3 * Function graph tracer.
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>
9 #include <linux/uaccess.h>
10 #include <linux/ftrace.h>
11 #include <linux/interrupt.h>
12 #include <linux/slab.h>
16 #include "trace_output.h"
18 static bool kill_ftrace_graph;
21 * ftrace_graph_is_dead - returns true if ftrace_graph_stop() was called
23 * ftrace_graph_stop() is called when a severe error is detected in
24 * the function graph tracing. This function is called by the critical
25 * paths of function graph to keep those paths from doing any more harm.
27 bool ftrace_graph_is_dead(void)
29 return kill_ftrace_graph;
33 * ftrace_graph_stop - set to permanently disable function graph tracincg
35 * In case of an error int function graph tracing, this is called
36 * to try to keep function graph tracing from causing any more harm.
37 * Usually this is pretty severe and this is called to try to at least
38 * get a warning out to the user.
40 void ftrace_graph_stop(void)
42 kill_ftrace_graph = true;
45 /* When set, irq functions will be ignored */
46 static int ftrace_graph_skip_irqs;
48 struct fgraph_cpu_data {
53 unsigned long enter_funcs[FTRACE_RETFUNC_DEPTH];
57 struct fgraph_cpu_data __percpu *cpu_data;
59 /* Place to preserve last processed entry. */
60 struct ftrace_graph_ent_entry ent;
61 struct ftrace_graph_ret_entry ret;
66 #define TRACE_GRAPH_INDENT 2
68 static unsigned int max_depth;
70 static struct tracer_opt trace_opts[] = {
71 /* Display overruns? (for self-debug purpose) */
72 { TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
74 { TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
75 /* Display Overhead ? */
76 { TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
77 /* Display proc name/pid */
78 { TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
79 /* Display duration of execution */
80 { TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
81 /* Display absolute time of an entry */
82 { TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
83 /* Display interrupts */
84 { TRACER_OPT(funcgraph-irqs, TRACE_GRAPH_PRINT_IRQS) },
85 /* Display function name after trailing } */
86 { TRACER_OPT(funcgraph-tail, TRACE_GRAPH_PRINT_TAIL) },
87 /* Include sleep time (scheduled out) between entry and return */
88 { TRACER_OPT(sleep-time, TRACE_GRAPH_SLEEP_TIME) },
89 /* Include time within nested functions */
90 { TRACER_OPT(graph-time, TRACE_GRAPH_GRAPH_TIME) },
94 static struct tracer_flags tracer_flags = {
95 /* Don't display overruns, proc, or tail by default */
96 .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
97 TRACE_GRAPH_PRINT_DURATION | TRACE_GRAPH_PRINT_IRQS |
98 TRACE_GRAPH_SLEEP_TIME | TRACE_GRAPH_GRAPH_TIME,
102 static struct trace_array *graph_array;
105 * DURATION column is being also used to display IRQ signs,
106 * following values are used by print_graph_irq and others
107 * to fill in space into DURATION column.
110 FLAGS_FILL_FULL = 1 << TRACE_GRAPH_PRINT_FILL_SHIFT,
111 FLAGS_FILL_START = 2 << TRACE_GRAPH_PRINT_FILL_SHIFT,
112 FLAGS_FILL_END = 3 << TRACE_GRAPH_PRINT_FILL_SHIFT,
116 print_graph_duration(struct trace_array *tr, unsigned long long duration,
117 struct trace_seq *s, u32 flags);
119 /* Add a function return address to the trace stack on thread info.*/
121 ftrace_push_return_trace(unsigned long ret, unsigned long func, int *depth,
122 unsigned long frame_pointer)
124 unsigned long long calltime;
127 if (unlikely(ftrace_graph_is_dead()))
130 if (!current->ret_stack)
134 * We must make sure the ret_stack is tested before we read
139 /* The return trace stack is full */
140 if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
141 atomic_inc(¤t->trace_overrun);
146 * The curr_ret_stack is an index to ftrace return stack of
147 * current task. Its value should be in [0, FTRACE_RETFUNC_
148 * DEPTH) when the function graph tracer is used. To support
149 * filtering out specific functions, it makes the index
150 * negative by subtracting huge value (FTRACE_NOTRACE_DEPTH)
151 * so when it sees a negative index the ftrace will ignore
152 * the record. And the index gets recovered when returning
153 * from the filtered function by adding the FTRACE_NOTRACE_
154 * DEPTH and then it'll continue to record functions normally.
156 * The curr_ret_stack is initialized to -1 and get increased
157 * in this function. So it can be less than -1 only if it was
158 * filtered out via ftrace_graph_notrace_addr() which can be
159 * set from set_graph_notrace file in tracefs by user.
161 if (current->curr_ret_stack < -1)
164 calltime = trace_clock_local();
166 index = ++current->curr_ret_stack;
167 if (ftrace_graph_notrace_addr(func))
168 current->curr_ret_stack -= FTRACE_NOTRACE_DEPTH;
170 current->ret_stack[index].ret = ret;
171 current->ret_stack[index].func = func;
172 current->ret_stack[index].calltime = calltime;
173 current->ret_stack[index].subtime = 0;
174 current->ret_stack[index].fp = frame_pointer;
175 *depth = current->curr_ret_stack;
180 /* Retrieve a function return address to the trace stack on thread info.*/
182 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
183 unsigned long frame_pointer)
187 index = current->curr_ret_stack;
190 * A negative index here means that it's just returned from a
191 * notrace'd function. Recover index to get an original
192 * return address. See ftrace_push_return_trace().
194 * TODO: Need to check whether the stack gets corrupted.
197 index += FTRACE_NOTRACE_DEPTH;
199 if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
202 /* Might as well panic, otherwise we have no where to go */
203 *ret = (unsigned long)panic;
207 #if defined(CONFIG_HAVE_FUNCTION_GRAPH_FP_TEST) && !defined(CC_USING_FENTRY)
209 * The arch may choose to record the frame pointer used
210 * and check it here to make sure that it is what we expect it
211 * to be. If gcc does not set the place holder of the return
212 * address in the frame pointer, and does a copy instead, then
213 * the function graph trace will fail. This test detects this
216 * Currently, x86_32 with optimize for size (-Os) makes the latest
219 * Note, -mfentry does not use frame pointers, and this test
220 * is not needed if CC_USING_FENTRY is set.
222 if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
224 WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
225 " from func %ps return to %lx\n",
226 current->ret_stack[index].fp,
228 (void *)current->ret_stack[index].func,
229 current->ret_stack[index].ret);
230 *ret = (unsigned long)panic;
235 *ret = current->ret_stack[index].ret;
236 trace->func = current->ret_stack[index].func;
237 trace->calltime = current->ret_stack[index].calltime;
238 trace->overrun = atomic_read(¤t->trace_overrun);
239 trace->depth = index;
243 * Send the trace to the ring-buffer.
244 * @return the original return address.
246 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
248 struct ftrace_graph_ret trace;
251 ftrace_pop_return_trace(&trace, &ret, frame_pointer);
252 trace.rettime = trace_clock_local();
254 current->curr_ret_stack--;
256 * The curr_ret_stack can be less than -1 only if it was
257 * filtered out and it's about to return from the function.
258 * Recover the index and continue to trace normal functions.
260 if (current->curr_ret_stack < -1) {
261 current->curr_ret_stack += FTRACE_NOTRACE_DEPTH;
266 * The trace should run after decrementing the ret counter
267 * in case an interrupt were to come in. We don't want to
268 * lose the interrupt if max_depth is set.
270 ftrace_graph_return(&trace);
272 if (unlikely(!ret)) {
275 /* Might as well panic. What else to do? */
276 ret = (unsigned long)panic;
282 int __trace_graph_entry(struct trace_array *tr,
283 struct ftrace_graph_ent *trace,
287 struct trace_event_call *call = &event_funcgraph_entry;
288 struct ring_buffer_event *event;
289 struct ring_buffer *buffer = tr->trace_buffer.buffer;
290 struct ftrace_graph_ent_entry *entry;
292 event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
293 sizeof(*entry), flags, pc);
296 entry = ring_buffer_event_data(event);
297 entry->graph_ent = *trace;
298 if (!call_filter_check_discard(call, entry, buffer, event))
299 __buffer_unlock_commit(buffer, event);
304 static inline int ftrace_graph_ignore_irqs(void)
306 if (!ftrace_graph_skip_irqs || trace_recursion_test(TRACE_IRQ_BIT))
312 int trace_graph_entry(struct ftrace_graph_ent *trace)
314 struct trace_array *tr = graph_array;
315 struct trace_array_cpu *data;
322 if (!ftrace_trace_task(tr))
325 /* trace it when it is-nested-in or is a function enabled. */
326 if ((!(trace->depth || ftrace_graph_addr(trace->func)) ||
327 ftrace_graph_ignore_irqs()) || (trace->depth < 0) ||
328 (max_depth && trace->depth >= max_depth))
332 * Do not trace a function if it's filtered by set_graph_notrace.
333 * Make the index of ret stack negative to indicate that it should
334 * ignore further functions. But it needs its own ret stack entry
335 * to recover the original index in order to continue tracing after
336 * returning from the function.
338 if (ftrace_graph_notrace_addr(trace->func))
342 * Stop here if tracing_threshold is set. We only write function return
343 * events to the ring buffer.
348 local_irq_save(flags);
349 cpu = raw_smp_processor_id();
350 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
351 disabled = atomic_inc_return(&data->disabled);
352 if (likely(disabled == 1)) {
353 pc = preempt_count();
354 ret = __trace_graph_entry(tr, trace, flags, pc);
359 atomic_dec(&data->disabled);
360 local_irq_restore(flags);
366 __trace_graph_function(struct trace_array *tr,
367 unsigned long ip, unsigned long flags, int pc)
369 u64 time = trace_clock_local();
370 struct ftrace_graph_ent ent = {
374 struct ftrace_graph_ret ret = {
381 __trace_graph_entry(tr, &ent, flags, pc);
382 __trace_graph_return(tr, &ret, flags, pc);
386 trace_graph_function(struct trace_array *tr,
387 unsigned long ip, unsigned long parent_ip,
388 unsigned long flags, int pc)
390 __trace_graph_function(tr, ip, flags, pc);
393 void __trace_graph_return(struct trace_array *tr,
394 struct ftrace_graph_ret *trace,
398 struct trace_event_call *call = &event_funcgraph_exit;
399 struct ring_buffer_event *event;
400 struct ring_buffer *buffer = tr->trace_buffer.buffer;
401 struct ftrace_graph_ret_entry *entry;
403 event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_RET,
404 sizeof(*entry), flags, pc);
407 entry = ring_buffer_event_data(event);
409 if (!call_filter_check_discard(call, entry, buffer, event))
410 __buffer_unlock_commit(buffer, event);
413 void trace_graph_return(struct ftrace_graph_ret *trace)
415 struct trace_array *tr = graph_array;
416 struct trace_array_cpu *data;
422 local_irq_save(flags);
423 cpu = raw_smp_processor_id();
424 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
425 disabled = atomic_inc_return(&data->disabled);
426 if (likely(disabled == 1)) {
427 pc = preempt_count();
428 __trace_graph_return(tr, trace, flags, pc);
430 atomic_dec(&data->disabled);
431 local_irq_restore(flags);
434 void set_graph_array(struct trace_array *tr)
438 /* Make graph_array visible before we start tracing */
443 static void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
445 if (tracing_thresh &&
446 (trace->rettime - trace->calltime < tracing_thresh))
449 trace_graph_return(trace);
452 static int graph_trace_init(struct trace_array *tr)
458 ret = register_ftrace_graph(&trace_graph_thresh_return,
461 ret = register_ftrace_graph(&trace_graph_return,
465 tracing_start_cmdline_record();
470 static void graph_trace_reset(struct trace_array *tr)
472 tracing_stop_cmdline_record();
473 unregister_ftrace_graph();
476 static int graph_trace_update_thresh(struct trace_array *tr)
478 graph_trace_reset(tr);
479 return graph_trace_init(tr);
482 static int max_bytes_for_cpu;
484 static void print_graph_cpu(struct trace_seq *s, int cpu)
487 * Start with a space character - to make it stand out
488 * to the right a bit when trace output is pasted into
491 trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
494 #define TRACE_GRAPH_PROCINFO_LENGTH 14
496 static void print_graph_proc(struct trace_seq *s, pid_t pid)
498 char comm[TASK_COMM_LEN];
499 /* sign + log10(MAX_INT) + '\0' */
505 trace_find_cmdline(pid, comm);
507 sprintf(pid_str, "%d", pid);
509 /* 1 stands for the "-" character */
510 len = strlen(comm) + strlen(pid_str) + 1;
512 if (len < TRACE_GRAPH_PROCINFO_LENGTH)
513 spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
515 /* First spaces to align center */
516 for (i = 0; i < spaces / 2; i++)
517 trace_seq_putc(s, ' ');
519 trace_seq_printf(s, "%s-%s", comm, pid_str);
521 /* Last spaces to align center */
522 for (i = 0; i < spaces - (spaces / 2); i++)
523 trace_seq_putc(s, ' ');
527 static void print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
529 trace_seq_putc(s, ' ');
530 trace_print_lat_fmt(s, entry);
533 /* If the pid changed since the last trace, output this event */
535 verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
543 last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
545 if (*last_pid == pid)
548 prev_pid = *last_pid;
554 * Context-switch trace line:
556 ------------------------------------------
557 | 1) migration/0--1 => sshd-1755
558 ------------------------------------------
561 trace_seq_puts(s, " ------------------------------------------\n");
562 print_graph_cpu(s, cpu);
563 print_graph_proc(s, prev_pid);
564 trace_seq_puts(s, " => ");
565 print_graph_proc(s, pid);
566 trace_seq_puts(s, "\n ------------------------------------------\n\n");
569 static struct ftrace_graph_ret_entry *
570 get_return_for_leaf(struct trace_iterator *iter,
571 struct ftrace_graph_ent_entry *curr)
573 struct fgraph_data *data = iter->private;
574 struct ring_buffer_iter *ring_iter = NULL;
575 struct ring_buffer_event *event;
576 struct ftrace_graph_ret_entry *next;
579 * If the previous output failed to write to the seq buffer,
580 * then we just reuse the data from before.
582 if (data && data->failed) {
587 ring_iter = trace_buffer_iter(iter, iter->cpu);
589 /* First peek to compare current entry and the next one */
591 event = ring_buffer_iter_peek(ring_iter, NULL);
594 * We need to consume the current entry to see
597 ring_buffer_consume(iter->trace_buffer->buffer, iter->cpu,
599 event = ring_buffer_peek(iter->trace_buffer->buffer, iter->cpu,
606 next = ring_buffer_event_data(event);
610 * Save current and next entries for later reference
611 * if the output fails.
615 * If the next event is not a return type, then
616 * we only care about what type it is. Otherwise we can
617 * safely copy the entire event.
619 if (next->ent.type == TRACE_GRAPH_RET)
622 data->ret.ent.type = next->ent.type;
626 if (next->ent.type != TRACE_GRAPH_RET)
629 if (curr->ent.pid != next->ent.pid ||
630 curr->graph_ent.func != next->ret.func)
633 /* this is a leaf, now advance the iterator */
635 ring_buffer_read(ring_iter, NULL);
640 static void print_graph_abs_time(u64 t, struct trace_seq *s)
642 unsigned long usecs_rem;
644 usecs_rem = do_div(t, NSEC_PER_SEC);
647 trace_seq_printf(s, "%5lu.%06lu | ",
648 (unsigned long)t, usecs_rem);
652 print_graph_irq(struct trace_iterator *iter, unsigned long addr,
653 enum trace_type type, int cpu, pid_t pid, u32 flags)
655 struct trace_array *tr = iter->tr;
656 struct trace_seq *s = &iter->seq;
657 struct trace_entry *ent = iter->ent;
659 if (addr < (unsigned long)__irqentry_text_start ||
660 addr >= (unsigned long)__irqentry_text_end)
663 if (tr->trace_flags & TRACE_ITER_CONTEXT_INFO) {
665 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
666 print_graph_abs_time(iter->ts, s);
669 if (flags & TRACE_GRAPH_PRINT_CPU)
670 print_graph_cpu(s, cpu);
673 if (flags & TRACE_GRAPH_PRINT_PROC) {
674 print_graph_proc(s, pid);
675 trace_seq_puts(s, " | ");
679 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
680 print_graph_lat_fmt(s, ent);
684 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_START);
686 if (type == TRACE_GRAPH_ENT)
687 trace_seq_puts(s, "==========>");
689 trace_seq_puts(s, "<==========");
691 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_END);
692 trace_seq_putc(s, '\n');
696 trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
698 unsigned long nsecs_rem = do_div(duration, 1000);
699 /* log10(ULONG_MAX) + '\0' */
705 sprintf(usecs_str, "%lu", (unsigned long) duration);
708 trace_seq_printf(s, "%s", usecs_str);
710 len = strlen(usecs_str);
712 /* Print nsecs (we don't want to exceed 7 numbers) */
714 size_t slen = min_t(size_t, sizeof(nsecs_str), 8UL - len);
716 snprintf(nsecs_str, slen, "%03lu", nsecs_rem);
717 trace_seq_printf(s, ".%s", nsecs_str);
718 len += strlen(nsecs_str) + 1;
721 trace_seq_puts(s, " us ");
723 /* Print remaining spaces to fit the row's width */
724 for (i = len; i < 8; i++)
725 trace_seq_putc(s, ' ');
729 print_graph_duration(struct trace_array *tr, unsigned long long duration,
730 struct trace_seq *s, u32 flags)
732 if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
733 !(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
736 /* No real adata, just filling the column with spaces */
737 switch (flags & TRACE_GRAPH_PRINT_FILL_MASK) {
738 case FLAGS_FILL_FULL:
739 trace_seq_puts(s, " | ");
741 case FLAGS_FILL_START:
742 trace_seq_puts(s, " ");
745 trace_seq_puts(s, " |");
749 /* Signal a overhead of time execution to the output */
750 if (flags & TRACE_GRAPH_PRINT_OVERHEAD)
751 trace_seq_printf(s, "%c ", trace_find_mark(duration));
753 trace_seq_puts(s, " ");
755 trace_print_graph_duration(duration, s);
756 trace_seq_puts(s, "| ");
759 /* Case of a leaf function on its call entry */
760 static enum print_line_t
761 print_graph_entry_leaf(struct trace_iterator *iter,
762 struct ftrace_graph_ent_entry *entry,
763 struct ftrace_graph_ret_entry *ret_entry,
764 struct trace_seq *s, u32 flags)
766 struct fgraph_data *data = iter->private;
767 struct trace_array *tr = iter->tr;
768 struct ftrace_graph_ret *graph_ret;
769 struct ftrace_graph_ent *call;
770 unsigned long long duration;
773 graph_ret = &ret_entry->ret;
774 call = &entry->graph_ent;
775 duration = graph_ret->rettime - graph_ret->calltime;
778 struct fgraph_cpu_data *cpu_data;
781 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
784 * Comments display at + 1 to depth. Since
785 * this is a leaf function, keep the comments
786 * equal to this depth.
788 cpu_data->depth = call->depth - 1;
790 /* No need to keep this function around for this depth */
791 if (call->depth < FTRACE_RETFUNC_DEPTH)
792 cpu_data->enter_funcs[call->depth] = 0;
795 /* Overhead and duration */
796 print_graph_duration(tr, duration, s, flags);
799 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
800 trace_seq_putc(s, ' ');
802 trace_seq_printf(s, "%ps();\n", (void *)call->func);
804 return trace_handle_return(s);
807 static enum print_line_t
808 print_graph_entry_nested(struct trace_iterator *iter,
809 struct ftrace_graph_ent_entry *entry,
810 struct trace_seq *s, int cpu, u32 flags)
812 struct ftrace_graph_ent *call = &entry->graph_ent;
813 struct fgraph_data *data = iter->private;
814 struct trace_array *tr = iter->tr;
818 struct fgraph_cpu_data *cpu_data;
821 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
822 cpu_data->depth = call->depth;
824 /* Save this function pointer to see if the exit matches */
825 if (call->depth < FTRACE_RETFUNC_DEPTH)
826 cpu_data->enter_funcs[call->depth] = call->func;
830 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
833 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
834 trace_seq_putc(s, ' ');
836 trace_seq_printf(s, "%ps() {\n", (void *)call->func);
838 if (trace_seq_has_overflowed(s))
839 return TRACE_TYPE_PARTIAL_LINE;
842 * we already consumed the current entry to check the next one
843 * and see if this is a leaf.
845 return TRACE_TYPE_NO_CONSUME;
849 print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
850 int type, unsigned long addr, u32 flags)
852 struct fgraph_data *data = iter->private;
853 struct trace_entry *ent = iter->ent;
854 struct trace_array *tr = iter->tr;
858 verif_pid(s, ent->pid, cpu, data);
862 print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
864 if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
868 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
869 print_graph_abs_time(iter->ts, s);
872 if (flags & TRACE_GRAPH_PRINT_CPU)
873 print_graph_cpu(s, cpu);
876 if (flags & TRACE_GRAPH_PRINT_PROC) {
877 print_graph_proc(s, ent->pid);
878 trace_seq_puts(s, " | ");
882 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
883 print_graph_lat_fmt(s, ent);
889 * Entry check for irq code
892 * - we are inside irq code
893 * - we just entered irq code
896 * - funcgraph-interrupts option is set
897 * - we are not inside irq code
900 check_irq_entry(struct trace_iterator *iter, u32 flags,
901 unsigned long addr, int depth)
905 struct fgraph_data *data = iter->private;
908 * If we are either displaying irqs, or we got called as
909 * a graph event and private data does not exist,
910 * then we bypass the irq check.
912 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
916 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
919 * We are inside the irq code
924 if ((addr < (unsigned long)__irqentry_text_start) ||
925 (addr >= (unsigned long)__irqentry_text_end))
929 * We are entering irq code.
936 * Return check for irq code
939 * - we are inside irq code
940 * - we just left irq code
943 * - funcgraph-interrupts option is set
944 * - we are not inside irq code
947 check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
951 struct fgraph_data *data = iter->private;
954 * If we are either displaying irqs, or we got called as
955 * a graph event and private data does not exist,
956 * then we bypass the irq check.
958 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
962 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
965 * We are not inside the irq code.
967 if (*depth_irq == -1)
971 * We are inside the irq code, and this is returning entry.
972 * Let's not trace it and clear the entry depth, since
973 * we are out of irq code.
975 * This condition ensures that we 'leave the irq code' once
976 * we are out of the entry depth. Thus protecting us from
977 * the RETURN entry loss.
979 if (*depth_irq >= depth) {
985 * We are inside the irq code, and this is not the entry.
990 static enum print_line_t
991 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
992 struct trace_iterator *iter, u32 flags)
994 struct fgraph_data *data = iter->private;
995 struct ftrace_graph_ent *call = &field->graph_ent;
996 struct ftrace_graph_ret_entry *leaf_ret;
997 static enum print_line_t ret;
1000 if (check_irq_entry(iter, flags, call->func, call->depth))
1001 return TRACE_TYPE_HANDLED;
1003 print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags);
1005 leaf_ret = get_return_for_leaf(iter, field);
1007 ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
1009 ret = print_graph_entry_nested(iter, field, s, cpu, flags);
1013 * If we failed to write our output, then we need to make
1014 * note of it. Because we already consumed our entry.
1026 static enum print_line_t
1027 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
1028 struct trace_entry *ent, struct trace_iterator *iter,
1031 unsigned long long duration = trace->rettime - trace->calltime;
1032 struct fgraph_data *data = iter->private;
1033 struct trace_array *tr = iter->tr;
1034 pid_t pid = ent->pid;
1035 int cpu = iter->cpu;
1039 if (check_irq_return(iter, flags, trace->depth))
1040 return TRACE_TYPE_HANDLED;
1043 struct fgraph_cpu_data *cpu_data;
1044 int cpu = iter->cpu;
1046 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
1049 * Comments display at + 1 to depth. This is the
1050 * return from a function, we now want the comments
1051 * to display at the same level of the bracket.
1053 cpu_data->depth = trace->depth - 1;
1055 if (trace->depth < FTRACE_RETFUNC_DEPTH) {
1056 if (cpu_data->enter_funcs[trace->depth] != trace->func)
1058 cpu_data->enter_funcs[trace->depth] = 0;
1062 print_graph_prologue(iter, s, 0, 0, flags);
1064 /* Overhead and duration */
1065 print_graph_duration(tr, duration, s, flags);
1068 for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++)
1069 trace_seq_putc(s, ' ');
1072 * If the return function does not have a matching entry,
1073 * then the entry was lost. Instead of just printing
1074 * the '}' and letting the user guess what function this
1075 * belongs to, write out the function name. Always do
1076 * that if the funcgraph-tail option is enabled.
1078 if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL))
1079 trace_seq_puts(s, "}\n");
1081 trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
1084 if (flags & TRACE_GRAPH_PRINT_OVERRUN)
1085 trace_seq_printf(s, " (Overruns: %lu)\n",
1088 print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
1091 return trace_handle_return(s);
1094 static enum print_line_t
1095 print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
1096 struct trace_iterator *iter, u32 flags)
1098 struct trace_array *tr = iter->tr;
1099 unsigned long sym_flags = (tr->trace_flags & TRACE_ITER_SYM_MASK);
1100 struct fgraph_data *data = iter->private;
1101 struct trace_event *event;
1107 depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
1109 print_graph_prologue(iter, s, 0, 0, flags);
1112 print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
1116 for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++)
1117 trace_seq_putc(s, ' ');
1120 trace_seq_puts(s, "/* ");
1122 switch (iter->ent->type) {
1124 ret = trace_print_bprintk_msg_only(iter);
1125 if (ret != TRACE_TYPE_HANDLED)
1129 ret = trace_print_printk_msg_only(iter);
1130 if (ret != TRACE_TYPE_HANDLED)
1134 event = ftrace_find_event(ent->type);
1136 return TRACE_TYPE_UNHANDLED;
1138 ret = event->funcs->trace(iter, sym_flags, event);
1139 if (ret != TRACE_TYPE_HANDLED)
1143 if (trace_seq_has_overflowed(s))
1146 /* Strip ending newline */
1147 if (s->buffer[s->seq.len - 1] == '\n') {
1148 s->buffer[s->seq.len - 1] = '\0';
1152 trace_seq_puts(s, " */\n");
1154 return trace_handle_return(s);
1159 print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1161 struct ftrace_graph_ent_entry *field;
1162 struct fgraph_data *data = iter->private;
1163 struct trace_entry *entry = iter->ent;
1164 struct trace_seq *s = &iter->seq;
1165 int cpu = iter->cpu;
1168 if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1169 per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1170 return TRACE_TYPE_HANDLED;
1174 * If the last output failed, there's a possibility we need
1175 * to print out the missing entry which would never go out.
1177 if (data && data->failed) {
1179 iter->cpu = data->cpu;
1180 ret = print_graph_entry(field, s, iter, flags);
1181 if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1182 per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1183 ret = TRACE_TYPE_NO_CONSUME;
1189 switch (entry->type) {
1190 case TRACE_GRAPH_ENT: {
1192 * print_graph_entry() may consume the current event,
1193 * thus @field may become invalid, so we need to save it.
1194 * sizeof(struct ftrace_graph_ent_entry) is very small,
1195 * it can be safely saved at the stack.
1197 struct ftrace_graph_ent_entry saved;
1198 trace_assign_type(field, entry);
1200 return print_graph_entry(&saved, s, iter, flags);
1202 case TRACE_GRAPH_RET: {
1203 struct ftrace_graph_ret_entry *field;
1204 trace_assign_type(field, entry);
1205 return print_graph_return(&field->ret, s, entry, iter, flags);
1209 /* dont trace stack and functions as comments */
1210 return TRACE_TYPE_UNHANDLED;
1213 return print_graph_comment(s, entry, iter, flags);
1216 return TRACE_TYPE_HANDLED;
1219 static enum print_line_t
1220 print_graph_function(struct trace_iterator *iter)
1222 return print_graph_function_flags(iter, tracer_flags.val);
1225 static enum print_line_t
1226 print_graph_function_event(struct trace_iterator *iter, int flags,
1227 struct trace_event *event)
1229 return print_graph_function(iter);
1232 static void print_lat_header(struct seq_file *s, u32 flags)
1234 static const char spaces[] = " " /* 16 spaces */
1236 " "; /* 17 spaces */
1239 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1241 if (flags & TRACE_GRAPH_PRINT_CPU)
1243 if (flags & TRACE_GRAPH_PRINT_PROC)
1246 seq_printf(s, "#%.*s _-----=> irqs-off \n", size, spaces);
1247 seq_printf(s, "#%.*s / _----=> need-resched \n", size, spaces);
1248 seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1249 seq_printf(s, "#%.*s|| / _--=> preempt-depth \n", size, spaces);
1250 seq_printf(s, "#%.*s||| / \n", size, spaces);
1253 static void __print_graph_headers_flags(struct trace_array *tr,
1254 struct seq_file *s, u32 flags)
1256 int lat = tr->trace_flags & TRACE_ITER_LATENCY_FMT;
1259 print_lat_header(s, flags);
1263 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1264 seq_puts(s, " TIME ");
1265 if (flags & TRACE_GRAPH_PRINT_CPU)
1266 seq_puts(s, " CPU");
1267 if (flags & TRACE_GRAPH_PRINT_PROC)
1268 seq_puts(s, " TASK/PID ");
1270 seq_puts(s, "||||");
1271 if (flags & TRACE_GRAPH_PRINT_DURATION)
1272 seq_puts(s, " DURATION ");
1273 seq_puts(s, " FUNCTION CALLS\n");
1277 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1279 if (flags & TRACE_GRAPH_PRINT_CPU)
1281 if (flags & TRACE_GRAPH_PRINT_PROC)
1282 seq_puts(s, " | | ");
1284 seq_puts(s, "||||");
1285 if (flags & TRACE_GRAPH_PRINT_DURATION)
1286 seq_puts(s, " | | ");
1287 seq_puts(s, " | | | |\n");
1290 static void print_graph_headers(struct seq_file *s)
1292 print_graph_headers_flags(s, tracer_flags.val);
1295 void print_graph_headers_flags(struct seq_file *s, u32 flags)
1297 struct trace_iterator *iter = s->private;
1298 struct trace_array *tr = iter->tr;
1300 if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
1303 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT) {
1304 /* print nothing if the buffers are empty */
1305 if (trace_empty(iter))
1308 print_trace_header(s, iter);
1311 __print_graph_headers_flags(tr, s, flags);
1314 void graph_trace_open(struct trace_iterator *iter)
1316 /* pid and depth on the last trace processed */
1317 struct fgraph_data *data;
1321 iter->private = NULL;
1323 /* We can be called in atomic context via ftrace_dump() */
1324 gfpflags = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
1326 data = kzalloc(sizeof(*data), gfpflags);
1330 data->cpu_data = alloc_percpu_gfp(struct fgraph_cpu_data, gfpflags);
1331 if (!data->cpu_data)
1334 for_each_possible_cpu(cpu) {
1335 pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1336 int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1337 int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1338 int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1346 iter->private = data;
1353 pr_warn("function graph tracer: not enough memory\n");
1356 void graph_trace_close(struct trace_iterator *iter)
1358 struct fgraph_data *data = iter->private;
1361 free_percpu(data->cpu_data);
1367 func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1369 if (bit == TRACE_GRAPH_PRINT_IRQS)
1370 ftrace_graph_skip_irqs = !set;
1372 if (bit == TRACE_GRAPH_SLEEP_TIME)
1373 ftrace_graph_sleep_time_control(set);
1375 if (bit == TRACE_GRAPH_GRAPH_TIME)
1376 ftrace_graph_graph_time_control(set);
1381 static struct trace_event_functions graph_functions = {
1382 .trace = print_graph_function_event,
1385 static struct trace_event graph_trace_entry_event = {
1386 .type = TRACE_GRAPH_ENT,
1387 .funcs = &graph_functions,
1390 static struct trace_event graph_trace_ret_event = {
1391 .type = TRACE_GRAPH_RET,
1392 .funcs = &graph_functions
1395 static struct tracer graph_trace __tracer_data = {
1396 .name = "function_graph",
1397 .update_thresh = graph_trace_update_thresh,
1398 .open = graph_trace_open,
1399 .pipe_open = graph_trace_open,
1400 .close = graph_trace_close,
1401 .pipe_close = graph_trace_close,
1402 .init = graph_trace_init,
1403 .reset = graph_trace_reset,
1404 .print_line = print_graph_function,
1405 .print_header = print_graph_headers,
1406 .flags = &tracer_flags,
1407 .set_flag = func_graph_set_flag,
1408 #ifdef CONFIG_FTRACE_SELFTEST
1409 .selftest = trace_selftest_startup_function_graph,
1415 graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1421 ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1433 graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1436 char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1439 n = sprintf(buf, "%d\n", max_depth);
1441 return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1444 static const struct file_operations graph_depth_fops = {
1445 .open = tracing_open_generic,
1446 .write = graph_depth_write,
1447 .read = graph_depth_read,
1448 .llseek = generic_file_llseek,
1451 static __init int init_graph_tracefs(void)
1453 struct dentry *d_tracer;
1455 d_tracer = tracing_init_dentry();
1456 if (IS_ERR(d_tracer))
1459 trace_create_file("max_graph_depth", 0644, d_tracer,
1460 NULL, &graph_depth_fops);
1464 fs_initcall(init_graph_tracefs);
1466 static __init int init_graph_trace(void)
1468 max_bytes_for_cpu = snprintf(NULL, 0, "%d", nr_cpu_ids - 1);
1470 if (!register_trace_event(&graph_trace_entry_event)) {
1471 pr_warn("Warning: could not register graph trace events\n");
1475 if (!register_trace_event(&graph_trace_ret_event)) {
1476 pr_warn("Warning: could not register graph trace events\n");
1480 return register_tracer(&graph_trace);
1483 core_initcall(init_graph_trace);