Merge branch 'main' into cxymds/fix-1926-owner-selection

crates/config/src/constants/runtime.rs

@@ -57,6 +57,13 @@ pub const DEFAULT_MAX_IO_EVENTS_PER_TICK: usize = 1024;
 pub const DEFAULT_EVENT_INTERVAL: u32 = 61;
 pub const DEFAULT_RNG_SEED: Option<u64> = None; // None means random
 
+/// Dedicated blocking thread pool for fsync/fdatasync operations.
+/// When > 1, fsync operations are isolated from the main blocking pool to
+/// prevent device-bound fsync from starving read operations (pread/stat/open).
+/// Default 0 means auto (no isolation, use main runtime).
+pub const ENV_FSYNC_BLOCKING_THREADS: &str = "RUSTFS_RUNTIME_FSYNC_BLOCKING_THREADS";
+pub const DEFAULT_FSYNC_BLOCKING_THREADS: usize = 0;
+
 // Dial9 Tokio Telemetry Default values
 pub const DEFAULT_RUNTIME_DIAL9_ENABLED: bool = false; // Disabled by default
 pub const DEFAULT_RUNTIME_DIAL9_OUTPUT_DIR: &str = "/var/log/rustfs/telemetry";

crates/ecstore/src/disk/os.rs

@@ -315,7 +315,7 @@ pub async fn fsync_dir(dir: impl AsRef<Path>) -> io::Result<()> {
     #[cfg(unix)]
     {
         let dir = dir.as_ref().to_path_buf();
-        tokio::task::spawn_blocking(move || fsync_dir_std(dir)).await?
+        fsync_spawn_blocking(move || fsync_dir_std(dir)).await?
     }
 
     #[cfg(not(unix))]
@@ -683,7 +683,7 @@ async fn fsync_open_dst_dir_group(group: &DstDirFsyncGroup) -> io::Result<()> {
     #[cfg(test)]
     let dir = group.dir.clone();
     let dir_file = group.dir_file.clone();
-    tokio::task::spawn_blocking(move || {
+    fsync_spawn_blocking(move || {
         #[cfg(test)]
         {
             if let Some(kind) = fsync_dir_recorder::take_grouped_failure(&dir) {
@@ -1080,6 +1080,44 @@ const TEST_GLOBAL_FILE_SYNCS: usize = 64;
 
 static FILE_SYNC_PERMITS: LazyLock<Semaphore> = LazyLock::new(|| Semaphore::new(global_file_sync_limit()));
 static DISK_FILE_SYNC_LIMITERS: LazyLock<Mutex<HashMap<PathBuf, Weak<Semaphore>>>> = LazyLock::new(|| Mutex::new(HashMap::new()));
+
+/// Dedicated tokio runtime for fsync/fdatasync blocking operations. When
+/// configured with >1 threads, isolates device-bound fsync from the main
+/// blocking pool so reads (pread/stat/open) are not starved. `None` means
+/// fall back to the main runtime (zero behavior change).
+static FSYNC_RUNTIME: LazyLock<Option<tokio::runtime::Runtime>> = LazyLock::new(|| {
+    let threads =
+        rustfs_utils::get_env_usize(rustfs_config::ENV_FSYNC_BLOCKING_THREADS, rustfs_config::DEFAULT_FSYNC_BLOCKING_THREADS);
+    if threads <= 1 {
+        return None;
+    }
+    let mut builder = tokio::runtime::Builder::new_multi_thread();
+    builder
+        .worker_threads(num_cpus::get().min(8))
+        .max_blocking_threads(threads)
+        .thread_name("rustfs-fsync")
+        .thread_stack_size(512 * 1024)
+        .enable_all();
+    match builder.build() {
+        Ok(rt) => {
+            tracing::info!(threads, "fsync dedicated blocking pool enabled");
+            Some(rt)
+        }
+        Err(err) => {
+            tracing::warn!(%err, "failed to build fsync runtime, falling back to main pool");
+            None
+        }
+    }
+});
+
+/// Spawn a blocking task on the fsync-dedicated runtime if configured,
+/// otherwise fall back to the main tokio blocking pool.
+fn fsync_spawn_blocking<T: Send + 'static>(f: impl FnOnce() -> T + Send + 'static) -> tokio::task::JoinHandle<T> {
+    match FSYNC_RUNTIME.as_ref() {
+        Some(rt) => rt.spawn_blocking(f),
+        None => tokio::task::spawn_blocking(f),
+    }
+}
 static DISK_VOLUME_MUTATION_LOCKS: LazyLock<Mutex<HashMap<PathBuf, Weak<RwLock<()>>>>> =
     LazyLock::new(|| Mutex::new(HashMap::new()));
 type NamespaceMutationLock = AsyncMutex<()>;
@@ -1217,7 +1255,7 @@ where
     F: FnOnce() -> io::Result<T> + Send + 'static,
 {
     let (disk_permit, global_permit) = acquire_file_sync_permits(disk_permits).await?;
-    let result = tokio::task::spawn_blocking(move || {
+    let result = fsync_spawn_blocking(move || {
         let _disk_permit = disk_permit;
         work()
     })
@@ -2146,7 +2184,7 @@ async fn run_blocking_namespace_file_sync_operation_with_global<T: Send + 'stati
         wait_started,
     );
     let disk_permit = admission.disk_permit.clone();
-    let result = tokio::task::spawn_blocking(move || {
+    let result = fsync_spawn_blocking(move || {
         let _lease = lease;
         let _disk_permit = disk_permit;
         operation()

crates/ecstore/src/runtime/instance.rs

@@ -160,10 +160,6 @@ pub struct InstanceContext {
     /// workers (scanner/heal/tier/lifecycle) without touching another instance.
     /// Replaces the process-global cancel-token static.
     background_cancel_token: OnceLock<CancellationToken>,
-    /// Serializes decommission data-movement operations with cancellation and
-    /// a subsequent restart. Readers are held across one object side effect;
-    /// the transition path takes the writer after cancelling the routine.
-    decommission_operation_gate: Arc<RwLock<()>>,
     /// Resolves object-encryption material at the application boundary.
     object_encryption_resolver: OnceLock<Arc<dyn ObjectEncryptionResolver>>,
     tier_delete_journal_recovery_stores: std::sync::Mutex<HashSet<Uuid>>,
@@ -204,7 +200,6 @@ impl InstanceContext {
             local_disk_set_drives: Arc::new(RwLock::new(Vec::new())),
             bucket_metadata_sys: std::sync::Mutex::new(None),
             background_cancel_token: OnceLock::new(),
-            decommission_operation_gate: Arc::new(RwLock::new(())),
             object_encryption_resolver: OnceLock::new(),
             tier_delete_journal_recovery_stores: std::sync::Mutex::new(HashSet::new()),
             transition_transaction_recovery_stores: std::sync::Mutex::new(HashSet::new()),
@@ -223,10 +218,6 @@ impl InstanceContext {
         self.lock_manager.clone()
     }
 
-    pub(crate) fn decommission_operation_gate(&self) -> Arc<RwLock<()>> {
-        Arc::clone(&self.decommission_operation_gate)
-    }
-
     /// Install the application-owned object-encryption resolver once.
     pub fn set_object_encryption_resolver(
         &self,

crates/ecstore/src/store/heal.rs

@@ -297,10 +297,16 @@ impl ECStore {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::bucket::metadata_sys;
     use crate::core::pools::{PoolDecommissionInfo, PoolStatus};
-    use crate::disk::{DiskOption, format::FormatV3, new_disk};
-    use crate::layout::endpoints::{Endpoints, PoolEndpoints};
+    use crate::disk::{DeleteOptions, DiskOption, format::FormatV3, new_disk};
+    use crate::layout::endpoints::{EndpointServerPools, Endpoints, PoolEndpoints};
+    use crate::runtime::instance::InstanceContext;
+    use crate::storage_api_contracts::bucket::{BucketOperations, MakeBucketOptions};
+    use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations};
     use crate::store::init_format::{load_format_erasure, save_format_file};
+    use crate::store::init_local_disks_with_instance_ctx;
+    use tokio_util::sync::CancellationToken;
 
     async fn minimal_heal_pool(pool_idx: usize) -> Arc<Sets> {
         let format = FormatV3::new(1, 1);
@@ -347,6 +353,51 @@ mod tests {
         }
     }
 
+    async fn multi_pool_heal_store() -> (tempfile::TempDir, Arc<ECStore>, CancellationToken) {
+        let temp_dir = tempfile::tempdir().expect("multi-pool heal test directory should be created");
+        let mut pool_endpoints = Vec::new();
+        for pool_index in 0..2 {
+            let mut endpoints = Vec::new();
+            for disk_index in 0..4 {
+                let disk_path = temp_dir.path().join(format!("pool{pool_index}-disk{disk_index}"));
+                tokio::fs::create_dir_all(&disk_path)
+                    .await
+                    .expect("multi-pool heal test disk should be created");
+                let mut endpoint = Endpoint::try_from(disk_path.to_str().expect("disk path should be utf8"))
+                    .expect("test endpoint should parse");
+                endpoint.set_pool_index(pool_index);
+                endpoint.set_set_index(0);
+                endpoint.set_disk_index(disk_index);
+                endpoints.push(endpoint);
+            }
+            pool_endpoints.push(PoolEndpoints {
+                legacy: false,
+                set_count: 1,
+                drives_per_set: 4,
+                endpoints: Endpoints::from(endpoints),
+                cmd_line: format!("heal-owner-pool-{pool_index}"),
+                platform: "test".to_string(),
+            });
+        }
+
+        let endpoint_pools = EndpointServerPools::from(pool_endpoints);
+        let instance_ctx = Arc::new(InstanceContext::new());
+        init_local_disks_with_instance_ctx(&instance_ctx, endpoint_pools.clone())
+            .await
+            .expect("multi-pool local disks should initialize");
+        let shutdown = CancellationToken::new();
+        let store = ECStore::new_with_instance_ctx(
+            "127.0.0.1:0".parse().expect("test address should parse"),
+            endpoint_pools,
+            shutdown.clone(),
+            instance_ctx,
+        )
+        .await
+        .expect("multi-pool test store should initialize");
+        metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        (temp_dir, store, shutdown)
+    }
+
     #[tokio::test]
     async fn heal_object_pool_scope_selects_only_requested_pool() {
         let store = minimal_heal_store().await;
@@ -506,6 +557,204 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn unscoped_heal_object_suspended_owner_semantics() {
+        let (_temp_dir, store, shutdown) = multi_pool_heal_store().await;
+        let bucket = format!("heal-owner-{}", Uuid::new_v4().simple());
+        let active_object = "active-owner";
+        let suspended_only_object = "suspended-only";
+        let duplicate_object = "duplicate-owner";
+        let marker_object = "marker-owner";
+        let quorum_object = "quorum-owner";
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("bucket should be created in all pools");
+
+        let mut active_reader = PutObjReader::from_vec(b"active owner".to_vec());
+        store.pools[0]
+            .put_object(&bucket, active_object, &mut active_reader, &ObjectOptions::default())
+            .await
+            .expect("active owner object should be written");
+        let active_disks = store.pools[0].disk_set[0].disks.read().await.clone();
+        let missing_active_disk = active_disks[0].clone().expect("active disk should be online");
+        missing_active_disk
+            .delete(
+                &bucket,
+                active_object,
+                DeleteOptions {
+                    recursive: true,
+                    immediate: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("active owner shard should be removed for repair");
+        assert!(
+            missing_active_disk.read_xl(&bucket, active_object, false).await.is_err(),
+            "the active owner fixture must start with one missing metadata copy"
+        );
+
+        let mut suspended_reader = PutObjReader::from_vec(b"suspended owner".to_vec());
+        store.pools[1]
+            .put_object(&bucket, suspended_only_object, &mut suspended_reader, &ObjectOptions::default())
+            .await
+            .expect("suspended owner object should be written");
+        for (pool_index, mod_time) in [1_i64, 2_i64].into_iter().enumerate() {
+            let mut duplicate_reader = PutObjReader::from_vec(format!("duplicate-pool-{pool_index}").into_bytes());
+            store.pools[pool_index]
+                .put_object(
+                    &bucket,
+                    duplicate_object,
+                    &mut duplicate_reader,
+                    &ObjectOptions {
+                        mod_time: Some(OffsetDateTime::UNIX_EPOCH + time::Duration::seconds(mod_time)),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("duplicate owner object should be written");
+        }
+        let history_version = Uuid::new_v4();
+        let mut history_reader = PutObjReader::from_vec(b"marker history".to_vec());
+        store.pools[0]
+            .put_object(
+                &bucket,
+                marker_object,
+                &mut history_reader,
+                &ObjectOptions {
+                    versioned: true,
+                    version_id: Some(history_version.to_string()),
+                    mod_time: Some(OffsetDateTime::UNIX_EPOCH + time::Duration::seconds(1)),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("versioned marker history should be written");
+        store.pools[0]
+            .delete_object(
+                &bucket,
+                marker_object,
+                ObjectOptions {
+                    versioned: true,
+                    mod_time: Some(OffsetDateTime::UNIX_EPOCH + time::Duration::seconds(2)),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("delete marker should be written");
+        let mut quorum_reader = PutObjReader::from_vec(b"quorum boundary".to_vec());
+        store.pools[0]
+            .put_object(&bucket, quorum_object, &mut quorum_reader, &ObjectOptions::default())
+            .await
+            .expect("quorum boundary object should be written");
+        {
+            let mut pool_meta = store.pool_meta.write().await;
+            let mut next = PoolMeta::new(&store.pools, &pool_meta);
+            next.pools[1].decommission = Some(PoolDecommissionInfo {
+                start_time: Some(OffsetDateTime::UNIX_EPOCH),
+                ..Default::default()
+            });
+            *pool_meta = next;
+        }
+
+        let (_, duplicate_owner) = store
+            .get_latest_object_info_with_idx(&bucket, duplicate_object, &ObjectOptions::default())
+            .await
+            .expect("duplicate owner should resolve");
+        assert_eq!(duplicate_owner, 1, "latest duplicate must win when all pools are eligible");
+        let (_, active_duplicate_owner) = store
+            .get_latest_object_info_with_idx(
+                &bucket,
+                duplicate_object,
+                &ObjectOptions {
+                    skip_decommissioned: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("active duplicate owner should resolve");
+        assert_eq!(
+            active_duplicate_owner, 0,
+            "suspended duplicate must be excluded from active owner selection"
+        );
+        let (marker_info, marker_owner) = store
+            .get_latest_object_info_with_idx(
+                &bucket,
+                marker_object,
+                &ObjectOptions {
+                    skip_decommissioned: true,
+                    versioned: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("latest delete marker should resolve");
+        assert_eq!(marker_owner, 0);
+        assert!(marker_info.delete_marker, "latest version must preserve delete-marker semantics");
+
+        let (active_result, active_err) = store
+            .handle_heal_object(&bucket, active_object, "", &HealOpts::default())
+            .await
+            .expect("unscoped active-owner heal should complete");
+        assert_eq!(active_result.object, active_object);
+        assert!(active_err.is_none(), "active owner must be selected even with a suspended pool");
+        assert!(
+            missing_active_disk.read_xl(&bucket, active_object, false).await.is_ok(),
+            "active owner heal must write the missing disk metadata: result={active_result:?}, err={active_err:?}"
+        );
+        assert!(
+            store.pools[1]
+                .get_object_info(&bucket, active_object, &ObjectOptions::default())
+                .await
+                .is_err(),
+            "the suspended pool must not be written for an active-owner object"
+        );
+
+        let (suspended_result, suspended_err) = store
+            .handle_heal_object(&bucket, suspended_only_object, "", &HealOpts::default())
+            .await
+            .expect("unscoped suspended-only heal should return a terminal result");
+        assert!(suspended_result.object.is_empty());
+        assert!(matches!(suspended_err, Some(Error::FileNotFound)));
+        assert!(
+            store.pools[1]
+                .get_object_info(&bucket, suspended_only_object, &ObjectOptions::default())
+                .await
+                .is_ok(),
+            "suspended-only data must remain untouched when unscoped heal reports absent"
+        );
+
+        let (_, explicit_err) = store
+            .handle_heal_object(
+                &bucket,
+                suspended_only_object,
+                "",
+                &HealOpts {
+                    pool: Some(1),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("explicit suspended-owner heal should return a mapped error");
+        assert!(matches!(explicit_err, Some(Error::SlowDown)));
+
+        let original_quorum_disks = store.pools[0].disk_set[0].disks.read().await.clone();
+        let surviving_quorum_disk = original_quorum_disks[3].clone();
+        *store.pools[0].disk_set[0].disks.write().await = vec![None, None, None, surviving_quorum_disk];
+        let (_, quorum_err) = store
+            .handle_heal_object(&bucket, quorum_object, "", &HealOpts::default())
+            .await
+            .expect("quorum boundary heal should return a mapped result");
+        *store.pools[0].disk_set[0].disks.write().await = original_quorum_disks;
+        assert!(
+            matches!(quorum_err, Some(Error::ErasureReadQuorum)),
+            "quorum-boundary heal must preserve quorum error, got {quorum_err:?}"
+        );
+        shutdown.cancel();
+    }
+
     #[tokio::test]
     async fn handle_heal_format_continues_after_a_pool_error() {
         let canonical_format = FormatV3::new(1, 3);