Merge remote-tracking branch 'origin/main' into overtrue/fix-1902-mpu-decommission

.config/e2e-full-selection.txt

@@ -1,2 +1,2 @@
-sha256-darwin=52a05fdfae8bcf6f5828cc2b1e91b2a268139d3f7e1fc47d7b875b55fffb3995
-sha256-linux=a22d8af72e250595ac4445e8c880f3f9706202e09ed196e5b7baac632dead8d8
+sha256-darwin=9f767b37ed8b1c82da62ea441462d75487785c8086e56f08fb6f6cd89c6e2e52
+sha256-linux=fbdaf42b220958d4b1e8880e0f8b5a7992d38e21051bb60596dd4538424757d6

.config/nextest.toml

@@ -107,10 +107,9 @@ filter = 'package(e2e_test) & test(/^inline_fast_path_cluster_test::/)'
 test-group = 'e2e-inline-boundaries'
 
 # Vault KMS tests share the fixed dev-server port 8200. serial_test's #[serial]
-# does not cross nextest process boundaries, so keep every Vault-backed test in
-# one group.
+# does not cross nextest process boundaries, so keep these tests in one group.
 [[profile.default.overrides]]
-filter = 'package(e2e_test) & (test(/^kms::kms_vault_test::/) | test(/^kms::configured_roundtrip_test::test_configured_vault_kms_admin_and_versioned_cleanup$/))'
+filter = 'package(e2e_test) & test(/^kms::kms_vault_test::/)'
 test-group = 'e2e-vault'
 
 # ---------------------------------------------------------------------------
@@ -444,5 +443,5 @@ filter = 'package(e2e_test) & test(/^inline_fast_path_cluster_test::/)'
 test-group = 'e2e-inline-boundaries'
 
 [[profile.e2e-full.overrides]]
-filter = 'package(e2e_test) & (test(/^kms::kms_vault_test::/) | test(/^kms::configured_roundtrip_test::test_configured_vault_kms_admin_and_versioned_cleanup$/))'
+filter = 'package(e2e_test) & test(/^kms::kms_vault_test::/)'
 test-group = 'e2e-vault'

Cargo.lock

@@ -1858,9 +1858,9 @@ dependencies = [
 
 [[package]]
 name = "cc"
-version = "1.4.4"
+version = "1.4.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0ad534f4357a5264cce5019c989cf66a4f0dc4e0d1b1d15f8aacec0ff7360273"
+checksum = "509591b7bcd67f4ef775afad7662703b4935daaa6ec0e5605cfb1090b32a2b6d"
 dependencies = [
  "find-msvc-tools",
  "jobserver",
@@ -2522,6 +2522,12 @@ dependencies = [
  "subtle",
 ]
 
+[[package]]
+name = "cty"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b365fabc795046672053e29c954733ec3b05e4be654ab130fe8f1f94d7051f35"
+
 [[package]]
 name = "curve25519-dalek"
 version = "4.1.3"
@@ -5982,6 +5988,15 @@ version = "0.2.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6d2cec3eae94f9f509c767b45932f1ada8350c4bdb85af2fcab4a3c14807981"
 
+[[package]]
+name = "libmimalloc-sys"
+version = "0.1.49"
+source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
+dependencies = [
+ "cc",
+ "cty",
+]
+
 [[package]]
 name = "libredox"
 version = "0.1.20"
@@ -6382,6 +6397,14 @@ dependencies = [
  "synstructure 0.13.2",
 ]
 
+[[package]]
+name = "mimalloc"
+version = "0.1.52"
+source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
+dependencies = [
+ "libmimalloc-sys",
+]
+
 [[package]]
 name = "mime"
 version = "0.3.17"
@@ -9139,11 +9162,13 @@ dependencies = [
  "insta",
  "jiff",
  "libc",
+ "libmimalloc-sys",
  "libsystemd",
  "matchit 0.9.2",
  "md-5 0.11.0",
  "metrics",
  "metrics-util",
+ "mimalloc",
  "mime_guess",
  "opentelemetry",
  "opentelemetry_sdk",
@@ -9179,8 +9204,6 @@ dependencies = [
  "rustfs-lock",
  "rustfs-log-analyzer",
  "rustfs-madmin",
- "rustfs-mimalloc",
- "rustfs-mimalloc-sys",
  "rustfs-notify",
  "rustfs-object-capacity",
  "rustfs-object-data-cache",
@@ -9852,24 +9875,6 @@ dependencies = [
  "tokio",
 ]
 
-[[package]]
-name = "rustfs-mimalloc"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a406f4aa07084301d485beec873af6dccc8e3f8762da244743df92038b1db1a6"
-dependencies = [
- "rustfs-mimalloc-sys",
-]
-
-[[package]]
-name = "rustfs-mimalloc-sys"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c3051b819175f58445d4c369a72f0ab88149f3885ba8bea2aff3be01f53fe7cd"
-dependencies = [
- "cc",
-]
-
 [[package]]
 name = "rustfs-notify"
 version = "1.0.0-rc.3"

Cargo.toml

@@ -350,8 +350,8 @@ russh-sftp = "2.4.0"
 dav-server = "0.11.0"
 
 # Performance Analysis and Memory Profiling
-rustfs-mimalloc = { version = "0.5.0" }
-rustfs-mimalloc-sys = { version = "0.5.0" }
+mimalloc = { version = "0.1.52", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11" }
+libmimalloc-sys = { version = "0.1.49", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11", features = ["extended"] }
 hotpath = { version = "0.23.3", default-features = false }
 # Snapshot testing for output format regression detection
 insta = { version = "1.48" }

crates/e2e_test/src/kms/configured_roundtrip_test.rs

@@ -432,6 +432,7 @@ async fn test_configured_local_kms_admin_and_versioned_cleanup() -> TestResult {
 }
 
 #[tokio::test]
+#[ignore = "requires a Vault binary"]
 async fn test_configured_vault_kms_admin_and_versioned_cleanup() -> TestResult {
     let mut env = VaultTestEnvironment::new().await?;
     env.start_vault().await?;

crates/ecstore/src/core/pools.rs

@@ -5143,6 +5143,9 @@ impl ECStore {
         let buckets = self.get_buckets_to_decommission().await?;
         let pool = self.pools[idx].clone();
 
+        self.ensure_decommission_multipart_uploads_drained(idx, &pool, &buckets)
+            .await?;
+
         for (set_index, set) in pool.disk_set.iter().enumerate() {
             for bucket_info in &buckets {
                 let mut lifecycle_config = None;
@@ -5256,6 +5259,49 @@ impl ECStore {
         Ok(())
     }
 
+    async fn ensure_decommission_multipart_uploads_drained(
+        &self,
+        idx: usize,
+        pool: &Sets,
+        buckets: &[DecomBucketInfo],
+    ) -> Result<()> {
+        let mut bucket_names = buckets
+            .iter()
+            .filter(|bucket| bucket.name != RUSTFS_META_BUCKET)
+            .map(|bucket| bucket.name.as_str())
+            .collect::<Vec<_>>();
+        bucket_names.sort_unstable();
+        bucket_names.dedup();
+
+        // Take one bucket fence at a time so cross-bucket COPY cannot form an
+        // ABBA cycle. Suspension prevents new source uploads after each fence.
+        for bucket in bucket_names {
+            let lifecycle_guard = self.acquire_bucket_lifecycle_write_lock(bucket).await?;
+            if lifecycle_guard.is_lock_lost() {
+                return Err(Error::other(format!(
+                    "decommission multipart drain lost the bucket lifecycle fence for `{bucket}`"
+                )));
+            }
+            for set in &pool.disk_set {
+                if let Some(upload_path) = set.first_multipart_upload_path_for_decommission(bucket).await? {
+                    return Err(Error::other(format!(
+                        "pool {idx} still contains multipart upload `{upload_path}` for bucket `{bucket}`; resolve it before retrying decommission"
+                    )));
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    #[cfg(test)]
+    pub(crate) async fn ensure_decommission_multipart_uploads_drained_for_test(self: &Arc<Self>, idx: usize) -> Result<()> {
+        let buckets = self.get_buckets_to_decommission().await?;
+        let pool = self.pools[idx].clone();
+        self.ensure_decommission_multipart_uploads_drained(idx, pool.as_ref(), &buckets)
+            .await
+    }
+
     #[tracing::instrument(skip(self, rd))]
     async fn decommission_object(
         self: Arc<Self>,

crates/ecstore/src/set_disk/ops/multipart.rs

@@ -556,6 +556,69 @@ async fn multipart_upload_paths_on_disk(disk: DiskStore, bucket: &str) -> disk::
 }
 
 impl SetDisks {
+    async fn discover_multipart_upload_paths(
+        &self,
+        orig_bucket: &str,
+        error_path: &str,
+    ) -> Result<(Vec<Option<DiskStore>>, Vec<String>, usize)> {
+        let disks = self.disks.read().await.clone();
+        if disks.is_empty() {
+            return Err(Error::ErasureReadQuorum);
+        }
+        let discovery_quorum = if self.default_parity_count == 0 {
+            disks.len()
+        } else {
+            (disks.len() / 2).max(1)
+        };
+        let mut discovery_errors = (0..disks.len()).map(|_| Some(DiskError::DiskNotFound)).collect::<Vec<_>>();
+        let mut candidate_counts = HashMap::<String, usize>::new();
+        let mut discovery_tasks = JoinSet::new();
+        for (index, disk) in disks.iter().enumerate() {
+            let disk = disk.clone();
+            let orig_bucket = orig_bucket.to_string();
+            discovery_tasks.spawn(async move {
+                let result = match disk {
+                    Some(disk) => multipart_upload_paths_on_disk(disk, &orig_bucket).await,
+                    None => Err(DiskError::DiskNotFound),
+                };
+                (index, result)
+            });
+        }
+
+        while let Some(task_result) = discovery_tasks.join_next().await {
+            let Ok((index, result)) = task_result else {
+                continue;
+            };
+            match result {
+                Ok(paths) => {
+                    discovery_errors[index] = None;
+                    for path in paths {
+                        *candidate_counts.entry(path).or_insert(0) += 1;
+                    }
+                }
+                Err(err) => discovery_errors[index] = Some(err),
+            }
+        }
+
+        if let Some(err) = reduce_read_quorum_errs(&discovery_errors, OBJECT_OP_IGNORED_ERRS, discovery_quorum) {
+            return Err(to_object_err(err.into(), vec![orig_bucket, error_path]));
+        }
+
+        let mut candidate_paths = candidate_counts
+            .into_iter()
+            .filter_map(|(path, count)| (count >= discovery_quorum).then_some(path))
+            .collect::<Vec<_>>();
+        candidate_paths.sort_unstable();
+        Ok((disks, candidate_paths, discovery_quorum))
+    }
+
+    pub(crate) async fn first_multipart_upload_path_for_decommission(&self, bucket: &str) -> Result<Option<String>> {
+        let (_, paths, _) = self
+            .discover_multipart_upload_paths(bucket, RUSTFS_META_MULTIPART_BUCKET)
+            .await?;
+        Ok(paths.into_iter().next())
+    }
+
     async fn acquire_multipart_upload_read_lock(
         &self,
         op: &'static str,
@@ -747,53 +810,7 @@ impl SetDisks {
         max_uploads: usize,
         expected_incarnation_id: Option<Uuid>,
     ) -> Result<ListMultipartsInfo> {
-        let disks = self.disks.read().await.clone();
-        if disks.is_empty() {
-            return Err(Error::ErasureReadQuorum);
-        }
-        let discovery_quorum = if self.default_parity_count == 0 {
-            disks.len()
-        } else {
-            (disks.len() / 2).max(1)
-        };
-        let mut discovery_errors = (0..disks.len()).map(|_| Some(DiskError::DiskNotFound)).collect::<Vec<_>>();
-        let mut candidate_counts = HashMap::<String, usize>::new();
-        let mut discovery_tasks = JoinSet::new();
-        for (index, disk) in disks.iter().enumerate() {
-            let disk = disk.clone();
-            let bucket = bucket.to_string();
-            discovery_tasks.spawn(async move {
-                let result = match disk {
-                    Some(disk) => multipart_upload_paths_on_disk(disk, &bucket).await,
-                    None => Err(DiskError::DiskNotFound),
-                };
-                (index, result)
-            });
-        }
-
-        while let Some(task_result) = discovery_tasks.join_next().await {
-            let Ok((index, result)) = task_result else {
-                continue;
-            };
-            match result {
-                Ok(paths) => {
-                    discovery_errors[index] = None;
-                    for path in paths {
-                        *candidate_counts.entry(path).or_insert(0) += 1;
-                    }
-                }
-                Err(err) => discovery_errors[index] = Some(err),
-            }
-        }
-
-        if let Some(err) = reduce_read_quorum_errs(&discovery_errors, OBJECT_OP_IGNORED_ERRS, discovery_quorum) {
-            return Err(to_object_err(err.into(), vec![bucket, prefix]));
-        }
-
-        let candidate_paths = candidate_counts
-            .into_iter()
-            .filter_map(|(path, count)| (count >= discovery_quorum).then_some(path))
-            .collect::<Vec<_>>();
+        let (disks, candidate_paths, discovery_quorum) = self.discover_multipart_upload_paths(bucket, prefix).await?;
         let listed_uploads = stream::iter(candidate_paths)
             .map(|upload_path| {
                 let disks = &disks;

crates/ecstore/src/set_disk/ops/object.rs

@@ -2124,13 +2124,26 @@ impl SetDisks {
 
             let put_object_size = known_put_object_storage_size(data.size());
             let shard_file_size_raw = erasure.shard_file_size(put_object_size);
-            let is_inline_buffer = storage_class_config.should_inline(shard_file_size_raw, erasure.data_shards, opts.versioned);
+            let is_inline_buffer =
+                storage_class_config.should_inline(shard_file_size_raw, erasure.data_shards, opts.versioned);
 
             let collect_stage_timing = rustfs_io_metrics::put_stage_metrics_enabled() || issue3031_diag_enabled();
             let shard_file_size = shard_file_size_raw;
             let shard_size = erasure.shard_size();
             let write_path = classify_put_write_path(is_inline_buffer, put_object_size, fi.erasure.block_size);
             let direct_inline_commit = matches!(write_path, SmallWritePath::Inline);
+            {
+                use std::io::Write;
+                let msg = format!(
+                    "INLINE_DEBUG: bucket={} obj={} size={} shard_fs={} ds={} bs={} inline={} direct={} path={} iblock={} ver={}\n",
+                    bucket, object, put_object_size, shard_file_size_raw, erasure.data_shards, fi.erasure.block_size,
+                    is_inline_buffer, direct_inline_commit, write_path.metric_label(), storage_class_config.inline_block(), opts.versioned
+                );
+                if let Ok(mut f) = std::fs::OpenOptions::new().create(true).append(true).open("/tmp/rustfs_inline_debug.log") {
+                    let _ = f.write_all(msg.as_bytes());
+                }
+                let _ = std::io::stderr().write_all(msg.as_bytes());
+            }
             rustfs_io_metrics::record_put_object_path(write_path.metric_label());
             let writer_setup_stage_start = collect_stage_timing.then(Instant::now);
             let (mut writers, errors) = if direct_inline_commit {

crates/ecstore/src/store/init.rs

@@ -1589,6 +1589,177 @@ mod tests {
         shutdown.cancel();
     }
 
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn suspended_decommission_source_multipart_remains_operable_until_drained() {
+        let temp_dir = tempfile::tempdir().expect("create decommission multipart drain store dir");
+        let (_ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "decommission-multipart-drain", &[4, 4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+        let bucket = format!("decommission-multipart-drain-{}", uuid::Uuid::new_v4());
+        let complete_object = "complete.bin";
+        let abort_object = "abort.bin";
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create decommission multipart drain bucket");
+
+        let incarnation = store.bucket_incarnation_id(&bucket).await.expect("read bucket incarnation");
+        let lifecycle_guard = store
+            .acquire_bucket_lifecycle_read_lock(&bucket)
+            .await
+            .expect("acquire multipart creation lifecycle fence");
+        let mut upload_opts = ObjectOptions {
+            expected_bucket_incarnation_id: Some(incarnation),
+            ..Default::default()
+        };
+        upload_opts.add_bucket_lifecycle_lock_guard(&lifecycle_guard);
+        let complete_upload = store.pools[0]
+            .new_multipart_upload(&bucket, complete_object, &upload_opts)
+            .await
+            .expect("create source upload to complete");
+        let abort_upload = store.pools[0]
+            .new_multipart_upload(&bucket, abort_object, &upload_opts)
+            .await
+            .expect("create source upload to abort");
+        drop(lifecycle_guard);
+
+        mark_test_pool_decommissioning(&store, 0).await;
+
+        let err = store
+            .ensure_decommission_multipart_uploads_drained_for_test(0)
+            .await
+            .expect_err("an unresolved source multipart upload must block final decommission");
+        let drain_error = err.to_string();
+        assert!(
+            drain_error.contains("still contains multipart upload") && drain_error.contains(&bucket),
+            "the drain error must identify both the upload path and user bucket: {drain_error}"
+        );
+
+        let listed = store
+            .list_multipart_uploads(&bucket, "", None, None, None, 100)
+            .await
+            .expect("list uploads from suspended decommission source");
+        assert!(
+            listed
+                .uploads
+                .iter()
+                .any(|upload| upload.upload_id.as_str() == complete_upload.upload_id.as_str()),
+            "the upload selected before suspension must remain visible"
+        );
+        store
+            .get_multipart_info(&bucket, complete_object, &complete_upload.upload_id, &ObjectOptions::default())
+            .await
+            .expect("read upload metadata from suspended decommission source");
+
+        let mut part_reader = PutObjReader::from_vec(b"multipart body".to_vec());
+        let part = store
+            .put_object_part(
+                &bucket,
+                complete_object,
+                &complete_upload.upload_id,
+                1,
+                &mut part_reader,
+                &ObjectOptions::default(),
+            )
+            .await
+            .expect("write part to suspended decommission source");
+        let parts = store
+            .list_object_parts(&bucket, complete_object, &complete_upload.upload_id, None, 100, &ObjectOptions::default())
+            .await
+            .expect("list parts from suspended decommission source");
+        assert_eq!(parts.parts.len(), 1);
+        assert_eq!(parts.parts[0].etag.as_deref(), part.etag.as_deref());
+
+        store
+            .clone()
+            .complete_multipart_upload(
+                &bucket,
+                complete_object,
+                &complete_upload.upload_id,
+                vec![crate::storage_api_contracts::multipart::CompletePart {
+                    part_num: part.part_num,
+                    etag: part.etag,
+                    ..Default::default()
+                }],
+                &ObjectOptions::default(),
+            )
+            .await
+            .expect("complete upload on suspended decommission source");
+        store
+            .abort_multipart_upload(&bucket, abort_object, &abort_upload.upload_id, &ObjectOptions::default())
+            .await
+            .expect("abort upload on suspended decommission source");
+
+        store
+            .ensure_decommission_multipart_uploads_drained_for_test(0)
+            .await
+            .expect("final decommission gate should open after all source uploads are resolved");
+        assert_pool_object_present(&store.pools[0], &bucket, complete_object).await;
+
+        shutdown.cancel();
+    }
+
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn active_multipart_upload_routes_before_faulted_suspended_source() {
+        let temp_dir = tempfile::tempdir().expect("create active-first multipart routing store dir");
+        let (_ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "active-first-multipart-routing", &[4, 4]))
+                .await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+        let bucket = format!("active-first-multipart-routing-{}", uuid::Uuid::new_v4());
+        let object = "target-upload.bin";
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create active-first multipart routing bucket");
+
+        let incarnation = store.bucket_incarnation_id(&bucket).await.expect("read bucket incarnation");
+        let lifecycle_guard = store
+            .acquire_bucket_lifecycle_read_lock(&bucket)
+            .await
+            .expect("acquire multipart creation lifecycle fence");
+        let mut upload_opts = ObjectOptions {
+            expected_bucket_incarnation_id: Some(incarnation),
+            ..Default::default()
+        };
+        upload_opts.add_bucket_lifecycle_lock_guard(&lifecycle_guard);
+        let upload = store.pools[1]
+            .new_multipart_upload(&bucket, object, &upload_opts)
+            .await
+            .expect("create upload in active target pool");
+        drop(lifecycle_guard);
+
+        mark_test_pool_decommissioning(&store, 0).await;
+        let source_set = store.pools[0].get_disks_by_key(object);
+        let original_source_disks = {
+            let mut disks = source_set.disks.write().await;
+            let original = disks.clone();
+            disks.fill(None);
+            original
+        };
+
+        let source_result = store.pools[0]
+            .get_multipart_info(&bucket, object, &upload.upload_id, &ObjectOptions::default())
+            .await;
+        let routed_result = store
+            .get_multipart_info(&bucket, object, &upload.upload_id, &ObjectOptions::default())
+            .await;
+        *source_set.disks.write().await = original_source_disks;
+
+        assert!(
+            matches!(&source_result, Err(StorageError::ErasureReadQuorum)),
+            "the suspended source must expose the injected hard read failure: {source_result:?}"
+        );
+        let routed = routed_result.expect("the active target UploadID must be resolved before the faulted suspended source");
+        assert_eq!(routed.upload_id, upload.upload_id);
+
+        shutdown.cancel();
+    }
+
     #[tokio::test]
     #[serial_test::serial(storage_class_env)]
     async fn delete_objects_skips_active_rebalance_source_pool() {

crates/ecstore/src/store/multipart.rs

@@ -196,6 +196,25 @@ async fn list_pool_multipart_uploads_for_incarnation(
 }
 
 impl ECStore {
+    async fn existing_multipart_pool_order(&self) -> Vec<usize> {
+        // A draining source must not hide a valid UploadID in an active target,
+        // while physical order within each phase preserves fail-closed errors.
+        let mut active = Vec::with_capacity(self.pools.len());
+        let mut draining = Vec::new();
+        for (idx, pool) in self.pools.iter().enumerate() {
+            if self.is_pool_rebalancing(pool.pool_idx).await {
+                continue;
+            }
+            if self.is_suspended(pool.pool_idx).await {
+                draining.push(idx);
+            } else {
+                active.push(idx);
+            }
+        }
+        active.extend(draining);
+        active
+    }
+
     #[allow(clippy::too_many_arguments)]
     pub async fn list_multipart_uploads_for_bucket_incarnation(
         &self,
@@ -290,10 +309,8 @@ impl ECStore {
                 .await;
         }
 
-        for pool in self.pools.iter() {
-            if self.is_suspended(pool.pool_idx).await || self.is_pool_rebalancing(pool.pool_idx).await {
-                continue;
-            }
+        for pool_idx in self.existing_multipart_pool_order().await {
+            let pool = &self.pools[pool_idx];
             return match pool
                 .list_object_parts(bucket, object, upload_id, part_number_marker, max_parts, opts)
                 .await
@@ -353,10 +370,8 @@ impl ECStore {
         let mut common_prefixes = HashSet::new();
         let mut source_truncated = false;
 
-        for pool in self.pools.iter() {
-            if self.is_suspended(pool.pool_idx).await || self.is_pool_rebalancing(pool.pool_idx).await {
-                continue;
-            }
+        for pool_idx in self.existing_multipart_pool_order().await {
+            let pool = &self.pools[pool_idx];
             let res = list_pool_multipart_uploads_for_incarnation(
                 pool,
                 bucket,
@@ -523,10 +538,8 @@ impl ECStore {
                 .await;
         }
 
-        for pool in self.pools.iter() {
-            if self.is_suspended(pool.pool_idx).await || self.is_pool_rebalancing(pool.pool_idx).await {
-                continue;
-            }
+        for pool_idx in self.existing_multipart_pool_order().await {
+            let pool = &self.pools[pool_idx];
             let err = match pool.put_object_part(bucket, object, upload_id, part_id, data, opts).await {
                 Ok(res) => return Ok(res),
                 Err(err) => {
@@ -586,10 +599,8 @@ impl ECStore {
             return self.pools[0].get_multipart_info(bucket, object, upload_id, opts).await;
         }
 
-        for pool in self.pools.iter() {
-            if self.is_suspended(pool.pool_idx).await || self.is_pool_rebalancing(pool.pool_idx).await {
-                continue;
-            }
+        for pool_idx in self.existing_multipart_pool_order().await {
+            let pool = &self.pools[pool_idx];
 
             return match pool.get_multipart_info(bucket, object, upload_id, opts).await {
                 Ok(res) => Ok(res),
@@ -624,10 +635,8 @@ impl ECStore {
             return self.pools[0].abort_multipart_upload(bucket, object, upload_id, opts).await;
         }
 
-        for pool in self.pools.iter() {
-            if self.is_suspended(pool.pool_idx).await || self.is_pool_rebalancing(pool.pool_idx).await {
-                continue;
-            }
+        for pool_idx in self.existing_multipart_pool_order().await {
+            let pool = &self.pools[pool_idx];
 
             let err = match pool.abort_multipart_upload(bucket, object, upload_id, opts).await {
                 Ok(_) => return Ok(()),
@@ -685,10 +694,8 @@ impl ECStore {
                 .await;
         }
 
-        for pool in self.pools.iter() {
-            if self.is_suspended(pool.pool_idx).await || self.is_pool_rebalancing(pool.pool_idx).await {
-                continue;
-            }
+        for pool_idx in self.existing_multipart_pool_order().await {
+            let pool = &self.pools[pool_idx];
 
             let pool = pool.clone();
             let err = match pool

crates/ecstore/src/store/object.rs

@@ -3194,7 +3194,7 @@ impl ECStore {
 
         // Default return value
         let mut del_objects = vec![DeletedObject::default(); objects.len()];
-        let accounting = vec![None; objects.len()];
+        let mut accounting = vec![None; objects.len()];
 
         let mut del_errs = Vec::with_capacity(objects.len());
         for _ in 0..objects.len() {

crates/ecstore/src/store/rebalance/support.rs

@@ -271,7 +271,7 @@ pub(super) fn resolve_latest_object_info_candidates(
             .filter(|candidate| latest_candidate_mod_time(candidate) == Some(latest_mod_time))
             .collect::<Vec<_>>();
 
-        latest_candidates.sort_by_key(|candidate| std::cmp::Reverse(candidate.idx));
+        latest_candidates.sort_by(|left, right| right.idx.cmp(&left.idx));
 
         let Some(winner) = latest_candidates.first() else {
             return Err(Error::ErasureReadQuorum);

deny.toml

@@ -43,6 +43,9 @@ allow-git = [
     # RustFS fork carrying presigned expiry and constant-time authentication fixes.
     # owner: rustfs-maintainers  review: 2026-10
     "https://github.com/rustfs/s3s.git",
+    # MiMalloc fork pinned for hotpath allocation counting support.
+    # owner: houseme  review: 2026-10
+    "https://github.com/xonatius/mimalloc_rust.git",
 ]
 
 [bans]

docs/testing/e2e-suite-inventory.md

@@ -58,7 +58,7 @@
 | heal_erasure_disk_rebuild_test | 4 | 🌙 |
 | inline_fast_path_cluster_test | 16 |  |
 | internode_rpc_signature_e2e_test | 5 |  |
-| kms | 47 |  |
+| kms | 46 |  |
 | leading_slash_key_test | 2 | ✅ |
 | lifecycle_regression_test | 4 |  |
 | list_buckets_auth_test | 1 | ✅ |
@@ -99,4 +99,4 @@
 | tls_hot_reload_test | 1 | ✅ |
 | version_id_regression_test | 10 | ✅ |
 
-**Total listed: 576 tests across 82 modules · PR smoke: 163 tests / 36 modules · merge/main full: 454 tests / 73 modules · nightly replication: 55 tests · nightly cluster faults: 28 tests / 7 modules · nightly protocols: 16 tests** · updated 2026-08-23.
+**Total listed: 575 tests across 82 modules · PR smoke: 163 tests / 36 modules · merge/main full: 453 tests / 73 modules · nightly replication: 55 tests · nightly cluster faults: 28 tests / 7 modules · nightly protocols: 16 tests** · updated 2026-08-23.

rustfs/Cargo.toml

@@ -336,13 +336,13 @@ opentelemetry = { workspace = true }
 tracing-opentelemetry = { workspace = true }
 # Data structures
 hashbrown = { workspace = true, features = ["serde", "rayon"] }
-rustfs-mimalloc = { workspace = true }
+mimalloc = { workspace = true }
 
 [target.'cfg(target_os = "linux")'.dependencies]
 libsystemd.workspace = true
 
 [target.'cfg(not(target_os = "windows"))'.dependencies]
-rustfs-mimalloc-sys.workspace = true
+libmimalloc-sys.workspace = true
 
 [dev-dependencies]
 uuid = { workspace = true, features = ["v4", "v5", "fast-rng", "macro-diagnostics"] }

rustfs/src/allocator_reclaim.rs

@@ -369,8 +369,14 @@ pub fn allocator_reclaim_controller_snapshot(ctx: &CancellationToken) -> Allocat
 }
 
 #[cfg(not(target_os = "windows"))]
+#[allow(unsafe_code)]
 fn collect_allocator_memory(force: bool) -> Result<(), String> {
-    rustfs_mimalloc::MiMalloc::collect(force);
+    // SAFETY: `mi_collect` is provided by the active global allocator backend
+    // on this target family. It is explicitly intended to reclaim retained
+    // pages/segments and does not require additional invariants from the caller.
+    unsafe {
+        libmimalloc_sys::mi_collect(force);
+    }
     Ok(())
 }
 

rustfs/src/main.rs

@@ -26,22 +26,22 @@ struct MiMallocAllocator;
 unsafe impl GlobalAlloc for MiMallocAllocator {
     unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
         // SAFETY: the caller upholds GlobalAlloc's contract for layout.
-        unsafe { rustfs_mimalloc::MiMalloc.alloc(layout) }
+        unsafe { mimalloc::MiMalloc.alloc(layout) }
     }
 
     unsafe fn alloc_zeroed(&self, layout: Layout) -> *mut u8 {
         // SAFETY: the caller upholds GlobalAlloc's contract for layout.
-        unsafe { rustfs_mimalloc::MiMalloc.alloc_zeroed(layout) }
+        unsafe { mimalloc::MiMalloc.alloc_zeroed(layout) }
     }
 
     unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
         // SAFETY: ptr and layout came from this allocator and are forwarded unchanged.
-        unsafe { rustfs_mimalloc::MiMalloc.dealloc(ptr, layout) }
+        unsafe { mimalloc::MiMalloc.dealloc(ptr, layout) }
     }
 
     unsafe fn realloc(&self, ptr: *mut u8, layout: Layout, new_size: usize) -> *mut u8 {
         // SAFETY: ptr and layout came from this allocator and are forwarded unchanged.
-        unsafe { rustfs_mimalloc::MiMalloc.realloc(ptr, layout, new_size) }
+        unsafe { mimalloc::MiMalloc.realloc(ptr, layout, new_size) }
     }
 }
 
@@ -51,7 +51,7 @@ static GLOBAL: hotpath::CountingAllocator<MiMallocAllocator> = hotpath::Counting
 
 #[cfg(not(all(feature = "hotpath", feature = "hotpath-alloc")))]
 #[global_allocator]
-static GLOBAL: rustfs_mimalloc::MiMalloc = rustfs_mimalloc::MiMalloc;
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
 
 fn main() {
     let _hotpath_guard = hotpath::HotpathGuardBuilder::new("main").build();
@@ -71,9 +71,8 @@ mod tests {
         allocation.extend_from_slice(&[7_u8; 64]);
 
         assert_eq!(allocation.len(), 64);
-        let heap = rustfs_mimalloc::heap::Heap::main();
         // SAFETY: the live Vec pointer is valid to inspect for heap ownership.
-        assert!(unsafe { heap.contains(allocation.as_ptr()) });
+        assert!(unsafe { libmimalloc_sys::mi_is_in_heap_region(allocation.as_ptr().cast()) });
     }
 
     #[test]
@@ -86,13 +85,12 @@ mod tests {
         let layout = Layout::from_size_align(32, 8).expect("valid test allocation layout");
         let grown_layout = Layout::from_size_align(64, 8).expect("valid grown test allocation layout");
         let allocator = super::MiMallocAllocator;
-        let heap = rustfs_mimalloc::heap::Heap::main();
 
         // SAFETY: The pointer is checked for null before use and later released
         // through the same allocator with the corresponding layout.
         let ptr = unsafe { allocator.alloc_zeroed(layout) };
         assert!(!ptr.is_null());
-        assert!(unsafe { heap.contains(ptr) });
+        assert!(unsafe { libmimalloc_sys::mi_is_in_heap_region(ptr.cast()) });
         assert!(unsafe { std::slice::from_raw_parts(ptr, 32).iter().all(|byte| *byte == 0) });
 
         // SAFETY: `ptr` was allocated by `allocator` with `layout`; on failure
@@ -104,7 +102,7 @@ mod tests {
             panic!("mimalloc realloc failed in allocator smoke test");
         }
 
-        assert!(unsafe { heap.contains(grown_ptr) });
+        assert!(unsafe { libmimalloc_sys::mi_is_in_heap_region(grown_ptr.cast()) });
         // SAFETY: `grown_ptr` was reallocated by `allocator` and is released
         // with the matching grown layout.
         unsafe { allocator.dealloc(grown_ptr, grown_layout) };

rustfs/src/memory_observability.rs

@@ -17,7 +17,10 @@ use rustfs_io_metrics::{
     record_cpu_usage, record_memory_usage, record_process_memory_split,
 };
 use serde::Serialize;
+#[cfg(any(test, not(target_os = "windows")))]
 use serde_json::Value;
+#[cfg(not(target_os = "windows"))]
+use std::ffi::CStr;
 use std::path::Path;
 use std::sync::{Arc, Mutex, OnceLock};
 use std::time::Duration;
@@ -228,18 +231,7 @@ fn read_cgroup_memory_snapshot() -> Option<CgroupMemorySnapshot> {
     read_cgroup_v2().or_else(read_cgroup_v1)
 }
 
-fn read_allocator_memory_snapshot() -> Option<AllocatorMemorySnapshot> {
-    let json = rustfs_mimalloc::MiMalloc::stats_json();
-    if json.is_empty() {
-        return None;
-    }
-    let observation = parse_mimalloc_stats_json(&json)?;
-    Some(AllocatorMemorySnapshot {
-        backend: crate::allocator_reclaim::allocator_backend(),
-        observation,
-    })
-}
-
+#[cfg(any(test, not(target_os = "windows")))]
 fn numeric_json_value(value: &Value) -> Option<u64> {
     match value {
         Value::Number(number) => number
@@ -250,6 +242,7 @@ fn numeric_json_value(value: &Value) -> Option<u64> {
     }
 }
 
+#[cfg(any(test, not(target_os = "windows")))]
 fn numeric_json_field(value: &Value, field: &str) -> Option<u64> {
     match value {
         Value::Object(fields) => fields
@@ -261,6 +254,7 @@ fn numeric_json_field(value: &Value, field: &str) -> Option<u64> {
     }
 }
 
+#[cfg(any(test, not(target_os = "windows")))]
 fn mimalloc_stat_field(value: &Value, metric: &str, field: &str) -> Option<u64> {
     match value {
         Value::Object(fields) => {
@@ -277,10 +271,12 @@ fn mimalloc_stat_field(value: &Value, metric: &str, field: &str) -> Option<u64>
     }
 }
 
+#[cfg(any(test, not(target_os = "windows")))]
 fn mimalloc_stat_current(value: &Value, metric: &str) -> Option<u64> {
     mimalloc_stat_field(value, metric, "current")
 }
 
+#[cfg(any(test, not(target_os = "windows")))]
 fn mimalloc_stat_sum(value: &Value, metrics: &[&str], field: &str) -> Option<u64> {
     metrics
         .iter()
@@ -289,6 +285,7 @@ fn mimalloc_stat_sum(value: &Value, metrics: &[&str], field: &str) -> Option<u64
         .filter(|value| *value > 0)
 }
 
+#[cfg(any(test, not(target_os = "windows")))]
 fn parse_mimalloc_stats_json(stats_json: &str) -> Option<AllocatorMemoryObservation> {
     let value = serde_json::from_str::<Value>(stats_json).ok()?;
     let malloc_metrics = ["malloc_normal", "malloc_huge"];
@@ -315,6 +312,33 @@ fn parse_mimalloc_stats_json(stats_json: &str) -> Option<AllocatorMemoryObservat
     }
 }
 
+#[cfg(not(target_os = "windows"))]
+#[allow(unsafe_code)]
+fn read_allocator_memory_snapshot() -> Option<AllocatorMemorySnapshot> {
+    // SAFETY: `mi_stats_get_json` returns a null-terminated JSON buffer owned by
+    // mimalloc when called with a null input buffer. The mimalloc API requires
+    // freeing that buffer with `mi_free`; parsing finishes before the buffer is freed.
+    let observation = unsafe {
+        let stats_ptr = libmimalloc_sys::mi_stats_get_json(0, std::ptr::null_mut());
+        if stats_ptr.is_null() {
+            return None;
+        }
+
+        let observation = CStr::from_ptr(stats_ptr).to_str().ok().and_then(parse_mimalloc_stats_json);
+        libmimalloc_sys::mi_free(stats_ptr.cast());
+        observation?
+    };
+    Some(AllocatorMemorySnapshot {
+        backend: crate::allocator_reclaim::allocator_backend(),
+        observation,
+    })
+}
+
+#[cfg(target_os = "windows")]
+fn read_allocator_memory_snapshot() -> Option<AllocatorMemorySnapshot> {
+    None
+}
+
 fn configured_memory_observability_interval_secs() -> u64 {
     rustfs_utils::get_env_u64(ENV_MEMORY_OBSERVABILITY_INTERVAL_SECS, DEFAULT_MEMORY_OBSERVABILITY_INTERVAL_SECS).max(1)
 }
@@ -542,13 +566,6 @@ mod tests {
         assert_eq!(parse_mimalloc_stats_json(r#"{ "allocator": "unknown" }"#), None);
     }
 
-    #[test]
-    fn read_allocator_memory_snapshot_uses_mimalloc_stats_json() {
-        let snapshot = super::read_allocator_memory_snapshot();
-        #[cfg(not(target_os = "windows"))]
-        assert!(snapshot.is_some(), "allocator snapshot should be available on non-Windows");
-    }
-
     #[test]
     fn memory_observability_snapshot_reports_disabled_when_metrics_are_disabled() {
         let snapshot = build_memory_observability_status_snapshot(false, 15, false);