Merge branch 'main' into fix/b5-t6-a4-metadata-lww

Replace the upstream xonatius/mimalloc_rust.git fork (mimalloc + libmimalloc-sys)
with the published rustfs-mimalloc (v0.5.0) and rustfs-mimalloc-sys (v0.5.0) crates
from crates.io.

The new crates are based on mimalloc V3 (v3.5.0) and provide:
- MiMalloc global allocator with safe API (collect, stats_json, process_info)
- Heap management and arena operations (heap module)
- Full FFI bindings to mimalloc V3

Changes:
- Workspace deps: mimalloc + libmimalloc-sys (git) → rustfs-mimalloc + rustfs-mimalloc-sys (crates.io)
- allocator_reclaim.rs: libmimalloc_sys::mi_collect → rustfs_mimalloc::MiMalloc::collect
- memory_observability.rs: raw FFI mi_stats_get_json → MiMalloc::stats_json()
- main.rs: heap ownership tests use Heap::contains() (V3 API)
- deny.toml: remove xonatius/mimalloc_rust.git from allow-git

Co-authored-by: heihutu <heihutu@gmail.com>

CLAUDE.md

@@ -30,7 +30,8 @@ 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 gates: `.github/workflows/ci.yml` (source of truth; never copy its steps into docs)
+- CI workflow steps: `.github/workflows/`; event, timeout, and required-status
+  matrix: [docs/testing/ci-gates.md](docs/testing/ci-gates.md)
 - 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

@@ -70,6 +70,8 @@ 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 = "1.4.3"
+version = "1.4.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "509591b7bcd67f4ef775afad7662703b4935daaa6ec0e5605cfb1090b32a2b6d"
+checksum = "0ad534f4357a5264cce5019c989cf66a4f0dc4e0d1b1d15f8aacec0ff7360273"
 dependencies = [
  "find-msvc-tools",
  "jobserver",
@@ -2522,12 +2522,6 @@ 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"
@@ -5988,15 +5982,6 @@ 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"
@@ -6397,14 +6382,6 @@ 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"
@@ -9162,13 +9139,11 @@ 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",
@@ -9204,6 +9179,8 @@ dependencies = [
  "rustfs-lock",
  "rustfs-log-analyzer",
  "rustfs-madmin",
+ "rustfs-mimalloc",
+ "rustfs-mimalloc-sys",
  "rustfs-notify",
  "rustfs-object-capacity",
  "rustfs-object-data-cache",
@@ -9875,6 +9852,24 @@ 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
-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"] }
+rustfs-mimalloc = { version = "0.5.0" }
+rustfs-mimalloc-sys = { version = "0.5.0" }
 hotpath = { version = "0.23.3", default-features = false }
 # Snapshot testing for output format regression detection
 insta = { version = "1.48" }

crates/e2e_test/src/lib.rs

@@ -61,6 +61,11 @@ mod get_codec_streaming_compat_test;
 #[cfg(test)]
 mod version_id_regression_test;
 
+// Receiver-side replication LWW (rustfs/backlog#1953): stale inbound
+// replication metadata must not overwrite a newer local category state.
+#[cfg(test)]
+mod replication_lww_receiver_test;
+
 // Data usage regression tests
 #[cfg(test)]
 mod data_usage_test;

crates/e2e_test/src/replication_lww_receiver_test.rs

@@ -0,0 +1,150 @@
+#![cfg(test)]
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Receiver-side replication LWW over the wire (rustfs/backlog#1953, audit
+//! A4/P1-6).
+//!
+//! In an active-active topology both sites' metadata states arrive at the
+//! peer as authorized replication PUTs carrying per-category source
+//! timestamps (`x-rustfs-source-replication-tagging-timestamp` header
+//! family). Before the fix the receiver applied them unconditionally, so a
+//! stale delivery overwrote a newer local state and the two sites diverged
+//! permanently while both reported COMPLETED. This test drives one live
+//! `rustfs` server with simulated inbound replication PUTs for the same
+//! object version and asserts the newer tagging state wins regardless of
+//! delivery order, while a stale delivery still succeeds at the object level
+//! (a failure would loop through MRF re-delivering the stale value).
+
+use crate::common::{RustFSTestEnvironment, init_logging};
+use aws_sdk_s3::Client;
+use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{BucketVersioningStatus, VersioningConfiguration};
+
+type TestResult = Result<(), Box<dyn std::error::Error + Send + Sync>>;
+
+const HDR_SOURCE_REPLICATION_REQUEST: &str = "x-rustfs-source-replication-request";
+const HDR_SOURCE_VERSION_ID: &str = "x-rustfs-source-version-id";
+const HDR_SOURCE_MTIME: &str = "x-rustfs-source-mtime";
+const HDR_SOURCE_TAGGING_TIMESTAMP: &str = "x-rustfs-source-replication-tagging-timestamp";
+
+const SOURCE_MTIME: &str = "2026-01-01T00:00:00Z";
+const T_STALE: &str = "2026-01-01T00:00:01Z";
+const T_LOCAL: &str = "2026-02-01T00:00:00Z";
+const T_NEWER: &str = "2026-03-01T00:00:00Z";
+
+/// Simulated inbound authorized replication PUT: same object version, tags and
+/// the source-authored tagging timestamp carried in transport headers.
+async fn inbound_replication_put(
+    client: &Client,
+    bucket: &str,
+    key: &str,
+    version_id: &str,
+    tags: &str,
+    tagging_timestamp: &str,
+) -> TestResult {
+    let version_id = version_id.to_string();
+    let tagging_timestamp = tagging_timestamp.to_string();
+    client
+        .put_object()
+        .bucket(bucket)
+        .key(key)
+        .body(ByteStream::from_static(b"lww-e2e-body"))
+        .tagging(tags)
+        .customize()
+        .mutate_request(move |req| {
+            req.headers_mut().insert(HDR_SOURCE_REPLICATION_REQUEST, "true");
+            req.headers_mut().insert(HDR_SOURCE_VERSION_ID, version_id.clone());
+            req.headers_mut().insert(HDR_SOURCE_MTIME, SOURCE_MTIME);
+            req.headers_mut()
+                .insert(HDR_SOURCE_TAGGING_TIMESTAMP, tagging_timestamp.clone());
+        })
+        .send()
+        .await?;
+    Ok(())
+}
+
+async fn tag_value(client: &Client, bucket: &str, key: &str, version_id: &str, tag_key: &str) -> Option<String> {
+    let tagging = client
+        .get_object_tagging()
+        .bucket(bucket)
+        .key(key)
+        .version_id(version_id)
+        .send()
+        .await
+        .expect("object tagging should be readable");
+    tagging
+        .tag_set()
+        .iter()
+        .find(|tag| tag.key() == tag_key)
+        .map(|tag| tag.value().to_string())
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn receiver_lww_keeps_newer_tags_across_delivery_orders() -> TestResult {
+    init_logging();
+    let mut env = RustFSTestEnvironment::new().await?;
+    env.start_rustfs_server(vec![]).await?;
+    let client = env.create_s3_client();
+
+    let bucket = "replication-lww-receiver";
+    let key = "object";
+    client.create_bucket().bucket(bucket).send().await?;
+    client
+        .put_bucket_versioning()
+        .bucket(bucket)
+        .versioning_configuration(
+            VersioningConfiguration::builder()
+                .status(BucketVersioningStatus::Enabled)
+                .build(),
+        )
+        .send()
+        .await?;
+
+    // First delivery establishes version V with tags stamped T_LOCAL.
+    let version_id = uuid::Uuid::new_v4().to_string();
+    inbound_replication_put(&client, bucket, key, &version_id, "site=local", T_LOCAL).await?;
+    assert_eq!(
+        tag_value(&client, bucket, key, &version_id, "site").await.as_deref(),
+        Some("local"),
+        "the first delivery must establish the tagged version"
+    );
+
+    // A stale delivery (older source timestamp) must succeed at the object
+    // level but must NOT overwrite the newer tags.
+    inbound_replication_put(&client, bucket, key, &version_id, "site=stale", T_STALE).await?;
+    assert_eq!(
+        tag_value(&client, bucket, key, &version_id, "site").await.as_deref(),
+        Some("local"),
+        "a stale inbound delivery must not overwrite newer tags (rustfs/backlog#1953)"
+    );
+
+    // A newer delivery still converges the version onto the newest state.
+    inbound_replication_put(&client, bucket, key, &version_id, "site=newer", T_NEWER).await?;
+    assert_eq!(
+        tag_value(&client, bucket, key, &version_id, "site").await.as_deref(),
+        Some("newer"),
+        "a newer inbound delivery must overwrite older tags"
+    );
+
+    client
+        .delete_object()
+        .bucket(bucket)
+        .key(key)
+        .version_id(&version_id)
+        .send()
+        .await?;
+    env.delete_test_bucket(bucket).await.ok();
+    Ok(())
+}

crates/ecstore/src/bucket/replication/replication_resyncer.rs

@@ -3868,6 +3868,7 @@ async fn replicate_object_with_multipart<S: ReplicationObjectIO>(ctx: MultipartR
                 actual_size,
                 object_info.etag.clone().unwrap_or_default(),
                 object_info.mod_time,
+                &put_opts.internal,
             ),
         )
         .await

crates/ecstore/src/bucket/replication/replication_target_boundary.rs

@@ -472,6 +472,7 @@ pub(crate) fn replication_complete_multipart_options(
     actual_size: String,
     source_etag: String,
     source_mtime: Option<OffsetDateTime>,
+    source_internal: &AdvancedPutOptions,
 ) -> PutObjectOptions {
     let mut user_metadata = HashMap::new();
     insert_header_map(&mut user_metadata, SUFFIX_REPLICATION_ACTUAL_OBJECT_SIZE, actual_size);
@@ -484,6 +485,14 @@ pub(crate) fn replication_complete_multipart_options(
             // mtime must degrade to epoch so header() suppresses the header
             // instead of asserting the replication time as the object's mtime.
             source_mtime: source_mtime.unwrap_or(OffsetDateTime::UNIX_EPOCH),
+            // Carry the per-category LWW timestamps on the complete request as
+            // well: the receiver's CompleteMultipartUpload options builder
+            // parses the same headers, so the multipart transport gets the
+            // same receiver-side LWW as the single-PUT transport
+            // (rustfs/backlog#1953). Epoch values keep the headers suppressed.
+            tagging_timestamp: source_internal.tagging_timestamp,
+            retention_timestamp: source_internal.retention_timestamp,
+            legalhold_timestamp: source_internal.legalhold_timestamp,
             replication_status: ReplicationStatusType::Replica,
             replication_request: true,
             ..Default::default()
@@ -663,20 +672,39 @@ mod tests {
     #[test]
     fn replication_complete_multipart_options_sets_actual_size() {
         let source_mtime = OffsetDateTime::from_unix_timestamp(1_716_170_000).expect("valid test timestamp");
+        let source_internal = AdvancedPutOptions {
+            tagging_timestamp: OffsetDateTime::from_unix_timestamp(1_716_170_100).expect("valid test timestamp"),
+            retention_timestamp: OffsetDateTime::from_unix_timestamp(1_716_170_200).expect("valid test timestamp"),
+            legalhold_timestamp: OffsetDateTime::from_unix_timestamp(1_716_170_300).expect("valid test timestamp"),
+            ..Default::default()
+        };
         let options = replication_complete_multipart_options(
             "1024".to_string(),
             "0123456789abcdef0123456789abcdef-3".to_string(),
             Some(source_mtime),
+            &source_internal,
         );
         assert_eq!(options.internal.source_etag, "0123456789abcdef0123456789abcdef-3");
         assert_eq!(options.internal.source_mtime, source_mtime);
 
+        // The complete request must carry the same per-category LWW timestamps
+        // as the initiate request; the receiver reads them from the complete
+        // headers (rustfs/backlog#1953).
+        assert_eq!(options.internal.tagging_timestamp, source_internal.tagging_timestamp);
+        assert_eq!(options.internal.retention_timestamp, source_internal.retention_timestamp);
+        assert_eq!(options.internal.legalhold_timestamp, source_internal.legalhold_timestamp);
+
         // Absent source mtime must degrade to epoch (header suppressed), not
         // the AdvancedPutOptions default of now_utc() — that default would
         // stamp the replication time as the replica's mtime and break the
-        // multipart HEAD convergence.
-        let options_no_mtime = replication_complete_multipart_options("1024".to_string(), String::new(), None);
+        // multipart HEAD convergence. Unset category timestamps stay epoch so
+        // header() keeps suppressing them.
+        let options_no_mtime =
+            replication_complete_multipart_options("1024".to_string(), String::new(), None, &AdvancedPutOptions::default());
         assert_eq!(options_no_mtime.internal.source_mtime.unix_timestamp(), 0);
+        assert_eq!(options_no_mtime.internal.tagging_timestamp.unix_timestamp(), 0);
+        assert_eq!(options_no_mtime.internal.retention_timestamp.unix_timestamp(), 0);
+        assert_eq!(options_no_mtime.internal.legalhold_timestamp.unix_timestamp(), 0);
 
         assert_eq!(
             get_header_map(&options.user_metadata, SUFFIX_REPLICATION_ACTUAL_OBJECT_SIZE).as_deref(),

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

@@ -2318,6 +2318,57 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
             fi.set_data_moved();
         }
 
+        // Receiver-side LWW (rustfs/backlog#1953): the multipart replication
+        // transport carries the category values at CreateMultipartUpload (in
+        // the staged upload metadata) and the source category timestamps on
+        // the complete request. Read the destination version under the held
+        // object write lock and keep any category this site modified more
+        // recently. Read failures (version absent on first replication, quorum
+        // errors) keep today's overwrite semantics: failing the complete would
+        // loop through MRF, re-delivering the stale value forever.
+        if crate::set_disk::ops::object::replication_lww_applicable(opts)
+            && let Some(version_id) = fi.version_id
+        {
+            match self
+                .get_object_info(
+                    bucket,
+                    object,
+                    &ObjectOptions {
+                        version_id: Some(version_id.to_string()),
+                        no_lock: true,
+                        metadata_cache_safe: false,
+                        versioned: opts.versioned,
+                        version_suspended: opts.version_suspended,
+                        ..Default::default()
+                    },
+                )
+                .await
+            {
+                Ok(existing) => {
+                    let stored = crate::set_disk::ops::object::stored_replication_category_metadata(&existing);
+                    crate::set_disk::ops::object::merge_replication_metadata_lww(&mut fi.metadata, &stored, opts);
+                }
+                // Version absent: first replication of this version, nothing
+                // local to compare — the normal path, not a degraded one.
+                Err(err) if is_err_object_not_found(&err) || is_err_version_not_found(&err) => {}
+                Err(err) => {
+                    // Degraded path: without the stored state the inbound
+                    // metadata is applied unchanged — exactly the overwrite
+                    // LWW exists to prevent — so this must be operator-visible.
+                    warn!(
+                        component = LOG_COMPONENT_ECSTORE,
+                        subsystem = LOG_SUBSYSTEM_SET_DISK,
+                        bucket,
+                        object,
+                        version_id = %version_id,
+                        error = %err,
+                        state = "replication_lww_read_unavailable",
+                        "SetDisk multipart replication LWW read skipped; inbound metadata applied without comparison"
+                    );
+                }
+            }
+        }
+
         for meta in parts_metadatas.iter_mut() {
             if meta.has_valid_erasure_geometry() {
                 meta.size = fi.size;
@@ -7055,6 +7106,97 @@ mod tests {
                 .await
         }
 
+        /// Receiver-side LWW on the multipart replication transport
+        /// (rustfs/backlog#1953): a metadata-only replication of a multipart
+        /// source object rides CreateMultipartUpload (category values in the
+        /// upload metadata) + CompleteMultipartUpload (category timestamps in
+        /// the complete options). A stale inbound tagging timestamp must not
+        /// overwrite a newer locally-tagged destination version.
+        #[tokio::test]
+        #[serial]
+        async fn complete_multipart_upload_stale_replication_tags_keep_local() {
+            use rustfs_utils::http::headers::AMZ_OBJECT_TAGGING;
+            use rustfs_utils::http::{SUFFIX_TAGGING_TIMESTAMP, get_str};
+            use time::format_description::well_known::Rfc3339;
+
+            const T_OLD: &str = "2026-01-01T00:00:00Z";
+            const T_LOCAL: &str = "2026-02-01T00:00:00Z";
+
+            let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+            let bucket = "multipart-replication-lww-bucket";
+            let object = "object";
+            make_bucket_on_all(&disk_stores, bucket).await;
+
+            // Local destination version with newer tags.
+            let version_id = Uuid::new_v4();
+            let mut local_metadata = HashMap::new();
+            local_metadata.insert(AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string());
+            rustfs_utils::http::insert_str(&mut local_metadata, SUFFIX_TAGGING_TIMESTAMP, T_LOCAL.to_string());
+            let mut local_reader = PutObjReader::from_vec(b"local body".to_vec());
+            set_disks
+                .put_object(
+                    bucket,
+                    object,
+                    &mut local_reader,
+                    &ObjectOptions {
+                        versioned: true,
+                        version_id: Some(version_id.to_string()),
+                        user_defined: local_metadata,
+                        // Explicit-version PUTs require the bucket Object Lock snapshot.
+                        object_lock_config_snapshot: Some(Arc::new(crate::set_disk::ObjectLockConfigSnapshot::new(
+                            crate::bucket::metadata_sys::ObjectLockConfigState::ConfirmedAbsent,
+                        ))),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("local versioned put should commit");
+
+            // Inbound replication upload carrying older tags for the same version.
+            let mut inbound_metadata = HashMap::new();
+            inbound_metadata.insert(AMZ_OBJECT_TAGGING.to_string(), "site=remote".to_string());
+            rustfs_utils::http::insert_str(&mut inbound_metadata, SUFFIX_TAGGING_TIMESTAMP, T_OLD.to_string());
+            let create_opts = ObjectOptions {
+                versioned: true,
+                user_defined: inbound_metadata,
+                ..Default::default()
+            };
+            let (upload_id, parts) =
+                stage_upload_with_create_opts(&set_disks, bucket, object, &payload(0x5a), &create_opts).await;
+            rewrite_staged_upload_version_id(&set_disks, bucket, object, &upload_id, Some(version_id)).await;
+
+            let complete_opts = ObjectOptions {
+                versioned: true,
+                replication_request: true,
+                replication_tagging_timestamp: Some(OffsetDateTime::parse(T_OLD, &Rfc3339).expect("test timestamp should parse")),
+                ..Default::default()
+            };
+            set_disks
+                .clone()
+                .complete_multipart_upload(bucket, object, &upload_id, parts, &complete_opts)
+                .await
+                .expect("replication multipart completion should succeed even when a category keeps local values");
+
+            let info = set_disks
+                .get_object_info(
+                    bucket,
+                    object,
+                    &ObjectOptions {
+                        versioned: true,
+                        version_id: Some(version_id.to_string()),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("completed version should be readable");
+            assert_eq!(
+                info.user_tags.as_str(),
+                "site=local",
+                "older inbound multipart tags must not overwrite newer local tags"
+            );
+            assert_eq!(get_str(&info.user_defined, SUFFIX_TAGGING_TIMESTAMP).as_deref(), Some(T_LOCAL));
+        }
+
         #[tokio::test]
         #[serial]
         async fn complete_multipart_upload_assigns_completion_version_id() {

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

@@ -1881,6 +1881,110 @@ fn delete_file_info_with_replication_transport_metadata(fi: &FileInfo) -> FileIn
     transported
 }
 
+/// True when an authorized replication write carries at least one per-category
+/// source timestamp, i.e. receiver-side LWW has something to judge.
+pub(in crate::set_disk) fn replication_lww_applicable(opts: &ObjectOptions) -> bool {
+    opts.replication_request
+        && (opts.replication_tagging_timestamp.is_some()
+            || opts.replication_retention_timestamp.is_some()
+            || opts.replication_legalhold_timestamp.is_some())
+}
+
+/// The stored per-category state of a destination version, as compared by
+/// [`merge_replication_metadata_lww`]. `ObjectInfo::from_file_info`
+/// externalizes tags into `user_tags` (stripping the metadata key), so the
+/// tag value is folded back into map form here.
+pub(in crate::set_disk) fn stored_replication_category_metadata(existing: &ObjectInfo) -> HashMap<String, String> {
+    let mut stored = (*existing.user_defined).clone();
+    if !existing.user_tags.is_empty() {
+        stored.insert(rustfs_utils::http::headers::AMZ_OBJECT_TAGGING.to_string(), (*existing.user_tags).clone());
+    }
+    stored
+}
+
+/// Receiver-side last-writer-wins for authorized replication writes
+/// (rustfs/backlog#1953, audit A4/P1-6). Metadata-only replication reuses the
+/// whole-object transports, so in active-active topologies an inbound write
+/// carries the source's tags / retention / legal hold verbatim and would
+/// otherwise overwrite a category the destination modified more recently —
+/// both sites end up permanently diverged while reporting COMPLETED.
+///
+/// Judged per category, only when the inbound request carries that category's
+/// source timestamp (`ObjectOptions::replication_*_timestamp`):
+/// - stored timestamp newer than inbound: the local category values and
+///   timestamp are kept; the rest of the write proceeds per the inbound
+///   metadata and the object-level result stays successful (failing the write
+///   instead would loop through MRF, re-delivering the stale value forever);
+/// - otherwise the inbound category wins and its internal timestamp key is
+///   pinned to the source-authored time — the PUT path re-stamps the
+///   object-lock timestamps with the receiver's clock
+///   (`parse_object_lock_retention` / `parse_object_lock_legal_hold` insert
+///   `now()` via `eval_metadata`), which would make the replica's clock the
+///   LWW authority and wedge later convergence;
+/// - no stored timestamp (pre-P1-6 data) or no inbound timestamp: the current
+///   overwrite behavior is preserved.
+///
+/// Returns whether `inbound` was modified. Callers must hold the object write
+/// lock so the stored values compared here are the ones being replaced.
+pub(in crate::set_disk) fn merge_replication_metadata_lww(
+    inbound: &mut HashMap<String, String>,
+    existing: &HashMap<String, String>,
+    opts: &ObjectOptions,
+) -> bool {
+    use rustfs_utils::http::headers::{
+        AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER, AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER, AMZ_OBJECT_TAGGING,
+    };
+    use rustfs_utils::http::metadata_compat::{
+        SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, SUFFIX_TAGGING_TIMESTAMP, get_str,
+        remove_str,
+    };
+    use time::format_description::well_known::Rfc3339;
+
+    let categories: [(Option<OffsetDateTime>, &str, &[&str]); 3] = [
+        (opts.replication_tagging_timestamp, SUFFIX_TAGGING_TIMESTAMP, &[AMZ_OBJECT_TAGGING]),
+        (
+            opts.replication_retention_timestamp,
+            SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP,
+            &[AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER],
+        ),
+        (
+            opts.replication_legalhold_timestamp,
+            SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP,
+            &[AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER],
+        ),
+    ];
+
+    let mut changed = false;
+    for (inbound_timestamp, timestamp_suffix, value_keys) in categories {
+        let Some(inbound_timestamp) = inbound_timestamp else { continue };
+        let is_category_value_key = |key: &str| value_keys.iter().any(|value_key| key.eq_ignore_ascii_case(value_key));
+        let stored_timestamp = get_str(existing, timestamp_suffix).and_then(|value| OffsetDateTime::parse(&value, &Rfc3339).ok());
+        if stored_timestamp.is_some_and(|stored| stored > inbound_timestamp) {
+            inbound.retain(|key, _| !is_category_value_key(key));
+            remove_str(inbound, timestamp_suffix);
+            for (key, value) in existing {
+                if is_category_value_key(key) {
+                    inbound.insert(key.clone(), value.clone());
+                }
+            }
+            // Restore the winning timestamp via insert_str, not a verbatim key
+            // copy: a MinIO-written version may carry only the
+            // x-minio-internal- key, and the dual-key invariant requires every
+            // write to produce both keys.
+            if let Some(stored_value) = get_str(existing, timestamp_suffix) {
+                rustfs_utils::http::insert_str(inbound, timestamp_suffix, stored_value);
+            }
+            changed = true;
+        } else if let Ok(source_authored) = inbound_timestamp.format(&Rfc3339)
+            && get_str(inbound, timestamp_suffix).as_deref() != Some(source_authored.as_str())
+        {
+            rustfs_utils::http::insert_str(inbound, timestamp_suffix, source_authored);
+            changed = true;
+        }
+    }
+    changed
+}
+
 impl SetDisks {
     pub(in crate::set_disk) async fn persist_old_data_cleanup_receipts(
         &self,
@@ -2124,26 +2228,13 @@ 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 {
@@ -2575,6 +2666,22 @@ impl SetDisks {
                         if check_object_lock_for_deletion_with_state(object_lock_config.state(), &existing, false)?.is_some() {
                             return Err(StorageError::PrefixAccessDenied(bucket.to_string(), object.to_string()));
                         }
+                        // Receiver-side LWW (rustfs/backlog#1953): reuse this
+                        // commit-lock read of the destination version so a
+                        // category (tags / retention / legal hold) modified
+                        // more recently on this site is kept instead of being
+                        // overwritten by the inbound replication metadata.
+                        if replication_lww_applicable(opts) {
+                            let stored = stored_replication_category_metadata(&existing);
+                            let mut merged = parts_metadatas[response_metadata_slot].metadata.clone();
+                            if merge_replication_metadata_lww(&mut merged, &stored, opts) {
+                                for (pfi, disk) in parts_metadatas.iter_mut().zip(shuffle_disks.iter()) {
+                                    if disk.is_some() {
+                                        pfi.metadata = merged.clone();
+                                    }
+                                }
+                            }
+                        }
                     }
                     Err(err) if is_err_object_not_found(&err) || is_err_version_not_found(&err) => {}
                     Err(err) => return Err(err),
@@ -8079,6 +8186,357 @@ mod replication_quota_safety_tests {
     }
 }
 
+#[cfg(test)]
+mod replication_lww_tests {
+    //! Receiver-side LWW for authorized replication writes (rustfs/backlog#1953,
+    //! audit A4/P1-6): an inbound replication PUT whose per-category timestamp
+    //! (tags / retention / legal hold) is older than the destination version's
+    //! stored timestamp must keep the local category values instead of
+    //! overwriting them; categories are judged independently and the write
+    //! itself still succeeds.
+
+    use super::hermetic_set_disks_support::hermetic_set_disks_isolated as hermetic_set_disks;
+    use super::*;
+    use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
+    use rustfs_utils::http::headers::{
+        AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER, AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER, AMZ_OBJECT_TAGGING,
+    };
+    use rustfs_utils::http::{
+        SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, SUFFIX_TAGGING_TIMESTAMP, get_str,
+        insert_str,
+    };
+    use time::format_description::well_known::Rfc3339;
+
+    const T_OLD: &str = "2026-01-01T00:00:00Z";
+    const T_LOCAL: &str = "2026-02-01T00:00:00Z";
+    const T_NEW: &str = "2026-03-01T00:00:00Z";
+
+    fn parse_ts(value: &str) -> OffsetDateTime {
+        OffsetDateTime::parse(value, &Rfc3339).expect("test timestamp should parse")
+    }
+
+    async fn make_bucket(disks: &[DiskStore], bucket: &str) {
+        for disk in disks {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+    }
+
+    async fn put_version(set_disks: &Arc<SetDisks>, bucket: &str, object: &str, version_id: &str, opts: &ObjectOptions) {
+        let mut reader = PutObjReader::from_vec(b"lww-body".to_vec());
+        set_disks
+            .put_object(bucket, object, &mut reader, opts)
+            .await
+            .expect("versioned put should commit");
+        assert_eq!(opts.version_id.as_deref(), Some(version_id));
+    }
+
+    fn versioned_opts(version_id: &str, user_defined: HashMap<String, String>) -> ObjectOptions {
+        ObjectOptions {
+            versioned: true,
+            version_id: Some(version_id.to_string()),
+            user_defined,
+            // Explicit-version PUTs require the bucket Object Lock snapshot.
+            object_lock_config_snapshot: Some(Arc::new(ObjectLockConfigSnapshot::new(
+                crate::bucket::metadata_sys::ObjectLockConfigState::ConfirmedAbsent,
+            ))),
+            ..Default::default()
+        }
+    }
+
+    /// Local state: version `version_id` with tags "site=local" stamped `T_LOCAL`.
+    async fn seed_local_tagged_version(set_disks: &Arc<SetDisks>, bucket: &str, object: &str, version_id: &str) {
+        let mut user_defined = HashMap::new();
+        user_defined.insert(AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string());
+        insert_str(&mut user_defined, SUFFIX_TAGGING_TIMESTAMP, T_LOCAL.to_string());
+        put_version(set_disks, bucket, object, version_id, &versioned_opts(version_id, user_defined)).await;
+    }
+
+    fn inbound_tagging_opts(version_id: &str, tags: &str, timestamp: &str) -> ObjectOptions {
+        let mut user_defined = HashMap::new();
+        user_defined.insert(AMZ_OBJECT_TAGGING.to_string(), tags.to_string());
+        insert_str(&mut user_defined, SUFFIX_TAGGING_TIMESTAMP, timestamp.to_string());
+        ObjectOptions {
+            replication_request: true,
+            replication_tagging_timestamp: Some(parse_ts(timestamp)),
+            ..versioned_opts(version_id, user_defined)
+        }
+    }
+
+    async fn version_info(set_disks: &Arc<SetDisks>, bucket: &str, object: &str, version_id: &str) -> ObjectInfo {
+        set_disks
+            .get_object_info(bucket, object, &versioned_opts(version_id, HashMap::new()))
+            .await
+            .expect("version should be readable")
+    }
+
+    #[tokio::test]
+    async fn inbound_stale_tagging_keeps_newer_local_tags() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-tagging-stale";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        seed_local_tagged_version(&set_disks, bucket, object, &version_id).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_tagging_opts(&version_id, "site=remote", T_OLD),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_tags.as_str(),
+            "site=local",
+            "older inbound tags must not overwrite newer local tags"
+        );
+        assert_eq!(
+            get_str(&info.user_defined, SUFFIX_TAGGING_TIMESTAMP).as_deref(),
+            Some(T_LOCAL),
+            "the winning local tagging timestamp must be preserved"
+        );
+    }
+
+    #[tokio::test]
+    async fn inbound_newer_tagging_overwrites_local_tags() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-tagging-newer";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        seed_local_tagged_version(&set_disks, bucket, object, &version_id).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_tagging_opts(&version_id, "site=remote", T_NEW),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_tags.as_str(),
+            "site=remote",
+            "newer inbound tags must overwrite older local tags"
+        );
+        assert_eq!(get_str(&info.user_defined, SUFFIX_TAGGING_TIMESTAMP).as_deref(), Some(T_NEW));
+    }
+
+    #[tokio::test]
+    async fn inbound_wins_when_local_has_no_tagging_timestamp() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-tagging-no-local-ts";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        // Pre-P1-6 data: local tags without a stored tagging timestamp.
+        let mut user_defined = HashMap::new();
+        user_defined.insert(AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, user_defined)).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_tagging_opts(&version_id, "site=remote", T_OLD),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_tags.as_str(),
+            "site=remote",
+            "without a local timestamp the inbound category must win (pre-LWW data compatibility)"
+        );
+    }
+
+    #[tokio::test]
+    async fn categories_are_judged_independently() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-category-independent";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+
+        // Local: newer tags (T_LOCAL), older *cleared* retention (T_OLD) —
+        // timestamp key only, the shape a replicated retention clear stores.
+        // (An active local retention would already block the overwrite at the
+        // WORM gate; the LWW-reachable retention states are cleared/expired.)
+        let mut local = HashMap::new();
+        local.insert(AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string());
+        insert_str(&mut local, SUFFIX_TAGGING_TIMESTAMP, T_LOCAL.to_string());
+        insert_str(&mut local, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, T_OLD.to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, local)).await;
+
+        // Inbound: older tags (T_OLD), newer retention (T_NEW).
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_TAGGING.to_string(), "site=remote".to_string());
+        insert_str(&mut inbound, SUFFIX_TAGGING_TIMESTAMP, T_OLD.to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_MODE_LOWER.to_string(), "COMPLIANCE".to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER.to_string(), "2028-01-01T00:00:00Z".to_string());
+        insert_str(&mut inbound, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, T_NEW.to_string());
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_tagging_timestamp: Some(parse_ts(T_OLD)),
+            replication_retention_timestamp: Some(parse_ts(T_NEW)),
+            ..versioned_opts(&version_id, inbound)
+        };
+        put_version(&set_disks, bucket, object, &version_id, &opts).await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(info.user_tags.as_str(), "site=local", "the stale tagging category must keep local values");
+        assert_eq!(
+            info.user_defined.get(AMZ_OBJECT_LOCK_MODE_LOWER).map(String::as_str),
+            Some("COMPLIANCE"),
+            "the newer retention category must be applied in the same write"
+        );
+        assert_eq!(get_str(&info.user_defined, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP).as_deref(), Some(T_NEW));
+    }
+
+    #[tokio::test]
+    async fn inbound_stale_legal_hold_keeps_local_value() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-legalhold-stale";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+
+        // Local: legal hold released (OFF) at T_LOCAL. (A local hold that is
+        // still ON already blocks the overwrite at the WORM gate; the
+        // LWW-reachable divergence is a stale inbound ON resurrecting a hold
+        // that was released more recently on this site.)
+        let mut local = HashMap::new();
+        local.insert(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER.to_string(), "OFF".to_string());
+        insert_str(&mut local, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, T_LOCAL.to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, local)).await;
+
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER.to_string(), "ON".to_string());
+        insert_str(&mut inbound, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, T_OLD.to_string());
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_legalhold_timestamp: Some(parse_ts(T_OLD)),
+            ..versioned_opts(&version_id, inbound)
+        };
+        put_version(&set_disks, bucket, object, &version_id, &opts).await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_defined.get(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER).map(String::as_str),
+            Some("OFF"),
+            "a stale inbound legal hold must not resurrect a hold released more recently"
+        );
+        assert_eq!(
+            get_str(&info.user_defined, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP).as_deref(),
+            Some(T_LOCAL)
+        );
+    }
+
+    /// Dual-key invariant under LWW: a MinIO-written destination version may
+    /// carry only the x-minio-internal timestamp key; when the local category
+    /// wins, the restored map must still hold BOTH compatibility keys.
+    #[test]
+    fn local_win_restores_both_internal_timestamp_keys_for_minio_only_metadata() {
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_TAGGING.to_string(), "site=remote".to_string());
+        insert_str(&mut inbound, SUFFIX_TAGGING_TIMESTAMP, T_OLD.to_string());
+        let existing = HashMap::from([
+            (AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string()),
+            ("X-Minio-Internal-Tagging-Timestamp".to_string(), T_LOCAL.to_string()),
+        ]);
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_tagging_timestamp: Some(parse_ts(T_OLD)),
+            ..Default::default()
+        };
+
+        assert!(merge_replication_metadata_lww(&mut inbound, &existing, &opts));
+        assert_eq!(inbound.get(AMZ_OBJECT_TAGGING).map(String::as_str), Some("site=local"));
+        assert_eq!(
+            inbound.get("x-rustfs-internal-tagging-timestamp").map(String::as_str),
+            Some(T_LOCAL),
+            "the RustFS twin key must be materialized even when the source version only had the MinIO key"
+        );
+        assert_eq!(inbound.get("x-minio-internal-tagging-timestamp").map(String::as_str), Some(T_LOCAL));
+    }
+
+    /// When the inbound category wins, the stored timestamp must be the
+    /// source-authored one: the PUT path's eval_metadata stamps the
+    /// object-lock timestamps with the receiver's clock
+    /// (`parse_object_lock_retention`), which would otherwise make this
+    /// replica's clock the LWW authority and wedge later convergence.
+    #[tokio::test]
+    async fn inbound_win_pins_stored_timestamp_to_source_authored_value() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-retention-ts-pinned";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+
+        // Local cleared retention at T_OLD.
+        let mut local = HashMap::new();
+        insert_str(&mut local, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, T_OLD.to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, local)).await;
+
+        // Inbound newer retention: the source authored T_LOCAL, but the PUT
+        // path's eval_metadata stomped the metadata key with receiver-now
+        // (simulated by T_NEW here).
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_LOCK_MODE_LOWER.to_string(), "GOVERNANCE".to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER.to_string(), "2028-01-01T00:00:00Z".to_string());
+        insert_str(&mut inbound, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, T_NEW.to_string());
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_retention_timestamp: Some(parse_ts(T_LOCAL)),
+            ..versioned_opts(&version_id, inbound)
+        };
+        put_version(&set_disks, bucket, object, &version_id, &opts).await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            get_str(&info.user_defined, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP).as_deref(),
+            Some(T_LOCAL),
+            "the stored category timestamp must be the source-authored time, not the receiver's clock"
+        );
+        assert_eq!(info.user_defined.get(AMZ_OBJECT_LOCK_MODE_LOWER).map(String::as_str), Some("GOVERNANCE"));
+    }
+
+    #[tokio::test]
+    async fn newer_local_tag_deletion_survives_stale_inbound_tags() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-tagging-deleted";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        // Local DeleteObjectTagging state: no tags, but a newer tagging timestamp.
+        let mut local = HashMap::new();
+        insert_str(&mut local, SUFFIX_TAGGING_TIMESTAMP, T_LOCAL.to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, local)).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_tagging_opts(&version_id, "site=remote", T_OLD),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert!(
+            info.user_tags.is_empty(),
+            "a newer local tag deletion must not be resurrected by older inbound tags"
+        );
+        assert_eq!(get_str(&info.user_defined, SUFFIX_TAGGING_TIMESTAMP).as_deref(), Some(T_LOCAL));
+    }
+}
+
 #[cfg(test)]
 mod inline_put_commit_path_tests {
     use super::hermetic_set_disks_support::hermetic_set_disks_isolated as hermetic_set_disks;

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 mut accounting = vec![None; objects.len()];
+        let 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(|left, right| right.idx.cmp(&left.idx));
+        latest_candidates.sort_by_key(|candidate| std::cmp::Reverse(candidate.idx));
 
         let Some(winner) = latest_candidates.first() else {
             return Err(Error::ErasureReadQuorum);

deny.toml

@@ -43,9 +43,6 @@ 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

@@ -0,0 +1,149 @@
+# 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"] }
-mimalloc = { workspace = true }
+rustfs-mimalloc = { workspace = true }
 
 [target.'cfg(target_os = "linux")'.dependencies]
 libsystemd.workspace = true
 
 [target.'cfg(not(target_os = "windows"))'.dependencies]
-libmimalloc-sys.workspace = true
+rustfs-mimalloc-sys.workspace = true
 
 [dev-dependencies]
 uuid = { workspace = true, features = ["v4", "v5", "fast-rng", "macro-diagnostics"] }

rustfs/src/allocator_reclaim.rs

@@ -369,14 +369,8 @@ 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> {
-    // 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);
-    }
+    rustfs_mimalloc::MiMalloc::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 { mimalloc::MiMalloc.alloc(layout) }
+        unsafe { rustfs_mimalloc::MiMalloc.alloc(layout) }
     }
 
     unsafe fn alloc_zeroed(&self, layout: Layout) -> *mut u8 {
         // SAFETY: the caller upholds GlobalAlloc's contract for layout.
-        unsafe { mimalloc::MiMalloc.alloc_zeroed(layout) }
+        unsafe { rustfs_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 { mimalloc::MiMalloc.dealloc(ptr, layout) }
+        unsafe { rustfs_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 { mimalloc::MiMalloc.realloc(ptr, layout, new_size) }
+        unsafe { rustfs_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: mimalloc::MiMalloc = mimalloc::MiMalloc;
+static GLOBAL: rustfs_mimalloc::MiMalloc = rustfs_mimalloc::MiMalloc;
 
 fn main() {
     let _hotpath_guard = hotpath::HotpathGuardBuilder::new("main").build();
@@ -71,8 +71,9 @@ 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 { libmimalloc_sys::mi_is_in_heap_region(allocation.as_ptr().cast()) });
+        assert!(unsafe { heap.contains(allocation.as_ptr()) });
     }
 
     #[test]
@@ -85,12 +86,13 @@ 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 { libmimalloc_sys::mi_is_in_heap_region(ptr.cast()) });
+        assert!(unsafe { heap.contains(ptr) });
         assert!(unsafe { std::slice::from_raw_parts(ptr, 32).iter().all(|byte| *byte == 0) });
 
         // SAFETY: `ptr` was allocated by `allocator` with `layout`; on failure
@@ -102,7 +104,7 @@ mod tests {
             panic!("mimalloc realloc failed in allocator smoke test");
         }
 
-        assert!(unsafe { libmimalloc_sys::mi_is_in_heap_region(grown_ptr.cast()) });
+        assert!(unsafe { heap.contains(grown_ptr) });
         // 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

@@ -17,10 +17,7 @@ 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;
@@ -231,7 +228,18 @@ fn read_cgroup_memory_snapshot() -> Option<CgroupMemorySnapshot> {
     read_cgroup_v2().or_else(read_cgroup_v1)
 }
 
-#[cfg(any(test, not(target_os = "windows")))]
+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,
+    })
+}
+
 fn numeric_json_value(value: &Value) -> Option<u64> {
     match value {
         Value::Number(number) => number
@@ -242,7 +250,6 @@ 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
@@ -254,7 +261,6 @@ 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) => {
@@ -271,12 +277,10 @@ 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()
@@ -285,7 +289,6 @@ 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"];
@@ -312,33 +315,6 @@ 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)
 }
@@ -566,6 +542,13 @@ 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);

rustfs/src/storage/options.rs

@@ -554,10 +554,11 @@ fn apply_replication_timestamps_from_headers(headers: &HeaderMap<HeaderValue>, o
 
     // Persist into the internal metadata keys so a later outbound replication
     // pass (replication_target_boundary) reads the source's modification
-    // times instead of falling back to mod_time.
-    // TODO(P1-6): receiver-side LWW is still missing — when the stored
-    // per-category timestamp is newer than the inbound one, the existing
-    // tags/retention/legal-hold should win instead of being overwritten.
+    // times instead of falling back to mod_time. Receiver-side LWW happens at
+    // the set layer under the object write lock
+    // (ecstore set_disk::ops::object::merge_replication_metadata_lww,
+    // rustfs/backlog#1953): a category whose stored timestamp is newer than
+    // the inbound one keeps the local values.
     for (timestamp, suffix) in [
         (opts.replication_tagging_timestamp, SUFFIX_TAGGING_TIMESTAMP),
         (opts.replication_retention_timestamp, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP),