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         local_irq_save(flags);
  95         arch_spin_lock(&stack_trace_max_lock);
  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 <= stack_trace_max_size)
 103                 goto out;
 104
 105         stack_trace_max_size = this_size;
 106
 107         stack_trace_max.nr_entries = 0;
 108         stack_trace_max.skip = 3;
 109
 110         save_stack_trace(&stack_trace_max);
 111
 112         /* Skip over the overhead of the stack tracer itself */
 113         for (i = 0; i < stack_trace_max.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 < stack_trace_max.nr_entries) {
 134                 int found = 0;
 135
 136                 stack_trace_index[x] = this_size;
 137                 p = start;
 138
 139                 for (; p < top && i < stack_trace_max.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_trace_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                                         stack_trace_max_size -= tracer_frame;
 160                                 }
 161                         }
 162                 }
 163
 164                 if (!found)
 165                         i++;
 166         }
 167
 168         stack_trace_max.nr_entries = x;
 169         for (; x < i; x++)
 170                 stack_dump_trace[x] = ULONG_MAX;
 171
 172         if (task_stack_end_corrupted(current)) {
 173                 stack_trace_print();
 174                 BUG();
 175         }
 176
 177  out:
 178         arch_spin_unlock(&stack_trace_max_lock);
 179         local_irq_restore(flags);
 180 }
 181
 182 static void
 183 stack_trace_call(unsigned long ip, unsigned long parent_ip,
 184                  struct ftrace_ops *op, struct pt_regs *pt_regs)
 185 {
 186         unsigned long stack;
 187         int cpu;
 188
 189         preempt_disable_notrace();
 190
 191         cpu = raw_smp_processor_id();
 192         /* no atomic needed, we only modify this variable by this cpu */
 193         if (per_cpu(trace_active, cpu)++ != 0)
 194                 goto out;
 195
 196         ip += MCOUNT_INSN_SIZE;
 197
 198         check_stack(ip, &stack);
 199
 200  out:
 201         per_cpu(trace_active, cpu)--;
 202         /* prevent recursion in schedule */
 203         preempt_enable_notrace();
 204 }
 205
 206 static struct ftrace_ops trace_ops __read_mostly =
 207 {
 208         .func = stack_trace_call,
 209         .flags = FTRACE_OPS_FL_RECURSION_SAFE,
 210 };
 211
 212 static ssize_t
 213 stack_max_size_read(struct file *filp, char __user *ubuf,
 214                     size_t count, loff_t *ppos)
 215 {
 216         unsigned long *ptr = filp->private_data;
 217         char buf[64];
 218         int r;
 219
 220         r = snprintf(buf, sizeof(buf), "%ld\n", *ptr);
 221         if (r > sizeof(buf))
 222                 r = sizeof(buf);
 223         return simple_read_from_buffer(ubuf, count, ppos, buf, r);
 224 }
 225
 226 static ssize_t
 227 stack_max_size_write(struct file *filp, const char __user *ubuf,
 228                      size_t count, loff_t *ppos)
 229 {
 230         long *ptr = filp->private_data;
 231         unsigned long val, flags;
 232         int ret;
 233         int cpu;
 234
 235         ret = kstrtoul_from_user(ubuf, count, 10, &val);
 236         if (ret)
 237                 return ret;
 238
 239         local_irq_save(flags);
 240
 241         /*
 242          * In case we trace inside arch_spin_lock() or after (NMI),
 243          * we will cause circular lock, so we also need to increase
 244          * the percpu trace_active here.
 245          */
 246         cpu = smp_processor_id();
 247         per_cpu(trace_active, cpu)++;
 248
 249         arch_spin_lock(&stack_trace_max_lock);
 250         *ptr = val;
 251         arch_spin_unlock(&stack_trace_max_lock);
 252
 253         per_cpu(trace_active, cpu)--;
 254         local_irq_restore(flags);
 255
 256         return count;
 257 }
 258
 259 static const struct file_operations stack_max_size_fops = {
 260         .open           = tracing_open_generic,
 261         .read           = stack_max_size_read,
 262         .write          = stack_max_size_write,
 263         .llseek         = default_llseek,
 264 };
 265
 266 static void *
 267 __next(struct seq_file *m, loff_t *pos)
 268 {
 269         long n = *pos - 1;
 270
 271         if (n > stack_trace_max.nr_entries || stack_dump_trace[n] == ULONG_MAX)
 272                 return NULL;
 273
 274         m->private = (void *)n;
 275         return &m->private;
 276 }
 277
 278 static void *
 279 t_next(struct seq_file *m, void *v, loff_t *pos)
 280 {
 281         (*pos)++;
 282         return __next(m, pos);
 283 }
 284
 285 static void *t_start(struct seq_file *m, loff_t *pos)
 286 {
 287         int cpu;
 288
 289         local_irq_disable();
 290
 291         cpu = smp_processor_id();
 292         per_cpu(trace_active, cpu)++;
 293
 294         arch_spin_lock(&stack_trace_max_lock);
 295
 296         if (*pos == 0)
 297                 return SEQ_START_TOKEN;
 298
 299         return __next(m, pos);
 300 }
 301
 302 static void t_stop(struct seq_file *m, void *p)
 303 {
 304         int cpu;
 305
 306         arch_spin_unlock(&stack_trace_max_lock);
 307
 308         cpu = smp_processor_id();
 309         per_cpu(trace_active, cpu)--;
 310
 311         local_irq_enable();
 312 }
 313
 314 static void trace_lookup_stack(struct seq_file *m, long i)
 315 {
 316         unsigned long addr = stack_dump_trace[i];
 317
 318         seq_printf(m, "%pS\n", (void *)addr);
 319 }
 320
 321 static void print_disabled(struct seq_file *m)
 322 {
 323         seq_puts(m, "#\n"
 324                  "#  Stack tracer disabled\n"
 325                  "#\n"
 326                  "# To enable the stack tracer, either add 'stacktrace' to the\n"
 327                  "# kernel command line\n"
 328                  "# or 'echo 1 > /proc/sys/kernel/stack_tracer_enabled'\n"
 329                  "#\n");
 330 }
 331
 332 static int t_show(struct seq_file *m, void *v)
 333 {
 334         long i;
 335         int size;
 336
 337         if (v == SEQ_START_TOKEN) {
 338                 seq_printf(m, "        Depth    Size   Location"
 339                            "    (%d entries)\n"
 340                            "        -----    ----   --------\n",
 341                            stack_trace_max.nr_entries);
 342
 343                 if (!stack_tracer_enabled && !stack_trace_max_size)
 344                         print_disabled(m);
 345
 346                 return 0;
 347         }
 348
 349         i = *(long *)v;
 350
 351         if (i >= stack_trace_max.nr_entries ||
 352             stack_dump_trace[i] == ULONG_MAX)
 353                 return 0;
 354
 355         if (i+1 == stack_trace_max.nr_entries ||
 356             stack_dump_trace[i+1] == ULONG_MAX)
 357                 size = stack_trace_index[i];
 358         else
 359                 size = stack_trace_index[i] - stack_trace_index[i+1];
 360
 361         seq_printf(m, "%3ld) %8d   %5d   ", i, stack_trace_index[i], size);
 362
 363         trace_lookup_stack(m, i);
 364
 365         return 0;
 366 }
 367
 368 static const struct seq_operations stack_trace_seq_ops = {
 369         .start          = t_start,
 370         .next           = t_next,
 371         .stop           = t_stop,
 372         .show           = t_show,
 373 };
 374
 375 static int stack_trace_open(struct inode *inode, struct file *file)
 376 {
 377         return seq_open(file, &stack_trace_seq_ops);
 378 }
 379
 380 static const struct file_operations stack_trace_fops = {
 381         .open           = stack_trace_open,
 382         .read           = seq_read,
 383         .llseek         = seq_lseek,
 384         .release        = seq_release,
 385 };
 386
 387 static int
 388 stack_trace_filter_open(struct inode *inode, struct file *file)
 389 {
 390         return ftrace_regex_open(&trace_ops, FTRACE_ITER_FILTER,
 391                                  inode, file);
 392 }
 393
 394 static const struct file_operations stack_trace_filter_fops = {
 395         .open = stack_trace_filter_open,
 396         .read = seq_read,
 397         .write = ftrace_filter_write,
 398         .llseek = tracing_lseek,
 399         .release = ftrace_regex_release,
 400 };
 401
 402 int
 403 stack_trace_sysctl(struct ctl_table *table, int write,
 404                    void __user *buffer, size_t *lenp,
 405                    loff_t *ppos)
 406 {
 407         int ret;
 408
 409         mutex_lock(&stack_sysctl_mutex);
 410
 411         ret = proc_dointvec(table, write, buffer, lenp, ppos);
 412
 413         if (ret || !write ||
 414             (last_stack_tracer_enabled == !!stack_tracer_enabled))
 415                 goto out;
 416
 417         last_stack_tracer_enabled = !!stack_tracer_enabled;
 418
 419         if (stack_tracer_enabled)
 420                 register_ftrace_function(&trace_ops);
 421         else
 422                 unregister_ftrace_function(&trace_ops);
 423
 424  out:
 425         mutex_unlock(&stack_sysctl_mutex);
 426         return ret;
 427 }
 428
 429 static char stack_trace_filter_buf[COMMAND_LINE_SIZE+1] __initdata;
 430
 431 static __init int enable_stacktrace(char *str)
 432 {
 433         if (strncmp(str, "_filter=", 8) == 0)
 434                 strncpy(stack_trace_filter_buf, str+8, COMMAND_LINE_SIZE);
 435
 436         stack_tracer_enabled = 1;
 437         last_stack_tracer_enabled = 1;
 438         return 1;
 439 }
 440 __setup("stacktrace", enable_stacktrace);
 441
 442 static __init int stack_trace_init(void)
 443 {
 444         struct dentry *d_tracer;
 445
 446         d_tracer = tracing_init_dentry();
 447         if (IS_ERR(d_tracer))
 448                 return 0;
 449
 450         trace_create_file("stack_max_size", 0644, d_tracer,
 451                         &stack_trace_max_size, &stack_max_size_fops);
 452
 453         trace_create_file("stack_trace", 0444, d_tracer,
 454                         NULL, &stack_trace_fops);
 455
 456         trace_create_file("stack_trace_filter", 0444, d_tracer,
 457                         NULL, &stack_trace_filter_fops);
 458
 459         if (stack_trace_filter_buf[0])
 460                 ftrace_set_early_filter(&trace_ops, stack_trace_filter_buf, 1);
 461
 462         if (stack_tracer_enabled)
 463                 register_ftrace_function(&trace_ops);
 464
 465         return 0;
 466 }
 467
 468 device_initcall(stack_trace_init);