Re: Remove dependency on libjansson

[Rui Ueyama <rui314@gmail.com>]

We have discussed various topics already, and I don't think there's a
single answer because this is all about engineering tradeoffs.

I'd like to hear from other devs who are following this thread if there are any.

On Tue, Apr 2, 2024 at 9:54 AM Luca Boccassi <bluca@debian.org> wrote:
>
> On Mon, 1 Apr 2024 at 14:23, Rui Ueyama <rui314@gmail.com> wrote:
> >
> > On Mon, Apr 1, 2024 at 9:16 PM Luca Boccassi <bluca@debian.org> wrote:
> > >
> > > On Mon, 1 Apr 2024 at 12:44, Rui Ueyama <rui314@gmail.com> wrote:
> > > >
> > > > On Mon, Apr 1, 2024 at 7:38 PM Luca Boccassi <bluca@debian.org> wrote:
> > > > >
> > > > > On Mon, 1 Apr 2024 at 10:50, Sam James <sam@gentoo.org> wrote:
> > > > > >
> > > > > > Fangrui Song <i@maskray.me> writes:
> > > > > >
> > > > > > > On Sun, Mar 31, 2024 at 8:31 PM Rui Ueyama <rui314@gmail.com> wrote:
> > > > > > >>
> > > > > > >> Hi,
> > > > > > >>
> > > > > > >> The recent xz incident demonstrated that supply chain attacks are a
> > > > > > >> real threat, and dependence on third-party libraries can have
> > > > > > >> significant consequences.
> > > > > > >>
> > > > > > >> In the wake of the incident, I propose we remove the dependency on
> > > > > > >> libjansson from GNU ld.
> > > > > > >>
> > > > > > >> First of all, why does GNU ld depend on libjansson which is a JSON
> > > > > > >> parsing library? GNU ld gained the `--package-metadata` option in May
> > > > > > >> 2022 to embed a JSON string into a .note section for package
> > > > > > >> management for Fedora and other Linux distributions. At the same time,
> > > > > > >> the dependency on libjansson, a library for parsing JSON-format
> > > > > > >> strings, was introduced to validate an argument for that option. If an
> > > > > > >> argument is not a valid JSON string, ld reports an error. If the
> > > > > > >> library is unavailable, or if `--disable-jansson` was passed to the
> > > > > > >> configure script, the library will not be linked and the error check
> > > > > > >> will be disabled. By default, the library will be linked if it exists.
> > > > > > >>
> > > > > > >> I opposed adding an extra dependency to GNU ld just for string
> > > > > > >> verification purposes because it didn't seem worth adding extra
> > > > > > >> dependency to the linker. LLVM lld and the mold linker also support
> > > > > > >> the option, but they do not verify if the argument is a valid JSON
> > > > > > >> string -- they simply treat it as an opaque string. If libjansson is
> > > > > > >> unavailable, even GNU ld doesn't verify arguments. Therefore, the
> > > > > > >> verification is not trustworthy, and the reader must be prepared for a
> > > > > > >> malformed JSON string when reading a .note section. Moreover,
> > > > > > >> verifying a string is straightforward without the feature; you can
> > > > > > >> simply `echo` the string to pipe it to `jq` for verification before
> > > > > > >> passing it to GNU ld.
> > > > > > >>
> > > > > > >> I just checked /usr/bin/ld on Ubuntu 24.04, which is set to be
> > > > > > >> released this month, and the dependency on libjansson was indeed
> > > > > > >> present.
> > > > > > >>
> > > > > > >> How much risk does it pose? Probably not much, as long as the library
> > > > > > >> is maintained properly. However, the stakes are high; if someone takes
> > > > > > >> control of the library and introduces malicious code, they could
> > > > > > >> execute a Ken Thompson-style supply chain attack. Since GNU ld is used
> > > > > > >> to build essentially everything, the attacker could in theory gain the
> > > > > > >> power to not just contaminate a specific program such as openssh, but
> > > > > > >> every executable in an official Linux distribution image. I think the
> > > > > > >> risk is not worth taking. I believe we just should remove the string
> > > > > > >> verification code and the dependency on the library from GNU ld.
> > > > > > >>
> > > > > > >> Rui Ueyama
> > > > > > >
> > > > > > > Thanks for bringing this up again. I support removing the json dependency.
> > > > > > >
> > > > > > > I lightly expressed my concern
> > > > > > > https://sourceware.org/pipermail/binutils/2022-May/120846.html and
> > > > > > > there might be others unsure about the dependency as well.
> > > > > >
> > > > > > I'd like to hear bluca's take before making up my mind. Note that it's
> > > > > > also automagic right now IIRC (enabled if installed, not opt-in).
> > > > > >
> > > > > > But my take on it so far is that it doesn't sound worth it.
> > > > >
> > > > > $ apt-cache rdepends libjansson4 | wc -l
> > > > > 310
> > > > >
> > > > > Sorry, but this looks to me like a knee-jerk reaction that misses the
> > > > > wood for the trees. The xz issue is that an extremely complex and
> > > > > sophisticated social engineering attack was planned and executed over
> > > > > multiple years. If it hadn't been on xz, it would have been on another
> > > > > project. Dropping dependencies left and right without thinking will
> > > > > not stop the next attack, as there is no shortage of overworked
> > > > > maintainers to try and abuse. In fact, avoiding common libraries and
> > > > > everyone reimplementing the wheel _increases_ the attack
> > > > > opportunities.
> > > > >
> > > > > The reason for this integration is that finding issues at build time
> > > > > is cheaper than finding them at runtime or release time. We rely on
> > > > > this in Fedora, Debian and Ubuntu to ensure valid data is written (in
> > > > > all packages in the former, in an increasing set of opt-ins in the
> > > > > latters) - in fact, validating input _before_ shipping binaries
> > > > > _improves_ security, rather than decreasing it, given the payload is
> > > > > parsed by a privileged component.
> > > > >
> > > > > So if it was removed upstream, we'd just have to patch it back in
> > > > > downstream in each interested distro - which if you recall, was
> > > > > exactly how sshd was exploited, due to the lack of an upstream common
> > > > > implementation. This is opt-in, so distributions that are not
> > > > > interested can just avoid enabling it, with no extra effort required.
> > > > > So please, leave it as-is. Thank you.
> > > >
> > > > I'd like to say that that's not an instinctive reaction to the xz
> > > > incident, as I expressed a concern when the dependency was being
> > > > introduced. I also want to say that I'm not trying to drop
> > > > dependencies left and right without thinking.
> > >
> > > Well, the email did start with:
> > >
> > > > The recent xz incident demonstrated that supply chain attacks are a real threat, and dependence on third-party libraries can have significant consequences.
> > >
> > > so it should be understandable why that's the impression it gave me.
> > > This feature has been present in binutils for 2 years now, and AFAIK
> > > caused zero issues so far (probably going to jinx it with this
> > > comment, but still).
> > >
> > > > I understand that the error check would be sometimes useful, but I'd
> > > > like to propose reevaluating its cost and benefit. IIUC, the
> > > > --package-metadata option is, as its name suggests, used by packaging
> > > > tools such as dpkg-dev or rpmbuild. My understanding is that this
> > > > option is not intended to be used directly by developers at large, but
> > > > rather, the JSON string is generated by a very limited set of
> > > > packaging tools from package description files and passed to the
> > > > linker. If that's the case, the string format check itself is an extra
> > > > measure to catch rare string construction bugs, given that JSON
> > > > strings are machine-generated and not manually crafted.
> > >
> > > The metadata comes from each package, so it is very much possible for
> > > errors to appear in one build but not another. And they would be hard
> > > to spot and expensive to fix, as they'd likely only be found when
> > > things go really wrong and the data is needed at runtime, or worse.
> >
> > Correct me if I'm wrong, but I believe the main concern in creating a
> > malformed JSON string is forgetting to escape certain characters when
> > embedding it in a JSON string. Correctly handling it shouldn't be too
> > difficult.
>
> No, it's about anything that results in an invalid payload.
>
> > I understand that it is sometimes hard to notice an error in the .note
> > section as it's just embedded data sitting there, and errors in the
> > section would often be noticed exactly in the situation when you
> > really need it. But if you are very cautious about the sanity of the
> > .note section, why would you stop here and only focus on catching JSON
> > format errors? If you do not trust your own tool and aim to catch bugs
> > before its generated data goes into production, you should read the
> > .note section back and compare the JSON contents with its package
> > metadata file. There are tons of different ways for the tool to
> > semantically go wrong such as missing field or something like that,
> > and just catching JSON format errors (or, maybe just string escaping
> > bugs) seems insufficient to me. And if you think that that kind of
> > error would be unlikely, then I wonder why you are worried about the
> > JSON format error.
>
> No, that is not a problem in reality, I'm not sure where you got that
> impression. What really matters is ensuring the payload is valid json,
> and the best way to do that is the current setup, because it costs
> nothing, it's already implemented, it works out of the box
> automatically for any caller regardless of the distribution and the
> packaging format, it's inoffensive as it is not enabled by default and
> requires opt-in, and has zero drawbacks or issues as evidenced by the
> 2 years it's been available and used. I don't see why it's necessary
> to create non-existing problems out of thin air and break existing
> workflows, as a knee-jerk reaction driven by a fundamental
> misunderstanding about what happened to a completely separate and
> independent project. There is literally nothing to gain and all to
> lose.
>
> > > > While we may still do extra checks for catching possible tooling bugs,
> > > > I wonder how difficult it would be to assign this additional
> > > > verification to jq or a similar tool instead of the linker. IIRC, dpkg
> > > > comes with a tool to scan all executables in a directory to find
> > > > packaging errors. Isn't it a better place to run the extra format
> > > > check? Would it be too complex to output the .note section contents
> > > > after linking and pipe them to jq to confirm the embedded string is a
> > > > valid JSON string?
> > >
> > > There's no such thing in dpkg. You might be thinking of Lintian, but
> > > that's an entirely separate tool, that is optional, runs long after
> > > the package is published and distributed, and is only seldomly checked
> > > (most maintainers just ignore it as it's informational only), and
> > > doesn't stop bad builds from being generated and released. It's also
> > > package-format-specific, so again every distribution would have to
> > > reimplement exactly the same check in different ways, which is just
> > > duplication of work for no gain. The intention here was to have bad
> > > input fail the build immediately, with no extra effort required from
> > > distributions or packagers, and it works very well for that.
> > >
> > > This is akin to suggesting that compilers shoudn't output warnings,
> > > but the job should just be left to static analyzers. The earlier
> > > problems are detected, the cheaper they are to fix, and the less
> > > impact they have. This is one of the few truisms of engineering.
> >
> > If a --package-metadata argument is machine-generated, a better
> > analogy would be the relationship between the compiler and the linker,
> > rather than the human programmer and the compiler. The linker
> > sometimes verifies object file contents, but it doesn't disassemble
> > the .text section to catch rare assembler bugs.
>
> No, it's actually a good analogy. Removing warnings and telling people
> to just do post-processing would be equally nonsensical, as warnings
> are already implemented, they provide value if one wants to use them
> and are harmless if one wants to ignore them, and do not cause any
> real issue. Moreover, as already explained, this is not how
> distributions and packaging actually work in the real world, there is
> no such postprocessing, so it's even more nonsensical as a suggestion.