On 11/20/2014 04:09 PM, Steven Rostedt wrote:
> On Thu, 13 Nov 2014 09:23:45 +0000
> Markos Chandras <Markos.Chandras@xxxxxxxxxx> wrote:
>
>> Hi,
>>
>> I am trying to understand why ftrace function graph doesn't work when
>> using static ftrace on MIPS. So, what happens when I do 'echo
>> function_graph > current_tracer' is that the ftrace_graph_caller from
>> mcount.S is executed once. The function that called it is
>> 'core_kernel_data()' from __register_ftrace_function in
>> kernel/trace/ftrace.c
>>
>> However, this is the only function that is reported in the trace file
>>
>> # cat trace
>> # tracer: function_graph
>> #
>> # CPU DURATION FUNCTION CALLS
>> # | | | | | | |
>> 0) | core_kernel_data() {
>> 0) 0.000 us | } /* core_kernel_data */
>>
>> The reason that the ftrace_graph_caller is never executed after that is
>> that the following as far as I understand:
>>
>> NESTED(_mcount, PT_SIZE, ra)
>> PTR_LA t1, ftrace_stub
>> PTR_L t2, ftrace_trace_function /* Prepare t2 for (1) */
>> bne t1, t2, static_trace
>> nop
>>
>> #ifdef CONFIG_FUNCTION_GRAPH_TRACER
>> PTR_L t3, ftrace_graph_return
>> bne t1, t3, ftrace_graph_caller
>> nop
>> PTR_LA t1, ftrace_graph_entry_stub
>> PTR_L t3, ftrace_graph_entry
>> bne t1, t3, ftrace_graph_caller
>> nop
>> #endif
>>
>> The previous 3 conditionals exists in arch/mips/kernel/mcount.S.
>> Originally, ftrace_trace_function == ftrace_stub, so the first
>> conditional is not taken and we end up executed the ftrace_graph_caller.
>> All good.
>> However, later on, ftrace_trace_function is set to 'ftrace_ops_no_ops',
>> so the first 'bne' is taken and the ftrace_graph_caller is never
>> executed after that. It is not clear to me if this behaviour is expected
>> so I used QEMU to get a backtrace when ftrace_trace_function is set to
>> ftrace_ops_no_ops.
>
> Looks like the static tracing code in wrong. The function graph code
> should be tested every time.
>
> -- Steve
>
Hi Steven,
I had a look on
https://www.kernel.org/doc/Documentation/trace/ftrace-design.txt
and section "HAVE_FUNCTION_GRAPH_TRACER"
According to the sample code, first we check "ftrace_trace_function !=
ftrace_stub" and if that's true, then we never check the ftrace_graph_*
functions which is exactly what MIPS does. So the graph functions are
not always checked. x86 seems to do something similar in mcount_64.S
--
markos
