Commit | Line | Data |
---|---|---|
f06c3810 IM |
1 | /* |
2 | * trace stack traces | |
3 | * | |
8a9e94c1 | 4 | * Copyright (C) 2004-2008, Soeren Sandmann |
f06c3810 IM |
5 | * Copyright (C) 2007 Steven Rostedt <srostedt@redhat.com> |
6 | * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com> | |
f06c3810 | 7 | */ |
f06c3810 | 8 | #include <linux/kallsyms.h> |
0075fa80 IM |
9 | #include <linux/debugfs.h> |
10 | #include <linux/hrtimer.h> | |
f06c3810 | 11 | #include <linux/uaccess.h> |
f06c3810 | 12 | #include <linux/ftrace.h> |
0075fa80 | 13 | #include <linux/module.h> |
56a08bdc | 14 | #include <linux/irq.h> |
0075fa80 | 15 | #include <linux/fs.h> |
f06c3810 | 16 | |
cd2134b1 SSP |
17 | #include <asm/stacktrace.h> |
18 | ||
f06c3810 IM |
19 | #include "trace.h" |
20 | ||
56a08bdc | 21 | static struct trace_array *sysprof_trace; |
f06c3810 IM |
22 | static int __read_mostly tracer_enabled; |
23 | ||
56a08bdc | 24 | /* |
d618b3e6 | 25 | * 1 msec sample interval by default: |
56a08bdc | 26 | */ |
d618b3e6 | 27 | static unsigned long sample_period = 1000000; |
842af315 | 28 | static const unsigned int sample_max_depth = 512; |
0075fa80 | 29 | |
d618b3e6 | 30 | static DEFINE_MUTEX(sample_timer_lock); |
0075fa80 IM |
31 | /* |
32 | * Per CPU hrtimers that do the profiling: | |
33 | */ | |
34 | static DEFINE_PER_CPU(struct hrtimer, stack_trace_hrtimer); | |
35 | ||
56a08bdc IM |
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 | { | |
9f6b4e3f IM |
43 | int ret; |
44 | ||
56a08bdc IM |
45 | if (!access_ok(VERIFY_READ, fp, sizeof(*frame))) |
46 | return 0; | |
47 | ||
9f6b4e3f IM |
48 | ret = 1; |
49 | pagefault_disable(); | |
50 | if (__copy_from_user_inatomic(frame, fp, sizeof(*frame))) | |
51 | ret = 0; | |
52 | pagefault_enable(); | |
56a08bdc | 53 | |
9f6b4e3f | 54 | return ret; |
56a08bdc IM |
55 | } |
56 | ||
cd2134b1 SSP |
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 | ||
4543ae7c | 91 | static const struct stacktrace_ops backtrace_ops = { |
cd2134b1 SSP |
92 | .warning = backtrace_warning, |
93 | .warning_symbol = backtrace_warning_symbol, | |
94 | .stack = backtrace_stack, | |
95 | .address = backtrace_address, | |
96 | }; | |
97 | ||
cf3271a7 | 98 | static int |
cd2134b1 SSP |
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; | |
cd2134b1 SSP |
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 | ||
cf3271a7 | 121 | return info.pos; |
cd2134b1 SSP |
122 | } |
123 | ||
56a08bdc IM |
124 | static void timer_notify(struct pt_regs *regs, int cpu) |
125 | { | |
56a08bdc IM |
126 | struct trace_array_cpu *data; |
127 | struct stack_frame frame; | |
128 | struct trace_array *tr; | |
9f6b4e3f | 129 | const void __user *fp; |
56a08bdc IM |
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 | ||
cd2134b1 | 146 | __trace_special(tr, data, 0, 0, current->pid); |
56a08bdc | 147 | |
cd2134b1 | 148 | if (!is_user) |
cf3271a7 SS |
149 | i = trace_kernel(regs, tr, data); |
150 | else | |
151 | i = 0; | |
56a08bdc | 152 | |
cf3271a7 SS |
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; | |
56a08bdc | 158 | |
cf3271a7 | 159 | fp = (void __user *)regs->bp; |
cd2134b1 | 160 | |
cf3271a7 | 161 | __trace_special(tr, data, 2, regs->ip, 0); |
56a08bdc | 162 | |
cf3271a7 | 163 | while (i < sample_max_depth) { |
a89cc195 | 164 | frame.next_fp = NULL; |
cf3271a7 SS |
165 | frame.return_address = 0; |
166 | if (!copy_stack_frame(fp, &frame)) | |
167 | break; | |
168 | if ((unsigned long)fp < regs->sp) | |
169 | break; | |
56a08bdc | 170 | |
cf3271a7 SS |
171 | __trace_special(tr, data, 2, frame.return_address, |
172 | (unsigned long)fp); | |
173 | fp = frame.next_fp; | |
174 | ||
175 | i++; | |
176 | } | |
177 | ||
178 | } | |
56a08bdc | 179 | |
9f6b4e3f IM |
180 | /* |
181 | * Special trace entry if we overflow the max depth: | |
182 | */ | |
842af315 | 183 | if (i == sample_max_depth) |
9caee613 | 184 | __trace_special(tr, data, -1, -1, -1); |
cf3271a7 SS |
185 | |
186 | __trace_special(tr, data, 3, current->pid, i); | |
56a08bdc IM |
187 | } |
188 | ||
0075fa80 IM |
189 | static enum hrtimer_restart stack_trace_timer_fn(struct hrtimer *hrtimer) |
190 | { | |
191 | /* trace here */ | |
56a08bdc | 192 | timer_notify(get_irq_regs(), smp_processor_id()); |
0075fa80 IM |
193 | |
194 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); | |
195 | ||
196 | return HRTIMER_RESTART; | |
197 | } | |
198 | ||
9e01c1b7 | 199 | static void start_stack_timer(void *unused) |
0075fa80 | 200 | { |
9e01c1b7 | 201 | struct hrtimer *hrtimer = &__get_cpu_var(stack_trace_hrtimer); |
0075fa80 IM |
202 | |
203 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
204 | hrtimer->function = stack_trace_timer_fn; | |
0075fa80 | 205 | |
5c333864 AB |
206 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
207 | HRTIMER_MODE_REL_PINNED); | |
0075fa80 IM |
208 | } |
209 | ||
210 | static void start_stack_timers(void) | |
211 | { | |
9e01c1b7 | 212 | on_each_cpu(start_stack_timer, NULL, 1); |
0075fa80 IM |
213 | } |
214 | ||
215 | static void stop_stack_timer(int cpu) | |
216 | { | |
217 | struct hrtimer *hrtimer = &per_cpu(stack_trace_hrtimer, cpu); | |
218 | ||
219 | hrtimer_cancel(hrtimer); | |
0075fa80 IM |
220 | } |
221 | ||
222 | static void stop_stack_timers(void) | |
223 | { | |
224 | int cpu; | |
225 | ||
226 | for_each_online_cpu(cpu) | |
227 | stop_stack_timer(cpu); | |
228 | } | |
229 | ||
ada6b835 | 230 | static void stop_stack_trace(struct trace_array *tr) |
f06c3810 | 231 | { |
d618b3e6 | 232 | mutex_lock(&sample_timer_lock); |
0075fa80 | 233 | stop_stack_timers(); |
f06c3810 | 234 | tracer_enabled = 0; |
d618b3e6 | 235 | mutex_unlock(&sample_timer_lock); |
f06c3810 IM |
236 | } |
237 | ||
1c80025a | 238 | static int stack_trace_init(struct trace_array *tr) |
f06c3810 | 239 | { |
56a08bdc | 240 | sysprof_trace = tr; |
f06c3810 | 241 | |
b22f4858 FW |
242 | tracing_start_cmdline_record(); |
243 | ||
b6f11df2 ACM |
244 | mutex_lock(&sample_timer_lock); |
245 | start_stack_timers(); | |
246 | tracer_enabled = 1; | |
247 | mutex_unlock(&sample_timer_lock); | |
1c80025a | 248 | return 0; |
f06c3810 IM |
249 | } |
250 | ||
ada6b835 | 251 | static void stack_trace_reset(struct trace_array *tr) |
f06c3810 | 252 | { |
b22f4858 | 253 | tracing_stop_cmdline_record(); |
c76f0694 | 254 | stop_stack_trace(tr); |
f06c3810 IM |
255 | } |
256 | ||
f06c3810 IM |
257 | static struct tracer stack_trace __read_mostly = |
258 | { | |
259 | .name = "sysprof", | |
260 | .init = stack_trace_init, | |
261 | .reset = stack_trace_reset, | |
f06c3810 | 262 | #ifdef CONFIG_FTRACE_SELFTEST |
a6dd24f8 | 263 | .selftest = trace_selftest_startup_sysprof, |
f06c3810 IM |
264 | #endif |
265 | }; | |
266 | ||
267 | __init static int init_stack_trace(void) | |
268 | { | |
269 | return register_tracer(&stack_trace); | |
270 | } | |
271 | device_initcall(init_stack_trace); | |
d618b3e6 IM |
272 | |
273 | #define MAX_LONG_DIGITS 22 | |
274 | ||
275 | static ssize_t | |
276 | sysprof_sample_read(struct file *filp, char __user *ubuf, | |
277 | size_t cnt, loff_t *ppos) | |
278 | { | |
279 | char buf[MAX_LONG_DIGITS]; | |
280 | int r; | |
281 | ||
282 | r = sprintf(buf, "%ld\n", nsecs_to_usecs(sample_period)); | |
283 | ||
284 | return simple_read_from_buffer(ubuf, cnt, ppos, buf, r); | |
285 | } | |
286 | ||
287 | static ssize_t | |
288 | sysprof_sample_write(struct file *filp, const char __user *ubuf, | |
289 | size_t cnt, loff_t *ppos) | |
290 | { | |
291 | char buf[MAX_LONG_DIGITS]; | |
292 | unsigned long val; | |
293 | ||
294 | if (cnt > MAX_LONG_DIGITS-1) | |
295 | cnt = MAX_LONG_DIGITS-1; | |
296 | ||
297 | if (copy_from_user(&buf, ubuf, cnt)) | |
298 | return -EFAULT; | |
299 | ||
300 | buf[cnt] = 0; | |
301 | ||
302 | val = simple_strtoul(buf, NULL, 10); | |
303 | /* | |
304 | * Enforce a minimum sample period of 100 usecs: | |
305 | */ | |
306 | if (val < 100) | |
307 | val = 100; | |
308 | ||
309 | mutex_lock(&sample_timer_lock); | |
310 | stop_stack_timers(); | |
311 | sample_period = val * 1000; | |
312 | start_stack_timers(); | |
313 | mutex_unlock(&sample_timer_lock); | |
314 | ||
315 | return cnt; | |
316 | } | |
317 | ||
5e2336a0 | 318 | static const struct file_operations sysprof_sample_fops = { |
d618b3e6 IM |
319 | .read = sysprof_sample_read, |
320 | .write = sysprof_sample_write, | |
321 | }; | |
322 | ||
323 | void init_tracer_sysprof_debugfs(struct dentry *d_tracer) | |
324 | { | |
d618b3e6 | 325 | |
5452af66 | 326 | trace_create_file("sysprof_sample_period", 0644, |
d618b3e6 | 327 | d_tracer, NULL, &sysprof_sample_fops); |
d618b3e6 | 328 | } |