Hi Magnus,
As mentioned before, the core idea here is really just providing a hook
for a script to be invoked after the native linker has been called.
So while thinking about making this more generic and not Windows
specific, it struck me that it might in fact be desirable to also drop
the "signing" notion from this altogether and instead consider it a way
to extend to the build system by plugging in custom scripts? And maybe
extend it with other strategically placed hooks?
Do you think this is a direction worth exploring further? Or that the
original, more focused scope is preferable?
Thanks,
Frederic
On 07/03/2025 16:49, Magnus Ihse Bursie wrote:
Hi Frederic,
I think your proposal makes sense, and as Erik says, the patch is very
unobtrusive.
I'm not at all sure if it is even possible, but if the signing logic
for macOS and Windows can be sort of unified and made more abstract,
that would certainly be a win.
/Magnus
On 2025-02-26 20:46, Erik Joelsson wrote:
On 2/26/25 10:40 AM, Frederic Thevenet wrote:
Hi Erik,
Thanks for your comments.
As a matter of fact, we at Red Hat also face similar constraints
that prevent us from being able to sign the files on the machine
that build them (and I suspect this is quite common-place).
My initial idea to accommodate that is to have the jdk build invoke
a user supplied script (via a --with configure argument) rather than
having it call signtool directly, so as to provide a way for
everyone to easily plug their specific configuration and logic (in
our case, the script uploads the file to sign to the signing machine
and download it once it is done, in a synchronous fashion).
In effect, it's just a hook that is invoked after the native linker
has been called.
I have created a PR[1] where you can see what I've done in details.
That's a small and unintrusive solution. Our solution is a lot more
involved as I figured I wanted to expose a set of top level "phase"
targets for signing for each module, so that building libs or
launchers can be done independently from signing (e.g.
java.base-libs, java.base-launchers and java.base-sign). This
necessitates a separate set of output directories for signed libs and
signed launchers, where we first build like normal to the regular
output dirs and then the sign target copies and signs binaries into
the *-signed outputdirs, and when building jmods, we need to pick up
the files from those *-signed directories. It's quite messy, so I've
kept it all in our internal makefile extensions. I'm also questioning
the usefulness of this flexibility given how much complexity it adds,
so your proposal is probably preferred.
That said, I'm wondering if we could make it a bit more generic and
not define it as "windows" specific? Then perhaps bake in the macos
signing logic into this as well. The default implementation could be
using the native OS tools while there is a configure option to
completely replace it.
/Erik
Thanks,
Frederic
[1] https://github.com/openjdk/jdk/pull/23807
On 26/02/2025 15:29, Erik Joelsson wrote:
Hello,
I think this is a reasonable idea and you are welcome to implement
it. At Oracle we have been solving this using custom extension
makefiles. That will unfortunately not change with this proposal as
our signing process does not involve direct access to the native
tools of the OS. Perhaps this implementation could be made
plug-able enough so that we could leverage it, I'm not sure. What
I'm trying to say is that we are not against this, but we are also
not inclined to invest resources into it.
/Erik
On 2/26/25 12:52 AM, Frederic Thevenet wrote:
Hi,
I would like to start a discussion about adding built-in support
for code signing native executable file and dynamic library on
Windows directly as part of the JDK build, in a similar fashion to
what already exists for macOS.
Most, if not all vendors already ensure that every native
executable files and dynamic libraries that are part of the binary
builds of OpenJDK that they distribute to their customers for the
Windows platform are digitally signed using a set of Microsoft
APIs and tools called Authenticode[1].
Since this is not a part of the existing JDK build system,
however, this means that each vendor has had to come up with their
own way to integrate the code signing step into their build pipeline.
As the shape of the JDK binary deliverable evolved to accommodate
features like modules, signing binaries as an after-the-fact
process has gradually become more complicated and error prone.
For instance, with the introduction of JEP 493 (Linking Run-Time
Images without JMODs)[2], adding a digital signature to files
after jlink runs as part of the build to create the JDK image when
the new '--generate-linkable-runtime' feature is enabled will
cause any subsequent uses of jlink from this image to fail with a
"Error:<xxxx> has been modified" error.
The solution to that particular problem is to ensure that the
signature is applied before jlink generates the checksums it will
later use to detect , and I believe the best way to achieve that
is to integrate code signing as an option in the jdk build process.
This would also offer vendors who decide to opt into using this
new feature to potentially simplify their overall build process.
For instance, this could get rid of extra steps such as filtering
out dlls that are part of the bundle but were not built as part of
the jdk (e.g. the Microsoft C and C++ runtime redistributables),
or making sure that 'exe' and 'dll' files packaged in the .jmod
files are also signed.
As a stated before a very similar option already exists for code
signing on the macOS platform, which further reinforce my thinking
that such a feature would make sense for Windows as well.
I have drafted a prototype for this and would be happy to create
JBS issue and submit a PR to move this forward, provided there is
interest.
Thanks,
Frederic Thevenet (fthevenet)
[1]
https://learn.microsoft.com/en-us/previous-versions/windows/internet-explorer/ie-developer/platform-apis/ms537359(v=vs.85)
[2] https://openjdk.org/jeps/493
--
Frederic Thevenet
Senior Software Engineer - OpenJDK
Red Hat France <https://www.redhat.com>
BAF5 C2D2 0BE0 1715 5EE1 0815 2065 AD47 B326 EB92
