xref: /linux/kernel/trace/trace_stack.c (revision 856e7c4b619af622d56b3b454f7bec32a170ac99)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
4  *
5  */
6 #include <linux/sched/task_stack.h>
7 #include <linux/stacktrace.h>
8 #include <linux/kallsyms.h>
9 #include <linux/seq_file.h>
10 #include <linux/spinlock.h>
11 #include <linux/uaccess.h>
12 #include <linux/ftrace.h>
13 #include <linux/module.h>
14 #include <linux/sysctl.h>
15 #include <linux/init.h>
16 
17 #include <asm/setup.h>
18 
19 #include "trace.h"
20 
21 static unsigned long stack_dump_trace[STACK_TRACE_ENTRIES+1] =
22 	 { [0 ... (STACK_TRACE_ENTRIES)] = ULONG_MAX };
23 unsigned stack_trace_index[STACK_TRACE_ENTRIES];
24 
25 /*
26  * Reserve one entry for the passed in ip. This will allow
27  * us to remove most or all of the stack size overhead
28  * added by the stack tracer itself.
29  */
30 struct stack_trace stack_trace_max = {
31 	.max_entries		= STACK_TRACE_ENTRIES - 1,
32 	.entries		= &stack_dump_trace[0],
33 };
34 
35 unsigned long stack_trace_max_size;
36 arch_spinlock_t stack_trace_max_lock =
37 	(arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
38 
39 DEFINE_PER_CPU(int, disable_stack_tracer);
40 static DEFINE_MUTEX(stack_sysctl_mutex);
41 
42 int stack_tracer_enabled;
43 static int last_stack_tracer_enabled;
44 
45 void stack_trace_print(void)
46 {
47 	long i;
48 	int size;
49 
50 	pr_emerg("        Depth    Size   Location    (%d entries)\n"
51 			   "        -----    ----   --------\n",
52 			   stack_trace_max.nr_entries);
53 
54 	for (i = 0; i < stack_trace_max.nr_entries; i++) {
55 		if (stack_dump_trace[i] == ULONG_MAX)
56 			break;
57 		if (i+1 == stack_trace_max.nr_entries ||
58 				stack_dump_trace[i+1] == ULONG_MAX)
59 			size = stack_trace_index[i];
60 		else
61 			size = stack_trace_index[i] - stack_trace_index[i+1];
62 
63 		pr_emerg("%3ld) %8d   %5d   %pS\n", i, stack_trace_index[i],
64 				size, (void *)stack_dump_trace[i]);
65 	}
66 }
67 
68 /*
69  * When arch-specific code overrides this function, the following
70  * data should be filled up, assuming stack_trace_max_lock is held to
71  * prevent concurrent updates.
72  *     stack_trace_index[]
73  *     stack_trace_max
74  *     stack_trace_max_size
75  */
76 void __weak
77 check_stack(unsigned long ip, unsigned long *stack)
78 {
79 	unsigned long this_size, flags; unsigned long *p, *top, *start;
80 	static int tracer_frame;
81 	int frame_size = READ_ONCE(tracer_frame);
82 	int i, x;
83 
84 	this_size = ((unsigned long)stack) & (THREAD_SIZE-1);
85 	this_size = THREAD_SIZE - this_size;
86 	/* Remove the frame of the tracer */
87 	this_size -= frame_size;
88 
89 	if (this_size <= stack_trace_max_size)
90 		return;
91 
92 	/* we do not handle interrupt stacks yet */
93 	if (!object_is_on_stack(stack))
94 		return;
95 
96 	/* Can't do this from NMI context (can cause deadlocks) */
97 	if (in_nmi())
98 		return;
99 
100 	local_irq_save(flags);
101 	arch_spin_lock(&stack_trace_max_lock);
102 
103 	/* In case another CPU set the tracer_frame on us */
104 	if (unlikely(!frame_size))
105 		this_size -= tracer_frame;
106 
107 	/* a race could have already updated it */
108 	if (this_size <= stack_trace_max_size)
109 		goto out;
110 
111 	stack_trace_max_size = this_size;
112 
113 	stack_trace_max.nr_entries = 0;
114 	stack_trace_max.skip = 3;
115 
116 	save_stack_trace(&stack_trace_max);
117 
118 	/* Skip over the overhead of the stack tracer itself */
119 	for (i = 0; i < stack_trace_max.nr_entries; i++) {
120 		if (stack_dump_trace[i] == ip)
121 			break;
122 	}
123 
124 	/*
125 	 * Some archs may not have the passed in ip in the dump.
126 	 * If that happens, we need to show everything.
127 	 */
128 	if (i == stack_trace_max.nr_entries)
129 		i = 0;
130 
131 	/*
132 	 * Now find where in the stack these are.
133 	 */
134 	x = 0;
135 	start = stack;
136 	top = (unsigned long *)
137 		(((unsigned long)start & ~(THREAD_SIZE-1)) + THREAD_SIZE);
138 
139 	/*
140 	 * Loop through all the entries. One of the entries may
141 	 * for some reason be missed on the stack, so we may
142 	 * have to account for them. If they are all there, this
143 	 * loop will only happen once. This code only takes place
144 	 * on a new max, so it is far from a fast path.
145 	 */
146 	while (i < stack_trace_max.nr_entries) {
147 		int found = 0;
148 
149 		stack_trace_index[x] = this_size;
150 		p = start;
151 
152 		for (; p < top && i < stack_trace_max.nr_entries; p++) {
153 			if (stack_dump_trace[i] == ULONG_MAX)
154 				break;
155 			/*
156 			 * The READ_ONCE_NOCHECK is used to let KASAN know that
157 			 * this is not a stack-out-of-bounds error.
158 			 */
159 			if ((READ_ONCE_NOCHECK(*p)) == stack_dump_trace[i]) {
160 				stack_dump_trace[x] = stack_dump_trace[i++];
161 				this_size = stack_trace_index[x++] =
162 					(top - p) * sizeof(unsigned long);
163 				found = 1;
164 				/* Start the search from here */
165 				start = p + 1;
166 				/*
167 				 * We do not want to show the overhead
168 				 * of the stack tracer stack in the
169 				 * max stack. If we haven't figured
170 				 * out what that is, then figure it out
171 				 * now.
172 				 */
173 				if (unlikely(!tracer_frame)) {
174 					tracer_frame = (p - stack) *
175 						sizeof(unsigned long);
176 					stack_trace_max_size -= tracer_frame;
177 				}
178 			}
179 		}
180 
181 		if (!found)
182 			i++;
183 	}
184 
185 	stack_trace_max.nr_entries = x;
186 	for (; x < i; x++)
187 		stack_dump_trace[x] = ULONG_MAX;
188 
189 	if (task_stack_end_corrupted(current)) {
190 		stack_trace_print();
191 		BUG();
192 	}
193 
194  out:
195 	arch_spin_unlock(&stack_trace_max_lock);
196 	local_irq_restore(flags);
197 }
198 
199 static void
200 stack_trace_call(unsigned long ip, unsigned long parent_ip,
201 		 struct ftrace_ops *op, struct pt_regs *pt_regs)
202 {
203 	unsigned long stack;
204 
205 	preempt_disable_notrace();
206 
207 	/* no atomic needed, we only modify this variable by this cpu */
208 	__this_cpu_inc(disable_stack_tracer);
209 	if (__this_cpu_read(disable_stack_tracer) != 1)
210 		goto out;
211 
212 	/* If rcu is not watching, then save stack trace can fail */
213 	if (!rcu_is_watching())
214 		goto out;
215 
216 	ip += MCOUNT_INSN_SIZE;
217 
218 	check_stack(ip, &stack);
219 
220  out:
221 	__this_cpu_dec(disable_stack_tracer);
222 	/* prevent recursion in schedule */
223 	preempt_enable_notrace();
224 }
225 
226 static struct ftrace_ops trace_ops __read_mostly =
227 {
228 	.func = stack_trace_call,
229 	.flags = FTRACE_OPS_FL_RECURSION_SAFE,
230 };
231 
232 static ssize_t
233 stack_max_size_read(struct file *filp, char __user *ubuf,
234 		    size_t count, loff_t *ppos)
235 {
236 	unsigned long *ptr = filp->private_data;
237 	char buf[64];
238 	int r;
239 
240 	r = snprintf(buf, sizeof(buf), "%ld\n", *ptr);
241 	if (r > sizeof(buf))
242 		r = sizeof(buf);
243 	return simple_read_from_buffer(ubuf, count, ppos, buf, r);
244 }
245 
246 static ssize_t
247 stack_max_size_write(struct file *filp, const char __user *ubuf,
248 		     size_t count, loff_t *ppos)
249 {
250 	long *ptr = filp->private_data;
251 	unsigned long val, flags;
252 	int ret;
253 
254 	ret = kstrtoul_from_user(ubuf, count, 10, &val);
255 	if (ret)
256 		return ret;
257 
258 	local_irq_save(flags);
259 
260 	/*
261 	 * In case we trace inside arch_spin_lock() or after (NMI),
262 	 * we will cause circular lock, so we also need to increase
263 	 * the percpu disable_stack_tracer here.
264 	 */
265 	__this_cpu_inc(disable_stack_tracer);
266 
267 	arch_spin_lock(&stack_trace_max_lock);
268 	*ptr = val;
269 	arch_spin_unlock(&stack_trace_max_lock);
270 
271 	__this_cpu_dec(disable_stack_tracer);
272 	local_irq_restore(flags);
273 
274 	return count;
275 }
276 
277 static const struct file_operations stack_max_size_fops = {
278 	.open		= tracing_open_generic,
279 	.read		= stack_max_size_read,
280 	.write		= stack_max_size_write,
281 	.llseek		= default_llseek,
282 };
283 
284 static void *
285 __next(struct seq_file *m, loff_t *pos)
286 {
287 	long n = *pos - 1;
288 
289 	if (n > stack_trace_max.nr_entries || stack_dump_trace[n] == ULONG_MAX)
290 		return NULL;
291 
292 	m->private = (void *)n;
293 	return &m->private;
294 }
295 
296 static void *
297 t_next(struct seq_file *m, void *v, loff_t *pos)
298 {
299 	(*pos)++;
300 	return __next(m, pos);
301 }
302 
303 static void *t_start(struct seq_file *m, loff_t *pos)
304 {
305 	local_irq_disable();
306 
307 	__this_cpu_inc(disable_stack_tracer);
308 
309 	arch_spin_lock(&stack_trace_max_lock);
310 
311 	if (*pos == 0)
312 		return SEQ_START_TOKEN;
313 
314 	return __next(m, pos);
315 }
316 
317 static void t_stop(struct seq_file *m, void *p)
318 {
319 	arch_spin_unlock(&stack_trace_max_lock);
320 
321 	__this_cpu_dec(disable_stack_tracer);
322 
323 	local_irq_enable();
324 }
325 
326 static void trace_lookup_stack(struct seq_file *m, long i)
327 {
328 	unsigned long addr = stack_dump_trace[i];
329 
330 	seq_printf(m, "%pS\n", (void *)addr);
331 }
332 
333 static void print_disabled(struct seq_file *m)
334 {
335 	seq_puts(m, "#\n"
336 		 "#  Stack tracer disabled\n"
337 		 "#\n"
338 		 "# To enable the stack tracer, either add 'stacktrace' to the\n"
339 		 "# kernel command line\n"
340 		 "# or 'echo 1 > /proc/sys/kernel/stack_tracer_enabled'\n"
341 		 "#\n");
342 }
343 
344 static int t_show(struct seq_file *m, void *v)
345 {
346 	long i;
347 	int size;
348 
349 	if (v == SEQ_START_TOKEN) {
350 		seq_printf(m, "        Depth    Size   Location"
351 			   "    (%d entries)\n"
352 			   "        -----    ----   --------\n",
353 			   stack_trace_max.nr_entries);
354 
355 		if (!stack_tracer_enabled && !stack_trace_max_size)
356 			print_disabled(m);
357 
358 		return 0;
359 	}
360 
361 	i = *(long *)v;
362 
363 	if (i >= stack_trace_max.nr_entries ||
364 	    stack_dump_trace[i] == ULONG_MAX)
365 		return 0;
366 
367 	if (i+1 == stack_trace_max.nr_entries ||
368 	    stack_dump_trace[i+1] == ULONG_MAX)
369 		size = stack_trace_index[i];
370 	else
371 		size = stack_trace_index[i] - stack_trace_index[i+1];
372 
373 	seq_printf(m, "%3ld) %8d   %5d   ", i, stack_trace_index[i], size);
374 
375 	trace_lookup_stack(m, i);
376 
377 	return 0;
378 }
379 
380 static const struct seq_operations stack_trace_seq_ops = {
381 	.start		= t_start,
382 	.next		= t_next,
383 	.stop		= t_stop,
384 	.show		= t_show,
385 };
386 
387 static int stack_trace_open(struct inode *inode, struct file *file)
388 {
389 	return seq_open(file, &stack_trace_seq_ops);
390 }
391 
392 static const struct file_operations stack_trace_fops = {
393 	.open		= stack_trace_open,
394 	.read		= seq_read,
395 	.llseek		= seq_lseek,
396 	.release	= seq_release,
397 };
398 
399 #ifdef CONFIG_DYNAMIC_FTRACE
400 
401 static int
402 stack_trace_filter_open(struct inode *inode, struct file *file)
403 {
404 	struct ftrace_ops *ops = inode->i_private;
405 
406 	return ftrace_regex_open(ops, FTRACE_ITER_FILTER,
407 				 inode, file);
408 }
409 
410 static const struct file_operations stack_trace_filter_fops = {
411 	.open = stack_trace_filter_open,
412 	.read = seq_read,
413 	.write = ftrace_filter_write,
414 	.llseek = tracing_lseek,
415 	.release = ftrace_regex_release,
416 };
417 
418 #endif /* CONFIG_DYNAMIC_FTRACE */
419 
420 int
421 stack_trace_sysctl(struct ctl_table *table, int write,
422 		   void __user *buffer, size_t *lenp,
423 		   loff_t *ppos)
424 {
425 	int ret;
426 
427 	mutex_lock(&stack_sysctl_mutex);
428 
429 	ret = proc_dointvec(table, write, buffer, lenp, ppos);
430 
431 	if (ret || !write ||
432 	    (last_stack_tracer_enabled == !!stack_tracer_enabled))
433 		goto out;
434 
435 	last_stack_tracer_enabled = !!stack_tracer_enabled;
436 
437 	if (stack_tracer_enabled)
438 		register_ftrace_function(&trace_ops);
439 	else
440 		unregister_ftrace_function(&trace_ops);
441 
442  out:
443 	mutex_unlock(&stack_sysctl_mutex);
444 	return ret;
445 }
446 
447 static char stack_trace_filter_buf[COMMAND_LINE_SIZE+1] __initdata;
448 
449 static __init int enable_stacktrace(char *str)
450 {
451 	if (strncmp(str, "_filter=", 8) == 0)
452 		strncpy(stack_trace_filter_buf, str+8, COMMAND_LINE_SIZE);
453 
454 	stack_tracer_enabled = 1;
455 	last_stack_tracer_enabled = 1;
456 	return 1;
457 }
458 __setup("stacktrace", enable_stacktrace);
459 
460 static __init int stack_trace_init(void)
461 {
462 	struct dentry *d_tracer;
463 
464 	d_tracer = tracing_init_dentry();
465 	if (IS_ERR(d_tracer))
466 		return 0;
467 
468 	trace_create_file("stack_max_size", 0644, d_tracer,
469 			&stack_trace_max_size, &stack_max_size_fops);
470 
471 	trace_create_file("stack_trace", 0444, d_tracer,
472 			NULL, &stack_trace_fops);
473 
474 #ifdef CONFIG_DYNAMIC_FTRACE
475 	trace_create_file("stack_trace_filter", 0644, d_tracer,
476 			  &trace_ops, &stack_trace_filter_fops);
477 #endif
478 
479 	if (stack_trace_filter_buf[0])
480 		ftrace_set_early_filter(&trace_ops, stack_trace_filter_buf, 1);
481 
482 	if (stack_tracer_enabled)
483 		register_ftrace_function(&trace_ops);
484 
485 	return 0;
486 }
487 
488 device_initcall(stack_trace_init);
489