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/debugfs.h>
  11 #include <linux/ftrace.h>
  12 #include <linux/module.h>
  13 #include <linux/sysctl.h>
  14 #include <linux/init.h>
  15 #include <linux/fs.h>
  16 #include <linux/magic.h>
  17
  18 #include <asm/setup.h>
  19
  20 #include "trace.h"
  21
  22 #define STACK_TRACE_ENTRIES 500
  23
  24 #ifdef CC_USING_FENTRY
  25 # define fentry         1
  26 #else
  27 # define fentry         0
  28 #endif
  29
  30 static unsigned long stack_dump_trace[STACK_TRACE_ENTRIES+1] =
  31          { [0 ... (STACK_TRACE_ENTRIES)] = ULONG_MAX };
  32 static unsigned stack_dump_index[STACK_TRACE_ENTRIES];
  33
  34 /*
  35  * Reserve one entry for the passed in ip. This will allow
  36  * us to remove most or all of the stack size overhead
  37  * added by the stack tracer itself.
  38  */
  39 static struct stack_trace max_stack_trace = {
  40         .max_entries            = STACK_TRACE_ENTRIES - 1,
  41         .entries                = &stack_dump_trace[1],
  42 };
  43
  44 static unsigned long max_stack_size;
  45 static arch_spinlock_t max_stack_lock =
  46         (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
  47
  48 static DEFINE_PER_CPU(int, trace_active);
  49 static DEFINE_MUTEX(stack_sysctl_mutex);
  50
  51 int stack_tracer_enabled;
  52 static int last_stack_tracer_enabled;
  53
  54 static inline void print_max_stack(void)
  55 {
  56         long i;
  57         int size;
  58
  59         pr_emerg("        Depth    Size   Location    (%d entries)\n"
  60                            "        -----    ----   --------\n",
  61                            max_stack_trace.nr_entries - 1);
  62
  63         for (i = 0; i < max_stack_trace.nr_entries; i++) {
  64                 if (stack_dump_trace[i] == ULONG_MAX)
  65                         break;
  66                 if (i+1 == max_stack_trace.nr_entries ||
  67                                 stack_dump_trace[i+1] == ULONG_MAX)
  68                         size = stack_dump_index[i];
  69                 else
  70                         size = stack_dump_index[i] - stack_dump_index[i+1];
  71
  72                 pr_emerg("%3ld) %8d   %5d   %pS\n", i, stack_dump_index[i],
  73                                 size, (void *)stack_dump_trace[i]);
  74         }
  75 }
  76
  77 static inline void
  78 check_stack(unsigned long ip, unsigned long *stack)
  79 {
  80         unsigned long this_size, flags; unsigned long *p, *top, *start;
  81         static int tracer_frame;
  82         int frame_size = ACCESS_ONCE(tracer_frame);
  83         int i;
  84
  85         this_size = ((unsigned long)stack) & (THREAD_SIZE-1);
  86         this_size = THREAD_SIZE - this_size;
  87         /* Remove the frame of the tracer */
  88         this_size -= frame_size;
  89
  90         if (this_size <= max_stack_size)
  91                 return;
  92
  93         /* we do not handle interrupt stacks yet */
  94         if (!object_is_on_stack(stack))
  95                 return;
  96
  97         local_irq_save(flags);
  98         arch_spin_lock(&max_stack_lock);
  99
 100         /* In case another CPU set the tracer_frame on us */
 101         if (unlikely(!frame_size))
 102                 this_size -= tracer_frame;
 103
 104         /* a race could have already updated it */
 105         if (this_size <= max_stack_size)
 106                 goto out;
 107
 108         max_stack_size = this_size;
 109
 110         max_stack_trace.nr_entries = 0;
 111
 112         if (using_ftrace_ops_list_func())
 113                 max_stack_trace.skip = 4;
 114         else
 115                 max_stack_trace.skip = 3;
 116
 117         save_stack_trace(&max_stack_trace);
 118
 119         /*
 120          * Add the passed in ip from the function tracer.
 121          * Searching for this on the stack will skip over
 122          * most of the overhead from the stack tracer itself.
 123          */
 124         stack_dump_trace[0] = ip;
 125         max_stack_trace.nr_entries++;
 126
 127         /*
 128          * Now find where in the stack these are.
 129          */
 130         i = 0;
 131         start = stack;
 132         top = (unsigned long *)
 133                 (((unsigned long)start & ~(THREAD_SIZE-1)) + THREAD_SIZE);
 134
 135         /*
 136          * Loop through all the entries. One of the entries may
 137          * for some reason be missed on the stack, so we may
 138          * have to account for them. If they are all there, this
 139          * loop will only happen once. This code only takes place
 140          * on a new max, so it is far from a fast path.
 141          */
 142         while (i < max_stack_trace.nr_entries) {
 143                 int found = 0;
 144
 145                 stack_dump_index[i] = this_size;
 146                 p = start;
 147
 148                 for (; p < top && i < max_stack_trace.nr_entries; p++) {
 149                         if (*p == stack_dump_trace[i]) {
 150                                 this_size = stack_dump_index[i++] =
 151                                         (top - p) * sizeof(unsigned long);
 152                                 found = 1;
 153                                 /* Start the search from here */
 154                                 start = p + 1;
 155                                 /*
 156                                  * We do not want to show the overhead
 157                                  * of the stack tracer stack in the
 158                                  * max stack. If we haven't figured
 159                                  * out what that is, then figure it out
 160                                  * now.
 161                                  */
 162                                 if (unlikely(!tracer_frame) && i == 1) {
 163                                         tracer_frame = (p - stack) *
 164                                                 sizeof(unsigned long);
 165                                         max_stack_size -= tracer_frame;
 166                                 }
 167                         }
 168                 }
 169
 170                 if (!found)
 171                         i++;
 172         }
 173
 174         if ((current != &init_task &&
 175                 *(end_of_stack(current)) != STACK_END_MAGIC)) {
 176                 print_max_stack();
 177                 BUG();
 178         }
 179
 180  out:
 181         arch_spin_unlock(&max_stack_lock);
 182         local_irq_restore(flags);
 183 }
 184
 185 static void
 186 stack_trace_call(unsigned long ip, unsigned long parent_ip,
 187                  struct ftrace_ops *op, struct pt_regs *pt_regs)
 188 {
 189         unsigned long stack;
 190         int cpu;
 191
 192         preempt_disable_notrace();
 193
 194         cpu = raw_smp_processor_id();
 195         /* no atomic needed, we only modify this variable by this cpu */
 196         if (per_cpu(trace_active, cpu)++ != 0)
 197                 goto out;
 198
 199         /*
 200          * When fentry is used, the traced function does not get
 201          * its stack frame set up, and we lose the parent.
 202          * The ip is pretty useless because the function tracer
 203          * was called before that function set up its stack frame.
 204          * In this case, we use the parent ip.
 205          *
 206          * By adding the return address of either the parent ip
 207          * or the current ip we can disregard most of the stack usage
 208          * caused by the stack tracer itself.
 209          *
 210          * The function tracer always reports the address of where the
 211          * mcount call was, but the stack will hold the return address.
 212          */
 213         if (fentry)
 214                 ip = parent_ip;
 215         else
 216                 ip += MCOUNT_INSN_SIZE;
 217
 218         check_stack(ip, &stack);
 219
 220  out:
 221         per_cpu(trace_active, cpu)--;
 222         /* prevent recursion in schedule */
 223         preempt_enable_notrace();
 224 }
 225
 226 static struct ftrace_ops trace_ops __read_mostly =
 227 {
 228         .func = stack_trace_call,
 229         .flags = FTRACE_OPS_FL_RECURSION_SAFE,
 230 };
 231
 232 static ssize_t
 233 stack_max_size_read(struct file *filp, char __user *ubuf,
 234                     size_t count, loff_t *ppos)
 235 {
 236         unsigned long *ptr = filp->private_data;
 237         char buf[64];
 238         int r;
 239
 240         r = snprintf(buf, sizeof(buf), "%ld\n", *ptr);
 241         if (r > sizeof(buf))
 242                 r = sizeof(buf);
 243         return simple_read_from_buffer(ubuf, count, ppos, buf, r);
 244 }
 245
 246 static ssize_t
 247 stack_max_size_write(struct file *filp, const char __user *ubuf,
 248                      size_t count, loff_t *ppos)
 249 {
 250         long *ptr = filp->private_data;
 251         unsigned long val, flags;
 252         int ret;
 253         int cpu;
 254
 255         ret = kstrtoul_from_user(ubuf, count, 10, &val);
 256         if (ret)
 257                 return ret;
 258
 259         local_irq_save(flags);
 260
 261         /*
 262          * In case we trace inside arch_spin_lock() or after (NMI),
 263          * we will cause circular lock, so we also need to increase
 264          * the percpu trace_active here.
 265          */
 266         cpu = smp_processor_id();
 267         per_cpu(trace_active, cpu)++;
 268
 269         arch_spin_lock(&max_stack_lock);
 270         *ptr = val;
 271         arch_spin_unlock(&max_stack_lock);
 272
 273         per_cpu(trace_active, cpu)--;
 274         local_irq_restore(flags);
 275
 276         return count;
 277 }
 278
 279 static const struct file_operations stack_max_size_fops = {
 280         .open           = tracing_open_generic,
 281         .read           = stack_max_size_read,
 282         .write          = stack_max_size_write,
 283         .llseek         = default_llseek,
 284 };
 285
 286 static void *
 287 __next(struct seq_file *m, loff_t *pos)
 288 {
 289         long n = *pos - 1;
 290
 291         if (n >= max_stack_trace.nr_entries || stack_dump_trace[n] == ULONG_MAX)
 292                 return NULL;
 293
 294         m->private = (void *)n;
 295         return &m->private;
 296 }
 297
 298 static void *
 299 t_next(struct seq_file *m, void *v, loff_t *pos)
 300 {
 301         (*pos)++;
 302         return __next(m, pos);
 303 }
 304
 305 static void *t_start(struct seq_file *m, loff_t *pos)
 306 {
 307         int cpu;
 308
 309         local_irq_disable();
 310
 311         cpu = smp_processor_id();
 312         per_cpu(trace_active, cpu)++;
 313
 314         arch_spin_lock(&max_stack_lock);
 315
 316         if (*pos == 0)
 317                 return SEQ_START_TOKEN;
 318
 319         return __next(m, pos);
 320 }
 321
 322 static void t_stop(struct seq_file *m, void *p)
 323 {
 324         int cpu;
 325
 326         arch_spin_unlock(&max_stack_lock);
 327
 328         cpu = smp_processor_id();
 329         per_cpu(trace_active, cpu)--;
 330
 331         local_irq_enable();
 332 }
 333
 334 static int trace_lookup_stack(struct seq_file *m, long i)
 335 {
 336         unsigned long addr = stack_dump_trace[i];
 337
 338         return seq_printf(m, "%pS\n", (void *)addr);
 339 }
 340
 341 static void print_disabled(struct seq_file *m)
 342 {
 343         seq_puts(m, "#\n"
 344                  "#  Stack tracer disabled\n"
 345                  "#\n"
 346                  "# To enable the stack tracer, either add 'stacktrace' to the\n"
 347                  "# kernel command line\n"
 348                  "# or 'echo 1 > /proc/sys/kernel/stack_tracer_enabled'\n"
 349                  "#\n");
 350 }
 351
 352 static int t_show(struct seq_file *m, void *v)
 353 {
 354         long i;
 355         int size;
 356
 357         if (v == SEQ_START_TOKEN) {
 358                 seq_printf(m, "        Depth    Size   Location"
 359                            "    (%d entries)\n"
 360                            "        -----    ----   --------\n",
 361                            max_stack_trace.nr_entries - 1);
 362
 363                 if (!stack_tracer_enabled && !max_stack_size)
 364                         print_disabled(m);
 365
 366                 return 0;
 367         }
 368
 369         i = *(long *)v;
 370
 371         if (i >= max_stack_trace.nr_entries ||
 372             stack_dump_trace[i] == ULONG_MAX)
 373                 return 0;
 374
 375         if (i+1 == max_stack_trace.nr_entries ||
 376             stack_dump_trace[i+1] == ULONG_MAX)
 377                 size = stack_dump_index[i];
 378         else
 379                 size = stack_dump_index[i] - stack_dump_index[i+1];
 380
 381         seq_printf(m, "%3ld) %8d   %5d   ", i, stack_dump_index[i], size);
 382
 383         trace_lookup_stack(m, i);
 384
 385         return 0;
 386 }
 387
 388 static const struct seq_operations stack_trace_seq_ops = {
 389         .start          = t_start,
 390         .next           = t_next,
 391         .stop           = t_stop,
 392         .show           = t_show,
 393 };
 394
 395 static int stack_trace_open(struct inode *inode, struct file *file)
 396 {
 397         return seq_open(file, &stack_trace_seq_ops);
 398 }
 399
 400 static const struct file_operations stack_trace_fops = {
 401         .open           = stack_trace_open,
 402         .read           = seq_read,
 403         .llseek         = seq_lseek,
 404         .release        = seq_release,
 405 };
 406
 407 static int
 408 stack_trace_filter_open(struct inode *inode, struct file *file)
 409 {
 410         return ftrace_regex_open(&trace_ops, FTRACE_ITER_FILTER,
 411                                  inode, file);
 412 }
 413
 414 static const struct file_operations stack_trace_filter_fops = {
 415         .open = stack_trace_filter_open,
 416         .read = seq_read,
 417         .write = ftrace_filter_write,
 418         .llseek = tracing_lseek,
 419         .release = ftrace_regex_release,
 420 };
 421
 422 int
 423 stack_trace_sysctl(struct ctl_table *table, int write,
 424                    void __user *buffer, size_t *lenp,
 425                    loff_t *ppos)
 426 {
 427         int ret;
 428
 429         mutex_lock(&stack_sysctl_mutex);
 430
 431         ret = proc_dointvec(table, write, buffer, lenp, ppos);
 432
 433         if (ret || !write ||
 434             (last_stack_tracer_enabled == !!stack_tracer_enabled))
 435                 goto out;
 436
 437         last_stack_tracer_enabled = !!stack_tracer_enabled;
 438
 439         if (stack_tracer_enabled)
 440                 register_ftrace_function(&trace_ops);
 441         else
 442                 unregister_ftrace_function(&trace_ops);
 443
 444  out:
 445         mutex_unlock(&stack_sysctl_mutex);
 446         return ret;
 447 }
 448
 449 static char stack_trace_filter_buf[COMMAND_LINE_SIZE+1] __initdata;
 450
 451 static __init int enable_stacktrace(char *str)
 452 {
 453         if (strncmp(str, "_filter=", 8) == 0)
 454                 strncpy(stack_trace_filter_buf, str+8, COMMAND_LINE_SIZE);
 455
 456         stack_tracer_enabled = 1;
 457         last_stack_tracer_enabled = 1;
 458         return 1;
 459 }
 460 __setup("stacktrace", enable_stacktrace);
 461
 462 static __init int stack_trace_init(void)
 463 {
 464         struct dentry *d_tracer;
 465
 466         d_tracer = tracing_init_dentry();
 467         if (!d_tracer)
 468                 return 0;
 469
 470         trace_create_file("stack_max_size", 0644, d_tracer,
 471                         &max_stack_size, &stack_max_size_fops);
 472
 473         trace_create_file("stack_trace", 0444, d_tracer,
 474                         NULL, &stack_trace_fops);
 475
 476         trace_create_file("stack_trace_filter", 0444, d_tracer,
 477                         NULL, &stack_trace_filter_fops);
 478
 479         if (stack_trace_filter_buf[0])
 480                 ftrace_set_early_filter(&trace_ops, stack_trace_filter_buf, 1);
 481
 482         if (stack_tracer_enabled)
 483                 register_ftrace_function(&trace_ops);
 484
 485         return 0;
 486 }
 487
 488 device_initcall(stack_trace_init);