Re: arm64 uprobes support showing signs of life

[Pratyush Anand <panand at redhat dot com>]

On Wednesday 07 January 2015 06:40 PM, William Cohen wrote:
> On 01/07/2015 07:07 AM, Pratyush Anand wrote:
> > Hi Will,
> >
> >
> > On Monday 05 January 2015 09:36 PM, William Cohen wrote:
> > > Hi All,
> > >
> > > Pratyush has been implementing uprobes support for aarch64 and has
> > > posted a set of patches
> > > (http://comments.gmane.org/gmane.linux.ports.arm.kernel/382237) for
> > > review.  The really new Linux kernels requires a patch to the
> > > systemtap runtime because the f_dentry macro has been removed.  With
> > > the patched systemtap I was able to run the systemtap testsuite and
> > > get some test results exercising the uprobes support:
> > >
> > > https://web.elastic.org/~dejazilla/viewsummary.php?summary=%3D%27%3C54A9E21C.2080504%40redhat.com%3E%27
> > >
> > > There looked to be a localized fixes for plt support to eliminate the
> > > unsupported systemtap.base/list.exp plt-* tests.
> > >
> > > A number of the tests appear to fail because of userspace arguments
> > > cannot be found for sdt probes.
> > >
> > > The uprobe patches still need some refinement.  On a number of
> > > systemtap "make installcheck" runs the kernel would get stuck spewing
> > > out:
> > >
> > >     Unexpected kernel single-step exception at EL1
> >
> > There are at least two functions of arm64/uprobe implementation uprobe_breakpoint_handler and uprobe_single_step_handler, where a kprobe insertion might be causing above issue. Currently I have qualified these functions with __kprobe, so that one can not insert a kprobe there. With this you should not be able to see the above message.
> >
> > However, I am still investigating if a kprobe insertion be allowed to these functions or any other function which is called directly from debug exception handler (do_debug_exception)
>
> Hi Pratyush,
>
> Are there any other places that need to be protected from kprobe probing in the arm64 code?

yes, I tried inserting kprobe in 
arch/arm64/kernel/debug-monitors.c:brk_handler and it failed. So, I 
think as of now, we can not probe any function which comes in the path 
of do_debug_exception.

Looking into kprobe code, it seems that it allows  REENTER, so it should 
have worked. Do not have much idea at present. Still, studying these 
part of code.

~Pratyush


> > Update code is here:
> >
> > https://github.com/pratyushanand/linux.git : ml_arm64_uprobe_devel_v2
> >
> > ~Pratyush
>
> I will give the new kernel a try.
>
> -Will
>
> > > However, things are looking better for user-space probing on arm64.
> > >
> > > -Will