From mboxrd@z Thu Jan 1 00:00:00 1970 From: David Miller Subject: Re: [PATCH] l2tp: Restore socket refcount when sendmsg succeeds Date: Tue, 12 Mar 2013 10:21:21 -0400 (EDT) Message-ID: <20130312.102121.1471200685376534204.davem@davemloft.net> References: <20130301150202.GA3649@alphalink.fr> <20130301.141252.1152926224695074046.davem@davemloft.net> <20130312103650.GA4512@alphalink.fr> Mime-Version: 1.0 Content-Type: Text/Plain; charset=us-ascii Content-Transfer-Encoding: 7bit Cc: jchapman@katalix.com, linux-kernel@vger.kernel.org, netdev@vger.kernel.org To: g.nault@alphalink.fr Return-path: In-Reply-To: <20130312103650.GA4512@alphalink.fr> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org From: Guillaume Nault Date: Tue, 12 Mar 2013 11:36:50 +0100 > On Fri, Mar 01, 2013 at 02:12:52PM -0500, David Miller wrote: >> Looking at how this code works, it is such a terrible design. This >> whole reference counting issue exists purely because >> pppol2tp_sock_to_session() grabs the 'sk' reference. >> >> In all but one case, it need not do this. >> >> The socket system calls have an implicit reference to 'sk' via >> socket->sk. If you can get into the system call and socket->sk >> is non-NULL then 'sk' is NOT going anywhere. >> >> And all of these system call handlers have this pattern: >> >> session = pppol2tp_sock_to_session(sk); >> ... >> sock_put(sk); >> >> The only case where the reference count is really needed is that >> sequence in pppol2tp_release(). >> >> Long term the right thing to do here is stop having this session >> grabber function take the 'sk' reference. Then in pppol2tp_release >> we'll grab a reference explicitly. At all the other call sites we >> then blast aweay all of the sock_put(sk) paths. >> > Could this also apply to l2tp_sock_to_tunnel() (in l2tp_core.h)? As per > my understanding, none of its callers needs to take a socket reference. > So sock_hold() could be removed in both pppol2tp_sock_to_session() and > l2tp_sock_to_tunnel() functions. The corresponding sock_put() calls > would then be removed from all calling functions but pppol2tp_release(). > If this is correct, I'll send a patch for net-next. Yes, it could be simplified in this way too. Just make sure that this interface is only used in system call / user context, where we know the underlying socket cannot go away on us.