From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752430AbcGSGOC (ORCPT ); Tue, 19 Jul 2016 02:14:02 -0400 Received: from mail-pa0-f66.google.com ([209.85.220.66]:33642 "EHLO mail-pa0-f66.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752049AbcGSGOA (ORCPT ); Tue, 19 Jul 2016 02:14:00 -0400 Date: Mon, 18 Jul 2016 23:13:54 -0700 From: Alexei Starovoitov To: Sargun Dhillon Cc: linux-kernel@vger.kernel.org, netdev@vger.kernel.org, Daniel Borkmann Subject: Re: [PATCH 1/1] tracing, bpf: Implement function bpf_probe_write Message-ID: <20160719061352.GA55865@ast-mbp.thefacebook.com> References: <20160713170849.GA76615@ast-mbp.thefacebook.com> <20160715054025.GA99435@ast-mbp> <20160716023035.GA19373@ast-mbp.thefacebook.com> <20160718041105.GA36253@ast-mbp.thefacebook.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.5.24 (2015-08-30) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Mon, Jul 18, 2016 at 03:57:17AM -0700, Sargun Dhillon wrote: > > > On Sun, 17 Jul 2016, Alexei Starovoitov wrote: > > >On Sun, Jul 17, 2016 at 03:19:13AM -0700, Sargun Dhillon wrote: > >> > >>+static u64 bpf_copy_to_user(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) > >>+{ > >>+ void *to = (void *) (long) r1; > >>+ void *from = (void *) (long) r2; > >>+ int size = (int) r3; > >>+ > >>+ /* check if we're in a user context */ > >>+ if (unlikely(in_interrupt())) > >>+ return -EINVAL; > >>+ if (unlikely(!current->pid)) > >>+ return -EINVAL; > >>+ > >>+ return copy_to_user(to, from, size); > >>+} > > > >thanks for the patch, unfortunately it's not that straightforward. > >copy_to_user might fault. Try enabling CONFIG_DEBUG_ATOMIC_SLEEP and > >you'll see the splat since bpf programs are protected by rcu. > >Also 'current' can be null and I'm not sure what current->pid does. > >So the writing to user memory either has to be verified to avoid > >sleeping and faults or we need to use something like task_work_add > >mechanism. Ideas are certainly welcome. > > > > > From casual inspection, I can't find where current can be null when > in_interrupt() is false. Although, we can check before dereferencing it. > When not in a user context, the pid of the task struct returns 0. > > As far as preventing sleep, would the following alteration do? Or do we > actually need something more sophisticated? > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c > index be89c148..45878f3 100644 > --- a/kernel/trace/bpf_trace.c > +++ b/kernel/trace/bpf_trace.c > @@ -86,14 +86,19 @@ static u64 bpf_copy_to_user(u64 r1, u64 r2, u64 r3, u64 > r4, u64 r5) > void *to = (void *) (long) r1; > void *from = (void *) (long) r2; > int size = (int) r3; > + struct task_struct *task = current; > > /* check if we're in a user context */ > if (unlikely(in_interrupt())) > return -EINVAL; > - if (unlikely(!current->pid)) > + if (unlikely(!task || !task->pid)) > return -EINVAL; > > - return copy_to_user(to, from, size); > + /* Is this a user address, or a kernel address? */ > + if (!access_ok(VERIFY_WRITE, to, size)) > + return -EINVAL; > + > + return probe_kernel_write(to, from, size); > } I think it can actually work. The only concern is that comment in access_ok() says that it may sleep whereas I couldn't find any arch where that would be the case. Could you please send an official patch with detailed commit log?