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