tracing/function-graph-tracer: various fixes and features
authorFrederic Weisbecker <fweisbec@gmail.com>
Fri, 23 Jan 2009 01:04:53 +0000 (17:04 -0800)
committerIngo Molnar <mingo@elte.hu>
Fri, 23 Jan 2009 10:18:08 +0000 (11:18 +0100)
This patch brings various bugfixes:

- Drop the first irrelevant task switch on the very beginning of a trace.

- Drop the OVERHEAD word from the headers, the DURATION word is sufficient
  and will not overlap other columns.

- Make the headers fit well their respective columns whatever the
  selected options.

Ie, default options:

 # tracer: function_graph
 #
 # CPU  DURATION                  FUNCTION CALLS
 # |     |   |                     |   |   |   |

  1)   0.646 us    |                    }
  1)               |                    mem_cgroup_del_lru_list() {
  1)   0.624 us    |                      lookup_page_cgroup();
  1)   1.970 us    |                    }

 echo funcgraph-proc > trace_options

 # tracer: function_graph
 #
 # CPU  TASK/PID        DURATION                  FUNCTION CALLS
 # |    |    |           |   |                     |   |   |   |

  0)   bash-2937    |   0.895 us    |                }
  0)   bash-2937    |   0.888 us    |                __rcu_read_unlock();
  0)   bash-2937    |   0.864 us    |                conv_uni_to_pc();
  0)   bash-2937    |   1.015 us    |                __rcu_read_lock();

 echo nofuncgraph-cpu > trace_options
 echo nofuncgraph-proc > trace_options

 # tracer: function_graph
 #
 #   DURATION                  FUNCTION CALLS
 #    |   |                     |   |   |   |

   3.752 us    |                  native_pud_val();
   0.616 us    |                  native_pud_val();
   0.624 us    |                  native_pmd_val();

About features, one can now disable the duration (this will hide the
overhead too for convenient reasons and because on  doesn't need
overhead if it hasn't the duration):

 echo nofuncgraph-duration > trace_options

 # tracer: function_graph
 #
 #                FUNCTION CALLS
 #                |   |   |   |

           cap_vm_enough_memory() {
             __vm_enough_memory() {
               vm_acct_memory();
             }
           }
         }

And at last, an option to print the absolute time:

 //Restart from default options
 echo funcgraph-abstime > trace_options

 # tracer: function_graph
 #
 #      TIME       CPU  DURATION                  FUNCTION CALLS
 #       |         |     |   |                     |   |   |   |

   261.339774 |   1) + 42.823 us   |    }
   261.339775 |   1)   1.045 us    |    _spin_lock_irq();
   261.339777 |   1)   0.940 us    |    _spin_lock_irqsave();
   261.339778 |   1)   0.752 us    |    _spin_unlock_irqrestore();
   261.339780 |   1)   0.857 us    |    _spin_unlock_irq();
   261.339782 |   1)               |    flush_to_ldisc() {
   261.339783 |   1)               |      tty_ldisc_ref() {
   261.339783 |   1)               |        tty_ldisc_try() {
   261.339784 |   1)   1.075 us    |          _spin_lock_irqsave();
   261.339786 |   1)   0.842 us    |          _spin_unlock_irqrestore();
   261.339788 |   1)   4.211 us    |        }
   261.339788 |   1)   5.662 us    |      }

The format is seconds.usecs.

I guess no one needs the nanosec precision here, the main goal is to have
an overview about the general timings of events, and to see the place when
the trace switches from one cpu to another.

ie:

   274.874760 |   1)   0.676 us    |      _spin_unlock();
   274.874762 |   1)   0.609 us    |      native_load_sp0();
   274.874763 |   1)   0.602 us    |      native_load_tls();
   274.878739 |   0)   0.722 us    |                  }
   274.878740 |   0)   0.714 us    |                  native_pmd_val();
   274.878741 |   0)   0.730 us    |                  native_pmd_val();

Here there is a 4000 usecs difference when we switch the cpu.

Changes in V2:

- Completely fix the first pointless task switch.

Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/trace/trace_functions_graph.c

index 3c545984816f35654b5fb8a72735bec75e9950f8..66fc7b806a38ea7d1b73c446012338d95e7148b2 100644 (file)
@@ -1,7 +1,7 @@
 /*
  *
  * Function graph tracer.
- * Copyright (c) 2008 Frederic Weisbecker <fweisbec@gmail.com>
+ * Copyright (c) 2008-2009 Frederic Weisbecker <fweisbec@gmail.com>
  * Mostly borrowed from function tracer which
  * is Copyright (c) Steven Rostedt <srostedt@redhat.com>
  *
 #define TRACE_GRAPH_PRINT_CPU          0x2
 #define TRACE_GRAPH_PRINT_OVERHEAD     0x4
 #define TRACE_GRAPH_PRINT_PROC         0x8
+#define TRACE_GRAPH_PRINT_DURATION     0x10
+#define TRACE_GRAPH_PRINT_ABS_TIME     0X20
 
 static struct tracer_opt trace_opts[] = {
-       /* Display overruns ? */
+       /* Display overruns? (for self-debug purpose) */
        { TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
        /* Display CPU ? */
        { TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
@@ -31,17 +33,22 @@ static struct tracer_opt trace_opts[] = {
        { TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
        /* Display proc name/pid */
        { TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
+       /* Display duration of execution */
+       { TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
+       /* Display absolute time of an entry */
+       { TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
        { } /* Empty entry */
 };
 
 static struct tracer_flags tracer_flags = {
        /* Don't display overruns and proc by default */
-       .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD,
+       .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
+              TRACE_GRAPH_PRINT_DURATION,
        .opts = trace_opts
 };
 
 /* pid on the last trace processed */
-static pid_t last_pid[NR_CPUS] = { [0 ... NR_CPUS-1] = -1 };
+
 
 static int graph_trace_init(struct trace_array *tr)
 {
@@ -154,17 +161,25 @@ print_graph_proc(struct trace_seq *s, pid_t pid)
 
 /* If the pid changed since the last trace, output this event */
 static enum print_line_t
-verif_pid(struct trace_seq *s, pid_t pid, int cpu)
+verif_pid(struct trace_seq *s, pid_t pid, int cpu, pid_t *last_pids_cpu)
 {
        pid_t prev_pid;
+       pid_t *last_pid;
        int ret;
 
-       if (last_pid[cpu] != -1 && last_pid[cpu] == pid)
+       if (!last_pids_cpu)
+               return TRACE_TYPE_HANDLED;
+
+       last_pid = per_cpu_ptr(last_pids_cpu, cpu);
+
+       if (*last_pid == pid)
                return TRACE_TYPE_HANDLED;
 
-       prev_pid = last_pid[cpu];
-       last_pid[cpu] = pid;
+       prev_pid = *last_pid;
+       *last_pid = pid;
 
+       if (prev_pid == -1)
+               return TRACE_TYPE_HANDLED;
 /*
  * Context-switch trace line:
 
@@ -232,9 +247,34 @@ trace_branch_is_leaf(struct trace_iterator *iter,
        return true;
 }
 
+/* Signal a overhead of time execution to the output */
+static int
+print_graph_overhead(unsigned long long duration, struct trace_seq *s)
+{
+       /* If duration disappear, we don't need anything */
+       if (!(tracer_flags.val & TRACE_GRAPH_PRINT_DURATION))
+               return 1;
+
+       /* Non nested entry or return */
+       if (duration == -1)
+               return trace_seq_printf(s, "  ");
+
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_OVERHEAD) {
+               /* Duration exceeded 100 msecs */
+               if (duration > 100000ULL)
+                       return trace_seq_printf(s, "! ");
+
+               /* Duration exceeded 10 msecs */
+               if (duration > 10000ULL)
+                       return trace_seq_printf(s, "+ ");
+       }
+
+       return trace_seq_printf(s, "  ");
+}
+
 static enum print_line_t
 print_graph_irq(struct trace_seq *s, unsigned long addr,
-                               enum trace_type type, int cpu, pid_t pid)
+               enum trace_type type, int cpu, pid_t pid)
 {
        int ret;
 
@@ -242,35 +282,40 @@ print_graph_irq(struct trace_seq *s, unsigned long addr,
                addr >= (unsigned long)__irqentry_text_end)
                return TRACE_TYPE_UNHANDLED;
 
-       if (type == TRACE_GRAPH_ENT) {
-               ret = trace_seq_printf(s, "==========> |  ");
-       } else {
-               /* Cpu */
-               if (tracer_flags.val & TRACE_GRAPH_PRINT_CPU) {
-                       ret = print_graph_cpu(s, cpu);
-                       if (ret == TRACE_TYPE_PARTIAL_LINE)
-                               return TRACE_TYPE_PARTIAL_LINE;
-               }
-               /* Proc */
-               if (tracer_flags.val & TRACE_GRAPH_PRINT_PROC) {
-                       ret = print_graph_proc(s, pid);
-                       if (ret == TRACE_TYPE_PARTIAL_LINE)
-                               return TRACE_TYPE_PARTIAL_LINE;
+       /* Cpu */
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_CPU) {
+               ret = print_graph_cpu(s, cpu);
+               if (ret == TRACE_TYPE_PARTIAL_LINE)
+                       return TRACE_TYPE_PARTIAL_LINE;
+       }
+       /* Proc */
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_PROC) {
+               ret = print_graph_proc(s, pid);
+               if (ret == TRACE_TYPE_PARTIAL_LINE)
+                       return TRACE_TYPE_PARTIAL_LINE;
+               ret = trace_seq_printf(s, " | ");
+               if (!ret)
+                       return TRACE_TYPE_PARTIAL_LINE;
+       }
 
-                       ret = trace_seq_printf(s, " | ");
-                       if (!ret)
-                               return TRACE_TYPE_PARTIAL_LINE;
-               }
+       /* No overhead */
+       ret = print_graph_overhead(-1, s);
+       if (!ret)
+               return TRACE_TYPE_PARTIAL_LINE;
 
-               /* No overhead */
-               if (tracer_flags.val & TRACE_GRAPH_PRINT_OVERHEAD) {
-                       ret = trace_seq_printf(s, "  ");
-                       if (!ret)
-                               return TRACE_TYPE_PARTIAL_LINE;
-               }
+       if (type == TRACE_GRAPH_ENT)
+               ret = trace_seq_printf(s, "==========>");
+       else
+               ret = trace_seq_printf(s, "<==========");
+
+       if (!ret)
+               return TRACE_TYPE_PARTIAL_LINE;
+
+       /* Don't close the duration column if haven't one */
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_DURATION)
+               trace_seq_printf(s, " |");
+       ret = trace_seq_printf(s, "\n");
 
-               ret = trace_seq_printf(s, "<========== |\n");
-       }
        if (!ret)
                return TRACE_TYPE_PARTIAL_LINE;
        return TRACE_TYPE_HANDLED;
@@ -289,7 +334,7 @@ print_graph_duration(unsigned long long duration, struct trace_seq *s)
        sprintf(msecs_str, "%lu", (unsigned long) duration);
 
        /* Print msecs */
-       ret = trace_seq_printf(s, msecs_str);
+       ret = trace_seq_printf(s, "%s", msecs_str);
        if (!ret)
                return TRACE_TYPE_PARTIAL_LINE;
 
@@ -322,19 +367,15 @@ print_graph_duration(unsigned long long duration, struct trace_seq *s)
 
 }
 
-/* Signal a overhead of time execution to the output */
-static int
-print_graph_overhead(unsigned long long duration, struct trace_seq *s)
+static int print_graph_abs_time(u64 t, struct trace_seq *s)
 {
-       /* Duration exceeded 100 msecs */
-       if (duration > 100000ULL)
-               return trace_seq_printf(s, "! ");
+       unsigned long usecs_rem;
 
-       /* Duration exceeded 10 msecs */
-       if (duration > 10000ULL)
-               return trace_seq_printf(s, "+ ");
+       usecs_rem = do_div(t, 1000000000);
+       usecs_rem /= 1000;
 
-       return trace_seq_printf(s, "  ");
+       return trace_seq_printf(s, "%5lu.%06lu |  ",
+                       (unsigned long)t, usecs_rem);
 }
 
 /* Case of a leaf function on its call entry */
@@ -357,16 +398,16 @@ print_graph_entry_leaf(struct trace_iterator *iter,
        duration = graph_ret->rettime - graph_ret->calltime;
 
        /* Overhead */
-       if (tracer_flags.val & TRACE_GRAPH_PRINT_OVERHEAD) {
-               ret = print_graph_overhead(duration, s);
-               if (!ret)
-                       return TRACE_TYPE_PARTIAL_LINE;
-       }
+       ret = print_graph_overhead(duration, s);
+       if (!ret)
+               return TRACE_TYPE_PARTIAL_LINE;
 
        /* Duration */
-       ret = print_graph_duration(duration, s);
-       if (ret == TRACE_TYPE_PARTIAL_LINE)
-               return TRACE_TYPE_PARTIAL_LINE;
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_DURATION) {
+               ret = print_graph_duration(duration, s);
+               if (ret == TRACE_TYPE_PARTIAL_LINE)
+                       return TRACE_TYPE_PARTIAL_LINE;
+       }
 
        /* Function */
        for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++) {
@@ -395,25 +436,17 @@ print_graph_entry_nested(struct ftrace_graph_ent_entry *entry,
        struct ftrace_graph_ent *call = &entry->graph_ent;
 
        /* No overhead */
-       if (tracer_flags.val & TRACE_GRAPH_PRINT_OVERHEAD) {
-               ret = trace_seq_printf(s, "  ");
-               if (!ret)
-                       return TRACE_TYPE_PARTIAL_LINE;
-       }
+       ret = print_graph_overhead(-1, s);
+       if (!ret)
+               return TRACE_TYPE_PARTIAL_LINE;
 
-       /* Interrupt */
-       ret = print_graph_irq(s, call->func, TRACE_GRAPH_ENT, cpu, pid);
-       if (ret == TRACE_TYPE_UNHANDLED) {
-               /* No time */
+       /* No time */
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_DURATION) {
                ret = trace_seq_printf(s, "            |  ");
                if (!ret)
                        return TRACE_TYPE_PARTIAL_LINE;
-       } else {
-               if (ret == TRACE_TYPE_PARTIAL_LINE)
-                       return TRACE_TYPE_PARTIAL_LINE;
        }
 
-
        /* Function */
        for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++) {
                ret = trace_seq_printf(s, " ");
@@ -434,15 +467,30 @@ print_graph_entry_nested(struct ftrace_graph_ent_entry *entry,
 
 static enum print_line_t
 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
-                       struct trace_iterator *iter, int cpu)
+                       struct trace_iterator *iter)
 {
        int ret;
+       int cpu = iter->cpu;
+       pid_t *last_entry = iter->private;
        struct trace_entry *ent = iter->ent;
+       struct ftrace_graph_ent *call = &field->graph_ent;
 
        /* Pid */
-       if (verif_pid(s, ent->pid, cpu) == TRACE_TYPE_PARTIAL_LINE)
+       if (verif_pid(s, ent->pid, cpu, last_entry) == TRACE_TYPE_PARTIAL_LINE)
+               return TRACE_TYPE_PARTIAL_LINE;
+
+       /* Interrupt */
+       ret = print_graph_irq(s, call->func, TRACE_GRAPH_ENT, cpu, ent->pid);
+       if (ret == TRACE_TYPE_PARTIAL_LINE)
                return TRACE_TYPE_PARTIAL_LINE;
 
+       /* Absolute time */
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_ABS_TIME) {
+               ret = print_graph_abs_time(iter->ts, s);
+               if (!ret)
+                       return TRACE_TYPE_PARTIAL_LINE;
+       }
+
        /* Cpu */
        if (tracer_flags.val & TRACE_GRAPH_PRINT_CPU) {
                ret = print_graph_cpu(s, cpu);
@@ -470,16 +518,25 @@ print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
 
 static enum print_line_t
 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
-                  struct trace_entry *ent, int cpu)
+                  struct trace_entry *ent, struct trace_iterator *iter)
 {
        int i;
        int ret;
+       int cpu = iter->cpu;
+       pid_t *last_pid = iter->private;
        unsigned long long duration = trace->rettime - trace->calltime;
 
        /* Pid */
-       if (verif_pid(s, ent->pid, cpu) == TRACE_TYPE_PARTIAL_LINE)
+       if (verif_pid(s, ent->pid, cpu, last_pid) == TRACE_TYPE_PARTIAL_LINE)
                return TRACE_TYPE_PARTIAL_LINE;
 
+       /* Absolute time */
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_ABS_TIME) {
+               ret = print_graph_abs_time(iter->ts, s);
+               if (!ret)
+                       return TRACE_TYPE_PARTIAL_LINE;
+       }
+
        /* Cpu */
        if (tracer_flags.val & TRACE_GRAPH_PRINT_CPU) {
                ret = print_graph_cpu(s, cpu);
@@ -499,16 +556,16 @@ print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
        }
 
        /* Overhead */
-       if (tracer_flags.val & TRACE_GRAPH_PRINT_OVERHEAD) {
-               ret = print_graph_overhead(duration, s);
-               if (!ret)
-                       return TRACE_TYPE_PARTIAL_LINE;
-       }
+       ret = print_graph_overhead(duration, s);
+       if (!ret)
+               return TRACE_TYPE_PARTIAL_LINE;
 
        /* Duration */
-       ret = print_graph_duration(duration, s);
-       if (ret == TRACE_TYPE_PARTIAL_LINE)
-               return TRACE_TYPE_PARTIAL_LINE;
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_DURATION) {
+               ret = print_graph_duration(duration, s);
+               if (ret == TRACE_TYPE_PARTIAL_LINE)
+                       return TRACE_TYPE_PARTIAL_LINE;
+       }
 
        /* Closing brace */
        for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++) {
@@ -542,14 +599,23 @@ print_graph_comment(struct print_entry *trace, struct trace_seq *s,
 {
        int i;
        int ret;
+       int cpu = iter->cpu;
+       pid_t *last_pid = iter->private;
+
+       /* Absolute time */
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_ABS_TIME) {
+               ret = print_graph_abs_time(iter->ts, s);
+               if (!ret)
+                       return TRACE_TYPE_PARTIAL_LINE;
+       }
 
        /* Pid */
-       if (verif_pid(s, ent->pid, iter->cpu) == TRACE_TYPE_PARTIAL_LINE)
+       if (verif_pid(s, ent->pid, cpu, last_pid) == TRACE_TYPE_PARTIAL_LINE)
                return TRACE_TYPE_PARTIAL_LINE;
 
        /* Cpu */
        if (tracer_flags.val & TRACE_GRAPH_PRINT_CPU) {
-               ret = print_graph_cpu(s, iter->cpu);
+               ret = print_graph_cpu(s, cpu);
                if (ret == TRACE_TYPE_PARTIAL_LINE)
                        return TRACE_TYPE_PARTIAL_LINE;
        }
@@ -566,17 +632,17 @@ print_graph_comment(struct print_entry *trace, struct trace_seq *s,
        }
 
        /* No overhead */
-       if (tracer_flags.val & TRACE_GRAPH_PRINT_OVERHEAD) {
-               ret = trace_seq_printf(s, "  ");
+       ret = print_graph_overhead(-1, s);
+       if (!ret)
+               return TRACE_TYPE_PARTIAL_LINE;
+
+       /* No time */
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_DURATION) {
+               ret = trace_seq_printf(s, "            |  ");
                if (!ret)
                        return TRACE_TYPE_PARTIAL_LINE;
        }
 
-       /* No time */
-       ret = trace_seq_printf(s, "            |  ");
-       if (!ret)
-               return TRACE_TYPE_PARTIAL_LINE;
-
        /* Indentation */
        if (trace->depth > 0)
                for (i = 0; i < (trace->depth + 1) * TRACE_GRAPH_INDENT; i++) {
@@ -614,13 +680,12 @@ print_graph_function(struct trace_iterator *iter)
        case TRACE_GRAPH_ENT: {
                struct ftrace_graph_ent_entry *field;
                trace_assign_type(field, entry);
-               return print_graph_entry(field, s, iter,
-                                        iter->cpu);
+               return print_graph_entry(field, s, iter);
        }
        case TRACE_GRAPH_RET: {
                struct ftrace_graph_ret_entry *field;
                trace_assign_type(field, entry);
-               return print_graph_return(&field->ret, s, entry, iter->cpu);
+               return print_graph_return(&field->ret, s, entry, iter);
        }
        case TRACE_PRINT: {
                struct print_entry *field;
@@ -636,28 +701,55 @@ static void print_graph_headers(struct seq_file *s)
 {
        /* 1st line */
        seq_printf(s, "# ");
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_ABS_TIME)
+               seq_printf(s, "     TIME       ");
        if (tracer_flags.val & TRACE_GRAPH_PRINT_CPU)
-               seq_printf(s, "CPU ");
+               seq_printf(s, "CPU");
        if (tracer_flags.val & TRACE_GRAPH_PRINT_PROC)
-               seq_printf(s, "TASK/PID     ");
-       if (tracer_flags.val & TRACE_GRAPH_PRINT_OVERHEAD)
-               seq_printf(s, "OVERHEAD/");
-       seq_printf(s, "DURATION            FUNCTION CALLS\n");
+               seq_printf(s, "  TASK/PID      ");
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_DURATION)
+               seq_printf(s, "  DURATION   ");
+       seq_printf(s, "               FUNCTION CALLS\n");
 
        /* 2nd line */
        seq_printf(s, "# ");
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_ABS_TIME)
+               seq_printf(s, "      |         ");
        if (tracer_flags.val & TRACE_GRAPH_PRINT_CPU)
-               seq_printf(s, "|   ");
+               seq_printf(s, "|  ");
        if (tracer_flags.val & TRACE_GRAPH_PRINT_PROC)
-               seq_printf(s, "|      |     ");
-       if (tracer_flags.val & TRACE_GRAPH_PRINT_OVERHEAD) {
-               seq_printf(s, "|        ");
-               seq_printf(s, "|                   |   |   |   |\n");
-       } else
-               seq_printf(s, "    |               |   |   |   |\n");
+               seq_printf(s, "  |    |        ");
+       if (tracer_flags.val & TRACE_GRAPH_PRINT_DURATION)
+               seq_printf(s, "   |   |      ");
+       seq_printf(s, "               |   |   |   |\n");
 }
+
+static void graph_trace_open(struct trace_iterator *iter)
+{
+       /* pid on the last trace processed */
+       pid_t *last_pid = alloc_percpu(pid_t);
+       int cpu;
+
+       if (!last_pid)
+               pr_warning("function graph tracer: not enough memory\n");
+       else
+               for_each_possible_cpu(cpu) {
+                       pid_t *pid = per_cpu_ptr(last_pid, cpu);
+                       *pid = -1;
+               }
+
+       iter->private = last_pid;
+}
+
+static void graph_trace_close(struct trace_iterator *iter)
+{
+       percpu_free(iter->private);
+}
+
 static struct tracer graph_trace __read_mostly = {
        .name           = "function_graph",
+       .open           = graph_trace_open,
+       .close          = graph_trace_close,
        .init           = graph_trace_init,
        .reset          = graph_trace_reset,
        .print_line     = print_graph_function,