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 #define STACK_TRACE_ENTRIES 500
  20
  21 static unsigned long stack_dump_trace[STACK_TRACE_ENTRIES+1] =
  22          { [0 ... (STACK_TRACE_ENTRIES)] = ULONG_MAX };
  23 static unsigned stack_dump_index[STACK_TRACE_ENTRIES];
  24
  25 /*
  26  * Reserve one entry for the passed in ip. This will allow
  27  * us to remove most or all of the stack size overhead
  28  * added by the stack tracer itself.
  29  */
  30 static struct stack_trace max_stack_trace = {
  31         .max_entries            = STACK_TRACE_ENTRIES - 1,
  32         .entries                = &stack_dump_trace[0],
  33 };
  34
  35 static unsigned long max_stack_size;
  36 static arch_spinlock_t max_stack_lock =
  37         (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
  38
  39 static DEFINE_PER_CPU(int, trace_active);
  40 static DEFINE_MUTEX(stack_sysctl_mutex);
  41
  42 int stack_tracer_enabled;
  43 static int last_stack_tracer_enabled;
  44
  45 static inline void print_max_stack(void)
  46 {
  47         long i;
  48         int size;
  49
  50         pr_emerg("        Depth    Size   Location    (%d entries)\n"
  51                            "        -----    ----   --------\n",
  52                            max_stack_trace.nr_entries);
  53
  54         for (i = 0; i < max_stack_trace.nr_entries; i++) {
  55                 if (stack_dump_trace[i] == ULONG_MAX)
  56                         break;
  57                 if (i+1 == max_stack_trace.nr_entries ||
  58                                 stack_dump_trace[i+1] == ULONG_MAX)
  59                         size = stack_dump_index[i];
  60                 else
  61                         size = stack_dump_index[i] - stack_dump_index[i+1];
  62
  63                 pr_emerg("%3ld) %8d   %5d   %pS\n", i, stack_dump_index[i],
  64                                 size, (void *)stack_dump_trace[i]);
  65         }
  66 }
  67
  68 static inline void
  69 check_stack(unsigned long ip, unsigned long *stack)
  70 {
  71         unsigned long this_size, flags; unsigned long *p, *top, *start;
  72         static int tracer_frame;
  73         int frame_size = ACCESS_ONCE(tracer_frame);
  74         int i, x;
  75
  76         this_size = ((unsigned long)stack) & (THREAD_SIZE-1);
  77         this_size = THREAD_SIZE - this_size;
  78         /* Remove the frame of the tracer */
  79         this_size -= frame_size;
  80
  81         if (this_size <= max_stack_size)
  82                 return;
  83
  84         /* we do not handle interrupt stacks yet */
  85         if (!object_is_on_stack(stack))
  86                 return;
  87
  88         local_irq_save(flags);
  89         arch_spin_lock(&max_stack_lock);
  90
  91         /*
  92          * RCU may not be watching, make it see us.
  93          * The stack trace code uses rcu_sched.
  94          */
  95         rcu_irq_enter();
  96
  97         /* In case another CPU set the tracer_frame on us */
  98         if (unlikely(!frame_size))
  99                 this_size -= tracer_frame;
 100
 101         /* a race could have already updated it */
 102         if (this_size <= max_stack_size)
 103                 goto out;
 104
 105         max_stack_size = this_size;
 106
 107         max_stack_trace.nr_entries = 0;
 108         max_stack_trace.skip = 3;
 109
 110         save_stack_trace(&max_stack_trace);
 111
 112         /* Skip over the overhead of the stack tracer itself */
 113         for (i = 0; i < max_stack_trace.nr_entries; i++) {
 114                 if (stack_dump_trace[i] == ip)
 115                         break;
 116         }
 117
 118         /*
 119          * Now find where in the stack these are.
 120          */
 121         x = 0;
 122         start = stack;
 123         top = (unsigned long *)
 124                 (((unsigned long)start & ~(THREAD_SIZE-1)) + THREAD_SIZE);
 125
 126         /*
 127          * Loop through all the entries. One of the entries may
 128          * for some reason be missed on the stack, so we may
 129          * have to account for them. If they are all there, this
 130          * loop will only happen once. This code only takes place
 131          * on a new max, so it is far from a fast path.
 132          */
 133         while (i < max_stack_trace.nr_entries) {
 134                 int found = 0;
 135
 136                 stack_dump_index[x] = this_size;
 137                 p = start;
 138
 139                 for (; p < top && i < max_stack_trace.nr_entries; p++) {
 140                         if (stack_dump_trace[i] == ULONG_MAX)
 141                                 break;
 142                         if (*p == stack_dump_trace[i]) {
 143                                 stack_dump_trace[x] = stack_dump_trace[i++];
 144                                 this_size = stack_dump_index[x++] =
 145                                         (top - p) * sizeof(unsigned long);
 146                                 found = 1;
 147                                 /* Start the search from here */
 148                                 start = p + 1;
 149                                 /*
 150                                  * We do not want to show the overhead
 151                                  * of the stack tracer stack in the
 152                                  * max stack. If we haven't figured
 153                                  * out what that is, then figure it out
 154                                  * now.
 155                                  */
 156                                 if (unlikely(!tracer_frame)) {
 157                                         tracer_frame = (p - stack) *
 158                                                 sizeof(unsigned long);
 159                                         max_stack_size -= tracer_frame;
 160                                 }
 161                         }
 162                 }
 163
 164                 if (!found)
 165                         i++;
 166         }
 167
 168         max_stack_trace.nr_entries = x;
 169         for (; x < i; x++)
 170                 stack_dump_trace[x] = ULONG_MAX;
 171
 172         if (task_stack_end_corrupted(current)) {
 173                 print_max_stack();
 174                 BUG();
 175         }
 176
 177  out:
 178         rcu_irq_exit();
 179         arch_spin_unlock(&max_stack_lock);
 180         local_irq_restore(flags);
 181 }
 182
 183 static void
 184 stack_trace_call(unsigned long ip, unsigned long parent_ip,
 185                  struct ftrace_ops *op, struct pt_regs *pt_regs)
 186 {
 187         unsigned long stack;
 188         int cpu;
 189
 190         preempt_disable_notrace();
 191
 192         cpu = raw_smp_processor_id();
 193         /* no atomic needed, we only modify this variable by this cpu */
 194         if (per_cpu(trace_active, cpu)++ != 0)
 195                 goto out;
 196
 197         ip += MCOUNT_INSN_SIZE;
 198
 199         check_stack(ip, &stack);
 200
 201  out:
 202         per_cpu(trace_active, cpu)--;
 203         /* prevent recursion in schedule */
 204         preempt_enable_notrace();
 205 }
 206
 207 static struct ftrace_ops trace_ops __read_mostly =
 208 {
 209         .func = stack_trace_call,
 210         .flags = FTRACE_OPS_FL_RECURSION_SAFE,
 211 };
 212
 213 static ssize_t
 214 stack_max_size_read(struct file *filp, char __user *ubuf,
 215                     size_t count, loff_t *ppos)
 216 {
 217         unsigned long *ptr = filp->private_data;
 218         char buf[64];
 219         int r;
 220
 221         r = snprintf(buf, sizeof(buf), "%ld\n", *ptr);
 222         if (r > sizeof(buf))
 223                 r = sizeof(buf);
 224         return simple_read_from_buffer(ubuf, count, ppos, buf, r);
 225 }
 226
 227 static ssize_t
 228 stack_max_size_write(struct file *filp, const char __user *ubuf,
 229                      size_t count, loff_t *ppos)
 230 {
 231         long *ptr = filp->private_data;
 232         unsigned long val, flags;
 233         int ret;
 234         int cpu;
 235
 236         ret = kstrtoul_from_user(ubuf, count, 10, &val);
 237         if (ret)
 238                 return ret;
 239
 240         local_irq_save(flags);
 241
 242         /*
 243          * In case we trace inside arch_spin_lock() or after (NMI),
 244          * we will cause circular lock, so we also need to increase
 245          * the percpu trace_active here.
 246          */
 247         cpu = smp_processor_id();
 248         per_cpu(trace_active, cpu)++;
 249
 250         arch_spin_lock(&max_stack_lock);
 251         *ptr = val;
 252         arch_spin_unlock(&max_stack_lock);
 253
 254         per_cpu(trace_active, cpu)--;
 255         local_irq_restore(flags);
 256
 257         return count;
 258 }
 259
 260 static const struct file_operations stack_max_size_fops = {
 261         .open           = tracing_open_generic,
 262         .read           = stack_max_size_read,
 263         .write          = stack_max_size_write,
 264         .llseek         = default_llseek,
 265 };
 266
 267 static void *
 268 __next(struct seq_file *m, loff_t *pos)
 269 {
 270         long n = *pos - 1;
 271
 272         if (n > max_stack_trace.nr_entries || stack_dump_trace[n] == ULONG_MAX)
 273                 return NULL;
 274
 275         m->private = (void *)n;
 276         return &m->private;
 277 }
 278
 279 static void *
 280 t_next(struct seq_file *m, void *v, loff_t *pos)
 281 {
 282         (*pos)++;
 283         return __next(m, pos);
 284 }
 285
 286 static void *t_start(struct seq_file *m, loff_t *pos)
 287 {
 288         int cpu;
 289
 290         local_irq_disable();
 291
 292         cpu = smp_processor_id();
 293         per_cpu(trace_active, cpu)++;
 294
 295         arch_spin_lock(&max_stack_lock);
 296
 297         if (*pos == 0)
 298                 return SEQ_START_TOKEN;
 299
 300         return __next(m, pos);
 301 }
 302
 303 static void t_stop(struct seq_file *m, void *p)
 304 {
 305         int cpu;
 306
 307         arch_spin_unlock(&max_stack_lock);
 308
 309         cpu = smp_processor_id();
 310         per_cpu(trace_active, cpu)--;
 311
 312         local_irq_enable();
 313 }
 314
 315 static void trace_lookup_stack(struct seq_file *m, long i)
 316 {
 317         unsigned long addr = stack_dump_trace[i];
 318
 319         seq_printf(m, "%pS\n", (void *)addr);
 320 }
 321
 322 static void print_disabled(struct seq_file *m)
 323 {
 324         seq_puts(m, "#\n"
 325                  "#  Stack tracer disabled\n"
 326                  "#\n"
 327                  "# To enable the stack tracer, either add 'stacktrace' to the\n"
 328                  "# kernel command line\n"
 329                  "# or 'echo 1 > /proc/sys/kernel/stack_tracer_enabled'\n"
 330                  "#\n");
 331 }
 332
 333 static int t_show(struct seq_file *m, void *v)
 334 {
 335         long i;
 336         int size;
 337
 338         if (v == SEQ_START_TOKEN) {
 339                 seq_printf(m, "        Depth    Size   Location"
 340                            "    (%d entries)\n"
 341                            "        -----    ----   --------\n",
 342                            max_stack_trace.nr_entries);
 343
 344                 if (!stack_tracer_enabled && !max_stack_size)
 345                         print_disabled(m);
 346
 347                 return 0;
 348         }
 349
 350         i = *(long *)v;
 351
 352         if (i >= max_stack_trace.nr_entries ||
 353             stack_dump_trace[i] == ULONG_MAX)
 354                 return 0;
 355
 356         if (i+1 == max_stack_trace.nr_entries ||
 357             stack_dump_trace[i+1] == ULONG_MAX)
 358                 size = stack_dump_index[i];
 359         else
 360                 size = stack_dump_index[i] - stack_dump_index[i+1];
 361
 362         seq_printf(m, "%3ld) %8d   %5d   ", i, stack_dump_index[i], size);
 363
 364         trace_lookup_stack(m, i);
 365
 366         return 0;
 367 }
 368
 369 static const struct seq_operations stack_trace_seq_ops = {
 370         .start          = t_start,
 371         .next           = t_next,
 372         .stop           = t_stop,
 373         .show           = t_show,
 374 };
 375
 376 static int stack_trace_open(struct inode *inode, struct file *file)
 377 {
 378         return seq_open(file, &stack_trace_seq_ops);
 379 }
 380
 381 static const struct file_operations stack_trace_fops = {
 382         .open           = stack_trace_open,
 383         .read           = seq_read,
 384         .llseek         = seq_lseek,
 385         .release        = seq_release,
 386 };
 387
 388 static int
 389 stack_trace_filter_open(struct inode *inode, struct file *file)
 390 {
 391         return ftrace_regex_open(&trace_ops, FTRACE_ITER_FILTER,
 392                                  inode, file);
 393 }
 394
 395 static const struct file_operations stack_trace_filter_fops = {
 396         .open = stack_trace_filter_open,
 397         .read = seq_read,
 398         .write = ftrace_filter_write,
 399         .llseek = tracing_lseek,
 400         .release = ftrace_regex_release,
 401 };
 402
 403 int
 404 stack_trace_sysctl(struct ctl_table *table, int write,
 405                    void __user *buffer, size_t *lenp,
 406                    loff_t *ppos)
 407 {
 408         int ret;
 409
 410         mutex_lock(&stack_sysctl_mutex);
 411
 412         ret = proc_dointvec(table, write, buffer, lenp, ppos);
 413
 414         if (ret || !write ||
 415             (last_stack_tracer_enabled == !!stack_tracer_enabled))
 416                 goto out;
 417
 418         last_stack_tracer_enabled = !!stack_tracer_enabled;
 419
 420         if (stack_tracer_enabled)
 421                 register_ftrace_function(&trace_ops);
 422         else
 423                 unregister_ftrace_function(&trace_ops);
 424
 425  out:
 426         mutex_unlock(&stack_sysctl_mutex);
 427         return ret;
 428 }
 429
 430 static char stack_trace_filter_buf[COMMAND_LINE_SIZE+1] __initdata;
 431
 432 static __init int enable_stacktrace(char *str)
 433 {
 434         if (strncmp(str, "_filter=", 8) == 0)
 435                 strncpy(stack_trace_filter_buf, str+8, COMMAND_LINE_SIZE);
 436
 437         stack_tracer_enabled = 1;
 438         last_stack_tracer_enabled = 1;
 439         return 1;
 440 }
 441 __setup("stacktrace", enable_stacktrace);
 442
 443 static __init int stack_trace_init(void)
 444 {
 445         struct dentry *d_tracer;
 446
 447         d_tracer = tracing_init_dentry();
 448         if (IS_ERR(d_tracer))
 449                 return 0;
 450
 451         trace_create_file("stack_max_size", 0644, d_tracer,
 452                         &max_stack_size, &stack_max_size_fops);
 453
 454         trace_create_file("stack_trace", 0444, d_tracer,
 455                         NULL, &stack_trace_fops);
 456
 457         trace_create_file("stack_trace_filter", 0444, d_tracer,
 458                         NULL, &stack_trace_filter_fops);
 459
 460         if (stack_trace_filter_buf[0])
 461                 ftrace_set_early_filter(&trace_ops, stack_trace_filter_buf, 1);
 462
 463         if (stack_tracer_enabled)
 464                 register_ftrace_function(&trace_ops);
 465
 466         return 0;
 467 }
 468
 469 device_initcall(stack_trace_init);