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_t prof_cpu_mask = CPU_MASK_ALL; 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 if (!slab_is_available()) { 115 prof_buffer = alloc_bootmem(buffer_bytes); 116 return 0; 117 } 118 119 prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL); 120 if (prof_buffer) 121 return 0; 122 123 prof_buffer = alloc_pages_exact(buffer_bytes, GFP_KERNEL|__GFP_ZERO); 124 if (prof_buffer) 125 return 0; 126 127 prof_buffer = vmalloc(buffer_bytes); 128 if (prof_buffer) 129 return 0; 130 131 return -ENOMEM; 132 } 133 134 /* Profile event notifications */ 135 136 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier); 137 static ATOMIC_NOTIFIER_HEAD(task_free_notifier); 138 static BLOCKING_NOTIFIER_HEAD(munmap_notifier); 139 140 void profile_task_exit(struct task_struct *task) 141 { 142 blocking_notifier_call_chain(&task_exit_notifier, 0, task); 143 } 144 145 int profile_handoff_task(struct task_struct *task) 146 { 147 int ret; 148 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task); 149 return (ret == NOTIFY_OK) ? 1 : 0; 150 } 151 152 void profile_munmap(unsigned long addr) 153 { 154 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr); 155 } 156 157 int task_handoff_register(struct notifier_block *n) 158 { 159 return atomic_notifier_chain_register(&task_free_notifier, n); 160 } 161 EXPORT_SYMBOL_GPL(task_handoff_register); 162 163 int task_handoff_unregister(struct notifier_block *n) 164 { 165 return atomic_notifier_chain_unregister(&task_free_notifier, n); 166 } 167 EXPORT_SYMBOL_GPL(task_handoff_unregister); 168 169 int profile_event_register(enum profile_type type, struct notifier_block *n) 170 { 171 int err = -EINVAL; 172 173 switch (type) { 174 case PROFILE_TASK_EXIT: 175 err = blocking_notifier_chain_register( 176 &task_exit_notifier, n); 177 break; 178 case PROFILE_MUNMAP: 179 err = blocking_notifier_chain_register( 180 &munmap_notifier, n); 181 break; 182 } 183 184 return err; 185 } 186 EXPORT_SYMBOL_GPL(profile_event_register); 187 188 int profile_event_unregister(enum profile_type type, struct notifier_block *n) 189 { 190 int err = -EINVAL; 191 192 switch (type) { 193 case PROFILE_TASK_EXIT: 194 err = blocking_notifier_chain_unregister( 195 &task_exit_notifier, n); 196 break; 197 case PROFILE_MUNMAP: 198 err = blocking_notifier_chain_unregister( 199 &munmap_notifier, n); 200 break; 201 } 202 203 return err; 204 } 205 EXPORT_SYMBOL_GPL(profile_event_unregister); 206 207 int register_timer_hook(int (*hook)(struct pt_regs *)) 208 { 209 if (timer_hook) 210 return -EBUSY; 211 timer_hook = hook; 212 return 0; 213 } 214 EXPORT_SYMBOL_GPL(register_timer_hook); 215 216 void unregister_timer_hook(int (*hook)(struct pt_regs *)) 217 { 218 WARN_ON(hook != timer_hook); 219 timer_hook = NULL; 220 /* make sure all CPUs see the NULL hook */ 221 synchronize_sched(); /* Allow ongoing interrupts to complete. */ 222 } 223 EXPORT_SYMBOL_GPL(unregister_timer_hook); 224 225 226 #ifdef CONFIG_SMP 227 /* 228 * Each cpu has a pair of open-addressed hashtables for pending 229 * profile hits. read_profile() IPI's all cpus to request them 230 * to flip buffers and flushes their contents to prof_buffer itself. 231 * Flip requests are serialized by the profile_flip_mutex. The sole 232 * use of having a second hashtable is for avoiding cacheline 233 * contention that would otherwise happen during flushes of pending 234 * profile hits required for the accuracy of reported profile hits 235 * and so resurrect the interrupt livelock issue. 236 * 237 * The open-addressed hashtables are indexed by profile buffer slot 238 * and hold the number of pending hits to that profile buffer slot on 239 * a cpu in an entry. When the hashtable overflows, all pending hits 240 * are accounted to their corresponding profile buffer slots with 241 * atomic_add() and the hashtable emptied. As numerous pending hits 242 * may be accounted to a profile buffer slot in a hashtable entry, 243 * this amortizes a number of atomic profile buffer increments likely 244 * to be far larger than the number of entries in the hashtable, 245 * particularly given that the number of distinct profile buffer 246 * positions to which hits are accounted during short intervals (e.g. 247 * several seconds) is usually very small. Exclusion from buffer 248 * flipping is provided by interrupt disablement (note that for 249 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from 250 * process context). 251 * The hash function is meant to be lightweight as opposed to strong, 252 * and was vaguely inspired by ppc64 firmware-supported inverted 253 * pagetable hash functions, but uses a full hashtable full of finite 254 * collision chains, not just pairs of them. 255 * 256 * -- wli 257 */ 258 static void __profile_flip_buffers(void *unused) 259 { 260 int cpu = smp_processor_id(); 261 262 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu); 263 } 264 265 static void profile_flip_buffers(void) 266 { 267 int i, j, cpu; 268 269 mutex_lock(&profile_flip_mutex); 270 j = per_cpu(cpu_profile_flip, get_cpu()); 271 put_cpu(); 272 on_each_cpu(__profile_flip_buffers, NULL, 1); 273 for_each_online_cpu(cpu) { 274 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j]; 275 for (i = 0; i < NR_PROFILE_HIT; ++i) { 276 if (!hits[i].hits) { 277 if (hits[i].pc) 278 hits[i].pc = 0; 279 continue; 280 } 281 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]); 282 hits[i].hits = hits[i].pc = 0; 283 } 284 } 285 mutex_unlock(&profile_flip_mutex); 286 } 287 288 static void profile_discard_flip_buffers(void) 289 { 290 int i, cpu; 291 292 mutex_lock(&profile_flip_mutex); 293 i = per_cpu(cpu_profile_flip, get_cpu()); 294 put_cpu(); 295 on_each_cpu(__profile_flip_buffers, NULL, 1); 296 for_each_online_cpu(cpu) { 297 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i]; 298 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit)); 299 } 300 mutex_unlock(&profile_flip_mutex); 301 } 302 303 void profile_hits(int type, void *__pc, unsigned int nr_hits) 304 { 305 unsigned long primary, secondary, flags, pc = (unsigned long)__pc; 306 int i, j, cpu; 307 struct profile_hit *hits; 308 309 if (prof_on != type || !prof_buffer) 310 return; 311 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1); 312 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT; 313 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT; 314 cpu = get_cpu(); 315 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)]; 316 if (!hits) { 317 put_cpu(); 318 return; 319 } 320 /* 321 * We buffer the global profiler buffer into a per-CPU 322 * queue and thus reduce the number of global (and possibly 323 * NUMA-alien) accesses. The write-queue is self-coalescing: 324 */ 325 local_irq_save(flags); 326 do { 327 for (j = 0; j < PROFILE_GRPSZ; ++j) { 328 if (hits[i + j].pc == pc) { 329 hits[i + j].hits += nr_hits; 330 goto out; 331 } else if (!hits[i + j].hits) { 332 hits[i + j].pc = pc; 333 hits[i + j].hits = nr_hits; 334 goto out; 335 } 336 } 337 i = (i + secondary) & (NR_PROFILE_HIT - 1); 338 } while (i != primary); 339 340 /* 341 * Add the current hit(s) and flush the write-queue out 342 * to the global buffer: 343 */ 344 atomic_add(nr_hits, &prof_buffer[pc]); 345 for (i = 0; i < NR_PROFILE_HIT; ++i) { 346 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]); 347 hits[i].pc = hits[i].hits = 0; 348 } 349 out: 350 local_irq_restore(flags); 351 put_cpu(); 352 } 353 354 static int __devinit profile_cpu_callback(struct notifier_block *info, 355 unsigned long action, void *__cpu) 356 { 357 int node, cpu = (unsigned long)__cpu; 358 struct page *page; 359 360 switch (action) { 361 case CPU_UP_PREPARE: 362 case CPU_UP_PREPARE_FROZEN: 363 node = cpu_to_node(cpu); 364 per_cpu(cpu_profile_flip, cpu) = 0; 365 if (!per_cpu(cpu_profile_hits, cpu)[1]) { 366 page = alloc_pages_node(node, 367 GFP_KERNEL | __GFP_ZERO, 368 0); 369 if (!page) 370 return NOTIFY_BAD; 371 per_cpu(cpu_profile_hits, cpu)[1] = page_address(page); 372 } 373 if (!per_cpu(cpu_profile_hits, cpu)[0]) { 374 page = alloc_pages_node(node, 375 GFP_KERNEL | __GFP_ZERO, 376 0); 377 if (!page) 378 goto out_free; 379 per_cpu(cpu_profile_hits, cpu)[0] = page_address(page); 380 } 381 break; 382 out_free: 383 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); 384 per_cpu(cpu_profile_hits, cpu)[1] = NULL; 385 __free_page(page); 386 return NOTIFY_BAD; 387 case CPU_ONLINE: 388 case CPU_ONLINE_FROZEN: 389 cpu_set(cpu, prof_cpu_mask); 390 break; 391 case CPU_UP_CANCELED: 392 case CPU_UP_CANCELED_FROZEN: 393 case CPU_DEAD: 394 case CPU_DEAD_FROZEN: 395 cpu_clear(cpu, prof_cpu_mask); 396 if (per_cpu(cpu_profile_hits, cpu)[0]) { 397 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]); 398 per_cpu(cpu_profile_hits, cpu)[0] = NULL; 399 __free_page(page); 400 } 401 if (per_cpu(cpu_profile_hits, cpu)[1]) { 402 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); 403 per_cpu(cpu_profile_hits, cpu)[1] = NULL; 404 __free_page(page); 405 } 406 break; 407 } 408 return NOTIFY_OK; 409 } 410 #else /* !CONFIG_SMP */ 411 #define profile_flip_buffers() do { } while (0) 412 #define profile_discard_flip_buffers() do { } while (0) 413 #define profile_cpu_callback NULL 414 415 void profile_hits(int type, void *__pc, unsigned int nr_hits) 416 { 417 unsigned long pc; 418 419 if (prof_on != type || !prof_buffer) 420 return; 421 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift; 422 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]); 423 } 424 #endif /* !CONFIG_SMP */ 425 EXPORT_SYMBOL_GPL(profile_hits); 426 427 void profile_tick(int type) 428 { 429 struct pt_regs *regs = get_irq_regs(); 430 431 if (type == CPU_PROFILING && timer_hook) 432 timer_hook(regs); 433 if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask)) 434 profile_hit(type, (void *)profile_pc(regs)); 435 } 436 437 #ifdef CONFIG_PROC_FS 438 #include <linux/proc_fs.h> 439 #include <asm/uaccess.h> 440 #include <asm/ptrace.h> 441 442 static int prof_cpu_mask_read_proc(char *page, char **start, off_t off, 443 int count, int *eof, void *data) 444 { 445 int len = cpumask_scnprintf(page, count, *(cpumask_t *)data); 446 if (count - len < 2) 447 return -EINVAL; 448 len += sprintf(page + len, "\n"); 449 return len; 450 } 451 452 static int prof_cpu_mask_write_proc(struct file *file, 453 const char __user *buffer, unsigned long count, void *data) 454 { 455 cpumask_t *mask = (cpumask_t *)data; 456 unsigned long full_count = count, err; 457 cpumask_t new_value; 458 459 err = cpumask_parse_user(buffer, count, new_value); 460 if (err) 461 return err; 462 463 *mask = new_value; 464 return full_count; 465 } 466 467 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir) 468 { 469 struct proc_dir_entry *entry; 470 471 /* create /proc/irq/prof_cpu_mask */ 472 entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir); 473 if (!entry) 474 return; 475 entry->data = (void *)&prof_cpu_mask; 476 entry->read_proc = prof_cpu_mask_read_proc; 477 entry->write_proc = prof_cpu_mask_write_proc; 478 } 479 480 /* 481 * This function accesses profiling information. The returned data is 482 * binary: the sampling step and the actual contents of the profile 483 * buffer. Use of the program readprofile is recommended in order to 484 * get meaningful info out of these data. 485 */ 486 static ssize_t 487 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos) 488 { 489 unsigned long p = *ppos; 490 ssize_t read; 491 char *pnt; 492 unsigned int sample_step = 1 << prof_shift; 493 494 profile_flip_buffers(); 495 if (p >= (prof_len+1)*sizeof(unsigned int)) 496 return 0; 497 if (count > (prof_len+1)*sizeof(unsigned int) - p) 498 count = (prof_len+1)*sizeof(unsigned int) - p; 499 read = 0; 500 501 while (p < sizeof(unsigned int) && count > 0) { 502 if (put_user(*((char *)(&sample_step)+p), buf)) 503 return -EFAULT; 504 buf++; p++; count--; read++; 505 } 506 pnt = (char *)prof_buffer + p - sizeof(atomic_t); 507 if (copy_to_user(buf, (void *)pnt, count)) 508 return -EFAULT; 509 read += count; 510 *ppos += read; 511 return read; 512 } 513 514 /* 515 * Writing to /proc/profile resets the counters 516 * 517 * Writing a 'profiling multiplier' value into it also re-sets the profiling 518 * interrupt frequency, on architectures that support this. 519 */ 520 static ssize_t write_profile(struct file *file, const char __user *buf, 521 size_t count, loff_t *ppos) 522 { 523 #ifdef CONFIG_SMP 524 extern int setup_profiling_timer(unsigned int multiplier); 525 526 if (count == sizeof(int)) { 527 unsigned int multiplier; 528 529 if (copy_from_user(&multiplier, buf, sizeof(int))) 530 return -EFAULT; 531 532 if (setup_profiling_timer(multiplier)) 533 return -EINVAL; 534 } 535 #endif 536 profile_discard_flip_buffers(); 537 memset(prof_buffer, 0, prof_len * sizeof(atomic_t)); 538 return count; 539 } 540 541 static const struct file_operations proc_profile_operations = { 542 .read = read_profile, 543 .write = write_profile, 544 }; 545 546 #ifdef CONFIG_SMP 547 static void __init profile_nop(void *unused) 548 { 549 } 550 551 static int create_hash_tables(void) 552 { 553 int cpu; 554 555 for_each_online_cpu(cpu) { 556 int node = cpu_to_node(cpu); 557 struct page *page; 558 559 page = alloc_pages_node(node, 560 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE, 561 0); 562 if (!page) 563 goto out_cleanup; 564 per_cpu(cpu_profile_hits, cpu)[1] 565 = (struct profile_hit *)page_address(page); 566 page = alloc_pages_node(node, 567 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE, 568 0); 569 if (!page) 570 goto out_cleanup; 571 per_cpu(cpu_profile_hits, cpu)[0] 572 = (struct profile_hit *)page_address(page); 573 } 574 return 0; 575 out_cleanup: 576 prof_on = 0; 577 smp_mb(); 578 on_each_cpu(profile_nop, NULL, 1); 579 for_each_online_cpu(cpu) { 580 struct page *page; 581 582 if (per_cpu(cpu_profile_hits, cpu)[0]) { 583 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]); 584 per_cpu(cpu_profile_hits, cpu)[0] = NULL; 585 __free_page(page); 586 } 587 if (per_cpu(cpu_profile_hits, cpu)[1]) { 588 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); 589 per_cpu(cpu_profile_hits, cpu)[1] = NULL; 590 __free_page(page); 591 } 592 } 593 return -1; 594 } 595 #else 596 #define create_hash_tables() ({ 0; }) 597 #endif 598 599 int create_proc_profile(void) 600 { 601 struct proc_dir_entry *entry; 602 603 if (!prof_on) 604 return 0; 605 if (create_hash_tables()) 606 return -ENOMEM; 607 entry = proc_create("profile", S_IWUSR | S_IRUGO, 608 NULL, &proc_profile_operations); 609 if (!entry) 610 return 0; 611 entry->size = (1+prof_len) * sizeof(atomic_t); 612 hotcpu_notifier(profile_cpu_callback, 0); 613 return 0; 614 } 615 module_init(create_proc_profile); 616 #endif /* CONFIG_PROC_FS */ 617