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  *      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 <linux/sched/stat.h>
  29
  30 #include <asm/sections.h>
  31 #include <asm/irq_regs.h>
  32 #include <asm/ptrace.h>
  33
  34 struct profile_hit {
  35         u32 pc, hits;
  36 };
  37 #define PROFILE_GRPSHIFT        3
  38 #define PROFILE_GRPSZ           (1 << PROFILE_GRPSHIFT)
  39 #define NR_PROFILE_HIT          (PAGE_SIZE/sizeof(struct profile_hit))
  40 #define NR_PROFILE_GRP          (NR_PROFILE_HIT/PROFILE_GRPSZ)
  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 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
  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 const char schedstr[] = "schedule";
  58         static const char sleepstr[] = "sleep";
  59         static const char kvmstr[] = "kvm";
  60         int par;
  61
  62         if (!strncmp(str, sleepstr, strlen(sleepstr))) {
  63 #ifdef CONFIG_SCHEDSTATS
  64                 force_schedstat_enabled();
  65                 prof_on = SLEEP_PROFILING;
  66                 if (str[strlen(sleepstr)] == ',')
  67                         str += strlen(sleepstr) + 1;
  68                 if (get_option(&str, &par))
  69                         prof_shift = par;
  70                 pr_info("kernel sleep profiling enabled (shift: %ld)\n",
  71                         prof_shift);
  72 #else
  73                 pr_warn("kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
  74 #endif /* CONFIG_SCHEDSTATS */
  75         } else if (!strncmp(str, schedstr, strlen(schedstr))) {
  76                 prof_on = SCHED_PROFILING;
  77                 if (str[strlen(schedstr)] == ',')
  78                         str += strlen(schedstr) + 1;
  79                 if (get_option(&str, &par))
  80                         prof_shift = par;
  81                 pr_info("kernel schedule profiling enabled (shift: %ld)\n",
  82                         prof_shift);
  83         } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
  84                 prof_on = KVM_PROFILING;
  85                 if (str[strlen(kvmstr)] == ',')
  86                         str += strlen(kvmstr) + 1;
  87                 if (get_option(&str, &par))
  88                         prof_shift = par;
  89                 pr_info("kernel KVM profiling enabled (shift: %ld)\n",
  90                         prof_shift);
  91         } else if (get_option(&str, &par)) {
  92                 prof_shift = par;
  93                 prof_on = CPU_PROFILING;
  94                 pr_info("kernel profiling enabled (shift: %ld)\n",
  95                         prof_shift);
  96         }
  97         return 1;
  98 }
  99 __setup("profile=", profile_setup);
 100
 101
 102 int __ref profile_init(void)
 103 {
 104         int buffer_bytes;
 105         if (!prof_on)
 106                 return 0;
 107
 108         /* only text is profiled */
 109         prof_len = (_etext - _stext) >> prof_shift;
 110         buffer_bytes = prof_len*sizeof(atomic_t);
 111
 112         if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
 113                 return -ENOMEM;
 114
 115         cpumask_copy(prof_cpu_mask, cpu_possible_mask);
 116
 117         prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
 118         if (prof_buffer)
 119                 return 0;
 120
 121         prof_buffer = alloc_pages_exact(buffer_bytes,
 122                                         GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
 123         if (prof_buffer)
 124                 return 0;
 125
 126         prof_buffer = vzalloc(buffer_bytes);
 127         if (prof_buffer)
 128                 return 0;
 129
 130         free_cpumask_var(prof_cpu_mask);
 131         return -ENOMEM;
 132 }
 133
 134 /* Profile event notifications */
 135
 136 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
 137 static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
 138 static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
 139
 140 void profile_task_exit(struct task_struct *task)
 141 {
 142         blocking_notifier_call_chain(&task_exit_notifier, 0, task);
 143 }
 144
 145 int profile_handoff_task(struct task_struct *task)
 146 {
 147         int ret;
 148         ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
 149         return (ret == NOTIFY_OK) ? 1 : 0;
 150 }
 151
 152 void profile_munmap(unsigned long addr)
 153 {
 154         blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
 155 }
 156
 157 int task_handoff_register(struct notifier_block *n)
 158 {
 159         return atomic_notifier_chain_register(&task_free_notifier, n);
 160 }
 161 EXPORT_SYMBOL_GPL(task_handoff_register);
 162
 163 int task_handoff_unregister(struct notifier_block *n)
 164 {
 165         return atomic_notifier_chain_unregister(&task_free_notifier, n);
 166 }
 167 EXPORT_SYMBOL_GPL(task_handoff_unregister);
 168
 169 int profile_event_register(enum profile_type type, struct notifier_block *n)
 170 {
 171         int err = -EINVAL;
 172
 173         switch (type) {
 174         case PROFILE_TASK_EXIT:
 175                 err = blocking_notifier_chain_register(
 176                                 &task_exit_notifier, n);
 177                 break;
 178         case PROFILE_MUNMAP:
 179                 err = blocking_notifier_chain_register(
 180                                 &munmap_notifier, n);
 181                 break;
 182         }
 183
 184         return err;
 185 }
 186 EXPORT_SYMBOL_GPL(profile_event_register);
 187
 188 int profile_event_unregister(enum profile_type type, struct notifier_block *n)
 189 {
 190         int err = -EINVAL;
 191
 192         switch (type) {
 193         case PROFILE_TASK_EXIT:
 194                 err = blocking_notifier_chain_unregister(
 195                                 &task_exit_notifier, n);
 196                 break;
 197         case PROFILE_MUNMAP:
 198                 err = blocking_notifier_chain_unregister(
 199                                 &munmap_notifier, n);
 200                 break;
 201         }
 202
 203         return err;
 204 }
 205 EXPORT_SYMBOL_GPL(profile_event_unregister);
 206
 207 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
 208 /*
 209  * Each cpu has a pair of open-addressed hashtables for pending
 210  * profile hits. read_profile() IPI's all cpus to request them
 211  * to flip buffers and flushes their contents to prof_buffer itself.
 212  * Flip requests are serialized by the profile_flip_mutex. The sole
 213  * use of having a second hashtable is for avoiding cacheline
 214  * contention that would otherwise happen during flushes of pending
 215  * profile hits required for the accuracy of reported profile hits
 216  * and so resurrect the interrupt livelock issue.
 217  *
 218  * The open-addressed hashtables are indexed by profile buffer slot
 219  * and hold the number of pending hits to that profile buffer slot on
 220  * a cpu in an entry. When the hashtable overflows, all pending hits
 221  * are accounted to their corresponding profile buffer slots with
 222  * atomic_add() and the hashtable emptied. As numerous pending hits
 223  * may be accounted to a profile buffer slot in a hashtable entry,
 224  * this amortizes a number of atomic profile buffer increments likely
 225  * to be far larger than the number of entries in the hashtable,
 226  * particularly given that the number of distinct profile buffer
 227  * positions to which hits are accounted during short intervals (e.g.
 228  * several seconds) is usually very small. Exclusion from buffer
 229  * flipping is provided by interrupt disablement (note that for
 230  * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
 231  * process context).
 232  * The hash function is meant to be lightweight as opposed to strong,
 233  * and was vaguely inspired by ppc64 firmware-supported inverted
 234  * pagetable hash functions, but uses a full hashtable full of finite
 235  * collision chains, not just pairs of them.
 236  *
 237  * -- nyc
 238  */
 239 static void __profile_flip_buffers(void *unused)
 240 {
 241         int cpu = smp_processor_id();
 242
 243         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
 244 }
 245
 246 static void profile_flip_buffers(void)
 247 {
 248         int i, j, cpu;
 249
 250         mutex_lock(&profile_flip_mutex);
 251         j = per_cpu(cpu_profile_flip, get_cpu());
 252         put_cpu();
 253         on_each_cpu(__profile_flip_buffers, NULL, 1);
 254         for_each_online_cpu(cpu) {
 255                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
 256                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
 257                         if (!hits[i].hits) {
 258                                 if (hits[i].pc)
 259                                         hits[i].pc = 0;
 260                                 continue;
 261                         }
 262                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
 263                         hits[i].hits = hits[i].pc = 0;
 264                 }
 265         }
 266         mutex_unlock(&profile_flip_mutex);
 267 }
 268
 269 static void profile_discard_flip_buffers(void)
 270 {
 271         int i, cpu;
 272
 273         mutex_lock(&profile_flip_mutex);
 274         i = per_cpu(cpu_profile_flip, get_cpu());
 275         put_cpu();
 276         on_each_cpu(__profile_flip_buffers, NULL, 1);
 277         for_each_online_cpu(cpu) {
 278                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
 279                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
 280         }
 281         mutex_unlock(&profile_flip_mutex);
 282 }
 283
 284 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
 285 {
 286         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
 287         int i, j, cpu;
 288         struct profile_hit *hits;
 289
 290         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
 291         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
 292         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
 293         cpu = get_cpu();
 294         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
 295         if (!hits) {
 296                 put_cpu();
 297                 return;
 298         }
 299         /*
 300          * We buffer the global profiler buffer into a per-CPU
 301          * queue and thus reduce the number of global (and possibly
 302          * NUMA-alien) accesses. The write-queue is self-coalescing:
 303          */
 304         local_irq_save(flags);
 305         do {
 306                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
 307                         if (hits[i + j].pc == pc) {
 308                                 hits[i + j].hits += nr_hits;
 309                                 goto out;
 310                         } else if (!hits[i + j].hits) {
 311                                 hits[i + j].pc = pc;
 312                                 hits[i + j].hits = nr_hits;
 313                                 goto out;
 314                         }
 315                 }
 316                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
 317         } while (i != primary);
 318
 319         /*
 320          * Add the current hit(s) and flush the write-queue out
 321          * to the global buffer:
 322          */
 323         atomic_add(nr_hits, &prof_buffer[pc]);
 324         for (i = 0; i < NR_PROFILE_HIT; ++i) {
 325                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
 326                 hits[i].pc = hits[i].hits = 0;
 327         }
 328 out:
 329         local_irq_restore(flags);
 330         put_cpu();
 331 }
 332
 333 static int profile_dead_cpu(unsigned int cpu)
 334 {
 335         struct page *page;
 336         int i;
 337
 338         if (prof_cpu_mask != NULL)
 339                 cpumask_clear_cpu(cpu, prof_cpu_mask);
 340
 341         for (i = 0; i < 2; i++) {
 342                 if (per_cpu(cpu_profile_hits, cpu)[i]) {
 343                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[i]);
 344                         per_cpu(cpu_profile_hits, cpu)[i] = NULL;
 345                         __free_page(page);
 346                 }
 347         }
 348         return 0;
 349 }
 350
 351 static int profile_prepare_cpu(unsigned int cpu)
 352 {
 353         int i, node = cpu_to_mem(cpu);
 354         struct page *page;
 355
 356         per_cpu(cpu_profile_flip, cpu) = 0;
 357
 358         for (i = 0; i < 2; i++) {
 359                 if (per_cpu(cpu_profile_hits, cpu)[i])
 360                         continue;
 361
 362                 page = __alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 363                 if (!page) {
 364                         profile_dead_cpu(cpu);
 365                         return -ENOMEM;
 366                 }
 367                 per_cpu(cpu_profile_hits, cpu)[i] = page_address(page);
 368
 369         }
 370         return 0;
 371 }
 372
 373 static int profile_online_cpu(unsigned int cpu)
 374 {
 375         if (prof_cpu_mask != NULL)
 376                 cpumask_set_cpu(cpu, prof_cpu_mask);
 377
 378         return 0;
 379 }
 380
 381 #else /* !CONFIG_SMP */
 382 #define profile_flip_buffers()          do { } while (0)
 383 #define profile_discard_flip_buffers()  do { } while (0)
 384
 385 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
 386 {
 387         unsigned long pc;
 388         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
 389         atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
 390 }
 391 #endif /* !CONFIG_SMP */
 392
 393 void profile_hits(int type, void *__pc, unsigned int nr_hits)
 394 {
 395         if (prof_on != type || !prof_buffer)
 396                 return;
 397         do_profile_hits(type, __pc, nr_hits);
 398 }
 399 EXPORT_SYMBOL_GPL(profile_hits);
 400
 401 void profile_tick(int type)
 402 {
 403         struct pt_regs *regs = get_irq_regs();
 404
 405         if (!user_mode(regs) && prof_cpu_mask != NULL &&
 406             cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
 407                 profile_hit(type, (void *)profile_pc(regs));
 408 }
 409
 410 #ifdef CONFIG_PROC_FS
 411 #include <linux/proc_fs.h>
 412 #include <linux/seq_file.h>
 413 #include <linux/uaccess.h>
 414
 415 static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
 416 {
 417         seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
 418         return 0;
 419 }
 420
 421 static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
 422 {
 423         return single_open(file, prof_cpu_mask_proc_show, NULL);
 424 }
 425
 426 static ssize_t prof_cpu_mask_proc_write(struct file *file,
 427         const char __user *buffer, size_t count, loff_t *pos)
 428 {
 429         cpumask_var_t new_value;
 430         int err;
 431
 432         if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
 433                 return -ENOMEM;
 434
 435         err = cpumask_parse_user(buffer, count, new_value);
 436         if (!err) {
 437                 cpumask_copy(prof_cpu_mask, new_value);
 438                 err = count;
 439         }
 440         free_cpumask_var(new_value);
 441         return err;
 442 }
 443
 444 static const struct file_operations prof_cpu_mask_proc_fops = {
 445         .open           = prof_cpu_mask_proc_open,
 446         .read           = seq_read,
 447         .llseek         = seq_lseek,
 448         .release        = single_release,
 449         .write          = prof_cpu_mask_proc_write,
 450 };
 451
 452 void create_prof_cpu_mask(void)
 453 {
 454         /* create /proc/irq/prof_cpu_mask */
 455         proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_fops);
 456 }
 457
 458 /*
 459  * This function accesses profiling information. The returned data is
 460  * binary: the sampling step and the actual contents of the profile
 461  * buffer. Use of the program readprofile is recommended in order to
 462  * get meaningful info out of these data.
 463  */
 464 static ssize_t
 465 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 466 {
 467         unsigned long p = *ppos;
 468         ssize_t read;
 469         char *pnt;
 470         unsigned int sample_step = 1 << prof_shift;
 471
 472         profile_flip_buffers();
 473         if (p >= (prof_len+1)*sizeof(unsigned int))
 474                 return 0;
 475         if (count > (prof_len+1)*sizeof(unsigned int) - p)
 476                 count = (prof_len+1)*sizeof(unsigned int) - p;
 477         read = 0;
 478
 479         while (p < sizeof(unsigned int) && count > 0) {
 480                 if (put_user(*((char *)(&sample_step)+p), buf))
 481                         return -EFAULT;
 482                 buf++; p++; count--; read++;
 483         }
 484         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
 485         if (copy_to_user(buf, (void *)pnt, count))
 486                 return -EFAULT;
 487         read += count;
 488         *ppos += read;
 489         return read;
 490 }
 491
 492 /*
 493  * Writing to /proc/profile resets the counters
 494  *
 495  * Writing a 'profiling multiplier' value into it also re-sets the profiling
 496  * interrupt frequency, on architectures that support this.
 497  */
 498 static ssize_t write_profile(struct file *file, const char __user *buf,
 499                              size_t count, loff_t *ppos)
 500 {
 501 #ifdef CONFIG_SMP
 502         extern int setup_profiling_timer(unsigned int multiplier);
 503
 504         if (count == sizeof(int)) {
 505                 unsigned int multiplier;
 506
 507                 if (copy_from_user(&multiplier, buf, sizeof(int)))
 508                         return -EFAULT;
 509
 510                 if (setup_profiling_timer(multiplier))
 511                         return -EINVAL;
 512         }
 513 #endif
 514         profile_discard_flip_buffers();
 515         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
 516         return count;
 517 }
 518
 519 static const struct file_operations proc_profile_operations = {
 520         .read           = read_profile,
 521         .write          = write_profile,
 522         .llseek         = default_llseek,
 523 };
 524
 525 int __ref create_proc_profile(void)
 526 {
 527         struct proc_dir_entry *entry;
 528 #ifdef CONFIG_SMP
 529         enum cpuhp_state online_state;
 530 #endif
 531
 532         int err = 0;
 533
 534         if (!prof_on)
 535                 return 0;
 536 #ifdef CONFIG_SMP
 537         err = cpuhp_setup_state(CPUHP_PROFILE_PREPARE, "PROFILE_PREPARE",
 538                                 profile_prepare_cpu, profile_dead_cpu);
 539         if (err)
 540                 return err;
 541
 542         err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "AP_PROFILE_ONLINE",
 543                                 profile_online_cpu, NULL);
 544         if (err < 0)
 545                 goto err_state_prep;
 546         online_state = err;
 547         err = 0;
 548 #endif
 549         entry = proc_create("profile", S_IWUSR | S_IRUGO,
 550                             NULL, &proc_profile_operations);
 551         if (!entry)
 552                 goto err_state_onl;
 553         proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
 554
 555         return err;
 556 err_state_onl:
 557 #ifdef CONFIG_SMP
 558         cpuhp_remove_state(online_state);
 559 err_state_prep:
 560         cpuhp_remove_state(CPUHP_PROFILE_PREPARE);
 561 #endif
 562         return err;
 563 }
 564 subsys_initcall(create_proc_profile);
 565 #endif /* CONFIG_PROC_FS */