xref: /openbmc/linux/kernel/tracepoint.c (revision 22d55f02)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) 2008-2014 Mathieu Desnoyers
4  */
5 #include <linux/module.h>
6 #include <linux/mutex.h>
7 #include <linux/types.h>
8 #include <linux/jhash.h>
9 #include <linux/list.h>
10 #include <linux/rcupdate.h>
11 #include <linux/tracepoint.h>
12 #include <linux/err.h>
13 #include <linux/slab.h>
14 #include <linux/sched/signal.h>
15 #include <linux/sched/task.h>
16 #include <linux/static_key.h>
17 
18 extern tracepoint_ptr_t __start___tracepoints_ptrs[];
19 extern tracepoint_ptr_t __stop___tracepoints_ptrs[];
20 
21 DEFINE_SRCU(tracepoint_srcu);
22 EXPORT_SYMBOL_GPL(tracepoint_srcu);
23 
24 /* Set to 1 to enable tracepoint debug output */
25 static const int tracepoint_debug;
26 
27 #ifdef CONFIG_MODULES
28 /*
29  * Tracepoint module list mutex protects the local module list.
30  */
31 static DEFINE_MUTEX(tracepoint_module_list_mutex);
32 
33 /* Local list of struct tp_module */
34 static LIST_HEAD(tracepoint_module_list);
35 #endif /* CONFIG_MODULES */
36 
37 /*
38  * tracepoints_mutex protects the builtin and module tracepoints.
39  * tracepoints_mutex nests inside tracepoint_module_list_mutex.
40  */
41 static DEFINE_MUTEX(tracepoints_mutex);
42 
43 static struct rcu_head *early_probes;
44 static bool ok_to_free_tracepoints;
45 
46 /*
47  * Note about RCU :
48  * It is used to delay the free of multiple probes array until a quiescent
49  * state is reached.
50  */
51 struct tp_probes {
52 	struct rcu_head rcu;
53 	struct tracepoint_func probes[0];
54 };
55 
56 static inline void *allocate_probes(int count)
57 {
58 	struct tp_probes *p  = kmalloc(count * sizeof(struct tracepoint_func)
59 			+ sizeof(struct tp_probes), GFP_KERNEL);
60 	return p == NULL ? NULL : p->probes;
61 }
62 
63 static void srcu_free_old_probes(struct rcu_head *head)
64 {
65 	kfree(container_of(head, struct tp_probes, rcu));
66 }
67 
68 static void rcu_free_old_probes(struct rcu_head *head)
69 {
70 	call_srcu(&tracepoint_srcu, head, srcu_free_old_probes);
71 }
72 
73 static __init int release_early_probes(void)
74 {
75 	struct rcu_head *tmp;
76 
77 	ok_to_free_tracepoints = true;
78 
79 	while (early_probes) {
80 		tmp = early_probes;
81 		early_probes = tmp->next;
82 		call_rcu(tmp, rcu_free_old_probes);
83 	}
84 
85 	return 0;
86 }
87 
88 /* SRCU is initialized at core_initcall */
89 postcore_initcall(release_early_probes);
90 
91 static inline void release_probes(struct tracepoint_func *old)
92 {
93 	if (old) {
94 		struct tp_probes *tp_probes = container_of(old,
95 			struct tp_probes, probes[0]);
96 
97 		/*
98 		 * We can't free probes if SRCU is not initialized yet.
99 		 * Postpone the freeing till after SRCU is initialized.
100 		 */
101 		if (unlikely(!ok_to_free_tracepoints)) {
102 			tp_probes->rcu.next = early_probes;
103 			early_probes = &tp_probes->rcu;
104 			return;
105 		}
106 
107 		/*
108 		 * Tracepoint probes are protected by both sched RCU and SRCU,
109 		 * by calling the SRCU callback in the sched RCU callback we
110 		 * cover both cases. So let us chain the SRCU and sched RCU
111 		 * callbacks to wait for both grace periods.
112 		 */
113 		call_rcu(&tp_probes->rcu, rcu_free_old_probes);
114 	}
115 }
116 
117 static void debug_print_probes(struct tracepoint_func *funcs)
118 {
119 	int i;
120 
121 	if (!tracepoint_debug || !funcs)
122 		return;
123 
124 	for (i = 0; funcs[i].func; i++)
125 		printk(KERN_DEBUG "Probe %d : %p\n", i, funcs[i].func);
126 }
127 
128 static struct tracepoint_func *
129 func_add(struct tracepoint_func **funcs, struct tracepoint_func *tp_func,
130 	 int prio)
131 {
132 	struct tracepoint_func *old, *new;
133 	int nr_probes = 0;
134 	int pos = -1;
135 
136 	if (WARN_ON(!tp_func->func))
137 		return ERR_PTR(-EINVAL);
138 
139 	debug_print_probes(*funcs);
140 	old = *funcs;
141 	if (old) {
142 		/* (N -> N+1), (N != 0, 1) probes */
143 		for (nr_probes = 0; old[nr_probes].func; nr_probes++) {
144 			/* Insert before probes of lower priority */
145 			if (pos < 0 && old[nr_probes].prio < prio)
146 				pos = nr_probes;
147 			if (old[nr_probes].func == tp_func->func &&
148 			    old[nr_probes].data == tp_func->data)
149 				return ERR_PTR(-EEXIST);
150 		}
151 	}
152 	/* + 2 : one for new probe, one for NULL func */
153 	new = allocate_probes(nr_probes + 2);
154 	if (new == NULL)
155 		return ERR_PTR(-ENOMEM);
156 	if (old) {
157 		if (pos < 0) {
158 			pos = nr_probes;
159 			memcpy(new, old, nr_probes * sizeof(struct tracepoint_func));
160 		} else {
161 			/* Copy higher priority probes ahead of the new probe */
162 			memcpy(new, old, pos * sizeof(struct tracepoint_func));
163 			/* Copy the rest after it. */
164 			memcpy(new + pos + 1, old + pos,
165 			       (nr_probes - pos) * sizeof(struct tracepoint_func));
166 		}
167 	} else
168 		pos = 0;
169 	new[pos] = *tp_func;
170 	new[nr_probes + 1].func = NULL;
171 	*funcs = new;
172 	debug_print_probes(*funcs);
173 	return old;
174 }
175 
176 static void *func_remove(struct tracepoint_func **funcs,
177 		struct tracepoint_func *tp_func)
178 {
179 	int nr_probes = 0, nr_del = 0, i;
180 	struct tracepoint_func *old, *new;
181 
182 	old = *funcs;
183 
184 	if (!old)
185 		return ERR_PTR(-ENOENT);
186 
187 	debug_print_probes(*funcs);
188 	/* (N -> M), (N > 1, M >= 0) probes */
189 	if (tp_func->func) {
190 		for (nr_probes = 0; old[nr_probes].func; nr_probes++) {
191 			if (old[nr_probes].func == tp_func->func &&
192 			     old[nr_probes].data == tp_func->data)
193 				nr_del++;
194 		}
195 	}
196 
197 	/*
198 	 * If probe is NULL, then nr_probes = nr_del = 0, and then the
199 	 * entire entry will be removed.
200 	 */
201 	if (nr_probes - nr_del == 0) {
202 		/* N -> 0, (N > 1) */
203 		*funcs = NULL;
204 		debug_print_probes(*funcs);
205 		return old;
206 	} else {
207 		int j = 0;
208 		/* N -> M, (N > 1, M > 0) */
209 		/* + 1 for NULL */
210 		new = allocate_probes(nr_probes - nr_del + 1);
211 		if (new == NULL)
212 			return ERR_PTR(-ENOMEM);
213 		for (i = 0; old[i].func; i++)
214 			if (old[i].func != tp_func->func
215 					|| old[i].data != tp_func->data)
216 				new[j++] = old[i];
217 		new[nr_probes - nr_del].func = NULL;
218 		*funcs = new;
219 	}
220 	debug_print_probes(*funcs);
221 	return old;
222 }
223 
224 /*
225  * Add the probe function to a tracepoint.
226  */
227 static int tracepoint_add_func(struct tracepoint *tp,
228 			       struct tracepoint_func *func, int prio)
229 {
230 	struct tracepoint_func *old, *tp_funcs;
231 	int ret;
232 
233 	if (tp->regfunc && !static_key_enabled(&tp->key)) {
234 		ret = tp->regfunc();
235 		if (ret < 0)
236 			return ret;
237 	}
238 
239 	tp_funcs = rcu_dereference_protected(tp->funcs,
240 			lockdep_is_held(&tracepoints_mutex));
241 	old = func_add(&tp_funcs, func, prio);
242 	if (IS_ERR(old)) {
243 		WARN_ON_ONCE(PTR_ERR(old) != -ENOMEM);
244 		return PTR_ERR(old);
245 	}
246 
247 	/*
248 	 * rcu_assign_pointer has as smp_store_release() which makes sure
249 	 * that the new probe callbacks array is consistent before setting
250 	 * a pointer to it.  This array is referenced by __DO_TRACE from
251 	 * include/linux/tracepoint.h using rcu_dereference_sched().
252 	 */
253 	rcu_assign_pointer(tp->funcs, tp_funcs);
254 	if (!static_key_enabled(&tp->key))
255 		static_key_slow_inc(&tp->key);
256 	release_probes(old);
257 	return 0;
258 }
259 
260 /*
261  * Remove a probe function from a tracepoint.
262  * Note: only waiting an RCU period after setting elem->call to the empty
263  * function insures that the original callback is not used anymore. This insured
264  * by preempt_disable around the call site.
265  */
266 static int tracepoint_remove_func(struct tracepoint *tp,
267 		struct tracepoint_func *func)
268 {
269 	struct tracepoint_func *old, *tp_funcs;
270 
271 	tp_funcs = rcu_dereference_protected(tp->funcs,
272 			lockdep_is_held(&tracepoints_mutex));
273 	old = func_remove(&tp_funcs, func);
274 	if (IS_ERR(old)) {
275 		WARN_ON_ONCE(PTR_ERR(old) != -ENOMEM);
276 		return PTR_ERR(old);
277 	}
278 
279 	if (!tp_funcs) {
280 		/* Removed last function */
281 		if (tp->unregfunc && static_key_enabled(&tp->key))
282 			tp->unregfunc();
283 
284 		if (static_key_enabled(&tp->key))
285 			static_key_slow_dec(&tp->key);
286 	}
287 	rcu_assign_pointer(tp->funcs, tp_funcs);
288 	release_probes(old);
289 	return 0;
290 }
291 
292 /**
293  * tracepoint_probe_register_prio -  Connect a probe to a tracepoint with priority
294  * @tp: tracepoint
295  * @probe: probe handler
296  * @data: tracepoint data
297  * @prio: priority of this function over other registered functions
298  *
299  * Returns 0 if ok, error value on error.
300  * Note: if @tp is within a module, the caller is responsible for
301  * unregistering the probe before the module is gone. This can be
302  * performed either with a tracepoint module going notifier, or from
303  * within module exit functions.
304  */
305 int tracepoint_probe_register_prio(struct tracepoint *tp, void *probe,
306 				   void *data, int prio)
307 {
308 	struct tracepoint_func tp_func;
309 	int ret;
310 
311 	mutex_lock(&tracepoints_mutex);
312 	tp_func.func = probe;
313 	tp_func.data = data;
314 	tp_func.prio = prio;
315 	ret = tracepoint_add_func(tp, &tp_func, prio);
316 	mutex_unlock(&tracepoints_mutex);
317 	return ret;
318 }
319 EXPORT_SYMBOL_GPL(tracepoint_probe_register_prio);
320 
321 /**
322  * tracepoint_probe_register -  Connect a probe to a tracepoint
323  * @tp: tracepoint
324  * @probe: probe handler
325  * @data: tracepoint data
326  *
327  * Returns 0 if ok, error value on error.
328  * Note: if @tp is within a module, the caller is responsible for
329  * unregistering the probe before the module is gone. This can be
330  * performed either with a tracepoint module going notifier, or from
331  * within module exit functions.
332  */
333 int tracepoint_probe_register(struct tracepoint *tp, void *probe, void *data)
334 {
335 	return tracepoint_probe_register_prio(tp, probe, data, TRACEPOINT_DEFAULT_PRIO);
336 }
337 EXPORT_SYMBOL_GPL(tracepoint_probe_register);
338 
339 /**
340  * tracepoint_probe_unregister -  Disconnect a probe from a tracepoint
341  * @tp: tracepoint
342  * @probe: probe function pointer
343  * @data: tracepoint data
344  *
345  * Returns 0 if ok, error value on error.
346  */
347 int tracepoint_probe_unregister(struct tracepoint *tp, void *probe, void *data)
348 {
349 	struct tracepoint_func tp_func;
350 	int ret;
351 
352 	mutex_lock(&tracepoints_mutex);
353 	tp_func.func = probe;
354 	tp_func.data = data;
355 	ret = tracepoint_remove_func(tp, &tp_func);
356 	mutex_unlock(&tracepoints_mutex);
357 	return ret;
358 }
359 EXPORT_SYMBOL_GPL(tracepoint_probe_unregister);
360 
361 static void for_each_tracepoint_range(
362 		tracepoint_ptr_t *begin, tracepoint_ptr_t *end,
363 		void (*fct)(struct tracepoint *tp, void *priv),
364 		void *priv)
365 {
366 	tracepoint_ptr_t *iter;
367 
368 	if (!begin)
369 		return;
370 	for (iter = begin; iter < end; iter++)
371 		fct(tracepoint_ptr_deref(iter), priv);
372 }
373 
374 #ifdef CONFIG_MODULES
375 bool trace_module_has_bad_taint(struct module *mod)
376 {
377 	return mod->taints & ~((1 << TAINT_OOT_MODULE) | (1 << TAINT_CRAP) |
378 			       (1 << TAINT_UNSIGNED_MODULE));
379 }
380 
381 static BLOCKING_NOTIFIER_HEAD(tracepoint_notify_list);
382 
383 /**
384  * register_tracepoint_notifier - register tracepoint coming/going notifier
385  * @nb: notifier block
386  *
387  * Notifiers registered with this function are called on module
388  * coming/going with the tracepoint_module_list_mutex held.
389  * The notifier block callback should expect a "struct tp_module" data
390  * pointer.
391  */
392 int register_tracepoint_module_notifier(struct notifier_block *nb)
393 {
394 	struct tp_module *tp_mod;
395 	int ret;
396 
397 	mutex_lock(&tracepoint_module_list_mutex);
398 	ret = blocking_notifier_chain_register(&tracepoint_notify_list, nb);
399 	if (ret)
400 		goto end;
401 	list_for_each_entry(tp_mod, &tracepoint_module_list, list)
402 		(void) nb->notifier_call(nb, MODULE_STATE_COMING, tp_mod);
403 end:
404 	mutex_unlock(&tracepoint_module_list_mutex);
405 	return ret;
406 }
407 EXPORT_SYMBOL_GPL(register_tracepoint_module_notifier);
408 
409 /**
410  * unregister_tracepoint_notifier - unregister tracepoint coming/going notifier
411  * @nb: notifier block
412  *
413  * The notifier block callback should expect a "struct tp_module" data
414  * pointer.
415  */
416 int unregister_tracepoint_module_notifier(struct notifier_block *nb)
417 {
418 	struct tp_module *tp_mod;
419 	int ret;
420 
421 	mutex_lock(&tracepoint_module_list_mutex);
422 	ret = blocking_notifier_chain_unregister(&tracepoint_notify_list, nb);
423 	if (ret)
424 		goto end;
425 	list_for_each_entry(tp_mod, &tracepoint_module_list, list)
426 		(void) nb->notifier_call(nb, MODULE_STATE_GOING, tp_mod);
427 end:
428 	mutex_unlock(&tracepoint_module_list_mutex);
429 	return ret;
430 
431 }
432 EXPORT_SYMBOL_GPL(unregister_tracepoint_module_notifier);
433 
434 /*
435  * Ensure the tracer unregistered the module's probes before the module
436  * teardown is performed. Prevents leaks of probe and data pointers.
437  */
438 static void tp_module_going_check_quiescent(struct tracepoint *tp, void *priv)
439 {
440 	WARN_ON_ONCE(tp->funcs);
441 }
442 
443 static int tracepoint_module_coming(struct module *mod)
444 {
445 	struct tp_module *tp_mod;
446 	int ret = 0;
447 
448 	if (!mod->num_tracepoints)
449 		return 0;
450 
451 	/*
452 	 * We skip modules that taint the kernel, especially those with different
453 	 * module headers (for forced load), to make sure we don't cause a crash.
454 	 * Staging, out-of-tree, and unsigned GPL modules are fine.
455 	 */
456 	if (trace_module_has_bad_taint(mod))
457 		return 0;
458 	mutex_lock(&tracepoint_module_list_mutex);
459 	tp_mod = kmalloc(sizeof(struct tp_module), GFP_KERNEL);
460 	if (!tp_mod) {
461 		ret = -ENOMEM;
462 		goto end;
463 	}
464 	tp_mod->mod = mod;
465 	list_add_tail(&tp_mod->list, &tracepoint_module_list);
466 	blocking_notifier_call_chain(&tracepoint_notify_list,
467 			MODULE_STATE_COMING, tp_mod);
468 end:
469 	mutex_unlock(&tracepoint_module_list_mutex);
470 	return ret;
471 }
472 
473 static void tracepoint_module_going(struct module *mod)
474 {
475 	struct tp_module *tp_mod;
476 
477 	if (!mod->num_tracepoints)
478 		return;
479 
480 	mutex_lock(&tracepoint_module_list_mutex);
481 	list_for_each_entry(tp_mod, &tracepoint_module_list, list) {
482 		if (tp_mod->mod == mod) {
483 			blocking_notifier_call_chain(&tracepoint_notify_list,
484 					MODULE_STATE_GOING, tp_mod);
485 			list_del(&tp_mod->list);
486 			kfree(tp_mod);
487 			/*
488 			 * Called the going notifier before checking for
489 			 * quiescence.
490 			 */
491 			for_each_tracepoint_range(mod->tracepoints_ptrs,
492 				mod->tracepoints_ptrs + mod->num_tracepoints,
493 				tp_module_going_check_quiescent, NULL);
494 			break;
495 		}
496 	}
497 	/*
498 	 * In the case of modules that were tainted at "coming", we'll simply
499 	 * walk through the list without finding it. We cannot use the "tainted"
500 	 * flag on "going", in case a module taints the kernel only after being
501 	 * loaded.
502 	 */
503 	mutex_unlock(&tracepoint_module_list_mutex);
504 }
505 
506 static int tracepoint_module_notify(struct notifier_block *self,
507 		unsigned long val, void *data)
508 {
509 	struct module *mod = data;
510 	int ret = 0;
511 
512 	switch (val) {
513 	case MODULE_STATE_COMING:
514 		ret = tracepoint_module_coming(mod);
515 		break;
516 	case MODULE_STATE_LIVE:
517 		break;
518 	case MODULE_STATE_GOING:
519 		tracepoint_module_going(mod);
520 		break;
521 	case MODULE_STATE_UNFORMED:
522 		break;
523 	}
524 	return ret;
525 }
526 
527 static struct notifier_block tracepoint_module_nb = {
528 	.notifier_call = tracepoint_module_notify,
529 	.priority = 0,
530 };
531 
532 static __init int init_tracepoints(void)
533 {
534 	int ret;
535 
536 	ret = register_module_notifier(&tracepoint_module_nb);
537 	if (ret)
538 		pr_warn("Failed to register tracepoint module enter notifier\n");
539 
540 	return ret;
541 }
542 __initcall(init_tracepoints);
543 #endif /* CONFIG_MODULES */
544 
545 /**
546  * for_each_kernel_tracepoint - iteration on all kernel tracepoints
547  * @fct: callback
548  * @priv: private data
549  */
550 void for_each_kernel_tracepoint(void (*fct)(struct tracepoint *tp, void *priv),
551 		void *priv)
552 {
553 	for_each_tracepoint_range(__start___tracepoints_ptrs,
554 		__stop___tracepoints_ptrs, fct, priv);
555 }
556 EXPORT_SYMBOL_GPL(for_each_kernel_tracepoint);
557 
558 #ifdef CONFIG_HAVE_SYSCALL_TRACEPOINTS
559 
560 /* NB: reg/unreg are called while guarded with the tracepoints_mutex */
561 static int sys_tracepoint_refcount;
562 
563 int syscall_regfunc(void)
564 {
565 	struct task_struct *p, *t;
566 
567 	if (!sys_tracepoint_refcount) {
568 		read_lock(&tasklist_lock);
569 		for_each_process_thread(p, t) {
570 			set_tsk_thread_flag(t, TIF_SYSCALL_TRACEPOINT);
571 		}
572 		read_unlock(&tasklist_lock);
573 	}
574 	sys_tracepoint_refcount++;
575 
576 	return 0;
577 }
578 
579 void syscall_unregfunc(void)
580 {
581 	struct task_struct *p, *t;
582 
583 	sys_tracepoint_refcount--;
584 	if (!sys_tracepoint_refcount) {
585 		read_lock(&tasklist_lock);
586 		for_each_process_thread(p, t) {
587 			clear_tsk_thread_flag(t, TIF_SYSCALL_TRACEPOINT);
588 		}
589 		read_unlock(&tasklist_lock);
590 	}
591 }
592 #endif
593