X-Forwarded-For
and other "real" client IP headers are often used incorrectly, resulting in bugs and security vulnerabilities. This library is an attempt to create a reference implementation of the correct ways to use such headers.
This library is written in Go, but the hope is that it will be reimplemented in other languages. Please open an issue if you would like to create such an implementation.
This library is freely licensed. You may use it as a dependency or copy it or modify it or anything else you want. It has no dependencies, is written in pure Go, and supports Go versions as far back as 1.13.
This library provides strategies for extracting the desired "real" client IP from various headers or from http.Request.RemoteAddr
(the client socket IP).
strategy, err := realclientip.NewRightmostTrustedCountStrategy("X-Forwarded-For", 2)
...
clientIP := strategy.ClientIP(req.Header, req.RemoteAddr)
Try it out in the playground.
There are a number of different strategies available -- the right one will depend on your network configuration. See the documentation to find out what's available and which you should use.
ClientIP
is threadsafe for all strategies. The same strategy instance can be used for handling all HTTP requests, for example.
There are examples of use in the documentation and _examples
directory.
The strategy used must be chosen and tuned for your network configuration. This should result in the strategy never returning an empty string -- i.e., never failing to find a candidate for the "real" IP. Consequently, getting an empty-string result should be treated as an application error, perhaps even worthy of panicking.
For example, if you have 2 levels of trusted reverse proxies, you would probably use RightmostTrustedCountStrategy
and it should work every time. If you're directly connected to the internet, you would probably use RemoteAddrStrategy
or something like ChainStrategy(LeftmostNonPrivateStrategy(...), RemoteAddrStrategy)
and you will be sure to get a value every time. If you're behind Cloudflare, you would probably use SingleIPHeaderStrategy("Cf-Connecting-IP")
and it should work every time.
So if an empty string is returned, it is either because the strategy choice or configuration is incorrect or your network configuration has changed. In either case, immediate remediation is required.
Leftmost-ish and rightmost-ish strategies support the X-Forwarded-For
and Forwarded
headers.
SingleIPHeaderStrategy
supports any header containing a single IP address or IP:port. For a list of some common headers, see the Single-IP Headers wiki page.
You must choose exactly the correct header for your configuration. Choosing the wrong header can result in failing to get the client IP or falling victim to IP spoofing.
Do not abuse ChainStrategy
to check multiple headers. There is likely only one header you should be checking, and checking more can leave you vulnerable to IP spoofing.
Support for the Forwarded
header should be sufficient for the vast majority of rightmost-ish uses, but it is not complete and doesn't completely adhere to RFC 7239. See the Test_forwardedHeaderRFCDeviations
test for details on deviations.
IPv6 zone identifiers are retained in the IP address returned by the strategies. Whether you should keep the zone depends on your specific use case. As a general rule, if you are not immediately using the IP address (for example, if you are appending it to the X-Forwarded-For
header and passing it on), then you should include the zone. This allows downstream consumers the option to use it. If your code is the final consumer of the IP address, then keeping the zone will depend on your specific case (for example: if you're logging the IP, then you probably want the zone; if you are rate limiting by IP, then you probably want to discard it).
To split the zone off and discard it, you may use realclientip.SplitHostZone
.
There is a copy of Cloudflare's IP ranges under ranges.Cloudflare
. This can be used with realclientip.RightmostTrustedRangeStrategy
. We may add more known cloud provider ranges in the future. Contributions are welcome to add new providers or update existing ones.
(It might be preferable to use provider APIs to retrieve the ranges, as they are guaranteed to be up-to-date.)
At the time of writing this library, Go 1.18 was only just released. It made sense to use the older net
package rather than the newer netip
, so that the required Go version wouldn't be so high as to exclude some users of the library.
In the future we may wish to switch to using netip
, but it will require API changes to AddressesAndRangesToIPNets
, RightmostTrustedRangeStrategy
, and ParseIPAddr
.
The values 0.0.0.0
(zero) and ::
(unspecified) are valid IPs, strictly speaking. However, this library treats them as invalid as they don't make sense to its intended uses. If you have a valid use case for them, please open an issue.
All IPs output by the library are first converted to a structure (like net.IP
) and then stringified. This helps normalize the cases where there are multiple ways of encoding the same IP -- like 192.0.2.1
and ::ffff:192.0.2.1
, and the various zero-collapsed states of IPv6 (fe80::1
vs fe80::0:0:0:1
, etc.).
Some input is allowed that isn't strictly correct. Some examples:
- IPv4 with brackets:
[2.2.2.2]:1234
- IPv4 with zone:
2.2.2.2%eth0
- Non-numeric port values:
2.2.2.2:nope
- Other
Forwarded
header deviations
It could be argued that it would be better to be absolutely strict in what is accepted.
As this library aspires to be a "reference implementation", the code is heavily commented. Perhaps more than is strictly necessary.
Strategies are created by calling a constructor, like NewRightmostTrustedCountStrategy("Forwarded", 2)
. That can make it awkward to create-and-call at the same time, like NewRightmostTrustedCountStrategy("Forwarded", 2).ClientIP(r.Header, r.RemoteAddr)
. We could have instead implemented non-pre-created functions, like RightmostTrustedCountStrategy("Forwarded", 2, r.Header, r.RemoteAddr)
. The reasons for the way we did it include:
- A consistent interface. This enables
ChainStrategy
. It also enables library users to have code paths that aren't strategy-dependent, in case they want the strategy to be configurable. - Pre-creation allows us to put as much of the invariant processing as possible into the creation step. (Although, in practice, so far, this is only the header name canonicalization.)
- No error return is required from the strategy
ClientIP
calls. (Although they can -- but should not -- return empty string.) All error-prone processing is done in the pre-creation.
An alternative approach could be using functions like:
func RightmostTrustedCountStrategy(headerName string, trustedRanges []*net.IPNet, headers http.Header, remoteAddr string) (Strategy, ip, error) {
...
strat, _, err := RightmostTrustedRangeStrategy("Forward", 2, "", "") // pre-create
_, ip, err := RightmostTrustedRangeStrategy("Forward", 2, r.Header, r.RemoteAddr) // use direct
But perhaps that's no less awkward.
A pre-release implementation of this library constructed functions rather than structs that implement an interface. The switch to the latter was made for a few reasons:
- It seems slightly more Go-idiomatic.
- It allows for adding new methods in the future without breaking the API. (Such as
String()
.) - It allows for configuration information to appear in a printf of a strategy struct. This can be useful for logging.
- The function approach is still easy to use, with the bound
ClientIP
method:getClientIP := NewRightmostTrustedCountStrategy("Forwarded", 2).ClientIP
If you want to reproduce this implementation in another language, please create an issue and we'll make a repo under this organization for you to use.