kernel/trace/trace_stack.c

   1 /*
   2  * Copyright (C) 2008 Steven Rostedt <[email protected]>
   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         /*
 100          * There's a slight chance that we are tracing inside the
 101          * RCU infrastructure, and rcu_irq_enter() will not work
 102          * as expected.
 103          */
 104         if (unlikely(rcu_irq_enter_disabled()))
 105                 return;
 106
 107         local_irq_save(flags);
 108         arch_spin_lock(&stack_trace_max_lock);
 109
 110         /*
 111          * RCU may not be watching, make it see us.
 112          * The stack trace code uses rcu_sched.
 113          */
 114         rcu_irq_enter();
 115
 116         /* In case another CPU set the tracer_frame on us */
 117         if (unlikely(!frame_size))
 118                 this_size -= tracer_frame;
 119
 120         /* a race could have already updated it */
 121         if (this_size <= stack_trace_max_size)
 122                 goto out;
 123
 124         stack_trace_max_size = this_size;
 125
 126         stack_trace_max.nr_entries = 0;
 127         stack_trace_max.skip = 3;
 128
 129         save_stack_trace(&stack_trace_max);
 130
 131         /* Skip over the overhead of the stack tracer itself */
 132         for (i = 0; i < stack_trace_max.nr_entries; i++) {
 133                 if (stack_dump_trace[i] == ip)
 134                         break;
 135         }
 136
 137         /*
 138          * Some archs may not have the passed in ip in the dump.
 139          * If that happens, we need to show everything.
 140          */
 141         if (i == stack_trace_max.nr_entries)
 142                 i = 0;
 143
 144         /*
 145          * Now find where in the stack these are.
 146          */
 147         x = 0;
 148         start = stack;
 149         top = (unsigned long *)
 150                 (((unsigned long)start & ~(THREAD_SIZE-1)) + THREAD_SIZE);
 151
 152         /*
 153          * Loop through all the entries. One of the entries may
 154          * for some reason be missed on the stack, so we may
 155          * have to account for them. If they are all there, this
 156          * loop will only happen once. This code only takes place
 157          * on a new max, so it is far from a fast path.
 158          */
 159         while (i < stack_trace_max.nr_entries) {
 160                 int found = 0;
 161
 162                 stack_trace_index[x] = this_size;
 163                 p = start;
 164
 165                 for (; p < top && i < stack_trace_max.nr_entries; p++) {
 166                         if (stack_dump_trace[i] == ULONG_MAX)
 167                                 break;
 168                         /*
 169                          * The READ_ONCE_NOCHECK is used to let KASAN know that
 170                          * this is not a stack-out-of-bounds error.
 171                          */
 172                         if ((READ_ONCE_NOCHECK(*p)) == stack_dump_trace[i]) {
 173                                 stack_dump_trace[x] = stack_dump_trace[i++];
 174                                 this_size = stack_trace_index[x++] =
 175                                         (top - p) * sizeof(unsigned long);
 176                                 found = 1;
 177                                 /* Start the search from here */
 178                                 start = p + 1;
 179                                 /*
 180                                  * We do not want to show the overhead
 181                                  * of the stack tracer stack in the
 182                                  * max stack. If we haven't figured
 183                                  * out what that is, then figure it out
 184                                  * now.
 185                                  */
 186                                 if (unlikely(!tracer_frame)) {
 187                                         tracer_frame = (p - stack) *
 188                                                 sizeof(unsigned long);
 189                                         stack_trace_max_size -= tracer_frame;
 190                                 }
 191                         }
 192                 }
 193
 194                 if (!found)
 195                         i++;
 196         }
 197
 198         stack_trace_max.nr_entries = x;
 199         for (; x < i; x++)
 200                 stack_dump_trace[x] = ULONG_MAX;
 201
 202         if (task_stack_end_corrupted(current)) {
 203                 stack_trace_print();
 204                 BUG();
 205         }
 206
 207  out:
 208         rcu_irq_exit();
 209         arch_spin_unlock(&stack_trace_max_lock);
 210         local_irq_restore(flags);
 211 }
 212
 213 static void
 214 stack_trace_call(unsigned long ip, unsigned long parent_ip,
 215                  struct ftrace_ops *op, struct pt_regs *pt_regs)
 216 {
 217         unsigned long stack;
 218
 219         preempt_disable_notrace();
 220
 221         /* no atomic needed, we only modify this variable by this cpu */
 222         __this_cpu_inc(disable_stack_tracer);
 223         if (__this_cpu_read(disable_stack_tracer) != 1)
 224                 goto out;
 225
 226         ip += MCOUNT_INSN_SIZE;
 227
 228         check_stack(ip, &stack);
 229
 230  out:
 231         __this_cpu_dec(disable_stack_tracer);
 232         /* prevent recursion in schedule */
 233         preempt_enable_notrace();
 234 }
 235
 236 static struct ftrace_ops trace_ops __read_mostly =
 237 {
 238         .func = stack_trace_call,
 239         .flags = FTRACE_OPS_FL_RECURSION_SAFE,
 240 };
 241
 242 static ssize_t
 243 stack_max_size_read(struct file *filp, char __user *ubuf,
 244                     size_t count, loff_t *ppos)
 245 {
 246         unsigned long *ptr = filp->private_data;
 247         char buf[64];
 248         int r;
 249
 250         r = snprintf(buf, sizeof(buf), "%ld\n", *ptr);
 251         if (r > sizeof(buf))
 252                 r = sizeof(buf);
 253         return simple_read_from_buffer(ubuf, count, ppos, buf, r);
 254 }
 255
 256 static ssize_t
 257 stack_max_size_write(struct file *filp, const char __user *ubuf,
 258                      size_t count, loff_t *ppos)
 259 {
 260         long *ptr = filp->private_data;
 261         unsigned long val, flags;
 262         int ret;
 263
 264         ret = kstrtoul_from_user(ubuf, count, 10, &val);
 265         if (ret)
 266                 return ret;
 267
 268         local_irq_save(flags);
 269
 270         /*
 271          * In case we trace inside arch_spin_lock() or after (NMI),
 272          * we will cause circular lock, so we also need to increase
 273          * the percpu disable_stack_tracer here.
 274          */
 275         __this_cpu_inc(disable_stack_tracer);
 276
 277         arch_spin_lock(&stack_trace_max_lock);
 278         *ptr = val;
 279         arch_spin_unlock(&stack_trace_max_lock);
 280
 281         __this_cpu_dec(disable_stack_tracer);
 282         local_irq_restore(flags);
 283
 284         return count;
 285 }
 286
 287 static const struct file_operations stack_max_size_fops = {
 288         .open           = tracing_open_generic,
 289         .read           = stack_max_size_read,
 290         .write          = stack_max_size_write,
 291         .llseek         = default_llseek,
 292 };
 293
 294 static void *
 295 __next(struct seq_file *m, loff_t *pos)
 296 {
 297         long n = *pos - 1;
 298
 299         if (n > stack_trace_max.nr_entries || stack_dump_trace[n] == ULONG_MAX)
 300                 return NULL;
 301
 302         m->private = (void *)n;
 303         return &m->private;
 304 }
 305
 306 static void *
 307 t_next(struct seq_file *m, void *v, loff_t *pos)
 308 {
 309         (*pos)++;
 310         return __next(m, pos);
 311 }
 312
 313 static void *t_start(struct seq_file *m, loff_t *pos)
 314 {
 315         local_irq_disable();
 316
 317         __this_cpu_inc(disable_stack_tracer);
 318
 319         arch_spin_lock(&stack_trace_max_lock);
 320
 321         if (*pos == 0)
 322                 return SEQ_START_TOKEN;
 323
 324         return __next(m, pos);
 325 }
 326
 327 static void t_stop(struct seq_file *m, void *p)
 328 {
 329         arch_spin_unlock(&stack_trace_max_lock);
 330
 331         __this_cpu_dec(disable_stack_tracer);
 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);