svn commit: r361303 - in head: lib/libc/gen libexec/rtld-elf sys/sys

[Mark Johnston]

On Thu, May 21, 2020 at 08:09:58PM +0300, Konstantin Belousov wrote:
> On Thu, May 21, 2020 at 01:01:24PM -0400, Mark Johnston wrote:
> > On Thu, May 21, 2020 at 07:56:46PM +0300, Konstantin Belousov wrote:
> > > On Thu, May 21, 2020 at 09:03:44AM -0700, John Baldwin wrote:
> > > > On 5/21/20 8:12 AM, Mark Johnston wrote:
> > > > > On Thu, May 21, 2020 at 04:41:52PM +0300, Konstantin Belousov wrote:
> > > > >> On Thu, May 21, 2020 at 03:02:07PM +0200, Antoine Brodin wrote:
> > > > >>> On Thu, May 21, 2020 at 12:08 AM Konstantin Belousov <kib at freebsd.org> wrote:
> > > > >>>>
> > > > >>>> Author: kib
> > > > >>>> Date: Wed May 20 22:08:26 2020
> > > > >>>> New Revision: 361303
> > > > >>>> URL: https://svnweb.freebsd.org/changeset/base/361303
> > > > >>>>
> > > > >>>> Log:
> > > > >>>>   Change the samantic of struct link_map l_addr member.
> > > > >>>>
> > > > >>>>   It previously returned the object map base address, while all other
> > > > >>>>   ELF operating systems return load offset, i.e. the difference between
> > > > >>>>   map base and the link base.
> > > > >>>>
> > > > >>>>   Explain the meaning of the field in the man page.
> > > > >>>>
> > > > >>>>   Stop filling the mips-only l_offs member, which is apparently unused.
> > > > >>>>
> > > > >>>>   PR:   246561
> > > > >>>>   Requested by: Damjan Jovanovic <damjan.jov at gmail.com>
> > > > >>>>   Reviewed by:  emaste, jhb, cem (previous version)
> > > > >>>>   Sponsored by: The FreeBSD Foundation
> > > > >>>>   MFC after:    1 week
> > > > >>>>   Differential revision:        https://reviews.freebsd.org/D24918
> > > > >>>>
> > > > >>>> Modified:
> > > > >>>>   head/lib/libc/gen/dlinfo.3
> > > > >>>>   head/libexec/rtld-elf/rtld.c
> > > > >>>>   head/sys/sys/link_elf.h
> > > > >>>
> > > > >>> Hi,
> > > > >>>
> > > > >>> After this commit,  some ports fail to build with signal 11.
> > > > >>> For instance lang/perl5.30 fails to build with default options (DTRACE on)
> > > > >>> Disabling the DTRACE option makes it able to build again.
> > > > >>>
> > > > >> I see, thank you for reporting.
> > > > >>
> > > > >> So drti.c:dtrace_dof_init() does read l_addr, and the dtrace code assumes
> > > > >> that l_addr is the base, not relocbase.
> > > > >>
> > > > >> Mark, was dofhp_addr initialization changed comparing to Solaris ?
> > > > > 
> > > > > It appears it has been the same since DTrace was imported.  illumos
> > > > > still has similar code.
> > > > > 
> > > > > Note that drti.o is linked into any executable and shlib that defines
> > > > > static probes, so the ABI change affects more than just dtrace(1).
> > > > > Would it be possible to define a new value for RTLD_DI_LINKMAP, and
> > > > > preserve the old behaviour for the old value?
> > > > 
> > > > I think a bigger question is if Solaris/illumos treat l_addr as mapbase
> > > > (absolute address) or relocbase (relative address).  In the discussion
> > > > in the phabricator I had assumed that all other OS's treated l_addr as
> > > > the relative offset (relocbase).  Does the code for illumos assume an
> > > > absolute address or does it assume a relative address in l_addr?
> > > 
> > > It is rather clear, since the dtrace code was pristine, that Solaris
> > > provides the mapbase.  I do not have Solaris/Illumos box anymore
> > > (for quite some time), so I cannot check directly.
> > > 
> > > My current PoV is that l_addr semantic must be restored, and relocbase
> > > provided by newly added member.
> > > 
> > > BTW, it is strange that perl triggers it, is it linked as PIE on HEAD ?
> > 
> > Isn't the problem when perl is *not* linked as PIE?  In this case
> > relocbase is 0, so the ELF header access becomes a NULL pointer
> > dereference.
> drti checks for ET_DYN, only then it uses l_addr at all.

The problem is before that, where it treats l_addr as a pointer to the
ELF header so it can check the type.

> PIE binaries are dso with non-zero base, non-PIE binaries are ET_EXEC,
> which should make dtri.c ignore the l_addr value.