/* 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 . */ #include "benc/String.h" #include "dht/Address.h" #include "dht/dhtcore/RouterModule.h" #include "dht/dhtcore/RouterModule_pvt.h" #include "dht/dhtcore/Node.h" #include "dht/dhtcore/NodeList.h" #include "dht/dhtcore/NodeStore.h" #include "dht/dhtcore/VersionList.h" #include "dht/CJDHTConstants.h" #include "dht/DHTMessage.h" #include "dht/DHTModule.h" #include "dht/DHTModuleRegistry.h" #include "util/log/Log.h" #include "memory/Allocator.h" #include "switch/LabelSplicer.h" #include "switch/NumberCompress.h" #include "util/events/EventBase.h" #include "util/AverageRoller.h" #include "util/Bits.h" #include "util/Hex.h" #include "util/Endian.h" #include "util/Pinger.h" #include "util/events/Time.h" #include "util/events/Timeout.h" #include "util/version/Version.h" #include "wire/Message.h" /* * The router module is the central part of the DHT engine. * It's job is to maintain a routing table which is updated by all incoming packets. * When it gets an incoming query, its job is to add nodes to the reply so that the asking node * can find other nodes which are closer to its target than us. * * This implementation does not split nodes explicitly into buckets not does it explicitly try to * distinguish between "good" and "bad" nodes. Instead it tries to determine which node will help * get to the requested record the fastest. Instead of periodicly pinging a random node in each * "bucket", this implementation periodically searches for a random[1] hash. When a node is sent a * query, the the distance[2] between it and the first node is divided by the amount of time it * takes the node to respond, for each successful search, this number is added to an attribute of * the node called "reach". * * Visually representing a node as an area whose location is defined by the node id and its size is * defined by the node reach, you can see that there is a possibility for a record to be closer in * key space to node2 while is is still further inside of node1's reach thus node1 is a better * choice for the next node to ask. * * |<--------- Node 1 ---------->| * |<--- Node 2 ---->| * ^----- Desired record location. * * New nodes are inserted into the table but with a reach of 0. It is up to the search client to * send search requests to them so they can prove their validity and have their reach number * updated. * * Reach of a node is incremented by 2 every time the node responds to a query and incremented by 1 * every time a node sends a query of its own. This has almost no effect except that it means a * node which has recently sent data will be preferred over one which has not. * * When a search is carried out, the next K returned nodes are not necessarily the closest known * nodes to the id of the record. The nodes returned will be the nodes with the lowest * distance:reach ratio. The distance:reach ratio is calculated by dividing the distance between * the node and the record by the node's reach number. Actually it is done by multiplying * UINT32_MAX minus the distance by the reach so that it does not need to use slower divison. * See: NodeCollector.h * * Since information about a node becomes stale over time, all reach numbers are decreased by * the constant REACH_DECREASE_PER_SECOND times the number of seconds since the last cycle, * this operation is performed periodicly every LOCAL_MAINTENANCE_SEARCH_MILLISECONDS unless * a local maintainence search is being run which is not often once the network is stable. * * TODO(cjd): --- * In order to have the nodes with least distance:reach ratio ready to handle any incoming search, * we precompute the borders where the "best next node" changes. This computation is best understood * by graphing the nodes with their location in keyspace on the X axis and their reach on the Y * axis. The border between two nodes, nodeA and nodeB is the location where a line drawn from the * X axis up to either node location would be the same angle. * * ^ ^ * | nodeA | nodeA * | |\ | |\__ * | | \ | | \__ * | | \ nodeB | | \nodeB * | | \ /| | | \__ * | | \ / | | | | \__ * | | \/ | | | | \__ * +---------------------------------------> +---------------------------------------> * ^-- border ^-- border2 * * Everything to the left of the border and everything to the right of border2 is to be serviced by * nodeA. Everything between the two borders is serviced by nodeB. Border2 is found by * drawing a line from the point given for nodeA to through the point given for nodeB and finding * the intersection of that line with the Y axis. border and border2 are shown on different graphs * only to limit clutter, they are the same nodeA and nodeB. * * When resolving a search, this implementation will lookup the location of the searched for record * and return the nodes which belong to the insides of the nearest K borders, this guarantees return * of the nodes whose distance:reach ratio is the lowest for that location. * --- * * This implementation must never respond to a search by sending any node who's id is not closer * to the target than its own. Such an event would lead to the possibility of "routing loops" and * must be prevented. Searches for which this node has the lowest distance:reach ratio will be * replied to with nodes which have 0 reach but are closer than this node or, if there are no such * nodes, no nodes at all. * * The search consumer in this routing module tries to minimize the amount of traffic sent when * doing a lookup. To achieve this, it sends a request only to the last node in the search response * packet, after the global mean response time has passed without it getting a response, it sends * requests to the second to last and so forth, working backward. Nodes which fail to respond in * time have their reach immedietly set to zero. * * The global mean response time is the average amount of time it takes a node to respond to a * search query. It is a rolling average over the past 256 seconds. * * To maximize the quality of service offered by this node this implementation will repeat * searches which it handles every number of seconds given by the constant: * GLOBAL_MAINTENANCE_SEARCH_MILLISECONDS. * Any incoming search with a get_peers request is eligable to be repeated. * * [1] The implementation runs periodic searches for random hashes but unless the search target * is closer in keyspace to this node than it is to any node with non-zero reach, the search * is not performed. This means that the node will send out lots of searches when it doesn't * know many reliable nodes but it will taper off like a governer as it becomes more * integrated in the network. These searches are run every number of milliseconds given * by the constant LOCAL_MAINTENANCE_SEARCH_MILLISECONDS. * * [2] If a response "overshoots" the record requested then it is calculated as if it had undershot * by the same amount so as not to provide arbitrage advantage to nodes who return results which * are very far away yet very inaccurate. If it overshoots by more than the distance between the * node and the searched for location (this should never happen), it is considered to be 0. */ /*--------------------Constants--------------------*/ /** The number of seconds of time overwhich to calculate the global mean response time. */ #define GMRT_SECONDS 256 /** * The number to initialize the global mean response time averager with so that it will * return sane results early on, this number can be much higher than the expected average. */ #define GMRT_INITAL_MILLISECONDS 5000 /** The number of nodes which we will keep track of. */ #define NODE_STORE_SIZE 8192 /** The number of milliseconds between attempting local maintenance searches. */ #define LOCAL_MAINTENANCE_SEARCH_MILLISECONDS 1000 /** * The number of milliseconds to pass between global maintainence searches. * These are searches for random targets which are used to discover new nodes. */ #define GLOBAL_MAINTENANCE_SEARCH_MILLISECONDS 30000 #define SEARCH_REPEAT_MILLISECONDS 7500 /** The number of times the GMRT before pings should be timed out. */ #define PING_TIMEOUT_GMRT_MULTIPLIER 100 /** The minimum amount of time before a ping should timeout. */ #define PING_TIMEOUT_MINIMUM 3000 /** You are not expected to understand this. */ #define LINK_STATE_MULTIPLIER 536870 /** All searches will be killed after this amount of time nomatter how bad the GMRT is. */ #define MAX_TIMEOUT 10000 /** Never allow a search to be timed out in less than this number of milliseconds. */ #define MIN_TIMEOUT 10 /*--------------------Prototypes--------------------*/ static int handleIncoming(struct DHTMessage* message, void* vcontext); static int handleOutgoing(struct DHTMessage* message, void* vcontext); /*--------------------Interface--------------------*/ /** * Register a new RouterModule. * * @param registry the DHT module registry for signal handling. * @param allocator a means to allocate memory. * @param myAddress the address for this DHT node. * @param nodeStore the place to put the nodes * @return the RouterModule. */ struct RouterModule* RouterModule_register(struct DHTModuleRegistry* registry, struct Allocator* allocator, const uint8_t myAddress[Address_KEY_SIZE], struct EventBase* eventBase, struct Log* logger, struct Random* rand, struct NodeStore* nodeStore) { struct RouterModule* const out = Allocator_calloc(allocator, sizeof(struct RouterModule), 1); struct DHTModule* dm = Allocator_clone(allocator, (&(struct DHTModule) { .name = "RouterModule", .context = out, .handleIncoming = handleIncoming, .handleOutgoing = handleOutgoing })); DHTModuleRegistry_register(dm, registry); Address_forKey(&out->address, myAddress); out->gmrtRoller = AverageRoller_new(GMRT_SECONDS, eventBase, allocator); AverageRoller_update(out->gmrtRoller, GMRT_INITAL_MILLISECONDS); out->nodeStore = nodeStore; out->registry = registry; out->eventBase = eventBase; out->logger = logger; out->allocator = allocator; out->rand = rand; out->pinger = Pinger_new(eventBase, rand, logger, allocator); Identity_set(out); return out; } /** * The amount of time to wait before skipping over the first node and trying another in a search. * Any node which can't beat this time will have its reach set to 0. * * @param module this module. * @return the timeout time. */ uint64_t RouterModule_searchTimeoutMilliseconds(struct RouterModule* module) { uint64_t x = (((uint64_t) AverageRoller_getAverage(module->gmrtRoller)) * 4); x = x + (Random_uint32(module->rand) % (x | 1)) / 2; return (x > MAX_TIMEOUT) ? MAX_TIMEOUT : (x < MIN_TIMEOUT) ? MIN_TIMEOUT : x; } static inline int sendNodes(struct NodeList* nodeList, struct DHTMessage* message, struct RouterModule* module, uint32_t askerVersion) { struct DHTMessage* query = message->replyTo; String* nodes = String_newBinary(NULL, nodeList->size * Address_SERIALIZED_SIZE, message->allocator); struct VersionList* versions = VersionList_new(nodeList->size, message->allocator); int i = 0; for (; i < (int)nodeList->size; i++) { // We have to modify the reply in case this node uses a longer label discriminator // in our switch than its target address, the target address *must* have the same // length or longer. struct Address addr; Bits_memcpyConst(&addr, &nodeList->nodes[i]->address, sizeof(struct Address)); addr.path = LabelSplicer_getLabelFor(addr.path, query->address->path); Address_serialize(&nodes->bytes[i * Address_SERIALIZED_SIZE], &addr); versions->versions[i] = nodeList->nodes[i]->address.protocolVersion; Assert_ifParanoid(!Bits_isZero(&nodes->bytes[i * Address_SERIALIZED_SIZE], Address_SERIALIZED_SIZE)); } nodes->len = i * Address_SERIALIZED_SIZE; versions->length = i; if (i > 0) { Dict_putString(message->asDict, CJDHTConstants_NODES, nodes, message->allocator); String* versionsStr = VersionList_stringify(versions, message->allocator); Dict_putString(message->asDict, CJDHTConstants_NODE_PROTOCOLS, versionsStr, message->allocator); } return 0; } /** * Handle an incoming search query. * This is setup to handle the outgoing *response* to the query, it should * be called from handleOutgoing() and populate the response with nodes. * * @param message the empty response message to populate. * @param replyArgs the arguments dictionary in the response (to be populated). * @param module the routing module context. * @return 0 as long as the packet should not be stopped (at this point always 0). */ static inline int handleQuery(struct DHTMessage* message, struct RouterModule* module) { struct DHTMessage* query = message->replyTo; int64_t* versionPtr = Dict_getInt(query->asDict, CJDHTConstants_PROTOCOL); uint32_t version = (versionPtr && *versionPtr <= UINT32_MAX) ? *versionPtr : 0; struct NodeList* nodeList = NULL; String* queryType = Dict_getString(query->asDict, CJDHTConstants_QUERY); if (String_equals(queryType, CJDHTConstants_QUERY_FN)) { // get the target String* target = Dict_getString(query->asDict, CJDHTConstants_TARGET); if (target == NULL || target->len != Address_SEARCH_TARGET_SIZE) { return 0; } struct Address targetAddr = { .path = 0 }; Bits_memcpyConst(targetAddr.ip6.bytes, target->bytes, Address_SEARCH_TARGET_SIZE); // send the closest nodes nodeList = NodeStore_getClosestNodes(module->nodeStore, &targetAddr, RouterModule_K, version, message->allocator); } else if (String_equals(queryType, CJDHTConstants_QUERY_GP)) { // get the target String* target = Dict_getString(query->asDict, CJDHTConstants_TARGET); if (target == NULL || target->len != 8) { return 0; } uint64_t targetPath; Bits_memcpyConst(&targetPath, target->bytes, 8); targetPath = Endian_bigEndianToHost64(targetPath); nodeList = NodeStore_getPeers(targetPath, RouterModule_K, message->allocator, module->nodeStore); } return (nodeList) ? sendNodes(nodeList, message, module, version) : 0; } /** * We handle 2 kinds of packets on the outgoing. * 1. our requests * 2. our replies to others' requests. * Everything is tagged with our address, replies to requests which are not ping requests * will also be given a list of nodes. */ static int handleOutgoing(struct DHTMessage* message, void* vcontext) { struct RouterModule* module = Identity_check((struct RouterModule*) vcontext); Dict_putInt(message->asDict, CJDHTConstants_PROTOCOL, Version_CURRENT_PROTOCOL, message->allocator); if (message->replyTo != NULL) { return handleQuery(message, module); } return 0; } struct PingContext { struct RouterModule_Promise pub; /** nonNull if this ping is part of a search. */ struct SearchContext* search; struct RouterModule* router; struct Address address; /** The internal ping structure */ struct Pinger_Ping* pp; /** A template of the message to be sent. */ Dict* messageDict; Identity }; static void sendMsg(String* txid, void* vpingContext) { struct PingContext* pc = Identity_check((struct PingContext*) vpingContext); // "t":"1234" Dict_putString(pc->messageDict, CJDHTConstants_TXID, txid, pc->pp->pingAlloc); struct Allocator* temp = Allocator_child(pc->pp->pingAlloc); struct Message* msg = Message_new(0, DHTMessage_MAX_SIZE + 512, temp); struct DHTMessage* dmesg = Allocator_calloc(temp, sizeof(struct DHTMessage), 1); dmesg->binMessage = msg; dmesg->address = &pc->address; dmesg->asDict = pc->messageDict; dmesg->allocator = temp; DHTModuleRegistry_handleOutgoing(dmesg, pc->router->registry); } static void onTimeout(uint32_t milliseconds, struct PingContext* pctx) { struct Node_Two* n = NodeStore_closestNode(pctx->router->nodeStore, pctx->address.path); // Ping timeout -> decrease reach if (n && !Bits_memcmp(pctx->address.key, n->address.key, 32)) { NodeStore_pathTimeout(pctx->router->nodeStore, pctx->address.path); } if (pctx->pub.callback) { pctx->pub.callback(&pctx->pub, milliseconds, NULL, NULL); } } static uint64_t pingTimeoutMilliseconds(struct RouterModule* module) { uint64_t out = AverageRoller_getAverage(module->gmrtRoller) * PING_TIMEOUT_GMRT_MULTIPLIER; return (out < PING_TIMEOUT_MINIMUM) ? PING_TIMEOUT_MINIMUM : out; } /** * The only type of message we handle on the incoming side is * a response to one of our queries. */ static int handleIncoming(struct DHTMessage* message, void* vcontext) { String* txid = Dict_getString(message->asDict, CJDHTConstants_TXID); String* query = Dict_getString(message->asDict, CJDHTConstants_QUERY); if (query || !txid) { return 0; } struct RouterModule* module = Identity_check((struct RouterModule*) vcontext); // This is retreived below by onResponseOrTimeout() module->currentMessage = message; Pinger_pongReceived(txid, module->pinger); module->currentMessage = NULL; return 0; } // ping or search response came in static void onResponseOrTimeout(String* data, uint32_t milliseconds, void* vping) { struct PingContext* pctx = Identity_check((struct PingContext*) vping); if (data == NULL) { // This is how Pinger signals a timeout. onTimeout(milliseconds, pctx); return; } struct RouterModule* module = pctx->router; // Grab out the message which was put here in handleIncoming() struct DHTMessage* message = module->currentMessage; module->currentMessage = NULL; // This should never happen if (!Address_isSameIp(&pctx->address, message->address)) { #ifdef Log_WARN uint8_t expectedAddr[60]; Address_print(expectedAddr, &pctx->address); uint8_t receivedAddr[60]; Address_print(receivedAddr, message->address); Log_warn(module->logger, "Got return packet from different address than search was sent!\n" "Expected:%s\n" " Got:%s\n", expectedAddr, receivedAddr); #endif return; } // update the GMRT AverageRoller_update(pctx->router->gmrtRoller, milliseconds); /* Log_debug(pctx->router->logger, "Received response in %u milliseconds, gmrt now %u\n", milliseconds, AverageRoller_getAverage(pctx->router->gmrtRoller)); */ // prevent division by zero if (milliseconds == 0) { milliseconds++; } struct Node_Two* node = NodeStore_closestNode(module->nodeStore, message->address->path); if (node && !Bits_memcmp(node->address.key, message->address->key, 32)) { NodeStore_pathResponse(module->nodeStore, message->address->path, milliseconds); } else { struct Node_Link* link = NodeStore_discoverNode(module->nodeStore, message->address, message->encodingScheme, message->encIndex, milliseconds); node = (link) ? link->child : NULL; } // EncodingSchemeModule should have added this node to the store, check it. if (!node) { #ifdef Log_DEBUG uint8_t printedAddr[60]; Address_print(printedAddr, message->address); Log_info(module->logger, "Got message from nonexistant node! [%s]\n", printedAddr); #endif return; } #ifdef Log_DEBUG String* versionBin = Dict_getString(message->asDict, CJDHTConstants_VERSION); if (versionBin && versionBin->len == 20) { uint8_t printedAddr[60]; Address_print(printedAddr, message->address); uint8_t versionStr[41]; Hex_encode(versionStr, 41, (uint8_t*) versionBin->bytes, 20); Log_debug(module->logger, "Got pong! [%s] ver[%s]\n", printedAddr, versionStr); } #endif if (pctx->pub.callback) { pctx->pub.callback(&pctx->pub, milliseconds, message->address, message->asDict); } } struct RouterModule_Promise* RouterModule_newMessage(struct Address* addr, uint32_t timeoutMilliseconds, struct RouterModule* module, struct Allocator* alloc) { // sending yourself a ping? // Assert_true(Bits_memcmp(addr->key, module->address.key, 32)); Assert_ifParanoid(addr->path == EncodingScheme_convertLabel(module->nodeStore->selfNode->encodingScheme, addr->path, EncodingScheme_convertLabel_convertTo_CANNONICAL)); if (timeoutMilliseconds == 0) { timeoutMilliseconds = pingTimeoutMilliseconds(module); } struct Pinger_Ping* pp = Pinger_newPing(NULL, onResponseOrTimeout, sendMsg, timeoutMilliseconds, alloc, module->pinger); struct PingContext* pctx = Allocator_clone(pp->pingAlloc, (&(struct PingContext) { .pub = { .alloc = pp->pingAlloc }, .router = module, .pp = pp })); Identity_set(pctx); Bits_memcpyConst(&pctx->address, addr, sizeof(struct Address)); pp->context = pctx; return &pctx->pub; } void RouterModule_sendMessage(struct RouterModule_Promise* promise, Dict* request) { struct PingContext* pctx = Identity_check((struct PingContext*)promise); pctx->messageDict = request; // actual send is triggered asynchronously } struct RouterModule_Promise* RouterModule_pingNode(struct Address* addr, uint32_t timeoutMilliseconds, struct RouterModule* module, struct Allocator* alloc) { struct RouterModule_Promise* promise = RouterModule_newMessage(addr, timeoutMilliseconds, module, alloc); Dict* d = Dict_new(promise->alloc); Dict_putString(d, CJDHTConstants_QUERY, CJDHTConstants_QUERY_PING, promise->alloc); RouterModule_sendMessage(promise, d); #ifdef Log_DEBUG uint8_t buff[60]; Address_print(buff, addr); Log_debug(module->logger, "Sending ping [%u] to [%s]", ((struct PingContext*)promise)->pp->handle, buff); #endif Assert_true(addr->path != 0); return promise; } struct RouterModule_Promise* RouterModule_getPeers(struct Address* addr, uint64_t nearbyLabel, uint32_t timeoutMilliseconds, struct RouterModule* module, struct Allocator* alloc) { struct RouterModule_Promise* promise = RouterModule_newMessage(addr, timeoutMilliseconds, module, alloc); Dict* d = Dict_new(promise->alloc); Dict_putString(d, CJDHTConstants_QUERY, CJDHTConstants_QUERY_GP, promise->alloc); uint64_t nearbyLabel_be = Endian_hostToBigEndian64(nearbyLabel); String* target = String_newBinary((char*)&nearbyLabel_be, 8, promise->alloc); Dict_putString(d, CJDHTConstants_TARGET, target, promise->alloc); RouterModule_sendMessage(promise, d); return promise; } struct Node_Two* RouterModule_lookup(uint8_t targetAddr[Address_SEARCH_TARGET_SIZE], struct RouterModule* module) { struct Address addr = { .path = 0 }; Bits_memcpyConst(addr.ip6.bytes, targetAddr, Address_SEARCH_TARGET_SIZE); return NodeStore_getBest(&addr, module->nodeStore); } struct Node_Two* RouterModule_nodeForPath(uint64_t path, struct RouterModule* module) { struct Node_Link* link = NodeStore_linkForPath(module->nodeStore, path); if (!link) { return NULL; } return link->child; } void RouterModule_brokenPath(const uint64_t path, struct RouterModule* module) { NodeStore_brokenPath(path, module->nodeStore); } uint32_t RouterModule_globalMeanResponseTime(struct RouterModule* module) { return (uint32_t) AverageRoller_getAverage(module->gmrtRoller); } void RouterModule_peerIsReachable(uint64_t pathToPeer, uint64_t lagMilliseconds, struct RouterModule* module) { Assert_ifParanoid(LabelSplicer_isOneHop(pathToPeer)); struct Node_Two* nn = RouterModule_nodeForPath(pathToPeer, module); for (struct Node_Link* peerLink = nn->reversePeers; peerLink; peerLink = peerLink->nextPeer) { if (peerLink->parent != module->nodeStore->selfNode) { continue; } if (peerLink->cannonicalLabel != pathToPeer) { continue; } struct Address address = { .path = 0 }; Bits_memcpyConst(&address, &nn->address, sizeof(struct Address)); address.path = pathToPeer; NodeStore_discoverNode(module->nodeStore, &address, nn->encodingScheme, peerLink->inverseLinkEncodingFormNumber, lagMilliseconds); return; } Assert_true(0); }