kernel/trace/trace_stack.c

   1 /*
   2  * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
   3  *
   4  */
   5 #include <linux/stacktrace.h>
   6 #include <linux/kallsyms.h>
   7 #include <linux/seq_file.h>
   8 #include <linux/spinlock.h>
   9 #include <linux/uaccess.h>
  10 #include <linux/ftrace.h>
  11 #include <linux/module.h>
  12 #include <linux/sysctl.h>
  13 #include <linux/init.h>
  14
  15 #include <asm/setup.h>
  16
  17 #include "trace.h"
  18
  19 static unsigned long stack_dump_trace[STACK_TRACE_ENTRIES+1] =
  20          { [0 ... (STACK_TRACE_ENTRIES)] = ULONG_MAX };
  21 unsigned stack_trace_index[STACK_TRACE_ENTRIES];
  22
  23 /*
  24  * Reserve one entry for the passed in ip. This will allow
  25  * us to remove most or all of the stack size overhead
  26  * added by the stack tracer itself.
  27  */
  28 struct stack_trace stack_trace_max = {
  29         .max_entries            = STACK_TRACE_ENTRIES - 1,
  30         .entries                = &stack_dump_trace[0],
  31 };
  32
  33 unsigned long stack_trace_max_size;
  34 arch_spinlock_t stack_trace_max_lock =
  35         (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
  36
  37 static DEFINE_PER_CPU(int, trace_active);
  38 static DEFINE_MUTEX(stack_sysctl_mutex);
  39
  40 int stack_tracer_enabled;
  41 static int last_stack_tracer_enabled;
  42
  43 void stack_trace_print(void)
  44 {
  45         long i;
  46         int size;
  47
  48         pr_emerg("        Depth    Size   Location    (%d entries)\n"
  49                            "        -----    ----   --------\n",
  50                            stack_trace_max.nr_entries);
  51
  52         for (i = 0; i < stack_trace_max.nr_entries; i++) {
  53                 if (stack_dump_trace[i] == ULONG_MAX)
  54                         break;
  55                 if (i+1 == stack_trace_max.nr_entries ||
  56                                 stack_dump_trace[i+1] == ULONG_MAX)
  57                         size = stack_trace_index[i];
  58                 else
  59                         size = stack_trace_index[i] - stack_trace_index[i+1];
  60
  61                 pr_emerg("%3ld) %8d   %5d   %pS\n", i, stack_trace_index[i],
  62                                 size, (void *)stack_dump_trace[i]);
  63         }
  64 }
  65
  66 /*
  67  * When arch-specific code overides this function, the following
  68  * data should be filled up, assuming stack_trace_max_lock is held to
  69  * prevent concurrent updates.
  70  *     stack_trace_index[]
  71  *     stack_trace_max
  72  *     stack_trace_max_size
  73  */
  74 void __weak
  75 check_stack(unsigned long ip, unsigned long *stack)
  76 {
  77         unsigned long this_size, flags; unsigned long *p, *top, *start;
  78         static int tracer_frame;
  79         int frame_size = ACCESS_ONCE(tracer_frame);
  80         int i, x;
  81
  82         this_size = ((unsigned long)stack) & (THREAD_SIZE-1);
  83         this_size = THREAD_SIZE - this_size;
  84         /* Remove the frame of the tracer */
  85         this_size -= frame_size;
  86
  87         if (this_size <= stack_trace_max_size)
  88                 return;
  89
  90         /* we do not handle interrupt stacks yet */
  91         if (!object_is_on_stack(stack))
  92                 return;
  93
  94         /* Can't do this from NMI context (can cause deadlocks) */
  95         if (in_nmi())
  96                 return;
  97
  98         local_irq_save(flags);
  99         arch_spin_lock(&stack_trace_max_lock);
 100
 101         /*
 102          * RCU may not be watching, make it see us.
 103          * The stack trace code uses rcu_sched.
 104          */
 105         rcu_irq_enter();
 106
 107         /* In case another CPU set the tracer_frame on us */
 108         if (unlikely(!frame_size))
 109                 this_size -= tracer_frame;
 110
 111         /* a race could have already updated it */
 112         if (this_size <= stack_trace_max_size)
 113                 goto out;
 114
 115         stack_trace_max_size = this_size;
 116
 117         stack_trace_max.nr_entries = 0;
 118         stack_trace_max.skip = 3;
 119
 120         save_stack_trace(&stack_trace_max);
 121
 122         /* Skip over the overhead of the stack tracer itself */
 123         for (i = 0; i < stack_trace_max.nr_entries; i++) {
 124                 if (stack_dump_trace[i] == ip)
 125                         break;
 126         }
 127
 128         /*
 129          * Some archs may not have the passed in ip in the dump.
 130          * If that happens, we need to show everything.
 131          */
 132         if (i == stack_trace_max.nr_entries)
 133                 i = 0;
 134
 135         /*
 136          * Now find where in the stack these are.
 137          */
 138         x = 0;
 139         start = stack;
 140         top = (unsigned long *)
 141                 (((unsigned long)start & ~(THREAD_SIZE-1)) + THREAD_SIZE);
 142
 143         /*
 144          * Loop through all the entries. One of the entries may
 145          * for some reason be missed on the stack, so we may
 146          * have to account for them. If they are all there, this
 147          * loop will only happen once. This code only takes place
 148          * on a new max, so it is far from a fast path.
 149          */
 150         while (i < stack_trace_max.nr_entries) {
 151                 int found = 0;
 152
 153                 stack_trace_index[x] = this_size;
 154                 p = start;
 155
 156                 for (; p < top && i < stack_trace_max.nr_entries; p++) {
 157                         if (stack_dump_trace[i] == ULONG_MAX)
 158                                 break;
 159                         /*
 160                          * The READ_ONCE_NOCHECK is used to let KASAN know that
 161                          * this is not a stack-out-of-bounds error.
 162                          */
 163                         if ((READ_ONCE_NOCHECK(*p)) == stack_dump_trace[i]) {
 164                                 stack_dump_trace[x] = stack_dump_trace[i++];
 165                                 this_size = stack_trace_index[x++] =
 166                                         (top - p) * sizeof(unsigned long);
 167                                 found = 1;
 168                                 /* Start the search from here */
 169                                 start = p + 1;
 170                                 /*
 171                                  * We do not want to show the overhead
 172                                  * of the stack tracer stack in the
 173                                  * max stack. If we haven't figured
 174                                  * out what that is, then figure it out
 175                                  * now.
 176                                  */
 177                                 if (unlikely(!tracer_frame)) {
 178                                         tracer_frame = (p - stack) *
 179                                                 sizeof(unsigned long);
 180                                         stack_trace_max_size -= tracer_frame;
 181                                 }
 182                         }
 183                 }
 184
 185                 if (!found)
 186                         i++;
 187         }
 188
 189         stack_trace_max.nr_entries = x;
 190         for (; x < i; x++)
 191                 stack_dump_trace[x] = ULONG_MAX;
 192
 193         if (task_stack_end_corrupted(current)) {
 194                 stack_trace_print();
 195                 BUG();
 196         }
 197
 198  out:
 199         rcu_irq_exit();
 200         arch_spin_unlock(&stack_trace_max_lock);
 201         local_irq_restore(flags);
 202 }
 203
 204 static void
 205 stack_trace_call(unsigned long ip, unsigned long parent_ip,
 206                  struct ftrace_ops *op, struct pt_regs *pt_regs)
 207 {
 208         unsigned long stack;
 209         int cpu;
 210
 211         preempt_disable_notrace();
 212
 213         cpu = raw_smp_processor_id();
 214         /* no atomic needed, we only modify this variable by this cpu */
 215         if (per_cpu(trace_active, cpu)++ != 0)
 216                 goto out;
 217
 218         ip += MCOUNT_INSN_SIZE;
 219
 220         check_stack(ip, &stack);
 221
 222  out:
 223         per_cpu(trace_active, cpu)--;
 224         /* prevent recursion in schedule */
 225         preempt_enable_notrace();
 226 }
 227
 228 static struct ftrace_ops trace_ops __read_mostly =
 229 {
 230         .func = stack_trace_call,
 231         .flags = FTRACE_OPS_FL_RECURSION_SAFE,
 232 };
 233
 234 static ssize_t
 235 stack_max_size_read(struct file *filp, char __user *ubuf,
 236                     size_t count, loff_t *ppos)
 237 {
 238         unsigned long *ptr = filp->private_data;
 239         char buf[64];
 240         int r;
 241
 242         r = snprintf(buf, sizeof(buf), "%ld\n", *ptr);
 243         if (r > sizeof(buf))
 244                 r = sizeof(buf);
 245         return simple_read_from_buffer(ubuf, count, ppos, buf, r);
 246 }
 247
 248 static ssize_t
 249 stack_max_size_write(struct file *filp, const char __user *ubuf,
 250                      size_t count, loff_t *ppos)
 251 {
 252         long *ptr = filp->private_data;
 253         unsigned long val, flags;
 254         int ret;
 255         int cpu;
 256
 257         ret = kstrtoul_from_user(ubuf, count, 10, &val);
 258         if (ret)
 259                 return ret;
 260
 261         local_irq_save(flags);
 262
 263         /*
 264          * In case we trace inside arch_spin_lock() or after (NMI),
 265          * we will cause circular lock, so we also need to increase
 266          * the percpu trace_active here.
 267          */
 268         cpu = smp_processor_id();
 269         per_cpu(trace_active, cpu)++;
 270
 271         arch_spin_lock(&stack_trace_max_lock);
 272         *ptr = val;
 273         arch_spin_unlock(&stack_trace_max_lock);
 274
 275         per_cpu(trace_active, cpu)--;
 276         local_irq_restore(flags);
 277
 278         return count;
 279 }
 280
 281 static const struct file_operations stack_max_size_fops = {
 282         .open           = tracing_open_generic,
 283         .read           = stack_max_size_read,
 284         .write          = stack_max_size_write,
 285         .llseek         = default_llseek,
 286 };
 287
 288 static void *
 289 __next(struct seq_file *m, loff_t *pos)
 290 {
 291         long n = *pos - 1;
 292
 293         if (n > stack_trace_max.nr_entries || stack_dump_trace[n] == ULONG_MAX)
 294                 return NULL;
 295
 296         m->private = (void *)n;
 297         return &m->private;
 298 }
 299
 300 static void *
 301 t_next(struct seq_file *m, void *v, loff_t *pos)
 302 {
 303         (*pos)++;
 304         return __next(m, pos);
 305 }
 306
 307 static void *t_start(struct seq_file *m, loff_t *pos)
 308 {
 309         int cpu;
 310
 311         local_irq_disable();
 312
 313         cpu = smp_processor_id();
 314         per_cpu(trace_active, cpu)++;
 315
 316         arch_spin_lock(&stack_trace_max_lock);
 317
 318         if (*pos == 0)
 319                 return SEQ_START_TOKEN;
 320
 321         return __next(m, pos);
 322 }
 323
 324 static void t_stop(struct seq_file *m, void *p)
 325 {
 326         int cpu;
 327
 328         arch_spin_unlock(&stack_trace_max_lock);
 329
 330         cpu = smp_processor_id();
 331         per_cpu(trace_active, cpu)--;
 332
 333         local_irq_enable();
 334 }
 335
 336 static void trace_lookup_stack(struct seq_file *m, long i)
 337 {
 338         unsigned long addr = stack_dump_trace[i];
 339
 340         seq_printf(m, "%pS\n", (void *)addr);
 341 }
 342
 343 static void print_disabled(struct seq_file *m)
 344 {
 345         seq_puts(m, "#\n"
 346                  "#  Stack tracer disabled\n"
 347                  "#\n"
 348                  "# To enable the stack tracer, either add 'stacktrace' to the\n"
 349                  "# kernel command line\n"
 350                  "# or 'echo 1 > /proc/sys/kernel/stack_tracer_enabled'\n"
 351                  "#\n");
 352 }
 353
 354 static int t_show(struct seq_file *m, void *v)
 355 {
 356         long i;
 357         int size;
 358
 359         if (v == SEQ_START_TOKEN) {
 360                 seq_printf(m, "        Depth    Size   Location"
 361                            "    (%d entries)\n"
 362                            "        -----    ----   --------\n",
 363                            stack_trace_max.nr_entries);
 364
 365                 if (!stack_tracer_enabled && !stack_trace_max_size)
 366                         print_disabled(m);
 367
 368                 return 0;
 369         }
 370
 371         i = *(long *)v;
 372
 373         if (i >= stack_trace_max.nr_entries ||
 374             stack_dump_trace[i] == ULONG_MAX)
 375                 return 0;
 376
 377         if (i+1 == stack_trace_max.nr_entries ||
 378             stack_dump_trace[i+1] == ULONG_MAX)
 379                 size = stack_trace_index[i];
 380         else
 381                 size = stack_trace_index[i] - stack_trace_index[i+1];
 382
 383         seq_printf(m, "%3ld) %8d   %5d   ", i, stack_trace_index[i], size);
 384
 385         trace_lookup_stack(m, i);
 386
 387         return 0;
 388 }
 389
 390 static const struct seq_operations stack_trace_seq_ops = {
 391         .start          = t_start,
 392         .next           = t_next,
 393         .stop           = t_stop,
 394         .show           = t_show,
 395 };
 396
 397 static int stack_trace_open(struct inode *inode, struct file *file)
 398 {
 399         return seq_open(file, &stack_trace_seq_ops);
 400 }
 401
 402 static const struct file_operations stack_trace_fops = {
 403         .open           = stack_trace_open,
 404         .read           = seq_read,
 405         .llseek         = seq_lseek,
 406         .release        = seq_release,
 407 };
 408
 409 static int
 410 stack_trace_filter_open(struct inode *inode, struct file *file)
 411 {
 412         return ftrace_regex_open(&trace_ops, FTRACE_ITER_FILTER,
 413                                  inode, file);
 414 }
 415
 416 static const struct file_operations stack_trace_filter_fops = {
 417         .open = stack_trace_filter_open,
 418         .read = seq_read,
 419         .write = ftrace_filter_write,
 420         .llseek = tracing_lseek,
 421         .release = ftrace_regex_release,
 422 };
 423
 424 int
 425 stack_trace_sysctl(struct ctl_table *table, int write,
 426                    void __user *buffer, size_t *lenp,
 427                    loff_t *ppos)
 428 {
 429         int ret;
 430
 431         mutex_lock(&stack_sysctl_mutex);
 432
 433         ret = proc_dointvec(table, write, buffer, lenp, ppos);
 434
 435         if (ret || !write ||
 436             (last_stack_tracer_enabled == !!stack_tracer_enabled))
 437                 goto out;
 438
 439         last_stack_tracer_enabled = !!stack_tracer_enabled;
 440
 441         if (stack_tracer_enabled)
 442                 register_ftrace_function(&trace_ops);
 443         else
 444                 unregister_ftrace_function(&trace_ops);
 445
 446  out:
 447         mutex_unlock(&stack_sysctl_mutex);
 448         return ret;
 449 }
 450
 451 static char stack_trace_filter_buf[COMMAND_LINE_SIZE+1] __initdata;
 452
 453 static __init int enable_stacktrace(char *str)
 454 {
 455         if (strncmp(str, "_filter=", 8) == 0)
 456                 strncpy(stack_trace_filter_buf, str+8, COMMAND_LINE_SIZE);
 457
 458         stack_tracer_enabled = 1;
 459         last_stack_tracer_enabled = 1;
 460         return 1;
 461 }
 462 __setup("stacktrace", enable_stacktrace);
 463
 464 static __init int stack_trace_init(void)
 465 {
 466         struct dentry *d_tracer;
 467
 468         d_tracer = tracing_init_dentry();
 469         if (IS_ERR(d_tracer))
 470                 return 0;
 471
 472         trace_create_file("stack_max_size", 0644, d_tracer,
 473                         &stack_trace_max_size, &stack_max_size_fops);
 474
 475         trace_create_file("stack_trace", 0444, d_tracer,
 476                         NULL, &stack_trace_fops);
 477
 478         trace_create_file("stack_trace_filter", 0444, d_tracer,
 479                         NULL, &stack_trace_filter_fops);
 480
 481         if (stack_trace_filter_buf[0])
 482                 ftrace_set_early_filter(&trace_ops, stack_trace_filter_buf, 1);
 483
 484         if (stack_tracer_enabled)
 485                 register_ftrace_function(&trace_ops);
 486
 487         return 0;
 488 }
 489
 490 device_initcall(stack_trace_init);