On 06/19/15 12:58, Kees Cook wrote:

On Fri, Jun 19, 2015 at 7:12 AM, David Long [off-list ref] wrote:

quoted

On 06/19/15 00:19, Michael Ellerman wrote:

quoted

On Mon, 2015-06-15 at 12:42 -0400, David Long wrote:

quoted

From: "David A. Long" <redacted>

The pt_regs_offset structure is used for HAVE_REGS_AND_STACK_ACCESS_API
   feature and has identical definitions in four different arch ptrace.h
include files. It seems unlikely that definition would ever need to be
changed regardless of architecture so lets move it into
include/linux/ptrace.h.

Signed-off-by: David A. Long <redacted>
---
   arch/powerpc/kernel/ptrace.c | 5 -----

Built and booted on powerpc, but is there an easy way to actually test the
code
paths in question?

There is an easy way to "smoke test" it on all archiectures that also
implement kprobes (which powerpc does).  If I'm understanding the powerpc
code correctly (WRT register naming conventions) just do the following:

cd /sys/kernel/debug/tracing
echo 'p do_fork %gpr0' > kprobe_events
echo 1 > events/kprobes/enable
ls
cat trace
echo 0 > events/kprobes/enable

Every fork() call done on the system between those two echo commands (hence
the "ls") should append a line to the trace file.  For a more exhaustive
test one could repeat this sequence for every register in the architecture.

This should work the same on all architectures supporting kprobes.  You just
have to use the appropriate register names for your architecture after the
"%".

Is this something we could codify into the selftests directory? It
seems like a great thing to capture in a single place somewhere (the
register lists, that is).
e
-Kees

Due to the architecture-specific naming of registers this would have to 
be added to the architecture subdirectories.  I only see powerpc and x86 
subdirs at this time so extending that infrastructure would have to be 
part of this.  Verifying the register contents would also require some 
change to the kernel, possibly a simple test module, which would have to 
be unique to each architecture.  Without that we could only check for 
recognition of the register name, although maybe that's good enough.

quoted

quoted

Acked-by: Michael Ellerman <mpe@ellerman.id.au>

cheers

Thanks,
-dl

-dl