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 #ifdef CC_USING_FENTRY
  24 # define fentry         1
  25 #else
  26 # define fentry         0
  27 #endif
  28
  29 static unsigned long stack_dump_trace[STACK_TRACE_ENTRIES+1] =
  30          { [0 ... (STACK_TRACE_ENTRIES)] = ULONG_MAX };
  31 static unsigned stack_dump_index[STACK_TRACE_ENTRIES];
  32
  33 /*
  34  * Reserve one entry for the passed in ip. This will allow
  35  * us to remove most or all of the stack size overhead
  36  * added by the stack tracer itself.
  37  */
  38 static struct stack_trace max_stack_trace = {
  39         .max_entries            = STACK_TRACE_ENTRIES - 1,
  40         .entries                = &stack_dump_trace[1],
  41 };
  42
  43 static unsigned long max_stack_size;
  44 static arch_spinlock_t max_stack_lock =
  45         (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
  46
  47 static DEFINE_PER_CPU(int, trace_active);
  48 static DEFINE_MUTEX(stack_sysctl_mutex);
  49
  50 int stack_tracer_enabled;
  51 static int last_stack_tracer_enabled;
  52
  53 static inline void
  54 check_stack(unsigned long ip, unsigned long *stack)
  55 {
  56         unsigned long this_size, flags;
  57         unsigned long *p, *top, *start;
  58         static int tracer_frame;
  59         int frame_size = ACCESS_ONCE(tracer_frame);
  60         int i;
  61
  62         this_size = ((unsigned long)stack) & (THREAD_SIZE-1);
  63         this_size = THREAD_SIZE - this_size;
  64         /* Remove the frame of the tracer */
  65         this_size -= frame_size;
  66
  67         if (this_size <= max_stack_size)
  68                 return;
  69
  70         /* we do not handle interrupt stacks yet */
  71         if (!object_is_on_stack(stack))
  72                 return;
  73
  74         local_irq_save(flags);
  75         arch_spin_lock(&max_stack_lock);
  76
  77         /* In case another CPU set the tracer_frame on us */
  78         if (unlikely(!frame_size))
  79                 this_size -= tracer_frame;
  80
  81         /* a race could have already updated it */
  82         if (this_size <= max_stack_size)
  83                 goto out;
  84
  85         max_stack_size = this_size;
  86
  87         max_stack_trace.nr_entries      = 0;
  88         max_stack_trace.skip            = 3;
  89
  90         save_stack_trace(&max_stack_trace);
  91
  92         /*
  93          * Add the passed in ip from the function tracer.
  94          * Searching for this on the stack will skip over
  95          * most of the overhead from the stack tracer itself.
  96          */
  97         stack_dump_trace[0] = ip;
  98         max_stack_trace.nr_entries++;
  99
 100         /*
 101          * Now find where in the stack these are.
 102          */
 103         i = 0;
 104         start = stack;
 105         top = (unsigned long *)
 106                 (((unsigned long)start & ~(THREAD_SIZE-1)) + THREAD_SIZE);
 107
 108         /*
 109          * Loop through all the entries. One of the entries may
 110          * for some reason be missed on the stack, so we may
 111          * have to account for them. If they are all there, this
 112          * loop will only happen once. This code only takes place
 113          * on a new max, so it is far from a fast path.
 114          */
 115         while (i < max_stack_trace.nr_entries) {
 116                 int found = 0;
 117
 118                 stack_dump_index[i] = this_size;
 119                 p = start;
 120
 121                 for (; p < top && i < max_stack_trace.nr_entries; p++) {
 122                         if (*p == stack_dump_trace[i]) {
 123                                 this_size = stack_dump_index[i++] =
 124                                         (top - p) * sizeof(unsigned long);
 125                                 found = 1;
 126                                 /* Start the search from here */
 127                                 start = p + 1;
 128                                 /*
 129                                  * We do not want to show the overhead
 130                                  * of the stack tracer stack in the
 131                                  * max stack. If we haven't figured
 132                                  * out what that is, then figure it out
 133                                  * now.
 134                                  */
 135                                 if (unlikely(!tracer_frame) && i == 1) {
 136                                         tracer_frame = (p - stack) *
 137                                                 sizeof(unsigned long);
 138                                         max_stack_size -= tracer_frame;
 139                                 }
 140                         }
 141                 }
 142
 143                 if (!found)
 144                         i++;
 145         }
 146
 147  out:
 148         arch_spin_unlock(&max_stack_lock);
 149         local_irq_restore(flags);
 150 }
 151
 152 static void
 153 stack_trace_call(unsigned long ip, unsigned long parent_ip,
 154                  struct ftrace_ops *op, struct pt_regs *pt_regs)
 155 {
 156         unsigned long stack;
 157         int cpu;
 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_RECURSION_SAFE,
 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         int ret;
 220         int cpu;
 221
 222         ret = kstrtoul_from_user(ubuf, count, 10, &val);
 223         if (ret)
 224                 return ret;
 225
 226         local_irq_save(flags);
 227
 228         /*
 229          * In case we trace inside arch_spin_lock() or after (NMI),
 230          * we will cause circular lock, so we also need to increase
 231          * the percpu trace_active here.
 232          */
 233         cpu = smp_processor_id();
 234         per_cpu(trace_active, cpu)++;
 235
 236         arch_spin_lock(&max_stack_lock);
 237         *ptr = val;
 238         arch_spin_unlock(&max_stack_lock);
 239
 240         per_cpu(trace_active, cpu)--;
 241         local_irq_restore(flags);
 242
 243         return count;
 244 }
 245
 246 static const struct file_operations stack_max_size_fops = {
 247         .open           = tracing_open_generic,
 248         .read           = stack_max_size_read,
 249         .write          = stack_max_size_write,
 250         .llseek         = default_llseek,
 251 };
 252
 253 static void *
 254 __next(struct seq_file *m, loff_t *pos)
 255 {
 256         long n = *pos - 1;
 257
 258         if (n >= max_stack_trace.nr_entries || stack_dump_trace[n] == ULONG_MAX)
 259                 return NULL;
 260
 261         m->private = (void *)n;
 262         return &m->private;
 263 }
 264
 265 static void *
 266 t_next(struct seq_file *m, void *v, loff_t *pos)
 267 {
 268         (*pos)++;
 269         return __next(m, pos);
 270 }
 271
 272 static void *t_start(struct seq_file *m, loff_t *pos)
 273 {
 274         int cpu;
 275
 276         local_irq_disable();
 277
 278         cpu = smp_processor_id();
 279         per_cpu(trace_active, cpu)++;
 280
 281         arch_spin_lock(&max_stack_lock);
 282
 283         if (*pos == 0)
 284                 return SEQ_START_TOKEN;
 285
 286         return __next(m, pos);
 287 }
 288
 289 static void t_stop(struct seq_file *m, void *p)
 290 {
 291         int cpu;
 292
 293         arch_spin_unlock(&max_stack_lock);
 294
 295         cpu = smp_processor_id();
 296         per_cpu(trace_active, cpu)--;
 297
 298         local_irq_enable();
 299 }
 300
 301 static int trace_lookup_stack(struct seq_file *m, long i)
 302 {
 303         unsigned long addr = stack_dump_trace[i];
 304
 305         return seq_printf(m, "%pS\n", (void *)addr);
 306 }
 307
 308 static void print_disabled(struct seq_file *m)
 309 {
 310         seq_puts(m, "#\n"
 311                  "#  Stack tracer disabled\n"
 312                  "#\n"
 313                  "# To enable the stack tracer, either add 'stacktrace' to the\n"
 314                  "# kernel command line\n"
 315                  "# or 'echo 1 > /proc/sys/kernel/stack_tracer_enabled'\n"
 316                  "#\n");
 317 }
 318
 319 static int t_show(struct seq_file *m, void *v)
 320 {
 321         long i;
 322         int size;
 323
 324         if (v == SEQ_START_TOKEN) {
 325                 seq_printf(m, "        Depth    Size   Location"
 326                            "    (%d entries)\n"
 327                            "        -----    ----   --------\n",
 328                            max_stack_trace.nr_entries - 1);
 329
 330                 if (!stack_tracer_enabled && !max_stack_size)
 331                         print_disabled(m);
 332
 333                 return 0;
 334         }
 335
 336         i = *(long *)v;
 337
 338         if (i >= max_stack_trace.nr_entries ||
 339             stack_dump_trace[i] == ULONG_MAX)
 340                 return 0;
 341
 342         if (i+1 == max_stack_trace.nr_entries ||
 343             stack_dump_trace[i+1] == ULONG_MAX)
 344                 size = stack_dump_index[i];
 345         else
 346                 size = stack_dump_index[i] - stack_dump_index[i+1];
 347
 348         seq_printf(m, "%3ld) %8d   %5d   ", i, stack_dump_index[i], size);
 349
 350         trace_lookup_stack(m, i);
 351
 352         return 0;
 353 }
 354
 355 static const struct seq_operations stack_trace_seq_ops = {
 356         .start          = t_start,
 357         .next           = t_next,
 358         .stop           = t_stop,
 359         .show           = t_show,
 360 };
 361
 362 static int stack_trace_open(struct inode *inode, struct file *file)
 363 {
 364         return seq_open(file, &stack_trace_seq_ops);
 365 }
 366
 367 static const struct file_operations stack_trace_fops = {
 368         .open           = stack_trace_open,
 369         .read           = seq_read,
 370         .llseek         = seq_lseek,
 371         .release        = seq_release,
 372 };
 373
 374 static int
 375 stack_trace_filter_open(struct inode *inode, struct file *file)
 376 {
 377         return ftrace_regex_open(&trace_ops, FTRACE_ITER_FILTER,
 378                                  inode, file);
 379 }
 380
 381 static const struct file_operations stack_trace_filter_fops = {
 382         .open = stack_trace_filter_open,
 383         .read = seq_read,
 384         .write = ftrace_filter_write,
 385         .llseek = tracing_lseek,
 386         .release = ftrace_regex_release,
 387 };
 388
 389 int
 390 stack_trace_sysctl(struct ctl_table *table, int write,
 391                    void __user *buffer, size_t *lenp,
 392                    loff_t *ppos)
 393 {
 394         int ret;
 395
 396         mutex_lock(&stack_sysctl_mutex);
 397
 398         ret = proc_dointvec(table, write, buffer, lenp, ppos);
 399
 400         if (ret || !write ||
 401             (last_stack_tracer_enabled == !!stack_tracer_enabled))
 402                 goto out;
 403
 404         last_stack_tracer_enabled = !!stack_tracer_enabled;
 405
 406         if (stack_tracer_enabled)
 407                 register_ftrace_function(&trace_ops);
 408         else
 409                 unregister_ftrace_function(&trace_ops);
 410
 411  out:
 412         mutex_unlock(&stack_sysctl_mutex);
 413         return ret;
 414 }
 415
 416 static char stack_trace_filter_buf[COMMAND_LINE_SIZE+1] __initdata;
 417
 418 static __init int enable_stacktrace(char *str)
 419 {
 420         if (strncmp(str, "_filter=", 8) == 0)
 421                 strncpy(stack_trace_filter_buf, str+8, COMMAND_LINE_SIZE);
 422
 423         stack_tracer_enabled = 1;
 424         last_stack_tracer_enabled = 1;
 425         return 1;
 426 }
 427 __setup("stacktrace", enable_stacktrace);
 428
 429 static __init int stack_trace_init(void)
 430 {
 431         struct dentry *d_tracer;
 432
 433         d_tracer = tracing_init_dentry();
 434         if (!d_tracer)
 435                 return 0;
 436
 437         trace_create_file("stack_max_size", 0644, d_tracer,
 438                         &max_stack_size, &stack_max_size_fops);
 439
 440         trace_create_file("stack_trace", 0444, d_tracer,
 441                         NULL, &stack_trace_fops);
 442
 443         trace_create_file("stack_trace_filter", 0444, d_tracer,
 444                         NULL, &stack_trace_filter_fops);
 445
 446         if (stack_trace_filter_buf[0])
 447                 ftrace_set_early_filter(&trace_ops, stack_trace_filter_buf, 1);
 448
 449         if (stack_tracer_enabled)
 450                 register_ftrace_function(&trace_ops);
 451
 452         return 0;
 453 }
 454
 455 device_initcall(stack_trace_init);