torspec/proposals/306-ipv6-happy-eyeballs.txt
teor e27cbca306
Prop 306: Restore some statistics from an earlier draft
These statistics are optional, but they may be useful:
* client and relay connections
* statistics in the heartbeat logs

Also improve the explanation of the "extra prop 306 connections"
statistic.

Part of 29801.
2020-01-15 22:44:36 +10:00

347 lines
14 KiB
Plaintext

Filename: 306-ipv6-happy-eyeballs.txt
Title: A Tor Implementation of IPv6 Happy Eyeballs
Author: Neel Chauhan
Created: 25-Jun-2019
Supercedes: 299
Status: Open
Ticket: https://trac.torproject.org/projects/tor/ticket/29801
1. Introduction
As IPv4 address space becomes scarce, ISPs and organizations will deploy
IPv6 in their networks. Right now, Tor clients connect to entry nodes using
IPv4 connectivity by default.
When networks first transition to IPv6, both IPv4 and IPv6 will be enabled
on most networks in a so-called "dual-stack" configuration. This is to not
break existing IPv4-only applications while enabling IPv6 connectivity.
However, IPv6 connectivity may be unreliable and clients should be able
to connect to the entry node using the most reliable technology, whether
IPv4 or IPv6.
In ticket #27490, we introduced the option ClientAutoIPv6ORPort which
lets a client randomly choose between IPv4 or IPv6. However, this
random decision does not take into account unreliable connectivity
or falling back to the alternate IP version should one be unreliable
or unavailable.
One way to select between IPv4 and IPv6 on a dual-stack network is a
so-called "Happy Eyeballs" algorithm as per RFC 8305. In one, a client
attempts the preferred IP family, whether IPv4 or IPv6. Should it work,
the client sticks with the preferred IP family. Otherwise, the client
attempts the alternate version. This means if a dual-stack client has
both IPv4 and IPv6, and IPv6 is unreliable, preferred or not, the
client uses IPv4, and vice versa. However, if IPv4 and IPv6 are both
equally reliable, and IPv6 is preferred, we use IPv6.
In Proposal 299, we have attempted a IP fallback mechanism using failure
counters and preferring IPv4 and IPv6 based on the state of the counters.
However, Prop299 was not standard Happy Eyeballs and an alternative,
standards-compliant proposal was requested in [P299-TRAC] to avoid issues
from complexity caused by randomness.
This proposal describes a Tor implementation of Happy Eyeballs and is
intended as a successor to Proposal 299.
2. Address/Relay Selection
This section describes the necessary changes for address selection to
implement Prop306.
2.1. Address Handling Changes
To be able to handle Happy Eyeballs in Tor, we will need to modify the
data structures used for connections to entry nodes, namely the extend info
structure.
Entry nodes are usually guards, but some clients don't use guards:
* Bootstrapping clients can connect to fallback directory mirrors or
authorities
* v3 single onion services can use IPv4 or IPv6 addresses to connect
to introduction and rendezvous points, and
* Clients can be configured to disable entry guards
Bridges are out of scope for this proposal, because Tor does not support
multiple IP addresses in a single bridge line.
The extend info structure should contain both an IPv4 and an IPv6 address.
This will allow us to try IPv4 and the IPv6 addresses should both be
available on a relay and the client is dual-stack.
When processing:
* relay descriptors,
* hard-coded authority and fallback directory lists,
* onion service descriptors, or
* onion service introduce cells,
and filling in the extend info data structure, we need to fill in both the
IPv4 and IPv6 address if both are available. If only one family is
available for a relay (IPv4 or IPv6), we should leave the other family null.
2.2 Bootstrap Changes
Tor's hard-coded authority and fallback directory mirror lists contain
some entries with IPv6 ORPorts. As of January 2020, 56% of authorities and
47% of fallback directories have IPv6.
During bootstrapping, we should have an option for the maximum number of
IPv4-only nodes, before the next node must have an IPv6 ORPort. The
parameter is as follows:
* MaxNumIPv4BootstrapAttempts NUM
During bootstrap, the minimum fraction of nodes with IPv6 ORPorts will be
1/(1 + MaxNumIPv4BootstrapAttempts). And the average fraction will be
larger than both minimum fraction, and the actual proportion of IPv6
ORPorts in the fallback directory list. (Clients mainly use fallback
directories for bootstrapping.)
Since this option is used during bootstrapping, it can not have a
corresponding consensus parameter.
The default value for MaxNumIPv4BootstrapAttempts should be 2. This
means that every third bootstrap node must have an IPv6 ORPort. And on
average, just over half of bootstrap nodes chosen by clients will have an
IPv6 ORPort. This change won't have much impact on load-balancing, because
almost half the fallback directory mirrors have IPv6 ORPorts.
The minimum value of MaxNumIPv4BootstrapAttempts is 0. (Every bootstrap
node must have an IPv6 ORPort. This setting is equivalent to
ClientPreferIPv6ORPort 1.)
The maximum value of MaxNumIPv4BootstrapAttempts should be 100. (Since
most clients only make a few bootstrap connections, bootstrap nodes will
be chosen at random, regardless of their IPv6 ORPorts.)
2.3. Guard Selection Changes
When we select guard candidates, we should have an option for the number of
primary IPv6 entry guards. The parameter is as follows:
* NumIPv6Guards NUM
If UseEntryGuards is set to 1, we will select exactly this number of IPv6
relays for our primary guard list, which is the set of relays we strongly
prefer when connecting to the Tor network. (This number should also apply
to all of Tor's other guard lists, scaled up based on the relative size of
the list.)
If NUM is -1, we try to learn the number from the NumIPv6Guards
consensus parameter. If the consensus parameter isn't set, we should
default to 1.
The default value for NumIPv6Guards should be -1. (Use the consensus
parameter, or the underlying default value of 1.)
As of September 2019, approximately 20% of Tor's guards supported IPv6,
by consensus weight. (Excluding exits that are also guards, because
clients avoid choosing exits in their guard lists.)
If all Tor clients implement NumIPv6Guards, then these 20% of guards will
handle approximately 33% of Tor's traffic. (Because the default value of
NumPrimaryGuards is 3.) This may have a significant impact on Tor's
load-balancing. Therefore, we should deploy this feature gradually, and try
to increase the number of relays that support IPv6 to at least 33%.
To minimise the impact on load-balancing, IPv6 support should only be
required for exactly NumIPv6Guards during guard list selection. All other
guards should be IPv4-only guards. Once approximately 50% of guards support
IPv6, NumIPv6Guards can become a minimum requirement, rather than an exact
requirement.
The minimum configurable value of NumIPv6Guards is -1. (Use the consensus
parameter, or the underlying default.)
The minimum resulting value of NumIPv6Guards is 0. (Guards will be chosen
at random, regardless of their IPv6 ORPorts.)
The maximum value of NumIPv6Guards should be the configured value of
NumPrimaryGuards. (Every guard must have an IPv6 ORPort. This setting is
equivalent to ClientPreferIPv6ORPort 1.)
3. Relay Connections
If there is an existing authenticated connection, we must use it similar
to how we used it pre-Prop306.
If there is no existing authenticated connection for an entry node, tor
currently attempts to connect using the first available, allowed, and
preferred address. (Determined using the existing Client IPv4 and IPv6
options.)
We should also allow falling back to the alternate address. For this,
a design will be given in Section 3.1.
3.1. TCP Connection to Preferred Address On First TCP Success
In this design, we will connect via TCP to the first preferred address.
On a failure or after a 250 msec delay, we attempt to connect via TCP to
the alternate address. On a success, Tor attempts to authenticate and
closes the other connection.
This design is close to RFC 8305 and is similar to how Happy Eyeballs
is implemented in a web browser.
3.2. Handling Connection Successes And Failures
Should a connection to a entry node succeed and is authenticated via TLS,
we can then use the connection. In this case, we should cancel all other
connection timers and in-progress connections. Cancelling the timers is
necessary so we don't attempt new unnecessary connections when our
existing connection is successful, preventing denial-of-service risks.
However, if we fail all available and allowed connections, we should tell
the rest of Tor that the connection has failed. This is so we can attempt
another entry node.
3.3. Connection Attempt Delays
As mentioned in [TEOR-P306-REP], initially, clients should prefer IPv4
by default. The Connection Attempt Delay, or delay between IPv4 and IPv6
connections should be 250 msec. This is to avoid the overhead from tunneled
IPv6 connections.
The Connection Attempt Delay should not be dynamically adjusted, as it adds
privacy risks. This value should be fixed, and could be manually adjusted
using this torrc option or consensus parameter:
* ConnectionAttemptDelay N [msec|second]
The Minimum and Maximum Connection Attempt Delays should also not be
dynamically adjusted for privacy reasons. The Minimum should be fixed at
10 msec as per RFC 8305. But the maximum should be higher than the RFC 8305
recommendation of 2 seconds. For Tor, we should make this timeout value 30
seconds to match Tor's existing timeout.
We need to make it possible for users to set the Maximum Connection Attempt
Delay value higher for slower and higher-latency networks such as dial-up
and satellite.
4. Option Changes
As we enable IPv6-enabled clients to connect out of the box, we should
adjust the default options to enable IPv6 while not breaking IPv4-only
clients.
The new default options should be:
* ClientUseIPv4 1 (to enable IPv4)
* ClientUseIPv6 1 (to enable IPv6)
* ClientPreferIPv6ORPort 0 (for load-balancing reasons so we don't
overload IPv6-only guards)
* ConnectionAttemptDelay 250 msec (the recommended delay between IPv4
and IPv6, as per RFC 8305)
One thing to note is that clients should be able to connect with the above
options on IPv4-only, dual-stack, and IPv6-only networks, and they should
also work if ClientPreferIPv6ORPort is 1. But we shouldn't expect
IPv4 or IPv6 to work if ClientUseIPv4 or ClientUseIPv6 is set to 0.
When the majority of clients and relay are IPv6-capable, we could set the
default value of ClientPreferIPv6ORPort to 1, in order to take advantage
of IPv6. We could add a ClientPreferIPv6ORPort consensus parameter, so we
can make this change network-wide.
5. Relay Statistics
Entry nodes could measure the following statistics for both IPv4 and IPv6:
* Number of successful connections
* Number of extra Prop306 connections (unsuccessful or cancelled)
* Client closes the connection before completing TLS
* Client closes the connection before sending any circuit or data cells
* Number of client and relay connections
* We can distinguish between authenticated (relay, authority
reachability) and unauthenticated (client, bridge) connections
Should we implement Section 5:
* We can send this information to the directory authorities using relay
extra-info descriptors
* We should consider the privacy implications of these statistics, and
how much noise we need to add to them
* We can include these statistics in the Heartbeat logs
6. Initial Feasibility Testing
We should test this proposal with the following scenarios:
* Different combinations of values for the options ClientUseIPv4,
ClientUseIPv6, and ClientPreferIPv6ORPort on IPv4-only, IPv6-only,
and dual-stack connections
* Dual-stack connections of different technologies, including
high-bandwidth and low-latency (e.g. FTTH), moderate-bandwidth and
moderate-latency (e.g. DSL, LTE), and high-latency and low-bandwidth
(e.g. satellite, dial-up) to see if Prop306 is reliable and feasible
7. Minimum Viable Prop306 Product
The mimumum viable product for Prop306 must include the following:
* The address handling, bootstrap, and entry guard changes described in
Section 2. (Single Onion Services are optional, Bridge Clients are out
of scope. The consensus parameter and torrc options are optional.)
* The alternative address retry algorithm in Section 3.1.
* The Connection Success/Failure mechanism in Section 3.2.
* The Connection Delay mechanism in Section 3.3. (The
ConnectionAttemptDelay torrc option and consensus parameter are
optional.)
* A default setup capable of both IPv4 and IPv6 connections with the
options described in Section 4. (The ClientPreferIPv6ORPort consensus
parameter is optional.)
8. Optional Features
Some features which are optional include:
* Single Onion services: extend info address changes for onion service
descriptors and introduce cells. (Section 2.1.)
* Bridge clients are out of scope: they would require bridge line format
changes, internal bridge data structure changes, and extend info address
changes. (Section 2.1.)
* MaxNumIPv4BootstrapAttempts torrc option. We may need this option if
the proposed default doesn't work for some clients. (Section 2.2.)
* NumIPv6Guards torrc option and consensus parameter. We may need this
option if the proposed default doesn't work for some clients.
(Section 2.3.)
* ConnectionAttemptDelay torrc option and consensus parameter. We will need
this option if the Connection Attempt Delay needs to be manually
adjusted, for instance, if clients often fail IPv6 connections.
(Section 3.3.)
* ClientPreferIPv6ORPort consensus parameter. (Section 4.)
* IPv4, IPv6, client, relay, and extra Prop306 connection statistics.
While optional, these statistics may be useful for debugging and
reliability testing, and metrics on IPv4 vs IPv6. (Section 5.)
9. Acknowledgments
Thank you so much to teor for your discussion on this happy eyeballs
proposal. I wouldn't have been able to do this has it not been for
your help.
10. Refrences
[P299-TRAC]: https://trac.torproject.org/projects/tor/ticket/29801
[TEOR-P306-REP]: https://lists.torproject.org/pipermail/tor-dev/2019-July/013919.html