ring-buffer: Make benchmark handle missed events
[linux-2.6.git] / 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         .walk_stack             = print_context_stack,
97 };
98
99 static int
100 trace_kernel(struct pt_regs *regs, struct trace_array *tr,
101              struct trace_array_cpu *data)
102 {
103         struct backtrace_info info;
104         unsigned long bp;
105         char *stack;
106
107         info.tr = tr;
108         info.data = data;
109         info.pos = 1;
110
111         __trace_special(info.tr, info.data, 1, regs->ip, 0);
112
113         stack = ((char *)regs + sizeof(struct pt_regs));
114 #ifdef CONFIG_FRAME_POINTER
115         bp = regs->bp;
116 #else
117         bp = 0;
118 #endif
119
120         dump_trace(NULL, regs, (void *)stack, bp, &backtrace_ops, &info);
121
122         return info.pos;
123 }
124
125 static void timer_notify(struct pt_regs *regs, int cpu)
126 {
127         struct trace_array_cpu *data;
128         struct stack_frame frame;
129         struct trace_array *tr;
130         const void __user *fp;
131         int is_user;
132         int i;
133
134         if (!regs)
135                 return;
136
137         tr = sysprof_trace;
138         data = tr->data[cpu];
139         is_user = user_mode(regs);
140
141         if (!current || current->pid == 0)
142                 return;
143
144         if (is_user && current->state != TASK_RUNNING)
145                 return;
146
147         __trace_special(tr, data, 0, 0, current->pid);
148
149         if (!is_user)
150                 i = trace_kernel(regs, tr, data);
151         else
152                 i = 0;
153
154         /*
155          * Trace user stack if we are not a kernel thread
156          */
157         if (current->mm && i < sample_max_depth) {
158                 regs = (struct pt_regs *)current->thread.sp0 - 1;
159
160                 fp = (void __user *)regs->bp;
161
162                 __trace_special(tr, data, 2, regs->ip, 0);
163
164                 while (i < sample_max_depth) {
165                         frame.next_fp = NULL;
166                         frame.return_address = 0;
167                         if (!copy_stack_frame(fp, &frame))
168                                 break;
169                         if ((unsigned long)fp < regs->sp)
170                                 break;
171
172                         __trace_special(tr, data, 2, frame.return_address,
173                                         (unsigned long)fp);
174                         fp = frame.next_fp;
175
176                         i++;
177                 }
178
179         }
180
181         /*
182          * Special trace entry if we overflow the max depth:
183          */
184         if (i == sample_max_depth)
185                 __trace_special(tr, data, -1, -1, -1);
186
187         __trace_special(tr, data, 3, current->pid, i);
188 }
189
190 static enum hrtimer_restart stack_trace_timer_fn(struct hrtimer *hrtimer)
191 {
192         /* trace here */
193         timer_notify(get_irq_regs(), smp_processor_id());
194
195         hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
196
197         return HRTIMER_RESTART;
198 }
199
200 static void start_stack_timer(void *unused)
201 {
202         struct hrtimer *hrtimer = &__get_cpu_var(stack_trace_hrtimer);
203
204         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
205         hrtimer->function = stack_trace_timer_fn;
206
207         hrtimer_start(hrtimer, ns_to_ktime(sample_period),
208                       HRTIMER_MODE_REL_PINNED);
209 }
210
211 static void start_stack_timers(void)
212 {
213         on_each_cpu(start_stack_timer, NULL, 1);
214 }
215
216 static void stop_stack_timer(int cpu)
217 {
218         struct hrtimer *hrtimer = &per_cpu(stack_trace_hrtimer, cpu);
219
220         hrtimer_cancel(hrtimer);
221 }
222
223 static void stop_stack_timers(void)
224 {
225         int cpu;
226
227         for_each_online_cpu(cpu)
228                 stop_stack_timer(cpu);
229 }
230
231 static void stop_stack_trace(struct trace_array *tr)
232 {
233         mutex_lock(&sample_timer_lock);
234         stop_stack_timers();
235         tracer_enabled = 0;
236         mutex_unlock(&sample_timer_lock);
237 }
238
239 static int stack_trace_init(struct trace_array *tr)
240 {
241         sysprof_trace = tr;
242
243         tracing_start_cmdline_record();
244
245         mutex_lock(&sample_timer_lock);
246         start_stack_timers();
247         tracer_enabled = 1;
248         mutex_unlock(&sample_timer_lock);
249         return 0;
250 }
251
252 static void stack_trace_reset(struct trace_array *tr)
253 {
254         tracing_stop_cmdline_record();
255         stop_stack_trace(tr);
256 }
257
258 static struct tracer stack_trace __read_mostly =
259 {
260         .name           = "sysprof",
261         .init           = stack_trace_init,
262         .reset          = stack_trace_reset,
263 #ifdef CONFIG_FTRACE_SELFTEST
264         .selftest    = trace_selftest_startup_sysprof,
265 #endif
266 };
267
268 __init static int init_stack_trace(void)
269 {
270         return register_tracer(&stack_trace);
271 }
272 device_initcall(init_stack_trace);
273
274 #define MAX_LONG_DIGITS 22
275
276 static ssize_t
277 sysprof_sample_read(struct file *filp, char __user *ubuf,
278                     size_t cnt, loff_t *ppos)
279 {
280         char buf[MAX_LONG_DIGITS];
281         int r;
282
283         r = sprintf(buf, "%ld\n", nsecs_to_usecs(sample_period));
284
285         return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
286 }
287
288 static ssize_t
289 sysprof_sample_write(struct file *filp, const char __user *ubuf,
290                      size_t cnt, loff_t *ppos)
291 {
292         char buf[MAX_LONG_DIGITS];
293         unsigned long val;
294
295         if (cnt > MAX_LONG_DIGITS-1)
296                 cnt = MAX_LONG_DIGITS-1;
297
298         if (copy_from_user(&buf, ubuf, cnt))
299                 return -EFAULT;
300
301         buf[cnt] = 0;
302
303         val = simple_strtoul(buf, NULL, 10);
304         /*
305          * Enforce a minimum sample period of 100 usecs:
306          */
307         if (val < 100)
308                 val = 100;
309
310         mutex_lock(&sample_timer_lock);
311         stop_stack_timers();
312         sample_period = val * 1000;
313         start_stack_timers();
314         mutex_unlock(&sample_timer_lock);
315
316         return cnt;
317 }
318
319 static const struct file_operations sysprof_sample_fops = {
320         .read           = sysprof_sample_read,
321         .write          = sysprof_sample_write,
322 };
323
324 void init_tracer_sysprof_debugfs(struct dentry *d_tracer)
325 {
326
327         trace_create_file("sysprof_sample_period", 0644,
328                         d_tracer, NULL, &sysprof_sample_fops);
329 }