authn
This README outlines how we acquire and use credentials when interacting with a registry.
As much as possible, we attempt to emulate docker
's authentication behavior and configuration so that this library "just works" if you've already configured credentials that work with docker
; however, when things don't work, a basic understanding of what's going on can help with debugging.
The official documentation for how authentication with docker
works is (reasonably) scattered across several different sites and GitHub repositories, so we've tried to summarize the relevant bits here.
tl;dr for consumers of this package
By default, pkg/v1/remote
uses Anonymous
credentials (i.e. none), which for most registries will only allow read access to public images.
To use the credentials found in your Docker config file, you can use the DefaultKeychain
, e.g.:
package main
import (
"fmt"
"github.com/google/go-containerregistry/pkg/authn"
"github.com/google/go-containerregistry/pkg/name"
"github.com/google/go-containerregistry/pkg/v1/remote"
)
func main() {
ref, err := name.ParseReference("registry.example.com/private/repo")
if err != nil {
panic(err)
}
// Fetch the manifest using default credentials.
img, err := remote.Get(ref, remote.WithAuthFromKeychain(authn.DefaultKeychain))
if err != nil {
panic(err)
}
// Prints the digest of registry.example.com/private/repo
fmt.Println(img.Digest)
}
The DefaultKeychain
will use credentials as described in your Docker config file -- usually ~/.docker/config.json
, or %USERPROFILE%\.docker\config.json
on Windows -- or the location described by the DOCKER_CONFIG
environment variable, if set.
If those are not found, DefaultKeychain
will look for credentials configured using Podman's expectation that these are found in ${XDG_RUNTIME_DIR}/containers/auth.json
.
See below for more information about what is configured in this file.
Emulating Cloud Provider Credential Helpers
pkg/v1/google.Keychain
provides a Keychain
implementation that emulates docker-credential-gcr
to find credentials in the environment.
See google.NewEnvAuthenticator
and google.NewGcloudAuthenticator
for more information.
To emulate other credential helpers without requiring them to be available as executables, NewKeychainFromHelper
provides an adapter that takes a Go implementation satisfying a subset of the credentials.Helper
interface, and makes it available as a Keychain
.
This means that you can emulate, for example, Amazon ECR's docker-credential-ecr-login
credential helper using the same implementation:
import (
ecr "github.com/awslabs/amazon-ecr-credential-helper/ecr-login"
"github.com/awslabs/amazon-ecr-credential-helper/ecr-login/api"
"github.com/google/go-containerregistry/pkg/authn"
"github.com/google/go-containerregistry/pkg/v1/remote"
)
func main() {
// ...
ecrHelper := ecr.ECRHelper{ClientFactory: api.DefaultClientFactory()}
img, err := remote.Get(ref, remote.WithAuthFromKeychain(authn.NewKeychainFromHelper(ecrHelper)))
if err != nil {
panic(err)
}
// ...
}
Likewise, you can emulate Azure's ACR docker-credential-acr-env
credential helper:
import (
"github.com/chrismellard/docker-credential-acr-env/pkg/credhelper"
"github.com/google/go-containerregistry/pkg/authn"
"github.com/google/go-containerregistry/pkg/v1/remote"
)
func main() {
// ...
acrHelper := credhelper.NewACRCredentialsHelper()
img, err := remote.Get(ref, remote.WithAuthFromKeychain(authn.NewKeychainFromHelper(acrHelper)))
if err != nil {
panic(err)
}
// ...
}
Using Multiple Keychain
s
NewMultiKeychain
allows you to specify multiple Keychain
implementations, which will be checked in order when credentials are needed.
For example:
kc := authn.NewMultiKeychain(
authn.DefaultKeychain,
google.Keychain,
authn.NewFromHelper(ecr.ECRHelper{ClientFactory: api.DefaultClientFactory{}}),
authn.NewFromHelper(acr.ACRCredHelper{}),
)
This multi-keychain will:
- first check for credentials found in the Docker config file, as describe above, then
- check for GCP credentials available in the environment, as described above, then
- check for ECR credentials by emulating the ECR credential helper, then
- check for ACR credentials by emulating the ACR credential helper.
If any keychain implementation is able to provide credentials for the request, they will be used, and further keychain implementations will not be consulted.
If no implementations are able to provide credentials, Anonymous
credentials will be used.
Docker Config Auth
What follows attempts to gather useful information about Docker's config.json and make it available in one place.
If you have questions, please file an issue.
Plaintext
The config file is where your credentials are stored when you invoke docker login
, e.g. the contents may look something like this:
{
"auths": {
"registry.example.com": {
"auth": "QXp1cmVEaWFtb25kOmh1bnRlcjI="
}
}
}
The auths
map has an entry per registry, and the auth
field contains your username and password encoded as HTTP 'Basic' Auth.
NOTE: This means that your credentials are stored in plaintext:
$ echo "QXp1cmVEaWFtb25kOmh1bnRlcjI=" | base64 -d
AzureDiamond:hunter2
For what it's worth, this config file is equivalent to:
{
"auths": {
"registry.example.com": {
"username": "AzureDiamond",
"password": "hunter2"
}
}
}
... which is useful to know if e.g. your CI system provides you a registry username and password via environment variables and you want to populate this file manually without invoking docker login
.
Helpers
If you log in like this, docker
will warn you that you should use a credential helper, and you should!
To configure a global credential helper:
{
"credsStore": "osxkeychain"
}
To configure a per-registry credential helper:
{
"credHelpers": {
"gcr.io": "gcr"
}
}
We use github.com/docker/cli/cli/config.Load
to parse the config file and invoke any necessary credential helpers. This handles the logic of taking a ConfigFile
+ registry domain and producing an AuthConfig
, which determines how we authenticate to the registry.
Credential Helpers
The credential helper protocol allows you to configure a binary that supplies credentials for the registry, rather than hard-coding them in the config file.
The protocol has several verbs, but the one we most care about is get
.
For example, using the following config file:
{
"credHelpers": {
"gcr.io": "gcr",
"eu.gcr.io": "gcr"
}
}
To acquire credentials for gcr.io
, we look in the credHelpers
map to find
the credential helper for gcr.io
is gcr
. By appending that value to
docker-credential-
, we can get the name of the binary we need to use.
For this example, that's docker-credential-gcr
, which must be on our $PATH
.
We'll then invoke that binary to get credentials:
$ echo "gcr.io" | docker-credential-gcr get
{"Username":"_token","Secret":"<long access token>"}
You can configure the same credential helper for multiple registries, which is
why we need to pass the domain in via STDIN, e.g. if we were trying to access
eu.gcr.io
, we'd do this instead:
$ echo "eu.gcr.io" | docker-credential-gcr get
{"Username":"_token","Secret":"<long access token>"}
Debugging credential helpers
If a credential helper is configured but doesn't seem to be working, it can be
challenging to debug. Implementing a fake credential helper lets you poke around
to make it easier to see where the failure is happening.
This "implements" a credential helper with hard-coded values:
#!/usr/bin/env bash
echo '{"Username":"<token>","Secret":"hunter2"}'
This implements a credential helper that prints the output of
docker-credential-gcr
to both stderr and whatever called it, which allows you
to snoop on another credential helper:
#!/usr/bin/env bash
docker-credential-gcr $@ | tee >(cat 1>&2)
Put those files somewhere on your path, naming them e.g.
docker-credential-hardcoded
and docker-credential-tee
, then modify the
config file to use them:
{
"credHelpers": {
"gcr.io": "tee",
"eu.gcr.io": "hardcoded"
}
}
The docker-credential-tee
trick works with both crane
and docker
:
$ crane manifest gcr.io/google-containers/pause > /dev/null
{"ServerURL":"","Username":"_dcgcr_1_5_0_token","Secret":"<redacted>"}
$ docker pull gcr.io/google-containers/pause
Using default tag: latest
{"ServerURL":"","Username":"_dcgcr_1_5_0_token","Secret":"<redacted>"}
latest: Pulling from google-containers/pause
a3ed95caeb02: Pull complete
4964c72cd024: Pull complete
Digest: sha256:a78c2d6208eff9b672de43f880093100050983047b7b0afe0217d3656e1b0d5f
Status: Downloaded newer image for gcr.io/google-containers/pause:latest
gcr.io/google-containers/pause:latest
The Registry
There are two methods for authenticating against a registry:
token and
oauth2.
Both methods are used to acquire an opaque Bearer
token (or
RegistryToken)
to use in the Authorization
header. The registry will return a 401 Unauthorized
during the version
check
(or during normal operations) with
Www-Authenticate challenge
indicating how to proceed.
Token
If we get back an AuthConfig
containing a Username/Password
or
Auth
,
we'll use the token method for authentication:
OAuth 2
If we get back an AuthConfig
containing an IdentityToken
we'll use the oauth2 method for authentication:
This happens when a credential helper returns a response with the
Username
set to <token>
(no, that's not a placeholder, the literal string "<token>"
).
It is unclear why: moby/moby#36926.
We only support the oauth2 grant_type
for refresh_token
(#629),
since it's impossible to determine from the registry response whether we should
use oauth, and the token method for authentication is widely implemented by
registries.