Re: [RFC PATCH v2 1/6] fprobe: Use fprobe_regs in fprobe entry handler

[Florent Revest]

On Wed, Aug 9, 2023 at 6:09 PM Florent Revest <revest@xxxxxxxxxxxx> wrote:
>
> On Wed, Aug 9, 2023 at 4:10 PM Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> >
> > Hi Florent,
> >
> > On Wed, 9 Aug 2023 12:28:38 +0200
> > Florent Revest <revest@xxxxxxxxxxxx> wrote:
> >
> > > On Mon, Aug 7, 2023 at 8:48 AM Masami Hiramatsu (Google)
> > > <mhiramat@xxxxxxxxxx> wrote:
> > > >
> > > > From: Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
> > > >
> > > > This allows fprobes to be available with CONFIG_DYNAMIC_FTRACE_WITH_ARGS
> > > > instead of CONFIG_DYNAMIC_FTRACE_WITH_REGS, then we can enable fprobe
> > > > on arm64.
> > >
> > > This patch lets fprobe code build on configs WITH_ARGS and !WITH_REGS
> > > but fprobe wouldn't run on these builds because fprobe still registers
> > > to ftrace with FTRACE_OPS_FL_SAVE_REGS, which would fail on
> > > !WITH_REGS. Shouldn't we also let the fprobe_init callers decide
> > > whether they want REGS or not ?
> >
> > Ah, I think you meant FPROBE_EVENTS? Yes I forgot to add the dependency
> > on it. But fprobe itself can work because fprobe just pass the ftrace_regs
> > to the handlers. (Note that exit callback may not work until next patch)
>
> No, I mean that fprobe still registers its ftrace ops with the
> FTRACE_OPS_FL_SAVE_REGS flag:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/mhiramat/linux.git/tree/kernel/trace/fprobe.c?h=topic/fprobe-ftrace-regs&id=2ca022b2753ae0d2a2513c95f7ed5b5b727fb2c4#n185
>
> Which means that __register_ftrace_function will return -EINVAL on
> builds !WITH_REGS:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/mhiramat/linux.git/tree/kernel/trace/ftrace.c?h=topic/fprobe-ftrace-regs&id=2ca022b2753ae0d2a2513c95f7ed5b5b727fb2c4#n338
>
> As documented here:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/mhiramat/linux.git/tree/include/linux/ftrace.h?h=topic/fprobe-ftrace-regs&id=2ca022b2753ae0d2a2513c95f7ed5b5b727fb2c4#n188
>
> There are two parts to using sparse pt_regs. One is "static": having
> WITH_ARGS in the config, the second one is "dynamic": a ftrace ops
> needs to specify that it doesn't want to go through the ftrace
> trampoline that saves a full pt_regs, by not giving
> FTRACE_OPS_FL_SAVE_REGS. If we want fprobe to work on builds
> !WITH_REGS then we should both remove Kconfig dependencies to
> WITH_REGS (like you've done) but also stop passing this ftrace ops
> flag.

Said in a different way: there are arches that support both WITH_ARGS
and WITH_REGS (like x86 actually). They have two ftrace trampolines
compiled in: ftrace_caller and ftrace_regs_caller, one for each
usecase. If you register to ftrace with the FTRACE_OPS_FL_SAVE_REGS
flag you are telling it that what you want is a pt_regs. If you are
trying to move away from pt_regs and support ftrace_regs in the more
general case (meaning, in the case where it can contain a sparse
pt_regs) then you should stop passing that flag so you go through the
lighter, faster trampoline and test your code in the circumstances
where ftrace_regs isn't just a regular pt_regs but an actually sparse
or light data structure.

I hope that makes my thoughts clearer? It's a hairy topic ahah