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