Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <CA+G9fYs9sg69JgmQNZhutQnbijb4GzcO03XF66EjkQ6CTpXXxA@mail.gmail.com>
 <CAK8P3a1Lx1MMQ3s1uWjevsi2wqFo2r=k1hhrxf1spUxEQX_Rag@mail.gmail.com>
In-Reply-To: <CAK8P3a1Lx1MMQ3s1uWjevsi2wqFo2r=k1hhrxf1spUxEQX_Rag@mail.gmail.com>
From:   Jann Horn <jannh@google.com>
Date:   Mon, 23 Nov 2020 15:02:10 +0100
Message-ID: <CAG48ez17CKBMO4193wxuWLRQWQ+q6EV=Qr5oTWiKivMxEi0zQw@mail.gmail.com>
Subject: Re: [arm64] kernel BUG at kernel/seccomp.c:1309!
To:     Arnd Bergmann <arnd@kernel.org>,
        Gabriel Krisman Bertazi <krisman@collabora.com>,
        Andy Lutomirski <luto@kernel.org>,
        Thomas Gleixner <tglx@linutronix.de>
Cc:     Naresh Kamboju <naresh.kamboju@linaro.org>,
        open list <linux-kernel@vger.kernel.org>,
        Netdev <netdev@vger.kernel.org>, bpf <bpf@vger.kernel.org>,
        lkft-triage@lists.linaro.org,
        Linux ARM <linux-arm-kernel@lists.infradead.org>,
        Daniel Borkmann <daniel@iogearbox.net>,
        Kees Cook <keescook@chromium.org>,
        Andrii Nakryiko <andriin@fb.com>,
        Song Liu <songliubraving@fb.com>, Yonghong Song <yhs@fb.com>,
        Andy Lutomirski <luto@amacapital.net>,
        Sumit Semwal <sumit.semwal@linaro.org>,
        Arnd Bergmann <arnd@arndb.de>,
        YiFei Zhu <yifeifz2@illinois.edu>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Nov 23, 2020 at 2:45 PM Arnd Bergmann <arnd@kernel.org> wrote:
> On Mon, Nov 23, 2020 at 12:15 PM Naresh Kamboju
> <naresh.kamboju@linaro.org> wrote:
> >
> > While booting arm64 kernel the following kernel BUG noticed on several arm64
> > devices running linux next 20201123 tag kernel.
> >
> >
> > $ git log --oneline next-20201120..next-20201123 -- kernel/seccomp.c
> > 5c5c5fa055ea Merge remote-tracking branch 'seccomp/for-next/seccomp'
> > bce6a8cba7bf Merge branch 'linus'
> > 7ef95e3dbcee Merge branch 'for-linus/seccomp' into for-next/seccomp
> > fab686eb0307 seccomp: Remove bogus __user annotations
> > 0d8315dddd28 seccomp/cache: Report cache data through /proc/pid/seccomp_cache
> > 8e01b51a31a1 seccomp/cache: Add "emulator" to check if filter is constant allow
> > f9d480b6ffbe seccomp/cache: Lookup syscall allowlist bitmap for fast path
> > 23d67a54857a seccomp: Migrate to use SYSCALL_WORK flag
> >
> >
> > Please find these easy steps to reproduce the kernel build and boot.
>
> Adding Gabriel Krisman Bertazi to Cc, as the last patch (23d67a54857a) here
> seems suspicious: it changes
>
> diff --git a/include/linux/seccomp.h b/include/linux/seccomp.h
> index 02aef2844c38..47763f3999f7 100644
> --- a/include/linux/seccomp.h
> +++ b/include/linux/seccomp.h
> @@ -42,7 +42,7 @@ struct seccomp {
>  extern int __secure_computing(const struct seccomp_data *sd);
>  static inline int secure_computing(void)
>  {
> -       if (unlikely(test_thread_flag(TIF_SECCOMP)))
> +       if (unlikely(test_syscall_work(SECCOMP)))
>                 return  __secure_computing(NULL);
>         return 0;
>  }
>
> which is in the call chain directly before
>
> int __secure_computing(const struct seccomp_data *sd)
> {
>        int mode = current->seccomp.mode;
>
> ...
>         switch (mode) {
>         case SECCOMP_MODE_STRICT:
>                 __secure_computing_strict(this_syscall);  /* may call do_exit */
>                 return 0;
>         case SECCOMP_MODE_FILTER:
>                 return __seccomp_filter(this_syscall, sd, false);
>         default:
>                 BUG();
>         }
> }
>
> Clearly, current->seccomp.mode is set to something other
> than SECCOMP_MODE_STRICT or SECCOMP_MODE_FILTER
> while the test_syscall_work(SECCOMP) returns true, and this
> must have not been the case earlier.

Ah, I think the problem is actually in
3136b93c3fb2b7c19e853e049203ff8f2b9dd2cd ("entry: Expose helpers to
migrate TIF to SYSCALL_WORK flag"). In the !GENERIC_ENTRY case, it
adds this code:

+#define set_syscall_work(fl)                                           \
+       set_ti_thread_flag(current_thread_info(), SYSCALL_WORK_##fl)
+#define test_syscall_work(fl) \
+       test_ti_thread_flag(current_thread_info(), SYSCALL_WORK_##fl)
+#define clear_syscall_work(fl) \
+       clear_ti_thread_flag(current_thread_info(), SYSCALL_WORK_##fl)
+
+#define set_task_syscall_work(t, fl) \
+       set_ti_thread_flag(task_thread_info(t), TIF_##fl)
+#define test_task_syscall_work(t, fl) \
+       test_ti_thread_flag(task_thread_info(t), TIF_##fl)
+#define clear_task_syscall_work(t, fl) \
+       clear_ti_thread_flag(task_thread_info(t), TIF_##fl)

but the SYSCALL_WORK_FLAGS are not valid on !GENERIC_ENTRY, we'll mix
up (on arm64) SYSCALL_WORK_BIT_SECCOMP (==0) and TIF_SIGPENDING (==0).

As part of fixing this, it might be a good idea to put "enum
syscall_work_bit" behind a "#ifdef CONFIG_GENERIC_ENTRY" to avoid
future accidents like this?