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