Re: [oe] INCOMPATIBLE_LICENSES and WHITELIST_<license> usage


Peter Kjellerstedt
 

-----Original Message-----
From: Richard Purdie <richard.purdie@...>
Sent: den 19 februari 2022 01:45
To: Peter Kjellerstedt <peter.kjellerstedt@...>; Saul Wold
<Saul.Wold@...>; openembedded-
architecture@...; OE-core <openembedded-
core@...>; OpenEmbedded Devel List <openembedded-
devel@...>
Subject: Re: [oe] INCOMPATIBLE_LICENSES and WHITELIST_<license> usage

On Fri, 2022-02-18 at 23:34 +0000, Peter Kjellerstedt wrote:
Warning: wall of text ahead. Sorry about that, but I believe it is
important that we get this right if we are redesigning it.

TL;DR: see the end for the new suggested licensing variables.

-----Original Message-----
From: openembedded-devel@... <openembedded-
devel@...> On Behalf Of Richard Purdie
Sent: den 18 februari 2022 15:14
To: Saul Wold <Saul.Wold@...>; openembedded-
architecture@...; OE-core <openembedded-
core@...>; OpenEmbedded Devel List <openembedded-
devel@...>
Subject: Re: [oe] INCOMPATIBLE_LICENSES and WHITELIST_<license> usage

On Thu, 2022-02-17 at 15:01 -0800, Saul Wold wrote:
I am working on a proposal to re-write how INCOMPATIBLE_LICENSES is
used and processed to possibly include a COMPATIBLE_LICENSES variable
as well, see PeterK's email [0]

I am trying to determine the usage of WHITELIST_<license> which
would be used to override a license that might be listed in
INCOMPATIBLE_LICENSES variable.

Randy and I have done a quick and dirty survey of a 100 or so layers
(thanks Randy) and could not find any real usage other than what's
currently in OE-Core for WHITELIST_GPL-3.0.

If you are using WHITELIST_<license>, please let me reply with your
usage.

[0] https://lists.openembedded.org/g/openembedded-devel/message/95166
We need to be mindful that we need to resolve this to unblock the
other language changes and feature creep here is potentially
problematic. I do think it is worth trying to improve things rather
than blindly allowing the horrible syntax in this variable to
continue though.

The test case we have for this currently is:

WHITELIST_GPL-3.0:pn-core-image-minimal = "bash"

so I'd wondered about an alternative of:

INCOMPATIBLE_LICENSE_EXCEPTIONS:pn-core-image-minimal = "bash:GPL-3.0"
You do not really need the license here (more than possibly as
information to anyone reading the code). Specifying that a package is
allowed in an image regardless of its licenses can just as well be
done by allowing the variable to take a list of packages:

INCOMPATIBLE_LICENSE_EXCEPTIONS:pn-core-image-minimal = "bash readline"
I'm not sure I agree with that. I get the feeling some people want to
allow these packages as long as they're with license X. Licenses can
change over time and this allows a safety net that if it changes, the
exception has to be updated too.
Well, to complicate matters then, if you really want to be explicit
with the license exceptions you make for a recipe or a package, I will
argue that you need to do it for all incompatible licenses a recipe
uses. With today's code, as soon as you allow one incompatible license
for a recipe, the recipe is allowed to be built, even if it has
specified multiple incompatible licenses. E.g., if we take the gnupg
recipe, which specifies LICENSE = "GPLv3 & LGPLv3", and we have an
INCOMPATIBLE_LICENSE = "GPL-3.0-only LGPL-3.0-only", with your example
I would then expect the following:

INCOMPATIBLE_LICENSE_EXCEPTIONS:pn-core-image-minimal = "gnupg:GPL-3.0-only gnupg:LGPL-3.0-only"

to allow the gnupg package to be installed in the image. However, I
guess you could combine our suggestions by making the license part
of the value optional, i.e., the following would then also allow the
package to be included under the same conditions:

INCOMPATIBLE_LICENSE_EXCEPTIONS:pn-core-image-minimal = "gnupg"

An alternative (and in my opinion better) variable name would be:

IMAGE_ALLOW_PACKAGES:pn-core-image-minimal = "bash readline"

focusing on that this is a list of _packages_ allowed for a given
_image_. This does not explicitly mention licenses, but I do not
believe that is needed.
If you assume this isn't being done on license reasons, sure. Except
see above, I think this does need to account for licenses, at least
the way many use it.

After all there could be multiple reasons a package is not allowed in
an image and this variable would allow to
ignore them all because that is typically what you want to do:
specify that you really want that package in that image. I guess
this is in some sense the opposite of PACKAGE_EXCLUDE. And I guess
like :append vs :remove, if someone for some reason specifies a
package in both IMAGE_ALLOW_PACKAGES and PACKAGE_EXCLUDE, then the
later should win (to err on the side of caution).
How would IMAGE_ALLOW_PACKAGES work and be different to IMAGE_INSTALL?
Well, IMAGE_INSTALL determines what packages you want in your image,
while IMAGE_ALLOW_PACKAGES allows packages to be installed that would
otherwise render an error if you tried. To me it seems perfectly
natural, but I am probably a bit biased. ;)

This interface would likely confuse more users than it would help.
Well, it is really based on that only package names need to be specified
and not the licenses. If both are needed (as you believe), then this name
makes less sense.

For the case where you want to allow a recipe to be built despite
it using licenses that are otherwise not allowed you can simply use

INCOMPATIBLE_LICENSE:pn-foo = ""

and don't really need a separate variable for it.
That is a good point and perhaps should influence how an
INCOMPATIBLE_LICENSE_EXCEPTIONS should be package based rather than
recipe. I doubt the pn- override existed when that variable was
originally added.
I definitely believe that we need to separate the case for building
a recipe from the case for installing a package into an image. For the
recipe case, I believe that manipulating RECIPE_INCOMPATIBLE_LICENSES
(or whatever the variable ends up being called) using normal variable
manipulations should be enough, without having to introduce yet
another variable. I.e., my example above would correspond to giving
a blanket exception for all licenses for the given recipe, while a
case where a specific license is excepted would be something like
RECIPE_INCOMPATIBLE_LICENSES:remove:pn-foo = "GPL-3.0-only"

For the installing a package into an image case, I don't see a way
to avoid an extra variable as discussed above. However, the name
INCOMPATIBLE_LICENSE_EXCEPTIONS gives no indications that it is
image specific, which I think is bad. If we introduce the other
names I suggest below, then this would instead map to
IMAGE_INCOMPATIBLE_LICENSES_EXCEPTIONS (though I dislike that name
because it is so long, but maybe it is inevitable).

I'm still of the opinion the AVAILABLE_LICENSES variable is something
we should be aiming to remove ultimately too, it has horrible issues
with layers changing hashes for all recipes.
Since I was the one to introduce it, I will answer to that. It was
introduced so that it is possible to specify compatible licenses
instead of incompatible licenses by basically calculating the
incompatible licenses by taking the set of available licenses minus
the set of compatible licenses. If a mechanism to do that is
introduced, e.g., by adding support for a COMPATIBLE_LICENSES in
addition to INCOMPATIBLE_LICENSE, then I am of the opinion that we
can remove AVAILABLE_LICENSES again.

We also need this mechanism in the code for handling allowed vs
disallowed licenses because the current code really cannot handled a
list of many hundreds of incompatible licenses, which is what we got
after all SPDX licenses were added to OE-Core. The code is extremely
inefficient and evaluates the list of incompatible licenses over and
over again. In our case that meant the recipe parsing time tripled.
Ironically, part of the reason I want to change the design of
WHITELIST_XXX is that it forces code that doesn't perform well. You
may be surprised to find that this change actually improves that 3x
issue you've seen. If not, I think it would lead us in a direction
where it can certainly help.

That said, we really need two sets of variables. One for blocking
the building of recipes because of its license(s), and one for
preventing packages with disallowed licenses to be included in a
given image. These are very different use cases and they take
totally different lists of items (recipes vs packages). The current
mess where the same variables are used for both cases needs to be
resolved.

So why do we need both cases? The first case where recipes are
prevented from being built makes it possible to, e.g., prevent a
newer versioned recipe that uses GPL-3.0-or-later to be built and
instead build an older version that uses GPL-2.0-or-later (i.e.,
what meta-gplv2 can be used for today).

The second case allows to build code that uses disallowed licenses,
as long as the packages are not added to the image. Why is this
useful? Because many recipes are built for some packages that are
never used in the given image, and it is then much easier to allow
them to be built as long as their output is not used. E.g., very
many recipes depend on bash which is GPL-3.0-or-later, making it
near impossible to avoid having to build it. However, it is
perfectly possible to build production images that do not need bash
to be installed.

I believe for this second case we should have two variables,
IMAGE_COMPATIBLE_LICENSES and IMAGE_INCOMPATIBLE_LICENSES. And to
make the naming clear for the first case, I would suggest calling
those variables RECIPE_COMPATIBLE_LICENSES and
RECIPE_INCOMPATIBLE_LICENSES.

Also note that the use of RECIPE_COMPATIBLE_LICENSES and
RECIPE_INCOMPATIBLE_LICENSES is mutually exclusive, as is the use
of IMAGE_COMPATIBLE_LICENSES and IMAGE_INCOMPATIBLE_LICENSES. I.e.,
you will have to choose whether to specify what licenses to allow
or what licenses to disallow. You cannot do both. This is because
specifying a list of compatible licenses means that all other
licenses are by definition incompatible, and vice versa. However,
this also means that it makes perfect sense to be able to, e.g.,
specify a few RECIPE_INCOMPATIBLE_LICENSES together with a list of
IMAGE_COMPATIBLE_LICENSES.
I suspect there will be people who will want configurations where they
specify that they are happy with X and definitely don't want Y to work.
The unspecified licenses would effectively be compatible but not be
explicitly specified. It seems like an odd setup but I can imagine
people configuring it and wanting it to work.

I am very worried about the performance implications in this "every
license" in a compatible list. Obviously for the case your legal
department cares about, you have to do it but I'm not sure we want
to force it onto everyone (and AVAILABLE_LICENSES already heads in
that direction).
By splitting the list of licenses into two, to handle the opposite
use cases of allowed vs disallowed licenses, we avoid the need to know
the full list of potential licenses as we only need to know what is
allowed or disallowed and can exclude anything else.

Thus the expectation is that the lists in either *_COMPATIBLE_LICENSES
or *_INCOMPATIBLE_LICENSES will be relatively short, compared to lists
based on all licenses. In our case, we have 80 licenses in our list
of allowed licenses. That is something that I think even the current
code can handle reasonably well. And without pattern matching and
mappings of old license names, the code would basically have to do
something like:

compatible_licenses = (d.getVar('COMPATIBLE_LICENSES') or '').split()
incompatible_licenses = (d.getVar('INCOMPATIBLE_LICENSES') or '').split()
disallowed_licenses = set()
...
if compatible_licenses and license not in compatible_licenses or \
incompatible_licenses and license in incompatible_licenses:
disallowed_licenses.add(license)

Also, based on the assumption that in most cases you want to be able to
build as much as possible, I would guess that you would typically use
RECIPES_INCOMPATIBLE_LICENSES to exclude only a select few licenses that
cannot be used at all (e.g., GPL-3.0), and then use
IMAGE_COMPATIBLE_LICENSES or IMAGE_INCOMPATIBLE_LICENSES to limit what
goes into the images. Thus the longer lists would only affect the image
recipes compared to all recipes as it is today.

Oh, and another thing I would like to take the opportunity to raise
is whether we should continue to support patterns in these list, or
if we should change it so that the lists of licenses need to
explicitly specify all licenses. The latter would greatly simplify
the code, especially when combined with only allowing SPDX names.
If we decide to remove support for patterns, and based on the
assumption that the pattern is typically used to specify "*GPL-3.0*",
we could make available a variable or two that contain the typical
sets of GPL licenses. This would also allow us to remove the code
that handles how an or-later licenses specified as '<license>+'
should be treated in combination with patterns.
I'd love to remove it but it is something which people want and now
expect from the code. You might not like it, others do. How do we
please everyone? I don't think we'd be able to remove that, only
perhaps limit it a little more.
But if we redesign the license handling anyway, can't we change their
expectations at the same time? There really aren't that many groups
of SPDX licenses where it is useful to use pattern matching anyway
and that you are likely to put in these variables. Basically I think
it comes down to some set of GPL licenses. I.e., replacing "*GPL-3.0*"
with the corresponding licenses really doesn't yield a very long list.
And as I suggested above, we could even provide a variable or two with
the expanded lists of the most typical use cases. E.g., if we added a
GPL3_LICENSES variable, the INCOMPATIBLE_LICENSES = "*GPL-3.0*" would
become INCOMPATIBLE_LICENSES = "${GPL3_LICENSES}". Yes, it is a little
less flexible, but it would do wonders for the code...

Anyway, this is a minor gripe compared to the rest, so if we can't
get rid of it, so be it.

So, to reiterate, these are the variables that I suggest we adopt
to be able to handle the various use cases regarding licenses:

RECIPE_COMPATIBLE_LICENSES - list of compatible licenses for
recipes
RECIPE_INCOMPATIBLE_LICENSES - list of incompatible licenses for
recipes
IMAGE_COMPATIBLE_LICENSES - list of compatible licenses for
packages in images
IMAGE_INCOMPATIBLE_LICENSES - list of incompatible licenses for
packages in images
IMAGE_ALLOW_PACKAGES - list of packages allowed in the
image regardless of licenses and
other restrictions

This also means that the old WHITELIST_*, INCOMPATIBLE_LICENSE
and AVAILABLE_LICENSES variables are removed. I also suggest we
remove the support for patterns in the new variables.
Whilst I don't just want to map WHITELIST_* to something renamed, I think
the above is a bit too radical to get into this late in the release cycle
so I suspect we'll have to compromise. This discussion needed to happen
early in the cycle with people actively working on it.
And I really, really wish I had had the time then to do exactly that. :(
I know I even said I would after I noticed the problem with the prolonged
recipe parsing times due to the huge increase in available licenses, but
alas time wasn't on my side. So I did what anyone would do and made a
workaround and went on to handle more pressing problems. I am very sorry
about that.

Just to be really clear, in particular I detest IMAGE_ALLOW_PACKAGES as a
variable name. I think you're trying to be too wide in scope with "other
restrictions" and it will just confuse people.
Ouch, and I thought it was pretty ok. :( Especially compared to
INCOMPATIBLE_LICENSE_EXCEPTIONS, which I don't like very much (though that
is mostly based on its length). Anyway, this has been discussed at length
above, so I will not delve on it more here.

Cheers,

Richard
//Peter

Join openembedded-core@lists.openembedded.org to automatically receive all group messages.