Skip to main content

fakecloud_core/
dispatch.rs

1use axum::body::Body;
2use axum::extract::{ConnectInfo, Extension, Query};
3use axum::http::{Request, StatusCode};
4use axum::response::Response;
5use bytes::Bytes;
6use std::collections::HashMap;
7use std::net::SocketAddr;
8use std::sync::Arc;
9
10use crate::auth::{
11    is_root_bypass, ConditionContext, CredentialResolver, IamMode, IamPolicyEvaluator, Principal,
12    PrincipalType, ResourcePolicyProvider,
13};
14use crate::protocol::{self, AwsProtocol};
15use crate::registry::ServiceRegistry;
16use crate::service::{AwsRequest, ResponseBody};
17
18/// The main dispatch handler. All HTTP requests come through here.
19pub async fn dispatch(
20    ConnectInfo(remote_addr): ConnectInfo<SocketAddr>,
21    Extension(registry): Extension<Arc<ServiceRegistry>>,
22    Extension(config): Extension<Arc<DispatchConfig>>,
23    Query(query_params): Query<HashMap<String, String>>,
24    request: Request<Body>,
25) -> Response<Body> {
26    let remote_addr = Some(remote_addr);
27    let request_id = uuid::Uuid::new_v4().to_string();
28
29    let (parts, body) = request.into_parts();
30
31    // Streaming opt-in: if the route is a known large-body S3 / ECR
32    // upload, we skip the buffered `to_bytes` step entirely and hand
33    // the raw body to the service handler. The handler spills it to
34    // disk on the fly. Header-only detection covers every streaming
35    // candidate (none of them rely on form-body sniffing).
36    let stream_route = streaming_route(
37        &parts.method,
38        parts.uri.path(),
39        &parts.headers,
40        &query_params,
41    );
42    let header_only = protocol::detect_service_headers_only(&parts.headers, &query_params);
43    let stream_dispatch = match (&stream_route, &header_only) {
44        // Header-only detection agrees with the URL match — covers S3
45        // PUT object (SigV4 service=s3 in Authorization).
46        (Some(sr), Some(detected)) if sr.0 == detected.service => Some(detected.clone()),
47        // ECR OCI v2 blob upload has no AWS auth header; the path
48        // alone (`/v2/.../blobs/uploads/...`) tells us the route is
49        // ECR. Synthesize a DetectedRequest so dispatch picks the
50        // streaming path. Same special-case the buffered branch
51        // applies on detect_service None (see below).
52        (Some((service, _)), None) if *service == "ecr" => Some(protocol::DetectedRequest {
53            service: "ecr".to_string(),
54            action: String::new(),
55            protocol: AwsProtocol::Rest,
56        }),
57        _ => None,
58    };
59
60    let (body_bytes, body_stream) = if stream_dispatch.is_some() {
61        (Bytes::new(), Some(body))
62    } else {
63        // Buffered path: materialize the body into memory under the
64        // configured cap. `FAKECLOUD_MAX_REQUEST_BODY_BYTES` (default
65        // 1 GiB) caps non-streaming requests; streaming routes have no
66        // cap because nothing materializes the entire body in RAM.
67        let max_body_bytes = max_request_body_bytes();
68        match axum::body::to_bytes(body, max_body_bytes).await {
69            Ok(b) => (b, None),
70            Err(_) => {
71                return build_error_response(
72                    StatusCode::PAYLOAD_TOO_LARGE,
73                    "RequestEntityTooLarge",
74                    "Request body too large",
75                    &request_id,
76                    AwsProtocol::Query,
77                );
78            }
79        }
80    };
81
82    // Detect service and action
83    let detected = if let Some(d) = stream_dispatch {
84        d
85    } else {
86        match protocol::detect_service(&parts.headers, &query_params, &body_bytes) {
87            Some(d) => d,
88            None => {
89                // A request carrying X-Amz-Target is unambiguously an awsJson
90                // call whose operation we couldn't map to a known service. AWS
91                // answers these with UnknownOperationException; routing them to
92                // the apigateway catch-all below would 404 with a misleading
93                // "Stage not found" instead.
94                if let Some(target) = parts
95                    .headers
96                    .get("x-amz-target")
97                    .and_then(|v| v.to_str().ok())
98                {
99                    return build_error_response(
100                        StatusCode::BAD_REQUEST,
101                        "UnknownOperationException",
102                        &format!("The operation {target} is not recognized."),
103                        &request_id,
104                        AwsProtocol::Json,
105                    );
106                }
107                // OPTIONS requests (CORS preflight) don't carry Authorization headers.
108                // Route them to S3 since S3 is the only REST service that handles CORS.
109                // Note: API Gateway CORS preflight is not fully supported in this emulator
110                // because we can't distinguish between S3 and API Gateway OPTIONS requests
111                // without additional context (in real AWS, they have different domains).
112                if parts.method == http::Method::OPTIONS {
113                    protocol::DetectedRequest {
114                        service: "s3".to_string(),
115                        action: String::new(),
116                        protocol: AwsProtocol::Rest,
117                    }
118                } else if parts.uri.path() == "/v2" || parts.uri.path().starts_with("/v2/") {
119                    // OCI Distribution v2 protocol. Docker CLI / OCI clients
120                    // use Basic auth (not SigV4) and GET /v2/ with no body,
121                    // so this must be matched before the apigateway fallback.
122                    protocol::DetectedRequest {
123                        service: "ecr".to_string(),
124                        action: String::new(),
125                        protocol: AwsProtocol::Rest,
126                    }
127                } else if let Some(bucket) = anonymous_s3_bucket(&parts.uri, &config) {
128                    // Unsigned request whose first path segment names an
129                    // existing S3 bucket: an anonymous path-style S3 access
130                    // (e.g. serving a public-read object to a browser). Without
131                    // this it would fall through to the apigateway catch-all
132                    // below and 404 with "Stage not found" (#1707). Authorization
133                    // for the anonymous caller still runs in the IAM block.
134                    tracing::debug!(bucket = %bucket, "routing unsigned request to S3 (existing bucket)");
135                    protocol::DetectedRequest {
136                        service: "s3".to_string(),
137                        action: String::new(),
138                        protocol: AwsProtocol::Rest,
139                    }
140                } else if !parts.uri.path().starts_with("/_") {
141                    // Requests without AWS auth that don't match any service might be
142                    // API Gateway execute API calls (plain HTTP without signatures).
143                    // Route them to apigateway service which will validate if a matching
144                    // API/stage exists. Skip special FakeCloud endpoints (/_*).
145                    protocol::DetectedRequest {
146                        service: "apigateway".to_string(),
147                        action: String::new(),
148                        protocol: AwsProtocol::RestJson,
149                    }
150                } else {
151                    return build_error_response(
152                        StatusCode::BAD_REQUEST,
153                        "MissingAction",
154                        "Could not determine target service or action from request",
155                        &request_id,
156                        AwsProtocol::Query,
157                    );
158                }
159            }
160        }
161    };
162
163    // Bedrock-agent and bedrock-runtime both send `bedrock` in the SigV4
164    // credential scope, but bedrock-agent has its own service handler.
165    // Disambiguate based on the request path.
166    let detected = if detected.service == "bedrock" {
167        let first_seg = parts.uri.path().split('/').nth(1);
168        if matches!(
169            first_seg,
170            Some(
171                "agents"
172                    | "knowledgebases"
173                    | "flows"
174                    | "prompts"
175                    | "tags"
176                    | "retrieveAndGenerate"
177                    | "retrieveAndGenerateStream"
178                    | "optimize-prompt"
179                    | "sessions"
180                    | "invocations"
181                    | "generate-query"
182                    | "rerank"
183            )
184        ) {
185            // Further disambiguate runtime vs control plane for agents/flows paths
186            let segs: Vec<&str> = parts.uri.path().split('/').collect();
187            let is_runtime = matches!(
188                segs.as_slice(),
189                ["", "agents", _, "agentAliases", _, ..]  // InvokeAgent
190                    | ["", "flows", _, "aliases", _]   // InvokeFlow
191                    | ["", "knowledgebases", _, "retrieve"] // Retrieve
192                    | ["", "retrieveAndGenerate"]
193                    | ["", "retrieveAndGenerateStream"]
194                    | ["", "optimize-prompt"]
195                    | ["", "sessions", ..]
196                    | ["", "invocations", ..]
197                    | ["", "generate-query"]
198                    | ["", "rerank"]
199            );
200            if is_runtime {
201                protocol::DetectedRequest {
202                    service: "bedrock-agent-runtime".to_string(),
203                    ..detected
204                }
205            } else {
206                protocol::DetectedRequest {
207                    service: "bedrock-agent".to_string(),
208                    ..detected
209                }
210            }
211        } else {
212            detected
213        }
214    } else {
215        detected
216    };
217
218    // Amazon DocumentDB shares RDS's Query wire protocol, `rds` SigV4
219    // signing scope, and `rds.<region>.amazonaws.com` endpoint, so a real
220    // `aws-sdk-docdb` request is indistinguishable from `aws-sdk-rds` by
221    // signing name or host alone. The DocumentDB SDK does stamp an
222    // `api/docdb` token into its `user-agent`; use it to route the request
223    // to the dedicated `docdb` handler. (The conformance probe signs the
224    // `docdb` scope directly, so it never reaches this branch.) Requests
225    // without the token stay on `rds`.
226    let detected = if detected.service == "rds" && user_agent_indicates_docdb(&parts.headers) {
227        protocol::DetectedRequest {
228            service: "docdb".to_string(),
229            ..detected
230        }
231    } else {
232        detected
233    };
234
235    // Amazon Neptune shares RDS's Query wire protocol, `rds` SigV4 signing
236    // scope, and `rds.<region>.amazonaws.com` endpoint, so a real
237    // `aws-sdk-neptune` request is indistinguishable from `aws-sdk-rds` by
238    // signing name or host alone. The Neptune SDK does stamp an
239    // `api/neptune` token into its `user-agent`; use it to route the request
240    // to the dedicated `neptune` handler. (The conformance probe signs the
241    // `neptune` scope directly, so it never reaches this branch.) Requests
242    // without the token stay on `rds`.
243    let detected = if detected.service == "rds" && user_agent_indicates_neptune(&parts.headers) {
244        protocol::DetectedRequest {
245            service: "neptune".to_string(),
246            ..detected
247        }
248    } else {
249        detected
250    };
251
252    // Look up service
253    let service = match registry.get(&detected.service) {
254        Some(s) => s,
255        None => {
256            return build_error_response(
257                detected.protocol.error_status(),
258                "UnknownService",
259                &format!("Service '{}' is not available", detected.service),
260                &request_id,
261                detected.protocol,
262            );
263        }
264    };
265
266    // Extract region and access key from auth header (or presigned query).
267    let auth_header = parts
268        .headers
269        .get("authorization")
270        .and_then(|v| v.to_str().ok())
271        .unwrap_or("");
272    let header_info = fakecloud_aws::sigv4::parse_sigv4(auth_header);
273    let presigned_info = if header_info.is_none() {
274        // Presigned URL: credentials live in the query string.
275        fakecloud_aws::sigv4::parse_sigv4_presigned(&query_params).map(|p| p.as_info())
276    } else {
277        None
278    };
279    let sigv4_info = header_info.or(presigned_info);
280    // SigV2 presigned URLs (`AWSAccessKeyId` + `Signature` + `Expires` query
281    // parameters) carry the access key outside the SigV4 grammar, so the SigV4
282    // parsers above return None. Recover the key here so a SigV2-presigned
283    // request is attributed to its caller instead of being treated as
284    // anonymous (which would deny it under the object read auth gate).
285    let access_key_id = sigv4_info
286        .as_ref()
287        .map(|info| info.access_key.clone())
288        .or_else(|| sigv2_presigned_access_key(&query_params));
289
290    // Host-header routing hint: LocalStack-shaped
291    // `<svc>.<region>.localhost.localstack.cloud[:port]`, real-AWS
292    // `<svc>.<region>.amazonaws.com`, and every S3 virtual-hosted variant
293    // of both. Secondary region source and carries the bucket for
294    // virtual-hosted S3 path rewrite.
295    let host_info = protocol::parse_routing_host_from_headers(&parts.headers);
296
297    let region = sigv4_info
298        .map(|info| info.region)
299        .or_else(|| host_info.as_ref().map(|h| h.region.clone()))
300        .or_else(|| extract_region_from_user_agent(&parts.headers))
301        .unwrap_or_else(|| config.region.clone());
302
303    // Resolve the caller's principal up front so both SigV4 verification
304    // (which needs the secret) and the service handler (which needs the
305    // identity for GetCallerIdentity and IAM enforcement) share a single
306    // lookup. The root-bypass AKID skips resolution entirely — `test`
307    // credentials have no backing identity and must always pass.
308    let caller_akid = access_key_id.as_deref().unwrap_or("");
309    let resolved = if !caller_akid.is_empty() && !is_root_bypass(caller_akid) {
310        config
311            .credential_resolver
312            .as_ref()
313            .and_then(|r| r.resolve(caller_akid))
314    } else {
315        None
316    };
317    let caller_principal = resolved.as_ref().map(|r| r.principal.clone());
318    let caller_session_policies = resolved
319        .as_ref()
320        .map(|r| r.session_policies.clone())
321        .unwrap_or_default();
322
323    // Opt-in SigV4 cryptographic verification. Runs before the service
324    // handler so a failing signature never reaches business logic. The
325    // reserved `test*` root identity short-circuits verification to keep
326    // local-dev workflows frictionless.
327    //
328    // A fully anonymous request — no `Authorization` header AND no presigned
329    // credential (SigV4 `X-Amz-Credential` or SigV2 `AWSAccessKeyId`) — carries
330    // no signature to verify, so it must NOT be hard-403'd here even under
331    // `--verify-sigv4`. AWS treats an unsigned request as the anonymous
332    // principal and lets a resource policy / public-read ACL authorize it
333    // (e.g. a public S3 object GET). Let it fall through to the
334    // anonymous-authorization path below, which allows a public resource and
335    // otherwise denies. A request that DID present a credential but failed to
336    // parse is NOT anonymous (`Authorization` non-empty or a presign query
337    // present) and still returns the malformed-signature error below.
338    let is_fully_anonymous = auth_header.is_empty()
339        && !query_params.contains_key("X-Amz-Credential")
340        && sigv2_presigned_access_key(&query_params).is_none();
341    if config.verify_sigv4
342        && !is_fully_anonymous
343        && !is_root_bypass(caller_akid)
344        && config.credential_resolver.is_some()
345    {
346        let amz_date = parts
347            .headers
348            .get("x-amz-date")
349            .and_then(|v| v.to_str().ok());
350        let parsed = fakecloud_aws::sigv4::parse_sigv4_header(auth_header, amz_date)
351            .or_else(|| fakecloud_aws::sigv4::parse_sigv4_presigned(&query_params));
352        let parsed = match parsed {
353            Some(p) => p,
354            None => {
355                return build_error_response(
356                    StatusCode::FORBIDDEN,
357                    "IncompleteSignature",
358                    "Request is missing or has a malformed AWS Signature",
359                    &request_id,
360                    detected.protocol,
361                );
362            }
363        };
364        let resolved_for_verify = match resolved.as_ref() {
365            Some(r) => r,
366            None => {
367                return build_error_response(
368                    StatusCode::FORBIDDEN,
369                    "InvalidClientTokenId",
370                    "The security token included in the request is invalid",
371                    &request_id,
372                    detected.protocol,
373                );
374            }
375        };
376        let headers_vec = fakecloud_aws::sigv4::headers_from_http(&parts.headers);
377        let raw_query_for_verify = parts.uri.query().unwrap_or("").to_string();
378        let verify_req = fakecloud_aws::sigv4::VerifyRequest {
379            method: parts.method.as_str(),
380            path: parts.uri.path(),
381            query: &raw_query_for_verify,
382            headers: &headers_vec,
383            body: &body_bytes,
384        };
385        match fakecloud_aws::sigv4::verify(
386            &parsed,
387            &verify_req,
388            &resolved_for_verify.secret_access_key,
389            chrono::Utc::now(),
390        ) {
391            Ok(()) => {
392                // Bind the buffered request body to the signed
393                // `x-amz-content-sha256`. The sigv4 canonical builder uses that
394                // header value verbatim (it is a signed header) and deliberately
395                // does NOT re-hash the body: for streaming / aws-chunked routes
396                // `body_bytes` is empty at that layer, so re-hashing there would
397                // reject legitimate signed requests (see the
398                // `feedback_sigv4_body_hash_wrong_layer` caveat). S3 rebinds the
399                // body hash in its own write path (`XAmzContentSHA256Mismatch`);
400                // every other service is bound HERE, where the full buffered body
401                // is available. Only a genuine 64-char lowercase-hex digest is
402                // checked — `UNSIGNED-PAYLOAD`, `STREAMING-*`, and presigned
403                // (UNSIGNED-PAYLOAD) requests are skipped, so correct clients are
404                // unaffected. A mismatch means the body was altered after signing;
405                // AWS would then compute a different canonical request, so return
406                // `SignatureDoesNotMatch`.
407                if !parsed.is_presigned && detected.service != "s3" {
408                    if let Some(signed_hash) = parts
409                        .headers
410                        .get("x-amz-content-sha256")
411                        .and_then(|v| v.to_str().ok())
412                        .filter(|h| is_hex_sha256(h))
413                    {
414                        if sha256_hex_lower(&body_bytes) != signed_hash {
415                            return build_error_response(
416                                StatusCode::FORBIDDEN,
417                                "SignatureDoesNotMatch",
418                                "The request signature we calculated does not match the signature you provided",
419                                &request_id,
420                                detected.protocol,
421                            );
422                        }
423                    }
424                }
425            }
426            Err(fakecloud_aws::sigv4::SigV4Error::RequestTimeTooSkewed { .. }) => {
427                return build_error_response(
428                    StatusCode::FORBIDDEN,
429                    "RequestTimeTooSkewed",
430                    "The difference between the request time and the current time is too large",
431                    &request_id,
432                    detected.protocol,
433                );
434            }
435            Err(fakecloud_aws::sigv4::SigV4Error::InvalidDate(msg)) => {
436                return build_error_response(
437                    StatusCode::FORBIDDEN,
438                    "IncompleteSignature",
439                    &format!("Invalid x-amz-date: {msg}"),
440                    &request_id,
441                    detected.protocol,
442                );
443            }
444            Err(fakecloud_aws::sigv4::SigV4Error::Malformed(msg)) => {
445                return build_error_response(
446                    StatusCode::FORBIDDEN,
447                    "IncompleteSignature",
448                    &format!("Malformed SigV4 signature: {msg}"),
449                    &request_id,
450                    detected.protocol,
451                );
452            }
453            Err(fakecloud_aws::sigv4::SigV4Error::SignatureMismatch) => {
454                return build_error_response(
455                    StatusCode::FORBIDDEN,
456                    "SignatureDoesNotMatch",
457                    "The request signature we calculated does not match the signature you provided",
458                    &request_id,
459                    detected.protocol,
460                );
461            }
462            Err(fakecloud_aws::sigv4::SigV4Error::PresignedUrlExpired { .. }) => {
463                return build_error_response(
464                    StatusCode::FORBIDDEN,
465                    "AccessDenied",
466                    "Request has expired",
467                    &request_id,
468                    detected.protocol,
469                );
470            }
471            Err(fakecloud_aws::sigv4::SigV4Error::InvalidPresignExpires(_)) => {
472                return build_error_response(
473                    StatusCode::BAD_REQUEST,
474                    "AuthorizationQueryParametersError",
475                    "X-Amz-Expires must be a number between 1 and 604800 seconds",
476                    &request_id,
477                    detected.protocol,
478                );
479            }
480        }
481    }
482
483    // Build path segments. For S3 virtual-hosted-style requests the bucket
484    // lives in the Host header, not the path — prepend it so the S3 handler
485    // sees a uniform path-style request. SigV4 verification above already
486    // ran against the wire path, so this rewrite is signature-safe.
487    let wire_path = parts.uri.path();
488    let path = if detected.service == "s3" {
489        if let Some(bucket) = host_info.as_ref().and_then(|h| h.bucket.as_deref()) {
490            let prefix_with_slash = format!("/{bucket}/");
491            let is_bucket_root = wire_path.trim_end_matches('/') == format!("/{bucket}");
492            if wire_path.starts_with(&prefix_with_slash) || is_bucket_root {
493                wire_path.to_string()
494            } else if wire_path == "/" || wire_path.is_empty() {
495                format!("/{bucket}")
496            } else {
497                format!("/{bucket}{wire_path}")
498            }
499        } else {
500            wire_path.to_string()
501        }
502    } else {
503        wire_path.to_string()
504    };
505    let raw_query = parts.uri.query().unwrap_or("").to_string();
506    let path_segments: Vec<String> = path
507        .split('/')
508        .filter(|s| !s.is_empty())
509        .map(|s| s.to_string())
510        .collect();
511
512    // For JSON protocol, validate that non-empty bodies are valid JSON
513    if detected.protocol == AwsProtocol::Json
514        && !body_bytes.is_empty()
515        && serde_json::from_slice::<serde_json::Value>(&body_bytes).is_err()
516    {
517        return build_error_response(
518            StatusCode::BAD_REQUEST,
519            "SerializationException",
520            "Start of structure or map found where not expected",
521            &request_id,
522            AwsProtocol::Json,
523        );
524    }
525
526    // Merge query params with form body params for the Query family (awsQuery
527    // and ec2Query share identical form-encoded request bodies).
528    let mut all_params = query_params;
529    if matches!(
530        detected.protocol,
531        AwsProtocol::Query | AwsProtocol::Ec2Query
532    ) {
533        let body_params = protocol::parse_query_body(&body_bytes);
534        for (k, v) in body_params {
535            all_params.entry(k).or_insert(v);
536        }
537    }
538
539    // CloudWatch (`monitoring`) advertises awsJson1_0 alongside awsQuery. Its
540    // handlers all read the flat awsQuery param map, so when a client uses the
541    // JSON protocol we flatten the JSON body into that same map, leaving the
542    // handlers unchanged. The handler emits a JSON response for JSON callers.
543    if detected.protocol == AwsProtocol::Json && detected.service == "monitoring" {
544        let body_params = protocol::flatten_json_to_query(&body_bytes);
545        for (k, v) in body_params {
546            all_params.entry(k).or_insert(v);
547        }
548    }
549
550    let aws_request = AwsRequest {
551        service: detected.service.clone(),
552        action: detected.action.clone(),
553        region,
554        account_id: caller_principal
555            .as_ref()
556            .map(|p| p.account_id.clone())
557            .unwrap_or_else(|| config.account_id.clone()),
558        request_id: request_id.clone(),
559        headers: parts.headers,
560        query_params: all_params,
561        body: body_bytes,
562        body_stream: parking_lot::Mutex::new(body_stream),
563        path_segments,
564        raw_path: path,
565        raw_query,
566        method: parts.method,
567        is_query_protocol: matches!(
568            detected.protocol,
569            AwsProtocol::Query | AwsProtocol::Ec2Query
570        ),
571        access_key_id,
572        principal: caller_principal,
573    };
574
575    tracing::info!(
576        service = %aws_request.service,
577        action = %aws_request.action,
578        request_id = %aws_request.request_id,
579        "handling request"
580    );
581
582    // Opt-in IAM identity-policy enforcement. Runs before the service
583    // handler so a deny never reaches business logic. Root principals
584    // (both `test*` bypass AKIDs and the account's IAM root) are exempt,
585    // matching AWS behavior. Services that haven't opted in via
586    // `iam_enforceable()` are transparently skipped — the startup log
587    // lists which services are under enforcement so users always know.
588    if config.iam_mode.is_enabled()
589        && service.iam_enforceable()
590        && !is_root_bypass(aws_request.access_key_id.as_deref().unwrap_or(""))
591    {
592        if let Some(evaluator) = config.policy_evaluator.as_ref() {
593            if let Some(principal) = aws_request.principal.as_ref() {
594                if !principal.is_root() {
595                    if let Some(iam_action) = service.iam_action_for(&aws_request) {
596                        let mut condition_context = build_condition_context(
597                            principal,
598                            remote_addr,
599                            &aws_request.region,
600                            is_secure_transport(&aws_request.headers),
601                        );
602                        // F3 keys riding on the resolved credential. STS
603                        // populates these at mint time so subsequent
604                        // requests under the credential can be evaluated
605                        // against `aws:MultiFactorAuthPresent`,
606                        // `aws:MultiFactorAuthAge`, `aws:TokenIssueTime`,
607                        // and `aws:FederatedProvider`. IAM user access
608                        // keys carry none of these, matching AWS.
609                        if let Some(rc) = resolved.as_ref() {
610                            condition_context.aws_mfa_present = Some(rc.mfa_present);
611                            condition_context.aws_token_issue_time = rc.token_issued_at;
612                            condition_context.aws_federated_provider =
613                                rc.federated_provider.clone();
614                            // `aws:MultiFactorAuthAge` is "seconds since
615                            // MFA was asserted" — computed at evaluation
616                            // time from the token issue moment so the
617                            // value increases monotonically as the session
618                            // ages. Only set when the session was actually
619                            // minted with MFA; otherwise the key is
620                            // absent, matching AWS.
621                            if rc.mfa_present {
622                                if let Some(issued) = rc.token_issued_at {
623                                    let age = chrono::Utc::now()
624                                        .signed_duration_since(issued)
625                                        .num_seconds()
626                                        .max(0);
627                                    condition_context.aws_mfa_age_seconds = Some(age);
628                                }
629                            }
630                        }
631                        condition_context.service_keys =
632                            service.iam_condition_keys_for(&aws_request, &iam_action);
633
634                        // ABAC: populate tag-based condition keys.
635                        // aws:ResourceTag/*
636                        match service.resource_tags_for(&iam_action.resource) {
637                            Some(tags) => condition_context.resource_tags = Some(tags),
638                            None => tracing::debug!(
639                                target: "fakecloud::iam::audit",
640                                service = %detected.service,
641                                resource = %iam_action.resource,
642                                "service does not expose resource tags for ABAC; skipping aws:ResourceTag/* evaluation"
643                            ),
644                        }
645                        // aws:RequestTag/* + aws:TagKeys
646                        match service.request_tags_from(&aws_request, iam_action.action) {
647                            Some(tags) => condition_context.request_tags = Some(tags),
648                            None => tracing::debug!(
649                                target: "fakecloud::iam::audit",
650                                service = %detected.service,
651                                action = %iam_action.action_string(),
652                                "service does not expose request tags for ABAC; skipping aws:RequestTag/* / aws:TagKeys evaluation"
653                            ),
654                        }
655                        // aws:PrincipalTag/*
656                        condition_context.principal_tags = principal.tags.clone();
657
658                        // Phase 2: fetch the resource-based policy (if
659                        // any) attached to the target resource and
660                        // pass it to the evaluator alongside the
661                        // principal's identity policies. The resource's
662                        // owning account is parsed from the ARN (#381
663                        // multi-account alignment); S3 ARNs have an
664                        // empty account field, so we fall back to the
665                        // server's configured account ID in that case.
666                        let resource_policy_json =
667                            config.resource_policy_provider.as_ref().and_then(|p| {
668                                p.resource_policy(&detected.service, &iam_action.resource)
669                            });
670                        // Derive the resource-owning account. Prefer a provider
671                        // lookup (S3 ARNs carry no account, so the bucket's
672                        // owner is resolved from state — without this, account
673                        // A reaching account B's bucket would be mis-read as
674                        // same-account and skip B's bucket-policy requirement,
675                        // bug-audit 2026-05-28, 5.3), then fall back to the
676                        // account embedded in the ARN (SQS/SNS/Lambda/…), then
677                        // to the caller's account for wildcard / unscoped
678                        // actions (ListQueues, GetCallerIdentity).
679                        let resource_account_id = config
680                            .resource_policy_provider
681                            .as_ref()
682                            .and_then(|p| {
683                                p.resource_owner_account(&detected.service, &iam_action.resource)
684                            })
685                            .or_else(|| parse_account_from_arn(&iam_action.resource))
686                            .unwrap_or_else(|| principal.account_id.clone());
687                        // SCP ceiling: resolve the inherited SCP chain
688                        // for this principal (management accounts and
689                        // service-linked roles come back as `None`, in
690                        // which case the evaluator treats the layer as
691                        // absent). Audit breadcrumbs emitted by the
692                        // resolver itself, not here.
693                        let scps = config
694                            .scp_resolver
695                            .as_ref()
696                            .and_then(|r| r.scps_for(principal));
697                        let decision = evaluator.evaluate_with_resource_policy(
698                            principal,
699                            &iam_action,
700                            &condition_context,
701                            resource_policy_json.as_deref(),
702                            &resource_account_id,
703                            &caller_session_policies,
704                            scps.as_deref(),
705                        );
706                        if !decision.is_allow() {
707                            tracing::warn!(
708                                target: "fakecloud::iam::audit",
709                                service = %detected.service,
710                                action = %iam_action.action_string(),
711                                resource = %iam_action.resource,
712                                principal = %principal.arn,
713                                resource_policy_present = resource_policy_json.is_some(),
714                                decision = ?decision,
715                                mode = %config.iam_mode,
716                                request_id = %request_id,
717                                "IAM policy evaluation denied request"
718                            );
719                            if config.iam_mode.is_strict() {
720                                // Real AWS includes an "Encoded
721                                // authorization failure message" suffix
722                                // on AccessDeniedException — an opaque
723                                // base64+zlib JSON blob that the caller
724                                // can pass to STS
725                                // `DecodeAuthorizationMessage` to
726                                // recover the structured deny reason
727                                // (action, principal, matched
728                                // statements, condition context). We
729                                // produce the same blob inline so
730                                // existing tooling that decodes deny
731                                // reasons works against fakecloud.
732                                let context_summary = serde_json::json!({
733                                    "aws:PrincipalArn": principal.arn,
734                                    "aws:PrincipalAccount": principal.account_id,
735                                    "aws:RequestedRegion": condition_context
736                                        .aws_requested_region
737                                        .clone()
738                                        .unwrap_or_default(),
739                                    "aws:SecureTransport": condition_context
740                                        .aws_secure_transport
741                                        .unwrap_or(false),
742                                    "aws:Action": iam_action.action_string(),
743                                    "aws:Resource": iam_action.resource,
744                                    "decision": format!("{:?}", decision),
745                                });
746                                let action_string = iam_action.action_string();
747                                let encoded = crate::auth_message::encode_deny(
748                                    matches!(decision, crate::auth::IamDecision::ExplicitDeny),
749                                    Some(&action_string),
750                                    Some(&principal.arn),
751                                    Vec::new(),
752                                    Some(context_summary),
753                                );
754                                return build_error_response(
755                                    StatusCode::FORBIDDEN,
756                                    "AccessDeniedException",
757                                    &format!(
758                                        "User: {} is not authorized to perform: {} on resource: {} Encoded authorization failure message: {}",
759                                        principal.arn,
760                                        iam_action.action_string(),
761                                        iam_action.resource,
762                                        encoded,
763                                    ),
764                                    &request_id,
765                                    detected.protocol,
766                                );
767                            }
768                            // Soft mode: audit log already emitted; fall
769                            // through to the handler.
770                        }
771                    } else {
772                        // Service opted in via `iam_enforceable()` but its
773                        // `iam_action_for` returned no `IamAction` for this
774                        // specific operation (e.g. S3's `s3_detect_action`
775                        // has `_ => return None` arms for unrecognized
776                        // sub-resources). Under strict enforcement that must
777                        // fail closed: an operation we cannot map to an IAM
778                        // action cannot be authorized, so serving it would be
779                        // a fail-open bypass of `--iam strict`. Deny by
780                        // default. In soft mode we preserve the historical
781                        // warn-and-allow so an incomplete mapping surfaces
782                        // during rollout without blocking traffic.
783                        tracing::warn!(
784                            target: "fakecloud::iam::audit",
785                            service = %detected.service,
786                            action = %aws_request.action,
787                            mode = %config.iam_mode,
788                            request_id = %request_id,
789                            "service is iam_enforceable but has no IamAction mapping for this action; denying under strict, allowing under soft"
790                        );
791                        if config.iam_mode.is_strict() {
792                            return build_error_response(
793                                StatusCode::FORBIDDEN,
794                                "AccessDeniedException",
795                                &format!(
796                                    "User: {} is not authorized to perform: {}: no IAM action mapping exists for this operation, so it cannot be authorized under strict IAM enforcement",
797                                    principal.arn, aws_request.action,
798                                ),
799                                &request_id,
800                                detected.protocol,
801                            );
802                        }
803                        // Soft mode: audit log emitted; fall through to the
804                        // handler.
805                    }
806                }
807            } else if aws_request.access_key_id.is_none() {
808                // Truly anonymous (unsigned) caller — no Authorization header at
809                // all. No identity policies exist, so authorization rests
810                // entirely on the resource policy (a bucket policy granting
811                // `Principal:"*"`) and public-read ACLs — mirroring AWS, which
812                // denies anonymous requests unless the resource is explicitly
813                // made public. Without this an anonymous request that reached an
814                // iam_enforceable service in enforcement mode would bypass
815                // authorization entirely.
816                //
817                // A request that carried an Authorization header but whose
818                // credential did not resolve (principal `None` with
819                // `access_key_id` `Some`) is intentionally left alone here: with
820                // SigV4 verification off, fakecloud does not reject unverified
821                // signed requests, and turning them into anonymous denials would
822                // change long-standing behavior.
823                if let Some(iam_action) = service.iam_action_for(&aws_request) {
824                    let now = chrono::Utc::now();
825                    let mut condition_context = ConditionContext {
826                        aws_source_ip: remote_addr.map(|sa| sa.ip()),
827                        aws_current_time: Some(now),
828                        aws_epoch_time: Some(now.timestamp()),
829                        aws_secure_transport: Some(is_secure_transport(&aws_request.headers)),
830                        aws_requested_region: Some(aws_request.region.clone()),
831                        ..Default::default()
832                    };
833                    condition_context.service_keys =
834                        service.iam_condition_keys_for(&aws_request, &iam_action);
835                    let resource_policy_json = config
836                        .resource_policy_provider
837                        .as_ref()
838                        .and_then(|p| p.resource_policy(&detected.service, &iam_action.resource));
839                    let policy_decision = evaluator.evaluate_anonymous(
840                        &iam_action,
841                        &condition_context,
842                        resource_policy_json.as_deref(),
843                    );
844                    let policy_allows = policy_decision.is_allow();
845                    // An explicit Deny in the resource policy always wins, even
846                    // over a public-read ACL — matching AWS's Deny-overrides
847                    // precedence. Collapsing the decision to a bool and ORing the
848                    // ACL let a public ACL override an explicit anonymous Deny.
849                    let policy_explicit_deny =
850                        matches!(policy_decision, crate::auth::IamDecision::ExplicitDeny);
851                    let acl_allows = !policy_explicit_deny
852                        && config.resource_policy_provider.as_ref().is_some_and(|p| {
853                            p.public_acl_allows(
854                                &detected.service,
855                                &iam_action.resource,
856                                iam_action.action,
857                            )
858                        });
859                    if !policy_allows && !acl_allows {
860                        tracing::warn!(
861                            target: "fakecloud::iam::audit",
862                            service = %detected.service,
863                            action = %iam_action.action_string(),
864                            resource = %iam_action.resource,
865                            resource_policy_present = resource_policy_json.is_some(),
866                            mode = %config.iam_mode,
867                            request_id = %request_id,
868                            "anonymous request denied: no public bucket policy or ACL grants the action"
869                        );
870                        if config.iam_mode.is_strict() {
871                            return build_error_response(
872                                StatusCode::FORBIDDEN,
873                                "AccessDenied",
874                                "Access Denied",
875                                &request_id,
876                                detected.protocol,
877                            );
878                        }
879                        // Soft mode: audit log emitted; fall through to the handler.
880                    }
881                } else {
882                    // Anonymous request to an iam_enforceable service whose
883                    // operation has no IamAction mapping. Mirror the signed
884                    // branch above: an operation we cannot map to an IAM action
885                    // cannot be authorized, so serving it would be a fail-open
886                    // bypass of `--iam strict`. Deny by default under strict;
887                    // soft mode warns and falls through.
888                    tracing::warn!(
889                        target: "fakecloud::iam::audit",
890                        service = %detected.service,
891                        action = %aws_request.action,
892                        mode = %config.iam_mode,
893                        request_id = %request_id,
894                        "anonymous request to iam_enforceable service has no IamAction mapping; denying under strict, allowing under soft"
895                    );
896                    if config.iam_mode.is_strict() {
897                        return build_error_response(
898                            StatusCode::FORBIDDEN,
899                            "AccessDenied",
900                            "Access Denied",
901                            &request_id,
902                            detected.protocol,
903                        );
904                    }
905                }
906            }
907        }
908    }
909
910    match service.handle(aws_request).await {
911        Ok(resp) => {
912            let mut builder = Response::builder()
913                .status(resp.status)
914                .header("x-amzn-requestid", &request_id)
915                .header("x-amz-request-id", &request_id);
916
917            if !resp.content_type.is_empty() {
918                builder = builder.header("content-type", &resp.content_type);
919            }
920
921            let has_content_length = resp
922                .headers
923                .iter()
924                .any(|(k, _)| k.as_str().eq_ignore_ascii_case("content-length"));
925
926            for (k, v) in &resp.headers {
927                builder = builder.header(k, v);
928            }
929
930            match resp.body {
931                ResponseBody::Bytes(b) => builder.body(Body::from(b)).unwrap(),
932                ResponseBody::File { file, size } => {
933                    let stream = tokio_util::io::ReaderStream::new(file);
934                    let body = Body::from_stream(stream);
935                    if !has_content_length {
936                        builder = builder.header("content-length", size.to_string());
937                    }
938                    builder.body(body).unwrap()
939                }
940            }
941        }
942        Err(err) => {
943            tracing::warn!(
944                service = %detected.service,
945                action = %detected.action,
946                error = %err,
947                "request failed"
948            );
949            let error_headers = err.response_headers().to_vec();
950            let mut resp = build_error_response_with_fields(
951                err.status(),
952                err.code(),
953                &err.message(),
954                &request_id,
955                detected.protocol,
956                err.extra_fields(),
957            );
958            for (k, v) in &error_headers {
959                if let (Ok(name), Ok(val)) = (
960                    k.parse::<http::header::HeaderName>(),
961                    v.parse::<http::header::HeaderValue>(),
962                ) {
963                    resp.headers_mut().insert(name, val);
964                }
965            }
966            resp
967        }
968    }
969}
970
971/// Configuration passed to the dispatch handler.
972#[derive(Clone)]
973pub struct DispatchConfig {
974    pub region: String,
975    pub account_id: String,
976    /// Whether to cryptographically verify SigV4 signatures on incoming
977    /// requests. Wired through from `--verify-sigv4` /
978    /// `FAKECLOUD_VERIFY_SIGV4`. Off by default.
979    pub verify_sigv4: bool,
980    /// IAM policy evaluation mode. Wired through from `--iam` /
981    /// `FAKECLOUD_IAM`. Defaults to [`IamMode::Off`]. Actual evaluation is
982    /// added in a later batch; today this field is plumbed but never
983    /// consulted.
984    pub iam_mode: IamMode,
985    /// Resolves access key IDs to their secrets and owning principals.
986    /// Required when `verify_sigv4` or `iam_mode != Off`. When `None`, both
987    /// features gracefully degrade to off-by-default behavior.
988    pub credential_resolver: Option<Arc<dyn CredentialResolver>>,
989    /// Evaluates IAM identity policies for a resolved principal + action.
990    /// Required when `iam_mode != Off`. When `None`, enforcement silently
991    /// degrades to off even if `iam_mode` is set.
992    pub policy_evaluator: Option<Arc<dyn IamPolicyEvaluator>>,
993    /// Resolves resource-based policies (S3 bucket policies in the
994    /// initial rollout) to hand to the evaluator alongside the
995    /// principal's identity policies. `None` means the server was
996    /// started without any resource-policy-owning service registered;
997    /// dispatch then behaves as if no resource policy is attached to
998    /// any resource, identical to the Phase 1 behavior.
999    pub resource_policy_provider: Option<Arc<dyn ResourcePolicyProvider>>,
1000    /// Resolves the ordered SCP chain that applies to a principal's
1001    /// account (root-OU first, account-direct last). `None` means no
1002    /// organizations resolver has been registered — SCPs never gate
1003    /// any request in that case. Off-by-default matches the Batch 4
1004    /// contract: zero behavior change until a user calls
1005    /// `CreateOrganization` and the resolver is wired.
1006    pub scp_resolver: Option<Arc<dyn crate::auth::ScpResolver>>,
1007}
1008
1009impl std::fmt::Debug for DispatchConfig {
1010    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1011        f.debug_struct("DispatchConfig")
1012            .field("region", &self.region)
1013            .field("account_id", &self.account_id)
1014            .field("verify_sigv4", &self.verify_sigv4)
1015            .field("iam_mode", &self.iam_mode)
1016            .field(
1017                "credential_resolver",
1018                &self
1019                    .credential_resolver
1020                    .as_ref()
1021                    .map(|_| "<CredentialResolver>"),
1022            )
1023            .field(
1024                "policy_evaluator",
1025                &self
1026                    .policy_evaluator
1027                    .as_ref()
1028                    .map(|_| "<IamPolicyEvaluator>"),
1029            )
1030            .field(
1031                "resource_policy_provider",
1032                &self
1033                    .resource_policy_provider
1034                    .as_ref()
1035                    .map(|_| "<ResourcePolicyProvider>"),
1036            )
1037            .field(
1038                "scp_resolver",
1039                &self.scp_resolver.as_ref().map(|_| "<ScpResolver>"),
1040            )
1041            .finish()
1042    }
1043}
1044
1045impl DispatchConfig {
1046    /// Minimal constructor for tests and call sites that don't care about the
1047    /// opt-in security features.
1048    pub fn new(region: impl Into<String>, account_id: impl Into<String>) -> Self {
1049        Self {
1050            region: region.into(),
1051            account_id: account_id.into(),
1052            verify_sigv4: false,
1053            iam_mode: IamMode::Off,
1054            credential_resolver: None,
1055            policy_evaluator: None,
1056            resource_policy_provider: None,
1057            scp_resolver: None,
1058        }
1059    }
1060}
1061
1062/// Extract the 12-digit account ID segment from an AWS ARN.
1063///
1064/// ARNs follow `arn:<partition>:<service>:<region>:<account>:<resource>`.
1065/// Identifies routes that opt into streaming request bodies. Returns
1066/// `Some((service, action_hint))` when the dispatch path should hand
1067/// the raw body to the service handler unbuffered, otherwise `None`
1068/// for the default buffered path. The handler reads the stream via
1069/// [`crate::service::AwsRequest::take_body_stream`].
1070///
1071/// Streaming-eligible routes today:
1072///
1073/// * `s3` PUT object — `PUT /<bucket>/<key>` with a SigV4 (or
1074///   presigned) auth header. Covers PutObject, UploadPart, and
1075///   UploadPartCopy. The S3 service spills to disk via
1076///   [`fakecloud_persistence::BodySource::File`] when the stream is
1077///   present.
1078/// * `ecr` OCI Distribution v2 blob upload — `PATCH` and `PUT` on
1079///   `/v2/{name}/blobs/uploads/{uuid}`. The ECR service spools the
1080///   stream into a per-upload temp file before computing the digest.
1081fn streaming_route(
1082    method: &http::Method,
1083    path: &str,
1084    headers: &http::HeaderMap,
1085    query_params: &HashMap<String, String>,
1086) -> Option<(&'static str, &'static str)> {
1087    // ECR OCI v2 blob upload (PATCH chunk + final PUT).
1088    if (method == http::Method::PATCH || method == http::Method::PUT)
1089        && path.starts_with("/v2/")
1090        && path.contains("/blobs/uploads/")
1091    {
1092        return Some(("ecr", ""));
1093    }
1094
1095    // S3 PutObject / UploadPart / UploadPartCopy. Detect either via
1096    // SigV4 service field in the Authorization header OR via a SigV4
1097    // presigned URL (X-Amz-Credential .../s3/...) OR a SigV2 presigned
1098    // URL (AWSAccessKeyId + Signature + Expires query parameters).
1099    if method == http::Method::PUT {
1100        let after = path.trim_start_matches('/');
1101        // Path-style PutObject is `PUT /<bucket>/<key>` (path contains a
1102        // slash); virtual-hosted-style is `PUT /<key>` with the bucket
1103        // in the Host header. For virtual-hosted, accept any non-empty
1104        // path so the key flows through the streaming dispatch — the
1105        // Host parser already routed this request to S3.
1106        let virtual_hosted_s3 = protocol::parse_routing_host_from_headers(headers)
1107            .filter(|h| h.service == "s3" && h.bucket.is_some())
1108            .is_some();
1109        if after.is_empty() || (!virtual_hosted_s3 && !after.contains('/')) {
1110            return None;
1111        }
1112        let header_s3 = headers
1113            .get("authorization")
1114            .and_then(|v| v.to_str().ok())
1115            .and_then(fakecloud_aws::sigv4::parse_sigv4)
1116            .map(|info| info.service == "s3")
1117            .unwrap_or(false);
1118        let presigned_v4_s3 = query_params
1119            .get("X-Amz-Credential")
1120            .and_then(|c| c.split('/').nth(3).map(|s| s.to_string()))
1121            .map(|service| service == "s3")
1122            .unwrap_or(false);
1123        let presigned_v2 = query_params.contains_key("AWSAccessKeyId")
1124            && query_params.contains_key("Signature")
1125            && query_params.contains_key("Expires");
1126        if header_s3 || presigned_v4_s3 || presigned_v2 {
1127            return Some(("s3", ""));
1128        }
1129    }
1130
1131    None
1132}
1133
1134/// Default request-body buffering cap. fakecloud reads the entire
1135/// request body into memory before handing it to a service handler,
1136/// so this ceiling caps RAM usage per in-flight request.
1137///
1138/// Default 1 GiB — comfortably above legitimate single S3 PutObject
1139/// payloads (AWS recommends multipart above ~100 MiB) and each
1140/// multipart part dispatches through here separately. Override with
1141/// `FAKECLOUD_MAX_REQUEST_BODY_BYTES` (decimal bytes) when running
1142/// stress tests that push past the default.
1143const DEFAULT_MAX_REQUEST_BODY_BYTES: usize = 1024 * 1024 * 1024;
1144
1145/// The server-wide buffered request body cap in bytes, from
1146/// `FAKECLOUD_MAX_REQUEST_BODY_BYTES` (default 1 GiB). Public so buffered
1147/// sub-proxies (e.g. the CloudFront viewer data plane) apply the SAME cap as
1148/// direct traffic rather than an inconsistent limit of their own.
1149pub fn max_request_body_bytes() -> usize {
1150    static CACHED: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
1151    *CACHED.get_or_init(|| {
1152        std::env::var("FAKECLOUD_MAX_REQUEST_BODY_BYTES")
1153            .ok()
1154            .and_then(|s| s.parse::<usize>().ok())
1155            .filter(|&n| n > 0)
1156            .unwrap_or(DEFAULT_MAX_REQUEST_BODY_BYTES)
1157    })
1158}
1159
1160/// For the cross-account decision in IAM enforcement, the "resource
1161/// account" is the ARN's account segment. Some services (notably S3)
1162/// produce ARNs with an empty account field — for those we return
1163/// `None` and let the caller fall back to the server's configured
1164/// account ID. Malformed or non-ARN strings also return `None`.
1165fn parse_account_from_arn(arn: &str) -> Option<String> {
1166    let mut parts = arn.splitn(6, ':');
1167    if parts.next()? != "arn" {
1168        return None;
1169    }
1170    let _partition = parts.next()?;
1171    let _service = parts.next()?;
1172    let _region = parts.next()?;
1173    let account = parts.next()?;
1174    // Resource segment must exist (parts.next().is_some()) for the ARN
1175    // to be well-formed, but we don't consume its value here.
1176    parts.next()?;
1177    if account.is_empty() {
1178        None
1179    } else {
1180        Some(account.to_string())
1181    }
1182}
1183
1184/// Whether the request's `user-agent` (or `x-amz-user-agent`) carries the
1185/// `api/neptune` SDK-metadata token the `aws-sdk-neptune` client stamps in.
1186/// Used to disambiguate Neptune from RDS, which share the `rds` SigV4 scope
1187/// and endpoint. Matches the `api/neptune` token that AWS SDKs emit (a
1188/// leading `api/neptune` optionally followed by `#`/`/` and a version).
1189fn user_agent_indicates_neptune(headers: &http::HeaderMap) -> bool {
1190    for name in ["user-agent", "x-amz-user-agent"] {
1191        if let Some(ua) = headers.get(name).and_then(|v| v.to_str().ok()) {
1192            for part in ua.split_whitespace() {
1193                if let Some(rest) = part.strip_prefix("api/neptune") {
1194                    if rest.is_empty() || rest.starts_with('#') || rest.starts_with('/') {
1195                        return true;
1196                    }
1197                }
1198            }
1199        }
1200    }
1201    false
1202}
1203
1204/// Extract region from User-Agent header suffix `region/<region>`.
1205/// Whether the request's `user-agent` (or `x-amz-user-agent`) carries the
1206/// `api/docdb` SDK-metadata token the `aws-sdk-docdb` client stamps in. Used
1207/// to disambiguate DocumentDB from RDS, which share the `rds` SigV4 scope
1208/// and endpoint. Matches the `api/docdb` token that AWS SDKs emit (a leading
1209/// `api/docdb` optionally followed by `#`/`/` and a version).
1210fn user_agent_indicates_docdb(headers: &http::HeaderMap) -> bool {
1211    for name in ["user-agent", "x-amz-user-agent"] {
1212        if let Some(ua) = headers.get(name).and_then(|v| v.to_str().ok()) {
1213            for part in ua.split_whitespace() {
1214                if let Some(rest) = part.strip_prefix("api/docdb") {
1215                    if rest.is_empty() || rest.starts_with('#') || rest.starts_with('/') {
1216                        return true;
1217                    }
1218                }
1219            }
1220        }
1221    }
1222    false
1223}
1224
1225fn extract_region_from_user_agent(headers: &http::HeaderMap) -> Option<String> {
1226    let ua = headers.get("user-agent")?.to_str().ok()?;
1227    for part in ua.split_whitespace() {
1228        if let Some(region) = part.strip_prefix("region/") {
1229            if !region.is_empty() {
1230                return Some(region.to_string());
1231            }
1232        }
1233    }
1234    None
1235}
1236
1237fn build_error_response(
1238    status: StatusCode,
1239    code: &str,
1240    message: &str,
1241    request_id: &str,
1242    protocol: AwsProtocol,
1243) -> Response<Body> {
1244    build_error_response_with_fields(status, code, message, request_id, protocol, &[])
1245}
1246
1247fn build_error_response_with_fields(
1248    status: StatusCode,
1249    code: &str,
1250    message: &str,
1251    request_id: &str,
1252    protocol: AwsProtocol,
1253    extra_fields: &[(String, String)],
1254) -> Response<Body> {
1255    let (status, content_type, body) = match protocol {
1256        // awsQuery services (SQS, SNS, IAM, STS, RDS, ELBv2, CloudWatch,
1257        // AutoScaling, ...) share the `<ErrorResponse>` envelope.
1258        AwsProtocol::Query => {
1259            fakecloud_aws::error::xml_error_response(status, code, message, request_id)
1260        }
1261        // EC2 uses the distinct ec2Query error envelope
1262        // (`<Response><Errors><Error>...</Errors><RequestID>`). Only EC2 is
1263        // classified `Ec2Query`, so the other Query-protocol services above are
1264        // untouched.
1265        AwsProtocol::Ec2Query => {
1266            fakecloud_aws::ec2query::ec2_error_response(status, code, message, request_id)
1267        }
1268        AwsProtocol::Rest => fakecloud_aws::error::s3_xml_error_response_with_fields(
1269            status,
1270            code,
1271            message,
1272            request_id,
1273            extra_fields,
1274        ),
1275        AwsProtocol::Json | AwsProtocol::RestJson => {
1276            fakecloud_aws::error::json_error_response_with_fields(
1277                status,
1278                code,
1279                message,
1280                extra_fields,
1281            )
1282        }
1283    };
1284
1285    // S3 (and other REST-XML services) place the error code in
1286    // `x-amz-error-code` so HEAD responses — which HTTP forbids from
1287    // carrying a body — still surface the code. AWS SDKs read this header
1288    // when the body is empty. Emit it on every error response so HEAD,
1289    // OPTIONS, and any client that strips the body still see the code.
1290    // Backend errors regularly include newlines (multi-line stderr from
1291    // docker/podman/etc.); HTTP header values reject control characters,
1292    // so sanitize before insertion or the builder rejects the response
1293    // and the connection drops.
1294    let safe_code = sanitize_header_value(code);
1295    let safe_message = sanitize_header_value(message);
1296    let mut builder = Response::builder()
1297        .status(status)
1298        .header("content-type", content_type)
1299        .header("x-amzn-requestid", request_id)
1300        .header("x-amz-request-id", request_id);
1301    if let Ok(v) = http::HeaderValue::from_str(&safe_code) {
1302        builder = builder.header("x-amz-error-code", v);
1303    }
1304    if let Ok(v) = http::HeaderValue::from_str(&safe_message) {
1305        builder = builder.header("x-amz-error-message", v);
1306    }
1307    builder.body(Body::from(body)).unwrap_or_else(|_| {
1308        // Builder only fails if a header is invalid; we sanitized the two
1309        // we control, so the remaining ones (content-type, request id) are
1310        // ASCII and safe. This fallback exists purely so we never panic.
1311        Response::new(Body::empty())
1312    })
1313}
1314
1315/// Strip characters that HTTP header values reject (control bytes, CR/LF/TAB)
1316/// and truncate to a length that AWS SDKs handle cleanly. Backend tools
1317/// (docker, podman, kubectl, …) emit multi-line stderr, and forwarding that
1318/// raw into `x-amz-error-message` previously panicked the dispatcher.
1319fn sanitize_header_value(s: &str) -> String {
1320    const MAX_LEN: usize = 1024;
1321    let mut out = String::with_capacity(s.len().min(MAX_LEN));
1322    for ch in s.chars() {
1323        if out.len() >= MAX_LEN {
1324            break;
1325        }
1326        // Header values forbid CR, LF, and other control bytes (RFC 9110).
1327        // Replace with a single space so multi-line messages stay readable.
1328        if ch.is_control() {
1329            if !out.ends_with(' ') {
1330                out.push(' ');
1331            }
1332        } else {
1333            out.push(ch);
1334        }
1335    }
1336    out.trim().to_string()
1337}
1338
1339/// Build the [`ConditionContext`] passed to the IAM evaluator for one
1340/// request. Populates the 10 global condition keys from the resolved
1341/// principal + the HTTP request. Service-specific keys are deferred to
1342/// a follow-up batch and left empty.
1343/// For an unsigned request that no other detection rule claimed, return the
1344/// bucket name when the first path segment names an existing S3 bucket.
1345///
1346/// fakecloud serves every service from one endpoint, so an anonymous
1347/// path-style S3 request (`GET /bucket/key`, no SigV4) is indistinguishable
1348/// from an API Gateway execute-api call by headers alone. Bucket existence is
1349/// the disambiguator: if the segment is a real bucket, route to S3; otherwise
1350/// fall through to the apigateway catch-all. Uses the already-wired
1351/// `resource_policy_provider`, which resolves S3 bucket ownership from state
1352/// (`Some` => the bucket exists). Returns `None` when no provider is wired.
1353/// Recover the access key from a SigV2 presigned URL. AWS SigV2 presigning
1354/// puts the key in the `AWSAccessKeyId` query parameter alongside `Signature`
1355/// and `Expires`; all three must be present for the URL to be a SigV2 presign.
1356/// Returns None for SigV4 presigns (which use `X-Amz-Credential`) or unsigned
1357/// requests.
1358fn sigv2_presigned_access_key(query_params: &HashMap<String, String>) -> Option<String> {
1359    if query_params.contains_key("Signature") && query_params.contains_key("Expires") {
1360        query_params.get("AWSAccessKeyId").cloned()
1361    } else {
1362        None
1363    }
1364}
1365
1366/// True when `s` is a 64-character lowercase-hex SHA-256 digest — the form a
1367/// signed `x-amz-content-sha256` payload hash takes. Distinguishes a real body
1368/// hash from the `UNSIGNED-PAYLOAD` / `STREAMING-*` markers (shorter and
1369/// containing non-hex characters), so only genuine hashes are re-bound to the
1370/// buffered body.
1371fn is_hex_sha256(s: &str) -> bool {
1372    s.len() == 64 && s.bytes().all(|b| matches!(b, b'0'..=b'9' | b'a'..=b'f'))
1373}
1374
1375/// Lowercase-hex SHA-256 of `bytes`, used to re-bind a signed
1376/// `x-amz-content-sha256` to the buffered request body for non-S3 services.
1377fn sha256_hex_lower(bytes: &[u8]) -> String {
1378    use sha2::{Digest, Sha256};
1379    let digest = Sha256::digest(bytes);
1380    const HEX: &[u8] = b"0123456789abcdef";
1381    let mut out = String::with_capacity(64);
1382    for b in digest {
1383        out.push(HEX[(b >> 4) as usize] as char);
1384        out.push(HEX[(b & 0x0f) as usize] as char);
1385    }
1386    out
1387}
1388
1389fn anonymous_s3_bucket(uri: &http::Uri, config: &DispatchConfig) -> Option<String> {
1390    let provider = config.resource_policy_provider.as_ref()?;
1391    let segment = uri.path().split('/').find(|s| !s.is_empty())?.to_string();
1392    let arn = format!("arn:aws:s3:::{segment}");
1393    provider.resource_owner_account("s3", &arn).map(|_| segment)
1394}
1395
1396fn build_condition_context(
1397    principal: &Principal,
1398    remote_addr: Option<SocketAddr>,
1399    region: &str,
1400    secure_transport: bool,
1401) -> ConditionContext {
1402    let now = chrono::Utc::now();
1403    ConditionContext {
1404        aws_username: aws_username_from_principal(principal),
1405        aws_userid: Some(principal.user_id.clone()),
1406        aws_principal_arn: Some(principal.arn.clone()),
1407        aws_principal_account: Some(principal.account_id.clone()),
1408        aws_principal_type: Some(principal_type_label(principal.principal_type).to_string()),
1409        aws_source_ip: remote_addr.map(|sa| sa.ip()),
1410        aws_current_time: Some(now),
1411        aws_epoch_time: Some(now.timestamp()),
1412        aws_secure_transport: Some(secure_transport),
1413        aws_requested_region: Some(region.to_string()),
1414        // F3 keys: populated from the caller's session context when STS
1415        // mints credentials with MFA / SAML / OIDC / VPC-endpoint hints.
1416        // Default-None here so tests/dispatch sites that don't set them
1417        // safe-fail any policy referencing them — matching AWS for keys
1418        // that aren't asserted.
1419        aws_mfa_present: None,
1420        aws_mfa_age_seconds: None,
1421        aws_called_via: Vec::new(),
1422        aws_source_vpce: None,
1423        aws_source_vpc: None,
1424        aws_vpc_source_ip: None,
1425        aws_federated_provider: None,
1426        aws_token_issue_time: None,
1427        service_keys: Default::default(),
1428        resource_tags: None,
1429        request_tags: None,
1430        principal_tags: None,
1431    }
1432}
1433
1434/// `aws:username` is only set for IAM users, matching AWS. For assumed
1435/// roles, federated users, root, and unknown principals the key is
1436/// absent — operators that reference it without `IfExists` safe-fail.
1437fn aws_username_from_principal(principal: &Principal) -> Option<String> {
1438    if principal.principal_type != PrincipalType::User {
1439        return None;
1440    }
1441    let after = principal.arn.rsplit_once(":user/").map(|(_, s)| s)?;
1442    // Strip any IAM path prefix; bare username is the last segment.
1443    Some(after.rsplit('/').next().unwrap_or(after).to_string())
1444}
1445
1446/// AWS's `aws:PrincipalType` uses PascalCase identifiers, distinct from
1447/// the lowercase ones [`PrincipalType::as_str`] returns for ARNs.
1448fn principal_type_label(t: PrincipalType) -> &'static str {
1449    match t {
1450        PrincipalType::User => "User",
1451        PrincipalType::AssumedRole => "AssumedRole",
1452        PrincipalType::FederatedUser => "FederatedUser",
1453        PrincipalType::Root => "Account",
1454        PrincipalType::Unknown => "Unknown",
1455    }
1456}
1457
1458/// Best-effort detection of TLS-terminated requests. Direct HTTPS
1459/// connections are not yet supported by the fakecloud server (it speaks
1460/// plain HTTP), so the only signal is an `x-forwarded-proto: https`
1461/// header set by an upstream proxy. Anything else evaluates to `false`,
1462/// which matches the typical local-dev setup.
1463fn is_secure_transport(headers: &http::HeaderMap) -> bool {
1464    headers
1465        .get("x-forwarded-proto")
1466        .and_then(|v| v.to_str().ok())
1467        .map(|s| s.eq_ignore_ascii_case("https"))
1468        .unwrap_or(false)
1469}
1470
1471trait ProtocolExt {
1472    fn error_status(&self) -> StatusCode;
1473}
1474
1475impl ProtocolExt for AwsProtocol {
1476    fn error_status(&self) -> StatusCode {
1477        StatusCode::BAD_REQUEST
1478    }
1479}
1480
1481#[cfg(test)]
1482mod tests {
1483    use super::*;
1484
1485    #[test]
1486    fn default_max_request_body_bytes_is_one_gib() {
1487        // Without the env override, the cap defaults to 1 GiB. The
1488        // public function caches via OnceLock so only the first call
1489        // in the process matters; we assert the constant directly.
1490        assert_eq!(DEFAULT_MAX_REQUEST_BODY_BYTES, 1024 * 1024 * 1024);
1491    }
1492
1493    #[test]
1494    fn sigv2_presigned_access_key_extracted_with_signature_and_expires() {
1495        let mut q = HashMap::new();
1496        q.insert("AWSAccessKeyId".to_string(), "AKIAEXAMPLE".to_string());
1497        q.insert("Signature".to_string(), "abc%2Bdef".to_string());
1498        q.insert("Expires".to_string(), "1700000000".to_string());
1499        assert_eq!(
1500            sigv2_presigned_access_key(&q).as_deref(),
1501            Some("AKIAEXAMPLE")
1502        );
1503    }
1504
1505    #[test]
1506    fn sigv2_presigned_access_key_none_without_signature_or_expires() {
1507        // AWSAccessKeyId alone (e.g. a stray query param) is not a SigV2
1508        // presign and must not be treated as a credential.
1509        let mut q = HashMap::new();
1510        q.insert("AWSAccessKeyId".to_string(), "AKIAEXAMPLE".to_string());
1511        assert_eq!(sigv2_presigned_access_key(&q), None);
1512
1513        q.insert("Expires".to_string(), "1700000000".to_string());
1514        assert_eq!(
1515            sigv2_presigned_access_key(&q),
1516            None,
1517            "missing Signature must not qualify"
1518        );
1519    }
1520
1521    #[test]
1522    fn sigv2_presigned_access_key_none_for_unsigned_request() {
1523        assert_eq!(sigv2_presigned_access_key(&HashMap::new()), None);
1524    }
1525
1526    #[test]
1527    fn is_hex_sha256_accepts_real_digest_rejects_markers() {
1528        // A genuine 64-char lowercase-hex digest is a bindable body hash.
1529        assert!(is_hex_sha256(&sha256_hex_lower(b"hello")));
1530        assert!(is_hex_sha256(
1531            "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
1532        ));
1533        // The SigV4 payload markers are NOT hashes and must be skipped.
1534        assert!(!is_hex_sha256("UNSIGNED-PAYLOAD"));
1535        assert!(!is_hex_sha256("STREAMING-AWS4-HMAC-SHA256-PAYLOAD"));
1536        assert!(!is_hex_sha256("STREAMING-UNSIGNED-PAYLOAD-TRAILER"));
1537        // Wrong length or non-lowercase-hex characters are rejected.
1538        assert!(!is_hex_sha256("abc123"));
1539        assert!(!is_hex_sha256(
1540            "E3B0C44298FC1C149AFBF4C8996FB92427AE41E4649B934CA495991B7852B855"
1541        ));
1542    }
1543
1544    #[test]
1545    fn sha256_hex_lower_matches_known_vectors() {
1546        // Empty input -> the well-known SHA-256 of "".
1547        assert_eq!(
1548            sha256_hex_lower(b""),
1549            "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
1550        );
1551        assert_eq!(
1552            sha256_hex_lower(b"abc"),
1553            "ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"
1554        );
1555        assert_eq!(sha256_hex_lower(b"abc").len(), 64);
1556    }
1557
1558    #[test]
1559    fn dispatch_config_new_defaults_to_off() {
1560        let cfg = DispatchConfig::new("us-east-1", "123456789012");
1561        assert_eq!(cfg.region, "us-east-1");
1562        assert_eq!(cfg.account_id, "123456789012");
1563        assert!(!cfg.verify_sigv4);
1564        assert_eq!(cfg.iam_mode, IamMode::Off);
1565    }
1566
1567    #[test]
1568    fn aws_username_strips_iam_path_for_users() {
1569        let p = Principal {
1570            arn: "arn:aws:iam::123456789012:user/engineering/alice".into(),
1571            user_id: "AIDAALICE".into(),
1572            account_id: "123456789012".into(),
1573            principal_type: PrincipalType::User,
1574            source_identity: None,
1575            tags: None,
1576        };
1577        assert_eq!(aws_username_from_principal(&p), Some("alice".into()));
1578    }
1579
1580    #[test]
1581    fn aws_username_unset_for_assumed_role() {
1582        let p = Principal {
1583            arn: "arn:aws:sts::123456789012:assumed-role/ops/session".into(),
1584            user_id: "AROAOPS:session".into(),
1585            account_id: "123456789012".into(),
1586            principal_type: PrincipalType::AssumedRole,
1587            source_identity: None,
1588            tags: None,
1589        };
1590        assert_eq!(aws_username_from_principal(&p), None);
1591    }
1592
1593    #[test]
1594    fn principal_type_label_matches_aws_casing() {
1595        assert_eq!(principal_type_label(PrincipalType::User), "User");
1596        assert_eq!(
1597            principal_type_label(PrincipalType::AssumedRole),
1598            "AssumedRole"
1599        );
1600        assert_eq!(principal_type_label(PrincipalType::Root), "Account");
1601    }
1602
1603    #[test]
1604    fn build_condition_context_populates_global_keys() {
1605        let p = Principal {
1606            arn: "arn:aws:iam::123456789012:user/alice".into(),
1607            user_id: "AIDAALICE".into(),
1608            account_id: "123456789012".into(),
1609            principal_type: PrincipalType::User,
1610            source_identity: None,
1611            tags: None,
1612        };
1613        let addr: SocketAddr = "10.0.0.1:54321".parse().unwrap();
1614        let ctx = build_condition_context(&p, Some(addr), "us-east-1", false);
1615        assert_eq!(ctx.aws_username.as_deref(), Some("alice"));
1616        assert_eq!(ctx.aws_userid.as_deref(), Some("AIDAALICE"));
1617        assert_eq!(
1618            ctx.aws_principal_arn.as_deref(),
1619            Some("arn:aws:iam::123456789012:user/alice")
1620        );
1621        assert_eq!(ctx.aws_principal_account.as_deref(), Some("123456789012"));
1622        assert_eq!(ctx.aws_principal_type.as_deref(), Some("User"));
1623        assert_eq!(
1624            ctx.aws_source_ip.map(|i| i.to_string()).as_deref(),
1625            Some("10.0.0.1")
1626        );
1627        assert_eq!(ctx.aws_requested_region.as_deref(), Some("us-east-1"));
1628        assert_eq!(ctx.aws_secure_transport, Some(false));
1629        assert!(ctx.aws_current_time.is_some());
1630        assert!(ctx.aws_epoch_time.is_some());
1631    }
1632
1633    #[test]
1634    fn is_secure_transport_reads_x_forwarded_proto() {
1635        let mut headers = http::HeaderMap::new();
1636        headers.insert("x-forwarded-proto", "https".parse().unwrap());
1637        assert!(is_secure_transport(&headers));
1638        headers.insert("x-forwarded-proto", "http".parse().unwrap());
1639        assert!(!is_secure_transport(&headers));
1640        let empty = http::HeaderMap::new();
1641        assert!(!is_secure_transport(&empty));
1642    }
1643
1644    #[test]
1645    fn parse_account_from_arn_extracts_standard_shapes() {
1646        assert_eq!(
1647            parse_account_from_arn("arn:aws:sqs:us-east-1:123456789012:queue"),
1648            Some("123456789012".to_string())
1649        );
1650        assert_eq!(
1651            parse_account_from_arn("arn:aws:iam::123456789012:user/alice"),
1652            Some("123456789012".to_string())
1653        );
1654    }
1655
1656    #[test]
1657    fn parse_account_from_arn_returns_none_for_s3_empty_account() {
1658        // S3 ARNs have both region and account empty.
1659        assert_eq!(parse_account_from_arn("arn:aws:s3:::my-bucket"), None);
1660        assert_eq!(
1661            parse_account_from_arn("arn:aws:s3:::my-bucket/path/to/key"),
1662            None
1663        );
1664    }
1665
1666    #[test]
1667    fn parse_account_from_arn_returns_none_for_malformed() {
1668        assert_eq!(parse_account_from_arn(""), None);
1669        assert_eq!(parse_account_from_arn("not-an-arn"), None);
1670        assert_eq!(parse_account_from_arn("arn:aws:sqs:us-east-1"), None);
1671        assert_eq!(parse_account_from_arn("arn:aws:sqs"), None);
1672    }
1673
1674    #[test]
1675    fn extract_region_from_user_agent_finds_region_segment() {
1676        let mut headers = http::HeaderMap::new();
1677        headers.insert(
1678            "user-agent",
1679            "aws-sdk-rust/1.0 os/linux region/eu-central-1"
1680                .parse()
1681                .unwrap(),
1682        );
1683        assert_eq!(
1684            extract_region_from_user_agent(&headers),
1685            Some("eu-central-1".to_string())
1686        );
1687    }
1688
1689    #[test]
1690    fn extract_region_from_user_agent_none_without_header() {
1691        let headers = http::HeaderMap::new();
1692        assert_eq!(extract_region_from_user_agent(&headers), None);
1693    }
1694
1695    #[test]
1696    fn extract_region_from_user_agent_ignores_empty_region() {
1697        let mut headers = http::HeaderMap::new();
1698        headers.insert("user-agent", "aws-sdk-java region/".parse().unwrap());
1699        assert_eq!(extract_region_from_user_agent(&headers), None);
1700    }
1701
1702    #[test]
1703    fn extract_region_from_user_agent_none_when_no_region_marker() {
1704        let mut headers = http::HeaderMap::new();
1705        headers.insert("user-agent", "curl/7.79.1".parse().unwrap());
1706        assert_eq!(extract_region_from_user_agent(&headers), None);
1707    }
1708
1709    #[test]
1710    fn aws_username_none_for_root() {
1711        let p = Principal {
1712            arn: "arn:aws:iam::123456789012:root".into(),
1713            user_id: "123456789012".into(),
1714            account_id: "123456789012".into(),
1715            principal_type: PrincipalType::Root,
1716            source_identity: None,
1717            tags: None,
1718        };
1719        assert_eq!(aws_username_from_principal(&p), None);
1720    }
1721
1722    #[test]
1723    fn aws_username_bare_no_path() {
1724        let p = Principal {
1725            arn: "arn:aws:iam::123456789012:user/bob".into(),
1726            user_id: "AIDABOB".into(),
1727            account_id: "123456789012".into(),
1728            principal_type: PrincipalType::User,
1729            source_identity: None,
1730            tags: None,
1731        };
1732        assert_eq!(aws_username_from_principal(&p), Some("bob".into()));
1733    }
1734
1735    #[test]
1736    fn principal_type_label_covers_federated_and_unknown() {
1737        assert_eq!(
1738            principal_type_label(PrincipalType::FederatedUser),
1739            "FederatedUser"
1740        );
1741        assert_eq!(principal_type_label(PrincipalType::Unknown), "Unknown");
1742    }
1743
1744    #[test]
1745    fn build_condition_context_marks_secure_when_flag_set() {
1746        let p = Principal {
1747            arn: "arn:aws:iam::123456789012:user/alice".into(),
1748            user_id: "AIDAALICE".into(),
1749            account_id: "123456789012".into(),
1750            principal_type: PrincipalType::User,
1751            source_identity: None,
1752            tags: None,
1753        };
1754        let ctx = build_condition_context(&p, None, "us-west-2", true);
1755        assert_eq!(ctx.aws_secure_transport, Some(true));
1756        assert!(ctx.aws_source_ip.is_none());
1757        assert_eq!(ctx.aws_requested_region.as_deref(), Some("us-west-2"));
1758    }
1759
1760    #[test]
1761    fn is_secure_transport_case_insensitive() {
1762        let mut headers = http::HeaderMap::new();
1763        headers.insert("x-forwarded-proto", "HTTPS".parse().unwrap());
1764        assert!(is_secure_transport(&headers));
1765    }
1766
1767    #[test]
1768    fn is_secure_transport_non_ascii_bytes_false() {
1769        let mut headers = http::HeaderMap::new();
1770        headers.insert(
1771            "x-forwarded-proto",
1772            http::HeaderValue::from_bytes(&[0xFF, 0xFE]).unwrap(),
1773        );
1774        assert!(!is_secure_transport(&headers));
1775    }
1776
1777    #[test]
1778    fn protocol_ext_error_status_is_bad_request() {
1779        assert_eq!(AwsProtocol::Query.error_status(), StatusCode::BAD_REQUEST);
1780        assert_eq!(AwsProtocol::Json.error_status(), StatusCode::BAD_REQUEST);
1781        assert_eq!(AwsProtocol::Rest.error_status(), StatusCode::BAD_REQUEST);
1782        assert_eq!(
1783            AwsProtocol::RestJson.error_status(),
1784            StatusCode::BAD_REQUEST
1785        );
1786    }
1787
1788    #[test]
1789    fn build_error_response_json_has_json_content_type() {
1790        let resp = build_error_response(
1791            StatusCode::BAD_REQUEST,
1792            "TestCode",
1793            "test msg",
1794            "req-1",
1795            AwsProtocol::Json,
1796        );
1797        assert_eq!(resp.status(), StatusCode::BAD_REQUEST);
1798        let ct = resp
1799            .headers()
1800            .get("content-type")
1801            .unwrap()
1802            .to_str()
1803            .unwrap();
1804        assert!(ct.contains("json"));
1805        let rid = resp
1806            .headers()
1807            .get("x-amzn-requestid")
1808            .unwrap()
1809            .to_str()
1810            .unwrap();
1811        assert_eq!(rid, "req-1");
1812    }
1813
1814    #[test]
1815    fn build_error_response_rest_returns_xml_content_type() {
1816        let resp = build_error_response(
1817            StatusCode::NOT_FOUND,
1818            "NoSuchBucket",
1819            "bucket missing",
1820            "req-2",
1821            AwsProtocol::Rest,
1822        );
1823        assert_eq!(resp.status(), StatusCode::NOT_FOUND);
1824        let ct = resp
1825            .headers()
1826            .get("content-type")
1827            .unwrap()
1828            .to_str()
1829            .unwrap();
1830        assert!(ct.contains("xml"));
1831    }
1832
1833    #[test]
1834    fn build_error_response_query_returns_xml() {
1835        let resp = build_error_response(
1836            StatusCode::BAD_REQUEST,
1837            "InvalidParameter",
1838            "bad param",
1839            "req-3",
1840            AwsProtocol::Query,
1841        );
1842        let ct = resp
1843            .headers()
1844            .get("content-type")
1845            .unwrap()
1846            .to_str()
1847            .unwrap();
1848        assert!(ct.contains("xml"));
1849    }
1850
1851    /// Regression for issue #1539: multi-line backend errors (e.g. podman
1852    /// stderr) used to panic the dispatcher when stuffed into the
1853    /// `x-amz-error-message` HTTP header. The response must build cleanly
1854    /// and the header value must not contain control characters.
1855    #[test]
1856    fn build_error_response_with_multiline_message_does_not_panic() {
1857        let resp = build_error_response(
1858            StatusCode::INTERNAL_SERVER_ERROR,
1859            "ServiceException",
1860            "Lambda execution failed: container failed to start: docker start failed: \
1861             Error: unable to start container \"abc\": \
1862             failed to create new hosts file:\nhost-gateway is empty\n",
1863            "req-multi",
1864            AwsProtocol::Json,
1865        );
1866        assert_eq!(resp.status(), StatusCode::INTERNAL_SERVER_ERROR);
1867        let msg = resp
1868            .headers()
1869            .get("x-amz-error-message")
1870            .expect("x-amz-error-message must be set even when input contains newlines")
1871            .to_str()
1872            .unwrap();
1873        assert!(!msg.contains('\n'));
1874        assert!(!msg.contains('\r'));
1875        assert!(msg.contains("Lambda execution failed"));
1876        assert!(msg.contains("host-gateway is empty"));
1877    }
1878
1879    #[test]
1880    fn build_error_response_with_control_chars_strips_them() {
1881        let resp = build_error_response(
1882            StatusCode::BAD_REQUEST,
1883            "Code\twith\ttabs",
1884            "msg\x00with\x01nulls",
1885            "req-ctrl",
1886            AwsProtocol::Json,
1887        );
1888        let code = resp
1889            .headers()
1890            .get("x-amz-error-code")
1891            .unwrap()
1892            .to_str()
1893            .unwrap();
1894        let msg = resp
1895            .headers()
1896            .get("x-amz-error-message")
1897            .unwrap()
1898            .to_str()
1899            .unwrap();
1900        assert!(!code.contains('\t'));
1901        assert!(!msg.contains('\x00'));
1902        assert!(!msg.contains('\x01'));
1903    }
1904
1905    #[test]
1906    fn sanitize_header_value_truncates_long_input() {
1907        let huge = "x".repeat(5_000);
1908        let out = sanitize_header_value(&huge);
1909        assert!(out.len() <= 1024);
1910    }
1911
1912    #[test]
1913    fn sanitize_header_value_collapses_consecutive_control_runs() {
1914        let out = sanitize_header_value("a\n\n\n\rb");
1915        assert_eq!(out, "a b");
1916    }
1917
1918    #[test]
1919    fn dispatch_config_carries_opt_in_flags() {
1920        let cfg = DispatchConfig {
1921            region: "eu-west-1".to_string(),
1922            account_id: "000000000000".to_string(),
1923            verify_sigv4: true,
1924            iam_mode: IamMode::Strict,
1925            credential_resolver: None,
1926            policy_evaluator: None,
1927            resource_policy_provider: None,
1928            scp_resolver: None,
1929        };
1930        assert!(cfg.verify_sigv4);
1931        assert!(cfg.iam_mode.is_strict());
1932        assert!(cfg.resource_policy_provider.is_none());
1933        assert!(cfg.scp_resolver.is_none());
1934    }
1935
1936    fn s3_sigv4_headers() -> http::HeaderMap {
1937        let mut headers = http::HeaderMap::new();
1938        headers.insert(
1939            "authorization",
1940            "AWS4-HMAC-SHA256 Credential=test/20240101/us-east-1/s3/aws4_request, \
1941             SignedHeaders=host, Signature=fake"
1942                .parse()
1943                .unwrap(),
1944        );
1945        headers
1946    }
1947
1948    #[test]
1949    fn streaming_route_path_style_s3_put_object() {
1950        let headers = s3_sigv4_headers();
1951        assert_eq!(
1952            streaming_route(
1953                &http::Method::PUT,
1954                "/my-bucket/key.txt",
1955                &headers,
1956                &HashMap::new(),
1957            ),
1958            Some(("s3", "")),
1959        );
1960    }
1961
1962    #[test]
1963    fn streaming_route_path_style_create_bucket_skipped() {
1964        // `PUT /bucket` (no trailing key) is CreateBucket — must NOT
1965        // hit the streaming path.
1966        let headers = s3_sigv4_headers();
1967        assert_eq!(
1968            streaming_route(&http::Method::PUT, "/my-bucket", &headers, &HashMap::new(),),
1969            None,
1970        );
1971    }
1972
1973    #[test]
1974    fn streaming_route_virtual_hosted_s3_put_object() {
1975        let mut headers = s3_sigv4_headers();
1976        headers.insert(
1977            "host",
1978            "vhost-bucket.s3.us-east-1.localhost.localstack.cloud:4566"
1979                .parse()
1980                .unwrap(),
1981        );
1982        // Virtual-hosted PUT has no bucket in the URL path (`/<key>`),
1983        // so the slash check used for path-style would reject it. The
1984        // Host parser confirms this is virtual-hosted S3 and the key
1985        // flows through the streaming dispatch.
1986        assert_eq!(
1987            streaming_route(&http::Method::PUT, "/hello.txt", &headers, &HashMap::new(),),
1988            Some(("s3", "")),
1989        );
1990    }
1991
1992    #[test]
1993    fn streaming_route_virtual_hosted_s3_root_skipped() {
1994        // `PUT /` against a virtual-hosted Host = CreateBucket, which
1995        // is handled buffered. Empty path-after-slash must short-circuit.
1996        let mut headers = s3_sigv4_headers();
1997        headers.insert(
1998            "host",
1999            "vhost-bucket.s3.us-east-1.localhost.localstack.cloud:4566"
2000                .parse()
2001                .unwrap(),
2002        );
2003        assert_eq!(
2004            streaming_route(&http::Method::PUT, "/", &headers, &HashMap::new()),
2005            None,
2006        );
2007    }
2008
2009    #[test]
2010    fn streaming_route_ecr_blob_upload() {
2011        let headers = http::HeaderMap::new();
2012        assert_eq!(
2013            streaming_route(
2014                &http::Method::PATCH,
2015                "/v2/my-repo/blobs/uploads/abcd1234",
2016                &headers,
2017                &HashMap::new(),
2018            ),
2019            Some(("ecr", "")),
2020        );
2021        assert_eq!(
2022            streaming_route(
2023                &http::Method::PUT,
2024                "/v2/my-repo/blobs/uploads/abcd1234",
2025                &headers,
2026                &HashMap::new(),
2027            ),
2028            Some(("ecr", "")),
2029        );
2030    }
2031
2032    #[test]
2033    fn streaming_route_presigned_v4_s3_put() {
2034        let headers = http::HeaderMap::new();
2035        let mut query_params = HashMap::new();
2036        query_params.insert(
2037            "X-Amz-Credential".to_string(),
2038            "test/20240101/us-east-1/s3/aws4_request".to_string(),
2039        );
2040        assert_eq!(
2041            streaming_route(
2042                &http::Method::PUT,
2043                "/my-bucket/key.txt",
2044                &headers,
2045                &query_params,
2046            ),
2047            Some(("s3", "")),
2048        );
2049    }
2050
2051    #[test]
2052    fn streaming_route_non_s3_auth_header_skipped() {
2053        // Same path shape but the SigV4 service is lambda — must not
2054        // wire the streaming dispatch (Lambda has its own buffered path).
2055        let mut headers = http::HeaderMap::new();
2056        headers.insert(
2057            "authorization",
2058            "AWS4-HMAC-SHA256 Credential=test/20240101/us-east-1/lambda/aws4_request, \
2059             SignedHeaders=host, Signature=fake"
2060                .parse()
2061                .unwrap(),
2062        );
2063        assert_eq!(
2064            streaming_route(
2065                &http::Method::PUT,
2066                "/my-bucket/key.txt",
2067                &headers,
2068                &HashMap::new(),
2069            ),
2070            None,
2071        );
2072    }
2073
2074    #[test]
2075    fn streaming_route_get_skipped() {
2076        let headers = s3_sigv4_headers();
2077        assert_eq!(
2078            streaming_route(
2079                &http::Method::GET,
2080                "/my-bucket/key.txt",
2081                &headers,
2082                &HashMap::new(),
2083            ),
2084            None,
2085        );
2086    }
2087}