1 /* 2 * linux/kernel/profile.c 3 * Simple profiling. Manages a direct-mapped profile hit count buffer, 4 * with configurable resolution, support for restricting the cpus on 5 * which profiling is done, and switching between cpu time and 6 * schedule() calls via kernel command line parameters passed at boot. 7 * 8 * Scheduler profiling support, Arjan van de Ven and Ingo Molnar, 9 * Red Hat, July 2004 10 * Consolidation of architecture support code for profiling, 11 * William Irwin, Oracle, July 2004 12 * Amortized hit count accounting via per-cpu open-addressed hashtables 13 * to resolve timer interrupt livelocks, William Irwin, Oracle, 2004 14 */ 15 16 #include <linux/module.h> 17 #include <linux/profile.h> 18 #include <linux/bootmem.h> 19 #include <linux/notifier.h> 20 #include <linux/mm.h> 21 #include <linux/cpumask.h> 22 #include <linux/cpu.h> 23 #include <linux/highmem.h> 24 #include <linux/mutex.h> 25 #include <linux/slab.h> 26 #include <linux/vmalloc.h> 27 #include <asm/sections.h> 28 #include <asm/irq_regs.h> 29 #include <asm/ptrace.h> 30 31 struct profile_hit { 32 u32 pc, hits; 33 }; 34 #define PROFILE_GRPSHIFT 3 35 #define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT) 36 #define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit)) 37 #define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ) 38 39 /* Oprofile timer tick hook */ 40 static int (*timer_hook)(struct pt_regs *) __read_mostly; 41 42 static atomic_t *prof_buffer; 43 static unsigned long prof_len, prof_shift; 44 45 int prof_on __read_mostly; 46 EXPORT_SYMBOL_GPL(prof_on); 47 48 static cpumask_var_t prof_cpu_mask; 49 #ifdef CONFIG_SMP 50 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits); 51 static DEFINE_PER_CPU(int, cpu_profile_flip); 52 static DEFINE_MUTEX(profile_flip_mutex); 53 #endif /* CONFIG_SMP */ 54 55 int profile_setup(char *str) 56 { 57 static char schedstr[] = "schedule"; 58 static char sleepstr[] = "sleep"; 59 static char kvmstr[] = "kvm"; 60 int par; 61 62 if (!strncmp(str, sleepstr, strlen(sleepstr))) { 63 #ifdef CONFIG_SCHEDSTATS 64 prof_on = SLEEP_PROFILING; 65 if (str[strlen(sleepstr)] == ',') 66 str += strlen(sleepstr) + 1; 67 if (get_option(&str, &par)) 68 prof_shift = par; 69 printk(KERN_INFO 70 "kernel sleep profiling enabled (shift: %ld)\n", 71 prof_shift); 72 #else 73 printk(KERN_WARNING 74 "kernel sleep profiling requires CONFIG_SCHEDSTATS\n"); 75 #endif /* CONFIG_SCHEDSTATS */ 76 } else if (!strncmp(str, schedstr, strlen(schedstr))) { 77 prof_on = SCHED_PROFILING; 78 if (str[strlen(schedstr)] == ',') 79 str += strlen(schedstr) + 1; 80 if (get_option(&str, &par)) 81 prof_shift = par; 82 printk(KERN_INFO 83 "kernel schedule profiling enabled (shift: %ld)\n", 84 prof_shift); 85 } else if (!strncmp(str, kvmstr, strlen(kvmstr))) { 86 prof_on = KVM_PROFILING; 87 if (str[strlen(kvmstr)] == ',') 88 str += strlen(kvmstr) + 1; 89 if (get_option(&str, &par)) 90 prof_shift = par; 91 printk(KERN_INFO 92 "kernel KVM profiling enabled (shift: %ld)\n", 93 prof_shift); 94 } else if (get_option(&str, &par)) { 95 prof_shift = par; 96 prof_on = CPU_PROFILING; 97 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n", 98 prof_shift); 99 } 100 return 1; 101 } 102 __setup("profile=", profile_setup); 103 104 105 int __ref profile_init(void) 106 { 107 int buffer_bytes; 108 if (!prof_on) 109 return 0; 110 111 /* only text is profiled */ 112 prof_len = (_etext - _stext) >> prof_shift; 113 buffer_bytes = prof_len*sizeof(atomic_t); 114 115 if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL)) 116 return -ENOMEM; 117 118 cpumask_copy(prof_cpu_mask, cpu_possible_mask); 119 120 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN); 121 if (prof_buffer) 122 return 0; 123 124 prof_buffer = alloc_pages_exact(buffer_bytes, 125 GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN); 126 if (prof_buffer) 127 return 0; 128 129 prof_buffer = vmalloc(buffer_bytes); 130 if (prof_buffer) 131 return 0; 132 133 free_cpumask_var(prof_cpu_mask); 134 return -ENOMEM; 135 } 136 137 /* Profile event notifications */ 138 139 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier); 140 static ATOMIC_NOTIFIER_HEAD(task_free_notifier); 141 static BLOCKING_NOTIFIER_HEAD(munmap_notifier); 142 143 void profile_task_exit(struct task_struct *task) 144 { 145 blocking_notifier_call_chain(&task_exit_notifier, 0, task); 146 } 147 148 int profile_handoff_task(struct task_struct *task) 149 { 150 int ret; 151 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task); 152 return (ret == NOTIFY_OK) ? 1 : 0; 153 } 154 155 void profile_munmap(unsigned long addr) 156 { 157 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr); 158 } 159 160 int task_handoff_register(struct notifier_block *n) 161 { 162 return atomic_notifier_chain_register(&task_free_notifier, n); 163 } 164 EXPORT_SYMBOL_GPL(task_handoff_register); 165 166 int task_handoff_unregister(struct notifier_block *n) 167 { 168 return atomic_notifier_chain_unregister(&task_free_notifier, n); 169 } 170 EXPORT_SYMBOL_GPL(task_handoff_unregister); 171 172 int profile_event_register(enum profile_type type, struct notifier_block *n) 173 { 174 int err = -EINVAL; 175 176 switch (type) { 177 case PROFILE_TASK_EXIT: 178 err = blocking_notifier_chain_register( 179 &task_exit_notifier, n); 180 break; 181 case PROFILE_MUNMAP: 182 err = blocking_notifier_chain_register( 183 &munmap_notifier, n); 184 break; 185 } 186 187 return err; 188 } 189 EXPORT_SYMBOL_GPL(profile_event_register); 190 191 int profile_event_unregister(enum profile_type type, struct notifier_block *n) 192 { 193 int err = -EINVAL; 194 195 switch (type) { 196 case PROFILE_TASK_EXIT: 197 err = blocking_notifier_chain_unregister( 198 &task_exit_notifier, n); 199 break; 200 case PROFILE_MUNMAP: 201 err = blocking_notifier_chain_unregister( 202 &munmap_notifier, n); 203 break; 204 } 205 206 return err; 207 } 208 EXPORT_SYMBOL_GPL(profile_event_unregister); 209 210 int register_timer_hook(int (*hook)(struct pt_regs *)) 211 { 212 if (timer_hook) 213 return -EBUSY; 214 timer_hook = hook; 215 return 0; 216 } 217 EXPORT_SYMBOL_GPL(register_timer_hook); 218 219 void unregister_timer_hook(int (*hook)(struct pt_regs *)) 220 { 221 WARN_ON(hook != timer_hook); 222 timer_hook = NULL; 223 /* make sure all CPUs see the NULL hook */ 224 synchronize_sched(); /* Allow ongoing interrupts to complete. */ 225 } 226 EXPORT_SYMBOL_GPL(unregister_timer_hook); 227 228 229 #ifdef CONFIG_SMP 230 /* 231 * Each cpu has a pair of open-addressed hashtables for pending 232 * profile hits. read_profile() IPI's all cpus to request them 233 * to flip buffers and flushes their contents to prof_buffer itself. 234 * Flip requests are serialized by the profile_flip_mutex. The sole 235 * use of having a second hashtable is for avoiding cacheline 236 * contention that would otherwise happen during flushes of pending 237 * profile hits required for the accuracy of reported profile hits 238 * and so resurrect the interrupt livelock issue. 239 * 240 * The open-addressed hashtables are indexed by profile buffer slot 241 * and hold the number of pending hits to that profile buffer slot on 242 * a cpu in an entry. When the hashtable overflows, all pending hits 243 * are accounted to their corresponding profile buffer slots with 244 * atomic_add() and the hashtable emptied. As numerous pending hits 245 * may be accounted to a profile buffer slot in a hashtable entry, 246 * this amortizes a number of atomic profile buffer increments likely 247 * to be far larger than the number of entries in the hashtable, 248 * particularly given that the number of distinct profile buffer 249 * positions to which hits are accounted during short intervals (e.g. 250 * several seconds) is usually very small. Exclusion from buffer 251 * flipping is provided by interrupt disablement (note that for 252 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from 253 * process context). 254 * The hash function is meant to be lightweight as opposed to strong, 255 * and was vaguely inspired by ppc64 firmware-supported inverted 256 * pagetable hash functions, but uses a full hashtable full of finite 257 * collision chains, not just pairs of them. 258 * 259 * -- wli 260 */ 261 static void __profile_flip_buffers(void *unused) 262 { 263 int cpu = smp_processor_id(); 264 265 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu); 266 } 267 268 static void profile_flip_buffers(void) 269 { 270 int i, j, cpu; 271 272 mutex_lock(&profile_flip_mutex); 273 j = per_cpu(cpu_profile_flip, get_cpu()); 274 put_cpu(); 275 on_each_cpu(__profile_flip_buffers, NULL, 1); 276 for_each_online_cpu(cpu) { 277 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j]; 278 for (i = 0; i < NR_PROFILE_HIT; ++i) { 279 if (!hits[i].hits) { 280 if (hits[i].pc) 281 hits[i].pc = 0; 282 continue; 283 } 284 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]); 285 hits[i].hits = hits[i].pc = 0; 286 } 287 } 288 mutex_unlock(&profile_flip_mutex); 289 } 290 291 static void profile_discard_flip_buffers(void) 292 { 293 int i, cpu; 294 295 mutex_lock(&profile_flip_mutex); 296 i = per_cpu(cpu_profile_flip, get_cpu()); 297 put_cpu(); 298 on_each_cpu(__profile_flip_buffers, NULL, 1); 299 for_each_online_cpu(cpu) { 300 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i]; 301 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit)); 302 } 303 mutex_unlock(&profile_flip_mutex); 304 } 305 306 void profile_hits(int type, void *__pc, unsigned int nr_hits) 307 { 308 unsigned long primary, secondary, flags, pc = (unsigned long)__pc; 309 int i, j, cpu; 310 struct profile_hit *hits; 311 312 if (prof_on != type || !prof_buffer) 313 return; 314 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1); 315 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT; 316 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT; 317 cpu = get_cpu(); 318 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)]; 319 if (!hits) { 320 put_cpu(); 321 return; 322 } 323 /* 324 * We buffer the global profiler buffer into a per-CPU 325 * queue and thus reduce the number of global (and possibly 326 * NUMA-alien) accesses. The write-queue is self-coalescing: 327 */ 328 local_irq_save(flags); 329 do { 330 for (j = 0; j < PROFILE_GRPSZ; ++j) { 331 if (hits[i + j].pc == pc) { 332 hits[i + j].hits += nr_hits; 333 goto out; 334 } else if (!hits[i + j].hits) { 335 hits[i + j].pc = pc; 336 hits[i + j].hits = nr_hits; 337 goto out; 338 } 339 } 340 i = (i + secondary) & (NR_PROFILE_HIT - 1); 341 } while (i != primary); 342 343 /* 344 * Add the current hit(s) and flush the write-queue out 345 * to the global buffer: 346 */ 347 atomic_add(nr_hits, &prof_buffer[pc]); 348 for (i = 0; i < NR_PROFILE_HIT; ++i) { 349 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]); 350 hits[i].pc = hits[i].hits = 0; 351 } 352 out: 353 local_irq_restore(flags); 354 put_cpu(); 355 } 356 357 static int __cpuinit profile_cpu_callback(struct notifier_block *info, 358 unsigned long action, void *__cpu) 359 { 360 int node, cpu = (unsigned long)__cpu; 361 struct page *page; 362 363 switch (action) { 364 case CPU_UP_PREPARE: 365 case CPU_UP_PREPARE_FROZEN: 366 node = cpu_to_node(cpu); 367 per_cpu(cpu_profile_flip, cpu) = 0; 368 if (!per_cpu(cpu_profile_hits, cpu)[1]) { 369 page = alloc_pages_exact_node(node, 370 GFP_KERNEL | __GFP_ZERO, 371 0); 372 if (!page) 373 return NOTIFY_BAD; 374 per_cpu(cpu_profile_hits, cpu)[1] = page_address(page); 375 } 376 if (!per_cpu(cpu_profile_hits, cpu)[0]) { 377 page = alloc_pages_exact_node(node, 378 GFP_KERNEL | __GFP_ZERO, 379 0); 380 if (!page) 381 goto out_free; 382 per_cpu(cpu_profile_hits, cpu)[0] = page_address(page); 383 } 384 break; 385 out_free: 386 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); 387 per_cpu(cpu_profile_hits, cpu)[1] = NULL; 388 __free_page(page); 389 return NOTIFY_BAD; 390 case CPU_ONLINE: 391 case CPU_ONLINE_FROZEN: 392 if (prof_cpu_mask != NULL) 393 cpumask_set_cpu(cpu, prof_cpu_mask); 394 break; 395 case CPU_UP_CANCELED: 396 case CPU_UP_CANCELED_FROZEN: 397 case CPU_DEAD: 398 case CPU_DEAD_FROZEN: 399 if (prof_cpu_mask != NULL) 400 cpumask_clear_cpu(cpu, prof_cpu_mask); 401 if (per_cpu(cpu_profile_hits, cpu)[0]) { 402 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]); 403 per_cpu(cpu_profile_hits, cpu)[0] = NULL; 404 __free_page(page); 405 } 406 if (per_cpu(cpu_profile_hits, cpu)[1]) { 407 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); 408 per_cpu(cpu_profile_hits, cpu)[1] = NULL; 409 __free_page(page); 410 } 411 break; 412 } 413 return NOTIFY_OK; 414 } 415 #else /* !CONFIG_SMP */ 416 #define profile_flip_buffers() do { } while (0) 417 #define profile_discard_flip_buffers() do { } while (0) 418 #define profile_cpu_callback NULL 419 420 void profile_hits(int type, void *__pc, unsigned int nr_hits) 421 { 422 unsigned long pc; 423 424 if (prof_on != type || !prof_buffer) 425 return; 426 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift; 427 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]); 428 } 429 #endif /* !CONFIG_SMP */ 430 EXPORT_SYMBOL_GPL(profile_hits); 431 432 void profile_tick(int type) 433 { 434 struct pt_regs *regs = get_irq_regs(); 435 436 if (type == CPU_PROFILING && timer_hook) 437 timer_hook(regs); 438 if (!user_mode(regs) && prof_cpu_mask != NULL && 439 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask)) 440 profile_hit(type, (void *)profile_pc(regs)); 441 } 442 443 #ifdef CONFIG_PROC_FS 444 #include <linux/proc_fs.h> 445 #include <asm/uaccess.h> 446 447 static int prof_cpu_mask_read_proc(char *page, char **start, off_t off, 448 int count, int *eof, void *data) 449 { 450 int len = cpumask_scnprintf(page, count, data); 451 if (count - len < 2) 452 return -EINVAL; 453 len += sprintf(page + len, "\n"); 454 return len; 455 } 456 457 static int prof_cpu_mask_write_proc(struct file *file, 458 const char __user *buffer, unsigned long count, void *data) 459 { 460 struct cpumask *mask = data; 461 unsigned long full_count = count, err; 462 cpumask_var_t new_value; 463 464 if (!alloc_cpumask_var(&new_value, GFP_KERNEL)) 465 return -ENOMEM; 466 467 err = cpumask_parse_user(buffer, count, new_value); 468 if (!err) { 469 cpumask_copy(mask, new_value); 470 err = full_count; 471 } 472 free_cpumask_var(new_value); 473 return err; 474 } 475 476 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir) 477 { 478 struct proc_dir_entry *entry; 479 480 /* create /proc/irq/prof_cpu_mask */ 481 entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir); 482 if (!entry) 483 return; 484 entry->data = prof_cpu_mask; 485 entry->read_proc = prof_cpu_mask_read_proc; 486 entry->write_proc = prof_cpu_mask_write_proc; 487 } 488 489 /* 490 * This function accesses profiling information. The returned data is 491 * binary: the sampling step and the actual contents of the profile 492 * buffer. Use of the program readprofile is recommended in order to 493 * get meaningful info out of these data. 494 */ 495 static ssize_t 496 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos) 497 { 498 unsigned long p = *ppos; 499 ssize_t read; 500 char *pnt; 501 unsigned int sample_step = 1 << prof_shift; 502 503 profile_flip_buffers(); 504 if (p >= (prof_len+1)*sizeof(unsigned int)) 505 return 0; 506 if (count > (prof_len+1)*sizeof(unsigned int) - p) 507 count = (prof_len+1)*sizeof(unsigned int) - p; 508 read = 0; 509 510 while (p < sizeof(unsigned int) && count > 0) { 511 if (put_user(*((char *)(&sample_step)+p), buf)) 512 return -EFAULT; 513 buf++; p++; count--; read++; 514 } 515 pnt = (char *)prof_buffer + p - sizeof(atomic_t); 516 if (copy_to_user(buf, (void *)pnt, count)) 517 return -EFAULT; 518 read += count; 519 *ppos += read; 520 return read; 521 } 522 523 /* 524 * Writing to /proc/profile resets the counters 525 * 526 * Writing a 'profiling multiplier' value into it also re-sets the profiling 527 * interrupt frequency, on architectures that support this. 528 */ 529 static ssize_t write_profile(struct file *file, const char __user *buf, 530 size_t count, loff_t *ppos) 531 { 532 #ifdef CONFIG_SMP 533 extern int setup_profiling_timer(unsigned int multiplier); 534 535 if (count == sizeof(int)) { 536 unsigned int multiplier; 537 538 if (copy_from_user(&multiplier, buf, sizeof(int))) 539 return -EFAULT; 540 541 if (setup_profiling_timer(multiplier)) 542 return -EINVAL; 543 } 544 #endif 545 profile_discard_flip_buffers(); 546 memset(prof_buffer, 0, prof_len * sizeof(atomic_t)); 547 return count; 548 } 549 550 static const struct file_operations proc_profile_operations = { 551 .read = read_profile, 552 .write = write_profile, 553 }; 554 555 #ifdef CONFIG_SMP 556 static void profile_nop(void *unused) 557 { 558 } 559 560 static int create_hash_tables(void) 561 { 562 int cpu; 563 564 for_each_online_cpu(cpu) { 565 int node = cpu_to_node(cpu); 566 struct page *page; 567 568 page = alloc_pages_exact_node(node, 569 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE, 570 0); 571 if (!page) 572 goto out_cleanup; 573 per_cpu(cpu_profile_hits, cpu)[1] 574 = (struct profile_hit *)page_address(page); 575 page = alloc_pages_exact_node(node, 576 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE, 577 0); 578 if (!page) 579 goto out_cleanup; 580 per_cpu(cpu_profile_hits, cpu)[0] 581 = (struct profile_hit *)page_address(page); 582 } 583 return 0; 584 out_cleanup: 585 prof_on = 0; 586 smp_mb(); 587 on_each_cpu(profile_nop, NULL, 1); 588 for_each_online_cpu(cpu) { 589 struct page *page; 590 591 if (per_cpu(cpu_profile_hits, cpu)[0]) { 592 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]); 593 per_cpu(cpu_profile_hits, cpu)[0] = NULL; 594 __free_page(page); 595 } 596 if (per_cpu(cpu_profile_hits, cpu)[1]) { 597 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); 598 per_cpu(cpu_profile_hits, cpu)[1] = NULL; 599 __free_page(page); 600 } 601 } 602 return -1; 603 } 604 #else 605 #define create_hash_tables() ({ 0; }) 606 #endif 607 608 int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */ 609 { 610 struct proc_dir_entry *entry; 611 612 if (!prof_on) 613 return 0; 614 if (create_hash_tables()) 615 return -ENOMEM; 616 entry = proc_create("profile", S_IWUSR | S_IRUGO, 617 NULL, &proc_profile_operations); 618 if (!entry) 619 return 0; 620 entry->size = (1+prof_len) * sizeof(atomic_t); 621 hotcpu_notifier(profile_cpu_callback, 0); 622 return 0; 623 } 624 module_init(create_proc_profile); 625 #endif /* CONFIG_PROC_FS */ 626