kernel/profile.c

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