Re: efficacy of MODULE_LICENSE

[Greg KH <greg@kroah.com>]

On Wed, Jul 10, 2019 at 07:41:03AM -0600, J Lovejoy wrote:
> 
> 
> > On Jul 10, 2019, at 3:38 AM, Greg KH <gregkh@linuxfoundation.org> wrote:
> > 
> > On Tue, Jul 09, 2019 at 10:28:59PM -0600, J Lovejoy wrote:
> >> Hi all,
> >> 
> >> We seem to have gone a bit quiet recently! Hopefully that’s just a
> >> symptom of nicer weather and holiday season, but we can still pick up
> >> some momentum :)
> >> 
> >> I wanted to get your input on the MODULE_LICENSE tag, which I have
> >> found to be a bit vexing in some instances. I am finding examples
> >> where there is a clearly identifiable license in the file, for example
> >> ISC, and then the MODULE_LICENSE tag is something like "Dual BSD/GPL”.
> >> There is absolutely no other reference to GPL whatsoever (or any BSD
> >> variant for that matter).
> > 
> > MODULE_LICENSE is used by the kernel itself, at runtime, to determine
> > the "license" of the module that is being loaded into it.
> > 
> > At that point in time, it is a dual-licensed chunk of code, as it
> > incorporated gplv2 bits into it in order to create that module image,
> > right?
> 
> by dual-licensed, I’m assuming you mean conjunctive (“AND”) by way of the incorporated GPL-2.0 bits into the ISC-licensed file, then yes, that sounds right.
> But this file, by itself, is just ISC for the license as the “incorporating” part only happens at runtime, right?
> 
> so for purposes of identifying the license of the file (and if someone wanted to re-use that file elsewhere), ISC would be the operative license info.
> > 
> >> Based on my understanding of
> >> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/module.h#n172
> >> <https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/module.h#n172>
> >> - the MODULE_LICENSE info was never meant to be definitive license
> >> info, but seemingly more of an approximation.  I’m wondering if others
> >> have a different view?
> > 
> > It is used at runtime to determine if the module has access to some
> > types of kernel symbols or not.
> > 
> > It can also be used at any time to extract the license from the module
> > image on a disk, you can see this by running the 'modinfo' program on
> > any kernel module:
> > 	$ modinfo visor | grep license
> > 	license:        GPL v2
> 
> “extract the license” is a bit concerning here - as I take that to
> mean the extraction part is the MODULE_LICENSE info, which isn’t
> really accurate. 

It seems pretty accurate for the aggregated work that is the object file
that is loaded into the kernel, right?  How is it not "accurate"?

> or is the intent here not necessarily accuracy, but a threshold
> determination that it’s open source/compatible (and not proprietary or
> something like that)?

That is what we use it for, yes.

It really is a yes/no test that we use to allow access to specific
internal kernel symbols or not.

But, it's also useful for when companies "lie" with that string.
Lawyers generally frown when technical people at a company try to do
"tricks" and fool the kernel by messing with that string.  I know at
least 2 VERY major and huge companies[1] that have turned their
treatment around for this very reason when they realized what they were
"caught" doing, and now are big contributors to our kernel community.

So it is a nice tool to be able to use at times :)

> >> More specifically - where we have specific license match (like the
> >> example above) - we can add the appropriate SPDX identifier, but if we
> >> leave the MODULE_LICENSE info, I suspect that scanners will pick that
> >> up and report a mix of licensing info (e.g., ISC, BSD, GPL, as in my
> >> above example), which kind of brings us to the same place we are now.
> >> Should we also remove the MODULE_LICENSE tag where it contradicts the
> >> actual license info in terms of an exact license match (i.e., there is
> >> nothing to match to GPL here, other than the MODULE_LICENSE tag, but
> >> there is an exact match to a different license, ISC, in this case).
> > 
> > MODULE_LICENSE predated SPDX by a decade or so, and was designed to
> > solve a totally different use case.  I would not try to mix the two, or
> > infer one from the other.
> > 
> > MODULE_LICENSE covers the "resulting image" of combining many different
> > files that can have different SPDX-identified licenses in them.
> > 
> > Does this help any?
> 
> yes. And I can understand the different use case, I guess my
> concern/question is does the existence of MODULE_LICENSE info that
> sort of contradicts the actual license info for the file (when looking
> just at that file, not the combined/resulting image) frustrate the
> goal of having clean licensing info for when people run scans over the
> kernel?

It shouldn't, which is why we clarified a number of these strings
recently with the documentation.  What else needs to be adjusted to make
this more accurate for what you want to use when scanning a binary
image?

Remember, that's what this is used for, the binary image, not for the
individual files.

thanks,

greg k-h