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/debugfs.h>
10 #include <linux/uaccess.h>
11 #include <linux/ftrace.h>
12 #include <linux/slab.h>
16 #include "trace_output.h"
18 /* When set, irq functions will be ignored */
19 static int ftrace_graph_skip_irqs;
21 struct fgraph_cpu_data {
26 unsigned long enter_funcs[FTRACE_RETFUNC_DEPTH];
30 struct fgraph_cpu_data __percpu *cpu_data;
32 /* Place to preserve last processed entry. */
33 struct ftrace_graph_ent_entry ent;
34 struct ftrace_graph_ret_entry ret;
39 #define TRACE_GRAPH_INDENT 2
41 static unsigned int max_depth;
43 static struct tracer_opt trace_opts[] = {
44 /* Display overruns? (for self-debug purpose) */
45 { TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
47 { TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
48 /* Display Overhead ? */
49 { TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
50 /* Display proc name/pid */
51 { TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
52 /* Display duration of execution */
53 { TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
54 /* Display absolute time of an entry */
55 { TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
56 /* Display interrupts */
57 { TRACER_OPT(funcgraph-irqs, TRACE_GRAPH_PRINT_IRQS) },
58 /* Display function name after trailing } */
59 { TRACER_OPT(funcgraph-tail, TRACE_GRAPH_PRINT_TAIL) },
63 static struct tracer_flags tracer_flags = {
64 /* Don't display overruns, proc, or tail by default */
65 .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
66 TRACE_GRAPH_PRINT_DURATION | TRACE_GRAPH_PRINT_IRQS,
70 static struct trace_array *graph_array;
73 * DURATION column is being also used to display IRQ signs,
74 * following values are used by print_graph_irq and others
75 * to fill in space into DURATION column.
78 FLAGS_FILL_FULL = 1 << TRACE_GRAPH_PRINT_FILL_SHIFT,
79 FLAGS_FILL_START = 2 << TRACE_GRAPH_PRINT_FILL_SHIFT,
80 FLAGS_FILL_END = 3 << TRACE_GRAPH_PRINT_FILL_SHIFT,
83 static enum print_line_t
84 print_graph_duration(unsigned long long duration, struct trace_seq *s,
87 /* Add a function return address to the trace stack on thread info.*/
89 ftrace_push_return_trace(unsigned long ret, unsigned long func, int *depth,
90 unsigned long frame_pointer)
92 unsigned long long calltime;
95 if (!current->ret_stack)
99 * We must make sure the ret_stack is tested before we read
104 /* The return trace stack is full */
105 if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
106 atomic_inc(¤t->trace_overrun);
111 * The curr_ret_stack is an index to ftrace return stack of
112 * current task. Its value should be in [0, FTRACE_RETFUNC_
113 * DEPTH) when the function graph tracer is used. To support
114 * filtering out specific functions, it makes the index
115 * negative by subtracting huge value (FTRACE_NOTRACE_DEPTH)
116 * so when it sees a negative index the ftrace will ignore
117 * the record. And the index gets recovered when returning
118 * from the filtered function by adding the FTRACE_NOTRACE_
119 * DEPTH and then it'll continue to record functions normally.
121 * The curr_ret_stack is initialized to -1 and get increased
122 * in this function. So it can be less than -1 only if it was
123 * filtered out via ftrace_graph_notrace_addr() which can be
124 * set from set_graph_notrace file in debugfs by user.
126 if (current->curr_ret_stack < -1)
129 calltime = trace_clock_local();
131 index = ++current->curr_ret_stack;
132 if (ftrace_graph_notrace_addr(func))
133 current->curr_ret_stack -= FTRACE_NOTRACE_DEPTH;
135 current->ret_stack[index].ret = ret;
136 current->ret_stack[index].func = func;
137 current->ret_stack[index].calltime = calltime;
138 current->ret_stack[index].subtime = 0;
139 current->ret_stack[index].fp = frame_pointer;
140 *depth = current->curr_ret_stack;
145 /* Retrieve a function return address to the trace stack on thread info.*/
147 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
148 unsigned long frame_pointer)
152 index = current->curr_ret_stack;
155 * A negative index here means that it's just returned from a
156 * notrace'd function. Recover index to get an original
157 * return address. See ftrace_push_return_trace().
159 * TODO: Need to check whether the stack gets corrupted.
162 index += FTRACE_NOTRACE_DEPTH;
164 if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
167 /* Might as well panic, otherwise we have no where to go */
168 *ret = (unsigned long)panic;
172 #if defined(CONFIG_HAVE_FUNCTION_GRAPH_FP_TEST) && !defined(CC_USING_FENTRY)
174 * The arch may choose to record the frame pointer used
175 * and check it here to make sure that it is what we expect it
176 * to be. If gcc does not set the place holder of the return
177 * address in the frame pointer, and does a copy instead, then
178 * the function graph trace will fail. This test detects this
181 * Currently, x86_32 with optimize for size (-Os) makes the latest
184 * Note, -mfentry does not use frame pointers, and this test
185 * is not needed if CC_USING_FENTRY is set.
187 if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
189 WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
190 " from func %ps return to %lx\n",
191 current->ret_stack[index].fp,
193 (void *)current->ret_stack[index].func,
194 current->ret_stack[index].ret);
195 *ret = (unsigned long)panic;
200 *ret = current->ret_stack[index].ret;
201 trace->func = current->ret_stack[index].func;
202 trace->calltime = current->ret_stack[index].calltime;
203 trace->overrun = atomic_read(¤t->trace_overrun);
204 trace->depth = index;
208 * Send the trace to the ring-buffer.
209 * @return the original return address.
211 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
213 struct ftrace_graph_ret trace;
216 ftrace_pop_return_trace(&trace, &ret, frame_pointer);
217 trace.rettime = trace_clock_local();
219 current->curr_ret_stack--;
221 * The curr_ret_stack can be less than -1 only if it was
222 * filtered out and it's about to return from the function.
223 * Recover the index and continue to trace normal functions.
225 if (current->curr_ret_stack < -1) {
226 current->curr_ret_stack += FTRACE_NOTRACE_DEPTH;
231 * The trace should run after decrementing the ret counter
232 * in case an interrupt were to come in. We don't want to
233 * lose the interrupt if max_depth is set.
235 ftrace_graph_return(&trace);
237 if (unlikely(!ret)) {
240 /* Might as well panic. What else to do? */
241 ret = (unsigned long)panic;
247 int __trace_graph_entry(struct trace_array *tr,
248 struct ftrace_graph_ent *trace,
252 struct ftrace_event_call *call = &event_funcgraph_entry;
253 struct ring_buffer_event *event;
254 struct ring_buffer *buffer = tr->trace_buffer.buffer;
255 struct ftrace_graph_ent_entry *entry;
257 if (unlikely(__this_cpu_read(ftrace_cpu_disabled)))
260 event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
261 sizeof(*entry), flags, pc);
264 entry = ring_buffer_event_data(event);
265 entry->graph_ent = *trace;
266 if (!call_filter_check_discard(call, entry, buffer, event))
267 __buffer_unlock_commit(buffer, event);
272 static inline int ftrace_graph_ignore_irqs(void)
274 if (!ftrace_graph_skip_irqs || trace_recursion_test(TRACE_IRQ_BIT))
280 int trace_graph_entry(struct ftrace_graph_ent *trace)
282 struct trace_array *tr = graph_array;
283 struct trace_array_cpu *data;
290 if (!ftrace_trace_task(current))
293 /* trace it when it is-nested-in or is a function enabled. */
294 if ((!(trace->depth || ftrace_graph_addr(trace->func)) ||
295 ftrace_graph_ignore_irqs()) || (trace->depth < 0) ||
296 (max_depth && trace->depth >= max_depth))
300 * Do not trace a function if it's filtered by set_graph_notrace.
301 * Make the index of ret stack negative to indicate that it should
302 * ignore further functions. But it needs its own ret stack entry
303 * to recover the original index in order to continue tracing after
304 * returning from the function.
306 if (ftrace_graph_notrace_addr(trace->func))
309 local_irq_save(flags);
310 cpu = raw_smp_processor_id();
311 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
312 disabled = atomic_inc_return(&data->disabled);
313 if (likely(disabled == 1)) {
314 pc = preempt_count();
315 ret = __trace_graph_entry(tr, trace, flags, pc);
320 atomic_dec(&data->disabled);
321 local_irq_restore(flags);
326 int trace_graph_thresh_entry(struct ftrace_graph_ent *trace)
331 return trace_graph_entry(trace);
335 __trace_graph_function(struct trace_array *tr,
336 unsigned long ip, unsigned long flags, int pc)
338 u64 time = trace_clock_local();
339 struct ftrace_graph_ent ent = {
343 struct ftrace_graph_ret ret = {
350 __trace_graph_entry(tr, &ent, flags, pc);
351 __trace_graph_return(tr, &ret, flags, pc);
355 trace_graph_function(struct trace_array *tr,
356 unsigned long ip, unsigned long parent_ip,
357 unsigned long flags, int pc)
359 __trace_graph_function(tr, ip, flags, pc);
362 void __trace_graph_return(struct trace_array *tr,
363 struct ftrace_graph_ret *trace,
367 struct ftrace_event_call *call = &event_funcgraph_exit;
368 struct ring_buffer_event *event;
369 struct ring_buffer *buffer = tr->trace_buffer.buffer;
370 struct ftrace_graph_ret_entry *entry;
372 if (unlikely(__this_cpu_read(ftrace_cpu_disabled)))
375 event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_RET,
376 sizeof(*entry), flags, pc);
379 entry = ring_buffer_event_data(event);
381 if (!call_filter_check_discard(call, entry, buffer, event))
382 __buffer_unlock_commit(buffer, event);
385 void trace_graph_return(struct ftrace_graph_ret *trace)
387 struct trace_array *tr = graph_array;
388 struct trace_array_cpu *data;
394 local_irq_save(flags);
395 cpu = raw_smp_processor_id();
396 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
397 disabled = atomic_inc_return(&data->disabled);
398 if (likely(disabled == 1)) {
399 pc = preempt_count();
400 __trace_graph_return(tr, trace, flags, pc);
402 atomic_dec(&data->disabled);
403 local_irq_restore(flags);
406 void set_graph_array(struct trace_array *tr)
410 /* Make graph_array visible before we start tracing */
415 void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
417 if (tracing_thresh &&
418 (trace->rettime - trace->calltime < tracing_thresh))
421 trace_graph_return(trace);
424 static int graph_trace_init(struct trace_array *tr)
430 ret = register_ftrace_graph(&trace_graph_thresh_return,
431 &trace_graph_thresh_entry);
433 ret = register_ftrace_graph(&trace_graph_return,
437 tracing_start_cmdline_record();
442 static void graph_trace_reset(struct trace_array *tr)
444 tracing_stop_cmdline_record();
445 unregister_ftrace_graph();
448 static int max_bytes_for_cpu;
450 static enum print_line_t
451 print_graph_cpu(struct trace_seq *s, int cpu)
456 * Start with a space character - to make it stand out
457 * to the right a bit when trace output is pasted into
460 ret = trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
462 return TRACE_TYPE_PARTIAL_LINE;
464 return TRACE_TYPE_HANDLED;
467 #define TRACE_GRAPH_PROCINFO_LENGTH 14
469 static enum print_line_t
470 print_graph_proc(struct trace_seq *s, pid_t pid)
472 char comm[TASK_COMM_LEN];
473 /* sign + log10(MAX_INT) + '\0' */
480 trace_find_cmdline(pid, comm);
482 sprintf(pid_str, "%d", pid);
484 /* 1 stands for the "-" character */
485 len = strlen(comm) + strlen(pid_str) + 1;
487 if (len < TRACE_GRAPH_PROCINFO_LENGTH)
488 spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
490 /* First spaces to align center */
491 for (i = 0; i < spaces / 2; i++) {
492 ret = trace_seq_putc(s, ' ');
494 return TRACE_TYPE_PARTIAL_LINE;
497 ret = trace_seq_printf(s, "%s-%s", comm, pid_str);
499 return TRACE_TYPE_PARTIAL_LINE;
501 /* Last spaces to align center */
502 for (i = 0; i < spaces - (spaces / 2); i++) {
503 ret = trace_seq_putc(s, ' ');
505 return TRACE_TYPE_PARTIAL_LINE;
507 return TRACE_TYPE_HANDLED;
511 static enum print_line_t
512 print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
514 if (!trace_seq_putc(s, ' '))
517 return trace_print_lat_fmt(s, entry);
520 /* If the pid changed since the last trace, output this event */
521 static enum print_line_t
522 verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
529 return TRACE_TYPE_HANDLED;
531 last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
533 if (*last_pid == pid)
534 return TRACE_TYPE_HANDLED;
536 prev_pid = *last_pid;
540 return TRACE_TYPE_HANDLED;
542 * Context-switch trace line:
544 ------------------------------------------
545 | 1) migration/0--1 => sshd-1755
546 ------------------------------------------
549 ret = trace_seq_puts(s,
550 " ------------------------------------------\n");
552 return TRACE_TYPE_PARTIAL_LINE;
554 ret = print_graph_cpu(s, cpu);
555 if (ret == TRACE_TYPE_PARTIAL_LINE)
556 return TRACE_TYPE_PARTIAL_LINE;
558 ret = print_graph_proc(s, prev_pid);
559 if (ret == TRACE_TYPE_PARTIAL_LINE)
560 return TRACE_TYPE_PARTIAL_LINE;
562 ret = trace_seq_puts(s, " => ");
564 return TRACE_TYPE_PARTIAL_LINE;
566 ret = print_graph_proc(s, pid);
567 if (ret == TRACE_TYPE_PARTIAL_LINE)
568 return TRACE_TYPE_PARTIAL_LINE;
570 ret = trace_seq_puts(s,
571 "\n ------------------------------------------\n\n");
573 return TRACE_TYPE_PARTIAL_LINE;
575 return TRACE_TYPE_HANDLED;
578 static struct ftrace_graph_ret_entry *
579 get_return_for_leaf(struct trace_iterator *iter,
580 struct ftrace_graph_ent_entry *curr)
582 struct fgraph_data *data = iter->private;
583 struct ring_buffer_iter *ring_iter = NULL;
584 struct ring_buffer_event *event;
585 struct ftrace_graph_ret_entry *next;
588 * If the previous output failed to write to the seq buffer,
589 * then we just reuse the data from before.
591 if (data && data->failed) {
596 ring_iter = trace_buffer_iter(iter, iter->cpu);
598 /* First peek to compare current entry and the next one */
600 event = ring_buffer_iter_peek(ring_iter, NULL);
603 * We need to consume the current entry to see
606 ring_buffer_consume(iter->trace_buffer->buffer, iter->cpu,
608 event = ring_buffer_peek(iter->trace_buffer->buffer, iter->cpu,
615 next = ring_buffer_event_data(event);
619 * Save current and next entries for later reference
620 * if the output fails.
624 * If the next event is not a return type, then
625 * we only care about what type it is. Otherwise we can
626 * safely copy the entire event.
628 if (next->ent.type == TRACE_GRAPH_RET)
631 data->ret.ent.type = next->ent.type;
635 if (next->ent.type != TRACE_GRAPH_RET)
638 if (curr->ent.pid != next->ent.pid ||
639 curr->graph_ent.func != next->ret.func)
642 /* this is a leaf, now advance the iterator */
644 ring_buffer_read(ring_iter, NULL);
649 static int print_graph_abs_time(u64 t, struct trace_seq *s)
651 unsigned long usecs_rem;
653 usecs_rem = do_div(t, NSEC_PER_SEC);
656 return trace_seq_printf(s, "%5lu.%06lu | ",
657 (unsigned long)t, usecs_rem);
660 static enum print_line_t
661 print_graph_irq(struct trace_iterator *iter, unsigned long addr,
662 enum trace_type type, int cpu, pid_t pid, u32 flags)
665 struct trace_seq *s = &iter->seq;
667 if (addr < (unsigned long)__irqentry_text_start ||
668 addr >= (unsigned long)__irqentry_text_end)
669 return TRACE_TYPE_UNHANDLED;
671 if (trace_flags & TRACE_ITER_CONTEXT_INFO) {
673 if (flags & TRACE_GRAPH_PRINT_ABS_TIME) {
674 ret = print_graph_abs_time(iter->ts, s);
676 return TRACE_TYPE_PARTIAL_LINE;
680 if (flags & TRACE_GRAPH_PRINT_CPU) {
681 ret = print_graph_cpu(s, cpu);
682 if (ret == TRACE_TYPE_PARTIAL_LINE)
683 return TRACE_TYPE_PARTIAL_LINE;
687 if (flags & TRACE_GRAPH_PRINT_PROC) {
688 ret = print_graph_proc(s, pid);
689 if (ret == TRACE_TYPE_PARTIAL_LINE)
690 return TRACE_TYPE_PARTIAL_LINE;
691 ret = trace_seq_puts(s, " | ");
693 return TRACE_TYPE_PARTIAL_LINE;
698 ret = print_graph_duration(0, s, flags | FLAGS_FILL_START);
699 if (ret != TRACE_TYPE_HANDLED)
702 if (type == TRACE_GRAPH_ENT)
703 ret = trace_seq_puts(s, "==========>");
705 ret = trace_seq_puts(s, "<==========");
708 return TRACE_TYPE_PARTIAL_LINE;
710 ret = print_graph_duration(0, s, flags | FLAGS_FILL_END);
711 if (ret != TRACE_TYPE_HANDLED)
714 ret = trace_seq_putc(s, '\n');
717 return TRACE_TYPE_PARTIAL_LINE;
718 return TRACE_TYPE_HANDLED;
722 trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
724 unsigned long nsecs_rem = do_div(duration, 1000);
725 /* log10(ULONG_MAX) + '\0' */
731 sprintf(msecs_str, "%lu", (unsigned long) duration);
734 ret = trace_seq_printf(s, "%s", msecs_str);
736 return TRACE_TYPE_PARTIAL_LINE;
738 len = strlen(msecs_str);
740 /* Print nsecs (we don't want to exceed 7 numbers) */
742 size_t slen = min_t(size_t, sizeof(nsecs_str), 8UL - len);
744 snprintf(nsecs_str, slen, "%03lu", nsecs_rem);
745 ret = trace_seq_printf(s, ".%s", nsecs_str);
747 return TRACE_TYPE_PARTIAL_LINE;
748 len += strlen(nsecs_str);
751 ret = trace_seq_puts(s, " us ");
753 return TRACE_TYPE_PARTIAL_LINE;
755 /* Print remaining spaces to fit the row's width */
756 for (i = len; i < 7; i++) {
757 ret = trace_seq_putc(s, ' ');
759 return TRACE_TYPE_PARTIAL_LINE;
761 return TRACE_TYPE_HANDLED;
764 static enum print_line_t
765 print_graph_duration(unsigned long long duration, struct trace_seq *s,
770 if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
771 !(trace_flags & TRACE_ITER_CONTEXT_INFO))
772 return TRACE_TYPE_HANDLED;
774 /* No real adata, just filling the column with spaces */
775 switch (flags & TRACE_GRAPH_PRINT_FILL_MASK) {
776 case FLAGS_FILL_FULL:
777 ret = trace_seq_puts(s, " | ");
778 return ret ? TRACE_TYPE_HANDLED : TRACE_TYPE_PARTIAL_LINE;
779 case FLAGS_FILL_START:
780 ret = trace_seq_puts(s, " ");
781 return ret ? TRACE_TYPE_HANDLED : TRACE_TYPE_PARTIAL_LINE;
783 ret = trace_seq_puts(s, " |");
784 return ret ? TRACE_TYPE_HANDLED : TRACE_TYPE_PARTIAL_LINE;
787 /* Signal a overhead of time execution to the output */
788 if (flags & TRACE_GRAPH_PRINT_OVERHEAD) {
789 /* Duration exceeded 100 msecs */
790 if (duration > 100000ULL)
791 ret = trace_seq_puts(s, "! ");
792 /* Duration exceeded 10 msecs */
793 else if (duration > 10000ULL)
794 ret = trace_seq_puts(s, "+ ");
798 * The -1 means we either did not exceed the duration tresholds
799 * or we dont want to print out the overhead. Either way we need
800 * to fill out the space.
803 ret = trace_seq_puts(s, " ");
805 /* Catching here any failure happenned above */
807 return TRACE_TYPE_PARTIAL_LINE;
809 ret = trace_print_graph_duration(duration, s);
810 if (ret != TRACE_TYPE_HANDLED)
813 ret = trace_seq_puts(s, "| ");
815 return TRACE_TYPE_PARTIAL_LINE;
817 return TRACE_TYPE_HANDLED;
820 /* Case of a leaf function on its call entry */
821 static enum print_line_t
822 print_graph_entry_leaf(struct trace_iterator *iter,
823 struct ftrace_graph_ent_entry *entry,
824 struct ftrace_graph_ret_entry *ret_entry,
825 struct trace_seq *s, u32 flags)
827 struct fgraph_data *data = iter->private;
828 struct ftrace_graph_ret *graph_ret;
829 struct ftrace_graph_ent *call;
830 unsigned long long duration;
834 graph_ret = &ret_entry->ret;
835 call = &entry->graph_ent;
836 duration = graph_ret->rettime - graph_ret->calltime;
839 struct fgraph_cpu_data *cpu_data;
842 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
845 * Comments display at + 1 to depth. Since
846 * this is a leaf function, keep the comments
847 * equal to this depth.
849 cpu_data->depth = call->depth - 1;
851 /* No need to keep this function around for this depth */
852 if (call->depth < FTRACE_RETFUNC_DEPTH)
853 cpu_data->enter_funcs[call->depth] = 0;
856 /* Overhead and duration */
857 ret = print_graph_duration(duration, s, flags);
858 if (ret == TRACE_TYPE_PARTIAL_LINE)
859 return TRACE_TYPE_PARTIAL_LINE;
862 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++) {
863 ret = trace_seq_putc(s, ' ');
865 return TRACE_TYPE_PARTIAL_LINE;
868 ret = trace_seq_printf(s, "%ps();\n", (void *)call->func);
870 return TRACE_TYPE_PARTIAL_LINE;
872 return TRACE_TYPE_HANDLED;
875 static enum print_line_t
876 print_graph_entry_nested(struct trace_iterator *iter,
877 struct ftrace_graph_ent_entry *entry,
878 struct trace_seq *s, int cpu, u32 flags)
880 struct ftrace_graph_ent *call = &entry->graph_ent;
881 struct fgraph_data *data = iter->private;
886 struct fgraph_cpu_data *cpu_data;
889 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
890 cpu_data->depth = call->depth;
892 /* Save this function pointer to see if the exit matches */
893 if (call->depth < FTRACE_RETFUNC_DEPTH)
894 cpu_data->enter_funcs[call->depth] = call->func;
898 ret = print_graph_duration(0, s, flags | FLAGS_FILL_FULL);
899 if (ret != TRACE_TYPE_HANDLED)
903 for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++) {
904 ret = trace_seq_putc(s, ' ');
906 return TRACE_TYPE_PARTIAL_LINE;
909 ret = trace_seq_printf(s, "%ps() {\n", (void *)call->func);
911 return TRACE_TYPE_PARTIAL_LINE;
914 * we already consumed the current entry to check the next one
915 * and see if this is a leaf.
917 return TRACE_TYPE_NO_CONSUME;
920 static enum print_line_t
921 print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
922 int type, unsigned long addr, u32 flags)
924 struct fgraph_data *data = iter->private;
925 struct trace_entry *ent = iter->ent;
930 if (verif_pid(s, ent->pid, cpu, data) == TRACE_TYPE_PARTIAL_LINE)
931 return TRACE_TYPE_PARTIAL_LINE;
935 ret = print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
936 if (ret == TRACE_TYPE_PARTIAL_LINE)
937 return TRACE_TYPE_PARTIAL_LINE;
940 if (!(trace_flags & TRACE_ITER_CONTEXT_INFO))
944 if (flags & TRACE_GRAPH_PRINT_ABS_TIME) {
945 ret = print_graph_abs_time(iter->ts, s);
947 return TRACE_TYPE_PARTIAL_LINE;
951 if (flags & TRACE_GRAPH_PRINT_CPU) {
952 ret = print_graph_cpu(s, cpu);
953 if (ret == TRACE_TYPE_PARTIAL_LINE)
954 return TRACE_TYPE_PARTIAL_LINE;
958 if (flags & TRACE_GRAPH_PRINT_PROC) {
959 ret = print_graph_proc(s, ent->pid);
960 if (ret == TRACE_TYPE_PARTIAL_LINE)
961 return TRACE_TYPE_PARTIAL_LINE;
963 ret = trace_seq_puts(s, " | ");
965 return TRACE_TYPE_PARTIAL_LINE;
969 if (trace_flags & TRACE_ITER_LATENCY_FMT) {
970 ret = print_graph_lat_fmt(s, ent);
971 if (ret == TRACE_TYPE_PARTIAL_LINE)
972 return TRACE_TYPE_PARTIAL_LINE;
979 * Entry check for irq code
982 * - we are inside irq code
983 * - we just entered irq code
986 * - funcgraph-interrupts option is set
987 * - we are not inside irq code
990 check_irq_entry(struct trace_iterator *iter, u32 flags,
991 unsigned long addr, int depth)
995 struct fgraph_data *data = iter->private;
998 * If we are either displaying irqs, or we got called as
999 * a graph event and private data does not exist,
1000 * then we bypass the irq check.
1002 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
1006 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1009 * We are inside the irq code
1011 if (*depth_irq >= 0)
1014 if ((addr < (unsigned long)__irqentry_text_start) ||
1015 (addr >= (unsigned long)__irqentry_text_end))
1019 * We are entering irq code.
1026 * Return check for irq code
1029 * - we are inside irq code
1030 * - we just left irq code
1033 * - funcgraph-interrupts option is set
1034 * - we are not inside irq code
1037 check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
1039 int cpu = iter->cpu;
1041 struct fgraph_data *data = iter->private;
1044 * If we are either displaying irqs, or we got called as
1045 * a graph event and private data does not exist,
1046 * then we bypass the irq check.
1048 if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
1052 depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1055 * We are not inside the irq code.
1057 if (*depth_irq == -1)
1061 * We are inside the irq code, and this is returning entry.
1062 * Let's not trace it and clear the entry depth, since
1063 * we are out of irq code.
1065 * This condition ensures that we 'leave the irq code' once
1066 * we are out of the entry depth. Thus protecting us from
1067 * the RETURN entry loss.
1069 if (*depth_irq >= depth) {
1075 * We are inside the irq code, and this is not the entry.
1080 static enum print_line_t
1081 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
1082 struct trace_iterator *iter, u32 flags)
1084 struct fgraph_data *data = iter->private;
1085 struct ftrace_graph_ent *call = &field->graph_ent;
1086 struct ftrace_graph_ret_entry *leaf_ret;
1087 static enum print_line_t ret;
1088 int cpu = iter->cpu;
1090 if (check_irq_entry(iter, flags, call->func, call->depth))
1091 return TRACE_TYPE_HANDLED;
1093 if (print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags))
1094 return TRACE_TYPE_PARTIAL_LINE;
1096 leaf_ret = get_return_for_leaf(iter, field);
1098 ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
1100 ret = print_graph_entry_nested(iter, field, s, cpu, flags);
1104 * If we failed to write our output, then we need to make
1105 * note of it. Because we already consumed our entry.
1117 static enum print_line_t
1118 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
1119 struct trace_entry *ent, struct trace_iterator *iter,
1122 unsigned long long duration = trace->rettime - trace->calltime;
1123 struct fgraph_data *data = iter->private;
1124 pid_t pid = ent->pid;
1125 int cpu = iter->cpu;
1130 if (check_irq_return(iter, flags, trace->depth))
1131 return TRACE_TYPE_HANDLED;
1134 struct fgraph_cpu_data *cpu_data;
1135 int cpu = iter->cpu;
1137 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
1140 * Comments display at + 1 to depth. This is the
1141 * return from a function, we now want the comments
1142 * to display at the same level of the bracket.
1144 cpu_data->depth = trace->depth - 1;
1146 if (trace->depth < FTRACE_RETFUNC_DEPTH) {
1147 if (cpu_data->enter_funcs[trace->depth] != trace->func)
1149 cpu_data->enter_funcs[trace->depth] = 0;
1153 if (print_graph_prologue(iter, s, 0, 0, flags))
1154 return TRACE_TYPE_PARTIAL_LINE;
1156 /* Overhead and duration */
1157 ret = print_graph_duration(duration, s, flags);
1158 if (ret == TRACE_TYPE_PARTIAL_LINE)
1159 return TRACE_TYPE_PARTIAL_LINE;
1162 for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++) {
1163 ret = trace_seq_putc(s, ' ');
1165 return TRACE_TYPE_PARTIAL_LINE;
1169 * If the return function does not have a matching entry,
1170 * then the entry was lost. Instead of just printing
1171 * the '}' and letting the user guess what function this
1172 * belongs to, write out the function name. Always do
1173 * that if the funcgraph-tail option is enabled.
1175 if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL)) {
1176 ret = trace_seq_puts(s, "}\n");
1178 return TRACE_TYPE_PARTIAL_LINE;
1180 ret = trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
1182 return TRACE_TYPE_PARTIAL_LINE;
1186 if (flags & TRACE_GRAPH_PRINT_OVERRUN) {
1187 ret = trace_seq_printf(s, " (Overruns: %lu)\n",
1190 return TRACE_TYPE_PARTIAL_LINE;
1193 ret = print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
1195 if (ret == TRACE_TYPE_PARTIAL_LINE)
1196 return TRACE_TYPE_PARTIAL_LINE;
1198 return TRACE_TYPE_HANDLED;
1201 static enum print_line_t
1202 print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
1203 struct trace_iterator *iter, u32 flags)
1205 unsigned long sym_flags = (trace_flags & TRACE_ITER_SYM_MASK);
1206 struct fgraph_data *data = iter->private;
1207 struct trace_event *event;
1213 depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
1215 if (print_graph_prologue(iter, s, 0, 0, flags))
1216 return TRACE_TYPE_PARTIAL_LINE;
1219 ret = print_graph_duration(0, s, flags | FLAGS_FILL_FULL);
1220 if (ret != TRACE_TYPE_HANDLED)
1225 for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++) {
1226 ret = trace_seq_putc(s, ' ');
1228 return TRACE_TYPE_PARTIAL_LINE;
1232 ret = trace_seq_puts(s, "/* ");
1234 return TRACE_TYPE_PARTIAL_LINE;
1236 switch (iter->ent->type) {
1238 ret = trace_print_bprintk_msg_only(iter);
1239 if (ret != TRACE_TYPE_HANDLED)
1243 ret = trace_print_printk_msg_only(iter);
1244 if (ret != TRACE_TYPE_HANDLED)
1248 event = ftrace_find_event(ent->type);
1250 return TRACE_TYPE_UNHANDLED;
1252 ret = event->funcs->trace(iter, sym_flags, event);
1253 if (ret != TRACE_TYPE_HANDLED)
1257 /* Strip ending newline */
1258 if (s->buffer[s->len - 1] == '\n') {
1259 s->buffer[s->len - 1] = '\0';
1263 ret = trace_seq_puts(s, " */\n");
1265 return TRACE_TYPE_PARTIAL_LINE;
1267 return TRACE_TYPE_HANDLED;
1272 print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1274 struct ftrace_graph_ent_entry *field;
1275 struct fgraph_data *data = iter->private;
1276 struct trace_entry *entry = iter->ent;
1277 struct trace_seq *s = &iter->seq;
1278 int cpu = iter->cpu;
1281 if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1282 per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1283 return TRACE_TYPE_HANDLED;
1287 * If the last output failed, there's a possibility we need
1288 * to print out the missing entry which would never go out.
1290 if (data && data->failed) {
1292 iter->cpu = data->cpu;
1293 ret = print_graph_entry(field, s, iter, flags);
1294 if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1295 per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1296 ret = TRACE_TYPE_NO_CONSUME;
1302 switch (entry->type) {
1303 case TRACE_GRAPH_ENT: {
1305 * print_graph_entry() may consume the current event,
1306 * thus @field may become invalid, so we need to save it.
1307 * sizeof(struct ftrace_graph_ent_entry) is very small,
1308 * it can be safely saved at the stack.
1310 struct ftrace_graph_ent_entry saved;
1311 trace_assign_type(field, entry);
1313 return print_graph_entry(&saved, s, iter, flags);
1315 case TRACE_GRAPH_RET: {
1316 struct ftrace_graph_ret_entry *field;
1317 trace_assign_type(field, entry);
1318 return print_graph_return(&field->ret, s, entry, iter, flags);
1322 /* dont trace stack and functions as comments */
1323 return TRACE_TYPE_UNHANDLED;
1326 return print_graph_comment(s, entry, iter, flags);
1329 return TRACE_TYPE_HANDLED;
1332 static enum print_line_t
1333 print_graph_function(struct trace_iterator *iter)
1335 return print_graph_function_flags(iter, tracer_flags.val);
1338 static enum print_line_t
1339 print_graph_function_event(struct trace_iterator *iter, int flags,
1340 struct trace_event *event)
1342 return print_graph_function(iter);
1345 static void print_lat_header(struct seq_file *s, u32 flags)
1347 static const char spaces[] = " " /* 16 spaces */
1349 " "; /* 17 spaces */
1352 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1354 if (flags & TRACE_GRAPH_PRINT_CPU)
1356 if (flags & TRACE_GRAPH_PRINT_PROC)
1359 seq_printf(s, "#%.*s _-----=> irqs-off \n", size, spaces);
1360 seq_printf(s, "#%.*s / _----=> need-resched \n", size, spaces);
1361 seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1362 seq_printf(s, "#%.*s|| / _--=> preempt-depth \n", size, spaces);
1363 seq_printf(s, "#%.*s||| / \n", size, spaces);
1366 static void __print_graph_headers_flags(struct seq_file *s, u32 flags)
1368 int lat = trace_flags & TRACE_ITER_LATENCY_FMT;
1371 print_lat_header(s, flags);
1375 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1376 seq_printf(s, " TIME ");
1377 if (flags & TRACE_GRAPH_PRINT_CPU)
1378 seq_printf(s, " CPU");
1379 if (flags & TRACE_GRAPH_PRINT_PROC)
1380 seq_printf(s, " TASK/PID ");
1382 seq_printf(s, "||||");
1383 if (flags & TRACE_GRAPH_PRINT_DURATION)
1384 seq_printf(s, " DURATION ");
1385 seq_printf(s, " FUNCTION CALLS\n");
1389 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1390 seq_printf(s, " | ");
1391 if (flags & TRACE_GRAPH_PRINT_CPU)
1392 seq_printf(s, " | ");
1393 if (flags & TRACE_GRAPH_PRINT_PROC)
1394 seq_printf(s, " | | ");
1396 seq_printf(s, "||||");
1397 if (flags & TRACE_GRAPH_PRINT_DURATION)
1398 seq_printf(s, " | | ");
1399 seq_printf(s, " | | | |\n");
1402 void print_graph_headers(struct seq_file *s)
1404 print_graph_headers_flags(s, tracer_flags.val);
1407 void print_graph_headers_flags(struct seq_file *s, u32 flags)
1409 struct trace_iterator *iter = s->private;
1411 if (!(trace_flags & TRACE_ITER_CONTEXT_INFO))
1414 if (trace_flags & TRACE_ITER_LATENCY_FMT) {
1415 /* print nothing if the buffers are empty */
1416 if (trace_empty(iter))
1419 print_trace_header(s, iter);
1422 __print_graph_headers_flags(s, flags);
1425 void graph_trace_open(struct trace_iterator *iter)
1427 /* pid and depth on the last trace processed */
1428 struct fgraph_data *data;
1431 iter->private = NULL;
1433 data = kzalloc(sizeof(*data), GFP_KERNEL);
1437 data->cpu_data = alloc_percpu(struct fgraph_cpu_data);
1438 if (!data->cpu_data)
1441 for_each_possible_cpu(cpu) {
1442 pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1443 int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1444 int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1445 int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1453 iter->private = data;
1460 pr_warning("function graph tracer: not enough memory\n");
1463 void graph_trace_close(struct trace_iterator *iter)
1465 struct fgraph_data *data = iter->private;
1468 free_percpu(data->cpu_data);
1474 func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1476 if (bit == TRACE_GRAPH_PRINT_IRQS)
1477 ftrace_graph_skip_irqs = !set;
1482 static struct trace_event_functions graph_functions = {
1483 .trace = print_graph_function_event,
1486 static struct trace_event graph_trace_entry_event = {
1487 .type = TRACE_GRAPH_ENT,
1488 .funcs = &graph_functions,
1491 static struct trace_event graph_trace_ret_event = {
1492 .type = TRACE_GRAPH_RET,
1493 .funcs = &graph_functions
1496 static struct tracer graph_trace __tracer_data = {
1497 .name = "function_graph",
1498 .open = graph_trace_open,
1499 .pipe_open = graph_trace_open,
1500 .close = graph_trace_close,
1501 .pipe_close = graph_trace_close,
1502 .init = graph_trace_init,
1503 .reset = graph_trace_reset,
1504 .print_line = print_graph_function,
1505 .print_header = print_graph_headers,
1506 .flags = &tracer_flags,
1507 .set_flag = func_graph_set_flag,
1508 #ifdef CONFIG_FTRACE_SELFTEST
1509 .selftest = trace_selftest_startup_function_graph,
1515 graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1521 ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1533 graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1536 char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1539 n = sprintf(buf, "%d\n", max_depth);
1541 return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1544 static const struct file_operations graph_depth_fops = {
1545 .open = tracing_open_generic,
1546 .write = graph_depth_write,
1547 .read = graph_depth_read,
1548 .llseek = generic_file_llseek,
1551 static __init int init_graph_debugfs(void)
1553 struct dentry *d_tracer;
1555 d_tracer = tracing_init_dentry();
1559 trace_create_file("max_graph_depth", 0644, d_tracer,
1560 NULL, &graph_depth_fops);
1564 fs_initcall(init_graph_debugfs);
1566 static __init int init_graph_trace(void)
1568 max_bytes_for_cpu = snprintf(NULL, 0, "%d", nr_cpu_ids - 1);
1570 if (!register_ftrace_event(&graph_trace_entry_event)) {
1571 pr_warning("Warning: could not register graph trace events\n");
1575 if (!register_ftrace_event(&graph_trace_ret_event)) {
1576 pr_warning("Warning: could not register graph trace events\n");
1580 return register_tracer(&graph_trace);
1583 core_initcall(init_graph_trace);