[RFC PATCH] bpf: preload: Fix build error when O= is set

[David Gow]

On Wed, Dec 16, 2020 at 10:53 PM Quentin Monnet <quentin at isovalent.com> wrote:
>
> 2020-11-21 17:48 UTC+0800 ~ David Gow <davidgow at google.com>
> > On Sat, Nov 21, 2020 at 3:38 PM Andrii Nakryiko
> > <andrii.nakryiko at gmail.com> wrote:
> >>
> >> On Thu, Nov 19, 2020 at 12:51 AM David Gow <davidgow at google.com> wrote:
> >>>
> >>> If BPF_PRELOAD is enabled, and an out-of-tree build is requested with
> >>> make O=<path>, compilation seems to fail with:
> >>>
> >>> tools/scripts/Makefile.include:4: *** O=.kunit does not exist.  Stop.
> >>> make[4]: *** [../kernel/bpf/preload/Makefile:8: kernel/bpf/preload/libbpf.a] Error 2
> >>> make[3]: *** [../scripts/Makefile.build:500: kernel/bpf/preload] Error 2
> >>> make[2]: *** [../scripts/Makefile.build:500: kernel/bpf] Error 2
> >>> make[2]: *** Waiting for unfinished jobs....
> >>> make[1]: *** [.../Makefile:1799: kernel] Error 2
> >>> make[1]: *** Waiting for unfinished jobs....
> >>> make: *** [Makefile:185: __sub-make] Error 2
> >>>
> >>> By the looks of things, this is because the (relative path) O= passed on
> >>> the command line is being passed to the libbpf Makefile, which then
> >>> can't find the directory. Given OUTPUT= is being passed anyway, we can
> >>> work around this by explicitly setting an empty O=, which will be
> >>> ignored in favour of OUTPUT= in tools/scripts/Makefile.include.
> >>
> >> Strange, but I can't repro it. I use make O=<absolute path> all the
> >> time with no issues. I just tried specifically with a make O=.build,
> >> where .build is inside Linux repo, and it still worked fine. See also
> >> be40920fbf10 ("tools: Let O= makes handle a relative path with -C
> >> option") which was supposed to address such an issue. So I'm wondering
> >> what exactly is causing this problem.
> >>
> > [+ linux-um list]
> >
> > Hmm... From a quick check, I can't reproduce this on x86, so it's
> > possibly a UML-specific issue.
> >
> > The problem here seems to be that $PWD is, for whatever reason, equal
> > to the srcdir on x86, but not on UML. In general, $PWD behaves pretty
> > weirdly -- I don't fully understand it -- but if I add a tactical "PWD
> > := $(shell pwd)" or use $(CURDIR) instead, the issue shows up on x86
> > as well. I guess this is because PWD only gets updated when set by a
> > shell or something, and UML does this somewhere?
> >
> > Thoughts?
> >
> > Cheers,
> > -- David
>
> Hi David, Andrii,
>
> David, did you use a different command for building for UML and x86? I'm
> asking because I reproduce on x86, but only for some targets, in
> particular when I tried bindeb-pkg.

I just ran "make ARCH={x86,um} O=.bpftest", with defconfig + enabling
BPF_PRELOAD and its dependencies. UML fails, x86 works. (Though I can
reproduce the failure if I make bindeb-pkg on x86).

(It also shows up when building UML with the allyesconfig-based KUnit
alltests option by running "./tools/testing/kunit/kunit.py run
--alltests", though this understandably takes a long time and is less
obvious)
>
> With "make O=.build vmlinux", I have:
> - $(O) for "dummy" check in tools/scripts/Makefile.include set to
> /linux/.build
> - $(PWD) for same check set to /linux/tools
> - Since $(O) is an absolute path, the "dummy" check passes
>
> With "make O=.build bindeb-pkg", I have instead:
> - $(O) set to .build (relative path)
> - $(PWD) set to /linux/.build
> - "dummy" check changes to /linux/.build and searches for .build in it,
> which fails and aborts the build
>
> (tools/scripts/Makefile.include is included from libbpf's Makefile,
> called from kernel/bpf/preload/Makefile.)
>
> I'm not sure how exactly the bindeb-pkg target ends up passing these values.

Yeah: I haven't been able to find where uml is changing them either:
I'm assuming there's something which changes directory and/or spawns a
shell/recursive make to change $(PWD) or something.

> For what it's worth, I have been solving this (before finding this
> thread) with a fix close to yours, I pass "O=$(abspath .)" on the
> command line for building libbpf in kernel/bpf/preload/Makefile. It
> looked consistent to me with the "tools/:" target from the main
> Makefile, where "O=$(abspath $(objtree))" is passed (and $(objtree) is ".").

Given that there are several targets being broken here, it's probably
worth having a fix like this which overrides O= rather than trying to
hunt down every target which could change $(PWD). I don't particularly
mind whether we use O= or O=$(abspath .), both are working in the UML
usecase as well.

Does anyone object to basically accepting either this patch as-is, or
using O=$(abspath .)?


Cheers,
-- David