fix(heal): fence format repair during pool transitions

crates/ecstore/src/core/pools.rs

@@ -1675,7 +1675,7 @@ impl PoolMeta {
         self.load_no_lock(pool).await
     }
 
-    async fn load_no_lock<S>(&mut self, pool: Arc<S>) -> Result<()>
+    pub(crate) async fn load_no_lock<S>(&mut self, pool: Arc<S>) -> Result<()>
     where
         S: EcstoreObjectIO,
     {

crates/ecstore/src/store/heal.rs

@@ -13,7 +13,12 @@
 // limitations under the License.
 
 use super::*;
+use crate::core::pools::POOL_META_NAME;
+use crate::services::rebalance::{REBAL_META_NAME, RebalStatus};
+use crate::set_disk::get_lock_acquire_timeout;
 use crate::storage_api_contracts::heal::HealOperations as _;
+use crate::storage_api_contracts::namespace::NamespaceLocking as _;
+use rustfs_lock::NamespaceLockGuard;
 use tracing::trace;
 
 const LOG_COMPONENT_ECSTORE: &str = "ecstore";
@@ -30,7 +35,119 @@ fn invalid_heal_pool_index(pool_idx: usize, pool_count: usize) -> Error {
     )
 }
 
+#[derive(Debug, Clone, Copy)]
+enum HealFormatPoolSkip {
+    Completed,
+    Retryable,
+}
+
+fn classify_heal_format_pool(
+    pool_idx: usize,
+    pool_cmd_line: &str,
+    pool_meta: &PoolMeta,
+    rebalance_meta: Option<&RebalanceMeta>,
+) -> Option<HealFormatPoolSkip> {
+    let Some(pool) = pool_meta.pools.get(pool_idx) else {
+        return Some(HealFormatPoolSkip::Retryable);
+    };
+
+    if pool.id != pool_idx || pool_cmd_line.is_empty() || pool.cmd_line.is_empty() || pool.cmd_line != pool_cmd_line {
+        return Some(HealFormatPoolSkip::Retryable);
+    }
+
+    if let Some(decommission) = pool.decommission.as_ref() {
+        if decommission.complete {
+            return Some(HealFormatPoolSkip::Completed);
+        }
+        if decommission.failed || decommission.canceled || decommission.queued || pool_meta.is_suspended(pool_idx) {
+            return Some(HealFormatPoolSkip::Retryable);
+        }
+    }
+
+    if let Some(meta) = rebalance_meta {
+        let Some(pool_stats) = meta.pool_stats.get(pool_idx) else {
+            return Some(HealFormatPoolSkip::Retryable);
+        };
+        if pool_stats.info.stopping || (pool_stats.participating && pool_stats.info.status == RebalStatus::Started) {
+            return Some(HealFormatPoolSkip::Retryable);
+        }
+    }
+
+    None
+}
+
+fn heal_format_pool_skip_error(skip: HealFormatPoolSkip) -> Error {
+    match skip {
+        HealFormatPoolSkip::Completed => StorageError::NoHealRequired,
+        HealFormatPoolSkip::Retryable => StorageError::SlowDown,
+    }
+}
+
+fn heal_format_fence_lost_error() -> Error {
+    StorageError::SlowDown
+}
+
 impl ECStore {
+    async fn acquire_heal_format_fence(
+        &self,
+    ) -> Result<(NamespaceLockGuard, NamespaceLockGuard, PoolMeta, Option<RebalanceMeta>)> {
+        let metadata_pool = self
+            .pools
+            .first()
+            .cloned()
+            .ok_or_else(|| Error::other("heal format requires at least one storage pool"))?;
+
+        // Metadata fence order is part of the decommission/rebalance protocol:
+        // pool.bin must always be acquired before rebalance.bin.
+        let pool_lock = metadata_pool.new_ns_lock(RUSTFS_META_BUCKET, POOL_META_NAME).await?;
+        let pool_guard = pool_lock.get_write_lock(get_lock_acquire_timeout()).await?;
+        let rebalance_lock = metadata_pool.new_ns_lock(RUSTFS_META_BUCKET, REBAL_META_NAME).await?;
+        let rebalance_guard = rebalance_lock.get_write_lock(get_lock_acquire_timeout()).await?;
+
+        if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        let mut pool_meta = PoolMeta::default();
+        pool_meta.load_no_lock(metadata_pool.clone()).await?;
+        if pool_meta.pools.len() != self.pools.len()
+            || pool_meta.pools.iter().enumerate().any(|(pool_idx, pool)| {
+                pool.id != pool_idx || pool.cmd_line.is_empty() || pool.cmd_line != self.pools[pool_idx].endpoints.cmd_line
+            })
+        {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        let mut rebalance_meta = RebalanceMeta::new();
+        let rebalance_meta = match rebalance_meta
+            .load_with_opts(
+                metadata_pool,
+                ObjectOptions {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+        {
+            Ok(()) => Some(rebalance_meta),
+            Err(Error::ConfigNotFound) => None,
+            Err(err) => return Err(err),
+        };
+
+        if rebalance_meta
+            .as_ref()
+            .is_some_and(|meta| meta.pool_stats.len() != self.pools.len())
+        {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        Ok((pool_guard, rebalance_guard, pool_meta, rebalance_meta))
+    }
+
     fn get_pools_for_heal_object(&self, opts: &HealOpts) -> Result<Vec<Arc<Sets>>> {
         match opts.pool {
             Some(pool_idx) => Ok(vec![
@@ -52,8 +169,24 @@ impl ECStore {
         };
 
         let mut count_no_heal = 0;
+        let mut count_completed = 0;
         let mut first_error = None;
-        for pool in self.pools.iter() {
+        for (pool_idx, pool) in self.pools.iter().enumerate() {
+            let (pool_guard, rebalance_guard, pool_meta, rebalance_meta) = self.acquire_heal_format_fence().await?;
+            if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+                first_error.get_or_insert(heal_format_fence_lost_error());
+                break;
+            }
+            if let Some(skip) = classify_heal_format_pool(pool_idx, &pool.endpoints.cmd_line, &pool_meta, rebalance_meta.as_ref())
+            {
+                if matches!(skip, HealFormatPoolSkip::Completed) {
+                    count_completed += 1;
+                } else {
+                    first_error.get_or_insert(heal_format_pool_skip_error(skip));
+                }
+                continue;
+            }
+
             let (mut result, err) = pool.heal_format(dry_run).await?;
             if let Some(err) = err {
                 match err {
@@ -69,11 +202,18 @@ impl ECStore {
             r.set_count += result.set_count;
             r.before.drives.append(&mut result.before.drives);
             r.after.drives.append(&mut result.after.drives);
+
+            // Sets::heal_format cannot observe this guard before each disk write;
+            // fail closed after the call if the lease was lost during format IO.
+            if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+                first_error.get_or_insert(heal_format_fence_lost_error());
+                break;
+            }
         }
         if let Some(err) = first_error {
             return Ok((r, Some(err)));
         }
-        if count_no_heal == self.pools.len() {
+        if count_no_heal + count_completed == self.pools.len() {
             info!(
                 event = EVENT_HEAL_FORMAT_COMPLETED,
                 component = LOG_COMPONENT_ECSTORE,
@@ -300,6 +440,7 @@ mod tests {
     use crate::core::pools::{PoolDecommissionInfo, PoolStatus};
     use crate::disk::{DiskOption, format::FormatV3, new_disk};
     use crate::layout::endpoints::{Endpoints, PoolEndpoints};
+    use crate::services::rebalance::{RebalanceInfo, RebalanceStats};
     use crate::store::init_format::{load_format_erasure, save_format_file};
 
     async fn minimal_heal_pool(pool_idx: usize) -> Arc<Sets> {
@@ -347,6 +488,164 @@ mod tests {
         }
     }
 
+    fn pool_meta_with_decommission(info: PoolDecommissionInfo) -> PoolMeta {
+        PoolMeta {
+            pools: vec![PoolStatus {
+                id: 0,
+                cmd_line: "pool-0".to_string(),
+                last_update: OffsetDateTime::UNIX_EPOCH,
+                decommission: Some(info),
+            }],
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn heal_format_pool_state_barriers_are_classified() {
+        let active = pool_meta_with_decommission(PoolDecommissionInfo {
+            start_time: Some(OffsetDateTime::UNIX_EPOCH),
+            ..Default::default()
+        });
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &active, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        for info in [
+            PoolDecommissionInfo {
+                failed: true,
+                ..Default::default()
+            },
+            PoolDecommissionInfo {
+                canceled: true,
+                ..Default::default()
+            },
+        ] {
+            assert!(matches!(
+                classify_heal_format_pool(0, "pool-0", &pool_meta_with_decommission(info), None),
+                Some(HealFormatPoolSkip::Retryable)
+            ));
+        }
+
+        let completed = pool_meta_with_decommission(PoolDecommissionInfo {
+            complete: true,
+            ..Default::default()
+        });
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &completed, None),
+            Some(HealFormatPoolSkip::Completed)
+        ));
+    }
+
+    #[test]
+    fn heal_format_pool_rebalance_barriers_and_identity_are_fail_closed() {
+        let identity_meta = pool_meta_with_decommission(PoolDecommissionInfo::default());
+        let rebalance = RebalanceMeta {
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Started,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&rebalance)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let stopping = RebalanceMeta {
+            pool_stats: vec![RebalanceStats {
+                info: RebalanceInfo {
+                    stopping: true,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopping)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let identity = pool_meta_with_decommission(PoolDecommissionInfo::default());
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-new", &identity, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let identity_without_decommission = PoolMeta {
+            pools: vec![PoolStatus {
+                id: 0,
+                cmd_line: "pool-0".to_string(),
+                last_update: OffsetDateTime::UNIX_EPOCH,
+                decommission: None,
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-new", &identity_without_decommission, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        assert!(matches!(
+            classify_heal_format_pool(0, "", &identity_meta, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &PoolMeta::default(), None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let stopped = RebalanceMeta {
+            stopped_at: Some(OffsetDateTime::UNIX_EPOCH),
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Stopped,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopped)).is_none());
+
+        let stopping_after_stop = RebalanceMeta {
+            stopped_at: Some(OffsetDateTime::UNIX_EPOCH),
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Started,
+                    stopping: true,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopping_after_stop)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+    }
+
+    #[test]
+    fn skipped_heal_format_pool_is_never_reported_as_success() {
+        assert!(matches!(
+            heal_format_pool_skip_error(HealFormatPoolSkip::Retryable),
+            StorageError::SlowDown
+        ));
+        assert!(matches!(
+            heal_format_pool_skip_error(HealFormatPoolSkip::Completed),
+            StorageError::NoHealRequired
+        ));
+    }
+
     #[tokio::test]
     async fn heal_object_pool_scope_selects_only_requested_pool() {
         let store = minimal_heal_store().await;
@@ -615,6 +914,18 @@ mod tests {
             bucket_fence_registry: std::sync::Arc::default(),
         };
 
+        let err = store
+            .handle_heal_format(false)
+            .await
+            .expect_err("missing pool metadata must fail closed before format writes");
+        assert!(matches!(err, StorageError::SlowDown));
+
+        let pool_meta = PoolMeta::new(&store.pools, &PoolMeta::default());
+        pool_meta
+            .save(store.pools.clone())
+            .await
+            .expect("pool metadata should be persisted before format heal");
+
         let (result, err) = store
             .handle_heal_format(false)
             .await
@@ -628,5 +939,22 @@ mod tests {
             .await
             .expect("the later pool should be healed despite the first pool error");
         assert_eq!(healed.erasure.this, recoverable_format.erasure.sets[0][2]);
+
+        let mut completed_meta = PoolMeta::new(&store.pools, &PoolMeta::default());
+        for status in &mut completed_meta.pools {
+            status.decommission = Some(PoolDecommissionInfo {
+                complete: true,
+                ..Default::default()
+            });
+        }
+        completed_meta
+            .save(store.pools.clone())
+            .await
+            .expect("completed pool metadata should be persisted");
+        let (_, err) = store
+            .handle_heal_format(false)
+            .await
+            .expect("completed pools should be reported as a no-op");
+        assert!(matches!(err, Some(StorageError::NoHealRequired)));
     }
 }

crates/heal/src/heal/task/heal_erasure_set.rs

@@ -231,6 +231,10 @@ impl HealTask {
                             "Heal erasure set format repair skipped because no format heal was required"
                         );
                     } else {
+                        let error = e;
+                        if error.is_recoverable_heal() {
+                            return Err(error);
+                        }
                         error!(
                             target: "rustfs::heal::task",
                             event = EVENT_HEAL_ERASURE_SET_RESULT,
@@ -239,7 +243,7 @@ impl HealTask {
                             task_id = %self.id,
                             set_disk_id,
                             result = "format_failed",
-                            error = %e,
+                            error = %error,
                             "Heal erasure set failed"
                         );
                         {
@@ -247,7 +251,7 @@ impl HealTask {
                             progress.update_progress(4, 4, 0, 0);
                         }
                         return Err(Error::TaskExecutionFailed {
-                            message: format!("Failed to heal disk format for {set_disk_id}: {e}"),
+                            message: format!("Failed to heal disk format for {set_disk_id}: {error}"),
                         });
                     }
                 } else {
@@ -284,6 +288,9 @@ impl HealTask {
             Err(Error::TaskCancelled) => return Err(Error::TaskCancelled),
             Err(Error::TaskTimeout) => return Err(Error::TaskTimeout),
             Err(e) => {
+                if e.is_recoverable_heal() {
+                    return Err(e);
+                }
                 error!(
                     target: "rustfs::heal::task",
                     event = EVENT_HEAL_ERASURE_SET_RESULT,

crates/heal/src/heal/task/tests.rs

@@ -547,6 +547,7 @@ struct MockStorage {
     heal_object_outcome: Mutex<Option<MockHealObjectOutcome>>,
     heal_object_outcomes: Mutex<HashMap<String, VecDeque<MockHealObjectOutcome>>>,
     format_no_heal_required: Mutex<bool>,
+    format_error: Mutex<Option<Error>>,
     global_format_calls: Mutex<u32>,
     replacement_format_calls: Mutex<Vec<(usize, usize, Vec<String>)>>,
     replacement_targets_ready: Mutex<bool>,
@@ -867,6 +868,9 @@ impl HealStorageAPI for MockStorage {
 
     async fn heal_format(&self, _dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
         *self.global_format_calls.lock().unwrap() += 1;
+        if let Some(error) = self.format_error.lock().unwrap().take() {
+            return Err(error);
+        }
         let no_heal_required = *self.format_no_heal_required.lock().unwrap();
         if no_heal_required {
             Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::NoHealRequired))))
@@ -2052,6 +2056,30 @@ async fn test_erasure_set_heal_continues_after_format_no_heal_required() {
     );
 }
 
+#[tokio::test]
+async fn erasure_set_format_slowdown_is_propagated() {
+    let storage = Arc::new(MockStorage {
+        format_error: Mutex::new(Some(Error::Storage(EcstoreError::SlowDown))),
+        ..Default::default()
+    });
+    let request = HealRequest::new(
+        HealType::ErasureSet {
+            buckets: Vec::new(),
+            set_disk_id: "pool_0_set_0".to_string(),
+        },
+        HealOptions::default(),
+        HealPriority::Normal,
+    );
+    let task = HealTask::from_request(request, storage);
+
+    let error = task
+        .execute()
+        .await
+        .expect_err("format SlowDown must remain recoverable for the task manager");
+
+    assert!(matches!(error, Error::Storage(EcstoreError::SlowDown)));
+}
+
 #[tokio::test]
 async fn erasure_set_bucket_prepass_failure_stops_before_object_heal() {
     let temp = TempDir::new().expect("temporary directory should be created");

crates/test-utils/src/lib.rs

@@ -245,6 +245,18 @@ impl TestECStoreEnvBuilder {
             .await
             .expect("build test ECStore");
 
+        // The production bootstrap only persists pool.bin from the elected
+        // first cluster node.  Test stores intentionally have no cluster
+        // election, but heal-format still requires that durable fence before
+        // it can write any disk format.  Materialize the validated topology
+        // here so the shared fixture models a ready single-node store.
+        let mut pool_meta = ecstore.pool_meta.read().await.clone();
+        pool_meta.dont_save = false;
+        pool_meta
+            .save(ecstore.pools.clone())
+            .await
+            .expect("persist test pool metadata");
+
         if self.init_bucket_metadata {
             let buckets_list = ecstore
                 .list_bucket(&BucketOptions {