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