From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752097AbaG0MJY (ORCPT ); Sun, 27 Jul 2014 08:09:24 -0400 Received: from mail-vc0-f171.google.com ([209.85.220.171]:50024 "EHLO mail-vc0-f171.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751978AbaG0MJV (ORCPT ); Sun, 27 Jul 2014 08:09:21 -0400 MIME-Version: 1.0 In-Reply-To: References: <1406296033-32693-1-git-send-email-drysdale@google.com> <1406296033-32693-12-git-send-email-drysdale@google.com> From: David Drysdale Date: Sun, 27 Jul 2014 13:09:00 +0100 Message-ID: Subject: Re: [PATCH 11/11] seccomp: Add tgid and tid into seccomp_data To: Andy Lutomirski Cc: Kees Cook , "Eric W. Biederman" , Julien Tinnes , Al Viro , Paolo Bonzini , LSM List , Greg Kroah-Hartman , Paul Moore , James Morris , Linux API , Meredydd Luff , Christoph Hellwig , "linux-kernel@vger.kernel.org" Content-Type: text/plain; charset=UTF-8 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Fri, Jul 25, 2014 at 6:18 PM, Andy Lutomirski wrote: > [cc: Eric Biederman] > > On Fri, Jul 25, 2014 at 10:10 AM, Kees Cook wrote: >> On Fri, Jul 25, 2014 at 8:59 AM, Andy Lutomirski wrote: >>> On Jul 25, 2014 6:48 AM, "David Drysdale" wrote: >>>> >>>> Add the current thread and thread group IDs into the data >>>> available for seccomp-bpf programs to work on. This allows >>>> installation of filters that police syscalls based on thread >>>> or process ID, e.g. tgkill(2)/kill(2)/prctl(2). >>>> >>>> Signed-off-by: David Drysdale >>>> --- >>>> include/uapi/linux/seccomp.h | 10 ++++++++++ >>>> kernel/seccomp.c | 2 ++ >>>> 2 files changed, 12 insertions(+) >>>> >>>> diff --git a/include/uapi/linux/seccomp.h b/include/uapi/linux/seccomp.h >>>> index ac2dc9f72973..b88370d6f6ca 100644 >>>> --- a/include/uapi/linux/seccomp.h >>>> +++ b/include/uapi/linux/seccomp.h >>>> @@ -36,12 +36,22 @@ >>>> * @instruction_pointer: at the time of the system call. >>>> * @args: up to 6 system call arguments always stored as 64-bit values >>>> * regardless of the architecture. >>>> + * @tgid: thread group ID of the thread executing the BPF program. >>>> + * @tid: thread ID of the thread executing the BPF program. >>>> + * The SECCOMP_DATA_TID_PRESENT macro indicates the presence of the >>>> + * tgid and tid fields; user programs may use this macro to conditionally >>>> + * compile code against older versions of the kernel. Note also that >>>> + * BPF programs should cope with the absence of these fields by testing >>>> + * the length of data available. >>>> */ >>>> struct seccomp_data { >>>> int nr; >>>> __u32 arch; >>>> __u64 instruction_pointer; >>>> __u64 args[6]; >>>> + __u32 tgid; >>>> + __u32 tid; >>>> }; >>>> +#define SECCOMP_DATA_TID_PRESENT 1 >>>> >>>> #endif /* _UAPI_LINUX_SECCOMP_H */ >>>> diff --git a/kernel/seccomp.c b/kernel/seccomp.c >>>> index 301bbc24739c..dd5146f15d6d 100644 >>>> --- a/kernel/seccomp.c >>>> +++ b/kernel/seccomp.c >>>> @@ -80,6 +80,8 @@ static void populate_seccomp_data(struct seccomp_data *sd) >>>> sd->args[4] = args[4]; >>>> sd->args[5] = args[5]; >>>> sd->instruction_pointer = KSTK_EIP(task); >>>> + sd->tgid = task_tgid_vnr(current); >>>> + sd->tid = task_pid_vnr(current); >>>> } >>> >>> This is, IMO, problematic. These should probably be relative to the >>> filter creator, not the filtered task. This will also hurt >>> performance. >> >> Yeah, we can't change the seccomp_data structure without a lot of >> care, and tgid/tid really should be encoded in the filter. However, it >> is tricky in the forking case. >> >>> >>> What's the use case? Can it be better achieved with a new eBPF function? The specific use case is to be able to write a filter that allows kill(2) or tgkill(2) to self, where the filter still works after forking. Capsicum capability mode in general locks down system calls that access PIDs (as they're a global namespace), but allows kill(self) as a pragmatic compromise to make it easier to migrate applications to use Capsicum. >> Julien had been wanting something like this too (though he'd suggested >> it via prctl): limit the signal functions to "self" only. I wonder if >> adding a prctl like done for O_BENEATH could work for signal sending? >> > > > Can we do one better and add a flag to prevent any non-self pid > lookups? This might actually be easy on top of the pid namespace work > (e.g. we could change the way that find_task_by_vpid works). That sounds like a good idea, as long as it's possible for non-CAP_SYS_ADMIN processes to do.... > It's far from just being signals. There's access_process_vm, ptrace, > all the signal functions, clock_gettime (see CPUCLOCK_PID -- yes, this > is ridiculous), and probably some others that I've forgotten about or > never noticed in the first place. For the Capsicum case in particular, most of these are restricted by the capability mode filter anyhow (although I need to fix it for CPUCLOCK_PID -- thanks for pointing that out); the kill(2) case was a special case to make migrations easier. But a more general mechanism seems sensible. > --Andy > >> -Kees >> >> -- >> Kees Cook >> Chrome OS Security > > > > -- > Andy Lutomirski > AMA Capital Management, LLC