Bug#924891: glibc: FTBFS: /<<PKGBUILDDIR>>/build-tree/amd64-libc/conform/UNIX98/ndbm.h/scratch/ndbm.h-test.c:1:10: fatal error: ndbm.h: No such file or directory
On 27/03/19 at 23:59 +0100, Florian Weimer wrote:
> retitle 924891 glibc: misc/tst-pkey fails due to cleared PKRU register after signal in amd64 32-bit compat mode
> thanks
>
> * Lucas Nussbaum:
>
> > On 27/03/19 at 08:48 +0100, Florian Weimer wrote:
> >> > If that's useful, I can easily provide access to an AWS VM to debug this
> >> > issue.
> >>
> >> Oh, that would be quite helpful indeed.
> >
> > Can you send your SSH key? (I thought there was a way to get the SSH key
> > for a DD, but I cannot find it anymore)
> >
> > Then you will be able to ssh to root@18.184.55.40.
> > There's sbuild and schroot setup on the VM.
> >
> > When you are done, please 'poweroff' the machine, which will terminate
> > it.
>
> The issue reproduces outside the chroot, with the stretch userland.
>
> What happens is that once we get out of the SIGUSR1 signal handler,
> the PKRU register has value zero. This happens around this code in
> the test:
>
> /* Check that in a signal handler, there is no access. */
> xsignal (SIGUSR1, &sigusr1_handler);
> xraise (SIGUSR1);
> xsignal (SIGUSR1, SIG_DFL);
> TEST_COMPARE (sigusr1_handler_ran, 1);
>
> I checked the following (via a breakpoint in pkey_get; I don't think
> GDB can read the PKRU register directly): Inside the SIGUSR1 signal
> handler, PKRU has value 0x55555554, as expected for this kernel, but
> after the return, we get zero. This is the first time a signal is
> delivered on the main thread, so it's consistent with fairly broken
> signal handling as far as the PKRU register is concerned. I guess
> clearing PKRU in this way might even constitute a minor security bug
> (because the zero value means no restrictions).
>
> This commit looks highly relevant:
>
> commit a4455082dc6f0b5d51a23523f77600e8ede47c79
> Author: Dave Hansen <dave.hansen@linux.intel.com>
> Date: Wed Jun 8 10:25:33 2016 -0700
>
> x86/signals: Add missing signal_compat code for x86 features
>
> The 32-bit siginfo is a different binary format than the 64-bit
> one. So, when running 32-bit binaries on 64-bit kernels, we have
> to convert the kernel's 64-bit version to a 32-bit version that
> userspace can grok.
>
> If the siginfo_t layout is incorrect (with regards to what the
> hardware writes), I expect that we might end up copying back the wrong
> PKRU value.
>
> I'm not sure what to do here. This really looks like a kernel bug.
> Maybe we should just verify that this is fixed in the buster kernel
> and move on?
>
> Lucas, can you run your rebuild tests on newer kernels?
Indeed. I upgraded the kernel to the stretch-backports one, and glibc
builds fine.
Lucas
Reply to: