Date: Tue, 28 Jun 2022 17:07:14 +0200 (CEST) Message-Id: <202206281507.25SF7EVp045610@...2054...> From: Gert Doering In-Reply-To: <20220624083809.23487-2-a@...2181...> References: <20220624083809.23487-2-a@...2181...> Subject: [Openvpn-devel] [PATCH applied] Re: dco: introduce low-level code for handling ovpn-dco in the Linux kernel List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: 8bit To: Antonio Quartulli Cc: openvpn-devel@lists.sourceforge.net I have stared at this code before, and it "seems to be reasonable". What I *can* attest is that it does not break existing "master" functionality on client or server, Linux or FreeBSD or "make distcheck" (autoconf changes, new modules, ...). I did not expect anything, but this was sort of the risk at this point. Calling configure with --enable-dco does compile the new stuff, so it is at least "compile safe" (and no warnings). Some review comments that could be fixed in a followup patch: - some of the functions (like "ovpn_dco_nlmsg_create()", "dco_new_peer()") could definitely use a bit of commenting so non-netlink-experts would understand what these functions do, and why... (and I'm sure I mentioned that before). - dco_new_peer() treats "localaddr" and "remoteaddr" differently when packing, which looks weird -> I think a comment that this is to reduce kernel code ("we need full sockaddr for the remote, and only the local IP address for local") would avoid confusion. - openvpn_dco_init_netlink() carries this TODO + /* TODO: Why are we setting this buffer size? */ ... so, someone should know, and add an appropriate comment :-) - a few words of explanation on "... multicast message sent by the ovpn-dco kernel module" might be good - I see you subscribe to it, and explain mcast_family_handler(), but I can not find any indication of "what sort of messages would we expect to see here?". - ovpn_dco_nlmsg_create() has + if (!nl_msg) + { + msg(M_ERR, "cannot allocate netlink message"); + return NULL; + } ... which never returns, as M_ERR is fatal. OTOH, *some* callers do + struct nl_msg *nl_msg = ovpn_dco_nlmsg_create(dco, OVPN_CMD_REGISTER_PACKET); + if (!nl_msg) + { + msg(M_ERR, "%s: cannot allocate message to register for control packets", ... but not *all* of them (dco_new_peer() does not check nl_msg at all). So this should be made consistent - like, either ovpn_dco_nlmsg_create() will M_ERR on allocation failure, or return NULL (and never M_ERR). If it does M_ERR itself, no need to check that in the callers... - open_tun_dco() reports an *error* top create DCO interface with D_DCO_DEBUG + msg(D_DCO_DEBUG, "Cannot create DCO interface %s: %d", dev, ret); ... and two lines down, an error on if_nametoindex() leads to a FATAL + msg(M_FATAL, "DCO: cannot retrieve ifindex for interface %s", dev); ... so what? Is this function allowed to fail, and return "ret < 0", or should it M_ERR/M_FATAL on errors? There is no pre-function comment that could clarify the intent... - dco_new_key() has code to support "none" + if (dco_cipher != OVPN_CIPHER_ALG_NONE) + { ... but dco_cipher can never be OVPN_CIPHER_ALG_NONE anymore... and if that indent is removed, the NLA_PUT calls for _iv do not need to wrap :-) - dco_set_peer() does not pass on the mss yet - in ovpn_handle_msg(), can this be called uninitialized? + if (!attrs[OVPN_ATTR_IFINDEX]) + { + msg(D_DCO, "ovpn-dco: Received message without ifindex"); + return NL_SKIP; + } (not sure if nla_parse() will ensure that there is enough attributes, or just "well-formed" - and whether *attrs actually gets zeroed by nla_parse(), as openvpn_handle_msg() has no CLEAR() ) - same function - under which conditions can this fire? Kernel getting confused? + uint32_t ifindex = nla_get_u32(attrs[OVPN_ATTR_IFINDEX]); + if (ifindex != dco->ifindex) + { + msg(D_DCO, "ovpn-dco: received message type %d with mismatched ifindex %d\n", Uncrustify has complained at me when I merged the patch (because in that moment, ovpn_dco_linux.h was "newly modified" and the exclusion rule does not match on the pre-commit-hook) - but as discussed, this is a bit complicated due to "kernel style" vs "openvpn style", so I've left it alone for the moment. Your patch has been applied to the master branch. commit e34437c26b764851555e4acbe2ccca6bec235c7e Author: Antonio Quartulli Date: Fri Jun 24 10:37:45 2022 +0200 dco: introduce low-level code for handling ovpn-dco in the Linux kernel Signed-off-by: Antonio Quartulli Acked-by: Arne Schwabe Message-Id: <20220624083809.23487-2-a@...2181...> URL: https://www.mail-archive.com/openvpn-devel@lists.sourceforge.net/msg24512.html Signed-off-by: Gert Doering -- kind regards, Gert Doering