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 /* Some archs may not define MCOUNT_INSN_SIZE */
 205 #ifndef MCOUNT_INSN_SIZE
 206 # define MCOUNT_INSN_SIZE 0
 207 #endif
 208
 209 static void
 210 stack_trace_call(unsigned long ip, unsigned long parent_ip,
 211                  struct ftrace_ops *op, struct pt_regs *pt_regs)
 212 {
 213         unsigned long stack;
 214         int cpu;
 215
 216         preempt_disable_notrace();
 217
 218         cpu = raw_smp_processor_id();
 219         /* no atomic needed, we only modify this variable by this cpu */
 220         if (per_cpu(trace_active, cpu)++ != 0)
 221                 goto out;
 222
 223         ip += MCOUNT_INSN_SIZE;
 224
 225         check_stack(ip, &stack);
 226
 227  out:
 228         per_cpu(trace_active, cpu)--;
 229         /* prevent recursion in schedule */
 230         preempt_enable_notrace();
 231 }
 232
 233 static struct ftrace_ops trace_ops __read_mostly =
 234 {
 235         .func = stack_trace_call,
 236         .flags = FTRACE_OPS_FL_RECURSION_SAFE,
 237 };
 238
 239 static ssize_t
 240 stack_max_size_read(struct file *filp, char __user *ubuf,
 241                     size_t count, loff_t *ppos)
 242 {
 243         unsigned long *ptr = filp->private_data;
 244         char buf[64];
 245         int r;
 246
 247         r = snprintf(buf, sizeof(buf), "%ld\n", *ptr);
 248         if (r > sizeof(buf))
 249                 r = sizeof(buf);
 250         return simple_read_from_buffer(ubuf, count, ppos, buf, r);
 251 }
 252
 253 static ssize_t
 254 stack_max_size_write(struct file *filp, const char __user *ubuf,
 255                      size_t count, loff_t *ppos)
 256 {
 257         long *ptr = filp->private_data;
 258         unsigned long val, flags;
 259         int ret;
 260         int cpu;
 261
 262         ret = kstrtoul_from_user(ubuf, count, 10, &val);
 263         if (ret)
 264                 return ret;
 265
 266         local_irq_save(flags);
 267
 268         /*
 269          * In case we trace inside arch_spin_lock() or after (NMI),
 270          * we will cause circular lock, so we also need to increase
 271          * the percpu trace_active here.
 272          */
 273         cpu = smp_processor_id();
 274         per_cpu(trace_active, cpu)++;
 275
 276         arch_spin_lock(&stack_trace_max_lock);
 277         *ptr = val;
 278         arch_spin_unlock(&stack_trace_max_lock);
 279
 280         per_cpu(trace_active, cpu)--;
 281         local_irq_restore(flags);
 282
 283         return count;
 284 }
 285
 286 static const struct file_operations stack_max_size_fops = {
 287         .open           = tracing_open_generic,
 288         .read           = stack_max_size_read,
 289         .write          = stack_max_size_write,
 290         .llseek         = default_llseek,
 291 };
 292
 293 static void *
 294 __next(struct seq_file *m, loff_t *pos)
 295 {
 296         long n = *pos - 1;
 297
 298         if (n > stack_trace_max.nr_entries || stack_dump_trace[n] == ULONG_MAX)
 299                 return NULL;
 300
 301         m->private = (void *)n;
 302         return &m->private;
 303 }
 304
 305 static void *
 306 t_next(struct seq_file *m, void *v, loff_t *pos)
 307 {
 308         (*pos)++;
 309         return __next(m, pos);
 310 }
 311
 312 static void *t_start(struct seq_file *m, loff_t *pos)
 313 {
 314         int cpu;
 315
 316         local_irq_disable();
 317
 318         cpu = smp_processor_id();
 319         per_cpu(trace_active, cpu)++;
 320
 321         arch_spin_lock(&stack_trace_max_lock);
 322
 323         if (*pos == 0)
 324                 return SEQ_START_TOKEN;
 325
 326         return __next(m, pos);
 327 }
 328
 329 static void t_stop(struct seq_file *m, void *p)
 330 {
 331         int cpu;
 332
 333         arch_spin_unlock(&stack_trace_max_lock);
 334
 335         cpu = smp_processor_id();
 336         per_cpu(trace_active, cpu)--;
 337
 338         local_irq_enable();
 339 }
 340
 341 static void trace_lookup_stack(struct seq_file *m, long i)
 342 {
 343         unsigned long addr = stack_dump_trace[i];
 344
 345         seq_printf(m, "%pS\n", (void *)addr);
 346 }
 347
 348 static void print_disabled(struct seq_file *m)
 349 {
 350         seq_puts(m, "#\n"
 351                  "#  Stack tracer disabled\n"
 352                  "#\n"
 353                  "# To enable the stack tracer, either add 'stacktrace' to the\n"
 354                  "# kernel command line\n"
 355                  "# or 'echo 1 > /proc/sys/kernel/stack_tracer_enabled'\n"
 356                  "#\n");
 357 }
 358
 359 static int t_show(struct seq_file *m, void *v)
 360 {
 361         long i;
 362         int size;
 363
 364         if (v == SEQ_START_TOKEN) {
 365                 seq_printf(m, "        Depth    Size   Location"
 366                            "    (%d entries)\n"
 367                            "        -----    ----   --------\n",
 368                            stack_trace_max.nr_entries);
 369
 370                 if (!stack_tracer_enabled && !stack_trace_max_size)
 371                         print_disabled(m);
 372
 373                 return 0;
 374         }
 375
 376         i = *(long *)v;
 377
 378         if (i >= stack_trace_max.nr_entries ||
 379             stack_dump_trace[i] == ULONG_MAX)
 380                 return 0;
 381
 382         if (i+1 == stack_trace_max.nr_entries ||
 383             stack_dump_trace[i+1] == ULONG_MAX)
 384                 size = stack_trace_index[i];
 385         else
 386                 size = stack_trace_index[i] - stack_trace_index[i+1];
 387
 388         seq_printf(m, "%3ld) %8d   %5d   ", i, stack_trace_index[i], size);
 389
 390         trace_lookup_stack(m, i);
 391
 392         return 0;
 393 }
 394
 395 static const struct seq_operations stack_trace_seq_ops = {
 396         .start          = t_start,
 397         .next           = t_next,
 398         .stop           = t_stop,
 399         .show           = t_show,
 400 };
 401
 402 static int stack_trace_open(struct inode *inode, struct file *file)
 403 {
 404         return seq_open(file, &stack_trace_seq_ops);
 405 }
 406
 407 static const struct file_operations stack_trace_fops = {
 408         .open           = stack_trace_open,
 409         .read           = seq_read,
 410         .llseek         = seq_lseek,
 411         .release        = seq_release,
 412 };
 413
 414 static int
 415 stack_trace_filter_open(struct inode *inode, struct file *file)
 416 {
 417         return ftrace_regex_open(&trace_ops, FTRACE_ITER_FILTER,
 418                                  inode, file);
 419 }
 420
 421 static const struct file_operations stack_trace_filter_fops = {
 422         .open = stack_trace_filter_open,
 423         .read = seq_read,
 424         .write = ftrace_filter_write,
 425         .llseek = tracing_lseek,
 426         .release = ftrace_regex_release,
 427 };
 428
 429 int
 430 stack_trace_sysctl(struct ctl_table *table, int write,
 431                    void __user *buffer, size_t *lenp,
 432                    loff_t *ppos)
 433 {
 434         int ret;
 435
 436         mutex_lock(&stack_sysctl_mutex);
 437
 438         ret = proc_dointvec(table, write, buffer, lenp, ppos);
 439
 440         if (ret || !write ||
 441             (last_stack_tracer_enabled == !!stack_tracer_enabled))
 442                 goto out;
 443
 444         last_stack_tracer_enabled = !!stack_tracer_enabled;
 445
 446         if (stack_tracer_enabled)
 447                 register_ftrace_function(&trace_ops);
 448         else
 449                 unregister_ftrace_function(&trace_ops);
 450
 451  out:
 452         mutex_unlock(&stack_sysctl_mutex);
 453         return ret;
 454 }
 455
 456 static char stack_trace_filter_buf[COMMAND_LINE_SIZE+1] __initdata;
 457
 458 static __init int enable_stacktrace(char *str)
 459 {
 460         if (strncmp(str, "_filter=", 8) == 0)
 461                 strncpy(stack_trace_filter_buf, str+8, COMMAND_LINE_SIZE);
 462
 463         stack_tracer_enabled = 1;
 464         last_stack_tracer_enabled = 1;
 465         return 1;
 466 }
 467 __setup("stacktrace", enable_stacktrace);
 468
 469 static __init int stack_trace_init(void)
 470 {
 471         struct dentry *d_tracer;
 472
 473         d_tracer = tracing_init_dentry();
 474         if (IS_ERR(d_tracer))
 475                 return 0;
 476
 477         trace_create_file("stack_max_size", 0644, d_tracer,
 478                         &stack_trace_max_size, &stack_max_size_fops);
 479
 480         trace_create_file("stack_trace", 0444, d_tracer,
 481                         NULL, &stack_trace_fops);
 482
 483         trace_create_file("stack_trace_filter", 0444, d_tracer,
 484                         NULL, &stack_trace_filter_fops);
 485
 486         if (stack_trace_filter_buf[0])
 487                 ftrace_set_early_filter(&trace_ops, stack_trace_filter_buf, 1);
 488
 489         if (stack_tracer_enabled)
 490                 register_ftrace_function(&trace_ops);
 491
 492         return 0;
 493 }
 494
 495 device_initcall(stack_trace_init);