RE: Debugging MIPS PIE executables

[Felix Radensky <felix dot radensky at broadcom dot com>]

Hi Mattew,

> -----Original Message-----
> From: Matthew Fortune [mailto:Matthew.Fortune@imgtec.com]
> Sent: Thursday, June 18, 2015 8:22 PM
> To: Felix Radensky; binutils@sourceware.org
> Subject: RE: Debugging MIPS PIE executables
> 
> Felix Radensky <felix.radensky@broadcom.com> writes:
> > > -----Original Message-----
> > > From: Matthew Fortune [mailto:Matthew.Fortune@imgtec.com]
> > > Sent: Tuesday, June 16, 2015 3:07 PM
> > > To: Felix Radensky; binutils@sourceware.org
> > > Subject: RE: Debugging MIPS PIE executables
> > >
> > > Felix Radensky <felix.radensky@broadcom.com> writes:
> > > > > -----Original Message-----
> > > > > From: Matthew Fortune [mailto:Matthew.Fortune@imgtec.com]
> > > > > Sent: Monday, June 15, 2015 9:03 PM
> > > > > To: Felix Radensky; binutils@sourceware.org
> > > > > Subject: RE: Debugging MIPS PIE executables
> > > > >
> > > > > Felix Radensky <felix.radensky@broadcom.com> writes:
> > > > > > Hi Mattew
> > > > > >
> > > > > > > -----Original Message-----
> > > > > > > From: Matthew Fortune [mailto:Matthew.Fortune@imgtec.com]
> > > > > > > Sent: Saturday, June 13, 2015 11:10 AM
> > > > > > > To: Felix Radensky; binutils@sourceware.org
> > > > > > > Subject: RE: Debugging MIPS PIE executables
> > > > > > >
> > > > > > > Felix Radensky <felix.radensky@broadcom.com> writes:
> > > > > > > > Hi Matthew,
> > > > > > > >
> > > > > > > > > -----Original Message-----
> > > > > > > > > From: Matthew Fortune
> > > > > > > > > [mailto:Matthew.Fortune@imgtec.com]
> > > > > > > > > Sent: Sunday, June 07, 2015 11:10 AM
> > > > > > > > > To: Felix Radensky; binutils@sourceware.org
> > > > > > > > > Subject: RE: Debugging MIPS PIE executables
> > > > > > > > >
> > > > > > > > > Hi Felix,
> > > > > > > > >
> > > > > > > > > Felix Radensky <felix.radensky@broadcom.com> writes:
> > > > > > > > > > I was wondering if any progress was made to resolve
> > > > > > > > > > the problem discussed in the following threads:
> > > > > > > > > >
> > > > > > > > > > https://sourceware.org/ml/binutils/2013-06/msg00183.ht
> > > > > > > > > > ml
> > > > > > > > > > https://sourceware.org/ml/binutils/2014-10/msg00200.ht
> > > > > > > > > > ml
> > > > > > > > > >
> > > > > > > > > > Are there any experimental patches for
> > > > > > > > > > binutils/gdb/glibc that
> > > > > > I
> > > > > > > > can
> > > > > > > > > > try ? PIE executables are must in many
> > > > > > > > > > security-conscious projects, and the lack of ability
> > > > > > > > > > to debug them is a serious
> > > > > > problem.
> > > > > > > > >
> > > > > > > > > This request has come up a few times again recently and
> > > > > > > > > I am hoping
> > > > > > > > to
> > > > > > > > > submit patches for this in the coming week. I still have
> > > > > > > > > not managed to build a PIE version of glibc which is a
> > > > > > > > > bit of a blocker. I'm currently trying to use HJ's
> > > > > > > > > recent patches to support PIE by default from GCC but
> > > > > > > > > now have link assertions in glibc instead of reloc
> > > > > > > > failures.
> > > > > > > > >
> > > > > > > > > I have binutils and gdb patches that have been used with
> > > > > > > > > bionic but they still need some cleaning up.
> > > > > > > >
> > > > > > > > Thanks a lot for a prompt reply. I'm looking forward to
> > > > > > > > test your patches.
> > > > > > >
> > > > > > > I've attached the patches I intend to submit but I still
> > > > > > > have to work through the binutils testsuite to account for
> > > > > > > the new tag and try to
> > > > > > write a useful test case.
> > > > > > > I'll also need to get advice on doing a GDB test for this as
> > > > > > > I don't know if/how to do that.
> > > > > > >
> > > > > > > The patches are manually tested for o32/n32/n64 PIE and
> > > > > > > ordinary executables with the expected behaviour that new
> > > > > > > PIEs with new ld.so show the correct library list.
> > > > > > >
> > > > > >
> > > > > > I've build a MIPS o32 little-endian toolchain using
> > > > > > crosstool-ng- 1.21.0, with gcc-5.1.0, binutils-2.25 and
> > > > > > gdb-7.9.1, and your patches applied. However the problem still
> > > > > > exists, I'm unable to debug PIE executables with shared
> > > > > > libraries or inspect core files produced by such executables.
> > > > > > The shared library is not listed by "info sharedlibrary", the
> > > > > > backtrace cannot be extracted from core file. I can send you
> > > > > > my crosstool-ng configuration file and the trivial application I use for
> testing.
> > > > >
> > > > > The support in the GDB side of this depends on the elf.h header
> > > > > having the
> > > > > DT_MIPS_RLD_MAP2 macro defined. Are you using native GDB or
> > > > > remote with the gdbserver? The gdbserver does the work for
> > > > > finding shared library debug if you are using a remote so has to
> > > > > be built from the patched
> > > sources.
> > > > >
> > > > > Debugging tips would be to firstly check that your PIE has the
> > > > > DT_MIPS_RLD_MAP2 tag using readelf. To get a textual description
> > > > > of the tag you will have to use readelf built from the patched binutils
> source.
> > > > >
> > > > > Then double check that GDB has been built against headers that
> > > > > define DT_MIPS_RLD_MAP2, the GLIBC patch includes changes to add
> > > > > this to the headers.
> > > > >
> > > >
> > > > Thanks for the tips. I've verified that my PIE executable has
> > > > DT_MIPS_RLD_MAP2 tag. Also recompiled cross GDB, gdbserver and
> > > > native GDB with DT_MIPS_RLD_MAP2 defined. I can now successfully
> > > > debug my PIE executable using gdbserver and cross GDB. What I
> > > > still cannot do is to debug
> > > using native GDB and to inspect core files using cross GDB.
> > >
> > > Core files would be new territory for me so it will take a bit of
> > > exploring to understand what piece of the puzzle is missing.
> > > Presumably the core file carries some indication of where the
> > > r_debug map is and that is not getting set correctly when generating the
> core file.
> > >
> > > I'll switch this over to the GDB list if I get stuck.
> > >
> > >
> >
> > Actually core file inspection works fine for me with native GDB on the
> > target, as well as debugging with native GDB. It's only the core file
> > inspection in cross GDB that doesn't work.
> 
> OK, the penny drops. For some reason I thought the cross build of GDB would
> use the local headers in binutils/gdb but having thought harder that couldn't
> possibly be true. The problem is that the host's elf.h won't have
> DT_MIPS_RLD_MAP2 defined so for a cross build this will have to be defined
> within the solib-svr4.c perhaps as an #ifndef DT_MIPS_RLD_MAP2 #define...
> 

Explicitly defining DT_MIPS_RLD_MAP2 in solib-svr4.c solves the problem. Huge thanks for your help, I really appreciate it. I hope your patches will make into mainline soon.

Felix.