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/slab.h>
15 #include "trace_output.h"
17 static bool kill_ftrace_graph;
20 * ftrace_graph_is_dead - returns true if ftrace_graph_stop() was called
22 * ftrace_graph_stop() is called when a severe error is detected in
23 * the function graph tracing. This function is called by the critical
24 * paths of function graph to keep those paths from doing any more harm.
26 bool ftrace_graph_is_dead(void)
28 return kill_ftrace_graph;
32 * ftrace_graph_stop - set to permanently disable function graph tracincg
34 * In case of an error int function graph tracing, this is called
35 * to try to keep function graph tracing from causing any more harm.
36 * Usually this is pretty severe and this is called to try to at least
37 * get a warning out to the user.
39 void ftrace_graph_stop(void)
41 kill_ftrace_graph = true;
44 /* When set, irq functions will be ignored */
45 static int ftrace_graph_skip_irqs;
47 struct fgraph_cpu_data {
52 unsigned long enter_funcs[FTRACE_RETFUNC_DEPTH];
56 struct fgraph_cpu_data __percpu *cpu_data;
58 /* Place to preserve last processed entry. */
59 struct ftrace_graph_ent_entry ent;
60 struct ftrace_graph_ret_entry ret;
65 #define TRACE_GRAPH_INDENT 2
67 static unsigned int max_depth;
69 static struct tracer_opt trace_opts[] = {
70 /* Display overruns? (for self-debug purpose) */
71 { TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
73 { TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
74 /* Display Overhead ? */
75 { TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
76 /* Display proc name/pid */
77 { TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
78 /* Display duration of execution */
79 { TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
80 /* Display absolute time of an entry */
81 { TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
82 /* Display interrupts */
83 { TRACER_OPT(funcgraph-irqs, TRACE_GRAPH_PRINT_IRQS) },
84 /* Display function name after trailing } */
85 { TRACER_OPT(funcgraph-tail, TRACE_GRAPH_PRINT_TAIL) },
89 static struct tracer_flags tracer_flags = {
90 /* Don't display overruns, proc, or tail by default */
91 .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
92 TRACE_GRAPH_PRINT_DURATION | TRACE_GRAPH_PRINT_IRQS,
96 static struct trace_array *graph_array;
99 * DURATION column is being also used to display IRQ signs,
100 * following values are used by print_graph_irq and others
101 * to fill in space into DURATION column.
104 FLAGS_FILL_FULL = 1 << TRACE_GRAPH_PRINT_FILL_SHIFT,
105 FLAGS_FILL_START = 2 << TRACE_GRAPH_PRINT_FILL_SHIFT,
106 FLAGS_FILL_END = 3 << TRACE_GRAPH_PRINT_FILL_SHIFT,
110 print_graph_duration(unsigned long long duration, struct trace_seq *s,
113 /* Add a function return address to the trace stack on thread info.*/
115 ftrace_push_return_trace(unsigned long ret, unsigned long func, int *depth,
116 unsigned long frame_pointer)
118 unsigned long long calltime;
121 if (unlikely(ftrace_graph_is_dead()))
124 if (!current->ret_stack)
128 * We must make sure the ret_stack is tested before we read
133 /* The return trace stack is full */
134 if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
135 atomic_inc(¤t->trace_overrun);
140 * The curr_ret_stack is an index to ftrace return stack of
141 * current task. Its value should be in [0, FTRACE_RETFUNC_
142 * DEPTH) when the function graph tracer is used. To support
143 * filtering out specific functions, it makes the index
144 * negative by subtracting huge value (FTRACE_NOTRACE_DEPTH)
145 * so when it sees a negative index the ftrace will ignore
146 * the record. And the index gets recovered when returning
147 * from the filtered function by adding the FTRACE_NOTRACE_
148 * DEPTH and then it'll continue to record functions normally.
150 * The curr_ret_stack is initialized to -1 and get increased
151 * in this function. So it can be less than -1 only if it was
152 * filtered out via ftrace_graph_notrace_addr() which can be
153 * set from set_graph_notrace file in tracefs by user.
155 if (current->curr_ret_stack < -1)
158 calltime = trace_clock_local();
160 index = ++current->curr_ret_stack;
161 if (ftrace_graph_notrace_addr(func))
162 current->curr_ret_stack -= FTRACE_NOTRACE_DEPTH;
164 current->ret_stack[index].ret = ret;
165 current->ret_stack[index].func = func;
166 current->ret_stack[index].calltime = calltime;
167 current->ret_stack[index].subtime = 0;
168 current->ret_stack[index].fp = frame_pointer;
169 *depth = current->curr_ret_stack;
174 /* Retrieve a function return address to the trace stack on thread info.*/
176 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
177 unsigned long frame_pointer)
181 index = current->curr_ret_stack;
184 * A negative index here means that it's just returned from a
185 * notrace'd function. Recover index to get an original
186 * return address. See ftrace_push_return_trace().
188 * TODO: Need to check whether the stack gets corrupted.
191 index += FTRACE_NOTRACE_DEPTH;
193 if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
196 /* Might as well panic, otherwise we have no where to go */
197 *ret = (unsigned long)panic;
201 #if defined(CONFIG_HAVE_FUNCTION_GRAPH_FP_TEST) && !defined(CC_USING_FENTRY)
203 * The arch may choose to record the frame pointer used
204 * and check it here to make sure that it is what we expect it
205 * to be. If gcc does not set the place holder of the return
206 * address in the frame pointer, and does a copy instead, then
207 * the function graph trace will fail. This test detects this
210 * Currently, x86_32 with optimize for size (-Os) makes the latest
213 * Note, -mfentry does not use frame pointers, and this test
214 * is not needed if CC_USING_FENTRY is set.
216 if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
218 WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
219 " from func %ps return to %lx\n",
220 current->ret_stack[index].fp,
222 (void *)current->ret_stack[index].func,
223 current->ret_stack[index].ret);
224 *ret = (unsigned long)panic;
229 *ret = current->ret_stack[index].ret;
230 trace->func = current->ret_stack[index].func;
231 trace->calltime = current->ret_stack[index].calltime;
232 trace->overrun = atomic_read(¤t->trace_overrun);
233 trace->depth = index;
237 * Send the trace to the ring-buffer.
238 * @return the original return address.
240 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
242 struct ftrace_graph_ret trace;
245 ftrace_pop_return_trace(&trace, &ret, frame_pointer);
246 trace.rettime = trace_clock_local();
248 current->curr_ret_stack--;
250 * The curr_ret_stack can be less than -1 only if it was
251 * filtered out and it's about to return from the function.
252 * Recover the index and continue to trace normal functions.
254 if (current->curr_ret_stack < -1) {
255 current->curr_ret_stack += FTRACE_NOTRACE_DEPTH;
260 * The trace should run after decrementing the ret counter
261 * in case an interrupt were to come in. We don't want to
262 * lose the interrupt if max_depth is set.
264 ftrace_graph_return(&trace);
266 if (unlikely(!ret)) {
269 /* Might as well panic. What else to do? */
270 ret = (unsigned long)panic;
276 int __trace_graph_entry(struct trace_array *tr,
277 struct ftrace_graph_ent *trace,
281 struct trace_event_call *call = &event_funcgraph_entry;
282 struct ring_buffer_event *event;
283 struct ring_buffer *buffer = tr->trace_buffer.buffer;
284 struct ftrace_graph_ent_entry *entry;
286 if (unlikely(__this_cpu_read(ftrace_cpu_disabled)))
289 event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
290 sizeof(*entry), flags, pc);
293 entry = ring_buffer_event_data(event);
294 entry->graph_ent = *trace;
295 if (!call_filter_check_discard(call, entry, buffer, event))
296 __buffer_unlock_commit(buffer, event);
301 static inline int ftrace_graph_ignore_irqs(void)
303 if (!ftrace_graph_skip_irqs || trace_recursion_test(TRACE_IRQ_BIT))
309 int trace_graph_entry(struct ftrace_graph_ent *trace)
311 struct trace_array *tr = graph_array;
312 struct trace_array_cpu *data;
319 if (!ftrace_trace_task(current))
322 /* trace it when it is-nested-in or is a function enabled. */
323 if ((!(trace->depth || ftrace_graph_addr(trace->func)) ||
324 ftrace_graph_ignore_irqs()) || (trace->depth < 0) ||
325 (max_depth && trace->depth >= max_depth))
329 * Do not trace a function if it's filtered by set_graph_notrace.
330 * Make the index of ret stack negative to indicate that it should
331 * ignore further functions. But it needs its own ret stack entry
332 * to recover the original index in order to continue tracing after
333 * returning from the function.
335 if (ftrace_graph_notrace_addr(trace->func))
338 local_irq_save(flags);
339 cpu = raw_smp_processor_id();
340 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
341 disabled = atomic_inc_return(&data->disabled);
342 if (likely(disabled == 1)) {
343 pc = preempt_count();
344 ret = __trace_graph_entry(tr, trace, flags, pc);
349 atomic_dec(&data->disabled);
350 local_irq_restore(flags);
355 static int trace_graph_thresh_entry(struct ftrace_graph_ent *trace)
360 return trace_graph_entry(trace);
364 __trace_graph_function(struct trace_array *tr,
365 unsigned long ip, unsigned long flags, int pc)
367 u64 time = trace_clock_local();
368 struct ftrace_graph_ent ent = {
372 struct ftrace_graph_ret ret = {
379 __trace_graph_entry(tr, &ent, flags, pc);
380 __trace_graph_return(tr, &ret, flags, pc);
384 trace_graph_function(struct trace_array *tr,
385 unsigned long ip, unsigned long parent_ip,
386 unsigned long flags, int pc)
388 __trace_graph_function(tr, ip, flags, pc);
391 void __trace_graph_return(struct trace_array *tr,
392 struct ftrace_graph_ret *trace,
396 struct trace_event_call *call = &event_funcgraph_exit;
397 struct ring_buffer_event *event;
398 struct ring_buffer *buffer = tr->trace_buffer.buffer;
399 struct ftrace_graph_ret_entry *entry;
401 if (unlikely(__this_cpu_read(ftrace_cpu_disabled)))
404 event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_RET,
405 sizeof(*entry), flags, pc);
408 entry = ring_buffer_event_data(event);
410 if (!call_filter_check_discard(call, entry, buffer, event))
411 __buffer_unlock_commit(buffer, event);
414 void trace_graph_return(struct ftrace_graph_ret *trace)
416 struct trace_array *tr = graph_array;
417 struct trace_array_cpu *data;
423 local_irq_save(flags);
424 cpu = raw_smp_processor_id();
425 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
426 disabled = atomic_inc_return(&data->disabled);
427 if (likely(disabled == 1)) {
428 pc = preempt_count();
429 __trace_graph_return(tr, trace, flags, pc);
431 atomic_dec(&data->disabled);
432 local_irq_restore(flags);
435 void set_graph_array(struct trace_array *tr)
439 /* Make graph_array visible before we start tracing */
444 static void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
446 if (tracing_thresh &&
447 (trace->rettime - trace->calltime < tracing_thresh))
450 trace_graph_return(trace);
453 static int graph_trace_init(struct trace_array *tr)
459 ret = register_ftrace_graph(&trace_graph_thresh_return,
460 &trace_graph_thresh_entry);
462 ret = register_ftrace_graph(&trace_graph_return,
466 tracing_start_cmdline_record();
471 static void graph_trace_reset(struct trace_array *tr)
473 tracing_stop_cmdline_record();
474 unregister_ftrace_graph();
477 static int graph_trace_update_thresh(struct trace_array *tr)
479 graph_trace_reset(tr);
480 return graph_trace_init(tr);
483 static int max_bytes_for_cpu;
485 static void print_graph_cpu(struct trace_seq *s, int cpu)
488 * Start with a space character - to make it stand out
489 * to the right a bit when trace output is pasted into
492 trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
495 #define TRACE_GRAPH_PROCINFO_LENGTH 14
497 static void print_graph_proc(struct trace_seq *s, pid_t pid)
499 char comm[TASK_COMM_LEN];
500 /* sign + log10(MAX_INT) + '\0' */
506 trace_find_cmdline(pid, comm);
508 sprintf(pid_str, "%d", pid);
510 /* 1 stands for the "-" character */
511 len = strlen(comm) + strlen(pid_str) + 1;
513 if (len < TRACE_GRAPH_PROCINFO_LENGTH)
514 spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
516 /* First spaces to align center */
517 for (i = 0; i < spaces / 2; i++)
518 trace_seq_putc(s, ' ');
520 trace_seq_printf(s, "%s-%s", comm, pid_str);
522 /* Last spaces to align center */
523 for (i = 0; i < spaces - (spaces / 2); i++)
524 trace_seq_putc(s, ' ');
528 static void print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
530 trace_seq_putc(s, ' ');
531 trace_print_lat_fmt(s, entry);
534 /* If the pid changed since the last trace, output this event */
536 verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
544 last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
546 if (*last_pid == pid)
549 prev_pid = *last_pid;
555 * Context-switch trace line:
557 ------------------------------------------
558 | 1) migration/0--1 => sshd-1755
559 ------------------------------------------
562 trace_seq_puts(s, " ------------------------------------------\n");
563 print_graph_cpu(s, cpu);
564 print_graph_proc(s, prev_pid);
565 trace_seq_puts(s, " => ");
566 print_graph_proc(s, pid);
567 trace_seq_puts(s, "\n ------------------------------------------\n\n");
570 static struct ftrace_graph_ret_entry *
571 get_return_for_leaf(struct trace_iterator *iter,
572 struct ftrace_graph_ent_entry *curr)
574 struct fgraph_data *data = iter->private;
575 struct ring_buffer_iter *ring_iter = NULL;
576 struct ring_buffer_event *event;
577 struct ftrace_graph_ret_entry *next;
580 * If the previous output failed to write to the seq buffer,
581 * then we just reuse the data from before.
583 if (data && data->failed) {
588 ring_iter = trace_buffer_iter(iter, iter->cpu);
590 /* First peek to compare current entry and the next one */
592 event = ring_buffer_iter_peek(ring_iter, NULL);
595 * We need to consume the current entry to see
598 ring_buffer_consume(iter->trace_buffer->buffer, iter->cpu,
600 event = ring_buffer_peek(iter->trace_buffer->buffer, iter->cpu,
607 next = ring_buffer_event_data(event);
611 * Save current and next entries for later reference
612 * if the output fails.
616 * If the next event is not a return type, then
617 * we only care about what type it is. Otherwise we can
618 * safely copy the entire event.
620 if (next->ent.type == TRACE_GRAPH_RET)
623 data->ret.ent.type = next->ent.type;
627 if (next->ent.type != TRACE_GRAPH_RET)
630 if (curr->ent.pid != next->ent.pid ||
631 curr->graph_ent.func != next->ret.func)
634 /* this is a leaf, now advance the iterator */
636 ring_buffer_read(ring_iter, NULL);
641 static void print_graph_abs_time(u64 t, struct trace_seq *s)
643 unsigned long usecs_rem;
645 usecs_rem = do_div(t, NSEC_PER_SEC);
648 trace_seq_printf(s, "%5lu.%06lu | ",
649 (unsigned long)t, usecs_rem);
653 print_graph_irq(struct trace_iterator *iter, unsigned long addr,
654 enum trace_type type, int cpu, pid_t pid, u32 flags)
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 (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 (trace_flags & TRACE_ITER_LATENCY_FMT)
680 print_graph_lat_fmt(s, ent);
684 print_graph_duration(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(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(unsigned long long duration, struct trace_seq *s,
732 if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
733 !(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 ftrace_graph_ret *graph_ret;
768 struct ftrace_graph_ent *call;
769 unsigned long long duration;
772 graph_ret = &ret_entry->ret;
773 call = &entry->graph_ent;
774 duration = graph_ret->rettime - graph_ret->calltime;
777 struct fgraph_cpu_data *cpu_data;
780 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
783 * Comments display at + 1 to depth. Since
784 * this is a leaf function, keep the comments
785 * equal to this depth.
787 cpu_data->depth = call->depth - 1;
789 /* No need to keep this function around for this depth */
790 if (call->depth < FTRACE_RETFUNC_DEPTH)
791 cpu_data->enter_funcs[call->depth] = 0;
794 /* Overhead and duration */
795 print_graph_duration(duration, s, flags);
798 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
799 trace_seq_putc(s, ' ');
801 trace_seq_printf(s, "%ps();\n", (void *)call->func);
803 return trace_handle_return(s);
806 static enum print_line_t
807 print_graph_entry_nested(struct trace_iterator *iter,
808 struct ftrace_graph_ent_entry *entry,
809 struct trace_seq *s, int cpu, u32 flags)
811 struct ftrace_graph_ent *call = &entry->graph_ent;
812 struct fgraph_data *data = iter->private;
816 struct fgraph_cpu_data *cpu_data;
819 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
820 cpu_data->depth = call->depth;
822 /* Save this function pointer to see if the exit matches */
823 if (call->depth < FTRACE_RETFUNC_DEPTH)
824 cpu_data->enter_funcs[call->depth] = call->func;
828 print_graph_duration(0, s, flags | FLAGS_FILL_FULL);
831 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
832 trace_seq_putc(s, ' ');
834 trace_seq_printf(s, "%ps() {\n", (void *)call->func);
836 if (trace_seq_has_overflowed(s))
837 return TRACE_TYPE_PARTIAL_LINE;
840 * we already consumed the current entry to check the next one
841 * and see if this is a leaf.
843 return TRACE_TYPE_NO_CONSUME;
847 print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
848 int type, unsigned long addr, u32 flags)
850 struct fgraph_data *data = iter->private;
851 struct trace_entry *ent = iter->ent;
855 verif_pid(s, ent->pid, cpu, data);
859 print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
861 if (!(trace_flags & TRACE_ITER_CONTEXT_INFO))
865 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
866 print_graph_abs_time(iter->ts, s);
869 if (flags & TRACE_GRAPH_PRINT_CPU)
870 print_graph_cpu(s, cpu);
873 if (flags & TRACE_GRAPH_PRINT_PROC) {
874 print_graph_proc(s, ent->pid);
875 trace_seq_puts(s, " | ");
879 if (trace_flags & TRACE_ITER_LATENCY_FMT)
880 print_graph_lat_fmt(s, ent);
886 * Entry check for irq code
889 * - we are inside irq code
890 * - we just entered irq code
893 * - funcgraph-interrupts option is set
894 * - we are not inside irq code
897 check_irq_entry(struct trace_iterator *iter, u32 flags,
898 unsigned long addr, int depth)
902 struct fgraph_data *data = iter->private;
905 * If we are either displaying irqs, or we got called as
906 * a graph event and private data does not exist,
907 * then we bypass the irq check.
909 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
913 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
916 * We are inside the irq code
921 if ((addr < (unsigned long)__irqentry_text_start) ||
922 (addr >= (unsigned long)__irqentry_text_end))
926 * We are entering irq code.
933 * Return check for irq code
936 * - we are inside irq code
937 * - we just left irq code
940 * - funcgraph-interrupts option is set
941 * - we are not inside irq code
944 check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
948 struct fgraph_data *data = iter->private;
951 * If we are either displaying irqs, or we got called as
952 * a graph event and private data does not exist,
953 * then we bypass the irq check.
955 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
959 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
962 * We are not inside the irq code.
964 if (*depth_irq == -1)
968 * We are inside the irq code, and this is returning entry.
969 * Let's not trace it and clear the entry depth, since
970 * we are out of irq code.
972 * This condition ensures that we 'leave the irq code' once
973 * we are out of the entry depth. Thus protecting us from
974 * the RETURN entry loss.
976 if (*depth_irq >= depth) {
982 * We are inside the irq code, and this is not the entry.
987 static enum print_line_t
988 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
989 struct trace_iterator *iter, u32 flags)
991 struct fgraph_data *data = iter->private;
992 struct ftrace_graph_ent *call = &field->graph_ent;
993 struct ftrace_graph_ret_entry *leaf_ret;
994 static enum print_line_t ret;
997 if (check_irq_entry(iter, flags, call->func, call->depth))
998 return TRACE_TYPE_HANDLED;
1000 print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags);
1002 leaf_ret = get_return_for_leaf(iter, field);
1004 ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
1006 ret = print_graph_entry_nested(iter, field, s, cpu, flags);
1010 * If we failed to write our output, then we need to make
1011 * note of it. Because we already consumed our entry.
1023 static enum print_line_t
1024 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
1025 struct trace_entry *ent, struct trace_iterator *iter,
1028 unsigned long long duration = trace->rettime - trace->calltime;
1029 struct fgraph_data *data = iter->private;
1030 pid_t pid = ent->pid;
1031 int cpu = iter->cpu;
1035 if (check_irq_return(iter, flags, trace->depth))
1036 return TRACE_TYPE_HANDLED;
1039 struct fgraph_cpu_data *cpu_data;
1040 int cpu = iter->cpu;
1042 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
1045 * Comments display at + 1 to depth. This is the
1046 * return from a function, we now want the comments
1047 * to display at the same level of the bracket.
1049 cpu_data->depth = trace->depth - 1;
1051 if (trace->depth < FTRACE_RETFUNC_DEPTH) {
1052 if (cpu_data->enter_funcs[trace->depth] != trace->func)
1054 cpu_data->enter_funcs[trace->depth] = 0;
1058 print_graph_prologue(iter, s, 0, 0, flags);
1060 /* Overhead and duration */
1061 print_graph_duration(duration, s, flags);
1064 for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++)
1065 trace_seq_putc(s, ' ');
1068 * If the return function does not have a matching entry,
1069 * then the entry was lost. Instead of just printing
1070 * the '}' and letting the user guess what function this
1071 * belongs to, write out the function name. Always do
1072 * that if the funcgraph-tail option is enabled.
1074 if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL))
1075 trace_seq_puts(s, "}\n");
1077 trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
1080 if (flags & TRACE_GRAPH_PRINT_OVERRUN)
1081 trace_seq_printf(s, " (Overruns: %lu)\n",
1084 print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
1087 return trace_handle_return(s);
1090 static enum print_line_t
1091 print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
1092 struct trace_iterator *iter, u32 flags)
1094 unsigned long sym_flags = (trace_flags & TRACE_ITER_SYM_MASK);
1095 struct fgraph_data *data = iter->private;
1096 struct trace_event *event;
1102 depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
1104 print_graph_prologue(iter, s, 0, 0, flags);
1107 print_graph_duration(0, s, flags | FLAGS_FILL_FULL);
1111 for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++)
1112 trace_seq_putc(s, ' ');
1115 trace_seq_puts(s, "/* ");
1117 switch (iter->ent->type) {
1119 ret = trace_print_bprintk_msg_only(iter);
1120 if (ret != TRACE_TYPE_HANDLED)
1124 ret = trace_print_printk_msg_only(iter);
1125 if (ret != TRACE_TYPE_HANDLED)
1129 event = ftrace_find_event(ent->type);
1131 return TRACE_TYPE_UNHANDLED;
1133 ret = event->funcs->trace(iter, sym_flags, event);
1134 if (ret != TRACE_TYPE_HANDLED)
1138 if (trace_seq_has_overflowed(s))
1141 /* Strip ending newline */
1142 if (s->buffer[s->seq.len - 1] == '\n') {
1143 s->buffer[s->seq.len - 1] = '\0';
1147 trace_seq_puts(s, " */\n");
1149 return trace_handle_return(s);
1154 print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1156 struct ftrace_graph_ent_entry *field;
1157 struct fgraph_data *data = iter->private;
1158 struct trace_entry *entry = iter->ent;
1159 struct trace_seq *s = &iter->seq;
1160 int cpu = iter->cpu;
1163 if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1164 per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1165 return TRACE_TYPE_HANDLED;
1169 * If the last output failed, there's a possibility we need
1170 * to print out the missing entry which would never go out.
1172 if (data && data->failed) {
1174 iter->cpu = data->cpu;
1175 ret = print_graph_entry(field, s, iter, flags);
1176 if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1177 per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1178 ret = TRACE_TYPE_NO_CONSUME;
1184 switch (entry->type) {
1185 case TRACE_GRAPH_ENT: {
1187 * print_graph_entry() may consume the current event,
1188 * thus @field may become invalid, so we need to save it.
1189 * sizeof(struct ftrace_graph_ent_entry) is very small,
1190 * it can be safely saved at the stack.
1192 struct ftrace_graph_ent_entry saved;
1193 trace_assign_type(field, entry);
1195 return print_graph_entry(&saved, s, iter, flags);
1197 case TRACE_GRAPH_RET: {
1198 struct ftrace_graph_ret_entry *field;
1199 trace_assign_type(field, entry);
1200 return print_graph_return(&field->ret, s, entry, iter, flags);
1204 /* dont trace stack and functions as comments */
1205 return TRACE_TYPE_UNHANDLED;
1208 return print_graph_comment(s, entry, iter, flags);
1211 return TRACE_TYPE_HANDLED;
1214 static enum print_line_t
1215 print_graph_function(struct trace_iterator *iter)
1217 return print_graph_function_flags(iter, tracer_flags.val);
1220 static enum print_line_t
1221 print_graph_function_event(struct trace_iterator *iter, int flags,
1222 struct trace_event *event)
1224 return print_graph_function(iter);
1227 static void print_lat_header(struct seq_file *s, u32 flags)
1229 static const char spaces[] = " " /* 16 spaces */
1231 " "; /* 17 spaces */
1234 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1236 if (flags & TRACE_GRAPH_PRINT_CPU)
1238 if (flags & TRACE_GRAPH_PRINT_PROC)
1241 seq_printf(s, "#%.*s _-----=> irqs-off \n", size, spaces);
1242 seq_printf(s, "#%.*s / _----=> need-resched \n", size, spaces);
1243 seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1244 seq_printf(s, "#%.*s|| / _--=> preempt-depth \n", size, spaces);
1245 seq_printf(s, "#%.*s||| / \n", size, spaces);
1248 static void __print_graph_headers_flags(struct seq_file *s, u32 flags)
1250 int lat = trace_flags & TRACE_ITER_LATENCY_FMT;
1253 print_lat_header(s, flags);
1257 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1258 seq_puts(s, " TIME ");
1259 if (flags & TRACE_GRAPH_PRINT_CPU)
1260 seq_puts(s, " CPU");
1261 if (flags & TRACE_GRAPH_PRINT_PROC)
1262 seq_puts(s, " TASK/PID ");
1264 seq_puts(s, "||||");
1265 if (flags & TRACE_GRAPH_PRINT_DURATION)
1266 seq_puts(s, " DURATION ");
1267 seq_puts(s, " FUNCTION CALLS\n");
1271 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1273 if (flags & TRACE_GRAPH_PRINT_CPU)
1275 if (flags & TRACE_GRAPH_PRINT_PROC)
1276 seq_puts(s, " | | ");
1278 seq_puts(s, "||||");
1279 if (flags & TRACE_GRAPH_PRINT_DURATION)
1280 seq_puts(s, " | | ");
1281 seq_puts(s, " | | | |\n");
1284 static void print_graph_headers(struct seq_file *s)
1286 print_graph_headers_flags(s, tracer_flags.val);
1289 void print_graph_headers_flags(struct seq_file *s, u32 flags)
1291 struct trace_iterator *iter = s->private;
1293 if (!(trace_flags & TRACE_ITER_CONTEXT_INFO))
1296 if (trace_flags & TRACE_ITER_LATENCY_FMT) {
1297 /* print nothing if the buffers are empty */
1298 if (trace_empty(iter))
1301 print_trace_header(s, iter);
1304 __print_graph_headers_flags(s, flags);
1307 void graph_trace_open(struct trace_iterator *iter)
1309 /* pid and depth on the last trace processed */
1310 struct fgraph_data *data;
1314 iter->private = NULL;
1316 /* We can be called in atomic context via ftrace_dump() */
1317 gfpflags = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
1319 data = kzalloc(sizeof(*data), gfpflags);
1323 data->cpu_data = alloc_percpu_gfp(struct fgraph_cpu_data, gfpflags);
1324 if (!data->cpu_data)
1327 for_each_possible_cpu(cpu) {
1328 pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1329 int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1330 int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1331 int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1339 iter->private = data;
1346 pr_warning("function graph tracer: not enough memory\n");
1349 void graph_trace_close(struct trace_iterator *iter)
1351 struct fgraph_data *data = iter->private;
1354 free_percpu(data->cpu_data);
1360 func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1362 if (bit == TRACE_GRAPH_PRINT_IRQS)
1363 ftrace_graph_skip_irqs = !set;
1368 static struct trace_event_functions graph_functions = {
1369 .trace = print_graph_function_event,
1372 static struct trace_event graph_trace_entry_event = {
1373 .type = TRACE_GRAPH_ENT,
1374 .funcs = &graph_functions,
1377 static struct trace_event graph_trace_ret_event = {
1378 .type = TRACE_GRAPH_RET,
1379 .funcs = &graph_functions
1382 static struct tracer graph_trace __tracer_data = {
1383 .name = "function_graph",
1384 .update_thresh = graph_trace_update_thresh,
1385 .open = graph_trace_open,
1386 .pipe_open = graph_trace_open,
1387 .close = graph_trace_close,
1388 .pipe_close = graph_trace_close,
1389 .init = graph_trace_init,
1390 .reset = graph_trace_reset,
1391 .print_line = print_graph_function,
1392 .print_header = print_graph_headers,
1393 .flags = &tracer_flags,
1394 .set_flag = func_graph_set_flag,
1395 #ifdef CONFIG_FTRACE_SELFTEST
1396 .selftest = trace_selftest_startup_function_graph,
1402 graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1408 ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1420 graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1423 char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1426 n = sprintf(buf, "%d\n", max_depth);
1428 return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1431 static const struct file_operations graph_depth_fops = {
1432 .open = tracing_open_generic,
1433 .write = graph_depth_write,
1434 .read = graph_depth_read,
1435 .llseek = generic_file_llseek,
1438 static __init int init_graph_tracefs(void)
1440 struct dentry *d_tracer;
1442 d_tracer = tracing_init_dentry();
1443 if (IS_ERR(d_tracer))
1446 trace_create_file("max_graph_depth", 0644, d_tracer,
1447 NULL, &graph_depth_fops);
1451 fs_initcall(init_graph_tracefs);
1453 static __init int init_graph_trace(void)
1455 max_bytes_for_cpu = snprintf(NULL, 0, "%d", nr_cpu_ids - 1);
1457 if (!register_trace_event(&graph_trace_entry_event)) {
1458 pr_warning("Warning: could not register graph trace events\n");
1462 if (!register_trace_event(&graph_trace_ret_event)) {
1463 pr_warning("Warning: could not register graph trace events\n");
1467 return register_tracer(&graph_trace);
1470 core_initcall(init_graph_trace);