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