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 31struct 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 */ 40static int (*timer_hook)(struct pt_regs *) __read_mostly; 41 42static atomic_t *prof_buffer; 43static unsigned long prof_len, prof_shift; 44 45int prof_on __read_mostly; 46EXPORT_SYMBOL_GPL(prof_on); 47 48static cpumask_var_t prof_cpu_mask; 49#ifdef CONFIG_SMP 50static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits); 51static DEFINE_PER_CPU(int, cpu_profile_flip); 52static DEFINE_MUTEX(profile_flip_mutex); 53#endif /* CONFIG_SMP */ 54 55int 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 105int __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 memset(prof_buffer, 0, buffer_bytes); 132 return 0; 133 } 134 135 free_cpumask_var(prof_cpu_mask); 136 return -ENOMEM; 137} 138 139/* Profile event notifications */ 140 141static BLOCKING_NOTIFIER_HEAD(task_exit_notifier); 142static ATOMIC_NOTIFIER_HEAD(task_free_notifier); 143static BLOCKING_NOTIFIER_HEAD(munmap_notifier); 144 145void profile_task_exit(struct task_struct *task) 146{ 147 blocking_notifier_call_chain(&task_exit_notifier, 0, task); 148} 149 150int profile_handoff_task(struct task_struct *task) 151{ 152 int ret; 153 ret = atomic_notifier_call_chain(&task_free_notifier, 0, task); 154 return (ret == NOTIFY_OK) ? 1 : 0; 155} 156 157void profile_munmap(unsigned long addr) 158{ 159 blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr); 160} 161 162int task_handoff_register(struct notifier_block *n) 163{ 164 return atomic_notifier_chain_register(&task_free_notifier, n); 165} 166EXPORT_SYMBOL_GPL(task_handoff_register); 167 168int task_handoff_unregister(struct notifier_block *n) 169{ 170 return atomic_notifier_chain_unregister(&task_free_notifier, n); 171} 172EXPORT_SYMBOL_GPL(task_handoff_unregister); 173 174int profile_event_register(enum profile_type type, struct notifier_block *n) 175{ 176 int err = -EINVAL; 177 178 switch (type) { 179 case PROFILE_TASK_EXIT: 180 err = blocking_notifier_chain_register( 181 &task_exit_notifier, n); 182 break; 183 case PROFILE_MUNMAP: 184 err = blocking_notifier_chain_register( 185 &munmap_notifier, n); 186 break; 187 } 188 189 return err; 190} 191EXPORT_SYMBOL_GPL(profile_event_register); 192 193int profile_event_unregister(enum profile_type type, struct notifier_block *n) 194{ 195 int err = -EINVAL; 196 197 switch (type) { 198 case PROFILE_TASK_EXIT: 199 err = blocking_notifier_chain_unregister( 200 &task_exit_notifier, n); 201 break; 202 case PROFILE_MUNMAP: 203 err = blocking_notifier_chain_unregister( 204 &munmap_notifier, n); 205 break; 206 } 207 208 return err; 209} 210EXPORT_SYMBOL_GPL(profile_event_unregister); 211 212int register_timer_hook(int (*hook)(struct pt_regs *)) 213{ 214 if (timer_hook) 215 return -EBUSY; 216 timer_hook = hook; 217 return 0; 218} 219EXPORT_SYMBOL_GPL(register_timer_hook); 220 221void unregister_timer_hook(int (*hook)(struct pt_regs *)) 222{ 223 WARN_ON(hook != timer_hook); 224 timer_hook = NULL; 225 /* make sure all CPUs see the NULL hook */ 226 synchronize_sched(); /* Allow ongoing interrupts to complete. */ 227} 228EXPORT_SYMBOL_GPL(unregister_timer_hook); 229 230 231#ifdef CONFIG_SMP 232/* 233 * Each cpu has a pair of open-addressed hashtables for pending 234 * profile hits. read_profile() IPI's all cpus to request them 235 * to flip buffers and flushes their contents to prof_buffer itself. 236 * Flip requests are serialized by the profile_flip_mutex. The sole 237 * use of having a second hashtable is for avoiding cacheline 238 * contention that would otherwise happen during flushes of pending 239 * profile hits required for the accuracy of reported profile hits 240 * and so resurrect the interrupt livelock issue. 241 * 242 * The open-addressed hashtables are indexed by profile buffer slot 243 * and hold the number of pending hits to that profile buffer slot on 244 * a cpu in an entry. When the hashtable overflows, all pending hits 245 * are accounted to their corresponding profile buffer slots with 246 * atomic_add() and the hashtable emptied. As numerous pending hits 247 * may be accounted to a profile buffer slot in a hashtable entry, 248 * this amortizes a number of atomic profile buffer increments likely 249 * to be far larger than the number of entries in the hashtable, 250 * particularly given that the number of distinct profile buffer 251 * positions to which hits are accounted during short intervals (e.g. 252 * several seconds) is usually very small. Exclusion from buffer 253 * flipping is provided by interrupt disablement (note that for 254 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from 255 * process context). 256 * The hash function is meant to be lightweight as opposed to strong, 257 * and was vaguely inspired by ppc64 firmware-supported inverted 258 * pagetable hash functions, but uses a full hashtable full of finite 259 * collision chains, not just pairs of them. 260 * 261 * -- wli 262 */ 263static void __profile_flip_buffers(void *unused) 264{ 265 int cpu = smp_processor_id(); 266 267 per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu); 268} 269 270static void profile_flip_buffers(void) 271{ 272 int i, j, cpu; 273 274 mutex_lock(&profile_flip_mutex); 275 j = per_cpu(cpu_profile_flip, get_cpu()); 276 put_cpu(); 277 on_each_cpu(__profile_flip_buffers, NULL, 1); 278 for_each_online_cpu(cpu) { 279 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j]; 280 for (i = 0; i < NR_PROFILE_HIT; ++i) { 281 if (!hits[i].hits) { 282 if (hits[i].pc) 283 hits[i].pc = 0; 284 continue; 285 } 286 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]); 287 hits[i].hits = hits[i].pc = 0; 288 } 289 } 290 mutex_unlock(&profile_flip_mutex); 291} 292 293static void profile_discard_flip_buffers(void) 294{ 295 int i, cpu; 296 297 mutex_lock(&profile_flip_mutex); 298 i = per_cpu(cpu_profile_flip, get_cpu()); 299 put_cpu(); 300 on_each_cpu(__profile_flip_buffers, NULL, 1); 301 for_each_online_cpu(cpu) { 302 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i]; 303 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit)); 304 } 305 mutex_unlock(&profile_flip_mutex); 306} 307 308void profile_hits(int type, void *__pc, unsigned int nr_hits) 309{ 310 unsigned long primary, secondary, flags, pc = (unsigned long)__pc; 311 int i, j, cpu; 312 struct profile_hit *hits; 313 314 if (prof_on != type || !prof_buffer) 315 return; 316 pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1); 317 i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT; 318 secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT; 319 cpu = get_cpu(); 320 hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)]; 321 if (!hits) { 322 put_cpu(); 323 return; 324 } 325 /* 326 * We buffer the global profiler buffer into a per-CPU 327 * queue and thus reduce the number of global (and possibly 328 * NUMA-alien) accesses. The write-queue is self-coalescing: 329 */ 330 local_irq_save(flags); 331 do { 332 for (j = 0; j < PROFILE_GRPSZ; ++j) { 333 if (hits[i + j].pc == pc) { 334 hits[i + j].hits += nr_hits; 335 goto out; 336 } else if (!hits[i + j].hits) { 337 hits[i + j].pc = pc; 338 hits[i + j].hits = nr_hits; 339 goto out; 340 } 341 } 342 i = (i + secondary) & (NR_PROFILE_HIT - 1); 343 } while (i != primary); 344 345 /* 346 * Add the current hit(s) and flush the write-queue out 347 * to the global buffer: 348 */ 349 atomic_add(nr_hits, &prof_buffer[pc]); 350 for (i = 0; i < NR_PROFILE_HIT; ++i) { 351 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]); 352 hits[i].pc = hits[i].hits = 0; 353 } 354out: 355 local_irq_restore(flags); 356 put_cpu(); 357} 358 359static int __cpuinit profile_cpu_callback(struct notifier_block *info, 360 unsigned long action, void *__cpu) 361{ 362 int node, cpu = (unsigned long)__cpu; 363 struct page *page; 364 365 switch (action) { 366 case CPU_UP_PREPARE: 367 case CPU_UP_PREPARE_FROZEN: 368 node = cpu_to_mem(cpu); 369 per_cpu(cpu_profile_flip, cpu) = 0; 370 if (!per_cpu(cpu_profile_hits, cpu)[1]) { 371 page = alloc_pages_exact_node(node, 372 GFP_KERNEL | __GFP_ZERO, 373 0); 374 if (!page) 375 return notifier_from_errno(-ENOMEM); 376 per_cpu(cpu_profile_hits, cpu)[1] = page_address(page); 377 } 378 if (!per_cpu(cpu_profile_hits, cpu)[0]) { 379 page = alloc_pages_exact_node(node, 380 GFP_KERNEL | __GFP_ZERO, 381 0); 382 if (!page) 383 goto out_free; 384 per_cpu(cpu_profile_hits, cpu)[0] = page_address(page); 385 } 386 break; 387out_free: 388 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); 389 per_cpu(cpu_profile_hits, cpu)[1] = NULL; 390 __free_page(page); 391 return notifier_from_errno(-ENOMEM); 392 case CPU_ONLINE: 393 case CPU_ONLINE_FROZEN: 394 if (prof_cpu_mask != NULL) 395 cpumask_set_cpu(cpu, prof_cpu_mask); 396 break; 397 case CPU_UP_CANCELED: 398 case CPU_UP_CANCELED_FROZEN: 399 case CPU_DEAD: 400 case CPU_DEAD_FROZEN: 401 if (prof_cpu_mask != NULL) 402 cpumask_clear_cpu(cpu, prof_cpu_mask); 403 if (per_cpu(cpu_profile_hits, cpu)[0]) { 404 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]); 405 per_cpu(cpu_profile_hits, cpu)[0] = NULL; 406 __free_page(page); 407 } 408 if (per_cpu(cpu_profile_hits, cpu)[1]) { 409 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); 410 per_cpu(cpu_profile_hits, cpu)[1] = NULL; 411 __free_page(page); 412 } 413 break; 414 } 415 return NOTIFY_OK; 416} 417#else /* !CONFIG_SMP */ 418#define profile_flip_buffers() do { } while (0) 419#define profile_discard_flip_buffers() do { } while (0) 420#define profile_cpu_callback NULL 421 422void profile_hits(int type, void *__pc, unsigned int nr_hits) 423{ 424 unsigned long pc; 425 426 if (prof_on != type || !prof_buffer) 427 return; 428 pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift; 429 atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]); 430} 431#endif /* !CONFIG_SMP */ 432EXPORT_SYMBOL_GPL(profile_hits); 433 434void profile_tick(int type) 435{ 436 struct pt_regs *regs = get_irq_regs(); 437 438 if (type == CPU_PROFILING && timer_hook) 439 timer_hook(regs); 440 if (!user_mode(regs) && prof_cpu_mask != NULL && 441 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask)) 442 profile_hit(type, (void *)profile_pc(regs)); 443} 444 445#ifdef CONFIG_PROC_FS 446#include <linux/proc_fs.h> 447#include <linux/seq_file.h> 448#include <asm/uaccess.h> 449 450static int prof_cpu_mask_proc_show(struct seq_file *m, void *v) 451{ 452 seq_cpumask(m, prof_cpu_mask); 453 seq_putc(m, '\n'); 454 return 0; 455} 456 457static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file) 458{ 459 return single_open(file, prof_cpu_mask_proc_show, NULL); 460} 461 462static ssize_t prof_cpu_mask_proc_write(struct file *file, 463 const char __user *buffer, size_t count, loff_t *pos) 464{ 465 cpumask_var_t new_value; 466 int err; 467 468 if (!alloc_cpumask_var(&new_value, GFP_KERNEL)) 469 return -ENOMEM; 470 471 err = cpumask_parse_user(buffer, count, new_value); 472 if (!err) { 473 cpumask_copy(prof_cpu_mask, new_value); 474 err = count; 475 } 476 free_cpumask_var(new_value); 477 return err; 478} 479 480static const struct file_operations prof_cpu_mask_proc_fops = { 481 .open = prof_cpu_mask_proc_open, 482 .read = seq_read, 483 .llseek = seq_lseek, 484 .release = single_release, 485 .write = prof_cpu_mask_proc_write, 486}; 487 488void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir) 489{ 490 /* create /proc/irq/prof_cpu_mask */ 491 proc_create("prof_cpu_mask", 0600, root_irq_dir, &prof_cpu_mask_proc_fops); 492} 493 494/* 495 * This function accesses profiling information. The returned data is 496 * binary: the sampling step and the actual contents of the profile 497 * buffer. Use of the program readprofile is recommended in order to 498 * get meaningful info out of these data. 499 */ 500static ssize_t 501read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos) 502{ 503 unsigned long p = *ppos; 504 ssize_t read; 505 char *pnt; 506 unsigned int sample_step = 1 << prof_shift; 507 508 profile_flip_buffers(); 509 if (p >= (prof_len+1)*sizeof(unsigned int)) 510 return 0; 511 if (count > (prof_len+1)*sizeof(unsigned int) - p) 512 count = (prof_len+1)*sizeof(unsigned int) - p; 513 read = 0; 514 515 while (p < sizeof(unsigned int) && count > 0) { 516 if (put_user(*((char *)(&sample_step)+p), buf)) 517 return -EFAULT; 518 buf++; p++; count--; read++; 519 } 520 pnt = (char *)prof_buffer + p - sizeof(atomic_t); 521 if (copy_to_user(buf, (void *)pnt, count)) 522 return -EFAULT; 523 read += count; 524 *ppos += read; 525 return read; 526} 527 528/* 529 * Writing to /proc/profile resets the counters 530 * 531 * Writing a 'profiling multiplier' value into it also re-sets the profiling 532 * interrupt frequency, on architectures that support this. 533 */ 534static ssize_t write_profile(struct file *file, const char __user *buf, 535 size_t count, loff_t *ppos) 536{ 537#ifdef CONFIG_SMP 538 extern int setup_profiling_timer(unsigned int multiplier); 539 540 if (count == sizeof(int)) { 541 unsigned int multiplier; 542 543 if (copy_from_user(&multiplier, buf, sizeof(int))) 544 return -EFAULT; 545 546 if (setup_profiling_timer(multiplier)) 547 return -EINVAL; 548 } 549#endif 550 profile_discard_flip_buffers(); 551 memset(prof_buffer, 0, prof_len * sizeof(atomic_t)); 552 return count; 553} 554 555static const struct file_operations proc_profile_operations = { 556 .read = read_profile, 557 .write = write_profile, 558}; 559 560#ifdef CONFIG_SMP 561static void profile_nop(void *unused) 562{ 563} 564 565static int create_hash_tables(void) 566{ 567 int cpu; 568 569 for_each_online_cpu(cpu) { 570 int node = cpu_to_mem(cpu); 571 struct page *page; 572 573 page = alloc_pages_exact_node(node, 574 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE, 575 0); 576 if (!page) 577 goto out_cleanup; 578 per_cpu(cpu_profile_hits, cpu)[1] 579 = (struct profile_hit *)page_address(page); 580 page = alloc_pages_exact_node(node, 581 GFP_KERNEL | __GFP_ZERO | GFP_THISNODE, 582 0); 583 if (!page) 584 goto out_cleanup; 585 per_cpu(cpu_profile_hits, cpu)[0] 586 = (struct profile_hit *)page_address(page); 587 } 588 return 0; 589out_cleanup: 590 prof_on = 0; 591 smp_mb(); 592 on_each_cpu(profile_nop, NULL, 1); 593 for_each_online_cpu(cpu) { 594 struct page *page; 595 596 if (per_cpu(cpu_profile_hits, cpu)[0]) { 597 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]); 598 per_cpu(cpu_profile_hits, cpu)[0] = NULL; 599 __free_page(page); 600 } 601 if (per_cpu(cpu_profile_hits, cpu)[1]) { 602 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]); 603 per_cpu(cpu_profile_hits, cpu)[1] = NULL; 604 __free_page(page); 605 } 606 } 607 return -1; 608} 609#else 610#define create_hash_tables() ({ 0; }) 611#endif 612 613int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */ 614{ 615 struct proc_dir_entry *entry; 616 617 if (!prof_on) 618 return 0; 619 if (create_hash_tables()) 620 return -ENOMEM; 621 entry = proc_create("profile", S_IWUSR | S_IRUGO, 622 NULL, &proc_profile_operations); 623 if (!entry) 624 return 0; 625 entry->size = (1+prof_len) * sizeof(atomic_t); 626 hotcpu_notifier(profile_cpu_callback, 0); 627 return 0; 628} 629module_init(create_proc_profile); 630#endif /* CONFIG_PROC_FS */ 631