Linux + Clang + execstack
Magnus Ihse Bursie
magnus.ihse.bursie at oracle.com
Wed Sep 5 06:50:04 UTC 2018
On 2018-09-05 01:01, Martin Buchholz wrote:
> I think we can all agree that passing flags to the linker to ensure
> non-executable stack is the right thing to do. But there's a question
> whether *also* adding something to our assembly language source files
> will be worth doing. Neither mechanism is sure to work. For the
> linker flag, we need to be aware of and test for the presence of the
> linker flag, but we might be using some other linker...
For the gcc toolchain this can not be the case:
# Minimum supported linker versions, empty means unspecified
We make sure we have an ld that supports the basic flags we assume.
We can add a similar check for the clang toolchain, if you want.
Mixing and matching compilers and linkers willy-nilly is not a supported
build option, and there's no point in adding extra guards against such a
thing. If you do that, and it turns out your build ended up broken,
So I'll insist that adding the linker flag is sufficient to make sure
this works, and that modifying the .s files are not necessary and will
provide no benefit.
> Similarly, we might end up using some other assembler, or we might
> need to mark the assembly source file in a different way than "GNU-stack".
> On Tue, Aug 21, 2018 at 4:14 AM, Magnus Ihse Bursie
> <magnus.ihse.bursie at oracle.com <mailto:magnus.ihse.bursie at oracle.com>>
> On 2018-08-21 02:03, David Holmes wrote:
> On 21/08/2018 9:39 AM, Arthur Eubanks wrote:
> On Mon, Aug 20, 2018 at 4:18 PM David Holmes
> <david.holmes at oracle.com <mailto:david.holmes at oracle.com>
> <mailto:david.holmes at oracle.com
> <mailto:david.holmes at oracle.com>>> wrote:
> Hi Arthur,
> cc'ing build-dev as this is currently a build issue.
> On 21/08/2018 3:11 AM, Arthur Eubanks wrote:
> > Hi,
> > At Google we're trying to build hotspot on Linux
> with clang. One
> thing that
> > happens is that the resulting libjvm.so is stack
> executable. When
> > hotspot we get warnings about the stack being
> > Compiling an assembly file into the final .so
> results in the
> stack being
> > executable. In this case the file is
> linux_x86_64.s. This doesn't
> > with gcc because "-Wl,-z,noexecstack" is passed as
> a hotspot
> linker flag
> > with gcc in flags-ldflags.m4. When using clang that
> linker flag isn't
> > passed.
> > Doing something like the solution in
> > fixes the problem without the use of linker flags.
> You mean the source code directives for the linker?
> Sorry, I wasn't specific enough, I meant the flags for the
> #if defined(__linux__) && defined(__ELF__)
> .section .note.GNU-stack, "", %progbits
> I think I prefer to see this handled explicitly in the
> build as is
> currently done. Can we just adjust
> ./make/autoconf/flags-ldflags.m4 to
> pass the linker flags for gcc and clang?
> I don't mind this solution, but it seems like the right
> thing to do is to fix things at the source level and
> remove extra unnecessary linker flags.
> Personally I see this as source code pollution. The concept of
> executable stacks has nothing to do with what is being
> expressed by the source code, or the language used for it.
> Just my 2c. I'll defer to build folk ... though they are still
> on vacation at the moment.
> I agree with David. The executable stack is a build option. Even
> if you change the source code so the compiler/assember does the
> right thing, we would still want to keep the compiler option.
> (Otherwise one day you'll end up with executable stacks due to
> someone adding a new asm file and forgetting the "magic incantation".)
> And, since we will keep the compiler option, there seems little
> point in also adding this stuff to the asm files.
> To address your concerns on clang: we should reasonably be giving
> the same options to clang. There is no good reason, except for
> oversight, that this is not done already. (Cleaning up and
> unifying the compiler flags is an ongoing, but slowly moving,
> process.) So the correct fix is to update flags-ldflags.m4.
> I removed "-Wl,-z,noexecstack" from the flags after adding
> the above assembler flags and libjvm.so is still correctly
> not stack executable. I don't really mind either way
> though. Maybe it's good to have an extra safeguard in the
> linker flags.
> > The jtreg test
> > checks for the stack being executable.
> > Any thoughts? If there are no objections, I can
> propose a patch
> that works
> > for both gcc and clang on Linux. Also, I'm not sure
> how/if macOS
> > this problem given that it uses clang.
> We don't seem to handle it at all on OS X. Does OS X
> prevent executable
> stacks itself?
> A quick search, according to Wikipedia
> 64-bit executables on macOS aren't stack or heap
> executable. Not sure if that information is accurate though.
> Seems to be:
> "macOS and iOS provide two features that can make it harder to
> exploit stack and buffer overflows: address space layout
> randomization (ASLR) and a non-executable stack and heap."
More information about the hotspot-dev