Re: [PATCH] binfmt_elf: Fail execution of shared objects with ELIBEXEC

Florian Weimer [off-list ref] writes:

* Mickaël Salaün:

quoted

On Sat, Jul 06, 2024 at 05:32:12PM +0200, Florian Weimer wrote:

quoted

* Mickaël Salaün:

quoted

On Fri, Jul 05, 2024 at 08:03:14PM +0200, Florian Weimer wrote:

quoted

* Mickaël Salaün:

quoted

Add a new AT_CHECK flag to execveat(2) to check if a file would be
allowed for execution.  The main use case is for script interpreters and
dynamic linkers to check execution permission according to the kernel's
security policy. Another use case is to add context to access logs e.g.,
which script (instead of interpreter) accessed a file.  As any
executable code, scripts could also use this check [1].

Some distributions no longer set executable bits on most shared objects,
which I assume would interfere with AT_CHECK probing for shared objects.

A file without the execute permission is not considered as executable by
the kernel.  The AT_CHECK flag doesn't change this semantic.  Please
note that this is just a check, not a restriction.  See the next patch
for the optional policy enforcement.

Anyway, we need to define the policy, and for Linux this is done with
the file permission bits.  So for systems willing to have a consistent
execution policy, we need to rely on the same bits.

Yes, that makes complete sense.  I just wanted to point out the odd
interaction with the old binutils bug and the (sadly still current)
kernel bug.

quoted

quoted

Removing the executable bit is attractive because of a combination of
two bugs: a binutils wart which until recently always set the entry
point address in the ELF header to zero, and the kernel not checking for
a zero entry point (maybe in combination with an absent program
interpreter) and failing the execve with ELIBEXEC, instead of doing the
execve and then faulting at virtual address zero.  Removing the
executable bit is currently the only way to avoid these confusing
crashes, so I understand the temptation.

Interesting.  Can you please point to the bug report and the fix?  I
don't see any ELIBEXEC in the kernel.

The kernel hasn't been fixed yet.  I do think this should be fixed, so
that distributions can bring back the executable bit.

Can you please point to the mailing list discussion or the bug report?

I'm not sure if this was ever reported upstream as an RFE to fail with
ELIBEXEC.  We have downstream bug report:

  Prevent executed .so files with e_entry == 0 from attempting to become
  a process.
  <https://bugzilla.redhat.com/show_bug.cgi?id=2004942>

I've put together a patch which seems to work, see below.

I don't think there's any impact on AT_CHECK with execveat because that
mode will never get to this point.

Thanks,
Florian

---8<-----------------------------------------------------------------
Subject: binfmt_elf: Fail execution of shared objects with ELIBEXEC
    
Historically, binutils has used the start of the text segment as the
entry point if _start was not defined.  Executing such files results
in crashes with random effects, depending on what code resides there.
However, starting with binutils 2.38, BFD ld uses a zero entry point,
due to commit 5226a6a892f922ea672e5775c61776830aaf27b7 ("Change the
linker's heuristic for computing the entry point for binaries so that
shared libraries default to an entry point of 0.").  This means
that shared objects with zero entry points are becoming more common,
and it makes sense for the kernel to recognize them and refuse
to execute them.

For backwards compatibility, if a load segment does not map the ELF
header at file offset zero, the kernel still proceeds as before, in
case the file is very non-standard and can actually start executing
at virtual offset zero.

As written I find the logic of the patch confusing, and slightly wrong.

The program header value e_entry is a virtual address, possibly adjusted
by load_bias.  Which makes testing it against the file offset of a
PT_LOAD segment wrong.  It needs to test against elf_ppnt->p_vaddr.

I think performing an early sanity check to avoid very confusing crashes
seems sensible (as long as it is inexpensive).  This appears inexpensive
enough that we don't care.  This code is also before begin_new_exec
so it is early enough to be meaningful.

I think the check should simply test if e_entry is mapped.  So a range
check please to see if e_entry falls in a PT_LOAD segment.

Having code start at virtual address 0 is a perfectly fine semantically
and might happen in embedded scenarios.

The program header is not required to be mapped or be first, (AKA
p_offset and p_vaddr can have a somewhat arbitrary relationship) so any
mention of the program header in your logic seems confusing to me.

I think your basic structure will work.  Just the first check needs to
check if e_entry is lands inside the virtual address of a PT_LOAD
segment.  The second check should just be checking a variable to see if
e_entry was inside any PT_LOAD segment, and there is no interpreter.

Does that make sense?

Eric