RE: [PATCH v2 0/3] elf: Allow dlopen of filter object to work [BZ #16272]

[David Kilroy <David.Kilroy@arm.com>]

Hi,

Ping. Just checking what the general feeling is about this patchset.

If it helps, I've locally tested a version which scans backwards through
new->l_prev and constructs map from the earliest non-relocated object (instead
of the current object). This passes the test. Would this version be preferred?



Thanks,

Dave.

Note: for the test case the walk of ->l_prev hits:
elf/tst-filterobj-lib.so
elf/ld-linux-x86-64.so.2
libc.so.6
dlfcn/libdl.so.2
linux-vdso.1
""                <- an object that doesn't have a ->l_name

> -----Original Message-----
> From: libc-alpha-owner at sourceware dot org <libc-alpha-owner at
> sourceware dot org> On Behalf Of David Kilroy
> Sent: 24 October 2019 11:34
> To: libc-alpha at sourceware dot org
> Cc: nd <nd at arm dot com>
> Subject: [PATCH v2 0/3] elf: Allow dlopen of filter object to work [BZ
> #16272]
> 
> This series fixes up the patchset for the comments raised so far.
> Patches 2 and 3 are unchanged.
> 
> v2:
>  - code formatting fixups
>  - add dependency of test output on filtee library
>  - tests changed to use the test framework
> 
> The main outstanding question is whether it is valid to use l_initfini
> to do the relocations, or whether we need to stick to following
> new->l_next (and new->l_prev). I'm reasonably confident that
> l_initfini contains all the objects in new->l_next. See
> https://sourceware.org/ml/libc-alpha/2019-10/msg00659.html
> 
> The majority of the discussion on this series has been around how
> filter objects are specified, and the alternatives to using filter
> objects. The most persuasive is to use a version_script to specify the
> ABI in the link library (.so), and setup the soname symlink (.so.1) to
> point to the filtee. This relies on:
> 
> * library versioning being used.
> 
> * dlopen() calls must use the soname.
> 
> * if there are multiple implementors, they need to be consistent in
>   versioning the library.
> 
> I think these are reasonable requirements, though may be difficult to
> ensure in the field.
> 
> Independent of the alternatives, I'd still like to advocate fixing
> dlopen for filter objects. glibc currently works (with various
> provisos) with filter objects when the application is linked against
> one. Not being able to work with the same library via dlopen is
> unexpected.
> 
> David Kilroy (3):
>   elf: Allow dlopen of filter object to work [BZ #16272]
>   elf: avoid redundant sort in dlopen
>   elf: avoid stack allocation in dl_open_worker
> 
>  elf/Makefile               | 13 +++++++++++--
>  elf/dl-deps.c              | 35 ++++++++++++++++++++++++++---------
>  elf/dl-open.c              | 32 +++++++++++++++-----------------
>  elf/tst-filterobj-dlopen.c | 39
> +++++++++++++++++++++++++++++++++++++++
>  elf/tst-filterobj-flt.c    | 24 ++++++++++++++++++++++++
>  elf/tst-filterobj-lib.c    | 24 ++++++++++++++++++++++++
>  elf/tst-filterobj-lib.h    | 18 ++++++++++++++++++
>  elf/tst-filterobj.c        | 36 ++++++++++++++++++++++++++++++++++++
>  8 files changed, 193 insertions(+), 28 deletions(-)
>  create mode 100644 elf/tst-filterobj-dlopen.c
>  create mode 100644 elf/tst-filterobj-flt.c
>  create mode 100644 elf/tst-filterobj-lib.c
>  create mode 100644 elf/tst-filterobj-lib.h
>  create mode 100644 elf/tst-filterobj.c
> 
> --
> 2.7.4