1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Infrastructure to took into function calls and returns.
4 * Copyright (c) 2008-2009 Frederic Weisbecker <fweisbec@gmail.com>
5 * Mostly borrowed from function tracer which
6 * is Copyright (c) Steven Rostedt <srostedt@redhat.com>
7 *
8 * Highly modified by Steven Rostedt (VMware).
9 */
10 #include <linux/suspend.h>
11 #include <linux/ftrace.h>
12 #include <linux/slab.h>
13
14 #include <trace/events/sched.h>
15
16 #include "ftrace_internal.h"
17
18 #ifdef CONFIG_DYNAMIC_FTRACE
19 #define ASSIGN_OPS_HASH(opsname, val) \
20 .func_hash = val, \
21 .local_hash.regex_lock = __MUTEX_INITIALIZER(opsname.local_hash.regex_lock),
22 #else
23 #define ASSIGN_OPS_HASH(opsname, val)
24 #endif
25
26 static bool kill_ftrace_graph;
27 int ftrace_graph_active;
28
29 /* Both enabled by default (can be cleared by function_graph tracer flags */
30 static bool fgraph_sleep_time = true;
31
32 /**
33 * ftrace_graph_is_dead - returns true if ftrace_graph_stop() was called
34 *
35 * ftrace_graph_stop() is called when a severe error is detected in
36 * the function graph tracing. This function is called by the critical
37 * paths of function graph to keep those paths from doing any more harm.
38 */
ftrace_graph_is_dead(void)39 bool ftrace_graph_is_dead(void)
40 {
41 return kill_ftrace_graph;
42 }
43
44 /**
45 * ftrace_graph_stop - set to permanently disable function graph tracincg
46 *
47 * In case of an error int function graph tracing, this is called
48 * to try to keep function graph tracing from causing any more harm.
49 * Usually this is pretty severe and this is called to try to at least
50 * get a warning out to the user.
51 */
ftrace_graph_stop(void)52 void ftrace_graph_stop(void)
53 {
54 kill_ftrace_graph = true;
55 }
56
57 /* Add a function return address to the trace stack on thread info.*/
58 static int
ftrace_push_return_trace(unsigned long ret,unsigned long func,unsigned long frame_pointer,unsigned long * retp)59 ftrace_push_return_trace(unsigned long ret, unsigned long func,
60 unsigned long frame_pointer, unsigned long *retp)
61 {
62 unsigned long long calltime;
63 int index;
64
65 if (unlikely(ftrace_graph_is_dead()))
66 return -EBUSY;
67
68 if (!current->ret_stack)
69 return -EBUSY;
70
71 /*
72 * We must make sure the ret_stack is tested before we read
73 * anything else.
74 */
75 smp_rmb();
76
77 /* The return trace stack is full */
78 if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
79 atomic_inc(¤t->trace_overrun);
80 return -EBUSY;
81 }
82
83 calltime = trace_clock_local();
84
85 index = ++current->curr_ret_stack;
86 barrier();
87 current->ret_stack[index].ret = ret;
88 current->ret_stack[index].func = func;
89 current->ret_stack[index].calltime = calltime;
90 #ifdef HAVE_FUNCTION_GRAPH_FP_TEST
91 current->ret_stack[index].fp = frame_pointer;
92 #endif
93 #ifdef HAVE_FUNCTION_GRAPH_RET_ADDR_PTR
94 current->ret_stack[index].retp = retp;
95 #endif
96 return 0;
97 }
98
99 /*
100 * Not all archs define MCOUNT_INSN_SIZE which is used to look for direct
101 * functions. But those archs currently don't support direct functions
102 * anyway, and ftrace_find_rec_direct() is just a stub for them.
103 * Define MCOUNT_INSN_SIZE to keep those archs compiling.
104 */
105 #ifndef MCOUNT_INSN_SIZE
106 /* Make sure this only works without direct calls */
107 # ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
108 # error MCOUNT_INSN_SIZE not defined with direct calls enabled
109 # endif
110 # define MCOUNT_INSN_SIZE 0
111 #endif
112
function_graph_enter(unsigned long ret,unsigned long func,unsigned long frame_pointer,unsigned long * retp)113 int function_graph_enter(unsigned long ret, unsigned long func,
114 unsigned long frame_pointer, unsigned long *retp)
115 {
116 struct ftrace_graph_ent trace;
117
118 /*
119 * Skip graph tracing if the return location is served by direct trampoline,
120 * since call sequence and return addresses is unpredicatable anymore.
121 * Ex: BPF trampoline may call original function and may skip frame
122 * depending on type of BPF programs attached.
123 */
124 if (ftrace_direct_func_count &&
125 ftrace_find_rec_direct(ret - MCOUNT_INSN_SIZE))
126 return -EBUSY;
127 trace.func = func;
128 trace.depth = ++current->curr_ret_depth;
129
130 if (ftrace_push_return_trace(ret, func, frame_pointer, retp))
131 goto out;
132
133 /* Only trace if the calling function expects to */
134 if (!ftrace_graph_entry(&trace))
135 goto out_ret;
136
137 return 0;
138 out_ret:
139 current->curr_ret_stack--;
140 out:
141 current->curr_ret_depth--;
142 return -EBUSY;
143 }
144
145 /* Retrieve a function return address to the trace stack on thread info.*/
146 static void
ftrace_pop_return_trace(struct ftrace_graph_ret * trace,unsigned long * ret,unsigned long frame_pointer)147 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
148 unsigned long frame_pointer)
149 {
150 int index;
151
152 index = current->curr_ret_stack;
153
154 if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
155 ftrace_graph_stop();
156 WARN_ON(1);
157 /* Might as well panic, otherwise we have no where to go */
158 *ret = (unsigned long)panic;
159 return;
160 }
161
162 #ifdef HAVE_FUNCTION_GRAPH_FP_TEST
163 /*
164 * The arch may choose to record the frame pointer used
165 * and check it here to make sure that it is what we expect it
166 * to be. If gcc does not set the place holder of the return
167 * address in the frame pointer, and does a copy instead, then
168 * the function graph trace will fail. This test detects this
169 * case.
170 *
171 * Currently, x86_32 with optimize for size (-Os) makes the latest
172 * gcc do the above.
173 *
174 * Note, -mfentry does not use frame pointers, and this test
175 * is not needed if CC_USING_FENTRY is set.
176 */
177 if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
178 ftrace_graph_stop();
179 WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
180 " from func %ps return to %lx\n",
181 current->ret_stack[index].fp,
182 frame_pointer,
183 (void *)current->ret_stack[index].func,
184 current->ret_stack[index].ret);
185 *ret = (unsigned long)panic;
186 return;
187 }
188 #endif
189
190 *ret = current->ret_stack[index].ret;
191 trace->func = current->ret_stack[index].func;
192 trace->calltime = current->ret_stack[index].calltime;
193 trace->overrun = atomic_read(¤t->trace_overrun);
194 trace->depth = current->curr_ret_depth--;
195 /*
196 * We still want to trace interrupts coming in if
197 * max_depth is set to 1. Make sure the decrement is
198 * seen before ftrace_graph_return.
199 */
200 barrier();
201 }
202
203 /*
204 * Hibernation protection.
205 * The state of the current task is too much unstable during
206 * suspend/restore to disk. We want to protect against that.
207 */
208 static int
ftrace_suspend_notifier_call(struct notifier_block * bl,unsigned long state,void * unused)209 ftrace_suspend_notifier_call(struct notifier_block *bl, unsigned long state,
210 void *unused)
211 {
212 switch (state) {
213 case PM_HIBERNATION_PREPARE:
214 pause_graph_tracing();
215 break;
216
217 case PM_POST_HIBERNATION:
218 unpause_graph_tracing();
219 break;
220 }
221 return NOTIFY_DONE;
222 }
223
224 static struct notifier_block ftrace_suspend_notifier = {
225 .notifier_call = ftrace_suspend_notifier_call,
226 };
227
228 /*
229 * Send the trace to the ring-buffer.
230 * @return the original return address.
231 */
ftrace_return_to_handler(unsigned long frame_pointer)232 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
233 {
234 struct ftrace_graph_ret trace;
235 unsigned long ret;
236
237 ftrace_pop_return_trace(&trace, &ret, frame_pointer);
238 trace.rettime = trace_clock_local();
239 ftrace_graph_return(&trace);
240 /*
241 * The ftrace_graph_return() may still access the current
242 * ret_stack structure, we need to make sure the update of
243 * curr_ret_stack is after that.
244 */
245 barrier();
246 current->curr_ret_stack--;
247
248 if (unlikely(!ret)) {
249 ftrace_graph_stop();
250 WARN_ON(1);
251 /* Might as well panic. What else to do? */
252 ret = (unsigned long)panic;
253 }
254
255 return ret;
256 }
257
258 /**
259 * ftrace_graph_get_ret_stack - return the entry of the shadow stack
260 * @task: The task to read the shadow stack from
261 * @idx: Index down the shadow stack
262 *
263 * Return the ret_struct on the shadow stack of the @task at the
264 * call graph at @idx starting with zero. If @idx is zero, it
265 * will return the last saved ret_stack entry. If it is greater than
266 * zero, it will return the corresponding ret_stack for the depth
267 * of saved return addresses.
268 */
269 struct ftrace_ret_stack *
ftrace_graph_get_ret_stack(struct task_struct * task,int idx)270 ftrace_graph_get_ret_stack(struct task_struct *task, int idx)
271 {
272 idx = task->curr_ret_stack - idx;
273
274 if (idx >= 0 && idx <= task->curr_ret_stack)
275 return &task->ret_stack[idx];
276
277 return NULL;
278 }
279
280 /**
281 * ftrace_graph_ret_addr - convert a potentially modified stack return address
282 * to its original value
283 *
284 * This function can be called by stack unwinding code to convert a found stack
285 * return address ('ret') to its original value, in case the function graph
286 * tracer has modified it to be 'return_to_handler'. If the address hasn't
287 * been modified, the unchanged value of 'ret' is returned.
288 *
289 * 'idx' is a state variable which should be initialized by the caller to zero
290 * before the first call.
291 *
292 * 'retp' is a pointer to the return address on the stack. It's ignored if
293 * the arch doesn't have HAVE_FUNCTION_GRAPH_RET_ADDR_PTR defined.
294 */
295 #ifdef HAVE_FUNCTION_GRAPH_RET_ADDR_PTR
ftrace_graph_ret_addr(struct task_struct * task,int * idx,unsigned long ret,unsigned long * retp)296 unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
297 unsigned long ret, unsigned long *retp)
298 {
299 int index = task->curr_ret_stack;
300 int i;
301
302 if (ret != (unsigned long)dereference_kernel_function_descriptor(return_to_handler))
303 return ret;
304
305 if (index < 0)
306 return ret;
307
308 for (i = 0; i <= index; i++)
309 if (task->ret_stack[i].retp == retp)
310 return task->ret_stack[i].ret;
311
312 return ret;
313 }
314 #else /* !HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
ftrace_graph_ret_addr(struct task_struct * task,int * idx,unsigned long ret,unsigned long * retp)315 unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
316 unsigned long ret, unsigned long *retp)
317 {
318 int task_idx;
319
320 if (ret != (unsigned long)dereference_kernel_function_descriptor(return_to_handler))
321 return ret;
322
323 task_idx = task->curr_ret_stack;
324
325 if (!task->ret_stack || task_idx < *idx)
326 return ret;
327
328 task_idx -= *idx;
329 (*idx)++;
330
331 return task->ret_stack[task_idx].ret;
332 }
333 #endif /* HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
334
335 static struct ftrace_ops graph_ops = {
336 .func = ftrace_stub,
337 .flags = FTRACE_OPS_FL_RECURSION_SAFE |
338 FTRACE_OPS_FL_INITIALIZED |
339 FTRACE_OPS_FL_PID |
340 FTRACE_OPS_FL_STUB,
341 #ifdef FTRACE_GRAPH_TRAMP_ADDR
342 .trampoline = FTRACE_GRAPH_TRAMP_ADDR,
343 /* trampoline_size is only needed for dynamically allocated tramps */
344 #endif
345 ASSIGN_OPS_HASH(graph_ops, &global_ops.local_hash)
346 };
347
ftrace_graph_sleep_time_control(bool enable)348 void ftrace_graph_sleep_time_control(bool enable)
349 {
350 fgraph_sleep_time = enable;
351 }
352
ftrace_graph_entry_stub(struct ftrace_graph_ent * trace)353 int ftrace_graph_entry_stub(struct ftrace_graph_ent *trace)
354 {
355 return 0;
356 }
357
358 /*
359 * Simply points to ftrace_stub, but with the proper protocol.
360 * Defined by the linker script in linux/vmlinux.lds.h
361 */
362 extern void ftrace_stub_graph(struct ftrace_graph_ret *);
363
364 /* The callbacks that hook a function */
365 trace_func_graph_ret_t ftrace_graph_return = ftrace_stub_graph;
366 trace_func_graph_ent_t ftrace_graph_entry = ftrace_graph_entry_stub;
367 static trace_func_graph_ent_t __ftrace_graph_entry = ftrace_graph_entry_stub;
368
369 /* Try to assign a return stack array on FTRACE_RETSTACK_ALLOC_SIZE tasks. */
alloc_retstack_tasklist(struct ftrace_ret_stack ** ret_stack_list)370 static int alloc_retstack_tasklist(struct ftrace_ret_stack **ret_stack_list)
371 {
372 int i;
373 int ret = 0;
374 int start = 0, end = FTRACE_RETSTACK_ALLOC_SIZE;
375 struct task_struct *g, *t;
376
377 for (i = 0; i < FTRACE_RETSTACK_ALLOC_SIZE; i++) {
378 ret_stack_list[i] =
379 kmalloc_array(FTRACE_RETFUNC_DEPTH,
380 sizeof(struct ftrace_ret_stack),
381 GFP_KERNEL);
382 if (!ret_stack_list[i]) {
383 start = 0;
384 end = i;
385 ret = -ENOMEM;
386 goto free;
387 }
388 }
389
390 rcu_read_lock();
391 for_each_process_thread(g, t) {
392 if (start == end) {
393 ret = -EAGAIN;
394 goto unlock;
395 }
396
397 if (t->ret_stack == NULL) {
398 atomic_set(&t->trace_overrun, 0);
399 t->curr_ret_stack = -1;
400 t->curr_ret_depth = -1;
401 /* Make sure the tasks see the -1 first: */
402 smp_wmb();
403 t->ret_stack = ret_stack_list[start++];
404 }
405 }
406
407 unlock:
408 rcu_read_unlock();
409 free:
410 for (i = start; i < end; i++)
411 kfree(ret_stack_list[i]);
412 return ret;
413 }
414
415 static void
ftrace_graph_probe_sched_switch(void * ignore,bool preempt,struct task_struct * prev,struct task_struct * next)416 ftrace_graph_probe_sched_switch(void *ignore, bool preempt,
417 struct task_struct *prev, struct task_struct *next)
418 {
419 unsigned long long timestamp;
420 int index;
421
422 /*
423 * Does the user want to count the time a function was asleep.
424 * If so, do not update the time stamps.
425 */
426 if (fgraph_sleep_time)
427 return;
428
429 timestamp = trace_clock_local();
430
431 prev->ftrace_timestamp = timestamp;
432
433 /* only process tasks that we timestamped */
434 if (!next->ftrace_timestamp)
435 return;
436
437 /*
438 * Update all the counters in next to make up for the
439 * time next was sleeping.
440 */
441 timestamp -= next->ftrace_timestamp;
442
443 for (index = next->curr_ret_stack; index >= 0; index--)
444 next->ret_stack[index].calltime += timestamp;
445 }
446
ftrace_graph_entry_test(struct ftrace_graph_ent * trace)447 static int ftrace_graph_entry_test(struct ftrace_graph_ent *trace)
448 {
449 if (!ftrace_ops_test(&global_ops, trace->func, NULL))
450 return 0;
451 return __ftrace_graph_entry(trace);
452 }
453
454 /*
455 * The function graph tracer should only trace the functions defined
456 * by set_ftrace_filter and set_ftrace_notrace. If another function
457 * tracer ops is registered, the graph tracer requires testing the
458 * function against the global ops, and not just trace any function
459 * that any ftrace_ops registered.
460 */
update_function_graph_func(void)461 void update_function_graph_func(void)
462 {
463 struct ftrace_ops *op;
464 bool do_test = false;
465
466 /*
467 * The graph and global ops share the same set of functions
468 * to test. If any other ops is on the list, then
469 * the graph tracing needs to test if its the function
470 * it should call.
471 */
472 do_for_each_ftrace_op(op, ftrace_ops_list) {
473 if (op != &global_ops && op != &graph_ops &&
474 op != &ftrace_list_end) {
475 do_test = true;
476 /* in double loop, break out with goto */
477 goto out;
478 }
479 } while_for_each_ftrace_op(op);
480 out:
481 if (do_test)
482 ftrace_graph_entry = ftrace_graph_entry_test;
483 else
484 ftrace_graph_entry = __ftrace_graph_entry;
485 }
486
487 static DEFINE_PER_CPU(struct ftrace_ret_stack *, idle_ret_stack);
488
489 static void
graph_init_task(struct task_struct * t,struct ftrace_ret_stack * ret_stack)490 graph_init_task(struct task_struct *t, struct ftrace_ret_stack *ret_stack)
491 {
492 atomic_set(&t->trace_overrun, 0);
493 t->ftrace_timestamp = 0;
494 /* make curr_ret_stack visible before we add the ret_stack */
495 smp_wmb();
496 t->ret_stack = ret_stack;
497 }
498
499 /*
500 * Allocate a return stack for the idle task. May be the first
501 * time through, or it may be done by CPU hotplug online.
502 */
ftrace_graph_init_idle_task(struct task_struct * t,int cpu)503 void ftrace_graph_init_idle_task(struct task_struct *t, int cpu)
504 {
505 t->curr_ret_stack = -1;
506 t->curr_ret_depth = -1;
507 /*
508 * The idle task has no parent, it either has its own
509 * stack or no stack at all.
510 */
511 if (t->ret_stack)
512 WARN_ON(t->ret_stack != per_cpu(idle_ret_stack, cpu));
513
514 if (ftrace_graph_active) {
515 struct ftrace_ret_stack *ret_stack;
516
517 ret_stack = per_cpu(idle_ret_stack, cpu);
518 if (!ret_stack) {
519 ret_stack =
520 kmalloc_array(FTRACE_RETFUNC_DEPTH,
521 sizeof(struct ftrace_ret_stack),
522 GFP_KERNEL);
523 if (!ret_stack)
524 return;
525 per_cpu(idle_ret_stack, cpu) = ret_stack;
526 }
527 graph_init_task(t, ret_stack);
528 }
529 }
530
531 /* Allocate a return stack for newly created task */
ftrace_graph_init_task(struct task_struct * t)532 void ftrace_graph_init_task(struct task_struct *t)
533 {
534 /* Make sure we do not use the parent ret_stack */
535 t->ret_stack = NULL;
536 t->curr_ret_stack = -1;
537 t->curr_ret_depth = -1;
538
539 if (ftrace_graph_active) {
540 struct ftrace_ret_stack *ret_stack;
541
542 ret_stack = kmalloc_array(FTRACE_RETFUNC_DEPTH,
543 sizeof(struct ftrace_ret_stack),
544 GFP_KERNEL);
545 if (!ret_stack)
546 return;
547 graph_init_task(t, ret_stack);
548 }
549 }
550
ftrace_graph_exit_task(struct task_struct * t)551 void ftrace_graph_exit_task(struct task_struct *t)
552 {
553 struct ftrace_ret_stack *ret_stack = t->ret_stack;
554
555 t->ret_stack = NULL;
556 /* NULL must become visible to IRQs before we free it: */
557 barrier();
558
559 kfree(ret_stack);
560 }
561
562 /* Allocate a return stack for each task */
start_graph_tracing(void)563 static int start_graph_tracing(void)
564 {
565 struct ftrace_ret_stack **ret_stack_list;
566 int ret, cpu;
567
568 ret_stack_list = kmalloc_array(FTRACE_RETSTACK_ALLOC_SIZE,
569 sizeof(struct ftrace_ret_stack *),
570 GFP_KERNEL);
571
572 if (!ret_stack_list)
573 return -ENOMEM;
574
575 /* The cpu_boot init_task->ret_stack will never be freed */
576 for_each_online_cpu(cpu) {
577 if (!idle_task(cpu)->ret_stack)
578 ftrace_graph_init_idle_task(idle_task(cpu), cpu);
579 }
580
581 do {
582 ret = alloc_retstack_tasklist(ret_stack_list);
583 } while (ret == -EAGAIN);
584
585 if (!ret) {
586 ret = register_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
587 if (ret)
588 pr_info("ftrace_graph: Couldn't activate tracepoint"
589 " probe to kernel_sched_switch\n");
590 }
591
592 kfree(ret_stack_list);
593 return ret;
594 }
595
register_ftrace_graph(struct fgraph_ops * gops)596 int register_ftrace_graph(struct fgraph_ops *gops)
597 {
598 int ret = 0;
599
600 mutex_lock(&ftrace_lock);
601
602 /* we currently allow only one tracer registered at a time */
603 if (ftrace_graph_active) {
604 ret = -EBUSY;
605 goto out;
606 }
607
608 register_pm_notifier(&ftrace_suspend_notifier);
609
610 ftrace_graph_active++;
611 ret = start_graph_tracing();
612 if (ret) {
613 ftrace_graph_active--;
614 goto out;
615 }
616
617 ftrace_graph_return = gops->retfunc;
618
619 /*
620 * Update the indirect function to the entryfunc, and the
621 * function that gets called to the entry_test first. Then
622 * call the update fgraph entry function to determine if
623 * the entryfunc should be called directly or not.
624 */
625 __ftrace_graph_entry = gops->entryfunc;
626 ftrace_graph_entry = ftrace_graph_entry_test;
627 update_function_graph_func();
628
629 ret = ftrace_startup(&graph_ops, FTRACE_START_FUNC_RET);
630 out:
631 mutex_unlock(&ftrace_lock);
632 return ret;
633 }
634
unregister_ftrace_graph(struct fgraph_ops * gops)635 void unregister_ftrace_graph(struct fgraph_ops *gops)
636 {
637 mutex_lock(&ftrace_lock);
638
639 if (unlikely(!ftrace_graph_active))
640 goto out;
641
642 ftrace_graph_active--;
643 ftrace_graph_return = ftrace_stub_graph;
644 ftrace_graph_entry = ftrace_graph_entry_stub;
645 __ftrace_graph_entry = ftrace_graph_entry_stub;
646 ftrace_shutdown(&graph_ops, FTRACE_STOP_FUNC_RET);
647 unregister_pm_notifier(&ftrace_suspend_notifier);
648 unregister_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
649
650 out:
651 mutex_unlock(&ftrace_lock);
652 }
653