| 1 | /** |
| 2 | * @file buffer_sync.c |
| 3 | * |
| 4 | * @remark Copyright 2002-2009 OProfile authors |
| 5 | * @remark Read the file COPYING |
| 6 | * |
| 7 | * @author John Levon <levon@movementarian.org> |
| 8 | * @author Barry Kasindorf |
| 9 | * @author Robert Richter <robert.richter@amd.com> |
| 10 | * |
| 11 | * This is the core of the buffer management. Each |
| 12 | * CPU buffer is processed and entered into the |
| 13 | * global event buffer. Such processing is necessary |
| 14 | * in several circumstances, mentioned below. |
| 15 | * |
| 16 | * The processing does the job of converting the |
| 17 | * transitory EIP value into a persistent dentry/offset |
| 18 | * value that the profiler can record at its leisure. |
| 19 | * |
| 20 | * See fs/dcookies.c for a description of the dentry/offset |
| 21 | * objects. |
| 22 | */ |
| 23 | |
| 24 | #include <linux/mm.h> |
| 25 | #include <linux/workqueue.h> |
| 26 | #include <linux/notifier.h> |
| 27 | #include <linux/dcookies.h> |
| 28 | #include <linux/profile.h> |
| 29 | #include <linux/module.h> |
| 30 | #include <linux/fs.h> |
| 31 | #include <linux/oprofile.h> |
| 32 | #include <linux/sched.h> |
| 33 | |
| 34 | #include "oprofile_stats.h" |
| 35 | #include "event_buffer.h" |
| 36 | #include "cpu_buffer.h" |
| 37 | #include "buffer_sync.h" |
| 38 | |
| 39 | static LIST_HEAD(dying_tasks); |
| 40 | static LIST_HEAD(dead_tasks); |
| 41 | static cpumask_var_t marked_cpus; |
| 42 | static DEFINE_SPINLOCK(task_mortuary); |
| 43 | static void process_task_mortuary(void); |
| 44 | |
| 45 | /* Take ownership of the task struct and place it on the |
| 46 | * list for processing. Only after two full buffer syncs |
| 47 | * does the task eventually get freed, because by then |
| 48 | * we are sure we will not reference it again. |
| 49 | * Can be invoked from softirq via RCU callback due to |
| 50 | * call_rcu() of the task struct, hence the _irqsave. |
| 51 | */ |
| 52 | static int |
| 53 | task_free_notify(struct notifier_block *self, unsigned long val, void *data) |
| 54 | { |
| 55 | unsigned long flags; |
| 56 | struct task_struct *task = data; |
| 57 | spin_lock_irqsave(&task_mortuary, flags); |
| 58 | list_add(&task->tasks, &dying_tasks); |
| 59 | spin_unlock_irqrestore(&task_mortuary, flags); |
| 60 | return NOTIFY_OK; |
| 61 | } |
| 62 | |
| 63 | |
| 64 | /* The task is on its way out. A sync of the buffer means we can catch |
| 65 | * any remaining samples for this task. |
| 66 | */ |
| 67 | static int |
| 68 | task_exit_notify(struct notifier_block *self, unsigned long val, void *data) |
| 69 | { |
| 70 | /* To avoid latency problems, we only process the current CPU, |
| 71 | * hoping that most samples for the task are on this CPU |
| 72 | */ |
| 73 | sync_buffer(raw_smp_processor_id()); |
| 74 | return 0; |
| 75 | } |
| 76 | |
| 77 | |
| 78 | /* The task is about to try a do_munmap(). We peek at what it's going to |
| 79 | * do, and if it's an executable region, process the samples first, so |
| 80 | * we don't lose any. This does not have to be exact, it's a QoI issue |
| 81 | * only. |
| 82 | */ |
| 83 | static int |
| 84 | munmap_notify(struct notifier_block *self, unsigned long val, void *data) |
| 85 | { |
| 86 | unsigned long addr = (unsigned long)data; |
| 87 | struct mm_struct *mm = current->mm; |
| 88 | struct vm_area_struct *mpnt; |
| 89 | |
| 90 | down_read(&mm->mmap_sem); |
| 91 | |
| 92 | mpnt = find_vma(mm, addr); |
| 93 | if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) { |
| 94 | up_read(&mm->mmap_sem); |
| 95 | /* To avoid latency problems, we only process the current CPU, |
| 96 | * hoping that most samples for the task are on this CPU |
| 97 | */ |
| 98 | sync_buffer(raw_smp_processor_id()); |
| 99 | return 0; |
| 100 | } |
| 101 | |
| 102 | up_read(&mm->mmap_sem); |
| 103 | return 0; |
| 104 | } |
| 105 | |
| 106 | |
| 107 | /* We need to be told about new modules so we don't attribute to a previously |
| 108 | * loaded module, or drop the samples on the floor. |
| 109 | */ |
| 110 | static int |
| 111 | module_load_notify(struct notifier_block *self, unsigned long val, void *data) |
| 112 | { |
| 113 | #ifdef CONFIG_MODULES |
| 114 | if (val != MODULE_STATE_COMING) |
| 115 | return 0; |
| 116 | |
| 117 | /* FIXME: should we process all CPU buffers ? */ |
| 118 | mutex_lock(&buffer_mutex); |
| 119 | add_event_entry(ESCAPE_CODE); |
| 120 | add_event_entry(MODULE_LOADED_CODE); |
| 121 | mutex_unlock(&buffer_mutex); |
| 122 | #endif |
| 123 | return 0; |
| 124 | } |
| 125 | |
| 126 | |
| 127 | static struct notifier_block task_free_nb = { |
| 128 | .notifier_call = task_free_notify, |
| 129 | }; |
| 130 | |
| 131 | static struct notifier_block task_exit_nb = { |
| 132 | .notifier_call = task_exit_notify, |
| 133 | }; |
| 134 | |
| 135 | static struct notifier_block munmap_nb = { |
| 136 | .notifier_call = munmap_notify, |
| 137 | }; |
| 138 | |
| 139 | static struct notifier_block module_load_nb = { |
| 140 | .notifier_call = module_load_notify, |
| 141 | }; |
| 142 | |
| 143 | |
| 144 | static void end_sync(void) |
| 145 | { |
| 146 | end_cpu_work(); |
| 147 | /* make sure we don't leak task structs */ |
| 148 | process_task_mortuary(); |
| 149 | process_task_mortuary(); |
| 150 | } |
| 151 | |
| 152 | |
| 153 | int sync_start(void) |
| 154 | { |
| 155 | int err; |
| 156 | |
| 157 | if (!zalloc_cpumask_var(&marked_cpus, GFP_KERNEL)) |
| 158 | return -ENOMEM; |
| 159 | |
| 160 | start_cpu_work(); |
| 161 | |
| 162 | err = task_handoff_register(&task_free_nb); |
| 163 | if (err) |
| 164 | goto out1; |
| 165 | err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb); |
| 166 | if (err) |
| 167 | goto out2; |
| 168 | err = profile_event_register(PROFILE_MUNMAP, &munmap_nb); |
| 169 | if (err) |
| 170 | goto out3; |
| 171 | err = register_module_notifier(&module_load_nb); |
| 172 | if (err) |
| 173 | goto out4; |
| 174 | |
| 175 | out: |
| 176 | return err; |
| 177 | out4: |
| 178 | profile_event_unregister(PROFILE_MUNMAP, &munmap_nb); |
| 179 | out3: |
| 180 | profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb); |
| 181 | out2: |
| 182 | task_handoff_unregister(&task_free_nb); |
| 183 | out1: |
| 184 | end_sync(); |
| 185 | free_cpumask_var(marked_cpus); |
| 186 | goto out; |
| 187 | } |
| 188 | |
| 189 | |
| 190 | void sync_stop(void) |
| 191 | { |
| 192 | unregister_module_notifier(&module_load_nb); |
| 193 | profile_event_unregister(PROFILE_MUNMAP, &munmap_nb); |
| 194 | profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb); |
| 195 | task_handoff_unregister(&task_free_nb); |
| 196 | end_sync(); |
| 197 | free_cpumask_var(marked_cpus); |
| 198 | } |
| 199 | |
| 200 | |
| 201 | /* Optimisation. We can manage without taking the dcookie sem |
| 202 | * because we cannot reach this code without at least one |
| 203 | * dcookie user still being registered (namely, the reader |
| 204 | * of the event buffer). */ |
| 205 | static inline unsigned long fast_get_dcookie(struct path *path) |
| 206 | { |
| 207 | unsigned long cookie; |
| 208 | |
| 209 | if (path->dentry->d_flags & DCACHE_COOKIE) |
| 210 | return (unsigned long)path->dentry; |
| 211 | get_dcookie(path, &cookie); |
| 212 | return cookie; |
| 213 | } |
| 214 | |
| 215 | |
| 216 | /* Look up the dcookie for the task's first VM_EXECUTABLE mapping, |
| 217 | * which corresponds loosely to "application name". This is |
| 218 | * not strictly necessary but allows oprofile to associate |
| 219 | * shared-library samples with particular applications |
| 220 | */ |
| 221 | static unsigned long get_exec_dcookie(struct mm_struct *mm) |
| 222 | { |
| 223 | unsigned long cookie = NO_COOKIE; |
| 224 | struct vm_area_struct *vma; |
| 225 | |
| 226 | if (!mm) |
| 227 | goto out; |
| 228 | |
| 229 | for (vma = mm->mmap; vma; vma = vma->vm_next) { |
| 230 | if (!vma->vm_file) |
| 231 | continue; |
| 232 | if (!(vma->vm_flags & VM_EXECUTABLE)) |
| 233 | continue; |
| 234 | cookie = fast_get_dcookie(&vma->vm_file->f_path); |
| 235 | break; |
| 236 | } |
| 237 | |
| 238 | out: |
| 239 | return cookie; |
| 240 | } |
| 241 | |
| 242 | |
| 243 | /* Convert the EIP value of a sample into a persistent dentry/offset |
| 244 | * pair that can then be added to the global event buffer. We make |
| 245 | * sure to do this lookup before a mm->mmap modification happens so |
| 246 | * we don't lose track. |
| 247 | */ |
| 248 | static unsigned long |
| 249 | lookup_dcookie(struct mm_struct *mm, unsigned long addr, off_t *offset) |
| 250 | { |
| 251 | unsigned long cookie = NO_COOKIE; |
| 252 | struct vm_area_struct *vma; |
| 253 | |
| 254 | for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) { |
| 255 | |
| 256 | if (addr < vma->vm_start || addr >= vma->vm_end) |
| 257 | continue; |
| 258 | |
| 259 | if (vma->vm_file) { |
| 260 | cookie = fast_get_dcookie(&vma->vm_file->f_path); |
| 261 | *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr - |
| 262 | vma->vm_start; |
| 263 | } else { |
| 264 | /* must be an anonymous map */ |
| 265 | *offset = addr; |
| 266 | } |
| 267 | |
| 268 | break; |
| 269 | } |
| 270 | |
| 271 | if (!vma) |
| 272 | cookie = INVALID_COOKIE; |
| 273 | |
| 274 | return cookie; |
| 275 | } |
| 276 | |
| 277 | static unsigned long last_cookie = INVALID_COOKIE; |
| 278 | |
| 279 | static void add_cpu_switch(int i) |
| 280 | { |
| 281 | add_event_entry(ESCAPE_CODE); |
| 282 | add_event_entry(CPU_SWITCH_CODE); |
| 283 | add_event_entry(i); |
| 284 | last_cookie = INVALID_COOKIE; |
| 285 | } |
| 286 | |
| 287 | static void add_kernel_ctx_switch(unsigned int in_kernel) |
| 288 | { |
| 289 | add_event_entry(ESCAPE_CODE); |
| 290 | if (in_kernel) |
| 291 | add_event_entry(KERNEL_ENTER_SWITCH_CODE); |
| 292 | else |
| 293 | add_event_entry(KERNEL_EXIT_SWITCH_CODE); |
| 294 | } |
| 295 | |
| 296 | static void |
| 297 | add_user_ctx_switch(struct task_struct const *task, unsigned long cookie) |
| 298 | { |
| 299 | add_event_entry(ESCAPE_CODE); |
| 300 | add_event_entry(CTX_SWITCH_CODE); |
| 301 | add_event_entry(task->pid); |
| 302 | add_event_entry(cookie); |
| 303 | /* Another code for daemon back-compat */ |
| 304 | add_event_entry(ESCAPE_CODE); |
| 305 | add_event_entry(CTX_TGID_CODE); |
| 306 | add_event_entry(task->tgid); |
| 307 | } |
| 308 | |
| 309 | |
| 310 | static void add_cookie_switch(unsigned long cookie) |
| 311 | { |
| 312 | add_event_entry(ESCAPE_CODE); |
| 313 | add_event_entry(COOKIE_SWITCH_CODE); |
| 314 | add_event_entry(cookie); |
| 315 | } |
| 316 | |
| 317 | |
| 318 | static void add_trace_begin(void) |
| 319 | { |
| 320 | add_event_entry(ESCAPE_CODE); |
| 321 | add_event_entry(TRACE_BEGIN_CODE); |
| 322 | } |
| 323 | |
| 324 | static void add_data(struct op_entry *entry, struct mm_struct *mm) |
| 325 | { |
| 326 | unsigned long code, pc, val; |
| 327 | unsigned long cookie; |
| 328 | off_t offset; |
| 329 | |
| 330 | if (!op_cpu_buffer_get_data(entry, &code)) |
| 331 | return; |
| 332 | if (!op_cpu_buffer_get_data(entry, &pc)) |
| 333 | return; |
| 334 | if (!op_cpu_buffer_get_size(entry)) |
| 335 | return; |
| 336 | |
| 337 | if (mm) { |
| 338 | cookie = lookup_dcookie(mm, pc, &offset); |
| 339 | |
| 340 | if (cookie == NO_COOKIE) |
| 341 | offset = pc; |
| 342 | if (cookie == INVALID_COOKIE) { |
| 343 | atomic_inc(&oprofile_stats.sample_lost_no_mapping); |
| 344 | offset = pc; |
| 345 | } |
| 346 | if (cookie != last_cookie) { |
| 347 | add_cookie_switch(cookie); |
| 348 | last_cookie = cookie; |
| 349 | } |
| 350 | } else |
| 351 | offset = pc; |
| 352 | |
| 353 | add_event_entry(ESCAPE_CODE); |
| 354 | add_event_entry(code); |
| 355 | add_event_entry(offset); /* Offset from Dcookie */ |
| 356 | |
| 357 | while (op_cpu_buffer_get_data(entry, &val)) |
| 358 | add_event_entry(val); |
| 359 | } |
| 360 | |
| 361 | static inline void add_sample_entry(unsigned long offset, unsigned long event) |
| 362 | { |
| 363 | add_event_entry(offset); |
| 364 | add_event_entry(event); |
| 365 | } |
| 366 | |
| 367 | |
| 368 | /* |
| 369 | * Add a sample to the global event buffer. If possible the |
| 370 | * sample is converted into a persistent dentry/offset pair |
| 371 | * for later lookup from userspace. Return 0 on failure. |
| 372 | */ |
| 373 | static int |
| 374 | add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel) |
| 375 | { |
| 376 | unsigned long cookie; |
| 377 | off_t offset; |
| 378 | |
| 379 | if (in_kernel) { |
| 380 | add_sample_entry(s->eip, s->event); |
| 381 | return 1; |
| 382 | } |
| 383 | |
| 384 | /* add userspace sample */ |
| 385 | |
| 386 | if (!mm) { |
| 387 | atomic_inc(&oprofile_stats.sample_lost_no_mm); |
| 388 | return 0; |
| 389 | } |
| 390 | |
| 391 | cookie = lookup_dcookie(mm, s->eip, &offset); |
| 392 | |
| 393 | if (cookie == INVALID_COOKIE) { |
| 394 | atomic_inc(&oprofile_stats.sample_lost_no_mapping); |
| 395 | return 0; |
| 396 | } |
| 397 | |
| 398 | if (cookie != last_cookie) { |
| 399 | add_cookie_switch(cookie); |
| 400 | last_cookie = cookie; |
| 401 | } |
| 402 | |
| 403 | add_sample_entry(offset, s->event); |
| 404 | |
| 405 | return 1; |
| 406 | } |
| 407 | |
| 408 | |
| 409 | static void release_mm(struct mm_struct *mm) |
| 410 | { |
| 411 | if (!mm) |
| 412 | return; |
| 413 | up_read(&mm->mmap_sem); |
| 414 | mmput(mm); |
| 415 | } |
| 416 | |
| 417 | |
| 418 | static struct mm_struct *take_tasks_mm(struct task_struct *task) |
| 419 | { |
| 420 | struct mm_struct *mm = get_task_mm(task); |
| 421 | if (mm) |
| 422 | down_read(&mm->mmap_sem); |
| 423 | return mm; |
| 424 | } |
| 425 | |
| 426 | |
| 427 | static inline int is_code(unsigned long val) |
| 428 | { |
| 429 | return val == ESCAPE_CODE; |
| 430 | } |
| 431 | |
| 432 | |
| 433 | /* Move tasks along towards death. Any tasks on dead_tasks |
| 434 | * will definitely have no remaining references in any |
| 435 | * CPU buffers at this point, because we use two lists, |
| 436 | * and to have reached the list, it must have gone through |
| 437 | * one full sync already. |
| 438 | */ |
| 439 | static void process_task_mortuary(void) |
| 440 | { |
| 441 | unsigned long flags; |
| 442 | LIST_HEAD(local_dead_tasks); |
| 443 | struct task_struct *task; |
| 444 | struct task_struct *ttask; |
| 445 | |
| 446 | spin_lock_irqsave(&task_mortuary, flags); |
| 447 | |
| 448 | list_splice_init(&dead_tasks, &local_dead_tasks); |
| 449 | list_splice_init(&dying_tasks, &dead_tasks); |
| 450 | |
| 451 | spin_unlock_irqrestore(&task_mortuary, flags); |
| 452 | |
| 453 | list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) { |
| 454 | list_del(&task->tasks); |
| 455 | free_task(task); |
| 456 | } |
| 457 | } |
| 458 | |
| 459 | |
| 460 | static void mark_done(int cpu) |
| 461 | { |
| 462 | int i; |
| 463 | |
| 464 | cpumask_set_cpu(cpu, marked_cpus); |
| 465 | |
| 466 | for_each_online_cpu(i) { |
| 467 | if (!cpumask_test_cpu(i, marked_cpus)) |
| 468 | return; |
| 469 | } |
| 470 | |
| 471 | /* All CPUs have been processed at least once, |
| 472 | * we can process the mortuary once |
| 473 | */ |
| 474 | process_task_mortuary(); |
| 475 | |
| 476 | cpumask_clear(marked_cpus); |
| 477 | } |
| 478 | |
| 479 | |
| 480 | /* FIXME: this is not sufficient if we implement syscall barrier backtrace |
| 481 | * traversal, the code switch to sb_sample_start at first kernel enter/exit |
| 482 | * switch so we need a fifth state and some special handling in sync_buffer() |
| 483 | */ |
| 484 | typedef enum { |
| 485 | sb_bt_ignore = -2, |
| 486 | sb_buffer_start, |
| 487 | sb_bt_start, |
| 488 | sb_sample_start, |
| 489 | } sync_buffer_state; |
| 490 | |
| 491 | /* Sync one of the CPU's buffers into the global event buffer. |
| 492 | * Here we need to go through each batch of samples punctuated |
| 493 | * by context switch notes, taking the task's mmap_sem and doing |
| 494 | * lookup in task->mm->mmap to convert EIP into dcookie/offset |
| 495 | * value. |
| 496 | */ |
| 497 | void sync_buffer(int cpu) |
| 498 | { |
| 499 | struct mm_struct *mm = NULL; |
| 500 | struct mm_struct *oldmm; |
| 501 | unsigned long val; |
| 502 | struct task_struct *new; |
| 503 | unsigned long cookie = 0; |
| 504 | int in_kernel = 1; |
| 505 | sync_buffer_state state = sb_buffer_start; |
| 506 | unsigned int i; |
| 507 | unsigned long available; |
| 508 | unsigned long flags; |
| 509 | struct op_entry entry; |
| 510 | struct op_sample *sample; |
| 511 | |
| 512 | mutex_lock(&buffer_mutex); |
| 513 | |
| 514 | add_cpu_switch(cpu); |
| 515 | |
| 516 | op_cpu_buffer_reset(cpu); |
| 517 | available = op_cpu_buffer_entries(cpu); |
| 518 | |
| 519 | for (i = 0; i < available; ++i) { |
| 520 | sample = op_cpu_buffer_read_entry(&entry, cpu); |
| 521 | if (!sample) |
| 522 | break; |
| 523 | |
| 524 | if (is_code(sample->eip)) { |
| 525 | flags = sample->event; |
| 526 | if (flags & TRACE_BEGIN) { |
| 527 | state = sb_bt_start; |
| 528 | add_trace_begin(); |
| 529 | } |
| 530 | if (flags & KERNEL_CTX_SWITCH) { |
| 531 | /* kernel/userspace switch */ |
| 532 | in_kernel = flags & IS_KERNEL; |
| 533 | if (state == sb_buffer_start) |
| 534 | state = sb_sample_start; |
| 535 | add_kernel_ctx_switch(flags & IS_KERNEL); |
| 536 | } |
| 537 | if (flags & USER_CTX_SWITCH |
| 538 | && op_cpu_buffer_get_data(&entry, &val)) { |
| 539 | /* userspace context switch */ |
| 540 | new = (struct task_struct *)val; |
| 541 | oldmm = mm; |
| 542 | release_mm(oldmm); |
| 543 | mm = take_tasks_mm(new); |
| 544 | if (mm != oldmm) |
| 545 | cookie = get_exec_dcookie(mm); |
| 546 | add_user_ctx_switch(new, cookie); |
| 547 | } |
| 548 | if (op_cpu_buffer_get_size(&entry)) |
| 549 | add_data(&entry, mm); |
| 550 | continue; |
| 551 | } |
| 552 | |
| 553 | if (state < sb_bt_start) |
| 554 | /* ignore sample */ |
| 555 | continue; |
| 556 | |
| 557 | if (add_sample(mm, sample, in_kernel)) |
| 558 | continue; |
| 559 | |
| 560 | /* ignore backtraces if failed to add a sample */ |
| 561 | if (state == sb_bt_start) { |
| 562 | state = sb_bt_ignore; |
| 563 | atomic_inc(&oprofile_stats.bt_lost_no_mapping); |
| 564 | } |
| 565 | } |
| 566 | release_mm(mm); |
| 567 | |
| 568 | mark_done(cpu); |
| 569 | |
| 570 | mutex_unlock(&buffer_mutex); |
| 571 | } |
| 572 | |
| 573 | /* The function can be used to add a buffer worth of data directly to |
| 574 | * the kernel buffer. The buffer is assumed to be a circular buffer. |
| 575 | * Take the entries from index start and end at index end, wrapping |
| 576 | * at max_entries. |
| 577 | */ |
| 578 | void oprofile_put_buff(unsigned long *buf, unsigned int start, |
| 579 | unsigned int stop, unsigned int max) |
| 580 | { |
| 581 | int i; |
| 582 | |
| 583 | i = start; |
| 584 | |
| 585 | mutex_lock(&buffer_mutex); |
| 586 | while (i != stop) { |
| 587 | add_event_entry(buf[i++]); |
| 588 | |
| 589 | if (i >= max) |
| 590 | i = 0; |
| 591 | } |
| 592 | |
| 593 | mutex_unlock(&buffer_mutex); |
| 594 | } |
| 595 | |