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;
  44 static unsigned short int prof_shift;
  45
  46 int prof_on __read_mostly;
  47 EXPORT_SYMBOL_GPL(prof_on);
  48
  49 static cpumask_var_t prof_cpu_mask;
  50 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
  51 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
  52 static DEFINE_PER_CPU(int, cpu_profile_flip);
  53 static DEFINE_MUTEX(profile_flip_mutex);
  54 #endif /* CONFIG_SMP */
  55
  56 int profile_setup(char *str)
  57 {
  58         static const char schedstr[] = "schedule";
  59         static const char sleepstr[] = "sleep";
  60         static const char kvmstr[] = "kvm";
  61         int par;
  62
  63         if (!strncmp(str, sleepstr, strlen(sleepstr))) {
  64 #ifdef CONFIG_SCHEDSTATS
  65                 force_schedstat_enabled();
  66                 prof_on = SLEEP_PROFILING;
  67                 if (str[strlen(sleepstr)] == ',')
  68                         str += strlen(sleepstr) + 1;
  69                 if (get_option(&str, &par))
  70                         prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
  71                 pr_info("kernel sleep profiling enabled (shift: %u)\n",
  72                         prof_shift);
  73 #else
  74                 pr_warn("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 = clamp(par, 0, BITS_PER_LONG - 1);
  82                 pr_info("kernel schedule profiling enabled (shift: %u)\n",
  83                         prof_shift);
  84         } else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
  85                 prof_on = KVM_PROFILING;
  86                 if (str[strlen(kvmstr)] == ',')
  87                         str += strlen(kvmstr) + 1;
  88                 if (get_option(&str, &par))
  89                         prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
  90                 pr_info("kernel KVM profiling enabled (shift: %u)\n",
  91                         prof_shift);
  92         } else if (get_option(&str, &par)) {
  93                 prof_shift = clamp(par, 0, BITS_PER_LONG - 1);
  94                 prof_on = CPU_PROFILING;
  95                 pr_info("kernel profiling enabled (shift: %u)\n",
  96                         prof_shift);
  97         }
  98         return 1;
  99 }
 100 __setup("profile=", profile_setup);
 101
 102
 103 int __ref profile_init(void)
 104 {
 105         int buffer_bytes;
 106         if (!prof_on)
 107                 return 0;
 108
 109         /* only text is profiled */
 110         prof_len = (_etext - _stext) >> prof_shift;
 111
 112         if (!prof_len) {
 113                 pr_warn("profiling shift: %u too large\n", prof_shift);
 114                 prof_on = 0;
 115                 return -EINVAL;
 116         }
 117
 118         buffer_bytes = prof_len*sizeof(atomic_t);
 119
 120         if (!alloc_cpumask_var(&prof_cpu_mask, GFP_KERNEL))
 121                 return -ENOMEM;
 122
 123         cpumask_copy(prof_cpu_mask, cpu_possible_mask);
 124
 125         prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL|__GFP_NOWARN);
 126         if (prof_buffer)
 127                 return 0;
 128
 129         prof_buffer = alloc_pages_exact(buffer_bytes,
 130                                         GFP_KERNEL|__GFP_ZERO|__GFP_NOWARN);
 131         if (prof_buffer)
 132                 return 0;
 133
 134         prof_buffer = vzalloc(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 #if defined(CONFIG_SMP) && defined(CONFIG_PROC_FS)
 216 /*
 217  * Each cpu has a pair of open-addressed hashtables for pending
 218  * profile hits. read_profile() IPI's all cpus to request them
 219  * to flip buffers and flushes their contents to prof_buffer itself.
 220  * Flip requests are serialized by the profile_flip_mutex. The sole
 221  * use of having a second hashtable is for avoiding cacheline
 222  * contention that would otherwise happen during flushes of pending
 223  * profile hits required for the accuracy of reported profile hits
 224  * and so resurrect the interrupt livelock issue.
 225  *
 226  * The open-addressed hashtables are indexed by profile buffer slot
 227  * and hold the number of pending hits to that profile buffer slot on
 228  * a cpu in an entry. When the hashtable overflows, all pending hits
 229  * are accounted to their corresponding profile buffer slots with
 230  * atomic_add() and the hashtable emptied. As numerous pending hits
 231  * may be accounted to a profile buffer slot in a hashtable entry,
 232  * this amortizes a number of atomic profile buffer increments likely
 233  * to be far larger than the number of entries in the hashtable,
 234  * particularly given that the number of distinct profile buffer
 235  * positions to which hits are accounted during short intervals (e.g.
 236  * several seconds) is usually very small. Exclusion from buffer
 237  * flipping is provided by interrupt disablement (note that for
 238  * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
 239  * process context).
 240  * The hash function is meant to be lightweight as opposed to strong,
 241  * and was vaguely inspired by ppc64 firmware-supported inverted
 242  * pagetable hash functions, but uses a full hashtable full of finite
 243  * collision chains, not just pairs of them.
 244  *
 245  * -- nyc
 246  */
 247 static void __profile_flip_buffers(void *unused)
 248 {
 249         int cpu = smp_processor_id();
 250
 251         per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
 252 }
 253
 254 static void profile_flip_buffers(void)
 255 {
 256         int i, j, cpu;
 257
 258         mutex_lock(&profile_flip_mutex);
 259         j = per_cpu(cpu_profile_flip, get_cpu());
 260         put_cpu();
 261         on_each_cpu(__profile_flip_buffers, NULL, 1);
 262         for_each_online_cpu(cpu) {
 263                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
 264                 for (i = 0; i < NR_PROFILE_HIT; ++i) {
 265                         if (!hits[i].hits) {
 266                                 if (hits[i].pc)
 267                                         hits[i].pc = 0;
 268                                 continue;
 269                         }
 270                         atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
 271                         hits[i].hits = hits[i].pc = 0;
 272                 }
 273         }
 274         mutex_unlock(&profile_flip_mutex);
 275 }
 276
 277 static void profile_discard_flip_buffers(void)
 278 {
 279         int i, cpu;
 280
 281         mutex_lock(&profile_flip_mutex);
 282         i = per_cpu(cpu_profile_flip, get_cpu());
 283         put_cpu();
 284         on_each_cpu(__profile_flip_buffers, NULL, 1);
 285         for_each_online_cpu(cpu) {
 286                 struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
 287                 memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
 288         }
 289         mutex_unlock(&profile_flip_mutex);
 290 }
 291
 292 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
 293 {
 294         unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
 295         int i, j, cpu;
 296         struct profile_hit *hits;
 297
 298         pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
 299         i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
 300         secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
 301         cpu = get_cpu();
 302         hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
 303         if (!hits) {
 304                 put_cpu();
 305                 return;
 306         }
 307         /*
 308          * We buffer the global profiler buffer into a per-CPU
 309          * queue and thus reduce the number of global (and possibly
 310          * NUMA-alien) accesses. The write-queue is self-coalescing:
 311          */
 312         local_irq_save(flags);
 313         do {
 314                 for (j = 0; j < PROFILE_GRPSZ; ++j) {
 315                         if (hits[i + j].pc == pc) {
 316                                 hits[i + j].hits += nr_hits;
 317                                 goto out;
 318                         } else if (!hits[i + j].hits) {
 319                                 hits[i + j].pc = pc;
 320                                 hits[i + j].hits = nr_hits;
 321                                 goto out;
 322                         }
 323                 }
 324                 i = (i + secondary) & (NR_PROFILE_HIT - 1);
 325         } while (i != primary);
 326
 327         /*
 328          * Add the current hit(s) and flush the write-queue out
 329          * to the global buffer:
 330          */
 331         atomic_add(nr_hits, &prof_buffer[pc]);
 332         for (i = 0; i < NR_PROFILE_HIT; ++i) {
 333                 atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
 334                 hits[i].pc = hits[i].hits = 0;
 335         }
 336 out:
 337         local_irq_restore(flags);
 338         put_cpu();
 339 }
 340
 341 static int profile_dead_cpu(unsigned int cpu)
 342 {
 343         struct page *page;
 344         int i;
 345
 346         if (prof_cpu_mask != NULL)
 347                 cpumask_clear_cpu(cpu, prof_cpu_mask);
 348
 349         for (i = 0; i < 2; i++) {
 350                 if (per_cpu(cpu_profile_hits, cpu)[i]) {
 351                         page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[i]);
 352                         per_cpu(cpu_profile_hits, cpu)[i] = NULL;
 353                         __free_page(page);
 354                 }
 355         }
 356         return 0;
 357 }
 358
 359 static int profile_prepare_cpu(unsigned int cpu)
 360 {
 361         int i, node = cpu_to_mem(cpu);
 362         struct page *page;
 363
 364         per_cpu(cpu_profile_flip, cpu) = 0;
 365
 366         for (i = 0; i < 2; i++) {
 367                 if (per_cpu(cpu_profile_hits, cpu)[i])
 368                         continue;
 369
 370                 page = __alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 371                 if (!page) {
 372                         profile_dead_cpu(cpu);
 373                         return -ENOMEM;
 374                 }
 375                 per_cpu(cpu_profile_hits, cpu)[i] = page_address(page);
 376
 377         }
 378         return 0;
 379 }
 380
 381 static int profile_online_cpu(unsigned int cpu)
 382 {
 383         if (prof_cpu_mask != NULL)
 384                 cpumask_set_cpu(cpu, prof_cpu_mask);
 385
 386         return 0;
 387 }
 388
 389 #else /* !CONFIG_SMP */
 390 #define profile_flip_buffers()          do { } while (0)
 391 #define profile_discard_flip_buffers()  do { } while (0)
 392
 393 static void do_profile_hits(int type, void *__pc, unsigned int nr_hits)
 394 {
 395         unsigned long pc;
 396         pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
 397         atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
 398 }
 399 #endif /* !CONFIG_SMP */
 400
 401 void profile_hits(int type, void *__pc, unsigned int nr_hits)
 402 {
 403         if (prof_on != type || !prof_buffer)
 404                 return;
 405         do_profile_hits(type, __pc, nr_hits);
 406 }
 407 EXPORT_SYMBOL_GPL(profile_hits);
 408
 409 void profile_tick(int type)
 410 {
 411         struct pt_regs *regs = get_irq_regs();
 412
 413         if (!user_mode(regs) && prof_cpu_mask != NULL &&
 414             cpumask_test_cpu(smp_processor_id(), prof_cpu_mask))
 415                 profile_hit(type, (void *)profile_pc(regs));
 416 }
 417
 418 #ifdef CONFIG_PROC_FS
 419 #include <linux/proc_fs.h>
 420 #include <linux/seq_file.h>
 421 #include <linux/uaccess.h>
 422
 423 static int prof_cpu_mask_proc_show(struct seq_file *m, void *v)
 424 {
 425         seq_printf(m, "%*pb\n", cpumask_pr_args(prof_cpu_mask));
 426         return 0;
 427 }
 428
 429 static int prof_cpu_mask_proc_open(struct inode *inode, struct file *file)
 430 {
 431         return single_open(file, prof_cpu_mask_proc_show, NULL);
 432 }
 433
 434 static ssize_t prof_cpu_mask_proc_write(struct file *file,
 435         const char __user *buffer, size_t count, loff_t *pos)
 436 {
 437         cpumask_var_t new_value;
 438         int err;
 439
 440         if (!alloc_cpumask_var(&new_value, GFP_KERNEL))
 441                 return -ENOMEM;
 442
 443         err = cpumask_parse_user(buffer, count, new_value);
 444         if (!err) {
 445                 cpumask_copy(prof_cpu_mask, new_value);
 446                 err = count;
 447         }
 448         free_cpumask_var(new_value);
 449         return err;
 450 }
 451
 452 static const struct file_operations prof_cpu_mask_proc_fops = {
 453         .open           = prof_cpu_mask_proc_open,
 454         .read           = seq_read,
 455         .llseek         = seq_lseek,
 456         .release        = single_release,
 457         .write          = prof_cpu_mask_proc_write,
 458 };
 459
 460 void create_prof_cpu_mask(void)
 461 {
 462         /* create /proc/irq/prof_cpu_mask */
 463         proc_create("irq/prof_cpu_mask", 0600, NULL, &prof_cpu_mask_proc_fops);
 464 }
 465
 466 /*
 467  * This function accesses profiling information. The returned data is
 468  * binary: the sampling step and the actual contents of the profile
 469  * buffer. Use of the program readprofile is recommended in order to
 470  * get meaningful info out of these data.
 471  */
 472 static ssize_t
 473 read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 474 {
 475         unsigned long p = *ppos;
 476         ssize_t read;
 477         char *pnt;
 478         unsigned long sample_step = 1UL << prof_shift;
 479
 480         profile_flip_buffers();
 481         if (p >= (prof_len+1)*sizeof(unsigned int))
 482                 return 0;
 483         if (count > (prof_len+1)*sizeof(unsigned int) - p)
 484                 count = (prof_len+1)*sizeof(unsigned int) - p;
 485         read = 0;
 486
 487         while (p < sizeof(unsigned int) && count > 0) {
 488                 if (put_user(*((char *)(&sample_step)+p), buf))
 489                         return -EFAULT;
 490                 buf++; p++; count--; read++;
 491         }
 492         pnt = (char *)prof_buffer + p - sizeof(atomic_t);
 493         if (copy_to_user(buf, (void *)pnt, count))
 494                 return -EFAULT;
 495         read += count;
 496         *ppos += read;
 497         return read;
 498 }
 499
 500 /*
 501  * Writing to /proc/profile resets the counters
 502  *
 503  * Writing a 'profiling multiplier' value into it also re-sets the profiling
 504  * interrupt frequency, on architectures that support this.
 505  */
 506 static ssize_t write_profile(struct file *file, const char __user *buf,
 507                              size_t count, loff_t *ppos)
 508 {
 509 #ifdef CONFIG_SMP
 510         extern int setup_profiling_timer(unsigned int multiplier);
 511
 512         if (count == sizeof(int)) {
 513                 unsigned int multiplier;
 514
 515                 if (copy_from_user(&multiplier, buf, sizeof(int)))
 516                         return -EFAULT;
 517
 518                 if (setup_profiling_timer(multiplier))
 519                         return -EINVAL;
 520         }
 521 #endif
 522         profile_discard_flip_buffers();
 523         memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
 524         return count;
 525 }
 526
 527 static const struct file_operations proc_profile_operations = {
 528         .read           = read_profile,
 529         .write          = write_profile,
 530         .llseek         = default_llseek,
 531 };
 532
 533 int __ref create_proc_profile(void)
 534 {
 535         struct proc_dir_entry *entry;
 536 #ifdef CONFIG_SMP
 537         enum cpuhp_state online_state;
 538 #endif
 539
 540         int err = 0;
 541
 542         if (!prof_on)
 543                 return 0;
 544 #ifdef CONFIG_SMP
 545         err = cpuhp_setup_state(CPUHP_PROFILE_PREPARE, "PROFILE_PREPARE",
 546                                 profile_prepare_cpu, profile_dead_cpu);
 547         if (err)
 548                 return err;
 549
 550         err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "AP_PROFILE_ONLINE",
 551                                 profile_online_cpu, NULL);
 552         if (err < 0)
 553                 goto err_state_prep;
 554         online_state = err;
 555         err = 0;
 556 #endif
 557         entry = proc_create("profile", S_IWUSR | S_IRUGO,
 558                             NULL, &proc_profile_operations);
 559         if (!entry)
 560                 goto err_state_onl;
 561         proc_set_size(entry, (1 + prof_len) * sizeof(atomic_t));
 562
 563         return err;
 564 err_state_onl:
 565 #ifdef CONFIG_SMP
 566         cpuhp_remove_state(online_state);
 567 err_state_prep:
 568         cpuhp_remove_state(CPUHP_PROFILE_PREPARE);
 569 #endif
 570         return err;
 571 }
 572 subsys_initcall(create_proc_profile);
 573 #endif /* CONFIG_PROC_FS */