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