function-graph: Add tracing_thresh support to function_graph tracer
Tim Bird [Thu, 25 Feb 2010 23:36:43 +0000 (15:36 -0800)]
Add support for tracing_thresh to the function_graph tracer.  This
version of this feature isolates the checks into new entry and
return functions, to avoid adding more conditional code into the
main function_graph paths.

When the tracing_thresh is set and the function graph tracer is
enabled, only the functions that took longer than the time in
microseconds that was set in tracing_thresh are recorded. To do this
efficiently, only the function exits are recorded:

 [tracing]# echo 100 > tracing_thresh
 [tracing]# echo function_graph > current_tracer
 [tracing]# cat trace
 # tracer: function_graph
 #
 # CPU  DURATION                  FUNCTION CALLS
 # |     |   |                     |   |   |   |
  1) ! 119.214 us  |  } /* smp_apic_timer_interrupt */
  1)   <========== |
  0) ! 101.527 us  |              } /* __rcu_process_callbacks */
  0) ! 126.461 us  |            } /* rcu_process_callbacks */
  0) ! 145.111 us  |          } /* __do_softirq */
  0) ! 149.667 us  |        } /* do_softirq */
  0) ! 168.817 us  |      } /* irq_exit */
  0) ! 248.254 us  |    } /* smp_apic_timer_interrupt */

Also, add support for specifying tracing_thresh on the kernel
command line.  When used like so: "tracing_thresh=200 ftrace=function_graph"
this can be used to analyse system startup.  It is important to disable
tracing soon after boot, in order to avoid losing the trace data.

Acked-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Tim Bird <tim.bird@am.sony.com>
LKML-Reference: <4B87098B.4040308@am.sony.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

kernel/trace/trace.c
kernel/trace/trace.h
kernel/trace/trace_functions_graph.c

index 6efd5cb..ababedb 100644 (file)
@@ -374,6 +374,21 @@ static int __init set_buf_size(char *str)
 }
 __setup("trace_buf_size=", set_buf_size);
 
+static int __init set_tracing_thresh(char *str)
+{
+       unsigned long threshhold;
+       int ret;
+
+       if (!str)
+               return 0;
+       ret = strict_strtoul(str, 0, &threshhold);
+       if (ret < 0)
+               return 0;
+       tracing_thresh = threshhold * 1000;
+       return 1;
+}
+__setup("tracing_thresh=", set_tracing_thresh);
+
 unsigned long nsecs_to_usecs(unsigned long nsecs)
 {
        return nsecs / 1000;
@@ -579,9 +594,10 @@ static ssize_t trace_seq_to_buffer(struct trace_seq *s, void *buf, size_t cnt)
 static arch_spinlock_t ftrace_max_lock =
        (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
 
+unsigned long __read_mostly    tracing_thresh;
+
 #ifdef CONFIG_TRACER_MAX_TRACE
 unsigned long __read_mostly    tracing_max_latency;
-unsigned long __read_mostly    tracing_thresh;
 
 /*
  * Copy the new maximum trace into the separate maximum-trace
@@ -4248,10 +4264,10 @@ static __init int tracer_init_debugfs(void)
 #ifdef CONFIG_TRACER_MAX_TRACE
        trace_create_file("tracing_max_latency", 0644, d_tracer,
                        &tracing_max_latency, &tracing_max_lat_fops);
+#endif
 
        trace_create_file("tracing_thresh", 0644, d_tracer,
                        &tracing_thresh, &tracing_max_lat_fops);
-#endif
 
        trace_create_file("README", 0444, d_tracer,
                        NULL, &tracing_readme_fops);
index fd05bca..1bc8cd1 100644 (file)
@@ -396,9 +396,10 @@ extern int process_new_ksym_entry(char *ksymname, int op, unsigned long addr);
 
 extern unsigned long nsecs_to_usecs(unsigned long nsecs);
 
+extern unsigned long tracing_thresh;
+
 #ifdef CONFIG_TRACER_MAX_TRACE
 extern unsigned long tracing_max_latency;
-extern unsigned long tracing_thresh;
 
 void update_max_tr(struct trace_array *tr, struct task_struct *tsk, int cpu);
 void update_max_tr_single(struct trace_array *tr,
index 7b1f246..e9df04b 100644 (file)
@@ -237,6 +237,14 @@ int trace_graph_entry(struct ftrace_graph_ent *trace)
        return ret;
 }
 
+int trace_graph_thresh_entry(struct ftrace_graph_ent *trace)
+{
+       if (tracing_thresh)
+               return 1;
+       else
+               return trace_graph_entry(trace);
+}
+
 static void __trace_graph_return(struct trace_array *tr,
                                struct ftrace_graph_ret *trace,
                                unsigned long flags,
@@ -290,13 +298,26 @@ void set_graph_array(struct trace_array *tr)
        smp_mb();
 }
 
+void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
+{
+       if (tracing_thresh &&
+           (trace->rettime - trace->calltime < tracing_thresh))
+               return;
+       else
+               trace_graph_return(trace);
+}
+
 static int graph_trace_init(struct trace_array *tr)
 {
        int ret;
 
        set_graph_array(tr);
-       ret = register_ftrace_graph(&trace_graph_return,
-                                   &trace_graph_entry);
+       if (tracing_thresh)
+               ret = register_ftrace_graph(&trace_graph_thresh_return,
+                                           &trace_graph_thresh_entry);
+       else
+               ret = register_ftrace_graph(&trace_graph_return,
+                                           &trace_graph_entry);
        if (ret)
                return ret;
        tracing_start_cmdline_record();