[PATCH] kprobes: Enable tracing for mololithic kernel images

[Masami Hiramatsu (Google)]

On Wed, 8 Jun 2022 11:19:19 -0700
Song Liu <song at kernel.org> wrote:

> On Wed, Jun 8, 2022 at 9:28 AM Ard Biesheuvel <ardb at kernel.org> wrote:
> >
> > Hello Jarkko,
> >
> > On Wed, 8 Jun 2022 at 02:02, Jarkko Sakkinen <jarkko at profian.com> wrote:
> > >
> > > Tracing with kprobes while running a monolithic kernel is currently
> > > impossible because CONFIG_KPROBES is dependent of CONFIG_MODULES.  This
> > > dependency is a result of kprobes code using the module allocator for the
> > > trampoline code.
> > >
> > > Detaching kprobes from modules helps to squeeze down the user space,
> > > e.g. when developing new core kernel features, while still having all
> > > the nice tracing capabilities.
> > >
> > > For kernel/ and arch/*, move module_alloc() and module_memfree() to
> > > module_alloc.c, and compile as part of vmlinux when either CONFIG_MODULES
> > > or CONFIG_KPROBES is enabled.  In addition, flag kernel module specific
> > > code with CONFIG_MODULES.
> > >
> > > As the result, kprobes can be used with a monolithic kernel.
> >
> > I think I may have mentioned this the previous time as well, but I
> > don't think this is the right approach.
> >
> > Kprobes uses alloc_insn_page() to allocate executable memory, but the
> > requirements for this memory are radically different compared to
> > loadable modules, which need to be within an arch-specific distance of
> > the core kernel, need KASAN backing etc etc.
> 
> I think the distance of core kernel requirement is the same for kprobe
> alloc_insn_page and modules, no?

This strongly depends on how kprobes (software breakpoint and
single-step) is implemented on the arch. For example, x86 implements
the so-called "kprobe-booster" which jumps back from the single
stepping trampoline buffer. Then the buffer address must be within
the range where it can jump to the original address.
However, if the arch implements single-step as an instruction
emulation, it has no such limitation. As far as I know, arm64
will do emulation for the instructions which change PC register
and will do direct execution with another software breakpoint
for other instructions.

Why I'm using module_alloc() for a generic function, is that
can cover the limitation most widely.
Thus, if we have CONFIG_ARCH_HAVE_ALLOC_INSN_PAGE flag and
kprobes can check it instead of using __weak function, the
kprobes may not need to depend on module_alloc() in general.

Thank you,

> 
> Thanks,
> Song
> 
> >
> > This is why arm64, for instance, does not implement alloc_insn_page()
> > in terms of module_alloc() [and likely does not belong in this patch
> > for that reason]
> 
> 
> 
> >
> > Is there any reason kprobes cannot simply use vmalloc()?
> >


-- 
Masami Hiramatsu (Google) <mhiramat at kernel.org>