cjdns/util/version/Version.h

/* vim: set expandtab ts=4 sw=4: */
/*
 * You may redistribute this program and/or modify it under the terms of
 * the GNU General Public License as published by the Free Software Foundation,
 * either version 3 of the License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program.  If not, see <https://www.gnu.org/licenses/>.
 */
#ifndef Version_H
#define Version_H

#include "util/Linker.h"
Linker_require("util/version/Version.c")

#include <stdint.h>

/*
 * Cjdns Protocol Versions
 *
 * The first argument to Version_COMPAT is the new version, the second argument is a list of
 * older versions with which it is compatible. All versions are obviously assumed to be compatible
 * with themselves and any version which is compatible with the most recent version to date is
 * assumed to be compatible with all future versions.
 */
#ifndef Version_COMPAT
    // defined otherwise in Version.c
    #define Version_COMPAT(one, twoRange)
#endif

/*
 * Version 0:
 * January 2012
 *
 * First version.
 */
Version_COMPAT(0, ([]))

/*
 * Version 1:
 * October 2012
 *
 * When you send someone a message through cjdns, it's encrypted.
 * When you send the first message to a router, it has a nice header on it which tells them your key
 * and allows them to establish a cryptographic session with you. For every message after that, they
 * need to remember that session and use it to decrypt the message.
 *
 * The way they remember which session a message is associated with is to look at the switching
 * label from that message and compare that with the label which was used when the first message was
 * sent (with the header).
 *
 * What I didn't think about at the time is that labels change. Nodes find a better path to a
 * destination and expect the same session to work. It would work if the other end knew which
 * session to use but it can't know, the label is different.
 *
 * This is a protocol bug.
 *
 * In my opinion, the best way to fix it is to send an additional header before the crypto nonce
 * which tells the foreign node who it came from. When the handshake message is sent, the node will
 * send a 4 byte integer which the other end will store. Every time the other end sends a
 * non-handshake message to this end, it will prepend that same integer to the encrypted message and
 * on recieving a message, a node will use that number to do a lookup of the correct session to use.
 * The number can be a pointer or index offset so this can be quite fast.
 * Integer 0xFFFFFFFF shall be reserved and handshake messages which contain this value must be
 * ignored or treated as protocol 0 messages.
 *
 * But this is a protocol break.
 *
 * If a node gets handed a number and doesn't know what to do with it, it will think it's a
 * CryptoAuth header and it will fail. New nodes would be able to try a message as the old form if
 * the new form doesn't work but old nodes will just fail if they are ever sent a message in the new
 * form.
 *
 * Proposed Solution:
 * A key will be sent in all findNodes responses in the router, this key will be "np" for nodes'
 * protocol (version).
 * It will contain a string representation of a list of the protocol version numbers for the nodes
 * which it is introducing. The numbers will all be the same number of bytes and the first byte will
 * give that number. The length of the string will always be equal to one, plus the number of nodes
 * times the value of the first byte.
 *
 * This example shows the p which would accompany a findNodes response containing 1 node with
 * protocol version zero and one node with protocol version 1.
 *
 * "np": "\x01\x00\x01"
 *
 * 2:np3:\x01\x00\x01
 *
 * This example shows the p which would accompany a findNodes response containing 1 node with
 * protocol version 300 and two nodes with protocol version 5.
 *
 * "np": "\x02\x01\x2c\x00\x05\x00\x05"
 *
 * 2:np7:\x02\x01\x2c\x00\x05\x00\x05
 *
 * All multi-byte numbers shall be encoded in big endian encoding.
 *
 * Each node will have an internal compatibility matrix giving protocol version numbers which
 * can communicate, a node shall not respond to a findNodes message with a response containing any
 * nodes which are known to be incompatible with the protocol version of the requesting node.
 * Versions which are beyond the highest version number in the compatability matrix will be assumed
 * to have the same compatability as the highest number in the table.
 *
 * All messages shall contain shall contain an "p" key as well but these will have a benc integer
 * representing the protocol version of the sending node.
 *
 * Change to the Ping switch control message:
 * New switch ping messages will all be 8 or more bytes long, they will begin with a magic field
 * and then a version number which is the version of the sending node.
 * The magic will be set to 0x09f91102 for all ping messages and to 0x9d74e35b in the response
 * messages (these numbers shall be big endian encoded). Following this number will be a 4 byte
 * field containing the protocol version. Nodes which only speak protocol version 0 will be
 * identifyable because they echo back 0x09f91102 rather than replacing it and they will be unlikely
 * to send a ping request whose content begins with 0x09f91102.
 *
 * Protocol1.1
 * In protocol0 there was a single pool of sessions shared between the outer and inner layer.
 * In protocol1 it was split because one pool needed to have handles and the other pool didn't.
 * The problem with this is communications do not necessarily travel back and forth along the
 * same path and protocol1 exhibited a pathology wherein one node was direct sending packets to
 * another while the other was routing the responses via an intermediary.
 * There were 2 CryptoAuth sessions between the two nodes and neither session was entering run
 * state. So protocol1 was broken and the new protocol1 (protocol1.1) sends session handles on top
 * of the CryptoAuth handshake headers even if the handshake is in the inner layer (under the Ipv6
 * header). It does not however send handles in the inner layer when the inner layer CryptoAuth
 * session is in HANDSHAKE3 state or above.
 * Protocol1.1 still identifies itself as Protocol1, it will not be able to communicate with
 * protocol0 in some circumstances. If it knows nothing about the other node and it forwards a
 * message via an intermediary, the message will be unreadable at the other end.
 *
 * Protocol1.2
 * Changes in protocol1.1 are reverted as a solution was found in the implementation.
 */
Version_COMPAT(1, ([0]))

/*
 * Version 2:
 * February 21, 2013
 *
 * Remove compatibility layer for communicating with version 0 nodes.
 *
 * August 22, 2014
 * The Version_2_COMPAT ifdefs were removed in e2a7ebc4d7c54b6b5fac2c0ba5c7882af8b238f2
 */
Version_COMPAT(2, ([1]))

/*
 * Version 3:
 * August 16, 2013
 *
 * In version 1, handles were introduced so that a session could be looked when a packet came in.
 * During the initiation of a session, the node's handle was placed before the CryptoAuth block
 * unless the message was a layer-3 (forwarded) message in which case it was at the beginning
 * inside of the innermost CryptoAuth block. This handle was transferred to the peer so they could
 * tell how *this* node identifies the session.
 *
 * When a layer-2 (non-forwarded) message was received which was *not* a session initiation packet,
 * the peer's handle was placed before the CryptoAuth block in order for the peer to be able to
 * lookup the session in their table.
 *
 * Unfortunately the handle outside of the CryptoAuth initiation block was not authenticated in any
 * way and a switch could alter it accidently or maliciously causing the wrong session identifier
 * to be stored leading to the session failing until it eventually times out.
 *
 * To fix this bug the handle attached to the initiation message has been moved into the CryptoAuth
 * block where it should have been in the first place. Obviously the peer's handle is still placed
 * outside of the CryptoAuth block so it can serve it's function.
 *
 * As of version 3, implementations must not send handles which are less than 4 so that when they
 * receive them back, they will not be confused with the initial 4 bytes of a CryptoAuth setup
 * packet which is not preceeded by a handle.
 */
Version_COMPAT(3, ([1,2]))

/*
 * Version 4:
 * August 27, 2013
 *
 * This version makes no protocol changes but fixes a nasty bug with forwarding which caused
 * messages to be forwarded to random nodes, updated to encourage nodes to forward via others
 * who do not have the bug.
 */
Version_COMPAT(4, ([1,2,3]))

/*
 * Version 5:
 * September 4, 2013
 *
 * This version introduces a new RPC call for getting directly connected peers from a node.
 * The new RPC call is called "gp" and it takes a target label called "tar" which must be an
 * 8 byte long benc string. It returns a list of peers exactly the same as a search but they
 * must all be direct peers and they are the peers whose labels have smallest XOR distance
 * from "tar".
 */
Version_COMPAT(5, ([1,2,3,4]))

/*
 * Version 6:
 * December 14, 2013
 *
 * Drop support for versions older than 5
 */
Version_COMPAT(6, ([5]))

/*
 * Version 7:
 * March 16, 2014
 *
 * A ceremonial version which has the pathfinder2 implemented.
 * this is still compatible with the same versions and doesn't expressly change the protocol
 * although it is nolonger able to communicate with *older* version 5 nodes which do not advertize
 * their encoding scheme.
 */
Version_COMPAT(7, ([5,6]))

/*
 * Version 8:
 * August 22, 2014
 *
 * Previous to version 8, switch headers had a "Type" field which indicated the data type.
 * The data type is now inferred from the content of the packet and in the case of need to send
 * a control (error) packet, the switch sends a "handle" which is set to 0xffffffff.
 * Also switches previous to v8 zeroed the priority field, now it is left untouched.
 */
Version_COMPAT(8, ([5,6,7]))

/*
 * Version 9:
 * September 12, 2014
 *
 * Version 8 was sending improper checksums on error frames, corrected in v9.
 */
Version_COMPAT(9, ([5,6,7,8]))

/*
 * Version 10:
 * September 18, 2014
 *
 * Drop support for pre-v7
 */
Version_COMPAT(10, ([7,8,9]))

/*
 * Version 11:
 * October 7, 2014
 *
 * New QoS system and changes to SwitchHeader structure, see SwitchHeader.h
 */
Version_COMPAT(11, ([7,8,9,10]))

/*
 * Version 12:
 * November 21, 2014
 *
 * New inter-router API "nh" which has same semantics as "fn" but gets the next hop in
 * a hypothetical packet forwarding operation.
 */
Version_COMPAT(12, ([7,8,9,10,11]))

/*
 * Version 13:
 * December 30, 2014
 *
 * 1. Drop nodes older than 12.
 *
 * 2. Change of ETHInterface protocol.
 * We determined that the ETHInterface frame format was not possible to evolve
 * so we determined to add the low 16 bits of the version number to the header.
 * This was brought about by a need to include the message length in the header
 * to fix ethernet cards which fail to strip the CRC from the received message.
 *
 * The beacon contains a version number so if an old node is beaconing then the
 * new node can still speak the old protocol. If the new node is beaconing and
 * an old node receives it and responds, it will fail as we have determined that
 * there is no reasonable way for the new node to determine that an old format
 * message is old. It will begin working again if the old node sends a beacon.
 */
Version_COMPAT(13, ([12]))

/*
 * Version 14:
 * January 23, 2014
 *
 * Cerimonial version 14 to indicate a bug-fix in the getPeers function which
 * previously returned the same list of peers every time. Furthermore the switch
 * pinger was fixed to set the labelShift field correctly, fixing a bug which
 * caused spurious loss of switch ping packets (culminating in peers "missing"
 * from the routing table).
 */
Version_COMPAT(14, ([12,13]))

/**
 * Version 15:
 * January 29, 2014
 *
 * Cerimonial version 15 to indicate a change in routing behavior, for each incoming
 * packet, remember the path it took, if no path to the return node is known,
 * use the path that the incoming packet took for the response.
 */
Version_COMPAT(15, ([12,13,14]))

/**
 * Version 16:
 * February 13, 2014
 *
 * Verschlumbesserung
 *
 * This version comprises a major refactoring both in the internal organizatin of and the behavior
 * of cjdns. First on the internal organization note, the file formerly known as Ducttape.c is no
 * more, it made cjdns when cjdns didn't work but now it's time has passed, long live Ducttape.
 * Ducttape has been broken up into a series of 5 modules, SwitchAdapter, ControlHandler,
 * SessionManager, UpperDistributor and TUNAdapter. Furthermore Interface.h has been removed and
 * replaced with Iface.h. Interface.h was "gendered", meaning the male side of one interface could
 * only be linked with the female side of the other, Iface is ungendered and is linked (potentially
 * to any other interface in the project) using Iface_plumb(). Furthermore Iface has been adapted
 * to facilitate manual tail-call optimization.
 *
 * A new protocol called PFChan seperates the Pathfinder from the (packet handling) core. The
 * pathfinder is nolonger required to answer any questions for the core synchronously and since it
 * now communicated using Message and Iface as opposed to function calls, it may be seperated into
 * an external process. Furthermore the EventEmitter (module to which the pathfinder connects) is
 * capable of accepting connections from multiple pathfinders, allowing advanced external
 * pathfinders to be developed outside of the main cjdns project.
 *
 * On the point of protocol, two new headers have been defined, one is sent over the wire to other
 * nodes and the other is merely internal protocol for communicating with the SessionManager.
 * DataHeader is a new header, sent over the wire between any two v16+ nodes, it replaces the faux
 * IPv6 headers which previously were sent over the wire. RouteHeader is used internally to tell
 * SessionManager where one wants the packet to go but is never seen on the wire. The final and
 * perhaps most significant change in this version is the loss of packet forwarding. No longer does
 * a v16 node attempt to forward a message to another node in case that it does not know a route to
 * the final destination instead it bufferes the packet and triggers a DHT search in the same way
 * that Ethernet buffers a frame and triggers an ARP request. This vastly simplifies the debugging
 * of weird routing behaviors.
 */
Version_COMPAT(16, ([12,13,14,15]))

/**
 * Version 17:
 * October 9, 2015
 *
 * Ouvrir le parapluie
 *
 * When a node connects to another node, before v17, the connection password is double-hashed and
 * sent while the password hash and a field called Derivations were hashed together (and then
 * hashed with the shared secret from Curve25519 crypto) so that theoretically the hash of the
 * password and a various value of Derivations could be passed to another node who could then
 * establish a more heavily secured session. After v17 Derivations now is meaningless and the
 * previously meaningless bit A which had been set in the CryptoHeader_Challenge is now cleared
 * (see CryptoHeader.h)
 *
 * Also v17 adds a new authType, authType 2 which is for logging in with name and password.
 * In pre-17 when a session setup packet is sent, the double-hash of the password is sent as a
 * key so that the server can lookup the correct password to test.
 * This means someone with the session setup packet could crack to find the actual password and
 * then connect to the server. The change introduces a "login" in addition to the password, this
 * instead is hashed and sent so if the attacker cracks the hash, they'll get the login and will
 * not be able to login to the server.
 */
Version_COMPAT(17, ([16]))

/**
 * Version 18:
 * May 13, 2016
 *
 * Rimpasto
 *
 * In early 2016 a fundimental flaw was discovered in the way in which the Pathfinder works.
 * This flaw was named The Drawbridge Bug. When a node starts, it populates it's switch table
 * in first-come-first-serve order. This is desired behavior, we must occasionally reshuffle
 * the switch table in order to avoid persistent holes in the table which cause unduly long
 * switch label directors.
 * Unfortunately when a node reshuffles it's switch table, this breaks labels which pass through
 * the node. More unfortunately, it is not always clear that they were broken. For instance:
 * A->B->C->D->E and C resets and now the same path goes A->B->C->X->Y, the traditional Pathfinder
 * would think that D->E is broken and might replace it with D->Y, which is an incorrect inference
 * but will work for making this particular path again - thus positively reinforcing the wrong
 * inferrence.
 * Furthermore this problem has caused the NodeStore to make conclusions which are provably
 * incorrect and this caused assertion failures which caused nodes to reset, causing more and more
 * reshuffling of switch tables and more and more crashes.
 * A secondary and more difficult problem is that the paths resolved by the DHT style routing
 * system are longer than they need to be because as the search wanders looking for a node which
 * knows the full path, so too does the traffic.
 * There exist many routing algorithms which are far more efficient and some are believed to scale
 * smoothly to millions of nodes but the routing table's resistance to poison is specific to cjdns
 * and other algorithms are mostly highly susceptable. A similar issue to poison-resiliance is that
 * cjdns nodes need not implement any highly complex behaviors in order to keep the network
 * functioning, they only answer a few simple questions such as "who do you know whose address is
 * numerically close to X" and "which of your peers has a path numerically close to Y". Most other
 * routing algorithms require complex implementations and even a small deviation in behavior on
 * one node could be disastrous.
 *
 * v18 Rimpasto is a rethink of the routing infrastructure. First we accept that the problem is
 * harder than it was intially thought to be. The old routing infrastructure is now optional and
 * is replaced by something known as subnode, subnode connects to what are called supernodes.
 * The supernode code is in a different codebase which is written in nodejs and is called cjdnsnode.
 * The subnode is mostly located in subnode directory in the project but it makes use of a few parts
 * of the old dht directory.
 * The subnode configuration contains a set of supernodes, not anyone can just start up a supernode
 * and use it to damage the network, subnodes need to trust a supernode in order to use it.
 * The subnode uses findNode and getPeers requests to find a path to a supernode and then it sends
 * a findPath query to the supernode in order to answer all of it's routing needs. The supernode
 * is expected to be able to construct a route label for the optimal path between any one of it's
 * clients and any other node in the network but how that is done is up to the supernode.
 * The subnode still answers all of the queries made by the old DHT based routing code but for
 * findNode queries, the subnode answers the query by requesting the information from the supernode,
 *
 * This version allows optional/experimental subnode using SUBNODE=1 in the build.
 */
Version_COMPAT(18, ([16,17]))

/**
 * Version 19:
 * February 21, 2017
 *
 * shibboleth
 *
 * This is a mostly cerimonial release, the supernode/subnode infrastructure is still not completely
 * ready. This release is largely cerimonial, there are no breaking changes to the protocol.
 * There have been improvements to the CryptoAuth handshake which should make it less likely that
 * the CryptoAuth session will go into a bad state and the nodes be unable to talk to eachother.
 */
Version_COMPAT(19, ([16,17,18]))

/**
 * Version 20:
 * March 16, 2017
 *
 * baratiner
 *
 * In this release the default behavior is to try to solicit supernodes by asking peers then to
 * attempt to announce to a supernode, however the old messages are still supported and they are
 * handled by the old nodestore. When something needs to be found, this version will use both the
 * old DHT and its supernode (if it has one) at the same time.
 */
Version_COMPAT(20, ([16,17,18,19]))

/**
 * Version 21:
 * June 23, 2020
 *
 * Disintermediated
 *
 * Changes to allow multiple peerings to the same node over different media. For example over
 * IPv4 and IPv6, or Ethernet and IPv4. To support this there is a new switch message called
 * RPATH which gets the reverse path for a connection. Perviously we used getpeers and then
 * found ourselves in the list, but since we may appear in the list multiple times, that is
 * nolonger acceptable.
 * Secondly, we changed the way announcements to the route server are filtered, filtering them
 * by label (actually peerNum) rather than by ipv6. This means the sync with snode is different.
 * Third, we added a new field "dnd" (do not disturb) to DHT messages, which indicates that our
 * DHT server is down and we would rather not be bothered with lots of find-node or get-peer
 * traffic. When running in SUBNODE mode, "dnd" is enabled.
 */
Version_COMPAT(21, ([20]))


/**
 * The current protocol version.
 */
#define Version_CURRENT_PROTOCOL 21
#define Version_20_COMPAT

#define Version_MINIMUM_COMPATIBLE 20
#define Version_DEFAULT_ASSUMPTION 20

/**
 * Check the compatibility matrix and return whether two versions are compatible.
 * If a version is not listed on the table, the highest version on the table is
 * substituted for it but if the return value is yes, it is changed to maybe.
 *
 * @param version1 the first version
 * @param version2 the second version
 * @return 1 meaning compatible or 0 meaning incompatible.
 */
int Version_isCompatible(uint32_t one, uint32_t two);

#endif