Re: [PATCHv3] arm: ftrace: Adds support for CONFIG_DYNAMIC_FTRACE_WITH_REGS

From: Abel Vesa
Date: Thu Feb 09 2017 - 14:11:04 EST


On Thu, Feb 09, 2017 at 07:01:10PM +0000, Abel Vesa wrote:
> On Thu, Feb 09, 2017 at 01:14:52PM -0500, Steven Rostedt wrote:
> >
> > [ sending again with Masami Cc'd ]
> >
> > On Thu, 9 Feb 2017 13:14:14 -0500
> > Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> >
> > > On Thu, 9 Feb 2017 18:06:44 +0000
> > > Russell King - ARM Linux <linux@xxxxxxxxxxxxxxx> wrote:
> > >
> > > > On Thu, Feb 09, 2017 at 12:13:22PM -0500, Steven Rostedt wrote:
> > > > > Then came along live kernel patching, which I believe this series is
> > > > > trying to support. What is needed by pt_regs is a way to "hijack" the
> > > > > function being called to instead call the patched function. That is,
> > > > > ftrace is not being used for tracing, but in reality, being used to
> > > > > modify the running kernel. It is being used to change what function
> > > > > gets called. ftrace is just a hook for that mechanism.
> > > >
> > > > So, would I be correct to assume that the only parts of pt_regs that
> > > > would be touched are those which contain arguments to the function,
> > > > and the register which would contain the return value?
> > > >
> > >
> > > For live kernel patching, perhaps.
> > >
> > > But for kprobes, I think they can touch anything. Matters what the
> > > creater of the kprobe wanted to do.
> > >
> Thing is, by saving all of them is the easiest way to ensure that the
> whole context is the same when the replacing function gets called, as
> I said before.
>
> We can't be sure that while __ftrace_ops_list_func is executing, any of
> the regs will have the value they had when the function-to-be-replaced
> was called. That's the reason I say we need to save them all.
Scratch that, I'm wrong, the reason is stupid. The context gets restored anyway after
__ftrace_ops_list_func is done.
> > > -- Steve
> >