As libnekit is still in its early alpha under heavy development, information in this page may be out of date.
libnekit is a cross-platform version of NEKit written in C to make building applications which route or analyze network traffic quick and simple. libnekit is built with security, efficiency and flexibility in mind.
Every connection is abstracted as a data flow chain (actually, two inter-connected chains) where data can be transformed at every hop. Data flow chain can be composed arbitrarily to provide flexibility for any scenario.
For example, you can put a HTTP debug data flow between some data flows where the data is plain text HTTP request and response. You then analyze the data and pass the data to the next hop intact, or rewrite the header adding DNT: 1
to every request and hope the websites behave, or whatever you want.
Accompanying with build-in data flows for common protocols and other tools and utilities for encryption, geoip support, HTTP header rewriting, etc., you can quickly create an application manipulating network traffic in a few lines.
Since libnekit is still in its early stage, many things tend to change. There will be minor API changes (even incompatible) in the coming period. But the architecture of it should remain unchanged.
Things supported now:
- Data flow chain to transform and analyze data with composition
- Efficient buffer implementation
- Define data handling logic by pre-defined or customized rules.
- Highly interfaced with dependency injection. You can provide you own implementation for basically everything including DNS resolver, data flow hop, rule, connection acceptor, etc.
- HTTP, SOCKS5 support (as client and server)
- Wrapper for many encryption and authentication algorithms for secured transportation
- HTTP stream rewriter for parsing and rewrite HTTP header as data stream in place
Things will be added or changed, ordered by priority:
- TLS support (as client)
- Minor API refinement
- Build for Windows
- Better error handling
- More detailed log
- Class level doc
- Use Conan as dependency management system
- Build for Android
- TLS support (as server, MITM)
- UDP support
- Python/Lua binding (not sure which one yet, please advise)
- Raw DNS request/response parsing
- IP stack support
Though libnekit itself has no platform dependent code, you still have to make sure its dependencies build on the target platform.
libnekit depends on Boost, OpenSSL, libsodium and maxminddb.
Only Boost is required. But currently the CMake script requires all dependencies present.
libnekit provides a script to build all the dependencies automatically and iOS(ios
), OSX(mac
) and linux x86/x64(linux
) are supported as of now. Windows, Android and linux ARM will also be supported later.
Build the dependencies requires Python 3 and non-antique CMake, install pipenv first then (choose the correct PLATFORM
):
pipenv run scripts/build_deps.py PLATFORM
The script will download all dependencies source codes and compile them.
It's very likely you don't have to compile libnekit independently, check next subsection.
After all dependencies are successfully built, use CMake to configure and build the project (choose correct platform):
mkdir build && cd build
cmake .. -DCMAKE_TOOLCHAIN_FILE=~/Projects/libnekit/cmake/toolchain/PLATFORM.cmake -DPLATFORM=PLATFORM
cmake --build . -- -j2
libnekit is built but not installed yet. If you want to distribute it, you need to copy all header files from include/
, dependencies from deps/PLATFORM/
and libnekit.a
to the proper place.
I may add an install
target later. But since libnekit requires Boost, the distribution would be too large.
If you project is built with CMake, then it should be very simple and straight forward, just add the root folder of libnekit to your project and configure it with the proper toolchain file. If you use your own toolchain config file, it's very likely libnekit will just compile with it so you don't need to change anything.
If you are not using CMake, then use CMake generator to generate a compatible project with the toolchain you are working with.
The idea is simple.
Everything begins with an implementation of ListenerInterface
where LocalDataInterface
is created when something happens, such as a new connection is accepted from a listening socket. But it does not have to be a typical TCP listener/acceptor, think of it as an event subscriber which reacts with generating LocalDataInterface
.
The LocalDataInterface
does whatever is needed to get enough meta information about the connection then ask a RuleManager
how to handle it. The RuleManager
will return a RemoteDataInterface
which will be responsible to process data for LocalDataInterface
. A Tunnel
is created to manage the whole procedure with this data flow pair.
Instance
is a run loop on a single thread. That's it. If that sounds confusing, think about how many threads you want to deal with. If you don't know, then you need just one thread which means you just create one instance of Instance
and stick with it. Otherwise, you need to create one instance for each thread.
Theoretically, it is possible to run one instance in several threads simultaneously, but that's too error-prone. It would be hard to understand every subtleties in libnekit in order to get the class involved thread-safe. So just follow this simple principle here, one thread per Instance
.
An Instance
can have many ProxyManager
s, where each ProxyManager
has a set of instances ListenerInterface
implementations accepting new connections and a RuleManager
decides how to handle these connections.
Each ProxyManager
has one and only one DNS resolver.
This seems like a unnecessary limit (and it is), but libnekit tries to be as efficient and secure as possible so there will be no resolution unless required.
This cause DNS resolution may happen in too many possible places that we have to provide a resolver for everything that may want to resolve the domain.
In order to avoid that, resolution should only happen by using an Endpoint
, which requires us to set a resolver for the Endpoint
if we want to resolve it. libnekit will do it for you whenever possible automatically, but it won't be able to know which resolver to use if there are multiple choices.
So the final rule is, ProxyManager
has only one resolver and everything inside it will use this resolver.
libnekit provides several rules and you can easily create new ones based on the property of the connection. Each rule instance is associated with a lambda that returns a RemoteDataFlowInterface*
that handles the data from local data flow. How you chain the remote data flows together is unlimited as long as it makes sense.
Now, with everything at hand, let's take a look at a real example.
#include <boost/asio.hpp>
#include <boost/log/common.hpp>
#include <boost/log/core.hpp>
#include <boost/log/expressions.hpp>
#include <boost/log/utility/setup/console.hpp>
#include <nekit/data_flow/http_data_flow.h>
#include <nekit/data_flow/http_server_data_flow.h>
#include <nekit/data_flow/socks5_data_flow.h>
#include <nekit/data_flow/socks5_server_data_flow.h>
#include <nekit/data_flow/speed_data_flow.h>
#include <nekit/init.h>
#include <nekit/instance.h>
#include <nekit/rule/all_rule.h>
#include <nekit/rule/dns_fail_rule.h>
#include <nekit/rule/geo_rule.h>
#include <nekit/transport/tcp_listener.h>
#include <nekit/transport/tcp_socket.h>
#include <nekit/utils/logger.h>
#include <nekit/utils/system_resolver.h>
using namespace nekit;
using namespace boost::asio;
// Setting up log, refer to Boost.Log doc.
namespace expr = boost::log::expressions;
BOOST_LOG_ATTRIBUTE_KEYWORD(severity, "Severity", utils::LogLevel);
BOOST_LOG_ATTRIBUTE_KEYWORD(instance_name, "Instance", std::string);
BOOST_LOG_ATTRIBUTE_KEYWORD(channel, "Channel", std::string);
int main() {
nekit::Initailize();
assert(utils::Maxmind::Initalize("test/GeoLite2-Country.mmdb"));
// Setting up log, refer to Boost.Log doc.
boost::log::add_console_log(
std::cout,
boost::log::keywords::filter = severity >= utils::LogLevel::Info,
boost::log::keywords::format =
(expr::stream
<< std::left << std::setw(7) << std::setfill(' ') << severity
<< expr::if_(expr::has_attr(
instance_name))[expr::stream << "[" << instance_name << "]"]
<< "[" << channel << "]" << expr::smessage));
// One instance is one run loop. You can create multiple instances if you want
// to use multi threads.
Instance instance{"Specht"};
// Proxy manager manages all the local proxies.
auto proxy_manager = std::make_unique<ProxyManager>();
// Rule manager handles rule matching.
auto rule_manager = std::make_unique<rule::RuleManager>(instance.io());
// Unless you understand everything, make sure there is only one resolver used
// for one instance.
auto resolver = std::make_unique<utils::SystemResolver>(instance.io(), 5);
// This is the lambda to be called to create a data flow chain connects to
// remote http proxy.
rule::RuleHandler http_proxy_handler =
[resolver(resolver.get())](std::shared_ptr<utils::Session> session) {
auto endpoint = std::make_shared<utils::Endpoint>("1.2.3.4", 9090);
endpoint->set_resolver(resolver);
// HttpDataFlow (1.2.3.4:9090) -> TcpSocket
return std::make_unique<data_flow::HttpDataFlow>(
endpoint, session, std::make_unique<transport::TcpSocket>(session),
nullptr);
};
// This is the lambda to be called to create a data flow chain automatically
// selects the fastest data flow becomes ready.
rule::RuleHandler speed_handler =
[resolver(resolver.get())](std::shared_ptr<utils::Session> session) {
auto endpoint = std::make_shared<utils::Endpoint>("1.2.3.4", 9091);
endpoint->set_resolver(resolver);
std::vector<
std::pair<std::unique_ptr<data_flow::RemoteDataFlowInterface>, int>>
data_flows;
data_flows.emplace_back(
// Socks5DataFlowv(1.2.3.4:9091) -> TcpSocket, delay 100
// milliseconds before connecting
std::make_pair(std::make_unique<data_flow::Socks5DataFlow>(
endpoint, session,
std::make_unique<transport::TcpSocket>(session)),
100));
data_flows.emplace_back(
// Start connecting directly with no delay
std::make_pair(std::make_unique<transport::TcpSocket>(session), 0));
return std::make_unique<data_flow::SpeedDataFlow>(
session, std::move(data_flows));
};
// This is the lambda to be called to connect directly.
rule::RuleHandler direct_handler =
[](std::shared_ptr<utils::Session> session) {
return std::make_unique<transport::TcpSocket>(session);
};
// If target host is in China, connect directly by using direct_handler.
std::shared_ptr<rule::RuleInterface> geo_cn_rule =
std::make_shared<rule::GeoRule>(utils::CountryIsoCode::CN, true,
direct_handler);
// If resolution fails, connect through proxy.
auto dns_fail_rule = std::make_shared<rule::DnsFailRule>(http_proxy_handler);
// Choose which ways is fast automatically.
auto all_rule = std::make_shared<rule::AllRule>(speed_handler);
// The rule will be matched in order.
rule_manager->AppendRule(geo_cn_rule);
rule_manager->AppendRule(dns_fail_rule);
rule_manager->AppendRule(all_rule);
proxy_manager->SetRuleManager(std::move(rule_manager));
proxy_manager->SetResolver(std::move(resolver));
// When a new connection comes in locally, we treat it as a connection to a
// SOCSK5 server.
auto socks5_listener = std::make_unique<transport::TcpListener>(
instance.io(), [](auto data_flow) {
return std::make_unique<data_flow::Socks5ServerDataFlow>(
std::move(data_flow), data_flow->Session());
});
// When a new connection comes in locally, we treat it as a connection to a
// http server.
auto http_listener = std::make_unique<transport::TcpListener>(
instance.io(), [](auto data_flow) {
return std::make_unique<data_flow::HttpServerDataFlow>(
std::move(data_flow), data_flow->Session());
});
socks5_listener->Bind("127.0.0.1", 8081);
http_listener->Bind("127.0.0.1", 8080);
proxy_manager->AddListener(std::move(socks5_listener));
proxy_manager->AddListener(std::move(http_listener));
// We can add as many proxy managers as we want, each listens on different
// ports with different rules (note rules are created with `std::shared_ptr`,
// they can be shared among different rule manages as along as they all
// belongs to the same instance.
instance.AddProxyManager(std::move(proxy_manager));
boost::asio::signal_set signals(*instance.io(), SIGINT, SIGTERM);
signals.async_wait([&instance](const boost::system::error_code& ec,
int signal_num) { instance.Stop(); });
instance.Run();
return 0;
}
Data flow is the core of libnekit. Let's take a look at how data flow is handled in libnekit first.
Below is the outline of DataFlowInterface
.
using DataEventHandler = std::function<void(utils::Buffer&&, utils::Error)>;
using EventHandler = std::function<void(utils::Error)>;
virtual utils::Cancelable Read(utils::Buffer&&, DataEventHandler) = 0;
virtual utils::Cancelable Write(utils::Buffer&&, EventHandler) = 0;
virtual utils::Cancelable CloseWrite(EventHandler) = 0;
virtual const FlowStateMachine& StateMachine() const = 0;
virtual DataFlowInterface* NextHop() const = 0;
The thing we want to archive with DataFlowInterface
is straight forward. We can read from and write to data flow. Each data flow has a FlowStateMachine
which encapsulates the state of the flow. The data flow is chained as NextHop()
will get us the next hop.
So to create a data flow, let's say, a data flow that communicates with a remote HTTP server, we can have a HTTP proxy data flow which has a TCP socket data flow as the next hop. When we write to the HTTP proxy data flow, it delegates the write request to the underlying transmission TCP socket.
This brings great flexibility and allows you to basically do anything you want with the data without interfering with anythings else.
Let's get back to the interface definition and take a closer look at the seemingly daunting definition of read method.
virtual utils::Cancelable Read(utils::Buffer&&, DataEventHandler) = 0;
We pass in a buffer and a handler, since almost everything in libnekit is asynchronous, we shouldn't expect the method will block until it really gets something. Instead, we give an event handler which will be called when there is some data read into the buffer. The DataEventHandler
which is defined as std::function<void(utils::Buffer&&, utils::Error)>
is simply a lambda which takes a buffer filled with data and an error code. We should check the if there is any error before we process the data.
The only thing left unclear is utils::Cancelable
, which is a little tricky.
You should get hold of how those libraries implementing event loop works first. It involves a lot of subtleties to implement them correctly but the principle is actually not very complex. Fortunately, you don't have to write one yourself so you don't have to worry about those subtleties.
The doc of Boost Asio provides a clear overview of how it works (which happens to be what libnekit is based on) and everyone should read it first before working with libnekit.
Now, as you have read and understood how asio works, let's have a quick recap and think about how the handler we provide is called.
It's very tempting to write a data flow that handle data encryption (with a FPE algorithm which keeps and ciphertext and plaintext and same length, like most block cipher) as this (only the read part):
class SecuredDataFlow {
public:
void Read(std::function<void(const boost::system::error_code &, std::size_t)>
handler) {
socket_.async_read_some(read_buffer_,
[this, handler](const boost::system::error_code &ec,
std::size_t bytes_transferred) {
if (ec) {
handler(ec, bytes_transferred);
return;
}
Decrypt(bytes_transferred);
handler(ec, bytes_transferred);
return;
});
}
void *ReadBuffer() { return read_buffer_; }
private:
void Decrypt(size_t len);
boost::asio::ip::tcp::socket socket_;
uint8_t read_buffer_[1024];
}
Everything seems fine, except for one, what if we decide we no longer need the socket? For example, this data flow is reading data from remote but our local application encountered an error when processing the stream and can't not proceed. It makes little sense to keep the socket (data flow) available, but can we just release it immediately?
As per the doc of Asio, when the socket is destructed it will "cancel any outstanding asynchronous operations associated with the socket", we should be able to just do it, and let boost cancel the pending read request if there is one.
Think. When the kernel notifies the io_context
that there is data available to read (on *nix) or the buffer is filled with data (on Windows), asio will put the callback handler we provide to a Completion Event Queue. The logic of observing event and executing callback are decoupled by utilizing this queue. There are several things that are probmatic:
- As we say event observation and callback execution is decoupled, we are actually saying Asio provides a very simple model for a multi thread application. Several thread can fetch task from the queue and execute them concurrently as long as the queue is thread-safe and the handler does not share anything not thread-safe. However,
uint8_t read_buffer_[1024]
is obviously not thread-safe, as kernel is filling data into the buffer, another thread may be trying to release theSecuredDataFlow
! (Cancellation in proactor model is really a little too complicate so we should try to avoid thinking about it.) - Another thing is a little more of an implementation choice: when a handler is pushed into the queue, there is no way to cancel it in Asio. If handler is pushed into the queue, then
SecuredDataFlow
is released, the handler won't know anything about it. When an execution thread fetch the handler from the queue, and call it, it can't dereferencethis
.
This tells us two things. First, the buffer's lifetime should exceeds the handler. Second, the handler needs to know the validity of this
and any other pointers it captures.
Usually, this is done by using std::shared_ptr
and std::weak_ptr
, the handler capture the shared pointer so the lifetime of the data flow exceeds the handler. It works, apart from the hassle that you have to inherit from std::enable_shared_from_this
and enforce that this class can only be managed with shared pointer, even if you never share it semantically. Also, working with std::weak_ptr
is really too cumbersome that you probably want to avoid it as much as possible.
Let's take a look at what is in the handler, usually, the handler capture the logic to proceed, which means it should also capture some shared (or weak) pointer of some data flows and other instances, as long as the handler exists, all of the instances will exist (or the memory will not be released, if you are using weak pointer, since you surely will use make_shared
), albeit not needed.
Apart from that, if you think carefully, this pattern actually creates a self-contained loop if you use shared pointer as in most Asio demo, where the handler keeps everything existing, and since the handler will proceed by creating more handler, this loop will never stop unless we tell every instances that they have to stop executing the logic in the handler since it is cancelled. The handler needs to check some state flag or all the weak pointers first to make sure it can proceed, which is cumbersome.
In order to avoid all of those issues, libnekit uses a very simple class holding a shared flag indicating whether this handler should be called or not.
The core of utils::Cancelable
is just a std::shared_ptr<bool>
.
class Cancelable {
public:
void Cancel();
...
bool
canceled() const;
private:
std::shared_ptr<bool> canceled_;
};
If we want to cancel the handler, we just call the Cancel()
method on the returned utils::Cancelable
. The data flow will first check if the captured utils::Cancelable
is cancelled or not.
Usually, the destructor of a data flow will explicitly cancel all the utils::Cancelable
returned by the next data flow hop, so the handler won't be called.
You can find more examples in tcp_socket.cc
.
Suppose we pass data by a plain memory block between the data flow chain, and some flow wants to encrypt the data with some authentication header in the front, it has to copy the whole block into an other new block or memmove
it. Consider, however, that iovec
is widely supported on nowadays OS, we can actually put the header in a standalone buffer and send the two buffers to the kernel in one call.
utils::Buffer
is created for this scenario (and some others). Internally, Buffer
is a linked list of Buf
s which is just a memory block. You can copy the data in and out as if the Buffer
is a monolithic block. You can insert and remove space from any offset and Buffer
will try to make the necessary copy minimal.
Tunnel
(just as what it does in NEKit) forwards data back and forth between local and remote. The data Tunnel
gets should be the raw data that you expect to transmit.
Tunnel
begins with an instance of LocalDataFlowInterface
where a newly accepted local (concrete or abstract) connection, it calls the Open()
method where the local data flow would fill in the necessary connect information into the utils::Session
bond to this connection.
Then the Tunnel
calls RuleManager
to get a RemoteDataFlowInterface
to handle the data. Tunnel
will call the Connect()
of the remote data flow to let it get ready for forwarding data, then call Continue()
on local data flow to notify it the whole connection tunnel is successfully established.
Then the Tunnel
forwards data back and forth between "local" and "remote" until both sides are closed or an error is encountered.
ProxyManager
will handle all of this for you so you don't have to write anything.
In order to know what to do when an event is triggered, we need to know the state of the data flow.
If you are familiar with the state transition of TCP, states of DataFlow
should be a piece of cake.
If you are not, you should at least get a general idea of it. But don't worry, DataFlow
uses FlowStateMachine
to handle states for you. All you have to do is tell the FlowStateMachine
what you want to do next. Internally, FlowStateMachine
uses assert
to make sure all the operations you are about to do are valid, so you will know if you have done anything right when debugging.
Below is the state diagram of data flows in libnekit:
┌────────────┐
│ │
┌──────────│ Init │
│ │ │
│ └────────────┘
│ │
│ │ Connect (remote)
│ │ Open (local)
│ ▼
│ ┌────────────┐
│ │ │
├──────────│Establishing│
│ │ │
│ └────────────┘
│ │
│ │ Connected (remote)
│ │ Continue Finished (local)
│ │
│ ▼
│ ┌────────────┐
│ │ │ Read
├──────────│Established │ Write
│ │ │
│ └────────────┘
│ │
│ │
│ │ Read EOF
│ │ CloseWrite
│ │
│ │
│ ▼
│ ┌────────────┐
│ │ │
├──────────│ Closing │
│ │ │
│ └────────────┘
│ │
│ │ Read
│ │ closed
│ │ &&
│ │ write
│ │ closed
│ │
▼ ▼
┌────────────┐ ┌────────────┐
│ │ │ │
│ Errored │ │ Closed │
│ │ │ │
└────────────┘ └────────────┘
It's a very simple DAG with single possible transition sequence (putting aside Errored
).
Any flow begins with Init
, and get into the Establishing
state when it is getting ready for forwarding data. When in Established
state, you can read or write data as long as there is not already a read or write request hasn't finished yet. When any one of the sides (read or write) is closed, or requested to be closed, the data flow transits into Closing
state and ends with Closed
state when both read and write are closed. Any error would put the data flow into Errored
state.
There are more helper methods provided by FlowStateMachine
to help you decide the readability, writability and other properties of the data flow. But most of the time you don't need to use them. Check the implementation of build-in data flow for examples.
The rule of thumb is: asynchronous methods don't throw, while synchronous ones may. (As of now, no method throws yet. But you should expect the error will be thrown instead of passed by method call.)
Throw exception from a callback handler does not make too much sense since the whole call stack is already gone. We can do very little about the error other than ignore it or terminate the application. However, I find throw exception significantly simplify the method signature though I don't like that C does not enforce labeling the "throwness" of method.
libnekit will use exception whenever possible. All exceptions from libnekit should be a subclass of utils::Error
.
MIT.