Date:   Tue, 14 Feb 2017 07:59:38 -0600
From:   Josh Poimboeuf <jpoimboe@...hat.com>
To:     Jessica Yu <jeyu@...hat.com>
Cc:     Marc MERLIN <marc@...lins.org>, linux-kernel@...r.kernel.org
Subject: Re: Please turn "Cannot use CONFIG_STACK_VALIDATION" into build error

On Mon, Feb 13, 2017 at 06:08:52PM -0500, Jessica Yu wrote:
> +++ Josh Poimboeuf [13/02/17 12:41 -0600]:
> > On Mon, Feb 13, 2017 at 12:07:09AM -0800, Marc MERLIN wrote:
> > > Hi Josh,
> > > 
> > > I'll start with the story as to why.
> > > i've lost more hours than I care to list, because I was unable to build
> > > the virtualbox kernel driver with newer kernels.
> > > Sadly, it gives no useful debug info outside of
> > > make[1]: *** No rule to make target '/tmp/vbox.0/linux/SUPDrv-linux.o', needed by '/tmp/vbox.0/vboxdrv.o'.  Stop.
> > > 
> > > It took some pretty deep debugging to finally see this:
> > >      Trying rule prerequisite 'tools/objtool/objtool'.
> > >      Looking for a rule with intermediate file 'tools/objtool/objtool'.
> > >       Avoiding implicit rule recursion.
> > > which look quite inoccuous and don't look as errors at all.
> > > When I filed a bug with the vbox folks, they were unable to find out why
> > > the module refused to build on my kernel, and I was stuck with older
> > > kernels as a result.
> > > 
> > > Then, I had another module, bbswitch, to turn off the nvidia chip on my
> > > laptop to save battery. That one also failed to build with newer
> > > kernels, but thankfully made it more clear that the problem was related
> > > to tools/objtool/objtool missing.
> > > 
> > > But why was it missing? No idea...
> > > I trace that down to CONFIG_STACK_VALIDATION which there seems to be no
> > > menu option for, so I manually disable it in .config, rebuild, and it's
> > > automatically re-enabled. Gah.
> > 
> > Hm, that doesn't sound right.  Nothing automatically enables
> > CONFIG_STACK_VALIDATION.  It should be disabled unless manually enabled.
> > Maybe you got it confused with CONFIG_HAVE_STACK_VALIDATION, which is
> > always enabled?
> > 
> > BTW, there is a config option for it in the menu:
> > 
> >  Kernel hacking
> >    Compile-time checks and compiler options
> >      Compile-time stack metadata validation
> > 
> > > More hair pulling, and finally I make a typo
> > > saruman:/usr/src/linux-block# make xonfig
> > > Makefile:1044: "Cannot use CONFIG_STACK_VALIDATION, please install libelf-dev, libelf-devel or elfutils-libelf-devel"
> > > scripts/kconfig/conf  --silentoldconfig Kconfig
> > > Makefile:1044: "Cannot use CONFIG_STACK_VALIDATION, please install libelf-dev, libelf-devel or elfutils-libelf-devel"
> > > make: *** No rule to make target 'xonfig'.  Stop.
> > > 
> > > Sure enough, this was my problem, but I never saw the error message
> > > because I build kernels with
> > > make-kpkg --revision 1gandalf kernel-image
> > > which does other stuff and hid that warning, which really should have
> > > been a fatal error in my opinion.
> > > 
> > > Given that
> > > 1) CONFIG_STACK_VALIDATION seems silently auto enabled.
> > > 2) without libelf-dev, the kernel will build but will leave a tree
> > > missing objtool, which in turn causes (all?) 3rd party modules to fail
> > > building.
> > 
> > Yes, this is a bug.
> > 
> > > 3) and that it's kind of non trivial to find out why if that happens,
> > > 
> > > Would you consider making
> > > "Cannot use CONFIG_STACK_VALIDATION, please install libelf-dev, libelf-devel or elfutils-libelf-devel"
> > > a build error as opposed to a warning?
> > > This sure would have saved me countless errors of debugging the wrong
> > > things.
> > 
> > Correct me if I'm wrong, but it sounds like make-kpkg suppressed stderr?
> > If so, that should be fixed.
> > 
> > When I try to build an OOT module with CONFIG_STACK_VALIDATION enabled
> > and elfutils-libelf-devel missing (on Fedora), I get:
> > 
> >  make: Entering directory '/home/jpoimboe/git/linux'
> >  make[1]: Entering directory '/home/jpoimboe/ktest/output'
> >    CC [M]  /home/jpoimboe/livepatch-test/1/livepatch2.o
> >  /bin/sh: ./tools/objtool/objtool: No such file or directory
> >  /home/jpoimboe/git/linux/scripts/Makefile.build:300: recipe for target '/home/jpoimboe/livepatch-test/1/livepatch2.o' failed
> >  make[2]: *** [/home/jpoimboe/livepatch-test/1/livepatch2.o] Error 1
> >  /home/jpoimboe/git/linux/Makefile:1490: recipe for target '_module_/home/jpoimboe/livepatch-test/1' failed
> >  make[1]: *** [_module_/home/jpoimboe/livepatch-test/1] Error 2
> >  make[1]: Leaving directory '/home/jpoimboe/ktest/output'
> >  Makefile:150: recipe for target 'sub-make' failed
> >  make: *** [sub-make] Error 2
> >  make: Leaving directory '/home/jpoimboe/git/linux'
> > 
> > It's not a perfect error message, but the
> > 
> >  '/bin/sh: ./tools/objtool/objtool: No such file or directory'
> > 
> > is at least a big clue.  I'm curious why you didn't see that.
> > 
> > Anyway, the above libelf-dev warning is just a warning and not a build
> > error because CONFIG_STACK_VALIDATION is enabled for allyesconfig, and
> > it's not a severe enough problem to warrant breaking the build.
> > 
> > Ideally the same warning should be printed when building OOT modules.
> > I'll try to figure out if there's a way to do that it.
> 
> Btw, it looks like the libelf warning is inside an `ifeq ($(KBUILD_EXTMOD),)`
> block, so it does not get applied to OOT modules. It would be
> possible to add the warning to the corresponding KBUILD_EXTMOD (else)
> block or somewhere else in the Makefile common to both cases (probably
> more preferable). I tried the latter case and the warning prints for
> all cases (vmlinux, intree, external modules).

Aha, that explains it, thanks.  Will look into fixing it.

-- 
Josh

Powered by blists - more mailing lists