projects/linux·blob v6.1 · refs

664 lines · 16.9 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/jump_label.h>
11#include <linux/suspend.h>
12#include <linux/ftrace.h>
13#include <linux/slab.h>
14
15#include <trace/events/sched.h>
16
17#include "ftrace_internal.h"
18
19#ifdef CONFIG_DYNAMIC_FTRACE
20#define ASSIGN_OPS_HASH(opsname, val) \
21 .func_hash = val, \
22 .local_hash.regex_lock = __MUTEX_INITIALIZER(opsname.local_hash.regex_lock),
23#else
24#define ASSIGN_OPS_HASH(opsname, val)
25#endif
26
27DEFINE_STATIC_KEY_FALSE(kill_ftrace_graph);
28int ftrace_graph_active;
29
30/* Both enabled by default (can be cleared by function_graph tracer flags */
31static bool fgraph_sleep_time = true;
32
33#ifdef CONFIG_DYNAMIC_FTRACE
34/*
35 * archs can override this function if they must do something
36 * to enable hook for graph tracer.
37 */
38int __weak ftrace_enable_ftrace_graph_caller(void)
39{
40 return 0;
41}
42
43/*
44 * archs can override this function if they must do something
45 * to disable hook for graph tracer.
46 */
47int __weak ftrace_disable_ftrace_graph_caller(void)
48{
49 return 0;
50}
51#endif
52
53/**
54 * ftrace_graph_stop - set to permanently disable function graph tracing
55 *
56 * In case of an error int function graph tracing, this is called
57 * to try to keep function graph tracing from causing any more harm.
58 * Usually this is pretty severe and this is called to try to at least
59 * get a warning out to the user.
60 */
61void ftrace_graph_stop(void)
62{
63 static_branch_enable(&kill_ftrace_graph);
64}
65
66/* Add a function return address to the trace stack on thread info.*/
67static int
68ftrace_push_return_trace(unsigned long ret, unsigned long func,
69 unsigned long frame_pointer, unsigned long *retp)
70{
71 unsigned long long calltime;
72 int index;
73
74 if (unlikely(ftrace_graph_is_dead()))
75 return -EBUSY;
76
77 if (!current->ret_stack)
78 return -EBUSY;
79
80 /*
81 * We must make sure the ret_stack is tested before we read
82 * anything else.
83 */
84 smp_rmb();
85
86 /* The return trace stack is full */
87 if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
88 atomic_inc(&current->trace_overrun);
89 return -EBUSY;
90 }
91
92 calltime = trace_clock_local();
93
94 index = ++current->curr_ret_stack;
95 barrier();
96 current->ret_stack[index].ret = ret;
97 current->ret_stack[index].func = func;
98 current->ret_stack[index].calltime = calltime;
99#ifdef HAVE_FUNCTION_GRAPH_FP_TEST
100 current->ret_stack[index].fp = frame_pointer;
101#endif
102#ifdef HAVE_FUNCTION_GRAPH_RET_ADDR_PTR
103 current->ret_stack[index].retp = retp;
104#endif
105 return 0;
106}
107
108/*
109 * Not all archs define MCOUNT_INSN_SIZE which is used to look for direct
110 * functions. But those archs currently don't support direct functions
111 * anyway, and ftrace_find_rec_direct() is just a stub for them.
112 * Define MCOUNT_INSN_SIZE to keep those archs compiling.
113 */
114#ifndef MCOUNT_INSN_SIZE
115/* Make sure this only works without direct calls */
116# ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
117# error MCOUNT_INSN_SIZE not defined with direct calls enabled
118# endif
119# define MCOUNT_INSN_SIZE 0
120#endif
121
122int function_graph_enter(unsigned long ret, unsigned long func,
123 unsigned long frame_pointer, unsigned long *retp)
124{
125 struct ftrace_graph_ent trace;
126
127#ifndef CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS
128 /*
129 * Skip graph tracing if the return location is served by direct trampoline,
130 * since call sequence and return addresses are unpredictable anyway.
131 * Ex: BPF trampoline may call original function and may skip frame
132 * depending on type of BPF programs attached.
133 */
134 if (ftrace_direct_func_count &&
135 ftrace_find_rec_direct(ret - MCOUNT_INSN_SIZE))
136 return -EBUSY;
137#endif
138 trace.func = func;
139 trace.depth = ++current->curr_ret_depth;
140
141 if (ftrace_push_return_trace(ret, func, frame_pointer, retp))
142 goto out;
143
144 /* Only trace if the calling function expects to */
145 if (!ftrace_graph_entry(&trace))
146 goto out_ret;
147
148 return 0;
149 out_ret:
150 current->curr_ret_stack--;
151 out:
152 current->curr_ret_depth--;
153 return -EBUSY;
154}
155
156/* Retrieve a function return address to the trace stack on thread info.*/
157static void
158ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
159 unsigned long frame_pointer)
160{
161 int index;
162
163 index = current->curr_ret_stack;
164
165 if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
166 ftrace_graph_stop();
167 WARN_ON(1);
168 /* Might as well panic, otherwise we have no where to go */
169 *ret = (unsigned long)panic;
170 return;
171 }
172
173#ifdef HAVE_FUNCTION_GRAPH_FP_TEST
174 /*
175 * The arch may choose to record the frame pointer used
176 * and check it here to make sure that it is what we expect it
177 * to be. If gcc does not set the place holder of the return
178 * address in the frame pointer, and does a copy instead, then
179 * the function graph trace will fail. This test detects this
180 * case.
181 *
182 * Currently, x86_32 with optimize for size (-Os) makes the latest
183 * gcc do the above.
184 *
185 * Note, -mfentry does not use frame pointers, and this test
186 * is not needed if CC_USING_FENTRY is set.
187 */
188 if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
189 ftrace_graph_stop();
190 WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
191 " from func %ps return to %lx\n",
192 current->ret_stack[index].fp,
193 frame_pointer,
194 (void *)current->ret_stack[index].func,
195 current->ret_stack[index].ret);
196 *ret = (unsigned long)panic;
197 return;
198 }
199#endif
200
201 *ret = current->ret_stack[index].ret;
202 trace->func = current->ret_stack[index].func;
203 trace->calltime = current->ret_stack[index].calltime;
204 trace->overrun = atomic_read(&current->trace_overrun);
205 trace->depth = current->curr_ret_depth--;
206 /*
207 * We still want to trace interrupts coming in if
208 * max_depth is set to 1. Make sure the decrement is
209 * seen before ftrace_graph_return.
210 */
211 barrier();
212}
213
214/*
215 * Hibernation protection.
216 * The state of the current task is too much unstable during
217 * suspend/restore to disk. We want to protect against that.
218 */
219static int
220ftrace_suspend_notifier_call(struct notifier_block *bl, unsigned long state,
221 void *unused)
222{
223 switch (state) {
224 case PM_HIBERNATION_PREPARE:
225 pause_graph_tracing();
226 break;
227
228 case PM_POST_HIBERNATION:
229 unpause_graph_tracing();
230 break;
231 }
232 return NOTIFY_DONE;
233}
234
235static struct notifier_block ftrace_suspend_notifier = {
236 .notifier_call = ftrace_suspend_notifier_call,
237};
238
239/*
240 * Send the trace to the ring-buffer.
241 * @return the original return address.
242 */
243unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
244{
245 struct ftrace_graph_ret trace;
246 unsigned long ret;
247
248 ftrace_pop_return_trace(&trace, &ret, frame_pointer);
249 trace.rettime = trace_clock_local();
250 ftrace_graph_return(&trace);
251 /*
252 * The ftrace_graph_return() may still access the current
253 * ret_stack structure, we need to make sure the update of
254 * curr_ret_stack is after that.
255 */
256 barrier();
257 current->curr_ret_stack--;
258
259 if (unlikely(!ret)) {
260 ftrace_graph_stop();
261 WARN_ON(1);
262 /* Might as well panic. What else to do? */
263 ret = (unsigned long)panic;
264 }
265
266 return ret;
267}
268
269/**
270 * ftrace_graph_get_ret_stack - return the entry of the shadow stack
271 * @task: The task to read the shadow stack from
272 * @idx: Index down the shadow stack
273 *
274 * Return the ret_struct on the shadow stack of the @task at the
275 * call graph at @idx starting with zero. If @idx is zero, it
276 * will return the last saved ret_stack entry. If it is greater than
277 * zero, it will return the corresponding ret_stack for the depth
278 * of saved return addresses.
279 */
280struct ftrace_ret_stack *
281ftrace_graph_get_ret_stack(struct task_struct *task, int idx)
282{
283 idx = task->curr_ret_stack - idx;
284
285 if (idx >= 0 && idx <= task->curr_ret_stack)
286 return &task->ret_stack[idx];
287
288 return NULL;
289}
290
291/**
292 * ftrace_graph_ret_addr - convert a potentially modified stack return address
293 * to its original value
294 *
295 * This function can be called by stack unwinding code to convert a found stack
296 * return address ('ret') to its original value, in case the function graph
297 * tracer has modified it to be 'return_to_handler'. If the address hasn't
298 * been modified, the unchanged value of 'ret' is returned.
299 *
300 * 'idx' is a state variable which should be initialized by the caller to zero
301 * before the first call.
302 *
303 * 'retp' is a pointer to the return address on the stack. It's ignored if
304 * the arch doesn't have HAVE_FUNCTION_GRAPH_RET_ADDR_PTR defined.
305 */
306#ifdef HAVE_FUNCTION_GRAPH_RET_ADDR_PTR
307unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
308 unsigned long ret, unsigned long *retp)
309{
310 int index = task->curr_ret_stack;
311 int i;
312
313 if (ret != (unsigned long)dereference_kernel_function_descriptor(return_to_handler))
314 return ret;
315
316 if (index < 0)
317 return ret;
318
319 for (i = 0; i <= index; i++)
320 if (task->ret_stack[i].retp == retp)
321 return task->ret_stack[i].ret;
322
323 return ret;
324}
325#else /* !HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
326unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
327 unsigned long ret, unsigned long *retp)
328{
329 int task_idx;
330
331 if (ret != (unsigned long)dereference_kernel_function_descriptor(return_to_handler))
332 return ret;
333
334 task_idx = task->curr_ret_stack;
335
336 if (!task->ret_stack || task_idx < *idx)
337 return ret;
338
339 task_idx -= *idx;
340 (*idx)++;
341
342 return task->ret_stack[task_idx].ret;
343}
344#endif /* HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
345
346static struct ftrace_ops graph_ops = {
347 .func = ftrace_graph_func,
348 .flags = FTRACE_OPS_FL_INITIALIZED |
349 FTRACE_OPS_FL_PID |
350 FTRACE_OPS_GRAPH_STUB,
351#ifdef FTRACE_GRAPH_TRAMP_ADDR
352 .trampoline = FTRACE_GRAPH_TRAMP_ADDR,
353 /* trampoline_size is only needed for dynamically allocated tramps */
354#endif
355 ASSIGN_OPS_HASH(graph_ops, &global_ops.local_hash)
356};
357
358void ftrace_graph_sleep_time_control(bool enable)
359{
360 fgraph_sleep_time = enable;
361}
362
363int ftrace_graph_entry_stub(struct ftrace_graph_ent *trace)
364{
365 return 0;
366}
367
368/*
369 * Simply points to ftrace_stub, but with the proper protocol.
370 * Defined by the linker script in linux/vmlinux.lds.h
371 */
372extern void ftrace_stub_graph(struct ftrace_graph_ret *);
373
374/* The callbacks that hook a function */
375trace_func_graph_ret_t ftrace_graph_return = ftrace_stub_graph;
376trace_func_graph_ent_t ftrace_graph_entry = ftrace_graph_entry_stub;
377static trace_func_graph_ent_t __ftrace_graph_entry = ftrace_graph_entry_stub;
378
379/* Try to assign a return stack array on FTRACE_RETSTACK_ALLOC_SIZE tasks. */
380static int alloc_retstack_tasklist(struct ftrace_ret_stack **ret_stack_list)
381{
382 int i;
383 int ret = 0;
384 int start = 0, end = FTRACE_RETSTACK_ALLOC_SIZE;
385 struct task_struct *g, *t;
386
387 for (i = 0; i < FTRACE_RETSTACK_ALLOC_SIZE; i++) {
388 ret_stack_list[i] =
389 kmalloc_array(FTRACE_RETFUNC_DEPTH,
390 sizeof(struct ftrace_ret_stack),
391 GFP_KERNEL);
392 if (!ret_stack_list[i]) {
393 start = 0;
394 end = i;
395 ret = -ENOMEM;
396 goto free;
397 }
398 }
399
400 rcu_read_lock();
401 for_each_process_thread(g, t) {
402 if (start == end) {
403 ret = -EAGAIN;
404 goto unlock;
405 }
406
407 if (t->ret_stack == NULL) {
408 atomic_set(&t->trace_overrun, 0);
409 t->curr_ret_stack = -1;
410 t->curr_ret_depth = -1;
411 /* Make sure the tasks see the -1 first: */
412 smp_wmb();
413 t->ret_stack = ret_stack_list[start++];
414 }
415 }
416
417unlock:
418 rcu_read_unlock();
419free:
420 for (i = start; i < end; i++)
421 kfree(ret_stack_list[i]);
422 return ret;
423}
424
425static void
426ftrace_graph_probe_sched_switch(void *ignore, bool preempt,
427 struct task_struct *prev,
428 struct task_struct *next,
429 unsigned int prev_state)
430{
431 unsigned long long timestamp;
432 int index;
433
434 /*
435 * Does the user want to count the time a function was asleep.
436 * If so, do not update the time stamps.
437 */
438 if (fgraph_sleep_time)
439 return;
440
441 timestamp = trace_clock_local();
442
443 prev->ftrace_timestamp = timestamp;
444
445 /* only process tasks that we timestamped */
446 if (!next->ftrace_timestamp)
447 return;
448
449 /*
450 * Update all the counters in next to make up for the
451 * time next was sleeping.
452 */
453 timestamp -= next->ftrace_timestamp;
454
455 for (index = next->curr_ret_stack; index >= 0; index--)
456 next->ret_stack[index].calltime += timestamp;
457}
458
459static int ftrace_graph_entry_test(struct ftrace_graph_ent *trace)
460{
461 if (!ftrace_ops_test(&global_ops, trace->func, NULL))
462 return 0;
463 return __ftrace_graph_entry(trace);
464}
465
466/*
467 * The function graph tracer should only trace the functions defined
468 * by set_ftrace_filter and set_ftrace_notrace. If another function
469 * tracer ops is registered, the graph tracer requires testing the
470 * function against the global ops, and not just trace any function
471 * that any ftrace_ops registered.
472 */
473void update_function_graph_func(void)
474{
475 struct ftrace_ops *op;
476 bool do_test = false;
477
478 /*
479 * The graph and global ops share the same set of functions
480 * to test. If any other ops is on the list, then
481 * the graph tracing needs to test if its the function
482 * it should call.
483 */
484 do_for_each_ftrace_op(op, ftrace_ops_list) {
485 if (op != &global_ops && op != &graph_ops &&
486 op != &ftrace_list_end) {
487 do_test = true;
488 /* in double loop, break out with goto */
489 goto out;
490 }
491 } while_for_each_ftrace_op(op);
492 out:
493 if (do_test)
494 ftrace_graph_entry = ftrace_graph_entry_test;
495 else
496 ftrace_graph_entry = __ftrace_graph_entry;
497}
498
499static DEFINE_PER_CPU(struct ftrace_ret_stack *, idle_ret_stack);
500
501static void
502graph_init_task(struct task_struct *t, struct ftrace_ret_stack *ret_stack)
503{
504 atomic_set(&t->trace_overrun, 0);
505 t->ftrace_timestamp = 0;
506 /* make curr_ret_stack visible before we add the ret_stack */
507 smp_wmb();
508 t->ret_stack = ret_stack;
509}
510
511/*
512 * Allocate a return stack for the idle task. May be the first
513 * time through, or it may be done by CPU hotplug online.
514 */
515void ftrace_graph_init_idle_task(struct task_struct *t, int cpu)
516{
517 t->curr_ret_stack = -1;
518 t->curr_ret_depth = -1;
519 /*
520 * The idle task has no parent, it either has its own
521 * stack or no stack at all.
522 */
523 if (t->ret_stack)
524 WARN_ON(t->ret_stack != per_cpu(idle_ret_stack, cpu));
525
526 if (ftrace_graph_active) {
527 struct ftrace_ret_stack *ret_stack;
528
529 ret_stack = per_cpu(idle_ret_stack, cpu);
530 if (!ret_stack) {
531 ret_stack =
532 kmalloc_array(FTRACE_RETFUNC_DEPTH,
533 sizeof(struct ftrace_ret_stack),
534 GFP_KERNEL);
535 if (!ret_stack)
536 return;
537 per_cpu(idle_ret_stack, cpu) = ret_stack;
538 }
539 graph_init_task(t, ret_stack);
540 }
541}
542
543/* Allocate a return stack for newly created task */
544void ftrace_graph_init_task(struct task_struct *t)
545{
546 /* Make sure we do not use the parent ret_stack */
547 t->ret_stack = NULL;
548 t->curr_ret_stack = -1;
549 t->curr_ret_depth = -1;
550
551 if (ftrace_graph_active) {
552 struct ftrace_ret_stack *ret_stack;
553
554 ret_stack = kmalloc_array(FTRACE_RETFUNC_DEPTH,
555 sizeof(struct ftrace_ret_stack),
556 GFP_KERNEL);
557 if (!ret_stack)
558 return;
559 graph_init_task(t, ret_stack);
560 }
561}
562
563void ftrace_graph_exit_task(struct task_struct *t)
564{
565 struct ftrace_ret_stack *ret_stack = t->ret_stack;
566
567 t->ret_stack = NULL;
568 /* NULL must become visible to IRQs before we free it: */
569 barrier();
570
571 kfree(ret_stack);
572}
573
574/* Allocate a return stack for each task */
575static int start_graph_tracing(void)
576{
577 struct ftrace_ret_stack **ret_stack_list;
578 int ret, cpu;
579
580 ret_stack_list = kmalloc_array(FTRACE_RETSTACK_ALLOC_SIZE,
581 sizeof(struct ftrace_ret_stack *),
582 GFP_KERNEL);
583
584 if (!ret_stack_list)
585 return -ENOMEM;
586
587 /* The cpu_boot init_task->ret_stack will never be freed */
588 for_each_online_cpu(cpu) {
589 if (!idle_task(cpu)->ret_stack)
590 ftrace_graph_init_idle_task(idle_task(cpu), cpu);
591 }
592
593 do {
594 ret = alloc_retstack_tasklist(ret_stack_list);
595 } while (ret == -EAGAIN);
596
597 if (!ret) {
598 ret = register_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
599 if (ret)
600 pr_info("ftrace_graph: Couldn't activate tracepoint"
601 " probe to kernel_sched_switch\n");
602 }
603
604 kfree(ret_stack_list);
605 return ret;
606}
607
608int register_ftrace_graph(struct fgraph_ops *gops)
609{
610 int ret = 0;
611
612 mutex_lock(&ftrace_lock);
613
614 /* we currently allow only one tracer registered at a time */
615 if (ftrace_graph_active) {
616 ret = -EBUSY;
617 goto out;
618 }
619
620 register_pm_notifier(&ftrace_suspend_notifier);
621
622 ftrace_graph_active++;
623 ret = start_graph_tracing();
624 if (ret) {
625 ftrace_graph_active--;
626 goto out;
627 }
628
629 ftrace_graph_return = gops->retfunc;
630
631 /*
632 * Update the indirect function to the entryfunc, and the
633 * function that gets called to the entry_test first. Then
634 * call the update fgraph entry function to determine if
635 * the entryfunc should be called directly or not.
636 */
637 __ftrace_graph_entry = gops->entryfunc;
638 ftrace_graph_entry = ftrace_graph_entry_test;
639 update_function_graph_func();
640
641 ret = ftrace_startup(&graph_ops, FTRACE_START_FUNC_RET);
642out:
643 mutex_unlock(&ftrace_lock);
644 return ret;
645}
646
647void unregister_ftrace_graph(struct fgraph_ops *gops)
648{
649 mutex_lock(&ftrace_lock);
650
651 if (unlikely(!ftrace_graph_active))
652 goto out;
653
654 ftrace_graph_active--;
655 ftrace_graph_return = ftrace_stub_graph;
656 ftrace_graph_entry = ftrace_graph_entry_stub;
657 __ftrace_graph_entry = ftrace_graph_entry_stub;
658 ftrace_shutdown(&graph_ops, FTRACE_STOP_FUNC_RET);
659 unregister_pm_notifier(&ftrace_suspend_notifier);
660 unregister_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
661
662 out:
663 mutex_unlock(&ftrace_lock);
664}