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/config.h>
  17 #include <linux/module.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/profile.h>
  25 #include <linux/highmem.h>
  26 #include <linux/mutex.h>
  27 #include <asm/sections.h>
  28 #include <asm/semaphore.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 static int prof_on __read_mostly;
  44 static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
  45 #ifdef CONFIG_SMP
  46 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
  47 static DEFINE_PER_CPU(int, cpu_profile_flip);
  48 static DEFINE_MUTEX(profile_flip_mutex);
  49 #endif /* CONFIG_SMP */
  50
  51 static int __init profile_setup(char * str)
  52 {
  53         static char __initdata schedstr[] = "schedule";
  54         int par;
  55
  56         if (!strncmp(str, schedstr, strlen(schedstr))) {
  57                 prof_on = SCHED_PROFILING;
  58                 if (str[strlen(schedstr)] == ',')
  59                         str += strlen(schedstr) + 1;
  60                 if (get_option(&str, &par))
  61                         prof_shift = par;
  62                 printk(KERN_INFO
  63                         "kernel schedule profiling enabled (shift: %ld)\n",
  64                         prof_shift);
  65         } else if (get_option(&str, &par)) {
  66                 prof_shift = par;
  67                 prof_on = CPU_PROFILING;
  68                 printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
  69                         prof_shift);
  70         }
  71         return 1;
  72 }
  73 __setup("profile=", profile_setup);
  74
  75
  76 void __init profile_init(void)
  77 {
  78         if (!prof_on)
  79                 return;
  80
  81         /* only text is profiled */
  82         prof_len = (_etext - _stext) >> prof_shift;
  83         prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
  84 }
  85
  86 /* Profile event notifications */
  87
  88 #ifdef CONFIG_PROFILING
  89
  90 static DECLARE_RWSEM(profile_rwsem);
  91 static DEFINE_RWLOCK(handoff_lock);
  92 static struct notifier_block * task_exit_notifier;
  93 static struct notifier_block * task_free_notifier;
  94 static struct notifier_block * munmap_notifier;
  95
  96 void profile_task_exit(struct task_struct * task)
  97 {
  98         down_read(&profile_rwsem);
  99         notifier_call_chain(&task_exit_notifier, 0, task);
 100         up_read(&profile_rwsem);
 101 }
 102
 103 int profile_handoff_task(struct task_struct * task)
 104 {
 105         int ret;
 106         read_lock(&handoff_lock);
 107         ret = notifier_call_chain(&task_free_notifier, 0, task);
 108         read_unlock(&handoff_lock);
 109         return (ret == NOTIFY_OK) ? 1 : 0;
 110 }
 111
 112 void profile_munmap(unsigned long addr)
 113 {
 114         down_read(&profile_rwsem);
 115         notifier_call_chain(&munmap_notifier, 0, (void *)addr);
 116         up_read(&profile_rwsem);
 117 }
 118
 119 int task_handoff_register(struct notifier_block * n)
 120 {
 121         int err = -EINVAL;
 122
 123         write_lock(&handoff_lock);
 124         err = notifier_chain_register(&task_free_notifier, n);
 125         write_unlock(&handoff_lock);
 126         return err;
 127 }
 128
 129 int task_handoff_unregister(struct notifier_block * n)
 130 {
 131         int err = -EINVAL;
 132
 133         write_lock(&handoff_lock);
 134         err = notifier_chain_unregister(&task_free_notifier, n);
 135         write_unlock(&handoff_lock);
 136         return err;
 137 }
 138
 139 int profile_event_register(enum profile_type type, struct notifier_block * n)
 140 {
 141         int err = -EINVAL;
 142
 143         down_write(&profile_rwsem);
 144
 145         switch (type) {
 146                 case PROFILE_TASK_EXIT:
 147                         err = notifier_chain_register(&task_exit_notifier, n);
 148                         break;
 149                 case PROFILE_MUNMAP:
 150                         err = notifier_chain_register(&munmap_notifier, n);
 151                         break;
 152         }
 153
 154         up_write(&profile_rwsem);
 155
 156         return err;
 157 }
 158
 159
 160 int profile_event_unregister(enum profile_type type, struct notifier_block * n)
 161 {
 162         int err = -EINVAL;
 163
 164         down_write(&profile_rwsem);
 165
 166         switch (type) {
 167                 case PROFILE_TASK_EXIT:
 168                         err = notifier_chain_unregister(&task_exit_notifier, n);
 169                         break;
 170                 case PROFILE_MUNMAP:
 171                         err = notifier_chain_unregister(&munmap_notifier, n);
 172                         break;
 173         }
 174
 175         up_write(&profile_rwsem);
 176         return err;
 177 }
 178
 179 int register_timer_hook(int (*hook)(struct pt_regs *))
 180 {
 181         if (timer_hook)
 182                 return -EBUSY;
 183         timer_hook = hook;
 184         return 0;
 185 }
 186
 187 void unregister_timer_hook(int (*hook)(struct pt_regs *))
 188 {
 189         WARN_ON(hook != timer_hook);
 190         timer_hook = NULL;
 191         /* make sure all CPUs see the NULL hook */
 192         synchronize_sched();  /* Allow ongoing interrupts to complete. */
 193 }
 194
 195 EXPORT_SYMBOL_GPL(register_timer_hook);
 196 EXPORT_SYMBOL_GPL(unregister_timer_hook);
 197 EXPORT_SYMBOL_GPL(task_handoff_register);
 198 EXPORT_SYMBOL_GPL(task_handoff_unregister);
 199
 200 #endif /* CONFIG_PROFILING */
 201
 202 EXPORT_SYMBOL_GPL(profile_event_register);
 203 EXPORT_SYMBOL_GPL(profile_event_unregister);
 204
 205 #ifdef CONFIG_SMP
 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 profile_hit() may be called from process context).
 229  * The hash function is meant to be lightweight as opposed to strong,
 230  * and was vaguely inspired by ppc64 firmware-supported inverted
 231  * pagetable hash functions, but uses a full hashtable full of finite
 232  * collision chains, not just pairs of them.
 233  *
 234  * -- wli
 235  */
 236 static void __profile_flip_buffers(void *unused)
 237 {
 238         int cpu = smp_processor_id();
 239
 240         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
 241 }
 242
 243 static void profile_flip_buffers(void)
 244 {
 245         int i, j, cpu;
 246
 247         mutex_lock(&profile_flip_mutex);
 248         j = per_cpu(cpu_profile_flip, get_cpu());
 249         put_cpu();
 250         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
 251         for_each_online_cpu(cpu) {
 252                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
 253                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
 254                         if (!hits[i].hits) {
 255                                 if (hits[i].pc)
 256                                         hits[i].pc = 0;
 257                                 continue;
 258                         }
 259                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
 260                         hits[i].hits = hits[i].pc = 0;
 261                 }
 262         }
 263         mutex_unlock(&profile_flip_mutex);
 264 }
 265
 266 static void profile_discard_flip_buffers(void)
 267 {
 268         int i, cpu;
 269
 270         mutex_lock(&profile_flip_mutex);
 271         i = per_cpu(cpu_profile_flip, get_cpu());
 272         put_cpu();
 273         on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
 274         for_each_online_cpu(cpu) {
 275                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
 276                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
 277         }
 278         mutex_unlock(&profile_flip_mutex);
 279 }
 280
 281 void profile_hit(int type, void *__pc)
 282 {
 283         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
 284         int i, j, cpu;
 285         struct profile_hit *hits;
 286
 287         if (prof_on != type || !prof_buffer)
 288                 return;
 289         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
 290         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
 291         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
 292         cpu = get_cpu();
 293         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
 294         if (!hits) {
 295                 put_cpu();
 296                 return;
 297         }
 298         local_irq_save(flags);
 299         do {
 300                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
 301                         if (hits[i + j].pc == pc) {
 302                                 hits[i + j].hits++;
 303                                 goto out;
 304                         } else if (!hits[i + j].hits) {
 305                                 hits[i + j].pc = pc;
 306                                 hits[i + j].hits = 1;
 307                                 goto out;
 308                         }
 309                 }
 310                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
 311         } while (i != primary);
 312         atomic_inc(&prof_buffer[pc]);
 313         for (i = 0; i < NR_PROFILE_HIT; ++i) {
 314                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
 315                 hits[i].pc = hits[i].hits = 0;
 316         }
 317 out:
 318         local_irq_restore(flags);
 319         put_cpu();
 320 }
 321
 322 #ifdef CONFIG_HOTPLUG_CPU
 323 static int __devinit profile_cpu_callback(struct notifier_block *info,
 324                                         unsigned long action, void *__cpu)
 325 {
 326         int node, cpu = (unsigned long)__cpu;
 327         struct page *page;
 328
 329         switch (action) {
 330         case CPU_UP_PREPARE:
 331                 node = cpu_to_node(cpu);
 332                 per_cpu(cpu_profile_flip, cpu) = 0;
 333                 if (!per_cpu(cpu_profile_hits, cpu)[1]) {
 334                         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 335                         if (!page)
 336                                 return NOTIFY_BAD;
 337                         per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
 338                 }
 339                 if (!per_cpu(cpu_profile_hits, cpu)[0]) {
 340                         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 341                         if (!page)
 342                                 goto out_free;
 343                         per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
 344                 }
 345                 break;
 346         out_free:
 347                 page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
 348                 per_cpu(cpu_profile_hits, cpu)[1] = NULL;
 349                 __free_page(page);
 350                 return NOTIFY_BAD;
 351         case CPU_ONLINE:
 352                 cpu_set(cpu, prof_cpu_mask);
 353                 break;
 354         case CPU_UP_CANCELED:
 355         case CPU_DEAD:
 356                 cpu_clear(cpu, prof_cpu_mask);
 357                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
 358                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
 359                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
 360                         __free_page(page);
 361                 }
 362                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
 363                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
 364                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
 365                         __free_page(page);
 366                 }
 367                 break;
 368         }
 369         return NOTIFY_OK;
 370 }
 371 #endif /* CONFIG_HOTPLUG_CPU */
 372 #else /* !CONFIG_SMP */
 373 #define profile_flip_buffers()          do { } while (0)
 374 #define profile_discard_flip_buffers()  do { } while (0)
 375
 376 void profile_hit(int type, void *__pc)
 377 {
 378         unsigned long pc;
 379
 380         if (prof_on != type || !prof_buffer)
 381                 return;
 382         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
 383         atomic_inc(&prof_buffer[min(pc, prof_len - 1)]);
 384 }
 385 #endif /* !CONFIG_SMP */
 386
 387 void profile_tick(int type, struct pt_regs *regs)
 388 {
 389         if (type == CPU_PROFILING && timer_hook)
 390                 timer_hook(regs);
 391         if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
 392                 profile_hit(type, (void *)profile_pc(regs));
 393 }
 394
 395 #ifdef CONFIG_PROC_FS
 396 #include <linux/proc_fs.h>
 397 #include <asm/uaccess.h>
 398 #include <asm/ptrace.h>
 399
 400 static int prof_cpu_mask_read_proc (char *page, char **start, off_t off,
 401                         int count, int *eof, void *data)
 402 {
 403         int len = cpumask_scnprintf(page, count, *(cpumask_t *)data);
 404         if (count - len < 2)
 405                 return -EINVAL;
 406         len += sprintf(page + len, "\n");
 407         return len;
 408 }
 409
 410 static int prof_cpu_mask_write_proc (struct file *file, const char __user *buffer,
 411                                         unsigned long count, void *data)
 412 {
 413         cpumask_t *mask = (cpumask_t *)data;
 414         unsigned long full_count = count, err;
 415         cpumask_t new_value;
 416
 417         err = cpumask_parse(buffer, count, new_value);
 418         if (err)
 419                 return err;
 420
 421         *mask = new_value;
 422         return full_count;
 423 }
 424
 425 void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
 426 {
 427         struct proc_dir_entry *entry;
 428
 429         /* create /proc/irq/prof_cpu_mask */
 430         if (!(entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir)))
 431                 return;
 432         entry->nlink = 1;
 433         entry->data = (void *)&prof_cpu_mask;
 434         entry->read_proc = prof_cpu_mask_read_proc;
 435         entry->write_proc = prof_cpu_mask_write_proc;
 436 }
 437
 438 /*
 439  * This function accesses profiling information. The returned data is
 440  * binary: the sampling step and the actual contents of the profile
 441  * buffer. Use of the program readprofile is recommended in order to
 442  * get meaningful info out of these data.
 443  */
 444 static ssize_t
 445 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 446 {
 447         unsigned long p = *ppos;
 448         ssize_t read;
 449         char * pnt;
 450         unsigned int sample_step = 1 << prof_shift;
 451
 452         profile_flip_buffers();
 453         if (p >= (prof_len+1)*sizeof(unsigned int))
 454                 return 0;
 455         if (count > (prof_len+1)*sizeof(unsigned int) - p)
 456                 count = (prof_len+1)*sizeof(unsigned int) - p;
 457         read = 0;
 458
 459         while (p < sizeof(unsigned int) && count > 0) {
 460                 put_user(*((char *)(&sample_step)+p),buf);
 461                 buf++; p++; count--; read++;
 462         }
 463         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
 464         if (copy_to_user(buf,(void *)pnt,count))
 465                 return -EFAULT;
 466         read += count;
 467         *ppos += read;
 468         return read;
 469 }
 470
 471 /*
 472  * Writing to /proc/profile resets the counters
 473  *
 474  * Writing a 'profiling multiplier' value into it also re-sets the profiling
 475  * interrupt frequency, on architectures that support this.
 476  */
 477 static ssize_t write_profile(struct file *file, const char __user *buf,
 478                              size_t count, loff_t *ppos)
 479 {
 480 #ifdef CONFIG_SMP
 481         extern int setup_profiling_timer (unsigned int multiplier);
 482
 483         if (count == sizeof(int)) {
 484                 unsigned int multiplier;
 485
 486                 if (copy_from_user(&multiplier, buf, sizeof(int)))
 487                         return -EFAULT;
 488
 489                 if (setup_profiling_timer(multiplier))
 490                         return -EINVAL;
 491         }
 492 #endif
 493         profile_discard_flip_buffers();
 494         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
 495         return count;
 496 }
 497
 498 static struct file_operations proc_profile_operations = {
 499         .read           = read_profile,
 500         .write          = write_profile,
 501 };
 502
 503 #ifdef CONFIG_SMP
 504 static void __init profile_nop(void *unused)
 505 {
 506 }
 507
 508 static int __init create_hash_tables(void)
 509 {
 510         int cpu;
 511
 512         for_each_online_cpu(cpu) {
 513                 int node = cpu_to_node(cpu);
 514                 struct page *page;
 515
 516                 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 517                 if (!page)
 518                         goto out_cleanup;
 519                 per_cpu(cpu_profile_hits, cpu)[1]
 520                                 = (struct profile_hit *)page_address(page);
 521                 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 522                 if (!page)
 523                         goto out_cleanup;
 524                 per_cpu(cpu_profile_hits, cpu)[0]
 525                                 = (struct profile_hit *)page_address(page);
 526         }
 527         return 0;
 528 out_cleanup:
 529         prof_on = 0;
 530         smp_mb();
 531         on_each_cpu(profile_nop, NULL, 0, 1);
 532         for_each_online_cpu(cpu) {
 533                 struct page *page;
 534
 535                 if (per_cpu(cpu_profile_hits, cpu)[0]) {
 536                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
 537                         per_cpu(cpu_profile_hits, cpu)[0] = NULL;
 538                         __free_page(page);
 539                 }
 540                 if (per_cpu(cpu_profile_hits, cpu)[1]) {
 541                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
 542                         per_cpu(cpu_profile_hits, cpu)[1] = NULL;
 543                         __free_page(page);
 544                 }
 545         }
 546         return -1;
 547 }
 548 #else
 549 #define create_hash_tables()                    ({ 0; })
 550 #endif
 551
 552 static int __init create_proc_profile(void)
 553 {
 554         struct proc_dir_entry *entry;
 555
 556         if (!prof_on)
 557                 return 0;
 558         if (create_hash_tables())
 559                 return -1;
 560         if (!(entry = create_proc_entry("profile", S_IWUSR | S_IRUGO, NULL)))
 561                 return 0;
 562         entry->proc_fops = &proc_profile_operations;
 563         entry->size = (1+prof_len) * sizeof(atomic_t);
 564         hotcpu_notifier(profile_cpu_callback, 0);
 565         return 0;
 566 }
 567 module_init(create_proc_profile);
 568 #endif /* CONFIG_PROC_FS */