1 // SPDX-License-Identifier: GPL-2.0
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>
8 * Highly modified by Steven Rostedt (VMware).
10 #include <linux/jump_label.h>
11 #include <linux/suspend.h>
12 #include <linux/ftrace.h>
13 #include <linux/slab.h>
15 #include <trace/events/sched.h>
17 #include "ftrace_internal.h"
19 #ifdef CONFIG_DYNAMIC_FTRACE
20 #define ASSIGN_OPS_HASH(opsname, val) \
22 .local_hash.regex_lock = __MUTEX_INITIALIZER(opsname.local_hash.regex_lock),
24 #define ASSIGN_OPS_HASH(opsname, val)
27 DEFINE_STATIC_KEY_FALSE(kill_ftrace_graph);
28 int ftrace_graph_active;
30 /* Both enabled by default (can be cleared by function_graph tracer flags */
31 static bool fgraph_sleep_time = true;
33 #ifdef CONFIG_DYNAMIC_FTRACE
35 * archs can override this function if they must do something
36 * to enable hook for graph tracer.
38 int __weak ftrace_enable_ftrace_graph_caller(void)
44 * archs can override this function if they must do something
45 * to disable hook for graph tracer.
47 int __weak ftrace_disable_ftrace_graph_caller(void)
54 * ftrace_graph_stop - set to permanently disable function graph tracing
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.
61 void ftrace_graph_stop(void)
63 static_branch_enable(&kill_ftrace_graph);
66 /* Add a function return address to the trace stack on thread info.*/
68 ftrace_push_return_trace(unsigned long ret, unsigned long func,
69 unsigned long frame_pointer, unsigned long *retp)
71 unsigned long long calltime;
74 if (unlikely(ftrace_graph_is_dead()))
77 if (!current->ret_stack)
81 * We must make sure the ret_stack is tested before we read
86 /* The return trace stack is full */
87 if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
88 atomic_inc(¤t->trace_overrun);
92 calltime = trace_clock_local();
94 index = ++current->curr_ret_stack;
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;
102 #ifdef HAVE_FUNCTION_GRAPH_RET_ADDR_PTR
103 current->ret_stack[index].retp = retp;
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.
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
119 # define MCOUNT_INSN_SIZE 0
122 int function_graph_enter(unsigned long ret, unsigned long func,
123 unsigned long frame_pointer, unsigned long *retp)
125 struct ftrace_graph_ent trace;
127 #ifndef CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS
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.
134 if (ftrace_direct_func_count &&
135 ftrace_find_rec_direct(ret - MCOUNT_INSN_SIZE))
139 trace.depth = ++current->curr_ret_depth;
141 if (ftrace_push_return_trace(ret, func, frame_pointer, retp))
144 /* Only trace if the calling function expects to */
145 if (!ftrace_graph_entry(&trace))
150 current->curr_ret_stack--;
152 current->curr_ret_depth--;
156 /* Retrieve a function return address to the trace stack on thread info.*/
158 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
159 unsigned long frame_pointer)
163 index = current->curr_ret_stack;
165 if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
168 /* Might as well panic, otherwise we have no where to go */
169 *ret = (unsigned long)panic;
173 #ifdef HAVE_FUNCTION_GRAPH_FP_TEST
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
182 * Currently, x86_32 with optimize for size (-Os) makes the latest
185 * Note, -mfentry does not use frame pointers, and this test
186 * is not needed if CC_USING_FENTRY is set.
188 if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
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,
194 (void *)current->ret_stack[index].func,
195 current->ret_stack[index].ret);
196 *ret = (unsigned long)panic;
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(¤t->trace_overrun);
205 trace->depth = current->curr_ret_depth--;
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.
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.
220 ftrace_suspend_notifier_call(struct notifier_block *bl, unsigned long state,
224 case PM_HIBERNATION_PREPARE:
225 pause_graph_tracing();
228 case PM_POST_HIBERNATION:
229 unpause_graph_tracing();
235 static struct notifier_block ftrace_suspend_notifier = {
236 .notifier_call = ftrace_suspend_notifier_call,
239 /* fgraph_ret_regs is not defined without CONFIG_FUNCTION_GRAPH_RETVAL */
240 struct fgraph_ret_regs;
243 * Send the trace to the ring-buffer.
244 * @return the original return address.
246 static unsigned long __ftrace_return_to_handler(struct fgraph_ret_regs *ret_regs,
247 unsigned long frame_pointer)
249 struct ftrace_graph_ret trace;
252 ftrace_pop_return_trace(&trace, &ret, frame_pointer);
253 #ifdef CONFIG_FUNCTION_GRAPH_RETVAL
254 trace.retval = fgraph_ret_regs_return_value(ret_regs);
256 trace.rettime = trace_clock_local();
257 ftrace_graph_return(&trace);
259 * The ftrace_graph_return() may still access the current
260 * ret_stack structure, we need to make sure the update of
261 * curr_ret_stack is after that.
264 current->curr_ret_stack--;
266 if (unlikely(!ret)) {
269 /* Might as well panic. What else to do? */
270 ret = (unsigned long)panic;
277 * After all architecures have selected HAVE_FUNCTION_GRAPH_RETVAL, we can
278 * leave only ftrace_return_to_handler(ret_regs).
280 #ifdef CONFIG_HAVE_FUNCTION_GRAPH_RETVAL
281 unsigned long ftrace_return_to_handler(struct fgraph_ret_regs *ret_regs)
283 return __ftrace_return_to_handler(ret_regs,
284 fgraph_ret_regs_frame_pointer(ret_regs));
287 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
289 return __ftrace_return_to_handler(NULL, frame_pointer);
294 * ftrace_graph_get_ret_stack - return the entry of the shadow stack
295 * @task: The task to read the shadow stack from
296 * @idx: Index down the shadow stack
298 * Return the ret_struct on the shadow stack of the @task at the
299 * call graph at @idx starting with zero. If @idx is zero, it
300 * will return the last saved ret_stack entry. If it is greater than
301 * zero, it will return the corresponding ret_stack for the depth
302 * of saved return addresses.
304 struct ftrace_ret_stack *
305 ftrace_graph_get_ret_stack(struct task_struct *task, int idx)
307 idx = task->curr_ret_stack - idx;
309 if (idx >= 0 && idx <= task->curr_ret_stack)
310 return &task->ret_stack[idx];
316 * ftrace_graph_ret_addr - convert a potentially modified stack return address
317 * to its original value
319 * This function can be called by stack unwinding code to convert a found stack
320 * return address ('ret') to its original value, in case the function graph
321 * tracer has modified it to be 'return_to_handler'. If the address hasn't
322 * been modified, the unchanged value of 'ret' is returned.
324 * 'idx' is a state variable which should be initialized by the caller to zero
325 * before the first call.
327 * 'retp' is a pointer to the return address on the stack. It's ignored if
328 * the arch doesn't have HAVE_FUNCTION_GRAPH_RET_ADDR_PTR defined.
330 #ifdef HAVE_FUNCTION_GRAPH_RET_ADDR_PTR
331 unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
332 unsigned long ret, unsigned long *retp)
334 int index = task->curr_ret_stack;
337 if (ret != (unsigned long)dereference_kernel_function_descriptor(return_to_handler))
343 for (i = 0; i <= index; i++)
344 if (task->ret_stack[i].retp == retp)
345 return task->ret_stack[i].ret;
349 #else /* !HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
350 unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
351 unsigned long ret, unsigned long *retp)
355 if (ret != (unsigned long)dereference_kernel_function_descriptor(return_to_handler))
358 task_idx = task->curr_ret_stack;
360 if (!task->ret_stack || task_idx < *idx)
366 return task->ret_stack[task_idx].ret;
368 #endif /* HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
370 static struct ftrace_ops graph_ops = {
371 .func = ftrace_graph_func,
372 .flags = FTRACE_OPS_FL_INITIALIZED |
374 FTRACE_OPS_GRAPH_STUB,
375 #ifdef FTRACE_GRAPH_TRAMP_ADDR
376 .trampoline = FTRACE_GRAPH_TRAMP_ADDR,
377 /* trampoline_size is only needed for dynamically allocated tramps */
379 ASSIGN_OPS_HASH(graph_ops, &global_ops.local_hash)
382 void ftrace_graph_sleep_time_control(bool enable)
384 fgraph_sleep_time = enable;
387 int ftrace_graph_entry_stub(struct ftrace_graph_ent *trace)
393 * Simply points to ftrace_stub, but with the proper protocol.
394 * Defined by the linker script in linux/vmlinux.lds.h
396 extern void ftrace_stub_graph(struct ftrace_graph_ret *);
398 /* The callbacks that hook a function */
399 trace_func_graph_ret_t ftrace_graph_return = ftrace_stub_graph;
400 trace_func_graph_ent_t ftrace_graph_entry = ftrace_graph_entry_stub;
401 static trace_func_graph_ent_t __ftrace_graph_entry = ftrace_graph_entry_stub;
403 /* Try to assign a return stack array on FTRACE_RETSTACK_ALLOC_SIZE tasks. */
404 static int alloc_retstack_tasklist(struct ftrace_ret_stack **ret_stack_list)
408 int start = 0, end = FTRACE_RETSTACK_ALLOC_SIZE;
409 struct task_struct *g, *t;
411 for (i = 0; i < FTRACE_RETSTACK_ALLOC_SIZE; i++) {
413 kmalloc_array(FTRACE_RETFUNC_DEPTH,
414 sizeof(struct ftrace_ret_stack),
416 if (!ret_stack_list[i]) {
425 for_each_process_thread(g, t) {
431 if (t->ret_stack == NULL) {
432 atomic_set(&t->trace_overrun, 0);
433 t->curr_ret_stack = -1;
434 t->curr_ret_depth = -1;
435 /* Make sure the tasks see the -1 first: */
437 t->ret_stack = ret_stack_list[start++];
444 for (i = start; i < end; i++)
445 kfree(ret_stack_list[i]);
450 ftrace_graph_probe_sched_switch(void *ignore, bool preempt,
451 struct task_struct *prev,
452 struct task_struct *next,
453 unsigned int prev_state)
455 unsigned long long timestamp;
459 * Does the user want to count the time a function was asleep.
460 * If so, do not update the time stamps.
462 if (fgraph_sleep_time)
465 timestamp = trace_clock_local();
467 prev->ftrace_timestamp = timestamp;
469 /* only process tasks that we timestamped */
470 if (!next->ftrace_timestamp)
474 * Update all the counters in next to make up for the
475 * time next was sleeping.
477 timestamp -= next->ftrace_timestamp;
479 for (index = next->curr_ret_stack; index >= 0; index--)
480 next->ret_stack[index].calltime += timestamp;
483 static int ftrace_graph_entry_test(struct ftrace_graph_ent *trace)
485 if (!ftrace_ops_test(&global_ops, trace->func, NULL))
487 return __ftrace_graph_entry(trace);
491 * The function graph tracer should only trace the functions defined
492 * by set_ftrace_filter and set_ftrace_notrace. If another function
493 * tracer ops is registered, the graph tracer requires testing the
494 * function against the global ops, and not just trace any function
495 * that any ftrace_ops registered.
497 void update_function_graph_func(void)
499 struct ftrace_ops *op;
500 bool do_test = false;
503 * The graph and global ops share the same set of functions
504 * to test. If any other ops is on the list, then
505 * the graph tracing needs to test if its the function
508 do_for_each_ftrace_op(op, ftrace_ops_list) {
509 if (op != &global_ops && op != &graph_ops &&
510 op != &ftrace_list_end) {
512 /* in double loop, break out with goto */
515 } while_for_each_ftrace_op(op);
518 ftrace_graph_entry = ftrace_graph_entry_test;
520 ftrace_graph_entry = __ftrace_graph_entry;
523 static DEFINE_PER_CPU(struct ftrace_ret_stack *, idle_ret_stack);
526 graph_init_task(struct task_struct *t, struct ftrace_ret_stack *ret_stack)
528 atomic_set(&t->trace_overrun, 0);
529 t->ftrace_timestamp = 0;
530 /* make curr_ret_stack visible before we add the ret_stack */
532 t->ret_stack = ret_stack;
536 * Allocate a return stack for the idle task. May be the first
537 * time through, or it may be done by CPU hotplug online.
539 void ftrace_graph_init_idle_task(struct task_struct *t, int cpu)
541 t->curr_ret_stack = -1;
542 t->curr_ret_depth = -1;
544 * The idle task has no parent, it either has its own
545 * stack or no stack at all.
548 WARN_ON(t->ret_stack != per_cpu(idle_ret_stack, cpu));
550 if (ftrace_graph_active) {
551 struct ftrace_ret_stack *ret_stack;
553 ret_stack = per_cpu(idle_ret_stack, cpu);
556 kmalloc_array(FTRACE_RETFUNC_DEPTH,
557 sizeof(struct ftrace_ret_stack),
561 per_cpu(idle_ret_stack, cpu) = ret_stack;
563 graph_init_task(t, ret_stack);
567 /* Allocate a return stack for newly created task */
568 void ftrace_graph_init_task(struct task_struct *t)
570 /* Make sure we do not use the parent ret_stack */
572 t->curr_ret_stack = -1;
573 t->curr_ret_depth = -1;
575 if (ftrace_graph_active) {
576 struct ftrace_ret_stack *ret_stack;
578 ret_stack = kmalloc_array(FTRACE_RETFUNC_DEPTH,
579 sizeof(struct ftrace_ret_stack),
583 graph_init_task(t, ret_stack);
587 void ftrace_graph_exit_task(struct task_struct *t)
589 struct ftrace_ret_stack *ret_stack = t->ret_stack;
592 /* NULL must become visible to IRQs before we free it: */
598 /* Allocate a return stack for each task */
599 static int start_graph_tracing(void)
601 struct ftrace_ret_stack **ret_stack_list;
604 ret_stack_list = kmalloc_array(FTRACE_RETSTACK_ALLOC_SIZE,
605 sizeof(struct ftrace_ret_stack *),
611 /* The cpu_boot init_task->ret_stack will never be freed */
612 for_each_online_cpu(cpu) {
613 if (!idle_task(cpu)->ret_stack)
614 ftrace_graph_init_idle_task(idle_task(cpu), cpu);
618 ret = alloc_retstack_tasklist(ret_stack_list);
619 } while (ret == -EAGAIN);
622 ret = register_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
624 pr_info("ftrace_graph: Couldn't activate tracepoint"
625 " probe to kernel_sched_switch\n");
628 kfree(ret_stack_list);
632 int register_ftrace_graph(struct fgraph_ops *gops)
636 mutex_lock(&ftrace_lock);
638 /* we currently allow only one tracer registered at a time */
639 if (ftrace_graph_active) {
644 register_pm_notifier(&ftrace_suspend_notifier);
646 ftrace_graph_active++;
647 ret = start_graph_tracing();
649 ftrace_graph_active--;
653 ftrace_graph_return = gops->retfunc;
656 * Update the indirect function to the entryfunc, and the
657 * function that gets called to the entry_test first. Then
658 * call the update fgraph entry function to determine if
659 * the entryfunc should be called directly or not.
661 __ftrace_graph_entry = gops->entryfunc;
662 ftrace_graph_entry = ftrace_graph_entry_test;
663 update_function_graph_func();
665 ret = ftrace_startup(&graph_ops, FTRACE_START_FUNC_RET);
667 mutex_unlock(&ftrace_lock);
671 void unregister_ftrace_graph(struct fgraph_ops *gops)
673 mutex_lock(&ftrace_lock);
675 if (unlikely(!ftrace_graph_active))
678 ftrace_graph_active--;
679 ftrace_graph_return = ftrace_stub_graph;
680 ftrace_graph_entry = ftrace_graph_entry_stub;
681 __ftrace_graph_entry = ftrace_graph_entry_stub;
682 ftrace_shutdown(&graph_ops, FTRACE_STOP_FUNC_RET);
683 unregister_pm_notifier(&ftrace_suspend_notifier);
684 unregister_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
687 mutex_unlock(&ftrace_lock);