On 3/9/19 7:11 AM, Greg KH wrote:
There is no licensing issue, see my follow-up comment about that.
It's all in ease-of-use here. You want to build a trace function against a running kernel, and now you have the header files for that specific kernel right there in the kernel itself to build against. It doesn't get easier than that.
Agreed.
It seems that opinions in this thread are split along two conceptions of a "Linux system". On one side there's that of the conventional "Linux distro" where the entity generating the distro has full control over the entire process and the parts thereof. On the other side there's the world that has evolved out of the multi-party ecosystem that Android fostered. In the later, there isn't a single party that controls all aspects of the "distro". Instead, there are a multitude of parties contributing to creating a fully-functional Linux-based system.
In the latter ecosystem, the kernel and the filesystems (*plural*) used with it do not necessarily come from the same place, are maintained by the same parties or even required to be in lock-step. For all I know, the filesystem images are coming from one party and the kernel is at one point from one party and at another point can be substituted, possibly for testing purposes, without userspace images ever changing. No licensing issues at all involved, either with regards to the headers or the distribution of the kernel itself, which, in as far as I've seen when speaking of known industry players (those that matter here), are always ultimately available in source from somewhere.
Instead the issue is that I want to be able to give user-space access to the headers that were used to build the kernel that they're running over, regardless of where this kernel came from. And since I have to assume that those two parts (kernel vs. user-space) are coming from separate parties and can be arbitrarily replaced, there needs to be some form of "contract" between them. The kernel's ABI already provides quite a bit of that and, as Joel pointed out, Google can enforce a few more things, such as default kernel configs, to "make things work". But one option that doesn't exist in the world Android evolves in is to somehow ensure that the filesystem images somehow have kernel-specific tool-required artifacts for every possible kernel they may run against. That's just not possible.
That's especially true in the case of modern-day Android where the final system is made of at least half a dozen filesystem images with each image possibly coming from a different party. The /vendor partition (with the hardware enablement), for example, could be coming from the device vendor while the /system partition (with the Android framework), for example, could be coming straight from Google in the form of "Generic System Image" -- an image meant to run as-is on any Treble-compliant system. /system might have the tools for eBPF, but /vendor might have modules, while still the "boot" partition might have the kernel. There's no reason I can't replace the boot partition *without* changing /vendor. And there's no reason I can't replace /vendor *without* replacing the boot partition. And all other combinations with all other images.
That, in my view, is a big part of the problem Joel's patch solves: in a system whose functionality requires multiple *independent* parties to work together, I can still get the necessary kernel headers for user-space tools to properly operate regardless of which part of the system id being substituted or replaced.
Cheers,