Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
References: <20191223130834.GA102399@zx2c4.com> <20191224135404.389039-1-Jason@zx2c4.com>
In-Reply-To: <20191224135404.389039-1-Jason@zx2c4.com>
From:   Arnd Bergmann <arnd@arndb.de>
Date:   Mon, 30 Dec 2019 12:57:59 +0100
Message-ID: <CAK8P3a1fVFDkHe=gLy55rHxwfZ8YKcUSYvnhSoMbcAgWy6Nm9w@mail.gmail.com>
Subject: Re: [PATCH] mips: vdso: conditionalize 32-bit time functions on COMPAT_32BIT_TIME
To:     "Jason A. Donenfeld" <Jason@zx2c4.com>
Cc:     "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "open list:BROADCOM NVRAM DRIVER" <linux-mips@vger.kernel.org>,
        Paul Burton <paulburton@kernel.org>,
        Vincenzo Frascino <vincenzo.frascino@arm.com>,
        Christian Brauner <christian.brauner@canonical.com>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Tue, Dec 24, 2019 at 2:54 PM Jason A. Donenfeld <Jason@zx2c4.com> wrote:
>
> When the VDSO falls back to 32-bit time functions on kernels with
> COMPAT_32BIT_TIME=n, userspace becomes corrupted and appears to crash
> shortly after, with something like:
>
> [    0.359617] do_page_fault(): sending SIGSEGV to init for invalid read access from 000000007ff790d0
> [    0.359843] epc = 0000000077e45df4 in libc.so[77da6000+de000]
> [    0.360319] ra  = 0000000010000c50 in init[10000000+2000]
> [    0.364456] Kernel panic - not syncing: Attempted to kill init! exitcode=0x0000000b
>
> This can be reproduced with simply calling `clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts)`,
> since `CLOCK_PROCESS_CPUTIME_ID` is not exported to the VDSO, invoking
> the syscall callback branch. This crash was observed with musl 1.20's
> clock_gettime implementation:

Thanks for the bug report! I'm not completely sure why this fails in
this particular
way though. I assume you are using musl-1.1.20, not a musl-1.2.0 snapshot
(the version 1.20 you list does not exist), so the combination you are testing
is supposed to just return -ENOSYS here to match the behavior of hte
system call.

> --- a/arch/mips/include/asm/vdso/gettimeofday.h
> +++ b/arch/mips/include/asm/vdso/gettimeofday.h
> @@ -107,7 +107,7 @@ static __always_inline int clock_getres_fallback(
>        return error ? -ret : ret;
> }
>
> -#if _MIPS_SIM != _MIPS_SIM_ABI64
> +#if _MIPS_SIM != _MIPS_SIM_ABI64 && defined(CONFIG_COMPAT_32BIT_TIME)
>
>  #define VDSO_HAS_32BIT_FALLBACK        1
>

I don't think this is the correct fix, it may actually make it worse
by changing the vdso implementation for clock_gettime32()
to fall back to clock_gettime64(), which would appear to work
correctly before y2038 but fail afterwards.  How about this one:

diff --git a/arch/mips/vdso/vdso.lds.S b/arch/mips/vdso/vdso.lds.S
index da4627430aba..0bdc6a026be8 100644
--- a/arch/mips/vdso/vdso.lds.S
+++ b/arch/mips/vdso/vdso.lds.S
@@ -93,9 +93,11 @@ VERSION
        LINUX_2.6 {
 #ifndef DISABLE_MIPS_VDSO
        global:
+#if (_MIPS_SIM == _MIPS_SIM_ABI64) || defined(CONFIG_COMPAT_32BIT_TIME)
                __vdso_clock_gettime;
                __vdso_gettimeofday;
                __vdso_clock_getres;
+#endif
 #if _MIPS_SIM != _MIPS_SIM_ABI64
                __vdso_clock_gettime64;
 #endif

That should ensure that no user space can call the old vdso
functions on a kernel that intentionally breaks the actual
syscalls.

      Arnd