projects/linux·blob v5.13 · refs

651 lines · 16.8 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Infrastructure to took into function calls and returns.
4 * Copyright (c) 2008-2009 Frederic Weisbecker <[email protected]>
5 * Mostly borrowed from function tracer which
6 * is Copyright (c) Steven Rostedt <[email protected]>
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
26static bool kill_ftrace_graph;
27int ftrace_graph_active;
28
29/* Both enabled by default (can be cleared by function_graph tracer flags */
30static 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 */
39bool ftrace_graph_is_dead(void)
40{
41 return kill_ftrace_graph;
42}
43
44/**
45 * ftrace_graph_stop - set to permanently disable function graph tracing
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 */
52void 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.*/
58static int
59ftrace_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(&current->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
113int 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 are unpredictable anyway.
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.*/
146static void
147ftrace_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(&current->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 */
208static int
209ftrace_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
224static 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 */
232unsigned 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 */
269struct ftrace_ret_stack *
270ftrace_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
296unsigned 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 */
315unsigned 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
335static struct ftrace_ops graph_ops = {
336 .func = ftrace_stub,
337 .flags = FTRACE_OPS_FL_INITIALIZED |
338 FTRACE_OPS_FL_PID |
339 FTRACE_OPS_FL_STUB,
340#ifdef FTRACE_GRAPH_TRAMP_ADDR
341 .trampoline = FTRACE_GRAPH_TRAMP_ADDR,
342 /* trampoline_size is only needed for dynamically allocated tramps */
343#endif
344 ASSIGN_OPS_HASH(graph_ops, &global_ops.local_hash)
345};
346
347void ftrace_graph_sleep_time_control(bool enable)
348{
349 fgraph_sleep_time = enable;
350}
351
352int ftrace_graph_entry_stub(struct ftrace_graph_ent *trace)
353{
354 return 0;
355}
356
357/*
358 * Simply points to ftrace_stub, but with the proper protocol.
359 * Defined by the linker script in linux/vmlinux.lds.h
360 */
361extern void ftrace_stub_graph(struct ftrace_graph_ret *);
362
363/* The callbacks that hook a function */
364trace_func_graph_ret_t ftrace_graph_return = ftrace_stub_graph;
365trace_func_graph_ent_t ftrace_graph_entry = ftrace_graph_entry_stub;
366static trace_func_graph_ent_t __ftrace_graph_entry = ftrace_graph_entry_stub;
367
368/* Try to assign a return stack array on FTRACE_RETSTACK_ALLOC_SIZE tasks. */
369static int alloc_retstack_tasklist(struct ftrace_ret_stack **ret_stack_list)
370{
371 int i;
372 int ret = 0;
373 int start = 0, end = FTRACE_RETSTACK_ALLOC_SIZE;
374 struct task_struct *g, *t;
375
376 for (i = 0; i < FTRACE_RETSTACK_ALLOC_SIZE; i++) {
377 ret_stack_list[i] =
378 kmalloc_array(FTRACE_RETFUNC_DEPTH,
379 sizeof(struct ftrace_ret_stack),
380 GFP_KERNEL);
381 if (!ret_stack_list[i]) {
382 start = 0;
383 end = i;
384 ret = -ENOMEM;
385 goto free;
386 }
387 }
388
389 rcu_read_lock();
390 for_each_process_thread(g, t) {
391 if (start == end) {
392 ret = -EAGAIN;
393 goto unlock;
394 }
395
396 if (t->ret_stack == NULL) {
397 atomic_set(&t->trace_overrun, 0);
398 t->curr_ret_stack = -1;
399 t->curr_ret_depth = -1;
400 /* Make sure the tasks see the -1 first: */
401 smp_wmb();
402 t->ret_stack = ret_stack_list[start++];
403 }
404 }
405
406unlock:
407 rcu_read_unlock();
408free:
409 for (i = start; i < end; i++)
410 kfree(ret_stack_list[i]);
411 return ret;
412}
413
414static void
415ftrace_graph_probe_sched_switch(void *ignore, bool preempt,
416 struct task_struct *prev, struct task_struct *next)
417{
418 unsigned long long timestamp;
419 int index;
420
421 /*
422 * Does the user want to count the time a function was asleep.
423 * If so, do not update the time stamps.
424 */
425 if (fgraph_sleep_time)
426 return;
427
428 timestamp = trace_clock_local();
429
430 prev->ftrace_timestamp = timestamp;
431
432 /* only process tasks that we timestamped */
433 if (!next->ftrace_timestamp)
434 return;
435
436 /*
437 * Update all the counters in next to make up for the
438 * time next was sleeping.
439 */
440 timestamp -= next->ftrace_timestamp;
441
442 for (index = next->curr_ret_stack; index >= 0; index--)
443 next->ret_stack[index].calltime += timestamp;
444}
445
446static int ftrace_graph_entry_test(struct ftrace_graph_ent *trace)
447{
448 if (!ftrace_ops_test(&global_ops, trace->func, NULL))
449 return 0;
450 return __ftrace_graph_entry(trace);
451}
452
453/*
454 * The function graph tracer should only trace the functions defined
455 * by set_ftrace_filter and set_ftrace_notrace. If another function
456 * tracer ops is registered, the graph tracer requires testing the
457 * function against the global ops, and not just trace any function
458 * that any ftrace_ops registered.
459 */
460void update_function_graph_func(void)
461{
462 struct ftrace_ops *op;
463 bool do_test = false;
464
465 /*
466 * The graph and global ops share the same set of functions
467 * to test. If any other ops is on the list, then
468 * the graph tracing needs to test if its the function
469 * it should call.
470 */
471 do_for_each_ftrace_op(op, ftrace_ops_list) {
472 if (op != &global_ops && op != &graph_ops &&
473 op != &ftrace_list_end) {
474 do_test = true;
475 /* in double loop, break out with goto */
476 goto out;
477 }
478 } while_for_each_ftrace_op(op);
479 out:
480 if (do_test)
481 ftrace_graph_entry = ftrace_graph_entry_test;
482 else
483 ftrace_graph_entry = __ftrace_graph_entry;
484}
485
486static DEFINE_PER_CPU(struct ftrace_ret_stack *, idle_ret_stack);
487
488static void
489graph_init_task(struct task_struct *t, struct ftrace_ret_stack *ret_stack)
490{
491 atomic_set(&t->trace_overrun, 0);
492 t->ftrace_timestamp = 0;
493 /* make curr_ret_stack visible before we add the ret_stack */
494 smp_wmb();
495 t->ret_stack = ret_stack;
496}
497
498/*
499 * Allocate a return stack for the idle task. May be the first
500 * time through, or it may be done by CPU hotplug online.
501 */
502void ftrace_graph_init_idle_task(struct task_struct *t, int cpu)
503{
504 t->curr_ret_stack = -1;
505 t->curr_ret_depth = -1;
506 /*
507 * The idle task has no parent, it either has its own
508 * stack or no stack at all.
509 */
510 if (t->ret_stack)
511 WARN_ON(t->ret_stack != per_cpu(idle_ret_stack, cpu));
512
513 if (ftrace_graph_active) {
514 struct ftrace_ret_stack *ret_stack;
515
516 ret_stack = per_cpu(idle_ret_stack, cpu);
517 if (!ret_stack) {
518 ret_stack =
519 kmalloc_array(FTRACE_RETFUNC_DEPTH,
520 sizeof(struct ftrace_ret_stack),
521 GFP_KERNEL);
522 if (!ret_stack)
523 return;
524 per_cpu(idle_ret_stack, cpu) = ret_stack;
525 }
526 graph_init_task(t, ret_stack);
527 }
528}
529
530/* Allocate a return stack for newly created task */
531void ftrace_graph_init_task(struct task_struct *t)
532{
533 /* Make sure we do not use the parent ret_stack */
534 t->ret_stack = NULL;
535 t->curr_ret_stack = -1;
536 t->curr_ret_depth = -1;
537
538 if (ftrace_graph_active) {
539 struct ftrace_ret_stack *ret_stack;
540
541 ret_stack = kmalloc_array(FTRACE_RETFUNC_DEPTH,
542 sizeof(struct ftrace_ret_stack),
543 GFP_KERNEL);
544 if (!ret_stack)
545 return;
546 graph_init_task(t, ret_stack);
547 }
548}
549
550void ftrace_graph_exit_task(struct task_struct *t)
551{
552 struct ftrace_ret_stack *ret_stack = t->ret_stack;
553
554 t->ret_stack = NULL;
555 /* NULL must become visible to IRQs before we free it: */
556 barrier();
557
558 kfree(ret_stack);
559}
560
561/* Allocate a return stack for each task */
562static int start_graph_tracing(void)
563{
564 struct ftrace_ret_stack **ret_stack_list;
565 int ret, cpu;
566
567 ret_stack_list = kmalloc_array(FTRACE_RETSTACK_ALLOC_SIZE,
568 sizeof(struct ftrace_ret_stack *),
569 GFP_KERNEL);
570
571 if (!ret_stack_list)
572 return -ENOMEM;
573
574 /* The cpu_boot init_task->ret_stack will never be freed */
575 for_each_online_cpu(cpu) {
576 if (!idle_task(cpu)->ret_stack)
577 ftrace_graph_init_idle_task(idle_task(cpu), cpu);
578 }
579
580 do {
581 ret = alloc_retstack_tasklist(ret_stack_list);
582 } while (ret == -EAGAIN);
583
584 if (!ret) {
585 ret = register_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
586 if (ret)
587 pr_info("ftrace_graph: Couldn't activate tracepoint"
588 " probe to kernel_sched_switch\n");
589 }
590
591 kfree(ret_stack_list);
592 return ret;
593}
594
595int register_ftrace_graph(struct fgraph_ops *gops)
596{
597 int ret = 0;
598
599 mutex_lock(&ftrace_lock);
600
601 /* we currently allow only one tracer registered at a time */
602 if (ftrace_graph_active) {
603 ret = -EBUSY;
604 goto out;
605 }
606
607 register_pm_notifier(&ftrace_suspend_notifier);
608
609 ftrace_graph_active++;
610 ret = start_graph_tracing();
611 if (ret) {
612 ftrace_graph_active--;
613 goto out;
614 }
615
616 ftrace_graph_return = gops->retfunc;
617
618 /*
619 * Update the indirect function to the entryfunc, and the
620 * function that gets called to the entry_test first. Then
621 * call the update fgraph entry function to determine if
622 * the entryfunc should be called directly or not.
623 */
624 __ftrace_graph_entry = gops->entryfunc;
625 ftrace_graph_entry = ftrace_graph_entry_test;
626 update_function_graph_func();
627
628 ret = ftrace_startup(&graph_ops, FTRACE_START_FUNC_RET);
629out:
630 mutex_unlock(&ftrace_lock);
631 return ret;
632}
633
634void unregister_ftrace_graph(struct fgraph_ops *gops)
635{
636 mutex_lock(&ftrace_lock);
637
638 if (unlikely(!ftrace_graph_active))
639 goto out;
640
641 ftrace_graph_active--;
642 ftrace_graph_return = ftrace_stub_graph;
643 ftrace_graph_entry = ftrace_graph_entry_stub;
644 __ftrace_graph_entry = ftrace_graph_entry_stub;
645 ftrace_shutdown(&graph_ops, FTRACE_STOP_FUNC_RET);
646 unregister_pm_notifier(&ftrace_suspend_notifier);
647 unregister_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
648
649 out:
650 mutex_unlock(&ftrace_lock);
651}