Tor Path Specification Roger Dingledine Nick Mathewson Note: This is an attempt to specify Tor as currently implemented. Future versions of Tor will implement improved algorithms. This document tries to cover how Tor chooses to build circuits and assign streams to circuits. Other implementations MAY take other approaches, but implementors should be aware of the anonymity and load-balancing implications of their choices. THIS SPEC ISN'T DONE YET. The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119. 1. General operation Tor begins building circuits as soon as it has enough directory information to do so (see section 5 of dir-spec.txt). Some circuits are built preemptively because we expect to need them later (for user traffic), and some are built because of immediate need (for user traffic that no current circuit can handle, for testing the network or our reachability, and so on). When a client application creates a new stream (by opening a SOCKS connection or launching a resolve request), we attach it to an appropriate open circuit if one exists, or wait if an appropriate circuit is in-progress. We launch a new circuit only if no current circuit can handle the request. We rotate circuits over time to avoid some profiling attacks. To build a circuit, we choose all the nodes we want to use, and then construct the circuit. Sometimes, when we want a circuit that ends at a given hop, and we have an appropriate unused circuit, we "cannibalize" the existing circuit and extend it to the new terminus. These processes are described in more detail below. This document describes Tor's automatic path selection logic only; path selection can be overridden by a controller (with the EXTENDCIRCUIT and ATTACHSTREAM commands). Paths constructed through these means may violate some constraints given below. 1.1. Terminology A "path" is an ordered sequence of nodes, not yet built as a circuit. A "clean" circuit is one that has not yet been used for any traffic. A "fast" or "stable" or "valid" node is one that has the 'Fast' or 'Stable' or 'Valid' flag set respectively, based on our current directory information. A "fast" or "stable" circuit is one consisting only of "fast" or "stable" nodes. In an "exit" circuit, the final node is chosen based on waiting stream requests if any, and in any case it avoids nodes with exit policy of "reject *:*". An "internal" circuit, on the other hand, is one where the final node is chosen just like a middle node (ignoring its exit policy). A "request" is a client-side stream or DNS resolve that needs to be served by a circuit. A "pending" circuit is one that we have started to build, but which has not yet completed. A circuit or path "supports" a request if it is okay to use the circuit/path to fulfill the request, according to the rules given below. A circuit or path "might support" a request if some aspect of the request is unknown (usually its target IP), but we believe the path probably supports the request according to the rules given below. 1.1. A server's bandwidth Old versions of Tor did not report bandwidths in network status documents, so clients had to learn them from the routers' advertised server descriptors. For versions of Tor prior to 0.2.1.17-rc, everywhere below where we refer to a server's "bandwidth", we mean its clipped advertised bandwidth, computed by taking the smaller of the 'rate' and 'observed' arguments to the "bandwidth" element in the server's descriptor. If a router's advertised bandwidth is greater than MAX_BELIEVABLE_BANDWIDTH (currently 10 MB/s), we clipped to that value. For more recent versions of Tor, we take the bandwidth value declared in the consensus, and fall back to the clipped advertised bandwidth only if the consensus does not have bandwidths listed. 2. Building circuits 2.1. When we build 2.1.1. Clients build circuits preemptively When running as a client, Tor tries to maintain at least a certain number of clean circuits, so that new streams can be handled quickly. To increase the likelihood of success, Tor tries to predict what circuits will be useful by choosing from among nodes that support the ports we have used in the recent past (by default one hour). Specifically, on startup Tor tries to maintain one clean fast exit circuit that allows connections to port 80, and at least two fast clean stable internal circuits in case we get a resolve request or hidden service request (at least three if we _run_ a hidden service). After that, Tor will adapt the circuits that it preemptively builds based on the requests it sees from the user: it tries to have two fast clean exit circuits available for every port seen within the past hour (each circuit can be adequate for many predicted ports -- it doesn't need two separate circuits for each port), and it tries to have the above internal circuits available if we've seen resolves or hidden service activity within the past hour. If there are 12 or more clean circuits open, it doesn't open more even if it has more predictions. Only stable circuits can "cover" a port that is listed in the LongLivedPorts config option. Similarly, hidden service requests to ports listed in LongLivedPorts make us create stable internal circuits. Note that if there are no requests from the user for an hour, Tor will predict no use and build no preemptive circuits. The Tor client SHOULD NOT store its list of predicted requests to a persistent medium. 2.1.2. Clients build circuits on demand Additionally, when a client request exists that no circuit (built or pending) might support, we create a new circuit to support the request. For exit connections, we pick an exit node that will handle the most pending requests (choosing arbitrarily among ties), launch a circuit to end there, and repeat until every unattached request might be supported by a pending or built circuit. For internal circuits, we pick an arbitrary acceptable path, repeating as needed. In some cases we can reuse an already established circuit if it's clean; see Section 2.3 (cannibalizing circuits) for details. 2.1.3. Servers build circuits for testing reachability and bandwidth Tor servers test reachability of their ORPort once they have successfully built a circuit (on start and whenever their IP address changes). They build an ordinary fast internal circuit with themselves as the last hop. As soon as any testing circuit succeeds, the Tor server decides it's reachable and is willing to publish a descriptor. We launch multiple testing circuits (one at a time), until we have NUM_PARALLEL_TESTING_CIRC (4) such circuits open. Then we do a "bandwidth test" by sending a certain number of relay drop cells down each circuit: BandwidthRate * 10 / CELL_NETWORK_SIZE total cells divided across the four circuits, but never more than CIRCWINDOW_START (1000) cells total. This exercises both outgoing and incoming bandwidth, and helps to jumpstart the observed bandwidth (see dir-spec.txt). Tor servers also test reachability of their DirPort once they have established a circuit, but they use an ordinary exit circuit for this purpose. 2.1.4. Hidden-service circuits See section 4 below. 2.1.5. Rate limiting of failed circuits If we fail to build a circuit N times in a X second period (see Section 2.3 for how this works), we stop building circuits until the X seconds have elapsed. XXXX 2.1.6. When to tear down circuits XXXX 2.2. Path selection and constraints We choose the path for each new circuit before we build it. We choose the exit node first, followed by the other nodes in the circuit. All paths we generate obey the following constraints: - We do not choose the same router twice for the same path. - We do not choose any router in the same family as another in the same path. - We do not choose more than one router in a given /16 subnet (unless EnforceDistinctSubnets is 0). - We don't choose any non-running or non-valid router unless we have been configured to do so. By default, we are configured to allow non-valid routers in "middle" and "rendezvous" positions. - If we're using Guard nodes, the first node must be a Guard (see 5 below) - XXXX Choosing the length For "fast" circuits, we only choose nodes with the Fast flag. For non-"fast" circuits, all nodes are eligible. For all circuits, we weight node selection according to router bandwidth. We also weight the bandwidth of Exit and Guard flagged nodes depending on the fraction of total bandwidth that they make up and depending upon the position they are being selected for. These weights are published in the consensus, and are computed as described in Section 3.4.3 of dir-spec.txt. They are: Wgg - Weight for Guard-flagged nodes in the guard position Wgm - Weight for non-flagged nodes in the guard Position Wgd - Weight for Guard+Exit-flagged nodes in the guard Position Wmg - Weight for Guard-flagged nodes in the middle Position Wmm - Weight for non-flagged nodes in the middle Position Wme - Weight for Exit-flagged nodes in the middle Position Wmd - Weight for Guard+Exit flagged nodes in the middle Position Weg - Weight for Guard flagged nodes in the exit Position Wem - Weight for non-flagged nodes in the exit Position Wee - Weight for Exit-flagged nodes in the exit Position Wed - Weight for Guard+Exit-flagged nodes in the exit Position Wgb - Weight for BEGIN_DIR-supporting Guard-flagged nodes Wmb - Weight for BEGIN_DIR-supporting non-flagged nodes Web - Weight for BEGIN_DIR-supporting Exit-flagged nodes Wdb - Weight for BEGIN_DIR-supporting Guard+Exit-flagged nodes Wbg - Weight for Guard+Exit-flagged nodes for BEGIN_DIR requests Wbm - Weight for Guard+Exit-flagged nodes for BEGIN_DIR requests Wbe - Weight for Guard+Exit-flagged nodes for BEGIN_DIR requests Wbd - Weight for Guard+Exit-flagged nodes for BEGIN_DIR requests Additionally, we may be building circuits with one or more requests in mind. Each kind of request puts certain constraints on paths: - All service-side introduction circuits and all rendezvous paths should be Stable. - All connection requests for connections that we think will need to stay open a long time require Stable circuits. Currently, Tor decides this by examining the request's target port, and comparing it to a list of "long-lived" ports. (Default: 21, 22, 706, 1863, 5050, 5190, 5222, 5223, 6667, 6697, 8300.) - DNS resolves require an exit node whose exit policy is not equivalent to "reject *:*". - Reverse DNS resolves require a version of Tor with advertised eventdns support (available in Tor 0.1.2.1-alpha-dev and later). - All connection requests require an exit node whose exit policy supports their target address and port (if known), or which "might support it" (if the address isn't known). See 2.2.1. - Rules for Fast? XXXXX 2.2.1. Choosing an exit If we know what IP address we want to connect to or resolve, we can trivially tell whether a given router will support it by simulating its declared exit policy. Because we often connect to addresses of the form hostname:port, we do not always know the target IP address when we select an exit node. In these cases, we need to pick an exit node that "might support" connections to a given address port with an unknown address. An exit node "might support" such a connection if any clause that accepts any connections to that port precedes all clauses (if any) that reject all connections to that port. Unless requested to do so by the user, we never choose an exit server flagged as "BadExit" by more than half of the authorities who advertise themselves as listing bad exits. 2.2.2. User configuration Users can alter the default behavior for path selection with configuration options. - If "ExitNodes" is provided, then every request requires an exit node on the ExitNodes list. (If a request is supported by no nodes on that list, and StrictExitNodes is false, then Tor treats that request as if ExitNodes were not provided.) - "EntryNodes" and "StrictEntryNodes" behave analogously. - If a user tries to connect to or resolve a hostname of the form ..exit, the request is rewritten to a request for , and the request is only supported by the exit whose nickname or fingerprint is . 2.3. Cannibalizing circuits If we need a circuit and have a clean one already established, in some cases we can adapt the clean circuit for our new purpose. Specifically, For hidden service interactions, we can "cannibalize" a clean internal circuit if one is available, so we don't need to build those circuits from scratch on demand. We can also cannibalize clean circuits when the client asks to exit at a given node -- either via the ".exit" notation or because the destination is running at the same location as an exit node. 2.4. Learning when to give up ("timeout") on circuit construction Since version 0.2.2.8-alpha, Tor attempts to learn when to give up on circuits based on network conditions. 2.4.1 Distribution choice and parameter estimation Based on studies of build times, we found that the distribution of circuit build times appears to be a Frechet distribution. However, estimators and quantile functions of the Frechet distribution are difficult to work with and slow to converge. So instead, since we are only interested in the accuracy of the tail, we approximate the tail of the distribution with a Pareto curve. We calculate the parameters for a Pareto distribution fitting the data using the estimators in equation 4 from: http://portal.acm.org/citation.cfm?id=1647962.1648139 This is: alpha_m = s/(ln(U(X)/Xm^n)) where s is the total number of completed circuits we have seen, and U(X) = x_max^u * Prod_s{x_i} with x_i as our i-th completed circuit time, x_max as the longest completed circuit build time we have yet observed, u as the number of unobserved timeouts that have no exact value recorded, and n as u+s, the total number of circuits that either timeout or complete. Using log laws, we compute this as the sum of logs to avoid overflow and ln(1.0+epsilon) precision issues: alpha_m = s/(u*ln(x_max) + Sum_s{ln(x_i)} - n*ln(Xm)) This estimator is closely related to the parameters present in: http://en.wikipedia.org/wiki/Pareto_distribution#Parameter_estimation except they are adjusted to handle the fact that our samples are right-censored at the timeout cutoff. Additionally, because this is not a true Pareto distribution, we alter how Xm is computed. The Xm parameter is computed as the midpoint of the most frequently occurring 50ms histogram bin, until the point where 1000 circuits are recorded. After this point, the weighted average of the top 'cbtnummodes' (default: 3) midpoint modes is used as Xm. All times below this value are counted as having the midpoint value of this weighted average bin. The timeout itself is calculated by using the Pareto Quantile function (the inverted CDF) to give us the value on the CDF such that 80% of the mass of the distribution is below the timeout value. Thus, we expect that the Tor client will accept the fastest 80% of the total number of paths on the network. 2.4.2. How much data to record From our observations, the minimum number of circuit build times for a reasonable fit appears to be on the order of 100. However, to keep a good fit over the long term, we store 1000 most recent circuit build times in a circular array. The Tor client should build test circuits at a rate of one per minute up until 100 circuits are built. This allows a fresh Tor to have a CircuitBuildTimeout estimated within 1.5 hours after install, upgrade, or network change (see below). Timeouts are stored on disk in a histogram of 50ms bin width, the same width used to calculate the Xm value above. This histogram must be shuffled after being read from disk, to preserve a proper expiration of old values after restart. 2.4.3. How to record timeouts Circuits that pass the timeout threshold should be allowed to continue building until a time corresponding to the point 'cbtclosequantile' (default 95) on the Pareto curve, or 60 seconds, whichever is greater. The actual completion times for these circuits should be recorded. Implementations should completely abandon a circuit and record a value as an 'unknown' timeout if the total build time exceeds this threshold. The reason for this is that right-censored pareto estimators begin to lose their accuracy if more than approximately 5% of the values are censored. Since we wish to set the cutoff at 20%, we must allow circuits to continue building past this cutoff point up to the 95th percentile. 2.4.4. Detecting Changing Network Conditions We attempt to detect both network connectivity loss and drastic changes in the timeout characteristics. We assume that we've had network connectivity loss if 3 circuits timeout and we've received no cells or TLS handshakes since those circuits began. We then temporarily set the timeout to 60 seconds and stop counting timeouts. If 3 more circuits timeout and the network still has not been live within this new 60 second timeout window, we then discard the previous timeouts during this period from our history. To detect changing network conditions, we keep a history of the timeout or non-timeout status of the past 20 circuits that successfully completed at least one hop. If more than 90% of these circuits timeout, we discard all buildtimes history, reset the timeout to 60, and then begin recomputing the timeout. If the timeout was already 60 or higher, we double the timeout. 2.4.5. Consensus parameters governing behavior Clients that implement circuit build timeout learning should obey the following consensus parameters that govern behavior, in order to allow us to handle bugs or other emergent behaviors due to client circuit construction. If these parameters are not present in the consensus, the listed default values should be used instead. cbtdisabled Default: 0 Effect: If non-zero, all CircuitBuildTime learning code should be disabled and history should be discarded. For use in emergency situations only. cbtnummodes Default: 3 Effect: This value governs how many modes to use in the weighted average calculation of Pareto paramter Xm. A value of 3 introduces some bias (2-5% of CDF) under ideal conditions, but allows for better performance in the event that a client chooses guard nodes of radically different performance characteristics. cbtrecentcount Default: 20 Effect: This is the number of circuit build times to keep track of for the following option. cbtmaxtimeouts Default: 18 Effect: When this many timeouts happen in the last 'cbtrecentcount' circuit attempts, the client should discard all of its history and begin learning a fresh timeout value. cbtmincircs Default: 100 Effect: This is the minimum number of circuits to build before computing a timeout. cbtquantile Default: 80 Effect: This is the position on the quantile curve to use to set the timeout value. It is a percent (0-99). cbtclosequantile Default: 95 Effect: This is the position on the quantile curve to use to set the timeout value to use to actually close circuits. It is a percent (0-99). cbttestfreq Default: 60 Effect: Describes how often in seconds to build a test circuit to gather timeout values. Only applies if less than 'cbtmincircs' have been recorded. cbtmintimeout Default: 2000 Effect: This is the minimum allowed timeout value in milliseconds. cbtinitialtimeout Default: 60000 Effect: This is the timeout value to use before computing a timeout, in milliseconds. 2.5. Handling failure If an attempt to extend a circuit fails (either because the first create failed or a subsequent extend failed) then the circuit is torn down and is no longer pending. (XXXX really?) Requests that might have been supported by the pending circuit thus become unsupported, and a new circuit needs to be constructed. If a stream "begin" attempt fails with an EXITPOLICY error, we decide that the exit node's exit policy is not correctly advertised, so we treat the exit node as if it were a non-exit until we retrieve a fresh descriptor for it. XXXX 3. Attaching streams to circuits When a circuit that might support a request is built, Tor tries to attach the request's stream to the circuit and sends a BEGIN, BEGIN_DIR, or RESOLVE relay cell as appropriate. If the request completes unsuccessfully, Tor considers the reason given in the CLOSE relay cell. [XXX yes, and?] After a request has remained unattached for SocksTimeout (2 minutes by default), Tor abandons the attempt and signals an error to the client as appropriate (e.g., by closing the SOCKS connection). XXX Timeouts and when Tor auto-retries. * What stream-end-reasons are appropriate for retrying. If no reply to BEGIN/RESOLVE, then the stream will timeout and fail. 4. Hidden-service related circuits XXX Tracking expected hidden service use (client-side and hidserv-side) 5. Guard nodes We use Guard nodes (also called "helper nodes" in the literature) to prevent certain profiling attacks. Here's the risk: if we choose entry and exit nodes at random, and an attacker controls C out of N servers (ignoring bandwidth), then the attacker will control the entry and exit node of any given circuit with probability (C/N)^2. But as we make many different circuits over time, then the probability that the attacker will see a sample of about (C/N)^2 of our traffic goes to 1. Since statistical sampling works, the attacker can be sure of learning a profile of our behavior. If, on the other hand, we picked an entry node and held it fixed, we would have probability C/N of choosing a bad entry and being profiled, and probability (N-C)/N of choosing a good entry and not being profiled. When guard nodes are enabled, Tor maintains an ordered list of entry nodes as our chosen guards, and stores this list persistently to disk. If a Guard node becomes unusable, rather than replacing it, Tor adds new guards to the end of the list. When choosing the first hop of a circuit, Tor chooses at random from among the first NumEntryGuards (default 3) usable guards on the list. If there are not at least 2 usable guards on the list, Tor adds routers until there are, or until there are no more usable routers to add. A guard is unusable if any of the following hold: - it is not marked as a Guard by the networkstatuses, - it is not marked Valid (and the user hasn't set AllowInvalid entry) - it is not marked Running - Tor couldn't reach it the last time it tried to connect A guard is unusable for a particular circuit if any of the rules for path selection in 2.2 are not met. In particular, if the circuit is "fast" and the guard is not Fast, or if the circuit is "stable" and the guard is not Stable, or if the guard has already been chosen as the exit node in that circuit, Tor can't use it as a guard node for that circuit. If the guard is excluded because of its status in the networkstatuses for over 30 days, Tor removes it from the list entirely, preserving order. If Tor fails to connect to an otherwise usable guard, it retries periodically: every hour for six hours, every 4 hours for 3 days, every 18 hours for a week, and every 36 hours thereafter. Additionally, Tor retries unreachable guards the first time it adds a new guard to the list, since it is possible that the old guards were only marked as unreachable because the network was unreachable or down. Tor does not add a guard persistently to the list until the first time we have connected to it successfully. 6. Router descriptor purposes There are currently three "purposes" supported for router descriptors: general, controller, and bridge. Most descriptors are of type general -- these are the ones listed in the consensus, and the ones fetched and used in normal cases. Controller-purpose descriptors are those delivered by the controller and labelled as such: they will be kept around (and expire like normal descriptors), and they can be used by the controller in its CIRCUITEXTEND commands. Otherwise they are ignored by Tor when it chooses paths. Bridge-purpose descriptors are for routers that are used as bridges. See doc/design-paper/blocking.pdf for more design explanation, or proposal 125 for specific details. Currently bridge descriptors are used in place of normal entry guards, for Tor clients that have UseBridges enabled. X. Old notes X.1. Do we actually do this? How to deal with network down. - While all helpers are down/unreachable and there are no established or on-the-way testing circuits, launch a testing circuit. (Do this periodically in the same way we try to establish normal circuits when things are working normally.) (Testing circuits are a special type of circuit, that streams won't attach to by accident.) - When a testing circuit succeeds, mark all helpers up and hold the testing circuit open. - If a connection to a helper succeeds, close all testing circuits. Else mark that helper down and try another. - If the last helper is marked down and we already have a testing circuit established, then add the first hop of that testing circuit to the end of our helper node list, close that testing circuit, and go back to square one. (Actually, rather than closing the testing circuit, can we get away with converting it to a normal circuit and beginning to use it immediately?) [Do we actually do any of the above? If so, let's spec it. If not, let's remove it. -NM] X.2. A thing we could do to deal with reachability. And as a bonus, it leads to an answer to Nick's attack ("If I pick my helper nodes all on 18.0.0.0:*, then I move, you'll know where I bootstrapped") -- the answer is to pick your original three helper nodes without regard for reachability. Then the above algorithm will add some more that are reachable for you, and if you move somewhere, it's more likely (though not certain) that some of the originals will become useful. Is that smart or just complex? X.3. Some stuff that worries me about entry guards. 2006 Jun, Nickm. It is unlikely for two users to have the same set of entry guards. Observing a user is sufficient to learn its entry guards. So, as we move around, entry guards make us linkable. If we want to change guards when our location (IP? subnet?) changes, we have two bad options. We could - Drop the old guards. But if we go back to our old location, we'll not use our old guards. For a laptop that sometimes gets used from work and sometimes from home, this is pretty fatal. - Remember the old guards as associated with the old location, and use them again if we ever go back to the old location. This would be nasty, since it would force us to record where we've been. [Do we do any of this now? If not, this should move into 099-misc or 098-todo. -NM]