Merge branch 'main' into fix/b5-t5-c2-preflight-lock

The site replication add preflight probed peer sites serially while
holding the process-wide lifecycle lock, so k unreachable sites held the
lock for k peer-request timeouts, and every concurrent
add/remove/refresh waited on an unbounded lock acquire for the whole
time. Probe all sites concurrently (matching the file's other peer
fan-outs) so k unreachable sites cost roughly one timeout, and bound the
lifecycle lock acquire at 30s, returning a retryable 503 to waiters
instead of hanging indefinitely.

Regression tests pin the preflight fan-out concurrency, the bounded
acquire's 503, and the 10s/3s peer client timeout constants.

Refs rustfs/backlog#1952, rustfs/backlog#1946, rustfs/backlog#1889

.github/workflows/nightly-gnu.yml

@@ -39,10 +39,11 @@ jobs:
     env:
       FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
     steps:
-      - name: Checkout repository
+      - name: Checkout main branch
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
         with:
           persist-credentials: false
+          ref: main
 
       - name: Setup Rust environment
         uses: ./.github/actions/setup
@@ -88,10 +89,11 @@ jobs:
       # either casing.
       NO_PROXY: 127.0.0.1,localhost
     steps:
-      - name: Checkout repository
+      - name: Checkout main branch
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
         with:
           persist-credentials: false
+          ref: main
 
       - name: Setup Rust environment
         uses: ./.github/actions/setup
@@ -176,10 +178,11 @@ jobs:
       FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
       NO_PROXY: 127.0.0.1,localhost
     steps:
-      - name: Checkout repository
+      - name: Checkout main branch
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
         with:
           persist-credentials: false
+          ref: main
 
       - name: Setup Rust environment
         uses: ./.github/actions/setup

crates/ecstore/src/core/pools.rs

@@ -462,16 +462,19 @@ pub(crate) fn local_decommission_queue_prefix(endpoints: &EndpointServerPools, i
     Ok(local)
 }
 
-fn resumable_decommission_queue_indices(meta: &PoolMeta) -> Vec<usize> {
+fn first_resumable_decommission_queue_indices(meta: &PoolMeta) -> Vec<usize> {
     let mut indices = Vec::new();
     for (idx, pool) in meta.pools.iter().enumerate() {
         if let Some(decommission) = &pool.decommission {
             if !decommission.has_decommission_state() {
                 continue;
             }
-            if decommission.complete || decommission.failed || decommission.canceled {
+            if decommission.complete {
                 continue;
             }
+            if decommission.failed || decommission.canceled {
+                break;
+            }
             indices.push(idx);
         }
     }
@@ -2403,10 +2406,24 @@ impl PoolMeta {
     }
 
     pub fn return_resumable_pools(&self) -> Vec<PoolStatus> {
-        resumable_decommission_queue_indices(self)
+        let mut new_pools = Vec::new();
-            .into_iter()
+        for pool in &self.pools {
-            .map(|idx| self.pools[idx].clone())
+            if let Some(decommission) = &pool.decommission {
-            .collect()
+                if !decommission.has_decommission_state() {
+                    continue;
+                }
+                if decommission.complete || decommission.failed || decommission.canceled {
+                    // Recovery is not required when:
+                    // - Decommissioning completed
+                    // - Decommissioning failed and must be explicitly restarted or cleared
+                    // - Decommissioning was cancelled
+                    continue;
+                }
+                // All other scenarios require recovery
+                new_pools.push(pool.clone());
+            }
+        }
+        new_pools
     }
 }
 
@@ -3315,7 +3332,7 @@ impl ECStore {
         let _start_guard = self.start_gate.lock().await;
         let indices = {
             let pool_meta = self.pool_meta.read().await;
-            resumable_decommission_queue_indices(&pool_meta)
+            first_resumable_decommission_queue_indices(&pool_meta)
                 .into_iter()
                 .filter(|idx| indices.contains(idx))
                 .collect::<Vec<_>>()
@@ -3360,7 +3377,7 @@ impl ECStore {
     pub async fn spawn_missing_local_decommission_routines(self: &Arc<Self>) -> Result<()> {
         let indices = {
             let pool_meta = self.pool_meta.read().await;
-            resumable_decommission_queue_indices(&pool_meta)
+            first_resumable_decommission_queue_indices(&pool_meta)
         };
         let indices = local_decommission_queue_prefix(&self.endpoints(), &indices)?;
         if indices.is_empty() {
@@ -6346,12 +6363,12 @@ mod pools_tests {
         ensure_decommission_start_keeps_active_pool, ensure_decommission_start_local_leader,
         ensure_decommission_start_pool_states, ensure_decommission_start_rebalance_meta_allowed,
         ensure_decommission_start_target_capacity, ensure_decommission_terminal_operation_supported,
-        ensure_local_decommission_pool_leaders, ensure_valid_decommission_pool_index, get_by_index, guard_decommission_cancelers,
+        ensure_local_decommission_pool_leaders, ensure_valid_decommission_pool_index, first_resumable_decommission_queue_indices,
-        has_active_decommission_canceler, is_decommission_active, is_decommission_cancel_requested,
+        get_by_index, guard_decommission_cancelers, has_active_decommission_canceler, is_decommission_active,
-        load_decommission_entry_versions, local_decommission_queue_prefix, mark_decommission_bucket_done,
+        is_decommission_cancel_requested, load_decommission_entry_versions, local_decommission_queue_prefix,
-        merge_pool_status_refresh, missing_decommission_worker_prefix, observe_decommission_terminal_reload_result,
+        mark_decommission_bucket_done, merge_pool_status_refresh, missing_decommission_worker_prefix,
-        pool_meta_has_active_decommission, require_decommission_store, reserve_decommission_start_cancelers,
+        observe_decommission_terminal_reload_result, pool_meta_has_active_decommission, require_decommission_store,
-        resolve_decommission_bucket_done_save_result, resolve_decommission_bucket_state,
+        reserve_decommission_start_cancelers, resolve_decommission_bucket_done_save_result, resolve_decommission_bucket_state,
         resolve_decommission_check_after_list_result, resolve_decommission_entry_cleanup_delete_result,
         resolve_decommission_entry_exact_versions, resolve_decommission_entry_reload_result,
         resolve_decommission_listing_worker_result, resolve_decommission_optional_bucket_config_result,
@@ -6359,9 +6376,9 @@ mod pools_tests {
         resolve_decommission_preflight_heal_result, resolve_decommission_progress_save_result,
         resolve_decommission_terminal_mark_after_error_result, resolve_decommission_terminal_mark_result,
         resolve_decommission_update_after_result, resolve_start_decommission_pool_meta_reload_result,
-        resumable_decommission_queue_indices, rollback_start_decommission_pool_meta, run_decommission_buckets_bounded,
+        rollback_start_decommission_pool_meta, run_decommission_buckets_bounded, run_decommission_listing_with_retry,
-        run_decommission_listing_with_retry, run_decommission_listing_with_retry_and_drain, run_decommission_side_effect,
+        run_decommission_listing_with_retry_and_drain, run_decommission_side_effect, should_cleanup_decommission_source_entry,
-        should_cleanup_decommission_source_entry, should_continue_decommission_queue, should_count_decommission_version_complete,
+        should_continue_decommission_queue, should_count_decommission_version_complete,
         should_preserve_decommission_canceled_state, should_reject_decommission_cancel_as_terminal,
         should_retry_decommission_cancel_reload, should_retry_decommission_listing, should_skip_canceled_decommission_routine,
         spawn_decommission_index_cancelers, split_decommission_buckets, take_and_cancel_decommission_canceler,
@@ -9370,7 +9387,7 @@ mod pools_tests {
     }
 
     #[test]
-    fn test_resumable_decommission_queue_indices_skip_terminal_predecessors() {
+    fn test_first_resumable_decommission_queue_indices_stops_at_failed_or_canceled_state() {
         let meta = PoolMeta {
             pools: vec![
                 decommission_test_pool_status(
@@ -9406,11 +9423,11 @@ mod pools_tests {
             ..Default::default()
         };
 
-        assert_eq!(resumable_decommission_queue_indices(&meta), vec![3]);
+        assert!(first_resumable_decommission_queue_indices(&meta).is_empty());
     }
 
     #[test]
-    fn test_resumable_decommission_queue_indices_preserve_active_predecessor_order() {
+    fn test_first_resumable_decommission_queue_indices_allows_after_completed_prefix() {
         let meta = PoolMeta {
             pools: vec![
                 decommission_test_pool_status(
@@ -9423,7 +9440,7 @@ mod pools_tests {
                 decommission_test_pool_status(
                     1,
                     Some(PoolDecommissionInfo {
-                        start_time: Some(OffsetDateTime::UNIX_EPOCH),
+                        queued: true,
                         ..Default::default()
                     }),
                 ),
@@ -9438,11 +9455,11 @@ mod pools_tests {
             ..Default::default()
         };
 
-        assert_eq!(resumable_decommission_queue_indices(&meta), vec![1, 2]);
+        assert_eq!(first_resumable_decommission_queue_indices(&meta), vec![1, 2]);
     }
 
     #[test]
-    fn test_runtime_and_startup_use_the_same_resumable_queue() {
+    fn test_return_resumable_pools_skips_failed_decommission() {
         let meta = PoolMeta {
             pools: vec![
                 decommission_test_pool_status(
@@ -9454,20 +9471,6 @@ mod pools_tests {
                 ),
                 decommission_test_pool_status(
                     1,
-                    Some(PoolDecommissionInfo {
-                        canceled: true,
-                        ..Default::default()
-                    }),
-                ),
-                decommission_test_pool_status(
-                    2,
-                    Some(PoolDecommissionInfo {
-                        start_time: Some(OffsetDateTime::UNIX_EPOCH),
-                        ..Default::default()
-                    }),
-                ),
-                decommission_test_pool_status(
-                    3,
                     Some(PoolDecommissionInfo {
                         queued: true,
                         ..Default::default()
@@ -9477,15 +9480,10 @@ mod pools_tests {
             ..Default::default()
         };
 
-        let runtime_indices = resumable_decommission_queue_indices(&meta);
+        let resumable = meta.return_resumable_pools();
-        let startup_ids = meta
-            .return_resumable_pools()
-            .into_iter()
-            .map(|pool| pool.id)
-            .collect::<Vec<_>>();
 
-        assert_eq!(runtime_indices, vec![2, 3]);
+        assert_eq!(resumable.len(), 1);
-        assert_eq!(startup_ids, vec![2, 3]);
+        assert_eq!(resumable[0].id, 1);
     }
 
     #[test]

crates/ecstore/src/io_support/bitrot.rs

@@ -784,24 +784,6 @@ pub(crate) fn create_deferred_bitrot_reader_with_stripe_handle(
 ///
 /// # Returns
 /// A Result containing the BitrotWriterWrapper or an error
-/// Size hint handed to `DiskAPI::create_file` for a bitrot-wrapped shard.
-///
-/// A known length is grown by one checksum per shard so the on-disk file size
-/// matches what the bitrot writer emits. A negative length is the
-/// unknown-size sentinel (`HashReader::SIZE_PRESERVE_LAYER`, used by SSE and
-/// compression) and must be preserved: `RemoteDisk::create_file` forwards it
-/// in the `put_file_stream` query, and the receiver only treats `size > 0` as
-/// a fixed body length when locating the authenticated trailer. Clamping it
-/// to `0` would claim an empty body and misframe the stream. `0` stays `0`
-/// because a genuinely empty object still means an empty body.
-fn bitrot_create_file_size(length: i64, shard_size: usize, checksum_algo: &HashAlgorithm) -> i64 {
-    if length <= 0 {
-        return length;
-    }
-    let length = length as usize;
-    (length.div_ceil(shard_size) * checksum_algo.size() + length) as i64
-}
-
 pub async fn create_bitrot_writer(
     is_inline_buffer: bool,
     disk: Option<&DiskStore>,
@@ -814,7 +796,12 @@ pub async fn create_bitrot_writer(
     let writer = if is_inline_buffer {
         CustomWriter::new_inline_buffer()
     } else if let Some(disk) = disk {
-        let length = bitrot_create_file_size(length, shard_size, &checksum_algo);
+        let length = if length > 0 {
+            let length = length as usize;
+            (length.div_ceil(shard_size) * checksum_algo.size() + length) as i64
+        } else {
+            0
+        };
 
         let file = disk.create_file("", volume, path, length).await?;
         #[cfg(feature = "hotpath")]
@@ -833,25 +820,6 @@ mod tests {
     use rustfs_rio::ChunkReader;
     use std::collections::VecDeque;
 
-    #[test]
-    fn bitrot_create_file_size_grows_known_length_by_checksums() {
-        // 10 bytes over 4-byte shards = 3 shards, each followed by a 32-byte hash.
-        assert_eq!(bitrot_create_file_size(10, 4, &HashAlgorithm::HighwayHash256), 10 + 3 * 32);
-        assert_eq!(bitrot_create_file_size(10, 4, &HashAlgorithm::None), 10);
-    }
-
-    #[test]
-    fn bitrot_create_file_size_keeps_empty_and_unknown_distinct() {
-        assert_eq!(bitrot_create_file_size(0, 4, &HashAlgorithm::HighwayHash256), 0);
-        // SSE/compression streams advertise SIZE_PRESERVE_LAYER (-1); the remote
-        // put_file_stream receiver relies on a non-positive size to parse the auth
-        // trailer from the stream tail, so the sentinel must survive untouched.
-        assert_eq!(
-            bitrot_create_file_size(rustfs_rio::HashReader::SIZE_PRESERVE_LAYER, 4, &HashAlgorithm::HighwayHash256),
-            rustfs_rio::HashReader::SIZE_PRESERVE_LAYER
-        );
-    }
-
     struct TestChunkReader {
         chunks: VecDeque<Bytes>,
     }

crates/kms/tests/vault_ha_failover_live.rs

@@ -16,14 +16,14 @@
 //!
 //! `scripts/test/vault_ha_kms_live.sh` owns the official Vault containers and
 //! kills the active node while this test continuously decrypts through a
-//! surviving standby. KV2 and Transit must recover after the bounded circuit
+//! surviving standby. KV2 and Transit requests must remain successful, use a
-//! interval, use a bounded number of attempts, and leave the circuit and
+//! bounded number of attempts, and leave the circuit and in-flight gauges at
-//! in-flight gauges at zero after a new leader is elected.
+//! zero after a new leader is elected.
 
 use std::collections::HashMap;
 use std::path::{Path, PathBuf};
+use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
-use std::sync::{Arc, Mutex};
 use std::time::Duration;
 
 use metrics_util::MetricKind;
@@ -43,11 +43,6 @@ const OPERATION_ATTEMPTS: &str = "rustfs_kms_backend_operation_attempts";
 const IN_FLIGHT: &str = "rustfs_kms_backend_in_flight";
 const CIRCUIT_OPEN: &str = "rustfs_kms_backend_circuit_open";
 const MAX_ATTEMPTS: u32 = 10;
-const ATTEMPT_TIMEOUT: Duration = Duration::from_secs(2);
-const HEALTHY_PROGRESS_TIMEOUT: Duration = Duration::from_secs(20);
-// The circuit remains open for 30s after five failed attempts.
-const POST_FAILOVER_PROGRESS_TIMEOUT: Duration = Duration::from_secs(35);
-const FAILOVER_ERROR_POLL_INTERVAL: Duration = Duration::from_millis(100);
 
 type MetricEntry = (
     metrics_util::CompositeKey,
@@ -69,7 +64,7 @@ fn config(backend: KmsBackend, backend_config: BackendConfig) -> KmsConfig {
         backend,
         backend_config,
         allow_insecure_dev_defaults: true,
-        timeout: ATTEMPT_TIMEOUT,
+        timeout: Duration::from_secs(2),
         retry_attempts: MAX_ATTEMPTS,
         enable_cache: false,
         ..KmsConfig::default()
@@ -169,31 +164,14 @@ fn retryable_failures(snapshot: &[MetricEntry], operation: &str) -> u64 {
         .sum()
 }
 
-async fn wait_for_count(
+async fn wait_for_count(counter: &AtomicU64, minimum: u64, description: &str) {
-    counter: &AtomicU64,
+    tokio::time::timeout(Duration::from_secs(20), async {
-    failure: &Mutex<Option<String>>,
-    minimum: u64,
-    description: &str,
-    timeout: Duration,
-) {
-    tokio::time::timeout(timeout, async {
         while counter.load(Ordering::SeqCst) < minimum {
-            if let Some(error) = failure.lock().expect("decrypt failure lock poisoned").as_ref() {
-                panic!(
-                    "{description} worker failed after {} successful decrypts: {error}",
-                    counter.load(Ordering::SeqCst)
-                );
-            }
             tokio::time::sleep(Duration::from_millis(25)).await;
         }
     })
     .await
-    .unwrap_or_else(|_| {
+    .unwrap_or_else(|_| panic!("timed out waiting for {description}"));
-        panic!(
-            "timed out after {timeout:?} waiting for {description}: completed {}, expected {minimum}",
-            counter.load(Ordering::SeqCst)
-        )
-    });
 }
 
 async fn wait_for_file(path: &Path, description: &str) {
@@ -211,8 +189,7 @@ async fn decrypt_loop<B: KmsBackendTrait + Send + Sync + 'static>(
     request: DecryptRequest,
     expected: Vec<u8>,
     completed: Arc<AtomicU64>,
-    allow_failover_errors: Arc<AtomicBool>,
+    failed: Arc<AtomicBool>,
-    failure: Arc<Mutex<Option<String>>>,
     stop: CancellationToken,
 ) {
     while !stop.is_cancelled() {
@@ -220,18 +197,8 @@ async fn decrypt_loop<B: KmsBackendTrait + Send + Sync + 'static>(
             Ok(response) if response.plaintext == expected => {
                 completed.fetch_add(1, Ordering::SeqCst);
             }
-            Ok(_) => {
+            Ok(_) | Err(_) => {
-                *failure.lock().expect("decrypt failure lock poisoned") =
+                failed.store(true, Ordering::SeqCst);
-                    Some("decrypt returned unexpected plaintext".to_string());
-                return;
-            }
-            Err(rustfs_kms::KmsError::BackendError { .. } | rustfs_kms::KmsError::OperationTimedOut { .. })
-                if allow_failover_errors.load(Ordering::SeqCst) =>
-            {
-                tokio::time::sleep(FAILOVER_ERROR_POLL_INTERVAL).await;
-            }
-            Err(error) => {
-                *failure.lock().expect("decrypt failure lock poisoned") = Some(error.to_string());
                 return;
             }
         }
@@ -329,9 +296,7 @@ async fn exercise_failover(snapshotter: &Snapshotter) {
     );
 
     let stop = CancellationToken::new();
-    let allow_failover_errors = Arc::new(AtomicBool::new(false));
+    let failed = Arc::new(AtomicBool::new(false));
-    let kv2_failure = Arc::new(Mutex::new(None));
-    let transit_failure = Arc::new(Mutex::new(None));
     let kv2_completed = Arc::new(AtomicU64::new(0));
     let transit_completed = Arc::new(AtomicU64::new(0));
     let kv2_worker = tokio::spawn(decrypt_loop(
@@ -339,8 +304,7 @@ async fn exercise_failover(snapshotter: &Snapshotter) {
         kv2_request,
         kv2_data_key.plaintext_key,
         Arc::clone(&kv2_completed),
-        Arc::clone(&allow_failover_errors),
+        Arc::clone(&failed),
-        Arc::clone(&kv2_failure),
         stop.clone(),
     ));
     let transit_worker = tokio::spawn(decrypt_loop(
@@ -348,21 +312,12 @@ async fn exercise_failover(snapshotter: &Snapshotter) {
         transit_request,
         transit_data_key.plaintext_key,
         Arc::clone(&transit_completed),
-        Arc::clone(&allow_failover_errors),
+        Arc::clone(&failed),
-        Arc::clone(&transit_failure),
         stop.clone(),
     ));
 
-    wait_for_count(&kv2_completed, &kv2_failure, 2, "two healthy KV2 decrypts", HEALTHY_PROGRESS_TIMEOUT).await;
+    wait_for_count(&kv2_completed, 2, "two healthy KV2 decrypts").await;
-    wait_for_count(
+    wait_for_count(&transit_completed, 2, "two healthy Transit decrypts").await;
-        &transit_completed,
-        &transit_failure,
-        2,
-        "two healthy Transit decrypts",
-        HEALTHY_PROGRESS_TIMEOUT,
-    )
-    .await;
-    allow_failover_errors.store(true, Ordering::SeqCst);
     std::fs::write(&marker, b"ready").expect("publish failover readiness marker");
 
     wait_for_file(&elected, "the replacement Vault leader").await;
@@ -371,39 +326,18 @@ async fn exercise_failover(snapshotter: &Snapshotter) {
 
     let kv2_after_election = kv2_completed.load(Ordering::SeqCst) + 2;
     let transit_after_election = transit_completed.load(Ordering::SeqCst) + 2;
-    wait_for_count(
+    wait_for_count(&kv2_completed, kv2_after_election, "post-failover KV2 decrypts").await;
-        &kv2_completed,
+    wait_for_count(&transit_completed, transit_after_election, "post-failover Transit decrypts").await;
-        &kv2_failure,
-        kv2_after_election,
-        "post-failover KV2 decrypts",
-        POST_FAILOVER_PROGRESS_TIMEOUT,
-    )
-    .await;
-    wait_for_count(
-        &transit_completed,
-        &transit_failure,
-        transit_after_election,
-        "post-failover Transit decrypts",
-        POST_FAILOVER_PROGRESS_TIMEOUT,
-    )
-    .await;
 
     stop.cancel();
     kv2_worker.await.expect("KV2 decrypt worker must join");
     transit_worker.await.expect("Transit decrypt worker must join");
-    assert!(
+    assert!(!failed.load(Ordering::SeqCst), "no decrypt may fail or return different plaintext");
-        kv2_failure.lock().expect("KV2 failure lock poisoned").is_none(),
-        "no KV2 decrypt may fail or return different plaintext"
-    );
-    assert!(
-        transit_failure.lock().expect("Transit failure lock poisoned").is_none(),
-        "no Transit decrypt may fail or return different plaintext"
-    );
 }
 
 #[test]
 #[ignore = "requires a real three-node Vault Raft cluster; run scripts/test/vault_ha_kms_live.sh"]
-fn vault_raft_leader_failure_recovers_kv2_and_transit_decrypts() {
+fn vault_raft_leader_failure_preserves_kv2_and_transit_decrypts() {
     let recorder = DebuggingRecorder::new();
     let snapshotter = recorder.snapshotter();
     metrics::with_local_recorder(&recorder, || {
@@ -415,6 +349,11 @@ fn vault_raft_leader_failure_recovers_kv2_and_transit_decrypts() {
     });
     let snapshot = snapshotter.snapshot().into_vec();
 
+    assert_eq!(
+        counter_value(&snapshot, OPERATIONS_TOTAL, &[("outcome", "circuit_open")]),
+        0,
+        "a bounded leader election must not open the circuit"
+    );
     assert_eq!(
         counter_value(&snapshot, OPERATIONS_TOTAL, &[("outcome", "budget_exhausted")]),
         0,

rustfs/src/admin/handlers/site_replication.rs

@@ -138,6 +138,11 @@ const SITE_REPL_RESYNC_DEFAULT_PAGE_SIZE: usize = 100;
 const SITE_REPL_RESYNC_MAX_PAGE_SIZE: usize = 1000;
 const SITE_REPLICATION_PEER_REQUEST_TIMEOUT: Duration = Duration::from_secs(10);
 const SITE_REPLICATION_PEER_CONNECT_TIMEOUT: Duration = Duration::from_secs(3);
+/// Bound on waiting for the lifecycle lock (below). 3x the peer request
+/// timeout: outlives one full peer round of a healthy concurrent lifecycle
+/// operation, while converting a holder wedged on unreachable peers into a
+/// retryable 503 for the waiter instead of an unbounded hang.
+const SITE_REPLICATION_LIFECYCLE_LOCK_TIMEOUT: Duration = Duration::from_secs(30);
 const SITE_REPLICATION_PEER_ERROR_DETAIL_LIMIT: usize = 256;
 const SITE_REPLICATION_INITIAL_SYNC_ERROR_LIMIT: usize = 32;
 const MAX_PEER_CA_CERT_PEM_SIZE: usize = 256 * 1024;
@@ -388,9 +393,17 @@ struct SiteReplicationLifecycleGuard {
 }
 
 impl SiteReplicationLifecycleGuard {
-    async fn acquire() -> Self {
+    /// Bounded acquire: a holder wedged on unreachable peers (each probe
-        Self {
+    /// costs up to [`SITE_REPLICATION_PEER_REQUEST_TIMEOUT`]) must not hang
-            _guard: SITE_REPLICATION_LIFECYCLE_LOCK.lock().await,
+    /// every other lifecycle operation indefinitely, so waiters get a
+    /// retryable 503 after [`SITE_REPLICATION_LIFECYCLE_LOCK_TIMEOUT`].
+    async fn acquire() -> S3Result<Self> {
+        match tokio::time::timeout(SITE_REPLICATION_LIFECYCLE_LOCK_TIMEOUT, SITE_REPLICATION_LIFECYCLE_LOCK.lock()).await {
+            Ok(guard) => Ok(Self { _guard: guard }),
+            Err(_) => Err(S3Error::with_message(
+                S3ErrorCode::ServiceUnavailable,
+                "another site replication lifecycle operation is in progress; retry later".to_string(),
+            )),
         }
     }
 
@@ -2128,6 +2141,27 @@ async fn remote_add_preflight_info(site: &PeerSite) -> S3Result<SiteReplicationA
     add_preflight_info_from_sr_info(site, info, idp_settings)
 }
 
+/// Preflight every site in an add request while the lifecycle lock is held.
+/// Probes run concurrently (matching the other peer fan-outs in this file):
+/// k unreachable sites cost roughly one peer request timeout, not k of them.
+/// Results (and the first error, if any) are reported in request order.
+async fn add_preflight_infos(
+    sites: &[PeerSite],
+    current_state: &SiteReplicationState,
+    local_peer: &PeerInfo,
+) -> S3Result<Vec<SiteReplicationAddPreflightInfo>> {
+    futures::future::join_all(sites.iter().map(|site| async move {
+        if same_identity_endpoint(&site.endpoint, &local_peer.endpoint) {
+            local_add_preflight_info(current_state, local_peer, site).await
+        } else {
+            remote_add_preflight_info(site).await
+        }
+    }))
+    .await
+    .into_iter()
+    .collect()
+}
+
 fn validate_add_preflight_topology(infos: &[SiteReplicationAddPreflightInfo], local_peer: &PeerInfo) -> S3Result<()> {
     let mut deployment_ids = HashSet::new();
     let mut local_seen = false;
@@ -9872,7 +9906,7 @@ impl Operation for SiteReplicationAddHandler {
         let cred = validate_site_replication_admin_request(&req, AdminAction::SiteReplicationAddAction).await?;
         reject_site_replicator_on_public_admin(&cred)?;
         let replicate_ilm_expiry = sr_add_replicate_ilm_expiry(&req.uri);
-        let lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await;
+        let lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await?;
         // Everything up to the commit below is preflight: peer probes, IAM
         // work and the join fan-out all talk to the network, so none of it may
         // run inside the state transaction. The snapshot read here is what the
@@ -9887,14 +9921,7 @@ impl Operation for SiteReplicationAddHandler {
         // inject it so the add preflight (which requires the local deployment) succeeds. No-op for `mc`.
         ensure_local_site_present(&mut sites, &local_peer);
         validate_add_sites(&sites, &local_peer)?;
-        let mut preflight_infos = Vec::with_capacity(sites.len());
+        let preflight_infos = add_preflight_infos(&sites, &current_state, &local_peer).await?;
-        for site in &sites {
-            if same_identity_endpoint(&site.endpoint, &local_peer.endpoint) {
-                preflight_infos.push(local_add_preflight_info(&current_state, &local_peer, site).await?);
-            } else {
-                preflight_infos.push(remote_add_preflight_info(site).await?);
-            }
-        }
         validate_add_preflight_topology(&preflight_infos, &local_peer)?;
         let expected_updated_at = current_state.updated_at;
         require_add_peer_tls_capability(&sites, &local_peer).await?;
@@ -10102,7 +10129,7 @@ impl Operation for SiteReplicationRemoveHandler {
     async fn call(&self, req: S3Request<Body>, _params: Params<'_, '_>) -> S3Result<S3Response<(StatusCode, Body)>> {
         let cred = validate_site_replication_admin_request(&req, AdminAction::SiteReplicationRemoveAction).await?;
         reject_site_replicator_on_public_admin(&cred)?;
-        let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await;
+        let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await?;
         // The request body is read before the bucket-op guard and the state
         // transaction: a client that stalls mid-body must hold neither the
         // state-object lock nor the write half of the bucket-op RwLock (which
@@ -10312,7 +10339,7 @@ where
     F: FnOnce(SRPeerJoinReq) -> Fut + Send + 'static,
     Fut: std::future::Future<Output = S3Result<()>> + Send + 'static,
 {
-    let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await;
+    let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await?;
     admit_peer_join_across_nodes(local_endpoint, join_req, defer_sync_state_enable, apply_iam).await
 }
 
@@ -11142,7 +11169,7 @@ impl Operation for SRPeerRemoveHandler {
     async fn call(&self, req: S3Request<Body>, _params: Params<'_, '_>) -> S3Result<S3Response<(StatusCode, Body)>> {
         validate_site_replication_admin_request(&req, AdminAction::SiteReplicationRemoveAction).await?;
         let remove_req: SRRemoveReq = read_site_replication_json(req, "", false).await?;
-        let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await;
+        let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await?;
         let _bucket_op_guard = SITE_REPLICATION_BUCKET_OP_LOCK.write().await;
         let removed_deployment_ids = update_site_replication_state(move |state| {
             if pending_endpoint_refresh(state).is_some() {
@@ -11186,7 +11213,7 @@ impl Operation for SiteReplicationResyncOpHandler {
         let operation = query.get("operation").cloned().unwrap_or_default();
         let resolved_store = object_store_from_req(&req);
         let requested_peer: PeerInfo = read_site_replication_json(req, "", false).await?;
-        let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await;
+        let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await?;
         let (peer, existing_status) = {
             let state = load_site_replication_state().await?;
             let local_peer = current_local_runtime_peer(&state);
@@ -11478,7 +11505,7 @@ impl Operation for SRRotateServiceAccountHandler {
         // mid-repair and race its own IAM write against the reconciler's
         // stale one. (The removed process mutex used to provide this
         // exclusion as a side effect.)
-        let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await;
+        let _lifecycle_guard = SiteReplicationLifecycleGuard::acquire().await?;
         let local_endpoint = site_replication_local_endpoint(&req.uri, &req.headers);
         let rotation_parent = cred.access_key.clone();
         let (pending_rotation, local_peer, previous_access_key) = update_site_replication_state_when_changed(move |state| {
@@ -14019,7 +14046,9 @@ mod tests {
     async fn test_add_bootstrap_scope_only_allows_expected_bucket_setup_until_guard_drops() {
         let token;
         {
-            let lifecycle = SiteReplicationLifecycleGuard::acquire().await;
+            let lifecycle = SiteReplicationLifecycleGuard::acquire()
+                .await
+                .expect("acquire lifecycle guard");
             let guard = SiteReplicationAddInProgressGuard::start(lifecycle, HashSet::from(["legacy-bucket".to_string()]))
                 .expect("start site replication add guard");
             token = guard.token.to_string();
@@ -14075,14 +14104,18 @@ mod tests {
     #[tokio::test]
     #[serial]
     async fn test_add_lifecycle_allows_callback_before_remove_writer() {
-        let lifecycle = SiteReplicationLifecycleGuard::acquire().await;
+        let lifecycle = SiteReplicationLifecycleGuard::acquire()
+            .await
+            .expect("acquire lifecycle guard");
         let add_guard =
             SiteReplicationAddInProgressGuard::start(lifecycle, HashSet::new()).expect("start site replication add guard");
         let (started_tx, started_rx) = tokio::sync::oneshot::channel();
         let (entered_tx, mut entered_rx) = tokio::sync::oneshot::channel();
         let remove = tokio::spawn(async move {
             let _ = started_tx.send(());
-            let _lifecycle = SiteReplicationLifecycleGuard::acquire().await;
+            let _lifecycle = SiteReplicationLifecycleGuard::acquire()
+                .await
+                .expect("acquire lifecycle guard");
             let _bucket_op = SITE_REPLICATION_BUCKET_OP_LOCK.write().await;
             let _ = entered_tx.send(());
         });
@@ -14102,6 +14135,111 @@ mod tests {
         entered_rx.await.expect("remove entered lifecycle");
     }
 
+    /// Deleting either constant (or "simplifying" the client builders to
+    /// inline values) removes the only bound on how long a lifecycle
+    /// operation can be wedged per unreachable peer (#1889 C1 / #1952 C2).
+    #[test]
+    fn test_peer_timeout_constants_bound_unreachable_peer_probes() {
+        assert_eq!(SITE_REPLICATION_PEER_REQUEST_TIMEOUT, Duration::from_secs(10));
+        assert_eq!(SITE_REPLICATION_PEER_CONNECT_TIMEOUT, Duration::from_secs(3));
+        assert!(
+            SITE_REPLICATION_LIFECYCLE_LOCK_TIMEOUT >= SITE_REPLICATION_PEER_REQUEST_TIMEOUT,
+            "a waiter must not give up before the holder's single wedged peer probe can finish"
+        );
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial]
+    async fn test_lifecycle_guard_acquire_times_out_with_retryable_503() {
+        let holder = SiteReplicationLifecycleGuard::acquire().await.expect("first acquire");
+        let err =
+            match tokio::time::timeout(SITE_REPLICATION_LIFECYCLE_LOCK_TIMEOUT * 2, SiteReplicationLifecycleGuard::acquire())
+                .await
+                .expect("bounded acquire must not hang while the lock is held")
+            {
+                Ok(_) => panic!("acquire while the lock is held should time out"),
+                Err(err) => err,
+            };
+        assert_eq!(err.code(), &S3ErrorCode::ServiceUnavailable);
+
+        drop(holder);
+        tokio::time::timeout(Duration::from_secs(1), SiteReplicationLifecycleGuard::acquire())
+            .await
+            .expect("acquire after release must not wait")
+            .expect("acquire after release");
+    }
+
+    #[derive(Clone)]
+    struct PreflightFanoutTestState {
+        metainfo_barrier: Arc<tokio::sync::Barrier>,
+    }
+
+    async fn preflight_fanout_test_handler(State(state): State<PreflightFanoutTestState>, uri: Uri) -> (StatusCode, String) {
+        if uri.path().ends_with("/site-replication/metainfo") {
+            state.metainfo_barrier.wait().await;
+        }
+        (StatusCode::OK, "{}".to_string())
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn test_add_preflight_probes_sites_concurrently() {
+        temp_env::async_with_vars(
+            [(ALLOW_LOOPBACK_REPLICATION_TARGET_ENV, Some("true"))],
+            add_preflight_probes_sites_concurrently_inner(),
+        )
+        .await;
+    }
+
+    async fn add_preflight_probes_sites_concurrently_inner() {
+        const REMOTE_SITES: usize = 3;
+        let listener = match TcpListener::bind("127.0.0.1:0").await {
+            Ok(listener) => listener,
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return,
+            Err(err) => panic!("bind preflight test server: {err}"),
+        };
+        let endpoint = format!("http://{}", listener.local_addr().expect("preflight test address"));
+        let state = PreflightFanoutTestState {
+            metainfo_barrier: Arc::new(tokio::sync::Barrier::new(REMOTE_SITES)),
+        };
+        let server = tokio::spawn(async move {
+            axum::serve(listener, Router::new().fallback(any(preflight_fanout_test_handler)).with_state(state))
+                .await
+                .expect("serve preflight test requests");
+        });
+
+        let sites: Vec<PeerSite> = (0..REMOTE_SITES)
+            .map(|index| PeerSite {
+                name: format!("site-{index}"),
+                endpoint: endpoint.clone(),
+                access_key: "test-access".to_string(),
+                secret_key: "test-secret".to_string(),
+                ..Default::default()
+            })
+            .collect();
+        let local_peer = PeerInfo {
+            deployment_id: "local".to_string(),
+            endpoint: "http://192.0.2.1:9000".to_string(),
+            ..Default::default()
+        };
+        let current_state = SiteReplicationState::default();
+
+        // Each site's metainfo request parks on a barrier that only releases
+        // once every site's request has arrived: serial probing never sends
+        // the second request and dies on the peer request timeout, so
+        // finishing well inside that timeout proves the probes overlap —
+        // which is what caps k unreachable sites at one timeout, not k.
+        let infos = tokio::time::timeout(
+            SITE_REPLICATION_PEER_REQUEST_TIMEOUT / 2,
+            add_preflight_infos(&sites, &current_state, &local_peer),
+        )
+        .await
+        .expect("preflight probes must fan out concurrently, not serially")
+        .expect("preflight infos");
+        assert_eq!(infos.len(), REMOTE_SITES);
+        server.abort();
+    }
+
     #[test]
     fn test_merge_add_sites_propagates_replicate_ilm_expiry() {
         let state = merge_add_sites(

scripts/test/vault_ha_kms_live.sh

@@ -241,7 +241,7 @@ env \
   RUSTFS_TEST_VAULT_FAILOVER_MARKER="$MARKER" \
   RUSTFS_TEST_VAULT_OLD_LEADER="$OLD_LEADER" \
   cargo test -p rustfs-kms --test vault_ha_failover_live \
-    vault_raft_leader_failure_recovers_kv2_and_transit_decrypts -- \
+    vault_raft_leader_failure_preserves_kv2_and_transit_decrypts -- \
     --ignored --nocapture --test-threads=1 &
 TEST_PID=$!