On Wed, Dec 13, 2023 at 08:58:06AM +0500, Muhammad Usama Anjum wrote:
> On 12/13/23 7:14 AM, John Hubbard wrote:

> > Oh, this sounds like it would work nicely. No more "make headers"
> > required (hooray!). Instead, the new approach would be "selftests are
> > allowed to include from tools/include", and then we can just start
> > copying the files that we need to that location, and gradually fix up
> > all the selftests.

> No, this wouldn't work.

Note that we have a bunch of selftests (at least arm64, hid, kvm, rseq
and sgx from a quick grep) which already use and rely on the headers in
tools/include.

> * The selftests are applications which include default header files. The
> application don't care from where the header files are picked up at compile
> time. We should be able to build the application on normal system with
> latest headers installed without any changes.

I think there is much less interest in building out of the kernel than
there is in avoiding having to handle random userspace headers...

> * The header files cannot be included directly as they need to be processed
> first which is done by `make headers`. Here is a diff between kernel fs.h
> and processed header file to be used by applications:

I guess that's another reason why the sync is done manually.  There are
also a bunch of files in tools/include that are just completely
different implementations of things (not just uapi).