On Thu, 2008-08-07 at 09:38 -0400, Vivek Goyal wrote:
[...]
> What kind of problem we run into if we don't disable the ftracer?
>
> I think there are too many #ifdefs now and probably we can at least
> get rid if #ifdef CONFIG_FTRACE thing.
>
> I think ftracer needs to export the function to enable the tracer
> back (tracer_enable()) so that we don't directly play with ftrace_enabled
> variable. tracer_enable() can be do {} while{0} in case of CONFIG_FTRACE=n
> so that we can get rid of #ifdefs here.
The ftracer issue for kexec is reported by Dhaval Giani and fixed by
Ingo as in following thread:
http://lkml.org/lkml/2008/2/19/175
After some testing, I found that if we enable ftrace before
restore_processor_state(), system will hang. I think maybe ftracer
depends on some processor state that we destroyed during kexec and
restored by restore_processor_state(). So I move save_processor_state()
and restore_processor_state() into machine_kexec() and enable ftrace
after restore_processor_state().
The #ifdef CONFIG_FTRACE should be removed. I think an interface like
irq_save/restore is good for this.
saved_ftrace_enabled = ftrace_save_enabled()
<...>
ftrace_restore_enabled(saved_ftrace_enabled)
Best Regards,
Huang Ying
