[ +acme ]
On 03/06/2018 05:00 PM, David Miller wrote:
From: Jiri Benc <redacted>
Date: Tue, 6 Mar 2018 16:03:25 +0100
quoted
On Tue, 6 Mar 2018 15:39:07 +0100, Daniel Borkmann wrote:
quoted
Thanks for the fix, Jiri! The standard approach to resolve such header dependencies under
tools/ would be to add a copy of magic.h uapi header into tools/include/uapi/linux/magic.h.
Both bpftool and libbpf have tools/include/uapi/ in their include path from their
Makefile, so they would pull this in automatically and it would also allow to get rid
of the extra ifdef in libbpf then. Could you look into that?
That's what I tried at first. But honestly, this is a shortcut to hell.
Eventually, we'd end up with most of uapi headers duplicated under
tools/include/uapi and hopelessly out of sync.
The right approach would be to export uapi headers from the currently
built kernel somewhere (a temporary directory, perhaps) and use that to
build the tools. We should not have duplicated and out of sync headers
in the kernel tree. Just look at the git log for tools/include/uapi to
see what I mean by "out of sync".
I understand your frustration.
I'm really puzzled why doing "make headers_install" and then building
these tools does not pick those in-kernel headers up. That's what
really should happen.
Arnaldo, given this came out of tools/perf originally and duplicating/syncing
headers is common practice since about 2014 in kernel git tree, do you have
some context on why the above was/is not considered?
My current understanding is that the general preference would be on copying
the headers into tools/include/ infrastructure once there are dependencies
identified that would be missing on older/local system headers rather than
ifdef'ery of various bit and pieces in the code that need to make use of them.
Would be good to get some clarification on that in any case.
But that said, I'd also be fine taking the three-liner as is into bpf as a fix.
The kernel tree internally should be self-consistent.
It's one thing for an external tool like iproute2 to duplicate stuff
like this, but user programs inside the kernel have no excuse for
requiring things like that just to build.