
| 1861 | |
| 1862 | /* |
| 1863 | * The new state must be visible before we turn it on in the hardware: |
| 1864 | */ |
| 1865 | smp_wmb(); |
| 1866 | |
| 1867 | perf_pmu_disable(event->pmu); |
| 1868 | |
| 1869 | event->tstamp_running += tstamp - event->tstamp_stopped; |
| 1870 | |
| 1871 | perf_set_shadow_time(event, ctx, tstamp); |
| 1872 | |
| 1873 | perf_log_itrace_start(event); |
| 1874 | |
| 1875 | if (event->pmu->add(event, PERF_EF_START)) { |
| 1876 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 1877 | event->oncpu = -1; |
| 1878 | ret = -EAGAIN; |
| 1879 | goto out; |
| 1880 | } |
| 1881 | |
| 1882 | if (!is_software_event(event)) |
| 1883 | cpuctx->active_oncpu++; |
| 1884 | if (!ctx->nr_active++) |
| 1885 | perf_event_ctx_activate(ctx); |
| 1886 | if (event->attr.freq && event->attr.sample_freq) |
| 1887 | ctx->nr_freq++; |
| 1888 | |
| 1889 | if (event->attr.exclusive) |
| 1890 | cpuctx->exclusive = 1; |
| 1891 | |
| 1892 | if (is_orphaned_child(event)) |
| 1893 | schedule_orphans_remove(ctx); |
| 1894 | |
| 1895 | out: |
| 1896 | perf_pmu_enable(event->pmu); |
| 1897 | |
| 1898 | return ret; |
| 1899 | } |
| 1900 | |
| 1901 | static int |
| 1902 | group_sched_in(struct perf_event *group_event, |
| 1903 | struct perf_cpu_context *cpuctx, |
| 1904 | struct perf_event_context *ctx) |
| 1905 | { |
| 1906 | struct perf_event *event, *partial_group = NULL; |
| 1907 | struct pmu *pmu = ctx->pmu; |
| 1908 | u64 now = ctx->time; |
| 1909 | bool simulate = false; |
| 1910 | |
| 1911 | if (group_event->state == PERF_EVENT_STATE_OFF) |
| 1912 | return 0; |
| 1913 | |
| 1914 | pmu->start_txn(pmu); |
| 1915 | |
| 1916 | if (event_sched_in(group_event, cpuctx, ctx)) { |
| 1917 | pmu->cancel_txn(pmu); |
| 1918 | perf_cpu_hrtimer_restart(cpuctx); |
| 1919 | return -EAGAIN; |
| 1920 | } |
| 1921 | |
| 1922 | /* |
| 1923 | * Schedule in siblings as one group (if any): |
| 1924 | */ |
| 1925 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
| 1926 | if (event_sched_in(event, cpuctx, ctx)) { |
| 1927 | partial_group = event; |
| 1928 | goto group_error; |
| 1929 | } |
| 1930 | } |
| 1931 | |
| 1932 | if (!pmu->commit_txn(pmu)) |
| 1933 | return 0; |
| 1934 | |
| 1935 | group_error: |
| 1936 | /* |
| 1937 | * Groups can be scheduled in as one unit only, so undo any |
| 1938 | * partial group before returning: |
| 1939 | * The events up to the failed event are scheduled out normally, |
| 1940 | * tstamp_stopped will be updated. |
| 1941 | * |
| 1942 | * The failed events and the remaining siblings need to have |
| 1943 | * their timings updated as if they had gone thru event_sched_in() |
| 1944 | * and event_sched_out(). This is required to get consistent timings |
| 1945 | * across the group. This also takes care of the case where the group |
| 1946 | * could never be scheduled by ensuring tstamp_stopped is set to mark |
| 1947 | * the time the event was actually stopped, such that time delta |
| 1948 | * calculation in update_event_times() is correct. |
| 1949 | */ |
| 1950 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
| 1951 | if (event == partial_group) |
| 1952 | simulate = true; |
| 1953 | |
| 1954 | if (simulate) { |
| 1955 | event->tstamp_running += now - event->tstamp_stopped; |
| 1956 | event->tstamp_stopped = now; |
| 1957 | } else { |
| 1958 | event_sched_out(event, cpuctx, ctx); |
| 1959 | } |
| 1960 | } |
| 1961 | event_sched_out(group_event, cpuctx, ctx); |
| 1962 | |
| 1963 | pmu->cancel_txn(pmu); |
| 1964 | |
| 1965 | perf_cpu_hrtimer_restart(cpuctx); |
| 1966 | |
| 1967 | return -EAGAIN; |
| 1968 | } |
| 1969 | |
| 1970 | /* |
| 1971 | * Work out whether we can put this event group on the CPU now. |
| 1972 | */ |
| 1973 | static int group_can_go_on(struct perf_event *event, |
| 1974 | struct perf_cpu_context *cpuctx, |
| 1975 | int can_add_hw) |
| 1976 | { |
| 1977 | /* |
| 1978 | * Groups consisting entirely of software events can always go on. |
| 1979 | */ |
| 1980 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
| 1981 | return 1; |
| 1982 | /* |
| 1983 | * If an exclusive group is already on, no other hardware |
| 1984 | * events can go on. |
| 1985 | */ |
| 1986 | if (cpuctx->exclusive) |
| 1987 | return 0; |
| 1988 | /* |
| 1989 | * If this group is exclusive and there are already |
| 1990 | * events on the CPU, it can't go on. |
| 1991 | */ |
| 1992 | if (event->attr.exclusive && cpuctx->active_oncpu) |
| 1993 | return 0; |
| 1994 | /* |
| 1995 | * Otherwise, try to add it if all previous groups were able |
| 1996 | * to go on. |
| 1997 | */ |
| 1998 | return can_add_hw; |
| 1999 | } |
| 2000 | |
| 2001 | static void add_event_to_ctx(struct perf_event *event, |
| 2002 | struct perf_event_context *ctx) |
| 2003 | { |
| 2004 | u64 tstamp = perf_event_time(event); |
| 2005 | |
| 2006 | list_add_event(event, ctx); |
| 2007 | perf_group_attach(event); |
| 2008 | event->tstamp_enabled = tstamp; |
| 2009 | event->tstamp_running = tstamp; |
| 2010 | event->tstamp_stopped = tstamp; |
| 2011 | } |
| 2012 | |
| 2013 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
| 2014 | static void |
| 2015 | ctx_sched_in(struct perf_event_context *ctx, |
| 2016 | struct perf_cpu_context *cpuctx, |
| 2017 | enum event_type_t event_type, |
| 2018 | struct task_struct *task); |
| 2019 | |
| 2020 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
| 2021 | struct perf_event_context *ctx, |
| 2022 | struct task_struct *task) |
| 2023 | { |
| 2024 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); |
| 2025 | if (ctx) |
| 2026 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); |
| 2027 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); |
| 2028 | if (ctx) |
| 2029 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); |
| 2030 | } |
| 2031 | |
| 2032 | /* |
| 2033 | * Cross CPU call to install and enable a performance event |
| 2034 | * |
| 2035 | * Must be called with ctx->mutex held |
| 2036 | */ |
| 2037 | static int __perf_install_in_context(void *info) |
| 2038 | { |
| 2039 | struct perf_event *event = info; |
| 2040 | struct perf_event_context *ctx = event->ctx; |
| 2041 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
| 2042 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
| 2043 | struct task_struct *task = current; |
| 2044 | |
| 2045 | perf_ctx_lock(cpuctx, task_ctx); |
| 2046 | perf_pmu_disable(cpuctx->ctx.pmu); |
| 2047 | |
| 2048 | /* |
| 2049 | * If there was an active task_ctx schedule it out. |
| 2050 | */ |
| 2051 | if (task_ctx) |
| 2052 | task_ctx_sched_out(task_ctx); |
| 2053 | |
| 2054 | /* |
| 2055 | * If the context we're installing events in is not the |
| 2056 | * active task_ctx, flip them. |
| 2057 | */ |
| 2058 | if (ctx->task && task_ctx != ctx) { |
| 2059 | if (task_ctx) |
| 2060 | raw_spin_unlock(&task_ctx->lock); |
| 2061 | raw_spin_lock(&ctx->lock); |
| 2062 | task_ctx = ctx; |
| 2063 | } |
| 2064 | |
| 2065 | if (task_ctx) { |
| 2066 | cpuctx->task_ctx = task_ctx; |
| 2067 | task = task_ctx->task; |
| 2068 | } |
| 2069 | |
| 2070 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
| 2071 | |
| 2072 | update_context_time(ctx); |
| 2073 | /* |
| 2074 | * update cgrp time only if current cgrp |
| 2075 | * matches event->cgrp. Must be done before |
| 2076 | * calling add_event_to_ctx() |
| 2077 | */ |
| 2078 | update_cgrp_time_from_event(event); |
| 2079 | |
| 2080 | add_event_to_ctx(event, ctx); |
| 2081 | |
| 2082 | /* |
| 2083 | * Schedule everything back in |
| 2084 | */ |
| 2085 | perf_event_sched_in(cpuctx, task_ctx, task); |
| 2086 | |
| 2087 | perf_pmu_enable(cpuctx->ctx.pmu); |
| 2088 | perf_ctx_unlock(cpuctx, task_ctx); |
| 2089 | |
| 2090 | return 0; |
| 2091 | } |
| 2092 | |
| 2093 | /* |
| 2094 | * Attach a performance event to a context |
| 2095 | * |
| 2096 | * First we add the event to the list with the hardware enable bit |
| 2097 | * in event->hw_config cleared. |
| 2098 | * |
| 2099 | * If the event is attached to a task which is on a CPU we use a smp |
| 2100 | * call to enable it in the task context. The task might have been |
| 2101 | * scheduled away, but we check this in the smp call again. |
| 2102 | */ |
| 2103 | static void |
| 2104 | perf_install_in_context(struct perf_event_context *ctx, |
| 2105 | struct perf_event *event, |
| 2106 | int cpu) |
| 2107 | { |
| 2108 | struct task_struct *task = ctx->task; |
| 2109 | |
| 2110 | lockdep_assert_held(&ctx->mutex); |
| 2111 | |
| 2112 | event->ctx = ctx; |
| 2113 | if (event->cpu != -1) |
| 2114 | event->cpu = cpu; |
| 2115 | |
| 2116 | if (!task) { |
| 2117 | /* |
| 2118 | * Per cpu events are installed via an smp call and |
| 2119 | * the install is always successful. |
| 2120 | */ |
| 2121 | cpu_function_call(cpu, __perf_install_in_context, event); |
| 2122 | return; |
| 2123 | } |
| 2124 | |
| 2125 | retry: |
| 2126 | if (!task_function_call(task, __perf_install_in_context, event)) |
| 2127 | return; |
| 2128 | |
| 2129 | raw_spin_lock_irq(&ctx->lock); |
| 2130 | /* |
| 2131 | * If we failed to find a running task, but find the context active now |
| 2132 | * that we've acquired the ctx->lock, retry. |
| 2133 | */ |
| 2134 | if (ctx->is_active) { |
| 2135 | raw_spin_unlock_irq(&ctx->lock); |
| 2136 | /* |
| 2137 | * Reload the task pointer, it might have been changed by |
| 2138 | * a concurrent perf_event_context_sched_out(). |
| 2139 | */ |
| 2140 | task = ctx->task; |
| 2141 | goto retry; |
| 2142 | } |
| 2143 | |
| 2144 | /* |
| 2145 | * Since the task isn't running, its safe to add the event, us holding |
| 2146 | * the ctx->lock ensures the task won't get scheduled in. |
| 2147 | */ |
| 2148 | add_event_to_ctx(event, ctx); |
| 2149 | raw_spin_unlock_irq(&ctx->lock); |
| 2150 | } |
| 2151 | |
| 2152 | /* |
| 2153 | * Put a event into inactive state and update time fields. |
| 2154 | * Enabling the leader of a group effectively enables all |
| 2155 | * the group members that aren't explicitly disabled, so we |
| 2156 | * have to update their ->tstamp_enabled also. |
| 2157 | * Note: this works for group members as well as group leaders |
| 2158 | * since the non-leader members' sibling_lists will be empty. |
| 2159 | */ |
| 2160 | static void __perf_event_mark_enabled(struct perf_event *event) |
| 2161 | { |
| 2162 | struct perf_event *sub; |
| 2163 | u64 tstamp = perf_event_time(event); |
| 2164 | |
| 2165 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 2166 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
| 2167 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
| 2168 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
| 2169 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; |
| 2170 | } |
| 2171 | } |
| 2172 | |
| 2173 | /* |
| 2174 | * Cross CPU call to enable a performance event |
| 2175 | */ |
| 2176 | static int __perf_event_enable(void *info) |
| 2177 | { |
| 2178 | struct perf_event *event = info; |
| 2179 | struct perf_event_context *ctx = event->ctx; |
| 2180 | struct perf_event *leader = event->group_leader; |
| 2181 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
| 2182 | int err; |
| 2183 | |
| 2184 | /* |
| 2185 | * There's a time window between 'ctx->is_active' check |
| 2186 | * in perf_event_enable function and this place having: |
| 2187 | * - IRQs on |
| 2188 | * - ctx->lock unlocked |
| 2189 | * |
| 2190 | * where the task could be killed and 'ctx' deactivated |
| 2191 | * by perf_event_exit_task. |
| 2192 | */ |
| 2193 | if (!ctx->is_active) |
| 2194 | return -EINVAL; |
| 2195 | |
| 2196 | raw_spin_lock(&ctx->lock); |
| 2197 | update_context_time(ctx); |
| 2198 | |
| 2199 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
| 2200 | goto unlock; |
| 2201 | |
| 2202 | /* |
| 2203 | * set current task's cgroup time reference point |
| 2204 | */ |
| 2205 | perf_cgroup_set_timestamp(current, ctx); |
| 2206 | |
| 2207 | __perf_event_mark_enabled(event); |
| 2208 | |
| 2209 | if (!event_filter_match(event)) { |
| 2210 | if (is_cgroup_event(event)) |
| 2211 | perf_cgroup_defer_enabled(event); |
| 2212 | goto unlock; |
| 2213 | } |
| 2214 | |
| 2215 | /* |
| 2216 | * If the event is in a group and isn't the group leader, |
| 2217 | * then don't put it on unless the group is on. |
| 2218 | */ |
| 2219 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
| 2220 | goto unlock; |
| 2221 | |
| 2222 | if (!group_can_go_on(event, cpuctx, 1)) { |
| 2223 | err = -EEXIST; |
| 2224 | } else { |
| 2225 | if (event == leader) |
| 2226 | err = group_sched_in(event, cpuctx, ctx); |
| 2227 | else |
| 2228 | err = event_sched_in(event, cpuctx, ctx); |
| 2229 | } |
| 2230 | |
| 2231 | if (err) { |
| 2232 | /* |
| 2233 | * If this event can't go on and it's part of a |
| 2234 | * group, then the whole group has to come off. |
| 2235 | */ |
| 2236 | if (leader != event) { |
| 2237 | group_sched_out(leader, cpuctx, ctx); |
| 2238 | perf_cpu_hrtimer_restart(cpuctx); |
| 2239 | } |
| 2240 | if (leader->attr.pinned) { |
| 2241 | update_group_times(leader); |
| 2242 | leader->state = PERF_EVENT_STATE_ERROR; |
| 2243 | } |
| 2244 | } |
| 2245 | |
| 2246 | unlock: |
| 2247 | raw_spin_unlock(&ctx->lock); |
| 2248 | |
| 2249 | return 0; |
| 2250 | } |
| 2251 | |
| 2252 | /* |
| 2253 | * Enable a event. |
| 2254 | * |
| 2255 | * If event->ctx is a cloned context, callers must make sure that |
| 2256 | * every task struct that event->ctx->task could possibly point to |
| 2257 | * remains valid. This condition is satisfied when called through |
| 2258 | * perf_event_for_each_child or perf_event_for_each as described |
| 2259 | * for perf_event_disable. |
| 2260 | */ |
| 2261 | static void _perf_event_enable(struct perf_event *event) |
| 2262 | { |
| 2263 | struct perf_event_context *ctx = event->ctx; |
| 2264 | struct task_struct *task = ctx->task; |
| 2265 | |
| 2266 | if (!task) { |
| 2267 | /* |
| 2268 | * Enable the event on the cpu that it's on |
| 2269 | */ |
| 2270 | cpu_function_call(event->cpu, __perf_event_enable, event); |
| 2271 | return; |
| 2272 | } |
| 2273 | |
| 2274 | raw_spin_lock_irq(&ctx->lock); |
| 2275 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
| 2276 | goto out; |
| 2277 | |
| 2278 | /* |
| 2279 | * If the event is in error state, clear that first. |
| 2280 | * That way, if we see the event in error state below, we |
| 2281 | * know that it has gone back into error state, as distinct |
| 2282 | * from the task having been scheduled away before the |
| 2283 | * cross-call arrived. |
| 2284 | */ |
| 2285 | if (event->state == PERF_EVENT_STATE_ERROR) |
| 2286 | event->state = PERF_EVENT_STATE_OFF; |
| 2287 | |
| 2288 | retry: |
| 2289 | if (!ctx->is_active) { |
| 2290 | __perf_event_mark_enabled(event); |
| 2291 | goto out; |
| 2292 | } |
| 2293 | |
| 2294 | raw_spin_unlock_irq(&ctx->lock); |
| 2295 | |
| 2296 | if (!task_function_call(task, __perf_event_enable, event)) |
| 2297 | return; |
| 2298 | |
| 2299 | raw_spin_lock_irq(&ctx->lock); |
| 2300 | |
| 2301 | /* |
| 2302 | * If the context is active and the event is still off, |
| 2303 | * we need to retry the cross-call. |
| 2304 | */ |
| 2305 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
| 2306 | /* |
| 2307 | * task could have been flipped by a concurrent |
| 2308 | * perf_event_context_sched_out() |
| 2309 | */ |
| 2310 | task = ctx->task; |
| 2311 | goto retry; |
| 2312 | } |
| 2313 | |
| 2314 | out: |
| 2315 | raw_spin_unlock_irq(&ctx->lock); |
| 2316 | } |
| 2317 | |
| 2318 | /* |
| 2319 | * See perf_event_disable(); |
| 2320 | */ |
| 2321 | void perf_event_enable(struct perf_event *event) |
| 2322 | { |
| 2323 | struct perf_event_context *ctx; |
| 2324 | |
| 2325 | ctx = perf_event_ctx_lock(event); |
| 2326 | _perf_event_enable(event); |
| 2327 | perf_event_ctx_unlock(event, ctx); |
| 2328 | } |
| 2329 | EXPORT_SYMBOL_GPL(perf_event_enable); |
| 2330 | |
| 2331 | static int _perf_event_refresh(struct perf_event *event, int refresh) |
| 2332 | { |
| 2333 | /* |
| 2334 | * not supported on inherited events |
| 2335 | */ |
| 2336 | if (event->attr.inherit || !is_sampling_event(event)) |
| 2337 | return -EINVAL; |
| 2338 | |
| 2339 | atomic_add(refresh, &event->event_limit); |
| 2340 | _perf_event_enable(event); |
| 2341 | |
| 2342 | return 0; |
| 2343 | } |
| 2344 | |
| 2345 | /* |
| 2346 | * See perf_event_disable() |
| 2347 | */ |
| 2348 | int perf_event_refresh(struct perf_event *event, int refresh) |
| 2349 | { |
| 2350 | struct perf_event_context *ctx; |
| 2351 | int ret; |
| 2352 | |
| 2353 | ctx = perf_event_ctx_lock(event); |
| 2354 | ret = _perf_event_refresh(event, refresh); |
| 2355 | perf_event_ctx_unlock(event, ctx); |
| 2356 | |
| 2357 | return ret; |
| 2358 | } |
| 2359 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
| 2360 | |
| 2361 | static void ctx_sched_out(struct perf_event_context *ctx, |
| 2362 | struct perf_cpu_context *cpuctx, |
| 2363 | enum event_type_t event_type) |
| 2364 | { |
| 2365 | struct perf_event *event; |
| 2366 | int is_active = ctx->is_active; |
| 2367 | |
| 2368 | ctx->is_active &= ~event_type; |
| 2369 | if (likely(!ctx->nr_events)) |
| 2370 | return; |
| 2371 | |
| 2372 | update_context_time(ctx); |
| 2373 | update_cgrp_time_from_cpuctx(cpuctx); |
| 2374 | if (!ctx->nr_active) |
| 2375 | return; |
| 2376 | |
| 2377 | perf_pmu_disable(ctx->pmu); |
| 2378 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
| 2379 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
| 2380 | group_sched_out(event, cpuctx, ctx); |
| 2381 | } |
| 2382 | |
| 2383 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
| 2384 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
| 2385 | group_sched_out(event, cpuctx, ctx); |
| 2386 | } |
| 2387 | perf_pmu_enable(ctx->pmu); |
| 2388 | } |
| 2389 | |
| 2390 | /* |
| 2391 | * Test whether two contexts are equivalent, i.e. whether they have both been |
| 2392 | * cloned from the same version of the same context. |
| 2393 | * |
| 2394 | * Equivalence is measured using a generation number in the context that is |
| 2395 | * incremented on each modification to it; see unclone_ctx(), list_add_event() |
| 2396 | * and list_del_event(). |
| 2397 | */ |
| 2398 | static int context_equiv(struct perf_event_context *ctx1, |
| 2399 | struct perf_event_context *ctx2) |
| 2400 | { |
| 2401 | lockdep_assert_held(&ctx1->lock); |
| 2402 | lockdep_assert_held(&ctx2->lock); |
| 2403 | |
| 2404 | /* Pinning disables the swap optimization */ |
| 2405 | if (ctx1->pin_count || ctx2->pin_count) |
| 2406 | return 0; |
| 2407 | |
| 2408 | /* If ctx1 is the parent of ctx2 */ |
| 2409 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) |
| 2410 | return 1; |
| 2411 | |
| 2412 | /* If ctx2 is the parent of ctx1 */ |
| 2413 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) |
| 2414 | return 1; |
| 2415 | |
| 2416 | /* |
| 2417 | * If ctx1 and ctx2 have the same parent; we flatten the parent |
| 2418 | * hierarchy, see perf_event_init_context(). |
| 2419 | */ |
| 2420 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && |
| 2421 | ctx1->parent_gen == ctx2->parent_gen) |
| 2422 | return 1; |
| 2423 | |
| 2424 | /* Unmatched */ |
| 2425 | return 0; |
| 2426 | } |
| 2427 | |
| 2428 | static void __perf_event_sync_stat(struct perf_event *event, |
| 2429 | struct perf_event *next_event) |
| 2430 | { |
| 2431 | u64 value; |
| 2432 | |
| 2433 | if (!event->attr.inherit_stat) |
| 2434 | return; |
| 2435 | |
| 2436 | /* |
| 2437 | * Update the event value, we cannot use perf_event_read() |
| 2438 | * because we're in the middle of a context switch and have IRQs |
| 2439 | * disabled, which upsets smp_call_function_single(), however |
| 2440 | * we know the event must be on the current CPU, therefore we |
| 2441 | * don't need to use it. |
| 2442 | */ |
| 2443 | switch (event->state) { |
| 2444 | case PERF_EVENT_STATE_ACTIVE: |
| 2445 | event->pmu->read(event); |
| 2446 | /* fall-through */ |
| 2447 | |
| 2448 | case PERF_EVENT_STATE_INACTIVE: |
| 2449 | update_event_times(event); |
| 2450 | break; |
| 2451 | |
| 2452 | default: |
| 2453 | break; |
| 2454 | } |
| 2455 | |
| 2456 | /* |
| 2457 | * In order to keep per-task stats reliable we need to flip the event |
| 2458 | * values when we flip the contexts. |
| 2459 | */ |
| 2460 | value = local64_read(&next_event->count); |
| 2461 | value = local64_xchg(&event->count, value); |
| 2462 | local64_set(&next_event->count, value); |
| 2463 | |
| 2464 | swap(event->total_time_enabled, next_event->total_time_enabled); |
| 2465 | swap(event->total_time_running, next_event->total_time_running); |
| 2466 | |
| 2467 | /* |
| 2468 | * Since we swizzled the values, update the user visible data too. |
| 2469 | */ |
| 2470 | perf_event_update_userpage(event); |
| 2471 | perf_event_update_userpage(next_event); |
| 2472 | } |
| 2473 | |
| 2474 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
| 2475 | struct perf_event_context *next_ctx) |
| 2476 | { |
| 2477 | struct perf_event *event, *next_event; |
| 2478 | |
| 2479 | if (!ctx->nr_stat) |
| 2480 | return; |
| 2481 | |
| 2482 | update_context_time(ctx); |
| 2483 | |
| 2484 | event = list_first_entry(&ctx->event_list, |
| 2485 | struct perf_event, event_entry); |
| 2486 | |
| 2487 | next_event = list_first_entry(&next_ctx->event_list, |
| 2488 | struct perf_event, event_entry); |
| 2489 | |
| 2490 | while (&event->event_entry != &ctx->event_list && |
| 2491 | &next_event->event_entry != &next_ctx->event_list) { |
| 2492 | |
| 2493 | __perf_event_sync_stat(event, next_event); |
| 2494 | |
| 2495 | event = list_next_entry(event, event_entry); |
| 2496 | next_event = list_next_entry(next_event, event_entry); |
| 2497 | } |
| 2498 | } |
| 2499 | |
| 2500 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
| 2501 | struct task_struct *next) |
| 2502 | { |
| 2503 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
| 2504 | struct perf_event_context *next_ctx; |
| 2505 | struct perf_event_context *parent, *next_parent; |
| 2506 | struct perf_cpu_context *cpuctx; |
| 2507 | int do_switch = 1; |
| 2508 | |
| 2509 | if (likely(!ctx)) |
| 2510 | return; |
| 2511 | |
| 2512 | cpuctx = __get_cpu_context(ctx); |
| 2513 | if (!cpuctx->task_ctx) |
| 2514 | return; |
| 2515 | |
| 2516 | rcu_read_lock(); |
| 2517 | next_ctx = next->perf_event_ctxp[ctxn]; |
| 2518 | if (!next_ctx) |
| 2519 | goto unlock; |
| 2520 | |
| 2521 | parent = rcu_dereference(ctx->parent_ctx); |
| 2522 | next_parent = rcu_dereference(next_ctx->parent_ctx); |
| 2523 | |
| 2524 | /* If neither context have a parent context; they cannot be clones. */ |
| 2525 | if (!parent && !next_parent) |
| 2526 | goto unlock; |
| 2527 | |
| 2528 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { |
| 2529 | /* |
| 2530 | * Looks like the two contexts are clones, so we might be |
| 2531 | * able to optimize the context switch. We lock both |
| 2532 | * contexts and check that they are clones under the |
| 2533 | * lock (including re-checking that neither has been |
| 2534 | * uncloned in the meantime). It doesn't matter which |
| 2535 | * order we take the locks because no other cpu could |
| 2536 | * be trying to lock both of these tasks. |
| 2537 | */ |
| 2538 | raw_spin_lock(&ctx->lock); |
| 2539 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); |
| 2540 | if (context_equiv(ctx, next_ctx)) { |
| 2541 | /* |
| 2542 | * XXX do we need a memory barrier of sorts |
| 2543 | * wrt to rcu_dereference() of perf_event_ctxp |
| 2544 | */ |
| 2545 | task->perf_event_ctxp[ctxn] = next_ctx; |
| 2546 | next->perf_event_ctxp[ctxn] = ctx; |
| 2547 | ctx->task = next; |
| 2548 | next_ctx->task = task; |
| 2549 | |
| 2550 | swap(ctx->task_ctx_data, next_ctx->task_ctx_data); |
| 2551 | |
| 2552 | do_switch = 0; |
| 2553 | |
| 2554 | perf_event_sync_stat(ctx, next_ctx); |
| 2555 | } |
| 2556 | raw_spin_unlock(&next_ctx->lock); |
| 2557 | raw_spin_unlock(&ctx->lock); |
| 2558 | } |
| 2559 | unlock: |
| 2560 | rcu_read_unlock(); |
| 2561 | |
| 2562 | if (do_switch) { |
| 2563 | raw_spin_lock(&ctx->lock); |
| 2564 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
| 2565 | cpuctx->task_ctx = NULL; |
| 2566 | raw_spin_unlock(&ctx->lock); |
| 2567 | } |
| 2568 | } |
| 2569 | |
| 2570 | void perf_sched_cb_dec(struct pmu *pmu) |
| 2571 | { |
| 2572 | this_cpu_dec(perf_sched_cb_usages); |
| 2573 | } |
| 2574 | |
| 2575 | void perf_sched_cb_inc(struct pmu *pmu) |
| 2576 | { |
| 2577 | this_cpu_inc(perf_sched_cb_usages); |
| 2578 | } |
| 2579 | |
| 2580 | /* |
| 2581 | * This function provides the context switch callback to the lower code |
| 2582 | * layer. It is invoked ONLY when the context switch callback is enabled. |
| 2583 | */ |
| 2584 | static void perf_pmu_sched_task(struct task_struct *prev, |
| 2585 | struct task_struct *next, |
| 2586 | bool sched_in) |
| 2587 | { |
| 2588 | struct perf_cpu_context *cpuctx; |
| 2589 | struct pmu *pmu; |
| 2590 | unsigned long flags; |
| 2591 | |
| 2592 | if (prev == next) |
| 2593 | return; |
| 2594 | |
| 2595 | local_irq_save(flags); |
| 2596 | |
| 2597 | rcu_read_lock(); |
| 2598 | |
| 2599 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
| 2600 | if (pmu->sched_task) { |
| 2601 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
| 2602 | |
| 2603 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
| 2604 | |
| 2605 | perf_pmu_disable(pmu); |
| 2606 | |
| 2607 | pmu->sched_task(cpuctx->task_ctx, sched_in); |
| 2608 | |
| 2609 | perf_pmu_enable(pmu); |
| 2610 | |
| 2611 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); |
| 2612 | } |
| 2613 | } |
| 2614 | |
| 2615 | rcu_read_unlock(); |
| 2616 | |
| 2617 | local_irq_restore(flags); |
| 2618 | } |
| 2619 | |
| 2620 | #define for_each_task_context_nr(ctxn) \ |
| 2621 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) |
| 2622 | |
| 2623 | /* |
| 2624 | * Called from scheduler to remove the events of the current task, |
| 2625 | * with interrupts disabled. |
| 2626 | * |
| 2627 | * We stop each event and update the event value in event->count. |
| 2628 | * |
| 2629 | * This does not protect us against NMI, but disable() |
| 2630 | * sets the disabled bit in the control field of event _before_ |
| 2631 | * accessing the event control register. If a NMI hits, then it will |
| 2632 | * not restart the event. |
| 2633 | */ |
| 2634 | void __perf_event_task_sched_out(struct task_struct *task, |
| 2635 | struct task_struct *next) |
| 2636 | { |
| 2637 | int ctxn; |
| 2638 | |
| 2639 | if (__this_cpu_read(perf_sched_cb_usages)) |
| 2640 | perf_pmu_sched_task(task, next, false); |
| 2641 | |
| 2642 | for_each_task_context_nr(ctxn) |
| 2643 | perf_event_context_sched_out(task, ctxn, next); |
| 2644 | |
| 2645 | /* |
| 2646 | * if cgroup events exist on this CPU, then we need |
| 2647 | * to check if we have to switch out PMU state. |
| 2648 | * cgroup event are system-wide mode only |
| 2649 | */ |
| 2650 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
| 2651 | perf_cgroup_sched_out(task, next); |
| 2652 | } |
| 2653 | |
| 2654 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
| 2655 | { |
| 2656 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
| 2657 | |
| 2658 | if (!cpuctx->task_ctx) |
| 2659 | return; |
| 2660 | |
| 2661 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) |
| 2662 | return; |
| 2663 | |
| 2664 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
| 2665 | cpuctx->task_ctx = NULL; |
| 2666 | } |
| 2667 | |
| 2668 | /* |
| 2669 | * Called with IRQs disabled |
| 2670 | */ |
| 2671 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
| 2672 | enum event_type_t event_type) |
| 2673 | { |
| 2674 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); |
| 2675 | } |
| 2676 | |
| 2677 | static void |
| 2678 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
| 2679 | struct perf_cpu_context *cpuctx) |
| 2680 | { |
| 2681 | struct perf_event *event; |
| 2682 | |
| 2683 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
| 2684 | if (event->state <= PERF_EVENT_STATE_OFF) |
| 2685 | continue; |
| 2686 | if (!event_filter_match(event)) |
| 2687 | continue; |
| 2688 | |
| 2689 | /* may need to reset tstamp_enabled */ |
| 2690 | if (is_cgroup_event(event)) |
| 2691 | perf_cgroup_mark_enabled(event, ctx); |
| 2692 | |
| 2693 | if (group_can_go_on(event, cpuctx, 1)) |
| 2694 | group_sched_in(event, cpuctx, ctx); |
| 2695 | |
| 2696 | /* |
| 2697 | * If this pinned group hasn't been scheduled, |
| 2698 | * put it in error state. |
| 2699 | */ |
| 2700 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
| 2701 | update_group_times(event); |
| 2702 | event->state = PERF_EVENT_STATE_ERROR; |
| 2703 | } |
| 2704 | } |
| 2705 | } |
| 2706 | |
| 2707 | static void |
| 2708 | ctx_flexible_sched_in(struct perf_event_context *ctx, |
| 2709 | struct perf_cpu_context *cpuctx) |
| 2710 | { |
| 2711 | struct perf_event *event; |
| 2712 | int can_add_hw = 1; |
| 2713 | |
| 2714 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
| 2715 | /* Ignore events in OFF or ERROR state */ |
| 2716 | if (event->state <= PERF_EVENT_STATE_OFF) |
| 2717 | continue; |
| 2718 | /* |
| 2719 | * Listen to the 'cpu' scheduling filter constraint |
| 2720 | * of events: |
| 2721 | */ |
| 2722 | if (!event_filter_match(event)) |
| 2723 | continue; |
| 2724 | |
| 2725 | /* may need to reset tstamp_enabled */ |
| 2726 | if (is_cgroup_event(event)) |
| 2727 | perf_cgroup_mark_enabled(event, ctx); |
| 2728 | |
| 2729 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
| 2730 | if (group_sched_in(event, cpuctx, ctx)) |
| 2731 | can_add_hw = 0; |
| 2732 | } |
| 2733 | } |
| 2734 | } |
| 2735 | |
| 2736 | static void |
| 2737 | ctx_sched_in(struct perf_event_context *ctx, |
| 2738 | struct perf_cpu_context *cpuctx, |
| 2739 | enum event_type_t event_type, |
| 2740 | struct task_struct *task) |
| 2741 | { |
| 2742 | u64 now; |
| 2743 | int is_active = ctx->is_active; |
| 2744 | |
| 2745 | ctx->is_active |= event_type; |
| 2746 | if (likely(!ctx->nr_events)) |
| 2747 | return; |
| 2748 | |
| 2749 | now = perf_clock(); |
| 2750 | ctx->timestamp = now; |
| 2751 | perf_cgroup_set_timestamp(task, ctx); |
| 2752 | /* |
| 2753 | * First go through the list and put on any pinned groups |
| 2754 | * in order to give them the best chance of going on. |
| 2755 | */ |
| 2756 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
| 2757 | ctx_pinned_sched_in(ctx, cpuctx); |
| 2758 | |
| 2759 | /* Then walk through the lower prio flexible groups */ |
| 2760 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
| 2761 | ctx_flexible_sched_in(ctx, cpuctx); |
| 2762 | } |
| 2763 | |
| 2764 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
| 2765 | enum event_type_t event_type, |
| 2766 | struct task_struct *task) |
| 2767 | { |
| 2768 | struct perf_event_context *ctx = &cpuctx->ctx; |
| 2769 | |
| 2770 | ctx_sched_in(ctx, cpuctx, event_type, task); |
| 2771 | } |
| 2772 | |
| 2773 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
| 2774 | struct task_struct *task) |
| 2775 | { |
| 2776 | struct perf_cpu_context *cpuctx; |
| 2777 | |
| 2778 | cpuctx = __get_cpu_context(ctx); |
| 2779 | if (cpuctx->task_ctx == ctx) |
| 2780 | return; |
| 2781 | |
| 2782 | perf_ctx_lock(cpuctx, ctx); |
| 2783 | perf_pmu_disable(ctx->pmu); |
| 2784 | /* |
| 2785 | * We want to keep the following priority order: |
| 2786 | * cpu pinned (that don't need to move), task pinned, |
| 2787 | * cpu flexible, task flexible. |
| 2788 | */ |
| 2789 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
| 2790 | |
| 2791 | if (ctx->nr_events) |
| 2792 | cpuctx->task_ctx = ctx; |
| 2793 | |
| 2794 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
| 2795 | |
| 2796 | perf_pmu_enable(ctx->pmu); |
| 2797 | perf_ctx_unlock(cpuctx, ctx); |
| 2798 | } |
| 2799 | |
| 2800 | /* |
| 2801 | * Called from scheduler to add the events of the current task |
| 2802 | * with interrupts disabled. |
| 2803 | * |
| 2804 | * We restore the event value and then enable it. |
| 2805 | * |
| 2806 | * This does not protect us against NMI, but enable() |
| 2807 | * sets the enabled bit in the control field of event _before_ |
| 2808 | * accessing the event control register. If a NMI hits, then it will |
| 2809 | * keep the event running. |
| 2810 | */ |
| 2811 | void __perf_event_task_sched_in(struct task_struct *prev, |
| 2812 | struct task_struct *task) |
| 2813 | { |
| 2814 | struct perf_event_context *ctx; |
| 2815 | int ctxn; |
| 2816 | |
| 2817 | for_each_task_context_nr(ctxn) { |
| 2818 | ctx = task->perf_event_ctxp[ctxn]; |
| 2819 | if (likely(!ctx)) |
| 2820 | continue; |
| 2821 | |
| 2822 | perf_event_context_sched_in(ctx, task); |
| 2823 | } |
| 2824 | /* |
| 2825 | * if cgroup events exist on this CPU, then we need |
| 2826 | * to check if we have to switch in PMU state. |
| 2827 | * cgroup event are system-wide mode only |
| 2828 | */ |
| 2829 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
| 2830 | perf_cgroup_sched_in(prev, task); |
| 2831 | |
| 2832 | if (__this_cpu_read(perf_sched_cb_usages)) |
| 2833 | perf_pmu_sched_task(prev, task, true); |
| 2834 | } |
| 2835 | |
| 2836 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
| 2837 | { |
| 2838 | u64 frequency = event->attr.sample_freq; |
| 2839 | u64 sec = NSEC_PER_SEC; |
| 2840 | u64 divisor, dividend; |
| 2841 | |
| 2842 | int count_fls, nsec_fls, frequency_fls, sec_fls; |
| 2843 | |
| 2844 | count_fls = fls64(count); |
| 2845 | nsec_fls = fls64(nsec); |
| 2846 | frequency_fls = fls64(frequency); |
| 2847 | sec_fls = 30; |
| 2848 | |
| 2849 | /* |
| 2850 | * We got @count in @nsec, with a target of sample_freq HZ |
| 2851 | * the target period becomes: |
| 2852 | * |
| 2853 | * @count * 10^9 |
| 2854 | * period = ------------------- |
| 2855 | * @nsec * sample_freq |
| 2856 | * |
| 2857 | */ |
| 2858 | |
| 2859 | /* |
| 2860 | * Reduce accuracy by one bit such that @a and @b converge |
| 2861 | * to a similar magnitude. |
| 2862 | */ |
| 2863 | #define REDUCE_FLS(a, b) \ |
| 2864 | do { \ |
| 2865 | if (a##_fls > b##_fls) { \ |
| 2866 | a >>= 1; \ |
| 2867 | a##_fls--; \ |
| 2868 | } else { \ |
| 2869 | b >>= 1; \ |
| 2870 | b##_fls--; \ |
| 2871 | } \ |
| 2872 | } while (0) |
| 2873 | |
| 2874 | /* |
| 2875 | * Reduce accuracy until either term fits in a u64, then proceed with |
| 2876 | * the other, so that finally we can do a u64/u64 division. |
| 2877 | */ |
| 2878 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { |
| 2879 | REDUCE_FLS(nsec, frequency); |
| 2880 | REDUCE_FLS(sec, count); |
| 2881 | } |
| 2882 | |
| 2883 | if (count_fls + sec_fls > 64) { |
| 2884 | divisor = nsec * frequency; |
| 2885 | |
| 2886 | while (count_fls + sec_fls > 64) { |
| 2887 | REDUCE_FLS(count, sec); |
| 2888 | divisor >>= 1; |
| 2889 | } |
| 2890 | |
| 2891 | dividend = count * sec; |
| 2892 | } else { |
| 2893 | dividend = count * sec; |
| 2894 | |
| 2895 | while (nsec_fls + frequency_fls > 64) { |
| 2896 | REDUCE_FLS(nsec, frequency); |
| 2897 | dividend >>= 1; |
| 2898 | } |
| 2899 | |
| 2900 | divisor = nsec * frequency; |
| 2901 | } |
| 2902 | |
| 2903 | if (!divisor) |
| 2904 | return dividend; |
| 2905 | |
| 2906 | return div64_u64(dividend, divisor); |
| 2907 | } |
| 2908 | |
| 2909 | static DEFINE_PER_CPU(int, perf_throttled_count); |
| 2910 | static DEFINE_PER_CPU(u64, perf_throttled_seq); |
| 2911 | |
| 2912 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
| 2913 | { |
| 2914 | struct hw_perf_event *hwc = &event->hw; |
| 2915 | s64 period, sample_period; |
| 2916 | s64 delta; |
| 2917 | |
| 2918 | period = perf_calculate_period(event, nsec, count); |
| 2919 | |
| 2920 | delta = (s64)(period - hwc->sample_period); |
| 2921 | delta = (delta + 7) / 8; /* low pass filter */ |
| 2922 | |
| 2923 | sample_period = hwc->sample_period + delta; |
| 2924 | |
| 2925 | if (!sample_period) |
| 2926 | sample_period = 1; |
| 2927 | |
| 2928 | hwc->sample_period = sample_period; |
| 2929 | |
| 2930 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
| 2931 | if (disable) |
| 2932 | event->pmu->stop(event, PERF_EF_UPDATE); |
| 2933 | |
| 2934 | local64_set(&hwc->period_left, 0); |
| 2935 | |
| 2936 | if (disable) |
| 2937 | event->pmu->start(event, PERF_EF_RELOAD); |
| 2938 | } |
| 2939 | } |
| 2940 | |
| 2941 | /* |
| 2942 | * combine freq adjustment with unthrottling to avoid two passes over the |
| 2943 | * events. At the same time, make sure, having freq events does not change |
| 2944 | * the rate of unthrottling as that would introduce bias. |
| 2945 | */ |
| 2946 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, |
| 2947 | int needs_unthr) |
| 2948 | { |
| 2949 | struct perf_event *event; |
| 2950 | struct hw_perf_event *hwc; |
| 2951 | u64 now, period = TICK_NSEC; |
| 2952 | s64 delta; |
| 2953 | |
| 2954 | /* |
| 2955 | * only need to iterate over all events iff: |
| 2956 | * - context have events in frequency mode (needs freq adjust) |
| 2957 | * - there are events to unthrottle on this cpu |
| 2958 | */ |
| 2959 | if (!(ctx->nr_freq || needs_unthr)) |
| 2960 | return; |
| 2961 | |
| 2962 | raw_spin_lock(&ctx->lock); |
| 2963 | perf_pmu_disable(ctx->pmu); |
| 2964 | |
| 2965 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
| 2966 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
| 2967 | continue; |
| 2968 | |
| 2969 | if (!event_filter_match(event)) |
| 2970 | continue; |
| 2971 | |
| 2972 | perf_pmu_disable(event->pmu); |
| 2973 | |
| 2974 | hwc = &event->hw; |
| 2975 | |
| 2976 | if (hwc->interrupts == MAX_INTERRUPTS) { |
| 2977 | hwc->interrupts = 0; |
| 2978 | perf_log_throttle(event, 1); |
| 2979 | event->pmu->start(event, 0); |
| 2980 | } |
| 2981 | |
| 2982 | if (!event->attr.freq || !event->attr.sample_freq) |
| 2983 | goto next; |
| 2984 | |
| 2985 | /* |
| 2986 | * stop the event and update event->count |
| 2987 | */ |
| 2988 | event->pmu->stop(event, PERF_EF_UPDATE); |
| 2989 | |
| 2990 | now = local64_read(&event->count); |
| 2991 | delta = now - hwc->freq_count_stamp; |
| 2992 | hwc->freq_count_stamp = now; |
| 2993 | |
| 2994 | /* |
| 2995 | * restart the event |
| 2996 | * reload only if value has changed |
| 2997 | * we have stopped the event so tell that |
| 2998 | * to perf_adjust_period() to avoid stopping it |
| 2999 | * twice. |
| 3000 | */ |
| 3001 | if (delta > 0) |
| 3002 | perf_adjust_period(event, period, delta, false); |
| 3003 | |
| 3004 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); |
| 3005 | next: |
| 3006 | perf_pmu_enable(event->pmu); |
| 3007 | } |
| 3008 | |
| 3009 | perf_pmu_enable(ctx->pmu); |
| 3010 | raw_spin_unlock(&ctx->lock); |
| 3011 | } |
| 3012 | |
| 3013 | /* |
| 3014 | * Round-robin a context's events: |
| 3015 | */ |
| 3016 | static void rotate_ctx(struct perf_event_context *ctx) |
| 3017 | { |
| 3018 | /* |
| 3019 | * Rotate the first entry last of non-pinned groups. Rotation might be |
| 3020 | * disabled by the inheritance code. |
| 3021 | */ |
| 3022 | if (!ctx->rotate_disable) |
| 3023 | list_rotate_left(&ctx->flexible_groups); |
| 3024 | } |
| 3025 | |
| 3026 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
| 3027 | { |
| 3028 | struct perf_event_context *ctx = NULL; |
| 3029 | int rotate = 0; |
| 3030 | |
| 3031 | if (cpuctx->ctx.nr_events) { |
| 3032 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
| 3033 | rotate = 1; |
| 3034 | } |
| 3035 | |
| 3036 | ctx = cpuctx->task_ctx; |
| 3037 | if (ctx && ctx->nr_events) { |
| 3038 | if (ctx->nr_events != ctx->nr_active) |
| 3039 | rotate = 1; |
| 3040 | } |
| 3041 | |
| 3042 | if (!rotate) |
| 3043 | goto done; |
| 3044 | |
| 3045 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
| 3046 | perf_pmu_disable(cpuctx->ctx.pmu); |
| 3047 | |
| 3048 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
| 3049 | if (ctx) |
| 3050 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); |
| 3051 | |
| 3052 | rotate_ctx(&cpuctx->ctx); |
| 3053 | if (ctx) |
| 3054 | rotate_ctx(ctx); |
| 3055 | |
| 3056 | perf_event_sched_in(cpuctx, ctx, current); |
| 3057 | |
| 3058 | perf_pmu_enable(cpuctx->ctx.pmu); |
| 3059 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); |
| 3060 | done: |
| 3061 | |
| 3062 | return rotate; |
| 3063 | } |
| 3064 | |
| 3065 | #ifdef CONFIG_NO_HZ_FULL |
| 3066 | bool perf_event_can_stop_tick(void) |
| 3067 | { |
| 3068 | if (atomic_read(&nr_freq_events) || |
| 3069 | __this_cpu_read(perf_throttled_count)) |
| 3070 | return false; |
| 3071 | else |
| 3072 | return true; |
| 3073 | } |
| 3074 | #endif |
| 3075 | |
| 3076 | void perf_event_task_tick(void) |
| 3077 | { |
| 3078 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
| 3079 | struct perf_event_context *ctx, *tmp; |
| 3080 | int throttled; |
| 3081 | |
| 3082 | WARN_ON(!irqs_disabled()); |
| 3083 | |
| 3084 | __this_cpu_inc(perf_throttled_seq); |
| 3085 | throttled = __this_cpu_xchg(perf_throttled_count, 0); |
| 3086 | |
| 3087 | list_for_each_entry_safe(ctx, tmp, head, active_ctx_list) |
| 3088 | perf_adjust_freq_unthr_context(ctx, throttled); |
| 3089 | } |
| 3090 | |
| 3091 | static int event_enable_on_exec(struct perf_event *event, |
| 3092 | struct perf_event_context *ctx) |
| 3093 | { |
| 3094 | if (!event->attr.enable_on_exec) |
| 3095 | return 0; |
| 3096 | |
| 3097 | event->attr.enable_on_exec = 0; |
| 3098 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
| 3099 | return 0; |
| 3100 | |
| 3101 | __perf_event_mark_enabled(event); |
| 3102 | |
| 3103 | return 1; |
| 3104 | } |
| 3105 | |
| 3106 | /* |
| 3107 | * Enable all of a task's events that have been marked enable-on-exec. |
| 3108 | * This expects task == current. |
| 3109 | */ |
| 3110 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
| 3111 | { |
| 3112 | struct perf_event_context *clone_ctx = NULL; |
| 3113 | struct perf_event *event; |
| 3114 | unsigned long flags; |
| 3115 | int enabled = 0; |
| 3116 | int ret; |
| 3117 | |
| 3118 | local_irq_save(flags); |
| 3119 | if (!ctx || !ctx->nr_events) |
| 3120 | goto out; |
| 3121 | |
| 3122 | /* |
| 3123 | * We must ctxsw out cgroup events to avoid conflict |
| 3124 | * when invoking perf_task_event_sched_in() later on |
| 3125 | * in this function. Otherwise we end up trying to |
| 3126 | * ctxswin cgroup events which are already scheduled |
| 3127 | * in. |
| 3128 | */ |
| 3129 | perf_cgroup_sched_out(current, NULL); |
| 3130 | |
| 3131 | raw_spin_lock(&ctx->lock); |
| 3132 | task_ctx_sched_out(ctx); |
| 3133 | |
| 3134 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
| 3135 | ret = event_enable_on_exec(event, ctx); |
| 3136 | if (ret) |
| 3137 | enabled = 1; |
| 3138 | } |
| 3139 | |
| 3140 | /* |
| 3141 | * Unclone this context if we enabled any event. |
| 3142 | */ |
| 3143 | if (enabled) |
| 3144 | clone_ctx = unclone_ctx(ctx); |
| 3145 | |
| 3146 | raw_spin_unlock(&ctx->lock); |
| 3147 | |
| 3148 | /* |
| 3149 | * Also calls ctxswin for cgroup events, if any: |
| 3150 | */ |
| 3151 | perf_event_context_sched_in(ctx, ctx->task); |
| 3152 | out: |
| 3153 | local_irq_restore(flags); |
| 3154 | |
| 3155 | if (clone_ctx) |
| 3156 | put_ctx(clone_ctx); |
| 3157 | } |
| 3158 | |
| 3159 | void perf_event_exec(void) |
| 3160 | { |
| 3161 | struct perf_event_context *ctx; |
| 3162 | int ctxn; |
| 3163 | |
| 3164 | rcu_read_lock(); |
| 3165 | for_each_task_context_nr(ctxn) { |
| 3166 | ctx = current->perf_event_ctxp[ctxn]; |
| 3167 | if (!ctx) |
| 3168 | continue; |
| 3169 | |
| 3170 | perf_event_enable_on_exec(ctx); |
| 3171 | } |
| 3172 | rcu_read_unlock(); |
| 3173 | } |
| 3174 | |
| 3175 | /* |
| 3176 | * Cross CPU call to read the hardware event |
| 3177 | */ |
| 3178 | static void __perf_event_read(void *info) |
| 3179 | { |
| 3180 | struct perf_event *event = info; |
| 3181 | struct perf_event_context *ctx = event->ctx; |
| 3182 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
| 3183 | |
| 3184 | /* |
| 3185 | * If this is a task context, we need to check whether it is |
| 3186 | * the current task context of this cpu. If not it has been |
| 3187 | * scheduled out before the smp call arrived. In that case |
| 3188 | * event->count would have been updated to a recent sample |
| 3189 | * when the event was scheduled out. |
| 3190 | */ |
| 3191 | if (ctx->task && cpuctx->task_ctx != ctx) |
| 3192 | return; |
| 3193 | |
| 3194 | raw_spin_lock(&ctx->lock); |
| 3195 | if (ctx->is_active) { |
| 3196 | update_context_time(ctx); |
| 3197 | update_cgrp_time_from_event(event); |
| 3198 | } |
| 3199 | update_event_times(event); |
| 3200 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
| 3201 | event->pmu->read(event); |
| 3202 | raw_spin_unlock(&ctx->lock); |
| 3203 | } |
| 3204 | |
| 3205 | static inline u64 perf_event_count(struct perf_event *event) |
| 3206 | { |
| 3207 | if (event->pmu->count) |
| 3208 | return event->pmu->count(event); |
| 3209 | |
| 3210 | return __perf_event_count(event); |
| 3211 | } |
| 3212 | |
| 3213 | static u64 perf_event_read(struct perf_event *event) |
| 3214 | { |
| 3215 | /* |
| 3216 | * If event is enabled and currently active on a CPU, update the |
| 3217 | * value in the event structure: |
| 3218 | */ |
| 3219 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
| 3220 | smp_call_function_single(event->oncpu, |
| 3221 | __perf_event_read, event, 1); |
| 3222 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { |
| 3223 | struct perf_event_context *ctx = event->ctx; |
| 3224 | unsigned long flags; |
| 3225 | |
| 3226 | raw_spin_lock_irqsave(&ctx->lock, flags); |
| 3227 | /* |
| 3228 | * may read while context is not active |
| 3229 | * (e.g., thread is blocked), in that case |
| 3230 | * we cannot update context time |
| 3231 | */ |
| 3232 | if (ctx->is_active) { |
| 3233 | update_context_time(ctx); |
| 3234 | update_cgrp_time_from_event(event); |
| 3235 | } |
| 3236 | update_event_times(event); |
| 3237 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
| 3238 | } |
| 3239 | |
| 3240 | return perf_event_count(event); |
| 3241 | } |
| 3242 | |
| 3243 | /* |
| 3244 | * Initialize the perf_event context in a task_struct: |
| 3245 | */ |
| 3246 | static void __perf_event_init_context(struct perf_event_context *ctx) |
| 3247 | { |
| 3248 | raw_spin_lock_init(&ctx->lock); |
| 3249 | mutex_init(&ctx->mutex); |
| 3250 | INIT_LIST_HEAD(&ctx->active_ctx_list); |
| 3251 | INIT_LIST_HEAD(&ctx->pinned_groups); |
| 3252 | INIT_LIST_HEAD(&ctx->flexible_groups); |
| 3253 | INIT_LIST_HEAD(&ctx->event_list); |
| 3254 | atomic_set(&ctx->refcount, 1); |
| 3255 | INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work); |
| 3256 | } |
| 3257 | |
| 3258 | static struct perf_event_context * |
| 3259 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) |
| 3260 | { |
| 3261 | struct perf_event_context *ctx; |
| 3262 | |
| 3263 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); |
| 3264 | if (!ctx) |
| 3265 | return NULL; |
| 3266 | |
| 3267 | __perf_event_init_context(ctx); |
| 3268 | if (task) { |
| 3269 | ctx->task = task; |
| 3270 | get_task_struct(task); |
| 3271 | } |
| 3272 | ctx->pmu = pmu; |
| 3273 | |
| 3274 | return ctx; |
| 3275 | } |
| 3276 | |
| 3277 | static struct task_struct * |
| 3278 | find_lively_task_by_vpid(pid_t vpid) |
| 3279 | { |
| 3280 | struct task_struct *task; |
| 3281 | int err; |
| 3282 | |
| 3283 | rcu_read_lock(); |
| 3284 | if (!vpid) |
| 3285 | task = current; |
| 3286 | else |
| 3287 | task = find_task_by_vpid(vpid); |
| 3288 | if (task) |
| 3289 | get_task_struct(task); |
| 3290 | rcu_read_unlock(); |
| 3291 | |
| 3292 | if (!task) |
| 3293 | return ERR_PTR(-ESRCH); |
| 3294 | |
| 3295 | /* Reuse ptrace permission checks for now. */ |
| 3296 | err = -EACCES; |
| 3297 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) |
| 3298 | goto errout; |
| 3299 | |
| 3300 | return task; |
| 3301 | errout: |
| 3302 | put_task_struct(task); |
| 3303 | return ERR_PTR(err); |
| 3304 | |
| 3305 | } |
| 3306 | |
| 3307 | /* |
| 3308 | * Returns a matching context with refcount and pincount. |
| 3309 | */ |
| 3310 | static struct perf_event_context * |
| 3311 | find_get_context(struct pmu *pmu, struct task_struct *task, |
| 3312 | struct perf_event *event) |
| 3313 | { |
| 3314 | struct perf_event_context *ctx, *clone_ctx = NULL; |
| 3315 | struct perf_cpu_context *cpuctx; |
| 3316 | void *task_ctx_data = NULL; |
| 3317 | unsigned long flags; |
| 3318 | int ctxn, err; |
| 3319 | int cpu = event->cpu; |
| 3320 | |
| 3321 | if (!task) { |
| 3322 | /* Must be root to operate on a CPU event: */ |
| 3323 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
| 3324 | return ERR_PTR(-EACCES); |
| 3325 | |
| 3326 | /* |
| 3327 | * We could be clever and allow to attach a event to an |
| 3328 | * offline CPU and activate it when the CPU comes up, but |
| 3329 | * that's for later. |
| 3330 | */ |
| 3331 | if (!cpu_online(cpu)) |
| 3332 | return ERR_PTR(-ENODEV); |
| 3333 | |
| 3334 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
| 3335 | ctx = &cpuctx->ctx; |
| 3336 | get_ctx(ctx); |
| 3337 | ++ctx->pin_count; |
| 3338 | |
| 3339 | return ctx; |
| 3340 | } |
| 3341 | |
| 3342 | err = -EINVAL; |
| 3343 | ctxn = pmu->task_ctx_nr; |
| 3344 | if (ctxn < 0) |
| 3345 | goto errout; |
| 3346 | |
| 3347 | if (event->attach_state & PERF_ATTACH_TASK_DATA) { |
| 3348 | task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL); |
| 3349 | if (!task_ctx_data) { |
| 3350 | err = -ENOMEM; |
| 3351 | goto errout; |
| 3352 | } |
| 3353 | } |
| 3354 | |
| 3355 | retry: |
| 3356 | ctx = perf_lock_task_context(task, ctxn, &flags); |
| 3357 | if (ctx) { |
| 3358 | clone_ctx = unclone_ctx(ctx); |
| 3359 | ++ctx->pin_count; |
| 3360 | |
| 3361 | if (task_ctx_data && !ctx->task_ctx_data) { |
| 3362 | ctx->task_ctx_data = task_ctx_data; |
| 3363 | task_ctx_data = NULL; |
| 3364 | } |
| 3365 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
| 3366 | |
| 3367 | if (clone_ctx) |
| 3368 | put_ctx(clone_ctx); |
| 3369 | } else { |
| 3370 | ctx = alloc_perf_context(pmu, task); |
| 3371 | err = -ENOMEM; |
| 3372 | if (!ctx) |
| 3373 | goto errout; |
| 3374 | |
| 3375 | if (task_ctx_data) { |
| 3376 | ctx->task_ctx_data = task_ctx_data; |
| 3377 | task_ctx_data = NULL; |
| 3378 | } |
| 3379 | |
| 3380 | err = 0; |
| 3381 | mutex_lock(&task->perf_event_mutex); |
| 3382 | /* |
| 3383 | * If it has already passed perf_event_exit_task(). |
| 3384 | * we must see PF_EXITING, it takes this mutex too. |
| 3385 | */ |
| 3386 | if (task->flags & PF_EXITING) |
| 3387 | err = -ESRCH; |
| 3388 | else if (task->perf_event_ctxp[ctxn]) |
| 3389 | err = -EAGAIN; |
| 3390 | else { |
| 3391 | get_ctx(ctx); |
| 3392 | ++ctx->pin_count; |
| 3393 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
| 3394 | } |
| 3395 | mutex_unlock(&task->perf_event_mutex); |
| 3396 | |
| 3397 | if (unlikely(err)) { |
| 3398 | put_ctx(ctx); |
| 3399 | |
| 3400 | if (err == -EAGAIN) |
| 3401 | goto retry; |
| 3402 | goto errout; |
| 3403 | } |
| 3404 | } |
| 3405 | |
| 3406 | kfree(task_ctx_data); |
| 3407 | return ctx; |
| 3408 | |
| 3409 | errout: |
| 3410 | kfree(task_ctx_data); |
| 3411 | return ERR_PTR(err); |
| 3412 | } |
| 3413 | |
| 3414 | static void perf_event_free_filter(struct perf_event *event); |
| 3415 | static void perf_event_free_bpf_prog(struct perf_event *event); |
| 3416 | |
| 3417 | static void free_event_rcu(struct rcu_head *head) |
| 3418 | { |
| 3419 | struct perf_event *event; |
| 3420 | |
| 3421 | event = container_of(head, struct perf_event, rcu_head); |
| 3422 | if (event->ns) |
| 3423 | put_pid_ns(event->ns); |
| 3424 | perf_event_free_filter(event); |
| 3425 | perf_event_free_bpf_prog(event); |
| 3426 | kfree(event); |
| 3427 | } |
| 3428 | |
| 3429 | static void ring_buffer_attach(struct perf_event *event, |
| 3430 | struct ring_buffer *rb); |
| 3431 | |
| 3432 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
| 3433 | { |
| 3434 | if (event->parent) |
| 3435 | return; |
| 3436 | |
| 3437 | if (is_cgroup_event(event)) |
| 3438 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); |
| 3439 | } |
| 3440 | |
| 3441 | static void unaccount_event(struct perf_event *event) |
| 3442 | { |
| 3443 | if (event->parent) |
| 3444 | return; |
| 3445 | |
| 3446 | if (event->attach_state & PERF_ATTACH_TASK) |
| 3447 | static_key_slow_dec_deferred(&perf_sched_events); |
| 3448 | if (event->attr.mmap || event->attr.mmap_data) |
| 3449 | atomic_dec(&nr_mmap_events); |
| 3450 | if (event->attr.comm) |
| 3451 | atomic_dec(&nr_comm_events); |
| 3452 | if (event->attr.task) |
| 3453 | atomic_dec(&nr_task_events); |
| 3454 | if (event->attr.freq) |
| 3455 | atomic_dec(&nr_freq_events); |
| 3456 | if (is_cgroup_event(event)) |
| 3457 | static_key_slow_dec_deferred(&perf_sched_events); |
| 3458 | if (has_branch_stack(event)) |
| 3459 | static_key_slow_dec_deferred(&perf_sched_events); |
| 3460 | |
| 3461 | unaccount_event_cpu(event, event->cpu); |
| 3462 | } |
| 3463 | |
| 3464 | /* |
| 3465 | * The following implement mutual exclusion of events on "exclusive" pmus |
| 3466 | * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled |
| 3467 | * at a time, so we disallow creating events that might conflict, namely: |
| 3468 | * |
| 3469 | * 1) cpu-wide events in the presence of per-task events, |
| 3470 | * 2) per-task events in the presence of cpu-wide events, |
| 3471 | * 3) two matching events on the same context. |
| 3472 | * |
| 3473 | * The former two cases are handled in the allocation path (perf_event_alloc(), |
| 3474 | * __free_event()), the latter -- before the first perf_install_in_context(). |
| 3475 | */ |
| 3476 | static int exclusive_event_init(struct perf_event *event) |
| 3477 | { |
| 3478 | struct pmu *pmu = event->pmu; |
| 3479 | |
| 3480 | if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE)) |
| 3481 | return 0; |
| 3482 | |
| 3483 | /* |
| 3484 | * Prevent co-existence of per-task and cpu-wide events on the |
| 3485 | * same exclusive pmu. |
| 3486 | * |
| 3487 | * Negative pmu::exclusive_cnt means there are cpu-wide |
| 3488 | * events on this "exclusive" pmu, positive means there are |
| 3489 | * per-task events. |
| 3490 | * |
| 3491 | * Since this is called in perf_event_alloc() path, event::ctx |
| 3492 | * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK |
| 3493 | * to mean "per-task event", because unlike other attach states it |
| 3494 | * never gets cleared. |
| 3495 | */ |
| 3496 | if (event->attach_state & PERF_ATTACH_TASK) { |
| 3497 | if (!atomic_inc_unless_negative(&pmu->exclusive_cnt)) |
| 3498 | return -EBUSY; |
| 3499 | } else { |
| 3500 | if (!atomic_dec_unless_positive(&pmu->exclusive_cnt)) |
| 3501 | return -EBUSY; |
| 3502 | } |
| 3503 | |
| 3504 | return 0; |
| 3505 | } |
| 3506 | |
| 3507 | static void exclusive_event_destroy(struct perf_event *event) |
| 3508 | { |
| 3509 | struct pmu *pmu = event->pmu; |
| 3510 | |
| 3511 | if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE)) |
| 3512 | return; |
| 3513 | |
| 3514 | /* see comment in exclusive_event_init() */ |
| 3515 | if (event->attach_state & PERF_ATTACH_TASK) |
| 3516 | atomic_dec(&pmu->exclusive_cnt); |
| 3517 | else |
| 3518 | atomic_inc(&pmu->exclusive_cnt); |
| 3519 | } |
| 3520 | |
| 3521 | static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2) |
| 3522 | { |
| 3523 | if ((e1->pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && |
| 3524 | (e1->cpu == e2->cpu || |
| 3525 | e1->cpu == -1 || |
| 3526 | e2->cpu == -1)) |
| 3527 | return true; |
| 3528 | return false; |
| 3529 | } |
| 3530 | |
| 3531 | /* Called under the same ctx::mutex as perf_install_in_context() */ |
| 3532 | static bool exclusive_event_installable(struct perf_event *event, |
| 3533 | struct perf_event_context *ctx) |
| 3534 | { |
| 3535 | struct perf_event *iter_event; |
| 3536 | struct pmu *pmu = event->pmu; |
| 3537 | |
| 3538 | if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE)) |
| 3539 | return true; |
| 3540 | |
| 3541 | list_for_each_entry(iter_event, &ctx->event_list, event_entry) { |
| 3542 | if (exclusive_event_match(iter_event, event)) |
| 3543 | return false; |
| 3544 | } |
| 3545 | |
| 3546 | return true; |
| 3547 | } |
| 3548 | |
| 3549 | static void __free_event(struct perf_event *event) |
| 3550 | { |
| 3551 | if (!event->parent) { |
| 3552 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
| 3553 | put_callchain_buffers(); |
| 3554 | } |
| 3555 | |
| 3556 | if (event->destroy) |
| 3557 | event->destroy(event); |
| 3558 | |
| 3559 | if (event->ctx) |
| 3560 | put_ctx(event->ctx); |
| 3561 | |
| 3562 | if (event->pmu) { |
| 3563 | exclusive_event_destroy(event); |
| 3564 | module_put(event->pmu->module); |
| 3565 | } |
| 3566 | |
| 3567 | call_rcu(&event->rcu_head, free_event_rcu); |
| 3568 | } |
| 3569 | |
| 3570 | static void _free_event(struct perf_event *event) |
| 3571 | { |
| 3572 | irq_work_sync(&event->pending); |
| 3573 | |
| 3574 | unaccount_event(event); |
| 3575 | |
| 3576 | if (event->rb) { |
| 3577 | /* |
| 3578 | * Can happen when we close an event with re-directed output. |
| 3579 | * |
| 3580 | * Since we have a 0 refcount, perf_mmap_close() will skip |
| 3581 | * over us; possibly making our ring_buffer_put() the last. |
| 3582 | */ |
| 3583 | mutex_lock(&event->mmap_mutex); |
| 3584 | ring_buffer_attach(event, NULL); |
| 3585 | mutex_unlock(&event->mmap_mutex); |
| 3586 | } |
| 3587 | |
| 3588 | if (is_cgroup_event(event)) |
| 3589 | perf_detach_cgroup(event); |
| 3590 | |
| 3591 | __free_event(event); |
| 3592 | } |
| 3593 | |
| 3594 | /* |
| 3595 | * Used to free events which have a known refcount of 1, such as in error paths |
| 3596 | * where the event isn't exposed yet and inherited events. |
| 3597 | */ |
| 3598 | static void free_event(struct perf_event *event) |
| 3599 | { |
| 3600 | if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, |
| 3601 | "unexpected event refcount: %ld; ptr=%p\n", |
| 3602 | atomic_long_read(&event->refcount), event)) { |
| 3603 | /* leak to avoid use-after-free */ |
| 3604 | return; |
| 3605 | } |
| 3606 | |
| 3607 | _free_event(event); |
| 3608 | } |
| 3609 | |
| 3610 | /* |
| 3611 | * Remove user event from the owner task. |
| 3612 | */ |
| 3613 | static void perf_remove_from_owner(struct perf_event *event) |
| 3614 | { |
| 3615 | struct task_struct *owner; |
| 3616 | |
| 3617 | rcu_read_lock(); |
| 3618 | owner = ACCESS_ONCE(event->owner); |
| 3619 | /* |
| 3620 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe |
| 3621 | * !owner it means the list deletion is complete and we can indeed |
| 3622 | * free this event, otherwise we need to serialize on |
| 3623 | * owner->perf_event_mutex. |
| 3624 | */ |
| 3625 | smp_read_barrier_depends(); |
| 3626 | if (owner) { |
| 3627 | /* |
| 3628 | * Since delayed_put_task_struct() also drops the last |
| 3629 | * task reference we can safely take a new reference |
| 3630 | * while holding the rcu_read_lock(). |
| 3631 | */ |
| 3632 | get_task_struct(owner); |
| 3633 | } |
| 3634 | rcu_read_unlock(); |
| 3635 | |
| 3636 | if (owner) { |
| 3637 | /* |
| 3638 | * If we're here through perf_event_exit_task() we're already |
| 3639 | * holding ctx->mutex which would be an inversion wrt. the |
| 3640 | * normal lock order. |
| 3641 | * |
| 3642 | * However we can safely take this lock because its the child |
| 3643 | * ctx->mutex. |
| 3644 | */ |
| 3645 | mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING); |
| 3646 | |
| 3647 | /* |
| 3648 | * We have to re-check the event->owner field, if it is cleared |
| 3649 | * we raced with perf_event_exit_task(), acquiring the mutex |
| 3650 | * ensured they're done, and we can proceed with freeing the |
| 3651 | * event. |
| 3652 | */ |
| 3653 | if (event->owner) |
| 3654 | list_del_init(&event->owner_entry); |
| 3655 | mutex_unlock(&owner->perf_event_mutex); |
| 3656 | put_task_struct(owner); |
| 3657 | } |
| 3658 | } |
| 3659 | |
| 3660 | /* |
| 3661 | * Called when the last reference to the file is gone. |
| 3662 | */ |
| 3663 | static void put_event(struct perf_event *event) |
| 3664 | { |
| 3665 | struct perf_event_context *ctx; |
| 3666 | |
| 3667 | if (!atomic_long_dec_and_test(&event->refcount)) |
| 3668 | return; |
| 3669 | |
| 3670 | if (!is_kernel_event(event)) |
| 3671 | perf_remove_from_owner(event); |
| 3672 | |
| 3673 | /* |
| 3674 | * There are two ways this annotation is useful: |
| 3675 | * |
| 3676 | * 1) there is a lock recursion from perf_event_exit_task |
| 3677 | * see the comment there. |
| 3678 | * |
| 3679 | * 2) there is a lock-inversion with mmap_sem through |
| 3680 | * perf_event_read_group(), which takes faults while |
| 3681 | * holding ctx->mutex, however this is called after |
| 3682 | * the last filedesc died, so there is no possibility |
| 3683 | * to trigger the AB-BA case. |
| 3684 | */ |
| 3685 | ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING); |
| 3686 | WARN_ON_ONCE(ctx->parent_ctx); |
| 3687 | perf_remove_from_context(event, true); |
| 3688 | perf_event_ctx_unlock(event, ctx); |
| 3689 | |
| 3690 | _free_event(event); |
| 3691 | } |
| 3692 | |
| 3693 | int perf_event_release_kernel(struct perf_event *event) |
| 3694 | { |
| 3695 | put_event(event); |
| 3696 | return 0; |
| 3697 | } |
| 3698 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
| 3699 | |
| 3700 | static int perf_release(struct inode *inode, struct file *file) |
| 3701 | { |
| 3702 | put_event(file->private_data); |
| 3703 | return 0; |
| 3704 | } |
| 3705 | |
| 3706 | /* |
| 3707 | * Remove all orphanes events from the context. |
| 3708 | */ |
| 3709 | static void orphans_remove_work(struct work_struct *work) |
| 3710 | { |
| 3711 | struct perf_event_context *ctx; |
| 3712 | struct perf_event *event, *tmp; |
| 3713 | |
| 3714 | ctx = container_of(work, struct perf_event_context, |
| 3715 | orphans_remove.work); |
| 3716 | |
| 3717 | mutex_lock(&ctx->mutex); |
| 3718 | list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) { |
| 3719 | struct perf_event *parent_event = event->parent; |
| 3720 | |
| 3721 | if (!is_orphaned_child(event)) |
| 3722 | continue; |
| 3723 | |
| 3724 | perf_remove_from_context(event, true); |
| 3725 | |
| 3726 | mutex_lock(&parent_event->child_mutex); |
| 3727 | list_del_init(&event->child_list); |
| 3728 | mutex_unlock(&parent_event->child_mutex); |
| 3729 | |
| 3730 | free_event(event); |
| 3731 | put_event(parent_event); |
| 3732 | } |
| 3733 | |
| 3734 | raw_spin_lock_irq(&ctx->lock); |
| 3735 | ctx->orphans_remove_sched = false; |
| 3736 | raw_spin_unlock_irq(&ctx->lock); |
| 3737 | mutex_unlock(&ctx->mutex); |
| 3738 | |
| 3739 | put_ctx(ctx); |
| 3740 | } |
| 3741 | |
| 3742 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
| 3743 | { |
| 3744 | struct perf_event *child; |
| 3745 | u64 total = 0; |
| 3746 | |
| 3747 | *enabled = 0; |
| 3748 | *running = 0; |
| 3749 | |
| 3750 | mutex_lock(&event->child_mutex); |
| 3751 | total += perf_event_read(event); |
| 3752 | *enabled += event->total_time_enabled + |
| 3753 | atomic64_read(&event->child_total_time_enabled); |
| 3754 | *running += event->total_time_running + |
| 3755 | atomic64_read(&event->child_total_time_running); |
| 3756 | |
| 3757 | list_for_each_entry(child, &event->child_list, child_list) { |
| 3758 | total += perf_event_read(child); |
| 3759 | *enabled += child->total_time_enabled; |
| 3760 | *running += child->total_time_running; |
| 3761 | } |
| 3762 | mutex_unlock(&event->child_mutex); |
| 3763 | |
| 3764 | return total; |
| 3765 | } |
| 3766 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
| 3767 | |
| 3768 | static int perf_event_read_group(struct perf_event *event, |
| 3769 | u64 read_format, char __user *buf) |
| 3770 | { |
| 3771 | struct perf_event *leader = event->group_leader, *sub; |
| 3772 | struct perf_event_context *ctx = leader->ctx; |
| 3773 | int n = 0, size = 0, ret; |
| 3774 | u64 count, enabled, running; |
| 3775 | u64 values[5]; |
| 3776 | |
| 3777 | lockdep_assert_held(&ctx->mutex); |
| 3778 | |
| 3779 | count = perf_event_read_value(leader, &enabled, &running); |
| 3780 | |
| 3781 | values[n++] = 1 + leader->nr_siblings; |
| 3782 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
| 3783 | values[n++] = enabled; |
| 3784 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
| 3785 | values[n++] = running; |
| 3786 | values[n++] = count; |
| 3787 | if (read_format & PERF_FORMAT_ID) |
| 3788 | values[n++] = primary_event_id(leader); |
| 3789 | |
| 3790 | size = n * sizeof(u64); |
| 3791 | |
| 3792 | if (copy_to_user(buf, values, size)) |
| 3793 | return -EFAULT; |
| 3794 | |
| 3795 | ret = size; |
| 3796 | |
| 3797 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
| 3798 | n = 0; |
| 3799 | |
| 3800 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
| 3801 | if (read_format & PERF_FORMAT_ID) |
| 3802 | values[n++] = primary_event_id(sub); |
| 3803 | |
| 3804 | size = n * sizeof(u64); |
| 3805 | |
| 3806 | if (copy_to_user(buf + ret, values, size)) { |
| 3807 | return -EFAULT; |
| 3808 | } |
| 3809 | |
| 3810 | ret += size; |
| 3811 | } |
| 3812 | |
| 3813 | return ret; |
| 3814 | } |
| 3815 | |
| 3816 | static int perf_event_read_one(struct perf_event *event, |
| 3817 | u64 read_format, char __user *buf) |
| 3818 | { |
| 3819 | u64 enabled, running; |
| 3820 | u64 values[4]; |
| 3821 | int n = 0; |
| 3822 | |
| 3823 | values[n++] = perf_event_read_value(event, &enabled, &running); |
| 3824 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
| 3825 | values[n++] = enabled; |
| 3826 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
| 3827 | values[n++] = running; |
| 3828 | if (read_format & PERF_FORMAT_ID) |
| 3829 | values[n++] = primary_event_id(event); |
| 3830 | |
| 3831 | if (copy_to_user(buf, values, n * sizeof(u64))) |
| 3832 | return -EFAULT; |
| 3833 | |
| 3834 | return n * sizeof(u64); |
| 3835 | } |
| 3836 | |
| 3837 | static bool is_event_hup(struct perf_event *event) |
| 3838 | { |
| 3839 | bool no_children; |
| 3840 | |
| 3841 | if (event->state != PERF_EVENT_STATE_EXIT) |
| 3842 | return false; |
| 3843 | |
| 3844 | mutex_lock(&event->child_mutex); |
| 3845 | no_children = list_empty(&event->child_list); |
| 3846 | mutex_unlock(&event->child_mutex); |
| 3847 | return no_children; |
| 3848 | } |
| 3849 | |
| 3850 | /* |
| 3851 | * Read the performance event - simple non blocking version for now |
| 3852 | */ |
| 3853 | static ssize_t |
| 3854 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
| 3855 | { |
| 3856 | u64 read_format = event->attr.read_format; |
| 3857 | int ret; |
| 3858 | |
| 3859 | /* |
| 3860 | * Return end-of-file for a read on a event that is in |
| 3861 | * error state (i.e. because it was pinned but it couldn't be |
| 3862 | * scheduled on to the CPU at some point). |
| 3863 | */ |
| 3864 | if (event->state == PERF_EVENT_STATE_ERROR) |
| 3865 | return 0; |
| 3866 | |
| 3867 | if (count < event->read_size) |
| 3868 | return -ENOSPC; |
| 3869 | |
| 3870 | WARN_ON_ONCE(event->ctx->parent_ctx); |
| 3871 | if (read_format & PERF_FORMAT_GROUP) |
| 3872 | ret = perf_event_read_group(event, read_format, buf); |
| 3873 | else |
| 3874 | ret = perf_event_read_one(event, read_format, buf); |
| 3875 | |
| 3876 | return ret; |
| 3877 | } |
| 3878 | |
| 3879 | static ssize_t |
| 3880 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) |
| 3881 | { |
| 3882 | struct perf_event *event = file->private_data; |
| 3883 | struct perf_event_context *ctx; |
| 3884 | int ret; |
| 3885 | |
| 3886 | ctx = perf_event_ctx_lock(event); |
| 3887 | ret = perf_read_hw(event, buf, count); |
| 3888 | perf_event_ctx_unlock(event, ctx); |
| 3889 | |
| 3890 | return ret; |
| 3891 | } |
| 3892 | |
| 3893 | static unsigned int perf_poll(struct file *file, poll_table *wait) |
| 3894 | { |
| 3895 | struct perf_event *event = file->private_data; |
| 3896 | struct ring_buffer *rb; |
| 3897 | unsigned int events = POLLHUP; |
| 3898 | |
| 3899 | poll_wait(file, &event->waitq, wait); |
| 3900 | |
| 3901 | if (is_event_hup(event)) |
| 3902 | return events; |
| 3903 | |
| 3904 | /* |
| 3905 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
| 3906 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. |
| 3907 | */ |
| 3908 | mutex_lock(&event->mmap_mutex); |
| 3909 | rb = event->rb; |
| 3910 | if (rb) |
| 3911 | events = atomic_xchg(&rb->poll, 0); |
| 3912 | mutex_unlock(&event->mmap_mutex); |
| 3913 | return events; |
| 3914 | } |
| 3915 | |
| 3916 | static void _perf_event_reset(struct perf_event *event) |
| 3917 | { |
| 3918 | (void)perf_event_read(event); |
| 3919 | local64_set(&event->count, 0); |
| 3920 | perf_event_update_userpage(event); |
| 3921 | } |
| 3922 | |
| 3923 | /* |
| 3924 | * Holding the top-level event's child_mutex means that any |
| 3925 | * descendant process that has inherited this event will block |
| 3926 | * in sync_child_event if it goes to exit, thus satisfying the |
| 3927 | * task existence requirements of perf_event_enable/disable. |
| 3928 | */ |
| 3929 | static void perf_event_for_each_child(struct perf_event *event, |
| 3930 | void (*func)(struct perf_event *)) |
| 3931 | { |
| 3932 | struct perf_event *child; |
| 3933 | |
| 3934 | WARN_ON_ONCE(event->ctx->parent_ctx); |
| 3935 | |
| 3936 | mutex_lock(&event->child_mutex); |
| 3937 | func(event); |
| 3938 | list_for_each_entry(child, &event->child_list, child_list) |
| 3939 | func(child); |
| 3940 | mutex_unlock(&event->child_mutex); |
| 3941 | } |
| 3942 | |
| 3943 | static void perf_event_for_each(struct perf_event *event, |
| 3944 | void (*func)(struct perf_event *)) |
| 3945 | { |
| 3946 | struct perf_event_context *ctx = event->ctx; |
| 3947 | struct perf_event *sibling; |
| 3948 | |
| 3949 | lockdep_assert_held(&ctx->mutex); |
| 3950 | |
| 3951 | event = event->group_leader; |
| 3952 | |
| 3953 | perf_event_for_each_child(event, func); |
| 3954 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
| 3955 | perf_event_for_each_child(sibling, func); |
| 3956 | } |
| 3957 | |
| 3958 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
| 3959 | { |
| 3960 | struct perf_event_context *ctx = event->ctx; |
| 3961 | int ret = 0, active; |
| 3962 | u64 value; |
| 3963 | |
| 3964 | if (!is_sampling_event(event)) |
| 3965 | return -EINVAL; |
| 3966 | |
| 3967 | if (copy_from_user(&value, arg, sizeof(value))) |
| 3968 | return -EFAULT; |
| 3969 | |
| 3970 | if (!value) |
| 3971 | return -EINVAL; |
| 3972 | |
| 3973 | raw_spin_lock_irq(&ctx->lock); |
| 3974 | if (event->attr.freq) { |
| 3975 | if (value > sysctl_perf_event_sample_rate) { |
| 3976 | ret = -EINVAL; |
| 3977 | goto unlock; |
| 3978 | } |
| 3979 | |
| 3980 | event->attr.sample_freq = value; |
| 3981 | } else { |
| 3982 | event->attr.sample_period = value; |
| 3983 | event->hw.sample_period = value; |
| 3984 | } |
| 3985 | |
| 3986 | active = (event->state == PERF_EVENT_STATE_ACTIVE); |
| 3987 | if (active) { |
| 3988 | perf_pmu_disable(ctx->pmu); |
| 3989 | event->pmu->stop(event, PERF_EF_UPDATE); |
| 3990 | } |
| 3991 | |
| 3992 | local64_set(&event->hw.period_left, 0); |
| 3993 | |
| 3994 | if (active) { |
| 3995 | event->pmu->start(event, PERF_EF_RELOAD); |
| 3996 | perf_pmu_enable(ctx->pmu); |
| 3997 | } |
| 3998 | |
| 3999 | unlock: |
| 4000 | raw_spin_unlock_irq(&ctx->lock); |
| 4001 | |
| 4002 | return ret; |
| 4003 | } |
| 4004 | |
| 4005 | static const struct file_operations perf_fops; |
| 4006 | |
| 4007 | static inline int perf_fget_light(int fd, struct fd *p) |
| 4008 | { |
| 4009 | struct fd f = fdget(fd); |
| 4010 | if (!f.file) |
| 4011 | return -EBADF; |
| 4012 | |
| 4013 | if (f.file->f_op != &perf_fops) { |
| 4014 | fdput(f); |
| 4015 | return -EBADF; |
| 4016 | } |
| 4017 | *p = f; |
| 4018 | return 0; |
| 4019 | } |
| 4020 | |
| 4021 | static int perf_event_set_output(struct perf_event *event, |
| 4022 | struct perf_event *output_event); |
| 4023 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
| 4024 | static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd); |
| 4025 | |
| 4026 | static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg) |
| 4027 | { |
| 4028 | void (*func)(struct perf_event *); |
| 4029 | u32 flags = arg; |
| 4030 | |
| 4031 | switch (cmd) { |
| 4032 | case PERF_EVENT_IOC_ENABLE: |
| 4033 | func = _perf_event_enable; |
| 4034 | break; |
| 4035 | case PERF_EVENT_IOC_DISABLE: |
| 4036 | func = _perf_event_disable; |
| 4037 | break; |
| 4038 | case PERF_EVENT_IOC_RESET: |
| 4039 | func = _perf_event_reset; |
| 4040 | break; |
| 4041 | |
| 4042 | case PERF_EVENT_IOC_REFRESH: |
| 4043 | return _perf_event_refresh(event, arg); |
| 4044 | |
| 4045 | case PERF_EVENT_IOC_PERIOD: |
| 4046 | return perf_event_period(event, (u64 __user *)arg); |
| 4047 | |
| 4048 | case PERF_EVENT_IOC_ID: |
| 4049 | { |
| 4050 | u64 id = primary_event_id(event); |
| 4051 | |
| 4052 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) |
| 4053 | return -EFAULT; |
| 4054 | return 0; |
| 4055 | } |
| 4056 | |
| 4057 | case PERF_EVENT_IOC_SET_OUTPUT: |
| 4058 | { |
| 4059 | int ret; |
| 4060 | if (arg != -1) { |
| 4061 | struct perf_event *output_event; |
| 4062 | struct fd output; |
| 4063 | ret = perf_fget_light(arg, &output); |
| 4064 | if (ret) |
| 4065 | return ret; |
| 4066 | output_event = output.file->private_data; |
| 4067 | ret = perf_event_set_output(event, output_event); |
| 4068 | fdput(output); |
| 4069 | } else { |
| 4070 | ret = perf_event_set_output(event, NULL); |
| 4071 | } |
| 4072 | return ret; |
| 4073 | } |
| 4074 | |
| 4075 | case PERF_EVENT_IOC_SET_FILTER: |
| 4076 | return perf_event_set_filter(event, (void __user *)arg); |
| 4077 | |
| 4078 | case PERF_EVENT_IOC_SET_BPF: |
| 4079 | return perf_event_set_bpf_prog(event, arg); |
| 4080 | |
| 4081 | default: |
| 4082 | return -ENOTTY; |
| 4083 | } |
| 4084 | |
| 4085 | if (flags & PERF_IOC_FLAG_GROUP) |
| 4086 | perf_event_for_each(event, func); |
| 4087 | else |
| 4088 | perf_event_for_each_child(event, func); |
| 4089 | |
| 4090 | return 0; |
| 4091 | } |
| 4092 | |
| 4093 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
| 4094 | { |
| 4095 | struct perf_event *event = file->private_data; |
| 4096 | struct perf_event_context *ctx; |
| 4097 | long ret; |
| 4098 | |
| 4099 | ctx = perf_event_ctx_lock(event); |
| 4100 | ret = _perf_ioctl(event, cmd, arg); |
| 4101 | perf_event_ctx_unlock(event, ctx); |
| 4102 | |
| 4103 | return ret; |
| 4104 | } |
| 4105 | |
| 4106 | #ifdef CONFIG_COMPAT |
| 4107 | static long perf_compat_ioctl(struct file *file, unsigned int cmd, |
| 4108 | unsigned long arg) |
| 4109 | { |
| 4110 | switch (_IOC_NR(cmd)) { |
| 4111 | case _IOC_NR(PERF_EVENT_IOC_SET_FILTER): |
| 4112 | case _IOC_NR(PERF_EVENT_IOC_ID): |
| 4113 | /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ |
| 4114 | if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) { |
| 4115 | cmd &= ~IOCSIZE_MASK; |
| 4116 | cmd |= sizeof(void *) << IOCSIZE_SHIFT; |
| 4117 | } |
| 4118 | break; |
| 4119 | } |
| 4120 | return perf_ioctl(file, cmd, arg); |
| 4121 | } |
| 4122 | #else |
| 4123 | # define perf_compat_ioctl NULL |
| 4124 | #endif |
| 4125 | |
| 4126 | int perf_event_task_enable(void) |
| 4127 | { |
| 4128 | struct perf_event_context *ctx; |
| 4129 | struct perf_event *event; |
| 4130 | |
| 4131 | mutex_lock(¤t->perf_event_mutex); |
| 4132 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
| 4133 | ctx = perf_event_ctx_lock(event); |
| 4134 | perf_event_for_each_child(event, _perf_event_enable); |
| 4135 | perf_event_ctx_unlock(event, ctx); |
| 4136 | } |
| 4137 | mutex_unlock(¤t->perf_event_mutex); |
| 4138 | |
| 4139 | return 0; |
| 4140 | } |
| 4141 | |
| 4142 | int perf_event_task_disable(void) |
| 4143 | { |
| 4144 | struct perf_event_context *ctx; |
| 4145 | struct perf_event *event; |
| 4146 | |
| 4147 | mutex_lock(¤t->perf_event_mutex); |
| 4148 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
| 4149 | ctx = perf_event_ctx_lock(event); |
| 4150 | perf_event_for_each_child(event, _perf_event_disable); |
| 4151 | perf_event_ctx_unlock(event, ctx); |
| 4152 | } |
| 4153 | mutex_unlock(¤t->perf_event_mutex); |
| 4154 | |
| 4155 | return 0; |
| 4156 | } |
| 4157 | |
| 4158 | static int perf_event_index(struct perf_event *event) |
| 4159 | { |
| 4160 | if (event->hw.state & PERF_HES_STOPPED) |
| 4161 | return 0; |
| 4162 | |
| 4163 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
| 4164 | return 0; |
| 4165 | |
| 4166 | return event->pmu->event_idx(event); |
| 4167 | } |
| 4168 | |
| 4169 | static void calc_timer_values(struct perf_event *event, |
| 4170 | u64 *now, |
| 4171 | u64 *enabled, |
| 4172 | u64 *running) |
| 4173 | { |
| 4174 | u64 ctx_time; |
| 4175 | |
| 4176 | *now = perf_clock(); |
| 4177 | ctx_time = event->shadow_ctx_time + *now; |
| 4178 | *enabled = ctx_time - event->tstamp_enabled; |
| 4179 | *running = ctx_time - event->tstamp_running; |
| 4180 | } |
| 4181 | |
| 4182 | static void perf_event_init_userpage(struct perf_event *event) |
| 4183 | { |
| 4184 | struct perf_event_mmap_page *userpg; |
| 4185 | struct ring_buffer *rb; |
| 4186 | |
| 4187 | rcu_read_lock(); |
| 4188 | rb = rcu_dereference(event->rb); |
| 4189 | if (!rb) |
| 4190 | goto unlock; |
| 4191 | |
| 4192 | userpg = rb->user_page; |
| 4193 | |
| 4194 | /* Allow new userspace to detect that bit 0 is deprecated */ |
| 4195 | userpg->cap_bit0_is_deprecated = 1; |
| 4196 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); |
| 4197 | userpg->data_offset = PAGE_SIZE; |
| 4198 | userpg->data_size = perf_data_size(rb); |
| 4199 | |
| 4200 | unlock: |
| 4201 | rcu_read_unlock(); |
| 4202 | } |
| 4203 | |
| 4204 | void __weak arch_perf_update_userpage( |
| 4205 | struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now) |
| 4206 | { |
| 4207 | } |
| 4208 | |
| 4209 | /* |
| 4210 | * Callers need to ensure there can be no nesting of this function, otherwise |
| 4211 | * the seqlock logic goes bad. We can not serialize this because the arch |
| 4212 | * code calls this from NMI context. |
| 4213 | */ |
| 4214 | void perf_event_update_userpage(struct perf_event *event) |
| 4215 | { |
| 4216 | struct perf_event_mmap_page *userpg; |
| 4217 | struct ring_buffer *rb; |
| 4218 | u64 enabled, running, now; |
| 4219 | |
| 4220 | rcu_read_lock(); |
| 4221 | rb = rcu_dereference(event->rb); |
| 4222 | if (!rb) |
| 4223 | goto unlock; |
| 4224 | |
| 4225 | /* |
| 4226 | * compute total_time_enabled, total_time_running |
| 4227 | * based on snapshot values taken when the event |
| 4228 | * was last scheduled in. |
| 4229 | * |
| 4230 | * we cannot simply called update_context_time() |
| 4231 | * because of locking issue as we can be called in |
| 4232 | * NMI context |
| 4233 | */ |
| 4234 | calc_timer_values(event, &now, &enabled, &running); |
| 4235 | |
| 4236 | userpg = rb->user_page; |
| 4237 | /* |
| 4238 | * Disable preemption so as to not let the corresponding user-space |
| 4239 | * spin too long if we get preempted. |
| 4240 | */ |
| 4241 | preempt_disable(); |
| 4242 | ++userpg->lock; |
| 4243 | barrier(); |
| 4244 | userpg->index = perf_event_index(event); |
| 4245 | userpg->offset = perf_event_count(event); |
| 4246 | if (userpg->index) |
| 4247 | userpg->offset -= local64_read(&event->hw.prev_count); |
| 4248 | |
| 4249 | userpg->time_enabled = enabled + |
| 4250 | atomic64_read(&event->child_total_time_enabled); |
| 4251 | |
| 4252 | userpg->time_running = running + |
| 4253 | atomic64_read(&event->child_total_time_running); |
| 4254 | |
| 4255 | arch_perf_update_userpage(event, userpg, now); |
| 4256 | |
| 4257 | barrier(); |
| 4258 | ++userpg->lock; |
| 4259 | preempt_enable(); |
| 4260 | unlock: |
| 4261 | rcu_read_unlock(); |
| 4262 | } |
| 4263 | |
| 4264 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
| 4265 | { |
| 4266 | struct perf_event *event = vma->vm_file->private_data; |
| 4267 | struct ring_buffer *rb; |
| 4268 | int ret = VM_FAULT_SIGBUS; |
| 4269 | |
| 4270 | if (vmf->flags & FAULT_FLAG_MKWRITE) { |
| 4271 | if (vmf->pgoff == 0) |
| 4272 | ret = 0; |
| 4273 | return ret; |
| 4274 | } |
| 4275 | |
| 4276 | rcu_read_lock(); |
| 4277 | rb = rcu_dereference(event->rb); |
| 4278 | if (!rb) |
| 4279 | goto unlock; |
| 4280 | |
| 4281 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) |
| 4282 | goto unlock; |
| 4283 | |
| 4284 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
| 4285 | if (!vmf->page) |
| 4286 | goto unlock; |
| 4287 | |
| 4288 | get_page(vmf->page); |
| 4289 | vmf->page->mapping = vma->vm_file->f_mapping; |
| 4290 | vmf->page->index = vmf->pgoff; |
| 4291 | |
| 4292 | ret = 0; |
| 4293 | unlock: |
| 4294 | rcu_read_unlock(); |
| 4295 | |
| 4296 | return ret; |
| 4297 | } |
| 4298 | |
| 4299 | static void ring_buffer_attach(struct perf_event *event, |
| 4300 | struct ring_buffer *rb) |
| 4301 | { |
| 4302 | struct ring_buffer *old_rb = NULL; |
| 4303 | unsigned long flags; |
| 4304 | |
| 4305 | if (event->rb) { |
| 4306 | /* |
| 4307 | * Should be impossible, we set this when removing |
| 4308 | * event->rb_entry and wait/clear when adding event->rb_entry. |
| 4309 | */ |
| 4310 | WARN_ON_ONCE(event->rcu_pending); |
| 4311 | |
| 4312 | old_rb = event->rb; |
| 4313 | event->rcu_batches = get_state_synchronize_rcu(); |
| 4314 | event->rcu_pending = 1; |
| 4315 | |
| 4316 | spin_lock_irqsave(&old_rb->event_lock, flags); |
| 4317 | list_del_rcu(&event->rb_entry); |
| 4318 | spin_unlock_irqrestore(&old_rb->event_lock, flags); |
| 4319 | } |
| 4320 | |
| 4321 | if (event->rcu_pending && rb) { |
| 4322 | cond_synchronize_rcu(event->rcu_batches); |
| 4323 | event->rcu_pending = 0; |
| 4324 | } |
| 4325 | |
| 4326 | if (rb) { |
| 4327 | spin_lock_irqsave(&rb->event_lock, flags); |
| 4328 | list_add_rcu(&event->rb_entry, &rb->event_list); |
| 4329 | spin_unlock_irqrestore(&rb->event_lock, flags); |
| 4330 | } |
| 4331 | |
| 4332 | rcu_assign_pointer(event->rb, rb); |
| 4333 | |
| 4334 | if (old_rb) { |
| 4335 | ring_buffer_put(old_rb); |
| 4336 | /* |
| 4337 | * Since we detached before setting the new rb, so that we |
| 4338 | * could attach the new rb, we could have missed a wakeup. |
| 4339 | * Provide it now. |
| 4340 | */ |
| 4341 | wake_up_all(&event->waitq); |
| 4342 | } |
| 4343 | } |
| 4344 | |
| 4345 | static void ring_buffer_wakeup(struct perf_event *event) |
| 4346 | { |
| 4347 | struct ring_buffer *rb; |
| 4348 | |
| 4349 | rcu_read_lock(); |
| 4350 | rb = rcu_dereference(event->rb); |
| 4351 | if (rb) { |
| 4352 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) |
| 4353 | wake_up_all(&event->waitq); |
| 4354 | } |
| 4355 | rcu_read_unlock(); |
| 4356 | } |
| 4357 | |
| 4358 | static void rb_free_rcu(struct rcu_head *rcu_head) |
| 4359 | { |
| 4360 | struct ring_buffer *rb; |
| 4361 | |
| 4362 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
| 4363 | rb_free(rb); |
| 4364 | } |
| 4365 | |
| 4366 | struct ring_buffer *ring_buffer_get(struct perf_event *event) |
| 4367 | { |
| 4368 | struct ring_buffer *rb; |
| 4369 | |
| 4370 | rcu_read_lock(); |
| 4371 | rb = rcu_dereference(event->rb); |
| 4372 | if (rb) { |
| 4373 | if (!atomic_inc_not_zero(&rb->refcount)) |
| 4374 | rb = NULL; |
| 4375 | } |
| 4376 | rcu_read_unlock(); |
| 4377 | |
| 4378 | return rb; |
| 4379 | } |
| 4380 | |
| 4381 | void ring_buffer_put(struct ring_buffer *rb) |
| 4382 | { |
| 4383 | if (!atomic_dec_and_test(&rb->refcount)) |
| 4384 | return; |
| 4385 | |
| 4386 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
| 4387 | |
| 4388 | call_rcu(&rb->rcu_head, rb_free_rcu); |
| 4389 | } |
| 4390 | |
| 4391 | static void perf_mmap_open(struct vm_area_struct *vma) |
| 4392 | { |
| 4393 | struct perf_event *event = vma->vm_file->private_data; |
| 4394 | |
| 4395 | atomic_inc(&event->mmap_count); |
| 4396 | atomic_inc(&event->rb->mmap_count); |
| 4397 | |
| 4398 | if (vma->vm_pgoff) |
| 4399 | atomic_inc(&event->rb->aux_mmap_count); |
| 4400 | |
| 4401 | if (event->pmu->event_mapped) |
| 4402 | event->pmu->event_mapped(event); |
| 4403 | } |
| 4404 | |
| 4405 | /* |
| 4406 | * A buffer can be mmap()ed multiple times; either directly through the same |
| 4407 | * event, or through other events by use of perf_event_set_output(). |
| 4408 | * |
| 4409 | * In order to undo the VM accounting done by perf_mmap() we need to destroy |
| 4410 | * the buffer here, where we still have a VM context. This means we need |
| 4411 | * to detach all events redirecting to us. |
| 4412 | */ |
| 4413 | static void perf_mmap_close(struct vm_area_struct *vma) |
| 4414 | { |
| 4415 | struct perf_event *event = vma->vm_file->private_data; |
| 4416 | |
| 4417 | struct ring_buffer *rb = ring_buffer_get(event); |
| 4418 | struct user_struct *mmap_user = rb->mmap_user; |
| 4419 | int mmap_locked = rb->mmap_locked; |
| 4420 | unsigned long size = perf_data_size(rb); |
| 4421 | |
| 4422 | if (event->pmu->event_unmapped) |
| 4423 | event->pmu->event_unmapped(event); |
| 4424 | |
| 4425 | /* |
| 4426 | * rb->aux_mmap_count will always drop before rb->mmap_count and |
| 4427 | * event->mmap_count, so it is ok to use event->mmap_mutex to |
| 4428 | * serialize with perf_mmap here. |
| 4429 | */ |
| 4430 | if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff && |
| 4431 | atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) { |
| 4432 | atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm); |
| 4433 | vma->vm_mm->pinned_vm -= rb->aux_mmap_locked; |
| 4434 | |
| 4435 | rb_free_aux(rb); |
| 4436 | mutex_unlock(&event->mmap_mutex); |
| 4437 | } |
| 4438 | |
| 4439 | atomic_dec(&rb->mmap_count); |
| 4440 | |
| 4441 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) |
| 4442 | goto out_put; |
| 4443 | |
| 4444 | ring_buffer_attach(event, NULL); |
| 4445 | mutex_unlock(&event->mmap_mutex); |
| 4446 | |
| 4447 | /* If there's still other mmap()s of this buffer, we're done. */ |
| 4448 | if (atomic_read(&rb->mmap_count)) |
| 4449 | goto out_put; |
| 4450 | |
| 4451 | /* |
| 4452 | * No other mmap()s, detach from all other events that might redirect |
| 4453 | * into the now unreachable buffer. Somewhat complicated by the |
| 4454 | * fact that rb::event_lock otherwise nests inside mmap_mutex. |
| 4455 | */ |
| 4456 | again: |
| 4457 | rcu_read_lock(); |
| 4458 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { |
| 4459 | if (!atomic_long_inc_not_zero(&event->refcount)) { |
| 4460 | /* |
| 4461 | * This event is en-route to free_event() which will |
| 4462 | * detach it and remove it from the list. |
| 4463 | */ |
| 4464 | continue; |
| 4465 | } |
| 4466 | rcu_read_unlock(); |
| 4467 | |
| 4468 | mutex_lock(&event->mmap_mutex); |
| 4469 | /* |
| 4470 | * Check we didn't race with perf_event_set_output() which can |
| 4471 | * swizzle the rb from under us while we were waiting to |
| 4472 | * acquire mmap_mutex. |
| 4473 | * |
| 4474 | * If we find a different rb; ignore this event, a next |
| 4475 | * iteration will no longer find it on the list. We have to |
| 4476 | * still restart the iteration to make sure we're not now |
| 4477 | * iterating the wrong list. |
| 4478 | */ |
| 4479 | if (event->rb == rb) |
| 4480 | ring_buffer_attach(event, NULL); |
| 4481 | |
| 4482 | mutex_unlock(&event->mmap_mutex); |
| 4483 | put_event(event); |
| 4484 | |
| 4485 | /* |
| 4486 | * Restart the iteration; either we're on the wrong list or |
| 4487 | * destroyed its integrity by doing a deletion. |
| 4488 | */ |
| 4489 | goto again; |
| 4490 | } |
| 4491 | rcu_read_unlock(); |
| 4492 | |
| 4493 | /* |
| 4494 | * It could be there's still a few 0-ref events on the list; they'll |
| 4495 | * get cleaned up by free_event() -- they'll also still have their |
| 4496 | * ref on the rb and will free it whenever they are done with it. |
| 4497 | * |
| 4498 | * Aside from that, this buffer is 'fully' detached and unmapped, |
| 4499 | * undo the VM accounting. |
| 4500 | */ |
| 4501 | |
| 4502 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); |
| 4503 | vma->vm_mm->pinned_vm -= mmap_locked; |
| 4504 | free_uid(mmap_user); |
| 4505 | |
| 4506 | out_put: |
| 4507 | ring_buffer_put(rb); /* could be last */ |
| 4508 | } |
| 4509 | |
| 4510 | static const struct vm_operations_struct perf_mmap_vmops = { |
| 4511 | .open = perf_mmap_open, |
| 4512 | .close = perf_mmap_close, /* non mergable */ |
| 4513 | .fault = perf_mmap_fault, |
| 4514 | .page_mkwrite = perf_mmap_fault, |
| 4515 | }; |
| 4516 | |
| 4517 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) |
| 4518 | { |
| 4519 | struct perf_event *event = file->private_data; |
| 4520 | unsigned long user_locked, user_lock_limit; |
| 4521 | struct user_struct *user = current_user(); |
| 4522 | unsigned long locked, lock_limit; |
| 4523 | struct ring_buffer *rb = NULL; |
| 4524 | unsigned long vma_size; |
| 4525 | unsigned long nr_pages; |
| 4526 | long user_extra = 0, extra = 0; |
| 4527 | int ret = 0, flags = 0; |
| 4528 | |
| 4529 | /* |
| 4530 | * Don't allow mmap() of inherited per-task counters. This would |
| 4531 | * create a performance issue due to all children writing to the |
| 4532 | * same rb. |
| 4533 | */ |
| 4534 | if (event->cpu == -1 && event->attr.inherit) |
| 4535 | return -EINVAL; |
| 4536 | |
| 4537 | if (!(vma->vm_flags & VM_SHARED)) |
| 4538 | return -EINVAL; |
| 4539 | |
| 4540 | vma_size = vma->vm_end - vma->vm_start; |
| 4541 | |
| 4542 | if (vma->vm_pgoff == 0) { |
| 4543 | nr_pages = (vma_size / PAGE_SIZE) - 1; |
| 4544 | } else { |
| 4545 | /* |
| 4546 | * AUX area mapping: if rb->aux_nr_pages != 0, it's already |
| 4547 | * mapped, all subsequent mappings should have the same size |
| 4548 | * and offset. Must be above the normal perf buffer. |
| 4549 | */ |
| 4550 | u64 aux_offset, aux_size; |
| 4551 | |
| 4552 | if (!event->rb) |
| 4553 | return -EINVAL; |
| 4554 | |
| 4555 | nr_pages = vma_size / PAGE_SIZE; |
| 4556 | |
| 4557 | mutex_lock(&event->mmap_mutex); |
| 4558 | ret = -EINVAL; |
| 4559 | |
| 4560 | rb = event->rb; |
| 4561 | if (!rb) |
| 4562 | goto aux_unlock; |
| 4563 | |
| 4564 | aux_offset = ACCESS_ONCE(rb->user_page->aux_offset); |
| 4565 | aux_size = ACCESS_ONCE(rb->user_page->aux_size); |
| 4566 | |
| 4567 | if (aux_offset < perf_data_size(rb) + PAGE_SIZE) |
| 4568 | goto aux_unlock; |
| 4569 | |
| 4570 | if (aux_offset != vma->vm_pgoff << PAGE_SHIFT) |
| 4571 | goto aux_unlock; |
| 4572 | |
| 4573 | /* already mapped with a different offset */ |
| 4574 | if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff) |
| 4575 | goto aux_unlock; |
| 4576 | |
| 4577 | if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE) |
| 4578 | goto aux_unlock; |
| 4579 | |
| 4580 | /* already mapped with a different size */ |
| 4581 | if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages) |
| 4582 | goto aux_unlock; |
| 4583 | |
| 4584 | if (!is_power_of_2(nr_pages)) |
| 4585 | goto aux_unlock; |
| 4586 | |
| 4587 | if (!atomic_inc_not_zero(&rb->mmap_count)) |
| 4588 | goto aux_unlock; |
| 4589 | |
| 4590 | if (rb_has_aux(rb)) { |
| 4591 | atomic_inc(&rb->aux_mmap_count); |
| 4592 | ret = 0; |
| 4593 | goto unlock; |
| 4594 | } |
| 4595 | |
| 4596 | atomic_set(&rb->aux_mmap_count, 1); |
| 4597 | user_extra = nr_pages; |
| 4598 | |
| 4599 | goto accounting; |
| 4600 | } |
| 4601 | |
| 4602 | /* |
| 4603 | * If we have rb pages ensure they're a power-of-two number, so we |
| 4604 | * can do bitmasks instead of modulo. |
| 4605 | */ |
| 4606 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) |
| 4607 | return -EINVAL; |
| 4608 | |
| 4609 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
| 4610 | return -EINVAL; |
| 4611 | |
| 4612 | WARN_ON_ONCE(event->ctx->parent_ctx); |
| 4613 | again: |
| 4614 | mutex_lock(&event->mmap_mutex); |
| 4615 | if (event->rb) { |
| 4616 | if (event->rb->nr_pages != nr_pages) { |
| 4617 | ret = -EINVAL; |
| 4618 | goto unlock; |
| 4619 | } |
| 4620 | |
| 4621 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { |
| 4622 | /* |
| 4623 | * Raced against perf_mmap_close() through |
| 4624 | * perf_event_set_output(). Try again, hope for better |
| 4625 | * luck. |
| 4626 | */ |
| 4627 | mutex_unlock(&event->mmap_mutex); |
| 4628 | goto again; |
| 4629 | } |
| 4630 | |
| 4631 | goto unlock; |
| 4632 | } |
| 4633 | |
| 4634 | user_extra = nr_pages + 1; |
| 4635 | |
| 4636 | accounting: |
| 4637 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
| 4638 | |
| 4639 | /* |
| 4640 | * Increase the limit linearly with more CPUs: |
| 4641 | */ |
| 4642 | user_lock_limit *= num_online_cpus(); |
| 4643 | |
| 4644 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
| 4645 | |
| 4646 | if (user_locked > user_lock_limit) |
| 4647 | extra = user_locked - user_lock_limit; |
| 4648 | |
| 4649 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
| 4650 | lock_limit >>= PAGE_SHIFT; |
| 4651 | locked = vma->vm_mm->pinned_vm + extra; |
| 4652 | |
| 4653 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
| 4654 | !capable(CAP_IPC_LOCK)) { |
| 4655 | ret = -EPERM; |
| 4656 | goto unlock; |
| 4657 | } |
| 4658 | |
| 4659 | WARN_ON(!rb && event->rb); |
| 4660 | |
| 4661 | if (vma->vm_flags & VM_WRITE) |
| 4662 | flags |= RING_BUFFER_WRITABLE; |
| 4663 | |
| 4664 | if (!rb) { |
| 4665 | rb = rb_alloc(nr_pages, |
| 4666 | event->attr.watermark ? event->attr.wakeup_watermark : 0, |
| 4667 | event->cpu, flags); |
| 4668 | |
| 4669 | if (!rb) { |
| 4670 | ret = -ENOMEM; |
| 4671 | goto unlock; |
| 4672 | } |
| 4673 | |
| 4674 | atomic_set(&rb->mmap_count, 1); |
| 4675 | rb->mmap_user = get_current_user(); |
| 4676 | rb->mmap_locked = extra; |
| 4677 | |
| 4678 | ring_buffer_attach(event, rb); |
| 4679 | |
| 4680 | perf_event_init_userpage(event); |
| 4681 | perf_event_update_userpage(event); |
| 4682 | } else { |
| 4683 | ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages, |
| 4684 | event->attr.aux_watermark, flags); |
| 4685 | if (!ret) |
| 4686 | rb->aux_mmap_locked = extra; |
| 4687 | } |
| 4688 | |
| 4689 | unlock: |
| 4690 | if (!ret) { |
| 4691 | atomic_long_add(user_extra, &user->locked_vm); |
| 4692 | vma->vm_mm->pinned_vm += extra; |
| 4693 | |
| 4694 | atomic_inc(&event->mmap_count); |
| 4695 | } else if (rb) { |
| 4696 | atomic_dec(&rb->mmap_count); |
| 4697 | } |
| 4698 | aux_unlock: |
| 4699 | mutex_unlock(&event->mmap_mutex); |
| 4700 | |
| 4701 | /* |
| 4702 | * Since pinned accounting is per vm we cannot allow fork() to copy our |
| 4703 | * vma. |
| 4704 | */ |
| 4705 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
| 4706 | vma->vm_ops = &perf_mmap_vmops; |
| 4707 | |
| 4708 | if (event->pmu->event_mapped) |
| 4709 | event->pmu->event_mapped(event); |
| 4710 | |
| 4711 | return ret; |
| 4712 | } |
| 4713 | |
| 4714 | static int perf_fasync(int fd, struct file *filp, int on) |
| 4715 | { |
| 4716 | struct inode *inode = file_inode(filp); |
| 4717 | struct perf_event *event = filp->private_data; |
| 4718 | int retval; |
| 4719 | |
| 4720 | mutex_lock(&inode->i_mutex); |
| 4721 | retval = fasync_helper(fd, filp, on, &event->fasync); |
| 4722 | mutex_unlock(&inode->i_mutex); |
| 4723 | |
| 4724 | if (retval < 0) |
| 4725 | return retval; |
| 4726 | |
| 4727 | return 0; |
| 4728 | } |
| 4729 | |
| 4730 | static const struct file_operations perf_fops = { |
| 4731 | .llseek = no_llseek, |
| 4732 | .release = perf_release, |
| 4733 | .read = perf_read, |
| 4734 | .poll = perf_poll, |
| 4735 | .unlocked_ioctl = perf_ioctl, |
| 4736 | .compat_ioctl = perf_compat_ioctl, |
| 4737 | .mmap = perf_mmap, |
| 4738 | .fasync = perf_fasync, |
| 4739 | }; |
| 4740 | |
| 4741 | /* |
| 4742 | * Perf event wakeup |
| 4743 | * |
| 4744 | * If there's data, ensure we set the poll() state and publish everything |
| 4745 | * to user-space before waking everybody up. |
| 4746 | */ |
| 4747 | |
| 4748 | void perf_event_wakeup(struct perf_event *event) |
| 4749 | { |
| 4750 | ring_buffer_wakeup(event); |
| 4751 | |
| 4752 | if (event->pending_kill) { |
| 4753 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); |
| 4754 | event->pending_kill = 0; |
| 4755 | } |
| 4756 | } |
| 4757 | |
| 4758 | static void perf_pending_event(struct irq_work *entry) |
| 4759 | { |
| 4760 | struct perf_event *event = container_of(entry, |
| 4761 | struct perf_event, pending); |
| 4762 | int rctx; |
| 4763 | |
| 4764 | rctx = perf_swevent_get_recursion_context(); |
| 4765 | /* |
| 4766 | * If we 'fail' here, that's OK, it means recursion is already disabled |
| 4767 | * and we won't recurse 'further'. |
| 4768 | */ |
| 4769 | |
| 4770 | if (event->pending_disable) { |
| 4771 | event->pending_disable = 0; |
| 4772 | __perf_event_disable(event); |
| 4773 | } |
| 4774 | |
| 4775 | if (event->pending_wakeup) { |
| 4776 | event->pending_wakeup = 0; |
| 4777 | perf_event_wakeup(event); |
| 4778 | } |
| 4779 | |
| 4780 | if (rctx >= 0) |
| 4781 | perf_swevent_put_recursion_context(rctx); |
| 4782 | } |
| 4783 | |
| 4784 | /* |
| 4785 | * We assume there is only KVM supporting the callbacks. |
| 4786 | * Later on, we might change it to a list if there is |
| 4787 | * another virtualization implementation supporting the callbacks. |
| 4788 | */ |
| 4789 | struct perf_guest_info_callbacks *perf_guest_cbs; |
| 4790 | |
| 4791 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) |
| 4792 | { |
| 4793 | perf_guest_cbs = cbs; |
| 4794 | return 0; |
| 4795 | } |
| 4796 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); |
| 4797 | |
| 4798 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) |
| 4799 | { |
| 4800 | perf_guest_cbs = NULL; |
| 4801 | return 0; |
| 4802 | } |
| 4803 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); |
| 4804 | |
| 4805 | static void |
| 4806 | perf_output_sample_regs(struct perf_output_handle *handle, |
| 4807 | struct pt_regs *regs, u64 mask) |
| 4808 | { |
| 4809 | int bit; |
| 4810 | |
| 4811 | for_each_set_bit(bit, (const unsigned long *) &mask, |
| 4812 | sizeof(mask) * BITS_PER_BYTE) { |
| 4813 | u64 val; |
| 4814 | |
| 4815 | val = perf_reg_value(regs, bit); |
| 4816 | perf_output_put(handle, val); |
| 4817 | } |
| 4818 | } |
| 4819 | |
| 4820 | static void perf_sample_regs_user(struct perf_regs *regs_user, |
| 4821 | struct pt_regs *regs, |
| 4822 | struct pt_regs *regs_user_copy) |
| 4823 | { |
| 4824 | if (user_mode(regs)) { |
| 4825 | regs_user->abi = perf_reg_abi(current); |
| 4826 | regs_user->regs = regs; |
| 4827 | } else if (current->mm) { |
| 4828 | perf_get_regs_user(regs_user, regs, regs_user_copy); |
| 4829 | } else { |
| 4830 | regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE; |
| 4831 | regs_user->regs = NULL; |
| 4832 | } |
| 4833 | } |
| 4834 | |
| 4835 | static void perf_sample_regs_intr(struct perf_regs *regs_intr, |
| 4836 | struct pt_regs *regs) |
| 4837 | { |
| 4838 | regs_intr->regs = regs; |
| 4839 | regs_intr->abi = perf_reg_abi(current); |
| 4840 | } |
| 4841 | |
| 4842 | |
| 4843 | /* |
| 4844 | * Get remaining task size from user stack pointer. |
| 4845 | * |
| 4846 | * It'd be better to take stack vma map and limit this more |
| 4847 | * precisly, but there's no way to get it safely under interrupt, |
| 4848 | * so using TASK_SIZE as limit. |
| 4849 | */ |
| 4850 | static u64 perf_ustack_task_size(struct pt_regs *regs) |
| 4851 | { |
| 4852 | unsigned long addr = perf_user_stack_pointer(regs); |
| 4853 | |
| 4854 | if (!addr || addr >= TASK_SIZE) |
| 4855 | return 0; |
| 4856 | |
| 4857 | return TASK_SIZE - addr; |
| 4858 | } |
| 4859 | |
| 4860 | static u16 |
| 4861 | perf_sample_ustack_size(u16 stack_size, u16 header_size, |
| 4862 | struct pt_regs *regs) |
| 4863 | { |
| 4864 | u64 task_size; |
| 4865 | |
| 4866 | /* No regs, no stack pointer, no dump. */ |
| 4867 | if (!regs) |
| 4868 | return 0; |
| 4869 | |
| 4870 | /* |
| 4871 | * Check if we fit in with the requested stack size into the: |
| 4872 | * - TASK_SIZE |
| 4873 | * If we don't, we limit the size to the TASK_SIZE. |
| 4874 | * |
| 4875 | * - remaining sample size |
| 4876 | * If we don't, we customize the stack size to |
| 4877 | * fit in to the remaining sample size. |
| 4878 | */ |
| 4879 | |
| 4880 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); |
| 4881 | stack_size = min(stack_size, (u16) task_size); |
| 4882 | |
| 4883 | /* Current header size plus static size and dynamic size. */ |
| 4884 | header_size += 2 * sizeof(u64); |
| 4885 | |
| 4886 | /* Do we fit in with the current stack dump size? */ |
| 4887 | if ((u16) (header_size + stack_size) < header_size) { |
| 4888 | /* |
| 4889 | * If we overflow the maximum size for the sample, |
| 4890 | * we customize the stack dump size to fit in. |
| 4891 | */ |
| 4892 | stack_size = USHRT_MAX - header_size - sizeof(u64); |
| 4893 | stack_size = round_up(stack_size, sizeof(u64)); |
| 4894 | } |
| 4895 | |
| 4896 | return stack_size; |
| 4897 | } |
| 4898 | |
| 4899 | static void |
| 4900 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, |
| 4901 | struct pt_regs *regs) |
| 4902 | { |
| 4903 | /* Case of a kernel thread, nothing to dump */ |
| 4904 | if (!regs) { |
| 4905 | u64 size = 0; |
| 4906 | perf_output_put(handle, size); |
| 4907 | } else { |
| 4908 | unsigned long sp; |
| 4909 | unsigned int rem; |
| 4910 | u64 dyn_size; |
| 4911 | |
| 4912 | /* |
| 4913 | * We dump: |
| 4914 | * static size |
| 4915 | * - the size requested by user or the best one we can fit |
| 4916 | * in to the sample max size |
| 4917 | * data |
| 4918 | * - user stack dump data |
| 4919 | * dynamic size |
| 4920 | * - the actual dumped size |
| 4921 | */ |
| 4922 | |
| 4923 | /* Static size. */ |
| 4924 | perf_output_put(handle, dump_size); |
| 4925 | |
| 4926 | /* Data. */ |
| 4927 | sp = perf_user_stack_pointer(regs); |
| 4928 | rem = __output_copy_user(handle, (void *) sp, dump_size); |
| 4929 | dyn_size = dump_size - rem; |
| 4930 | |
| 4931 | perf_output_skip(handle, rem); |
| 4932 | |
| 4933 | /* Dynamic size. */ |
| 4934 | perf_output_put(handle, dyn_size); |
| 4935 | } |
| 4936 | } |
| 4937 | |
| 4938 | static void __perf_event_header__init_id(struct perf_event_header *header, |
| 4939 | struct perf_sample_data *data, |
| 4940 | struct perf_event *event) |
| 4941 | { |
| 4942 | u64 sample_type = event->attr.sample_type; |
| 4943 | |
| 4944 | data->type = sample_type; |
| 4945 | header->size += event->id_header_size; |
| 4946 | |
| 4947 | if (sample_type & PERF_SAMPLE_TID) { |
| 4948 | /* namespace issues */ |
| 4949 | data->tid_entry.pid = perf_event_pid(event, current); |
| 4950 | data->tid_entry.tid = perf_event_tid(event, current); |
| 4951 | } |
| 4952 | |
| 4953 | if (sample_type & PERF_SAMPLE_TIME) |
| 4954 | data->time = perf_event_clock(event); |
| 4955 | |
| 4956 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
| 4957 | data->id = primary_event_id(event); |
| 4958 | |
| 4959 | if (sample_type & PERF_SAMPLE_STREAM_ID) |
| 4960 | data->stream_id = event->id; |
| 4961 | |
| 4962 | if (sample_type & PERF_SAMPLE_CPU) { |
| 4963 | data->cpu_entry.cpu = raw_smp_processor_id(); |
| 4964 | data->cpu_entry.reserved = 0; |
| 4965 | } |
| 4966 | } |
| 4967 | |
| 4968 | void perf_event_header__init_id(struct perf_event_header *header, |
| 4969 | struct perf_sample_data *data, |
| 4970 | struct perf_event *event) |
| 4971 | { |
| 4972 | if (event->attr.sample_id_all) |
| 4973 | __perf_event_header__init_id(header, data, event); |
| 4974 | } |
| 4975 | |
| 4976 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, |
| 4977 | struct perf_sample_data *data) |
| 4978 | { |
| 4979 | u64 sample_type = data->type; |
| 4980 | |
| 4981 | if (sample_type & PERF_SAMPLE_TID) |
| 4982 | perf_output_put(handle, data->tid_entry); |
| 4983 | |
| 4984 | if (sample_type & PERF_SAMPLE_TIME) |
| 4985 | perf_output_put(handle, data->time); |
| 4986 | |
| 4987 | if (sample_type & PERF_SAMPLE_ID) |
| 4988 | perf_output_put(handle, data->id); |
| 4989 | |
| 4990 | if (sample_type & PERF_SAMPLE_STREAM_ID) |
| 4991 | perf_output_put(handle, data->stream_id); |
| 4992 | |
| 4993 | if (sample_type & PERF_SAMPLE_CPU) |
| 4994 | perf_output_put(handle, data->cpu_entry); |
| 4995 | |
| 4996 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
| 4997 | perf_output_put(handle, data->id); |
| 4998 | } |
| 4999 | |
| 5000 | void perf_event__output_id_sample(struct perf_event *event, |
| 5001 | struct perf_output_handle *handle, |
| 5002 | struct perf_sample_data *sample) |
| 5003 | { |
| 5004 | if (event->attr.sample_id_all) |
| 5005 | __perf_event__output_id_sample(handle, sample); |
| 5006 | } |
| 5007 | |
| 5008 | static void perf_output_read_one(struct perf_output_handle *handle, |
| 5009 | struct perf_event *event, |
| 5010 | u64 enabled, u64 running) |
| 5011 | { |
| 5012 | u64 read_format = event->attr.read_format; |
| 5013 | u64 values[4]; |
| 5014 | int n = 0; |
| 5015 | |
| 5016 | values[n++] = perf_event_count(event); |
| 5017 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
| 5018 | values[n++] = enabled + |
| 5019 | atomic64_read(&event->child_total_time_enabled); |
| 5020 | } |
| 5021 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { |
| 5022 | values[n++] = running + |
| 5023 | atomic64_read(&event->child_total_time_running); |
| 5024 | } |
| 5025 | if (read_format & PERF_FORMAT_ID) |
| 5026 | values[n++] = primary_event_id(event); |
| 5027 | |
| 5028 | __output_copy(handle, values, n * sizeof(u64)); |
| 5029 | } |
| 5030 | |
| 5031 | /* |
| 5032 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
| 5033 | */ |
| 5034 | static void perf_output_read_group(struct perf_output_handle *handle, |
| 5035 | struct perf_event *event, |
| 5036 | u64 enabled, u64 running) |
| 5037 | { |
| 5038 | struct perf_event *leader = event->group_leader, *sub; |
| 5039 | u64 read_format = event->attr.read_format; |
| 5040 | u64 values[5]; |
| 5041 | int n = 0; |
| 5042 | |
| 5043 | values[n++] = 1 + leader->nr_siblings; |
| 5044 | |
| 5045 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
| 5046 | values[n++] = enabled; |
| 5047 | |
| 5048 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
| 5049 | values[n++] = running; |
| 5050 | |
| 5051 | if (leader != event) |
| 5052 | leader->pmu->read(leader); |
| 5053 | |
| 5054 | values[n++] = perf_event_count(leader); |
| 5055 | if (read_format & PERF_FORMAT_ID) |
| 5056 | values[n++] = primary_event_id(leader); |
| 5057 | |
| 5058 | __output_copy(handle, values, n * sizeof(u64)); |
| 5059 | |
| 5060 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
| 5061 | n = 0; |
| 5062 | |
| 5063 | if ((sub != event) && |
| 5064 | (sub->state == PERF_EVENT_STATE_ACTIVE)) |
| 5065 | sub->pmu->read(sub); |
| 5066 | |
| 5067 | values[n++] = perf_event_count(sub); |
| 5068 | if (read_format & PERF_FORMAT_ID) |
| 5069 | values[n++] = primary_event_id(sub); |
| 5070 | |
| 5071 | __output_copy(handle, values, n * sizeof(u64)); |
| 5072 | } |
| 5073 | } |
| 5074 | |
| 5075 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
| 5076 | PERF_FORMAT_TOTAL_TIME_RUNNING) |
| 5077 | |
| 5078 | static void perf_output_read(struct perf_output_handle *handle, |
| 5079 | struct perf_event *event) |
| 5080 | { |
| 5081 | u64 enabled = 0, running = 0, now; |
| 5082 | u64 read_format = event->attr.read_format; |
| 5083 | |
| 5084 | /* |
| 5085 | * compute total_time_enabled, total_time_running |
| 5086 | * based on snapshot values taken when the event |
| 5087 | * was last scheduled in. |
| 5088 | * |
| 5089 | * we cannot simply called update_context_time() |
| 5090 | * because of locking issue as we are called in |
| 5091 | * NMI context |
| 5092 | */ |
| 5093 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
| 5094 | calc_timer_values(event, &now, &enabled, &running); |
| 5095 | |
| 5096 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
| 5097 | perf_output_read_group(handle, event, enabled, running); |
| 5098 | else |
| 5099 | perf_output_read_one(handle, event, enabled, running); |
| 5100 | } |
| 5101 | |
| 5102 | void perf_output_sample(struct perf_output_handle *handle, |
| 5103 | struct perf_event_header *header, |
| 5104 | struct perf_sample_data *data, |
| 5105 | struct perf_event *event) |
| 5106 | { |
| 5107 | u64 sample_type = data->type; |
| 5108 | |
| 5109 | perf_output_put(handle, *header); |
| 5110 | |
| 5111 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
| 5112 | perf_output_put(handle, data->id); |
| 5113 | |
| 5114 | if (sample_type & PERF_SAMPLE_IP) |
| 5115 | perf_output_put(handle, data->ip); |
| 5116 | |
| 5117 | if (sample_type & PERF_SAMPLE_TID) |
| 5118 | perf_output_put(handle, data->tid_entry); |
| 5119 | |
| 5120 | if (sample_type & PERF_SAMPLE_TIME) |
| 5121 | perf_output_put(handle, data->time); |
| 5122 | |
| 5123 | if (sample_type & PERF_SAMPLE_ADDR) |
| 5124 | perf_output_put(handle, data->addr); |
| 5125 | |
| 5126 | if (sample_type & PERF_SAMPLE_ID) |
| 5127 | perf_output_put(handle, data->id); |
| 5128 | |
| 5129 | if (sample_type & PERF_SAMPLE_STREAM_ID) |
| 5130 | perf_output_put(handle, data->stream_id); |
| 5131 | |
| 5132 | if (sample_type & PERF_SAMPLE_CPU) |
| 5133 | perf_output_put(handle, data->cpu_entry); |
| 5134 | |
| 5135 | if (sample_type & PERF_SAMPLE_PERIOD) |
| 5136 | perf_output_put(handle, data->period); |
| 5137 | |
| 5138 | if (sample_type & PERF_SAMPLE_READ) |
| 5139 | perf_output_read(handle, event); |
| 5140 | |
| 5141 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
| 5142 | if (data->callchain) { |
| 5143 | int size = 1; |
| 5144 | |
| 5145 | if (data->callchain) |
| 5146 | size += data->callchain->nr; |
| 5147 | |
| 5148 | size *= sizeof(u64); |
| 5149 | |
| 5150 | __output_copy(handle, data->callchain, size); |
| 5151 | } else { |
| 5152 | u64 nr = 0; |
| 5153 | perf_output_put(handle, nr); |
| 5154 | } |
| 5155 | } |
| 5156 | |
| 5157 | if (sample_type & PERF_SAMPLE_RAW) { |
| 5158 | if (data->raw) { |
| 5159 | perf_output_put(handle, data->raw->size); |
| 5160 | __output_copy(handle, data->raw->data, |
| 5161 | data->raw->size); |
| 5162 | } else { |
| 5163 | struct { |
| 5164 | u32 size; |
| 5165 | u32 data; |
| 5166 | } raw = { |
| 5167 | .size = sizeof(u32), |
| 5168 | .data = 0, |
| 5169 | }; |
| 5170 | perf_output_put(handle, raw); |
| 5171 | } |
| 5172 | } |
| 5173 | |
| 5174 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
| 5175 | if (data->br_stack) { |
| 5176 | size_t size; |
| 5177 | |
| 5178 | size = data->br_stack->nr |
| 5179 | * sizeof(struct perf_branch_entry); |
| 5180 | |
| 5181 | perf_output_put(handle, data->br_stack->nr); |
| 5182 | perf_output_copy(handle, data->br_stack->entries, size); |
| 5183 | } else { |
| 5184 | /* |
| 5185 | * we always store at least the value of nr |
| 5186 | */ |
| 5187 | u64 nr = 0; |
| 5188 | perf_output_put(handle, nr); |
| 5189 | } |
| 5190 | } |
| 5191 | |
| 5192 | if (sample_type & PERF_SAMPLE_REGS_USER) { |
| 5193 | u64 abi = data->regs_user.abi; |
| 5194 | |
| 5195 | /* |
| 5196 | * If there are no regs to dump, notice it through |
| 5197 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). |
| 5198 | */ |
| 5199 | perf_output_put(handle, abi); |
| 5200 | |
| 5201 | if (abi) { |
| 5202 | u64 mask = event->attr.sample_regs_user; |
| 5203 | perf_output_sample_regs(handle, |
| 5204 | data->regs_user.regs, |
| 5205 | mask); |
| 5206 | } |
| 5207 | } |
| 5208 | |
| 5209 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
| 5210 | perf_output_sample_ustack(handle, |
| 5211 | data->stack_user_size, |
| 5212 | data->regs_user.regs); |
| 5213 | } |
| 5214 | |
| 5215 | if (sample_type & PERF_SAMPLE_WEIGHT) |
| 5216 | perf_output_put(handle, data->weight); |
| 5217 | |
| 5218 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
| 5219 | perf_output_put(handle, data->data_src.val); |
| 5220 | |
| 5221 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
| 5222 | perf_output_put(handle, data->txn); |
| 5223 | |
| 5224 | if (sample_type & PERF_SAMPLE_REGS_INTR) { |
| 5225 | u64 abi = data->regs_intr.abi; |
| 5226 | /* |
| 5227 | * If there are no regs to dump, notice it through |
| 5228 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). |
| 5229 | */ |
| 5230 | perf_output_put(handle, abi); |
| 5231 | |
| 5232 | if (abi) { |
| 5233 | u64 mask = event->attr.sample_regs_intr; |
| 5234 | |
| 5235 | perf_output_sample_regs(handle, |
| 5236 | data->regs_intr.regs, |
| 5237 | mask); |
| 5238 | } |
| 5239 | } |
| 5240 | |
| 5241 | if (!event->attr.watermark) { |
| 5242 | int wakeup_events = event->attr.wakeup_events; |
| 5243 | |
| 5244 | if (wakeup_events) { |
| 5245 | struct ring_buffer *rb = handle->rb; |
| 5246 | int events = local_inc_return(&rb->events); |
| 5247 | |
| 5248 | if (events >= wakeup_events) { |
| 5249 | local_sub(wakeup_events, &rb->events); |
| 5250 | local_inc(&rb->wakeup); |
| 5251 | } |
| 5252 | } |
| 5253 | } |
| 5254 | } |
| 5255 | |
| 5256 | void perf_prepare_sample(struct perf_event_header *header, |
| 5257 | struct perf_sample_data *data, |
| 5258 | struct perf_event *event, |
| 5259 | struct pt_regs *regs) |
| 5260 | { |
| 5261 | u64 sample_type = event->attr.sample_type; |
| 5262 | |
| 5263 | header->type = PERF_RECORD_SAMPLE; |
| 5264 | header->size = sizeof(*header) + event->header_size; |
| 5265 | |
| 5266 | header->misc = 0; |
| 5267 | header->misc |= perf_misc_flags(regs); |
| 5268 | |
| 5269 | __perf_event_header__init_id(header, data, event); |
| 5270 | |
| 5271 | if (sample_type & PERF_SAMPLE_IP) |
| 5272 | data->ip = perf_instruction_pointer(regs); |
| 5273 | |
| 5274 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
| 5275 | int size = 1; |
| 5276 | |
| 5277 | data->callchain = perf_callchain(event, regs); |
| 5278 | |
| 5279 | if (data->callchain) |
| 5280 | size += data->callchain->nr; |
| 5281 | |
| 5282 | header->size += size * sizeof(u64); |
| 5283 | } |
| 5284 | |
| 5285 | if (sample_type & PERF_SAMPLE_RAW) { |
| 5286 | int size = sizeof(u32); |
| 5287 | |
| 5288 | if (data->raw) |
| 5289 | size += data->raw->size; |
| 5290 | else |
| 5291 | size += sizeof(u32); |
| 5292 | |
| 5293 | WARN_ON_ONCE(size & (sizeof(u64)-1)); |
| 5294 | header->size += size; |
| 5295 | } |
| 5296 | |
| 5297 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
| 5298 | int size = sizeof(u64); /* nr */ |
| 5299 | if (data->br_stack) { |
| 5300 | size += data->br_stack->nr |
| 5301 | * sizeof(struct perf_branch_entry); |
| 5302 | } |
| 5303 | header->size += size; |
| 5304 | } |
| 5305 | |
| 5306 | if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER)) |
| 5307 | perf_sample_regs_user(&data->regs_user, regs, |
| 5308 | &data->regs_user_copy); |
| 5309 | |
| 5310 | if (sample_type & PERF_SAMPLE_REGS_USER) { |
| 5311 | /* regs dump ABI info */ |
| 5312 | int size = sizeof(u64); |
| 5313 | |
| 5314 | if (data->regs_user.regs) { |
| 5315 | u64 mask = event->attr.sample_regs_user; |
| 5316 | size += hweight64(mask) * sizeof(u64); |
| 5317 | } |
| 5318 | |
| 5319 | header->size += size; |
| 5320 | } |
| 5321 | |
| 5322 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
| 5323 | /* |
| 5324 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways |
| 5325 | * processed as the last one or have additional check added |
| 5326 | * in case new sample type is added, because we could eat |
| 5327 | * up the rest of the sample size. |
| 5328 | */ |
| 5329 | u16 stack_size = event->attr.sample_stack_user; |
| 5330 | u16 size = sizeof(u64); |
| 5331 | |
| 5332 | stack_size = perf_sample_ustack_size(stack_size, header->size, |
| 5333 | data->regs_user.regs); |
| 5334 | |
| 5335 | /* |
| 5336 | * If there is something to dump, add space for the dump |
| 5337 | * itself and for the field that tells the dynamic size, |
| 5338 | * which is how many have been actually dumped. |
| 5339 | */ |
| 5340 | if (stack_size) |
| 5341 | size += sizeof(u64) + stack_size; |
| 5342 | |
| 5343 | data->stack_user_size = stack_size; |
| 5344 | header->size += size; |
| 5345 | } |
| 5346 | |
| 5347 | if (sample_type & PERF_SAMPLE_REGS_INTR) { |
| 5348 | /* regs dump ABI info */ |
| 5349 | int size = sizeof(u64); |
| 5350 | |
| 5351 | perf_sample_regs_intr(&data->regs_intr, regs); |
| 5352 | |
| 5353 | if (data->regs_intr.regs) { |
| 5354 | u64 mask = event->attr.sample_regs_intr; |
| 5355 | |
| 5356 | size += hweight64(mask) * sizeof(u64); |
| 5357 | } |
| 5358 | |
| 5359 | header->size += size; |
| 5360 | } |
| 5361 | } |
| 5362 | |
| 5363 | static void perf_event_output(struct perf_event *event, |
| 5364 | struct perf_sample_data *data, |
| 5365 | struct pt_regs *regs) |
| 5366 | { |
| 5367 | struct perf_output_handle handle; |
| 5368 | struct perf_event_header header; |
| 5369 | |
| 5370 | /* protect the callchain buffers */ |
| 5371 | rcu_read_lock(); |
| 5372 | |
| 5373 | perf_prepare_sample(&header, data, event, regs); |
| 5374 | |
| 5375 | if (perf_output_begin(&handle, event, header.size)) |
| 5376 | goto exit; |
| 5377 | |
| 5378 | perf_output_sample(&handle, &header, data, event); |
| 5379 | |
| 5380 | perf_output_end(&handle); |
| 5381 | |
| 5382 | exit: |
| 5383 | rcu_read_unlock(); |
| 5384 | } |
| 5385 | |
| 5386 | /* |
| 5387 | * read event_id |
| 5388 | */ |
| 5389 | |
| 5390 | struct perf_read_event { |
| 5391 | struct perf_event_header header; |
| 5392 | |
| 5393 | u32 pid; |
| 5394 | u32 tid; |
| 5395 | }; |
| 5396 | |
| 5397 | static void |
| 5398 | perf_event_read_event(struct perf_event *event, |
| 5399 | struct task_struct *task) |
| 5400 | { |
| 5401 | struct perf_output_handle handle; |
| 5402 | struct perf_sample_data sample; |
| 5403 | struct perf_read_event read_event = { |
| 5404 | .header = { |
| 5405 | .type = PERF_RECORD_READ, |
| 5406 | .misc = 0, |
| 5407 | .size = sizeof(read_event) + event->read_size, |
| 5408 | }, |
| 5409 | .pid = perf_event_pid(event, task), |
| 5410 | .tid = perf_event_tid(event, task), |
| 5411 | }; |
| 5412 | int ret; |
| 5413 | |
| 5414 | perf_event_header__init_id(&read_event.header, &sample, event); |
| 5415 | ret = perf_output_begin(&handle, event, read_event.header.size); |
| 5416 | if (ret) |
| 5417 | return; |
| 5418 | |
| 5419 | perf_output_put(&handle, read_event); |
| 5420 | perf_output_read(&handle, event); |
| 5421 | perf_event__output_id_sample(event, &handle, &sample); |
| 5422 | |
| 5423 | perf_output_end(&handle); |
| 5424 | } |
| 5425 | |
| 5426 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
| 5427 | |
| 5428 | static void |
| 5429 | perf_event_aux_ctx(struct perf_event_context *ctx, |
| 5430 | perf_event_aux_output_cb output, |
| 5431 | void *data) |
| 5432 | { |
| 5433 | struct perf_event *event; |
| 5434 | |
| 5435 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
| 5436 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
| 5437 | continue; |
| 5438 | if (!event_filter_match(event)) |
| 5439 | continue; |
| 5440 | output(event, data); |
| 5441 | } |
| 5442 | } |
| 5443 | |
| 5444 | static void |
| 5445 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
| 5446 | struct perf_event_context *task_ctx) |
| 5447 | { |
| 5448 | struct perf_cpu_context *cpuctx; |
| 5449 | struct perf_event_context *ctx; |
| 5450 | struct pmu *pmu; |
| 5451 | int ctxn; |
| 5452 | |
| 5453 | rcu_read_lock(); |
| 5454 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
| 5455 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); |
| 5456 | if (cpuctx->unique_pmu != pmu) |
| 5457 | goto next; |
| 5458 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
| 5459 | if (task_ctx) |
| 5460 | goto next; |
| 5461 | ctxn = pmu->task_ctx_nr; |
| 5462 | if (ctxn < 0) |
| 5463 | goto next; |
| 5464 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); |
| 5465 | if (ctx) |
| 5466 | perf_event_aux_ctx(ctx, output, data); |
| 5467 | next: |
| 5468 | put_cpu_ptr(pmu->pmu_cpu_context); |
| 5469 | } |
| 5470 | |
| 5471 | if (task_ctx) { |
| 5472 | preempt_disable(); |
| 5473 | perf_event_aux_ctx(task_ctx, output, data); |
| 5474 | preempt_enable(); |
| 5475 | } |
| 5476 | rcu_read_unlock(); |
| 5477 | } |
| 5478 | |
| 5479 | /* |
| 5480 | * task tracking -- fork/exit |
| 5481 | * |
| 5482 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
| 5483 | */ |
| 5484 | |
| 5485 | struct perf_task_event { |
| 5486 | struct task_struct *task; |
| 5487 | struct perf_event_context *task_ctx; |
| 5488 | |
| 5489 | struct { |
| 5490 | struct perf_event_header header; |
| 5491 | |
| 5492 | u32 pid; |
| 5493 | u32 ppid; |
| 5494 | u32 tid; |
| 5495 | u32 ptid; |
| 5496 | u64 time; |
| 5497 | } event_id; |
| 5498 | }; |
| 5499 | |
| 5500 | static int perf_event_task_match(struct perf_event *event) |
| 5501 | { |
| 5502 | return event->attr.comm || event->attr.mmap || |
| 5503 | event->attr.mmap2 || event->attr.mmap_data || |
| 5504 | event->attr.task; |
| 5505 | } |
| 5506 | |
| 5507 | static void perf_event_task_output(struct perf_event *event, |
| 5508 | void *data) |
| 5509 | { |
| 5510 | struct perf_task_event *task_event = data; |
| 5511 | struct perf_output_handle handle; |
| 5512 | struct perf_sample_data sample; |
| 5513 | struct task_struct *task = task_event->task; |
| 5514 | int ret, size = task_event->event_id.header.size; |
| 5515 | |
| 5516 | if (!perf_event_task_match(event)) |
| 5517 | return; |
| 5518 | |
| 5519 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
| 5520 | |
| 5521 | ret = perf_output_begin(&handle, event, |
| 5522 | task_event->event_id.header.size); |
| 5523 | if (ret) |
| 5524 | goto out; |
| 5525 | |
| 5526 | task_event->event_id.pid = perf_event_pid(event, task); |
| 5527 | task_event->event_id.ppid = perf_event_pid(event, current); |
| 5528 | |
| 5529 | task_event->event_id.tid = perf_event_tid(event, task); |
| 5530 | task_event->event_id.ptid = perf_event_tid(event, current); |
| 5531 | |
| 5532 | task_event->event_id.time = perf_event_clock(event); |
| 5533 | |
| 5534 | perf_output_put(&handle, task_event->event_id); |
| 5535 | |
| 5536 | perf_event__output_id_sample(event, &handle, &sample); |
| 5537 | |
| 5538 | perf_output_end(&handle); |
| 5539 | out: |
| 5540 | task_event->event_id.header.size = size; |
| 5541 | } |
| 5542 | |
| 5543 | static void perf_event_task(struct task_struct *task, |
| 5544 | struct perf_event_context *task_ctx, |
| 5545 | int new) |
| 5546 | { |
| 5547 | struct perf_task_event task_event; |
| 5548 | |
| 5549 | if (!atomic_read(&nr_comm_events) && |
| 5550 | !atomic_read(&nr_mmap_events) && |
| 5551 | !atomic_read(&nr_task_events)) |
| 5552 | return; |
| 5553 | |
| 5554 | task_event = (struct perf_task_event){ |
| 5555 | .task = task, |
| 5556 | .task_ctx = task_ctx, |
| 5557 | .event_id = { |
| 5558 | .header = { |
| 5559 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
| 5560 | .misc = 0, |
| 5561 | .size = sizeof(task_event.event_id), |
| 5562 | }, |
| 5563 | /* .pid */ |
| 5564 | /* .ppid */ |
| 5565 | /* .tid */ |
| 5566 | /* .ptid */ |
| 5567 | /* .time */ |
| 5568 | }, |
| 5569 | }; |
| 5570 | |
| 5571 | perf_event_aux(perf_event_task_output, |
| 5572 | &task_event, |
| 5573 | task_ctx); |
| 5574 | } |
| 5575 | |
| 5576 | void perf_event_fork(struct task_struct *task) |
| 5577 | { |
| 5578 | perf_event_task(task, NULL, 1); |
| 5579 | } |
| 5580 | |
| 5581 | /* |
| 5582 | * comm tracking |
| 5583 | */ |
| 5584 | |
| 5585 | struct perf_comm_event { |
| 5586 | struct task_struct *task; |
| 5587 | char *comm; |
| 5588 | int comm_size; |
| 5589 | |
| 5590 | struct { |
| 5591 | struct perf_event_header header; |
| 5592 | |
| 5593 | u32 pid; |
| 5594 | u32 tid; |
| 5595 | } event_id; |
| 5596 | }; |
| 5597 | |
| 5598 | static int perf_event_comm_match(struct perf_event *event) |
| 5599 | { |
| 5600 | return event->attr.comm; |
| 5601 | } |
| 5602 | |
| 5603 | static void perf_event_comm_output(struct perf_event *event, |
| 5604 | void *data) |
| 5605 | { |
| 5606 | struct perf_comm_event *comm_event = data; |
| 5607 | struct perf_output_handle handle; |
| 5608 | struct perf_sample_data sample; |
| 5609 | int size = comm_event->event_id.header.size; |
| 5610 | int ret; |
| 5611 | |
| 5612 | if (!perf_event_comm_match(event)) |
| 5613 | return; |
| 5614 | |
| 5615 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
| 5616 | ret = perf_output_begin(&handle, event, |
| 5617 | comm_event->event_id.header.size); |
| 5618 | |
| 5619 | if (ret) |
| 5620 | goto out; |
| 5621 | |
| 5622 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
| 5623 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); |
| 5624 | |
| 5625 | perf_output_put(&handle, comm_event->event_id); |
| 5626 | __output_copy(&handle, comm_event->comm, |
| 5627 | comm_event->comm_size); |
| 5628 | |
| 5629 | perf_event__output_id_sample(event, &handle, &sample); |
| 5630 | |
| 5631 | perf_output_end(&handle); |
| 5632 | out: |
| 5633 | comm_event->event_id.header.size = size; |
| 5634 | } |
| 5635 | |
| 5636 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
| 5637 | { |
| 5638 | char comm[TASK_COMM_LEN]; |
| 5639 | unsigned int size; |
| 5640 | |
| 5641 | memset(comm, 0, sizeof(comm)); |
| 5642 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
| 5643 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
| 5644 | |
| 5645 | comm_event->comm = comm; |
| 5646 | comm_event->comm_size = size; |
| 5647 | |
| 5648 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
| 5649 | |
| 5650 | perf_event_aux(perf_event_comm_output, |
| 5651 | comm_event, |
| 5652 | NULL); |
| 5653 | } |
| 5654 | |
| 5655 | void perf_event_comm(struct task_struct *task, bool exec) |
| 5656 | { |
| 5657 | struct perf_comm_event comm_event; |
| 5658 | |
| 5659 | if (!atomic_read(&nr_comm_events)) |
| 5660 | return; |
| 5661 | |
| 5662 | comm_event = (struct perf_comm_event){ |
| 5663 | .task = task, |
| 5664 | /* .comm */ |
| 5665 | /* .comm_size */ |
| 5666 | .event_id = { |
| 5667 | .header = { |
| 5668 | .type = PERF_RECORD_COMM, |
| 5669 | .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0, |
| 5670 | /* .size */ |
| 5671 | }, |
| 5672 | /* .pid */ |
| 5673 | /* .tid */ |
| 5674 | }, |
| 5675 | }; |
| 5676 | |
| 5677 | perf_event_comm_event(&comm_event); |
| 5678 | } |
| 5679 | |
| 5680 | /* |
| 5681 | * mmap tracking |
| 5682 | */ |
| 5683 | |
| 5684 | struct perf_mmap_event { |
| 5685 | struct vm_area_struct *vma; |
| 5686 | |
| 5687 | const char *file_name; |
| 5688 | int file_size; |
| 5689 | int maj, min; |
| 5690 | u64 ino; |
| 5691 | u64 ino_generation; |
| 5692 | u32 prot, flags; |
| 5693 | |
| 5694 | struct { |
| 5695 | struct perf_event_header header; |
| 5696 | |
| 5697 | u32 pid; |
| 5698 | u32 tid; |
| 5699 | u64 start; |
| 5700 | u64 len; |
| 5701 | u64 pgoff; |
| 5702 | } event_id; |
| 5703 | }; |
| 5704 | |
| 5705 | static int perf_event_mmap_match(struct perf_event *event, |
| 5706 | void *data) |
| 5707 | { |
| 5708 | struct perf_mmap_event *mmap_event = data; |
| 5709 | struct vm_area_struct *vma = mmap_event->vma; |
| 5710 | int executable = vma->vm_flags & VM_EXEC; |
| 5711 | |
| 5712 | return (!executable && event->attr.mmap_data) || |
| 5713 | (executable && (event->attr.mmap || event->attr.mmap2)); |
| 5714 | } |
| 5715 | |
| 5716 | static void perf_event_mmap_output(struct perf_event *event, |
| 5717 | void *data) |
| 5718 | { |
| 5719 | struct perf_mmap_event *mmap_event = data; |
| 5720 | struct perf_output_handle handle; |
| 5721 | struct perf_sample_data sample; |
| 5722 | int size = mmap_event->event_id.header.size; |
| 5723 | int ret; |
| 5724 | |
| 5725 | if (!perf_event_mmap_match(event, data)) |
| 5726 | return; |
| 5727 | |
| 5728 | if (event->attr.mmap2) { |
| 5729 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; |
| 5730 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); |
| 5731 | mmap_event->event_id.header.size += sizeof(mmap_event->min); |
| 5732 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); |
| 5733 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
| 5734 | mmap_event->event_id.header.size += sizeof(mmap_event->prot); |
| 5735 | mmap_event->event_id.header.size += sizeof(mmap_event->flags); |
| 5736 | } |
| 5737 | |
| 5738 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
| 5739 | ret = perf_output_begin(&handle, event, |
| 5740 | mmap_event->event_id.header.size); |
| 5741 | if (ret) |
| 5742 | goto out; |
| 5743 | |
| 5744 | mmap_event->event_id.pid = perf_event_pid(event, current); |
| 5745 | mmap_event->event_id.tid = perf_event_tid(event, current); |
| 5746 | |
| 5747 | perf_output_put(&handle, mmap_event->event_id); |
| 5748 | |
| 5749 | if (event->attr.mmap2) { |
| 5750 | perf_output_put(&handle, mmap_event->maj); |
| 5751 | perf_output_put(&handle, mmap_event->min); |
| 5752 | perf_output_put(&handle, mmap_event->ino); |
| 5753 | perf_output_put(&handle, mmap_event->ino_generation); |
| 5754 | perf_output_put(&handle, mmap_event->prot); |
| 5755 | perf_output_put(&handle, mmap_event->flags); |
| 5756 | } |
| 5757 | |
| 5758 | __output_copy(&handle, mmap_event->file_name, |
| 5759 | mmap_event->file_size); |
| 5760 | |
| 5761 | perf_event__output_id_sample(event, &handle, &sample); |
| 5762 | |
| 5763 | perf_output_end(&handle); |
| 5764 | out: |
| 5765 | mmap_event->event_id.header.size = size; |
| 5766 | } |
| 5767 | |
| 5768 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
| 5769 | { |
| 5770 | struct vm_area_struct *vma = mmap_event->vma; |
| 5771 | struct file *file = vma->vm_file; |
| 5772 | int maj = 0, min = 0; |
| 5773 | u64 ino = 0, gen = 0; |
| 5774 | u32 prot = 0, flags = 0; |
| 5775 | unsigned int size; |
| 5776 | char tmp[16]; |
| 5777 | char *buf = NULL; |
| 5778 | char *name; |
| 5779 | |
| 5780 | if (file) { |
| 5781 | struct inode *inode; |
| 5782 | dev_t dev; |
| 5783 | |
| 5784 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
| 5785 | if (!buf) { |
| 5786 | name = "//enomem"; |
| 5787 | goto cpy_name; |
| 5788 | } |
| 5789 | /* |
| 5790 | * d_path() works from the end of the rb backwards, so we |
| 5791 | * need to add enough zero bytes after the string to handle |
| 5792 | * the 64bit alignment we do later. |
| 5793 | */ |
| 5794 | name = file_path(file, buf, PATH_MAX - sizeof(u64)); |
| 5795 | if (IS_ERR(name)) { |
| 5796 | name = "//toolong"; |
| 5797 | goto cpy_name; |
| 5798 | } |
| 5799 | inode = file_inode(vma->vm_file); |
| 5800 | dev = inode->i_sb->s_dev; |
| 5801 | ino = inode->i_ino; |
| 5802 | gen = inode->i_generation; |
| 5803 | maj = MAJOR(dev); |
| 5804 | min = MINOR(dev); |
| 5805 | |
| 5806 | if (vma->vm_flags & VM_READ) |
| 5807 | prot |= PROT_READ; |
| 5808 | if (vma->vm_flags & VM_WRITE) |
| 5809 | prot |= PROT_WRITE; |
| 5810 | if (vma->vm_flags & VM_EXEC) |
| 5811 | prot |= PROT_EXEC; |
| 5812 | |
| 5813 | if (vma->vm_flags & VM_MAYSHARE) |
| 5814 | flags = MAP_SHARED; |
| 5815 | else |
| 5816 | flags = MAP_PRIVATE; |
| 5817 | |
| 5818 | if (vma->vm_flags & VM_DENYWRITE) |
| 5819 | flags |= MAP_DENYWRITE; |
| 5820 | if (vma->vm_flags & VM_MAYEXEC) |
| 5821 | flags |= MAP_EXECUTABLE; |
| 5822 | if (vma->vm_flags & VM_LOCKED) |
| 5823 | flags |= MAP_LOCKED; |
| 5824 | if (vma->vm_flags & VM_HUGETLB) |
| 5825 | flags |= MAP_HUGETLB; |
| 5826 | |
| 5827 | goto got_name; |
| 5828 | } else { |
| 5829 | if (vma->vm_ops && vma->vm_ops->name) { |
| 5830 | name = (char *) vma->vm_ops->name(vma); |
| 5831 | if (name) |
| 5832 | goto cpy_name; |
| 5833 | } |
| 5834 | |
| 5835 | name = (char *)arch_vma_name(vma); |
| 5836 | if (name) |
| 5837 | goto cpy_name; |
| 5838 | |
| 5839 | if (vma->vm_start <= vma->vm_mm->start_brk && |
| 5840 | vma->vm_end >= vma->vm_mm->brk) { |
| 5841 | name = "[heap]"; |
| 5842 | goto cpy_name; |
| 5843 | } |
| 5844 | if (vma->vm_start <= vma->vm_mm->start_stack && |
| 5845 | vma->vm_end >= vma->vm_mm->start_stack) { |
| 5846 | name = "[stack]"; |
| 5847 | goto cpy_name; |
| 5848 | } |
| 5849 | |
| 5850 | name = "//anon"; |
| 5851 | goto cpy_name; |
| 5852 | } |
| 5853 | |
| 5854 | cpy_name: |
| 5855 | strlcpy(tmp, name, sizeof(tmp)); |
| 5856 | name = tmp; |
| 5857 | got_name: |
| 5858 | /* |
| 5859 | * Since our buffer works in 8 byte units we need to align our string |
| 5860 | * size to a multiple of 8. However, we must guarantee the tail end is |
| 5861 | * zero'd out to avoid leaking random bits to userspace. |
| 5862 | */ |
| 5863 | size = strlen(name)+1; |
| 5864 | while (!IS_ALIGNED(size, sizeof(u64))) |
| 5865 | name[size++] = '\0'; |
| 5866 | |
| 5867 | mmap_event->file_name = name; |
| 5868 | mmap_event->file_size = size; |
| 5869 | mmap_event->maj = maj; |
| 5870 | mmap_event->min = min; |
| 5871 | mmap_event->ino = ino; |
| 5872 | mmap_event->ino_generation = gen; |
| 5873 | mmap_event->prot = prot; |
| 5874 | mmap_event->flags = flags; |
| 5875 | |
| 5876 | if (!(vma->vm_flags & VM_EXEC)) |
| 5877 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; |
| 5878 | |
| 5879 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
| 5880 | |
| 5881 | perf_event_aux(perf_event_mmap_output, |
| 5882 | mmap_event, |
| 5883 | NULL); |
| 5884 | |
| 5885 | kfree(buf); |
| 5886 | } |
| 5887 | |
| 5888 | void perf_event_mmap(struct vm_area_struct *vma) |
| 5889 | { |
| 5890 | struct perf_mmap_event mmap_event; |
| 5891 | |
| 5892 | if (!atomic_read(&nr_mmap_events)) |
| 5893 | return; |
| 5894 | |
| 5895 | mmap_event = (struct perf_mmap_event){ |
| 5896 | .vma = vma, |
| 5897 | /* .file_name */ |
| 5898 | /* .file_size */ |
| 5899 | .event_id = { |
| 5900 | .header = { |
| 5901 | .type = PERF_RECORD_MMAP, |
| 5902 | .misc = PERF_RECORD_MISC_USER, |
| 5903 | /* .size */ |
| 5904 | }, |
| 5905 | /* .pid */ |
| 5906 | /* .tid */ |
| 5907 | .start = vma->vm_start, |
| 5908 | .len = vma->vm_end - vma->vm_start, |
| 5909 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
| 5910 | }, |
| 5911 | /* .maj (attr_mmap2 only) */ |
| 5912 | /* .min (attr_mmap2 only) */ |
| 5913 | /* .ino (attr_mmap2 only) */ |
| 5914 | /* .ino_generation (attr_mmap2 only) */ |
| 5915 | /* .prot (attr_mmap2 only) */ |
| 5916 | /* .flags (attr_mmap2 only) */ |
| 5917 | }; |
| 5918 | |
| 5919 | perf_event_mmap_event(&mmap_event); |
| 5920 | } |
| 5921 | |
| 5922 | void perf_event_aux_event(struct perf_event *event, unsigned long head, |
| 5923 | unsigned long size, u64 flags) |
| 5924 | { |
| 5925 | struct perf_output_handle handle; |
| 5926 | struct perf_sample_data sample; |
| 5927 | struct perf_aux_event { |
| 5928 | struct perf_event_header header; |
| 5929 | u64 offset; |
| 5930 | u64 size; |
| 5931 | u64 flags; |
| 5932 | } rec = { |
| 5933 | .header = { |
| 5934 | .type = PERF_RECORD_AUX, |
| 5935 | .misc = 0, |
| 5936 | .size = sizeof(rec), |
| 5937 | }, |
| 5938 | .offset = head, |
| 5939 | .size = size, |
| 5940 | .flags = flags, |
| 5941 | }; |
| 5942 | int ret; |
| 5943 | |
| 5944 | perf_event_header__init_id(&rec.header, &sample, event); |
| 5945 | ret = perf_output_begin(&handle, event, rec.header.size); |
| 5946 | |
| 5947 | if (ret) |
| 5948 | return; |
| 5949 | |
| 5950 | perf_output_put(&handle, rec); |
| 5951 | perf_event__output_id_sample(event, &handle, &sample); |
| 5952 | |
| 5953 | perf_output_end(&handle); |
| 5954 | } |
| 5955 | |
| 5956 | /* |
| 5957 | * IRQ throttle logging |
| 5958 | */ |
| 5959 | |
| 5960 | static void perf_log_throttle(struct perf_event *event, int enable) |
| 5961 | { |
| 5962 | struct perf_output_handle handle; |
| 5963 | struct perf_sample_data sample; |
| 5964 | int ret; |
| 5965 | |
| 5966 | struct { |
| 5967 | struct perf_event_header header; |
| 5968 | u64 time; |
| 5969 | u64 id; |
| 5970 | u64 stream_id; |
| 5971 | } throttle_event = { |
| 5972 | .header = { |
| 5973 | .type = PERF_RECORD_THROTTLE, |
| 5974 | .misc = 0, |
| 5975 | .size = sizeof(throttle_event), |
| 5976 | }, |
| 5977 | .time = perf_event_clock(event), |
| 5978 | .id = primary_event_id(event), |
| 5979 | .stream_id = event->id, |
| 5980 | }; |
| 5981 | |
| 5982 | if (enable) |
| 5983 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
| 5984 | |
| 5985 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
| 5986 | |
| 5987 | ret = perf_output_begin(&handle, event, |
| 5988 | throttle_event.header.size); |
| 5989 | if (ret) |
| 5990 | return; |
| 5991 | |
| 5992 | perf_output_put(&handle, throttle_event); |
| 5993 | perf_event__output_id_sample(event, &handle, &sample); |
| 5994 | perf_output_end(&handle); |
| 5995 | } |
| 5996 | |
| 5997 | static void perf_log_itrace_start(struct perf_event *event) |
| 5998 | { |
| 5999 | struct perf_output_handle handle; |
| 6000 | struct perf_sample_data sample; |
| 6001 | struct perf_aux_event { |
| 6002 | struct perf_event_header header; |
| 6003 | u32 pid; |
| 6004 | u32 tid; |
| 6005 | } rec; |
| 6006 | int ret; |
| 6007 | |
| 6008 | if (event->parent) |
| 6009 | event = event->parent; |
| 6010 | |
| 6011 | if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) || |
| 6012 | event->hw.itrace_started) |
| 6013 | return; |
| 6014 | |
| 6015 | event->hw.itrace_started = 1; |
| 6016 | |
| 6017 | rec.header.type = PERF_RECORD_ITRACE_START; |
| 6018 | rec.header.misc = 0; |
| 6019 | rec.header.size = sizeof(rec); |
| 6020 | rec.pid = perf_event_pid(event, current); |
| 6021 | rec.tid = perf_event_tid(event, current); |
| 6022 | |
| 6023 | perf_event_header__init_id(&rec.header, &sample, event); |
| 6024 | ret = perf_output_begin(&handle, event, rec.header.size); |
| 6025 | |
| 6026 | if (ret) |
| 6027 | return; |
| 6028 | |
| 6029 | perf_output_put(&handle, rec); |
| 6030 | perf_event__output_id_sample(event, &handle, &sample); |
| 6031 | |
| 6032 | perf_output_end(&handle); |
| 6033 | } |
| 6034 | |
| 6035 | /* |
| 6036 | * Generic event overflow handling, sampling. |
| 6037 | */ |
| 6038 | |
| 6039 | static int __perf_event_overflow(struct perf_event *event, |
| 6040 | int throttle, struct perf_sample_data *data, |
| 6041 | struct pt_regs *regs) |
| 6042 | { |
| 6043 | int events = atomic_read(&event->event_limit); |
| 6044 | struct hw_perf_event *hwc = &event->hw; |
| 6045 | u64 seq; |
| 6046 | int ret = 0; |
| 6047 | |
| 6048 | /* |
| 6049 | * Non-sampling counters might still use the PMI to fold short |
| 6050 | * hardware counters, ignore those. |
| 6051 | */ |
| 6052 | if (unlikely(!is_sampling_event(event))) |
| 6053 | return 0; |
| 6054 | |
| 6055 | seq = __this_cpu_read(perf_throttled_seq); |
| 6056 | if (seq != hwc->interrupts_seq) { |
| 6057 | hwc->interrupts_seq = seq; |
| 6058 | hwc->interrupts = 1; |
| 6059 | } else { |
| 6060 | hwc->interrupts++; |
| 6061 | if (unlikely(throttle |
| 6062 | && hwc->interrupts >= max_samples_per_tick)) { |
| 6063 | __this_cpu_inc(perf_throttled_count); |
| 6064 | hwc->interrupts = MAX_INTERRUPTS; |
| 6065 | perf_log_throttle(event, 0); |
| 6066 | tick_nohz_full_kick(); |
| 6067 | ret = 1; |
| 6068 | } |
| 6069 | } |
| 6070 | |
| 6071 | if (event->attr.freq) { |
| 6072 | u64 now = perf_clock(); |
| 6073 | s64 delta = now - hwc->freq_time_stamp; |
| 6074 | |
| 6075 | hwc->freq_time_stamp = now; |
| 6076 | |
| 6077 | if (delta > 0 && delta < 2*TICK_NSEC) |
| 6078 | perf_adjust_period(event, delta, hwc->last_period, true); |
| 6079 | } |
| 6080 | |
| 6081 | /* |
| 6082 | * XXX event_limit might not quite work as expected on inherited |
| 6083 | * events |
| 6084 | */ |
| 6085 | |
| 6086 | event->pending_kill = POLL_IN; |
| 6087 | if (events && atomic_dec_and_test(&event->event_limit)) { |
| 6088 | ret = 1; |
| 6089 | event->pending_kill = POLL_HUP; |
| 6090 | event->pending_disable = 1; |
| 6091 | irq_work_queue(&event->pending); |
| 6092 | } |
| 6093 | |
| 6094 | if (event->overflow_handler) |
| 6095 | event->overflow_handler(event, data, regs); |
| 6096 | else |
| 6097 | perf_event_output(event, data, regs); |
| 6098 | |
| 6099 | if (event->fasync && event->pending_kill) { |
| 6100 | event->pending_wakeup = 1; |
| 6101 | irq_work_queue(&event->pending); |
| 6102 | } |
| 6103 | |
| 6104 | return ret; |
| 6105 | } |
| 6106 | |
| 6107 | int perf_event_overflow(struct perf_event *event, |
| 6108 | struct perf_sample_data *data, |
| 6109 | struct pt_regs *regs) |
| 6110 | { |
| 6111 | return __perf_event_overflow(event, 1, data, regs); |
| 6112 | } |
| 6113 | |
| 6114 | /* |
| 6115 | * Generic software event infrastructure |
| 6116 | */ |
| 6117 | |
| 6118 | struct swevent_htable { |
| 6119 | struct swevent_hlist *swevent_hlist; |
| 6120 | struct mutex hlist_mutex; |
| 6121 | int hlist_refcount; |
| 6122 | |
| 6123 | /* Recursion avoidance in each contexts */ |
| 6124 | int recursion[PERF_NR_CONTEXTS]; |
| 6125 | |
| 6126 | /* Keeps track of cpu being initialized/exited */ |
| 6127 | bool online; |
| 6128 | }; |
| 6129 | |
| 6130 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); |
| 6131 | |
| 6132 | /* |
| 6133 | * We directly increment event->count and keep a second value in |
| 6134 | * event->hw.period_left to count intervals. This period event |
| 6135 | * is kept in the range [-sample_period, 0] so that we can use the |
| 6136 | * sign as trigger. |
| 6137 | */ |
| 6138 | |
| 6139 | u64 perf_swevent_set_period(struct perf_event *event) |
| 6140 | { |
| 6141 | struct hw_perf_event *hwc = &event->hw; |
| 6142 | u64 period = hwc->last_period; |
| 6143 | u64 nr, offset; |
| 6144 | s64 old, val; |
| 6145 | |
| 6146 | hwc->last_period = hwc->sample_period; |
| 6147 | |
| 6148 | again: |
| 6149 | old = val = local64_read(&hwc->period_left); |
| 6150 | if (val < 0) |
| 6151 | return 0; |
| 6152 | |
| 6153 | nr = div64_u64(period + val, period); |
| 6154 | offset = nr * period; |
| 6155 | val -= offset; |
| 6156 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
| 6157 | goto again; |
| 6158 | |
| 6159 | return nr; |
| 6160 | } |
| 6161 | |
| 6162 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
| 6163 | struct perf_sample_data *data, |
| 6164 | struct pt_regs *regs) |
| 6165 | { |
| 6166 | struct hw_perf_event *hwc = &event->hw; |
| 6167 | int throttle = 0; |
| 6168 | |
| 6169 | if (!overflow) |
| 6170 | overflow = perf_swevent_set_period(event); |
| 6171 | |
| 6172 | if (hwc->interrupts == MAX_INTERRUPTS) |
| 6173 | return; |
| 6174 | |
| 6175 | for (; overflow; overflow--) { |
| 6176 | if (__perf_event_overflow(event, throttle, |
| 6177 | data, regs)) { |
| 6178 | /* |
| 6179 | * We inhibit the overflow from happening when |
| 6180 | * hwc->interrupts == MAX_INTERRUPTS. |
| 6181 | */ |
| 6182 | break; |
| 6183 | } |
| 6184 | throttle = 1; |
| 6185 | } |
| 6186 | } |
| 6187 | |
| 6188 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
| 6189 | struct perf_sample_data *data, |
| 6190 | struct pt_regs *regs) |
| 6191 | { |
| 6192 | struct hw_perf_event *hwc = &event->hw; |
| 6193 | |
| 6194 | local64_add(nr, &event->count); |
| 6195 | |
| 6196 | if (!regs) |
| 6197 | return; |
| 6198 | |
| 6199 | if (!is_sampling_event(event)) |
| 6200 | return; |
| 6201 | |
| 6202 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
| 6203 | data->period = nr; |
| 6204 | return perf_swevent_overflow(event, 1, data, regs); |
| 6205 | } else |
| 6206 | data->period = event->hw.last_period; |
| 6207 | |
| 6208 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
| 6209 | return perf_swevent_overflow(event, 1, data, regs); |
| 6210 | |
| 6211 | if (local64_add_negative(nr, &hwc->period_left)) |
| 6212 | return; |
| 6213 | |
| 6214 | perf_swevent_overflow(event, 0, data, regs); |
| 6215 | } |
| 6216 | |
| 6217 | static int perf_exclude_event(struct perf_event *event, |
| 6218 | struct pt_regs *regs) |
| 6219 | { |
| 6220 | if (event->hw.state & PERF_HES_STOPPED) |
| 6221 | return 1; |
| 6222 | |
| 6223 | if (regs) { |
| 6224 | if (event->attr.exclude_user && user_mode(regs)) |
| 6225 | return 1; |
| 6226 | |
| 6227 | if (event->attr.exclude_kernel && !user_mode(regs)) |
| 6228 | return 1; |
| 6229 | } |
| 6230 | |
| 6231 | return 0; |
| 6232 | } |
| 6233 | |
| 6234 | static int perf_swevent_match(struct perf_event *event, |
| 6235 | enum perf_type_id type, |
| 6236 | u32 event_id, |
| 6237 | struct perf_sample_data *data, |
| 6238 | struct pt_regs *regs) |
| 6239 | { |
| 6240 | if (event->attr.type != type) |
| 6241 | return 0; |
| 6242 | |
| 6243 | if (event->attr.config != event_id) |
| 6244 | return 0; |
| 6245 | |
| 6246 | if (perf_exclude_event(event, regs)) |
| 6247 | return 0; |
| 6248 | |
| 6249 | return 1; |
| 6250 | } |
| 6251 | |
| 6252 | static inline u64 swevent_hash(u64 type, u32 event_id) |
| 6253 | { |
| 6254 | u64 val = event_id | (type << 32); |
| 6255 | |
| 6256 | return hash_64(val, SWEVENT_HLIST_BITS); |
| 6257 | } |
| 6258 | |
| 6259 | static inline struct hlist_head * |
| 6260 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) |
| 6261 | { |
| 6262 | u64 hash = swevent_hash(type, event_id); |
| 6263 | |
| 6264 | return &hlist->heads[hash]; |
| 6265 | } |
| 6266 | |
| 6267 | /* For the read side: events when they trigger */ |
| 6268 | static inline struct hlist_head * |
| 6269 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
| 6270 | { |
| 6271 | struct swevent_hlist *hlist; |
| 6272 | |
| 6273 | hlist = rcu_dereference(swhash->swevent_hlist); |
| 6274 | if (!hlist) |
| 6275 | return NULL; |
| 6276 | |
| 6277 | return __find_swevent_head(hlist, type, event_id); |
| 6278 | } |
| 6279 | |
| 6280 | /* For the event head insertion and removal in the hlist */ |
| 6281 | static inline struct hlist_head * |
| 6282 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
| 6283 | { |
| 6284 | struct swevent_hlist *hlist; |
| 6285 | u32 event_id = event->attr.config; |
| 6286 | u64 type = event->attr.type; |
| 6287 | |
| 6288 | /* |
| 6289 | * Event scheduling is always serialized against hlist allocation |
| 6290 | * and release. Which makes the protected version suitable here. |
| 6291 | * The context lock guarantees that. |
| 6292 | */ |
| 6293 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
| 6294 | lockdep_is_held(&event->ctx->lock)); |
| 6295 | if (!hlist) |
| 6296 | return NULL; |
| 6297 | |
| 6298 | return __find_swevent_head(hlist, type, event_id); |
| 6299 | } |
| 6300 | |
| 6301 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, |
| 6302 | u64 nr, |
| 6303 | struct perf_sample_data *data, |
| 6304 | struct pt_regs *regs) |
| 6305 | { |
| 6306 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
| 6307 | struct perf_event *event; |
| 6308 | struct hlist_head *head; |
| 6309 | |
| 6310 | rcu_read_lock(); |
| 6311 | head = find_swevent_head_rcu(swhash, type, event_id); |
| 6312 | if (!head) |
| 6313 | goto end; |
| 6314 | |
| 6315 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
| 6316 | if (perf_swevent_match(event, type, event_id, data, regs)) |
| 6317 | perf_swevent_event(event, nr, data, regs); |
| 6318 | } |
| 6319 | end: |
| 6320 | rcu_read_unlock(); |
| 6321 | } |
| 6322 | |
| 6323 | DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]); |
| 6324 | |
| 6325 | int perf_swevent_get_recursion_context(void) |
| 6326 | { |
| 6327 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
| 6328 | |
| 6329 | return get_recursion_context(swhash->recursion); |
| 6330 | } |
| 6331 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
| 6332 | |
| 6333 | inline void perf_swevent_put_recursion_context(int rctx) |
| 6334 | { |
| 6335 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
| 6336 | |
| 6337 | put_recursion_context(swhash->recursion, rctx); |
| 6338 | } |
| 6339 | |
| 6340 | void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
| 6341 | { |
| 6342 | struct perf_sample_data data; |
| 6343 | |
| 6344 | if (WARN_ON_ONCE(!regs)) |
| 6345 | return; |
| 6346 | |
| 6347 | perf_sample_data_init(&data, addr, 0); |
| 6348 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
| 6349 | } |
| 6350 | |
| 6351 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
| 6352 | { |
| 6353 | int rctx; |
| 6354 | |
| 6355 | preempt_disable_notrace(); |
| 6356 | rctx = perf_swevent_get_recursion_context(); |
| 6357 | if (unlikely(rctx < 0)) |
| 6358 | goto fail; |
| 6359 | |
| 6360 | ___perf_sw_event(event_id, nr, regs, addr); |
| 6361 | |
| 6362 | perf_swevent_put_recursion_context(rctx); |
| 6363 | fail: |
| 6364 | preempt_enable_notrace(); |
| 6365 | } |
| 6366 | |
| 6367 | static void perf_swevent_read(struct perf_event *event) |
| 6368 | { |
| 6369 | } |
| 6370 | |
| 6371 | static int perf_swevent_add(struct perf_event *event, int flags) |
| 6372 | { |
| 6373 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
| 6374 | struct hw_perf_event *hwc = &event->hw; |
| 6375 | struct hlist_head *head; |
| 6376 | |
| 6377 | if (is_sampling_event(event)) { |
| 6378 | hwc->last_period = hwc->sample_period; |
| 6379 | perf_swevent_set_period(event); |
| 6380 | } |
| 6381 | |
| 6382 | hwc->state = !(flags & PERF_EF_START); |
| 6383 | |
| 6384 | head = find_swevent_head(swhash, event); |
| 6385 | if (!head) { |
| 6386 | /* |
| 6387 | * We can race with cpu hotplug code. Do not |
| 6388 | * WARN if the cpu just got unplugged. |
| 6389 | */ |
| 6390 | WARN_ON_ONCE(swhash->online); |
| 6391 | return -EINVAL; |
| 6392 | } |
| 6393 | |
| 6394 | hlist_add_head_rcu(&event->hlist_entry, head); |
| 6395 | perf_event_update_userpage(event); |
| 6396 | |
| 6397 | return 0; |
| 6398 | } |
| 6399 | |
| 6400 | static void perf_swevent_del(struct perf_event *event, int flags) |
| 6401 | { |
| 6402 | hlist_del_rcu(&event->hlist_entry); |
| 6403 | } |
| 6404 | |
| 6405 | static void perf_swevent_start(struct perf_event *event, int flags) |
| 6406 | { |
| 6407 | event->hw.state = 0; |
| 6408 | } |
| 6409 | |
| 6410 | static void perf_swevent_stop(struct perf_event *event, int flags) |
| 6411 | { |
| 6412 | event->hw.state = PERF_HES_STOPPED; |
| 6413 | } |
| 6414 | |
| 6415 | /* Deref the hlist from the update side */ |
| 6416 | static inline struct swevent_hlist * |
| 6417 | swevent_hlist_deref(struct swevent_htable *swhash) |
| 6418 | { |
| 6419 | return rcu_dereference_protected(swhash->swevent_hlist, |
| 6420 | lockdep_is_held(&swhash->hlist_mutex)); |
| 6421 | } |
| 6422 | |
| 6423 | static void swevent_hlist_release(struct swevent_htable *swhash) |
| 6424 | { |
| 6425 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
| 6426 | |
| 6427 | if (!hlist) |
| 6428 | return; |
| 6429 | |
| 6430 | RCU_INIT_POINTER(swhash->swevent_hlist, NULL); |
| 6431 | kfree_rcu(hlist, rcu_head); |
| 6432 | } |
| 6433 | |
| 6434 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) |
| 6435 | { |
| 6436 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
| 6437 | |
| 6438 | mutex_lock(&swhash->hlist_mutex); |
| 6439 | |
| 6440 | if (!--swhash->hlist_refcount) |
| 6441 | swevent_hlist_release(swhash); |
| 6442 | |
| 6443 | mutex_unlock(&swhash->hlist_mutex); |
| 6444 | } |
| 6445 | |
| 6446 | static void swevent_hlist_put(struct perf_event *event) |
| 6447 | { |
| 6448 | int cpu; |
| 6449 | |
| 6450 | for_each_possible_cpu(cpu) |
| 6451 | swevent_hlist_put_cpu(event, cpu); |
| 6452 | } |
| 6453 | |
| 6454 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) |
| 6455 | { |
| 6456 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
| 6457 | int err = 0; |
| 6458 | |
| 6459 | mutex_lock(&swhash->hlist_mutex); |
| 6460 | |
| 6461 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
| 6462 | struct swevent_hlist *hlist; |
| 6463 | |
| 6464 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); |
| 6465 | if (!hlist) { |
| 6466 | err = -ENOMEM; |
| 6467 | goto exit; |
| 6468 | } |
| 6469 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
| 6470 | } |
| 6471 | swhash->hlist_refcount++; |
| 6472 | exit: |
| 6473 | mutex_unlock(&swhash->hlist_mutex); |
| 6474 | |
| 6475 | return err; |
| 6476 | } |
| 6477 | |
| 6478 | static int swevent_hlist_get(struct perf_event *event) |
| 6479 | { |
| 6480 | int err; |
| 6481 | int cpu, failed_cpu; |
| 6482 | |
| 6483 | get_online_cpus(); |
| 6484 | for_each_possible_cpu(cpu) { |
| 6485 | err = swevent_hlist_get_cpu(event, cpu); |
| 6486 | if (err) { |
| 6487 | failed_cpu = cpu; |
| 6488 | goto fail; |
| 6489 | } |
| 6490 | } |
| 6491 | put_online_cpus(); |
| 6492 | |
| 6493 | return 0; |
| 6494 | fail: |
| 6495 | for_each_possible_cpu(cpu) { |
| 6496 | if (cpu == failed_cpu) |
| 6497 | break; |
| 6498 | swevent_hlist_put_cpu(event, cpu); |
| 6499 | } |
| 6500 | |
| 6501 | put_online_cpus(); |
| 6502 | return err; |
| 6503 | } |
| 6504 | |
| 6505 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
| 6506 | |
| 6507 | static void sw_perf_event_destroy(struct perf_event *event) |
| 6508 | { |
| 6509 | u64 event_id = event->attr.config; |
| 6510 | |
| 6511 | WARN_ON(event->parent); |
| 6512 | |
| 6513 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
| 6514 | swevent_hlist_put(event); |
| 6515 | } |
| 6516 | |
| 6517 | static int perf_swevent_init(struct perf_event *event) |
| 6518 | { |
| 6519 | u64 event_id = event->attr.config; |
| 6520 | |
| 6521 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
| 6522 | return -ENOENT; |
| 6523 | |
| 6524 | /* |
| 6525 | * no branch sampling for software events |
| 6526 | */ |
| 6527 | if (has_branch_stack(event)) |
| 6528 | return -EOPNOTSUPP; |
| 6529 | |
| 6530 | switch (event_id) { |
| 6531 | case PERF_COUNT_SW_CPU_CLOCK: |
| 6532 | case PERF_COUNT_SW_TASK_CLOCK: |
| 6533 | return -ENOENT; |
| 6534 | |
| 6535 | default: |
| 6536 | break; |
| 6537 | } |
| 6538 | |
| 6539 | if (event_id >= PERF_COUNT_SW_MAX) |
| 6540 | return -ENOENT; |
| 6541 | |
| 6542 | if (!event->parent) { |
| 6543 | int err; |
| 6544 | |
| 6545 | err = swevent_hlist_get(event); |
| 6546 | if (err) |
| 6547 | return err; |
| 6548 | |
| 6549 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
| 6550 | event->destroy = sw_perf_event_destroy; |
| 6551 | } |
| 6552 | |
| 6553 | return 0; |
| 6554 | } |
| 6555 | |
| 6556 | static struct pmu perf_swevent = { |
| 6557 | .task_ctx_nr = perf_sw_context, |
| 6558 | |
| 6559 | .capabilities = PERF_PMU_CAP_NO_NMI, |
| 6560 | |
| 6561 | .event_init = perf_swevent_init, |
| 6562 | .add = perf_swevent_add, |
| 6563 | .del = perf_swevent_del, |
| 6564 | .start = perf_swevent_start, |
| 6565 | .stop = perf_swevent_stop, |
| 6566 | .read = perf_swevent_read, |
| 6567 | }; |
| 6568 | |
| 6569 | #ifdef CONFIG_EVENT_TRACING |
| 6570 | |
| 6571 | static int perf_tp_filter_match(struct perf_event *event, |
| 6572 | struct perf_sample_data *data) |
| 6573 | { |
| 6574 | void *record = data->raw->data; |
| 6575 | |
| 6576 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) |
| 6577 | return 1; |
| 6578 | return 0; |
| 6579 | } |
| 6580 | |
| 6581 | static int perf_tp_event_match(struct perf_event *event, |
| 6582 | struct perf_sample_data *data, |
| 6583 | struct pt_regs *regs) |
| 6584 | { |
| 6585 | if (event->hw.state & PERF_HES_STOPPED) |
| 6586 | return 0; |
| 6587 | /* |
| 6588 | * All tracepoints are from kernel-space. |
| 6589 | */ |
| 6590 | if (event->attr.exclude_kernel) |
| 6591 | return 0; |
| 6592 | |
| 6593 | if (!perf_tp_filter_match(event, data)) |
| 6594 | return 0; |
| 6595 | |
| 6596 | return 1; |
| 6597 | } |
| 6598 | |
| 6599 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, |
| 6600 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
| 6601 | struct task_struct *task) |
| 6602 | { |
| 6603 | struct perf_sample_data data; |
| 6604 | struct perf_event *event; |
| 6605 | |
| 6606 | struct perf_raw_record raw = { |
| 6607 | .size = entry_size, |
| 6608 | .data = record, |
| 6609 | }; |
| 6610 | |
| 6611 | perf_sample_data_init(&data, addr, 0); |
| 6612 | data.raw = &raw; |
| 6613 | |
| 6614 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
| 6615 | if (perf_tp_event_match(event, &data, regs)) |
| 6616 | perf_swevent_event(event, count, &data, regs); |
| 6617 | } |
| 6618 | |
| 6619 | /* |
| 6620 | * If we got specified a target task, also iterate its context and |
| 6621 | * deliver this event there too. |
| 6622 | */ |
| 6623 | if (task && task != current) { |
| 6624 | struct perf_event_context *ctx; |
| 6625 | struct trace_entry *entry = record; |
| 6626 | |
| 6627 | rcu_read_lock(); |
| 6628 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); |
| 6629 | if (!ctx) |
| 6630 | goto unlock; |
| 6631 | |
| 6632 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
| 6633 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
| 6634 | continue; |
| 6635 | if (event->attr.config != entry->type) |
| 6636 | continue; |
| 6637 | if (perf_tp_event_match(event, &data, regs)) |
| 6638 | perf_swevent_event(event, count, &data, regs); |
| 6639 | } |
| 6640 | unlock: |
| 6641 | rcu_read_unlock(); |
| 6642 | } |
| 6643 | |
| 6644 | perf_swevent_put_recursion_context(rctx); |
| 6645 | } |
| 6646 | EXPORT_SYMBOL_GPL(perf_tp_event); |
| 6647 | |
| 6648 | static void tp_perf_event_destroy(struct perf_event *event) |
| 6649 | { |
| 6650 | perf_trace_destroy(event); |
| 6651 | } |
| 6652 | |
| 6653 | static int perf_tp_event_init(struct perf_event *event) |
| 6654 | { |
| 6655 | int err; |
| 6656 | |
| 6657 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
| 6658 | return -ENOENT; |
| 6659 | |
| 6660 | /* |
| 6661 | * no branch sampling for tracepoint events |
| 6662 | */ |
| 6663 | if (has_branch_stack(event)) |
| 6664 | return -EOPNOTSUPP; |
| 6665 | |
| 6666 | err = perf_trace_init(event); |
| 6667 | if (err) |
| 6668 | return err; |
| 6669 | |
| 6670 | event->destroy = tp_perf_event_destroy; |
| 6671 | |
| 6672 | return 0; |
| 6673 | } |
| 6674 | |
| 6675 | static struct pmu perf_tracepoint = { |
| 6676 | .task_ctx_nr = perf_sw_context, |
| 6677 | |
| 6678 | .event_init = perf_tp_event_init, |
| 6679 | .add = perf_trace_add, |
| 6680 | .del = perf_trace_del, |
| 6681 | .start = perf_swevent_start, |
| 6682 | .stop = perf_swevent_stop, |
| 6683 | .read = perf_swevent_read, |
| 6684 | }; |
| 6685 | |
| 6686 | static inline void perf_tp_register(void) |
| 6687 | { |
| 6688 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
| 6689 | } |
| 6690 | |
| 6691 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) |
| 6692 | { |
| 6693 | char *filter_str; |
| 6694 | int ret; |
| 6695 | |
| 6696 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
| 6697 | return -EINVAL; |
| 6698 | |
| 6699 | filter_str = strndup_user(arg, PAGE_SIZE); |
| 6700 | if (IS_ERR(filter_str)) |
| 6701 | return PTR_ERR(filter_str); |
| 6702 | |
| 6703 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); |
| 6704 | |
| 6705 | kfree(filter_str); |
| 6706 | return ret; |
| 6707 | } |
| 6708 | |
| 6709 | static void perf_event_free_filter(struct perf_event *event) |
| 6710 | { |
| 6711 | ftrace_profile_free_filter(event); |
| 6712 | } |
| 6713 | |
| 6714 | static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd) |
| 6715 | { |
| 6716 | struct bpf_prog *prog; |
| 6717 | |
| 6718 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
| 6719 | return -EINVAL; |
| 6720 | |
| 6721 | if (event->tp_event->prog) |
| 6722 | return -EEXIST; |
| 6723 | |
| 6724 | if (!(event->tp_event->flags & TRACE_EVENT_FL_KPROBE)) |
| 6725 | /* bpf programs can only be attached to kprobes */ |
| 6726 | return -EINVAL; |
| 6727 | |
| 6728 | prog = bpf_prog_get(prog_fd); |
| 6729 | if (IS_ERR(prog)) |
| 6730 | return PTR_ERR(prog); |
| 6731 | |
| 6732 | if (prog->type != BPF_PROG_TYPE_KPROBE) { |
| 6733 | /* valid fd, but invalid bpf program type */ |
| 6734 | bpf_prog_put(prog); |
| 6735 | return -EINVAL; |
| 6736 | } |
| 6737 | |
| 6738 | event->tp_event->prog = prog; |
| 6739 | |
| 6740 | return 0; |
| 6741 | } |
| 6742 | |
| 6743 | static void perf_event_free_bpf_prog(struct perf_event *event) |
| 6744 | { |
| 6745 | struct bpf_prog *prog; |
| 6746 | |
| 6747 | if (!event->tp_event) |
| 6748 | return; |
| 6749 | |
| 6750 | prog = event->tp_event->prog; |
| 6751 | if (prog) { |
| 6752 | event->tp_event->prog = NULL; |
| 6753 | bpf_prog_put(prog); |
| 6754 | } |
| 6755 | } |
| 6756 | |
| 6757 | #else |
| 6758 | |
| 6759 | static inline void perf_tp_register(void) |
| 6760 | { |
| 6761 | } |
| 6762 | |
| 6763 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) |
| 6764 | { |
| 6765 | return -ENOENT; |
| 6766 | } |
| 6767 | |
| 6768 | static void perf_event_free_filter(struct perf_event *event) |
| 6769 | { |
| 6770 | } |
| 6771 | |
| 6772 | static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd) |
| 6773 | { |
| 6774 | return -ENOENT; |
| 6775 | } |
| 6776 | |
| 6777 | static void perf_event_free_bpf_prog(struct perf_event *event) |
| 6778 | { |
| 6779 | } |
| 6780 | #endif /* CONFIG_EVENT_TRACING */ |
| 6781 | |
| 6782 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
| 6783 | void perf_bp_event(struct perf_event *bp, void *data) |
| 6784 | { |
| 6785 | struct perf_sample_data sample; |
| 6786 | struct pt_regs *regs = data; |
| 6787 | |
| 6788 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
| 6789 | |
| 6790 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
| 6791 | perf_swevent_event(bp, 1, &sample, regs); |
| 6792 | } |
| 6793 | #endif |
| 6794 | |
| 6795 | /* |
| 6796 | * hrtimer based swevent callback |
| 6797 | */ |
| 6798 | |
| 6799 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
| 6800 | { |
| 6801 | enum hrtimer_restart ret = HRTIMER_RESTART; |
| 6802 | struct perf_sample_data data; |
| 6803 | struct pt_regs *regs; |
| 6804 | struct perf_event *event; |
| 6805 | u64 period; |
| 6806 | |
| 6807 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
| 6808 | |
| 6809 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
| 6810 | return HRTIMER_NORESTART; |
| 6811 | |
| 6812 | event->pmu->read(event); |
| 6813 | |
| 6814 | perf_sample_data_init(&data, 0, event->hw.last_period); |
| 6815 | regs = get_irq_regs(); |
| 6816 | |
| 6817 | if (regs && !perf_exclude_event(event, regs)) { |
| 6818 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
| 6819 | if (__perf_event_overflow(event, 1, &data, regs)) |
| 6820 | ret = HRTIMER_NORESTART; |
| 6821 | } |
| 6822 | |
| 6823 | period = max_t(u64, 10000, event->hw.sample_period); |
| 6824 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); |
| 6825 | |
| 6826 | return ret; |
| 6827 | } |
| 6828 | |
| 6829 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
| 6830 | { |
| 6831 | struct hw_perf_event *hwc = &event->hw; |
| 6832 | s64 period; |
| 6833 | |
| 6834 | if (!is_sampling_event(event)) |
| 6835 | return; |
| 6836 | |
| 6837 | period = local64_read(&hwc->period_left); |
| 6838 | if (period) { |
| 6839 | if (period < 0) |
| 6840 | period = 10000; |
| 6841 | |
| 6842 | local64_set(&hwc->period_left, 0); |
| 6843 | } else { |
| 6844 | period = max_t(u64, 10000, hwc->sample_period); |
| 6845 | } |
| 6846 | __hrtimer_start_range_ns(&hwc->hrtimer, |
| 6847 | ns_to_ktime(period), 0, |
| 6848 | HRTIMER_MODE_REL_PINNED, 0); |
| 6849 | } |
| 6850 | |
| 6851 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) |
| 6852 | { |
| 6853 | struct hw_perf_event *hwc = &event->hw; |
| 6854 | |
| 6855 | if (is_sampling_event(event)) { |
| 6856 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
| 6857 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
| 6858 | |
| 6859 | hrtimer_cancel(&hwc->hrtimer); |
| 6860 | } |
| 6861 | } |
| 6862 | |
| 6863 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
| 6864 | { |
| 6865 | struct hw_perf_event *hwc = &event->hw; |
| 6866 | |
| 6867 | if (!is_sampling_event(event)) |
| 6868 | return; |
| 6869 | |
| 6870 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
| 6871 | hwc->hrtimer.function = perf_swevent_hrtimer; |
| 6872 | |
| 6873 | /* |
| 6874 | * Since hrtimers have a fixed rate, we can do a static freq->period |
| 6875 | * mapping and avoid the whole period adjust feedback stuff. |
| 6876 | */ |
| 6877 | if (event->attr.freq) { |
| 6878 | long freq = event->attr.sample_freq; |
| 6879 | |
| 6880 | event->attr.sample_period = NSEC_PER_SEC / freq; |
| 6881 | hwc->sample_period = event->attr.sample_period; |
| 6882 | local64_set(&hwc->period_left, hwc->sample_period); |
| 6883 | hwc->last_period = hwc->sample_period; |
| 6884 | event->attr.freq = 0; |
| 6885 | } |
| 6886 | } |
| 6887 | |
| 6888 | /* |
| 6889 | * Software event: cpu wall time clock |
| 6890 | */ |
| 6891 | |
| 6892 | static void cpu_clock_event_update(struct perf_event *event) |
| 6893 | { |
| 6894 | s64 prev; |
| 6895 | u64 now; |
| 6896 | |
| 6897 | now = local_clock(); |
| 6898 | prev = local64_xchg(&event->hw.prev_count, now); |
| 6899 | local64_add(now - prev, &event->count); |
| 6900 | } |
| 6901 | |
| 6902 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
| 6903 | { |
| 6904 | local64_set(&event->hw.prev_count, local_clock()); |
| 6905 | perf_swevent_start_hrtimer(event); |
| 6906 | } |
| 6907 | |
| 6908 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
| 6909 | { |
| 6910 | perf_swevent_cancel_hrtimer(event); |
| 6911 | cpu_clock_event_update(event); |
| 6912 | } |
| 6913 | |
| 6914 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
| 6915 | { |
| 6916 | if (flags & PERF_EF_START) |
| 6917 | cpu_clock_event_start(event, flags); |
| 6918 | perf_event_update_userpage(event); |
| 6919 | |
| 6920 | return 0; |
| 6921 | } |
| 6922 | |
| 6923 | static void cpu_clock_event_del(struct perf_event *event, int flags) |
| 6924 | { |
| 6925 | cpu_clock_event_stop(event, flags); |
| 6926 | } |
| 6927 | |
| 6928 | static void cpu_clock_event_read(struct perf_event *event) |
| 6929 | { |
| 6930 | cpu_clock_event_update(event); |
| 6931 | } |
| 6932 | |
| 6933 | static int cpu_clock_event_init(struct perf_event *event) |
| 6934 | { |
| 6935 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
| 6936 | return -ENOENT; |
| 6937 | |
| 6938 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) |
| 6939 | return -ENOENT; |
| 6940 | |
| 6941 | /* |
| 6942 | * no branch sampling for software events |
| 6943 | */ |
| 6944 | if (has_branch_stack(event)) |
| 6945 | return -EOPNOTSUPP; |
| 6946 | |
| 6947 | perf_swevent_init_hrtimer(event); |
| 6948 | |
| 6949 | return 0; |
| 6950 | } |
| 6951 | |
| 6952 | static struct pmu perf_cpu_clock = { |
| 6953 | .task_ctx_nr = perf_sw_context, |
| 6954 | |
| 6955 | .capabilities = PERF_PMU_CAP_NO_NMI, |
| 6956 | |
| 6957 | .event_init = cpu_clock_event_init, |
| 6958 | .add = cpu_clock_event_add, |
| 6959 | .del = cpu_clock_event_del, |
| 6960 | .start = cpu_clock_event_start, |
| 6961 | .stop = cpu_clock_event_stop, |
| 6962 | .read = cpu_clock_event_read, |
| 6963 | }; |
| 6964 | |
| 6965 | /* |
| 6966 | * Software event: task time clock |
| 6967 | */ |
| 6968 | |
| 6969 | static void task_clock_event_update(struct perf_event *event, u64 now) |
| 6970 | { |
| 6971 | u64 prev; |
| 6972 | s64 delta; |
| 6973 | |
| 6974 | prev = local64_xchg(&event->hw.prev_count, now); |
| 6975 | delta = now - prev; |
| 6976 | local64_add(delta, &event->count); |
| 6977 | } |
| 6978 | |
| 6979 | static void task_clock_event_start(struct perf_event *event, int flags) |
| 6980 | { |
| 6981 | local64_set(&event->hw.prev_count, event->ctx->time); |
| 6982 | perf_swevent_start_hrtimer(event); |
| 6983 | } |
| 6984 | |
| 6985 | static void task_clock_event_stop(struct perf_event *event, int flags) |
| 6986 | { |
| 6987 | perf_swevent_cancel_hrtimer(event); |
| 6988 | task_clock_event_update(event, event->ctx->time); |
| 6989 | } |
| 6990 | |
| 6991 | static int task_clock_event_add(struct perf_event *event, int flags) |
| 6992 | { |
| 6993 | if (flags & PERF_EF_START) |
| 6994 | task_clock_event_start(event, flags); |
| 6995 | perf_event_update_userpage(event); |
| 6996 | |
| 6997 | return 0; |
| 6998 | } |
| 6999 | |
| 7000 | static void task_clock_event_del(struct perf_event *event, int flags) |
| 7001 | { |
| 7002 | task_clock_event_stop(event, PERF_EF_UPDATE); |
| 7003 | } |
| 7004 | |
| 7005 | static void task_clock_event_read(struct perf_event *event) |
| 7006 | { |
| 7007 | u64 now = perf_clock(); |
| 7008 | u64 delta = now - event->ctx->timestamp; |
| 7009 | u64 time = event->ctx->time + delta; |
| 7010 | |
| 7011 | task_clock_event_update(event, time); |
| 7012 | } |
| 7013 | |
| 7014 | static int task_clock_event_init(struct perf_event *event) |
| 7015 | { |
| 7016 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
| 7017 | return -ENOENT; |
| 7018 | |
| 7019 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) |
| 7020 | return -ENOENT; |
| 7021 | |
| 7022 | /* |
| 7023 | * no branch sampling for software events |
| 7024 | */ |
| 7025 | if (has_branch_stack(event)) |
| 7026 | return -EOPNOTSUPP; |
| 7027 | |
| 7028 | perf_swevent_init_hrtimer(event); |
| 7029 | |
| 7030 | return 0; |
| 7031 | } |
| 7032 | |
| 7033 | static struct pmu perf_task_clock = { |
| 7034 | .task_ctx_nr = perf_sw_context, |
| 7035 | |
| 7036 | .capabilities = PERF_PMU_CAP_NO_NMI, |
| 7037 | |
| 7038 | .event_init = task_clock_event_init, |
| 7039 | .add = task_clock_event_add, |
| 7040 | .del = task_clock_event_del, |
| 7041 | .start = task_clock_event_start, |
| 7042 | .stop = task_clock_event_stop, |
| 7043 | .read = task_clock_event_read, |
| 7044 | }; |
| 7045 | |
| 7046 | static void perf_pmu_nop_void(struct pmu *pmu) |
| 7047 | { |
| 7048 | } |
| 7049 | |
| 7050 | static int perf_pmu_nop_int(struct pmu *pmu) |
| 7051 | { |
| 7052 | return 0; |
| 7053 | } |
| 7054 | |
| 7055 | static void perf_pmu_start_txn(struct pmu *pmu) |
| 7056 | { |
| 7057 | perf_pmu_disable(pmu); |
| 7058 | } |
| 7059 | |
| 7060 | static int perf_pmu_commit_txn(struct pmu *pmu) |
| 7061 | { |
| 7062 | perf_pmu_enable(pmu); |
| 7063 | return 0; |
| 7064 | } |
| 7065 | |
| 7066 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
| 7067 | { |
| 7068 | perf_pmu_enable(pmu); |
| 7069 | } |
| 7070 | |
| 7071 | static int perf_event_idx_default(struct perf_event *event) |
| 7072 | { |
| 7073 | return 0; |
| 7074 | } |
| 7075 | |
| 7076 | /* |
| 7077 | * Ensures all contexts with the same task_ctx_nr have the same |
| 7078 | * pmu_cpu_context too. |
| 7079 | */ |
| 7080 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
| 7081 | { |
| 7082 | struct pmu *pmu; |
| 7083 | |
| 7084 | if (ctxn < 0) |
| 7085 | return NULL; |
| 7086 | |
| 7087 | list_for_each_entry(pmu, &pmus, entry) { |
| 7088 | if (pmu->task_ctx_nr == ctxn) |
| 7089 | return pmu->pmu_cpu_context; |
| 7090 | } |
| 7091 | |
| 7092 | return NULL; |
| 7093 | } |
| 7094 | |
| 7095 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
| 7096 | { |
| 7097 | int cpu; |
| 7098 | |
| 7099 | for_each_possible_cpu(cpu) { |
| 7100 | struct perf_cpu_context *cpuctx; |
| 7101 | |
| 7102 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
| 7103 | |
| 7104 | if (cpuctx->unique_pmu == old_pmu) |
| 7105 | cpuctx->unique_pmu = pmu; |
| 7106 | } |
| 7107 | } |
| 7108 | |
| 7109 | static void free_pmu_context(struct pmu *pmu) |
| 7110 | { |
| 7111 | struct pmu *i; |
| 7112 | |
| 7113 | mutex_lock(&pmus_lock); |
| 7114 | /* |
| 7115 | * Like a real lame refcount. |
| 7116 | */ |
| 7117 | list_for_each_entry(i, &pmus, entry) { |
| 7118 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { |
| 7119 | update_pmu_context(i, pmu); |
| 7120 | goto out; |
| 7121 | } |
| 7122 | } |
| 7123 | |
| 7124 | free_percpu(pmu->pmu_cpu_context); |
| 7125 | out: |
| 7126 | mutex_unlock(&pmus_lock); |
| 7127 | } |
| 7128 | static struct idr pmu_idr; |
| 7129 | |
| 7130 | static ssize_t |
| 7131 | type_show(struct device *dev, struct device_attribute *attr, char *page) |
| 7132 | { |
| 7133 | struct pmu *pmu = dev_get_drvdata(dev); |
| 7134 | |
| 7135 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); |
| 7136 | } |
| 7137 | static DEVICE_ATTR_RO(type); |
| 7138 | |
| 7139 | static ssize_t |
| 7140 | perf_event_mux_interval_ms_show(struct device *dev, |
| 7141 | struct device_attribute *attr, |
| 7142 | char *page) |
| 7143 | { |
| 7144 | struct pmu *pmu = dev_get_drvdata(dev); |
| 7145 | |
| 7146 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); |
| 7147 | } |
| 7148 | |
| 7149 | static ssize_t |
| 7150 | perf_event_mux_interval_ms_store(struct device *dev, |
| 7151 | struct device_attribute *attr, |
| 7152 | const char *buf, size_t count) |
| 7153 | { |
| 7154 | struct pmu *pmu = dev_get_drvdata(dev); |
| 7155 | int timer, cpu, ret; |
| 7156 | |
| 7157 | ret = kstrtoint(buf, 0, &timer); |
| 7158 | if (ret) |
| 7159 | return ret; |
| 7160 | |
| 7161 | if (timer < 1) |
| 7162 | return -EINVAL; |
| 7163 | |
| 7164 | /* same value, noting to do */ |
| 7165 | if (timer == pmu->hrtimer_interval_ms) |
| 7166 | return count; |
| 7167 | |
| 7168 | pmu->hrtimer_interval_ms = timer; |
| 7169 | |
| 7170 | /* update all cpuctx for this PMU */ |
| 7171 | for_each_possible_cpu(cpu) { |
| 7172 | struct perf_cpu_context *cpuctx; |
| 7173 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
| 7174 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); |
| 7175 | |
| 7176 | if (hrtimer_active(&cpuctx->hrtimer)) |
| 7177 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); |
| 7178 | } |
| 7179 | |
| 7180 | return count; |
| 7181 | } |
| 7182 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
| 7183 | |
| 7184 | static struct attribute *pmu_dev_attrs[] = { |
| 7185 | &dev_attr_type.attr, |
| 7186 | &dev_attr_perf_event_mux_interval_ms.attr, |
| 7187 | NULL, |
| 7188 | }; |
| 7189 | ATTRIBUTE_GROUPS(pmu_dev); |
| 7190 | |
| 7191 | static int pmu_bus_running; |
| 7192 | static struct bus_type pmu_bus = { |
| 7193 | .name = "event_source", |
| 7194 | .dev_groups = pmu_dev_groups, |
| 7195 | }; |
| 7196 | |
| 7197 | static void pmu_dev_release(struct device *dev) |
| 7198 | { |
| 7199 | kfree(dev); |
| 7200 | } |
| 7201 | |
| 7202 | static int pmu_dev_alloc(struct pmu *pmu) |
| 7203 | { |
| 7204 | int ret = -ENOMEM; |
| 7205 | |
| 7206 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); |
| 7207 | if (!pmu->dev) |
| 7208 | goto out; |
| 7209 | |
| 7210 | pmu->dev->groups = pmu->attr_groups; |
| 7211 | device_initialize(pmu->dev); |
| 7212 | ret = dev_set_name(pmu->dev, "%s", pmu->name); |
| 7213 | if (ret) |
| 7214 | goto free_dev; |
| 7215 | |
| 7216 | dev_set_drvdata(pmu->dev, pmu); |
| 7217 | pmu->dev->bus = &pmu_bus; |
| 7218 | pmu->dev->release = pmu_dev_release; |
| 7219 | ret = device_add(pmu->dev); |
| 7220 | if (ret) |
| 7221 | goto free_dev; |
| 7222 | |
| 7223 | out: |
| 7224 | return ret; |
| 7225 | |
| 7226 | free_dev: |
| 7227 | put_device(pmu->dev); |
| 7228 | goto out; |
| 7229 | } |
| 7230 | |
| 7231 | static struct lock_class_key cpuctx_mutex; |
| 7232 | static struct lock_class_key cpuctx_lock; |
| 7233 | |
| 7234 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
| 7235 | { |
| 7236 | int cpu, ret; |
| 7237 | |
| 7238 | mutex_lock(&pmus_lock); |
| 7239 | ret = -ENOMEM; |
| 7240 | pmu->pmu_disable_count = alloc_percpu(int); |
| 7241 | if (!pmu->pmu_disable_count) |
| 7242 | goto unlock; |
| 7243 | |
| 7244 | pmu->type = -1; |
| 7245 | if (!name) |
| 7246 | goto skip_type; |
| 7247 | pmu->name = name; |
| 7248 | |
| 7249 | if (type < 0) { |
| 7250 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
| 7251 | if (type < 0) { |
| 7252 | ret = type; |
| 7253 | goto free_pdc; |
| 7254 | } |
| 7255 | } |
| 7256 | pmu->type = type; |
| 7257 | |
| 7258 | if (pmu_bus_running) { |
| 7259 | ret = pmu_dev_alloc(pmu); |
| 7260 | if (ret) |
| 7261 | goto free_idr; |
| 7262 | } |
| 7263 | |
| 7264 | skip_type: |
| 7265 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
| 7266 | if (pmu->pmu_cpu_context) |
| 7267 | goto got_cpu_context; |
| 7268 | |
| 7269 | ret = -ENOMEM; |
| 7270 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
| 7271 | if (!pmu->pmu_cpu_context) |
| 7272 | goto free_dev; |
| 7273 | |
| 7274 | for_each_possible_cpu(cpu) { |
| 7275 | struct perf_cpu_context *cpuctx; |
| 7276 | |
| 7277 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
| 7278 | __perf_event_init_context(&cpuctx->ctx); |
| 7279 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
| 7280 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
| 7281 | cpuctx->ctx.pmu = pmu; |
| 7282 | |
| 7283 | __perf_cpu_hrtimer_init(cpuctx, cpu); |
| 7284 | |
| 7285 | cpuctx->unique_pmu = pmu; |
| 7286 | } |
| 7287 | |
| 7288 | got_cpu_context: |
| 7289 | if (!pmu->start_txn) { |
| 7290 | if (pmu->pmu_enable) { |
| 7291 | /* |
| 7292 | * If we have pmu_enable/pmu_disable calls, install |
| 7293 | * transaction stubs that use that to try and batch |
| 7294 | * hardware accesses. |
| 7295 | */ |
| 7296 | pmu->start_txn = perf_pmu_start_txn; |
| 7297 | pmu->commit_txn = perf_pmu_commit_txn; |
| 7298 | pmu->cancel_txn = perf_pmu_cancel_txn; |
| 7299 | } else { |
| 7300 | pmu->start_txn = perf_pmu_nop_void; |
| 7301 | pmu->commit_txn = perf_pmu_nop_int; |
| 7302 | pmu->cancel_txn = perf_pmu_nop_void; |
| 7303 | } |
| 7304 | } |
| 7305 | |
| 7306 | if (!pmu->pmu_enable) { |
| 7307 | pmu->pmu_enable = perf_pmu_nop_void; |
| 7308 | pmu->pmu_disable = perf_pmu_nop_void; |
| 7309 | } |
| 7310 | |
| 7311 | if (!pmu->event_idx) |
| 7312 | pmu->event_idx = perf_event_idx_default; |
| 7313 | |
| 7314 | list_add_rcu(&pmu->entry, &pmus); |
| 7315 | atomic_set(&pmu->exclusive_cnt, 0); |
| 7316 | ret = 0; |
| 7317 | unlock: |
| 7318 | mutex_unlock(&pmus_lock); |
| 7319 | |
| 7320 | return ret; |
| 7321 | |
| 7322 | free_dev: |
| 7323 | device_del(pmu->dev); |
| 7324 | put_device(pmu->dev); |
| 7325 | |
| 7326 | free_idr: |
| 7327 | if (pmu->type >= PERF_TYPE_MAX) |
| 7328 | idr_remove(&pmu_idr, pmu->type); |
| 7329 | |
| 7330 | free_pdc: |
| 7331 | free_percpu(pmu->pmu_disable_count); |
| 7332 | goto unlock; |
| 7333 | } |
| 7334 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
| 7335 | |
| 7336 | void perf_pmu_unregister(struct pmu *pmu) |
| 7337 | { |
| 7338 | mutex_lock(&pmus_lock); |
| 7339 | list_del_rcu(&pmu->entry); |
| 7340 | mutex_unlock(&pmus_lock); |
| 7341 | |
| 7342 | /* |
| 7343 | * We dereference the pmu list under both SRCU and regular RCU, so |
| 7344 | * synchronize against both of those. |
| 7345 | */ |
| 7346 | synchronize_srcu(&pmus_srcu); |
| 7347 | synchronize_rcu(); |
| 7348 | |
| 7349 | free_percpu(pmu->pmu_disable_count); |
| 7350 | if (pmu->type >= PERF_TYPE_MAX) |
| 7351 | idr_remove(&pmu_idr, pmu->type); |
| 7352 | device_del(pmu->dev); |
| 7353 | put_device(pmu->dev); |
| 7354 | free_pmu_context(pmu); |
| 7355 | } |
| 7356 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
| 7357 | |
| 7358 | static int perf_try_init_event(struct pmu *pmu, struct perf_event *event) |
| 7359 | { |
| 7360 | struct perf_event_context *ctx = NULL; |
| 7361 | int ret; |
| 7362 | |
| 7363 | if (!try_module_get(pmu->module)) |
| 7364 | return -ENODEV; |
| 7365 | |
| 7366 | if (event->group_leader != event) { |
| 7367 | ctx = perf_event_ctx_lock(event->group_leader); |
| 7368 | BUG_ON(!ctx); |
| 7369 | } |
| 7370 | |
| 7371 | event->pmu = pmu; |
| 7372 | ret = pmu->event_init(event); |
| 7373 | |
| 7374 | if (ctx) |
| 7375 | perf_event_ctx_unlock(event->group_leader, ctx); |
| 7376 | |
| 7377 | if (ret) |
| 7378 | module_put(pmu->module); |
| 7379 | |
| 7380 | return ret; |
| 7381 | } |
| 7382 | |
| 7383 | struct pmu *perf_init_event(struct perf_event *event) |
| 7384 | { |
| 7385 | struct pmu *pmu = NULL; |
| 7386 | int idx; |
| 7387 | int ret; |
| 7388 | |
| 7389 | idx = srcu_read_lock(&pmus_srcu); |
| 7390 | |
| 7391 | rcu_read_lock(); |
| 7392 | pmu = idr_find(&pmu_idr, event->attr.type); |
| 7393 | rcu_read_unlock(); |
| 7394 | if (pmu) { |
| 7395 | ret = perf_try_init_event(pmu, event); |
| 7396 | if (ret) |
| 7397 | pmu = ERR_PTR(ret); |
| 7398 | goto unlock; |
| 7399 | } |
| 7400 | |
| 7401 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
| 7402 | ret = perf_try_init_event(pmu, event); |
| 7403 | if (!ret) |
| 7404 | goto unlock; |
| 7405 | |
| 7406 | if (ret != -ENOENT) { |
| 7407 | pmu = ERR_PTR(ret); |
| 7408 | goto unlock; |
| 7409 | } |
| 7410 | } |
| 7411 | pmu = ERR_PTR(-ENOENT); |
| 7412 | unlock: |
| 7413 | srcu_read_unlock(&pmus_srcu, idx); |
| 7414 | |
| 7415 | return pmu; |
| 7416 | } |
| 7417 | |
| 7418 | static void account_event_cpu(struct perf_event *event, int cpu) |
| 7419 | { |
| 7420 | if (event->parent) |
| 7421 | return; |
| 7422 | |
| 7423 | if (is_cgroup_event(event)) |
| 7424 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); |
| 7425 | } |
| 7426 | |
| 7427 | static void account_event(struct perf_event *event) |
| 7428 | { |
| 7429 | if (event->parent) |
| 7430 | return; |
| 7431 | |
| 7432 | if (event->attach_state & PERF_ATTACH_TASK) |
| 7433 | static_key_slow_inc(&perf_sched_events.key); |
| 7434 | if (event->attr.mmap || event->attr.mmap_data) |
| 7435 | atomic_inc(&nr_mmap_events); |
| 7436 | if (event->attr.comm) |
| 7437 | atomic_inc(&nr_comm_events); |
| 7438 | if (event->attr.task) |
| 7439 | atomic_inc(&nr_task_events); |
| 7440 | if (event->attr.freq) { |
| 7441 | if (atomic_inc_return(&nr_freq_events) == 1) |
| 7442 | tick_nohz_full_kick_all(); |
| 7443 | } |
| 7444 | if (has_branch_stack(event)) |
| 7445 | static_key_slow_inc(&perf_sched_events.key); |
| 7446 | if (is_cgroup_event(event)) |
| 7447 | static_key_slow_inc(&perf_sched_events.key); |
| 7448 | |
| 7449 | account_event_cpu(event, event->cpu); |
| 7450 | } |
| 7451 | |
| 7452 | /* |
| 7453 | * Allocate and initialize a event structure |
| 7454 | */ |
| 7455 | static struct perf_event * |
| 7456 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
| 7457 | struct task_struct *task, |
| 7458 | struct perf_event *group_leader, |
| 7459 | struct perf_event *parent_event, |
| 7460 | perf_overflow_handler_t overflow_handler, |
| 7461 | void *context, int cgroup_fd) |
| 7462 | { |
| 7463 | struct pmu *pmu; |
| 7464 | struct perf_event *event; |
| 7465 | struct hw_perf_event *hwc; |
| 7466 | long err = -EINVAL; |
| 7467 | |
| 7468 | if ((unsigned)cpu >= nr_cpu_ids) { |
| 7469 | if (!task || cpu != -1) |
| 7470 | return ERR_PTR(-EINVAL); |
| 7471 | } |
| 7472 | |
| 7473 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
| 7474 | if (!event) |
| 7475 | return ERR_PTR(-ENOMEM); |
| 7476 | |
| 7477 | /* |
| 7478 | * Single events are their own group leaders, with an |
| 7479 | * empty sibling list: |
| 7480 | */ |
| 7481 | if (!group_leader) |
| 7482 | group_leader = event; |
| 7483 | |
| 7484 | mutex_init(&event->child_mutex); |
| 7485 | INIT_LIST_HEAD(&event->child_list); |
| 7486 | |
| 7487 | INIT_LIST_HEAD(&event->group_entry); |
| 7488 | INIT_LIST_HEAD(&event->event_entry); |
| 7489 | INIT_LIST_HEAD(&event->sibling_list); |
| 7490 | INIT_LIST_HEAD(&event->rb_entry); |
| 7491 | INIT_LIST_HEAD(&event->active_entry); |
| 7492 | INIT_HLIST_NODE(&event->hlist_entry); |
| 7493 | |
| 7494 | |
| 7495 | init_waitqueue_head(&event->waitq); |
| 7496 | init_irq_work(&event->pending, perf_pending_event); |
| 7497 | |
| 7498 | mutex_init(&event->mmap_mutex); |
| 7499 | |
| 7500 | atomic_long_set(&event->refcount, 1); |
| 7501 | event->cpu = cpu; |
| 7502 | event->attr = *attr; |
| 7503 | event->group_leader = group_leader; |
| 7504 | event->pmu = NULL; |
| 7505 | event->oncpu = -1; |
| 7506 | |
| 7507 | event->parent = parent_event; |
| 7508 | |
| 7509 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
| 7510 | event->id = atomic64_inc_return(&perf_event_id); |
| 7511 | |
| 7512 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 7513 | |
| 7514 | if (task) { |
| 7515 | event->attach_state = PERF_ATTACH_TASK; |
| 7516 | /* |
| 7517 | * XXX pmu::event_init needs to know what task to account to |
| 7518 | * and we cannot use the ctx information because we need the |
| 7519 | * pmu before we get a ctx. |
| 7520 | */ |
| 7521 | event->hw.target = task; |
| 7522 | } |
| 7523 | |
| 7524 | event->clock = &local_clock; |
| 7525 | if (parent_event) |
| 7526 | event->clock = parent_event->clock; |
| 7527 | |
| 7528 | if (!overflow_handler && parent_event) { |
| 7529 | overflow_handler = parent_event->overflow_handler; |
| 7530 | context = parent_event->overflow_handler_context; |
| 7531 | } |
| 7532 | |
| 7533 | event->overflow_handler = overflow_handler; |
| 7534 | event->overflow_handler_context = context; |
| 7535 | |
| 7536 | perf_event__state_init(event); |
| 7537 | |
| 7538 | pmu = NULL; |
| 7539 | |
| 7540 | hwc = &event->hw; |
| 7541 | hwc->sample_period = attr->sample_period; |
| 7542 | if (attr->freq && attr->sample_freq) |
| 7543 | hwc->sample_period = 1; |
| 7544 | hwc->last_period = hwc->sample_period; |
| 7545 | |
| 7546 | local64_set(&hwc->period_left, hwc->sample_period); |
| 7547 | |
| 7548 | /* |
| 7549 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
| 7550 | */ |
| 7551 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
| 7552 | goto err_ns; |
| 7553 | |
| 7554 | if (!has_branch_stack(event)) |
| 7555 | event->attr.branch_sample_type = 0; |
| 7556 | |
| 7557 | if (cgroup_fd != -1) { |
| 7558 | err = perf_cgroup_connect(cgroup_fd, event, attr, group_leader); |
| 7559 | if (err) |
| 7560 | goto err_ns; |
| 7561 | } |
| 7562 | |
| 7563 | pmu = perf_init_event(event); |
| 7564 | if (!pmu) |
| 7565 | goto err_ns; |
| 7566 | else if (IS_ERR(pmu)) { |
| 7567 | err = PTR_ERR(pmu); |
| 7568 | goto err_ns; |
| 7569 | } |
| 7570 | |
| 7571 | err = exclusive_event_init(event); |
| 7572 | if (err) |
| 7573 | goto err_pmu; |
| 7574 | |
| 7575 | if (!event->parent) { |
| 7576 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
| 7577 | err = get_callchain_buffers(); |
| 7578 | if (err) |
| 7579 | goto err_per_task; |
| 7580 | } |
| 7581 | } |
| 7582 | |
| 7583 | return event; |
| 7584 | |
| 7585 | err_per_task: |
| 7586 | exclusive_event_destroy(event); |
| 7587 | |
| 7588 | err_pmu: |
| 7589 | if (event->destroy) |
| 7590 | event->destroy(event); |
| 7591 | module_put(pmu->module); |
| 7592 | err_ns: |
| 7593 | if (is_cgroup_event(event)) |
| 7594 | perf_detach_cgroup(event); |
| 7595 | if (event->ns) |
| 7596 | put_pid_ns(event->ns); |
| 7597 | kfree(event); |
| 7598 | |
| 7599 | return ERR_PTR(err); |
| 7600 | } |
| 7601 | |
| 7602 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
| 7603 | struct perf_event_attr *attr) |
| 7604 | { |
| 7605 | u32 size; |
| 7606 | int ret; |
| 7607 | |
| 7608 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) |
| 7609 | return -EFAULT; |
| 7610 | |
| 7611 | /* |
| 7612 | * zero the full structure, so that a short copy will be nice. |
| 7613 | */ |
| 7614 | memset(attr, 0, sizeof(*attr)); |
| 7615 | |
| 7616 | ret = get_user(size, &uattr->size); |
| 7617 | if (ret) |
| 7618 | return ret; |
| 7619 | |
| 7620 | if (size > PAGE_SIZE) /* silly large */ |
| 7621 | goto err_size; |
| 7622 | |
| 7623 | if (!size) /* abi compat */ |
| 7624 | size = PERF_ATTR_SIZE_VER0; |
| 7625 | |
| 7626 | if (size < PERF_ATTR_SIZE_VER0) |
| 7627 | goto err_size; |
| 7628 | |
| 7629 | /* |
| 7630 | * If we're handed a bigger struct than we know of, |
| 7631 | * ensure all the unknown bits are 0 - i.e. new |
| 7632 | * user-space does not rely on any kernel feature |
| 7633 | * extensions we dont know about yet. |
| 7634 | */ |
| 7635 | if (size > sizeof(*attr)) { |
| 7636 | unsigned char __user *addr; |
| 7637 | unsigned char __user *end; |
| 7638 | unsigned char val; |
| 7639 | |
| 7640 | addr = (void __user *)uattr + sizeof(*attr); |
| 7641 | end = (void __user *)uattr + size; |
| 7642 | |
| 7643 | for (; addr < end; addr++) { |
| 7644 | ret = get_user(val, addr); |
| 7645 | if (ret) |
| 7646 | return ret; |
| 7647 | if (val) |
| 7648 | goto err_size; |
| 7649 | } |
| 7650 | size = sizeof(*attr); |
| 7651 | } |
| 7652 | |
| 7653 | ret = copy_from_user(attr, uattr, size); |
| 7654 | if (ret) |
| 7655 | return -EFAULT; |
| 7656 | |
| 7657 | if (attr->__reserved_1) |
| 7658 | return -EINVAL; |
| 7659 | |
| 7660 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) |
| 7661 | return -EINVAL; |
| 7662 | |
| 7663 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) |
| 7664 | return -EINVAL; |
| 7665 | |
| 7666 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
| 7667 | u64 mask = attr->branch_sample_type; |
| 7668 | |
| 7669 | /* only using defined bits */ |
| 7670 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) |
| 7671 | return -EINVAL; |
| 7672 | |
| 7673 | /* at least one branch bit must be set */ |
| 7674 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) |
| 7675 | return -EINVAL; |
| 7676 | |
| 7677 | /* propagate priv level, when not set for branch */ |
| 7678 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { |
| 7679 | |
| 7680 | /* exclude_kernel checked on syscall entry */ |
| 7681 | if (!attr->exclude_kernel) |
| 7682 | mask |= PERF_SAMPLE_BRANCH_KERNEL; |
| 7683 | |
| 7684 | if (!attr->exclude_user) |
| 7685 | mask |= PERF_SAMPLE_BRANCH_USER; |
| 7686 | |
| 7687 | if (!attr->exclude_hv) |
| 7688 | mask |= PERF_SAMPLE_BRANCH_HV; |
| 7689 | /* |
| 7690 | * adjust user setting (for HW filter setup) |
| 7691 | */ |
| 7692 | attr->branch_sample_type = mask; |
| 7693 | } |
| 7694 | /* privileged levels capture (kernel, hv): check permissions */ |
| 7695 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) |
| 7696 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
| 7697 | return -EACCES; |
| 7698 | } |
| 7699 | |
| 7700 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
| 7701 | ret = perf_reg_validate(attr->sample_regs_user); |
| 7702 | if (ret) |
| 7703 | return ret; |
| 7704 | } |
| 7705 | |
| 7706 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { |
| 7707 | if (!arch_perf_have_user_stack_dump()) |
| 7708 | return -ENOSYS; |
| 7709 | |
| 7710 | /* |
| 7711 | * We have __u32 type for the size, but so far |
| 7712 | * we can only use __u16 as maximum due to the |
| 7713 | * __u16 sample size limit. |
| 7714 | */ |
| 7715 | if (attr->sample_stack_user >= USHRT_MAX) |
| 7716 | ret = -EINVAL; |
| 7717 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) |
| 7718 | ret = -EINVAL; |
| 7719 | } |
| 7720 | |
| 7721 | if (attr->sample_type & PERF_SAMPLE_REGS_INTR) |
| 7722 | ret = perf_reg_validate(attr->sample_regs_intr); |
| 7723 | out: |
| 7724 | return ret; |
| 7725 | |
| 7726 | err_size: |
| 7727 | put_user(sizeof(*attr), &uattr->size); |
| 7728 | ret = -E2BIG; |
| 7729 | goto out; |
| 7730 | } |
| 7731 | |
| 7732 | static int |
| 7733 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) |
| 7734 | { |
| 7735 | struct ring_buffer *rb = NULL; |
| 7736 | int ret = -EINVAL; |
| 7737 | |
| 7738 | if (!output_event) |
| 7739 | goto set; |
| 7740 | |
| 7741 | /* don't allow circular references */ |
| 7742 | if (event == output_event) |
| 7743 | goto out; |
| 7744 | |
| 7745 | /* |
| 7746 | * Don't allow cross-cpu buffers |
| 7747 | */ |
| 7748 | if (output_event->cpu != event->cpu) |
| 7749 | goto out; |
| 7750 | |
| 7751 | /* |
| 7752 | * If its not a per-cpu rb, it must be the same task. |
| 7753 | */ |
| 7754 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) |
| 7755 | goto out; |
| 7756 | |
| 7757 | /* |
| 7758 | * Mixing clocks in the same buffer is trouble you don't need. |
| 7759 | */ |
| 7760 | if (output_event->clock != event->clock) |
| 7761 | goto out; |
| 7762 | |
| 7763 | /* |
| 7764 | * If both events generate aux data, they must be on the same PMU |
| 7765 | */ |
| 7766 | if (has_aux(event) && has_aux(output_event) && |
| 7767 | event->pmu != output_event->pmu) |
| 7768 | goto out; |
| 7769 | |
| 7770 | set: |
| 7771 | mutex_lock(&event->mmap_mutex); |
| 7772 | /* Can't redirect output if we've got an active mmap() */ |
| 7773 | if (atomic_read(&event->mmap_count)) |
| 7774 | goto unlock; |
| 7775 | |
| 7776 | if (output_event) { |
| 7777 | /* get the rb we want to redirect to */ |
| 7778 | rb = ring_buffer_get(output_event); |
| 7779 | if (!rb) |
| 7780 | goto unlock; |
| 7781 | } |
| 7782 | |
| 7783 | ring_buffer_attach(event, rb); |
| 7784 | |
| 7785 | ret = 0; |
| 7786 | unlock: |
| 7787 | mutex_unlock(&event->mmap_mutex); |
| 7788 | |
| 7789 | out: |
| 7790 | return ret; |
| 7791 | } |
| 7792 | |
| 7793 | static void mutex_lock_double(struct mutex *a, struct mutex *b) |
| 7794 | { |
| 7795 | if (b < a) |
| 7796 | swap(a, b); |
| 7797 | |
| 7798 | mutex_lock(a); |
| 7799 | mutex_lock_nested(b, SINGLE_DEPTH_NESTING); |
| 7800 | } |
| 7801 | |
| 7802 | static int perf_event_set_clock(struct perf_event *event, clockid_t clk_id) |
| 7803 | { |
| 7804 | bool nmi_safe = false; |
| 7805 | |
| 7806 | switch (clk_id) { |
| 7807 | case CLOCK_MONOTONIC: |
| 7808 | event->clock = &ktime_get_mono_fast_ns; |
| 7809 | nmi_safe = true; |
| 7810 | break; |
| 7811 | |
| 7812 | case CLOCK_MONOTONIC_RAW: |
| 7813 | event->clock = &ktime_get_raw_fast_ns; |
| 7814 | nmi_safe = true; |
| 7815 | break; |
| 7816 | |
| 7817 | case CLOCK_REALTIME: |
| 7818 | event->clock = &ktime_get_real_ns; |
| 7819 | break; |
| 7820 | |
| 7821 | case CLOCK_BOOTTIME: |
| 7822 | event->clock = &ktime_get_boot_ns; |
| 7823 | break; |
| 7824 | |
| 7825 | case CLOCK_TAI: |
| 7826 | event->clock = &ktime_get_tai_ns; |
| 7827 | break; |
| 7828 | |
| 7829 | default: |
| 7830 | return -EINVAL; |
| 7831 | } |
| 7832 | |
| 7833 | if (!nmi_safe && !(event->pmu->capabilities & PERF_PMU_CAP_NO_NMI)) |
| 7834 | return -EINVAL; |
| 7835 | |
| 7836 | return 0; |
| 7837 | } |
| 7838 | |
| 7839 | /** |
| 7840 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
| 7841 | * |
| 7842 | * @attr_uptr: event_id type attributes for monitoring/sampling |
| 7843 | * @pid: target pid |
| 7844 | * @cpu: target cpu |
| 7845 | * @group_fd: group leader event fd |
| 7846 | */ |
| 7847 | SYSCALL_DEFINE5(perf_event_open, |
| 7848 | struct perf_event_attr __user *, attr_uptr, |
| 7849 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
| 7850 | { |
| 7851 | struct perf_event *group_leader = NULL, *output_event = NULL; |
| 7852 | struct perf_event *event, *sibling; |
| 7853 | struct perf_event_attr attr; |
| 7854 | struct perf_event_context *ctx, *uninitialized_var(gctx); |
| 7855 | struct file *event_file = NULL; |
| 7856 | struct fd group = {NULL, 0}; |
| 7857 | struct task_struct *task = NULL; |
| 7858 | struct pmu *pmu; |
| 7859 | int event_fd; |
| 7860 | int move_group = 0; |
| 7861 | int err; |
| 7862 | int f_flags = O_RDWR; |
| 7863 | int cgroup_fd = -1; |
| 7864 | |
| 7865 | /* for future expandability... */ |
| 7866 | if (flags & ~PERF_FLAG_ALL) |
| 7867 | return -EINVAL; |
| 7868 | |
| 7869 | err = perf_copy_attr(attr_uptr, &attr); |
| 7870 | if (err) |
| 7871 | return err; |
| 7872 | |
| 7873 | if (!attr.exclude_kernel) { |
| 7874 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
| 7875 | return -EACCES; |
| 7876 | } |
| 7877 | |
| 7878 | if (attr.freq) { |
| 7879 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
| 7880 | return -EINVAL; |
| 7881 | } else { |
| 7882 | if (attr.sample_period & (1ULL << 63)) |
| 7883 | return -EINVAL; |
| 7884 | } |
| 7885 | |
| 7886 | /* |
| 7887 | * In cgroup mode, the pid argument is used to pass the fd |
| 7888 | * opened to the cgroup directory in cgroupfs. The cpu argument |
| 7889 | * designates the cpu on which to monitor threads from that |
| 7890 | * cgroup. |
| 7891 | */ |
| 7892 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) |
| 7893 | return -EINVAL; |
| 7894 | |
| 7895 | if (flags & PERF_FLAG_FD_CLOEXEC) |
| 7896 | f_flags |= O_CLOEXEC; |
| 7897 | |
| 7898 | event_fd = get_unused_fd_flags(f_flags); |
| 7899 | if (event_fd < 0) |
| 7900 | return event_fd; |
| 7901 | |
| 7902 | if (group_fd != -1) { |
| 7903 | err = perf_fget_light(group_fd, &group); |
| 7904 | if (err) |
| 7905 | goto err_fd; |
| 7906 | group_leader = group.file->private_data; |
| 7907 | if (flags & PERF_FLAG_FD_OUTPUT) |
| 7908 | output_event = group_leader; |
| 7909 | if (flags & PERF_FLAG_FD_NO_GROUP) |
| 7910 | group_leader = NULL; |
| 7911 | } |
| 7912 | |
| 7913 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
| 7914 | task = find_lively_task_by_vpid(pid); |
| 7915 | if (IS_ERR(task)) { |
| 7916 | err = PTR_ERR(task); |
| 7917 | goto err_group_fd; |
| 7918 | } |
| 7919 | } |
| 7920 | |
| 7921 | if (task && group_leader && |
| 7922 | group_leader->attr.inherit != attr.inherit) { |
| 7923 | err = -EINVAL; |
| 7924 | goto err_task; |
| 7925 | } |
| 7926 | |
| 7927 | get_online_cpus(); |
| 7928 | |
| 7929 | if (flags & PERF_FLAG_PID_CGROUP) |
| 7930 | cgroup_fd = pid; |
| 7931 | |
| 7932 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
| 7933 | NULL, NULL, cgroup_fd); |
| 7934 | if (IS_ERR(event)) { |
| 7935 | err = PTR_ERR(event); |
| 7936 | goto err_cpus; |
| 7937 | } |
| 7938 | |
| 7939 | if (is_sampling_event(event)) { |
| 7940 | if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { |
| 7941 | err = -ENOTSUPP; |
| 7942 | goto err_alloc; |
| 7943 | } |
| 7944 | } |
| 7945 | |
| 7946 | account_event(event); |
| 7947 | |
| 7948 | /* |
| 7949 | * Special case software events and allow them to be part of |
| 7950 | * any hardware group. |
| 7951 | */ |
| 7952 | pmu = event->pmu; |
| 7953 | |
| 7954 | if (attr.use_clockid) { |
| 7955 | err = perf_event_set_clock(event, attr.clockid); |
| 7956 | if (err) |
| 7957 | goto err_alloc; |
| 7958 | } |
| 7959 | |
| 7960 | if (group_leader && |
| 7961 | (is_software_event(event) != is_software_event(group_leader))) { |
| 7962 | if (is_software_event(event)) { |
| 7963 | /* |
| 7964 | * If event and group_leader are not both a software |
| 7965 | * event, and event is, then group leader is not. |
| 7966 | * |
| 7967 | * Allow the addition of software events to !software |
| 7968 | * groups, this is safe because software events never |
| 7969 | * fail to schedule. |
| 7970 | */ |
| 7971 | pmu = group_leader->pmu; |
| 7972 | } else if (is_software_event(group_leader) && |
| 7973 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { |
| 7974 | /* |
| 7975 | * In case the group is a pure software group, and we |
| 7976 | * try to add a hardware event, move the whole group to |
| 7977 | * the hardware context. |
| 7978 | */ |
| 7979 | move_group = 1; |
| 7980 | } |
| 7981 | } |
| 7982 | |
| 7983 | /* |
| 7984 | * Get the target context (task or percpu): |
| 7985 | */ |
| 7986 | ctx = find_get_context(pmu, task, event); |
| 7987 | if (IS_ERR(ctx)) { |
| 7988 | err = PTR_ERR(ctx); |
| 7989 | goto err_alloc; |
| 7990 | } |
| 7991 | |
| 7992 | if ((pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && group_leader) { |
| 7993 | err = -EBUSY; |
| 7994 | goto err_context; |
| 7995 | } |
| 7996 | |
| 7997 | if (task) { |
| 7998 | put_task_struct(task); |
| 7999 | task = NULL; |
| 8000 | } |
| 8001 | |
| 8002 | /* |
| 8003 | * Look up the group leader (we will attach this event to it): |
| 8004 | */ |
| 8005 | if (group_leader) { |
| 8006 | err = -EINVAL; |
| 8007 | |
| 8008 | /* |
| 8009 | * Do not allow a recursive hierarchy (this new sibling |
| 8010 | * becoming part of another group-sibling): |
| 8011 | */ |
| 8012 | if (group_leader->group_leader != group_leader) |
| 8013 | goto err_context; |
| 8014 | |
| 8015 | /* All events in a group should have the same clock */ |
| 8016 | if (group_leader->clock != event->clock) |
| 8017 | goto err_context; |
| 8018 | |
| 8019 | /* |
| 8020 | * Do not allow to attach to a group in a different |
| 8021 | * task or CPU context: |
| 8022 | */ |
| 8023 | if (move_group) { |
| 8024 | /* |
| 8025 | * Make sure we're both on the same task, or both |
| 8026 | * per-cpu events. |
| 8027 | */ |
| 8028 | if (group_leader->ctx->task != ctx->task) |
| 8029 | goto err_context; |
| 8030 | |
| 8031 | /* |
| 8032 | * Make sure we're both events for the same CPU; |
| 8033 | * grouping events for different CPUs is broken; since |
| 8034 | * you can never concurrently schedule them anyhow. |
| 8035 | */ |
| 8036 | if (group_leader->cpu != event->cpu) |
| 8037 | goto err_context; |
| 8038 | } else { |
| 8039 | if (group_leader->ctx != ctx) |
| 8040 | goto err_context; |
| 8041 | } |
| 8042 | |
| 8043 | /* |
| 8044 | * Only a group leader can be exclusive or pinned |
| 8045 | */ |
| 8046 | if (attr.exclusive || attr.pinned) |
| 8047 | goto err_context; |
| 8048 | } |
| 8049 | |
| 8050 | if (output_event) { |
| 8051 | err = perf_event_set_output(event, output_event); |
| 8052 | if (err) |
| 8053 | goto err_context; |
| 8054 | } |
| 8055 | |
| 8056 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
| 8057 | f_flags); |
| 8058 | if (IS_ERR(event_file)) { |
| 8059 | err = PTR_ERR(event_file); |
| 8060 | goto err_context; |
| 8061 | } |
| 8062 | |
| 8063 | if (move_group) { |
| 8064 | gctx = group_leader->ctx; |
| 8065 | |
| 8066 | /* |
| 8067 | * See perf_event_ctx_lock() for comments on the details |
| 8068 | * of swizzling perf_event::ctx. |
| 8069 | */ |
| 8070 | mutex_lock_double(&gctx->mutex, &ctx->mutex); |
| 8071 | |
| 8072 | perf_remove_from_context(group_leader, false); |
| 8073 | |
| 8074 | list_for_each_entry(sibling, &group_leader->sibling_list, |
| 8075 | group_entry) { |
| 8076 | perf_remove_from_context(sibling, false); |
| 8077 | put_ctx(gctx); |
| 8078 | } |
| 8079 | } else { |
| 8080 | mutex_lock(&ctx->mutex); |
| 8081 | } |
| 8082 | |
| 8083 | WARN_ON_ONCE(ctx->parent_ctx); |
| 8084 | |
| 8085 | if (move_group) { |
| 8086 | /* |
| 8087 | * Wait for everybody to stop referencing the events through |
| 8088 | * the old lists, before installing it on new lists. |
| 8089 | */ |
| 8090 | synchronize_rcu(); |
| 8091 | |
| 8092 | /* |
| 8093 | * Install the group siblings before the group leader. |
| 8094 | * |
| 8095 | * Because a group leader will try and install the entire group |
| 8096 | * (through the sibling list, which is still in-tact), we can |
| 8097 | * end up with siblings installed in the wrong context. |
| 8098 | * |
| 8099 | * By installing siblings first we NO-OP because they're not |
| 8100 | * reachable through the group lists. |
| 8101 | */ |
| 8102 | list_for_each_entry(sibling, &group_leader->sibling_list, |
| 8103 | group_entry) { |
| 8104 | perf_event__state_init(sibling); |
| 8105 | perf_install_in_context(ctx, sibling, sibling->cpu); |
| 8106 | get_ctx(ctx); |
| 8107 | } |
| 8108 | |
| 8109 | /* |
| 8110 | * Removing from the context ends up with disabled |
| 8111 | * event. What we want here is event in the initial |
| 8112 | * startup state, ready to be add into new context. |
| 8113 | */ |
| 8114 | perf_event__state_init(group_leader); |
| 8115 | perf_install_in_context(ctx, group_leader, group_leader->cpu); |
| 8116 | get_ctx(ctx); |
| 8117 | } |
| 8118 | |
| 8119 | if (!exclusive_event_installable(event, ctx)) { |
| 8120 | err = -EBUSY; |
| 8121 | mutex_unlock(&ctx->mutex); |
| 8122 | fput(event_file); |
| 8123 | goto err_context; |
| 8124 | } |
| 8125 | |
| 8126 | perf_install_in_context(ctx, event, event->cpu); |
| 8127 | perf_unpin_context(ctx); |
| 8128 | |
| 8129 | if (move_group) { |
| 8130 | mutex_unlock(&gctx->mutex); |
| 8131 | put_ctx(gctx); |
| 8132 | } |
| 8133 | mutex_unlock(&ctx->mutex); |
| 8134 | |
| 8135 | put_online_cpus(); |
| 8136 | |
| 8137 | event->owner = current; |
| 8138 | |
| 8139 | mutex_lock(¤t->perf_event_mutex); |
| 8140 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); |
| 8141 | mutex_unlock(¤t->perf_event_mutex); |
| 8142 | |
| 8143 | /* |
| 8144 | * Precalculate sample_data sizes |
| 8145 | */ |
| 8146 | perf_event__header_size(event); |
| 8147 | perf_event__id_header_size(event); |
| 8148 | |
| 8149 | /* |
| 8150 | * Drop the reference on the group_event after placing the |
| 8151 | * new event on the sibling_list. This ensures destruction |
| 8152 | * of the group leader will find the pointer to itself in |
| 8153 | * perf_group_detach(). |
| 8154 | */ |
| 8155 | fdput(group); |
| 8156 | fd_install(event_fd, event_file); |
| 8157 | return event_fd; |
| 8158 | |
| 8159 | err_context: |
| 8160 | perf_unpin_context(ctx); |
| 8161 | put_ctx(ctx); |
| 8162 | err_alloc: |
| 8163 | free_event(event); |
| 8164 | err_cpus: |
| 8165 | put_online_cpus(); |
| 8166 | err_task: |
| 8167 | if (task) |
| 8168 | put_task_struct(task); |
| 8169 | err_group_fd: |
| 8170 | fdput(group); |
| 8171 | err_fd: |
| 8172 | put_unused_fd(event_fd); |
| 8173 | return err; |
| 8174 | } |
| 8175 | |
| 8176 | /** |
| 8177 | * perf_event_create_kernel_counter |
| 8178 | * |
| 8179 | * @attr: attributes of the counter to create |
| 8180 | * @cpu: cpu in which the counter is bound |
| 8181 | * @task: task to profile (NULL for percpu) |
| 8182 | */ |
| 8183 | struct perf_event * |
| 8184 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, |
| 8185 | struct task_struct *task, |
| 8186 | perf_overflow_handler_t overflow_handler, |
| 8187 | void *context) |
| 8188 | { |
| 8189 | struct perf_event_context *ctx; |
| 8190 | struct perf_event *event; |
| 8191 | int err; |
| 8192 | |
| 8193 | /* |
| 8194 | * Get the target context (task or percpu): |
| 8195 | */ |
| 8196 | |
| 8197 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
| 8198 | overflow_handler, context, -1); |
| 8199 | if (IS_ERR(event)) { |
| 8200 | err = PTR_ERR(event); |
| 8201 | goto err; |
| 8202 | } |
| 8203 | |
| 8204 | /* Mark owner so we could distinguish it from user events. */ |
| 8205 | event->owner = EVENT_OWNER_KERNEL; |
| 8206 | |
| 8207 | account_event(event); |
| 8208 | |
| 8209 | ctx = find_get_context(event->pmu, task, event); |
| 8210 | if (IS_ERR(ctx)) { |
| 8211 | err = PTR_ERR(ctx); |
| 8212 | goto err_free; |
| 8213 | } |
| 8214 | |
| 8215 | WARN_ON_ONCE(ctx->parent_ctx); |
| 8216 | mutex_lock(&ctx->mutex); |
| 8217 | if (!exclusive_event_installable(event, ctx)) { |
| 8218 | mutex_unlock(&ctx->mutex); |
| 8219 | perf_unpin_context(ctx); |
| 8220 | put_ctx(ctx); |
| 8221 | err = -EBUSY; |
| 8222 | goto err_free; |
| 8223 | } |
| 8224 | |
| 8225 | perf_install_in_context(ctx, event, cpu); |
| 8226 | perf_unpin_context(ctx); |
| 8227 | mutex_unlock(&ctx->mutex); |
| 8228 | |
| 8229 | return event; |
| 8230 | |
| 8231 | err_free: |
| 8232 | free_event(event); |
| 8233 | err: |
| 8234 | return ERR_PTR(err); |
| 8235 | } |
| 8236 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
| 8237 | |
| 8238 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
| 8239 | { |
| 8240 | struct perf_event_context *src_ctx; |
| 8241 | struct perf_event_context *dst_ctx; |
| 8242 | struct perf_event *event, *tmp; |
| 8243 | LIST_HEAD(events); |
| 8244 | |
| 8245 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; |
| 8246 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; |
| 8247 | |
| 8248 | /* |
| 8249 | * See perf_event_ctx_lock() for comments on the details |
| 8250 | * of swizzling perf_event::ctx. |
| 8251 | */ |
| 8252 | mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex); |
| 8253 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, |
| 8254 | event_entry) { |
| 8255 | perf_remove_from_context(event, false); |
| 8256 | unaccount_event_cpu(event, src_cpu); |
| 8257 | put_ctx(src_ctx); |
| 8258 | list_add(&event->migrate_entry, &events); |
| 8259 | } |
| 8260 | |
| 8261 | /* |
| 8262 | * Wait for the events to quiesce before re-instating them. |
| 8263 | */ |
| 8264 | synchronize_rcu(); |
| 8265 | |
| 8266 | /* |
| 8267 | * Re-instate events in 2 passes. |
| 8268 | * |
| 8269 | * Skip over group leaders and only install siblings on this first |
| 8270 | * pass, siblings will not get enabled without a leader, however a |
| 8271 | * leader will enable its siblings, even if those are still on the old |
| 8272 | * context. |
| 8273 | */ |
| 8274 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
| 8275 | if (event->group_leader == event) |
| 8276 | continue; |
| 8277 | |
| 8278 | list_del(&event->migrate_entry); |
| 8279 | if (event->state >= PERF_EVENT_STATE_OFF) |
| 8280 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 8281 | account_event_cpu(event, dst_cpu); |
| 8282 | perf_install_in_context(dst_ctx, event, dst_cpu); |
| 8283 | get_ctx(dst_ctx); |
| 8284 | } |
| 8285 | |
| 8286 | /* |
| 8287 | * Once all the siblings are setup properly, install the group leaders |
| 8288 | * to make it go. |
| 8289 | */ |
| 8290 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
| 8291 | list_del(&event->migrate_entry); |
| 8292 | if (event->state >= PERF_EVENT_STATE_OFF) |
| 8293 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 8294 | account_event_cpu(event, dst_cpu); |
| 8295 | perf_install_in_context(dst_ctx, event, dst_cpu); |
| 8296 | get_ctx(dst_ctx); |
| 8297 | } |
| 8298 | mutex_unlock(&dst_ctx->mutex); |
| 8299 | mutex_unlock(&src_ctx->mutex); |
| 8300 | } |
| 8301 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); |
| 8302 | |
| 8303 | static void sync_child_event(struct perf_event *child_event, |
| 8304 | struct task_struct *child) |
| 8305 | { |
| 8306 | struct perf_event *parent_event = child_event->parent; |
| 8307 | u64 child_val; |
| 8308 | |
| 8309 | if (child_event->attr.inherit_stat) |
| 8310 | perf_event_read_event(child_event, child); |
| 8311 | |
| 8312 | child_val = perf_event_count(child_event); |
| 8313 | |
| 8314 | /* |
| 8315 | * Add back the child's count to the parent's count: |
| 8316 | */ |
| 8317 | atomic64_add(child_val, &parent_event->child_count); |
| 8318 | atomic64_add(child_event->total_time_enabled, |
| 8319 | &parent_event->child_total_time_enabled); |
| 8320 | atomic64_add(child_event->total_time_running, |
| 8321 | &parent_event->child_total_time_running); |
| 8322 | |
| 8323 | /* |
| 8324 | * Remove this event from the parent's list |
| 8325 | */ |
| 8326 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
| 8327 | mutex_lock(&parent_event->child_mutex); |
| 8328 | list_del_init(&child_event->child_list); |
| 8329 | mutex_unlock(&parent_event->child_mutex); |
| 8330 | |
| 8331 | /* |
| 8332 | * Make sure user/parent get notified, that we just |
| 8333 | * lost one event. |
| 8334 | */ |
| 8335 | perf_event_wakeup(parent_event); |
| 8336 | |
| 8337 | /* |
| 8338 | * Release the parent event, if this was the last |
| 8339 | * reference to it. |
| 8340 | */ |
| 8341 | put_event(parent_event); |
| 8342 | } |
| 8343 | |
| 8344 | static void |
| 8345 | __perf_event_exit_task(struct perf_event *child_event, |
| 8346 | struct perf_event_context *child_ctx, |
| 8347 | struct task_struct *child) |
| 8348 | { |
| 8349 | /* |
| 8350 | * Do not destroy the 'original' grouping; because of the context |
| 8351 | * switch optimization the original events could've ended up in a |
| 8352 | * random child task. |
| 8353 | * |
| 8354 | * If we were to destroy the original group, all group related |
| 8355 | * operations would cease to function properly after this random |
| 8356 | * child dies. |
| 8357 | * |
| 8358 | * Do destroy all inherited groups, we don't care about those |
| 8359 | * and being thorough is better. |
| 8360 | */ |
| 8361 | perf_remove_from_context(child_event, !!child_event->parent); |
| 8362 | |
| 8363 | /* |
| 8364 | * It can happen that the parent exits first, and has events |
| 8365 | * that are still around due to the child reference. These |
| 8366 | * events need to be zapped. |
| 8367 | */ |
| 8368 | if (child_event->parent) { |
| 8369 | sync_child_event(child_event, child); |
| 8370 | free_event(child_event); |
| 8371 | } else { |
| 8372 | child_event->state = PERF_EVENT_STATE_EXIT; |
| 8373 | perf_event_wakeup(child_event); |
| 8374 | } |
| 8375 | } |
| 8376 | |
| 8377 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
| 8378 | { |
| 8379 | struct perf_event *child_event, *next; |
| 8380 | struct perf_event_context *child_ctx, *clone_ctx = NULL; |
| 8381 | unsigned long flags; |
| 8382 | |
| 8383 | if (likely(!child->perf_event_ctxp[ctxn])) { |
| 8384 | perf_event_task(child, NULL, 0); |
| 8385 | return; |
| 8386 | } |
| 8387 | |
| 8388 | local_irq_save(flags); |
| 8389 | /* |
| 8390 | * We can't reschedule here because interrupts are disabled, |
| 8391 | * and either child is current or it is a task that can't be |
| 8392 | * scheduled, so we are now safe from rescheduling changing |
| 8393 | * our context. |
| 8394 | */ |
| 8395 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
| 8396 | |
| 8397 | /* |
| 8398 | * Take the context lock here so that if find_get_context is |
| 8399 | * reading child->perf_event_ctxp, we wait until it has |
| 8400 | * incremented the context's refcount before we do put_ctx below. |
| 8401 | */ |
| 8402 | raw_spin_lock(&child_ctx->lock); |
| 8403 | task_ctx_sched_out(child_ctx); |
| 8404 | child->perf_event_ctxp[ctxn] = NULL; |
| 8405 | |
| 8406 | /* |
| 8407 | * If this context is a clone; unclone it so it can't get |
| 8408 | * swapped to another process while we're removing all |
| 8409 | * the events from it. |
| 8410 | */ |
| 8411 | clone_ctx = unclone_ctx(child_ctx); |
| 8412 | update_context_time(child_ctx); |
| 8413 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
| 8414 | |
| 8415 | if (clone_ctx) |
| 8416 | put_ctx(clone_ctx); |
| 8417 | |
| 8418 | /* |
| 8419 | * Report the task dead after unscheduling the events so that we |
| 8420 | * won't get any samples after PERF_RECORD_EXIT. We can however still |
| 8421 | * get a few PERF_RECORD_READ events. |
| 8422 | */ |
| 8423 | perf_event_task(child, child_ctx, 0); |
| 8424 | |
| 8425 | /* |
| 8426 | * We can recurse on the same lock type through: |
| 8427 | * |
| 8428 | * __perf_event_exit_task() |
| 8429 | * sync_child_event() |
| 8430 | * put_event() |
| 8431 | * mutex_lock(&ctx->mutex) |
| 8432 | * |
| 8433 | * But since its the parent context it won't be the same instance. |
| 8434 | */ |
| 8435 | mutex_lock(&child_ctx->mutex); |
| 8436 | |
| 8437 | list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) |
| 8438 | __perf_event_exit_task(child_event, child_ctx, child); |
| 8439 | |
| 8440 | mutex_unlock(&child_ctx->mutex); |
| 8441 | |
| 8442 | put_ctx(child_ctx); |
| 8443 | } |
| 8444 | |
| 8445 | /* |
| 8446 | * When a child task exits, feed back event values to parent events. |
| 8447 | */ |
| 8448 | void perf_event_exit_task(struct task_struct *child) |
| 8449 | { |
| 8450 | struct perf_event *event, *tmp; |
| 8451 | int ctxn; |
| 8452 | |
| 8453 | mutex_lock(&child->perf_event_mutex); |
| 8454 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, |
| 8455 | owner_entry) { |
| 8456 | list_del_init(&event->owner_entry); |
| 8457 | |
| 8458 | /* |
| 8459 | * Ensure the list deletion is visible before we clear |
| 8460 | * the owner, closes a race against perf_release() where |
| 8461 | * we need to serialize on the owner->perf_event_mutex. |
| 8462 | */ |
| 8463 | smp_wmb(); |
| 8464 | event->owner = NULL; |
| 8465 | } |
| 8466 | mutex_unlock(&child->perf_event_mutex); |
| 8467 | |
| 8468 | for_each_task_context_nr(ctxn) |
| 8469 | perf_event_exit_task_context(child, ctxn); |
| 8470 | } |
| 8471 | |
| 8472 | static void perf_free_event(struct perf_event *event, |
| 8473 | struct perf_event_context *ctx) |
| 8474 | { |
| 8475 | struct perf_event *parent = event->parent; |
| 8476 | |
| 8477 | if (WARN_ON_ONCE(!parent)) |
| 8478 | return; |
| 8479 | |
| 8480 | mutex_lock(&parent->child_mutex); |
| 8481 | list_del_init(&event->child_list); |
| 8482 | mutex_unlock(&parent->child_mutex); |
| 8483 | |
| 8484 | put_event(parent); |
| 8485 | |
| 8486 | raw_spin_lock_irq(&ctx->lock); |
| 8487 | perf_group_detach(event); |
| 8488 | list_del_event(event, ctx); |
| 8489 | raw_spin_unlock_irq(&ctx->lock); |
| 8490 | free_event(event); |
| 8491 | } |
| 8492 | |
| 8493 | /* |
| 8494 | * Free an unexposed, unused context as created by inheritance by |
| 8495 | * perf_event_init_task below, used by fork() in case of fail. |
| 8496 | * |
| 8497 | * Not all locks are strictly required, but take them anyway to be nice and |
| 8498 | * help out with the lockdep assertions. |
| 8499 | */ |
| 8500 | void perf_event_free_task(struct task_struct *task) |
| 8501 | { |
| 8502 | struct perf_event_context *ctx; |
| 8503 | struct perf_event *event, *tmp; |
| 8504 | int ctxn; |
| 8505 | |
| 8506 | for_each_task_context_nr(ctxn) { |
| 8507 | ctx = task->perf_event_ctxp[ctxn]; |
| 8508 | if (!ctx) |
| 8509 | continue; |
| 8510 | |
| 8511 | mutex_lock(&ctx->mutex); |
| 8512 | again: |
| 8513 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
| 8514 | group_entry) |
| 8515 | perf_free_event(event, ctx); |
| 8516 | |
| 8517 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
| 8518 | group_entry) |
| 8519 | perf_free_event(event, ctx); |
| 8520 | |
| 8521 | if (!list_empty(&ctx->pinned_groups) || |
| 8522 | !list_empty(&ctx->flexible_groups)) |
| 8523 | goto again; |
| 8524 | |
| 8525 | mutex_unlock(&ctx->mutex); |
| 8526 | |
| 8527 | put_ctx(ctx); |
| 8528 | } |
| 8529 | } |
| 8530 | |
| 8531 | void perf_event_delayed_put(struct task_struct *task) |
| 8532 | { |
| 8533 | int ctxn; |
| 8534 | |
| 8535 | for_each_task_context_nr(ctxn) |
| 8536 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); |
| 8537 | } |
| 8538 | |
| 8539 | /* |
| 8540 | * inherit a event from parent task to child task: |
| 8541 | */ |
| 8542 | static struct perf_event * |
| 8543 | inherit_event(struct perf_event *parent_event, |
| 8544 | struct task_struct *parent, |
| 8545 | struct perf_event_context *parent_ctx, |
| 8546 | struct task_struct *child, |
| 8547 | struct perf_event *group_leader, |
| 8548 | struct perf_event_context *child_ctx) |
| 8549 | { |
| 8550 | enum perf_event_active_state parent_state = parent_event->state; |
| 8551 | struct perf_event *child_event; |
| 8552 | unsigned long flags; |
| 8553 | |
| 8554 | /* |
| 8555 | * Instead of creating recursive hierarchies of events, |
| 8556 | * we link inherited events back to the original parent, |
| 8557 | * which has a filp for sure, which we use as the reference |
| 8558 | * count: |
| 8559 | */ |
| 8560 | if (parent_event->parent) |
| 8561 | parent_event = parent_event->parent; |
| 8562 | |
| 8563 | child_event = perf_event_alloc(&parent_event->attr, |
| 8564 | parent_event->cpu, |
| 8565 | child, |
| 8566 | group_leader, parent_event, |
| 8567 | NULL, NULL, -1); |
| 8568 | if (IS_ERR(child_event)) |
| 8569 | return child_event; |
| 8570 | |
| 8571 | if (is_orphaned_event(parent_event) || |
| 8572 | !atomic_long_inc_not_zero(&parent_event->refcount)) { |
| 8573 | free_event(child_event); |
| 8574 | return NULL; |
| 8575 | } |
| 8576 | |
| 8577 | get_ctx(child_ctx); |
| 8578 | |
| 8579 | /* |
| 8580 | * Make the child state follow the state of the parent event, |
| 8581 | * not its attr.disabled bit. We hold the parent's mutex, |
| 8582 | * so we won't race with perf_event_{en, dis}able_family. |
| 8583 | */ |
| 8584 | if (parent_state >= PERF_EVENT_STATE_INACTIVE) |
| 8585 | child_event->state = PERF_EVENT_STATE_INACTIVE; |
| 8586 | else |
| 8587 | child_event->state = PERF_EVENT_STATE_OFF; |
| 8588 | |
| 8589 | if (parent_event->attr.freq) { |
| 8590 | u64 sample_period = parent_event->hw.sample_period; |
| 8591 | struct hw_perf_event *hwc = &child_event->hw; |
| 8592 | |
| 8593 | hwc->sample_period = sample_period; |
| 8594 | hwc->last_period = sample_period; |
| 8595 | |
| 8596 | local64_set(&hwc->period_left, sample_period); |
| 8597 | } |
| 8598 | |
| 8599 | child_event->ctx = child_ctx; |
| 8600 | child_event->overflow_handler = parent_event->overflow_handler; |
| 8601 | child_event->overflow_handler_context |
| 8602 | = parent_event->overflow_handler_context; |
| 8603 | |
| 8604 | /* |
| 8605 | * Precalculate sample_data sizes |
| 8606 | */ |
| 8607 | perf_event__header_size(child_event); |
| 8608 | perf_event__id_header_size(child_event); |
| 8609 | |
| 8610 | /* |
| 8611 | * Link it up in the child's context: |
| 8612 | */ |
| 8613 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
| 8614 | add_event_to_ctx(child_event, child_ctx); |
| 8615 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
| 8616 | |
| 8617 | /* |
| 8618 | * Link this into the parent event's child list |
| 8619 | */ |
| 8620 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
| 8621 | mutex_lock(&parent_event->child_mutex); |
| 8622 | list_add_tail(&child_event->child_list, &parent_event->child_list); |
| 8623 | mutex_unlock(&parent_event->child_mutex); |
| 8624 | |
| 8625 | return child_event; |
| 8626 | } |
| 8627 | |
| 8628 | static int inherit_group(struct perf_event *parent_event, |
| 8629 | struct task_struct *parent, |
| 8630 | struct perf_event_context *parent_ctx, |
| 8631 | struct task_struct *child, |
| 8632 | struct perf_event_context *child_ctx) |
| 8633 | { |
| 8634 | struct perf_event *leader; |
| 8635 | struct perf_event *sub; |
| 8636 | struct perf_event *child_ctr; |
| 8637 | |
| 8638 | leader = inherit_event(parent_event, parent, parent_ctx, |
| 8639 | child, NULL, child_ctx); |
| 8640 | if (IS_ERR(leader)) |
| 8641 | return PTR_ERR(leader); |
| 8642 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { |
| 8643 | child_ctr = inherit_event(sub, parent, parent_ctx, |
| 8644 | child, leader, child_ctx); |
| 8645 | if (IS_ERR(child_ctr)) |
| 8646 | return PTR_ERR(child_ctr); |
| 8647 | } |
| 8648 | return 0; |
| 8649 | } |
| 8650 | |
| 8651 | static int |
| 8652 | inherit_task_group(struct perf_event *event, struct task_struct *parent, |
| 8653 | struct perf_event_context *parent_ctx, |
| 8654 | struct task_struct *child, int ctxn, |
| 8655 | int *inherited_all) |
| 8656 | { |
| 8657 | int ret; |
| 8658 | struct perf_event_context *child_ctx; |
| 8659 | |
| 8660 | if (!event->attr.inherit) { |
| 8661 | *inherited_all = 0; |
| 8662 | return 0; |
| 8663 | } |
| 8664 | |
| 8665 | child_ctx = child->perf_event_ctxp[ctxn]; |
| 8666 | if (!child_ctx) { |
| 8667 | /* |
| 8668 | * This is executed from the parent task context, so |
| 8669 | * inherit events that have been marked for cloning. |
| 8670 | * First allocate and initialize a context for the |
| 8671 | * child. |
| 8672 | */ |
| 8673 | |
| 8674 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
| 8675 | if (!child_ctx) |
| 8676 | return -ENOMEM; |
| 8677 | |
| 8678 | child->perf_event_ctxp[ctxn] = child_ctx; |
| 8679 | } |
| 8680 | |
| 8681 | ret = inherit_group(event, parent, parent_ctx, |
| 8682 | child, child_ctx); |
| 8683 | |
| 8684 | if (ret) |
| 8685 | *inherited_all = 0; |
| 8686 | |
| 8687 | return ret; |
| 8688 | } |
| 8689 | |
| 8690 | /* |
| 8691 | * Initialize the perf_event context in task_struct |
| 8692 | */ |
| 8693 | static int perf_event_init_context(struct task_struct *child, int ctxn) |
| 8694 | { |
| 8695 | struct perf_event_context *child_ctx, *parent_ctx; |
| 8696 | struct perf_event_context *cloned_ctx; |
| 8697 | struct perf_event *event; |
| 8698 | struct task_struct *parent = current; |
| 8699 | int inherited_all = 1; |
| 8700 | unsigned long flags; |
| 8701 | int ret = 0; |
| 8702 | |
| 8703 | if (likely(!parent->perf_event_ctxp[ctxn])) |
| 8704 | return 0; |
| 8705 | |
| 8706 | /* |
| 8707 | * If the parent's context is a clone, pin it so it won't get |
| 8708 | * swapped under us. |
| 8709 | */ |
| 8710 | parent_ctx = perf_pin_task_context(parent, ctxn); |
| 8711 | if (!parent_ctx) |
| 8712 | return 0; |
| 8713 | |
| 8714 | /* |
| 8715 | * No need to check if parent_ctx != NULL here; since we saw |
| 8716 | * it non-NULL earlier, the only reason for it to become NULL |
| 8717 | * is if we exit, and since we're currently in the middle of |
| 8718 | * a fork we can't be exiting at the same time. |
| 8719 | */ |
| 8720 | |
| 8721 | /* |
| 8722 | * Lock the parent list. No need to lock the child - not PID |
| 8723 | * hashed yet and not running, so nobody can access it. |
| 8724 | */ |
| 8725 | mutex_lock(&parent_ctx->mutex); |
| 8726 | |
| 8727 | /* |
| 8728 | * We dont have to disable NMIs - we are only looking at |
| 8729 | * the list, not manipulating it: |
| 8730 | */ |
| 8731 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
| 8732 | ret = inherit_task_group(event, parent, parent_ctx, |
| 8733 | child, ctxn, &inherited_all); |
| 8734 | if (ret) |
| 8735 | break; |
| 8736 | } |
| 8737 | |
| 8738 | /* |
| 8739 | * We can't hold ctx->lock when iterating the ->flexible_group list due |
| 8740 | * to allocations, but we need to prevent rotation because |
| 8741 | * rotate_ctx() will change the list from interrupt context. |
| 8742 | */ |
| 8743 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
| 8744 | parent_ctx->rotate_disable = 1; |
| 8745 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
| 8746 | |
| 8747 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
| 8748 | ret = inherit_task_group(event, parent, parent_ctx, |
| 8749 | child, ctxn, &inherited_all); |
| 8750 | if (ret) |
| 8751 | break; |
| 8752 | } |
| 8753 | |
| 8754 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
| 8755 | parent_ctx->rotate_disable = 0; |
| 8756 | |
| 8757 | child_ctx = child->perf_event_ctxp[ctxn]; |
| 8758 | |
| 8759 | if (child_ctx && inherited_all) { |
| 8760 | /* |
| 8761 | * Mark the child context as a clone of the parent |
| 8762 | * context, or of whatever the parent is a clone of. |
| 8763 | * |
| 8764 | * Note that if the parent is a clone, the holding of |
| 8765 | * parent_ctx->lock avoids it from being uncloned. |
| 8766 | */ |
| 8767 | cloned_ctx = parent_ctx->parent_ctx; |
| 8768 | if (cloned_ctx) { |
| 8769 | child_ctx->parent_ctx = cloned_ctx; |
| 8770 | child_ctx->parent_gen = parent_ctx->parent_gen; |
| 8771 | } else { |
| 8772 | child_ctx->parent_ctx = parent_ctx; |
| 8773 | child_ctx->parent_gen = parent_ctx->generation; |
| 8774 | } |
| 8775 | get_ctx(child_ctx->parent_ctx); |
| 8776 | } |
| 8777 | |
| 8778 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
| 8779 | mutex_unlock(&parent_ctx->mutex); |
| 8780 | |
| 8781 | perf_unpin_context(parent_ctx); |
| 8782 | put_ctx(parent_ctx); |
| 8783 | |
| 8784 | return ret; |
| 8785 | } |
| 8786 | |
| 8787 | /* |
| 8788 | * Initialize the perf_event context in task_struct |
| 8789 | */ |
| 8790 | int perf_event_init_task(struct task_struct *child) |
| 8791 | { |
| 8792 | int ctxn, ret; |
| 8793 | |
| 8794 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
| 8795 | mutex_init(&child->perf_event_mutex); |
| 8796 | INIT_LIST_HEAD(&child->perf_event_list); |
| 8797 | |
| 8798 | for_each_task_context_nr(ctxn) { |
| 8799 | ret = perf_event_init_context(child, ctxn); |
| 8800 | if (ret) { |
| 8801 | perf_event_free_task(child); |
| 8802 | return ret; |
| 8803 | } |
| 8804 | } |
| 8805 | |
| 8806 | return 0; |
| 8807 | } |
| 8808 | |
| 8809 | static void __init perf_event_init_all_cpus(void) |
| 8810 | { |
| 8811 | struct swevent_htable *swhash; |
| 8812 | int cpu; |
| 8813 | |
| 8814 | for_each_possible_cpu(cpu) { |
| 8815 | swhash = &per_cpu(swevent_htable, cpu); |
| 8816 | mutex_init(&swhash->hlist_mutex); |
| 8817 | INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu)); |
| 8818 | } |
| 8819 | } |
| 8820 | |
| 8821 | static void perf_event_init_cpu(int cpu) |
| 8822 | { |
| 8823 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
| 8824 | |
| 8825 | mutex_lock(&swhash->hlist_mutex); |
| 8826 | swhash->online = true; |
| 8827 | if (swhash->hlist_refcount > 0) { |
| 8828 | struct swevent_hlist *hlist; |
| 8829 | |
| 8830 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
| 8831 | WARN_ON(!hlist); |
| 8832 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
| 8833 | } |
| 8834 | mutex_unlock(&swhash->hlist_mutex); |
| 8835 | } |
| 8836 | |
| 8837 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
| 8838 | static void __perf_event_exit_context(void *__info) |
| 8839 | { |
| 8840 | struct remove_event re = { .detach_group = true }; |
| 8841 | struct perf_event_context *ctx = __info; |
| 8842 | |
| 8843 | rcu_read_lock(); |
| 8844 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
| 8845 | __perf_remove_from_context(&re); |
| 8846 | rcu_read_unlock(); |
| 8847 | } |
| 8848 | |
| 8849 | static void perf_event_exit_cpu_context(int cpu) |
| 8850 | { |
| 8851 | struct perf_event_context *ctx; |
| 8852 | struct pmu *pmu; |
| 8853 | int idx; |
| 8854 | |
| 8855 | idx = srcu_read_lock(&pmus_srcu); |
| 8856 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
| 8857 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
| 8858 | |
| 8859 | mutex_lock(&ctx->mutex); |
| 8860 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); |
| 8861 | mutex_unlock(&ctx->mutex); |
| 8862 | } |
| 8863 | srcu_read_unlock(&pmus_srcu, idx); |
| 8864 | } |
| 8865 | |
| 8866 | static void perf_event_exit_cpu(int cpu) |
| 8867 | { |
| 8868 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
| 8869 | |
| 8870 | perf_event_exit_cpu_context(cpu); |
| 8871 | |
| 8872 | mutex_lock(&swhash->hlist_mutex); |
| 8873 | swhash->online = false; |
| 8874 | swevent_hlist_release(swhash); |
| 8875 | mutex_unlock(&swhash->hlist_mutex); |
| 8876 | } |
| 8877 | #else |
| 8878 | static inline void perf_event_exit_cpu(int cpu) { } |
| 8879 | #endif |
| 8880 | |
| 8881 | static int |
| 8882 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) |
| 8883 | { |
| 8884 | int cpu; |
| 8885 | |
| 8886 | for_each_online_cpu(cpu) |
| 8887 | perf_event_exit_cpu(cpu); |
| 8888 | |
| 8889 | return NOTIFY_OK; |
| 8890 | } |
| 8891 | |
| 8892 | /* |
| 8893 | * Run the perf reboot notifier at the very last possible moment so that |
| 8894 | * the generic watchdog code runs as long as possible. |
| 8895 | */ |
| 8896 | static struct notifier_block perf_reboot_notifier = { |
| 8897 | .notifier_call = perf_reboot, |
| 8898 | .priority = INT_MIN, |
| 8899 | }; |
| 8900 | |
| 8901 | static int |
| 8902 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
| 8903 | { |
| 8904 | unsigned int cpu = (long)hcpu; |
| 8905 | |
| 8906 | switch (action & ~CPU_TASKS_FROZEN) { |
| 8907 | |
| 8908 | case CPU_UP_PREPARE: |
| 8909 | case CPU_DOWN_FAILED: |
| 8910 | perf_event_init_cpu(cpu); |
| 8911 | break; |
| 8912 | |
| 8913 | case CPU_UP_CANCELED: |
| 8914 | case CPU_DOWN_PREPARE: |
| 8915 | perf_event_exit_cpu(cpu); |
| 8916 | break; |
| 8917 | default: |
| 8918 | break; |
| 8919 | } |
| 8920 | |
| 8921 | return NOTIFY_OK; |
| 8922 | } |
| 8923 | |
| 8924 | void __init perf_event_init(void) |
| 8925 | { |
| 8926 | int ret; |
| 8927 | |
| 8928 | idr_init(&pmu_idr); |
| 8929 | |
| 8930 | perf_event_init_all_cpus(); |
| 8931 | init_srcu_struct(&pmus_srcu); |
| 8932 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
| 8933 | perf_pmu_register(&perf_cpu_clock, NULL, -1); |
| 8934 | perf_pmu_register(&perf_task_clock, NULL, -1); |
| 8935 | perf_tp_register(); |
| 8936 | perf_cpu_notifier(perf_cpu_notify); |
| 8937 | register_reboot_notifier(&perf_reboot_notifier); |
| 8938 | |
| 8939 | ret = init_hw_breakpoint(); |
| 8940 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); |
| 8941 | |
| 8942 | /* do not patch jump label more than once per second */ |
| 8943 | jump_label_rate_limit(&perf_sched_events, HZ); |
| 8944 | |
| 8945 | /* |
| 8946 | * Build time assertion that we keep the data_head at the intended |
| 8947 | * location. IOW, validation we got the __reserved[] size right. |
| 8948 | */ |
| 8949 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) |
| 8950 | != 1024); |
| 8951 | } |
| 8952 | |
| 8953 | ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr, |
| 8954 | char *page) |
| 8955 | { |
| 8956 | struct perf_pmu_events_attr *pmu_attr = |
| 8957 | container_of(attr, struct perf_pmu_events_attr, attr); |
| 8958 | |
| 8959 | if (pmu_attr->event_str) |
| 8960 | return sprintf(page, "%s\n", pmu_attr->event_str); |
| 8961 | |
| 8962 | return 0; |
| 8963 | } |
| 8964 | |
| 8965 | static int __init perf_event_sysfs_init(void) |
| 8966 | { |
| 8967 | struct pmu *pmu; |
| 8968 | int ret; |
| 8969 | |
| 8970 | mutex_lock(&pmus_lock); |
| 8971 | |
| 8972 | ret = bus_register(&pmu_bus); |
| 8973 | if (ret) |
| 8974 | goto unlock; |
| 8975 | |
| 8976 | list_for_each_entry(pmu, &pmus, entry) { |
| 8977 | if (!pmu->name || pmu->type < 0) |
| 8978 | continue; |
| 8979 | |
| 8980 | ret = pmu_dev_alloc(pmu); |
| 8981 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); |
| 8982 | } |
| 8983 | pmu_bus_running = 1; |
| 8984 | ret = 0; |
| 8985 | |
| 8986 | unlock: |
| 8987 | mutex_unlock(&pmus_lock); |
| 8988 | |
| 8989 | return ret; |
| 8990 | } |
| 8991 | device_initcall(perf_event_sysfs_init); |
| 8992 | |
| 8993 | #ifdef CONFIG_CGROUP_PERF |
| 8994 | static struct cgroup_subsys_state * |
| 8995 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) |
| 8996 | { |
| 8997 | struct perf_cgroup *jc; |
| 8998 | |
| 8999 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
| 9000 | if (!jc) |
| 9001 | return ERR_PTR(-ENOMEM); |
| 9002 | |
| 9003 | jc->info = alloc_percpu(struct perf_cgroup_info); |
| 9004 | if (!jc->info) { |
| 9005 | kfree(jc); |
| 9006 | return ERR_PTR(-ENOMEM); |
| 9007 | } |
| 9008 | |
| 9009 | return &jc->css; |
| 9010 | } |
| 9011 | |
| 9012 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
| 9013 | { |
| 9014 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
| 9015 | |
| 9016 | free_percpu(jc->info); |
| 9017 | kfree(jc); |
| 9018 | } |
| 9019 | |
| 9020 | static int __perf_cgroup_move(void *info) |
| 9021 | { |
| 9022 | struct task_struct *task = info; |
| 9023 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); |
| 9024 | return 0; |
| 9025 | } |
| 9026 | |
| 9027 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
| 9028 | struct cgroup_taskset *tset) |
| 9029 | { |
| 9030 | struct task_struct *task; |
| 9031 | |
| 9032 | cgroup_taskset_for_each(task, tset) |
| 9033 | task_function_call(task, __perf_cgroup_move, task); |
| 9034 | } |
| 9035 | |
| 9036 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
| 9037 | struct cgroup_subsys_state *old_css, |
| 9038 | struct task_struct *task) |
| 9039 | { |
| 9040 | /* |
| 9041 | * cgroup_exit() is called in the copy_process() failure path. |
| 9042 | * Ignore this case since the task hasn't ran yet, this avoids |
| 9043 | * trying to poke a half freed task state from generic code. |
| 9044 | */ |
| 9045 | if (!(task->flags & PF_EXITING)) |
| 9046 | return; |
| 9047 | |
| 9048 | task_function_call(task, __perf_cgroup_move, task); |
| 9049 | } |
| 9050 | |
| 9051 | struct cgroup_subsys perf_event_cgrp_subsys = { |
| 9052 | .css_alloc = perf_cgroup_css_alloc, |
| 9053 | .css_free = perf_cgroup_css_free, |
| 9054 | .exit = perf_cgroup_exit, |
| 9055 | .attach = perf_cgroup_attach, |
| 9056 | }; |
| 9057 | #endif /* CONFIG_CGROUP_PERF */ |