[lkml]   [2018]   [Feb]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RFC PATCH 4/7] kconfig: support new special property shell=
On Sun, Feb 11, 2018 at 2:34 AM, Ulf Magnusson <> wrote:
> Looks to me like there's a few unrelated issues here:
> 1. The stack protector support test scripts
> Worthwhile IMO if they (*in practice*) prevent hard-to-debug build errors or a
> subtly broken kernel from being built.
> A few questions:
> - How do things fail with a broken stack protector implementation?

There have been several ways I've seen:
- resulting kernel silently doesn't provide the stack protection at all
- resulting build fails at the end trying to link against a missing
global stack canary
- resulting kernel doesn't boot at all due to insane function preamble
on first use of canary

> - How common are those broken compilers?

I *thought* it was rare (i.e. gcc 4.2) but while working on ..._AUTO I
found breakage in akpm's 4.4 gcc, and all of Arnd's gccs due to some
very strange misconfiguration between the gcc build environment and
other options. So, it turns out this is unfortunately common. The good
news is that it does NOT appear to happen with most distro compilers,
though I've seen Android's compiler regress the global vs %gs at least
once about a year ago.

> - Do you really need to pass $(KBUILD_CPPFLAGS) when testing for breakage,
> or would a simpler static test work in practice?
> I don't know how messy it would be to get $(KBUILD_CPPFLAGS) into
> Kconfig, but should make sure it's actually needed in any case.
> The scripts are already split up as
> scripts/gcc-$(SRCARCH)_$(BITS)
> by the way, though only and
> exist.

I think it would work to skip KBUILD_CPPFLAGS right up until it
didn't. Since we have the arch split, we can already add -m32 to the
32-bit case, etc. However, I worry about interaction with other
selected build options. For example, while retpoline doesn't interact
stack-protector, it's easy to imagine things that might.

(And in thinking about this, does Kconfig know the true $CC in use?
i.e. the configured cross compiler, etc?)

> - How old do you need to go with GCC for -fno-stack-protector to give an
> error (i.e., for not even the option to be recognized)? Is it still
> warranted to test for it?

Old? That's not the case. The check for -fno-stack-protector will
likely be needed forever, as some distro compilers enable
stack-protector by default. So when someone wants to explicitly build
without stack-protector (or if the compiler's stack-protector is
detected as broken), we must force it off for the kernel build.

> Adding some CCs who worked on the stack protector test scripts.
> And yeah, I was assuming that needing support scripts would be rare, and that
> you'd usually just check whether gcc accepts the flag.

That would have been nice, yes. :(

> When you Google "gcc broken stack protector", the top hits about are about the
> scripts/ script in the kernel throwing a false
> positive by the way (fixed in 82031ea29e45 ("scripts/has-stack-protector: add
> -fno-PIE")).

That's just the most recent case (from the case where some distro
compilers enabled PIE by default). And was why I was hoping to drop
the scripts entirely.

> 2. Whether to hide the Kconfig stack protector alternatives or always show them
> Or equivalently, whether to automatically fall back on other stack protector
> alternatives (including no stack protector) if the one specified in the .config
> isn't available.
> I'll let you battle this one out. In any case, as a user, I'd want a
> super-clear message telling me what to change if the build breaks because of
> missing stack protector support.

Yes, exactly.

The reason I built the _AUTO support was to make this easier for
people to not have to think about. I wanted to get rid of explicit
support (i.e. _REGULAR or _STRONG) but some people didn't want _STRONG
(since it has greater code-size impact than _REGULAR), so we've had to
keep that too.

> 3. Whether to implement CC_STACKPROTECTOR_AUTO in Kconfig or the Makefiles
> I'd just go with whatever is simplest here. I don't find the Kconfig version
> too bad, but I'm already very familiar with Kconfig, so it's harder for me to
> tell how it looks to other people.
> I'd add some comments to explain the idea in the final version.
> @Kees:
> I was assuming that the Makefiles would error out with a message if none of the
> CC_STACKPROTECTOR_* variables are set, in addition to the Kconfig warning.

That doesn't happy either before nor after _AUTO. The default value
before was _NONE, and the default value after is _AUTO, which will use
whatever is available.

> You could offload part of that check to Kconfig with something like
> CHOSEN_STACKPROTECTOR_AVAILABLE could then be checked in the Makefile.
> It has the advantage of making the constraint clear in the Kconfig file
> at least.
> You could add some kind of assert feature to Kconfig too, but IMO it's not
> warranted purely for one-offs like this at least.

Agreed; I want to do whatever we can to simplify things. :)

> That's details though. I'd want to explain it with a comment in any case if we
> go with something like this, since it's slightly kludgy and subtle
> (CC_STACKPROTECTOR_{STRONG,REGULAR,NONE} form a kind of choice, only you can't
> express it like that directly, since it's derived from other symbols).
> Here's an overview of the current Kconfig layout by the way, assuming
> the old no-fallback behavior and CC_STACKPROTECTOR_AUTO being
> implemented in Kconfig:
> # Feature tests

As long as the feature tests include actual compiler output tests,
yeah, this seems fine.

> # User request
> # The actual "output" to the Makefiles

This should be fine too (though by this point, I think Kconfig would
already know the specific option, so it could just pass it with a
single output (CC_OPT_STACKPROTECTOR below?)

> # Some possible output "nicities"
> Does anyone have objections to the naming or other things? I saw some
> references to "Santa's wish list" in messages of commits that dealt with other
> variables named WANT_*, though I didn't look into those cases. ;)

Another case I mentioned before that I just want to make sure we don't
reintroduce the problem of getting "stuck" with a bad .config file.
While adding _STRONG support, I discovered the two-phase Kconfig
resolution that happens during the build. If you selected _STRONG with
a strong-capable compiler, everything was fine. If you then tried to
build with an older compiler, you'd get stuck since _STRONG wasn't
support (as detected during the first Kconfig phase) so the
generated/autoconf.h would never get updated with the newly selected
_REGULAR). I moved the Makefile analysis of available stack-protector
options into the second phase (i.e. after all the Kconfig runs), and
that worked to both unstick such configs and provide a clear message
early in the build about what wasn't available.

If all this detection is getting moved up into Kconfig, I'm worried
we'll end up in this state again. If the answer is "you have to delete
autoconf.h if you change compilers", then that's fine, but it sure
seems unfriendly. :)


Kees Cook
Pixel Security

 \ /
  Last update: 2018-02-11 18:57    [W:0.225 / U:2.036 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site