From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-4.0 required=3.0 tests=HEADER_FROM_DIFFERENT_DOMAINS, INCLUDES_PATCH,MAILING_LIST_MULTI,SPF_PASS autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id C8A87C282C2 for ; Wed, 13 Feb 2019 03:59:14 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id 8C47321934 for ; Wed, 13 Feb 2019 03:59:14 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S2387605AbfBMD7N (ORCPT ); Tue, 12 Feb 2019 22:59:13 -0500 Received: from out03.mta.xmission.com ([166.70.13.233]:52121 "EHLO out03.mta.xmission.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1730772AbfBMD7M (ORCPT ); Tue, 12 Feb 2019 22:59:12 -0500 Received: from in01.mta.xmission.com ([166.70.13.51]) by out03.mta.xmission.com with esmtps (TLS1.2:ECDHE_RSA_AES_128_GCM_SHA256:128) (Exim 4.87) (envelope-from ) id 1gtlhU-0004Cv-DJ; Tue, 12 Feb 2019 20:59:08 -0700 Received: from ip68-227-174-240.om.om.cox.net ([68.227.174.240] helo=x220.xmission.com) by in01.mta.xmission.com with esmtpsa (TLS1.2:ECDHE_RSA_AES_128_GCM_SHA256:128) (Exim 4.87) (envelope-from ) id 1gtlhJ-0008Ap-2c; Tue, 12 Feb 2019 20:59:08 -0700 From: ebiederm@xmission.com (Eric W. Biederman) To: Oleg Nesterov Cc: Dmitry Vyukov , Thomas Gleixner , Ingo Molnar , Peter Zijlstra , LKML , Arnaldo Carvalho de Melo , Alexander Shishkin , jolsa@redhat.com, Namhyung Kim , luca abeni , syzkaller , Ivan Delalande References: <878syu7tcm.fsf@xmission.com> <87tvhi4vl7.fsf@xmission.com> <87o97q1cky.fsf_-_@xmission.com> <20190206180754.GA23476@redhat.com> <87imxwv9jp.fsf@xmission.com> <875ztwt7yy.fsf_-_@xmission.com> <87zhr8rtd6.fsf_-_@xmission.com> <20190211141340.GA21430@redhat.com> <87zhr1g7ls.fsf@xmission.com> <871s4dctci.fsf@xmission.com> <20190212165022.GA29263@redhat.com> Date: Tue, 12 Feb 2019 21:58:51 -0600 In-Reply-To: <20190212165022.GA29263@redhat.com> (Oleg Nesterov's message of "Tue, 12 Feb 2019 17:50:23 +0100") Message-ID: <87mun05og4.fsf@xmission.com> User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/25.1 (gnu/linux) MIME-Version: 1.0 Content-Type: text/plain X-XM-SPF: eid=1gtlhJ-0008Ap-2c;;;mid=<87mun05og4.fsf@xmission.com>;;;hst=in01.mta.xmission.com;;;ip=68.227.174.240;;;frm=ebiederm@xmission.com;;;spf=neutral X-XM-AID: U2FsdGVkX19bxTii4t1DJXvgJf3FAuTv64DoCJfA598= X-SA-Exim-Connect-IP: 68.227.174.240 X-SA-Exim-Mail-From: ebiederm@xmission.com Subject: Re: [PATCH 1/2] signal: Always notice exiting tasks X-SA-Exim-Version: 4.2.1 (built Thu, 05 May 2016 13:38:54 -0600) X-SA-Exim-Scanned: Yes (on in01.mta.xmission.com) Sender: linux-kernel-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Oleg Nesterov writes: > On 02/12, Eric W. Biederman wrote: >> >> > Here I was trying for the simple minimal change and I hit this landmine. >> > Which leaves me with the question of what should be semantics of signal >> > handling after exit. > > Yes, currently it is undefined. Even signal_pending() is random. > >> > I think from dim memory of previous conversations the desired semantics >> > look like: >> > a) Ignore all signal state except for SIGKILL. >> > b) Letting SIGKILL wake up the process should be sufficient. > > signal_wake_up(true) to make fatal_signal_pending() == T, I think. > >> Oleg any ideas on how to make PTRACE_EVENT_EXIT reliably killable? > > My answer is very simple: PTRACE_EVENT_EXIT must not stop if the tracee was > killed by the "real" SIGKILL (not by group_exit/etc), that is all. But this > is another user-visible change, it can equally confuse, say, strace (albeit > not too much iiuc). > > But this needs another discussion. Yes. Quite. I will just point out that as described that logic will rebreak Ivan's program. >> diff --git a/kernel/signal.c b/kernel/signal.c >> index 99fa8ff06fd9..a1f154dca73c 100644 >> --- a/kernel/signal.c >> +++ b/kernel/signal.c >> @@ -2544,6 +2544,9 @@ bool get_signal(struct ksignal *ksig) >> } >> >> fatal: >> + /* No more signals can be pending past this point */ >> + sigdelset(¤t->pending.signal, SIGKILL); > > Well, this is very confusing. In fact, this is not really correct. Say, we should > not remove the pending SIGKILL if we are going to call do_coredump(). This is > possible if ptrace_signal() was called, or after is_current_pgrp_orphaned() returns > false. I don't see bugs in it. But it is certainly subtle and that is not what is needed right now. The subtlety is that we will never have a per thread SIGKILL pending unless signal_group_exit is true. So removing when it is not there is harmless. >> + clear_tsk_thread_flag(current, TIF_SIGPENDING); > > I don't understand this change, it looks irrelevant. Possibly makes sense, but > this connects to "semantics of signal handling after exit". As on the other the location is too subtle for the regression fix. The primary motivation is that dequeue_signal calls recalc_sigpending. And in the common case that will result clearing the TIF_SIGPENDING which will result in signal_pending being false. I have not found a location that cares enough to cause a misbehavior if we don't clear TIF_SIGPENDING but it is a practical change and there might be. So if the word of the day is be very conservative and avoid landminds I expect we need the clearing of TIF_SIGPENDING. Hmm. Probably using recalc_sigpending() now that I think about it. > OK, we need a minimal incremental fix for now. I'd suggest to replace > > ksig->info.si_signo = signr = SIGKILL; > if (signal_group_exit(signal)) > goto fatal; > > added by this patch with > > if (__fatal_signal_pending(current)) { > ksig->info.si_signo = signr = SIGKILL; > sigdelset(¤t->pending.signal, SIGKILL); > goto fatal; > } > > __fatal_signal_pending() is cheaper and looks more understandable. I definitely agree that it is much less likely to cause a problem if we move all of the work before jumping to fatal. The cost of both __fatal_signal_pending and signal_group_exit is just a cache line read. So not a big deal wither way. On the other hand __fatal_signal_pending as currently implemented is insanely subtle and arguably a bit confusing. It tests for a SIGKILL in the current pending sigset, to discover the signal group property of if a process as started exiting. In the long run we need our data structures not to be subtle and tricky to use. To do that we need a test of something in signal_struct because it is a per signal group property. Further we need to remove the abuse of the per thread SIGKILL. Since signal_group_exit always implies __fatal_signal_pending in this case and the reverse. I see no reason to use a function that requires we maintain a huge amount of confusing and unnecessary machinery to keep working. All of that plus the signal_group_exit test has been tested and shown to fix an ignored SIGKILL and the only practical problem is it doesn't do one or two little things that dequeue_signal has done that made it impossible to stop in PTRACE_EVENT_EXIT. So for the regression fix let's just do the few little things that dequeue_signal used to do. That gives us a strong guarantee that nothing else was missed. Eric