kernel/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                         if (*p == stack_dump_trace[i]) {
 160                                 stack_dump_trace[x] = stack_dump_trace[i++];
 161                                 this_size = stack_trace_index[x++] =
 162                                         (top - p) * sizeof(unsigned long);
 163                                 found = 1;
 164                                 /* Start the search from here */
 165                                 start = p + 1;
 166                                 /*
 167                                  * We do not want to show the overhead
 168                                  * of the stack tracer stack in the
 169                                  * max stack. If we haven't figured
 170                                  * out what that is, then figure it out
 171                                  * now.
 172                                  */
 173                                 if (unlikely(!tracer_frame)) {
 174                                         tracer_frame = (p - stack) *
 175                                                 sizeof(unsigned long);
 176                                         stack_trace_max_size -= tracer_frame;
 177                                 }
 178                         }
 179                 }
 180
 181                 if (!found)
 182                         i++;
 183         }
 184
 185         stack_trace_max.nr_entries = x;
 186         for (; x < i; x++)
 187                 stack_dump_trace[x] = ULONG_MAX;
 188
 189         if (task_stack_end_corrupted(current)) {
 190                 stack_trace_print();
 191                 BUG();
 192         }
 193
 194  out:
 195         rcu_irq_exit();
 196         arch_spin_unlock(&stack_trace_max_lock);
 197         local_irq_restore(flags);
 198 }
 199
 200 static void
 201 stack_trace_call(unsigned long ip, unsigned long parent_ip,
 202                  struct ftrace_ops *op, struct pt_regs *pt_regs)
 203 {
 204         unsigned long stack;
 205         int cpu;
 206
 207         preempt_disable_notrace();
 208
 209         cpu = raw_smp_processor_id();
 210         /* no atomic needed, we only modify this variable by this cpu */
 211         if (per_cpu(trace_active, cpu)++ != 0)
 212                 goto out;
 213
 214         ip += MCOUNT_INSN_SIZE;
 215
 216         check_stack(ip, &stack);
 217
 218  out:
 219         per_cpu(trace_active, cpu)--;
 220         /* prevent recursion in schedule */
 221         preempt_enable_notrace();
 222 }
 223
 224 static struct ftrace_ops trace_ops __read_mostly =
 225 {
 226         .func = stack_trace_call,
 227         .flags = FTRACE_OPS_FL_RECURSION_SAFE,
 228 };
 229
 230 static ssize_t
 231 stack_max_size_read(struct file *filp, char __user *ubuf,
 232                     size_t count, loff_t *ppos)
 233 {
 234         unsigned long *ptr = filp->private_data;
 235         char buf[64];
 236         int r;
 237
 238         r = snprintf(buf, sizeof(buf), "%ld\n", *ptr);
 239         if (r > sizeof(buf))
 240                 r = sizeof(buf);
 241         return simple_read_from_buffer(ubuf, count, ppos, buf, r);
 242 }
 243
 244 static ssize_t
 245 stack_max_size_write(struct file *filp, const char __user *ubuf,
 246                      size_t count, loff_t *ppos)
 247 {
 248         long *ptr = filp->private_data;
 249         unsigned long val, flags;
 250         int ret;
 251         int cpu;
 252
 253         ret = kstrtoul_from_user(ubuf, count, 10, &val);
 254         if (ret)
 255                 return ret;
 256
 257         local_irq_save(flags);
 258
 259         /*
 260          * In case we trace inside arch_spin_lock() or after (NMI),
 261          * we will cause circular lock, so we also need to increase
 262          * the percpu trace_active here.
 263          */
 264         cpu = smp_processor_id();
 265         per_cpu(trace_active, cpu)++;
 266
 267         arch_spin_lock(&stack_trace_max_lock);
 268         *ptr = val;
 269         arch_spin_unlock(&stack_trace_max_lock);
 270
 271         per_cpu(trace_active, cpu)--;
 272         local_irq_restore(flags);
 273
 274         return count;
 275 }
 276
 277 static const struct file_operations stack_max_size_fops = {
 278         .open           = tracing_open_generic,
 279         .read           = stack_max_size_read,
 280         .write          = stack_max_size_write,
 281         .llseek         = default_llseek,
 282 };
 283
 284 static void *
 285 __next(struct seq_file *m, loff_t *pos)
 286 {
 287         long n = *pos - 1;
 288
 289         if (n > stack_trace_max.nr_entries || stack_dump_trace[n] == ULONG_MAX)
 290                 return NULL;
 291
 292         m->private = (void *)n;
 293         return &m->private;
 294 }
 295
 296 static void *
 297 t_next(struct seq_file *m, void *v, loff_t *pos)
 298 {
 299         (*pos)++;
 300         return __next(m, pos);
 301 }
 302
 303 static void *t_start(struct seq_file *m, loff_t *pos)
 304 {
 305         int cpu;
 306
 307         local_irq_disable();
 308
 309         cpu = smp_processor_id();
 310         per_cpu(trace_active, cpu)++;
 311
 312         arch_spin_lock(&stack_trace_max_lock);
 313
 314         if (*pos == 0)
 315                 return SEQ_START_TOKEN;
 316
 317         return __next(m, pos);
 318 }
 319
 320 static void t_stop(struct seq_file *m, void *p)
 321 {
 322         int cpu;
 323
 324         arch_spin_unlock(&stack_trace_max_lock);
 325
 326         cpu = smp_processor_id();
 327         per_cpu(trace_active, cpu)--;
 328
 329         local_irq_enable();
 330 }
 331
 332 static void trace_lookup_stack(struct seq_file *m, long i)
 333 {
 334         unsigned long addr = stack_dump_trace[i];
 335
 336         seq_printf(m, "%pS\n", (void *)addr);
 337 }
 338
 339 static void print_disabled(struct seq_file *m)
 340 {
 341         seq_puts(m, "#\n"
 342                  "#  Stack tracer disabled\n"
 343                  "#\n"
 344                  "# To enable the stack tracer, either add 'stacktrace' to the\n"
 345                  "# kernel command line\n"
 346                  "# or 'echo 1 > /proc/sys/kernel/stack_tracer_enabled'\n"
 347                  "#\n");
 348 }
 349
 350 static int t_show(struct seq_file *m, void *v)
 351 {
 352         long i;
 353         int size;
 354
 355         if (v == SEQ_START_TOKEN) {
 356                 seq_printf(m, "        Depth    Size   Location"
 357                            "    (%d entries)\n"
 358                            "        -----    ----   --------\n",
 359                            stack_trace_max.nr_entries);
 360
 361                 if (!stack_tracer_enabled && !stack_trace_max_size)
 362                         print_disabled(m);
 363
 364                 return 0;
 365         }
 366
 367         i = *(long *)v;
 368
 369         if (i >= stack_trace_max.nr_entries ||
 370             stack_dump_trace[i] == ULONG_MAX)
 371                 return 0;
 372
 373         if (i+1 == stack_trace_max.nr_entries ||
 374             stack_dump_trace[i+1] == ULONG_MAX)
 375                 size = stack_trace_index[i];
 376         else
 377                 size = stack_trace_index[i] - stack_trace_index[i+1];
 378
 379         seq_printf(m, "%3ld) %8d   %5d   ", i, stack_trace_index[i], size);
 380
 381         trace_lookup_stack(m, i);
 382
 383         return 0;
 384 }
 385
 386 static const struct seq_operations stack_trace_seq_ops = {
 387         .start          = t_start,
 388         .next           = t_next,
 389         .stop           = t_stop,
 390         .show           = t_show,
 391 };
 392
 393 static int stack_trace_open(struct inode *inode, struct file *file)
 394 {
 395         return seq_open(file, &stack_trace_seq_ops);
 396 }
 397
 398 static const struct file_operations stack_trace_fops = {
 399         .open           = stack_trace_open,
 400         .read           = seq_read,
 401         .llseek         = seq_lseek,
 402         .release        = seq_release,
 403 };
 404
 405 static int
 406 stack_trace_filter_open(struct inode *inode, struct file *file)
 407 {
 408         return ftrace_regex_open(&trace_ops, FTRACE_ITER_FILTER,
 409                                  inode, file);
 410 }
 411
 412 static const struct file_operations stack_trace_filter_fops = {
 413         .open = stack_trace_filter_open,
 414         .read = seq_read,
 415         .write = ftrace_filter_write,
 416         .llseek = tracing_lseek,
 417         .release = ftrace_regex_release,
 418 };
 419
 420 int
 421 stack_trace_sysctl(struct ctl_table *table, int write,
 422                    void __user *buffer, size_t *lenp,
 423                    loff_t *ppos)
 424 {
 425         int ret;
 426
 427         mutex_lock(&stack_sysctl_mutex);
 428
 429         ret = proc_dointvec(table, write, buffer, lenp, ppos);
 430
 431         if (ret || !write ||
 432             (last_stack_tracer_enabled == !!stack_tracer_enabled))
 433                 goto out;
 434
 435         last_stack_tracer_enabled = !!stack_tracer_enabled;
 436
 437         if (stack_tracer_enabled)
 438                 register_ftrace_function(&trace_ops);
 439         else
 440                 unregister_ftrace_function(&trace_ops);
 441
 442  out:
 443         mutex_unlock(&stack_sysctl_mutex);
 444         return ret;
 445 }
 446
 447 static char stack_trace_filter_buf[COMMAND_LINE_SIZE+1] __initdata;
 448
 449 static __init int enable_stacktrace(char *str)
 450 {
 451         if (strncmp(str, "_filter=", 8) == 0)
 452                 strncpy(stack_trace_filter_buf, str+8, COMMAND_LINE_SIZE);
 453
 454         stack_tracer_enabled = 1;
 455         last_stack_tracer_enabled = 1;
 456         return 1;
 457 }
 458 __setup("stacktrace", enable_stacktrace);
 459
 460 static __init int stack_trace_init(void)
 461 {
 462         struct dentry *d_tracer;
 463
 464         d_tracer = tracing_init_dentry();
 465         if (IS_ERR(d_tracer))
 466                 return 0;
 467
 468         trace_create_file("stack_max_size", 0644, d_tracer,
 469                         &stack_trace_max_size, &stack_max_size_fops);
 470
 471         trace_create_file("stack_trace", 0444, d_tracer,
 472                         NULL, &stack_trace_fops);
 473
 474         trace_create_file("stack_trace_filter", 0444, d_tracer,
 475                         NULL, &stack_trace_filter_fops);
 476
 477         if (stack_trace_filter_buf[0])
 478                 ftrace_set_early_filter(&trace_ops, stack_trace_filter_buf, 1);
 479
 480         if (stack_tracer_enabled)
 481                 register_ftrace_function(&trace_ops);
 482
 483         return 0;
 484 }
 485
 486 device_initcall(stack_trace_init);