Build: use scoped credentials for interacting with S3

docs/dev/aws-temporary-credentials.rst

@@ -0,0 +1,30 @@
+AWS temporary credentials
+=========================
+
+Builders run arbitrary commands provided by the user, while we run the commands in a sandboxed environment (docker),
+that shouln't be the only line of defense, as we still interact with the files generated by the user outside docker for some operations.
+
+This is why instead of using credentials that have access to all the resources in AWS,
+we are using credentials that are generated by the `AWS STS service <https://docs.aws.amazon.com/STS/latest/APIReference/welcome.html>`__,
+which are temporary and scoped to the resources that are needed for the build.
+
+Local development
+-----------------
+
+In order to make use of STS, you need:
+
+- Create a role in IAM with a trusted entity type set to the AWS account that is going to be used to generate the temporary credentials.
+- Create an inline policy for the role, the policy should allow access to all S3 buckets and paths that are going to be used.
+
+You can use :ref:`environment variables <settings:AWS configuration>` to set the credentials for AWS, make sure to set the value of ``RTD_S3_PROVIDER`` to ``AWS``.
+
+.. note::
+
+   If you are part of the development team, you should be able to use the credentials from the ``storage-dev``` user,
+   which is already configured to make use of STS, and the ARN from the ``RTDSTSAssumeRoleDev`` role.
+
+.. note::
+
+   You should use AWS only when you are testing the AWS integration,
+   use the default minio provider for local development.
+   Otherwise, files may be overridden if multiple developers are using the same credentials.

docs/dev/index.rst

@@ -28,6 +28,7 @@ or taking the open source Read the Docs codebase for your own custom installatio
    migrations
    server-side-search
    search-integration
+   aws-temporary-credentials
    subscriptions
    settings
    tests

docs/dev/settings.rst

@@ -167,6 +167,22 @@ providers using the following environment variables:
 .. envvar:: RTD_SOCIALACCOUNT_PROVIDERS_GOOGLE_CLIENT_ID
 .. envvar:: RTD_SOCIALACCOUNT_PROVIDERS_GOOGLE_SECRET
 
+AWS configuration
+~~~~~~~~~~~~~~~~~
+
+The following variables can be used to use AWS in your local environment.
+Useful for testing :doc:`temporary credentials </aws-temporary-credentials>`.
+
+.. envvar:: RTD_S3_PROVIDER
+.. envvar:: RTD_AWS_ACCESS_KEY_ID
+.. envvar:: RTD_AWS_SECRET_ACCESS_KEY
+.. envvar:: RTD_AWS_STS_ASSUME_ROLE_ARN
+.. envvar:: RTD_S3_MEDIA_STORAGE_BUCKET
+.. envvar:: RTD_S3_BUILD_COMMANDS_STORAGE_BUCKET
+.. envvar:: RTD_S3_BUILD_TOOLS_STORAGE_BUCKET
+.. envvar:: RTD_S3_STATIC_STORAGE_BUCKET
+.. envvar:: RTD_AWS_S3_REGION_NAME
+
 Stripe secrets
 ~~~~~~~~~~~~~~
 

readthedocs/api/v2/views/model_views.py

@@ -1,6 +1,7 @@
 """Endpoints for listing Projects, Versions, Builds, etc."""
 
 import json
+from dataclasses import asdict
 
 import structlog
 from allauth.socialaccount.models import SocialAccount
@@ -27,6 +28,9 @@
 from readthedocs.api.v2.permissions import IsOwner
 from readthedocs.api.v2.permissions import ReadOnlyPermission
 from readthedocs.api.v2.utils import normalize_build_command
+from readthedocs.aws.security_token_service import AWSTemporaryCredentialsError
+from readthedocs.aws.security_token_service import get_s3_build_media_scoped_credentials
+from readthedocs.aws.security_token_service import get_s3_build_tools_scoped_credentials
 from readthedocs.builds.constants import INTERNAL
 from readthedocs.builds.models import Build
 from readthedocs.builds.models import BuildCommandResult
@@ -345,6 +349,44 @@ def reset(self, request, **kwargs):
     def get_queryset_for_api_key(self, api_key):
         return self.model.objects.filter(project=api_key.project)
 
+    @decorators.action(
+        detail=True,
+        permission_classes=[HasBuildAPIKey],
+        methods=["post"],
+        url_path="temporary-credentials/storage",
+    )
+    def temporary_credentials_for_storage(self, request, **kwargs):
+        """
+        Generate temporary credentials for interacting with storage.
+
+        This can generate temporary credentials for interacting with S3 only for now.
+        """
+        build = self.get_object()
+        credentials_type = request.data.get("type")
+
+        if credentials_type == "build_media":
+            method = get_s3_build_media_scoped_credentials
+            # 30 minutes should be enough for uploading build artifacts.
+            duration = 30 * 60
+        elif credentials_type == "build_tools":
+            method = get_s3_build_tools_scoped_credentials
+            # 30 minutes should be enough for downloading build tools.
+            duration = 30 * 60
+        else:
+            return Response(
+                {"error": "Invalid storage type"},
+                status=status.HTTP_400_BAD_REQUEST,
+            )
+
+        try:
+            credentials = method(build=build, duration=duration)
+        except AWSTemporaryCredentialsError:
+            return Response(
+                {"error": "Failed to generate temporary credentials"},
+                status=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            )
+        return Response({"s3": asdict(credentials)})
+
 
 class BuildCommandViewSet(DisableListEndpoint, CreateModelMixin, UserSelectViewSet):
     parser_classes = [JSONParser, MultiPartParser]

readthedocs/aws/security_token_service.py

@@ -0,0 +1,233 @@
+"""
+Module to interact with AWS STS (Security Token Service) to assume a role and get temporary scoped credentials.
+
+This is mainly used to generate temporary credentials to interact with S3 buckets from the builders.
+
+In order to make use of STS, we need:
+
+- Create a role in IAM with a trusted entity type set to the AWS account that is going to be used to generate the temporary credentials.
+- A policy that allows access to all S3 buckets and paths that are going to be used.
+  Which should be attached to the role.
+- The permissions of the temporary credentials are the result of the intersection of the role policy and the inline policy that is passed to the AssumeRole API.
+  This means that the inline policy can be used to limit the permissions of the temporary credentials, but not to expand them.
+
+See:
+
+- https://docs.aws.amazon.com/STS/latest/APIReference/API_AssumeRole.html
+- https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_sts-comparison.html
+- https://docs.aws.amazon.com/IAM/latest/UserGuide/id_credentials_temp_control-access_assumerole.html
+- https://docs.readthedocs.com/dev/latest/aws-temporary-credentials.html
+"""
+
+import json
+from dataclasses import dataclass
+
+import boto3
+import structlog
+from django.conf import settings
+
+
+log = structlog.get_logger(__name__)
+
+
+class AWSTemporaryCredentialsError(Exception):
+    """Exception raised when there is an error getting AWS S3 credentials."""
+
+
+@dataclass
+class AWSTemporaryCredentials:
+    """Dataclass to hold AWS temporary credentials."""
+
+    access_key_id: str
+    secret_access_key: str
+    session_token: str | None
+
+
+@dataclass
+class AWSS3TemporaryCredentials(AWSTemporaryCredentials):
+    """Subclass of AWSTemporaryCredentials to include S3 specific fields."""
+
+    bucket_name: str
+    region_name: str
+
+
+def get_sts_client():
+    return boto3.client(
+        "sts",
+        aws_access_key_id=settings.AWS_ACCESS_KEY_ID,
+        aws_secret_access_key=settings.AWS_SECRET_ACCESS_KEY,
+        # TODO: should this be its own setting?
+        region_name=settings.AWS_S3_REGION_NAME,
+    )
+
+
+def _get_scoped_credentials(*, session_name, policy, duration) -> AWSTemporaryCredentials:
+    """
+    :param session_name: An identifier to attach to the generated credentials, useful to identify who requested them.
+     AWS limits the session name to 64 characters, so if the session_name is too long, it will be truncated.
+    :duration: The duration of the credentials in seconds. Default is 15 minutes.
+     Note that the minimum duration time is 15 minutes and the maximum is given by the role (defaults to 1 hour).
+
+    .. note::
+
+       If USING_AWS is set to False, this function will return
+       the values of the AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY settings.
+       Useful for local development where we don't have a service like AWS STS.
+    """
+    if not settings.USING_AWS:
+        return AWSTemporaryCredentials(
+            access_key_id=settings.AWS_ACCESS_KEY_ID,
+            secret_access_key=settings.AWS_SECRET_ACCESS_KEY,
+            # A session token is not needed for the default credentials.
+            session_token=None,
+        )
+
+    # Limit to 64 characters, as per AWS limitations.
+    session_name = session_name[:64]
+    try:
+        sts_client = get_sts_client()
+        response = sts_client.assume_role(
+            RoleArn=settings.AWS_STS_ASSUME_ROLE_ARN,
+            RoleSessionName=session_name,
+            Policy=json.dumps(policy),
+            DurationSeconds=duration,
+        )
+    except Exception:
+        log.exception(
+            "Error while assuming role to generate temporary credentials",
+            session_name=session_name,
+            policy=policy,
+            duration=duration,
+        )
+        raise AWSTemporaryCredentialsError
+
+    credentials = response["Credentials"]
+    return AWSTemporaryCredentials(
+        access_key_id=credentials["AccessKeyId"],
+        secret_access_key=credentials["SecretAccessKey"],
+        session_token=credentials["SessionToken"],
+    )
+
+
+def get_s3_build_media_scoped_credentials(
+    *,
+    build,
+    duration=60 * 15,
+) -> AWSS3TemporaryCredentials:
+    """
+    Get temporary credentials with read/write access to the build media bucket.
+
+    The credentials are scoped to the paths that the build needs to access.
+
+    :duration: The duration of the credentials in seconds. Default is 15 minutes.
+     Note that the minimum duration time is 15 minutes and the maximum is given by the role (defaults to 1 hour).
+    """
+    project = build.project
+    version = build.version
+    bucket_arn = f"arn:aws:s3:::{settings.S3_MEDIA_STORAGE_BUCKET}"
+    storage_paths = version.get_storage_paths()
+    # Generate the list of allowed prefix resources
+    # The resulting prefix looks like:
+    # - html/project/latest/*
+    # - pdf/project/latest/*
+    allowed_prefixes = [f"{storage_path}/*" for storage_path in storage_paths]
+
+    # Generate the list of allowed object resources in ARN format.
+    # The resulting ARN looks like:
+    # arn:aws:s3:::readthedocs-media/html/project/latest/*
+    # arn:aws:s3:::readthedocs-media/pdf/project/latest/*
+    allowed_objects_arn = [f"{bucket_arn}/{prefix}" for prefix in allowed_prefixes]
+
+    # Inline policy document to limit the permissions of the temporary credentials.
+    policy = {
+        "Version": "2012-10-17",
+        "Statement": [
+            {
+                "Effect": "Allow",
+                "Action": [
+                    "s3:GetObject",
+                    "s3:PutObject",
+                    "s3:DeleteObject",
+                ],
+                "Resource": allowed_objects_arn,
+            },
+            # In order to list the objects in a path, we need to allow the ListBucket action.
+            # But since that action is not scoped to a path, we need to limit it using a condition.
+            {
+                "Effect": "Allow",
+                "Action": ["s3:ListBucket"],
+                "Resource": [
+                    bucket_arn,
+                ],
+                "Condition": {
+                    "StringLike": {
+                        "s3:prefix": allowed_prefixes,
+                    }
+                },
+            },
+        ],
+    }
+
+    session_name = f"rtd-{build.id}-{project.slug}-{version.slug}"
+    credentials = _get_scoped_credentials(
+        session_name=session_name,
+        policy=policy,
+        duration=duration,
+    )
+    return AWSS3TemporaryCredentials(
+        access_key_id=credentials.access_key_id,
+        secret_access_key=credentials.secret_access_key,
+        session_token=credentials.session_token,
+        region_name=settings.AWS_S3_REGION_NAME,
+        bucket_name=settings.S3_MEDIA_STORAGE_BUCKET,
+    )
+
+
+def get_s3_build_tools_scoped_credentials(
+    *,
+    build,
+    duration=60 * 15,
+) -> AWSS3TemporaryCredentials:
+    """
+    Get temporary credentials with read-only access to the build-tools bucket.
+
+    :param build: The build to get the credentials for.
+    :duration: The duration of the credentials in seconds. Default is 15 minutes.
+     Note that the minimum duration time is 15 minutes and the maximum is given by the role (defaults to 1 hour).
+    """
+    project = build.project
+    version = build.version
+    bucket = settings.S3_BUILD_TOOLS_STORAGE_BUCKET
+    bucket_arn = f"arn:aws:s3:::{bucket}"
+
+    # Inline policy to limit the permissions of the temporary credentials.
+    # The build-tools bucket is publicly readable, so we don't need to limit the permissions to a specific path.
+    policy = {
+        "Version": "2012-10-17",
+        "Statement": [
+            {
+                "Effect": "Allow",
+                "Action": [
+                    "s3:GetObject",
+                    "s3:ListBucket",
+                ],
+                "Resource": [
+                    bucket_arn,
+                    f"{bucket_arn}/*",
+                ],
+            },
+        ],
+    }
+    session_name = f"rtd-{build.id}-{project.slug}-{version.slug}"
+    credentials = _get_scoped_credentials(
+        session_name=session_name,
+        policy=policy,
+        duration=duration,
+    )
+    return AWSS3TemporaryCredentials(
+        access_key_id=credentials.access_key_id,
+        secret_access_key=credentials.secret_access_key,
+        session_token=credentials.session_token,
+        region_name=settings.AWS_S3_REGION_NAME,
+        bucket_name=bucket,
+    )