fix(rpc): bound internode channel concurrency

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/core/pools.rs

@@ -36,8 +36,7 @@ use crate::disk::error::DiskError;
 use crate::disk::{BUCKET_META_PREFIX, RUSTFS_META_BUCKET};
 use crate::error::{Error, Result};
 use crate::error::{
-    StorageError, is_err_bucket_exists, is_err_bucket_not_found, is_err_object_not_found, is_err_operation_canceled,
-    is_err_version_not_found,
+    StorageError, is_err_bucket_exists, is_err_bucket_not_found, is_err_object_not_found, is_err_version_not_found,
 };
 use crate::layout::endpoints::EndpointServerPools;
 use crate::object_api::{GetObjectReader, ObjectOptions};
@@ -774,76 +773,7 @@ async fn load_decommission_entry_exact_versions(
 }
 
 fn resolve_decommission_check_after_list_result(list_result: Result<()>, entry_error: Option<Error>) -> Result<()> {
-    match list_result {
-        Ok(()) => entry_error.map_or(Ok(()), Err),
-        Err(list_err) => resolve_decommission_listing_error(Some(list_err), entry_error).map_or(Ok(()), Err),
-    }
-}
-
-fn resolve_decommission_listing_error(listing_error: Option<Error>, entry_error: Option<Error>) -> Option<Error> {
-    match (listing_error, entry_error) {
-        (Some(listing_error), Some(entry_error)) if is_err_operation_canceled(&listing_error) => Some(entry_error),
-        (Some(listing_error), Some(entry_error)) if is_err_operation_canceled(&entry_error) => Some(listing_error),
-        (Some(listing_error), _) => Some(listing_error),
-        (None, entry_error) => entry_error,
-    }
-}
-
-fn decommission_unresolved_listing_error(
-    bucket: &str,
-    prefix: &str,
-    candidate: Option<&str>,
-    candidate_count: usize,
-    disk_error_count: usize,
-    pool_index: usize,
-    set_index: usize,
-) -> Error {
-    let location = candidate.unwrap_or(prefix);
-    Error::other(format!(
-        "decommission listing could not resolve metadata for {bucket}/{location} on pool {pool_index} set {set_index} ({candidate_count} candidate(s), {disk_error_count} disk error(s))"
-    ))
-}
-
-fn resolve_decommission_partial_listing_entry(
-    entries: MetaCacheEntries,
-    resolver: MetadataResolutionParams,
-    bucket: &str,
-    prefix: &str,
-    disk_error_count: usize,
-    pool_index: usize,
-    set_index: usize,
-) -> Result<MetaCacheEntry> {
-    let candidate_count = entries.as_ref().iter().flatten().count();
-    if let Some(entry) = entries.resolve(resolver) {
-        return Ok(entry);
-    }
-
-    let candidate = entries.as_ref().iter().flatten().map(|entry| entry.name.as_str()).next();
-    Err(decommission_unresolved_listing_error(
-        bucket,
-        prefix,
-        candidate,
-        candidate_count,
-        disk_error_count,
-        pool_index,
-        set_index,
-    ))
-}
-
-async fn record_decommission_entry_error(
-    entry_error: &Arc<tokio::sync::Mutex<Option<Error>>>,
-    rx: &CancellationToken,
-    err: Error,
-) {
-    if rx.is_cancelled() {
-        return;
-    }
-
-    let mut first_err = entry_error.lock().await;
-    if first_err.is_none() && !rx.is_cancelled() {
-        *first_err = Some(err);
-        rx.cancel();
-    }
+    if let Some(err) = entry_error { Err(err) } else { list_result }
 }
 
 fn resolve_decommission_pool_meta_reload_result(result: Result<()>, stage: &str) -> Result<()> {
@@ -3608,7 +3538,6 @@ impl ECStore {
             let rx_clone = rx.clone();
             let bi = bi.clone();
             let set_id = set_idx;
-            let listing_entry_error = entry_error.clone();
             let worker = tokio::spawn(async move {
                 let _listing_permit = listing_permit;
                 run_decommission_listing_with_retry(
@@ -3622,11 +3551,7 @@ impl ECStore {
                         let set = set.clone();
                         let rx = rx_clone.clone();
                         let bucket = bi.clone();
-                        let entry_error = listing_entry_error.clone();
-                        async move {
-                            set.list_objects_to_decommission(rx, bucket, callback, entry_error.clone(), idx, set_id)
-                                .await
-                        }
+                        async move { set.list_objects_to_decommission(rx, bucket, callback).await }
                     },
                 )
                 .await
@@ -3656,7 +3581,11 @@ impl ECStore {
 
         wait_decommission_worker_drain(&workers, worker_limit).await?;
 
-        if let Some(err) = resolve_decommission_listing_error(listing_worker_error, entry_error.lock().await.clone()) {
+        if let Some(err) = listing_worker_error {
+            return Err(err);
+        }
+
+        if let Some(err) = entry_error.lock().await.clone() {
             return Err(err);
         }
 
@@ -4262,7 +4191,7 @@ impl ECStore {
         let buckets = self.get_buckets_to_decommission().await?;
         let pool = self.pools[idx].clone();
 
-        for (set_index, set) in pool.disk_set.iter().enumerate() {
+        for set in &pool.disk_set {
             for bucket_info in &buckets {
                 let mut lifecycle_config = None;
                 let mut object_lock_config = None;
@@ -4357,7 +4286,7 @@ impl ECStore {
                 });
 
                 let list_result = set
-                    .list_objects_to_decommission(callback_rx, bucket_info.clone(), callback, entry_error.clone(), idx, set_index)
+                    .list_objects_to_decommission(callback_rx, bucket_info.clone(), callback)
                     .await;
                 let entry_error = entry_error.lock().await.clone();
                 resolve_decommission_check_after_list_result(list_result, entry_error)?;
@@ -5092,15 +5021,12 @@ mod tests {
 pub type ListCallback = Arc<dyn Fn(MetaCacheEntry) -> BoxFuture<'static, ()> + Send + Sync + 'static>;
 
 impl SetDisks {
-    #[tracing::instrument(skip(self, rx, cb_func, entry_error))]
+    #[tracing::instrument(skip(self, rx, cb_func))]
     async fn list_objects_to_decommission(
         self: &Arc<Self>,
         rx: CancellationToken,
         bucket_info: DecomBucketInfo,
         cb_func: ListCallback,
-        entry_error: Arc<tokio::sync::Mutex<Option<Error>>>,
-        pool_index: usize,
-        set_index: usize,
     ) -> Result<()> {
         let (disks, _) = self.get_online_disks_with_healing(false).await;
         ensure_decommission_listing_disks_available(!disks.is_empty(), &bucket_info.name)?;
@@ -5115,12 +5041,6 @@ impl SetDisks {
         };
 
         let cb1 = cb_func.clone();
-        let unresolved_error = entry_error.clone();
-        let unresolved_rx = rx.clone();
-        let unresolved_bucket = bucket_info.name.clone();
-        let unresolved_prefix = bucket_info.prefix.clone();
-        let unresolved_pool_index = pool_index;
-        let unresolved_set_index = set_index;
 
         list_path_raw(
             rx,
@@ -5133,51 +5053,20 @@ impl SetDisks {
                 skip_walkdir_total_timeout: true,
                 walkdir_stall_timeout: Some(DECOMMISSION_BACKGROUND_WALKDIR_STALL_TIMEOUT),
                 agreed: Some(Box::new(move |entry: MetaCacheEntry| Box::pin(cb1(entry)))),
-                partial: Some(Box::new(move |entries: MetaCacheEntries, errs: &[Option<DiskError>]| {
+                partial: Some(Box::new(move |entries: MetaCacheEntries, _: &[Option<DiskError>]| {
                     let resolver = resolver.clone();
                     let cb_func = cb_func.clone();
-                    let bucket = unresolved_bucket.clone();
-                    let prefix = unresolved_prefix.clone();
-                    let unresolved_error = unresolved_error.clone();
-                    let unresolved_rx = unresolved_rx.clone();
-                    let pool_index = unresolved_pool_index;
-                    let set_index = unresolved_set_index;
-                    let disk_error_count = errs.iter().flatten().count();
-                    if unresolved_rx.is_cancelled() {
-                        return Box::pin(async {});
-                    }
-
-                    match resolve_decommission_partial_listing_entry(
-                        entries,
-                        resolver,
-                        &bucket,
-                        &prefix,
-                        disk_error_count,
-                        pool_index,
-                        set_index,
-                    ) {
-                        Ok(entry) => {
+                    match entries.resolve(resolver) {
+                        Some(entry) => {
                             warn!("decommission_pool: list_objects_to_decommission get {}", &entry.name);
                             Box::pin(async move {
                                 cb_func(entry).await;
                             })
                         }
-                        Err(err) => Box::pin(async move {
-                            if unresolved_rx.is_cancelled() {
-                                return;
-                            }
-                            warn!(
-                                event = EVENT_DECOMMISSION_BUCKET,
-                                component = LOG_COMPONENT_ECSTORE,
-                                subsystem = LOG_SUBSYSTEM_POOLS,
-                                bucket = %bucket,
-                                prefix = %prefix,
-                                state = "unresolved_entry",
-                                error = %err,
-                                "Decommission listing failed closed on unresolved metadata"
-                            );
-                            record_decommission_entry_error(&unresolved_error, &unresolved_rx, err).await;
-                        }),
+                        None => {
+                            warn!("decommission_pool: list_objects_to_decommission get none");
+                            Box::pin(async {})
+                        }
                     }
                 })),
                 ..Default::default()
@@ -5185,10 +5074,6 @@ impl SetDisks {
         )
         .await?;
 
-        if let Some(err) = entry_error.lock().await.clone() {
-            return Err(err);
-        }
-
         Ok(())
     }
 }
@@ -5394,12 +5279,11 @@ mod pools_tests {
         has_active_decommission_canceler, is_decommission_active, is_decommission_cancel_requested,
         load_decommission_entry_versions, local_decommission_queue_prefix, mark_decommission_bucket_done,
         merge_pool_status_refresh, missing_decommission_worker_prefix, observe_decommission_terminal_reload_result,
-        pool_meta_has_active_decommission, record_decommission_entry_error, require_decommission_store,
-        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_error,
-        resolve_decommission_listing_worker_result, resolve_decommission_optional_bucket_config_result,
-        resolve_decommission_partial_listing_entry, resolve_decommission_pool_meta_reload_result,
+        pool_meta_has_active_decommission, require_decommission_store, 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, resolve_decommission_pool_meta_reload_result,
         resolve_decommission_preflight_heal_result, resolve_decommission_progress_save_result,
         resolve_decommission_spawn_failure_result, resolve_decommission_terminal_mark_after_error_result,
         resolve_decommission_terminal_mark_result, resolve_decommission_update_after_result,
@@ -5418,9 +5302,7 @@ mod pools_tests {
     use crate::error::{Error, StorageError};
     use crate::layout::endpoints::{EndpointServerPools, Endpoints, PoolEndpoints};
     use crate::services::rebalance::{RebalStatus, RebalanceInfo, RebalanceMeta, RebalanceStats};
-    use rustfs_filemeta::{
-        FileInfo, FileInfoVersions, MetaCacheEntries, MetaCacheEntry, MetadataResolutionParams, ObjectPartInfo,
-    };
+    use rustfs_filemeta::{FileInfo, FileInfoVersions, MetaCacheEntry, ObjectPartInfo};
     use rustfs_rio::Index;
     use std::sync::{
         Arc,
@@ -6439,65 +6321,6 @@ mod pools_tests {
         assert!(matches!(err, Error::SlowDown));
     }
 
-    #[test]
-    fn test_resolve_decommission_partial_listing_entry_rejects_unresolved_metadata() {
-        let err = resolve_decommission_partial_listing_entry(
-            MetaCacheEntries(vec![None]),
-            MetadataResolutionParams {
-                dir_quorum: 2,
-                obj_quorum: 2,
-                bucket: "bucket-a".to_string(),
-                ..Default::default()
-            },
-            "bucket-a",
-            "prefix/",
-            1,
-            2,
-            3,
-        )
-        .expect_err("unresolved partial listing must fail closed");
-
-        let message = err.to_string();
-        assert!(message.contains("decommission listing could not resolve metadata"));
-        assert!(message.contains("bucket-a/prefix/"));
-        assert!(message.contains("pool 2 set 3"));
-        assert!(message.contains("1 disk error(s)"));
-    }
-
-    #[tokio::test]
-    async fn test_record_decommission_entry_error_cancels_listing_and_preserves_first_error() {
-        let entry_error = Arc::new(tokio::sync::Mutex::new(None));
-        let rx = CancellationToken::new();
-
-        record_decommission_entry_error(&entry_error, &rx, Error::SlowDown).await;
-        record_decommission_entry_error(&entry_error, &rx, Error::OperationCanceled).await;
-
-        assert!(rx.is_cancelled());
-        assert!(matches!(*entry_error.lock().await, Some(Error::SlowDown)));
-    }
-
-    #[tokio::test]
-    async fn test_record_decommission_entry_error_ignores_already_canceled_listing() {
-        let entry_error = Arc::new(tokio::sync::Mutex::new(None));
-        let rx = CancellationToken::new();
-        rx.cancel();
-
-        record_decommission_entry_error(&entry_error, &rx, Error::SlowDown).await;
-
-        assert!(entry_error.lock().await.is_none());
-    }
-
-    #[test]
-    fn test_resolve_decommission_listing_error_preserves_real_listing_failure() {
-        let err = resolve_decommission_listing_error(Some(Error::SlowDown), Some(Error::OperationCanceled))
-            .expect("listing failure should be returned");
-        assert!(matches!(err, Error::SlowDown));
-
-        let err = resolve_decommission_listing_error(Some(Error::OperationCanceled), Some(Error::SlowDown))
-            .expect("entry failure should be returned");
-        assert!(matches!(err, Error::SlowDown));
-    }
-
     #[test]
     fn test_resolve_decommission_check_after_list_result_returns_list_result_without_entry_error() {
         let err = resolve_decommission_check_after_list_result(Err(Error::OperationCanceled), None)

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/protos/src/lib.rs

@@ -2106,6 +2106,9 @@ pub enum ChannelClass {
     Bulk,
 }
 
+// Keep multiplexed unary RPCs below h2's per-connection small-frame budget.
+const INTERNODE_RPC_CONCURRENCY_LIMIT: usize = 64;
+
 /// Whether control/bulk channel isolation is enabled (env-gated, default off for safe rollout).
 fn channel_isolation_enabled() -> bool {
     rustfs_utils::get_env_bool(
@@ -2188,6 +2191,7 @@ async fn build_channel(dial_addr: &str, cache_key: &str) -> Result<Channel, Box<
     let mut connector = Endpoint::from_shared(dial_addr.to_string())?
         // Fast connection timeout for dead peer detection
         .connect_timeout(connect_timeout)
+        .concurrency_limit(INTERNODE_RPC_CONCURRENCY_LIMIT)
         // TCP-level keepalive - OS will probe connection
         .tcp_keepalive(Some(tcp_keepalive))
         // Disable Nagle so latency-sensitive control-plane RPCs (locks/health) are not batched