fix(kms): handle concurrent Vault KV2 baseline races

This commit is contained in:
overtrue
2026-08-02 18:39:42 +08:00
parent f440a14e53
commit 1bafc14a7a
2 changed files with 399 additions and 6 deletions
+241 -6
View File
@@ -21,6 +21,7 @@
//! assertions. It intentionally implements just enough HTTP/1.1 for the
//! `vaultrs` reqwest client: no keep-alive, no chunked bodies.
use std::collections::BTreeMap;
use std::sync::{Arc, Mutex};
use tokio::io::{AsyncReadExt, AsyncWriteExt};
@@ -67,6 +68,7 @@ pub(crate) struct ScriptedVault {
/// Base address (`http://127.0.0.1:port`) to point a Vault client at.
pub(crate) address: String,
requests: Arc<Mutex<Vec<(String, String)>>>,
kv2_state: Option<Arc<Mutex<Kv2State>>>,
}
impl ScriptedVault {
@@ -83,13 +85,16 @@ impl ScriptedVault {
tokio::spawn(async move {
let mut responses = responses.into_iter();
loop {
let Ok((mut stream, _)) = listener.accept().await else {
let Ok((stream, _)) = listener.accept().await else {
return;
};
let Some(request) = read_request(&mut stream).await else {
let Some((request_line, body, mut stream)) = read_request(stream).await else {
continue;
};
recorded.lock().expect("scripted vault request log poisoned").push(request);
recorded
.lock()
.expect("scripted vault request log poisoned")
.push((request_line, body));
let response = responses
.next()
.unwrap_or_else(|| ScriptedResponse::error(599, "scripted vault: script exhausted"));
@@ -104,7 +109,59 @@ impl ScriptedVault {
}
});
Self { address, requests }
Self {
address,
requests,
kv2_state: None,
}
}
/// Bind a small stateful KV2 responder for concurrency tests.
///
/// Unlike Self::serve, this responder evaluates CAS writes against a
/// shared in-memory record and handles connections concurrently. It models
/// only the KV2 data and metadata paths used by the rotation protocol; an
/// unknown request receives a 599 response so a test cannot silently
/// under-specify the Vault exchange.
pub(crate) async fn serve_kv2(key_path: &str, key_data: serde_json::Value) -> Self {
let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind scripted vault listener");
let address = format!("http://{}", listener.local_addr().expect("scripted vault local addr"));
let requests = Arc::new(Mutex::new(Vec::new()));
let state = Arc::new(Mutex::new(Kv2State::new(key_data)));
let recorded = Arc::clone(&requests);
let state_for_server = Arc::clone(&state);
let key_path = key_path.to_string();
tokio::spawn(async move {
loop {
let Ok((stream, _)) = listener.accept().await else {
return;
};
let recorded = Arc::clone(&recorded);
let state = Arc::clone(&state_for_server);
let key_path = key_path.clone();
tokio::spawn(async move {
let Some((request_line, body, stream)) = read_request(stream).await else {
return;
};
recorded
.lock()
.expect("scripted vault request log poisoned")
.push((request_line.clone(), body.clone()));
let response = state
.lock()
.expect("scripted KV2 state poisoned")
.respond(&key_path, &request_line, &body);
write_response(stream, response).await;
});
}
});
Self {
address,
requests,
kv2_state: Some(state),
}
}
/// The `METHOD /path` lines of every request served so far, in order.
@@ -128,13 +185,191 @@ impl ScriptedVault {
.map(|(_, body)| body.clone())
.collect()
}
/// Snapshot the in-memory KV2 state used by Self::serve_kv2.
pub(crate) fn kv2_snapshot(&self) -> Option<Kv2Snapshot> {
self.kv2_state.as_ref().map(|state| {
let state = state.lock().expect("scripted KV2 state poisoned");
Kv2Snapshot {
current_data: state.current_data.clone(),
current_secret_version: state.current_secret_version,
version_records: state.version_records.clone(),
}
})
}
}
/// State captured by the stateful KV2 responder for assertions in wiring tests.
#[derive(Debug, Clone)]
pub(crate) struct Kv2Snapshot {
pub(crate) current_data: serde_json::Value,
pub(crate) current_secret_version: u64,
pub(crate) version_records: BTreeMap<u32, serde_json::Value>,
}
#[derive(Debug)]
struct Kv2State {
current_data: serde_json::Value,
current_secret_version: u64,
history: BTreeMap<u64, serde_json::Value>,
version_records: BTreeMap<u32, serde_json::Value>,
}
impl Kv2State {
fn new(current_data: serde_json::Value) -> Self {
Self {
history: BTreeMap::from([(1, current_data.clone())]),
current_data,
current_secret_version: 1,
version_records: BTreeMap::new(),
}
}
fn respond(&mut self, key_path: &str, request_line: &str, body: &str) -> ScriptedResponse {
let Some((method, path)) = request_line.split_once(' ') else {
return ScriptedResponse::error(599, "scripted KV2: malformed request line");
};
let data_path = format!("/v1/secret/data/{key_path}");
let metadata_path = format!("/v1/secret/metadata/{key_path}");
let version_data_prefix = format!("{data_path}/versions/");
let version_metadata_path = format!("{metadata_path}/versions");
if method == "GET" && path == metadata_path {
return ScriptedResponse::ok(metadata_data(self.current_secret_version));
}
if method == "LIST" && path == version_metadata_path {
let keys = self.version_records.keys().map(u32::to_string).collect::<Vec<_>>();
return ScriptedResponse::ok(serde_json::json!({ "keys": keys }));
}
if let Some(version) = path
.strip_prefix(&version_data_prefix)
.and_then(|value| value.parse::<u32>().ok())
{
return match method {
"GET" => self
.version_records
.get(&version)
.cloned()
.map(read_data)
.unwrap_or_else(|| ScriptedResponse::error(404, "not found")),
"POST" => self.create_version_record(version, body),
_ => ScriptedResponse::error(599, "scripted KV2: unsupported version request"),
};
}
if method == "GET" && path.strip_prefix(&data_path).is_some() {
let version = path
.split_once("?version=")
.and_then(|(_, value)| value.parse::<u64>().ok())
.unwrap_or(self.current_secret_version);
return self
.history
.get(&version)
.cloned()
.map(read_data)
.unwrap_or_else(|| ScriptedResponse::error(404, "not found"));
}
if method == "POST" && path == data_path {
return self.write_current_record(body);
}
ScriptedResponse::error(599, "scripted KV2: unexpected request")
}
fn create_version_record(&mut self, version: u32, body: &str) -> ScriptedResponse {
let body: serde_json::Value = match serde_json::from_str(body) {
Ok(body) => body,
Err(_) => return ScriptedResponse::error(400, "invalid JSON"),
};
if body["options"]["cas"].as_u64() != Some(0) {
return ScriptedResponse::error(400, "version record requires create-only CAS");
}
if self.version_records.contains_key(&version) {
return ScriptedResponse::error(400, CAS_CONFLICT_MESSAGE);
}
let Some(data) = body.get("data").cloned() else {
return ScriptedResponse::error(400, "missing data");
};
self.version_records.insert(version, data);
write_ack(1)
}
fn write_current_record(&mut self, body: &str) -> ScriptedResponse {
let body: serde_json::Value = match serde_json::from_str(body) {
Ok(body) => body,
Err(_) => return ScriptedResponse::error(400, "invalid JSON"),
};
if body["options"]["cas"].as_u64() != Some(self.current_secret_version) {
return ScriptedResponse::error(400, CAS_CONFLICT_MESSAGE);
}
let Some(data) = body.get("data").cloned() else {
return ScriptedResponse::error(400, "missing data");
};
self.current_secret_version += 1;
self.current_data = data.clone();
self.history.insert(self.current_secret_version, data);
write_ack(self.current_secret_version)
}
}
const CAS_CONFLICT_MESSAGE: &str = "check-and-set parameter did not match the current version";
fn metadata_data(current_version: u64) -> serde_json::Value {
serde_json::json!({
"cas_required": false,
"created_time": "2026-01-01T00:00:00Z",
"current_version": current_version,
"delete_version_after": "0s",
"max_versions": 0,
"oldest_version": 0,
"updated_time": "2026-01-01T00:00:00Z",
"custom_metadata": null,
"versions": {},
})
}
fn read_data(data: serde_json::Value) -> ScriptedResponse {
ScriptedResponse::ok(serde_json::json!({
"data": data,
"metadata": {
"created_time": "2026-01-01T00:00:00Z",
"deletion_time": "",
"custom_metadata": null,
"destroyed": false,
"version": 1,
},
}))
}
fn write_ack(version: u64) -> ScriptedResponse {
ScriptedResponse::ok(serde_json::json!({
"created_time": "2026-01-01T00:00:00Z",
"custom_metadata": null,
"deletion_time": "",
"destroyed": false,
"version": version,
}))
}
async fn write_response(mut stream: TcpStream, response: ScriptedResponse) {
if let ScriptedResponse::Http { status, body } = response {
let payload = format!(
"HTTP/1.1 {status} Scripted\r\ncontent-type: application/json\r\ncontent-length: {}\r\nconnection: close\r\n\r\n{body}",
body.len(),
);
let _ = stream.write_all(payload.as_bytes()).await;
let _ = stream.shutdown().await;
}
}
/// Read one HTTP/1.1 request (head plus content-length body) and return its
/// `METHOD /path` line together with the body. Draining the body before
/// responding keeps the client from seeing a connection reset while it is
/// still writing.
async fn read_request(stream: &mut TcpStream) -> Option<(String, String)> {
async fn read_request(mut stream: TcpStream) -> Option<(String, String, TcpStream)> {
let mut buffer = Vec::new();
let mut chunk = [0u8; 4096];
let head_end = loop {
@@ -178,5 +413,5 @@ async fn read_request(stream: &mut TcpStream) -> Option<(String, String)> {
}
body.truncate(content_length);
Some((format!("{method} {path}"), String::from_utf8_lossy(&body).into_owned()))
Some((format!("{method} {path}"), String::from_utf8_lossy(&body).into_owned(), stream))
}
+158
View File
@@ -696,8 +696,15 @@ impl VaultKmsClient {
async fn ensure_version_history_consistent(&self, key_id: &str, key_data: &VaultKeyData) -> Result<()> {
let recorded = self.recorded_version_numbers(key_id).await?;
// A first-rotation record set (the current v1 record and, at most, the
// next v2 record) can be in flight while another caller is between its
// create-only writes and the baseline CAS. The material check below
// validates the v1 record before adopting it; a later key with a
// missing persisted baseline still indicates a lost baseline and fails
// closed.
if key_data.baseline_version.is_none()
&& let Some(oldest_recorded) = recorded.iter().min().copied()
&& !(key_data.version == 1 && recorded.iter().all(|version| *version <= 2))
{
warn!(
key_id,
@@ -1895,6 +1902,19 @@ mod tests {
(vault, client)
}
async fn scripted_kv2_client(key_data: &VaultKeyData) -> (ScriptedVault, VaultKmsClient) {
let vault = ScriptedVault::serve_kv2(
"rustfs/kms/keys/wired-key",
serde_json::to_value(key_data).expect("serialize scripted KV2 key"),
)
.await;
let (vault_config, kms_config) = scripted_configs(&vault.address);
let client = VaultKmsClient::new(vault_config, &kms_config)
.await
.expect("scripted Vault client");
(vault, client)
}
fn healthy_key_data() -> VaultKeyData {
VaultKeyData {
algorithm: "AES_256".to_string(),
@@ -3223,6 +3243,100 @@ mod tests {
);
}
/// Concurrent rotations use KV2 create-only records and a CAS pointer
/// switch. The stateful scripted Vault applies those preconditions to real
/// HTTP requests, so this test proves committed versions are unique and
/// contiguous instead of only checking that several calls returned.
#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
async fn wired_concurrent_kv2_rotations_commit_unique_monotonic_versions() {
use std::collections::HashSet;
use std::sync::Arc;
const ATTEMPTS: usize = 8;
let (vault, client) = scripted_kv2_client(&healthy_key_data()).await;
let client = Arc::new(client);
let barrier = Arc::new(tokio::sync::Barrier::new(ATTEMPTS));
let tasks: Vec<_> = (0..ATTEMPTS)
.map(|_| {
let client = Arc::clone(&client);
let barrier = Arc::clone(&barrier);
tokio::spawn(async move {
barrier.wait().await;
client.rotate_key("wired-key", None).await
})
})
.collect();
let mut committed_versions = Vec::new();
let mut errors = Vec::new();
for task in tasks {
match task.await.expect("join concurrent rotation task") {
Ok(result) => committed_versions.push(result.version),
Err(error) => errors.push(error),
}
}
assert!(
!committed_versions.is_empty(),
"at least one concurrent rotation must commit; errors: {errors:?}"
);
assert!(
errors.iter().all(
|error| matches!(error, KmsError::InvalidOperation { message } if message.contains("Concurrent modification"))
),
"a lost CAS race must be the only expected failure: {errors:?}"
);
let mut sorted_versions = committed_versions.clone();
sorted_versions.sort_unstable();
let unique_versions: HashSet<_> = sorted_versions.iter().copied().collect();
assert_eq!(
unique_versions.len(),
sorted_versions.len(),
"concurrent rotations must never return a version twice: {committed_versions:?}"
);
let successful_rotations = u32::try_from(sorted_versions.len()).expect("test attempts fit u32");
let current_version = 1u32 + successful_rotations;
assert_eq!(
sorted_versions,
(2..=current_version).collect::<Vec<_>>(),
"committed versions must form one monotonic sequence: {sorted_versions:?}"
);
let snapshot = vault.kv2_snapshot().expect("stateful KV2 snapshot");
let persisted_version = snapshot.current_data["version"]
.as_u64()
.and_then(|version| u32::try_from(version).ok())
.expect("current KV2 record must carry a u32 key version");
assert_eq!(persisted_version, current_version);
assert!(
snapshot.current_secret_version >= u64::from(current_version),
"the KV2 secret version must advance with each committed pointer switch"
);
assert_eq!(
snapshot.version_records.keys().copied().collect::<Vec<_>>(),
(1..=current_version).collect::<Vec<_>>(),
"every committed KMS version must have exactly one immutable record"
);
let mut materials = HashSet::new();
for (version, record) in &snapshot.version_records {
let material = record["encrypted_key_material"]
.as_str()
.expect("version record must carry encrypted key material");
assert!(materials.insert(material), "version {version} reuses another version's material");
}
assert_eq!(
snapshot.current_data["encrypted_key_material"],
snapshot
.version_records
.get(&current_version)
.expect("current version record")["encrypted_key_material"],
"the top-level fast path must match the current immutable version record"
);
}
/// The tags write-back after a create is a check-and-set read-modify-write
/// that carries the key material over from the freshly read record.
#[tokio::test]
@@ -3465,6 +3579,50 @@ mod tests {
);
}
/// A concurrent first rotation may leave both immutable records visible
/// before this caller observes the baseline CAS. That transient state must
/// continue through the create-only/adopt path rather than being reported
/// as a permanently lost baseline.
#[tokio::test]
async fn wired_rotate_recovers_concurrent_first_rotation_without_baseline() {
let key_data = healthy_key_data();
let record_v2 = VaultKeyVersionRecord {
version: 2,
encrypted_key_material: rotated_material(),
created_at: Zoned::now(),
};
let (vault, client) = scripted_client(vec![
ScriptedResponse::ok(kv2_metadata_read_data(1)),
ScriptedResponse::ok(kv2_read_data(&key_data)),
ScriptedResponse::ok(serde_json::json!({ "keys": ["1", "2"] })),
// Another rotation already created the baseline record.
ScriptedResponse::error(400, CAS_CONFLICT_MESSAGE),
ScriptedResponse::ok(kv2_read_version_record_data(&VaultKeyVersionRecord {
version: 1,
encrypted_key_material: key_data.encrypted_key_material.clone(),
created_at: key_data.created_at.clone(),
})),
// Persist the baseline pointer, then adopt the already-created v2.
ScriptedResponse::ok(kv2_write_ack()),
ScriptedResponse::error(400, CAS_CONFLICT_MESSAGE),
ScriptedResponse::ok(kv2_read_version_record_data(&record_v2)),
ScriptedResponse::ok(kv2_write_ack()),
])
.await;
let rotated = client
.rotate_key("wired-key", None)
.await
.expect("an in-flight first rotation must be recoverable");
assert_eq!(rotated.version, 2);
assert_eq!(
parse_write_body(&vault.request_bodies()[8])["data"]["encrypted_key_material"],
serde_json::json!(record_v2.encrypted_key_material),
"the pointer must adopt the immutable v2 material"
);
}
/// The mixed-version corruption path: a node older than versioned rotation
/// performed a lifecycle write, which rewrites the whole key record and
/// silently drops the `baseline_version` it does not know. Version records