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