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 DEFINE_PER_CPU(int, disable_stack_tracer);
  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 overrides 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         /* In case another CPU set the tracer_frame on us */
 103         if (unlikely(!frame_size))
 104                 this_size -= tracer_frame;
 105
 106         /* a race could have already updated it */
 107         if (this_size <= stack_trace_max_size)
 108                 goto out;
 109
 110         stack_trace_max_size = this_size;
 111
 112         stack_trace_max.nr_entries = 0;
 113         stack_trace_max.skip = 3;
 114
 115         save_stack_trace(&stack_trace_max);
 116
 117         /* Skip over the overhead of the stack tracer itself */
 118         for (i = 0; i < stack_trace_max.nr_entries; i++) {
 119                 if (stack_dump_trace[i] == ip)
 120                         break;
 121         }
 122
 123         /*
 124          * Some archs may not have the passed in ip in the dump.
 125          * If that happens, we need to show everything.
 126          */
 127         if (i == stack_trace_max.nr_entries)
 128                 i = 0;
 129
 130         /*
 131          * Now find where in the stack these are.
 132          */
 133         x = 0;
 134         start = stack;
 135         top = (unsigned long *)
 136                 (((unsigned long)start & ~(THREAD_SIZE-1)) + THREAD_SIZE);
 137
 138         /*
 139          * Loop through all the entries. One of the entries may
 140          * for some reason be missed on the stack, so we may
 141          * have to account for them. If they are all there, this
 142          * loop will only happen once. This code only takes place
 143          * on a new max, so it is far from a fast path.
 144          */
 145         while (i < stack_trace_max.nr_entries) {
 146                 int found = 0;
 147
 148                 stack_trace_index[x] = this_size;
 149                 p = start;
 150
 151                 for (; p < top && i < stack_trace_max.nr_entries; p++) {
 152                         if (stack_dump_trace[i] == ULONG_MAX)
 153                                 break;
 154                         /*
 155                          * The READ_ONCE_NOCHECK is used to let KASAN know that
 156                          * this is not a stack-out-of-bounds error.
 157                          */
 158                         if ((READ_ONCE_NOCHECK(*p)) == stack_dump_trace[i]) {
 159                                 stack_dump_trace[x] = stack_dump_trace[i++];
 160                                 this_size = stack_trace_index[x++] =
 161                                         (top - p) * sizeof(unsigned long);
 162                                 found = 1;
 163                                 /* Start the search from here */
 164                                 start = p + 1;
 165                                 /*
 166                                  * We do not want to show the overhead
 167                                  * of the stack tracer stack in the
 168                                  * max stack. If we haven't figured
 169                                  * out what that is, then figure it out
 170                                  * now.
 171                                  */
 172                                 if (unlikely(!tracer_frame)) {
 173                                         tracer_frame = (p - stack) *
 174                                                 sizeof(unsigned long);
 175                                         stack_trace_max_size -= tracer_frame;
 176                                 }
 177                         }
 178                 }
 179
 180                 if (!found)
 181                         i++;
 182         }
 183
 184         stack_trace_max.nr_entries = x;
 185         for (; x < i; x++)
 186                 stack_dump_trace[x] = ULONG_MAX;
 187
 188         if (task_stack_end_corrupted(current)) {
 189                 stack_trace_print();
 190                 BUG();
 191         }
 192
 193  out:
 194         arch_spin_unlock(&stack_trace_max_lock);
 195         local_irq_restore(flags);
 196 }
 197
 198 static void
 199 stack_trace_call(unsigned long ip, unsigned long parent_ip,
 200                  struct ftrace_ops *op, struct pt_regs *pt_regs)
 201 {
 202         unsigned long stack;
 203
 204         preempt_disable_notrace();
 205
 206         /* no atomic needed, we only modify this variable by this cpu */
 207         __this_cpu_inc(disable_stack_tracer);
 208         if (__this_cpu_read(disable_stack_tracer) != 1)
 209                 goto out;
 210
 211         ip += MCOUNT_INSN_SIZE;
 212
 213         check_stack(ip, &stack);
 214
 215  out:
 216         __this_cpu_dec(disable_stack_tracer);
 217         /* prevent recursion in schedule */
 218         preempt_enable_notrace();
 219 }
 220
 221 static struct ftrace_ops trace_ops __read_mostly =
 222 {
 223         .func = stack_trace_call,
 224         .flags = FTRACE_OPS_FL_RECURSION_SAFE,
 225 };
 226
 227 static ssize_t
 228 stack_max_size_read(struct file *filp, char __user *ubuf,
 229                     size_t count, loff_t *ppos)
 230 {
 231         unsigned long *ptr = filp->private_data;
 232         char buf[64];
 233         int r;
 234
 235         r = snprintf(buf, sizeof(buf), "%ld\n", *ptr);
 236         if (r > sizeof(buf))
 237                 r = sizeof(buf);
 238         return simple_read_from_buffer(ubuf, count, ppos, buf, r);
 239 }
 240
 241 static ssize_t
 242 stack_max_size_write(struct file *filp, const char __user *ubuf,
 243                      size_t count, loff_t *ppos)
 244 {
 245         long *ptr = filp->private_data;
 246         unsigned long val, flags;
 247         int ret;
 248
 249         ret = kstrtoul_from_user(ubuf, count, 10, &val);
 250         if (ret)
 251                 return ret;
 252
 253         local_irq_save(flags);
 254
 255         /*
 256          * In case we trace inside arch_spin_lock() or after (NMI),
 257          * we will cause circular lock, so we also need to increase
 258          * the percpu disable_stack_tracer here.
 259          */
 260         __this_cpu_inc(disable_stack_tracer);
 261
 262         arch_spin_lock(&stack_trace_max_lock);
 263         *ptr = val;
 264         arch_spin_unlock(&stack_trace_max_lock);
 265
 266         __this_cpu_dec(disable_stack_tracer);
 267         local_irq_restore(flags);
 268
 269         return count;
 270 }
 271
 272 static const struct file_operations stack_max_size_fops = {
 273         .open           = tracing_open_generic,
 274         .read           = stack_max_size_read,
 275         .write          = stack_max_size_write,
 276         .llseek         = default_llseek,
 277 };
 278
 279 static void *
 280 __next(struct seq_file *m, loff_t *pos)
 281 {
 282         long n = *pos - 1;
 283
 284         if (n > stack_trace_max.nr_entries || stack_dump_trace[n] == ULONG_MAX)
 285                 return NULL;
 286
 287         m->private = (void *)n;
 288         return &m->private;
 289 }
 290
 291 static void *
 292 t_next(struct seq_file *m, void *v, loff_t *pos)
 293 {
 294         (*pos)++;
 295         return __next(m, pos);
 296 }
 297
 298 static void *t_start(struct seq_file *m, loff_t *pos)
 299 {
 300         local_irq_disable();
 301
 302         __this_cpu_inc(disable_stack_tracer);
 303
 304         arch_spin_lock(&stack_trace_max_lock);
 305
 306         if (*pos == 0)
 307                 return SEQ_START_TOKEN;
 308
 309         return __next(m, pos);
 310 }
 311
 312 static void t_stop(struct seq_file *m, void *p)
 313 {
 314         arch_spin_unlock(&stack_trace_max_lock);
 315
 316         __this_cpu_dec(disable_stack_tracer);
 317
 318         local_irq_enable();
 319 }
 320
 321 static void trace_lookup_stack(struct seq_file *m, long i)
 322 {
 323         unsigned long addr = stack_dump_trace[i];
 324
 325         seq_printf(m, "%pS\n", (void *)addr);
 326 }
 327
 328 static void print_disabled(struct seq_file *m)
 329 {
 330         seq_puts(m, "#\n"
 331                  "#  Stack tracer disabled\n"
 332                  "#\n"
 333                  "# To enable the stack tracer, either add 'stacktrace' to the\n"
 334                  "# kernel command line\n"
 335                  "# or 'echo 1 > /proc/sys/kernel/stack_tracer_enabled'\n"
 336                  "#\n");
 337 }
 338
 339 static int t_show(struct seq_file *m, void *v)
 340 {
 341         long i;
 342         int size;
 343
 344         if (v == SEQ_START_TOKEN) {
 345                 seq_printf(m, "        Depth    Size   Location"
 346                            "    (%d entries)\n"
 347                            "        -----    ----   --------\n",
 348                            stack_trace_max.nr_entries);
 349
 350                 if (!stack_tracer_enabled && !stack_trace_max_size)
 351                         print_disabled(m);
 352
 353                 return 0;
 354         }
 355
 356         i = *(long *)v;
 357
 358         if (i >= stack_trace_max.nr_entries ||
 359             stack_dump_trace[i] == ULONG_MAX)
 360                 return 0;
 361
 362         if (i+1 == stack_trace_max.nr_entries ||
 363             stack_dump_trace[i+1] == ULONG_MAX)
 364                 size = stack_trace_index[i];
 365         else
 366                 size = stack_trace_index[i] - stack_trace_index[i+1];
 367
 368         seq_printf(m, "%3ld) %8d   %5d   ", i, stack_trace_index[i], size);
 369
 370         trace_lookup_stack(m, i);
 371
 372         return 0;
 373 }
 374
 375 static const struct seq_operations stack_trace_seq_ops = {
 376         .start          = t_start,
 377         .next           = t_next,
 378         .stop           = t_stop,
 379         .show           = t_show,
 380 };
 381
 382 static int stack_trace_open(struct inode *inode, struct file *file)
 383 {
 384         return seq_open(file, &stack_trace_seq_ops);
 385 }
 386
 387 static const struct file_operations stack_trace_fops = {
 388         .open           = stack_trace_open,
 389         .read           = seq_read,
 390         .llseek         = seq_lseek,
 391         .release        = seq_release,
 392 };
 393
 394 #ifdef CONFIG_DYNAMIC_FTRACE
 395
 396 static int
 397 stack_trace_filter_open(struct inode *inode, struct file *file)
 398 {
 399         struct ftrace_ops *ops = inode->i_private;
 400
 401         return ftrace_regex_open(ops, FTRACE_ITER_FILTER,
 402                                  inode, file);
 403 }
 404
 405 static const struct file_operations stack_trace_filter_fops = {
 406         .open = stack_trace_filter_open,
 407         .read = seq_read,
 408         .write = ftrace_filter_write,
 409         .llseek = tracing_lseek,
 410         .release = ftrace_regex_release,
 411 };
 412
 413 #endif /* CONFIG_DYNAMIC_FTRACE */
 414
 415 int
 416 stack_trace_sysctl(struct ctl_table *table, int write,
 417                    void __user *buffer, size_t *lenp,
 418                    loff_t *ppos)
 419 {
 420         int ret;
 421
 422         mutex_lock(&stack_sysctl_mutex);
 423
 424         ret = proc_dointvec(table, write, buffer, lenp, ppos);
 425
 426         if (ret || !write ||
 427             (last_stack_tracer_enabled == !!stack_tracer_enabled))
 428                 goto out;
 429
 430         last_stack_tracer_enabled = !!stack_tracer_enabled;
 431
 432         if (stack_tracer_enabled)
 433                 register_ftrace_function(&trace_ops);
 434         else
 435                 unregister_ftrace_function(&trace_ops);
 436
 437  out:
 438         mutex_unlock(&stack_sysctl_mutex);
 439         return ret;
 440 }
 441
 442 static char stack_trace_filter_buf[COMMAND_LINE_SIZE+1] __initdata;
 443
 444 static __init int enable_stacktrace(char *str)
 445 {
 446         if (strncmp(str, "_filter=", 8) == 0)
 447                 strncpy(stack_trace_filter_buf, str+8, COMMAND_LINE_SIZE);
 448
 449         stack_tracer_enabled = 1;
 450         last_stack_tracer_enabled = 1;
 451         return 1;
 452 }
 453 __setup("stacktrace", enable_stacktrace);
 454
 455 static __init int stack_trace_init(void)
 456 {
 457         struct dentry *d_tracer;
 458
 459         d_tracer = tracing_init_dentry();
 460         if (IS_ERR(d_tracer))
 461                 return 0;
 462
 463         trace_create_file("stack_max_size", 0644, d_tracer,
 464                         &stack_trace_max_size, &stack_max_size_fops);
 465
 466         trace_create_file("stack_trace", 0444, d_tracer,
 467                         NULL, &stack_trace_fops);
 468
 469 #ifdef CONFIG_DYNAMIC_FTRACE
 470         trace_create_file("stack_trace_filter", 0444, d_tracer,
 471                           &trace_ops, &stack_trace_filter_fops);
 472 #endif
 473
 474         if (stack_trace_filter_buf[0])
 475                 ftrace_set_early_filter(&trace_ops, stack_trace_filter_buf, 1);
 476
 477         if (stack_tracer_enabled)
 478                 register_ftrace_function(&trace_ops);
 479
 480         return 0;
 481 }
 482
 483 device_initcall(stack_trace_init);