Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1 | /* |
Ingo Molnar | 57c0c15 | 2009-09-21 12:20:38 +0200 | [diff] [blame] | 2 | * Performance events core code: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3 | * |
| 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
| 5 | * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar |
| 6 | * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> |
| 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
| 8 | * |
Ingo Molnar | 57c0c15 | 2009-09-21 12:20:38 +0200 | [diff] [blame] | 9 | * For licensing details see kernel-base/COPYING |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 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/file.h> |
| 17 | #include <linux/poll.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 18 | #include <linux/slab.h> |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 19 | #include <linux/hash.h> |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 20 | #include <linux/sysfs.h> |
| 21 | #include <linux/dcache.h> |
| 22 | #include <linux/percpu.h> |
| 23 | #include <linux/ptrace.h> |
| 24 | #include <linux/vmstat.h> |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 25 | #include <linux/vmalloc.h> |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 26 | #include <linux/hardirq.h> |
| 27 | #include <linux/rculist.h> |
| 28 | #include <linux/uaccess.h> |
| 29 | #include <linux/syscalls.h> |
| 30 | #include <linux/anon_inodes.h> |
| 31 | #include <linux/kernel_stat.h> |
| 32 | #include <linux/perf_event.h> |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 33 | #include <linux/ftrace_event.h> |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 34 | |
| 35 | #include <asm/irq_regs.h> |
| 36 | |
| 37 | /* |
| 38 | * Each CPU has a list of per CPU events: |
| 39 | */ |
Xiao Guangrong | aa5452d | 2009-12-09 11:28:13 +0800 | [diff] [blame] | 40 | static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 41 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 42 | static atomic_t nr_events __read_mostly; |
| 43 | static atomic_t nr_mmap_events __read_mostly; |
| 44 | static atomic_t nr_comm_events __read_mostly; |
| 45 | static atomic_t nr_task_events __read_mostly; |
| 46 | |
| 47 | /* |
| 48 | * perf event paranoia level: |
| 49 | * -1 - not paranoid at all |
| 50 | * 0 - disallow raw tracepoint access for unpriv |
| 51 | * 1 - disallow cpu events for unpriv |
| 52 | * 2 - disallow kernel profiling for unpriv |
| 53 | */ |
| 54 | int sysctl_perf_event_paranoid __read_mostly = 1; |
| 55 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 56 | int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */ |
| 57 | |
| 58 | /* |
| 59 | * max perf event sample rate |
| 60 | */ |
| 61 | int sysctl_perf_event_sample_rate __read_mostly = 100000; |
| 62 | |
| 63 | static atomic64_t perf_event_id; |
| 64 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 65 | void __weak perf_event_print_debug(void) { } |
| 66 | |
Peter Zijlstra | 33696fc | 2010-06-14 08:49:00 +0200 | [diff] [blame] | 67 | void perf_pmu_disable(struct pmu *pmu) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 68 | { |
Peter Zijlstra | 33696fc | 2010-06-14 08:49:00 +0200 | [diff] [blame] | 69 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
| 70 | if (!(*count)++) |
| 71 | pmu->pmu_disable(pmu); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 72 | } |
| 73 | |
Peter Zijlstra | 33696fc | 2010-06-14 08:49:00 +0200 | [diff] [blame] | 74 | void perf_pmu_enable(struct pmu *pmu) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 75 | { |
Peter Zijlstra | 33696fc | 2010-06-14 08:49:00 +0200 | [diff] [blame] | 76 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
| 77 | if (!--(*count)) |
| 78 | pmu->pmu_enable(pmu); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 79 | } |
| 80 | |
| 81 | static void get_ctx(struct perf_event_context *ctx) |
| 82 | { |
| 83 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
| 84 | } |
| 85 | |
| 86 | static void free_ctx(struct rcu_head *head) |
| 87 | { |
| 88 | struct perf_event_context *ctx; |
| 89 | |
| 90 | ctx = container_of(head, struct perf_event_context, rcu_head); |
| 91 | kfree(ctx); |
| 92 | } |
| 93 | |
| 94 | static void put_ctx(struct perf_event_context *ctx) |
| 95 | { |
| 96 | if (atomic_dec_and_test(&ctx->refcount)) { |
| 97 | if (ctx->parent_ctx) |
| 98 | put_ctx(ctx->parent_ctx); |
| 99 | if (ctx->task) |
| 100 | put_task_struct(ctx->task); |
| 101 | call_rcu(&ctx->rcu_head, free_ctx); |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | static void unclone_ctx(struct perf_event_context *ctx) |
| 106 | { |
| 107 | if (ctx->parent_ctx) { |
| 108 | put_ctx(ctx->parent_ctx); |
| 109 | ctx->parent_ctx = NULL; |
| 110 | } |
| 111 | } |
| 112 | |
| 113 | /* |
| 114 | * If we inherit events we want to return the parent event id |
| 115 | * to userspace. |
| 116 | */ |
| 117 | static u64 primary_event_id(struct perf_event *event) |
| 118 | { |
| 119 | u64 id = event->id; |
| 120 | |
| 121 | if (event->parent) |
| 122 | id = event->parent->id; |
| 123 | |
| 124 | return id; |
| 125 | } |
| 126 | |
| 127 | /* |
| 128 | * Get the perf_event_context for a task and lock it. |
| 129 | * This has to cope with with the fact that until it is locked, |
| 130 | * the context could get moved to another task. |
| 131 | */ |
| 132 | static struct perf_event_context * |
| 133 | perf_lock_task_context(struct task_struct *task, unsigned long *flags) |
| 134 | { |
| 135 | struct perf_event_context *ctx; |
| 136 | |
| 137 | rcu_read_lock(); |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 138 | retry: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 139 | ctx = rcu_dereference(task->perf_event_ctxp); |
| 140 | if (ctx) { |
| 141 | /* |
| 142 | * If this context is a clone of another, it might |
| 143 | * get swapped for another underneath us by |
| 144 | * perf_event_task_sched_out, though the |
| 145 | * rcu_read_lock() protects us from any context |
| 146 | * getting freed. Lock the context and check if it |
| 147 | * got swapped before we could get the lock, and retry |
| 148 | * if so. If we locked the right context, then it |
| 149 | * can't get swapped on us any more. |
| 150 | */ |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 151 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 152 | if (ctx != rcu_dereference(task->perf_event_ctxp)) { |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 153 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 154 | goto retry; |
| 155 | } |
| 156 | |
| 157 | if (!atomic_inc_not_zero(&ctx->refcount)) { |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 158 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 159 | ctx = NULL; |
| 160 | } |
| 161 | } |
| 162 | rcu_read_unlock(); |
| 163 | return ctx; |
| 164 | } |
| 165 | |
| 166 | /* |
| 167 | * Get the context for a task and increment its pin_count so it |
| 168 | * can't get swapped to another task. This also increments its |
| 169 | * reference count so that the context can't get freed. |
| 170 | */ |
| 171 | static struct perf_event_context *perf_pin_task_context(struct task_struct *task) |
| 172 | { |
| 173 | struct perf_event_context *ctx; |
| 174 | unsigned long flags; |
| 175 | |
| 176 | ctx = perf_lock_task_context(task, &flags); |
| 177 | if (ctx) { |
| 178 | ++ctx->pin_count; |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 179 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 180 | } |
| 181 | return ctx; |
| 182 | } |
| 183 | |
| 184 | static void perf_unpin_context(struct perf_event_context *ctx) |
| 185 | { |
| 186 | unsigned long flags; |
| 187 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 188 | raw_spin_lock_irqsave(&ctx->lock, flags); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 189 | --ctx->pin_count; |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 190 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 191 | put_ctx(ctx); |
| 192 | } |
| 193 | |
Peter Zijlstra | f67218c | 2009-11-23 11:37:27 +0100 | [diff] [blame] | 194 | static inline u64 perf_clock(void) |
| 195 | { |
Peter Zijlstra | c676329 | 2010-05-25 10:48:51 +0200 | [diff] [blame] | 196 | return local_clock(); |
Peter Zijlstra | f67218c | 2009-11-23 11:37:27 +0100 | [diff] [blame] | 197 | } |
| 198 | |
| 199 | /* |
| 200 | * Update the record of the current time in a context. |
| 201 | */ |
| 202 | static void update_context_time(struct perf_event_context *ctx) |
| 203 | { |
| 204 | u64 now = perf_clock(); |
| 205 | |
| 206 | ctx->time += now - ctx->timestamp; |
| 207 | ctx->timestamp = now; |
| 208 | } |
| 209 | |
| 210 | /* |
| 211 | * Update the total_time_enabled and total_time_running fields for a event. |
| 212 | */ |
| 213 | static void update_event_times(struct perf_event *event) |
| 214 | { |
| 215 | struct perf_event_context *ctx = event->ctx; |
| 216 | u64 run_end; |
| 217 | |
| 218 | if (event->state < PERF_EVENT_STATE_INACTIVE || |
| 219 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) |
| 220 | return; |
| 221 | |
Peter Zijlstra | acd1d7c | 2009-11-23 15:00:36 +0100 | [diff] [blame] | 222 | if (ctx->is_active) |
| 223 | run_end = ctx->time; |
| 224 | else |
| 225 | run_end = event->tstamp_stopped; |
| 226 | |
| 227 | event->total_time_enabled = run_end - event->tstamp_enabled; |
Peter Zijlstra | f67218c | 2009-11-23 11:37:27 +0100 | [diff] [blame] | 228 | |
| 229 | if (event->state == PERF_EVENT_STATE_INACTIVE) |
| 230 | run_end = event->tstamp_stopped; |
| 231 | else |
| 232 | run_end = ctx->time; |
| 233 | |
| 234 | event->total_time_running = run_end - event->tstamp_running; |
| 235 | } |
| 236 | |
Peter Zijlstra | 96c21a4 | 2010-05-11 16:19:10 +0200 | [diff] [blame] | 237 | /* |
| 238 | * Update total_time_enabled and total_time_running for all events in a group. |
| 239 | */ |
| 240 | static void update_group_times(struct perf_event *leader) |
| 241 | { |
| 242 | struct perf_event *event; |
| 243 | |
| 244 | update_event_times(leader); |
| 245 | list_for_each_entry(event, &leader->sibling_list, group_entry) |
| 246 | update_event_times(event); |
| 247 | } |
| 248 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 249 | static struct list_head * |
| 250 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) |
| 251 | { |
| 252 | if (event->attr.pinned) |
| 253 | return &ctx->pinned_groups; |
| 254 | else |
| 255 | return &ctx->flexible_groups; |
| 256 | } |
| 257 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 258 | /* |
| 259 | * Add a event from the lists for its context. |
| 260 | * Must be called with ctx->mutex and ctx->lock held. |
| 261 | */ |
| 262 | static void |
| 263 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
| 264 | { |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 265 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
| 266 | event->attach_state |= PERF_ATTACH_CONTEXT; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 267 | |
| 268 | /* |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 269 | * If we're a stand alone event or group leader, we go to the context |
| 270 | * list, group events are kept attached to the group so that |
| 271 | * perf_group_detach can, at all times, locate all siblings. |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 272 | */ |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 273 | if (event->group_leader == event) { |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 274 | struct list_head *list; |
| 275 | |
Frederic Weisbecker | d6f962b | 2010-01-10 01:25:51 +0100 | [diff] [blame] | 276 | if (is_software_event(event)) |
| 277 | event->group_flags |= PERF_GROUP_SOFTWARE; |
| 278 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 279 | list = ctx_group_list(event, ctx); |
| 280 | list_add_tail(&event->group_entry, list); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 281 | } |
| 282 | |
| 283 | list_add_rcu(&event->event_entry, &ctx->event_list); |
| 284 | ctx->nr_events++; |
| 285 | if (event->attr.inherit_stat) |
| 286 | ctx->nr_stat++; |
| 287 | } |
| 288 | |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 289 | static void perf_group_attach(struct perf_event *event) |
| 290 | { |
| 291 | struct perf_event *group_leader = event->group_leader; |
| 292 | |
| 293 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP); |
| 294 | event->attach_state |= PERF_ATTACH_GROUP; |
| 295 | |
| 296 | if (group_leader == event) |
| 297 | return; |
| 298 | |
| 299 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && |
| 300 | !is_software_event(event)) |
| 301 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; |
| 302 | |
| 303 | list_add_tail(&event->group_entry, &group_leader->sibling_list); |
| 304 | group_leader->nr_siblings++; |
| 305 | } |
| 306 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 307 | /* |
| 308 | * Remove a event from the lists for its context. |
| 309 | * Must be called with ctx->mutex and ctx->lock held. |
| 310 | */ |
| 311 | static void |
| 312 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
| 313 | { |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 314 | /* |
| 315 | * We can have double detach due to exit/hot-unplug + close. |
| 316 | */ |
| 317 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 318 | return; |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 319 | |
| 320 | event->attach_state &= ~PERF_ATTACH_CONTEXT; |
| 321 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 322 | ctx->nr_events--; |
| 323 | if (event->attr.inherit_stat) |
| 324 | ctx->nr_stat--; |
| 325 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 326 | list_del_rcu(&event->event_entry); |
| 327 | |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 328 | if (event->group_leader == event) |
| 329 | list_del_init(&event->group_entry); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 330 | |
Peter Zijlstra | 96c21a4 | 2010-05-11 16:19:10 +0200 | [diff] [blame] | 331 | update_group_times(event); |
Stephane Eranian | b2e74a2 | 2009-11-26 09:24:30 -0800 | [diff] [blame] | 332 | |
| 333 | /* |
| 334 | * If event was in error state, then keep it |
| 335 | * that way, otherwise bogus counts will be |
| 336 | * returned on read(). The only way to get out |
| 337 | * of error state is by explicit re-enabling |
| 338 | * of the event |
| 339 | */ |
| 340 | if (event->state > PERF_EVENT_STATE_OFF) |
| 341 | event->state = PERF_EVENT_STATE_OFF; |
Peter Zijlstra | 050735b | 2010-05-11 11:51:53 +0200 | [diff] [blame] | 342 | } |
| 343 | |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 344 | static void perf_group_detach(struct perf_event *event) |
Peter Zijlstra | 050735b | 2010-05-11 11:51:53 +0200 | [diff] [blame] | 345 | { |
| 346 | struct perf_event *sibling, *tmp; |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 347 | struct list_head *list = NULL; |
| 348 | |
| 349 | /* |
| 350 | * We can have double detach due to exit/hot-unplug + close. |
| 351 | */ |
| 352 | if (!(event->attach_state & PERF_ATTACH_GROUP)) |
| 353 | return; |
| 354 | |
| 355 | event->attach_state &= ~PERF_ATTACH_GROUP; |
| 356 | |
| 357 | /* |
| 358 | * If this is a sibling, remove it from its group. |
| 359 | */ |
| 360 | if (event->group_leader != event) { |
| 361 | list_del_init(&event->group_entry); |
| 362 | event->group_leader->nr_siblings--; |
| 363 | return; |
| 364 | } |
| 365 | |
| 366 | if (!list_empty(&event->group_entry)) |
| 367 | list = &event->group_entry; |
Peter Zijlstra | 2e2af50 | 2009-11-23 11:37:25 +0100 | [diff] [blame] | 368 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 369 | /* |
| 370 | * If this was a group event with sibling events then |
| 371 | * upgrade the siblings to singleton events by adding them |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 372 | * to whatever list we are on. |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 373 | */ |
| 374 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 375 | if (list) |
| 376 | list_move_tail(&sibling->group_entry, list); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 377 | sibling->group_leader = sibling; |
Frederic Weisbecker | d6f962b | 2010-01-10 01:25:51 +0100 | [diff] [blame] | 378 | |
| 379 | /* Inherit group flags from the previous leader */ |
| 380 | sibling->group_flags = event->group_flags; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 381 | } |
| 382 | } |
| 383 | |
Stephane Eranian | fa66f07 | 2010-08-26 16:40:01 +0200 | [diff] [blame] | 384 | static inline int |
| 385 | event_filter_match(struct perf_event *event) |
| 386 | { |
| 387 | return event->cpu == -1 || event->cpu == smp_processor_id(); |
| 388 | } |
| 389 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 390 | static void |
| 391 | event_sched_out(struct perf_event *event, |
| 392 | struct perf_cpu_context *cpuctx, |
| 393 | struct perf_event_context *ctx) |
| 394 | { |
Stephane Eranian | fa66f07 | 2010-08-26 16:40:01 +0200 | [diff] [blame] | 395 | u64 delta; |
| 396 | /* |
| 397 | * An event which could not be activated because of |
| 398 | * filter mismatch still needs to have its timings |
| 399 | * maintained, otherwise bogus information is return |
| 400 | * via read() for time_enabled, time_running: |
| 401 | */ |
| 402 | if (event->state == PERF_EVENT_STATE_INACTIVE |
| 403 | && !event_filter_match(event)) { |
| 404 | delta = ctx->time - event->tstamp_stopped; |
| 405 | event->tstamp_running += delta; |
| 406 | event->tstamp_stopped = ctx->time; |
| 407 | } |
| 408 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 409 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
| 410 | return; |
| 411 | |
| 412 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 413 | if (event->pending_disable) { |
| 414 | event->pending_disable = 0; |
| 415 | event->state = PERF_EVENT_STATE_OFF; |
| 416 | } |
| 417 | event->tstamp_stopped = ctx->time; |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 418 | event->pmu->del(event, 0); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 419 | event->oncpu = -1; |
| 420 | |
| 421 | if (!is_software_event(event)) |
| 422 | cpuctx->active_oncpu--; |
| 423 | ctx->nr_active--; |
| 424 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
| 425 | cpuctx->exclusive = 0; |
| 426 | } |
| 427 | |
| 428 | static void |
| 429 | group_sched_out(struct perf_event *group_event, |
| 430 | struct perf_cpu_context *cpuctx, |
| 431 | struct perf_event_context *ctx) |
| 432 | { |
| 433 | struct perf_event *event; |
Stephane Eranian | fa66f07 | 2010-08-26 16:40:01 +0200 | [diff] [blame] | 434 | int state = group_event->state; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 435 | |
| 436 | event_sched_out(group_event, cpuctx, ctx); |
| 437 | |
| 438 | /* |
| 439 | * Schedule out siblings (if any): |
| 440 | */ |
| 441 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
| 442 | event_sched_out(event, cpuctx, ctx); |
| 443 | |
Stephane Eranian | fa66f07 | 2010-08-26 16:40:01 +0200 | [diff] [blame] | 444 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 445 | cpuctx->exclusive = 0; |
| 446 | } |
| 447 | |
| 448 | /* |
| 449 | * Cross CPU call to remove a performance event |
| 450 | * |
| 451 | * We disable the event on the hardware level first. After that we |
| 452 | * remove it from the context list. |
| 453 | */ |
| 454 | static void __perf_event_remove_from_context(void *info) |
| 455 | { |
| 456 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
| 457 | struct perf_event *event = info; |
| 458 | struct perf_event_context *ctx = event->ctx; |
| 459 | |
| 460 | /* |
| 461 | * If this is a task context, we need to check whether it is |
| 462 | * the current task context of this cpu. If not it has been |
| 463 | * scheduled out before the smp call arrived. |
| 464 | */ |
| 465 | if (ctx->task && cpuctx->task_ctx != ctx) |
| 466 | return; |
| 467 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 468 | raw_spin_lock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 469 | |
| 470 | event_sched_out(event, cpuctx, ctx); |
| 471 | |
| 472 | list_del_event(event, ctx); |
| 473 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 474 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 475 | } |
| 476 | |
| 477 | |
| 478 | /* |
| 479 | * Remove the event from a task's (or a CPU's) list of events. |
| 480 | * |
| 481 | * Must be called with ctx->mutex held. |
| 482 | * |
| 483 | * CPU events are removed with a smp call. For task events we only |
| 484 | * call when the task is on a CPU. |
| 485 | * |
| 486 | * If event->ctx is a cloned context, callers must make sure that |
| 487 | * every task struct that event->ctx->task could possibly point to |
| 488 | * remains valid. This is OK when called from perf_release since |
| 489 | * that only calls us on the top-level context, which can't be a clone. |
| 490 | * When called from perf_event_exit_task, it's OK because the |
| 491 | * context has been detached from its task. |
| 492 | */ |
| 493 | static void perf_event_remove_from_context(struct perf_event *event) |
| 494 | { |
| 495 | struct perf_event_context *ctx = event->ctx; |
| 496 | struct task_struct *task = ctx->task; |
| 497 | |
| 498 | if (!task) { |
| 499 | /* |
| 500 | * Per cpu events are removed via an smp call and |
André Goddard Rosa | af901ca | 2009-11-14 13:09:05 -0200 | [diff] [blame] | 501 | * the removal is always successful. |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 502 | */ |
| 503 | smp_call_function_single(event->cpu, |
| 504 | __perf_event_remove_from_context, |
| 505 | event, 1); |
| 506 | return; |
| 507 | } |
| 508 | |
| 509 | retry: |
| 510 | task_oncpu_function_call(task, __perf_event_remove_from_context, |
| 511 | event); |
| 512 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 513 | raw_spin_lock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 514 | /* |
| 515 | * If the context is active we need to retry the smp call. |
| 516 | */ |
| 517 | if (ctx->nr_active && !list_empty(&event->group_entry)) { |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 518 | raw_spin_unlock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 519 | goto retry; |
| 520 | } |
| 521 | |
| 522 | /* |
| 523 | * The lock prevents that this context is scheduled in so we |
| 524 | * can remove the event safely, if the call above did not |
| 525 | * succeed. |
| 526 | */ |
Peter Zijlstra | 6c2bfcb | 2009-11-23 11:37:24 +0100 | [diff] [blame] | 527 | if (!list_empty(&event->group_entry)) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 528 | list_del_event(event, ctx); |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 529 | raw_spin_unlock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 530 | } |
| 531 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 532 | /* |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 533 | * Cross CPU call to disable a performance event |
| 534 | */ |
| 535 | static void __perf_event_disable(void *info) |
| 536 | { |
| 537 | struct perf_event *event = info; |
| 538 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
| 539 | struct perf_event_context *ctx = event->ctx; |
| 540 | |
| 541 | /* |
| 542 | * If this is a per-task event, need to check whether this |
| 543 | * event's task is the current task on this cpu. |
| 544 | */ |
| 545 | if (ctx->task && cpuctx->task_ctx != ctx) |
| 546 | return; |
| 547 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 548 | raw_spin_lock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 549 | |
| 550 | /* |
| 551 | * If the event is on, turn it off. |
| 552 | * If it is in error state, leave it in error state. |
| 553 | */ |
| 554 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
| 555 | update_context_time(ctx); |
| 556 | update_group_times(event); |
| 557 | if (event == event->group_leader) |
| 558 | group_sched_out(event, cpuctx, ctx); |
| 559 | else |
| 560 | event_sched_out(event, cpuctx, ctx); |
| 561 | event->state = PERF_EVENT_STATE_OFF; |
| 562 | } |
| 563 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 564 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 565 | } |
| 566 | |
| 567 | /* |
| 568 | * Disable a event. |
| 569 | * |
| 570 | * If event->ctx is a cloned context, callers must make sure that |
| 571 | * every task struct that event->ctx->task could possibly point to |
| 572 | * remains valid. This condition is satisifed when called through |
| 573 | * perf_event_for_each_child or perf_event_for_each because they |
| 574 | * hold the top-level event's child_mutex, so any descendant that |
| 575 | * goes to exit will block in sync_child_event. |
| 576 | * When called from perf_pending_event it's OK because event->ctx |
| 577 | * is the current context on this CPU and preemption is disabled, |
| 578 | * hence we can't get into perf_event_task_sched_out for this context. |
| 579 | */ |
Frederic Weisbecker | 44234ad | 2009-12-09 09:25:48 +0100 | [diff] [blame] | 580 | void perf_event_disable(struct perf_event *event) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 581 | { |
| 582 | struct perf_event_context *ctx = event->ctx; |
| 583 | struct task_struct *task = ctx->task; |
| 584 | |
| 585 | if (!task) { |
| 586 | /* |
| 587 | * Disable the event on the cpu that it's on |
| 588 | */ |
| 589 | smp_call_function_single(event->cpu, __perf_event_disable, |
| 590 | event, 1); |
| 591 | return; |
| 592 | } |
| 593 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 594 | retry: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 595 | task_oncpu_function_call(task, __perf_event_disable, event); |
| 596 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 597 | raw_spin_lock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 598 | /* |
| 599 | * If the event is still active, we need to retry the cross-call. |
| 600 | */ |
| 601 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 602 | raw_spin_unlock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 603 | goto retry; |
| 604 | } |
| 605 | |
| 606 | /* |
| 607 | * Since we have the lock this context can't be scheduled |
| 608 | * in, so we can change the state safely. |
| 609 | */ |
| 610 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
| 611 | update_group_times(event); |
| 612 | event->state = PERF_EVENT_STATE_OFF; |
| 613 | } |
| 614 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 615 | raw_spin_unlock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 616 | } |
| 617 | |
| 618 | static int |
| 619 | event_sched_in(struct perf_event *event, |
| 620 | struct perf_cpu_context *cpuctx, |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 621 | struct perf_event_context *ctx) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 622 | { |
| 623 | if (event->state <= PERF_EVENT_STATE_OFF) |
| 624 | return 0; |
| 625 | |
| 626 | event->state = PERF_EVENT_STATE_ACTIVE; |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 627 | event->oncpu = smp_processor_id(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 628 | /* |
| 629 | * The new state must be visible before we turn it on in the hardware: |
| 630 | */ |
| 631 | smp_wmb(); |
| 632 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 633 | if (event->pmu->add(event, PERF_EF_START)) { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 634 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 635 | event->oncpu = -1; |
| 636 | return -EAGAIN; |
| 637 | } |
| 638 | |
| 639 | event->tstamp_running += ctx->time - event->tstamp_stopped; |
| 640 | |
| 641 | if (!is_software_event(event)) |
| 642 | cpuctx->active_oncpu++; |
| 643 | ctx->nr_active++; |
| 644 | |
| 645 | if (event->attr.exclusive) |
| 646 | cpuctx->exclusive = 1; |
| 647 | |
| 648 | return 0; |
| 649 | } |
| 650 | |
| 651 | static int |
| 652 | group_sched_in(struct perf_event *group_event, |
| 653 | struct perf_cpu_context *cpuctx, |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 654 | struct perf_event_context *ctx) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 655 | { |
Lin Ming | 6bde9b6 | 2010-04-23 13:56:00 +0800 | [diff] [blame] | 656 | struct perf_event *event, *partial_group = NULL; |
Peter Zijlstra | 51b0fe3 | 2010-06-11 13:35:57 +0200 | [diff] [blame] | 657 | struct pmu *pmu = group_event->pmu; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 658 | |
| 659 | if (group_event->state == PERF_EVENT_STATE_OFF) |
| 660 | return 0; |
| 661 | |
Peter Zijlstra | ad5133b | 2010-06-15 12:22:39 +0200 | [diff] [blame] | 662 | pmu->start_txn(pmu); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 663 | |
Stephane Eranian | 90151c35 | 2010-05-25 16:23:10 +0200 | [diff] [blame] | 664 | if (event_sched_in(group_event, cpuctx, ctx)) { |
Peter Zijlstra | ad5133b | 2010-06-15 12:22:39 +0200 | [diff] [blame] | 665 | pmu->cancel_txn(pmu); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 666 | return -EAGAIN; |
Stephane Eranian | 90151c35 | 2010-05-25 16:23:10 +0200 | [diff] [blame] | 667 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 668 | |
| 669 | /* |
| 670 | * Schedule in siblings as one group (if any): |
| 671 | */ |
| 672 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 673 | if (event_sched_in(event, cpuctx, ctx)) { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 674 | partial_group = event; |
| 675 | goto group_error; |
| 676 | } |
| 677 | } |
| 678 | |
Peter Zijlstra | ad5133b | 2010-06-15 12:22:39 +0200 | [diff] [blame] | 679 | if (!pmu->commit_txn(pmu)) |
Paul Mackerras | 6e85158 | 2010-05-08 20:58:00 +1000 | [diff] [blame] | 680 | return 0; |
Lin Ming | 6bde9b6 | 2010-04-23 13:56:00 +0800 | [diff] [blame] | 681 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 682 | group_error: |
| 683 | /* |
| 684 | * Groups can be scheduled in as one unit only, so undo any |
| 685 | * partial group before returning: |
| 686 | */ |
| 687 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
| 688 | if (event == partial_group) |
| 689 | break; |
| 690 | event_sched_out(event, cpuctx, ctx); |
| 691 | } |
| 692 | event_sched_out(group_event, cpuctx, ctx); |
| 693 | |
Peter Zijlstra | ad5133b | 2010-06-15 12:22:39 +0200 | [diff] [blame] | 694 | pmu->cancel_txn(pmu); |
Stephane Eranian | 90151c35 | 2010-05-25 16:23:10 +0200 | [diff] [blame] | 695 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 696 | return -EAGAIN; |
| 697 | } |
| 698 | |
| 699 | /* |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 700 | * Work out whether we can put this event group on the CPU now. |
| 701 | */ |
| 702 | static int group_can_go_on(struct perf_event *event, |
| 703 | struct perf_cpu_context *cpuctx, |
| 704 | int can_add_hw) |
| 705 | { |
| 706 | /* |
| 707 | * Groups consisting entirely of software events can always go on. |
| 708 | */ |
Frederic Weisbecker | d6f962b | 2010-01-10 01:25:51 +0100 | [diff] [blame] | 709 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 710 | return 1; |
| 711 | /* |
| 712 | * If an exclusive group is already on, no other hardware |
| 713 | * events can go on. |
| 714 | */ |
| 715 | if (cpuctx->exclusive) |
| 716 | return 0; |
| 717 | /* |
| 718 | * If this group is exclusive and there are already |
| 719 | * events on the CPU, it can't go on. |
| 720 | */ |
| 721 | if (event->attr.exclusive && cpuctx->active_oncpu) |
| 722 | return 0; |
| 723 | /* |
| 724 | * Otherwise, try to add it if all previous groups were able |
| 725 | * to go on. |
| 726 | */ |
| 727 | return can_add_hw; |
| 728 | } |
| 729 | |
| 730 | static void add_event_to_ctx(struct perf_event *event, |
| 731 | struct perf_event_context *ctx) |
| 732 | { |
| 733 | list_add_event(event, ctx); |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 734 | perf_group_attach(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 735 | event->tstamp_enabled = ctx->time; |
| 736 | event->tstamp_running = ctx->time; |
| 737 | event->tstamp_stopped = ctx->time; |
| 738 | } |
| 739 | |
| 740 | /* |
| 741 | * Cross CPU call to install and enable a performance event |
| 742 | * |
| 743 | * Must be called with ctx->mutex held |
| 744 | */ |
| 745 | static void __perf_install_in_context(void *info) |
| 746 | { |
| 747 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
| 748 | struct perf_event *event = info; |
| 749 | struct perf_event_context *ctx = event->ctx; |
| 750 | struct perf_event *leader = event->group_leader; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 751 | int err; |
| 752 | |
| 753 | /* |
| 754 | * If this is a task context, we need to check whether it is |
| 755 | * the current task context of this cpu. If not it has been |
| 756 | * scheduled out before the smp call arrived. |
| 757 | * Or possibly this is the right context but it isn't |
| 758 | * on this cpu because it had no events. |
| 759 | */ |
| 760 | if (ctx->task && cpuctx->task_ctx != ctx) { |
| 761 | if (cpuctx->task_ctx || ctx->task != current) |
| 762 | return; |
| 763 | cpuctx->task_ctx = ctx; |
| 764 | } |
| 765 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 766 | raw_spin_lock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 767 | ctx->is_active = 1; |
| 768 | update_context_time(ctx); |
| 769 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 770 | add_event_to_ctx(event, ctx); |
| 771 | |
Peter Zijlstra | f4c4176 | 2009-12-16 17:55:54 +0100 | [diff] [blame] | 772 | if (event->cpu != -1 && event->cpu != smp_processor_id()) |
| 773 | goto unlock; |
| 774 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 775 | /* |
| 776 | * Don't put the event on if it is disabled or if |
| 777 | * it is in a group and the group isn't on. |
| 778 | */ |
| 779 | if (event->state != PERF_EVENT_STATE_INACTIVE || |
| 780 | (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)) |
| 781 | goto unlock; |
| 782 | |
| 783 | /* |
| 784 | * An exclusive event can't go on if there are already active |
| 785 | * hardware events, and no hardware event can go on if there |
| 786 | * is already an exclusive event on. |
| 787 | */ |
| 788 | if (!group_can_go_on(event, cpuctx, 1)) |
| 789 | err = -EEXIST; |
| 790 | else |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 791 | err = event_sched_in(event, cpuctx, ctx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 792 | |
| 793 | if (err) { |
| 794 | /* |
| 795 | * This event couldn't go on. If it is in a group |
| 796 | * then we have to pull the whole group off. |
| 797 | * If the event group is pinned then put it in error state. |
| 798 | */ |
| 799 | if (leader != event) |
| 800 | group_sched_out(leader, cpuctx, ctx); |
| 801 | if (leader->attr.pinned) { |
| 802 | update_group_times(leader); |
| 803 | leader->state = PERF_EVENT_STATE_ERROR; |
| 804 | } |
| 805 | } |
| 806 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 807 | unlock: |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 808 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 809 | } |
| 810 | |
| 811 | /* |
| 812 | * Attach a performance event to a context |
| 813 | * |
| 814 | * First we add the event to the list with the hardware enable bit |
| 815 | * in event->hw_config cleared. |
| 816 | * |
| 817 | * If the event is attached to a task which is on a CPU we use a smp |
| 818 | * call to enable it in the task context. The task might have been |
| 819 | * scheduled away, but we check this in the smp call again. |
| 820 | * |
| 821 | * Must be called with ctx->mutex held. |
| 822 | */ |
| 823 | static void |
| 824 | perf_install_in_context(struct perf_event_context *ctx, |
| 825 | struct perf_event *event, |
| 826 | int cpu) |
| 827 | { |
| 828 | struct task_struct *task = ctx->task; |
| 829 | |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 830 | event->ctx = ctx; |
| 831 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 832 | if (!task) { |
| 833 | /* |
| 834 | * Per cpu events are installed via an smp call and |
André Goddard Rosa | af901ca | 2009-11-14 13:09:05 -0200 | [diff] [blame] | 835 | * the install is always successful. |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 836 | */ |
| 837 | smp_call_function_single(cpu, __perf_install_in_context, |
| 838 | event, 1); |
| 839 | return; |
| 840 | } |
| 841 | |
| 842 | retry: |
| 843 | task_oncpu_function_call(task, __perf_install_in_context, |
| 844 | event); |
| 845 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 846 | raw_spin_lock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 847 | /* |
| 848 | * we need to retry the smp call. |
| 849 | */ |
| 850 | if (ctx->is_active && list_empty(&event->group_entry)) { |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 851 | raw_spin_unlock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 852 | goto retry; |
| 853 | } |
| 854 | |
| 855 | /* |
| 856 | * The lock prevents that this context is scheduled in so we |
| 857 | * can add the event safely, if it the call above did not |
| 858 | * succeed. |
| 859 | */ |
| 860 | if (list_empty(&event->group_entry)) |
| 861 | add_event_to_ctx(event, ctx); |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 862 | raw_spin_unlock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 863 | } |
| 864 | |
| 865 | /* |
| 866 | * Put a event into inactive state and update time fields. |
| 867 | * Enabling the leader of a group effectively enables all |
| 868 | * the group members that aren't explicitly disabled, so we |
| 869 | * have to update their ->tstamp_enabled also. |
| 870 | * Note: this works for group members as well as group leaders |
| 871 | * since the non-leader members' sibling_lists will be empty. |
| 872 | */ |
| 873 | static void __perf_event_mark_enabled(struct perf_event *event, |
| 874 | struct perf_event_context *ctx) |
| 875 | { |
| 876 | struct perf_event *sub; |
| 877 | |
| 878 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 879 | event->tstamp_enabled = ctx->time - event->total_time_enabled; |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 880 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
| 881 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 882 | sub->tstamp_enabled = |
| 883 | ctx->time - sub->total_time_enabled; |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 884 | } |
| 885 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 886 | } |
| 887 | |
| 888 | /* |
| 889 | * Cross CPU call to enable a performance event |
| 890 | */ |
| 891 | static void __perf_event_enable(void *info) |
| 892 | { |
| 893 | struct perf_event *event = info; |
| 894 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
| 895 | struct perf_event_context *ctx = event->ctx; |
| 896 | struct perf_event *leader = event->group_leader; |
| 897 | int err; |
| 898 | |
| 899 | /* |
| 900 | * If this is a per-task event, need to check whether this |
| 901 | * event's task is the current task on this cpu. |
| 902 | */ |
| 903 | if (ctx->task && cpuctx->task_ctx != ctx) { |
| 904 | if (cpuctx->task_ctx || ctx->task != current) |
| 905 | return; |
| 906 | cpuctx->task_ctx = ctx; |
| 907 | } |
| 908 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 909 | raw_spin_lock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 910 | ctx->is_active = 1; |
| 911 | update_context_time(ctx); |
| 912 | |
| 913 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
| 914 | goto unlock; |
| 915 | __perf_event_mark_enabled(event, ctx); |
| 916 | |
Peter Zijlstra | f4c4176 | 2009-12-16 17:55:54 +0100 | [diff] [blame] | 917 | if (event->cpu != -1 && event->cpu != smp_processor_id()) |
| 918 | goto unlock; |
| 919 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 920 | /* |
| 921 | * If the event is in a group and isn't the group leader, |
| 922 | * then don't put it on unless the group is on. |
| 923 | */ |
| 924 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
| 925 | goto unlock; |
| 926 | |
| 927 | if (!group_can_go_on(event, cpuctx, 1)) { |
| 928 | err = -EEXIST; |
| 929 | } else { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 930 | if (event == leader) |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 931 | err = group_sched_in(event, cpuctx, ctx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 932 | else |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 933 | err = event_sched_in(event, cpuctx, ctx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 934 | } |
| 935 | |
| 936 | if (err) { |
| 937 | /* |
| 938 | * If this event can't go on and it's part of a |
| 939 | * group, then the whole group has to come off. |
| 940 | */ |
| 941 | if (leader != event) |
| 942 | group_sched_out(leader, cpuctx, ctx); |
| 943 | if (leader->attr.pinned) { |
| 944 | update_group_times(leader); |
| 945 | leader->state = PERF_EVENT_STATE_ERROR; |
| 946 | } |
| 947 | } |
| 948 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 949 | unlock: |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 950 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 951 | } |
| 952 | |
| 953 | /* |
| 954 | * Enable a event. |
| 955 | * |
| 956 | * If event->ctx is a cloned context, callers must make sure that |
| 957 | * every task struct that event->ctx->task could possibly point to |
| 958 | * remains valid. This condition is satisfied when called through |
| 959 | * perf_event_for_each_child or perf_event_for_each as described |
| 960 | * for perf_event_disable. |
| 961 | */ |
Frederic Weisbecker | 44234ad | 2009-12-09 09:25:48 +0100 | [diff] [blame] | 962 | void perf_event_enable(struct perf_event *event) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 963 | { |
| 964 | struct perf_event_context *ctx = event->ctx; |
| 965 | struct task_struct *task = ctx->task; |
| 966 | |
| 967 | if (!task) { |
| 968 | /* |
| 969 | * Enable the event on the cpu that it's on |
| 970 | */ |
| 971 | smp_call_function_single(event->cpu, __perf_event_enable, |
| 972 | event, 1); |
| 973 | return; |
| 974 | } |
| 975 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 976 | raw_spin_lock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 977 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
| 978 | goto out; |
| 979 | |
| 980 | /* |
| 981 | * If the event is in error state, clear that first. |
| 982 | * That way, if we see the event in error state below, we |
| 983 | * know that it has gone back into error state, as distinct |
| 984 | * from the task having been scheduled away before the |
| 985 | * cross-call arrived. |
| 986 | */ |
| 987 | if (event->state == PERF_EVENT_STATE_ERROR) |
| 988 | event->state = PERF_EVENT_STATE_OFF; |
| 989 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 990 | retry: |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 991 | raw_spin_unlock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 992 | task_oncpu_function_call(task, __perf_event_enable, event); |
| 993 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 994 | raw_spin_lock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 995 | |
| 996 | /* |
| 997 | * If the context is active and the event is still off, |
| 998 | * we need to retry the cross-call. |
| 999 | */ |
| 1000 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) |
| 1001 | goto retry; |
| 1002 | |
| 1003 | /* |
| 1004 | * Since we have the lock this context can't be scheduled |
| 1005 | * in, so we can change the state safely. |
| 1006 | */ |
| 1007 | if (event->state == PERF_EVENT_STATE_OFF) |
| 1008 | __perf_event_mark_enabled(event, ctx); |
| 1009 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1010 | out: |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1011 | raw_spin_unlock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1012 | } |
| 1013 | |
| 1014 | static int perf_event_refresh(struct perf_event *event, int refresh) |
| 1015 | { |
| 1016 | /* |
| 1017 | * not supported on inherited events |
| 1018 | */ |
| 1019 | if (event->attr.inherit) |
| 1020 | return -EINVAL; |
| 1021 | |
| 1022 | atomic_add(refresh, &event->event_limit); |
| 1023 | perf_event_enable(event); |
| 1024 | |
| 1025 | return 0; |
| 1026 | } |
| 1027 | |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1028 | enum event_type_t { |
| 1029 | EVENT_FLEXIBLE = 0x1, |
| 1030 | EVENT_PINNED = 0x2, |
| 1031 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, |
| 1032 | }; |
| 1033 | |
| 1034 | static void ctx_sched_out(struct perf_event_context *ctx, |
| 1035 | struct perf_cpu_context *cpuctx, |
| 1036 | enum event_type_t event_type) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1037 | { |
| 1038 | struct perf_event *event; |
| 1039 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1040 | raw_spin_lock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1041 | ctx->is_active = 0; |
| 1042 | if (likely(!ctx->nr_events)) |
| 1043 | goto out; |
| 1044 | update_context_time(ctx); |
| 1045 | |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1046 | if (!ctx->nr_active) |
Peter Zijlstra | 24cd7f5 | 2010-06-11 17:32:03 +0200 | [diff] [blame] | 1047 | goto out; |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1048 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1049 | if (event_type & EVENT_PINNED) { |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1050 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
| 1051 | group_sched_out(event, cpuctx, ctx); |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1052 | } |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1053 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1054 | if (event_type & EVENT_FLEXIBLE) { |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1055 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
Xiao Guangrong | 8c9ed8e | 2009-09-25 13:51:17 +0800 | [diff] [blame] | 1056 | group_sched_out(event, cpuctx, ctx); |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1057 | } |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1058 | out: |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1059 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1060 | } |
| 1061 | |
| 1062 | /* |
| 1063 | * Test whether two contexts are equivalent, i.e. whether they |
| 1064 | * have both been cloned from the same version of the same context |
| 1065 | * and they both have the same number of enabled events. |
| 1066 | * If the number of enabled events is the same, then the set |
| 1067 | * of enabled events should be the same, because these are both |
| 1068 | * inherited contexts, therefore we can't access individual events |
| 1069 | * in them directly with an fd; we can only enable/disable all |
| 1070 | * events via prctl, or enable/disable all events in a family |
| 1071 | * via ioctl, which will have the same effect on both contexts. |
| 1072 | */ |
| 1073 | static int context_equiv(struct perf_event_context *ctx1, |
| 1074 | struct perf_event_context *ctx2) |
| 1075 | { |
| 1076 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx |
| 1077 | && ctx1->parent_gen == ctx2->parent_gen |
| 1078 | && !ctx1->pin_count && !ctx2->pin_count; |
| 1079 | } |
| 1080 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1081 | static void __perf_event_sync_stat(struct perf_event *event, |
| 1082 | struct perf_event *next_event) |
| 1083 | { |
| 1084 | u64 value; |
| 1085 | |
| 1086 | if (!event->attr.inherit_stat) |
| 1087 | return; |
| 1088 | |
| 1089 | /* |
| 1090 | * Update the event value, we cannot use perf_event_read() |
| 1091 | * because we're in the middle of a context switch and have IRQs |
| 1092 | * disabled, which upsets smp_call_function_single(), however |
| 1093 | * we know the event must be on the current CPU, therefore we |
| 1094 | * don't need to use it. |
| 1095 | */ |
| 1096 | switch (event->state) { |
| 1097 | case PERF_EVENT_STATE_ACTIVE: |
Peter Zijlstra | 3dbebf1 | 2009-11-20 22:19:52 +0100 | [diff] [blame] | 1098 | event->pmu->read(event); |
| 1099 | /* fall-through */ |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1100 | |
| 1101 | case PERF_EVENT_STATE_INACTIVE: |
| 1102 | update_event_times(event); |
| 1103 | break; |
| 1104 | |
| 1105 | default: |
| 1106 | break; |
| 1107 | } |
| 1108 | |
| 1109 | /* |
| 1110 | * In order to keep per-task stats reliable we need to flip the event |
| 1111 | * values when we flip the contexts. |
| 1112 | */ |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 1113 | value = local64_read(&next_event->count); |
| 1114 | value = local64_xchg(&event->count, value); |
| 1115 | local64_set(&next_event->count, value); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1116 | |
| 1117 | swap(event->total_time_enabled, next_event->total_time_enabled); |
| 1118 | swap(event->total_time_running, next_event->total_time_running); |
| 1119 | |
| 1120 | /* |
| 1121 | * Since we swizzled the values, update the user visible data too. |
| 1122 | */ |
| 1123 | perf_event_update_userpage(event); |
| 1124 | perf_event_update_userpage(next_event); |
| 1125 | } |
| 1126 | |
| 1127 | #define list_next_entry(pos, member) \ |
| 1128 | list_entry(pos->member.next, typeof(*pos), member) |
| 1129 | |
| 1130 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
| 1131 | struct perf_event_context *next_ctx) |
| 1132 | { |
| 1133 | struct perf_event *event, *next_event; |
| 1134 | |
| 1135 | if (!ctx->nr_stat) |
| 1136 | return; |
| 1137 | |
Peter Zijlstra | 02ffdbc | 2009-11-20 22:19:50 +0100 | [diff] [blame] | 1138 | update_context_time(ctx); |
| 1139 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1140 | event = list_first_entry(&ctx->event_list, |
| 1141 | struct perf_event, event_entry); |
| 1142 | |
| 1143 | next_event = list_first_entry(&next_ctx->event_list, |
| 1144 | struct perf_event, event_entry); |
| 1145 | |
| 1146 | while (&event->event_entry != &ctx->event_list && |
| 1147 | &next_event->event_entry != &next_ctx->event_list) { |
| 1148 | |
| 1149 | __perf_event_sync_stat(event, next_event); |
| 1150 | |
| 1151 | event = list_next_entry(event, event_entry); |
| 1152 | next_event = list_next_entry(next_event, event_entry); |
| 1153 | } |
| 1154 | } |
| 1155 | |
| 1156 | /* |
| 1157 | * Called from scheduler to remove the events of the current task, |
| 1158 | * with interrupts disabled. |
| 1159 | * |
| 1160 | * We stop each event and update the event value in event->count. |
| 1161 | * |
| 1162 | * This does not protect us against NMI, but disable() |
| 1163 | * sets the disabled bit in the control field of event _before_ |
| 1164 | * accessing the event control register. If a NMI hits, then it will |
| 1165 | * not restart the event. |
| 1166 | */ |
| 1167 | void perf_event_task_sched_out(struct task_struct *task, |
Peter Zijlstra | 49f4743 | 2009-12-27 11:51:52 +0100 | [diff] [blame] | 1168 | struct task_struct *next) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1169 | { |
Peter Zijlstra | 49f4743 | 2009-12-27 11:51:52 +0100 | [diff] [blame] | 1170 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1171 | struct perf_event_context *ctx = task->perf_event_ctxp; |
| 1172 | struct perf_event_context *next_ctx; |
| 1173 | struct perf_event_context *parent; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1174 | int do_switch = 1; |
| 1175 | |
Frederic Weisbecker | e49a5bd | 2010-03-22 19:40:03 +0100 | [diff] [blame] | 1176 | perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1177 | |
| 1178 | if (likely(!ctx || !cpuctx->task_ctx)) |
| 1179 | return; |
| 1180 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1181 | rcu_read_lock(); |
| 1182 | parent = rcu_dereference(ctx->parent_ctx); |
| 1183 | next_ctx = next->perf_event_ctxp; |
| 1184 | if (parent && next_ctx && |
| 1185 | rcu_dereference(next_ctx->parent_ctx) == parent) { |
| 1186 | /* |
| 1187 | * Looks like the two contexts are clones, so we might be |
| 1188 | * able to optimize the context switch. We lock both |
| 1189 | * contexts and check that they are clones under the |
| 1190 | * lock (including re-checking that neither has been |
| 1191 | * uncloned in the meantime). It doesn't matter which |
| 1192 | * order we take the locks because no other cpu could |
| 1193 | * be trying to lock both of these tasks. |
| 1194 | */ |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1195 | raw_spin_lock(&ctx->lock); |
| 1196 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1197 | if (context_equiv(ctx, next_ctx)) { |
| 1198 | /* |
| 1199 | * XXX do we need a memory barrier of sorts |
| 1200 | * wrt to rcu_dereference() of perf_event_ctxp |
| 1201 | */ |
| 1202 | task->perf_event_ctxp = next_ctx; |
| 1203 | next->perf_event_ctxp = ctx; |
| 1204 | ctx->task = next; |
| 1205 | next_ctx->task = task; |
| 1206 | do_switch = 0; |
| 1207 | |
| 1208 | perf_event_sync_stat(ctx, next_ctx); |
| 1209 | } |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1210 | raw_spin_unlock(&next_ctx->lock); |
| 1211 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1212 | } |
| 1213 | rcu_read_unlock(); |
| 1214 | |
| 1215 | if (do_switch) { |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1216 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1217 | cpuctx->task_ctx = NULL; |
| 1218 | } |
| 1219 | } |
| 1220 | |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1221 | static void task_ctx_sched_out(struct perf_event_context *ctx, |
| 1222 | enum event_type_t event_type) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1223 | { |
| 1224 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
| 1225 | |
| 1226 | if (!cpuctx->task_ctx) |
| 1227 | return; |
| 1228 | |
| 1229 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) |
| 1230 | return; |
| 1231 | |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1232 | ctx_sched_out(ctx, cpuctx, event_type); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1233 | cpuctx->task_ctx = NULL; |
| 1234 | } |
| 1235 | |
| 1236 | /* |
| 1237 | * Called with IRQs disabled |
| 1238 | */ |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1239 | static void __perf_event_task_sched_out(struct perf_event_context *ctx) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1240 | { |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1241 | task_ctx_sched_out(ctx, EVENT_ALL); |
| 1242 | } |
| 1243 | |
| 1244 | /* |
| 1245 | * Called with IRQs disabled |
| 1246 | */ |
| 1247 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
| 1248 | enum event_type_t event_type) |
| 1249 | { |
| 1250 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1251 | } |
| 1252 | |
| 1253 | static void |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1254 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 1255 | struct perf_cpu_context *cpuctx) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1256 | { |
| 1257 | struct perf_event *event; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1258 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1259 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
| 1260 | if (event->state <= PERF_EVENT_STATE_OFF) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1261 | continue; |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 1262 | if (event->cpu != -1 && event->cpu != smp_processor_id()) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1263 | continue; |
| 1264 | |
Xiao Guangrong | 8c9ed8e | 2009-09-25 13:51:17 +0800 | [diff] [blame] | 1265 | if (group_can_go_on(event, cpuctx, 1)) |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 1266 | group_sched_in(event, cpuctx, ctx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1267 | |
| 1268 | /* |
| 1269 | * If this pinned group hasn't been scheduled, |
| 1270 | * put it in error state. |
| 1271 | */ |
| 1272 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
| 1273 | update_group_times(event); |
| 1274 | event->state = PERF_EVENT_STATE_ERROR; |
| 1275 | } |
| 1276 | } |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1277 | } |
| 1278 | |
| 1279 | static void |
| 1280 | ctx_flexible_sched_in(struct perf_event_context *ctx, |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 1281 | struct perf_cpu_context *cpuctx) |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1282 | { |
| 1283 | struct perf_event *event; |
| 1284 | int can_add_hw = 1; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1285 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1286 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
| 1287 | /* Ignore events in OFF or ERROR state */ |
| 1288 | if (event->state <= PERF_EVENT_STATE_OFF) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1289 | continue; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1290 | /* |
| 1291 | * Listen to the 'cpu' scheduling filter constraint |
| 1292 | * of events: |
| 1293 | */ |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 1294 | if (event->cpu != -1 && event->cpu != smp_processor_id()) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1295 | continue; |
| 1296 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1297 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 1298 | if (group_sched_in(event, cpuctx, ctx)) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1299 | can_add_hw = 0; |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1300 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1301 | } |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1302 | } |
| 1303 | |
| 1304 | static void |
| 1305 | ctx_sched_in(struct perf_event_context *ctx, |
| 1306 | struct perf_cpu_context *cpuctx, |
| 1307 | enum event_type_t event_type) |
| 1308 | { |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1309 | raw_spin_lock(&ctx->lock); |
| 1310 | ctx->is_active = 1; |
| 1311 | if (likely(!ctx->nr_events)) |
| 1312 | goto out; |
| 1313 | |
| 1314 | ctx->timestamp = perf_clock(); |
| 1315 | |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1316 | /* |
| 1317 | * First go through the list and put on any pinned groups |
| 1318 | * in order to give them the best chance of going on. |
| 1319 | */ |
| 1320 | if (event_type & EVENT_PINNED) |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 1321 | ctx_pinned_sched_in(ctx, cpuctx); |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1322 | |
| 1323 | /* Then walk through the lower prio flexible groups */ |
| 1324 | if (event_type & EVENT_FLEXIBLE) |
Peter Zijlstra | 6e37738 | 2010-02-11 13:21:58 +0100 | [diff] [blame] | 1325 | ctx_flexible_sched_in(ctx, cpuctx); |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1326 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1327 | out: |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1328 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1329 | } |
| 1330 | |
Frederic Weisbecker | 329c0e0 | 2010-01-17 12:56:05 +0100 | [diff] [blame] | 1331 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
| 1332 | enum event_type_t event_type) |
| 1333 | { |
| 1334 | struct perf_event_context *ctx = &cpuctx->ctx; |
| 1335 | |
| 1336 | ctx_sched_in(ctx, cpuctx, event_type); |
| 1337 | } |
| 1338 | |
Frederic Weisbecker | 5b0311e | 2010-01-17 11:59:13 +0100 | [diff] [blame] | 1339 | static void task_ctx_sched_in(struct task_struct *task, |
| 1340 | enum event_type_t event_type) |
| 1341 | { |
| 1342 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
| 1343 | struct perf_event_context *ctx = task->perf_event_ctxp; |
| 1344 | |
| 1345 | if (likely(!ctx)) |
| 1346 | return; |
| 1347 | if (cpuctx->task_ctx == ctx) |
| 1348 | return; |
| 1349 | ctx_sched_in(ctx, cpuctx, event_type); |
| 1350 | cpuctx->task_ctx = ctx; |
| 1351 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1352 | /* |
| 1353 | * Called from scheduler to add the events of the current task |
| 1354 | * with interrupts disabled. |
| 1355 | * |
| 1356 | * We restore the event value and then enable it. |
| 1357 | * |
| 1358 | * This does not protect us against NMI, but enable() |
| 1359 | * sets the enabled bit in the control field of event _before_ |
| 1360 | * accessing the event control register. If a NMI hits, then it will |
| 1361 | * keep the event running. |
| 1362 | */ |
Peter Zijlstra | 49f4743 | 2009-12-27 11:51:52 +0100 | [diff] [blame] | 1363 | void perf_event_task_sched_in(struct task_struct *task) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1364 | { |
Frederic Weisbecker | 329c0e0 | 2010-01-17 12:56:05 +0100 | [diff] [blame] | 1365 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
| 1366 | struct perf_event_context *ctx = task->perf_event_ctxp; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1367 | |
Frederic Weisbecker | 329c0e0 | 2010-01-17 12:56:05 +0100 | [diff] [blame] | 1368 | if (likely(!ctx)) |
| 1369 | return; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1370 | |
Frederic Weisbecker | 329c0e0 | 2010-01-17 12:56:05 +0100 | [diff] [blame] | 1371 | if (cpuctx->task_ctx == ctx) |
| 1372 | return; |
| 1373 | |
| 1374 | /* |
| 1375 | * We want to keep the following priority order: |
| 1376 | * cpu pinned (that don't need to move), task pinned, |
| 1377 | * cpu flexible, task flexible. |
| 1378 | */ |
| 1379 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
| 1380 | |
| 1381 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED); |
| 1382 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE); |
| 1383 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE); |
| 1384 | |
| 1385 | cpuctx->task_ctx = ctx; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1386 | } |
| 1387 | |
| 1388 | #define MAX_INTERRUPTS (~0ULL) |
| 1389 | |
| 1390 | static void perf_log_throttle(struct perf_event *event, int enable); |
| 1391 | |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 1392 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
| 1393 | { |
| 1394 | u64 frequency = event->attr.sample_freq; |
| 1395 | u64 sec = NSEC_PER_SEC; |
| 1396 | u64 divisor, dividend; |
| 1397 | |
| 1398 | int count_fls, nsec_fls, frequency_fls, sec_fls; |
| 1399 | |
| 1400 | count_fls = fls64(count); |
| 1401 | nsec_fls = fls64(nsec); |
| 1402 | frequency_fls = fls64(frequency); |
| 1403 | sec_fls = 30; |
| 1404 | |
| 1405 | /* |
| 1406 | * We got @count in @nsec, with a target of sample_freq HZ |
| 1407 | * the target period becomes: |
| 1408 | * |
| 1409 | * @count * 10^9 |
| 1410 | * period = ------------------- |
| 1411 | * @nsec * sample_freq |
| 1412 | * |
| 1413 | */ |
| 1414 | |
| 1415 | /* |
| 1416 | * Reduce accuracy by one bit such that @a and @b converge |
| 1417 | * to a similar magnitude. |
| 1418 | */ |
| 1419 | #define REDUCE_FLS(a, b) \ |
| 1420 | do { \ |
| 1421 | if (a##_fls > b##_fls) { \ |
| 1422 | a >>= 1; \ |
| 1423 | a##_fls--; \ |
| 1424 | } else { \ |
| 1425 | b >>= 1; \ |
| 1426 | b##_fls--; \ |
| 1427 | } \ |
| 1428 | } while (0) |
| 1429 | |
| 1430 | /* |
| 1431 | * Reduce accuracy until either term fits in a u64, then proceed with |
| 1432 | * the other, so that finally we can do a u64/u64 division. |
| 1433 | */ |
| 1434 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { |
| 1435 | REDUCE_FLS(nsec, frequency); |
| 1436 | REDUCE_FLS(sec, count); |
| 1437 | } |
| 1438 | |
| 1439 | if (count_fls + sec_fls > 64) { |
| 1440 | divisor = nsec * frequency; |
| 1441 | |
| 1442 | while (count_fls + sec_fls > 64) { |
| 1443 | REDUCE_FLS(count, sec); |
| 1444 | divisor >>= 1; |
| 1445 | } |
| 1446 | |
| 1447 | dividend = count * sec; |
| 1448 | } else { |
| 1449 | dividend = count * sec; |
| 1450 | |
| 1451 | while (nsec_fls + frequency_fls > 64) { |
| 1452 | REDUCE_FLS(nsec, frequency); |
| 1453 | dividend >>= 1; |
| 1454 | } |
| 1455 | |
| 1456 | divisor = nsec * frequency; |
| 1457 | } |
| 1458 | |
Peter Zijlstra | f6ab91a | 2010-06-04 15:18:01 +0200 | [diff] [blame] | 1459 | if (!divisor) |
| 1460 | return dividend; |
| 1461 | |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 1462 | return div64_u64(dividend, divisor); |
| 1463 | } |
| 1464 | |
| 1465 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1466 | { |
| 1467 | struct hw_perf_event *hwc = &event->hw; |
Peter Zijlstra | f6ab91a | 2010-06-04 15:18:01 +0200 | [diff] [blame] | 1468 | s64 period, sample_period; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1469 | s64 delta; |
| 1470 | |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 1471 | period = perf_calculate_period(event, nsec, count); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1472 | |
| 1473 | delta = (s64)(period - hwc->sample_period); |
| 1474 | delta = (delta + 7) / 8; /* low pass filter */ |
| 1475 | |
| 1476 | sample_period = hwc->sample_period + delta; |
| 1477 | |
| 1478 | if (!sample_period) |
| 1479 | sample_period = 1; |
| 1480 | |
| 1481 | hwc->sample_period = sample_period; |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 1482 | |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 1483 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 1484 | event->pmu->stop(event, PERF_EF_UPDATE); |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 1485 | local64_set(&hwc->period_left, 0); |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 1486 | event->pmu->start(event, PERF_EF_RELOAD); |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 1487 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1488 | } |
| 1489 | |
| 1490 | static void perf_ctx_adjust_freq(struct perf_event_context *ctx) |
| 1491 | { |
| 1492 | struct perf_event *event; |
| 1493 | struct hw_perf_event *hwc; |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 1494 | u64 interrupts, now; |
| 1495 | s64 delta; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1496 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1497 | raw_spin_lock(&ctx->lock); |
Paul Mackerras | 03541f8 | 2009-10-14 16:58:03 +1100 | [diff] [blame] | 1498 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1499 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
| 1500 | continue; |
| 1501 | |
Peter Zijlstra | 5d27c23 | 2009-12-17 13:16:32 +0100 | [diff] [blame] | 1502 | if (event->cpu != -1 && event->cpu != smp_processor_id()) |
| 1503 | continue; |
| 1504 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1505 | hwc = &event->hw; |
| 1506 | |
| 1507 | interrupts = hwc->interrupts; |
| 1508 | hwc->interrupts = 0; |
| 1509 | |
| 1510 | /* |
| 1511 | * unthrottle events on the tick |
| 1512 | */ |
| 1513 | if (interrupts == MAX_INTERRUPTS) { |
| 1514 | perf_log_throttle(event, 1); |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 1515 | event->pmu->start(event, 0); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1516 | } |
| 1517 | |
| 1518 | if (!event->attr.freq || !event->attr.sample_freq) |
| 1519 | continue; |
| 1520 | |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 1521 | event->pmu->read(event); |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 1522 | now = local64_read(&event->count); |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 1523 | delta = now - hwc->freq_count_stamp; |
| 1524 | hwc->freq_count_stamp = now; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1525 | |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 1526 | if (delta > 0) |
| 1527 | perf_adjust_period(event, TICK_NSEC, delta); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1528 | } |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1529 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1530 | } |
| 1531 | |
| 1532 | /* |
| 1533 | * Round-robin a context's events: |
| 1534 | */ |
| 1535 | static void rotate_ctx(struct perf_event_context *ctx) |
| 1536 | { |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1537 | raw_spin_lock(&ctx->lock); |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1538 | |
Frederic Weisbecker | e286417 | 2010-01-09 21:05:28 +0100 | [diff] [blame] | 1539 | /* Rotate the first entry last of non-pinned groups */ |
Frederic Weisbecker | e286417 | 2010-01-09 21:05:28 +0100 | [diff] [blame] | 1540 | list_rotate_left(&ctx->flexible_groups); |
| 1541 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1542 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1543 | } |
| 1544 | |
Peter Zijlstra | 49f4743 | 2009-12-27 11:51:52 +0100 | [diff] [blame] | 1545 | void perf_event_task_tick(struct task_struct *curr) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1546 | { |
| 1547 | struct perf_cpu_context *cpuctx; |
| 1548 | struct perf_event_context *ctx; |
Peter Zijlstra | d4944a0 | 2010-03-08 13:51:20 +0100 | [diff] [blame] | 1549 | int rotate = 0; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1550 | |
| 1551 | if (!atomic_read(&nr_events)) |
| 1552 | return; |
| 1553 | |
Peter Zijlstra | 49f4743 | 2009-12-27 11:51:52 +0100 | [diff] [blame] | 1554 | cpuctx = &__get_cpu_var(perf_cpu_context); |
Peter Zijlstra | d4944a0 | 2010-03-08 13:51:20 +0100 | [diff] [blame] | 1555 | if (cpuctx->ctx.nr_events && |
| 1556 | cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
| 1557 | rotate = 1; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1558 | |
Peter Zijlstra | d4944a0 | 2010-03-08 13:51:20 +0100 | [diff] [blame] | 1559 | ctx = curr->perf_event_ctxp; |
| 1560 | if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active) |
| 1561 | rotate = 1; |
Peter Zijlstra | 9717e6c | 2010-01-28 13:57:44 +0100 | [diff] [blame] | 1562 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1563 | perf_ctx_adjust_freq(&cpuctx->ctx); |
| 1564 | if (ctx) |
| 1565 | perf_ctx_adjust_freq(ctx); |
| 1566 | |
Peter Zijlstra | d4944a0 | 2010-03-08 13:51:20 +0100 | [diff] [blame] | 1567 | if (!rotate) |
| 1568 | return; |
| 1569 | |
Frederic Weisbecker | 7defb0f | 2010-01-17 12:15:31 +0100 | [diff] [blame] | 1570 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1571 | if (ctx) |
Frederic Weisbecker | 7defb0f | 2010-01-17 12:15:31 +0100 | [diff] [blame] | 1572 | task_ctx_sched_out(ctx, EVENT_FLEXIBLE); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1573 | |
| 1574 | rotate_ctx(&cpuctx->ctx); |
| 1575 | if (ctx) |
| 1576 | rotate_ctx(ctx); |
| 1577 | |
Frederic Weisbecker | 7defb0f | 2010-01-17 12:15:31 +0100 | [diff] [blame] | 1578 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1579 | if (ctx) |
Frederic Weisbecker | 7defb0f | 2010-01-17 12:15:31 +0100 | [diff] [blame] | 1580 | task_ctx_sched_in(curr, EVENT_FLEXIBLE); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1581 | } |
| 1582 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1583 | static int event_enable_on_exec(struct perf_event *event, |
| 1584 | struct perf_event_context *ctx) |
| 1585 | { |
| 1586 | if (!event->attr.enable_on_exec) |
| 1587 | return 0; |
| 1588 | |
| 1589 | event->attr.enable_on_exec = 0; |
| 1590 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
| 1591 | return 0; |
| 1592 | |
| 1593 | __perf_event_mark_enabled(event, ctx); |
| 1594 | |
| 1595 | return 1; |
| 1596 | } |
| 1597 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1598 | /* |
| 1599 | * Enable all of a task's events that have been marked enable-on-exec. |
| 1600 | * This expects task == current. |
| 1601 | */ |
| 1602 | static void perf_event_enable_on_exec(struct task_struct *task) |
| 1603 | { |
| 1604 | struct perf_event_context *ctx; |
| 1605 | struct perf_event *event; |
| 1606 | unsigned long flags; |
| 1607 | int enabled = 0; |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1608 | int ret; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1609 | |
| 1610 | local_irq_save(flags); |
| 1611 | ctx = task->perf_event_ctxp; |
| 1612 | if (!ctx || !ctx->nr_events) |
| 1613 | goto out; |
| 1614 | |
| 1615 | __perf_event_task_sched_out(ctx); |
| 1616 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1617 | raw_spin_lock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1618 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1619 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
| 1620 | ret = event_enable_on_exec(event, ctx); |
| 1621 | if (ret) |
| 1622 | enabled = 1; |
| 1623 | } |
| 1624 | |
| 1625 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
| 1626 | ret = event_enable_on_exec(event, ctx); |
| 1627 | if (ret) |
| 1628 | enabled = 1; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1629 | } |
| 1630 | |
| 1631 | /* |
| 1632 | * Unclone this context if we enabled any event. |
| 1633 | */ |
| 1634 | if (enabled) |
| 1635 | unclone_ctx(ctx); |
| 1636 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1637 | raw_spin_unlock(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1638 | |
Peter Zijlstra | 49f4743 | 2009-12-27 11:51:52 +0100 | [diff] [blame] | 1639 | perf_event_task_sched_in(task); |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1640 | out: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1641 | local_irq_restore(flags); |
| 1642 | } |
| 1643 | |
| 1644 | /* |
| 1645 | * Cross CPU call to read the hardware event |
| 1646 | */ |
| 1647 | static void __perf_event_read(void *info) |
| 1648 | { |
| 1649 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
| 1650 | struct perf_event *event = info; |
| 1651 | struct perf_event_context *ctx = event->ctx; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1652 | |
| 1653 | /* |
| 1654 | * If this is a task context, we need to check whether it is |
| 1655 | * the current task context of this cpu. If not it has been |
| 1656 | * scheduled out before the smp call arrived. In that case |
| 1657 | * event->count would have been updated to a recent sample |
| 1658 | * when the event was scheduled out. |
| 1659 | */ |
| 1660 | if (ctx->task && cpuctx->task_ctx != ctx) |
| 1661 | return; |
| 1662 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1663 | raw_spin_lock(&ctx->lock); |
Peter Zijlstra | 58e5ad1 | 2009-11-20 22:19:53 +0100 | [diff] [blame] | 1664 | update_context_time(ctx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1665 | update_event_times(event); |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1666 | raw_spin_unlock(&ctx->lock); |
Peter Zijlstra | 2b8988c | 2009-11-20 22:19:54 +0100 | [diff] [blame] | 1667 | |
Peter Zijlstra | 58e5ad1 | 2009-11-20 22:19:53 +0100 | [diff] [blame] | 1668 | event->pmu->read(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1669 | } |
| 1670 | |
Peter Zijlstra | b5e5879 | 2010-05-21 14:43:12 +0200 | [diff] [blame] | 1671 | static inline u64 perf_event_count(struct perf_event *event) |
| 1672 | { |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 1673 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
Peter Zijlstra | b5e5879 | 2010-05-21 14:43:12 +0200 | [diff] [blame] | 1674 | } |
| 1675 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1676 | static u64 perf_event_read(struct perf_event *event) |
| 1677 | { |
| 1678 | /* |
| 1679 | * If event is enabled and currently active on a CPU, update the |
| 1680 | * value in the event structure: |
| 1681 | */ |
| 1682 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
| 1683 | smp_call_function_single(event->oncpu, |
| 1684 | __perf_event_read, event, 1); |
| 1685 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { |
Peter Zijlstra | 2b8988c | 2009-11-20 22:19:54 +0100 | [diff] [blame] | 1686 | struct perf_event_context *ctx = event->ctx; |
| 1687 | unsigned long flags; |
| 1688 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1689 | raw_spin_lock_irqsave(&ctx->lock, flags); |
Peter Zijlstra | 2b8988c | 2009-11-20 22:19:54 +0100 | [diff] [blame] | 1690 | update_context_time(ctx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1691 | update_event_times(event); |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1692 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1693 | } |
| 1694 | |
Peter Zijlstra | b5e5879 | 2010-05-21 14:43:12 +0200 | [diff] [blame] | 1695 | return perf_event_count(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1696 | } |
| 1697 | |
| 1698 | /* |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 1699 | * Callchain support |
| 1700 | */ |
| 1701 | |
| 1702 | struct callchain_cpus_entries { |
| 1703 | struct rcu_head rcu_head; |
| 1704 | struct perf_callchain_entry *cpu_entries[0]; |
| 1705 | }; |
| 1706 | |
Frederic Weisbecker | 7ae07ea | 2010-08-14 20:45:13 +0200 | [diff] [blame] | 1707 | static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]); |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 1708 | static atomic_t nr_callchain_events; |
| 1709 | static DEFINE_MUTEX(callchain_mutex); |
| 1710 | struct callchain_cpus_entries *callchain_cpus_entries; |
| 1711 | |
| 1712 | |
| 1713 | __weak void perf_callchain_kernel(struct perf_callchain_entry *entry, |
| 1714 | struct pt_regs *regs) |
| 1715 | { |
| 1716 | } |
| 1717 | |
| 1718 | __weak void perf_callchain_user(struct perf_callchain_entry *entry, |
| 1719 | struct pt_regs *regs) |
| 1720 | { |
| 1721 | } |
| 1722 | |
| 1723 | static void release_callchain_buffers_rcu(struct rcu_head *head) |
| 1724 | { |
| 1725 | struct callchain_cpus_entries *entries; |
| 1726 | int cpu; |
| 1727 | |
| 1728 | entries = container_of(head, struct callchain_cpus_entries, rcu_head); |
| 1729 | |
| 1730 | for_each_possible_cpu(cpu) |
| 1731 | kfree(entries->cpu_entries[cpu]); |
| 1732 | |
| 1733 | kfree(entries); |
| 1734 | } |
| 1735 | |
| 1736 | static void release_callchain_buffers(void) |
| 1737 | { |
| 1738 | struct callchain_cpus_entries *entries; |
| 1739 | |
| 1740 | entries = callchain_cpus_entries; |
| 1741 | rcu_assign_pointer(callchain_cpus_entries, NULL); |
| 1742 | call_rcu(&entries->rcu_head, release_callchain_buffers_rcu); |
| 1743 | } |
| 1744 | |
| 1745 | static int alloc_callchain_buffers(void) |
| 1746 | { |
| 1747 | int cpu; |
| 1748 | int size; |
| 1749 | struct callchain_cpus_entries *entries; |
| 1750 | |
| 1751 | /* |
| 1752 | * We can't use the percpu allocation API for data that can be |
| 1753 | * accessed from NMI. Use a temporary manual per cpu allocation |
| 1754 | * until that gets sorted out. |
| 1755 | */ |
| 1756 | size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) * |
| 1757 | num_possible_cpus(); |
| 1758 | |
| 1759 | entries = kzalloc(size, GFP_KERNEL); |
| 1760 | if (!entries) |
| 1761 | return -ENOMEM; |
| 1762 | |
Frederic Weisbecker | 7ae07ea | 2010-08-14 20:45:13 +0200 | [diff] [blame] | 1763 | size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS; |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 1764 | |
| 1765 | for_each_possible_cpu(cpu) { |
| 1766 | entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL, |
| 1767 | cpu_to_node(cpu)); |
| 1768 | if (!entries->cpu_entries[cpu]) |
| 1769 | goto fail; |
| 1770 | } |
| 1771 | |
| 1772 | rcu_assign_pointer(callchain_cpus_entries, entries); |
| 1773 | |
| 1774 | return 0; |
| 1775 | |
| 1776 | fail: |
| 1777 | for_each_possible_cpu(cpu) |
| 1778 | kfree(entries->cpu_entries[cpu]); |
| 1779 | kfree(entries); |
| 1780 | |
| 1781 | return -ENOMEM; |
| 1782 | } |
| 1783 | |
| 1784 | static int get_callchain_buffers(void) |
| 1785 | { |
| 1786 | int err = 0; |
| 1787 | int count; |
| 1788 | |
| 1789 | mutex_lock(&callchain_mutex); |
| 1790 | |
| 1791 | count = atomic_inc_return(&nr_callchain_events); |
| 1792 | if (WARN_ON_ONCE(count < 1)) { |
| 1793 | err = -EINVAL; |
| 1794 | goto exit; |
| 1795 | } |
| 1796 | |
| 1797 | if (count > 1) { |
| 1798 | /* If the allocation failed, give up */ |
| 1799 | if (!callchain_cpus_entries) |
| 1800 | err = -ENOMEM; |
| 1801 | goto exit; |
| 1802 | } |
| 1803 | |
| 1804 | err = alloc_callchain_buffers(); |
| 1805 | if (err) |
| 1806 | release_callchain_buffers(); |
| 1807 | exit: |
| 1808 | mutex_unlock(&callchain_mutex); |
| 1809 | |
| 1810 | return err; |
| 1811 | } |
| 1812 | |
| 1813 | static void put_callchain_buffers(void) |
| 1814 | { |
| 1815 | if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) { |
| 1816 | release_callchain_buffers(); |
| 1817 | mutex_unlock(&callchain_mutex); |
| 1818 | } |
| 1819 | } |
| 1820 | |
| 1821 | static int get_recursion_context(int *recursion) |
| 1822 | { |
| 1823 | int rctx; |
| 1824 | |
| 1825 | if (in_nmi()) |
| 1826 | rctx = 3; |
| 1827 | else if (in_irq()) |
| 1828 | rctx = 2; |
| 1829 | else if (in_softirq()) |
| 1830 | rctx = 1; |
| 1831 | else |
| 1832 | rctx = 0; |
| 1833 | |
| 1834 | if (recursion[rctx]) |
| 1835 | return -1; |
| 1836 | |
| 1837 | recursion[rctx]++; |
| 1838 | barrier(); |
| 1839 | |
| 1840 | return rctx; |
| 1841 | } |
| 1842 | |
| 1843 | static inline void put_recursion_context(int *recursion, int rctx) |
| 1844 | { |
| 1845 | barrier(); |
| 1846 | recursion[rctx]--; |
| 1847 | } |
| 1848 | |
| 1849 | static struct perf_callchain_entry *get_callchain_entry(int *rctx) |
| 1850 | { |
| 1851 | int cpu; |
| 1852 | struct callchain_cpus_entries *entries; |
| 1853 | |
| 1854 | *rctx = get_recursion_context(__get_cpu_var(callchain_recursion)); |
| 1855 | if (*rctx == -1) |
| 1856 | return NULL; |
| 1857 | |
| 1858 | entries = rcu_dereference(callchain_cpus_entries); |
| 1859 | if (!entries) |
| 1860 | return NULL; |
| 1861 | |
| 1862 | cpu = smp_processor_id(); |
| 1863 | |
| 1864 | return &entries->cpu_entries[cpu][*rctx]; |
| 1865 | } |
| 1866 | |
| 1867 | static void |
| 1868 | put_callchain_entry(int rctx) |
| 1869 | { |
| 1870 | put_recursion_context(__get_cpu_var(callchain_recursion), rctx); |
| 1871 | } |
| 1872 | |
| 1873 | static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs) |
| 1874 | { |
| 1875 | int rctx; |
| 1876 | struct perf_callchain_entry *entry; |
| 1877 | |
| 1878 | |
| 1879 | entry = get_callchain_entry(&rctx); |
| 1880 | if (rctx == -1) |
| 1881 | return NULL; |
| 1882 | |
| 1883 | if (!entry) |
| 1884 | goto exit_put; |
| 1885 | |
| 1886 | entry->nr = 0; |
| 1887 | |
| 1888 | if (!user_mode(regs)) { |
| 1889 | perf_callchain_store(entry, PERF_CONTEXT_KERNEL); |
| 1890 | perf_callchain_kernel(entry, regs); |
| 1891 | if (current->mm) |
| 1892 | regs = task_pt_regs(current); |
| 1893 | else |
| 1894 | regs = NULL; |
| 1895 | } |
| 1896 | |
| 1897 | if (regs) { |
| 1898 | perf_callchain_store(entry, PERF_CONTEXT_USER); |
| 1899 | perf_callchain_user(entry, regs); |
| 1900 | } |
| 1901 | |
| 1902 | exit_put: |
| 1903 | put_callchain_entry(rctx); |
| 1904 | |
| 1905 | return entry; |
| 1906 | } |
| 1907 | |
| 1908 | /* |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1909 | * Initialize the perf_event context in a task_struct: |
| 1910 | */ |
| 1911 | static void |
| 1912 | __perf_event_init_context(struct perf_event_context *ctx, |
| 1913 | struct task_struct *task) |
| 1914 | { |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1915 | raw_spin_lock_init(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1916 | mutex_init(&ctx->mutex); |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 1917 | INIT_LIST_HEAD(&ctx->pinned_groups); |
| 1918 | INIT_LIST_HEAD(&ctx->flexible_groups); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1919 | INIT_LIST_HEAD(&ctx->event_list); |
| 1920 | atomic_set(&ctx->refcount, 1); |
| 1921 | ctx->task = task; |
| 1922 | } |
| 1923 | |
| 1924 | static struct perf_event_context *find_get_context(pid_t pid, int cpu) |
| 1925 | { |
| 1926 | struct perf_event_context *ctx; |
| 1927 | struct perf_cpu_context *cpuctx; |
| 1928 | struct task_struct *task; |
| 1929 | unsigned long flags; |
| 1930 | int err; |
| 1931 | |
Peter Zijlstra | f4c4176 | 2009-12-16 17:55:54 +0100 | [diff] [blame] | 1932 | if (pid == -1 && cpu != -1) { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1933 | /* Must be root to operate on a CPU event: */ |
| 1934 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
| 1935 | return ERR_PTR(-EACCES); |
| 1936 | |
Paul Mackerras | 0f624e7 | 2009-12-15 19:40:32 +1100 | [diff] [blame] | 1937 | if (cpu < 0 || cpu >= nr_cpumask_bits) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1938 | return ERR_PTR(-EINVAL); |
| 1939 | |
| 1940 | /* |
| 1941 | * We could be clever and allow to attach a event to an |
| 1942 | * offline CPU and activate it when the CPU comes up, but |
| 1943 | * that's for later. |
| 1944 | */ |
Rusty Russell | f6325e3 | 2009-12-17 11:43:08 -0600 | [diff] [blame] | 1945 | if (!cpu_online(cpu)) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1946 | return ERR_PTR(-ENODEV); |
| 1947 | |
| 1948 | cpuctx = &per_cpu(perf_cpu_context, cpu); |
| 1949 | ctx = &cpuctx->ctx; |
| 1950 | get_ctx(ctx); |
| 1951 | |
| 1952 | return ctx; |
| 1953 | } |
| 1954 | |
| 1955 | rcu_read_lock(); |
| 1956 | if (!pid) |
| 1957 | task = current; |
| 1958 | else |
| 1959 | task = find_task_by_vpid(pid); |
| 1960 | if (task) |
| 1961 | get_task_struct(task); |
| 1962 | rcu_read_unlock(); |
| 1963 | |
| 1964 | if (!task) |
| 1965 | return ERR_PTR(-ESRCH); |
| 1966 | |
| 1967 | /* |
| 1968 | * Can't attach events to a dying task. |
| 1969 | */ |
| 1970 | err = -ESRCH; |
| 1971 | if (task->flags & PF_EXITING) |
| 1972 | goto errout; |
| 1973 | |
| 1974 | /* Reuse ptrace permission checks for now. */ |
| 1975 | err = -EACCES; |
| 1976 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) |
| 1977 | goto errout; |
| 1978 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 1979 | retry: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1980 | ctx = perf_lock_task_context(task, &flags); |
| 1981 | if (ctx) { |
| 1982 | unclone_ctx(ctx); |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 1983 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1984 | } |
| 1985 | |
| 1986 | if (!ctx) { |
Xiao Guangrong | aa5452d | 2009-12-09 11:28:13 +0800 | [diff] [blame] | 1987 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 1988 | err = -ENOMEM; |
| 1989 | if (!ctx) |
| 1990 | goto errout; |
| 1991 | __perf_event_init_context(ctx, task); |
| 1992 | get_ctx(ctx); |
| 1993 | if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) { |
| 1994 | /* |
| 1995 | * We raced with some other task; use |
| 1996 | * the context they set. |
| 1997 | */ |
| 1998 | kfree(ctx); |
| 1999 | goto retry; |
| 2000 | } |
| 2001 | get_task_struct(task); |
| 2002 | } |
| 2003 | |
| 2004 | put_task_struct(task); |
| 2005 | return ctx; |
| 2006 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 2007 | errout: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2008 | put_task_struct(task); |
| 2009 | return ERR_PTR(err); |
| 2010 | } |
| 2011 | |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 2012 | static void perf_event_free_filter(struct perf_event *event); |
| 2013 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2014 | static void free_event_rcu(struct rcu_head *head) |
| 2015 | { |
| 2016 | struct perf_event *event; |
| 2017 | |
| 2018 | event = container_of(head, struct perf_event, rcu_head); |
| 2019 | if (event->ns) |
| 2020 | put_pid_ns(event->ns); |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 2021 | perf_event_free_filter(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2022 | kfree(event); |
| 2023 | } |
| 2024 | |
| 2025 | static void perf_pending_sync(struct perf_event *event); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2026 | static void perf_buffer_put(struct perf_buffer *buffer); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2027 | |
| 2028 | static void free_event(struct perf_event *event) |
| 2029 | { |
| 2030 | perf_pending_sync(event); |
| 2031 | |
| 2032 | if (!event->parent) { |
| 2033 | atomic_dec(&nr_events); |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 2034 | if (event->attr.mmap || event->attr.mmap_data) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2035 | atomic_dec(&nr_mmap_events); |
| 2036 | if (event->attr.comm) |
| 2037 | atomic_dec(&nr_comm_events); |
| 2038 | if (event->attr.task) |
| 2039 | atomic_dec(&nr_task_events); |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 2040 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
| 2041 | put_callchain_buffers(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2042 | } |
| 2043 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2044 | if (event->buffer) { |
| 2045 | perf_buffer_put(event->buffer); |
| 2046 | event->buffer = NULL; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2047 | } |
| 2048 | |
| 2049 | if (event->destroy) |
| 2050 | event->destroy(event); |
| 2051 | |
| 2052 | put_ctx(event->ctx); |
| 2053 | call_rcu(&event->rcu_head, free_event_rcu); |
| 2054 | } |
| 2055 | |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 2056 | int perf_event_release_kernel(struct perf_event *event) |
| 2057 | { |
| 2058 | struct perf_event_context *ctx = event->ctx; |
| 2059 | |
Peter Zijlstra | 050735b | 2010-05-11 11:51:53 +0200 | [diff] [blame] | 2060 | /* |
| 2061 | * Remove from the PMU, can't get re-enabled since we got |
| 2062 | * here because the last ref went. |
| 2063 | */ |
| 2064 | perf_event_disable(event); |
| 2065 | |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 2066 | WARN_ON_ONCE(ctx->parent_ctx); |
Peter Zijlstra | a0507c8 | 2010-05-06 15:42:53 +0200 | [diff] [blame] | 2067 | /* |
| 2068 | * There are two ways this annotation is useful: |
| 2069 | * |
| 2070 | * 1) there is a lock recursion from perf_event_exit_task |
| 2071 | * see the comment there. |
| 2072 | * |
| 2073 | * 2) there is a lock-inversion with mmap_sem through |
| 2074 | * perf_event_read_group(), which takes faults while |
| 2075 | * holding ctx->mutex, however this is called after |
| 2076 | * the last filedesc died, so there is no possibility |
| 2077 | * to trigger the AB-BA case. |
| 2078 | */ |
| 2079 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); |
Peter Zijlstra | 050735b | 2010-05-11 11:51:53 +0200 | [diff] [blame] | 2080 | raw_spin_lock_irq(&ctx->lock); |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 2081 | perf_group_detach(event); |
Peter Zijlstra | 050735b | 2010-05-11 11:51:53 +0200 | [diff] [blame] | 2082 | list_del_event(event, ctx); |
Peter Zijlstra | 050735b | 2010-05-11 11:51:53 +0200 | [diff] [blame] | 2083 | raw_spin_unlock_irq(&ctx->lock); |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 2084 | mutex_unlock(&ctx->mutex); |
| 2085 | |
| 2086 | mutex_lock(&event->owner->perf_event_mutex); |
| 2087 | list_del_init(&event->owner_entry); |
| 2088 | mutex_unlock(&event->owner->perf_event_mutex); |
| 2089 | put_task_struct(event->owner); |
| 2090 | |
| 2091 | free_event(event); |
| 2092 | |
| 2093 | return 0; |
| 2094 | } |
| 2095 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
| 2096 | |
Peter Zijlstra | a66a305 | 2009-11-23 11:37:23 +0100 | [diff] [blame] | 2097 | /* |
| 2098 | * Called when the last reference to the file is gone. |
| 2099 | */ |
| 2100 | static int perf_release(struct inode *inode, struct file *file) |
| 2101 | { |
| 2102 | struct perf_event *event = file->private_data; |
| 2103 | |
| 2104 | file->private_data = NULL; |
| 2105 | |
| 2106 | return perf_event_release_kernel(event); |
| 2107 | } |
| 2108 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2109 | static int perf_event_read_size(struct perf_event *event) |
| 2110 | { |
| 2111 | int entry = sizeof(u64); /* value */ |
| 2112 | int size = 0; |
| 2113 | int nr = 1; |
| 2114 | |
| 2115 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
| 2116 | size += sizeof(u64); |
| 2117 | |
| 2118 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
| 2119 | size += sizeof(u64); |
| 2120 | |
| 2121 | if (event->attr.read_format & PERF_FORMAT_ID) |
| 2122 | entry += sizeof(u64); |
| 2123 | |
| 2124 | if (event->attr.read_format & PERF_FORMAT_GROUP) { |
| 2125 | nr += event->group_leader->nr_siblings; |
| 2126 | size += sizeof(u64); |
| 2127 | } |
| 2128 | |
| 2129 | size += entry * nr; |
| 2130 | |
| 2131 | return size; |
| 2132 | } |
| 2133 | |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2134 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2135 | { |
| 2136 | struct perf_event *child; |
| 2137 | u64 total = 0; |
| 2138 | |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2139 | *enabled = 0; |
| 2140 | *running = 0; |
| 2141 | |
Peter Zijlstra | 6f10581 | 2009-11-20 22:19:56 +0100 | [diff] [blame] | 2142 | mutex_lock(&event->child_mutex); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2143 | total += perf_event_read(event); |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2144 | *enabled += event->total_time_enabled + |
| 2145 | atomic64_read(&event->child_total_time_enabled); |
| 2146 | *running += event->total_time_running + |
| 2147 | atomic64_read(&event->child_total_time_running); |
| 2148 | |
| 2149 | list_for_each_entry(child, &event->child_list, child_list) { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2150 | total += perf_event_read(child); |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2151 | *enabled += child->total_time_enabled; |
| 2152 | *running += child->total_time_running; |
| 2153 | } |
Peter Zijlstra | 6f10581 | 2009-11-20 22:19:56 +0100 | [diff] [blame] | 2154 | mutex_unlock(&event->child_mutex); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2155 | |
| 2156 | return total; |
| 2157 | } |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 2158 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2159 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2160 | static int perf_event_read_group(struct perf_event *event, |
| 2161 | u64 read_format, char __user *buf) |
| 2162 | { |
| 2163 | struct perf_event *leader = event->group_leader, *sub; |
Peter Zijlstra | 6f10581 | 2009-11-20 22:19:56 +0100 | [diff] [blame] | 2164 | int n = 0, size = 0, ret = -EFAULT; |
| 2165 | struct perf_event_context *ctx = leader->ctx; |
Peter Zijlstra | abf4868 | 2009-11-20 22:19:49 +0100 | [diff] [blame] | 2166 | u64 values[5]; |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2167 | u64 count, enabled, running; |
Peter Zijlstra | abf4868 | 2009-11-20 22:19:49 +0100 | [diff] [blame] | 2168 | |
Peter Zijlstra | 6f10581 | 2009-11-20 22:19:56 +0100 | [diff] [blame] | 2169 | mutex_lock(&ctx->mutex); |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2170 | count = perf_event_read_value(leader, &enabled, &running); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2171 | |
| 2172 | values[n++] = 1 + leader->nr_siblings; |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2173 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
| 2174 | values[n++] = enabled; |
| 2175 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
| 2176 | values[n++] = running; |
Peter Zijlstra | abf4868 | 2009-11-20 22:19:49 +0100 | [diff] [blame] | 2177 | values[n++] = count; |
| 2178 | if (read_format & PERF_FORMAT_ID) |
| 2179 | values[n++] = primary_event_id(leader); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2180 | |
| 2181 | size = n * sizeof(u64); |
| 2182 | |
| 2183 | if (copy_to_user(buf, values, size)) |
Peter Zijlstra | 6f10581 | 2009-11-20 22:19:56 +0100 | [diff] [blame] | 2184 | goto unlock; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2185 | |
Peter Zijlstra | 6f10581 | 2009-11-20 22:19:56 +0100 | [diff] [blame] | 2186 | ret = size; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2187 | |
| 2188 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
Peter Zijlstra | abf4868 | 2009-11-20 22:19:49 +0100 | [diff] [blame] | 2189 | n = 0; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2190 | |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2191 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
Peter Zijlstra | abf4868 | 2009-11-20 22:19:49 +0100 | [diff] [blame] | 2192 | if (read_format & PERF_FORMAT_ID) |
| 2193 | values[n++] = primary_event_id(sub); |
| 2194 | |
| 2195 | size = n * sizeof(u64); |
| 2196 | |
Stephane Eranian | 184d3da | 2009-11-23 21:40:49 -0800 | [diff] [blame] | 2197 | if (copy_to_user(buf + ret, values, size)) { |
Peter Zijlstra | 6f10581 | 2009-11-20 22:19:56 +0100 | [diff] [blame] | 2198 | ret = -EFAULT; |
| 2199 | goto unlock; |
| 2200 | } |
Peter Zijlstra | abf4868 | 2009-11-20 22:19:49 +0100 | [diff] [blame] | 2201 | |
| 2202 | ret += size; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2203 | } |
Peter Zijlstra | 6f10581 | 2009-11-20 22:19:56 +0100 | [diff] [blame] | 2204 | unlock: |
| 2205 | mutex_unlock(&ctx->mutex); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2206 | |
Peter Zijlstra | abf4868 | 2009-11-20 22:19:49 +0100 | [diff] [blame] | 2207 | return ret; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2208 | } |
| 2209 | |
| 2210 | static int perf_event_read_one(struct perf_event *event, |
| 2211 | u64 read_format, char __user *buf) |
| 2212 | { |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2213 | u64 enabled, running; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2214 | u64 values[4]; |
| 2215 | int n = 0; |
| 2216 | |
Peter Zijlstra | 59ed446 | 2009-11-20 22:19:55 +0100 | [diff] [blame] | 2217 | values[n++] = perf_event_read_value(event, &enabled, &running); |
| 2218 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
| 2219 | values[n++] = enabled; |
| 2220 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
| 2221 | values[n++] = running; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2222 | if (read_format & PERF_FORMAT_ID) |
| 2223 | values[n++] = primary_event_id(event); |
| 2224 | |
| 2225 | if (copy_to_user(buf, values, n * sizeof(u64))) |
| 2226 | return -EFAULT; |
| 2227 | |
| 2228 | return n * sizeof(u64); |
| 2229 | } |
| 2230 | |
| 2231 | /* |
| 2232 | * Read the performance event - simple non blocking version for now |
| 2233 | */ |
| 2234 | static ssize_t |
| 2235 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
| 2236 | { |
| 2237 | u64 read_format = event->attr.read_format; |
| 2238 | int ret; |
| 2239 | |
| 2240 | /* |
| 2241 | * Return end-of-file for a read on a event that is in |
| 2242 | * error state (i.e. because it was pinned but it couldn't be |
| 2243 | * scheduled on to the CPU at some point). |
| 2244 | */ |
| 2245 | if (event->state == PERF_EVENT_STATE_ERROR) |
| 2246 | return 0; |
| 2247 | |
| 2248 | if (count < perf_event_read_size(event)) |
| 2249 | return -ENOSPC; |
| 2250 | |
| 2251 | WARN_ON_ONCE(event->ctx->parent_ctx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2252 | if (read_format & PERF_FORMAT_GROUP) |
| 2253 | ret = perf_event_read_group(event, read_format, buf); |
| 2254 | else |
| 2255 | ret = perf_event_read_one(event, read_format, buf); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2256 | |
| 2257 | return ret; |
| 2258 | } |
| 2259 | |
| 2260 | static ssize_t |
| 2261 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) |
| 2262 | { |
| 2263 | struct perf_event *event = file->private_data; |
| 2264 | |
| 2265 | return perf_read_hw(event, buf, count); |
| 2266 | } |
| 2267 | |
| 2268 | static unsigned int perf_poll(struct file *file, poll_table *wait) |
| 2269 | { |
| 2270 | struct perf_event *event = file->private_data; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2271 | struct perf_buffer *buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2272 | unsigned int events = POLL_HUP; |
| 2273 | |
| 2274 | rcu_read_lock(); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2275 | buffer = rcu_dereference(event->buffer); |
| 2276 | if (buffer) |
| 2277 | events = atomic_xchg(&buffer->poll, 0); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2278 | rcu_read_unlock(); |
| 2279 | |
| 2280 | poll_wait(file, &event->waitq, wait); |
| 2281 | |
| 2282 | return events; |
| 2283 | } |
| 2284 | |
| 2285 | static void perf_event_reset(struct perf_event *event) |
| 2286 | { |
| 2287 | (void)perf_event_read(event); |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 2288 | local64_set(&event->count, 0); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2289 | perf_event_update_userpage(event); |
| 2290 | } |
| 2291 | |
| 2292 | /* |
| 2293 | * Holding the top-level event's child_mutex means that any |
| 2294 | * descendant process that has inherited this event will block |
| 2295 | * in sync_child_event if it goes to exit, thus satisfying the |
| 2296 | * task existence requirements of perf_event_enable/disable. |
| 2297 | */ |
| 2298 | static void perf_event_for_each_child(struct perf_event *event, |
| 2299 | void (*func)(struct perf_event *)) |
| 2300 | { |
| 2301 | struct perf_event *child; |
| 2302 | |
| 2303 | WARN_ON_ONCE(event->ctx->parent_ctx); |
| 2304 | mutex_lock(&event->child_mutex); |
| 2305 | func(event); |
| 2306 | list_for_each_entry(child, &event->child_list, child_list) |
| 2307 | func(child); |
| 2308 | mutex_unlock(&event->child_mutex); |
| 2309 | } |
| 2310 | |
| 2311 | static void perf_event_for_each(struct perf_event *event, |
| 2312 | void (*func)(struct perf_event *)) |
| 2313 | { |
| 2314 | struct perf_event_context *ctx = event->ctx; |
| 2315 | struct perf_event *sibling; |
| 2316 | |
| 2317 | WARN_ON_ONCE(ctx->parent_ctx); |
| 2318 | mutex_lock(&ctx->mutex); |
| 2319 | event = event->group_leader; |
| 2320 | |
| 2321 | perf_event_for_each_child(event, func); |
| 2322 | func(event); |
| 2323 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
| 2324 | perf_event_for_each_child(event, func); |
| 2325 | mutex_unlock(&ctx->mutex); |
| 2326 | } |
| 2327 | |
| 2328 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
| 2329 | { |
| 2330 | struct perf_event_context *ctx = event->ctx; |
| 2331 | unsigned long size; |
| 2332 | int ret = 0; |
| 2333 | u64 value; |
| 2334 | |
| 2335 | if (!event->attr.sample_period) |
| 2336 | return -EINVAL; |
| 2337 | |
| 2338 | size = copy_from_user(&value, arg, sizeof(value)); |
| 2339 | if (size != sizeof(value)) |
| 2340 | return -EFAULT; |
| 2341 | |
| 2342 | if (!value) |
| 2343 | return -EINVAL; |
| 2344 | |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 2345 | raw_spin_lock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2346 | if (event->attr.freq) { |
| 2347 | if (value > sysctl_perf_event_sample_rate) { |
| 2348 | ret = -EINVAL; |
| 2349 | goto unlock; |
| 2350 | } |
| 2351 | |
| 2352 | event->attr.sample_freq = value; |
| 2353 | } else { |
| 2354 | event->attr.sample_period = value; |
| 2355 | event->hw.sample_period = value; |
| 2356 | } |
| 2357 | unlock: |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 2358 | raw_spin_unlock_irq(&ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2359 | |
| 2360 | return ret; |
| 2361 | } |
| 2362 | |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2363 | static const struct file_operations perf_fops; |
| 2364 | |
| 2365 | static struct perf_event *perf_fget_light(int fd, int *fput_needed) |
| 2366 | { |
| 2367 | struct file *file; |
| 2368 | |
| 2369 | file = fget_light(fd, fput_needed); |
| 2370 | if (!file) |
| 2371 | return ERR_PTR(-EBADF); |
| 2372 | |
| 2373 | if (file->f_op != &perf_fops) { |
| 2374 | fput_light(file, *fput_needed); |
| 2375 | *fput_needed = 0; |
| 2376 | return ERR_PTR(-EBADF); |
| 2377 | } |
| 2378 | |
| 2379 | return file->private_data; |
| 2380 | } |
| 2381 | |
| 2382 | static int perf_event_set_output(struct perf_event *event, |
| 2383 | struct perf_event *output_event); |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 2384 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2385 | |
| 2386 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
| 2387 | { |
| 2388 | struct perf_event *event = file->private_data; |
| 2389 | void (*func)(struct perf_event *); |
| 2390 | u32 flags = arg; |
| 2391 | |
| 2392 | switch (cmd) { |
| 2393 | case PERF_EVENT_IOC_ENABLE: |
| 2394 | func = perf_event_enable; |
| 2395 | break; |
| 2396 | case PERF_EVENT_IOC_DISABLE: |
| 2397 | func = perf_event_disable; |
| 2398 | break; |
| 2399 | case PERF_EVENT_IOC_RESET: |
| 2400 | func = perf_event_reset; |
| 2401 | break; |
| 2402 | |
| 2403 | case PERF_EVENT_IOC_REFRESH: |
| 2404 | return perf_event_refresh(event, arg); |
| 2405 | |
| 2406 | case PERF_EVENT_IOC_PERIOD: |
| 2407 | return perf_event_period(event, (u64 __user *)arg); |
| 2408 | |
| 2409 | case PERF_EVENT_IOC_SET_OUTPUT: |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2410 | { |
| 2411 | struct perf_event *output_event = NULL; |
| 2412 | int fput_needed = 0; |
| 2413 | int ret; |
| 2414 | |
| 2415 | if (arg != -1) { |
| 2416 | output_event = perf_fget_light(arg, &fput_needed); |
| 2417 | if (IS_ERR(output_event)) |
| 2418 | return PTR_ERR(output_event); |
| 2419 | } |
| 2420 | |
| 2421 | ret = perf_event_set_output(event, output_event); |
| 2422 | if (output_event) |
| 2423 | fput_light(output_event->filp, fput_needed); |
| 2424 | |
| 2425 | return ret; |
| 2426 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2427 | |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 2428 | case PERF_EVENT_IOC_SET_FILTER: |
| 2429 | return perf_event_set_filter(event, (void __user *)arg); |
| 2430 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2431 | default: |
| 2432 | return -ENOTTY; |
| 2433 | } |
| 2434 | |
| 2435 | if (flags & PERF_IOC_FLAG_GROUP) |
| 2436 | perf_event_for_each(event, func); |
| 2437 | else |
| 2438 | perf_event_for_each_child(event, func); |
| 2439 | |
| 2440 | return 0; |
| 2441 | } |
| 2442 | |
| 2443 | int perf_event_task_enable(void) |
| 2444 | { |
| 2445 | struct perf_event *event; |
| 2446 | |
| 2447 | mutex_lock(¤t->perf_event_mutex); |
| 2448 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) |
| 2449 | perf_event_for_each_child(event, perf_event_enable); |
| 2450 | mutex_unlock(¤t->perf_event_mutex); |
| 2451 | |
| 2452 | return 0; |
| 2453 | } |
| 2454 | |
| 2455 | int perf_event_task_disable(void) |
| 2456 | { |
| 2457 | struct perf_event *event; |
| 2458 | |
| 2459 | mutex_lock(¤t->perf_event_mutex); |
| 2460 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) |
| 2461 | perf_event_for_each_child(event, perf_event_disable); |
| 2462 | mutex_unlock(¤t->perf_event_mutex); |
| 2463 | |
| 2464 | return 0; |
| 2465 | } |
| 2466 | |
| 2467 | #ifndef PERF_EVENT_INDEX_OFFSET |
| 2468 | # define PERF_EVENT_INDEX_OFFSET 0 |
| 2469 | #endif |
| 2470 | |
| 2471 | static int perf_event_index(struct perf_event *event) |
| 2472 | { |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 2473 | if (event->hw.state & PERF_HES_STOPPED) |
| 2474 | return 0; |
| 2475 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2476 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
| 2477 | return 0; |
| 2478 | |
| 2479 | return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET; |
| 2480 | } |
| 2481 | |
| 2482 | /* |
| 2483 | * Callers need to ensure there can be no nesting of this function, otherwise |
| 2484 | * the seqlock logic goes bad. We can not serialize this because the arch |
| 2485 | * code calls this from NMI context. |
| 2486 | */ |
| 2487 | void perf_event_update_userpage(struct perf_event *event) |
| 2488 | { |
| 2489 | struct perf_event_mmap_page *userpg; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2490 | struct perf_buffer *buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2491 | |
| 2492 | rcu_read_lock(); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2493 | buffer = rcu_dereference(event->buffer); |
| 2494 | if (!buffer) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2495 | goto unlock; |
| 2496 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2497 | userpg = buffer->user_page; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2498 | |
| 2499 | /* |
| 2500 | * Disable preemption so as to not let the corresponding user-space |
| 2501 | * spin too long if we get preempted. |
| 2502 | */ |
| 2503 | preempt_disable(); |
| 2504 | ++userpg->lock; |
| 2505 | barrier(); |
| 2506 | userpg->index = perf_event_index(event); |
Peter Zijlstra | b5e5879 | 2010-05-21 14:43:12 +0200 | [diff] [blame] | 2507 | userpg->offset = perf_event_count(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2508 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 2509 | userpg->offset -= local64_read(&event->hw.prev_count); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2510 | |
| 2511 | userpg->time_enabled = event->total_time_enabled + |
| 2512 | atomic64_read(&event->child_total_time_enabled); |
| 2513 | |
| 2514 | userpg->time_running = event->total_time_running + |
| 2515 | atomic64_read(&event->child_total_time_running); |
| 2516 | |
| 2517 | barrier(); |
| 2518 | ++userpg->lock; |
| 2519 | preempt_enable(); |
| 2520 | unlock: |
| 2521 | rcu_read_unlock(); |
| 2522 | } |
| 2523 | |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2524 | static unsigned long perf_data_size(struct perf_buffer *buffer); |
| 2525 | |
| 2526 | static void |
| 2527 | perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags) |
| 2528 | { |
| 2529 | long max_size = perf_data_size(buffer); |
| 2530 | |
| 2531 | if (watermark) |
| 2532 | buffer->watermark = min(max_size, watermark); |
| 2533 | |
| 2534 | if (!buffer->watermark) |
| 2535 | buffer->watermark = max_size / 2; |
| 2536 | |
| 2537 | if (flags & PERF_BUFFER_WRITABLE) |
| 2538 | buffer->writable = 1; |
| 2539 | |
| 2540 | atomic_set(&buffer->refcount, 1); |
| 2541 | } |
| 2542 | |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2543 | #ifndef CONFIG_PERF_USE_VMALLOC |
| 2544 | |
| 2545 | /* |
| 2546 | * Back perf_mmap() with regular GFP_KERNEL-0 pages. |
| 2547 | */ |
| 2548 | |
| 2549 | static struct page * |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2550 | perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2551 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2552 | if (pgoff > buffer->nr_pages) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2553 | return NULL; |
| 2554 | |
| 2555 | if (pgoff == 0) |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2556 | return virt_to_page(buffer->user_page); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2557 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2558 | return virt_to_page(buffer->data_pages[pgoff - 1]); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2559 | } |
| 2560 | |
Peter Zijlstra | a19d35c | 2010-05-17 18:48:00 +0200 | [diff] [blame] | 2561 | static void *perf_mmap_alloc_page(int cpu) |
| 2562 | { |
| 2563 | struct page *page; |
| 2564 | int node; |
| 2565 | |
| 2566 | node = (cpu == -1) ? cpu : cpu_to_node(cpu); |
| 2567 | page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0); |
| 2568 | if (!page) |
| 2569 | return NULL; |
| 2570 | |
| 2571 | return page_address(page); |
| 2572 | } |
| 2573 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2574 | static struct perf_buffer * |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2575 | perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2576 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2577 | struct perf_buffer *buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2578 | unsigned long size; |
| 2579 | int i; |
| 2580 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2581 | size = sizeof(struct perf_buffer); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2582 | size += nr_pages * sizeof(void *); |
| 2583 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2584 | buffer = kzalloc(size, GFP_KERNEL); |
| 2585 | if (!buffer) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2586 | goto fail; |
| 2587 | |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2588 | buffer->user_page = perf_mmap_alloc_page(cpu); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2589 | if (!buffer->user_page) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2590 | goto fail_user_page; |
| 2591 | |
| 2592 | for (i = 0; i < nr_pages; i++) { |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2593 | buffer->data_pages[i] = perf_mmap_alloc_page(cpu); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2594 | if (!buffer->data_pages[i]) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2595 | goto fail_data_pages; |
| 2596 | } |
| 2597 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2598 | buffer->nr_pages = nr_pages; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2599 | |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2600 | perf_buffer_init(buffer, watermark, flags); |
| 2601 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2602 | return buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2603 | |
| 2604 | fail_data_pages: |
| 2605 | for (i--; i >= 0; i--) |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2606 | free_page((unsigned long)buffer->data_pages[i]); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2607 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2608 | free_page((unsigned long)buffer->user_page); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2609 | |
| 2610 | fail_user_page: |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2611 | kfree(buffer); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2612 | |
| 2613 | fail: |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2614 | return NULL; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2615 | } |
| 2616 | |
| 2617 | static void perf_mmap_free_page(unsigned long addr) |
| 2618 | { |
| 2619 | struct page *page = virt_to_page((void *)addr); |
| 2620 | |
| 2621 | page->mapping = NULL; |
| 2622 | __free_page(page); |
| 2623 | } |
| 2624 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2625 | static void perf_buffer_free(struct perf_buffer *buffer) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2626 | { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2627 | int i; |
| 2628 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2629 | perf_mmap_free_page((unsigned long)buffer->user_page); |
| 2630 | for (i = 0; i < buffer->nr_pages; i++) |
| 2631 | perf_mmap_free_page((unsigned long)buffer->data_pages[i]); |
| 2632 | kfree(buffer); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2633 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2634 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2635 | static inline int page_order(struct perf_buffer *buffer) |
Peter Zijlstra | 3cafa9f | 2010-05-20 19:07:56 +0200 | [diff] [blame] | 2636 | { |
| 2637 | return 0; |
| 2638 | } |
| 2639 | |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2640 | #else |
| 2641 | |
| 2642 | /* |
| 2643 | * Back perf_mmap() with vmalloc memory. |
| 2644 | * |
| 2645 | * Required for architectures that have d-cache aliasing issues. |
| 2646 | */ |
| 2647 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2648 | static inline int page_order(struct perf_buffer *buffer) |
Peter Zijlstra | 3cafa9f | 2010-05-20 19:07:56 +0200 | [diff] [blame] | 2649 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2650 | return buffer->page_order; |
Peter Zijlstra | 3cafa9f | 2010-05-20 19:07:56 +0200 | [diff] [blame] | 2651 | } |
| 2652 | |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2653 | static struct page * |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2654 | perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2655 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2656 | if (pgoff > (1UL << page_order(buffer))) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2657 | return NULL; |
| 2658 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2659 | return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2660 | } |
| 2661 | |
| 2662 | static void perf_mmap_unmark_page(void *addr) |
| 2663 | { |
| 2664 | struct page *page = vmalloc_to_page(addr); |
| 2665 | |
| 2666 | page->mapping = NULL; |
| 2667 | } |
| 2668 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2669 | static void perf_buffer_free_work(struct work_struct *work) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2670 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2671 | struct perf_buffer *buffer; |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2672 | void *base; |
| 2673 | int i, nr; |
| 2674 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2675 | buffer = container_of(work, struct perf_buffer, work); |
| 2676 | nr = 1 << page_order(buffer); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2677 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2678 | base = buffer->user_page; |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2679 | for (i = 0; i < nr + 1; i++) |
| 2680 | perf_mmap_unmark_page(base + (i * PAGE_SIZE)); |
| 2681 | |
| 2682 | vfree(base); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2683 | kfree(buffer); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2684 | } |
| 2685 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2686 | static void perf_buffer_free(struct perf_buffer *buffer) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2687 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2688 | schedule_work(&buffer->work); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2689 | } |
| 2690 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2691 | static struct perf_buffer * |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2692 | perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2693 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2694 | struct perf_buffer *buffer; |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2695 | unsigned long size; |
| 2696 | void *all_buf; |
| 2697 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2698 | size = sizeof(struct perf_buffer); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2699 | size += sizeof(void *); |
| 2700 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2701 | buffer = kzalloc(size, GFP_KERNEL); |
| 2702 | if (!buffer) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2703 | goto fail; |
| 2704 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2705 | INIT_WORK(&buffer->work, perf_buffer_free_work); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2706 | |
| 2707 | all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE); |
| 2708 | if (!all_buf) |
| 2709 | goto fail_all_buf; |
| 2710 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2711 | buffer->user_page = all_buf; |
| 2712 | buffer->data_pages[0] = all_buf + PAGE_SIZE; |
| 2713 | buffer->page_order = ilog2(nr_pages); |
| 2714 | buffer->nr_pages = 1; |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2715 | |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2716 | perf_buffer_init(buffer, watermark, flags); |
| 2717 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2718 | return buffer; |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2719 | |
| 2720 | fail_all_buf: |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2721 | kfree(buffer); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2722 | |
| 2723 | fail: |
| 2724 | return NULL; |
| 2725 | } |
| 2726 | |
| 2727 | #endif |
| 2728 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2729 | static unsigned long perf_data_size(struct perf_buffer *buffer) |
Peter Zijlstra | 3cafa9f | 2010-05-20 19:07:56 +0200 | [diff] [blame] | 2730 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2731 | return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer)); |
Peter Zijlstra | 3cafa9f | 2010-05-20 19:07:56 +0200 | [diff] [blame] | 2732 | } |
| 2733 | |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2734 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
| 2735 | { |
| 2736 | struct perf_event *event = vma->vm_file->private_data; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2737 | struct perf_buffer *buffer; |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2738 | int ret = VM_FAULT_SIGBUS; |
| 2739 | |
| 2740 | if (vmf->flags & FAULT_FLAG_MKWRITE) { |
| 2741 | if (vmf->pgoff == 0) |
| 2742 | ret = 0; |
| 2743 | return ret; |
| 2744 | } |
| 2745 | |
| 2746 | rcu_read_lock(); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2747 | buffer = rcu_dereference(event->buffer); |
| 2748 | if (!buffer) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2749 | goto unlock; |
| 2750 | |
| 2751 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) |
| 2752 | goto unlock; |
| 2753 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2754 | vmf->page = perf_mmap_to_page(buffer, vmf->pgoff); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2755 | if (!vmf->page) |
| 2756 | goto unlock; |
| 2757 | |
| 2758 | get_page(vmf->page); |
| 2759 | vmf->page->mapping = vma->vm_file->f_mapping; |
| 2760 | vmf->page->index = vmf->pgoff; |
| 2761 | |
| 2762 | ret = 0; |
| 2763 | unlock: |
| 2764 | rcu_read_unlock(); |
| 2765 | |
| 2766 | return ret; |
| 2767 | } |
| 2768 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2769 | static void perf_buffer_free_rcu(struct rcu_head *rcu_head) |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2770 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2771 | struct perf_buffer *buffer; |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2772 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2773 | buffer = container_of(rcu_head, struct perf_buffer, rcu_head); |
| 2774 | perf_buffer_free(buffer); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2775 | } |
| 2776 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2777 | static struct perf_buffer *perf_buffer_get(struct perf_event *event) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2778 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2779 | struct perf_buffer *buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2780 | |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2781 | rcu_read_lock(); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2782 | buffer = rcu_dereference(event->buffer); |
| 2783 | if (buffer) { |
| 2784 | if (!atomic_inc_not_zero(&buffer->refcount)) |
| 2785 | buffer = NULL; |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2786 | } |
| 2787 | rcu_read_unlock(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2788 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2789 | return buffer; |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2790 | } |
| 2791 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2792 | static void perf_buffer_put(struct perf_buffer *buffer) |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2793 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2794 | if (!atomic_dec_and_test(&buffer->refcount)) |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2795 | return; |
| 2796 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2797 | call_rcu(&buffer->rcu_head, perf_buffer_free_rcu); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2798 | } |
| 2799 | |
| 2800 | static void perf_mmap_open(struct vm_area_struct *vma) |
| 2801 | { |
| 2802 | struct perf_event *event = vma->vm_file->private_data; |
| 2803 | |
| 2804 | atomic_inc(&event->mmap_count); |
| 2805 | } |
| 2806 | |
| 2807 | static void perf_mmap_close(struct vm_area_struct *vma) |
| 2808 | { |
| 2809 | struct perf_event *event = vma->vm_file->private_data; |
| 2810 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2811 | if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2812 | unsigned long size = perf_data_size(event->buffer); |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2813 | struct user_struct *user = event->mmap_user; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2814 | struct perf_buffer *buffer = event->buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2815 | |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2816 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm); |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2817 | vma->vm_mm->locked_vm -= event->mmap_locked; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2818 | rcu_assign_pointer(event->buffer, NULL); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2819 | mutex_unlock(&event->mmap_mutex); |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2820 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2821 | perf_buffer_put(buffer); |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2822 | free_uid(user); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2823 | } |
| 2824 | } |
| 2825 | |
Alexey Dobriyan | f0f37e2 | 2009-09-27 22:29:37 +0400 | [diff] [blame] | 2826 | static const struct vm_operations_struct perf_mmap_vmops = { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2827 | .open = perf_mmap_open, |
| 2828 | .close = perf_mmap_close, |
| 2829 | .fault = perf_mmap_fault, |
| 2830 | .page_mkwrite = perf_mmap_fault, |
| 2831 | }; |
| 2832 | |
| 2833 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) |
| 2834 | { |
| 2835 | struct perf_event *event = file->private_data; |
| 2836 | unsigned long user_locked, user_lock_limit; |
| 2837 | struct user_struct *user = current_user(); |
| 2838 | unsigned long locked, lock_limit; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2839 | struct perf_buffer *buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2840 | unsigned long vma_size; |
| 2841 | unsigned long nr_pages; |
| 2842 | long user_extra, extra; |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2843 | int ret = 0, flags = 0; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2844 | |
Peter Zijlstra | c792061 | 2010-05-18 10:33:24 +0200 | [diff] [blame] | 2845 | /* |
| 2846 | * Don't allow mmap() of inherited per-task counters. This would |
| 2847 | * create a performance issue due to all children writing to the |
| 2848 | * same buffer. |
| 2849 | */ |
| 2850 | if (event->cpu == -1 && event->attr.inherit) |
| 2851 | return -EINVAL; |
| 2852 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2853 | if (!(vma->vm_flags & VM_SHARED)) |
| 2854 | return -EINVAL; |
| 2855 | |
| 2856 | vma_size = vma->vm_end - vma->vm_start; |
| 2857 | nr_pages = (vma_size / PAGE_SIZE) - 1; |
| 2858 | |
| 2859 | /* |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2860 | * If we have buffer pages ensure they're a power-of-two number, so we |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2861 | * can do bitmasks instead of modulo. |
| 2862 | */ |
| 2863 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) |
| 2864 | return -EINVAL; |
| 2865 | |
| 2866 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
| 2867 | return -EINVAL; |
| 2868 | |
| 2869 | if (vma->vm_pgoff != 0) |
| 2870 | return -EINVAL; |
| 2871 | |
| 2872 | WARN_ON_ONCE(event->ctx->parent_ctx); |
| 2873 | mutex_lock(&event->mmap_mutex); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2874 | if (event->buffer) { |
| 2875 | if (event->buffer->nr_pages == nr_pages) |
| 2876 | atomic_inc(&event->buffer->refcount); |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2877 | else |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2878 | ret = -EINVAL; |
| 2879 | goto unlock; |
| 2880 | } |
| 2881 | |
| 2882 | user_extra = nr_pages + 1; |
| 2883 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
| 2884 | |
| 2885 | /* |
| 2886 | * Increase the limit linearly with more CPUs: |
| 2887 | */ |
| 2888 | user_lock_limit *= num_online_cpus(); |
| 2889 | |
| 2890 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
| 2891 | |
| 2892 | extra = 0; |
| 2893 | if (user_locked > user_lock_limit) |
| 2894 | extra = user_locked - user_lock_limit; |
| 2895 | |
Jiri Slaby | 78d7d40 | 2010-03-05 13:42:54 -0800 | [diff] [blame] | 2896 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2897 | lock_limit >>= PAGE_SHIFT; |
| 2898 | locked = vma->vm_mm->locked_vm + extra; |
| 2899 | |
| 2900 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
| 2901 | !capable(CAP_IPC_LOCK)) { |
| 2902 | ret = -EPERM; |
| 2903 | goto unlock; |
| 2904 | } |
| 2905 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2906 | WARN_ON(event->buffer); |
Peter Zijlstra | 906010b | 2009-09-21 16:08:49 +0200 | [diff] [blame] | 2907 | |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2908 | if (vma->vm_flags & VM_WRITE) |
| 2909 | flags |= PERF_BUFFER_WRITABLE; |
| 2910 | |
| 2911 | buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark, |
| 2912 | event->cpu, flags); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 2913 | if (!buffer) { |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2914 | ret = -ENOMEM; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2915 | goto unlock; |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2916 | } |
Peter Zijlstra | d57e34f | 2010-05-28 19:41:35 +0200 | [diff] [blame] | 2917 | rcu_assign_pointer(event->buffer, buffer); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2918 | |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2919 | atomic_long_add(user_extra, &user->locked_vm); |
| 2920 | event->mmap_locked = extra; |
| 2921 | event->mmap_user = get_current_user(); |
| 2922 | vma->vm_mm->locked_vm += event->mmap_locked; |
| 2923 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2924 | unlock: |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 2925 | if (!ret) |
| 2926 | atomic_inc(&event->mmap_count); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2927 | mutex_unlock(&event->mmap_mutex); |
| 2928 | |
| 2929 | vma->vm_flags |= VM_RESERVED; |
| 2930 | vma->vm_ops = &perf_mmap_vmops; |
| 2931 | |
| 2932 | return ret; |
| 2933 | } |
| 2934 | |
| 2935 | static int perf_fasync(int fd, struct file *filp, int on) |
| 2936 | { |
| 2937 | struct inode *inode = filp->f_path.dentry->d_inode; |
| 2938 | struct perf_event *event = filp->private_data; |
| 2939 | int retval; |
| 2940 | |
| 2941 | mutex_lock(&inode->i_mutex); |
| 2942 | retval = fasync_helper(fd, filp, on, &event->fasync); |
| 2943 | mutex_unlock(&inode->i_mutex); |
| 2944 | |
| 2945 | if (retval < 0) |
| 2946 | return retval; |
| 2947 | |
| 2948 | return 0; |
| 2949 | } |
| 2950 | |
| 2951 | static const struct file_operations perf_fops = { |
Arnd Bergmann | 3326c1c | 2010-03-23 19:09:33 +0100 | [diff] [blame] | 2952 | .llseek = no_llseek, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 2953 | .release = perf_release, |
| 2954 | .read = perf_read, |
| 2955 | .poll = perf_poll, |
| 2956 | .unlocked_ioctl = perf_ioctl, |
| 2957 | .compat_ioctl = perf_ioctl, |
| 2958 | .mmap = perf_mmap, |
| 2959 | .fasync = perf_fasync, |
| 2960 | }; |
| 2961 | |
| 2962 | /* |
| 2963 | * Perf event wakeup |
| 2964 | * |
| 2965 | * If there's data, ensure we set the poll() state and publish everything |
| 2966 | * to user-space before waking everybody up. |
| 2967 | */ |
| 2968 | |
| 2969 | void perf_event_wakeup(struct perf_event *event) |
| 2970 | { |
| 2971 | wake_up_all(&event->waitq); |
| 2972 | |
| 2973 | if (event->pending_kill) { |
| 2974 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); |
| 2975 | event->pending_kill = 0; |
| 2976 | } |
| 2977 | } |
| 2978 | |
| 2979 | /* |
| 2980 | * Pending wakeups |
| 2981 | * |
| 2982 | * Handle the case where we need to wakeup up from NMI (or rq->lock) context. |
| 2983 | * |
| 2984 | * The NMI bit means we cannot possibly take locks. Therefore, maintain a |
| 2985 | * single linked list and use cmpxchg() to add entries lockless. |
| 2986 | */ |
| 2987 | |
| 2988 | static void perf_pending_event(struct perf_pending_entry *entry) |
| 2989 | { |
| 2990 | struct perf_event *event = container_of(entry, |
| 2991 | struct perf_event, pending); |
| 2992 | |
| 2993 | if (event->pending_disable) { |
| 2994 | event->pending_disable = 0; |
| 2995 | __perf_event_disable(event); |
| 2996 | } |
| 2997 | |
| 2998 | if (event->pending_wakeup) { |
| 2999 | event->pending_wakeup = 0; |
| 3000 | perf_event_wakeup(event); |
| 3001 | } |
| 3002 | } |
| 3003 | |
| 3004 | #define PENDING_TAIL ((struct perf_pending_entry *)-1UL) |
| 3005 | |
| 3006 | static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = { |
| 3007 | PENDING_TAIL, |
| 3008 | }; |
| 3009 | |
| 3010 | static void perf_pending_queue(struct perf_pending_entry *entry, |
| 3011 | void (*func)(struct perf_pending_entry *)) |
| 3012 | { |
| 3013 | struct perf_pending_entry **head; |
| 3014 | |
| 3015 | if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL) |
| 3016 | return; |
| 3017 | |
| 3018 | entry->func = func; |
| 3019 | |
| 3020 | head = &get_cpu_var(perf_pending_head); |
| 3021 | |
| 3022 | do { |
| 3023 | entry->next = *head; |
| 3024 | } while (cmpxchg(head, entry->next, entry) != entry->next); |
| 3025 | |
| 3026 | set_perf_event_pending(); |
| 3027 | |
| 3028 | put_cpu_var(perf_pending_head); |
| 3029 | } |
| 3030 | |
| 3031 | static int __perf_pending_run(void) |
| 3032 | { |
| 3033 | struct perf_pending_entry *list; |
| 3034 | int nr = 0; |
| 3035 | |
| 3036 | list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL); |
| 3037 | while (list != PENDING_TAIL) { |
| 3038 | void (*func)(struct perf_pending_entry *); |
| 3039 | struct perf_pending_entry *entry = list; |
| 3040 | |
| 3041 | list = list->next; |
| 3042 | |
| 3043 | func = entry->func; |
| 3044 | entry->next = NULL; |
| 3045 | /* |
| 3046 | * Ensure we observe the unqueue before we issue the wakeup, |
| 3047 | * so that we won't be waiting forever. |
| 3048 | * -- see perf_not_pending(). |
| 3049 | */ |
| 3050 | smp_wmb(); |
| 3051 | |
| 3052 | func(entry); |
| 3053 | nr++; |
| 3054 | } |
| 3055 | |
| 3056 | return nr; |
| 3057 | } |
| 3058 | |
| 3059 | static inline int perf_not_pending(struct perf_event *event) |
| 3060 | { |
| 3061 | /* |
| 3062 | * If we flush on whatever cpu we run, there is a chance we don't |
| 3063 | * need to wait. |
| 3064 | */ |
| 3065 | get_cpu(); |
| 3066 | __perf_pending_run(); |
| 3067 | put_cpu(); |
| 3068 | |
| 3069 | /* |
| 3070 | * Ensure we see the proper queue state before going to sleep |
| 3071 | * so that we do not miss the wakeup. -- see perf_pending_handle() |
| 3072 | */ |
| 3073 | smp_rmb(); |
| 3074 | return event->pending.next == NULL; |
| 3075 | } |
| 3076 | |
| 3077 | static void perf_pending_sync(struct perf_event *event) |
| 3078 | { |
| 3079 | wait_event(event->waitq, perf_not_pending(event)); |
| 3080 | } |
| 3081 | |
| 3082 | void perf_event_do_pending(void) |
| 3083 | { |
| 3084 | __perf_pending_run(); |
| 3085 | } |
| 3086 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3087 | /* |
Zhang, Yanmin | 39447b3 | 2010-04-19 13:32:41 +0800 | [diff] [blame] | 3088 | * We assume there is only KVM supporting the callbacks. |
| 3089 | * Later on, we might change it to a list if there is |
| 3090 | * another virtualization implementation supporting the callbacks. |
| 3091 | */ |
| 3092 | struct perf_guest_info_callbacks *perf_guest_cbs; |
| 3093 | |
| 3094 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) |
| 3095 | { |
| 3096 | perf_guest_cbs = cbs; |
| 3097 | return 0; |
| 3098 | } |
| 3099 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); |
| 3100 | |
| 3101 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) |
| 3102 | { |
| 3103 | perf_guest_cbs = NULL; |
| 3104 | return 0; |
| 3105 | } |
| 3106 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); |
| 3107 | |
| 3108 | /* |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3109 | * Output |
| 3110 | */ |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3111 | static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3112 | unsigned long offset, unsigned long head) |
| 3113 | { |
| 3114 | unsigned long mask; |
| 3115 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3116 | if (!buffer->writable) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3117 | return true; |
| 3118 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3119 | mask = perf_data_size(buffer) - 1; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3120 | |
| 3121 | offset = (offset - tail) & mask; |
| 3122 | head = (head - tail) & mask; |
| 3123 | |
| 3124 | if ((int)(head - offset) < 0) |
| 3125 | return false; |
| 3126 | |
| 3127 | return true; |
| 3128 | } |
| 3129 | |
| 3130 | static void perf_output_wakeup(struct perf_output_handle *handle) |
| 3131 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3132 | atomic_set(&handle->buffer->poll, POLL_IN); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3133 | |
| 3134 | if (handle->nmi) { |
| 3135 | handle->event->pending_wakeup = 1; |
| 3136 | perf_pending_queue(&handle->event->pending, |
| 3137 | perf_pending_event); |
| 3138 | } else |
| 3139 | perf_event_wakeup(handle->event); |
| 3140 | } |
| 3141 | |
| 3142 | /* |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3143 | * We need to ensure a later event_id doesn't publish a head when a former |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3144 | * event isn't done writing. However since we need to deal with NMIs we |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3145 | * cannot fully serialize things. |
| 3146 | * |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3147 | * We only publish the head (and generate a wakeup) when the outer-most |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3148 | * event completes. |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3149 | */ |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3150 | static void perf_output_get_handle(struct perf_output_handle *handle) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3151 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3152 | struct perf_buffer *buffer = handle->buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3153 | |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3154 | preempt_disable(); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3155 | local_inc(&buffer->nest); |
| 3156 | handle->wakeup = local_read(&buffer->wakeup); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3157 | } |
| 3158 | |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3159 | static void perf_output_put_handle(struct perf_output_handle *handle) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3160 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3161 | struct perf_buffer *buffer = handle->buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3162 | unsigned long head; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3163 | |
| 3164 | again: |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3165 | head = local_read(&buffer->head); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3166 | |
| 3167 | /* |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3168 | * IRQ/NMI can happen here, which means we can miss a head update. |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3169 | */ |
| 3170 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3171 | if (!local_dec_and_test(&buffer->nest)) |
Frederic Weisbecker | acd35a4 | 2010-05-20 21:28:34 +0200 | [diff] [blame] | 3172 | goto out; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3173 | |
| 3174 | /* |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3175 | * Publish the known good head. Rely on the full barrier implied |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3176 | * by atomic_dec_and_test() order the buffer->head read and this |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3177 | * write. |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3178 | */ |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3179 | buffer->user_page->data_head = head; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3180 | |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3181 | /* |
| 3182 | * Now check if we missed an update, rely on the (compiler) |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3183 | * barrier in atomic_dec_and_test() to re-read buffer->head. |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3184 | */ |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3185 | if (unlikely(head != local_read(&buffer->head))) { |
| 3186 | local_inc(&buffer->nest); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3187 | goto again; |
| 3188 | } |
| 3189 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3190 | if (handle->wakeup != local_read(&buffer->wakeup)) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3191 | perf_output_wakeup(handle); |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3192 | |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 3193 | out: |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3194 | preempt_enable(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3195 | } |
| 3196 | |
Peter Zijlstra | a94ffaa | 2010-05-20 19:50:07 +0200 | [diff] [blame] | 3197 | __always_inline void perf_output_copy(struct perf_output_handle *handle, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3198 | const void *buf, unsigned int len) |
| 3199 | { |
Peter Zijlstra | 5d967a8 | 2010-05-20 16:46:39 +0200 | [diff] [blame] | 3200 | do { |
Peter Zijlstra | a94ffaa | 2010-05-20 19:50:07 +0200 | [diff] [blame] | 3201 | unsigned long size = min_t(unsigned long, handle->size, len); |
Peter Zijlstra | 5d967a8 | 2010-05-20 16:46:39 +0200 | [diff] [blame] | 3202 | |
| 3203 | memcpy(handle->addr, buf, size); |
| 3204 | |
| 3205 | len -= size; |
| 3206 | handle->addr += size; |
Frederic Weisbecker | 74048f8 | 2010-05-27 21:34:58 +0200 | [diff] [blame] | 3207 | buf += size; |
Peter Zijlstra | 5d967a8 | 2010-05-20 16:46:39 +0200 | [diff] [blame] | 3208 | handle->size -= size; |
| 3209 | if (!handle->size) { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3210 | struct perf_buffer *buffer = handle->buffer; |
Peter Zijlstra | 3cafa9f | 2010-05-20 19:07:56 +0200 | [diff] [blame] | 3211 | |
Peter Zijlstra | 5d967a8 | 2010-05-20 16:46:39 +0200 | [diff] [blame] | 3212 | handle->page++; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3213 | handle->page &= buffer->nr_pages - 1; |
| 3214 | handle->addr = buffer->data_pages[handle->page]; |
| 3215 | handle->size = PAGE_SIZE << page_order(buffer); |
Peter Zijlstra | 5d967a8 | 2010-05-20 16:46:39 +0200 | [diff] [blame] | 3216 | } |
| 3217 | } while (len); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3218 | } |
| 3219 | |
| 3220 | int perf_output_begin(struct perf_output_handle *handle, |
| 3221 | struct perf_event *event, unsigned int size, |
| 3222 | int nmi, int sample) |
| 3223 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3224 | struct perf_buffer *buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3225 | unsigned long tail, offset, head; |
| 3226 | int have_lost; |
| 3227 | struct { |
| 3228 | struct perf_event_header header; |
| 3229 | u64 id; |
| 3230 | u64 lost; |
| 3231 | } lost_event; |
| 3232 | |
| 3233 | rcu_read_lock(); |
| 3234 | /* |
| 3235 | * For inherited events we send all the output towards the parent. |
| 3236 | */ |
| 3237 | if (event->parent) |
| 3238 | event = event->parent; |
| 3239 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3240 | buffer = rcu_dereference(event->buffer); |
| 3241 | if (!buffer) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3242 | goto out; |
| 3243 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3244 | handle->buffer = buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3245 | handle->event = event; |
| 3246 | handle->nmi = nmi; |
| 3247 | handle->sample = sample; |
| 3248 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3249 | if (!buffer->nr_pages) |
Stephane Eranian | 00d1d0b | 2010-05-17 12:46:01 +0200 | [diff] [blame] | 3250 | goto out; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3251 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3252 | have_lost = local_read(&buffer->lost); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3253 | if (have_lost) |
| 3254 | size += sizeof(lost_event); |
| 3255 | |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3256 | perf_output_get_handle(handle); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3257 | |
| 3258 | do { |
| 3259 | /* |
| 3260 | * Userspace could choose to issue a mb() before updating the |
| 3261 | * tail pointer. So that all reads will be completed before the |
| 3262 | * write is issued. |
| 3263 | */ |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3264 | tail = ACCESS_ONCE(buffer->user_page->data_tail); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3265 | smp_rmb(); |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3266 | offset = head = local_read(&buffer->head); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3267 | head += size; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3268 | if (unlikely(!perf_output_space(buffer, tail, offset, head))) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3269 | goto fail; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3270 | } while (local_cmpxchg(&buffer->head, offset, head) != offset); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3271 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3272 | if (head - local_read(&buffer->wakeup) > buffer->watermark) |
| 3273 | local_add(buffer->watermark, &buffer->wakeup); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3274 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3275 | handle->page = offset >> (PAGE_SHIFT + page_order(buffer)); |
| 3276 | handle->page &= buffer->nr_pages - 1; |
| 3277 | handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1); |
| 3278 | handle->addr = buffer->data_pages[handle->page]; |
Peter Zijlstra | 5d967a8 | 2010-05-20 16:46:39 +0200 | [diff] [blame] | 3279 | handle->addr += handle->size; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3280 | handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size; |
Peter Zijlstra | 5d967a8 | 2010-05-20 16:46:39 +0200 | [diff] [blame] | 3281 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3282 | if (have_lost) { |
| 3283 | lost_event.header.type = PERF_RECORD_LOST; |
| 3284 | lost_event.header.misc = 0; |
| 3285 | lost_event.header.size = sizeof(lost_event); |
| 3286 | lost_event.id = event->id; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3287 | lost_event.lost = local_xchg(&buffer->lost, 0); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3288 | |
| 3289 | perf_output_put(handle, lost_event); |
| 3290 | } |
| 3291 | |
| 3292 | return 0; |
| 3293 | |
| 3294 | fail: |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3295 | local_inc(&buffer->lost); |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3296 | perf_output_put_handle(handle); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3297 | out: |
| 3298 | rcu_read_unlock(); |
| 3299 | |
| 3300 | return -ENOSPC; |
| 3301 | } |
| 3302 | |
| 3303 | void perf_output_end(struct perf_output_handle *handle) |
| 3304 | { |
| 3305 | struct perf_event *event = handle->event; |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3306 | struct perf_buffer *buffer = handle->buffer; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3307 | |
| 3308 | int wakeup_events = event->attr.wakeup_events; |
| 3309 | |
| 3310 | if (handle->sample && wakeup_events) { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3311 | int events = local_inc_return(&buffer->events); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3312 | if (events >= wakeup_events) { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 3313 | local_sub(wakeup_events, &buffer->events); |
| 3314 | local_inc(&buffer->wakeup); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3315 | } |
| 3316 | } |
| 3317 | |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3318 | perf_output_put_handle(handle); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3319 | rcu_read_unlock(); |
| 3320 | } |
| 3321 | |
| 3322 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
| 3323 | { |
| 3324 | /* |
| 3325 | * only top level events have the pid namespace they were created in |
| 3326 | */ |
| 3327 | if (event->parent) |
| 3328 | event = event->parent; |
| 3329 | |
| 3330 | return task_tgid_nr_ns(p, event->ns); |
| 3331 | } |
| 3332 | |
| 3333 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) |
| 3334 | { |
| 3335 | /* |
| 3336 | * only top level events have the pid namespace they were created in |
| 3337 | */ |
| 3338 | if (event->parent) |
| 3339 | event = event->parent; |
| 3340 | |
| 3341 | return task_pid_nr_ns(p, event->ns); |
| 3342 | } |
| 3343 | |
| 3344 | static void perf_output_read_one(struct perf_output_handle *handle, |
| 3345 | struct perf_event *event) |
| 3346 | { |
| 3347 | u64 read_format = event->attr.read_format; |
| 3348 | u64 values[4]; |
| 3349 | int n = 0; |
| 3350 | |
Peter Zijlstra | b5e5879 | 2010-05-21 14:43:12 +0200 | [diff] [blame] | 3351 | values[n++] = perf_event_count(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3352 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
| 3353 | values[n++] = event->total_time_enabled + |
| 3354 | atomic64_read(&event->child_total_time_enabled); |
| 3355 | } |
| 3356 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { |
| 3357 | values[n++] = event->total_time_running + |
| 3358 | atomic64_read(&event->child_total_time_running); |
| 3359 | } |
| 3360 | if (read_format & PERF_FORMAT_ID) |
| 3361 | values[n++] = primary_event_id(event); |
| 3362 | |
| 3363 | perf_output_copy(handle, values, n * sizeof(u64)); |
| 3364 | } |
| 3365 | |
| 3366 | /* |
| 3367 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
| 3368 | */ |
| 3369 | static void perf_output_read_group(struct perf_output_handle *handle, |
| 3370 | struct perf_event *event) |
| 3371 | { |
| 3372 | struct perf_event *leader = event->group_leader, *sub; |
| 3373 | u64 read_format = event->attr.read_format; |
| 3374 | u64 values[5]; |
| 3375 | int n = 0; |
| 3376 | |
| 3377 | values[n++] = 1 + leader->nr_siblings; |
| 3378 | |
| 3379 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
| 3380 | values[n++] = leader->total_time_enabled; |
| 3381 | |
| 3382 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) |
| 3383 | values[n++] = leader->total_time_running; |
| 3384 | |
| 3385 | if (leader != event) |
| 3386 | leader->pmu->read(leader); |
| 3387 | |
Peter Zijlstra | b5e5879 | 2010-05-21 14:43:12 +0200 | [diff] [blame] | 3388 | values[n++] = perf_event_count(leader); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3389 | if (read_format & PERF_FORMAT_ID) |
| 3390 | values[n++] = primary_event_id(leader); |
| 3391 | |
| 3392 | perf_output_copy(handle, values, n * sizeof(u64)); |
| 3393 | |
| 3394 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
| 3395 | n = 0; |
| 3396 | |
| 3397 | if (sub != event) |
| 3398 | sub->pmu->read(sub); |
| 3399 | |
Peter Zijlstra | b5e5879 | 2010-05-21 14:43:12 +0200 | [diff] [blame] | 3400 | values[n++] = perf_event_count(sub); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3401 | if (read_format & PERF_FORMAT_ID) |
| 3402 | values[n++] = primary_event_id(sub); |
| 3403 | |
| 3404 | perf_output_copy(handle, values, n * sizeof(u64)); |
| 3405 | } |
| 3406 | } |
| 3407 | |
| 3408 | static void perf_output_read(struct perf_output_handle *handle, |
| 3409 | struct perf_event *event) |
| 3410 | { |
| 3411 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
| 3412 | perf_output_read_group(handle, event); |
| 3413 | else |
| 3414 | perf_output_read_one(handle, event); |
| 3415 | } |
| 3416 | |
| 3417 | void perf_output_sample(struct perf_output_handle *handle, |
| 3418 | struct perf_event_header *header, |
| 3419 | struct perf_sample_data *data, |
| 3420 | struct perf_event *event) |
| 3421 | { |
| 3422 | u64 sample_type = data->type; |
| 3423 | |
| 3424 | perf_output_put(handle, *header); |
| 3425 | |
| 3426 | if (sample_type & PERF_SAMPLE_IP) |
| 3427 | perf_output_put(handle, data->ip); |
| 3428 | |
| 3429 | if (sample_type & PERF_SAMPLE_TID) |
| 3430 | perf_output_put(handle, data->tid_entry); |
| 3431 | |
| 3432 | if (sample_type & PERF_SAMPLE_TIME) |
| 3433 | perf_output_put(handle, data->time); |
| 3434 | |
| 3435 | if (sample_type & PERF_SAMPLE_ADDR) |
| 3436 | perf_output_put(handle, data->addr); |
| 3437 | |
| 3438 | if (sample_type & PERF_SAMPLE_ID) |
| 3439 | perf_output_put(handle, data->id); |
| 3440 | |
| 3441 | if (sample_type & PERF_SAMPLE_STREAM_ID) |
| 3442 | perf_output_put(handle, data->stream_id); |
| 3443 | |
| 3444 | if (sample_type & PERF_SAMPLE_CPU) |
| 3445 | perf_output_put(handle, data->cpu_entry); |
| 3446 | |
| 3447 | if (sample_type & PERF_SAMPLE_PERIOD) |
| 3448 | perf_output_put(handle, data->period); |
| 3449 | |
| 3450 | if (sample_type & PERF_SAMPLE_READ) |
| 3451 | perf_output_read(handle, event); |
| 3452 | |
| 3453 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
| 3454 | if (data->callchain) { |
| 3455 | int size = 1; |
| 3456 | |
| 3457 | if (data->callchain) |
| 3458 | size += data->callchain->nr; |
| 3459 | |
| 3460 | size *= sizeof(u64); |
| 3461 | |
| 3462 | perf_output_copy(handle, data->callchain, size); |
| 3463 | } else { |
| 3464 | u64 nr = 0; |
| 3465 | perf_output_put(handle, nr); |
| 3466 | } |
| 3467 | } |
| 3468 | |
| 3469 | if (sample_type & PERF_SAMPLE_RAW) { |
| 3470 | if (data->raw) { |
| 3471 | perf_output_put(handle, data->raw->size); |
| 3472 | perf_output_copy(handle, data->raw->data, |
| 3473 | data->raw->size); |
| 3474 | } else { |
| 3475 | struct { |
| 3476 | u32 size; |
| 3477 | u32 data; |
| 3478 | } raw = { |
| 3479 | .size = sizeof(u32), |
| 3480 | .data = 0, |
| 3481 | }; |
| 3482 | perf_output_put(handle, raw); |
| 3483 | } |
| 3484 | } |
| 3485 | } |
| 3486 | |
| 3487 | void perf_prepare_sample(struct perf_event_header *header, |
| 3488 | struct perf_sample_data *data, |
| 3489 | struct perf_event *event, |
| 3490 | struct pt_regs *regs) |
| 3491 | { |
| 3492 | u64 sample_type = event->attr.sample_type; |
| 3493 | |
| 3494 | data->type = sample_type; |
| 3495 | |
| 3496 | header->type = PERF_RECORD_SAMPLE; |
| 3497 | header->size = sizeof(*header); |
| 3498 | |
| 3499 | header->misc = 0; |
| 3500 | header->misc |= perf_misc_flags(regs); |
| 3501 | |
| 3502 | if (sample_type & PERF_SAMPLE_IP) { |
| 3503 | data->ip = perf_instruction_pointer(regs); |
| 3504 | |
| 3505 | header->size += sizeof(data->ip); |
| 3506 | } |
| 3507 | |
| 3508 | if (sample_type & PERF_SAMPLE_TID) { |
| 3509 | /* namespace issues */ |
| 3510 | data->tid_entry.pid = perf_event_pid(event, current); |
| 3511 | data->tid_entry.tid = perf_event_tid(event, current); |
| 3512 | |
| 3513 | header->size += sizeof(data->tid_entry); |
| 3514 | } |
| 3515 | |
| 3516 | if (sample_type & PERF_SAMPLE_TIME) { |
| 3517 | data->time = perf_clock(); |
| 3518 | |
| 3519 | header->size += sizeof(data->time); |
| 3520 | } |
| 3521 | |
| 3522 | if (sample_type & PERF_SAMPLE_ADDR) |
| 3523 | header->size += sizeof(data->addr); |
| 3524 | |
| 3525 | if (sample_type & PERF_SAMPLE_ID) { |
| 3526 | data->id = primary_event_id(event); |
| 3527 | |
| 3528 | header->size += sizeof(data->id); |
| 3529 | } |
| 3530 | |
| 3531 | if (sample_type & PERF_SAMPLE_STREAM_ID) { |
| 3532 | data->stream_id = event->id; |
| 3533 | |
| 3534 | header->size += sizeof(data->stream_id); |
| 3535 | } |
| 3536 | |
| 3537 | if (sample_type & PERF_SAMPLE_CPU) { |
| 3538 | data->cpu_entry.cpu = raw_smp_processor_id(); |
| 3539 | data->cpu_entry.reserved = 0; |
| 3540 | |
| 3541 | header->size += sizeof(data->cpu_entry); |
| 3542 | } |
| 3543 | |
| 3544 | if (sample_type & PERF_SAMPLE_PERIOD) |
| 3545 | header->size += sizeof(data->period); |
| 3546 | |
| 3547 | if (sample_type & PERF_SAMPLE_READ) |
| 3548 | header->size += perf_event_read_size(event); |
| 3549 | |
| 3550 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
| 3551 | int size = 1; |
| 3552 | |
| 3553 | data->callchain = perf_callchain(regs); |
| 3554 | |
| 3555 | if (data->callchain) |
| 3556 | size += data->callchain->nr; |
| 3557 | |
| 3558 | header->size += size * sizeof(u64); |
| 3559 | } |
| 3560 | |
| 3561 | if (sample_type & PERF_SAMPLE_RAW) { |
| 3562 | int size = sizeof(u32); |
| 3563 | |
| 3564 | if (data->raw) |
| 3565 | size += data->raw->size; |
| 3566 | else |
| 3567 | size += sizeof(u32); |
| 3568 | |
| 3569 | WARN_ON_ONCE(size & (sizeof(u64)-1)); |
| 3570 | header->size += size; |
| 3571 | } |
| 3572 | } |
| 3573 | |
| 3574 | static void perf_event_output(struct perf_event *event, int nmi, |
| 3575 | struct perf_sample_data *data, |
| 3576 | struct pt_regs *regs) |
| 3577 | { |
| 3578 | struct perf_output_handle handle; |
| 3579 | struct perf_event_header header; |
| 3580 | |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 3581 | /* protect the callchain buffers */ |
| 3582 | rcu_read_lock(); |
| 3583 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3584 | perf_prepare_sample(&header, data, event, regs); |
| 3585 | |
| 3586 | if (perf_output_begin(&handle, event, header.size, nmi, 1)) |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 3587 | goto exit; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3588 | |
| 3589 | perf_output_sample(&handle, &header, data, event); |
| 3590 | |
| 3591 | perf_output_end(&handle); |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 3592 | |
| 3593 | exit: |
| 3594 | rcu_read_unlock(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3595 | } |
| 3596 | |
| 3597 | /* |
| 3598 | * read event_id |
| 3599 | */ |
| 3600 | |
| 3601 | struct perf_read_event { |
| 3602 | struct perf_event_header header; |
| 3603 | |
| 3604 | u32 pid; |
| 3605 | u32 tid; |
| 3606 | }; |
| 3607 | |
| 3608 | static void |
| 3609 | perf_event_read_event(struct perf_event *event, |
| 3610 | struct task_struct *task) |
| 3611 | { |
| 3612 | struct perf_output_handle handle; |
| 3613 | struct perf_read_event read_event = { |
| 3614 | .header = { |
| 3615 | .type = PERF_RECORD_READ, |
| 3616 | .misc = 0, |
| 3617 | .size = sizeof(read_event) + perf_event_read_size(event), |
| 3618 | }, |
| 3619 | .pid = perf_event_pid(event, task), |
| 3620 | .tid = perf_event_tid(event, task), |
| 3621 | }; |
| 3622 | int ret; |
| 3623 | |
| 3624 | ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0); |
| 3625 | if (ret) |
| 3626 | return; |
| 3627 | |
| 3628 | perf_output_put(&handle, read_event); |
| 3629 | perf_output_read(&handle, event); |
| 3630 | |
| 3631 | perf_output_end(&handle); |
| 3632 | } |
| 3633 | |
| 3634 | /* |
| 3635 | * task tracking -- fork/exit |
| 3636 | * |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 3637 | * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3638 | */ |
| 3639 | |
| 3640 | struct perf_task_event { |
| 3641 | struct task_struct *task; |
| 3642 | struct perf_event_context *task_ctx; |
| 3643 | |
| 3644 | struct { |
| 3645 | struct perf_event_header header; |
| 3646 | |
| 3647 | u32 pid; |
| 3648 | u32 ppid; |
| 3649 | u32 tid; |
| 3650 | u32 ptid; |
| 3651 | u64 time; |
| 3652 | } event_id; |
| 3653 | }; |
| 3654 | |
| 3655 | static void perf_event_task_output(struct perf_event *event, |
| 3656 | struct perf_task_event *task_event) |
| 3657 | { |
| 3658 | struct perf_output_handle handle; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3659 | struct task_struct *task = task_event->task; |
Mike Galbraith | 8bb39f9 | 2010-03-26 11:11:33 +0100 | [diff] [blame] | 3660 | int size, ret; |
| 3661 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3662 | size = task_event->event_id.header.size; |
| 3663 | ret = perf_output_begin(&handle, event, size, 0, 0); |
| 3664 | |
Peter Zijlstra | ef60777 | 2010-05-18 10:50:41 +0200 | [diff] [blame] | 3665 | if (ret) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3666 | return; |
| 3667 | |
| 3668 | task_event->event_id.pid = perf_event_pid(event, task); |
| 3669 | task_event->event_id.ppid = perf_event_pid(event, current); |
| 3670 | |
| 3671 | task_event->event_id.tid = perf_event_tid(event, task); |
| 3672 | task_event->event_id.ptid = perf_event_tid(event, current); |
| 3673 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3674 | perf_output_put(&handle, task_event->event_id); |
| 3675 | |
| 3676 | perf_output_end(&handle); |
| 3677 | } |
| 3678 | |
| 3679 | static int perf_event_task_match(struct perf_event *event) |
| 3680 | { |
Peter Zijlstra | 6f93d0a | 2010-02-14 11:12:04 +0100 | [diff] [blame] | 3681 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
Peter Zijlstra | 22e1908 | 2010-01-18 09:12:32 +0100 | [diff] [blame] | 3682 | return 0; |
| 3683 | |
Peter Zijlstra | 5d27c23 | 2009-12-17 13:16:32 +0100 | [diff] [blame] | 3684 | if (event->cpu != -1 && event->cpu != smp_processor_id()) |
| 3685 | return 0; |
| 3686 | |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 3687 | if (event->attr.comm || event->attr.mmap || |
| 3688 | event->attr.mmap_data || event->attr.task) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3689 | return 1; |
| 3690 | |
| 3691 | return 0; |
| 3692 | } |
| 3693 | |
| 3694 | static void perf_event_task_ctx(struct perf_event_context *ctx, |
| 3695 | struct perf_task_event *task_event) |
| 3696 | { |
| 3697 | struct perf_event *event; |
| 3698 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3699 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
| 3700 | if (perf_event_task_match(event)) |
| 3701 | perf_event_task_output(event, task_event); |
| 3702 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3703 | } |
| 3704 | |
| 3705 | static void perf_event_task_event(struct perf_task_event *task_event) |
| 3706 | { |
| 3707 | struct perf_cpu_context *cpuctx; |
| 3708 | struct perf_event_context *ctx = task_event->task_ctx; |
| 3709 | |
Peter Zijlstra | d6ff86c | 2009-11-20 22:19:46 +0100 | [diff] [blame] | 3710 | rcu_read_lock(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3711 | cpuctx = &get_cpu_var(perf_cpu_context); |
| 3712 | perf_event_task_ctx(&cpuctx->ctx, task_event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3713 | if (!ctx) |
Peter Zijlstra | 6f93d0a | 2010-02-14 11:12:04 +0100 | [diff] [blame] | 3714 | ctx = rcu_dereference(current->perf_event_ctxp); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3715 | if (ctx) |
| 3716 | perf_event_task_ctx(ctx, task_event); |
Peter Zijlstra | 5d27c23 | 2009-12-17 13:16:32 +0100 | [diff] [blame] | 3717 | put_cpu_var(perf_cpu_context); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3718 | rcu_read_unlock(); |
| 3719 | } |
| 3720 | |
| 3721 | static void perf_event_task(struct task_struct *task, |
| 3722 | struct perf_event_context *task_ctx, |
| 3723 | int new) |
| 3724 | { |
| 3725 | struct perf_task_event task_event; |
| 3726 | |
| 3727 | if (!atomic_read(&nr_comm_events) && |
| 3728 | !atomic_read(&nr_mmap_events) && |
| 3729 | !atomic_read(&nr_task_events)) |
| 3730 | return; |
| 3731 | |
| 3732 | task_event = (struct perf_task_event){ |
| 3733 | .task = task, |
| 3734 | .task_ctx = task_ctx, |
| 3735 | .event_id = { |
| 3736 | .header = { |
| 3737 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
| 3738 | .misc = 0, |
| 3739 | .size = sizeof(task_event.event_id), |
| 3740 | }, |
| 3741 | /* .pid */ |
| 3742 | /* .ppid */ |
| 3743 | /* .tid */ |
| 3744 | /* .ptid */ |
Peter Zijlstra | 6f93d0a | 2010-02-14 11:12:04 +0100 | [diff] [blame] | 3745 | .time = perf_clock(), |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3746 | }, |
| 3747 | }; |
| 3748 | |
| 3749 | perf_event_task_event(&task_event); |
| 3750 | } |
| 3751 | |
| 3752 | void perf_event_fork(struct task_struct *task) |
| 3753 | { |
| 3754 | perf_event_task(task, NULL, 1); |
| 3755 | } |
| 3756 | |
| 3757 | /* |
| 3758 | * comm tracking |
| 3759 | */ |
| 3760 | |
| 3761 | struct perf_comm_event { |
| 3762 | struct task_struct *task; |
| 3763 | char *comm; |
| 3764 | int comm_size; |
| 3765 | |
| 3766 | struct { |
| 3767 | struct perf_event_header header; |
| 3768 | |
| 3769 | u32 pid; |
| 3770 | u32 tid; |
| 3771 | } event_id; |
| 3772 | }; |
| 3773 | |
| 3774 | static void perf_event_comm_output(struct perf_event *event, |
| 3775 | struct perf_comm_event *comm_event) |
| 3776 | { |
| 3777 | struct perf_output_handle handle; |
| 3778 | int size = comm_event->event_id.header.size; |
| 3779 | int ret = perf_output_begin(&handle, event, size, 0, 0); |
| 3780 | |
| 3781 | if (ret) |
| 3782 | return; |
| 3783 | |
| 3784 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
| 3785 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); |
| 3786 | |
| 3787 | perf_output_put(&handle, comm_event->event_id); |
| 3788 | perf_output_copy(&handle, comm_event->comm, |
| 3789 | comm_event->comm_size); |
| 3790 | perf_output_end(&handle); |
| 3791 | } |
| 3792 | |
| 3793 | static int perf_event_comm_match(struct perf_event *event) |
| 3794 | { |
Peter Zijlstra | 6f93d0a | 2010-02-14 11:12:04 +0100 | [diff] [blame] | 3795 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
Peter Zijlstra | 22e1908 | 2010-01-18 09:12:32 +0100 | [diff] [blame] | 3796 | return 0; |
| 3797 | |
Peter Zijlstra | 5d27c23 | 2009-12-17 13:16:32 +0100 | [diff] [blame] | 3798 | if (event->cpu != -1 && event->cpu != smp_processor_id()) |
| 3799 | return 0; |
| 3800 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3801 | if (event->attr.comm) |
| 3802 | return 1; |
| 3803 | |
| 3804 | return 0; |
| 3805 | } |
| 3806 | |
| 3807 | static void perf_event_comm_ctx(struct perf_event_context *ctx, |
| 3808 | struct perf_comm_event *comm_event) |
| 3809 | { |
| 3810 | struct perf_event *event; |
| 3811 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3812 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
| 3813 | if (perf_event_comm_match(event)) |
| 3814 | perf_event_comm_output(event, comm_event); |
| 3815 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3816 | } |
| 3817 | |
| 3818 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
| 3819 | { |
| 3820 | struct perf_cpu_context *cpuctx; |
| 3821 | struct perf_event_context *ctx; |
| 3822 | unsigned int size; |
| 3823 | char comm[TASK_COMM_LEN]; |
| 3824 | |
| 3825 | memset(comm, 0, sizeof(comm)); |
Márton Németh | 96b02d7 | 2009-11-21 23:10:15 +0100 | [diff] [blame] | 3826 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3827 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
| 3828 | |
| 3829 | comm_event->comm = comm; |
| 3830 | comm_event->comm_size = size; |
| 3831 | |
| 3832 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
| 3833 | |
Peter Zijlstra | f6595f3 | 2009-11-20 22:19:47 +0100 | [diff] [blame] | 3834 | rcu_read_lock(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3835 | cpuctx = &get_cpu_var(perf_cpu_context); |
| 3836 | perf_event_comm_ctx(&cpuctx->ctx, comm_event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3837 | ctx = rcu_dereference(current->perf_event_ctxp); |
| 3838 | if (ctx) |
| 3839 | perf_event_comm_ctx(ctx, comm_event); |
Peter Zijlstra | 5d27c23 | 2009-12-17 13:16:32 +0100 | [diff] [blame] | 3840 | put_cpu_var(perf_cpu_context); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3841 | rcu_read_unlock(); |
| 3842 | } |
| 3843 | |
| 3844 | void perf_event_comm(struct task_struct *task) |
| 3845 | { |
| 3846 | struct perf_comm_event comm_event; |
| 3847 | |
| 3848 | if (task->perf_event_ctxp) |
| 3849 | perf_event_enable_on_exec(task); |
| 3850 | |
| 3851 | if (!atomic_read(&nr_comm_events)) |
| 3852 | return; |
| 3853 | |
| 3854 | comm_event = (struct perf_comm_event){ |
| 3855 | .task = task, |
| 3856 | /* .comm */ |
| 3857 | /* .comm_size */ |
| 3858 | .event_id = { |
| 3859 | .header = { |
| 3860 | .type = PERF_RECORD_COMM, |
| 3861 | .misc = 0, |
| 3862 | /* .size */ |
| 3863 | }, |
| 3864 | /* .pid */ |
| 3865 | /* .tid */ |
| 3866 | }, |
| 3867 | }; |
| 3868 | |
| 3869 | perf_event_comm_event(&comm_event); |
| 3870 | } |
| 3871 | |
| 3872 | /* |
| 3873 | * mmap tracking |
| 3874 | */ |
| 3875 | |
| 3876 | struct perf_mmap_event { |
| 3877 | struct vm_area_struct *vma; |
| 3878 | |
| 3879 | const char *file_name; |
| 3880 | int file_size; |
| 3881 | |
| 3882 | struct { |
| 3883 | struct perf_event_header header; |
| 3884 | |
| 3885 | u32 pid; |
| 3886 | u32 tid; |
| 3887 | u64 start; |
| 3888 | u64 len; |
| 3889 | u64 pgoff; |
| 3890 | } event_id; |
| 3891 | }; |
| 3892 | |
| 3893 | static void perf_event_mmap_output(struct perf_event *event, |
| 3894 | struct perf_mmap_event *mmap_event) |
| 3895 | { |
| 3896 | struct perf_output_handle handle; |
| 3897 | int size = mmap_event->event_id.header.size; |
| 3898 | int ret = perf_output_begin(&handle, event, size, 0, 0); |
| 3899 | |
| 3900 | if (ret) |
| 3901 | return; |
| 3902 | |
| 3903 | mmap_event->event_id.pid = perf_event_pid(event, current); |
| 3904 | mmap_event->event_id.tid = perf_event_tid(event, current); |
| 3905 | |
| 3906 | perf_output_put(&handle, mmap_event->event_id); |
| 3907 | perf_output_copy(&handle, mmap_event->file_name, |
| 3908 | mmap_event->file_size); |
| 3909 | perf_output_end(&handle); |
| 3910 | } |
| 3911 | |
| 3912 | static int perf_event_mmap_match(struct perf_event *event, |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 3913 | struct perf_mmap_event *mmap_event, |
| 3914 | int executable) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3915 | { |
Peter Zijlstra | 6f93d0a | 2010-02-14 11:12:04 +0100 | [diff] [blame] | 3916 | if (event->state < PERF_EVENT_STATE_INACTIVE) |
Peter Zijlstra | 22e1908 | 2010-01-18 09:12:32 +0100 | [diff] [blame] | 3917 | return 0; |
| 3918 | |
Peter Zijlstra | 5d27c23 | 2009-12-17 13:16:32 +0100 | [diff] [blame] | 3919 | if (event->cpu != -1 && event->cpu != smp_processor_id()) |
| 3920 | return 0; |
| 3921 | |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 3922 | if ((!executable && event->attr.mmap_data) || |
| 3923 | (executable && event->attr.mmap)) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3924 | return 1; |
| 3925 | |
| 3926 | return 0; |
| 3927 | } |
| 3928 | |
| 3929 | static void perf_event_mmap_ctx(struct perf_event_context *ctx, |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 3930 | struct perf_mmap_event *mmap_event, |
| 3931 | int executable) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3932 | { |
| 3933 | struct perf_event *event; |
| 3934 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3935 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 3936 | if (perf_event_mmap_match(event, mmap_event, executable)) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3937 | perf_event_mmap_output(event, mmap_event); |
| 3938 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3939 | } |
| 3940 | |
| 3941 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
| 3942 | { |
| 3943 | struct perf_cpu_context *cpuctx; |
| 3944 | struct perf_event_context *ctx; |
| 3945 | struct vm_area_struct *vma = mmap_event->vma; |
| 3946 | struct file *file = vma->vm_file; |
| 3947 | unsigned int size; |
| 3948 | char tmp[16]; |
| 3949 | char *buf = NULL; |
| 3950 | const char *name; |
| 3951 | |
| 3952 | memset(tmp, 0, sizeof(tmp)); |
| 3953 | |
| 3954 | if (file) { |
| 3955 | /* |
| 3956 | * d_path works from the end of the buffer backwards, so we |
| 3957 | * need to add enough zero bytes after the string to handle |
| 3958 | * the 64bit alignment we do later. |
| 3959 | */ |
| 3960 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); |
| 3961 | if (!buf) { |
| 3962 | name = strncpy(tmp, "//enomem", sizeof(tmp)); |
| 3963 | goto got_name; |
| 3964 | } |
| 3965 | name = d_path(&file->f_path, buf, PATH_MAX); |
| 3966 | if (IS_ERR(name)) { |
| 3967 | name = strncpy(tmp, "//toolong", sizeof(tmp)); |
| 3968 | goto got_name; |
| 3969 | } |
| 3970 | } else { |
| 3971 | if (arch_vma_name(mmap_event->vma)) { |
| 3972 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), |
| 3973 | sizeof(tmp)); |
| 3974 | goto got_name; |
| 3975 | } |
| 3976 | |
| 3977 | if (!vma->vm_mm) { |
| 3978 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); |
| 3979 | goto got_name; |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 3980 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
| 3981 | vma->vm_end >= vma->vm_mm->brk) { |
| 3982 | name = strncpy(tmp, "[heap]", sizeof(tmp)); |
| 3983 | goto got_name; |
| 3984 | } else if (vma->vm_start <= vma->vm_mm->start_stack && |
| 3985 | vma->vm_end >= vma->vm_mm->start_stack) { |
| 3986 | name = strncpy(tmp, "[stack]", sizeof(tmp)); |
| 3987 | goto got_name; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 3988 | } |
| 3989 | |
| 3990 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
| 3991 | goto got_name; |
| 3992 | } |
| 3993 | |
| 3994 | got_name: |
| 3995 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
| 3996 | |
| 3997 | mmap_event->file_name = name; |
| 3998 | mmap_event->file_size = size; |
| 3999 | |
| 4000 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
| 4001 | |
Peter Zijlstra | f6d9dd2 | 2009-11-20 22:19:48 +0100 | [diff] [blame] | 4002 | rcu_read_lock(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4003 | cpuctx = &get_cpu_var(perf_cpu_context); |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 4004 | perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, vma->vm_flags & VM_EXEC); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4005 | ctx = rcu_dereference(current->perf_event_ctxp); |
| 4006 | if (ctx) |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 4007 | perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC); |
Peter Zijlstra | 5d27c23 | 2009-12-17 13:16:32 +0100 | [diff] [blame] | 4008 | put_cpu_var(perf_cpu_context); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4009 | rcu_read_unlock(); |
| 4010 | |
| 4011 | kfree(buf); |
| 4012 | } |
| 4013 | |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 4014 | void perf_event_mmap(struct vm_area_struct *vma) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4015 | { |
| 4016 | struct perf_mmap_event mmap_event; |
| 4017 | |
| 4018 | if (!atomic_read(&nr_mmap_events)) |
| 4019 | return; |
| 4020 | |
| 4021 | mmap_event = (struct perf_mmap_event){ |
| 4022 | .vma = vma, |
| 4023 | /* .file_name */ |
| 4024 | /* .file_size */ |
| 4025 | .event_id = { |
| 4026 | .header = { |
| 4027 | .type = PERF_RECORD_MMAP, |
Zhang, Yanmin | 39447b3 | 2010-04-19 13:32:41 +0800 | [diff] [blame] | 4028 | .misc = PERF_RECORD_MISC_USER, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4029 | /* .size */ |
| 4030 | }, |
| 4031 | /* .pid */ |
| 4032 | /* .tid */ |
| 4033 | .start = vma->vm_start, |
| 4034 | .len = vma->vm_end - vma->vm_start, |
Peter Zijlstra | 3a0304e | 2010-02-26 10:33:41 +0100 | [diff] [blame] | 4035 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4036 | }, |
| 4037 | }; |
| 4038 | |
| 4039 | perf_event_mmap_event(&mmap_event); |
| 4040 | } |
| 4041 | |
| 4042 | /* |
| 4043 | * IRQ throttle logging |
| 4044 | */ |
| 4045 | |
| 4046 | static void perf_log_throttle(struct perf_event *event, int enable) |
| 4047 | { |
| 4048 | struct perf_output_handle handle; |
| 4049 | int ret; |
| 4050 | |
| 4051 | struct { |
| 4052 | struct perf_event_header header; |
| 4053 | u64 time; |
| 4054 | u64 id; |
| 4055 | u64 stream_id; |
| 4056 | } throttle_event = { |
| 4057 | .header = { |
| 4058 | .type = PERF_RECORD_THROTTLE, |
| 4059 | .misc = 0, |
| 4060 | .size = sizeof(throttle_event), |
| 4061 | }, |
| 4062 | .time = perf_clock(), |
| 4063 | .id = primary_event_id(event), |
| 4064 | .stream_id = event->id, |
| 4065 | }; |
| 4066 | |
| 4067 | if (enable) |
| 4068 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
| 4069 | |
| 4070 | ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0); |
| 4071 | if (ret) |
| 4072 | return; |
| 4073 | |
| 4074 | perf_output_put(&handle, throttle_event); |
| 4075 | perf_output_end(&handle); |
| 4076 | } |
| 4077 | |
| 4078 | /* |
| 4079 | * Generic event overflow handling, sampling. |
| 4080 | */ |
| 4081 | |
| 4082 | static int __perf_event_overflow(struct perf_event *event, int nmi, |
| 4083 | int throttle, struct perf_sample_data *data, |
| 4084 | struct pt_regs *regs) |
| 4085 | { |
| 4086 | int events = atomic_read(&event->event_limit); |
| 4087 | struct hw_perf_event *hwc = &event->hw; |
| 4088 | int ret = 0; |
| 4089 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4090 | if (!throttle) { |
| 4091 | hwc->interrupts++; |
| 4092 | } else { |
| 4093 | if (hwc->interrupts != MAX_INTERRUPTS) { |
| 4094 | hwc->interrupts++; |
| 4095 | if (HZ * hwc->interrupts > |
| 4096 | (u64)sysctl_perf_event_sample_rate) { |
| 4097 | hwc->interrupts = MAX_INTERRUPTS; |
| 4098 | perf_log_throttle(event, 0); |
| 4099 | ret = 1; |
| 4100 | } |
| 4101 | } else { |
| 4102 | /* |
| 4103 | * Keep re-disabling events even though on the previous |
| 4104 | * pass we disabled it - just in case we raced with a |
| 4105 | * sched-in and the event got enabled again: |
| 4106 | */ |
| 4107 | ret = 1; |
| 4108 | } |
| 4109 | } |
| 4110 | |
| 4111 | if (event->attr.freq) { |
| 4112 | u64 now = perf_clock(); |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 4113 | s64 delta = now - hwc->freq_time_stamp; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4114 | |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 4115 | hwc->freq_time_stamp = now; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4116 | |
Peter Zijlstra | abd5071 | 2010-01-26 18:50:16 +0100 | [diff] [blame] | 4117 | if (delta > 0 && delta < 2*TICK_NSEC) |
| 4118 | perf_adjust_period(event, delta, hwc->last_period); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4119 | } |
| 4120 | |
| 4121 | /* |
| 4122 | * XXX event_limit might not quite work as expected on inherited |
| 4123 | * events |
| 4124 | */ |
| 4125 | |
| 4126 | event->pending_kill = POLL_IN; |
| 4127 | if (events && atomic_dec_and_test(&event->event_limit)) { |
| 4128 | ret = 1; |
| 4129 | event->pending_kill = POLL_HUP; |
| 4130 | if (nmi) { |
| 4131 | event->pending_disable = 1; |
| 4132 | perf_pending_queue(&event->pending, |
| 4133 | perf_pending_event); |
| 4134 | } else |
| 4135 | perf_event_disable(event); |
| 4136 | } |
| 4137 | |
Peter Zijlstra | 453f19e | 2009-11-20 22:19:43 +0100 | [diff] [blame] | 4138 | if (event->overflow_handler) |
| 4139 | event->overflow_handler(event, nmi, data, regs); |
| 4140 | else |
| 4141 | perf_event_output(event, nmi, data, regs); |
| 4142 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4143 | return ret; |
| 4144 | } |
| 4145 | |
| 4146 | int perf_event_overflow(struct perf_event *event, int nmi, |
| 4147 | struct perf_sample_data *data, |
| 4148 | struct pt_regs *regs) |
| 4149 | { |
| 4150 | return __perf_event_overflow(event, nmi, 1, data, regs); |
| 4151 | } |
| 4152 | |
| 4153 | /* |
| 4154 | * Generic software event infrastructure |
| 4155 | */ |
| 4156 | |
| 4157 | /* |
| 4158 | * We directly increment event->count and keep a second value in |
| 4159 | * event->hw.period_left to count intervals. This period event |
| 4160 | * is kept in the range [-sample_period, 0] so that we can use the |
| 4161 | * sign as trigger. |
| 4162 | */ |
| 4163 | |
| 4164 | static u64 perf_swevent_set_period(struct perf_event *event) |
| 4165 | { |
| 4166 | struct hw_perf_event *hwc = &event->hw; |
| 4167 | u64 period = hwc->last_period; |
| 4168 | u64 nr, offset; |
| 4169 | s64 old, val; |
| 4170 | |
| 4171 | hwc->last_period = hwc->sample_period; |
| 4172 | |
| 4173 | again: |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 4174 | old = val = local64_read(&hwc->period_left); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4175 | if (val < 0) |
| 4176 | return 0; |
| 4177 | |
| 4178 | nr = div64_u64(period + val, period); |
| 4179 | offset = nr * period; |
| 4180 | val -= offset; |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 4181 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4182 | goto again; |
| 4183 | |
| 4184 | return nr; |
| 4185 | } |
| 4186 | |
Peter Zijlstra | 0cff784 | 2009-11-20 22:19:44 +0100 | [diff] [blame] | 4187 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4188 | int nmi, struct perf_sample_data *data, |
| 4189 | struct pt_regs *regs) |
| 4190 | { |
| 4191 | struct hw_perf_event *hwc = &event->hw; |
| 4192 | int throttle = 0; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4193 | |
| 4194 | data->period = event->hw.last_period; |
Peter Zijlstra | 0cff784 | 2009-11-20 22:19:44 +0100 | [diff] [blame] | 4195 | if (!overflow) |
| 4196 | overflow = perf_swevent_set_period(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4197 | |
| 4198 | if (hwc->interrupts == MAX_INTERRUPTS) |
| 4199 | return; |
| 4200 | |
| 4201 | for (; overflow; overflow--) { |
| 4202 | if (__perf_event_overflow(event, nmi, throttle, |
| 4203 | data, regs)) { |
| 4204 | /* |
| 4205 | * We inhibit the overflow from happening when |
| 4206 | * hwc->interrupts == MAX_INTERRUPTS. |
| 4207 | */ |
| 4208 | break; |
| 4209 | } |
| 4210 | throttle = 1; |
| 4211 | } |
| 4212 | } |
| 4213 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4214 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4215 | int nmi, struct perf_sample_data *data, |
| 4216 | struct pt_regs *regs) |
| 4217 | { |
| 4218 | struct hw_perf_event *hwc = &event->hw; |
| 4219 | |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 4220 | local64_add(nr, &event->count); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4221 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4222 | if (!regs) |
| 4223 | return; |
| 4224 | |
Peter Zijlstra | 0cff784 | 2009-11-20 22:19:44 +0100 | [diff] [blame] | 4225 | if (!hwc->sample_period) |
| 4226 | return; |
| 4227 | |
| 4228 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
| 4229 | return perf_swevent_overflow(event, 1, nmi, data, regs); |
| 4230 | |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 4231 | if (local64_add_negative(nr, &hwc->period_left)) |
Peter Zijlstra | 0cff784 | 2009-11-20 22:19:44 +0100 | [diff] [blame] | 4232 | return; |
| 4233 | |
| 4234 | perf_swevent_overflow(event, 0, nmi, data, regs); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4235 | } |
| 4236 | |
Frederic Weisbecker | f5ffe02 | 2009-11-23 15:42:34 +0100 | [diff] [blame] | 4237 | static int perf_exclude_event(struct perf_event *event, |
| 4238 | struct pt_regs *regs) |
| 4239 | { |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4240 | if (event->hw.state & PERF_HES_STOPPED) |
| 4241 | return 0; |
| 4242 | |
Frederic Weisbecker | f5ffe02 | 2009-11-23 15:42:34 +0100 | [diff] [blame] | 4243 | if (regs) { |
| 4244 | if (event->attr.exclude_user && user_mode(regs)) |
| 4245 | return 1; |
| 4246 | |
| 4247 | if (event->attr.exclude_kernel && !user_mode(regs)) |
| 4248 | return 1; |
| 4249 | } |
| 4250 | |
| 4251 | return 0; |
| 4252 | } |
| 4253 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4254 | static int perf_swevent_match(struct perf_event *event, |
| 4255 | enum perf_type_id type, |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 4256 | u32 event_id, |
| 4257 | struct perf_sample_data *data, |
| 4258 | struct pt_regs *regs) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4259 | { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4260 | if (event->attr.type != type) |
| 4261 | return 0; |
Frederic Weisbecker | f5ffe02 | 2009-11-23 15:42:34 +0100 | [diff] [blame] | 4262 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4263 | if (event->attr.config != event_id) |
| 4264 | return 0; |
| 4265 | |
Frederic Weisbecker | f5ffe02 | 2009-11-23 15:42:34 +0100 | [diff] [blame] | 4266 | if (perf_exclude_event(event, regs)) |
| 4267 | return 0; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4268 | |
| 4269 | return 1; |
| 4270 | } |
| 4271 | |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4272 | static inline u64 swevent_hash(u64 type, u32 event_id) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4273 | { |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4274 | u64 val = event_id | (type << 32); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4275 | |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4276 | return hash_64(val, SWEVENT_HLIST_BITS); |
| 4277 | } |
| 4278 | |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4279 | static inline struct hlist_head * |
| 4280 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4281 | { |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4282 | u64 hash = swevent_hash(type, event_id); |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4283 | |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4284 | return &hlist->heads[hash]; |
| 4285 | } |
| 4286 | |
| 4287 | /* For the read side: events when they trigger */ |
| 4288 | static inline struct hlist_head * |
| 4289 | find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id) |
| 4290 | { |
| 4291 | struct swevent_hlist *hlist; |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4292 | |
| 4293 | hlist = rcu_dereference(ctx->swevent_hlist); |
| 4294 | if (!hlist) |
| 4295 | return NULL; |
| 4296 | |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4297 | return __find_swevent_head(hlist, type, event_id); |
| 4298 | } |
| 4299 | |
| 4300 | /* For the event head insertion and removal in the hlist */ |
| 4301 | static inline struct hlist_head * |
| 4302 | find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event) |
| 4303 | { |
| 4304 | struct swevent_hlist *hlist; |
| 4305 | u32 event_id = event->attr.config; |
| 4306 | u64 type = event->attr.type; |
| 4307 | |
| 4308 | /* |
| 4309 | * Event scheduling is always serialized against hlist allocation |
| 4310 | * and release. Which makes the protected version suitable here. |
| 4311 | * The context lock guarantees that. |
| 4312 | */ |
| 4313 | hlist = rcu_dereference_protected(ctx->swevent_hlist, |
| 4314 | lockdep_is_held(&event->ctx->lock)); |
| 4315 | if (!hlist) |
| 4316 | return NULL; |
| 4317 | |
| 4318 | return __find_swevent_head(hlist, type, event_id); |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4319 | } |
| 4320 | |
| 4321 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, |
| 4322 | u64 nr, int nmi, |
| 4323 | struct perf_sample_data *data, |
| 4324 | struct pt_regs *regs) |
| 4325 | { |
| 4326 | struct perf_cpu_context *cpuctx; |
| 4327 | struct perf_event *event; |
| 4328 | struct hlist_node *node; |
| 4329 | struct hlist_head *head; |
| 4330 | |
| 4331 | cpuctx = &__get_cpu_var(perf_cpu_context); |
| 4332 | |
| 4333 | rcu_read_lock(); |
| 4334 | |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4335 | head = find_swevent_head_rcu(cpuctx, type, event_id); |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4336 | |
| 4337 | if (!head) |
| 4338 | goto end; |
| 4339 | |
| 4340 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 4341 | if (perf_swevent_match(event, type, event_id, data, regs)) |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4342 | perf_swevent_event(event, nr, nmi, data, regs); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4343 | } |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4344 | end: |
| 4345 | rcu_read_unlock(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4346 | } |
| 4347 | |
Peter Zijlstra | 4ed7c92 | 2009-11-23 11:37:29 +0100 | [diff] [blame] | 4348 | int perf_swevent_get_recursion_context(void) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4349 | { |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4350 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
Frederic Weisbecker | ce71b9d | 2009-11-22 05:26:55 +0100 | [diff] [blame] | 4351 | |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 4352 | return get_recursion_context(cpuctx->recursion); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4353 | } |
Ingo Molnar | 645e8cc | 2009-11-22 12:20:19 +0100 | [diff] [blame] | 4354 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4355 | |
Peter Zijlstra | ecc55f8 | 2010-05-21 15:11:34 +0200 | [diff] [blame] | 4356 | void inline perf_swevent_put_recursion_context(int rctx) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4357 | { |
Peter Zijlstra | 4ed7c92 | 2009-11-23 11:37:29 +0100 | [diff] [blame] | 4358 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 4359 | |
| 4360 | put_recursion_context(cpuctx->recursion, rctx); |
Frederic Weisbecker | ce71b9d | 2009-11-22 05:26:55 +0100 | [diff] [blame] | 4361 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4362 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4363 | void __perf_sw_event(u32 event_id, u64 nr, int nmi, |
| 4364 | struct pt_regs *regs, u64 addr) |
| 4365 | { |
Ingo Molnar | a4234bf | 2009-11-23 10:57:59 +0100 | [diff] [blame] | 4366 | struct perf_sample_data data; |
Peter Zijlstra | 4ed7c92 | 2009-11-23 11:37:29 +0100 | [diff] [blame] | 4367 | int rctx; |
| 4368 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4369 | preempt_disable_notrace(); |
Peter Zijlstra | 4ed7c92 | 2009-11-23 11:37:29 +0100 | [diff] [blame] | 4370 | rctx = perf_swevent_get_recursion_context(); |
| 4371 | if (rctx < 0) |
| 4372 | return; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4373 | |
Peter Zijlstra | dc1d628 | 2010-03-03 15:55:04 +0100 | [diff] [blame] | 4374 | perf_sample_data_init(&data, addr); |
Ingo Molnar | a4234bf | 2009-11-23 10:57:59 +0100 | [diff] [blame] | 4375 | |
| 4376 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs); |
Peter Zijlstra | 4ed7c92 | 2009-11-23 11:37:29 +0100 | [diff] [blame] | 4377 | |
| 4378 | perf_swevent_put_recursion_context(rctx); |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4379 | preempt_enable_notrace(); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4380 | } |
| 4381 | |
| 4382 | static void perf_swevent_read(struct perf_event *event) |
| 4383 | { |
| 4384 | } |
| 4385 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4386 | static int perf_swevent_add(struct perf_event *event, int flags) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4387 | { |
| 4388 | struct hw_perf_event *hwc = &event->hw; |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4389 | struct perf_cpu_context *cpuctx; |
| 4390 | struct hlist_head *head; |
| 4391 | |
| 4392 | cpuctx = &__get_cpu_var(perf_cpu_context); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4393 | |
| 4394 | if (hwc->sample_period) { |
| 4395 | hwc->last_period = hwc->sample_period; |
| 4396 | perf_swevent_set_period(event); |
| 4397 | } |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4398 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4399 | hwc->state = !(flags & PERF_EF_START); |
| 4400 | |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4401 | head = find_swevent_head(cpuctx, event); |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4402 | if (WARN_ON_ONCE(!head)) |
| 4403 | return -EINVAL; |
| 4404 | |
| 4405 | hlist_add_head_rcu(&event->hlist_entry, head); |
| 4406 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4407 | return 0; |
| 4408 | } |
| 4409 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4410 | static void perf_swevent_del(struct perf_event *event, int flags) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4411 | { |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4412 | hlist_del_rcu(&event->hlist_entry); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4413 | } |
| 4414 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4415 | static void perf_swevent_start(struct perf_event *event, int flags) |
Peter Zijlstra | c6df8d5 | 2010-06-03 11:21:20 +0200 | [diff] [blame] | 4416 | { |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4417 | event->hw.state = 0; |
Peter Zijlstra | c6df8d5 | 2010-06-03 11:21:20 +0200 | [diff] [blame] | 4418 | } |
| 4419 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4420 | static void perf_swevent_stop(struct perf_event *event, int flags) |
Peter Zijlstra | c6df8d5 | 2010-06-03 11:21:20 +0200 | [diff] [blame] | 4421 | { |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4422 | event->hw.state = PERF_HES_STOPPED; |
Peter Zijlstra | c6df8d5 | 2010-06-03 11:21:20 +0200 | [diff] [blame] | 4423 | } |
| 4424 | |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4425 | /* Deref the hlist from the update side */ |
| 4426 | static inline struct swevent_hlist * |
| 4427 | swevent_hlist_deref(struct perf_cpu_context *cpuctx) |
| 4428 | { |
| 4429 | return rcu_dereference_protected(cpuctx->swevent_hlist, |
| 4430 | lockdep_is_held(&cpuctx->hlist_mutex)); |
| 4431 | } |
| 4432 | |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4433 | static void swevent_hlist_release_rcu(struct rcu_head *rcu_head) |
| 4434 | { |
| 4435 | struct swevent_hlist *hlist; |
| 4436 | |
| 4437 | hlist = container_of(rcu_head, struct swevent_hlist, rcu_head); |
| 4438 | kfree(hlist); |
| 4439 | } |
| 4440 | |
| 4441 | static void swevent_hlist_release(struct perf_cpu_context *cpuctx) |
| 4442 | { |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4443 | struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx); |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4444 | |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4445 | if (!hlist) |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4446 | return; |
| 4447 | |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4448 | rcu_assign_pointer(cpuctx->swevent_hlist, NULL); |
| 4449 | call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu); |
| 4450 | } |
| 4451 | |
| 4452 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) |
| 4453 | { |
| 4454 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); |
| 4455 | |
| 4456 | mutex_lock(&cpuctx->hlist_mutex); |
| 4457 | |
| 4458 | if (!--cpuctx->hlist_refcount) |
| 4459 | swevent_hlist_release(cpuctx); |
| 4460 | |
| 4461 | mutex_unlock(&cpuctx->hlist_mutex); |
| 4462 | } |
| 4463 | |
| 4464 | static void swevent_hlist_put(struct perf_event *event) |
| 4465 | { |
| 4466 | int cpu; |
| 4467 | |
| 4468 | if (event->cpu != -1) { |
| 4469 | swevent_hlist_put_cpu(event, event->cpu); |
| 4470 | return; |
| 4471 | } |
| 4472 | |
| 4473 | for_each_possible_cpu(cpu) |
| 4474 | swevent_hlist_put_cpu(event, cpu); |
| 4475 | } |
| 4476 | |
| 4477 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) |
| 4478 | { |
| 4479 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); |
| 4480 | int err = 0; |
| 4481 | |
| 4482 | mutex_lock(&cpuctx->hlist_mutex); |
| 4483 | |
Frederic Weisbecker | 49f135e | 2010-05-20 10:17:46 +0200 | [diff] [blame] | 4484 | if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) { |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4485 | struct swevent_hlist *hlist; |
| 4486 | |
| 4487 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); |
| 4488 | if (!hlist) { |
| 4489 | err = -ENOMEM; |
| 4490 | goto exit; |
| 4491 | } |
| 4492 | rcu_assign_pointer(cpuctx->swevent_hlist, hlist); |
| 4493 | } |
| 4494 | cpuctx->hlist_refcount++; |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 4495 | exit: |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4496 | mutex_unlock(&cpuctx->hlist_mutex); |
| 4497 | |
| 4498 | return err; |
| 4499 | } |
| 4500 | |
| 4501 | static int swevent_hlist_get(struct perf_event *event) |
| 4502 | { |
| 4503 | int err; |
| 4504 | int cpu, failed_cpu; |
| 4505 | |
| 4506 | if (event->cpu != -1) |
| 4507 | return swevent_hlist_get_cpu(event, event->cpu); |
| 4508 | |
| 4509 | get_online_cpus(); |
| 4510 | for_each_possible_cpu(cpu) { |
| 4511 | err = swevent_hlist_get_cpu(event, cpu); |
| 4512 | if (err) { |
| 4513 | failed_cpu = cpu; |
| 4514 | goto fail; |
| 4515 | } |
| 4516 | } |
| 4517 | put_online_cpus(); |
| 4518 | |
| 4519 | return 0; |
Peter Zijlstra | 9ed6060 | 2010-06-11 17:36:35 +0200 | [diff] [blame] | 4520 | fail: |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4521 | for_each_possible_cpu(cpu) { |
| 4522 | if (cpu == failed_cpu) |
| 4523 | break; |
| 4524 | swevent_hlist_put_cpu(event, cpu); |
| 4525 | } |
| 4526 | |
| 4527 | put_online_cpus(); |
| 4528 | return err; |
| 4529 | } |
| 4530 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4531 | atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
Frederic Weisbecker | 95476b6 | 2010-04-14 23:42:18 +0200 | [diff] [blame] | 4532 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4533 | static void sw_perf_event_destroy(struct perf_event *event) |
| 4534 | { |
| 4535 | u64 event_id = event->attr.config; |
| 4536 | |
| 4537 | WARN_ON(event->parent); |
| 4538 | |
| 4539 | atomic_dec(&perf_swevent_enabled[event_id]); |
| 4540 | swevent_hlist_put(event); |
| 4541 | } |
| 4542 | |
| 4543 | static int perf_swevent_init(struct perf_event *event) |
| 4544 | { |
| 4545 | int event_id = event->attr.config; |
| 4546 | |
| 4547 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
| 4548 | return -ENOENT; |
| 4549 | |
| 4550 | switch (event_id) { |
| 4551 | case PERF_COUNT_SW_CPU_CLOCK: |
| 4552 | case PERF_COUNT_SW_TASK_CLOCK: |
| 4553 | return -ENOENT; |
| 4554 | |
| 4555 | default: |
| 4556 | break; |
| 4557 | } |
| 4558 | |
| 4559 | if (event_id > PERF_COUNT_SW_MAX) |
| 4560 | return -ENOENT; |
| 4561 | |
| 4562 | if (!event->parent) { |
| 4563 | int err; |
| 4564 | |
| 4565 | err = swevent_hlist_get(event); |
| 4566 | if (err) |
| 4567 | return err; |
| 4568 | |
| 4569 | atomic_inc(&perf_swevent_enabled[event_id]); |
| 4570 | event->destroy = sw_perf_event_destroy; |
| 4571 | } |
| 4572 | |
| 4573 | return 0; |
| 4574 | } |
| 4575 | |
| 4576 | static struct pmu perf_swevent = { |
| 4577 | .event_init = perf_swevent_init, |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4578 | .add = perf_swevent_add, |
| 4579 | .del = perf_swevent_del, |
| 4580 | .start = perf_swevent_start, |
| 4581 | .stop = perf_swevent_stop, |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4582 | .read = perf_swevent_read, |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4583 | }; |
Frederic Weisbecker | 95476b6 | 2010-04-14 23:42:18 +0200 | [diff] [blame] | 4584 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4585 | #ifdef CONFIG_EVENT_TRACING |
| 4586 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4587 | static int perf_tp_filter_match(struct perf_event *event, |
Frederic Weisbecker | 95476b6 | 2010-04-14 23:42:18 +0200 | [diff] [blame] | 4588 | struct perf_sample_data *data) |
| 4589 | { |
| 4590 | void *record = data->raw->data; |
| 4591 | |
| 4592 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) |
| 4593 | return 1; |
| 4594 | return 0; |
| 4595 | } |
| 4596 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4597 | static int perf_tp_event_match(struct perf_event *event, |
| 4598 | struct perf_sample_data *data, |
| 4599 | struct pt_regs *regs) |
| 4600 | { |
Peter Zijlstra | 580d607 | 2010-05-20 20:54:31 +0200 | [diff] [blame] | 4601 | /* |
| 4602 | * All tracepoints are from kernel-space. |
| 4603 | */ |
| 4604 | if (event->attr.exclude_kernel) |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4605 | return 0; |
| 4606 | |
| 4607 | if (!perf_tp_filter_match(event, data)) |
| 4608 | return 0; |
| 4609 | |
| 4610 | return 1; |
| 4611 | } |
| 4612 | |
| 4613 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, |
Peter Zijlstra | ecc55f8 | 2010-05-21 15:11:34 +0200 | [diff] [blame] | 4614 | struct pt_regs *regs, struct hlist_head *head, int rctx) |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4615 | { |
| 4616 | struct perf_sample_data data; |
| 4617 | struct perf_event *event; |
| 4618 | struct hlist_node *node; |
| 4619 | |
| 4620 | struct perf_raw_record raw = { |
| 4621 | .size = entry_size, |
| 4622 | .data = record, |
| 4623 | }; |
| 4624 | |
| 4625 | perf_sample_data_init(&data, addr); |
| 4626 | data.raw = &raw; |
| 4627 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4628 | hlist_for_each_entry_rcu(event, node, head, hlist_entry) { |
| 4629 | if (perf_tp_event_match(event, &data, regs)) |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4630 | perf_swevent_event(event, count, 1, &data, regs); |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4631 | } |
Peter Zijlstra | ecc55f8 | 2010-05-21 15:11:34 +0200 | [diff] [blame] | 4632 | |
| 4633 | perf_swevent_put_recursion_context(rctx); |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4634 | } |
| 4635 | EXPORT_SYMBOL_GPL(perf_tp_event); |
| 4636 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4637 | static void tp_perf_event_destroy(struct perf_event *event) |
| 4638 | { |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4639 | perf_trace_destroy(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4640 | } |
| 4641 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4642 | static int perf_tp_event_init(struct perf_event *event) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4643 | { |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 4644 | int err; |
| 4645 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4646 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
| 4647 | return -ENOENT; |
| 4648 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4649 | /* |
| 4650 | * Raw tracepoint data is a severe data leak, only allow root to |
| 4651 | * have these. |
| 4652 | */ |
| 4653 | if ((event->attr.sample_type & PERF_SAMPLE_RAW) && |
| 4654 | perf_paranoid_tracepoint_raw() && |
| 4655 | !capable(CAP_SYS_ADMIN)) |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4656 | return -EPERM; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4657 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 4658 | err = perf_trace_init(event); |
| 4659 | if (err) |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4660 | return err; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4661 | |
| 4662 | event->destroy = tp_perf_event_destroy; |
| 4663 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4664 | return 0; |
| 4665 | } |
| 4666 | |
| 4667 | static struct pmu perf_tracepoint = { |
| 4668 | .event_init = perf_tp_event_init, |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4669 | .add = perf_trace_add, |
| 4670 | .del = perf_trace_del, |
| 4671 | .start = perf_swevent_start, |
| 4672 | .stop = perf_swevent_stop, |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4673 | .read = perf_swevent_read, |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4674 | }; |
| 4675 | |
| 4676 | static inline void perf_tp_register(void) |
| 4677 | { |
| 4678 | perf_pmu_register(&perf_tracepoint); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4679 | } |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 4680 | |
| 4681 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) |
| 4682 | { |
| 4683 | char *filter_str; |
| 4684 | int ret; |
| 4685 | |
| 4686 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
| 4687 | return -EINVAL; |
| 4688 | |
| 4689 | filter_str = strndup_user(arg, PAGE_SIZE); |
| 4690 | if (IS_ERR(filter_str)) |
| 4691 | return PTR_ERR(filter_str); |
| 4692 | |
| 4693 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); |
| 4694 | |
| 4695 | kfree(filter_str); |
| 4696 | return ret; |
| 4697 | } |
| 4698 | |
| 4699 | static void perf_event_free_filter(struct perf_event *event) |
| 4700 | { |
| 4701 | ftrace_profile_free_filter(event); |
| 4702 | } |
| 4703 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4704 | #else |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 4705 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4706 | static inline void perf_tp_register(void) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4707 | { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4708 | } |
Li Zefan | 6fb2915 | 2009-10-15 11:21:42 +0800 | [diff] [blame] | 4709 | |
| 4710 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) |
| 4711 | { |
| 4712 | return -ENOENT; |
| 4713 | } |
| 4714 | |
| 4715 | static void perf_event_free_filter(struct perf_event *event) |
| 4716 | { |
| 4717 | } |
| 4718 | |
Li Zefan | 07b139c | 2009-12-21 14:27:35 +0800 | [diff] [blame] | 4719 | #endif /* CONFIG_EVENT_TRACING */ |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4720 | |
Frederic Weisbecker | 24f1e32c | 2009-09-09 19:22:48 +0200 | [diff] [blame] | 4721 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
Frederic Weisbecker | f5ffe02 | 2009-11-23 15:42:34 +0100 | [diff] [blame] | 4722 | void perf_bp_event(struct perf_event *bp, void *data) |
Frederic Weisbecker | 24f1e32c | 2009-09-09 19:22:48 +0200 | [diff] [blame] | 4723 | { |
Frederic Weisbecker | f5ffe02 | 2009-11-23 15:42:34 +0100 | [diff] [blame] | 4724 | struct perf_sample_data sample; |
| 4725 | struct pt_regs *regs = data; |
| 4726 | |
Peter Zijlstra | dc1d628 | 2010-03-03 15:55:04 +0100 | [diff] [blame] | 4727 | perf_sample_data_init(&sample, bp->attr.bp_addr); |
Frederic Weisbecker | f5ffe02 | 2009-11-23 15:42:34 +0100 | [diff] [blame] | 4728 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4729 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
| 4730 | perf_swevent_event(bp, 1, 1, &sample, regs); |
Frederic Weisbecker | 24f1e32c | 2009-09-09 19:22:48 +0200 | [diff] [blame] | 4731 | } |
Frederic Weisbecker | 24f1e32c | 2009-09-09 19:22:48 +0200 | [diff] [blame] | 4732 | #endif |
| 4733 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4734 | /* |
| 4735 | * hrtimer based swevent callback |
| 4736 | */ |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4737 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4738 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4739 | { |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4740 | enum hrtimer_restart ret = HRTIMER_RESTART; |
| 4741 | struct perf_sample_data data; |
| 4742 | struct pt_regs *regs; |
| 4743 | struct perf_event *event; |
| 4744 | u64 period; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4745 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4746 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
| 4747 | event->pmu->read(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4748 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4749 | perf_sample_data_init(&data, 0); |
| 4750 | data.period = event->hw.last_period; |
| 4751 | regs = get_irq_regs(); |
| 4752 | |
| 4753 | if (regs && !perf_exclude_event(event, regs)) { |
| 4754 | if (!(event->attr.exclude_idle && current->pid == 0)) |
| 4755 | if (perf_event_overflow(event, 0, &data, regs)) |
| 4756 | ret = HRTIMER_NORESTART; |
| 4757 | } |
| 4758 | |
| 4759 | period = max_t(u64, 10000, event->hw.sample_period); |
| 4760 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); |
| 4761 | |
| 4762 | return ret; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 4763 | } |
| 4764 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4765 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
| 4766 | { |
| 4767 | struct hw_perf_event *hwc = &event->hw; |
| 4768 | |
| 4769 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
| 4770 | hwc->hrtimer.function = perf_swevent_hrtimer; |
| 4771 | if (hwc->sample_period) { |
Peter Zijlstra | fa407f3 | 2010-06-24 12:35:12 +0200 | [diff] [blame] | 4772 | s64 period = local64_read(&hwc->period_left); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4773 | |
Peter Zijlstra | fa407f3 | 2010-06-24 12:35:12 +0200 | [diff] [blame] | 4774 | if (period) { |
| 4775 | if (period < 0) |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4776 | period = 10000; |
Peter Zijlstra | fa407f3 | 2010-06-24 12:35:12 +0200 | [diff] [blame] | 4777 | |
| 4778 | local64_set(&hwc->period_left, 0); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4779 | } else { |
| 4780 | period = max_t(u64, 10000, hwc->sample_period); |
| 4781 | } |
| 4782 | __hrtimer_start_range_ns(&hwc->hrtimer, |
| 4783 | ns_to_ktime(period), 0, |
| 4784 | HRTIMER_MODE_REL, 0); |
| 4785 | } |
| 4786 | } |
| 4787 | |
| 4788 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) |
| 4789 | { |
| 4790 | struct hw_perf_event *hwc = &event->hw; |
| 4791 | |
| 4792 | if (hwc->sample_period) { |
| 4793 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
Peter Zijlstra | fa407f3 | 2010-06-24 12:35:12 +0200 | [diff] [blame] | 4794 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4795 | |
| 4796 | hrtimer_cancel(&hwc->hrtimer); |
| 4797 | } |
| 4798 | } |
| 4799 | |
| 4800 | /* |
| 4801 | * Software event: cpu wall time clock |
| 4802 | */ |
| 4803 | |
| 4804 | static void cpu_clock_event_update(struct perf_event *event) |
| 4805 | { |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4806 | s64 prev; |
| 4807 | u64 now; |
| 4808 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4809 | now = local_clock(); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4810 | prev = local64_xchg(&event->hw.prev_count, now); |
| 4811 | local64_add(now - prev, &event->count); |
| 4812 | } |
| 4813 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4814 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4815 | { |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4816 | local64_set(&event->hw.prev_count, local_clock()); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4817 | perf_swevent_start_hrtimer(event); |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4818 | } |
| 4819 | |
| 4820 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
| 4821 | { |
| 4822 | perf_swevent_cancel_hrtimer(event); |
| 4823 | cpu_clock_event_update(event); |
| 4824 | } |
| 4825 | |
| 4826 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
| 4827 | { |
| 4828 | if (flags & PERF_EF_START) |
| 4829 | cpu_clock_event_start(event, flags); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4830 | |
| 4831 | return 0; |
| 4832 | } |
| 4833 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4834 | static void cpu_clock_event_del(struct perf_event *event, int flags) |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4835 | { |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4836 | cpu_clock_event_stop(event, flags); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4837 | } |
| 4838 | |
| 4839 | static void cpu_clock_event_read(struct perf_event *event) |
| 4840 | { |
| 4841 | cpu_clock_event_update(event); |
| 4842 | } |
| 4843 | |
| 4844 | static int cpu_clock_event_init(struct perf_event *event) |
| 4845 | { |
| 4846 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
| 4847 | return -ENOENT; |
| 4848 | |
| 4849 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) |
| 4850 | return -ENOENT; |
| 4851 | |
| 4852 | return 0; |
| 4853 | } |
| 4854 | |
| 4855 | static struct pmu perf_cpu_clock = { |
| 4856 | .event_init = cpu_clock_event_init, |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4857 | .add = cpu_clock_event_add, |
| 4858 | .del = cpu_clock_event_del, |
| 4859 | .start = cpu_clock_event_start, |
| 4860 | .stop = cpu_clock_event_stop, |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4861 | .read = cpu_clock_event_read, |
| 4862 | }; |
| 4863 | |
| 4864 | /* |
| 4865 | * Software event: task time clock |
| 4866 | */ |
| 4867 | |
| 4868 | static void task_clock_event_update(struct perf_event *event, u64 now) |
| 4869 | { |
| 4870 | u64 prev; |
| 4871 | s64 delta; |
| 4872 | |
| 4873 | prev = local64_xchg(&event->hw.prev_count, now); |
| 4874 | delta = now - prev; |
| 4875 | local64_add(delta, &event->count); |
| 4876 | } |
| 4877 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4878 | static void task_clock_event_start(struct perf_event *event, int flags) |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4879 | { |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4880 | local64_set(&event->hw.prev_count, event->ctx->time); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4881 | perf_swevent_start_hrtimer(event); |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4882 | } |
| 4883 | |
| 4884 | static void task_clock_event_stop(struct perf_event *event, int flags) |
| 4885 | { |
| 4886 | perf_swevent_cancel_hrtimer(event); |
| 4887 | task_clock_event_update(event, event->ctx->time); |
| 4888 | } |
| 4889 | |
| 4890 | static int task_clock_event_add(struct perf_event *event, int flags) |
| 4891 | { |
| 4892 | if (flags & PERF_EF_START) |
| 4893 | task_clock_event_start(event, flags); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4894 | |
| 4895 | return 0; |
| 4896 | } |
| 4897 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4898 | static void task_clock_event_del(struct perf_event *event, int flags) |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4899 | { |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4900 | task_clock_event_stop(event, PERF_EF_UPDATE); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4901 | } |
| 4902 | |
| 4903 | static void task_clock_event_read(struct perf_event *event) |
| 4904 | { |
| 4905 | u64 time; |
| 4906 | |
| 4907 | if (!in_nmi()) { |
| 4908 | update_context_time(event->ctx); |
| 4909 | time = event->ctx->time; |
| 4910 | } else { |
| 4911 | u64 now = perf_clock(); |
| 4912 | u64 delta = now - event->ctx->timestamp; |
| 4913 | time = event->ctx->time + delta; |
| 4914 | } |
| 4915 | |
| 4916 | task_clock_event_update(event, time); |
| 4917 | } |
| 4918 | |
| 4919 | static int task_clock_event_init(struct perf_event *event) |
| 4920 | { |
| 4921 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
| 4922 | return -ENOENT; |
| 4923 | |
| 4924 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) |
| 4925 | return -ENOENT; |
| 4926 | |
| 4927 | return 0; |
| 4928 | } |
| 4929 | |
| 4930 | static struct pmu perf_task_clock = { |
| 4931 | .event_init = task_clock_event_init, |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 4932 | .add = task_clock_event_add, |
| 4933 | .del = task_clock_event_del, |
| 4934 | .start = task_clock_event_start, |
| 4935 | .stop = task_clock_event_stop, |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4936 | .read = task_clock_event_read, |
| 4937 | }; |
| 4938 | |
| 4939 | static LIST_HEAD(pmus); |
| 4940 | static DEFINE_MUTEX(pmus_lock); |
| 4941 | static struct srcu_struct pmus_srcu; |
| 4942 | |
Peter Zijlstra | ad5133b | 2010-06-15 12:22:39 +0200 | [diff] [blame] | 4943 | static void perf_pmu_nop_void(struct pmu *pmu) |
| 4944 | { |
| 4945 | } |
| 4946 | |
| 4947 | static int perf_pmu_nop_int(struct pmu *pmu) |
| 4948 | { |
| 4949 | return 0; |
| 4950 | } |
| 4951 | |
| 4952 | static void perf_pmu_start_txn(struct pmu *pmu) |
| 4953 | { |
| 4954 | perf_pmu_disable(pmu); |
| 4955 | } |
| 4956 | |
| 4957 | static int perf_pmu_commit_txn(struct pmu *pmu) |
| 4958 | { |
| 4959 | perf_pmu_enable(pmu); |
| 4960 | return 0; |
| 4961 | } |
| 4962 | |
| 4963 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
| 4964 | { |
| 4965 | perf_pmu_enable(pmu); |
| 4966 | } |
| 4967 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4968 | int perf_pmu_register(struct pmu *pmu) |
| 4969 | { |
Peter Zijlstra | 33696fc | 2010-06-14 08:49:00 +0200 | [diff] [blame] | 4970 | int ret; |
| 4971 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 4972 | mutex_lock(&pmus_lock); |
Peter Zijlstra | 33696fc | 2010-06-14 08:49:00 +0200 | [diff] [blame] | 4973 | ret = -ENOMEM; |
| 4974 | pmu->pmu_disable_count = alloc_percpu(int); |
| 4975 | if (!pmu->pmu_disable_count) |
| 4976 | goto unlock; |
Peter Zijlstra | ad5133b | 2010-06-15 12:22:39 +0200 | [diff] [blame] | 4977 | |
| 4978 | if (!pmu->start_txn) { |
| 4979 | if (pmu->pmu_enable) { |
| 4980 | /* |
| 4981 | * If we have pmu_enable/pmu_disable calls, install |
| 4982 | * transaction stubs that use that to try and batch |
| 4983 | * hardware accesses. |
| 4984 | */ |
| 4985 | pmu->start_txn = perf_pmu_start_txn; |
| 4986 | pmu->commit_txn = perf_pmu_commit_txn; |
| 4987 | pmu->cancel_txn = perf_pmu_cancel_txn; |
| 4988 | } else { |
| 4989 | pmu->start_txn = perf_pmu_nop_void; |
| 4990 | pmu->commit_txn = perf_pmu_nop_int; |
| 4991 | pmu->cancel_txn = perf_pmu_nop_void; |
| 4992 | } |
| 4993 | } |
| 4994 | |
| 4995 | if (!pmu->pmu_enable) { |
| 4996 | pmu->pmu_enable = perf_pmu_nop_void; |
| 4997 | pmu->pmu_disable = perf_pmu_nop_void; |
| 4998 | } |
| 4999 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5000 | list_add_rcu(&pmu->entry, &pmus); |
Peter Zijlstra | 33696fc | 2010-06-14 08:49:00 +0200 | [diff] [blame] | 5001 | ret = 0; |
| 5002 | unlock: |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5003 | mutex_unlock(&pmus_lock); |
| 5004 | |
Peter Zijlstra | 33696fc | 2010-06-14 08:49:00 +0200 | [diff] [blame] | 5005 | return ret; |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5006 | } |
| 5007 | |
| 5008 | void perf_pmu_unregister(struct pmu *pmu) |
| 5009 | { |
| 5010 | mutex_lock(&pmus_lock); |
| 5011 | list_del_rcu(&pmu->entry); |
| 5012 | mutex_unlock(&pmus_lock); |
| 5013 | |
| 5014 | synchronize_srcu(&pmus_srcu); |
Peter Zijlstra | 33696fc | 2010-06-14 08:49:00 +0200 | [diff] [blame] | 5015 | |
| 5016 | free_percpu(pmu->pmu_disable_count); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5017 | } |
| 5018 | |
| 5019 | struct pmu *perf_init_event(struct perf_event *event) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5020 | { |
Peter Zijlstra | 51b0fe3 | 2010-06-11 13:35:57 +0200 | [diff] [blame] | 5021 | struct pmu *pmu = NULL; |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5022 | int idx; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5023 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5024 | idx = srcu_read_lock(&pmus_srcu); |
| 5025 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
| 5026 | int ret = pmu->event_init(event); |
| 5027 | if (!ret) |
| 5028 | break; |
| 5029 | if (ret != -ENOENT) { |
| 5030 | pmu = ERR_PTR(ret); |
| 5031 | break; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5032 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5033 | } |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5034 | srcu_read_unlock(&pmus_srcu, idx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5035 | |
| 5036 | return pmu; |
| 5037 | } |
| 5038 | |
| 5039 | /* |
| 5040 | * Allocate and initialize a event structure |
| 5041 | */ |
| 5042 | static struct perf_event * |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5043 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5044 | struct perf_event *group_leader, |
| 5045 | struct perf_event *parent_event, |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5046 | perf_overflow_handler_t overflow_handler) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5047 | { |
Peter Zijlstra | 51b0fe3 | 2010-06-11 13:35:57 +0200 | [diff] [blame] | 5048 | struct pmu *pmu; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5049 | struct perf_event *event; |
| 5050 | struct hw_perf_event *hwc; |
| 5051 | long err; |
| 5052 | |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5053 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5054 | if (!event) |
| 5055 | return ERR_PTR(-ENOMEM); |
| 5056 | |
| 5057 | /* |
| 5058 | * Single events are their own group leaders, with an |
| 5059 | * empty sibling list: |
| 5060 | */ |
| 5061 | if (!group_leader) |
| 5062 | group_leader = event; |
| 5063 | |
| 5064 | mutex_init(&event->child_mutex); |
| 5065 | INIT_LIST_HEAD(&event->child_list); |
| 5066 | |
| 5067 | INIT_LIST_HEAD(&event->group_entry); |
| 5068 | INIT_LIST_HEAD(&event->event_entry); |
| 5069 | INIT_LIST_HEAD(&event->sibling_list); |
| 5070 | init_waitqueue_head(&event->waitq); |
| 5071 | |
| 5072 | mutex_init(&event->mmap_mutex); |
| 5073 | |
| 5074 | event->cpu = cpu; |
| 5075 | event->attr = *attr; |
| 5076 | event->group_leader = group_leader; |
| 5077 | event->pmu = NULL; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5078 | event->oncpu = -1; |
| 5079 | |
| 5080 | event->parent = parent_event; |
| 5081 | |
| 5082 | event->ns = get_pid_ns(current->nsproxy->pid_ns); |
| 5083 | event->id = atomic64_inc_return(&perf_event_id); |
| 5084 | |
| 5085 | event->state = PERF_EVENT_STATE_INACTIVE; |
| 5086 | |
Frederic Weisbecker | b326e95 | 2009-12-05 09:44:31 +0100 | [diff] [blame] | 5087 | if (!overflow_handler && parent_event) |
| 5088 | overflow_handler = parent_event->overflow_handler; |
Frederic Weisbecker | 97eaf53 | 2009-10-18 15:33:50 +0200 | [diff] [blame] | 5089 | |
Frederic Weisbecker | b326e95 | 2009-12-05 09:44:31 +0100 | [diff] [blame] | 5090 | event->overflow_handler = overflow_handler; |
Frederic Weisbecker | 97eaf53 | 2009-10-18 15:33:50 +0200 | [diff] [blame] | 5091 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5092 | if (attr->disabled) |
| 5093 | event->state = PERF_EVENT_STATE_OFF; |
| 5094 | |
| 5095 | pmu = NULL; |
| 5096 | |
| 5097 | hwc = &event->hw; |
| 5098 | hwc->sample_period = attr->sample_period; |
| 5099 | if (attr->freq && attr->sample_freq) |
| 5100 | hwc->sample_period = 1; |
| 5101 | hwc->last_period = hwc->sample_period; |
| 5102 | |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 5103 | local64_set(&hwc->period_left, hwc->sample_period); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5104 | |
| 5105 | /* |
| 5106 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
| 5107 | */ |
| 5108 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
| 5109 | goto done; |
| 5110 | |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5111 | pmu = perf_init_event(event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5112 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5113 | done: |
| 5114 | err = 0; |
| 5115 | if (!pmu) |
| 5116 | err = -EINVAL; |
| 5117 | else if (IS_ERR(pmu)) |
| 5118 | err = PTR_ERR(pmu); |
| 5119 | |
| 5120 | if (err) { |
| 5121 | if (event->ns) |
| 5122 | put_pid_ns(event->ns); |
| 5123 | kfree(event); |
| 5124 | return ERR_PTR(err); |
| 5125 | } |
| 5126 | |
| 5127 | event->pmu = pmu; |
| 5128 | |
| 5129 | if (!event->parent) { |
| 5130 | atomic_inc(&nr_events); |
Eric B Munson | 3af9e85 | 2010-05-18 15:30:49 +0100 | [diff] [blame] | 5131 | if (event->attr.mmap || event->attr.mmap_data) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5132 | atomic_inc(&nr_mmap_events); |
| 5133 | if (event->attr.comm) |
| 5134 | atomic_inc(&nr_comm_events); |
| 5135 | if (event->attr.task) |
| 5136 | atomic_inc(&nr_task_events); |
Frederic Weisbecker | 927c7a9 | 2010-07-01 16:20:36 +0200 | [diff] [blame] | 5137 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
| 5138 | err = get_callchain_buffers(); |
| 5139 | if (err) { |
| 5140 | free_event(event); |
| 5141 | return ERR_PTR(err); |
| 5142 | } |
| 5143 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5144 | } |
| 5145 | |
| 5146 | return event; |
| 5147 | } |
| 5148 | |
| 5149 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
| 5150 | struct perf_event_attr *attr) |
| 5151 | { |
| 5152 | u32 size; |
| 5153 | int ret; |
| 5154 | |
| 5155 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) |
| 5156 | return -EFAULT; |
| 5157 | |
| 5158 | /* |
| 5159 | * zero the full structure, so that a short copy will be nice. |
| 5160 | */ |
| 5161 | memset(attr, 0, sizeof(*attr)); |
| 5162 | |
| 5163 | ret = get_user(size, &uattr->size); |
| 5164 | if (ret) |
| 5165 | return ret; |
| 5166 | |
| 5167 | if (size > PAGE_SIZE) /* silly large */ |
| 5168 | goto err_size; |
| 5169 | |
| 5170 | if (!size) /* abi compat */ |
| 5171 | size = PERF_ATTR_SIZE_VER0; |
| 5172 | |
| 5173 | if (size < PERF_ATTR_SIZE_VER0) |
| 5174 | goto err_size; |
| 5175 | |
| 5176 | /* |
| 5177 | * If we're handed a bigger struct than we know of, |
| 5178 | * ensure all the unknown bits are 0 - i.e. new |
| 5179 | * user-space does not rely on any kernel feature |
| 5180 | * extensions we dont know about yet. |
| 5181 | */ |
| 5182 | if (size > sizeof(*attr)) { |
| 5183 | unsigned char __user *addr; |
| 5184 | unsigned char __user *end; |
| 5185 | unsigned char val; |
| 5186 | |
| 5187 | addr = (void __user *)uattr + sizeof(*attr); |
| 5188 | end = (void __user *)uattr + size; |
| 5189 | |
| 5190 | for (; addr < end; addr++) { |
| 5191 | ret = get_user(val, addr); |
| 5192 | if (ret) |
| 5193 | return ret; |
| 5194 | if (val) |
| 5195 | goto err_size; |
| 5196 | } |
| 5197 | size = sizeof(*attr); |
| 5198 | } |
| 5199 | |
| 5200 | ret = copy_from_user(attr, uattr, size); |
| 5201 | if (ret) |
| 5202 | return -EFAULT; |
| 5203 | |
| 5204 | /* |
| 5205 | * If the type exists, the corresponding creation will verify |
| 5206 | * the attr->config. |
| 5207 | */ |
| 5208 | if (attr->type >= PERF_TYPE_MAX) |
| 5209 | return -EINVAL; |
| 5210 | |
Mahesh Salgaonkar | cd75764 | 2010-01-30 10:25:18 +0530 | [diff] [blame] | 5211 | if (attr->__reserved_1) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5212 | return -EINVAL; |
| 5213 | |
| 5214 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) |
| 5215 | return -EINVAL; |
| 5216 | |
| 5217 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) |
| 5218 | return -EINVAL; |
| 5219 | |
| 5220 | out: |
| 5221 | return ret; |
| 5222 | |
| 5223 | err_size: |
| 5224 | put_user(sizeof(*attr), &uattr->size); |
| 5225 | ret = -E2BIG; |
| 5226 | goto out; |
| 5227 | } |
| 5228 | |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5229 | static int |
| 5230 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5231 | { |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 5232 | struct perf_buffer *buffer = NULL, *old_buffer = NULL; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5233 | int ret = -EINVAL; |
| 5234 | |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5235 | if (!output_event) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5236 | goto set; |
| 5237 | |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5238 | /* don't allow circular references */ |
| 5239 | if (event == output_event) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5240 | goto out; |
| 5241 | |
Peter Zijlstra | 0f13930 | 2010-05-20 14:35:15 +0200 | [diff] [blame] | 5242 | /* |
| 5243 | * Don't allow cross-cpu buffers |
| 5244 | */ |
| 5245 | if (output_event->cpu != event->cpu) |
| 5246 | goto out; |
| 5247 | |
| 5248 | /* |
| 5249 | * If its not a per-cpu buffer, it must be the same task. |
| 5250 | */ |
| 5251 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) |
| 5252 | goto out; |
| 5253 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5254 | set: |
| 5255 | mutex_lock(&event->mmap_mutex); |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5256 | /* Can't redirect output if we've got an active mmap() */ |
| 5257 | if (atomic_read(&event->mmap_count)) |
| 5258 | goto unlock; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5259 | |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5260 | if (output_event) { |
| 5261 | /* get the buffer we want to redirect to */ |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 5262 | buffer = perf_buffer_get(output_event); |
| 5263 | if (!buffer) |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5264 | goto unlock; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5265 | } |
| 5266 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 5267 | old_buffer = event->buffer; |
| 5268 | rcu_assign_pointer(event->buffer, buffer); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5269 | ret = 0; |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5270 | unlock: |
| 5271 | mutex_unlock(&event->mmap_mutex); |
| 5272 | |
Peter Zijlstra | ca5135e | 2010-05-28 19:33:23 +0200 | [diff] [blame] | 5273 | if (old_buffer) |
| 5274 | perf_buffer_put(old_buffer); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5275 | out: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5276 | return ret; |
| 5277 | } |
| 5278 | |
| 5279 | /** |
| 5280 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
| 5281 | * |
| 5282 | * @attr_uptr: event_id type attributes for monitoring/sampling |
| 5283 | * @pid: target pid |
| 5284 | * @cpu: target cpu |
| 5285 | * @group_fd: group leader event fd |
| 5286 | */ |
| 5287 | SYSCALL_DEFINE5(perf_event_open, |
| 5288 | struct perf_event_attr __user *, attr_uptr, |
| 5289 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
| 5290 | { |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5291 | struct perf_event *event, *group_leader = NULL, *output_event = NULL; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5292 | struct perf_event_attr attr; |
| 5293 | struct perf_event_context *ctx; |
| 5294 | struct file *event_file = NULL; |
| 5295 | struct file *group_file = NULL; |
Al Viro | ea635c6 | 2010-05-26 17:40:29 -0400 | [diff] [blame] | 5296 | int event_fd; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5297 | int fput_needed = 0; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5298 | int err; |
| 5299 | |
| 5300 | /* for future expandability... */ |
| 5301 | if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT)) |
| 5302 | return -EINVAL; |
| 5303 | |
| 5304 | err = perf_copy_attr(attr_uptr, &attr); |
| 5305 | if (err) |
| 5306 | return err; |
| 5307 | |
| 5308 | if (!attr.exclude_kernel) { |
| 5309 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
| 5310 | return -EACCES; |
| 5311 | } |
| 5312 | |
| 5313 | if (attr.freq) { |
| 5314 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
| 5315 | return -EINVAL; |
| 5316 | } |
| 5317 | |
Al Viro | ea635c6 | 2010-05-26 17:40:29 -0400 | [diff] [blame] | 5318 | event_fd = get_unused_fd_flags(O_RDWR); |
| 5319 | if (event_fd < 0) |
| 5320 | return event_fd; |
| 5321 | |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5322 | event = perf_event_alloc(&attr, cpu, group_leader, NULL, NULL); |
| 5323 | if (IS_ERR(event)) { |
| 5324 | err = PTR_ERR(event); |
| 5325 | goto err_fd; |
| 5326 | } |
| 5327 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5328 | /* |
| 5329 | * Get the target context (task or percpu): |
| 5330 | */ |
| 5331 | ctx = find_get_context(pid, cpu); |
Al Viro | ea635c6 | 2010-05-26 17:40:29 -0400 | [diff] [blame] | 5332 | if (IS_ERR(ctx)) { |
| 5333 | err = PTR_ERR(ctx); |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5334 | goto err_alloc; |
Al Viro | ea635c6 | 2010-05-26 17:40:29 -0400 | [diff] [blame] | 5335 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5336 | |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5337 | if (group_fd != -1) { |
| 5338 | group_leader = perf_fget_light(group_fd, &fput_needed); |
| 5339 | if (IS_ERR(group_leader)) { |
| 5340 | err = PTR_ERR(group_leader); |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5341 | goto err_context; |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5342 | } |
| 5343 | group_file = group_leader->filp; |
| 5344 | if (flags & PERF_FLAG_FD_OUTPUT) |
| 5345 | output_event = group_leader; |
| 5346 | if (flags & PERF_FLAG_FD_NO_GROUP) |
| 5347 | group_leader = NULL; |
| 5348 | } |
| 5349 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5350 | /* |
| 5351 | * Look up the group leader (we will attach this event to it): |
| 5352 | */ |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5353 | if (group_leader) { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5354 | err = -EINVAL; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5355 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5356 | /* |
| 5357 | * Do not allow a recursive hierarchy (this new sibling |
| 5358 | * becoming part of another group-sibling): |
| 5359 | */ |
| 5360 | if (group_leader->group_leader != group_leader) |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5361 | goto err_context; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5362 | /* |
| 5363 | * Do not allow to attach to a group in a different |
| 5364 | * task or CPU context: |
| 5365 | */ |
| 5366 | if (group_leader->ctx != ctx) |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5367 | goto err_context; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5368 | /* |
| 5369 | * Only a group leader can be exclusive or pinned |
| 5370 | */ |
| 5371 | if (attr.exclusive || attr.pinned) |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5372 | goto err_context; |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5373 | } |
| 5374 | |
| 5375 | if (output_event) { |
| 5376 | err = perf_event_set_output(event, output_event); |
| 5377 | if (err) |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5378 | goto err_context; |
Peter Zijlstra | ac9721f | 2010-05-27 12:54:41 +0200 | [diff] [blame] | 5379 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5380 | |
Al Viro | ea635c6 | 2010-05-26 17:40:29 -0400 | [diff] [blame] | 5381 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
| 5382 | if (IS_ERR(event_file)) { |
| 5383 | err = PTR_ERR(event_file); |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5384 | goto err_context; |
Al Viro | ea635c6 | 2010-05-26 17:40:29 -0400 | [diff] [blame] | 5385 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5386 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5387 | event->filp = event_file; |
| 5388 | WARN_ON_ONCE(ctx->parent_ctx); |
| 5389 | mutex_lock(&ctx->mutex); |
| 5390 | perf_install_in_context(ctx, event, cpu); |
| 5391 | ++ctx->generation; |
| 5392 | mutex_unlock(&ctx->mutex); |
| 5393 | |
| 5394 | event->owner = current; |
| 5395 | get_task_struct(current); |
| 5396 | mutex_lock(¤t->perf_event_mutex); |
| 5397 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); |
| 5398 | mutex_unlock(¤t->perf_event_mutex); |
| 5399 | |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 5400 | /* |
| 5401 | * Drop the reference on the group_event after placing the |
| 5402 | * new event on the sibling_list. This ensures destruction |
| 5403 | * of the group leader will find the pointer to itself in |
| 5404 | * perf_group_detach(). |
| 5405 | */ |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5406 | fput_light(group_file, fput_needed); |
Al Viro | ea635c6 | 2010-05-26 17:40:29 -0400 | [diff] [blame] | 5407 | fd_install(event_fd, event_file); |
| 5408 | return event_fd; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5409 | |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5410 | err_context: |
Al Viro | ea635c6 | 2010-05-26 17:40:29 -0400 | [diff] [blame] | 5411 | fput_light(group_file, fput_needed); |
| 5412 | put_ctx(ctx); |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5413 | err_alloc: |
| 5414 | free_event(event); |
Al Viro | ea635c6 | 2010-05-26 17:40:29 -0400 | [diff] [blame] | 5415 | err_fd: |
| 5416 | put_unused_fd(event_fd); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5417 | return err; |
| 5418 | } |
| 5419 | |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 5420 | /** |
| 5421 | * perf_event_create_kernel_counter |
| 5422 | * |
| 5423 | * @attr: attributes of the counter to create |
| 5424 | * @cpu: cpu in which the counter is bound |
| 5425 | * @pid: task to profile |
| 5426 | */ |
| 5427 | struct perf_event * |
| 5428 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, |
Frederic Weisbecker | b326e95 | 2009-12-05 09:44:31 +0100 | [diff] [blame] | 5429 | pid_t pid, |
| 5430 | perf_overflow_handler_t overflow_handler) |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 5431 | { |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 5432 | struct perf_event_context *ctx; |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5433 | struct perf_event *event; |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 5434 | int err; |
| 5435 | |
| 5436 | /* |
| 5437 | * Get the target context (task or percpu): |
| 5438 | */ |
| 5439 | |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5440 | event = perf_event_alloc(attr, cpu, NULL, NULL, overflow_handler); |
| 5441 | if (IS_ERR(event)) { |
| 5442 | err = PTR_ERR(event); |
| 5443 | goto err; |
| 5444 | } |
| 5445 | |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 5446 | ctx = find_get_context(pid, cpu); |
Frederic Weisbecker | c6567f6 | 2009-11-26 05:35:41 +0100 | [diff] [blame] | 5447 | if (IS_ERR(ctx)) { |
| 5448 | err = PTR_ERR(ctx); |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5449 | goto err_free; |
Frederic Weisbecker | c6567f6 | 2009-11-26 05:35:41 +0100 | [diff] [blame] | 5450 | } |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 5451 | |
| 5452 | event->filp = NULL; |
| 5453 | WARN_ON_ONCE(ctx->parent_ctx); |
| 5454 | mutex_lock(&ctx->mutex); |
| 5455 | perf_install_in_context(ctx, event, cpu); |
| 5456 | ++ctx->generation; |
| 5457 | mutex_unlock(&ctx->mutex); |
| 5458 | |
| 5459 | event->owner = current; |
| 5460 | get_task_struct(current); |
| 5461 | mutex_lock(¤t->perf_event_mutex); |
| 5462 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); |
| 5463 | mutex_unlock(¤t->perf_event_mutex); |
| 5464 | |
| 5465 | return event; |
| 5466 | |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5467 | err_free: |
| 5468 | free_event(event); |
| 5469 | err: |
Frederic Weisbecker | c6567f6 | 2009-11-26 05:35:41 +0100 | [diff] [blame] | 5470 | return ERR_PTR(err); |
Arjan van de Ven | fb0459d | 2009-09-25 12:25:56 +0200 | [diff] [blame] | 5471 | } |
| 5472 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
| 5473 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5474 | /* |
| 5475 | * inherit a event from parent task to child task: |
| 5476 | */ |
| 5477 | static struct perf_event * |
| 5478 | inherit_event(struct perf_event *parent_event, |
| 5479 | struct task_struct *parent, |
| 5480 | struct perf_event_context *parent_ctx, |
| 5481 | struct task_struct *child, |
| 5482 | struct perf_event *group_leader, |
| 5483 | struct perf_event_context *child_ctx) |
| 5484 | { |
| 5485 | struct perf_event *child_event; |
| 5486 | |
| 5487 | /* |
| 5488 | * Instead of creating recursive hierarchies of events, |
| 5489 | * we link inherited events back to the original parent, |
| 5490 | * which has a filp for sure, which we use as the reference |
| 5491 | * count: |
| 5492 | */ |
| 5493 | if (parent_event->parent) |
| 5494 | parent_event = parent_event->parent; |
| 5495 | |
| 5496 | child_event = perf_event_alloc(&parent_event->attr, |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5497 | parent_event->cpu, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5498 | group_leader, parent_event, |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5499 | NULL); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5500 | if (IS_ERR(child_event)) |
| 5501 | return child_event; |
| 5502 | get_ctx(child_ctx); |
| 5503 | |
| 5504 | /* |
| 5505 | * Make the child state follow the state of the parent event, |
| 5506 | * not its attr.disabled bit. We hold the parent's mutex, |
| 5507 | * so we won't race with perf_event_{en, dis}able_family. |
| 5508 | */ |
| 5509 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) |
| 5510 | child_event->state = PERF_EVENT_STATE_INACTIVE; |
| 5511 | else |
| 5512 | child_event->state = PERF_EVENT_STATE_OFF; |
| 5513 | |
Peter Zijlstra | 75c9f32 | 2010-01-29 09:04:26 +0100 | [diff] [blame] | 5514 | if (parent_event->attr.freq) { |
| 5515 | u64 sample_period = parent_event->hw.sample_period; |
| 5516 | struct hw_perf_event *hwc = &child_event->hw; |
| 5517 | |
| 5518 | hwc->sample_period = sample_period; |
| 5519 | hwc->last_period = sample_period; |
| 5520 | |
Peter Zijlstra | e785059 | 2010-05-21 14:43:08 +0200 | [diff] [blame] | 5521 | local64_set(&hwc->period_left, sample_period); |
Peter Zijlstra | 75c9f32 | 2010-01-29 09:04:26 +0100 | [diff] [blame] | 5522 | } |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5523 | |
Peter Zijlstra | c3f00c7 | 2010-08-18 14:37:15 +0200 | [diff] [blame^] | 5524 | child_event->ctx = child_ctx; |
Peter Zijlstra | 453f19e | 2009-11-20 22:19:43 +0100 | [diff] [blame] | 5525 | child_event->overflow_handler = parent_event->overflow_handler; |
| 5526 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5527 | /* |
| 5528 | * Link it up in the child's context: |
| 5529 | */ |
| 5530 | add_event_to_ctx(child_event, child_ctx); |
| 5531 | |
| 5532 | /* |
| 5533 | * Get a reference to the parent filp - we will fput it |
| 5534 | * when the child event exits. This is safe to do because |
| 5535 | * we are in the parent and we know that the filp still |
| 5536 | * exists and has a nonzero count: |
| 5537 | */ |
| 5538 | atomic_long_inc(&parent_event->filp->f_count); |
| 5539 | |
| 5540 | /* |
| 5541 | * Link this into the parent event's child list |
| 5542 | */ |
| 5543 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
| 5544 | mutex_lock(&parent_event->child_mutex); |
| 5545 | list_add_tail(&child_event->child_list, &parent_event->child_list); |
| 5546 | mutex_unlock(&parent_event->child_mutex); |
| 5547 | |
| 5548 | return child_event; |
| 5549 | } |
| 5550 | |
| 5551 | static int inherit_group(struct perf_event *parent_event, |
| 5552 | struct task_struct *parent, |
| 5553 | struct perf_event_context *parent_ctx, |
| 5554 | struct task_struct *child, |
| 5555 | struct perf_event_context *child_ctx) |
| 5556 | { |
| 5557 | struct perf_event *leader; |
| 5558 | struct perf_event *sub; |
| 5559 | struct perf_event *child_ctr; |
| 5560 | |
| 5561 | leader = inherit_event(parent_event, parent, parent_ctx, |
| 5562 | child, NULL, child_ctx); |
| 5563 | if (IS_ERR(leader)) |
| 5564 | return PTR_ERR(leader); |
| 5565 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { |
| 5566 | child_ctr = inherit_event(sub, parent, parent_ctx, |
| 5567 | child, leader, child_ctx); |
| 5568 | if (IS_ERR(child_ctr)) |
| 5569 | return PTR_ERR(child_ctr); |
| 5570 | } |
| 5571 | return 0; |
| 5572 | } |
| 5573 | |
| 5574 | static void sync_child_event(struct perf_event *child_event, |
| 5575 | struct task_struct *child) |
| 5576 | { |
| 5577 | struct perf_event *parent_event = child_event->parent; |
| 5578 | u64 child_val; |
| 5579 | |
| 5580 | if (child_event->attr.inherit_stat) |
| 5581 | perf_event_read_event(child_event, child); |
| 5582 | |
Peter Zijlstra | b5e5879 | 2010-05-21 14:43:12 +0200 | [diff] [blame] | 5583 | child_val = perf_event_count(child_event); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5584 | |
| 5585 | /* |
| 5586 | * Add back the child's count to the parent's count: |
| 5587 | */ |
Peter Zijlstra | a6e6dea | 2010-05-21 14:27:58 +0200 | [diff] [blame] | 5588 | atomic64_add(child_val, &parent_event->child_count); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5589 | atomic64_add(child_event->total_time_enabled, |
| 5590 | &parent_event->child_total_time_enabled); |
| 5591 | atomic64_add(child_event->total_time_running, |
| 5592 | &parent_event->child_total_time_running); |
| 5593 | |
| 5594 | /* |
| 5595 | * Remove this event from the parent's list |
| 5596 | */ |
| 5597 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
| 5598 | mutex_lock(&parent_event->child_mutex); |
| 5599 | list_del_init(&child_event->child_list); |
| 5600 | mutex_unlock(&parent_event->child_mutex); |
| 5601 | |
| 5602 | /* |
| 5603 | * Release the parent event, if this was the last |
| 5604 | * reference to it. |
| 5605 | */ |
| 5606 | fput(parent_event->filp); |
| 5607 | } |
| 5608 | |
| 5609 | static void |
| 5610 | __perf_event_exit_task(struct perf_event *child_event, |
| 5611 | struct perf_event_context *child_ctx, |
| 5612 | struct task_struct *child) |
| 5613 | { |
| 5614 | struct perf_event *parent_event; |
| 5615 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5616 | perf_event_remove_from_context(child_event); |
| 5617 | |
| 5618 | parent_event = child_event->parent; |
| 5619 | /* |
| 5620 | * It can happen that parent exits first, and has events |
| 5621 | * that are still around due to the child reference. These |
| 5622 | * events need to be zapped - but otherwise linger. |
| 5623 | */ |
| 5624 | if (parent_event) { |
| 5625 | sync_child_event(child_event, child); |
| 5626 | free_event(child_event); |
| 5627 | } |
| 5628 | } |
| 5629 | |
| 5630 | /* |
| 5631 | * When a child task exits, feed back event values to parent events. |
| 5632 | */ |
| 5633 | void perf_event_exit_task(struct task_struct *child) |
| 5634 | { |
| 5635 | struct perf_event *child_event, *tmp; |
| 5636 | struct perf_event_context *child_ctx; |
| 5637 | unsigned long flags; |
| 5638 | |
| 5639 | if (likely(!child->perf_event_ctxp)) { |
| 5640 | perf_event_task(child, NULL, 0); |
| 5641 | return; |
| 5642 | } |
| 5643 | |
| 5644 | local_irq_save(flags); |
| 5645 | /* |
| 5646 | * We can't reschedule here because interrupts are disabled, |
| 5647 | * and either child is current or it is a task that can't be |
| 5648 | * scheduled, so we are now safe from rescheduling changing |
| 5649 | * our context. |
| 5650 | */ |
| 5651 | child_ctx = child->perf_event_ctxp; |
| 5652 | __perf_event_task_sched_out(child_ctx); |
| 5653 | |
| 5654 | /* |
| 5655 | * Take the context lock here so that if find_get_context is |
| 5656 | * reading child->perf_event_ctxp, we wait until it has |
| 5657 | * incremented the context's refcount before we do put_ctx below. |
| 5658 | */ |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 5659 | raw_spin_lock(&child_ctx->lock); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5660 | child->perf_event_ctxp = NULL; |
| 5661 | /* |
| 5662 | * If this context is a clone; unclone it so it can't get |
| 5663 | * swapped to another process while we're removing all |
| 5664 | * the events from it. |
| 5665 | */ |
| 5666 | unclone_ctx(child_ctx); |
Peter Zijlstra | 5e942bb | 2009-11-23 11:37:26 +0100 | [diff] [blame] | 5667 | update_context_time(child_ctx); |
Thomas Gleixner | e625cce1 | 2009-11-17 18:02:06 +0100 | [diff] [blame] | 5668 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5669 | |
| 5670 | /* |
| 5671 | * Report the task dead after unscheduling the events so that we |
| 5672 | * won't get any samples after PERF_RECORD_EXIT. We can however still |
| 5673 | * get a few PERF_RECORD_READ events. |
| 5674 | */ |
| 5675 | perf_event_task(child, child_ctx, 0); |
| 5676 | |
| 5677 | /* |
| 5678 | * We can recurse on the same lock type through: |
| 5679 | * |
| 5680 | * __perf_event_exit_task() |
| 5681 | * sync_child_event() |
| 5682 | * fput(parent_event->filp) |
| 5683 | * perf_release() |
| 5684 | * mutex_lock(&ctx->mutex) |
| 5685 | * |
| 5686 | * But since its the parent context it won't be the same instance. |
| 5687 | */ |
Peter Zijlstra | a0507c8 | 2010-05-06 15:42:53 +0200 | [diff] [blame] | 5688 | mutex_lock(&child_ctx->mutex); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5689 | |
| 5690 | again: |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5691 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
| 5692 | group_entry) |
| 5693 | __perf_event_exit_task(child_event, child_ctx, child); |
| 5694 | |
| 5695 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5696 | group_entry) |
| 5697 | __perf_event_exit_task(child_event, child_ctx, child); |
| 5698 | |
| 5699 | /* |
| 5700 | * If the last event was a group event, it will have appended all |
| 5701 | * its siblings to the list, but we obtained 'tmp' before that which |
| 5702 | * will still point to the list head terminating the iteration. |
| 5703 | */ |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5704 | if (!list_empty(&child_ctx->pinned_groups) || |
| 5705 | !list_empty(&child_ctx->flexible_groups)) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5706 | goto again; |
| 5707 | |
| 5708 | mutex_unlock(&child_ctx->mutex); |
| 5709 | |
| 5710 | put_ctx(child_ctx); |
| 5711 | } |
| 5712 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5713 | static void perf_free_event(struct perf_event *event, |
| 5714 | struct perf_event_context *ctx) |
| 5715 | { |
| 5716 | struct perf_event *parent = event->parent; |
| 5717 | |
| 5718 | if (WARN_ON_ONCE(!parent)) |
| 5719 | return; |
| 5720 | |
| 5721 | mutex_lock(&parent->child_mutex); |
| 5722 | list_del_init(&event->child_list); |
| 5723 | mutex_unlock(&parent->child_mutex); |
| 5724 | |
| 5725 | fput(parent->filp); |
| 5726 | |
Peter Zijlstra | 8a49542 | 2010-05-27 15:47:49 +0200 | [diff] [blame] | 5727 | perf_group_detach(event); |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5728 | list_del_event(event, ctx); |
| 5729 | free_event(event); |
| 5730 | } |
| 5731 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5732 | /* |
| 5733 | * free an unexposed, unused context as created by inheritance by |
| 5734 | * init_task below, used by fork() in case of fail. |
| 5735 | */ |
| 5736 | void perf_event_free_task(struct task_struct *task) |
| 5737 | { |
| 5738 | struct perf_event_context *ctx = task->perf_event_ctxp; |
| 5739 | struct perf_event *event, *tmp; |
| 5740 | |
| 5741 | if (!ctx) |
| 5742 | return; |
| 5743 | |
| 5744 | mutex_lock(&ctx->mutex); |
| 5745 | again: |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5746 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
| 5747 | perf_free_event(event, ctx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5748 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5749 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
| 5750 | group_entry) |
| 5751 | perf_free_event(event, ctx); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5752 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5753 | if (!list_empty(&ctx->pinned_groups) || |
| 5754 | !list_empty(&ctx->flexible_groups)) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5755 | goto again; |
| 5756 | |
| 5757 | mutex_unlock(&ctx->mutex); |
| 5758 | |
| 5759 | put_ctx(ctx); |
| 5760 | } |
| 5761 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5762 | static int |
| 5763 | inherit_task_group(struct perf_event *event, struct task_struct *parent, |
| 5764 | struct perf_event_context *parent_ctx, |
| 5765 | struct task_struct *child, |
| 5766 | int *inherited_all) |
| 5767 | { |
| 5768 | int ret; |
| 5769 | struct perf_event_context *child_ctx = child->perf_event_ctxp; |
| 5770 | |
| 5771 | if (!event->attr.inherit) { |
| 5772 | *inherited_all = 0; |
| 5773 | return 0; |
| 5774 | } |
| 5775 | |
| 5776 | if (!child_ctx) { |
| 5777 | /* |
| 5778 | * This is executed from the parent task context, so |
| 5779 | * inherit events that have been marked for cloning. |
| 5780 | * First allocate and initialize a context for the |
| 5781 | * child. |
| 5782 | */ |
| 5783 | |
| 5784 | child_ctx = kzalloc(sizeof(struct perf_event_context), |
| 5785 | GFP_KERNEL); |
| 5786 | if (!child_ctx) |
| 5787 | return -ENOMEM; |
| 5788 | |
| 5789 | __perf_event_init_context(child_ctx, child); |
| 5790 | child->perf_event_ctxp = child_ctx; |
| 5791 | get_task_struct(child); |
| 5792 | } |
| 5793 | |
| 5794 | ret = inherit_group(event, parent, parent_ctx, |
| 5795 | child, child_ctx); |
| 5796 | |
| 5797 | if (ret) |
| 5798 | *inherited_all = 0; |
| 5799 | |
| 5800 | return ret; |
| 5801 | } |
| 5802 | |
| 5803 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5804 | /* |
| 5805 | * Initialize the perf_event context in task_struct |
| 5806 | */ |
| 5807 | int perf_event_init_task(struct task_struct *child) |
| 5808 | { |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5809 | struct perf_event_context *child_ctx, *parent_ctx; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5810 | struct perf_event_context *cloned_ctx; |
| 5811 | struct perf_event *event; |
| 5812 | struct task_struct *parent = current; |
| 5813 | int inherited_all = 1; |
| 5814 | int ret = 0; |
| 5815 | |
| 5816 | child->perf_event_ctxp = NULL; |
| 5817 | |
| 5818 | mutex_init(&child->perf_event_mutex); |
| 5819 | INIT_LIST_HEAD(&child->perf_event_list); |
| 5820 | |
| 5821 | if (likely(!parent->perf_event_ctxp)) |
| 5822 | return 0; |
| 5823 | |
| 5824 | /* |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5825 | * If the parent's context is a clone, pin it so it won't get |
| 5826 | * swapped under us. |
| 5827 | */ |
| 5828 | parent_ctx = perf_pin_task_context(parent); |
| 5829 | |
| 5830 | /* |
| 5831 | * No need to check if parent_ctx != NULL here; since we saw |
| 5832 | * it non-NULL earlier, the only reason for it to become NULL |
| 5833 | * is if we exit, and since we're currently in the middle of |
| 5834 | * a fork we can't be exiting at the same time. |
| 5835 | */ |
| 5836 | |
| 5837 | /* |
| 5838 | * Lock the parent list. No need to lock the child - not PID |
| 5839 | * hashed yet and not running, so nobody can access it. |
| 5840 | */ |
| 5841 | mutex_lock(&parent_ctx->mutex); |
| 5842 | |
| 5843 | /* |
| 5844 | * We dont have to disable NMIs - we are only looking at |
| 5845 | * the list, not manipulating it: |
| 5846 | */ |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5847 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
| 5848 | ret = inherit_task_group(event, parent, parent_ctx, child, |
| 5849 | &inherited_all); |
| 5850 | if (ret) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5851 | break; |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5852 | } |
| 5853 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5854 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
| 5855 | ret = inherit_task_group(event, parent, parent_ctx, child, |
| 5856 | &inherited_all); |
| 5857 | if (ret) |
| 5858 | break; |
| 5859 | } |
| 5860 | |
| 5861 | child_ctx = child->perf_event_ctxp; |
| 5862 | |
Peter Zijlstra | 05cbaa2 | 2009-12-30 16:00:35 +0100 | [diff] [blame] | 5863 | if (child_ctx && inherited_all) { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5864 | /* |
| 5865 | * Mark the child context as a clone of the parent |
| 5866 | * context, or of whatever the parent is a clone of. |
| 5867 | * Note that if the parent is a clone, it could get |
| 5868 | * uncloned at any point, but that doesn't matter |
| 5869 | * because the list of events and the generation |
| 5870 | * count can't have changed since we took the mutex. |
| 5871 | */ |
| 5872 | cloned_ctx = rcu_dereference(parent_ctx->parent_ctx); |
| 5873 | if (cloned_ctx) { |
| 5874 | child_ctx->parent_ctx = cloned_ctx; |
| 5875 | child_ctx->parent_gen = parent_ctx->parent_gen; |
| 5876 | } else { |
| 5877 | child_ctx->parent_ctx = parent_ctx; |
| 5878 | child_ctx->parent_gen = parent_ctx->generation; |
| 5879 | } |
| 5880 | get_ctx(child_ctx->parent_ctx); |
| 5881 | } |
| 5882 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5883 | mutex_unlock(&parent_ctx->mutex); |
| 5884 | |
| 5885 | perf_unpin_context(parent_ctx); |
| 5886 | |
| 5887 | return ret; |
| 5888 | } |
| 5889 | |
Paul Mackerras | 220b140 | 2010-03-10 20:45:52 +1100 | [diff] [blame] | 5890 | static void __init perf_event_init_all_cpus(void) |
| 5891 | { |
| 5892 | int cpu; |
| 5893 | struct perf_cpu_context *cpuctx; |
| 5894 | |
| 5895 | for_each_possible_cpu(cpu) { |
| 5896 | cpuctx = &per_cpu(perf_cpu_context, cpu); |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 5897 | mutex_init(&cpuctx->hlist_mutex); |
Paul Mackerras | 220b140 | 2010-03-10 20:45:52 +1100 | [diff] [blame] | 5898 | __perf_event_init_context(&cpuctx->ctx, NULL); |
| 5899 | } |
| 5900 | } |
| 5901 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5902 | static void __cpuinit perf_event_init_cpu(int cpu) |
| 5903 | { |
| 5904 | struct perf_cpu_context *cpuctx; |
| 5905 | |
| 5906 | cpuctx = &per_cpu(perf_cpu_context, cpu); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5907 | |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 5908 | mutex_lock(&cpuctx->hlist_mutex); |
| 5909 | if (cpuctx->hlist_refcount > 0) { |
| 5910 | struct swevent_hlist *hlist; |
| 5911 | |
| 5912 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); |
| 5913 | WARN_ON_ONCE(!hlist); |
| 5914 | rcu_assign_pointer(cpuctx->swevent_hlist, hlist); |
| 5915 | } |
| 5916 | mutex_unlock(&cpuctx->hlist_mutex); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5917 | } |
| 5918 | |
| 5919 | #ifdef CONFIG_HOTPLUG_CPU |
| 5920 | static void __perf_event_exit_cpu(void *info) |
| 5921 | { |
| 5922 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); |
| 5923 | struct perf_event_context *ctx = &cpuctx->ctx; |
| 5924 | struct perf_event *event, *tmp; |
| 5925 | |
Frederic Weisbecker | 889ff01 | 2010-01-09 20:04:47 +0100 | [diff] [blame] | 5926 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
| 5927 | __perf_event_remove_from_context(event); |
| 5928 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5929 | __perf_event_remove_from_context(event); |
| 5930 | } |
| 5931 | static void perf_event_exit_cpu(int cpu) |
| 5932 | { |
| 5933 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); |
| 5934 | struct perf_event_context *ctx = &cpuctx->ctx; |
| 5935 | |
Frederic Weisbecker | 76e1d90 | 2010-04-05 15:35:57 +0200 | [diff] [blame] | 5936 | mutex_lock(&cpuctx->hlist_mutex); |
| 5937 | swevent_hlist_release(cpuctx); |
| 5938 | mutex_unlock(&cpuctx->hlist_mutex); |
| 5939 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5940 | mutex_lock(&ctx->mutex); |
| 5941 | smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1); |
| 5942 | mutex_unlock(&ctx->mutex); |
| 5943 | } |
| 5944 | #else |
| 5945 | static inline void perf_event_exit_cpu(int cpu) { } |
| 5946 | #endif |
| 5947 | |
| 5948 | static int __cpuinit |
| 5949 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
| 5950 | { |
| 5951 | unsigned int cpu = (long)hcpu; |
| 5952 | |
Peter Zijlstra | 5e11637 | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5953 | switch (action & ~CPU_TASKS_FROZEN) { |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5954 | |
| 5955 | case CPU_UP_PREPARE: |
Peter Zijlstra | 5e11637 | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5956 | case CPU_DOWN_FAILED: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5957 | perf_event_init_cpu(cpu); |
| 5958 | break; |
| 5959 | |
Peter Zijlstra | 5e11637 | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5960 | case CPU_UP_CANCELED: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5961 | case CPU_DOWN_PREPARE: |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5962 | perf_event_exit_cpu(cpu); |
| 5963 | break; |
| 5964 | |
| 5965 | default: |
| 5966 | break; |
| 5967 | } |
| 5968 | |
| 5969 | return NOTIFY_OK; |
| 5970 | } |
| 5971 | |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5972 | void __init perf_event_init(void) |
| 5973 | { |
Paul Mackerras | 220b140 | 2010-03-10 20:45:52 +1100 | [diff] [blame] | 5974 | perf_event_init_all_cpus(); |
Peter Zijlstra | b0a873e | 2010-06-11 13:35:08 +0200 | [diff] [blame] | 5975 | init_srcu_struct(&pmus_srcu); |
| 5976 | perf_pmu_register(&perf_swevent); |
| 5977 | perf_pmu_register(&perf_cpu_clock); |
| 5978 | perf_pmu_register(&perf_task_clock); |
| 5979 | perf_tp_register(); |
| 5980 | perf_cpu_notifier(perf_cpu_notify); |
Ingo Molnar | cdd6c48 | 2009-09-21 12:02:48 +0200 | [diff] [blame] | 5981 | } |