fix(ecstore): retry decommission entries safely on source changes

A single object's SourceChanged during decommission cleanup no longer
cancels the shared worker token and fails the whole pool operation.
Cleanup preflight and source-cleanup outcomes now retry per entry with
bounded attempts and cancellation-aware backoff, applied uniformly to
ordinary versions, delete markers, and tiered copies (removing the
try-once-only branches); every retry re-lists the entry and redoes
version multiset validation before touching the source. Only quorum
loss, unrecoverable system errors, or exceeding a pool-level
SourceChanged exhaustion threshold still fails the decommission, and
exhausted entries never delete their source versions.

Retry attempts, backoff, and deferred/exhausted reasons are logged per
entry for observability. Heavy regression tests spawn on dedicated
32MiB stacks following the existing store-test pattern.

Fixes rustfs/backlog#1913

CLAUDE.md

@@ -30,8 +30,7 @@ make build-docker BUILD_OS=ubuntu22.04
 - Crate membership: `Cargo.toml` `[workspace].members`
 - Architecture, layering, crate map: [ARCHITECTURE.md](ARCHITECTURE.md)
 - Migration guardrails & readiness contracts: [docs/architecture/](docs/architecture/README.md)
-- CI workflow steps: `.github/workflows/`; event, timeout, and required-status
-  matrix: [docs/testing/ci-gates.md](docs/testing/ci-gates.md)
+- CI gates: `.github/workflows/ci.yml` (source of truth; never copy its steps into docs)
 - Test-layer taxonomy, per-layer entry commands, serial/nextest rules, flake
   policy: [docs/testing/README.md](docs/testing/README.md)
 - Tier/ILM transition debugging (xl.meta inspection, versionId tracing):

CONTRIBUTING.md

@@ -70,8 +70,6 @@ make pre-pr
 
 > For the full test-layer taxonomy (unit / ecstore black-box / e2e / s3s-e2e / S3 compatibility / chaos / fuzz / bench), each layer's entry command, the naming conventions the migration gate depends on, and the serial/nextest rules, see [docs/testing/README.md](docs/testing/README.md).
 
-> For the event, timeout, required-status, and local reproduction matrix, see [docs/testing/ci-gates.md](docs/testing/ci-gates.md).
-
 ### 🔒 Automated Pre-commit Hooks
 #### What `make pre-commit` and `make pre-pr` actually run
 

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",
@@ -9564,7 +9587,6 @@ dependencies = [
  "serde",
  "serde_json",
  "serial_test",
- "sha2 0.11.0",
  "temp-env",
  "tempfile",
  "thiserror 2.0.20",
@@ -9853,24 +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/ecstore/src/disk/local.rs

@@ -8000,15 +8000,10 @@ impl DiskAPI for LocalDisk {
             use std::io::Write as _;
 
             let file_path = self.io_get_object_path(volume, path)?;
+            let lock_path = file_path.with_extension("rustfs-cas.lock");
             let path = path.to_string();
             let sync_metadata = effective_durability(volume).syncs_commit_metadata();
             return Ok(tokio::task::spawn_blocking(move || {
-                // A persistent directory lock bounds metadata growth. Removing
-                // per-target lock files can split flock ownership across inodes.
-                let lock_path = file_path
-                    .parent()
-                    .ok_or_else(|| std::io::Error::new(ErrorKind::InvalidInput, "conditional file has no parent"))?
-                    .join(".rustfs-cas.lock");
                 let lock = std::fs::OpenOptions::new()
                     .create(true)
                     .truncate(false)
@@ -8073,25 +8068,7 @@ impl DiskAPI for LocalDisk {
             .map_err(DiskError::from)??);
         }
 
-        #[cfg(windows)]
-        {
-            let file_path = self.io_get_object_path(volume, path)?;
-            let sync_metadata = effective_durability(volume).syncs_commit_metadata();
-            let publication_root = self.publication_root.clone();
-            return Ok(tokio::task::spawn_blocking(move || {
-                os::compare_and_update_control_file(
-                    &file_path,
-                    expected.as_deref(),
-                    replacement.as_deref(),
-                    sync_metadata,
-                    &publication_root,
-                )
-            })
-            .await
-            .map_err(DiskError::from)??);
-        }
-
-        #[cfg(not(any(unix, windows)))]
+        #[cfg(not(unix))]
         {
             let _ = (volume, path, expected, replacement);
             Err(DiskError::MethodNotAllowed)
@@ -21846,9 +21823,9 @@ mod test {
         assert!(matches!(results[1].as_ref().unwrap_err(), DiskError::Io(_)));
     }
 
-    #[cfg(any(unix, windows))]
+    #[cfg(unix)]
     #[tokio::test]
-    async fn windows_and_unix_conditional_file_update_never_deletes_a_new_owner() {
+    async fn conditional_file_update_never_deletes_a_new_owner() {
         use tempfile::tempdir;
 
         let dir = tempdir().expect("temp dir should be created");
@@ -21879,18 +21856,8 @@ mod test {
             disk.read_all(RUSTFS_META_BUCKET, HEALING_MARKER_PATH)
                 .await
                 .expect("new owner marker should remain"),
-            owner_b.clone()
+            owner_b
         );
-        assert_eq!(
-            disk.compare_and_update_file(RUSTFS_META_BUCKET, HEALING_MARKER_PATH, Some(owner_b), None)
-                .await
-                .expect("current owner should remove marker"),
-            ConditionalFileUpdate::Updated
-        );
-        assert!(matches!(
-            disk.read_all(RUSTFS_META_BUCKET, HEALING_MARKER_PATH).await,
-            Err(DiskError::FileNotFound)
-        ));
     }
 
     #[cfg(unix)]
@@ -21905,10 +21872,7 @@ mod test {
         let marker_path = disk
             .get_object_path(RUSTFS_META_BUCKET, HEALING_MARKER_PATH)
             .expect("marker path should resolve");
-        let lock_path = marker_path
-            .parent()
-            .expect("marker path should have a parent")
-            .join(".rustfs-cas.lock");
+        let lock_path = marker_path.with_extension("rustfs-cas.lock");
         let lock = std::fs::OpenOptions::new()
             .create(true)
             .truncate(false)
@@ -21929,40 +21893,6 @@ mod test {
         assert!(matches!(err, DiskError::Io(ref err) if err.kind() == ErrorKind::WouldBlock));
     }
 
-    #[cfg(windows)]
-    #[tokio::test]
-    async fn windows_conditional_file_update_returns_would_block_when_marker_lock_is_contended() {
-        let dir = tempfile::tempdir().expect("temp dir should be created");
-        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
-        let disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
-        ensure_test_volume(&disk, RUSTFS_META_BUCKET).await;
-        let marker_path = disk
-            .get_object_path(RUSTFS_META_BUCKET, HEALING_MARKER_PATH)
-            .expect("marker path should resolve");
-        let lock_path = marker_path
-            .parent()
-            .expect("marker path should have a parent")
-            .join(".rustfs-cas.lock");
-        let lock = std::fs::OpenOptions::new()
-            .create(true)
-            .truncate(false)
-            .read(true)
-            .write(true)
-            .open(lock_path)
-            .expect("marker lock should open");
-        lock.try_lock().expect("marker lock should be held");
-
-        let err = tokio::time::timeout(
-            Duration::from_secs(1),
-            disk.compare_and_update_file(RUSTFS_META_BUCKET, HEALING_MARKER_PATH, None, Some(Bytes::from_static(b"owner"))),
-        )
-        .await
-        .expect("contended conditional update must not block")
-        .expect_err("contended conditional update must retry");
-
-        assert!(matches!(err, DiskError::Io(ref err) if err.kind() == ErrorKind::WouldBlock));
-    }
-
     #[cfg(target_os = "linux")]
     #[tokio::test]
     async fn replacement_io_paths_stay_under_the_mount_lease() {

crates/ecstore/src/disk/os.rs

@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#[cfg(windows)]
-use crate::disk::ConditionalFileUpdate;
 use crate::disk::error::DiskError;
 use crate::disk::error::Result;
 use crate::disk::error_conv::to_file_error;
@@ -3460,89 +3458,6 @@ fn read_windows_relative_file(file_path: &Path, parent_guard: &ExistingBaseDirec
     Ok(Some(data))
 }
 
-#[cfg(windows)]
-pub(crate) fn compare_and_update_control_file(
-    file_path: &Path,
-    expected: Option<&[u8]>,
-    replacement: Option<&[u8]>,
-    sync_metadata: bool,
-    publication_root: &PublicationRoot,
-) -> io::Result<ConditionalFileUpdate> {
-    use windows_sys::{
-        Wdk::Storage::FileSystem::{
-            FILE_NON_DIRECTORY_FILE, FILE_OPEN, FILE_OPEN_IF, FILE_OPEN_REPARSE_POINT, FILE_SYNCHRONOUS_IO_NONALERT,
-        },
-        Win32::Storage::FileSystem::{
-            DELETE, FILE_ATTRIBUTE_NORMAL, FILE_READ_ATTRIBUTES, FILE_SHARE_READ, FILE_SHARE_WRITE, FILE_WRITE_DATA, SYNCHRONIZE,
-        },
-    };
-
-    let parent = file_path
-        .parent()
-        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "conditional file has no parent"))?;
-    let parent_guard = lock_windows_directory_tree(parent, Some(parent), publication_root)?;
-    let lock = open_windows_relative(
-        parent_guard.last_handle()?,
-        std::ffi::OsStr::new(".rustfs-cas.lock"),
-        SYNCHRONIZE | FILE_READ_ATTRIBUTES | FILE_WRITE_DATA,
-        FILE_SHARE_READ | FILE_SHARE_WRITE,
-        FILE_OPEN_IF,
-        FILE_NON_DIRECTORY_FILE | FILE_OPEN_REPARSE_POINT | FILE_SYNCHRONOUS_IO_NONALERT,
-        FILE_ATTRIBUTE_NORMAL,
-        true,
-    )?;
-    validate_windows_owned_file(&lock)?;
-    match lock.as_file().try_lock() {
-        Ok(()) => {}
-        Err(std::fs::TryLockError::WouldBlock) => return Err(io::Error::from(io::ErrorKind::WouldBlock)),
-        Err(std::fs::TryLockError::Error(err)) => return Err(err),
-    }
-
-    let current = read_windows_relative_file(file_path, &parent_guard)?;
-    let matches = match (&current, expected) {
-        (None, None) => true,
-        (Some(current), Some(expected)) => current.as_slice() == expected,
-        _ => false,
-    };
-    if !matches {
-        return Ok(match current {
-            None => ConditionalFileUpdate::Missing,
-            Some(_) => ConditionalFileUpdate::Mismatch,
-        });
-    }
-
-    match replacement {
-        Some(replacement) => RenameDestinationPathGuard {
-            directory: parent.to_path_buf(),
-            _directory_guard: parent_guard,
-        }
-        .write_file_for_path_access(file_path, replacement, sync_metadata, sync_metadata)?,
-        None => {
-            let file_name = file_path
-                .file_name()
-                .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "conditional file must have a name"))?;
-            let file = open_windows_relative(
-                parent_guard.last_handle()?,
-                file_name,
-                DELETE | SYNCHRONIZE | FILE_READ_ATTRIBUTES,
-                FILE_SHARE_READ,
-                FILE_OPEN,
-                FILE_NON_DIRECTORY_FILE | FILE_OPEN_REPARSE_POINT | FILE_SYNCHRONOUS_IO_NONALERT,
-                0,
-                true,
-            )?;
-            validate_windows_owned_file(&file)?;
-            set_windows_file_delete_on_close(&file, true)?;
-            drop(file);
-            if sync_metadata {
-                fsync_dir_std(parent)?;
-            }
-        }
-    }
-
-    Ok(ConditionalFileUpdate::Updated)
-}
-
 #[cfg(windows)]
 fn open_windows_directory_component(
     parent: &WindowsDirectoryHandle,

crates/ecstore/src/set_disk/mod.rs

@@ -4736,7 +4736,15 @@ impl SetDisks {
                 achieved: 0,
             });
         }
-        let parts_metadata = vec![fi.clone(); disks.len()];
+        // Rebuilt tiered metadata starts with index zero, but shuffling validates
+        // each source slot before assigning the shuffled index below.
+        let parts_metadata: Vec<FileInfo> = (0..disks.len())
+            .map(|disk_index| {
+                let mut part = fi.clone();
+                part.erasure.index = fi.erasure.distribution[disk_index];
+                part
+            })
+            .collect();
         let (shuffle_disks, parts_metadata) = Self::shuffle_disks_and_parts_metadata(&disks, &parts_metadata, &fi);
 
         let mut errs = Vec::with_capacity(shuffle_disks.len());

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

@@ -626,7 +626,7 @@ mod tests {
         io::Cursor,
         sync::{
             Arc,
-            atomic::{AtomicBool, Ordering},
+            atomic::{AtomicBool, AtomicUsize, Ordering},
         },
         time::Duration,
     };
@@ -1307,6 +1307,85 @@ mod tests {
         });
     }
 
+    const DECOMMISSION_TEST_FAULT_STAGE_DELETE_MARKER: &str = "delete_marker_copy";
+    const DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT: &str = "migrate_object";
+    #[cfg(feature = "test-util")]
+    const DECOMMISSION_TEST_FAULT_STAGE_TIERED: &str = "decommission_tiered_object";
+
+    async fn seed_decommission_source(
+        store: &Arc<crate::store::ECStore>,
+        bucket: &str,
+        object: &str,
+        body: Vec<u8>,
+        opts: &ObjectOptions,
+    ) {
+        let mut reader = PutObjReader::from_vec(body);
+        store.pools[0]
+            .put_object(bucket, object, &mut reader, opts)
+            .await
+            .expect("seed decommission source object");
+    }
+
+    async fn run_decommission_entry_retry_test(
+        store: &Arc<crate::store::ECStore>,
+        rx: CancellationToken,
+        bucket: &str,
+        object: &str,
+        expected_bucket_incarnation_id: Option<uuid::Uuid>,
+        source_changed_exhaustions: Arc<AtomicUsize>,
+    ) -> crate::error::Result<()> {
+        let source_set = store.pools[0].get_disks_by_key(object);
+        store
+            .decommission_entry_with_retry_state_for_test(
+                rx,
+                0,
+                MetaCacheEntry {
+                    name: object.to_string(),
+                    ..Default::default()
+                },
+                bucket.to_string(),
+                source_set,
+                expected_bucket_incarnation_id,
+                source_changed_exhaustions,
+            )
+            .await
+    }
+
+    async fn read_decommission_target_body(
+        store: &Arc<crate::store::ECStore>,
+        bucket: &str,
+        object: &str,
+        opts: &ObjectOptions,
+    ) -> Vec<u8> {
+        let mut reader = store.pools[1]
+            .get_object_reader(bucket, object, None, HeaderMap::new(), opts)
+            .await
+            .expect("read decommission target object");
+        let mut body = Vec::new();
+        reader
+            .stream
+            .read_to_end(&mut body)
+            .await
+            .expect("drain decommission target body");
+        body
+    }
+
+    async fn assert_decommission_source_absent(
+        store: &Arc<crate::store::ECStore>,
+        bucket: &str,
+        object: &str,
+        opts: &ObjectOptions,
+    ) {
+        let err = store.pools[0]
+            .get_object_info(bucket, object, opts)
+            .await
+            .expect_err("decommission source must be retained until target commit, then removed");
+        assert!(
+            matches!(err, StorageError::ObjectNotFound(_, _) | StorageError::VersionNotFound(_, _, _)),
+            "unexpected decommission source result: {err:?}"
+        );
+    }
+
     async fn write_decommission_test_multipart_source(
         store: &Arc<crate::store::ECStore>,
         pool_idx: usize,
@@ -3102,6 +3181,537 @@ mod tests {
         shutdown.cancel();
     }
 
+    #[test]
+    #[serial_test::serial(storage_class_env)]
+    fn decommission_entry_retries_source_changed_without_canceling_other_bucket() {
+        let handle = std::thread::Builder::new()
+            .name("decommission_entry_retries_source_changed_without_canceling_other_bucket".to_string())
+            .stack_size(32 * 1024 * 1024)
+            .spawn(|| {
+                let runtime = tokio::runtime::Builder::new_multi_thread()
+                    .enable_all()
+                    .worker_threads(2)
+                    .build()
+                    .expect("test runtime should build");
+                runtime.block_on(async {
+                    let temp_dir = tempfile::tempdir().expect("create decommission retry store dir");
+                    let (_ctx, store, shutdown) = without_storage_class_env(build_isolated_test_store(
+                        temp_dir.path(),
+                        "decommission-entry-retry",
+                        &[4, 4],
+                    ))
+                    .await;
+                    crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+                    let changed_bucket = format!("decom-retry-a-{}", uuid::Uuid::new_v4());
+                    let other_bucket = format!("decom-retry-b-{}", uuid::Uuid::new_v4());
+                    for bucket in [&changed_bucket, &other_bucket] {
+                        store
+                            .make_bucket(bucket, &MakeBucketOptions::default())
+                            .await
+                            .expect("create decommission retry bucket");
+                    }
+
+                    let changed_object = "changed.bin";
+                    let first_version = uuid::Uuid::new_v4();
+                    let second_version = uuid::Uuid::new_v4();
+                    let base_time = OffsetDateTime::now_utc();
+                    seed_decommission_source(
+                        &store,
+                        &changed_bucket,
+                        changed_object,
+                        b"first generation".to_vec(),
+                        &ObjectOptions {
+                            versioned: true,
+                            version_id: Some(first_version.to_string()),
+                            mod_time: Some(base_time),
+                            ..Default::default()
+                        },
+                    )
+                    .await;
+                    let other_object = "other.bin";
+                    seed_decommission_source(
+                        &store,
+                        &other_bucket,
+                        other_object,
+                        b"other bucket generation".to_vec(),
+                        &ObjectOptions::default(),
+                    )
+                    .await;
+                    mark_test_pool_decommissioning(&store, 0).await;
+
+                    let mutation_calls = Arc::new(AtomicUsize::new(0));
+                    let mutation_calls_for_hook = Arc::clone(&mutation_calls);
+                    let mutation_store = Arc::clone(&store);
+                    let mutation_bucket = changed_bucket.clone();
+                    let _mutation_guard = crate::core::pools::DecommissionCleanupMutationGuard::install(Arc::new(
+                        move |bucket, object, attempt| {
+                            let is_target = bucket == mutation_bucket.as_str() && object == changed_object;
+                            let calls = Arc::clone(&mutation_calls_for_hook);
+                            let store = Arc::clone(&mutation_store);
+                            let bucket = mutation_bucket.clone();
+                            Box::pin(async move {
+                                if !is_target {
+                                    return;
+                                }
+                                calls.fetch_add(1, Ordering::SeqCst);
+                                if attempt == 1 {
+                                    seed_decommission_source(
+                                        &store,
+                                        &bucket,
+                                        changed_object,
+                                        b"second generation".to_vec(),
+                                        &ObjectOptions {
+                                            versioned: true,
+                                            version_id: Some(second_version.to_string()),
+                                            mod_time: Some(base_time + time::Duration::seconds(1)),
+                                            ..Default::default()
+                                        },
+                                    )
+                                    .await;
+                                }
+                            })
+                        },
+                    ));
+
+                    let ordinary_faults = Arc::new(AtomicUsize::new(0));
+                    let ordinary_faults_for_hook = Arc::clone(&ordinary_faults);
+                    let fault_bucket = other_bucket.clone();
+                    let _fault_guard = crate::core::pools::DecommissionTestFaultGuard::install(Arc::new(
+                        move |stage, bucket, object, attempt| {
+                            let injected = stage == DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT
+                                && bucket == fault_bucket.as_str()
+                                && object == other_object
+                                && attempt < crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS;
+                            if injected {
+                                ordinary_faults_for_hook.fetch_add(1, Ordering::SeqCst);
+                            }
+                            injected
+                        },
+                    ));
+
+                    let rx = CancellationToken::new();
+                    let source_changed_exhaustions = Arc::new(AtomicUsize::new(0));
+                    let changed_incarnation = Some(
+                        store
+                            .bucket_incarnation_id(&changed_bucket)
+                            .await
+                            .expect("changed bucket incarnation"),
+                    );
+                    let other_incarnation = Some(
+                        store
+                            .bucket_incarnation_id(&other_bucket)
+                            .await
+                            .expect("other bucket incarnation"),
+                    );
+                    let (changed_result, other_result) = tokio::join!(
+                        run_decommission_entry_retry_test(
+                            &store,
+                            rx.clone(),
+                            &changed_bucket,
+                            changed_object,
+                            changed_incarnation,
+                            Arc::clone(&source_changed_exhaustions),
+                        ),
+                        run_decommission_entry_retry_test(
+                            &store,
+                            rx.clone(),
+                            &other_bucket,
+                            other_object,
+                            other_incarnation,
+                            Arc::clone(&source_changed_exhaustions),
+                        )
+                    );
+                    changed_result.expect("SourceChanged entry retry must converge");
+                    other_result.expect("other bucket entry must continue through ordinary copy retries");
+
+                    assert!(!rx.is_cancelled(), "entry-level SourceChanged must not cancel the shared worker token");
+                    assert_eq!(mutation_calls.load(Ordering::SeqCst), 2, "entry must be re-listed after SourceChanged");
+                    assert_eq!(ordinary_faults.load(Ordering::SeqCst), 2, "ordinary copy must consume the retry budget");
+                    assert_eq!(source_changed_exhaustions.load(Ordering::SeqCst), 0);
+
+                    for (version_id, expected_body) in [
+                        (first_version, b"first generation".as_slice()),
+                        (second_version, b"second generation".as_slice()),
+                    ] {
+                        let opts = ObjectOptions {
+                            versioned: true,
+                            version_id: Some(version_id.to_string()),
+                            ..Default::default()
+                        };
+                        assert_decommission_source_absent(&store, &changed_bucket, changed_object, &opts).await;
+                        assert_eq!(
+                            read_decommission_target_body(&store, &changed_bucket, changed_object, &opts).await,
+                            expected_body
+                        );
+                    }
+                    assert_decommission_source_absent(&store, &other_bucket, other_object, &ObjectOptions::default()).await;
+                    assert_eq!(
+                        read_decommission_target_body(&store, &other_bucket, other_object, &ObjectOptions::default()).await,
+                        b"other bucket generation"
+                    );
+
+                    shutdown.cancel();
+                });
+            })
+            .expect("spawn decommission retry test thread");
+        if let Err(payload) = handle.join() {
+            std::panic::resume_unwind(payload);
+        }
+    }
+
+    #[test]
+    #[serial_test::serial(storage_class_env)]
+    fn decommission_entry_exhausted_source_changed_retains_source_and_records_failure() {
+        let handle = std::thread::Builder::new()
+            .name("decommission_entry_exhausted_source_changed_retains_source_and_records_failure".to_string())
+            .stack_size(32 * 1024 * 1024)
+            .spawn(|| {
+                let runtime = tokio::runtime::Builder::new_multi_thread()
+                    .enable_all()
+                    .worker_threads(2)
+                    .build()
+                    .expect("test runtime should build");
+                runtime.block_on(async {
+                    let temp_dir = tempfile::tempdir().expect("create decommission exhaustion store dir");
+                    let (_ctx, store, shutdown) = without_storage_class_env(build_isolated_test_store(
+                        temp_dir.path(),
+                        "decommission-entry-exhaustion",
+                        &[4, 4],
+                    ))
+                    .await;
+                    crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+                    let bucket = format!("decom-exhausted-{}", uuid::Uuid::new_v4());
+                    let object = "exhausted.bin";
+                    store
+                        .make_bucket(&bucket, &MakeBucketOptions::default())
+                        .await
+                        .expect("create decommission exhaustion bucket");
+                    let original_version = uuid::Uuid::new_v4();
+                    let base_time = OffsetDateTime::now_utc();
+                    seed_decommission_source(
+                        &store,
+                        &bucket,
+                        object,
+                        b"original generation".to_vec(),
+                        &ObjectOptions {
+                            versioned: true,
+                            version_id: Some(original_version.to_string()),
+                            mod_time: Some(base_time),
+                            ..Default::default()
+                        },
+                    )
+                    .await;
+                    mark_test_pool_decommissioning(&store, 0).await;
+
+                    let mutation_calls = Arc::new(AtomicUsize::new(0));
+                    let mutation_calls_for_hook = Arc::clone(&mutation_calls);
+                    let mutation_store = Arc::clone(&store);
+                    let mutation_bucket = bucket.clone();
+                    let _mutation_guard = crate::core::pools::DecommissionCleanupMutationGuard::install(Arc::new(
+                        move |called_bucket, called_object, _attempt| {
+                            let is_target = called_bucket == mutation_bucket.as_str() && called_object == object;
+                            let calls = Arc::clone(&mutation_calls_for_hook);
+                            let store = Arc::clone(&mutation_store);
+                            let bucket = mutation_bucket.clone();
+                            Box::pin(async move {
+                                if !is_target {
+                                    return;
+                                }
+                                let call = calls.fetch_add(1, Ordering::SeqCst) + 1;
+                                let offset = i64::try_from(call).expect("entry retry count should fit i64");
+                                seed_decommission_source(
+                                    &store,
+                                    &bucket,
+                                    object,
+                                    format!("concurrent generation {call}").into_bytes(),
+                                    &ObjectOptions {
+                                        versioned: true,
+                                        version_id: Some(uuid::Uuid::new_v4().to_string()),
+                                        mod_time: Some(base_time + time::Duration::seconds(offset)),
+                                        ..Default::default()
+                                    },
+                                )
+                                .await;
+                            })
+                        },
+                    ));
+
+                    let rx = CancellationToken::new();
+                    let source_changed_exhaustions = Arc::new(AtomicUsize::new(0));
+                    let incarnation = Some(store.bucket_incarnation_id(&bucket).await.expect("bucket incarnation"));
+                    run_decommission_entry_retry_test(
+                        &store,
+                        rx.clone(),
+                        &bucket,
+                        object,
+                        incarnation,
+                        Arc::clone(&source_changed_exhaustions),
+                    )
+                    .await
+                    .expect("entry-level exhaustion must stay local below the pool threshold");
+
+                    assert!(!rx.is_cancelled(), "one exhausted entry must not cancel other bucket workers");
+                    assert_eq!(mutation_calls.load(Ordering::SeqCst), crate::core::pools::DECOMMISSION_ENTRY_MAX_ATTEMPTS);
+                    assert_eq!(source_changed_exhaustions.load(Ordering::SeqCst), 1);
+                    store.pools[0]
+                        .get_object_info(
+                            &bucket,
+                            object,
+                            &ObjectOptions {
+                                versioned: true,
+                                version_id: Some(original_version.to_string()),
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .expect("retry exhaustion must retain the original source version");
+                    let pool_meta = store.pool_meta.read().await;
+                    let info = pool_meta.pools[0]
+                        .decommission
+                        .as_ref()
+                        .expect("decommission progress must be initialized");
+                    assert_eq!(info.items_decommission_failed, 1, "exhausted entry must be visible as failed");
+                    drop(pool_meta);
+
+                    shutdown.cancel();
+                });
+            })
+            .expect("spawn decommission retry test thread");
+        if let Err(payload) = handle.join() {
+            std::panic::resume_unwind(payload);
+        }
+    }
+
+    #[test]
+    #[serial_test::serial(storage_class_env)]
+    fn decommission_entry_delete_marker_copy_retries_real_path() {
+        let handle = std::thread::Builder::new()
+            .name("decommission_entry_delete_marker_copy_retries_real_path".to_string())
+            .stack_size(32 * 1024 * 1024)
+            .spawn(|| {
+                let runtime = tokio::runtime::Builder::new_multi_thread()
+                    .enable_all()
+                    .worker_threads(2)
+                    .build()
+                    .expect("test runtime should build");
+                runtime.block_on(async {
+                    let temp_dir = tempfile::tempdir().expect("create delete marker retry store dir");
+                    let (_ctx, store, shutdown) = without_storage_class_env(build_isolated_test_store(
+                        temp_dir.path(),
+                        "decommission-delete-marker-retry",
+                        &[4, 4],
+                    ))
+                    .await;
+                    crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+                    let bucket = format!("decom-marker-{}", uuid::Uuid::new_v4());
+                    let object = "marker.bin";
+                    store
+                        .make_bucket(&bucket, &MakeBucketOptions::default())
+                        .await
+                        .expect("create delete marker retry bucket");
+                    let data_version = uuid::Uuid::new_v4();
+                    let marker_version = uuid::Uuid::new_v4();
+                    let base_time = OffsetDateTime::now_utc();
+                    seed_decommission_source(
+                        &store,
+                        &bucket,
+                        object,
+                        b"delete marker data".to_vec(),
+                        &ObjectOptions {
+                            versioned: true,
+                            version_id: Some(data_version.to_string()),
+                            mod_time: Some(base_time),
+                            ..Default::default()
+                        },
+                    )
+                    .await;
+                    store.pools[0]
+                        .delete_object(
+                            &bucket,
+                            object,
+                            ObjectOptions {
+                                versioned: true,
+                                version_id: Some(marker_version.to_string()),
+                                delete_marker: true,
+                                mod_time: Some(base_time + time::Duration::seconds(1)),
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .expect("seed source delete marker");
+                    mark_test_pool_decommissioning(&store, 0).await;
+
+                    let fault_calls = Arc::new(AtomicUsize::new(0));
+                    let fault_calls_for_hook = Arc::clone(&fault_calls);
+                    let fault_bucket = bucket.clone();
+                    let _fault_guard = crate::core::pools::DecommissionTestFaultGuard::install(Arc::new(
+                        move |stage, called_bucket, called_object, attempt| {
+                            let injected = stage == DECOMMISSION_TEST_FAULT_STAGE_DELETE_MARKER
+                                && called_bucket == fault_bucket.as_str()
+                                && called_object == object
+                                && attempt < crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS;
+                            if injected {
+                                fault_calls_for_hook.fetch_add(1, Ordering::SeqCst);
+                            }
+                            injected
+                        },
+                    ));
+
+                    let incarnation = Some(store.bucket_incarnation_id(&bucket).await.expect("bucket incarnation"));
+                    run_decommission_entry_retry_test(
+                        &store,
+                        CancellationToken::new(),
+                        &bucket,
+                        object,
+                        incarnation,
+                        Arc::new(AtomicUsize::new(0)),
+                    )
+                    .await
+                    .expect("delete marker copy retries must converge");
+
+                    assert_eq!(
+                        fault_calls.load(Ordering::SeqCst),
+                        crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS - 1
+                    );
+                    let marker_opts = ObjectOptions {
+                        versioned: true,
+                        version_id: Some(marker_version.to_string()),
+                        ..Default::default()
+                    };
+                    let target_marker = store.pools[1]
+                        .get_object_info(&bucket, object, &marker_opts)
+                        .await
+                        .expect("target delete marker must exist");
+                    assert!(target_marker.delete_marker);
+                    assert_decommission_source_absent(&store, &bucket, object, &marker_opts).await;
+
+                    let data_opts = ObjectOptions {
+                        versioned: true,
+                        version_id: Some(data_version.to_string()),
+                        ..Default::default()
+                    };
+                    assert_eq!(
+                        read_decommission_target_body(&store, &bucket, object, &data_opts).await,
+                        b"delete marker data"
+                    );
+                    assert_decommission_source_absent(&store, &bucket, object, &data_opts).await;
+
+                    shutdown.cancel();
+                });
+            })
+            .expect("spawn decommission retry test thread");
+        if let Err(payload) = handle.join() {
+            std::panic::resume_unwind(payload);
+        }
+    }
+
+    #[cfg(feature = "test-util")]
+    #[test]
+    #[serial_test::serial(storage_class_env)]
+    fn decommission_entry_tiered_copy_retries_real_path() {
+        let handle = std::thread::Builder::new()
+            .name("decommission_entry_tiered_copy_retries_real_path".to_string())
+            .stack_size(32 * 1024 * 1024)
+            .spawn(|| {
+                let runtime = tokio::runtime::Builder::new_multi_thread()
+                    .enable_all()
+                    .worker_threads(2)
+                    .build()
+                    .expect("test runtime should build");
+                runtime.block_on(async {
+                    let temp_dir = tempfile::tempdir().expect("create tiered retry store dir");
+                    let (ctx, store, shutdown) = without_storage_class_env(build_isolated_test_store(
+                        temp_dir.path(),
+                        "decommission-tiered-retry",
+                        &[4, 4],
+                    ))
+                    .await;
+                    crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+                    let bucket = format!("decom-tiered-{}", uuid::Uuid::new_v4());
+                    let object = "tiered.bin";
+                    store
+                        .make_bucket(&bucket, &MakeBucketOptions::default())
+                        .await
+                        .expect("create tiered retry bucket");
+                    let mut reader = PutObjReader::from_vec(b"tiered generation".to_vec());
+                    let original = store.pools[0]
+                        .put_object(&bucket, object, &mut reader, &ObjectOptions::default())
+                        .await
+                        .expect("seed tiered source object");
+                    let tier_name = format!("DECOM{}", &uuid::Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+                    register_mock_tier(&ctx.tier_config_mgr(), &tier_name).await;
+                    store.pools[0]
+                        .transition_object(
+                            &bucket,
+                            object,
+                            &ObjectOptions {
+                                transition: TransitionOptions {
+                                    status: TRANSITION_PENDING.to_string(),
+                                    tier: tier_name,
+                                    etag: original.etag.clone().expect("tiered source ETag"),
+                                    ..Default::default()
+                                },
+                                version_id: original.version_id.map(|version_id| version_id.to_string()),
+                                mod_time: original.mod_time,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .expect("transition source object to mock tier");
+                    mark_test_pool_decommissioning(&store, 0).await;
+
+                    let fault_calls = Arc::new(AtomicUsize::new(0));
+                    let fault_calls_for_hook = Arc::clone(&fault_calls);
+                    let fault_bucket = bucket.clone();
+                    let _fault_guard = crate::core::pools::DecommissionTestFaultGuard::install(Arc::new(
+                        move |stage, called_bucket, called_object, attempt| {
+                            let injected = stage == DECOMMISSION_TEST_FAULT_STAGE_TIERED
+                                && called_bucket == fault_bucket.as_str()
+                                && called_object == object
+                                && attempt < crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS;
+                            if injected {
+                                fault_calls_for_hook.fetch_add(1, Ordering::SeqCst);
+                            }
+                            injected
+                        },
+                    ));
+
+                    let incarnation = Some(store.bucket_incarnation_id(&bucket).await.expect("bucket incarnation"));
+                    run_decommission_entry_retry_test(
+                        &store,
+                        CancellationToken::new(),
+                        &bucket,
+                        object,
+                        incarnation,
+                        Arc::new(AtomicUsize::new(0)),
+                    )
+                    .await
+                    .expect("tiered copy retries must converge");
+
+                    assert_eq!(
+                        fault_calls.load(Ordering::SeqCst),
+                        crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS - 1
+                    );
+                    let target = store.pools[1]
+                        .get_object_info(&bucket, object, &ObjectOptions::default())
+                        .await
+                        .expect("tiered target metadata must exist");
+                    assert_eq!(target.transitioned_object.status, rustfs_filemeta::TRANSITION_COMPLETE);
+                    assert_decommission_source_absent(&store, &bucket, object, &ObjectOptions::default()).await;
+
+                    shutdown.cancel();
+                });
+            })
+            .expect("spawn decommission retry test thread");
+        if let Err(payload) = handle.join() {
+            std::panic::resume_unwind(payload);
+        }
+    }
+
     #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
     #[serial_test::serial(storage_class_env)]
     async fn decommission_outer_fence_loss_blocks_target_put_commit() {

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);

crates/heal/Cargo.toml

@@ -91,7 +91,6 @@ metrics = { workspace = true }
 base64 = { workspace = true }
 bytes = { workspace = true }
 crc-fast = { workspace = true }
-sha2 = { workspace = true }
 
 [dev-dependencies]
 serde_json = { workspace = true, features = ["raw_value"] }

crates/heal/src/heal/erasure_healer.rs

@@ -373,11 +373,6 @@ impl ErasureSetHealer {
         set_disk_id: &str,
         buckets: &[String],
     ) -> Result<(ResumeManager, CheckpointManager)> {
-        if self.replacement_task_id.is_none() && CheckpointManager::is_blocked(&self.disk, task_id).await {
-            return Err(Error::TaskExecutionFailed {
-                message: format!("Resume task {task_id} has a blocked checkpoint"),
-            });
-        }
         // check if resume state exists
         let has_resume_state = if self.replacement_task_id.is_some() {
             ResumeManager::has_replacement_intent(&self.disk, task_id).await

crates/heal/src/heal/resume.rs

@@ -51,7 +51,6 @@ const RESUME_STATE_FILE: &str = "ahm_resume_state.json";
 const REPLACEMENT_INTENT_FILE: &str = "ahm_replacement_intent.json";
 const RESUME_PROGRESS_FILE: &str = "ahm_progress.json";
 pub(super) const RESUME_CHECKPOINT_FILE: &str = "ahm_checkpoint.json";
-pub(super) const RESUME_CHECKPOINT_BLOCKED_FILE: &str = "ahm_checkpoint.blocked";
 const REPLACEMENT_COMPLETION_PROOF_FILE: &str = "ahm_replacement_completion_proof.json";
 const REPLACEMENT_RECOVERY_DIR: &str = "ahm-replacement";
 const REPLACEMENT_INTENT_SEAL_FILE: &str = "ahm_replacement_intent_seal";

crates/heal/src/heal/resume/checkpoint.rs

@@ -13,31 +13,26 @@
 // limitations under the License.
 
 use crate::{Error, Result};
-use base64::Engine as _;
 use serde::{Deserialize, Serialize};
-use sha2::{Digest, Sha256};
 use std::collections::HashSet;
 use std::path::Path;
 use std::sync::{Arc, Mutex};
 use std::time::{SystemTime, UNIX_EPOCH};
-use tokio::sync::{Mutex as AsyncMutex, RwLock};
+use tokio::sync::RwLock;
 use tracing::{debug, warn};
 
-use super::super::storage_api::owner::{EcstoreConditionalFileUpdate, EcstoreDiskAPI, EcstoreDiskBytes};
-use super::super::{BUCKET_META_PREFIX, DiskStore, HealDiskExt, RUSTFS_META_BUCKET};
+use super::super::{BUCKET_META_PREFIX, DiskStore, HealDiskExt as _, RUSTFS_META_BUCKET};
 use super::{
-    LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, PersistThrottle, RESUME_CHECKPOINT_BLOCKED_FILE, RESUME_CHECKPOINT_FILE,
-    delete_resume_file, path_to_str, validate_resume_task_id,
+    LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, PersistThrottle, RESUME_CHECKPOINT_FILE, delete_resume_file, path_to_str,
+    validate_resume_task_id,
 };
 
 const EVENT_HEAL_CHECKPOINT_STATE: &str = "heal_checkpoint_state";
-const RESUME_CHECKPOINT_DIGEST_FILE: &str = "ahm_checkpoint.sha256";
-const CHECKPOINT_PER_VERSION_SCHEMA: u32 = 5;
 
 /// Current on-disk schema version for `ResumeCheckpoint`. Same rationale as
 /// `CURRENT_RESUME_SCHEMA`: pre-per-version dedup identities are not comparable
 /// to the new `compose_key` identities, so a stale checkpoint is discarded.
-pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 6;
+pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 5;
 
 /// resume checkpoint
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -62,11 +57,6 @@ pub struct ResumeCheckpoint {
     pub failed_objects: HashSet<String>,
     /// skipped objects
     pub skipped_objects: HashSet<String>,
-    /// Integrity digest over the checkpoint with this field set to `None`.
-    /// Keeping it in the checkpoint makes the payload and its authentication
-    /// record one CAS generation instead of two independently-written files.
-    #[serde(default)]
-    pub integrity_digest: Option<String>,
 }
 
 impl ResumeCheckpoint {
@@ -80,7 +70,6 @@ impl ResumeCheckpoint {
             processed_objects: HashSet::new(),
             failed_objects: HashSet::new(),
             skipped_objects: HashSet::new(),
-            integrity_digest: None,
         }
     }
 
@@ -127,111 +116,17 @@ pub struct CheckpointManager {
     disk: DiskStore,
     checkpoint: Arc<RwLock<ResumeCheckpoint>>,
     throttle: Mutex<PersistThrottle>,
-    save_lock: AsyncMutex<()>,
-    last_saved: Mutex<Option<EcstoreDiskBytes>>,
 }
 
 impl CheckpointManager {
-    fn blocked_path(task_id: &str) -> std::path::PathBuf {
-        Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}"))
-    }
-
-    /// Return whether a checkpoint was permanently isolated after a malformed
-    /// or unsupported snapshot was observed.
-    pub(crate) async fn is_blocked(disk: &DiskStore, task_id: &str) -> bool {
-        if validate_resume_task_id(task_id).is_err() {
-            return false;
-        }
-        let blocked_path = Self::blocked_path(task_id);
-        let Ok(path) = path_to_str(&blocked_path) else {
-            return false;
-        };
-        match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
-            Ok(_) => true,
-            Err(crate::heal::DiskError::FileNotFound) => false,
-            Err(_) => true,
-        }
-    }
-
-    /// Validate the checkpoint while enumerating resumable state. This reads
-    /// the checkpoint once and also isolates malformed or unsupported data.
-    pub(crate) async fn is_resumable(disk: &DiskStore, task_id: &str) -> Result<bool> {
-        validate_resume_task_id(task_id)?;
-        if Self::is_blocked(disk, task_id).await {
-            return Err(Error::InvalidCheckpoint(format!("Resume task {task_id} has a blocked checkpoint")));
-        }
-        let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
-        let Ok(path) = path_to_str(&file_path) else {
-            return Err(Error::InvalidCheckpoint("Resume checkpoint path is not valid UTF-8".to_string()));
-        };
-        match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
-            Ok(bytes) if bytes.is_empty() => Ok(true),
-            Ok(bytes) => Self::load_from_data(disk.clone(), task_id, bytes.to_vec())
-                .await
-                .map(|_| true),
-            Err(crate::heal::DiskError::FileNotFound) => Ok(true),
-            Err(error) => Err(error.into()),
-        }
-    }
-
-    async fn block_invalid_snapshot(disk: &DiskStore, task_id: &str) {
-        // This marker is intentionally version-agnostic: an unsupported reader
-        // must stop selector retries until an operator cleans up the snapshot.
-        let blocked_path = Self::blocked_path(task_id);
-        let Ok(path) = path_to_str(&blocked_path) else {
-            return;
-        };
-        let result = EcstoreDiskAPI::compare_and_update_file(
-            disk.as_ref(),
-            RUSTFS_META_BUCKET,
-            path,
-            None,
-            Some(EcstoreDiskBytes::from_static(b"blocked")),
-        )
-        .await;
-        match result {
-            Ok(EcstoreConditionalFileUpdate::Updated | EcstoreConditionalFileUpdate::Mismatch) => {}
-            Ok(EcstoreConditionalFileUpdate::Missing) => warn!(
-                target: "rustfs::heal::resume",
-                event = EVENT_HEAL_CHECKPOINT_STATE,
-                component = LOG_COMPONENT_HEAL,
-                subsystem = LOG_SUBSYSTEM_RESUME,
-                task_id,
-                state = "blocked_marker_write_failed",
-                error = "marker target disappeared",
-                "Heal checkpoint could not persist its blocked marker"
-            ),
-            Err(error) => warn!(
-                target: "rustfs::heal::resume",
-                event = EVENT_HEAL_CHECKPOINT_STATE,
-                component = LOG_COMPONENT_HEAL,
-                subsystem = LOG_SUBSYSTEM_RESUME,
-                task_id,
-                state = "blocked_marker_write_failed",
-                error = %error,
-                "Heal checkpoint could not persist its blocked marker"
-            ),
-        }
-    }
-
     /// create new checkpoint manager
     pub async fn new(disk: DiskStore, task_id: String) -> Result<Self> {
         validate_resume_task_id(&task_id)?;
-        let checkpoint_volume = format!("{RUSTFS_META_BUCKET}/{BUCKET_META_PREFIX}");
-        if let Err(error) = EcstoreDiskAPI::make_volume(disk.as_ref(), &checkpoint_volume).await
-            && error != crate::heal::DiskError::VolumeExists
-        {
-            return Err(Error::TaskExecutionFailed {
-                message: format!("Failed to create checkpoint volume: {error}"),
-            });
-        }
         let checkpoint = ResumeCheckpoint::new(task_id);
         let manager = Self {
             disk,
             checkpoint: Arc::new(RwLock::new(checkpoint)),
             throttle: Mutex::new(PersistThrottle::new()),
-            save_lock: AsyncMutex::new(()),
-            last_saved: Mutex::new(None),
         };
 
         // save initial checkpoint
@@ -245,7 +140,6 @@ impl CheckpointManager {
                 error = %e,
                 "Heal checkpoint persistence failed"
             );
-            return Err(e);
         }
         Ok(manager)
     }
@@ -254,22 +148,11 @@ impl CheckpointManager {
     pub async fn load_from_disk(disk: DiskStore, task_id: &str) -> Result<Self> {
         validate_resume_task_id(task_id)?;
         let checkpoint_data = Self::read_checkpoint_file(&disk, task_id).await?;
-        Self::load_from_data(disk, task_id, checkpoint_data).await
-    }
-
-    async fn load_from_data(disk: DiskStore, task_id: &str, checkpoint_data: Vec<u8>) -> Result<Self> {
-        validate_resume_task_id(task_id)?;
-        let mut checkpoint: ResumeCheckpoint = match serde_json::from_slice(&checkpoint_data) {
-            Ok(checkpoint) => checkpoint,
-            Err(error) => {
-                Self::block_invalid_snapshot(&disk, task_id).await;
-                return Err(Error::TaskExecutionFailed {
-                    message: format!("Failed to deserialize checkpoint: {error}"),
-                });
-            }
-        };
+        let mut checkpoint: ResumeCheckpoint =
+            serde_json::from_slice(&checkpoint_data).map_err(|e| Error::TaskExecutionFailed {
+                message: format!("Failed to deserialize checkpoint: {e}"),
+            })?;
         if checkpoint.task_id != task_id {
-            Self::block_invalid_snapshot(&disk, task_id).await;
             return Err(Error::TaskExecutionFailed {
                 message: "Resume checkpoint task id does not match filename".to_string(),
             });
@@ -280,7 +163,6 @@ impl CheckpointManager {
         // identities. Discard the stale sets and position, then stamp the
         // current schema so the scan restarts cleanly.
         if checkpoint.schema_version > CURRENT_CHECKPOINT_SCHEMA {
-            Self::block_invalid_snapshot(&disk, task_id).await;
             return Err(Error::TaskExecutionFailed {
                 message: format!(
                     "Checkpoint schema {} is newer than supported schema {CURRENT_CHECKPOINT_SCHEMA}",
@@ -288,45 +170,7 @@ impl CheckpointManager {
                 ),
             });
         }
-
-        let integrity_verified = if let Some(expected) = checkpoint.integrity_digest.as_deref() {
-            let actual = Self::checkpoint_digest(&Self::serialize_without_digest(&checkpoint)?);
-            if expected != actual {
-                Self::block_invalid_snapshot(&disk, task_id).await;
-                return Err(Error::InvalidCheckpoint(format!(
-                    "Resume checkpoint digest does not match task {task_id}"
-                )));
-            }
-            true
-        } else if checkpoint.schema_version >= CURRENT_CHECKPOINT_SCHEMA {
-            Self::block_invalid_snapshot(&disk, task_id).await;
-            return Err(Error::InvalidCheckpoint(format!(
-                "Resume checkpoint digest is missing for task {task_id}"
-            )));
-        } else {
-            let digest_path = Self::digest_path(task_id);
-            let digest_path = path_to_str(&digest_path)?;
-            match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, digest_path).await {
-                Ok(expected) => {
-                    let actual = Self::checkpoint_digest(&checkpoint_data);
-                    if expected.as_ref() != actual.as_bytes() {
-                        Self::block_invalid_snapshot(&disk, task_id).await;
-                        return Err(Error::InvalidCheckpoint(format!(
-                            "Resume checkpoint digest does not match task {task_id}"
-                        )));
-                    }
-                    true
-                }
-                Err(crate::heal::DiskError::FileNotFound) => false,
-                Err(error) => {
-                    return Err(Error::TaskExecutionFailed {
-                        message: format!("Failed to read checkpoint digest: {error}"),
-                    });
-                }
-            }
-        };
-
-        if checkpoint.schema_version < CHECKPOINT_PER_VERSION_SCHEMA || !integrity_verified {
+        if checkpoint.schema_version < CURRENT_CHECKPOINT_SCHEMA {
             warn!(
                 target: "rustfs::heal::resume",
                 event = EVENT_HEAL_CHECKPOINT_STATE,
@@ -343,15 +187,13 @@ impl CheckpointManager {
             checkpoint.skipped_objects.clear();
             checkpoint.current_bucket_index = 0;
             checkpoint.current_object_index = 0;
+            checkpoint.schema_version = CURRENT_CHECKPOINT_SCHEMA;
         }
-        checkpoint.schema_version = CURRENT_CHECKPOINT_SCHEMA;
 
         Ok(Self {
             disk,
             checkpoint: Arc::new(RwLock::new(checkpoint)),
             throttle: Mutex::new(PersistThrottle::new()),
-            save_lock: AsyncMutex::new(()),
-            last_saved: Mutex::new(Some(EcstoreDiskBytes::from(checkpoint_data))),
         })
     }
 
@@ -362,7 +204,7 @@ impl CheckpointManager {
         }
         let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
         match path_to_str(&file_path) {
-            Ok(path_str) => match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path_str).await {
+            Ok(path_str) => match disk.read_all(RUSTFS_META_BUCKET, path_str).await {
                 Ok(data) => !data.is_empty(),
                 Err(_) => false,
             },
@@ -450,8 +292,6 @@ impl CheckpointManager {
 
         let checkpoint_file = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
         delete_resume_file(&self.disk, &checkpoint_file).await?;
-        delete_resume_file(&self.disk, &Self::digest_path(&task_id)).await?;
-        delete_resume_file(&self.disk, &Self::blocked_path(&task_id)).await?;
 
         debug!(
             target: "rustfs::heal::resume",
@@ -467,130 +307,21 @@ impl CheckpointManager {
 
     /// save checkpoint to disk
     async fn save_checkpoint(&self) -> Result<()> {
-        // Serialize saves and take the snapshot only after acquiring the lock:
-        // a slower writer must not publish a snapshot taken before a newer one.
-        let _save_guard = self.save_lock.lock().await;
-        let checkpoint = self.checkpoint.read().await.clone();
+        let checkpoint = self.checkpoint.read().await;
         validate_resume_task_id(&checkpoint.task_id)?;
-        let unsigned_checkpoint_data = Self::serialize_without_digest(&checkpoint)?;
-        let digest = Self::checkpoint_digest(&unsigned_checkpoint_data);
-        let mut persisted_checkpoint = checkpoint.clone();
-        persisted_checkpoint.integrity_digest = Some(digest);
-        let checkpoint_data =
-            EcstoreDiskBytes::from(serde_json::to_vec(&persisted_checkpoint).map_err(|e| Error::TaskExecutionFailed {
-                message: format!("Failed to serialize checkpoint: {e}"),
-            })?);
+        let checkpoint_data = serde_json::to_vec(&*checkpoint).map_err(|e| Error::TaskExecutionFailed {
+            message: format!("Failed to serialize checkpoint: {e}"),
+        })?;
 
         let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{}_{}", checkpoint.task_id, RESUME_CHECKPOINT_FILE));
 
         let path_str = path_to_str(&file_path)?;
-        let last_saved = self
-            .last_saved
-            .lock()
-            .map_err(|_| Error::TaskExecutionFailed {
-                message: "Checkpoint save state lock is poisoned; refusing to save".to_string(),
-            })?
-            .clone();
-        let update = EcstoreDiskAPI::compare_and_update_file(
-            self.disk.as_ref(),
-            RUSTFS_META_BUCKET,
-            path_str,
-            last_saved.clone(),
-            Some(checkpoint_data.clone()),
-        )
-        .await
-        .map_err(|e| Error::TaskExecutionFailed {
-            message: format!("Failed to save checkpoint: {e}"),
-        })?;
-
-        let expected = match update {
-            EcstoreConditionalFileUpdate::Updated => None,
-            EcstoreConditionalFileUpdate::Missing => {
-                return Err(Error::TaskExecutionFailed {
-                    message: "Checkpoint was removed after this manager saved it; refusing to recreate it".to_string(),
-                });
-            }
-            EcstoreConditionalFileUpdate::Mismatch => {
-                // A healthy manager normally completes the CAS above without
-                // another read or JSON parse. Inspect only after a mismatch so
-                // corruption and future schemas cannot be overwritten blindly.
-                let existing = match HealDiskExt::read_all(self.disk.as_ref(), RUSTFS_META_BUCKET, path_str).await {
-                    Ok(existing) => existing,
-                    Err(crate::heal::DiskError::FileNotFound) => {
-                        return Err(Error::TaskExecutionFailed {
-                            message: "Checkpoint was removed after this manager saved it; refusing to recreate it".to_string(),
-                        });
-                    }
-                    Err(error) => {
-                        return Err(Error::TaskExecutionFailed {
-                            message: format!("Failed to inspect checkpoint after CAS mismatch: {error}"),
-                        });
-                    }
-                };
-
-                if existing.is_empty() && last_saved.is_none() {
-                    Some(existing)
-                } else {
-                    let current: ResumeCheckpoint = match serde_json::from_slice(&existing) {
-                        Ok(current) => current,
-                        Err(error) => {
-                            Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
-                            return Err(Error::TaskExecutionFailed {
-                                message: format!("Existing checkpoint is corrupt: {error}"),
-                            });
-                        }
-                    };
-                    if current.task_id != checkpoint.task_id {
-                        Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
-                        return Err(Error::TaskExecutionFailed {
-                            message: "Existing checkpoint task id does not match filename".to_string(),
-                        });
-                    }
-                    if current.schema_version > CURRENT_CHECKPOINT_SCHEMA {
-                        Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
-                        return Err(Error::TaskExecutionFailed {
-                            message: format!(
-                                "Existing checkpoint schema {} is newer than supported schema {CURRENT_CHECKPOINT_SCHEMA}",
-                                current.schema_version
-                            ),
-                        });
-                    }
-                    if last_saved.as_ref().is_none_or(|saved| saved.as_ref() != existing.as_ref()) {
-                        return Err(Error::TaskExecutionFailed {
-                            message: "Checkpoint changed since this manager loaded it; refusing to overwrite newer progress"
-                                .to_string(),
-                        });
-                    }
-                    Some(existing)
-                }
-            }
-        };
-
-        if let Some(expected) = expected {
-            match EcstoreDiskAPI::compare_and_update_file(
-                self.disk.as_ref(),
-                RUSTFS_META_BUCKET,
-                path_str,
-                Some(expected),
-                Some(checkpoint_data.clone()),
-            )
+        self.disk
+            .write_all(RUSTFS_META_BUCKET, path_str, checkpoint_data.into())
             .await
             .map_err(|e| Error::TaskExecutionFailed {
-                message: format!("Failed to save checkpoint after CAS mismatch: {e}"),
-            })? {
-                EcstoreConditionalFileUpdate::Updated => {}
-                EcstoreConditionalFileUpdate::Missing | EcstoreConditionalFileUpdate::Mismatch => {
-                    return Err(Error::TaskExecutionFailed {
-                        message: "Checkpoint changed while saving; refusing to overwrite newer progress".to_string(),
-                    });
-                }
-            }
-        }
-
-        let mut last_saved = self.last_saved.lock().map_err(|_| Error::TaskExecutionFailed {
-            message: "Checkpoint save state lock is poisoned after save".to_string(),
-        })?;
-        *last_saved = Some(checkpoint_data);
+                message: format!("Failed to save checkpoint: {e}"),
+            })?;
 
         debug!(
             target: "rustfs::heal::resume",
@@ -610,38 +341,11 @@ impl CheckpointManager {
         let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
 
         let path_str = path_to_str(&file_path)?;
-        HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path_str)
+        disk.read_all(RUSTFS_META_BUCKET, path_str)
             .await
             .map(|bytes| bytes.to_vec())
             .map_err(|e| Error::TaskExecutionFailed {
                 message: format!("Failed to read checkpoint file: {e}"),
             })
     }
-
-    fn serialize_without_digest(checkpoint: &ResumeCheckpoint) -> Result<Vec<u8>> {
-        let mut unsigned = checkpoint.clone();
-        unsigned.integrity_digest = None;
-        let mut value = serde_json::to_value(&unsigned).map_err(|e| Error::TaskExecutionFailed {
-            message: format!("Failed to serialize checkpoint: {e}"),
-        })?;
-        for field in ["processed_objects", "failed_objects", "skipped_objects"] {
-            let Some(values) = value.get_mut(field).and_then(serde_json::Value::as_array_mut) else {
-                return Err(Error::TaskExecutionFailed {
-                    message: format!("Failed to canonicalize checkpoint field: {field}"),
-                });
-            };
-            values.sort_by(|left, right| left.as_str().cmp(&right.as_str()));
-        }
-        serde_json::to_vec(&value).map_err(|e| Error::TaskExecutionFailed {
-            message: format!("Failed to serialize checkpoint: {e}"),
-        })
-    }
-
-    fn checkpoint_digest(checkpoint_data: &[u8]) -> String {
-        base64::engine::general_purpose::STANDARD.encode(Sha256::digest(checkpoint_data))
-    }
-
-    fn digest_path(task_id: &str) -> std::path::PathBuf {
-        Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_DIGEST_FILE}"))
-    }
 }

crates/heal/src/heal/resume/tests.rs

@@ -1600,32 +1600,6 @@ async fn test_checkpoint_schema_v4_discarded_on_load() {
     temp_dir.close().expect("remove schema test directory");
 }
 
-#[tokio::test]
-async fn downgraded_unsigned_checkpoint_resets_untrusted_progress() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
-    manager.add_processed_object("victim-a".to_string()).await.unwrap();
-    manager.update_position(2, 500).await.unwrap();
-    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
-    let bytes = disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path).await.unwrap();
-    let mut downgraded: serde_json::Value = serde_json::from_slice(&bytes).unwrap();
-    downgraded["schema_version"] = serde_json::json!(CURRENT_CHECKPOINT_SCHEMA - 1);
-    downgraded.as_object_mut().unwrap().remove("integrity_digest");
-    downgraded["processed_objects"] = serde_json::json!(["victim-b"]);
-    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, serde_json::to_vec(&downgraded).unwrap().into())
-        .await
-        .expect("write downgraded checkpoint");
-
-    let manager = CheckpointManager::load_from_disk(disk, &task_id).await.unwrap();
-    let checkpoint = manager.get_checkpoint().await;
-    assert_eq!(checkpoint.schema_version, CURRENT_CHECKPOINT_SCHEMA);
-    assert_eq!(checkpoint.current_bucket_index, 0);
-    assert_eq!(checkpoint.current_object_index, 0);
-    assert!(checkpoint.processed_objects.is_empty());
-    temp_dir.close().unwrap();
-}
-
 #[tokio::test]
 async fn current_normal_resume_schema_preserves_progress() {
     let (temp_dir, disk) = schema_test_disk().await;
@@ -1701,369 +1675,6 @@ async fn future_resume_and_checkpoint_schemas_are_rejected() {
     temp_dir.close().expect("remove schema test directory");
 }
 
-#[tokio::test]
-async fn checkpoint_save_does_not_replace_a_non_empty_truncated_snapshot() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
-        .await
-        .expect("create checkpoint manager");
-    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
-    let truncated = b"{\"schema_version\":5,\"task_id\":";
-    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, truncated.as_slice().into())
-        .await
-        .expect("write truncated checkpoint fixture");
-
-    let error = manager
-        .update_position(2, 7)
-        .await
-        .expect_err("a truncated checkpoint must fail closed during save");
-    assert!(error.to_string().contains("Existing checkpoint is corrupt"));
-    assert_eq!(
-        disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
-            .await
-            .expect("read truncated checkpoint fixture"),
-        truncated.as_slice()
-    );
-    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
-    temp_dir.close().expect("remove checkpoint save test directory");
-}
-
-#[tokio::test]
-async fn checkpoint_save_does_not_replace_a_future_schema_snapshot() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
-        .await
-        .expect("create checkpoint manager");
-    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
-    let mut future = ResumeCheckpoint::new(task_id.clone());
-    future.schema_version = CURRENT_CHECKPOINT_SCHEMA + 1;
-    let future_bytes = serde_json::to_vec(&future).expect("serialize future checkpoint fixture");
-    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, future_bytes.clone().into())
-        .await
-        .expect("write future checkpoint fixture");
-
-    let error = manager
-        .update_position(2, 7)
-        .await
-        .expect_err("a future schema must fail closed during save");
-    assert!(error.to_string().contains("Existing checkpoint schema"));
-    assert_eq!(
-        disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
-            .await
-            .expect("read future checkpoint fixture"),
-        future_bytes
-    );
-    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
-    temp_dir.close().expect("remove future schema test directory");
-}
-
-#[tokio::test]
-async fn checkpoint_digest_rejects_same_length_progress_tampering() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
-        .await
-        .expect("create checkpoint manager");
-    manager
-        .add_processed_object("victim-a".to_string())
-        .await
-        .expect("persist checkpoint progress");
-    manager.update_position(1, 1).await.expect("flush checkpoint progress");
-    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
-    let original = disk
-        .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
-        .await
-        .expect("read checkpoint fixture");
-    let tampered = original
-        .windows(b"victim-a".len())
-        .position(|window| window == b"victim-a")
-        .map(|index| {
-            let mut bytes = original.to_vec();
-            bytes[index..index + b"victim-a".len()].copy_from_slice(b"victim-b");
-            bytes
-        })
-        .expect("checkpoint should contain the processed object");
-    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, tampered.into())
-        .await
-        .expect("write tampered checkpoint fixture");
-
-    assert!(CheckpointManager::load_from_disk(disk.clone(), &task_id).await.is_err());
-    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
-    temp_dir.close().expect("remove digest test directory");
-}
-
-#[tokio::test]
-async fn checkpoint_integrity_survives_missing_legacy_sidecar() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
-    manager.update_position(2, 9).await.unwrap();
-
-    let digest_path = format!("{BUCKET_META_PREFIX}/{task_id}_ahm_checkpoint.sha256");
-    delete_resume_file(&disk, Path::new(&digest_path)).await.unwrap();
-
-    let restored = CheckpointManager::load_from_disk(disk, &task_id).await.unwrap();
-    let checkpoint = restored.get_checkpoint().await;
-    assert_eq!(checkpoint.current_bucket_index, 2);
-    assert_eq!(checkpoint.current_object_index, 9);
-    assert!(checkpoint.integrity_digest.is_some());
-    temp_dir.close().unwrap();
-}
-
-#[tokio::test]
-async fn checkpoint_integrity_survives_multi_object_reload() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
-    for index in 0..32 {
-        manager.add_processed_object(format!("processed-{index}")).await.unwrap();
-        manager.add_failed_object(format!("failed-{index}")).await.unwrap();
-        manager.add_skipped_object(format!("skipped-{index}")).await.unwrap();
-    }
-    manager.update_position(2, 9).await.unwrap();
-
-    CheckpointManager::load_from_disk(disk, &task_id)
-        .await
-        .expect("a healthy multi-object checkpoint must survive reload");
-    temp_dir.close().unwrap();
-}
-
-#[tokio::test]
-async fn checkpoint_integrity_rejects_a_removed_embedded_digest() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
-    manager.update_position(2, 9).await.unwrap();
-
-    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
-    let bytes = disk
-        .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
-        .await
-        .expect("read checkpoint fixture");
-    let mut value: serde_json::Value = serde_json::from_slice(&bytes).unwrap();
-    value["current_object_index"] = serde_json::json!(10);
-    value.as_object_mut().unwrap().remove("integrity_digest");
-    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, serde_json::to_vec(&value).unwrap().into())
-        .await
-        .expect("write tampered checkpoint fixture");
-
-    assert!(
-        CheckpointManager::load_from_disk(disk.clone(), &task_id).await.is_err(),
-        "a current checkpoint without its embedded digest must fail closed"
-    );
-    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
-    temp_dir.close().unwrap();
-}
-
-#[tokio::test]
-async fn new_checkpoint_manager_rebuilds_an_empty_snapshot() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
-    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, EcstoreDiskBytes::new())
-        .await
-        .expect("write empty checkpoint fixture");
-
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
-        .await
-        .expect("a new manager must rebuild an empty checkpoint");
-    manager
-        .update_position(3, 11)
-        .await
-        .expect("rebuilt checkpoint must remain writable");
-    assert!(CheckpointManager::has_checkpoint(&disk, &task_id).await);
-    temp_dir.close().expect("remove empty checkpoint test directory");
-}
-
-#[tokio::test]
-async fn deleted_checkpoint_is_not_recreated_by_an_old_manager() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
-        .await
-        .expect("create checkpoint manager");
-    manager.cleanup().await.expect("delete checkpoint fixture");
-
-    let error = manager
-        .update_position(1, 2)
-        .await
-        .expect_err("an old manager must not resurrect a deleted checkpoint");
-    assert!(error.to_string().contains("removed after this manager saved it"));
-    assert!(!CheckpointManager::has_checkpoint(&disk, &task_id).await);
-    temp_dir.close().expect("remove deleted checkpoint test directory");
-}
-
-#[cfg(unix)]
-#[tokio::test]
-async fn checkpoint_cleanup_leaves_no_task_specific_lock_artifact() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
-        .await
-        .expect("create checkpoint manager");
-    let lock_path = Path::new(BUCKET_META_PREFIX)
-        .join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"))
-        .with_extension("rustfs-cas.lock");
-    let lock_path = temp_dir.path().join(RUSTFS_META_BUCKET).join(lock_path);
-
-    manager.cleanup().await.expect("delete checkpoint fixture");
-
-    assert!(
-        !lock_path.exists(),
-        "successful checkpoint cleanup must not leave a task-specific lock artifact"
-    );
-}
-
-#[tokio::test]
-async fn an_empty_blocked_marker_still_blocks_resume_selection() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
-        .await
-        .expect("create checkpoint manager");
-    let blocked_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
-    disk.write_all(RUSTFS_META_BUCKET, &blocked_path, EcstoreDiskBytes::new())
-        .await
-        .expect("write empty blocked marker fixture");
-
-    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
-    assert!(CheckpointManager::is_resumable(&disk, &task_id).await.is_err());
-    // Recovery requires replacing/cleaning the snapshot, then removing the
-    // marker; ordinary selector retries are intentionally not an unlock path.
-    manager.cleanup().await.expect("clean blocked checkpoint");
-    assert!(!CheckpointManager::is_blocked(&disk, &task_id).await);
-    temp_dir.close().expect("remove empty blocked marker test directory");
-}
-
-#[tokio::test]
-async fn resumable_selector_skips_healthy_tasks_with_blocked_markers() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let tasks = [
-        (ResumeUtils::generate_task_id(), EcstoreDiskBytes::new()),
-        (ResumeUtils::generate_task_id(), EcstoreDiskBytes::from_static(b"blocked")),
-    ];
-    for (task_id, marker) in &tasks {
-        ResumeManager::new(
-            disk.clone(),
-            task_id.clone(),
-            "erasure_set".to_string(),
-            "pool_0_set_0".to_string(),
-            vec!["bucket".to_string()],
-        )
-        .await
-        .expect("create healthy resume state");
-        CheckpointManager::new(disk.clone(), task_id.clone())
-            .await
-            .expect("create healthy checkpoint");
-        let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
-        let checkpoint_bytes = disk
-            .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
-            .await
-            .expect("read healthy checkpoint before blocking");
-        let marker_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
-        disk.write_all(RUSTFS_META_BUCKET, &marker_path, marker.clone())
-            .await
-            .expect("write blocked marker");
-
-        assert!(ResumeUtils::get_resumable_tasks(&disk).await.is_err());
-        assert_eq!(
-            disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
-                .await
-                .expect("read healthy checkpoint after blocking"),
-            checkpoint_bytes
-        );
-    }
-    temp_dir.close().expect("remove blocked selector test directory");
-}
-
-#[tokio::test]
-async fn stale_checkpoint_manager_cannot_overwrite_newer_progress() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let task_id = ResumeUtils::generate_task_id();
-    let first = CheckpointManager::new(disk.clone(), task_id.clone())
-        .await
-        .expect("create first checkpoint manager");
-    let second = CheckpointManager::load_from_disk(disk.clone(), &task_id)
-        .await
-        .expect("load second checkpoint manager");
-
-    second
-        .update_position(4, 20)
-        .await
-        .expect("persist newer checkpoint progress");
-    let error = first
-        .update_position(1, 3)
-        .await
-        .expect_err("stale checkpoint manager must not overwrite newer progress");
-    assert!(error.to_string().contains("newer progress"));
-
-    let persisted = CheckpointManager::load_from_disk(disk.clone(), &task_id)
-        .await
-        .expect("load newer checkpoint progress")
-        .get_checkpoint()
-        .await;
-    assert_eq!(persisted.current_bucket_index, 4);
-    assert_eq!(persisted.current_object_index, 20);
-    temp_dir.close().expect("remove stale manager test directory");
-}
-
-#[tokio::test]
-async fn resumable_selector_isolates_future_and_corrupt_checkpoints() {
-    let (temp_dir, disk) = schema_test_disk().await;
-    let future_task = ResumeUtils::generate_task_id();
-    let corrupt_task = ResumeUtils::generate_task_id();
-    for task_id in [&future_task, &corrupt_task] {
-        ResumeManager::new(
-            disk.clone(),
-            task_id.to_string(),
-            "erasure_set".to_string(),
-            "pool_0_set_0".to_string(),
-            vec!["bucket".to_string()],
-        )
-        .await
-        .expect("create resumable state fixture");
-    }
-
-    let future_path = format!("{BUCKET_META_PREFIX}/{future_task}_{RESUME_CHECKPOINT_FILE}");
-    let mut future = ResumeCheckpoint::new(future_task.clone());
-    future.schema_version = CURRENT_CHECKPOINT_SCHEMA + 1;
-    let future_bytes = serde_json::to_vec(&future).expect("serialize future checkpoint fixture");
-    disk.write_all(RUSTFS_META_BUCKET, &future_path, future_bytes.clone().into())
-        .await
-        .expect("write future checkpoint fixture");
-    let corrupt_path = format!("{BUCKET_META_PREFIX}/{corrupt_task}_{RESUME_CHECKPOINT_FILE}");
-    let corrupt_bytes = b"{truncated";
-    disk.write_all(RUSTFS_META_BUCKET, &corrupt_path, corrupt_bytes.as_slice().into())
-        .await
-        .expect("write corrupt checkpoint fixture");
-
-    assert!(CheckpointManager::is_resumable(&disk, &future_task).await.is_err());
-    assert!(CheckpointManager::is_resumable(&disk, &corrupt_task).await.is_err());
-    assert!(ResumeUtils::get_resumable_tasks(&disk).await.is_err());
-    for (task_id, path, bytes) in [
-        (&future_task, future_path, future_bytes),
-        (&corrupt_task, corrupt_path, corrupt_bytes.to_vec()),
-    ] {
-        assert_eq!(
-            disk.read_all(RUSTFS_META_BUCKET, &path)
-                .await
-                .expect("read isolated checkpoint bytes"),
-            bytes
-        );
-        let blocked_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
-        assert!(
-            !disk
-                .read_all(RUSTFS_META_BUCKET, &blocked_path)
-                .await
-                .expect("read checkpoint blocked marker")
-                .is_empty()
-        );
-    }
-    temp_dir.close().expect("remove selector isolation test directory");
-}
-
 #[test]
 fn test_persist_throttle_batches_until_threshold() {
     let mut throttle = PersistThrottle::new();

crates/heal/src/heal/resume/utils.rs

@@ -21,7 +21,7 @@ use uuid::Uuid;
 use super::super::{BUCKET_META_PREFIX, DiskError, DiskStore, HealDiskExt as _, RUSTFS_META_BUCKET};
 use super::replacement::{ReplacementPhase, ReplacementRecoveryRecord};
 use super::{
-    CheckpointManager, EVENT_HEAL_RESUME_STATE, LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, REPLACEMENT_COMPLETION_PROOF_FILE,
+    EVENT_HEAL_RESUME_STATE, LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, REPLACEMENT_COMPLETION_PROOF_FILE,
     REPLACEMENT_INTENT_FILE, RESUME_STATE_FILE, ResumeManager, ResumeStateFile, is_replacement_intent, path_to_str,
     replacement_recovery_corruption_for_state_load, replacement_recovery_dir, validate_resume_task_id,
 };
@@ -67,7 +67,6 @@ impl ResumeUtils {
                 // Extract task ID from filename: {task_id}_ahm_resume_state.json
                 if let Some(task_id) = entry.strip_suffix(&format!("_{RESUME_STATE_FILE}"))
                     && validate_resume_task_id(task_id).is_ok()
-                    && CheckpointManager::is_resumable(disk, task_id).await?
                 {
                     task_ids.push(task_id.to_string());
                 }

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/ci-gates.md

@@ -1,149 +0,0 @@
-# CI gate matrix
-
-This file is the source of truth for which validation runs on each event, its
-configured wall-clock budget, and whether it can block a merge. Test taxonomy,
-naming, and nextest serialization rules remain in [README.md](README.md); e2e
-membership and counts remain in
-[e2e-suite-inventory.md](e2e-suite-inventory.md).
-
-The distinction between **required** and **report-only** is load-bearing:
-a failing job blocks a merge only when its exact check name is present in the
-live `main` ruleset. A workflow name, a `merge_group` trigger, or a red PR check
-does not make a job required by itself.
-
-## Required merge checks
-
-The live `main` ruleset (`6436880`) currently requires exactly these contexts:
-
-| Required context | Producer | Validation |
-|---|---|---|
-| `CLA Check` | `.github/workflows/cla.yml` | Contributor agreement |
-| `Quick Checks` | `.github/workflows/ci.yml` | Formatting and repository guard scripts |
-| `Test and Lint` | `.github/workflows/ci.yml` | Clippy, workspace nextest excluding `e2e_test`, doctests, and migration proofs |
-
-For pull requests limited to the paths excluded by the main CI workflow,
-`.github/workflows/ci-docs-only.yml` reports `Quick Checks` and
-`Test and Lint` under the same names. It runs the real quick checks and the
-planning-document guard; it does not claim that Rust compilation or runtime
-tests ran. Despite the workflow name, these paths also include selected deploy,
-workflow, and lock files.
-
-Verify the live rule rather than trusting this snapshot before changing merge
-policy:
-
-```bash
-gh api repos/rustfs/rustfs/rulesets/6436880 \
-  --jq '.rules[] | select(.type == "required_status_checks") | .parameters'
-```
-
-The ruleset currently has `strict_required_status_checks_policy=false`.
-`Continuous Integration` accepts `merge_group` events and runs `e2e-full` for
-them, but `End-to-End Tests (full merge gate)` is not currently a required
-context. Therefore the repository is prepared to test a merge-queue SHA, but
-the workflow alone does not prove that every merge passed that lane.
-
-## Pull request and merge matrix
-
-Budgets below are job `timeout-minutes`, not typical runtimes. “Report-only”
-means the result is visible and actionable but is not in the live required
-context list.
-
-| Event | Validation | Budget | Merge status | Reproduction |
-|---|---|---:|---|---|
-| PR, non-doc change | `Quick Checks` | 10 min | Required | `make pre-commit` (broader local umbrella) |
-| PR, non-doc change | `Test and Lint` | 90 min | Required | `cargo nextest run --profile ci --all --exclude e2e_test` |
-| PR, non-doc change | `Typos` | 10 min | Report-only | `typos` |
-| PR, non-doc change | `ILM Integration (serial)` | 90 min | Report-only | Use the exact command in `.github/workflows/ci.yml` |
-| PR, non-doc change | rio-v2 / swift / sftp test-and-lint variants | 90 min each | Report-only | `cargo nextest run` with the workflow's feature set |
-| PR, non-doc change | `Build RustFS Debug Binary` | 30 min | Report-only; prerequisite for black-box lanes | `cargo build -p rustfs --bins` |
-| PR, non-doc change | `io_uring Integration (real)` | 30 min | Report-only | `cargo test -p rustfs-ecstore --lib uring_ -- --test-threads=1 --nocapture` |
-| PR, non-doc change | `End-to-End Tests` (`e2e-smoke` plus `s3s-e2e`) | 30 min | Report-only | `cargo nextest run --profile e2e-smoke -p e2e_test`; then `./scripts/e2e-run.sh ./target/debug/rustfs <data-dir>` |
-| PR, non-doc change | `S3 Implemented Tests` | 60 min | Report-only | Build `rustfs`, then run `scripts/s3-tests/run.sh` with `DEPLOY_MODE=binary`, `TEST_MODE=single`, and `MAXFAIL=0` |
-| PR, non-doc change | `S3 Lifecycle Behavior Tests` | 30 min | Report-only | Use the accelerated scanner environment in `.github/workflows/ci.yml` with `scripts/s3-tests/run.sh` |
-| PR touching dependency or workflow inputs | Cargo Deny / Workflow Pin Report / Dependency Review | 20 / 5 / 30 min | Report-only | `cargo deny check`; `scripts/security/check_workflow_pins.sh` |
-| PR touching architecture rules or architecture docs | `Architecture Migration Rules` | 10 min | Report-only | `scripts/check_architecture_migration_rules.sh` |
-| PR touching Nix or workspace manifests | `Nix Build & Check` | 60 min | Report-only | `nix flake check` |
-| PR limited to main-CI-excluded paths | companion `Quick Checks` and `Test and Lint` | 10 min each | Required | `git diff --check`; `make doc-paths-check` when documentation paths changed |
-| `merge_group` | Standard CI plus `e2e-full` | 55 min for `e2e-full` | Standard required contexts only; `e2e-full` report-only | `cargo nextest run --profile e2e-full -p e2e_test` |
-| Push to `main` | Standard CI plus `e2e-full` | 55 min for `e2e-full` | Post-merge detection | Same as `merge_group` |
-| PR touching fuzz inputs or harness paths | Build plus five 60-second fuzz smoke targets | 60 min build; 30 min per target | Report-only | `MAX_TOTAL_TIME=60 ./scripts/fuzz/run.sh` |
-| PR touching selected ecstore disk/format paths | `Rename Safety` on Windows | 60 min | Report-only | Run the four `cargo test -p rustfs-ecstore --lib <filter>` commands in `windows-filesystem.yml` on Windows |
-
-The authoritative e2e filters live in `.config/nextest.toml`; extend a profile
-instead of adding a second ad-hoc selector. Before a profile runs,
-`scripts/check_test_wiring.py` compares its exact membership to the committed
-digest so a silent test drop fails closed.
-
-## Scheduled and manual validation
-
-Scheduled lanes are independent fault domains. They do not block a pull
-request, but their workflow-local gate can fail the run and scheduled failures
-are routed to the shared failure-issue action. The scheduled-validation
-watchdog and freshness workflow separately detect incomplete runs and missing
-schedules.
-
-| Cadence (UTC unless noted) | Workflow / validation | Budget | Verdict and artifacts | Reproduction |
-|---|---|---:|---|---|
-| Daily 02:17 | Fuzz: five nightly corpus targets | 60 min build; 60 min per target | Gate; corpus/crash artifacts, scheduled failure alert | `MAX_TOTAL_TIME=<seconds> ./scripts/fuzz/run.sh` |
-| Daily 03:17 | MinIO interop (EC + SSE read parity) | 40 min | Gate; scheduled failure alert | Dispatch `minio-interop.yml` or follow its pinned Docker fixture steps |
-| Daily 04:29 | Replication / cluster-fault / protocol e2e | 45 / 90 / 90 min | Three independent gates; JUnit, membership, and server logs | `cargo nextest run --profile e2e-repl-nightly -p e2e_test`; `--profile e2e-nightly`; `-j 1 --profile e2e-protocols` |
-| Daily 06:31 | Warp performance A/B | 180 min | Regression budget gate; A/B summaries and server logs | `bash scripts/run_hotpath_warp_abba.sh --help` |
-| Daily 00:07 Asia/Shanghai (16:07 UTC previous day) | Nightly GNU build and Vault lanes | 150 / 90 / 60 min | Build, live Vault, and HA failover gates | Use the commands and pinned Vault images in `nightly-gnu.yml` |
-| Daily 03:23 | Security Audit | 20 / 5 min, plus 30 min on PR dependency review | Cargo Deny and workflow-pin gates; scheduled failure alert | `cargo deny check`; `scripts/security/check_workflow_pins.sh` |
-| Daily 23:47 | Scheduled Validation Freshness | 10 min | Fails when a critical schedule was never created or is stale | Dispatch `scheduled-validation-freshness.yml` |
-| Sunday 00:11 | Full `Continuous Integration` matrix | Per-job budgets above | Weekly variant coverage, including dormant rio-v2 binary/e2e lanes | Dispatch `ci.yml` |
-| Sunday 01:13 | Seven-platform build matrix | 150 min per platform | Build/package integrity; scheduled failure alert | Dispatch `build.yml` with an exact platform set |
-| Sunday 02:19 | Ceph s3-tests full sweep: single and real four-node, four shards each | 180 min per shard | Compatibility gate; report, JUnit, exact node IDs, and server logs | `scripts/s3-tests/run.sh` against an existing single or distributed target |
-| Sunday 06:41 | Mint | 120 min | **Report-only by design**; per-suite PASS/FAIL/NA and raw `log.json` | Reproduce the pinned Docker sequence in `mint.yml` or dispatch it |
-| Sunday 07:43 | Workspace line coverage | 120 min | Report-only trend; lcov and JSON retained 90 days | `make coverage` |
-| Monthly, day 1 06:37 | Runner Hygiene | 15 min | Validates runner ephemerality; scheduled failure alert | Dispatch `runner-hygiene.yml` |
-
-Manual `workflow_dispatch` exists for the scheduled workflows above. Manual
-runs are debugging evidence and intentionally do not open scheduled-failure
-issues. A manual performance run may explicitly allow a known regression; that
-override must not be treated as an ordinary passing baseline.
-
-## Release validation
-
-Release validation is post-merge and tag-driven; it does not substitute for a
-pull-request gate.
-
-| Event | Validation | Budget | Result |
-|---|---|---:|---|
-| Push to `main` or weekly schedule | `Build and Release` platform matrix | 150 min per platform | Build artifacts for all selected targets; no release publication on a main push |
-| Valid release or preview tag | `Build and Release` plus asset checks | 150 min per platform | Draft release, checksummed assets, and publish step |
-| Successful non-preview release-tag build | Docker image build and image scan | 60 min build; 30 min scan | Multi-architecture images plus vulnerability report |
-| Successful release-tag build | DEB/RPM packaging | 30 min per architecture | Packages and checksum files uploaded to the release |
-| Successful non-preview release-tag build | Helm template test and package | 30 min build; 30 min publish | Versioned chart and repository index |
-
-Use an exact preview tag for end-to-end release rehearsal. Manual dispatches
-are backfill/debug paths and do not prove the automatic `workflow_run` chain.
-
-## Evidence requirements
-
-A green check is useful only when it proves the intended behavior ran:
-
-- Record the exact commit SHA and run URL.
-- Separate product failure from runner prerequisites, service readiness, and
-  cancellation. Repair the precondition, then rerun the exact workload.
-- Preserve membership manifests, JUnit, raw compatibility logs, seeds, and
-  server logs where the workflow provides them.
-- For a bug fix or a new fault checker, provide sensitivity evidence: the old
-  behavior or an intentional mutation must fail the new oracle, and the fixed
-  behavior must pass it.
-- Never promote a report-only lane to required from one green run. Require at
-  least 14 days and 30 representative pull requests with at least 99% complete
-  execution, then update the ruleset and this table together.
-
-## Change checklist
-
-Update this file in the same pull request when any of these change:
-
-- workflow triggers, job names, timeouts, or nextest profile ownership;
-- required status contexts or strict/merge-queue policy;
-- scheduled cadence, alert routing, artifact contract, or local reproduction;
-- report-only versus gating semantics.
-
-Do not copy per-module test counts here. Update
-[e2e-suite-inventory.md](e2e-suite-inventory.md) and its enforced membership
-digest instead.

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);