kernel/trace/trace_sysprof.c

   1 /*
   2  * trace stack traces
   3  *
   4  * Copyright (C) 2004-2008, Soeren Sandmann
   5  * Copyright (C) 2007 Steven Rostedt <srostedt@redhat.com>
   6  * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
   7  */
   8 #include <linux/kallsyms.h>
   9 #include <linux/debugfs.h>
  10 #include <linux/hrtimer.h>
  11 #include <linux/uaccess.h>
  12 #include <linux/ftrace.h>
  13 #include <linux/module.h>
  14 #include <linux/irq.h>
  15 #include <linux/fs.h>
  16
  17 #include <asm/stacktrace.h>
  18
  19 #include "trace.h"
  20
  21 static struct trace_array       *sysprof_trace;
  22 static int __read_mostly        tracer_enabled;
  23
  24 /*
  25  * 1 msec sample interval by default:
  26  */
  27 static unsigned long sample_period = 1000000;
  28 static const unsigned int sample_max_depth = 512;
  29
  30 static DEFINE_MUTEX(sample_timer_lock);
  31 /*
  32  * Per CPU hrtimers that do the profiling:
  33  */
  34 static DEFINE_PER_CPU(struct hrtimer, stack_trace_hrtimer);
  35
  36 struct stack_frame {
  37         const void __user       *next_fp;
  38         unsigned long           return_address;
  39 };
  40
  41 static int copy_stack_frame(const void __user *fp, struct stack_frame *frame)
  42 {
  43         int ret;
  44
  45         if (!access_ok(VERIFY_READ, fp, sizeof(*frame)))
  46                 return 0;
  47
  48         ret = 1;
  49         pagefault_disable();
  50         if (__copy_from_user_inatomic(frame, fp, sizeof(*frame)))
  51                 ret = 0;
  52         pagefault_enable();
  53
  54         return ret;
  55 }
  56
  57 struct backtrace_info {
  58         struct trace_array_cpu  *data;
  59         struct trace_array      *tr;
  60         int                     pos;
  61 };
  62
  63 static void
  64 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
  65 {
  66         /* Ignore warnings */
  67 }
  68
  69 static void backtrace_warning(void *data, char *msg)
  70 {
  71         /* Ignore warnings */
  72 }
  73
  74 static int backtrace_stack(void *data, char *name)
  75 {
  76         /* Don't bother with IRQ stacks for now */
  77         return -1;
  78 }
  79
  80 static void backtrace_address(void *data, unsigned long addr, int reliable)
  81 {
  82         struct backtrace_info *info = data;
  83
  84         if (info->pos < sample_max_depth && reliable) {
  85                 __trace_special(info->tr, info->data, 1, addr, 0);
  86
  87                 info->pos++;
  88         }
  89 }
  90
  91 static const struct stacktrace_ops backtrace_ops = {
  92         .warning                = backtrace_warning,
  93         .warning_symbol         = backtrace_warning_symbol,
  94         .stack                  = backtrace_stack,
  95         .address                = backtrace_address,
  96 };
  97
  98 static int
  99 trace_kernel(struct pt_regs *regs, struct trace_array *tr,
 100              struct trace_array_cpu *data)
 101 {
 102         struct backtrace_info info;
 103         unsigned long bp;
 104         char *stack;
 105
 106         info.tr = tr;
 107         info.data = data;
 108         info.pos = 1;
 109
 110         __trace_special(info.tr, info.data, 1, regs->ip, 0);
 111
 112         stack = ((char *)regs + sizeof(struct pt_regs));
 113 #ifdef CONFIG_FRAME_POINTER
 114         bp = regs->bp;
 115 #else
 116         bp = 0;
 117 #endif
 118
 119         dump_trace(NULL, regs, (void *)stack, bp, &backtrace_ops, &info);
 120
 121         return info.pos;
 122 }
 123
 124 static void timer_notify(struct pt_regs *regs, int cpu)
 125 {
 126         struct trace_array_cpu *data;
 127         struct stack_frame frame;
 128         struct trace_array *tr;
 129         const void __user *fp;
 130         int is_user;
 131         int i;
 132
 133         if (!regs)
 134                 return;
 135
 136         tr = sysprof_trace;
 137         data = tr->data[cpu];
 138         is_user = user_mode(regs);
 139
 140         if (!current || current->pid == 0)
 141                 return;
 142
 143         if (is_user && current->state != TASK_RUNNING)
 144                 return;
 145
 146         __trace_special(tr, data, 0, 0, current->pid);
 147
 148         if (!is_user)
 149                 i = trace_kernel(regs, tr, data);
 150         else
 151                 i = 0;
 152
 153         /*
 154          * Trace user stack if we are not a kernel thread
 155          */
 156         if (current->mm && i < sample_max_depth) {
 157                 regs = (struct pt_regs *)current->thread.sp0 - 1;
 158
 159                 fp = (void __user *)regs->bp;
 160
 161                 __trace_special(tr, data, 2, regs->ip, 0);
 162
 163                 while (i < sample_max_depth) {
 164                         frame.next_fp = NULL;
 165                         frame.return_address = 0;
 166                         if (!copy_stack_frame(fp, &frame))
 167                                 break;
 168                         if ((unsigned long)fp < regs->sp)
 169                                 break;
 170
 171                         __trace_special(tr, data, 2, frame.return_address,
 172                                         (unsigned long)fp);
 173                         fp = frame.next_fp;
 174
 175                         i++;
 176                 }
 177
 178         }
 179
 180         /*
 181          * Special trace entry if we overflow the max depth:
 182          */
 183         if (i == sample_max_depth)
 184                 __trace_special(tr, data, -1, -1, -1);
 185
 186         __trace_special(tr, data, 3, current->pid, i);
 187 }
 188
 189 static enum hrtimer_restart stack_trace_timer_fn(struct hrtimer *hrtimer)
 190 {
 191         /* trace here */
 192         timer_notify(get_irq_regs(), smp_processor_id());
 193
 194         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 195
 196         return HRTIMER_RESTART;
 197 }
 198
 199 static void start_stack_timer(void *unused)
 200 {
 201         struct hrtimer *hrtimer = &__get_cpu_var(stack_trace_hrtimer);
 202
 203         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 204         hrtimer->function = stack_trace_timer_fn;
 205
 206         hrtimer_start(hrtimer, ns_to_ktime(sample_period), HRTIMER_MODE_REL);
 207 }
 208
 209 static void start_stack_timers(void)
 210 {
 211         on_each_cpu(start_stack_timer, NULL, 1);
 212 }
 213
 214 static void stop_stack_timer(int cpu)
 215 {
 216         struct hrtimer *hrtimer = &per_cpu(stack_trace_hrtimer, cpu);
 217
 218         hrtimer_cancel(hrtimer);
 219 }
 220
 221 static void stop_stack_timers(void)
 222 {
 223         int cpu;
 224
 225         for_each_online_cpu(cpu)
 226                 stop_stack_timer(cpu);
 227 }
 228
 229 static void stop_stack_trace(struct trace_array *tr)
 230 {
 231         mutex_lock(&sample_timer_lock);
 232         stop_stack_timers();
 233         tracer_enabled = 0;
 234         mutex_unlock(&sample_timer_lock);
 235 }
 236
 237 static int stack_trace_init(struct trace_array *tr)
 238 {
 239         sysprof_trace = tr;
 240
 241         tracing_start_cmdline_record();
 242
 243         mutex_lock(&sample_timer_lock);
 244         start_stack_timers();
 245         tracer_enabled = 1;
 246         mutex_unlock(&sample_timer_lock);
 247         return 0;
 248 }
 249
 250 static void stack_trace_reset(struct trace_array *tr)
 251 {
 252         tracing_stop_cmdline_record();
 253         stop_stack_trace(tr);
 254 }
 255
 256 static struct tracer stack_trace __read_mostly =
 257 {
 258         .name           = "sysprof",
 259         .init           = stack_trace_init,
 260         .reset          = stack_trace_reset,
 261 #ifdef CONFIG_FTRACE_SELFTEST
 262         .selftest    = trace_selftest_startup_sysprof,
 263 #endif
 264 };
 265
 266 __init static int init_stack_trace(void)
 267 {
 268         return register_tracer(&stack_trace);
 269 }
 270 device_initcall(init_stack_trace);
 271
 272 #define MAX_LONG_DIGITS 22
 273
 274 static ssize_t
 275 sysprof_sample_read(struct file *filp, char __user *ubuf,
 276                     size_t cnt, loff_t *ppos)
 277 {
 278         char buf[MAX_LONG_DIGITS];
 279         int r;
 280
 281         r = sprintf(buf, "%ld\n", nsecs_to_usecs(sample_period));
 282
 283         return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
 284 }
 285
 286 static ssize_t
 287 sysprof_sample_write(struct file *filp, const char __user *ubuf,
 288                      size_t cnt, loff_t *ppos)
 289 {
 290         char buf[MAX_LONG_DIGITS];
 291         unsigned long val;
 292
 293         if (cnt > MAX_LONG_DIGITS-1)
 294                 cnt = MAX_LONG_DIGITS-1;
 295
 296         if (copy_from_user(&buf, ubuf, cnt))
 297                 return -EFAULT;
 298
 299         buf[cnt] = 0;
 300
 301         val = simple_strtoul(buf, NULL, 10);
 302         /*
 303          * Enforce a minimum sample period of 100 usecs:
 304          */
 305         if (val < 100)
 306                 val = 100;
 307
 308         mutex_lock(&sample_timer_lock);
 309         stop_stack_timers();
 310         sample_period = val * 1000;
 311         start_stack_timers();
 312         mutex_unlock(&sample_timer_lock);
 313
 314         return cnt;
 315 }
 316
 317 static const struct file_operations sysprof_sample_fops = {
 318         .read           = sysprof_sample_read,
 319         .write          = sysprof_sample_write,
 320 };
 321
 322 void init_tracer_sysprof_debugfs(struct dentry *d_tracer)
 323 {
 324         struct dentry *entry;
 325
 326         entry = debugfs_create_file("sysprof_sample_period", 0644,
 327                         d_tracer, NULL, &sysprof_sample_fops);
 328         if (entry)
 329                 return;
 330         pr_warning("Could not create debugfs 'sysprof_sample_period' entry\n");
 331 }