kernel/trace/trace_stack.c

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