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